WORLD BANK LATIN AMERICAN AND CARIBBEAN STUDIES Raising the Bar for Productive Cities in Latin America and the Caribbean María Marta Ferreyra and Mark Roberts, editors Raising the Bar for Productive Cities in Latin America and the Caribbean RAISING THE BAR for Productive Cities in Latin America and the Caribbean María Marta Ferreyr a and Mark Roberts, Editors © 2018 International Bank for Reconstruction and Development / The World Bank 1818 H Street NW, Washington DC 20433 Telephone: 202-473-1000; Internet: www.worldbank.org Some rights reserved 1 2 3 4 21 20 19 18 This work is a product of the staff of The World Bank with external contributions. The findings, interpretations, and conclusions expressed in this work do not necessarily reflect the views of The World ­ Bank, its Board of Executive Directors, or the governments they represent. The World Bank does not guarantee the accuracy of the data included in this work. The boundaries, colors, denominations, and other information shown on any map in this work do not imply any judgment on the part of The World Bank concerning the legal status of any territory or the endorsement or acceptance of such boundaries. Nothing herein shall constitute or be considered to be a limitation upon or waiver of the privileges and immunities of The World Bank, all of which are specifically reserved. Rights and Permissions This work is available under the Creative Commons Attribution 3.0 IGO license (CC BY 3.0 IGO) http://creativecommons.org/licenses/by/3.0/igo. Under the Creative Commons Attribution license, you are free to copy, distribute, transmit, and adapt this work, including for commercial purposes, under the following conditions: Attribution—Please cite the work as follows: Ferreyra, María Marta, and Mark Roberts, editors. 2018. Raising the Bar for Productive Cities in Latin America and the Caribbean. Washington, DC: World Bank. doi:10.1596/978-1-4648-1258-3. License: Creative Commons Attribution CC BY 3.0 IGO. Translations—If you create a translation of this work, please add the following disclaimer along with the attribution: This translation was not created by The World Bank and should not be considered an official World Bank translation. The World Bank shall not be liable for any content or error in this translation. Adaptations—If you create an adaptation of this work, please add the following disclaimer along with the attribution: This is an adaptation of an original work by The World Bank. Views and opinions expressed in the adaptation are the sole responsibility of the author or authors of the adaptation and are not endorsed by The World Bank. Third-party content—The World Bank does not necessarily own each component of the content con- tained within the work. The World Bank therefore does not warrant that the use of any third-party- owned individual component or part contained in the work will not infringe on the rights of those third parties. The risk of claims resulting from such infringement rests solely with you. If you wish to reuse a component of the work, it is your responsibility to determine whether permission is needed for that reuse and to obtain permission from the copyright owner. Examples of components can include, but are not limited to, tables, figures, or images. All queries on rights and licenses should be addressed to World Bank Publications, The World Bank Group, 1818 H Street NW, Washington, DC 20433, USA; e-mail: pubrights@worldbank.org. ISBN (print): 978-1-4648-1258-3 ISBN (electronic): 978-1-4648-1270-5 DOI: 10.1596/978-1-4648-1258-3 Cover design: Bill Pragluski, Critical Stages, LLC. Cover art: Sunrise over The Andes, Santiago, Chile. © Getty. Used with permission; further permission required for reuse. Contents Foreword . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .xvii About the Authors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xix Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxi Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 The Productivity of LAC Cities Is Slightly above Average but below the Global Frontier . . . 2 What These Findings Might Mean for Policy . . . . . . . . . . . . . . . . . . . . . . . . 18 Annex OA: Productivity Measures Used in the Book to Assess LAC Cities . . . . . . . . . .19 Annex OB: The Need for Policy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .20 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .21 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .23 Part I.  Urbanization and Productivity in Latin America and the Caribbean . . . . . . 25 Chapter 1.  Urbanization, Economic Development, and Structural Transformation . . . . . 27 Paula Restrepo Cadavid and Grace Cineas Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .27 The Origins of Cities in Latin America and the Caribbean . . . . . . . . . . . . . . . . . .28 Urbanization in the LAC Region and the Rest of the World: Discrepancies between Consistent and Official Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 Urbanization, Economic Development, and Structural Transformation: How Does the LAC Region’s Performance Stack Up? . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .43 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .45 Chapter 2.   The Many Dimensions of Urbanization and the Productivity of Cities in Latin America and the Caribbean . . . . . . . . . . . . . . . . . . . . . . . . 49 Mark ­­Roberts Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .49 v vi   C o n t e n t s Defining a Global Data Set of Urban Areas . . . . . . . . . . . . . . . . . . . . . . . . .51 Urban Areas in the LAC Region Are More Densely Populated Than Those Elsewhere . . . .52 A Significant Share of Latin America and the Caribbean’s Urban Population Lives in Large MCAs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .59 A Third of LAC Countries Analyzed Suffer from Potentially Excessive Primacy . . . . . . .62 Implications for National Productivity: Density and MCAs Matter, but Urban Primacy Does Not . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .65 International Benchmarking of LAC Urban Areas’ Productivity: Better Than Average, but Lagging the Global Frontier . . . . . . . . . . . . . . . . . . . . . . 67 Productivity is Highly Dispersed across LAC Urban Areas . . . . . . . . . . . . . . . . . .72 ­­Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 Annex 2A: List of Comparator Countries for Each LAC Country . . . . . . . . . . . . . .77 Annex 2B: Statistical Tests of Differences in Population, Area, and Population Density between LAC Countries and Their Comparators . . . . . . . . . . . . . . . . . . . . . .78 Annex 2C: List of Multicity Agglomerations in the LAC Region . . . . . . . . . . . . . . .79 Annex 2D: Cross-Country Regression of Log(GDP per Capita) on Different Dimensions of Urbanization: Alternative Definition for a Multicity Agglomeration . . . . 81 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .81 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .84 Part II.  The Determinants of City Productivity in Latin America and the Caribbean . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 Chapter 3.  The Empirical Determinants of City Productivity . . . . . . . . . . . . . . . . .89 Mark ­Roberts Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .89 Cities Are More Productive Than Rural Areas . . . . . . . . . . . . . . . . . . . . . . . .91 Large Subnational Variations in Productivity, Explained Partly by Sorting . . . . . . . . . .96 Explaining Underlying Variations in Productivity: The Three Theories . . . . . . . . . . . .98 What about Firms? Evidence from World Bank Enterprise Surveys . . . . . . . . . . . . 104 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 Annex 3A: Results of Regressions on the Determinants of Underlying Productivity Variations Based on the Single-Stage Approach . . . . . . . . . . . . . . . 110 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 Chapter 4.  Transport Infrastructure and Agglomeration in Cities . . . . . . . . . . . . . . 117 Soumahoro Harris Selod and Souleymane ­ Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 Transport, Agglomeration, and Productivity: A Brief Review . . . . . . . . . . . . . . . 118 Transport in Latin America and the Caribbean: History, Current State, and Challenges . . 119 Roads and Agglomeration Economies: Evidence from Mexico . . . . . . . . . . . . . . . 129 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136 C o n t e n t s   vii Chapter 5.  Human Capital in Cities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141 María Marta Ferreyra Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141 Some Stylized Facts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 Returns to Aggregate Human Capital . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 Attracting Skilled Individuals to Cities . . . . . . . . . . . . . . . . . . . . . . . . . . . 156 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160 Annex 5A: Areas Used in the Stylized Facts . . . . . . . . . . . . . . . . . . . . . . . . 161 Annex 5B: Percentage of Employment in Services, by Educational Attainment . . . . . . . 162 Annex 5C: Probability of Working in the Service Sector for Skilled and Unskilled Workers, by Area Size . . . . . . . . . . . . . . . . . . . . . . . . . . . 163 Annex 5D: Percentage of Urban Population Born Abroad . . . . . . . . . . . . . . . . . .163 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165 Chapter 6.  Urban Form, Institutional Fragmentation, and Metropolitan Coordination . . .167 Nancy Lozano Gracia and Paula Restrepo Cadavid Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167 Urban Form and Productivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168 Institutional Fragmentation, Metropolitan Coordination, and Productivity . . . . . . . . 180 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187 Annex 6A: Seventy-Three Cities in Institutional Fragmentation and Coordination Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188 Annex 6B: Urban Form Indicators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189 Annex 6C: Correlation Matrix between Urban Form Variables . . . . . . . . . . . . . . 190 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192 Boxes O.1 Form, Skill, and Access. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 1.1 ­ Precolonial Densities, Location Fundamentals, and the Persistence of Subnational Population Densities. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 1.2 ­ Location Fundamentals and the Distribution of Economic Activity in Latin America and the Caribbean versus the Rest of the World. . . . . . . 31 1 ­ .3 The Agglomeration Index and the Cluster Algorithm. . . . . . . . . . . . . . . . . . . . . . . 35 ­ 1.4 Comparing the Population of Urban Areas: Cluster Algorithm versus Official Data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 2.1 ­­ Comparing Apples with Apples: Selecting Comparators for LAC Countries. . . . . . 55 2.2 ­­ Congestion Forces in LAC Urban Areas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 2.3 ­­ VIIRS Nighttime Lights Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 2.4 ­­ Cities and Aggregate Growth: United States and Brazil. . . . . . . . . . . . . . . . . . . . . . 75 ­ 3.1 SEDLAC: A Treasure Trove of Harmonized Data . . . . . . . . . . . . . . . . . . . . . . . . . 92 3.2 ­ Which Groups of Workers Benefit More?. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 3.3 The Determinants of Manufacturing Firm Productivity across Colombian Municipalities. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 viii   C o n t e n t s ­ .1 4 History of Road Development in Mexico. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 4.2 ­ Measuring Industrial Concentration: The Ellison and Glaeser Index. . . . . . . . . . . 130 ­ 4.3 Measuring Municipality Specialization: The Krugman Specialization ­ Index. . . . . 131 ­4.4 Market Access. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132 5.1 An Equilibrium Model of Household Sorting for Brazil . . . . . . . . . . . . . . . . . . . . 156 6.1 Outlining Urban Extents Using Nighttime Lights . . . . . . . . . . . . . . . . . . . . . . . . . 173 6.2 Constructing Institutional Fragmentation and Metropolitan Coordination Variables. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182 Figures O.1 LAC Countries Exhibit Average Productivity Given Their Urbanization Levels. . . . . 3 O.2 Productivity of LAC Cities Is above Average but Lags the Global Frontier. . . . . . . . 3 O.3 A High Percentage of LAC Cities Have Population Densities above the Global Median . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 O.4 LAC Cities Are Dense Because Their Areas Are Small. . . . . . . . . . . . . . . . . . . . . . . . 5 O.5 Productivity Varies Widely across Cities and Countries in Latin America and the Caribbean. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 O.6 Within-Country Productivity Dispersions Are High in LAC Countries. . . . . . . . . . . 7 O.7 More Populous LAC Cities Have Higher Shares of Skilled Labor. . . . . . . . . . . . . . . 8 O.8 Rail Is Not Prevalent in Latin America. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 O.9 Paved Road Density Has Been Stagnant in Latin America and the Caribbean. . . . . . 9 O.10 Unconditional and Conditional Effects of Density on Productivity Provide Insights into the Mechanisms for Agglomeration Effects . . . . . . . . . . . . . . 11 O.11 The Effects of Form, Skill, and Access on Productivity. . . . . . . . . . . . . . . . . . . . . . 12 O.12 In Most Countries, a City’s Population Density Does Not Have a Positive Significant Effect on Its Productivity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 O.13 Across Countries, Returns to Skill Are U-Shaped in Average City Skill. . . . . . . . . . 15 O.14 Individual Returns to Skill Fall and Then Rise with Own Education. . . . . . . . . . . . 16 O.15 Market Access Is Associated with City Productivity in Some Countries . . . . . . . . . 17 O.16 Countries with Better Road Coverage Have More Efficient Systems of Cities. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 1.1 ­ Strong Persistence in Subnational Population Densities in the LAC Region . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 1.2 ­ Urban Shares for Latin America and the Caribbean and Other World Regions, 1960–2015. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 ­ 1.3 Urban Shares for LAC Subregions, 1960–2015. . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 1.4 ­ Annual Growth of Urban Population, Worldwide and by Region, 1960–2005. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 1.5 ­ Urban Shares: Official versus Consistent Measures of Urbanization. . . . . . . . . . . . 35 B1.4.1 Comparison of Cluster Algorithm and WUP City Population Values, ­ by Region. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 ­ 1.4.2 Between and within Variation of the Relocation Fraction per Region. . . . . . . . . . . 37 B 1.6 ­ Relationship between Economic Development and the Urban Share on Official Measures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 1.7 ­ Change in the Structural Composition of the Economy, 1960–2009. . . . . . . . . . . . 42 2.1 ­­ Percentage of Urban Areas with Population Densities Higher Than the Global Median, by Region. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 2.2 ­­ Distribution of Area Size and Population across Urban Areas, Selected Regions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 C o n t e n t s   ix ­­ 2.3 Percentage of Urban Areas with Population Densities Higher Than the Global Median, by LAC Country . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 B2.2.1 Relationship between Traffic Congestion and Population Density, ­­ LAC Cities versus Non-LAC Cities. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 ­­ B2.2.2 Air Pollution in Cities in Latin America and the Caribbean and Other Regions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 ­­ 2.4 Percentage of Urban Population Living in Multicity Agglomerations, by Region. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 2.5 ­­ Multicity Agglomerations, by LAC Country. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62 2.6 ­­ Cross-Country Relationship between Urban Share and Share of National Population Living in Multicity Agglomerations . . . . . . . . . . . . . . . . . . . . 63 2.7 ­­ Urban Primacy, by Region. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 ­­ 2.8 Urban Primacy, LAC Countries and Comparators. . . . . . . . . . . . . . . . . . . . . . . . . . 64 2.9 ­­ Relationship between Log(Nighttime Lights) and Log(Population), All Urban Areas Globally . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 2.10 Mean Urban Area Productivity in LAC Countries Benchmarked ­­ against International Comparators. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 ­­ 2.11 Distribution of Productivity across Urban Areas, Selected Regions. . . . . . . . . . . . . 73 ­­ 2.12 Productivity Dispersion (Measured by the Coefficient of Variation) across Urban Areas in LAC Countries Benchmarked against High-Income International Comparators. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 ­­ 2.13 Productivity Dispersion across Urban Areas in a Country Is Negatively Correlated with National Road Density, 112 Countries. . . . . . . . . . . . . 75 3.1 ­ Ratio of Nominal Mean Urban to Nominal Mean Rural Wage in 15 LAC Countries, 2000–14. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 3.2 ­ Urban and Worker Premiums in 15 LAC Countries . . . . . . . . . . . . . . . . . . . . . . . . 95 3 ­ .3 Subnational Variations in Underlying Productivity in 16 LAC Countries. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98 3.4 ­ Correlation between Underlying Productivity and Population Density, Average Number of Years of Schooling, and Market Access . . . . . . . . . . . 99 3.5 ­ Cross-Country Heterogeneity in Estimated Elasticities of Underlying Productivity with Respect to Population Density, Average Number of Years of Schooling, and Market Access. . . . . . . . . . . . . . . . . . . . . . . . 102 3.6 ­ Different Dimensions of a City’s Business Environment. . . . . . . . . . . . . . . . . . . . 105 3.7 ­ Security Costs Incurred by Firms in Cities, Latin America and the Caribbean and Other Regions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 ­ 4.1 Length of Railroad Track in Service and Urban Share in Latin America and the Caribbean, 1900–2007. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 4.2 ­­ Export Density and Railroad Density in Selected Latin American Countries, 1900–30. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 4.3 Length of Roads and Urban Population Share in Latin America and the Caribbean, 1950–2000. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 4.4 ­ Modal Split of Surface Freight in Latin America and the Caribbean, 2012. . . . . . 124 ­ 4.5 Investment in Transport Infrastructure in Latin America and the Caribbean, 2000–13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124 ­ 4.6 Change in Transport Investment as a Share of GDP, 2008–15. . . . . . . . . . . . . . . . 125 ­ 4.7 Evolution of Paved Road Density, Selected Regions, 1961–2000. . . . . . . . . . . . . . 126 4.8 ­ Average and Per Capita Road Length in a 100-Kilometer Radius around Cities with at Least 1 Million Inhabitants. . . . . . . . . . . . . . . . . . . . . . . . . 127 x   C o n t e n t s 4.9 ­ Average Road Length in a 100-Kilometer Radius around Cities with at Least 1 Million Inhabitants. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 4.10 ­ Ad Valorem Freight and Real Tariffs for Intraregional Exports and Exports to the United States, 2005 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128 5.1 Distribution of Human Capital at the Area Level, circa 2014 . . . . . . . . . . . . . . . 144 5.2 Population and Human Capital in the Largest Areas, circa 2014 . . . . . . . . . . . . . 145 5.3 Percentage of the Adult Population Living in Urban Areas, circa 2014. . . . . . . . . 145 5.4 Percentage of Skilled Population, by Area Size, circa 2014. . . . . . . . . . . . . . . . . . 146 5.5 Average Gini Coefficient, by Area Size, circa 2014 . . . . . . . . . . . . . . . . . . . . . . . . 147 5.6 Percentage of Skilled Migrants, by Area Size, circa 2014. . . . . . . . . . . . . . . . . . . . 148 5.7 Percentage of Skilled Individuals Who Are Migrants, by Area Size, circa 2014. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 5.8 Percentage of Employment in Services, by Area Size, circa 2014. . . . . . . . . . . . . . 149 5.9 Percentage of Service Workers, by Sector in Large Areas and by Skill Level, circa 2014. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150 5.10 Returns to Aggregate Human Capital, 2000–14. . . . . . . . . . . . . . . . . . . . . . . . . . 153 5.11 Returns to Aggregate Human Capital, by Individual’s Own Education, 2000–14. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154 6.1 Urban Form in Latin America and the Caribbean Shows Great Variability. . . . . . 176 6.2 Change in Urban Form Indicators, 1996–2000. . . . . . . . . . . . . . . . . . . . . . . . . . . 177 6.3 What Levels of Fragmentation Are Needed to Reap the Benefits?. . . . . . . . . . . . . 186 Maps O.1 Multicity Agglomerations in Latin America and the Caribbean Span Multiple Municipalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 2.1 Examples of Multicity Agglomerations in Latin America and the Caribbean That Span Multiple Municipalities. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 3.1 Subnational Variations in Nominal Wages in South America . . . . . . . . . . . . . . . . . 97 3.2 Subnational Variations in Nominal Wages in Central America and the Caribbean . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 4.1 The Evolution of the Road Network in Mexico, 1985–2016 . . . . . . . . . . . . . . . . 122 4.2 ­ Spatial Distributions of Formal Establishments and Manufacturing Firms in Mexico, Overlaid on the Road Network, 2014. . . . . . . . . . . . . . . . . . . . 130 4.3 ­ Output Locality Specialization, Overlaid on the Road Network, 2014. . . . . . . . . 132 4.4 Changes in Market Access in Mexico, 1986–2014 . . . . . . . . . . . . . . . . . . . . . . . . 133 B6.1.1 Examples of Urban Extents over the DMSP-OLS Radiance-Calibrated 2010 Composite . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174 B6.2.1 Examples of Metropolitan Areas. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182 Tables ­B1.2.1 R-Squared Results for Relationship between Log(Radiance-Calibrated Nighttime Lights) and Base, Agriculture, and Trade Fundamentals . . . . . . . . . . . . 31 ­ 1.1 Regression Results for Relationship between Log(GDP per Capita) and the Official Urban Share. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 ­ 1.2 Regression Results for Relationship between Log(GDP per Capita) and the Urban Share, Using Consistent Measures. . . . . . . . . . . . . . . . . . . . . . . . . . 40 2.1 ­­ Summary Statistics for Global Sample of Urban Areas . . . . . . . . . . . . . . . . . . . . . . 52 ­­ 2.2 Number of Multicity Agglomerations, by Region. . . . . . . . . . . . . . . . . . . . . . . . . . 61 C o n t e n t s   xi ­­ 2.3 Cross-Country Regression of Log(GDP per Capita) on Different Dimensions of Urbanization. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 B2.3.1 Regression of Log(GDP) on VIIRS Nighttime Lights Data, 2015 . . . . . . . . . . . . . . 68 2.4 ­­ The 15 Urban Areas in the LAC Region with the Highest Estimated Economic Activity, as Measured by Nighttime Lights Data, 2015 . . . . . . . . . . . . . 69 2.5 Relationship between Log(Nighttime Lights) and Log(Population), All Urban Areas Globally . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 ­ 3.1 Differences in Characteristics between Urban and Rural Workers in 15 LAC Countries. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 3.2 ­ Results of Regressions on the Determinants of Underlying Productivity Variations across Subnational Areas. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 B3.2.1 Heterogeneous Effects of Determinants on Underlying Productivity ­ across Worker Subgroups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 ­ 3.3 The Effects of a City’s Business Environment on Firm Productivity. . . . . . . . . . . . 106 ­ 3A.1 Results of Regressions on the Determinants of Underlying Productivity Variations Based on the Single-Stage Approach. . . . . . . . . . . . . . . . . . . . . . . . . . . 110 ­ 4.1 Density of All Roads (Paved and Nonpaved) in Regions of the World, 1961–2000 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 4.2 ­ Modal Share of Surface Freight, by Region, 2015 . . . . . . . . . . . . . . . . . . . . . . . . 124 4.3 ­ Transport Infrastructure Investments, by Sector, 2008–15 . . . . . . . . . . . . . . . . . . 125 ­ 4.4 LAC Cities Are among the Top 100 Congested Places in the World . . . . . . . . . . . 128 ­ 4.5 The Effects of Market Access on Employment . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 ­ 4.6 The Effects of Market Access on Local Specialization. . . . . . . . . . . . . . . . . . . . . . 135 ­ 4.7 The Effects of Market Access on Nighttime Lights . . . . . . . . . . . . . . . . . . . . . . . . 135 5.1 Returns to Aggregate Human Capital, 2000–14. . . . . . . . . . . . . . . . . . . . . . . . . . 152 5.2 Local Labor Demand, Brazil, 2010. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158 5.3 The Effects of Raising Labor Demand for Higher Education Graduates in Feira de Santana, Brazil, 2010. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159 6.1 Examples of Urban Areas with High, Medium, and Low Values of the Indexes That Describe Urban Form. . . . . . . . . . . . . . . . . . . . . . . . . 170 6.2 Descriptive Statistics of Urban Form in LAC Cities. . . . . . . . . . . . . . . . . . . . . . . . 175 6.3 Regression Results for Urban Form and City Productivity with Outliers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179 6.4 Institutional Fragmentation and Metropolitan Coordination . . . . . . . . . . . . . . . . 181 6.5 Institutional Fragmentation and Metropolitan Coordination, LAC Region versus OECD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183 6.6 Top 15 Fragmented Metropolitan Areas, LAC Region versus OECD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183 6.7 Regression of a City Productivity Premium (ln) on Institutional Fragmentation and Metropolitan Coordination Variables. . . . . . . . . . . . . . . . . . . 185 Foreword S ince the early days of civilization, cannot be productive unless their cities human beings have come together in are also productive. Further, the region is cities. Cities (from the Latin civitas) and in need of greater productivity, as the civilization (from the Latin civilis) are inex- high growth rates of the first decade of the tricably linked. Throughout human history, new millennium have given way to low cities have been centers of civilization, cul- and uneven growth rates. Reviving ture, and human achievement. They have growth is thus at the top of policy makers’ also been powerful hubs of economic activ- agenda. ity, entrepreneurship, and innovation. As Despite the importance of LAC cities’ firms and workers gather in cities, opportu- productivity, surprisingly little is known nities emerge for employment and business. about it. The novel research conducted for As cities within a country become better this report tells us that while the productiv- connected, further opportunities emerge for ity of LAC cities is on par with the world’s production and trade among cities. average, it lags the world’s frontier, which is These opportunities, however, do not where LAC policy makers would wish to always come to full realization. Such is the be. Not only does the region lag other coun- case when cities are overwhelmed by conges- tries, but some cities lag others within the tion, or when people, goods, and services do same country. While human capital makes not flow freely across cities. Cities can only key contributions to cities’ productivity, realize their potential and their contribution other mechanisms, such as access to a larger to national productivity when policy makers market, seem rather muted. Closing these implement an enabling environment through productivity gaps calls for an enabling envi- a combination of policies at the local, state, ronment of adequate infrastructure, urban and national level. planning, public services, and metropoli- These issues are critical for the Latin tan governance. It also requires further America and the Caribbean (LAC) region investments in human capital and an econ- today. Since almost three-quarters of LAC’s omy that facilitates the flow of people, population lives in cities, LAC countries goods, and services across cities. xiii xiv   F o r e w o r d In its quest for greater productivity, LAC stimulate the type of insights and food for must seek to develop the full potential of its thought that leads to sound and progressive cities. We hope that the research presented in policy making. this report will enhance our knowledge and Jorge Familiar, Vice President Carlos Végh, Chief Economist Latin America and the Caribbean Region The World Bank Group Preface T his book investigates the contribu- great engine of growth—cities—cannot be tion of cities to productivit y in left untapped. Latin America and the Caribbean The book has two parts. Part I documents (LAC), a topic about which surprisingly overall urbanization patterns across the LAC little is known. The rapid economic growth region and their relationship to productivity that prevailed in the region during the first outcomes at the national and subnational lev- decade of the new millennium has, since els, compared with the rest of the world. Part the collapse of global commodity prices, II conducts a deeper, more rigorous analysis given way to low, uneven growth in recent of the underlying determinants of productiv- years. In this context, boosting productiv- ity differences across LAC cities focusing on ity is critical to reviving economic growth three key factors: city form, skills, and access in the region. And the potential of that to markets through transportation networks. xv Acknowledgments T his book was prepared by a team led by p e er re v ie wer s: U we D eich m a n n , G i l le s María Marta Ferreyra and Mark Roberts. Du ranton, William Maloney, and Forhad The core team also consisted of Nancy Shilpi. While the authoring team is very grate- Lozano Gracia, Paula Restrepo Cadavid, and ful for the guidance received, these reviewers Harris Selod, and received excellent research are not responsible for any remaining errors, assistance from Angelica Sanchez Diaz, Grace om issions, or i nterpretations. Add itional Cineas, Jane Park, and Souleymane Soumahoro. i nsig hts f rom Judy B a ker, M atias Busso, The work was conducted under the general Walker Hanlon, Maria Flavia Harari, Adam guidance of Augusto de la Torre, former Chief Storeygard, Daniel Sullivan, Matthew Turner, E conomist for the Latin A merica and the Daniel Xu, and other participants of a work- Caribbean (LAC) region of the World Bank, and shop on May 11 and 12, 2017, are gratefully Carlos Végh, current LAC Chief Economist of the acknowledged. World Bank, with substantial inputs from Daniel In preparing the book, the team benefitted from Lederman, former Deputy LAC Chief Economist, discussions with Peter Ellis, Somik Lall, and and Ming Zhang, LAC Practice Manager for the Horacio Cristian Terraza, while Anna Wellenstein Social, Urban, Rural and Resilience Global and Catalina Marulanda played an important role Practice of the World Bank. in early discussions relating to scoping out the Preparation of the book was informed by a series book. The team is also grateful for the support pro- of background papers. Authors of these background vided by senior management of the World Bank’s papers who have not already been named include Social, Urban, Rural, and Resilience Global Jorge Balat, Paulo Bastos, Brian Blankespoor, Practice, including not only Anna Welleinstein but Theophile Bougna, Maria Camila Casas, Chandan also Senior Director Ede Jorge Ijjasz-Vasquez and Deuskar, Juan Carlos Duque, Lin Fan, Rafael Director Sameh Naguib Wahba Tadros. Garduno, Jorge Patino, Luis Quintero, Daniel Bruce Ross-Larson was the principal editor, and Reyes, Benjamin Stewart, Christopher Timmins, Joe Caponio, Mike Crumplar, and John Wagley and Lixin C. Xu. Empirical work for the book was were the copyeditors. Additional editing work was underpinned by an extensive geospatial database performed by Joseph Coohill. Patricia Katayama for LAC that was developed in collaboration with a (acquisitions editor), Rumit Pancholi (production research team at the University of Southampton’s editor), and Deborah Appel-Barker (print coordina- GeoData Center led by Julia Branson and Chris tor), of the World Bank’s Formal Publishing Hill. Further support on data was generously pro- Program, were responsible for managing the design, vided by the World Bank’s Geospatial Operations typesetting, and printing of the book. Last, but not Support Team, as well as by Siobhan Murray. least, the authors thank Ruth Delgado, Ruth Eunice The team was fortunate to receive excellent Flores, Jacqueline Larrabure, and Michelle Chen advice and guidance from four distinguished for unfailing administrative support. xvii About the Authors María Marta Ferreyra is a senior economist in the University of Cambridge and a Fellow of the Office of the Chief Economist for Latin Murray Edwards College in Cambridge, A mer ic a a nd t he C a r ibb e a n of t he United Kingdom. Mark has published widely World Bank. Her research specializes in the in leading peer-reviewed journals on spatial economics of education, with special emphasis economic development and is a former coedi- on the effects of large-scale reforms. She has tor of the journal Spatial Economic Analysis. conducted research on charter schools, private He is the coauthor of the World Bank’s South school vouchers, public school accountability, Asia region’s flagship book, Leveraging and school finance reform for primary and Urbanization in South Asia, and the Latin secondary education in the United States and America and the Caribbean region’s flagship on higher education in Latin America and the report, Raising the Bar for Productive Cities Caribbean. Her research has been published in Latin America and the Caribbean. He is in journals such as the American Economic also currently leading the World Bank’s Review, the Journal of Public Economics, and Indonesia flagship report on urbanization. the American Economic Journal: Economic Mark holds a PhD in land economy and an Policy. She was the lead author of At a MA in economics from the University of Crossroads: Higher Education in Latin Cambridge, United Kingdom, as well as an America and the Caribbean. Before joining MSc in economics from Warwick University, the World Bank, she served as a faculty mem- United Kingdom. ber at the Tepper School of Business at Carnegie Mellon University. She holds a PhD Grace Cineas is a consultant in the Social, in economics from the Universit y of Urban, Rural, and Resilience Global Practice Wisconsin–Madison. of the World Bank, where her work has focused primarily on urban development and Mark Roberts is a senior urban economist in resilience in Europe and Central Asia. She the Social, Urban, Rural, and Resilience has also contributed to work in Latin Global Practice of the World Bank, where his America and Sub-Saharan Africa. Grace work primarily focuses on the East Asia and holds a master of international economics Pacific, and Latin America and the Caribbean and international relations from The Johns regions. Before joining the World Bank, Hopkins University School of Advanced Mark was a lecturer in spatial economics at International Studies. xix xx   About the Authors Harris Selod is a senior economist in the where she worked on models for measuring Development Research Group of the World capitalization of the value of local amenities Bank. His research focuses on the role of into housing prices. Her areas of work transport, property rights, and land markets include urban and regional economics, spa- on economic development. His papers have tial economic analysis, and spatial economet- been published in academic journals such as ric applications. the American Economic Journal: Economic Policy, the Economic Journal, the Journal of Paula Restrepo Cadavid is a senior urban Development Economics, the Journal of economist in the Social, Urban, Rural, and Public Economics, and the Journal of Urban Resilience Global Practice of the World Bank, Economics. He currently coordinates the where her work has primarily focused on the World Bank’s research program on Transport Eastern Europe and Central Asia and Latin Policies for Sustainable Growth and Poverty America and the Caribbean regions. At the Reduction and is the co-organizer of the World Bank, her work focuses on areas annual World Bank/George Washington related to urban and territorial development, Universit y Urbanization and Pover t y municipal finance, and housing. She is the Reduction Research Conference. Before join- lead author of the World Bank’s Cities in ing the World Bank, he served as an associate Eastern Europe and Central Asia: A Story of professor at the Paris School of Economics Urban Growth and Decline. She has also led and as a researcher at the French National or contributed to investment projects in Institute for Agricultural Research (Institut Albania, Azerbaijan, Colombia, Georgia, national de la recherche agronomique, or Honduras, Kyrgyz Republic, Moldova, Peru, INRA). He holds a PhD in economics from Tajikistan, and Uzbekistan. She holds a mas- Sorbonne University, an MSc in statistics ter’s degree in environmental and develop- f rom t he Pa r i s G radu ate S cho ol of ment economics from Ecole Polytechnique Economics, Statistics, and Finance (École and a PhD in economics from Ecole de Mines nationale de la statistique et de l’administra- de Paris, where she worked on assessing the tion économique, or ENSAE), and an MBA welfare effects of slum-upgrading policies. from ESCP Europe (École supérieure de com- Her areas of research span from urban and merce de Paris). regional economics to infrastructure financ- ing and environmental economics. Nancy Lozano Gracia is a senior economist in the Social, Urban, Rural, and Resilience Souleymane Soumahoro is an economist and Global Practice of the World Bank, where she consultant in the Development Research has worked extensively on designing and Group of the World Bank. His research using diagnostic tools to improve the under- focuses on the political economy of develop- standing of the challenges of rapid urbaniza- ment, access to infrastructure, and public ser- tion and city development and to help identify vice delivery. His research has been published priorities for action. As part of these efforts, in peer-reviewed academic journals such as she has led work using innovative data collec- Economic Development and Cultural tion methods such as satellite imagery, new Change and Applied Economic Letters. survey designs, and big data approaches, to Before joining the World Bank, he worked as build a better understanding of within-city a postdoctoral fellow at the Center for Global challenges. As a core member of the Global Development, a leading development think Solutions Group on Territorial Development, tank in Washington, DC. Also, he holds a her work has recently focused on using spa- PhD in economics from the University of tial analysis to identify priorities for action in Oklahoma and a master’s degree in interna- lagging regions. She holds a doctorate in tional economics from the University of applied economics from University of Illinois, Auvergne Clermont-Ferrand 1 in France. Abbreviations AAA American Automobile Association AI Agglomeration Index BE business environment CEDLAS Center for Distributive, Labor and Social Studies DENUE  Directorio Estadístico Nacional de Unidades Económicas (National Statistical Directory of Economic Units) DMSP-OLS Defense Meteorological Satellite Program–Operational Linescan System EAP East Asia and Pacific ECA Europe and Central Asia GDP gross domestic product GGDC Groningen Growth and Development Center GHSL Global Human Settlement Layer GIS Geographic Information System HCE human capital externality IBGE  Instituto Brasileiro de Geografia e Estatística (Brazilian Institute of Geography and Statistics) INEGI  Instituto Nacional de Estadísticas y Geografía (National Institute of Statistics and Geography) IV instrumental variable IPUMS Integrated Public Use Microdata Series LAC Latin America and the Caribbean MCA multicity agglomeration MENA Middle East and North Africa NTL nighttime lights OECD Organisation for Economic Co-operation and Development OLS ordinary least square PM particulate matter PPP purchasing power parity R&D research and development SA South Asia SEDLAC Socio-Economic Database for Latin America and the Caribbean xxi xxii   A b b r e v i a t i o n s SSA Sub-Saharan Africa TFP total factor productivity VIIRS Visible Infrared Imaging Radiometer Suite WAP working-age population WBES World Bank Enterprise Survey WDI World Development Indicators WHO World Health Organization WUP World Urbanization Prospects Overview I n modern economies, cities can be formi- productivity frontier, where LAC policy mak- dable engines of productivity and eco- ers want their cities to be. What accounts for nomic growth. By bringing people and the failure of LAC cities to reach the global firms together in close geographic proximity, frontier? First, although LAC cities benefit cities facilitate production, innovation, from strong positive agglomeration effects and trade. Historically, urbanization has associated with skills, they may lack the accompanied the productive transforma- “enabling environment” needed to fully lever- tion of economies—with the decline in age the wider benefits of agglomeration and low-productivity agricultural employment mitigate congestion costs. Thus, urban infra- and the rise of high-productivity manufactur- structure management and urban planning ing and services. Falling transportation may not be adequate to curb the congestion of costs—by facilitating trade by cities, both roads, basic urban services, and land and with one another and with rural areas—have housing markets associated with the high accelerated this process, further stimulating urban density in most LAC countries. Included both urbanization and development. in this is inadequate coordination across local Today, almost three-quarters of the popu- governments within fragmented metropolitan lation of Latin America and the Caribbean areas. Second, a lack of integration among (LAC)—or 433 million people—live in the cities within countries is associated with region’s 7,197 cities.1 Some are mega-­ cities, underinvestment in national transport net- such as São Paulo and Mexico City, each works, opening wide productivity gaps across boasting populations of about 20 million. 2 cities and undermining the aggregate contri- Others are small settlements in the gray area bution of cities to national productivity. between urban and rural. Some cities date The evidence also shows that human capi- back to precolonial times (Bogotá, Cuzco, tal is a bedrock source of productivity across Mexico City). Others were established by cities throughout the LAC region, but that Spanish and Portuguese conquistadores the skilled—who form a smaller share of the during colonial times (Asunción, Buenos workforce than in, say, the United States— Aires, São Paulo) or by the newly indepen- are also heavily concentrated in the largest dent cou nt ries in postcolonial times cities. This makes it a priority to close the (La Plata). Still others were established a few region’s shortfall of skills relative to the most decades ago (Brasilia, Puerto Ordaz). developed countries, and to ensure that both The productivity of LAC cities is on small and large cities can be attractive places par with the world average but lags the world for the skilled to live and work. Investing in 1 2   RAISING THE BAR infrastructure, transport, and human capital planning and management, and on policies in cities of all sizes, as well as developing effi- that influence the quality of the local busi- cient local governance institutions, will thus ness environment, including protection prove crucial to raising the bar for productiv- from crime. ity in the region’s cities—and ultimately in Because no city exists in isolation, its pro- the region’s countries as well. ductivity is related to that of other cities in The proximity of people and firms in cities the country. Any one city is part of a coun- can give rise to many benefits. The concen- try’s system of cities, where cities are con- tration of individuals, particularly the skilled, nected by transport and other networks. So can facilitate the exchange of ideas and the policies that affect the productivity of one sharing of knowledge, boosting innovation city will also have repercussions on other cit- and productivity. Firms located in a city enjoy ies. The easier the flow of goods, resources, the privilege of having access to a large local and people across cities, the greater the con- market, which may also be well connected to tribution of cities to national productivity. the markets of nearby cities. Access to a That is why maximizing the contribution of larger market can encourage a wider variety cities to a country’s productivity and growth of products and services, many of which are requires taking the whole system of cities inputs into the production of other firms. into account. The proximity of people and firms in cities also creates thick labor markets, which give firms access to larger and more diverse pools The Productivity of LAC Cities Is of workers, and workers access to a greater Slightly above Average but number and variety of potential employers, below the Global Frontier leading to better job matches. The proximity To compare LAC cities with those in the rest of people and firms also spreads the cost of of the world, an important complication is large-scale investments in transport and that countries differ in defining “urban.” infrastructure for basic services over many Overcoming this complication is critical for individuals. Cities thus generate productivity-​ cross-country comparisons. One crucial con- enhancing agglomeration effects. tribution of this book is to apply an algo- But cities also give rise to negative conges- rithm (the “cluster algorithm”) that allows tion effects. As the number of people and for a globally consistent definition of urban firms within a city grows, so does the areas. Rather than define urban areas on the demand for land, housing, and labor, raising basis of their official administrative boundar- the costs of living and conducting business. ies, which often fail to accurately delineate Without additional investments in infra- the actual extent of a city, this algorithm structure, or improvements in urban policy identifies cities as spatially contiguous dense and management, the city becomes more clusters of population, whose total popula- congested, roads and other public infrastruc- tion surpasses a well-defined threshold. 3 ture more crowded, and crime and grime With this definition, we calculate a variety of more prevalent. country-​ level urbanization metrics, the most All cities are subject to the opposing forces basic of which is a country’s urban share (the of agglomeration and congestion, but their percent of its population that lives in cities). net outcomes depend, at least in part, on a We also use the individual cities as units of city’s enabling environment for spurring ben- observation in their own right, which allows eficial agglomeration effects and mitigating us to benchmark the productivity of LAC cit- negative congestion effects. The enabling ies against those in the rest of the world. environment depends, in turn, on the extent The story of productivity in LAC cities in and quality of infrastructure provision within relation to the rest of the world has good news cities (such as roads, bridges, and utility and bad. Historically, the joint processes of and communications networks), on urban economic development and urbanization have O v er v ie w   3 given rise to a positive association between a FIGURE O.1  LAC Countries Exhibit Average Productivity Given country’s aggregate productivity (measured Their Urbanization Levels by gross domestic product [GDP] per capita) and the share of its population that lives in 12 GDP per capita PPP 2012 (log) urban areas (its urban share).4 Across coun- tries in the world, a 1 percentage point increase in the urban share is associated with 10 a 3.8 percent increase in GDP per capita. Without implying causality, this relationship, shown by the solid line in figure O.1, estab- 8 lishes a country’s expected productivity given its urban share. A country falling below the solid line underperforms, given its urban 6 20 40 60 80 100 share, and a country above it overperforms. As it turns out, LAC countries (indicated Urban share (%) by the orange, green, and red markers) on Rest of the world North America and Western Europe average perform as expected given their urban Caribbean Central America, without Mexico shares. This is true for the region, and for the South America Linear fit for the entire sample of 169 countries South America, Central America, and and Mexico Caribbean subregions. Nonetheless, LAC countries underperform relative to countries Source: Calculations based on WDI data and cities defined using the cluster algorithm of Dijkstra and Poelman (2014), as applied to Landscan 2012 gridded population data. in North America and Western Europe (blue Note: GDP per capita is measured in constant international dollars at 2012 PPP exchange rates. It is markers). The good news, then, is that LAC expressed in natural logs on the vertical axis. GDP = gross domestic product; LAC = Latin America and the Caribbean; PPP = purchasing power parity; WDI = World Development Indicators. countries perform as predicted given their urban shares; the bad news is that they are below the global productivity frontier.5 A similar conclusion emerges from using FIGURE O.2  Productivity of LAC Cities Is above Average but Lags city-level productivity measures. Figure O.2 the Global Frontier depicts the global relationship between a city’s level of economic activity—as proxied 15 by the intensity of the light it emits at night— and its level of population. As it turns out, Nighttime lights (log) 10 LAC cities overall perform above the global average—in other words, they are more 5 ­ p roductive than expected given their ­populations.6 This result is driven by South 0 American and Mexican cities (red markers); cities in the rest of the region tend to per- form around the global average. Yet, once −5 again, LAC cities fail to reach the global 8 10 12 14 16 18 frontier, given by the outer envelope of Population (log) points in the figure, representing mainly Rest of the world North America and Western Europe North American and Western European Caribbean Central America, without Mexico ­ cities (blue markers). South America Quadratic fit for the entire sample To summarize, LAC cities perform at or and Mexico of 63,089 cities above the global average, but they perform Source: Calculations based on nighttime lights data from the 2015 VIIRS annual composite below the global frontier. To provide insights product (https://ngdc.noaa.gov​/­eog/viirs​/­download_dnb_composites.html). Cities are defined into why LAC cities lag the global frontier, we using the cluster algorithm of Dijkstra and Poelman (2014), as applied to Landscan 2012 gridded population data. examine the distinctive features of LAC cities Note: Nighttime lights on the vertical axis is the sum of nighttime lights luminosity values within a relative to others in the world, and the role of given city. LAC = Latin America and the Caribbean; VIIRS = Visible Infrared Imaging Radiometer Suite. 4   RAISING THE BAR three critical, proximate determinants of city area relative to the population size. In rela- productivity—form, skill, and access. tion to the rest of the world, LAC cities are dense not because their populations are large but because their geographic areas are small, Distinctive Features of LAC Cities particularly compared with cities in ECA and Several features distinguish LAC cities from NAC (figure O.4). Given its potential to gen- others in the world—and can help in under- erate strong positive agglomeration effects, standing why they perform below the global high density can be a blessing. However, in frontier. the absence of an adequate enabling environ- ment to help manage congestion costs and Feature 1. LAC cities are relatively dense.  In foster these agglomeration effects, this bless- Bogotá, Colombia, almost 13,500 people ing can become a curse—which may help occupy each square kilometer of land, while explain why LAC cities lag the global pro- in Lima, Peru, nearly 9,000 people populate ductivity frontier.8 each square kilometer. More generally, with an average density of almost 2,400 people per Feature 2. Multicity agglomerations are square kilometer across all 7,197 of its cities, unusually prevalent.  The administrative the LAC region exhibits urban densities that definition of a city can differ quite radically are well above the world average of just over from the “true” urban extent of a city using 1,500. Although density is highest in South the cluster algorithm. Indeed, a city as defined American cities, followed by Central in this book can span multiple “cities” as American and Caribbean cities, it is high by defined from an administrative or jurisdictional international standards in all three subregions. viewpoint. We refer to such areas as multicity Further, 80 percent of LAC cities have a agglomerations (MCAs).9 By definition, population density above the global median, MCAs span multiple local government well above the percentage in regions such as jurisdictions. Take Mexico City and Santo Europe and Central Asia (ECA) and North Domingo: Mexico City’s urban area America (NAC) (figure O.3).7 encompasses 34 municipalities, and Santo Two factors can contribute to a city’s high Domingo’s covers 19 (map O.1).10 density. The first is a large population relative Of the world’s 295 MCAs, 54 are in the to the geographic area. The second is a small LAC region second only to East Asia FIGURE O.3  A High Percentage of LAC Cities Have Population Densities above the Global Median 100 80 60 Percent 40 20 0 South Caribbean Central MENA SSA SA EAP ECA North America America America Source: Calculations based on an analysis of cities defined using the cluster algorithm of Dijkstra and Poelman (2014), as applied to Landscan 2012 gridded population data. Note: A city is classified as dense if its mean population density exceeds the global median of 1,180 people per square kilometer. Central America includes Mexico. EAP = East Asia and the Pacific; ECA = Europe and Central Asia; LAC = Latin America and the Caribbean; MENA = Middle East and North Africa; SA = South Asia; SSA = Sub-Saharan Africa. O v er v ie w   5 FIGURE O.4  LAC Cities Are Dense Because Their Areas Are Small a. Size of city area b. City population 0.8 0.10 0.08 0.6 0.06 Density Density 0.4 0.04 0.2 0.02 0 0 0 20 40 60 80 100 8 10 12 14 16 18 Area (km2) Population (log) Caribbean Central America South America Europe and Central Asia North America Source: Calculations based on analysis of cities defined using the cluster algorithm of Dijkstra and Poelman (2014), as applied to Landscan 2012 gridded population data. Note: Panels a and b show, for different regions, the distribution of area (in square kilometers) and population (log), respectively, of cities using an Epanechnikov kernel. For expositional purposes, the distributions of area are trimmed at 100 square kilometers. Central America includes Mexico. LAC = Latin America and the Caribbean. MAP O.1  Multicity Agglomerations in Latin America and the Caribbean Span Multiple Municipalities a. Mexico City, Mexico b. Santo Domingo, Dominican Republic Caribbean Sea Mexico City cluster Other urban clusters Santo Domingo cluster Mexico City Other urban clusters 0 10 20 km 0 10 20 km Municipalities intersecting Mexico City cluster Municipalities intersecting Santo Domingo cluster Source: Calculations using Geographic Information Systems software and administrative boundary data from the LAC Geospatial Database (Branson et al. 2016). Note: In the maps, the red areas correspond to cities as defined using the cluster algorithm of Dijkstra and Poelman (2014), as applied to Landscan 2012 gridded population data. The yellow lines represent subnational administrative boundaries at the municipality level that belong to a city as officially defined. The dark blue lines represent the boundaries of municipalities that intersect with the city but that do not belong to the officially defined city. In the case of Mexico City, the officially defined city comprises several municipalities. LAC = Latin America and the Caribbean. 6   RAISING THE BAR FIGURE O.5  Productivity Varies Widely across Cities and Compare the productivity dispersion in Countries in Latin America and the Caribbean LAC countries with that of high-income countries. For each LAC country, the within-​ 0.8 country productivity dispersion is relatively high (figure O.6). So LAC systems of cities 0.6 are not well integrated and thus not fully productive.11 Density 0.4 Feature 4. Within countries, the skilled are unusually concentrated in large cities.  Skilled 0.2 people tend to sort into larger cities (figure O.7).12,13 This sorting takes place in the United States as well but is stronger in the 0 LAC region. In the United States, a 10 percent −10 −5 0 5 10 increase in a city’s population is associated with Productivity a 1.2 percent increase in the share of the city’s Caribbean Central America, without Mexico population that is skilled (Behrens and Robert- South America and Mexico Sub-Saharan Africa Nicoud 2015), but with a 2.9 percent increase in North America the LAC region.14 This indicates that, compared with the United States, skilled people are Source: Calculations based on nighttime lights data from the 2015 VIIRS annual composite product relatively more concentrated in a few large cities. (https://ngdc.noaa.gov/eog​/­viirs/download_dnb_composites.html). This concentration of skills may help to explain, Note: The figure shows density plots of the residuals from a regression at the city level where the dependent variable is the sum of nighttime lights (in logs) and the independent variable is the at least partly, the high productivity dispersion population (in logs). These residuals measure city-level productivity; cities have been identified across cities in LAC countries. by applying the cluster algorithm of Dijkstra and Poelman (2014) to Landscan 2012 gridded population data. VIIRS = Visible Infrared Imaging Radiometer Suite. Feature 5. Inequality in LAC cities is unusually high.  Not only are large LAC and the Pacific (EAP). About 40 percent of the cities more skilled but they are also more LAC region’s urban population resides in unequal. On average in the LAC region, MCAs, compared with a third of the world’s a 10 percent increase in city population is urban population. Thus, LAC cities may be associated with a 0.29 percent increase in particularly vulnerable to the shortcomings of income inequality, measured by the Gini MCAs, which arise when their local jurisdic- coefficient.15 The corresponding increase in tions fail to coordinate governance and the the United States is lower (0.12 percent), provision of public goods and services. indicating a stronger tendency toward income inequality in large LAC cities. Feature 3. Within countries, productivity Of the greater income inequality in the LAC varies widely across cities.  City labor region’s larger cities, 43 percent is due to skills. productivity, measured by the (log) intensity of Put differently, relative to smaller cities, large nighttime lights net of population, varies cities are more unequal because they are more widely across LAC cities (figure O.5). The skilled and have a greater share of high-­ earning LAC region’s most productive cities rival many individuals.16 A similar, yet weaker, finding North American cities, but the least productive holds for the United States, where skills explain are close to the top-performing African cities. only 25 percent of the association between city Within countries, productivity is widely population and income inequality.17 dispersed across cities. In a well-integrated That city population, skills, and inequal- system of cities, the flow of goods, people, ity are more strongly associated in the LAC and resources across cities closes productiv- region than in the United States may reflect ity gaps among cities and maximizes the the LAC region’s scarcity of skills. For cont ribution of t he system of cities. example, the share of individuals with some O v er v ie w   7 higher education in the average LAC coun- well-developed and well-used national rail try (18 percent) is roughly one-third of that networks. But, in Latin America, rail captures in the United States (59 percent), and only 22 percent of surface freight, close to the returns to higher education are concomi- 19 percent captured by rail in Africa (figure O.8). tantly higher (104 percent for the average A low share of freight shipped by rail LAC country, and more than twice that in would not be problematic if national road the United States).18,19 The stronger associa- networks were of high quality. But LAC tions in the LAC region may also reflect a roads are not, and the paved road density has more unequal distribution of amenities and been rather stagnant in the LAC region for public services—which serve to attract four decades (figure O.9). Although paved skilled people—across cities in the LAC road density in South Asia was only slightly region than in the United States. above that of the LAC region in the early 1960s, it is now much higher. Although EAP Feature 6. National transport networks remain and the Middle East and North Africa quite undeveloped.  In NAC, Asia Pacific, (MENA) regions started at virtually the same and Europe, about 40 percent or more of level as did the LAC region they too are surface freight is shipped by rail, reflecting ahead of the LAC region today. FIGURE O.6  Within-Country Productivity Dispersions Are High in LAC Countries a. Caribbean 200 Coe cient of variation (%) 150 100 50 0 Dominican Cuba Jamaica Haiti Trinidad and Republic Tobago b. Central America 800 150 700 Coe cient of variation (%) Coe cient of variation (%) 600 100 500 400 300 50 200 100 0 0 Honduras El Salvador Nicaragua Guatemala Panama Mexico Costa Rica LAC country Comparator 1 Comparator 2 Comparator 3 (continued) 8   RAISING THE BAR FIGURE O.6  Within-country Productivity Dispersions Are High in LAC Countries (continued) c. South America 600 Coe cient of variation (%) 500 400 300 200 100 0 Peru Guyana Colombia Bolivia Ecuador 100 Coe cient of variation (%) 80 60 40 20 0 Venezuela, RB Brazil Chile Paraguay Uruguay Argentina LAC country Comparator 1 Comparator 2 Comparator 3 Source: Calculations based on nighttime lights data from the 2015 VIIRS annual composite product (https://ngdc.noaa.gov/eog/viirs/download_dnb_composites.html). Note: City productivity is measured using the residuals from a regression at the city level where the dependent variable is the sum of nighttime lights (in logs) and the independent variable is the population (in logs). Productivity dispersion across a country’s cities is measured by the coefficient of variation (in percent). Comparators for each LAC country are restricted to high-income countries, but with no restrictions on the regions their comparators are drawn from. The methodology for selecting comparators is described in detail in box 2.1 in chapter 2. A full list of comparators for each LAC country is in annex 2A in chapter 2. LAC = Latin America and the Caribbean; VIIRS = Visible Infrared Imaging Radiometer Suite. FIGURE O.7  More Populous LAC Cities Have Higher Shares of Skilled Labor 45 40 35 30 Percent 25 20 15 10 5 0 a a ala as a ile ru r y lic ca y zil r ico ia do do ua ua bi gu tin liv Pe ur a ub Ri Ch m ex m Br lva ua g ug ra n Bo nd sta ep e ra lo ge M ca Ec at Ur Sa Ho Pa Co nR Co Ar Ni Gu El ica in m Do Small cities Medium cities Large cities Source: Calculations using SEDLAC for countries other than Brazil and IPUMS International for Brazil. Note: The figure shows the average percentage of adult population (age 25–64 years) with some higher education, by area size. The area size classification follows country-specific population thresholds, as explained in annex 5A of chapter 5. IPUMS = Integrated Public Use Microdata Series; LAC = Latin America and the Caribbean; SEDLAC = Socio-Economic Database for Latin America and the Caribbean. O v er v ie w   9 The underdevelopment of national trans- features of LAC cities help explain why these port networks in the LAC region reflects the cities lag the global productivity frontier. To lack of integration among cities in LAC further expand our understanding of this countries. Together with the unusually high question, we turn to empirical evidence on the concentration of skills in large cities, this proximate determinants of city productivity in underdevelopment may contribute to the the LAC region. wide dispersion of productivity across cities in LAC countries.20 FIGURE O.8  Rail Is Not Prevalent in Latin America To summarize, LAC cities have distinctive features. They are relatively dense, perhaps 50 exacerbating congestion forces given prevail- 45 ing infrastructure and policies. MCAs, with 40 their potential coordination and governance 35 problems, are unusually prevalent. Skilled 30 Percent human capital is highly concentrated in large 25 cities. Productivity dispersions across cities in 20 LAC countries are very high, indicating that 15 their systems of cities are not efficient. Such 10 dispersions may be linked to the underdevelop- 5 ment of national transport networks. The spa- 0 Europe Asia North Latin Africa Middle tial concentration of skills also means that two Pacific America America East issues—a deficit of skills in small cities, and inequality in large cities—may be particularly Source: International Transport Forum 2017. acute in the LAC region. These distinctive Note: The figure shows the percentage of goods transported by rail (as opposed to road). FIGURE O.9  Paved Road Density Has Been Stagnant in Latin America and the Caribbean 120 Paved road density (kilometers per 100 km2) 100 80 60 40 20 0 1961 1965 1970 1975 1980 1985 1990 1995 2000 Latin America and the Caribbean East Asia and Pacific United States Middle East and North Africa Western Europe South Asia Sub-Saharan Africa Source: Calculations based on data from the World Development Indicators. 10   RAISING THE BAR firms can contribute to knowledge spillovers The Role of Form, Skill, and Access in and allow all workers to learn from the most the Productivity of LAC Cities skilled ones. A city may be more productive than others As a result of these mechanisms, agglomer- because of sorting, as skilled and talented ation effects are associated with city form, people gravitate toward it. But its greater pro- skill, and access (box O.1). Form refers to the ductivity can also be due to agglomeration size and configuration of a city, skill to how effects, which operate through various mech- skilled individuals contribute to the productiv- anisms enabled by the proximity of firms and ity of others, and access to a city’s connected- individuals. For example, the greater number ness to other cities in the country through the of firms and workers that characterize cities transportation network. Figure O.10 shows can generate better matches between them. that, if we look at form alone (measured by The greater number of customers and firms population density), LAC cities experience can support a large and diversified array of positive agglomeration effects.21 Controlling suppliers of final and intermediate goods and for features of a city’s physical geography, a services, an effect facilitated by connections 1 percent increase in population density is to other cities and the markets they provide. associated with a 0.049 percent increase in It can also spread the cost of large-scale productivity (and nominal wages). 22 This investments in transport and infrastructure “unconditional” estimate is close to that for that underlie the supply of basic services. And the United States (0.046 percent), using a com- the interaction of workers within and across parable regression specification, but far BOX O.1  Form, Skill, and Access As the net outcome of the tussle between agglom- agglomerations is the governance challenges that eration and congestion forces, a city’s productivity they face in coordinating the provision of infra- depends on form, skill, and access. structure and basic services in a space fragmented Form refers to the size and configuration of a by local government administrative boundaries. city. A city’s size (usually measured by population The second dimension is geometric shape. For or density) influences the likelihood of interactions example, in a perfectly circular city, the average among individuals and firms. These interactions distance between two locations is lower than in an can stimulate a wide array of positive, productivity-​ elongated city of the same area. All else being equal, enhancing agglomeration effects. But they can therefore, interactions take place more easily in a also generate negative congestion effects, such as circular city than in an elongated city. increased crime and a heightened probability of The third dimension of form is internal struc- spreading communicable diseases. In the absence ture, which depends on a city’s road network among of offsetting investments and policies, density also other things. For example, mobility is easier in cities brings other negative congestion effects as markets with well-planned road networks that follow a reg- and infrastructure become crowded. ular pattern, such as a grid, than in cities with more A city’s configuration, meanwhile, encom- haphazard networks. passes several dimensions. The first dimension is The fourth dimension is land use. For exam- whether a city is, in fact, a broader metropolitan ple, building restrictions may favor sprawl, which area or multicity agglomeration. In an agglomera- in turn can increase the journey to work as well as tion, the boundaries between one administratively other travel times. Similarly, although land zoning defined “city” and another blur to such an extent is necessary—for example, to keep a chemical plant that it becomes difficult to tell where one ends and from locating in a residential area—overly strin- the other begins. A distinctive aspect of multicity gent zoning requirements may create unnecessary (continued) O v er v ie w   11 BOX O.1  Form, Skill, and Access (continued) d ­ istance between the places where people live and ­ thers in the city as well. For example, workers in a o the places where they work. city with a higher share of college-educated individuals Skill refers to a city’s aggregate stock of human cap- will likely be more productive because they will have ital, or aggregate skill. When individuals choose where greater opportunity to interact with such individuals to live in a country, they compare locations on the basis and learn from them, regardless of their own skill level. of such attributes as wages, job opportunities, hous- Access refers to a city’s connectedness to other ing values, natural amenities, manmade amenities— cities through the transportation network. a When including, for example, cultural attractions—and the a city is well connected to others, transporting peo- demographic composition of the population. Given ple or goods to and from other cities is not costly. In their preferences and personal characteristics (such as such a city, firms have access to markets that extend age, education, and place of birth)—they thus sort into beyond that of the city itself. By promoting trade with different cities. A city that attracts more skilled individ- other cities both domestically and internationally, this uals could be more productive simply because its resi- allows firms in a city to expand, become more special- dents are on average more productive—yet this is not ized, and benefit from economies of scale. And, when the meaning of “skill” as it relates to city productivity. firms and workers become freer to move between cit- Instead, skill refers here to the productivity contri- ies, they flock to more productive cities. In a system of bution of skilled individuals above and beyond their well-connected cities, the dispersion of productivity own productivities. This contribution arises because across cities is minimized, and cities maximize their a person’s human capital benefits not only her but overall contribution to national productivity. a. In this book, we focus mainly on access to other cities and areas in the same country. Hence, the results in both figures O.10 and O.11 are based on a measure of domestic market access. Some discussion of the importance of international market access through ports and airports, as well as the road and rail networks that connect cities to them, is contained in chapter 4. weaker than that for China (0.192 percent) or FIGURE O.10  Unconditional and Conditional Effects of Density India (0.076 percent).23 on Productivity Provide Insights into the Mechanisms for However, both skill and access are posi- Agglomeration Effects tively correlated with population density because more densely populated cities tend to 0.06 Elasticity of city productivity have higher average levels of human capital 0.05 and better access to other cities’ markets 0.04 through transportation networks. Figure O.10 shows “conditional” estimates of agglomera- 0.03 tion effects. When we control for skill (mea- 0.02 sured by average years of schooling), the estimated agglomeration effect shrinks to 0.01 0.013 percent. When we control for access 0 (measured by a market access index) as well Density Density and Density, skill, only skill and access as skill, it becomes almost zero.24 Comparing “conditional” agglomeration Source: Quintero and Roberts 2017. effects for LAC and other regions is difficult Note: The figure shows the sensitivity (elasticity) of city productivity to population density for because conditional estimates for other different regression specifications in which cities from 16 LAC countries are pooled. Productivity is measured as (log) city average nominal wage, controlling for worker characteristics. The first bar regions do not control for both skill and (“Density only”) shows the effect of population density on city productivity without controlling for market access. Yet, those that control for skill skill and access; the second bar (“Density and skill”) the effect of population density on city productivity when controlling for skill, but not access; the third bar (“Density, skill, and access”) the (measured by percent of the working-age effect of population density on city productivity when controlling for both skill and access. Skill is population with higher education), without measured as log average years of schooling. All three regression specifications control for features of a city’s physical geography (mean air temperature, terrain ruggedness, and total precipitation) controlling for access, paint a similar picture and include country fixed effects. The orange bar represents coefficients that are not significantly because the estimated agglomeration effect different from zero at the 10 percent level. 12   RAISING THE BAR for LAC (0.023 percent) is similar to that of To gain insight into the relative impor- the United States (0.024 percent), but lower tance of form, skill, and access, figure O.11 than that for China (0.112 percent) or India shows the sensitivity of productivity to form (0.052 percent). The LAC effect is, however, (density), skill, and access when all three are less precisely estimated than that for China, included in the same regression specification India, or the United States. 25 (along with features of a city’s physical This analysis indicates that agglomeration geography). effects in the LAC region operate mainly through skill (as workers in a city learn from Form.  Holding skill and access constant, skilled workers), and much less through density has, at best, no impact on productivity; access (as cities gain access to the markets of at worst, it has a negative impact. An increase other cities). By contrast, other positive in density is associated with almost no agglomeration effects in LAC cities that might change in city productivity. The response of be associated with population density seem to productivity to density varies across countries, be largely absent—such as those that might but its effect is significantly positive only for arise from better job matches, the growth of a Brazil, Dominican Republic, Ecuador, and large and local diversified array of specialized Peru. For Chile and Nicaragua, the effect of suppliers, spreading costs of large investments density on productivity is significant but in infrastructure and transport, and more negative (figure O.12). general knowledge spillovers beyond those ­ Other findings are also consistent with the associated with skilled workers. notion that, under current infrastructure and FIGURE O.11  The Effects of Form, Skill, and Access on Productivity a. Specification using average years b. Specification using percentage of the working-age population of schooling (age 14–65 years) with higher education 0.7 0.030 0.6 0.025 0.5 0.020 0.4 Elasticity Elasticity 0.015 0.3 0.010 0.2 0.1 0.005 0 0 Population Market Average years of Population Market Share of working-age density access schooling density access population with higher education Source: Quintero and Roberts 2017. Note: The figure shows the sensitivity (elasticity) of city productivity to density, skill, and access when all three are included in the same regression specification in which cities from 16 LAC countries are pooled. Productivity is measured as the (log) city average nominal wage, controlling for worker characteristics. Density, market access, and average years of schooling are in logs. For example, an increase in average years of schooling equal to 1 percent raises productivity (and thus wages) by 0.57 percent. The regression specification controls for features of a city’s physical geography (mean air temperature, terrain ruggedness, and total precipitation) and also includes country fixed effects. Orange represents coefficients that are not significantly different from zero at the 10 percent level. Panel a shows coefficients using average years of schooling as a measure of aggregate skill; in panel b, aggregate skill is measured through the percentage of the working-age population with higher education. O v er v ie w   13 FIGURE O.12  In Most Countries, a City’s Population Density Does Not Have a Positive Significant Effect on Its Productivity 0.08 0.06 0.04 0.02 0 –0.02 –0.04 –0.06 lic ico ala ia a ru r a as il ca ile r do do bi az gu liv Pe ur ub Ri Ch em ex m Br ua lva ra Bo nd sta ep lo M ca Ec at Sa Ho Co nR Co Ni Gu El ica in m Do Source: Quintero and Roberts 2017, background paper for this book. Note: The figure show the estimated elasticities to population density for each country derived from regressing, in country-level regressions, estimates of city productivity (measured in natural logs) on the following variables, expressed in natural logs: population density, average years of schooling, market access, mean air temperature, terrain ruggedness, and total precipitation. Productivity is measured as (log) city average nominal wage, controlling for worker characteristics. The orange bars represent coefficients that are not significantly different from zero at the 10 percent level. This figure excludes Argentina, Panama, and Uruguay because these countries lack a sufficient number of subnational locations (that is, observations for the regressions) to permit reliable estimation. policy conditions, density does not contribute an inadequate enabling environment associ- to city productivity in the LAC region. ated with a lack of infrastructure investment, A study for this book finds that, opposite to poor planning, and more generally poor what is found for the rest of the world, the urban management in cities. For example, LAC region’s labor productivity is lower in even if they have the same density, cities with large cities than in its smaller cities, after con- fewer vehicles on the road (perhaps because trolling for elements of a city’s business envi- of better public transportation) or with ronment and firm characteristics such as better traffic management systems will be industry, size and ownership structure, age, less congested. Indeed, four LAC cities— and whether the firm is an exporter (Reyes, Buenos Aires, Mexico City, Rio de Janeiro, Roberts, and Xu 2017).26 And, when consid- and Santiago de Chile—are among the ering all countries in the world, there is either world’s most congested, and Mexico City no association, or a negative one, between tops the chart.28 national levels of productivity, measured by Congestion effects in the form of crime GDP per capita, and density.27 might also be aggravated by there being little As mentioned above, the weak (or even basic protection from theft, kidnapping, and negative) contribution of density to city pro- other criminal activity. Across the world, labor ductivity in the LAC region suggests the productivity and firm total factor productivity absence of positive agglomeration effects (TFP) are lower in cities with higher private beyond those associated with skill and security costs, perhaps because firms must pay access. Because the region has relatively for private security to fill the void left by local dense cities, they may be suffering from neg- police (Reyes, Roberts, and Xu 2017). A case ative congestion effects, which more than study of Colombia for this book finds that offset positive agglomeration benefits. high levels of crime and violence have large, Congestion, in turn, may be aggravated by negative, and statistically significant effects on 14   RAISING THE BAR firm TFP, with large productivity losses asso- (approximately) raise their salaries by a ciated with the presence of paramilitary and remarkable 20 percent, coming in equal parts drug-trafficking groups in a city (Balat and from own and aggregate human capital. Casas 2017). Returns are not as high when skill is mea- Beyond density, other dimensions of a city’s sured by the share of higher education grad- form also bear on productivity. Most salient is u at e s i n a c i t y ’s p o p u l at i o n , w i t h the presence of MCAs. Although the associa- productivity rising by 2 percent for every tion between country productivity (measured 1 percentage point rise in share of graduates by log GDP per capita) and the share of a (figure O.11, panel b). country’s population that lives in MCAs is Regardless of the metric, returns to skill in positive in North American and Western cities are relatively high in the LAC region by European countries, it is virtually zero in LAC international standards. Although in other countries. This suggests that LAC countries parts of the world they are equal to 50–100 may not handle effectively the difficult coordi- percent of the private returns, in the LAC nation challenges that MCAs pose. Evidence region they are equal to 100 percent or more, from 73 large metropolitan areas in the LAC reflecting the region’s scarcity of skills region indicates that, although half of them (Duranton 2014). have a metropolitan-level governance body, Although returns to skill in cities are posi- the mere existence of such a body does not tive for all LAC countries, they vary across yield productivity gains, pointing to the need countries depending on average skill in the for better institutional arrangements. average city. The relationship is U-shaped, Lower productivity is also a feature of LAC indicating that, when a country’s cities have a cities with a long irregular shape (as opposed, low average skill level, returns fall as average say, to “round” cities). And it is a feature of skill rises yet increase after the cities reach a cities where segments of the street network critical skill level (figure O.13). are poorly connected (due, say, to dead ends, Returns to city skill are also U-shaped circular streets, and few street intersections). for an individual’s own level of education (figure O.14), indicating that, as an individ- Skill.  In the productivity race between ual’s skill rises, the return she or he enjoys density, skill, and access, skill emerges as the from city skill first falls and then rises. This clear winner. Holding density and access pattern likely reflects the interplay between constant, a 1 percent increase in skill the two sources of social returns to human (measured as average years of schooling) is capital: complementarities and human capi- associated with a 0.57 percent increase in city tal externalities. Complementarities arise productivity, much higher than the associated when skilled workers in a firm raise the pro- increase for density or access (figure O.11, ductivity of other workers (usually unskilled panel a). While the contribution of skill to ones) and are paid for it. For example, productivity varies across countries, it is skilled workers in a firm may streamline the significantly different from zero29 and positive production ­ process and thus enhance the for all of them, which is not the case for productivity of the firm’s unskilled workers. density or access. Complementarities also arise when the To understand the responsiveness of pro- greater presence of skilled individuals in a ductivity to skill, note that, when a LAC city raises demand for unskilled workers worker acquires an additional year of school- (who work at restaurants and drive cabs, for ing, his or her salary rises by 8.9 percent on example). In contrast, human capital exter- average; 30 when a city’s average years of nalities arise when skilled workers in a firm schooling rises by one year, salaries in the city raise the productivity of workers, perhaps in rise by 9.2 percent on average. This means other firms, but are not paid for it. For that, if all people within a city were to acquire example, skilled workers may exchange an extra year of education, this would knowledge and ideas with workers from O v er v ie w   15 FIGURE O.13  Across Countries, Returns to Skill Are U-Shaped in Average City Skill a. Returns to average years of schooling b. Returns to percent of higher education graduates BOL 0.10 0.20 HND HND 0.08 GTM Estimated returns to city skill Estimated returns to city skill 0.15 0.06 GTM 0.10 MEX CHL BOL BRA PER 0.04 BRA COL MEX 0.05 ECU CRI 0.02 ECU CHL NIC SLV SLV NIC COL PER DOM CRI 0 DOM 5 6 7 8 9 10 0 2 4 6 8 Average years of schooling Average percent of higher education graduates Source: Calculations using SEDLAC for countries other than Brazil and IPUMS International for Brazil. Sample covers 2000–2014. Note: The vertical axis shows, for each country, the estimated returns to city skill. The horizontal axis shows, for each country, the average of the corresponding variable; the average is calculated over the country’s cities. Average years of schooling, and percent of higher education graduates, correspond to individuals age 14–65 years. Returns can be expressed in percent if multiplied by 100. To obtain these returns, for each country we regress city-level productivity on the corresponding measure of city skill. These regressions control for area density, market access, air temperature, terrain ruggedness, and precipitation. City-level productivities are estimated by regressing, for each country, log wages on individual-level characteristics (age, age squared, years of schooling, gender, and marital status) and year fixed effects. We do not run these regressions for Argentina, Panama, and Uruguay because of their low number of cities. Coefficients from the quadratic specification in panel b are significantly different from zero. Coefficients from the quadratic specification in panel a are not significantly different from zero. IPUMS = Integrated Public Use Microdata Series; SEDLAC = Socio-Economic Database for Latin America and the Caribbean. For a list of country abbreviations, see annex 2A. other firms, either in formal settings such as Meanwhile, the least educated individuals conferences and public presentations or in enjoy the highest returns to the share of col- informal settings such as school meetings, lege graduates because, in addition to human c iv ic a s so c i at ion s , or neig hb orho o d capital externalities, they may benefit from interactions. complementarities as well. In general, an increase in city skill will The U-shaped pattern in figure O.14, raise salaries for unskilled workers because panel b, likely reflects a different balance of of both complementarities and human capital complementarities and human capital exter- externalities, yet it will have two opposing nalities. Because average years of schooling is effects on the salaries of skilled individuals: a about seven years (close to where returns negative effect due to greater relative supply reach a ­minimum in the figure) for the average of skilled individuals, and a positive effect LAC city, the average worker in this city is due to human capital externalities. So an unskilled. Thus, his or her impact on the pro- increase in city skill that leads to higher sala- ductivity of others is more likely to come from ries for skilled workers can be regarded as complementarities than from externalities. evidence of human capital externalities. Additional schooling for the average worker The U-shaped pattern in figure O.14, may hurt individuals with the least amount of panel a, provides evidence of human capital schooling, with whom he or she competes. externalities. The positive return to the share However, it may benefit individuals with more of higher education graduates among individ- schooling, by allowing them, for example, to uals with complete higher education suggests specialize in more complex tasks and leave the existence of human capital externalities. other tasks to the average worker. 16   RAISING THE BAR FIGURE O.14  Individual Returns to Skill Fall and Then Rise with Own Education a. Returns to percent of higher education graduates b. Return to average years of schooling 0.030 0.18 0.16 0.025 0.14 Returns to city skill Return to city skill 0.020 0.12 0.10 0.015 0.08 0.010 0.06 0.04 0.005 0.02 0 0 2 4 6 8 10 12 14 16 18 y y ry ry n at er ar ar tio da da uc gh Individual’s own years of schooling rim rim n ca on on io ed e hi du ep ep ec ec et re es es m et pl he pl So m et m m ig pl So Co Co eh m Co m So Source: Calculations using SEDLAC for countries other than Brazil and IPUMS International for Brazil. Sample is the same as that used by Quintero and Roberts (2017), covering 2000–14. Note: To construct panel a, we pool data from all countries and regress log wages on individual characteristics (age, age squared, indicators of educational attainment, gender, and marital status) interacted with country dummies, city-level characteristics (density, share of college graduates, market access, air temperature, terrain ruggedness, and precipitation), country-year fixed effects, and the interaction between indicators of individual educational attainment and the city share of college graduates. Individuals with complete primary (secondary) have not started secondary (higher) education. To construct panel b, we pool data from all countries and regress log wages on individual characteristics (age, age squared, years of schooling, years of schooling squared, gender, and marital status) interacted with country dummies, city-level characteristics (density, average years of schooling, market access, air temperature, terrain ruggedness, and precipitation), country-year fixed effects, the interaction between own years of schooling and average years of schooling, and the interaction between own years of schooling squared and average years of schooling. All relevant coefficients for these panels are significantly different from zero. IPUMS = Integrated Public Use Microdata Series; SEDLAC = Socio-Economic Database for Latin America and the Caribbean. Access.  Access to the markets of other cities with low economic potential. Indeed, a case in the same country through transportation study of Mexico for this book finds a stron- networks has a statistically significant ger effect of market access on city productiv- association with city productivity. Holding ity when adopting an estimation strategy that density and skill constant, a 1 percent controls for this potential bias (Blankespoor increase in access is associated with a 0.015– et al. 2017).31 It also finds that Mexico’s road 0.020 percent increase in productivity, well investment in recent decades was associated below the increase associated with skill but with local job growth and output, and with above the increase associated with density. increasing specialization among manufactur- The responsiveness of productivity to access ing firms. In other words, market access holds varies among countries and is significantly the promise of raising city productivity. different from zero in 6 out of 13 countries Second, even if cities have access to other (figure O.15). cities through the transportation network, Multiple factors may explain the low using the network may be costly in money impact of access on the productivity of LAC (due, say, to high toll prices, or to a noncom- cities. First, our estimate of access impact petitive transportation sector that limits sup- may be biased downward. This may be the ply and raises prices) or in difficulty (due, for case, for example, if transport investments example, to low road safety or to frequent have targeted cities in lagging regions, traffic disruptions created by protests). O v er v ie w   17 FIGURE O.15  Market Access Is Associated with City Productivity in Some Countries 0.05 0.04 0.03 0.02 Elasticity 0.01 0 –0.01 –0.02 –0.03 a r ca il lic ru a ico ile ia ala r as do do az gu bi liv Pe ur ub Ri Ch em ex m Br ua lva ra Bo nd sta ep lo M ca Ec at Sa Ho Co nR Co Ni Gu El ica in m Do Source: Quintero and Roberts 2017. Note: Figures show the estimated elasticities to market access for each country derived from regressing—in country-level regressions—estimated city productivity (measured in natural logs) on the following variables, measured in natural logs: population density, average years of schooling, market access, mean air temperature, terrain ruggedness, and total precipitation. Productivity is measured as (log) city average nominal wage, controlling for worker characteristics. The orange bars represent coefficients not significant at the 10 percent level. This figure excludes Argentina, Panama, and Uruguay because these countries lack a sufficient number of cities (that is, observations for the regressions) to permit reliable estimation. Access holds the promise of raising the some of the countries, although its esti- productivity not only of individual cities mated effect may be biased. These results but also of the whole system of cities. suggest that other types of agglomeration Indeed, improvements in national trans- effects associated with population density port networks can help create a more inte- are largely absent in LAC cities, which may g rate d s y s tem of c it ie s — w it h lower not have the necessary enabling environ- productivity dispersion across cities and ment. For example, current levels of infra- with a higher contribution to national pro- structure, urban management practices, du c t i v i t y. Ev i d e n c e f r o m c o u n t r i e s a nd t ra nspor t at ion pol icies may not throughout the world shows that the ­ adequately support LAC cities’ relatively within-country productivity dispersion high densities, resulting in congestion across cities is lower in countries with overwhelm positive agglomera- forces that ­ higher road density (figure O.16). tion effects. Institutional weaknesses that To summarize, agglomeration effects limit coordination across local govern- driven by skill—and, to much less extent, by ments in metropolitan areas may also access—are strong in LAC cities. Skill, dampen agglomeration effects. And the which has a positive effect in all countries, high within-country productivity disper- operates through complementarity between sion indicates that LAC city systems are not skilled and unskilled workers, and through efficient and do not maximize their contri- human capital externalities, mostly from bution to national productivity, likely skilled workers. Market access has a small because of poor intercity connectivity estimated positive impact, driven by only through the transportation network. 18   RAISING THE BAR FIGURE O.16  Countries with Better Road Coverage Have More Efficient Systems of Cities 3 Within-country dispersion of city productivity 2 PER BOL GTM NIC PAN SLV PRY COL CRI 1 CHL MEX ARG JAM BRA 0 DMA −2 0 2 4 6 8 Road density (log) LAC countries Non-LAC countries Source: Calculations based on nighttime lights data from the 2015 VIIRS annual composite product (https://ngdc.noaa.gov/eog/viirs/download_dnb​ _­composites.html) and road density data from the World Bank’s World Development Indicators database (http://data.worldbank.org/data-catalog​ /­world-development-indicators). Note: Productivity is measured using the residuals from a regression at the city level where the dependent variable is the sum of nighttime lights (in logs) and the independent variable is the population (in logs). Productivity dispersion across a country’s cities is measured by the interquartile range of the distribution of productivity. Road density is the ratio of the length of the country’s total road network to the country’s land area and is measured in kilometers per 100 km2 of land area. VIIRS = Visible Infrared Imaging Radiometer Suite. For a list of country abbreviations, see annex 2A. That LAC cities lag the world’s productiv- What These Findings Might ity frontier might be due not only to market Mean for Policy failures, but also to policy failures. 32 For Although this book is intended primarily as a example, although LAC cities benefit from research piece, its rich results can provide positive agglomeration effects, these effects food for thought for policy makers. As with are mainly associated with city skill—with any piece of applied research that makes use complementarities between skilled and of diverse data sets and a variety of methods, unskilled workers, as well as with spillovers extracting this food for thought is not of knowledge from skilled workers. By con- ­ necessarily straightforward. As might be trast, LAC cities largely lack other positive expected, not all results are consistent across agglomeration effects—such as those that the different methods and data sets. Even might arise from good job matches, a large when methods and data are consistent, not and diversified array of local suppliers of all results apply to all countries. In stepping intermediate inputs, the cost-sharing of large- back and viewing the body of research scale infrastructure and transport, and other presented in this book as a totality, several knowledge spillovers. Policy makers may thus policy-relevant insights emerge. need to improve the enabling environment O v er v ie w   19 for these broader agglomeration effects. migration of the unskilled to cities. Because Improvements may include carefully planned migration will increase the population (and infrastructure and public services to mitigate probably density) of cities, it may also the congestion created by current density. increase their congestion. It is all the more They may also include stronger coordination critical, then, for cities to create an enabling among municipalities in large metropolitan environment for strong agglomeration areas or MCAs, as well as effective policies effects. for deterring crime and improving security. At the same time, unskilled populations in LAC systems of cities do not seem to oper- cities mostly work in low-productivity, often ate efficiently. Within countries, cities seem local, services such as retail, hotels, and to be poorly integrated, and resources do not restaurants. 35 Under prevailing conditions, seem to flow to their most productive uses. further urbanization of unskilled workers Skilled people are strongly concentrated in may just continue shifting workers from the largest cities—even more than in the agriculture and manufacturing into low-­ United States. The concentration is in part productivity sectors. A better enabling envi- due to a relative shortage of skilled people at ronment for agglomeration effects, which the national level, which makes investing in operate more strongly for the formal sector, human capital across the board a priority. and for tradable goods and services may But it may also be due to an unequal distribu- reduce that effect. 36 To be productive, cities tion of basic services across areas that dispro- also need the enabling environment of sound portionately favors large cities. Although macroeconomic policies and efficient mar- improving this distribution would help in kets for goods and services, which are critical principle, great care must be exercised in the to the existence of productive firms, good design of relevant programs aiming to do jobs, and high national productivity. Without so—to make sure that gains for one city do this wider enabling environment, LAC cities not merely come at the expense of others.33 are not likely to reach the world’s productivity The inefficiency of city systems also frontier. appears related to the underdevelopment of Cities thus are lenses to consider a whole national transport networks and to barriers host of policies, including education, infra- to mobilit y across cities. E x pa nd i ng structure, transportation, and urban plan- transportation networks, and lowering the ning. Cities are the immediate context in pecuniary and nonpecuniary costs of their use, which people live and work. And, because would in principle raise cities’ productivity. almost three-quarters of the LAC population Eliminating obstacles that might constrain live and work in this context, raising the bar people from moving to the cities where they for the productivity of LAC cities is crucial. would be most productive might also help. For Although this book cannot provide all the example, a city’s inelastic housing supply can policy answers, we hope that, by taking a mean that, as the city grows, housing prices stride forward in our knowledge of the deter- rise rapidly but the housing stock does not, minants of productivity in LAC cities, the which limits people’s ability to move to the book can raise the bar for productive cities in city even if they would be more productive the region. there than in other places. Similarly, a city’s high crime rates might discourage people from moving to that city, even if they would be Annex OA: Productivity Measures more productive there.34 Used in the Book to Assess Whereas almost all skilled individuals in LAC Cities the region live in cities, many unskilled indi- viduals do not. Going forward, any addi- •  Per capita GDP at the national level proxies t io n a l u rb a n i z at io n t h at L AC m ay average labor productivity at the national experience will most likely be driven by the level and is relevant to the aggregate 20   RAISING THE BAR ­ ontribution of urbanization and cities to c others are negative. Aggregate skill, for national productivity. example, is subject to positive externalities. •  Nighttime lights (NTL) at the city level Although many workers in a city gain when measures output at the city level. Because aggregate skill rises, any one worker regards city-level GDP is typically not available, his or her contribution to the aggregate skill researchers have used the intensity of an level as negligible. Thus, when deciding area’s NTL as a proxy for its level of eco- whether to acquire more skill, individuals do nomic activity.37 not consider the benefit of their actions for •  NTL net of (log) population at the city the whole city and are thus less likely to level measures average labor productivity acquire additional skill. As a result, aggregate at the city level. skill in the city is below the socially optimal •  Average nominal wages at the city level level. Meanwhile, traffic congestion and pol- is a commonly used measure of a work- lution externalities represent classic textbook er’s productivity in the urban economics cases of congestion effects that are negative literature, especially in literature that esti- externalities. mates the strength of agglomeration econ- Cities are also notorious for their public omies (for example, see Duranton 2016 good problems. Cities typically contain and Chauvin et al. 2017). All other things infrastructure (such as bridges and roads) equal, a city that pays a higher average and public spaces (such as parks and town nominal wage can be considered to have a squares) that can be enjoyed by many indi- higher average level of labor productivity. viduals at once, without an easy mechanism •  Average nominal wages net of individual to exclude users. No individual is willing to worker characteristics at the city level mea- pay for public goods because all have an sures a city’s labor productivity, having con- incentive to free ride by letting others pay. As trolled for differences in the composition a result, no private firm is willing to provide of its workforce. If workers with the same public goods. observable characteristics (such as age, edu- Cities also suffer from coordination fail- cation, marital status, and gender) who live ures. Within a city, individual firms and in different cities within a country earn dif- workers may fail to coordinate. For example, ferent wages, it must be because their cities although all individuals may desire clean air, have different productivity levels.38 which could be more easily accomplished if •  TFP at the firm (establishment) level cap- more of them used public transportation tures firm productivity, net of the capital, rather than driving their own vehicles, many labor, and intermediate inputs used in the individuals may find it more convenient to production process. drive. Given the practical difficulties of coor- dinating among themselves, individuals may end up driving, thus raising pollution above Annex OB: The Need for Policy the socially optimal level. Similarly, many Cities represent potential engines of produc- firms might benefit from moving to a given tivity and growth. But, if cities are left location within the city if a sufficiently large to markets alone, this potential cannot be group of them moves there, yet no individual realized—for several reasons. firm might gain from moving alone. In the Externalities arise when a decision by an absence of a mechanism to coordinate their economic agent, such as a worker or firm, actions, firms might remain where they are, has consequences for other agents, yet the and might all be worse off. agent’s decision does not take such conse- Cities may fail to coordinate as well. The quences into account. In these cases, what is cities of an MCA can fail to coordinate, as best for the individual is not best for society discussed in the main text. More broadly, as a whole. Externalities are pervasive in cit- the cities in a system can fail to coordinate. ies. Some of these are positive, whereas many They may not have incentives to invest in O v er v ie w   21 human capital when workers are mobile which the frontier is implicitly defined by the across cities, because they may not reap the set of countries that exhibit the highest levels return to their investment if the workers of GDP per capita at given levels of urbaniza- move. And, if a public transit link benefits tion. Similar comments apply to figure O.2, where the frontier is defined by the set of cities two cities, neither city has an incentive to that exhibit the highest levels of economic invest in the link because the other city will activity at given levels of population. benefit as well. 6. This statement is based on the average perfor- More broadly, systems of cities can suffer mance of LAC cities—so LAC cities, on aver- from barriers to mobility that raise the cost age, exhibit higher output than we would of moving resources across cities. Whereas expect based on their populations. However, some of these barriers can be natural (a as also shown in figure O.2, LAC cities show mountain range), others arise from policy considerable variation around the average, regulations (overly restrictive building and with some exhibiting levels of output much planning regulations), or from coordination lower than we would expect based on their failures among cities (the two cities that could populations. We discuss the dispersion of pro- ductivity levels across LAC cities later in the benefit from a connecting transport link). overview. Such market failures justify policy inter- 7. Argentina, Barbados, and Grenada provide vention, both for cities and for systems of the most notable exceptions to the finding cities. that LAC countries have unusually dense cit- ies. Antigua and Barbuda, the Bahamas, Guyana, Jamaica, and St. Kitts and Nevis all Notes have a roughly 50:50 split between dense and 1. These figures are based on the globally consis- not dense cities. tent definition of urban areas that we intro- 8. Just as important as a city’s average density duce in chapter 1. They differ from from the perspective of fostering positive corresponding figures based on official agglomeration effects and mitigating conges- national definitions of urban areas, which, as tion is likely to be how that density is orga- discussed in detail in chapter 1, vary widely nized. This is discussed more in chapter 6. not only across countries within LAC but also 9. In this book, MCAs are defined as urban areas across countries globally. On the basis of identified by the cluster algorithm that encom- national definitions of urban areas, the share pass two or more cities as given by countries’ of the LAC region’s total population living in own definitions. Each component city must cities in 2016 was 80.1 percent. have at least 100,000 people. Nevertheless, as 2. A mega-city is generally defined as a city that discussed in chapter 2, our main regression has a population in excess of 10 million. results relating national productivity and the 3. The algorithm that we use is from Dijkstra share of population in MCAs also hold when and Poelman (2014). In total, we identify allowing for smaller component cities. almost 64,000 urban areas globally, of which 10. In the case of Mexico City, several of the just under 7,200 belong to LAC. For ease of municipalities in the officially defined city exposition, we refer to urban areas as “cities” (shown in map O.1 by the yellow lines) only throughout this overview, even though the overlap partially with its “true” urban extent. smaller and less densely populated urban 11. Within-country productivity dispersion does areas may perhaps be more aptly referred to not necessarily indicate inefficiency in the sys- as “towns.” To be classified as an urban area, tem of cities; it could also indicate a disparity a cluster must have a minimum density of 300 in the presence of amenities. For example, people per square km, and the cluster’s total some individuals may choose to live and work population must be at least 5,000. in a city where they do not maximize produc- 4. Multiple measures of productivity are used in tivity or wages simply because the city is close this book. See annex OA for a list of such to the beach. In these cases, productivity in measures. the country is not maximized, yet welfare is. 5. The concept of a global productivity frontier, However, assuming that the disparity in as presented here, is a purely empirical one in amenities accounts for a similar fraction of ­ 22   RAISING THE BAR productivity dispersion in LAC countries and Source for the United States: U.S. Census their comparators, we can view the high pro- Bureau, Current Population Survey 2010. ductivity dispersion within LAC countries Returns to higher education correspond to (relative to their comparators) as indicative of complete higher education. Source for LAC is inefficient systems of cities in the LAC region. Ferreyra et al. (2017); estimates for the United 12. An individual is defined as skilled who has at States are based on Card (2001) and Heckman, least some higher education. Lochner, and Todd (2006). 13. From figure O.7, Argentina is an exception to 19. Following Ferreyra et al. (2017), “higher edu- this pattern. cation” encompasses both bachelor’s programs 14. We obtain this result by regressing (log) share (akin to the bachelor’s programs in the United of the city population that is skilled on (log) States) and short-cycle programs (akin to asso- city population, pooling data for all LAC ciate degree programs in the United States). countries. Results are very similar when coun- 20. The high concentration of skills in large cities try fixed effects are included, or when we run may itself be a symptom of the underdevelop- a separate regression per country and average ment of national transport networks and, the country-specific coefficients. more generally, of a lack of integration 15. The Gini coefficient is a measure of inequal- between cities. Hence, high migration costs ity in the income distribution. It ranges associated with a lack of integration may limit between zero and 1. The larger the coeffi- migration for the unskilled more than for the cient, the greater the inequality. We obtain skilled, rendering the skilled more likely to the LAC elasticity (equal to 0.029) by migrate than the unskilled—as is the case in regressing (log) city Gini coefficient on (log) Brazil (Fan and Timmins 2017). city population, pooling data for all LAC 21. This discussion is based on the regressions, countries. When including country fixed which cover subnational areas in 16 LAC effects, the coefficient of this regression rises countries, reported in chapter 3. These regress from 0.029 to 0.042. The U.S. elasticity city productivity (in logs), as measured net of (equal to 0.012) comes from Behrens and individual worker characteristics, on (log) Robert-Nicoud (2015). population density, (log) mean air tempera- 16. Income inequality can be decomposed into ture, (log) terrain ruggedness, and (log) total two components: between-group and precipitation. within-​ group inequality. These correspond 22. To assess the role of form, skill, and access in to income inequality among individuals with city productivity, we measure productivity different skill levels, and among individuals ­ through average nominal wages, net of worker with the same skill level, respectively. Even if characteristics. See annex OA for further all ­ individuals in a city are skilled, income details on productivity measures used in this might be unequally distributed if income for book. For a discussion of why nominal wages the skilled is dispersed. In the LAC region, provide an appropriate measure of productiv- however, the greater inequality of larger cit- ity, see Combes and Gobillon (2015). ies is driven by between-group inequality— 23. Estimates from China, India, and the United by income inequality among individuals States come from Chauvin et al. (2017), who with different skill levels. do not control for cities’ physical geographic 17. The elasticity of the Gini coefficient with attributes (such as climate and terrain). As in respect to population falls from 0.012 to our case, they use individual-level data and 0.009 for the United States when controlling use density as a right-hand side ­variable. Using for city education (Behrens and Robert- aggregate data, Ciccone and Hall (1996) and Nicoud 2015). On average (across countries), Rosenthal and Strange (2008) ­ estimate this elasticity in the LAC region falls from agglomeration effects for the United States of 0.03 to 0.017. 0.04–0.05 percent. 18. These are private returns to higher education, 24. For each city, the market access index reflects accruing to the individual who attains it. the number of cities to which the city is con- Percent of skilled population is calculated rel- nected through the road network, the time it ative to the population ages 25–65 years in takes to travel to those cities, and those cities’ each country. Sources for LAC: SEDLAC for all population. See chapters 3 and 4 for further countries other than Brazil; IPUMS for Brazil. details. O v er v ie w   23 25. Although the effects for China, India, and the for some migrants to high-wage metro areas, United States reported by Chauvin et al. the poor quality of such housing may have (2017) are all statistically significant at the 1 deterred would-be migrants, thereby keeping percent level, the effect that we estimate for them “trapped” in less productive cities. LAC is significant at only the 10 percent level. 35. See chapter 5 for further details on the employ- 26. A large city here is defined as one that has a ment of skilled and unskilled individuals in population of more than 1 million or is a cities. As is well known, measuring productiv- national capital. ity in the service sector is remarkably difficult, 27. Country-level productivity is measured by partly because of the difficulties in measuring (log) GDP per capita, and average density is output. measured in two ways: as the weighted aver- 36. Chapter 3 presents evidence that agglomera- age of density levels across cities in a country tion effects are stronger in formal than in or as the percent of the population that lives informal economic activities. in dense cities. Findings are based on regres- 37. Among economists, the use of NTL to proxy sions that also control for a country’s urban for levels of economic activity has become share. For further details, see chapter 2. widespread since the work of Henderson, 28. This congestion ranking is based on TomTom Storeygard, and Weil (2011, 2012). Before data. See chapters 2 and 4 for further details. this, the ability of NTL to proxy for levels of 29. At the 5 percent significance level. economic activity had been highlighted in the 30. This is the average of country-specific field of remote sensing by, for example, Mincerian returns to schooling, controlling Elvidge et al. (1997). for individual characteristics. 38. Of course, the difference could also be due to 31. The bias associated with the endogenous systematic differences in their unobserved placement of transport infrastructure also has characteristics. We assume that controlling for the potential to go in the opposite direction. our set of observed individual characteristics Hence, the estimated coefficient on access minimizes the role of such differences. may be biased upward if transport invest- ments have been targeted at better connecting cities that policy makers anticipate will grow References rapidly. Balat, J., and C. Casas. 2017. “Firm Productivity 32. Annex OB describes the market failures asso- a nd C it ie s: T he C a s e of C olombi a .” ciated with cities. Background paper for this book, World Bank, 33. For example, place-based policies aiming at Washington, DC. boosting employment or economic activities Bastos, P. 2017. “Spatial Misallocation of Labor in specific areas have a mixed track record in Brazil.” Background paper for this book, (World Bank 2009). World Bank, Washington, DC. 34. In his case study of Brazil for this book, Behrens, K., and F. Robert-Nicoud. 2015. Bastos (2017) finds that the productivity dis- “Agglomeration Theory with Heterogeneous persion among workers in the formal sector, Agents.” In Handbook of Regional and Urban who made up two-thirds of the total Brazilian Economics, Volume 5, edited by Gilles workforce in 2013 (Messina and Silva 2018), Duranton, J. Vernon Henderson, and William has fallen in recent decades. This might have C. Strange, 171–87. Amsterdam: Elsevier. been prompted by the reduction of crime Blankespoor, B., T. Bougna, R. Garduno-Rivera, rates in the most productive metropolitan and H. Selod. 2017. “Roads and the Geography areas, which has served to attract workers of Economic Activities in Mexico.” Policy from other, less productive, areas. At the Research Working Paper 8226, World Bank, same time, productivity dispersion among cit- Washington, DC. ies in Brazil remains higher than in the United Branson, J., A. Campbell-Sutton, G. M. Hornby, States. One possible explanation is the short- D. D. Hornby, and C. Hill. 2016. “A Geospatial age of affordable housing in Brazil’s most Dat aba s e for L at i n A mer ic a a nd t he productive cities. On average, housing defi- Caribbean: Geodata.” Southampton, U.K.: cits have risen more in high-wage than in University of Southampton. low-wage metropolitan areas. Although Card, D. 2001. “Estimating the Return to informal housing presumably filled the gap Schooling: Progress on Some Persistent 24   RAISING THE BAR Econometric Problems.” Econometrica 69 (5): America and the Caribbean. Washington, DC: 1127–60. World Bank. Chauvin, J. P., E. Glaeser, Y. Ma, and K. Tobio. Heckman, J., L. Lochner, and P. Todd. 2006. 2017. “What Is Different about Urbanization “Earnings Functions, Rates of Return, and in Rich and Poor Countries? Cities in Brazil, Treatment Effects: The Mincer Equation and China, India, and the United States.” Journal Beyond.” In Handbook of the Economics of of Urban Economics 98: 17–49. Educ ation , Volume 1, edited by E . A. Ciccone, A., and R. Hall. 1996. “Productivity and Hanushek and F. Welch, 307–458. Amsterdam: the Density of Economic Activity.” American Elsevier. Economic Review 86 (1): 54–70. Henderson, J. V., A. Storeygard, and D. N. Weil. Combes, P. P., and L. Gobillon. 2015. “The 2011. “A Bright Idea for Measuring Economic Empirics of Agglomeration Economies.” In Growth.” American Economic Review 101 Handbook of Regional and Urban Economics, (3): 194–99. Volume 5, edited by Gilles Du ranton, ———. 2012. “Measuring Economic Growth J. Vernon Henderson, and William C. Strange, from Outer Space.” American Economic 247-348. Amsterdam: Elsevier. Review 102 (2): 994–1028. Dijkstra, L., and H. Poelman. 2014. “A Harmon­ ITF (International Transport Forum). 2017. ised Definition of Cities and Rural Areas: The “Capacity to Grow: Transport Infrastructure New Degree of Urbanization.” Regional Needs for Future Trade Growth.” Organisation Working Paper, Directorate-General for for Economic Co-operation and Development, Regional and Urban Policy, Eu ropean Paris. Commission, Brussels. Messina, J., and J. Silva. 2018. Wage Inequality Duranton, G. 2014. “Growing through Cities in in Latin America: Understanding the Past to Developing Countries.” World Bank Research Prepare for the Future . Washington, DC: Observer 30 (1): 39–73. World Bank. ———. 2016. “Agglomeration Effects in Colombia.” Quintero, L., and M. Roberts. 2017. “Explaining Journal of Regional Science 56 (2): 210–38. Spatial Variations in Productivity: Evidence Elvidge, C., K. Baugh, E. Kihn, H. Kroehl, E. from 16 LAC Countries.” Background paper Davis, and C. Davis. 1997. “Relation between for this book, World Bank, Washington, DC. Satellite Observed Visible-Near Infrared Reyes, J., M. Roberts, and L. C. Xu. 2017. “The Emissions, Population, Economic Activity Heterogeneous G row t h E f fe c t s of t he a n d E l e c t r i c P o w e r C o n s u m p t i o n .” Business Environment: Firm-Level Evidence International Journal of Remote Sensing for a Global Sample of Cities.” Policy 18 (6): 1373–79. Research Working Paper 8114, World Bank, Fan, L., and C. Timmins. 2017. “A Sorting Model Washington, DC. Approach to Valuing Urban Amenities in Rosenthal, S., and W. Strange. 2008. “The Brazil.” Background paper for this book, Attenuation of Human Capital Spillovers.” World Bank, Washington, DC. Journal of Urban Economics 64: 373–389. Ferreyra, M. M., C . Avitabile, J. Botero, World Bank. 2009. World Development Report: F. Haimovich, and S. Urzua. 2017. At a Reshaping Economic Geography. Washington, Crossroads: Higher Education in Latin DC: World Bank. PART Urbanization and Productivity in Latin America and the Caribbean I Cities in Latin America and the Caribbean and related trends of structural transforma- (LAC) are, on average, more productive than tion at the national ­ level. It also assesses those in many other regions of the w ­ orld. But whether GDP per capita levels among the they lag the global “frontier” of productivity region’s countries are higher or lower than performance, defined by North American might be expected given prevailing levels of and Western European c ­ ities. Considerable ­ u rbanization. Chapter 2 looks beyond the scope thus exists for “raising the bar” for share of a country’s population that lives in productive cities in the region, as well as cities to examine additional dimensions of increasing the contribution that cities, in urbanization within the region and their aggregate, make to national gross domestic links to national productivity p ­ erformance. product (GDP) per ­ capita. Part I of the book It also benchmarks the productivity of indi- provides an overview of major urbanization vidual LAC cities against those in the rest of trends and the productivity performance of the world, and analyzes the dispersion of pro- LAC c ­ ities. Chapter 1 analyzes urbanization cities. ductivity across the region’s ­ Urbanization, Economic Development, 1 and Structural Transformation Paula Restrepo Cadavid and Grace Cineas Introduction criteria, ­relying instead on official lists of cities. “Urban areas” mean different things ­ This chapter presents an overview of urban- in different countries, and at least some of ization trends in the Latin America and the apparent “underperformance” of LAC Caribbean (LAC) region and, where possible, u rbani zation m ig ht result from data their links to ­productivity. Unlike East Asia, ­artifacts. many of the LAC region’s largest urban cen- The chapter begins by briefly exploring the ters are inland, possibly a disadvantage for historical origins of population and economic international trade (Saavedra- Chanduvi and concentration in the region, reviewing the role Sennehauser 2009). LAC countries also of natural geographic endowments or location appear to have lower than expected produc- ­fundamentals.1 To do this, it builds on previ- levels. tivity, given their high urbanization ­ ous work by Maloney and Caicedo (2016) and Some analysts argue that policy distor- Henderson et a ­2016). When possible, it ­ l. ( tions have favored population concentrations assesses whether there are visible differences in in urban areas and capital cities, exacerbat- the way location fundamentals have influ- ing congestion forces and limiting the bene- enced the location of population and economic fits of agglomeration (Davis and Henderson activity in the LAC region and in the rest of 2003; Ades and Glaeser 1994; Krugman the ­world. and Elizondo ­ 1996). But most analysts use Focusing on national outcomes, the official (national) measures of urbanization chapter analyzes the relationship between a to support these findings, and these mea- country’s level of urbanization, as mea- sures are not consistent across ­countries. For sured by the share of its population that example, some countries use a minimum lives in urban areas, and its overall produc- population size to identify urban areas, but tivity, as measured by gross domestic prod- others do not use any explicitly stated uct (GDP) per ­ c apita. Drawing on, and The second section of this chapter, on the origins of cities, draws heavily on Maloney and Caicedo (2016) and on Henderson ­ l. (2017) produced for this book. et ­al. ­(2016). The third and fourth sections draw heavily on a background paper by Roberts et a Angelica Maria Sanchez Diaz also contributed to this ­ chapter. 27 28   RAISING THE BAR extending, a background study by Roberts The Origins of Cities in Latin et ­a l. (2017) undertaken for this book, it America and the Caribbean aims to establish how LAC cities are per- for m i ng i n relat ion to i nter n at iona l The Origins and Persistence of Cities: ­ benchmarks. The analysis first uses official Location Fundamentals and Historical measures of urbanization compiled by the “Accidents” United Nations in its World Urbanization The locations where cities emerge often have Prospects (WUP) database , which are not underlying natural advantages favorable to consistent across ­ c ountries. It then pro- ­ production. These natural advantages are poses an alternative method, the cluster usually referred to as location fundamentals algorithm, 2 which allows constructing and might include a favorable coastal loca- urbanization measures that are consistent tion or access to potentially navigable water- across ­countries. ways (or both), the presence of favorable This chapter has two main findings. terrain and climatic conditions, a (relative) First, unique historical features influence lack of vulnerability to natural disasters, and the location of cities in the LAC region the (relative) absence of disease ­ vectors. today. Compared with the rest of the world, Because of historical “accidents,” however, location fundamentals (natural advantages) cities can also emerge in locations that lack are not as relevant in determining where strong fundamentals (Arthur 1994; Krugman people and economic activity ­ concentrate. 1991a, ­ 1991b). Cities can, for example, be Globally, such f u ndamentals explain founded in certain places for administrative 57 percent of the variation in the location of purposes or to exert military control over a population, but this falls to 39 percent in ­territory. the LAC region. In addition, agriculture Furthermore, historical evidence from fundamentals (such as having a fertile hin- around the world suggests that, once a city is terland) better explain the concentration of founded, it tends to persist through time (Wahl population in the LAC region than trade 2016; Diamond 1997; Davis and Weinstein fundamentals (such as being close to the 2002; Olsson and Hibbs 2005; Comin, c oast). This might, however, be due to ­ Easterly, and Gong 2010; Spolaore and the inertia of density: once a city is created, Wacziarg 2013; Maloney and Caicedo ­ 2016). it tends to persist in time (Henderson et ­ al. City persistence can result from the perma- 2016), and many LAC countries urbanized ­ nence of strong location fundamentals, such as before the fall in global transport costs hinterland. Cities can also persist, a still-fertile ­ (when trade fundamentals were not as however, in locations that have lost their fun- ­relevant). 3 Many LAC cities might therefore damentals (perhaps because of the depletion of be in suboptimal locations, based on loca- nearby natural resources) or where the funda- tion ­f undamentals. mentals that originally underpinned develop- Second, LAC clearly is highly urbanized, ment are no longer ­ relevant. For example, independent of the measures (official or con- because of the increasing returns of agglomer- sistent) used to define u ­ rbanization. Using ation, a small settlement that emerged in an consistent urbanization measures, GDP per isolated and arid region for administrative or, capita in LAC countries no longer appears originally, resource-extraction purposes can urbanization. low for given prevailing levels of ­ be “locked in” to that location and grow to be But substantial room for improvement a large ­ city. Such “inertia of density” can lead ­ remains. The largest LAC economies are even cities to persist even when they are in places seeing their distance from the North whose location fundamentals are poor or no American productivity frontier widen, possi- longer relevant. Cities’ persistence in such bly because of below-average productivity places has long-term consequences for eco- gains from the LAC region’s structural nomic performance and overall productivity ­transformation. (Michaels and Rauch ­ 2013). U rbani z ati o n , E c o n o m i c D e v e l o p m ent , and S tru c tura l T rans f o r m ati o n   29 The Historical Origins and Persistence of fundamentals explains 43 percent of the vari- LAC Cities ation of subnational precolonial population Myriad factors influenced the location and densities in the LAC region.5 The remaining consolidation of LAC cities before, during, variation could be explained by unmeasured and after colonial ­ times. The emergence of location fundamentals (such as some natural precolonial settlements in the LAC region resources) or by historical “accidents” of can be traced to the adoption of sedentary the ­time. Some legends even suggest, as with agriculture, a shift from nomadic subsis- the origins of Tenochtitlan (today’s Mexico tence, and technological progress (Diamond City), that arbitrary forces might have played 1997; Bairoch and Braider ­ 1988).4 Driven by a role in the emergence of some precolonial these factors, early settlements located in ­settlements.6 places that had underlying natural advan- During colonial times, cities often consoli- ­ odies of water, fertile tages, including major b dated on top of existing s ­ ettlements. This land, and terrain configurations that pro- allowed European colonizers to benefit from vided natural defense from hostile tribes local labor and the existing infrastructure— (Saavedra-Chanduvi and Sennehauser ­ 2009). as well as, often, strong location fundamen- Following the same approach as Maloney tals (Maloney and Caicedo ­ 2 016). Some and Caicedo (2016)—described in box ­ 1.1— colonial cities, however, were created from we find that a small set of location scratch in places that the colonists deemed BOX ­1.1  Precolonial Densities, Location Fundamentals, and the Persistence of Subnational Population Densities Maloney and Caicedo (2016) study the persistence • Current population density refers to the total pop- of subnational population density in the LAC ulation in the year 2000 divided by the area of the region and North America using an innovative set state or province in square ­ kilometers. of historical data that include subnational data on • Location fundamentals include an index for cul- precolonial population density, recent population tivable land, the density of rivers as a share of ­ undamentals. As part density, and a set of location f area, temperature, altitude, rainfall, presence of of their study, they assess the role of location fun- malaria, terrain ruggedness, and distance to near- damentals in determining precolonial population est ­coast. densities, whether there are strong indications of • The units of observation are subnational admin- persistence—by comparing precolonial and recent istrative units that correspond to provinces, population densities in the region—and the role departments, or states depending on the ­ country. that extractive institutions (such as slavery) play in The sample includes 17 LAC ­ countries. Histor- reducing ­persistence. We use the same approach and ical information on precolonial densities is not data set as Maloney and Caicedo (2016) but restrict available at a lower scale (such as city ­ level). For the sample to cover only L­ AC.a Here is a description the purpose of our analysis, we run the following of the main variables used for the analysis: regressions: (i) the log of precolonial density on • Precolonial population density is the number of a set of location fundamentals and (ii) the log of indigenous people before the arrival of Colum- current population density on the log of precolo- bus divided by the area of the state or province in ­ undamentals nial density, with a set of locational f square ­kilometers. used as ­controls. ­ nalysis. a. The authors are grateful to Maloney and Caicedo for sharing their data and code to facilitate this a 30   RAISING THE BAR strategic for military or administrative rea- we find that location fundamentals explain sons (Saavedra-Chanduvi and Sennehauser 40 percent of the variation in subnational ­ 2009). Colonial cities were also often founded population densities in the region in ­ 2000. with the aim of facilitating commerce with The existence of precolonial settlements, ­ egion).7 Europe (not with cities in the r which captures persistence not explained by Although there are more recent examples location fundamentals, explains 11 percent of cities that have been created from scratch of the ­ variation. Furthermore, we find that in the region (such as Brasilia), the origins subnational precolonial densities in the of most of the LAC region’s present-day cities region are positively and significantly cor- can be traced back to precolonial or colonial related with subnational population density ­times.8 The reasons for this persistence may in 2000, even when controlling for location city. For some, it may reflect vary from city to ­ ­ fundamentals. This confirms that one of the persistence or strengthening of location central findings from Maloney and Caicedo ­f undamentals.9 For others, it likely stems (2016), the strong persistence of subnational from “lock-in” effects, as cities continue to population density, is maintained when prosper thanks to the accumulated returns to restricting the analysis to LAC countries agglomeration, even in locations that never (figure ­1.1).10 possessed, or have lost, good location The analysis of a second data set c ­ ompiled f undamentals. It is difficult to estimate the ­ by Henderson et ­ al. (2016; see box ­1.2), also relative importance of each of these factors or, suggests that there might be unique features more important, to assess the economic impli- that explain the location of economic and cations of this persistence; but two recent population density in the LAC region ­ today. studies and the data sets developed by their Current economic activity in the LAC region authors can provide us with some ­ insights. is less likely to be in places with strong loca- Using the same data and approach that tion fundamentals than it is in the rest of the Maloney and Caicedo (2016) used, but ­ world. The location of economic activity in restricting the analysis to LAC countries, the LAC region is also explained to a greater extent by agriculture location fundamentals than by trade location fundamentals, partly FIGURE 1.1  Strong Persistence in Subnational Population because of the region’s early urbanization Densities in the LAC Region process (before the fall of transport ­ costs). It is therefore possible that, if the LAC region’s 5 urban systems were to emerge today, many Population per square kilometer, 2000 (log) cities would be in different ­ places. These results do not, however, mean that some cit- ies in the region should be ­ relocated—rather 0 that policy makers should focus their efforts on maximizing cities’ ­ p roductivity given their ­location. −5 Urbanization in the LAC Region and the Rest of the World: Discrepancies between −10 Consistent and Official Measures −10 −5 0 Precolonial population per square kilometer (log) In this section, we explore how the share of the population that lives in urban areas Source: Estimated using the same database and approach as Maloney and Caicedo (2016), but (the urban population share) in the LAC ­ ample. restricting the sample to the 17 LAC countries included in their s Note: The units of observation are subnational units (provinces, departments, or s ­ tates). LAC = Latin region compares with that in the rest of the America and the Caribbean. world by examining some stylized facts, U rbani z ati o n , E c o n o m i c D e v e l o p m ent , and S tru c tura l T rans f o r m ati o n   31 BOX ­1.2  Location Fundamentals and the Distribution of Economic Activity in Latin America and the Caribbean versus the Rest of the World Henderson et ­ al. (2016) seek to determine how much Three results stand o ­ ut. First, there are subtle of the present spatial distribution of economic activity, differences in the features that influence the loca- at the global scale, is explained by what they call base, tion of economic activit y in the L AC region agriculture, and trade location ­ f undamentals. Base compared with the rest of the w ­ orld. As noted ear- fundamentals include elements such as malaria and lier, base, agriculture, and trade fundamentals, terrain ruggedness; agriculture fundamentals include although still statistically significant, explain only elements related to agriculture viability, and trade 39 percent of the variation in the location of eco- fundamentals include variables focusing on access to nomic activity in the LAC region but 57 percent water ­transport. They construct a global gridded sam- ­ worldwide. Second, agriculture fundamentals are ple of 250,000 cells (of about 1 km 2 at the ­ equator). better than trade fundamentals at explaining the The location of economic activity is proxied by night- location of economic a ­ ctivity. As described by time lights (specifically, by the log of nighttime light Henderson et ­ a l. (2016), however, this is expected radiance in ­ 2010). Using the R 2 statistic, they find to apply to countries that urbanized when trans- that base, agriculture, and trade fundamentals alone port costs were still high—the case of most LAC (see table ­B1.2.1 note), explain 57 percent of the ­ c ountries. In these countries, cities often located variation in the location of economic activity world- (and persist) in agricultural regions; however, in wide (column a, row 4, in table B ­ 1.2.1). We expand countries that urbanized when transport costs Henderson et ­ al.’s analysis to explore whether there are fell, cities often located in places with favorable differences in the factors driving population and eco- trade fundamentals (such as near the c ­ oast). nomic concentration in the LAC region (column ­ b). We Third, in the LAC region, South America is the also conduct a subregional analysis for the Caribbean, subregion where location fundamentals provide Central America, and South America (the last includes the least explanation for location of economic Mexico; columns c, d, and ­ e, respectively.). ­activity. TABLE ­B1.2.1  R-Squared Results for Relationship between Log(Radiance-Calibrated Nighttime Lights) and Base, Agriculture, and Trade Fundamentals (a) (b) (c) (d) (e) All LAC Caribbean Central America South America (1) Base + FE ­0.355 ­0.222 ­0.275 ­0.221 ­0.200 (2) Agriculture + base + FE ­0.566 ­0.343 ­0.381 ­0.386 ­0.328 (3) Trade + base + FE ­0.369 ­0.283 ­0.376 ­0.256 ­0.264 (4) Base + Agriculture + trade + FE ­0.568 ­0.386 ­0.433 ­0.456 ­0.373 Source: Column a presents results by Henderson et ­al. (2016); columns b through e show World Bank ­calculations. Note: The table presents the estimated R2 of different regressions where the dependent variable is the log (natural) of the radiance-calibrated nighttime lights; the independent variables are combinations of base, agriculture, and trade ­fundamentals. The unit of observation is the individual grid ­cell. Fixed effects (FE) are country fixed ­effects. Base location fundamentals include malaria and ruggedness; agriculture and trade covariates include 14 biome indicators (for agriculture) and five trade variables that focus on access to water ­transport. LAC = Latin America and the Caribbean. for which the choice of urbanization national definitions of urban areas, and o f f ic i a l or ot her w i s e — i s m e a su re s — ­ countries. Drawing on the thus vary across ­ ­ critical. We begin by showing these facts as background paper by Roberts et ­ a l. (2017), documented by official measures compiled we then revisit these facts by using urban- by t h e U n it e d N at io n s i n it s W U P ization measures that are consistent across ­database.11 These official measures rely on ­countries. 32   RAISING THE BAR On Official Measures, LAC Is Highly of the population living in urban a ­ reas). Urbanized Relative to Other Regions Among LAC subregions, South America By 1960, according to official urbanization (which we define in this chapter to include measures, half of the LAC region’s population Mexico) urbanized earlier and closely lived in urban areas, a milestone achieved glob- mirrored regional urbanization ­ t rends.12 ally only in 2008 (figure ­ 1.2). By 2015, more Countries in the Caribbean were predomi- than 80 percent of the region’s population lived nantly rural in the 1960s but underwent rapid in urban areas, making it the most urbanized urbanization and reached the 50 percent developing region and giving it urban popula- watershed by the late 1970s, well before their America. tion shares similar to those in North ­ counterparts in Central America, which The LAC region shows considerable reached it only in the late 1990s (figure ­ 1.3). heterogeneity in urban population ­ shares. Since 1960, the rate of growth of the Some of its countries, like Argentina and region’s urban population has been declining Uruguay, have very high levels (with more from ­4.49 percent annually between 1960 than 90 percent of the population living in and 1965 to ­ 1.61 percent between 2005 and officially defined urban a ­ reas). Others, such 2010 (figure 1­ .4).13 The decline in the urban as Antigua and Barbuda and Guyana, remain population growth rate is similar to that predominantly rural (with 30 percent or less observed in countries elsewhere in the world FIGURE ­1.2  Urban Shares for Latin America and the Caribbean and Other World Regions, 1960–2015 100 90 80 70 60 Urban share (%) 50 40 30 20 10 0 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 Latin America and the Caribbean East Asia and Pacific Europe and Central Asia Middle East and North Africa North America South Asia Sub-Saharan Africa World Source: Calculations based on World Development Indicators, July 2017 (https://data.worldbank.org/data-catalog/world-development-indicators), derived from World Urbanization ­Prospects. Note: On the y axis, the urban share is the total urban population, as defined by national statistics offices, as a percentage of the total population in each ­region. U rbani z ati o n , E c o n o m i c D e v e l o p m ent , and S tru c tura l T rans f o r m ati o n   33 FIGURE ­1.3  Urban Shares for LAC Subregions, 1960–2015 100 90 80 Urban share (%) 70 60 50 40 30 20 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 South America central America caribbean latin America and the caribbean Source: Calculations based on World Development Indicators, July 2017 (­https://data.worldbank​.org/data-catalog/world-development-indicators), derived from World Urbanization ­Prospects. Note: On the y axis, the urban share is the total urban population, as defined by national statistics offices, as a percentage of the total population in each ­region. LAC = Latin America and the ­Caribbean. FIGURE ­1.4  Annual Growth of Urban Population, Worldwide and by Region, 1960–2005 6 5 Growth in urban population (%) 4 3 2 1 0 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 Latin America and the Caribbean East Asia and Pacific Europe and Central Asia Middle East and North Africa North America South Asia Sub-Saharan Africa World Source: Calculations based on World Urbanization Prospects, July 2017 (­https://esa.un.org/unpd/wup/CD-ROM/). ­ ffices. Growth is calculated as the Note: The figure presents regional annual growth rates of the urban population, as defined by national statistics o compound annual growth rate between year x and year x+5 (World Urbanization Prospects estimates are available at five-year ­intervals). 34   RAISING THE BAR when they attain high urban s ­hares.14 average, LAC countries include smaller settle- However, the absolute size of urban popula- ments in their urban population ­ f igures. tion growth in the region remains large: These variations are problematic for cross- between 1960 and 1987, 181 million people and intracountry comparisons, limiting the were added to LAC cities, and slightly validity of comparing official measures of more—216 million—were added during the urbanization in LAC with those in the rest of subsequent 28 y ­ ears. As a reference, this is the ­world. greater than the entire population of Brazil The growing recognition of such prob- (207 million in 2016), the fifth most popu- lems has spawned methods that aim to world. lous country in the ­ establish consistent definitions of urban areas across countries, and so consistent measures of ­ urbanization. For example, the Official Measures of Urbanization Are Agglomeration Index (AI) originally devel- Problematic for Conducting Cross- oped by Uchida and Nelson (2008) and the Country Comparisons cluster algorithm developed by Dijkstra and Most published research on urbanization Poelman (2014) can be used, together with and development relies on official measures recently available global, gridded population of urbanization based on national definitions data, to produce consistent estimates of of urban areas (World Bank 2008; Chen et ­ al. urban areas for a large global cross-section 2014; Spence, Annez, and Buckley ­ 2009). of countries (box ­ 1.3). In their background As highlighted by Roberts et ­ al. (2017), these paper for this book, Roberts et ­ a l. (2017) definitions vary widely by c ­ ountry. From the implement both the cluster algorithm and 232 countries and territories included in the the AI using three different sources of glob- WUP database, 133 use one or more of four ally gridded population ­ data.17 In doing so, basic types of criteria to define their urban they show that the two algorithms generate a reas. By far the most common is a mini- ­ similar maps of urban areas, both for LAC mu m popu lation si ze th reshold (103 and ­globally. For consistency with chapter 2, ­ countries). For some countries, the criterion we focus mainly on results using the cluster consists of the availability of certain types of algorithm as applied to Landscan 2012 infrastructure (such as schools or piped globally gridded population ­ data. water), structure of the local economy, or a To see how different our urban area popu- minimum population density ­ t hreshold. lation estimates are with those produced Finally, a large number of countries (99) do by national definitions, we compared data not use any explicitly stated criteria, simply for a global sample of cities from the WUP listing urban areas by name or stating a (box ­1.4). designation of administrative units that con- ­ stitute ­cities.15 On Consistent Measures, Urbanization in In general, non-LAC countries have more LAC Is Closer to That in Other Regions stringent criteria than LAC countries for des- ignating areas as urban (World Bank 2008; Using the cluster algorithm, we revisited Ellis and Roberts 2015), with the upshot the cross-regional comparison of urbaniza- that, among the countries that use the mini- tion levels conducted using official measures mum population threshold to define urban of ­u rbanization.18 We found that urbaniza- areas, the mean threshold stands at about tion levels in the LAC region, using consis- 2,000 people for LAC but about 5,000 peo- tent measures, are lower than those estimated ple ­g lobally. This mean difference is not using official definitions but remain high, at driven by the small island nations of the about 73 ­ percent in 2012 (­ 1.5).19 We figure ­ Caribbean because that subregion’s mean also found that, when we use consistent population threshold is almost identical to urbanization measures, the LAC region’s level that for the LAC r ­egion.16 Instead, on of urbanization (as measured by its urban U rbani z ati o n , E c o n o m i c D e v e l o p m ent , and S tru c tura l T rans f o r m ati o n   35 FIGURE ­1.5  Urban Shares: Official versus Consistent Measures of Urbanization 100 90 Share of population in urban areas (%) 80 70 60 50 40 30 20 10 0 World Urbanization Prospects Cluster algorithm Agglomeration Index Latin America and the Caribbean North America Europe and Central Asia Middle East and North Africa East Asia and Pacific World Sub-Saharan Africa South Asia Source: Calculations based on Roberts et ­al. ­2017. Note: The results correspond to the combination of using the cluster algorithm and the Agglomeration Index with LandScan 2012 gridded population ­data. The background paper by ­ ere. Data are for ­2012. Roberts et ­al. (2017) also presents comparisons of urban shares using GHS–Pop gridded population data, and their results are broadly consistent with those h BOX ­1.3   The Agglomeration Index and the Cluster Algorithm The Agglomeration Index ­ ( A I). The A I algo- To implement the AI, we used estimated, not rithm defines urban areas from a labor market actual, commuting times (used by OECD) or min- ­ p erspective. It is similar to that proposed by the imum commuting thresholds (used by Duranton), Organisation for Economic Co-operation and because they are available ­ g lobally. Following Development (OECD) (2012), and the one used Uchida and Nelson (2008) and World Bank (2008), by Duranton ( ­2015). AI algorithms define func- an area in the commuting shed is defined as urban if tional urban areas as the spatial extent of the labor it has a population density of at least 150 people per ­ m arket. Approaches to delineating these areas typ- square kilometer and is located within a 60-minute ically involve identifying an “urban core” and a travel time radius of a settlement, which itself has a “commuting shed” around ­ it. population of at least ­ 50,000. Our AI algorithm starts from a database of set- The cluster ­ algorithm. This algorithm adopts a tlements and identifies those that are “sizable” on spatial-demographic approach to identifying urban the basis a population ­ threshold. Around each urban ­ areas. More specifically, it classifies cells in a population core, the algorithm then identifies the commut- grid according to their density, and then groups them ing shed as the areas located within a given radius into “urban ­ clusters.” A spatially contiguous group of (defined by travel time) of the urban core and meet- grid cells is classified as constituting an urban cluster ing a population density ­ t hreshold. Both the urban if each of these cells has a population density of at least core and the commuting shed are thus considered 300 people per square kilometer and if the aggregate urban ­areas. A country’s urban share is then defined population of the cells exceeds 5,000 ­ inhabitants. A as the share of the overall national population living country’s urban share is then defined as the share of the in such urban ­areas. overall national population living in “urban ­ clusters.” Note: For more details on these two methodologies and their implementation for this book, see Roberts et ­al. ­(2017). 36   RAISING THE BAR BOX ­1.4  Comparing the Population of Urban Areas: Cluster Algorithm versus Official Data We compared estimates of city population data pro- than 5 percent) were matched with more than one WUP duced using the cluster algorithm with figures from the c ­ ity. We compared the cluster algorithm city population World Urbanization Prospects (WUP) report “World with the largest WUP city matched, and with the total Urbanization Prospects: The 2014 Revision” for a global ­ population when several WUP cities matched with a sample of c­ ities. The WUP’s figures provide the most city. The results were similar in single cluster algorithm ­ up-to-date estimates of population for the largest cities both ­ cases. We mainly focused on the first c ­ omparison. in the world (cities with more than 300,000 inhabitants In comparing population, we proxied the degree of in 2014) and are based on city populations reported discrepancy using the “relocation f ­ raction.”b by national statistics o­ ffices. We used the city location What do we learn from this comparison? For the points provided by the WUP and matched them with largest cities in the world, our cluster algorithm– algorithm. the location of cities identified by the cluster ­ estimated population sizes are not very different We matched virtually all cities in the WUP data set from those estimated by national statistics o ­ ffices. with our urban areas (the sample for this comparison The global correlation between both populations is formed by 1,301 urban areas covering 1,484 cities in is 80 percent and increases to 90 percent when the WUP data set).a A small number of urban areas (less we include the population of all matched ­ c ities. FIGURE B1.4.1  Comparison of Cluster Algorithm and WUP City Population Values, by Region a. Caribbean b. Central America c. East Asia and Pacific 18 16 14 12 10 Cluster methodology population (log) d. Europe and Central Asia e. Middle East and North Africa f. North America 18 16 14 12 10 g. South America h. South Asia i. Sub-Saharan Africa 18 16 14 12 10 12 14 16 18 12 14 16 18 12 14 16 18 WUP population (log) Source: Cluster methodology populations are based on urban areas defined using the cluster algorithm of Dijkstra and Poelman (2014), as applied to LandScan 2012 gridded population ­data. WUP population figures are from the World Urbanization Prospects: The 2014 ­Revision. Note: We calculate the WUP population in 2012 using a linear interpolation of the WUP data in 2010 and ­2015. The orange line is a 45-degree line, thus the closer to the line each point is, the lower the difference between the two sources of ­population data. (continued) U rbani z ati o n , E c o n o m i c D e v e l o p m ent , and S tru c tura l T rans f o r m ati o n   37 BOX ­1.4  Comparing the Population of Urban Areas: Cluster Algorithm versus Official Data (continued) Moreover, the relocation fraction is low: on average, the estimation of the population within ­ countries. only 12 percent of the population in cities would We find the contrary when we compute, per region, have to be relocated to equalize the populations in the variation in the relocation fraction between both data ­ sets. and within ­ countries. For example, the coefficient Regional ­variations. As shown by figure ­ B1.4.1, of variation between countries is lower than that there are, however, important regional differences within countries in East Asia and the Pacific, Europe in the disagreement between the two measures, and Central Asia, and South ­ A merica. underscoring the different criteria countries have This means that we would expect to see a larger for defining cities and the need for a globally consis- difference in the disagreement between our calcu- tent ­ measure. The Americas tend to have the highest lation of population and that given by the WUP if agreement of all regions ­ g lobally. For example, the we choose two cities in the same country than if we correlation coefficients for South and North America choose the average disagreement of two different between the populations of cluster algorithm ­ cities countries in those regions (figure B ­ resents ­ 1.4.2 p and their corresponding WUP cities are 99 and 98 the data for all r ­ egions). The W UP compilation percent, respectively; and their average relocation ­ recognizes this discrepancy and classifies each city fractions are as low as 5 and 8 percent, ­ respectively. on three statistical concepts: urban agglomeration, Asian countries have less agreement, with correla- metropolitan area, and “city ­ proper.” For almost tions of 66 and 78 percent for South Asia and East 60 ­p ercent of the countries with more than one Asia and the Pacific, ­respectively. They have reloca- large city, their statistical concepts differ among tion fractions of 19 and 16 percent, ­ respectively. their ­cities. National ­variations. There is also considerable These results show that having a consistent mea- inconsistency in national delimitation of urban sure to define urban areas permits better compari- a reas. If countries defined their own cities consis- ­ sons of cities not only across countries or regions but tently, we would expect to have a constant bias in also within ­ countries. FIGURE ­B1.4.2  Between and within Variation of the Relocation Fraction per Region 140 120 Coefficient of variation (%) 100 80 60 40 20 0 l an ica cif d lA d Af and ica ica sia Af aran Al Pa an ra an be hA er er er ic sia a a ric ric sia nt pe rth ast h Am Am m rib Sa ut hA Ce ro A No E Ca So b- l rth e Eu st ra ut dl Su Ea nt No id So Ce M Between Within Source: Calculations based on urban areas defined using the cluster algorithm of Dijkstra and Poelman (2014), as applied to LandScan 2012 gridded population ­data. WUP population figures are from the World Urbanization Prospects: The 2014 ­Revision. ­ ample. Note: The figure shows the coefficient of variation of the relocation fraction between and within countries for each region and for the whole s The relocation fraction is calculated for each of the 1,301 cities ­matched. a. As in the “Defining a Global Data Set of Urban Areas” section of chapter 2, this sample excludes nine implausibly large urban extents formed by the cluster ­algorithm. ­ l. (2011), the relocation fraction Ri is defined as the fraction of population that needs to be relocated from a cluster city Siclus to the b. In similar fashion to Rozenfeld et a WUP SiWUP − Siclus corresponding WUP city Si (or vice versa), so that their populations are equalized: Ri def max(SiWUP , Siclus ) 38   RAISING THE BAR population share) is closer to that of the rest viewed in the literature as bound ­ together. of the w­ orld. In fact, whereas the overall Consistent with such findings, an analysis of estimated urban share does not change much cross-country data exhibits a strong and sig- for LAC, it does for other regions, notably nificant positive correlation between a coun- the Middle East and North Africa and South try’s development (measured by GDP per Asia (SA), where urbanization levels on con- capita) and its urban share (on official mea- sistent measures are much higher than on sures) (figure ­ 1.6). This correlation is often official m ­ easures. Thus, using consistent attributed to the structural change, associ- measures, urban shares in the LAC region ated with the movement of labor out of agri- are more aligned with those of other ­regions. culture and into manufacturing and services, In the LAC region, Central America and that accompanies urbanization (Henderson the Caribbean’s urban shares do not change 2003; Chenery and Taylor 1 ­ 968). The esti- ­much.20 However, we find changes in South mated elasticity between productivity (as America where urban shares seem overesti- measured by GDP per capita) and the share mated by official ­measures. of urbanization is so high, however, that it is questionable whether the relationship can Urbanization, Economic be at tributed solely to struct u ral Development, and Structural ­transformation.21 Transformation: How Does the LAC Similar to what is observed in the rest of Region's Performance Stack Up? the world, there is a strong and positive correlation between urban shares in the On Conventional Measures, LAC LAC region (on official measures) and Subregions Systematically Depart from GDP per c ­ apita. But is the relationship the Global Relationship between between the urban share and economic Urbanization and Economic development systematically different from Development the global one? Urbanization, economic development, and To find out, we regressed GDP per capita structural transformation have long been values (purchasing power parity, or PPP) on the urban share defined using official mea- FIGURE ­1.6  Relationship between Economic Development and the sures (table ­ 1.1). We find that LAC countries Urban Share on Official Measures are, on average, not behaving differently from 12 those in the rest of the world (­ column b in table ­1.1). However, there appear to be sys- tematic departures at the subregional level GDP per capita, PPP, 2012 (log) (column ­ c). Countries in South America 10 appear to have lower GDP per capita than one would predict from their urban ­ share. On the contrary, countries in the Caribbean appear to have higher GDP per capita than predicted by 8 their urbanization ­ levels. Countries in Central America present no systematic departure from the global ­relationship. 6 Although these results suggest that 0 20 40 60 80 100 South American cities are “underproductive” Urban share, World Development Indicators 2012 (and Caribbean cities are “overproductive”) Rest of the world Central America 95% confidence interval compared with the rest of the world, they South America Caribbean were obtained from official measures and so may simply reflect data artifacts arising from Source: Calculations based on World Urbanization Prospects data and World Development Indicators. the inconsistent definition of urban areas Note: Data are for ­2012. GDP = gross domestic product; PPP = purchasing power ­parity. across ­ countries. To examine this notion, U rbani z ati o n , E c o n o m i c D e v e l o p m ent , and S tru c tura l T rans f o r m ati o n   39 TABLE ­1.1  Regression Results for Relationship between Log(GDP per Capita) and the Official Urban Share   (a) (b) (c) (d) Urban population share ­4.200*** ­4.242*** ­4.346*** ­3.067*** ­(0.284) ­(0.288) ­(0.284) ­(0.299) Latin America and the Caribbean ­− 0.153 ­− 0.868* ­(0.166) ­(0.457) Caribbean ­0.603* ­(0.305) Central America ­− 0.274 ­(0.284) South America ­− 0.480** ­(0.226) Sub-Saharan Africa ­−1.590*** ­(0.465) South Asia ­− 0.821 ­(0.520) Middle East and North Africa ­− 0.610 ­(0.479) Europe and Central Asia ­− 0.338 ­(0.449) East Asia and Pacific ­− 0.614 ­(0.468) Constant ­6.769*** ­6.772*** ­6.714*** ­8.245*** ­(0.173) ­(0.173) ­(0.171) ­(0.500) No. of observations 146 146 146 146 R2 ­ .603 0 ­0.606 ­0.629 ­0.740 Source: Calculations using World Development I­ndicators. Note: The table shows the results of a regression at the country level where the dependent variable is the GDP per capita PPP in log, and the independent variable is the urbanization share (0–1) in column a and the urban share and regional dummies in columns ­b through d. Data are for ­2012. In columns ­ orld. In column d, the base category is North ­America. Standard errors are in ­parentheses. A white test for b and c, the base category is the rest of the w heteroscedasticity shows that we do not reject the null (homoscedasticity) at α = ­0.10. We use normal standard e ­ rrors. GDP = gross domestic product; PPP = purchasing power parity. ***p < 0.01. **p < 0.05. *p < 0.1. we revisited the analysis, using consistent official urbanization measures, we conducted urbanization. measures of ­ a second set of regressions using consistent ­ measures. Results can be found in table ­ 1.2.22 To start, we found that the statistical signif- On Consistent Measures, LAC icance in the global relationship between Subregions Perform Economically urbanization and economic development is as Predicted by the Global Relationship maintained—a country’s development level is, but Could Do Better to a significant degree, positively correlated To establish whether the systematic departure with its urban share (column a in table ­ 1.2). of LAC subregions stems from the use of However, the fit of the relationship between 40   RAISING THE BAR urbanization and economic development is not In addition, and unlike when we used as strong (R 2 values are much lower) as when official urbanization measures, we find that using official measures, and the estimated neither LAC (see column b in tables 1 ­ .1 and coefficient on the urban population share is 1.2) nor its subregions (see column c in ­ smaller (compare results in column a of table 1.1 and ­ tables ­ 1.2) appear to depart from the 1.2 with those in column a of table 1.1). ­ global relationship when we use consistent TABLE ­1.2  Regression Results for Relationship between Log(GDP per Capita) and the Urban Share, Using Consistent Measures   (a) (b) (c) (d) Urban population share ­3.743*** ­3.771*** ­3.752*** ­2.195*** ­(0.459) ­(0.468) ­(0.475) ­(0.435) Latin America and the Caribbean ­− 0.074 ­−1.235** ­(0.220) ­(0.555) Caribbean ­0.046 ­(0.420) Central America ­− 0.146 ­(0.385) South America ­− 0.088 ­(0.303) Sub-Saharan Africa ­−2.223*** ­(0.561) South Asia ­−2.291*** ­(0.605) Middle East and North Africa ­− 0.937 ­(0.585) Europe and Central Asia ­− 0.559 ­(0.547) East Asia and Pacific ­−1.119* ­(0.567) Constant ­6.861*** ­6.857*** ­6.868*** ­9.094*** ­(0.294) ­(0.295) ­(0.299) ­(0.625) No. of observations 146 146 146 146 2 R ­0.316 ­0.316 ­0.317 ­0.613 Source: Calculations based on World Development Indicators and urban population share based on urban areas defined using the cluster algorithm of Dijkstra and Poelman (2014), as applied to Landscan 2012 gridded population ­data. Note: The table shows the results of regressions at the country level where the dependent variable is the GDP per capita PPP in log (ln), and the independent variable is the urban share (0–1) in column a and the urban share and regional dummies in columns ­b through d. Data correspond to year ­2012. In columns b and c, the base category is the rest of the ­world. In column d, the base category is North ­America. Standard errors are in ­parentheses. A white test for heteroskedasticity shows that we do not reject the null (homoscedasticity) at alpha = ­0.10. We use normal standard ­errors. GDP = gross domestic product; PPP = purchasing power parity. ***p < 0.01. **p < 0.05. *p < 0.1. U rbani z ati o n , E c o n o m i c D e v e l o p m ent , and S tru c tura l T rans f o r m ati o n   41 urbanization ­ measures. South America as a between GDP per capita and urbanization group no longer appears to be “underpro- ­levels. ductive” and the Caribbean does not appear To examine structural transformation to be “overproductive” for their urbaniza- trends in the LAC region against those in the tion ­ levels. This finding suggests that the rest of the world, we relied on data from the narrative of underperforming cities in South Groningen Growth and Development Center America might be partly due, as suspected, ­(GGDC).25 These data have been widely used to data ­ artifacts. in economic literature for analyzing long- Having established that neither LAC nor term trends in the reallocation of labor and in its subregions differ from the global rela- output, and include data for 9 of the 33 LAC tionship between urbanization and eco- countries (including its four largest econo- nomic development, the task is to determine mies) and for subsets of countries around where LAC falls relative to other ­ regions. It the ­world. 26 For simplicity, we refer to the is quite possible that the global relationship nine-country LAC subset as L ­ AC-9. We also between urbanization and economic devel- use Asia and North America for the subsets opment is being “pushed down” by under- of countries from those two ­ regions. performing countries in some ­ regions. 23 We For LAC-9, we find that the reallocation of find that the LAC region seems to be per- people from rural to urban areas was accom- forming worse than North America (see col- panied by a reallocation of labor from agri- umn d of table 1 ­ .2, which compares all culture to manufacturing and services regions of the world; North America is (figure ­ 1.7, panel b), which tend to cluster in the omitted ­ c ategory). Similar results are or around ­ cities. For example, between 1960 obtained with official urbanization mea- and 2009 the share of the labor force in sures (see column d in table ­ 1.1). LAC-9 in agriculture fell from 47 to One hypothesis to explain the LAC 15 ­percent. Simultaneously, the share of the region’s underperformance relative to the labor force in services roughly doubled from global productivity frontier (consisting of 32 to 64 ­ percent. Few such dramatic changes North America in this case) is that the were seen in industry, which on average productivity gains expected from the reallo- absorbed 23 percent of the labor force over cation of labor from agriculture to manufac- the p­ eriod. A similar trend to that in LAC-9 is turing and services areas as workers migrated seen in Asia, which was also urbanizing in from rural to urban areas have not material- this period (see figure ­ a). 1.7, panel ­ ized in the LAC region (or at least not to the The large reallocation of labor in LAC-9 same ex tent as in today’s developed was, however, not coupled with a large shift ­ countries). The following subsection exam- in the composition of output, as measured by ines whether there is evidence to substantiate national value added (figure ­ e). 1.7, panel ­ this ­hypothesis. Although the share of people working in ser- vices in LAC-9 went up by 32 percentage points between 1960 and 2009, the value The LAC Region Has Mediocre added share of services in total output Productivity Gains from Structural increased only by 5 percentage points over the Transformation same period, passing from 51 to 56 ­ percent. The reallocation of labor between “rural” Over the same period, the share of output and “urban” sectors has important implica- coming from industry declined marginally tions for overall productivity and economic from 38 to 37 ­ percent. growth (Kuznets 1973; Alvarez-Cuadrado These results suggest that the reallocation and Poschke 2011; Herrendorf, Rogerson, of labor from agriculture to industry and ser- and Valentinyi 2013)24 and is posited to lie, vices in LAC-9 has not produced the expected at least in part, behind the global correlation labor productivity ­ gains. According to our 42   RAISING THE BAR FIGURE ­1.7  Change in the Structural Composition of the Economy, 1960–2009 a. Employment shares, b. Employment shares, c. Employment shares, Asia LAC-9 North America 100 100 100 Percent Percent Percent 50 50 50 0 0 0 1960 1970 1980 1990 2000 2009 1960 1970 1980 1990 2000 2009 1960 1970 1980 1990 2000 2009 d. Value added shares, e. Value added shares, f. Value added shares, Asia LAC-9 North America 100 100 100 Percent Percent Percent 50 50 50 0 0 0 1960 1970 1980 1990 2000 2009 1960 1970 1980 1990 2000 2009 1960 1970 1980 1990 2000 2009 Agriculture Services Industry Source: Calculations using the Groningen Growth and Development Center database. Note: LAC-9 = Argentina, Bolivia, Brazil, Chile, Colombia, Costa Rica, Mexico, Peru, and República Bolivariana de Venezuela. estimates, a 1 percent increase in the services ­ overnment.30 Although this “premature dein- g share of employment is linked to a ­ percent 3.8 ­ dustrialization” is attributed to external global increase in overall labor productivity in the forces, such as globalization and technological rest of the world but only a ­ 0 .7 percent progress (Rodrik 2015), cities and govern- increase in ­L AC-9. 27 A similar difference is ments can take steps to counteract the produc- found in industry, with an increase of tivity consequences of d ­ eindustrialization. 0.9 percent in productivity in LAC-9 against a ­ Cities can, for example, address the congestion 2 .2 percent gain in the rest of the ­ ­ world. forces that may inhibit firms from reaping the These findings, although they should be productivity gains linked to agglomeration treated with caution because of data limita- ­(chapter  2).31 They might also be able to help tions, reflect previous findings from Timmer, develop higher-value-added tradable services de Vries, and de Vries (2015) and Pages (2010) by addressing structural problems (such as who used the same data set,28 and recent find- local institutional and regulatory constraints) ings from Francis, Saliola, and Seker (2013) ­ hem.32 that stifle t time-span.29 based on firm-level data for a shorter ­ The lower than expected productivity gains obtained in LAC countries have widened the Conclusions labor–productivity gap with the productivity Among the multiple factors that shaped the frontier (North ­ A merica). Timmer, de Vries, emergence and persistence of LAC cities, and de Vries (2015) argue that this widening first, location fundamentals seemed to have comes from below-average productivity played less of a role in influencing the loca- growth of services and manufacturing in the tion of population and economic activities LAC region, and from the premature move- than in the rest of the world; and, second, ment of workers in the LAC region from man- agriculture fundamentals played a larger role ufacturing to lower-­ productivity services such than trade fundamentals in putting cities as retailing, wholesaling, construction, and where they are ­ today. These patterns likely U rbani z ati o n , E c o n o m i c D e v e l o p m ent , and S tru c tura l T rans f o r m ati o n   43 stem from a mix of historical “accidents” and 4. Around 1800 BCE, the growth of farming the fact that countries in the region urban- s ettlements—coupled with technological ­ ized before the fall of transport ­ costs. progress in modern-day Bolivia, Mexico, and On the links between urbanization and Peru—sustained the critical mass for later development of cities in Latin America (Haas, productivity, comparing LAC with the rest of Pozorski, and Pozorski ­ 1987). the world, we find that, using consistent mea- 5. The R2 goes up to 72 percent if one introduces sures of urbanization, the region’s productiv- country fixed effects, although it is not clear ity is within global expectations for its urban what these “fixed effects” might be capturing ­ share. Yet we also see substantial room for because the countries were not yet f ­ormed. improvement because the region’s gap with They might be capturing differences in popu- widening. the global productivity frontier is ­ lation density across the areas in the continent Chapter 2 shifts the focus to LAC cities because some were sparsely populated and (and other regions) as the unit of analysis, others were not, or the differences could be examining dimensions of urbanization (the due to specific characteristics of indigenous density of urban areas, the prevalence of mul- civilizations that settled in the different areas of what became the LAC ­ region. ticity agglomerations, and urban primacy) 6. According to legend, Huitzilopochtli, the god beyond those captured by a country’s urban of war, the sun, and human sacrifice, directed population share; the relationship between the Mexican people to settle on the island in these dimensions and national productivity the middle of lake T ­ excoco. He “ordered his performance; and productivity performance priests to look for the prickly pear cactus and at the individual urban area ­level. These first build a temple in his ­honor. They followed the two chapters provide a macrolevel founda- order and found the place on an island in the tion for the remaining chapters, which use middle of the lake” (de Rojas ­ 2012). microlevel data to analyze how the factors 7. For example, when Hernán Cortés arrived in and trends they discuss affect workers and 1519 searching for gold in what is now industries, transport infrastructures, human Mexico, he landed in what he named Villa Rica de la Vera Cruz (modern-day ­ Veracruz). capital, and the spatial form of ­ cities. He eventually declared the site a city to estab- lish his legitimacy and use it as a point from which to stage attacks against Montezuma, Notes the leader of Tenochtitlan (Saavedra-Chanduvi 1. Henderson et ­ al. (2016) classify three natural and Sennehauser 2009, 37; The Economist geographic endowments or location funda- ­2014). mentals: (i) base fundamentals (which include 8. The capital of Brazil, Brasilia, was founded in the presence of malaria and terrain rugged- 1960 to serve as the nation’s capital, replacing ness), (ii) agriculture fundamentals (which Rio de Janeiro, because Brasilia is in a more include a set of 14 biome indicators, among central ­location. others), and (iii) trade fundamentals (which 9. For example, being near the coast in precolo- include five trade variables that focus on nial times meant having nearby sources of access to water ­ transport). These three are settlements. food (fisheries) that could sustain ­ expected to capture natural advantages that In more recent times, a coastal location can are spatially concentrated and could favor the also be an advantage for ­ trade. concentration of ­ population. 10. The persistence of subnational population 2. The cluster algorithm defines urban areas as density is not observed across all c ­ountries. dense, spatially contiguous clusters of popula- Colombia and Mexico, for example, show tion. For details, see box ­ 1.2. ­ higher than average persistence, whereas 3. There were two waves of falling global trans- Argentina and Uruguay, for example, show a port ­cost. The first wave was from about reversal of density (because the areas that are 1840 to World War I ­. The second wave densely populated today were sparsely popu- occurred right after 1 ­ 950. Here we refer to lated in precolonial ­times). Historical features the second wave (1950). See World Bank lie behind many divergences in persistence, as (2008), Krugman (1991a, 1991b, and ­ 2007).­ covered by Maloney and Caicedo ­ (2016). ­ 44   RAISING THE BAR 11. WUP is compiled (from national sources) and 17. Namely, LandScan 2012, GHS Pop, and maintained by the Population Division of the WorldPop gridded population ­ data. United Nation’s Department of Economic and 18. Our analysis here is limited to urban shares Social ­ Affairs. We rely on data from the 2014 as opposed to trends because only one cross-­ revision of WUP ­ (https://esa.un.org/unpd/wup/). section (2012) of urban shares has been esti- 12. The literature adopting subregional classifica- mated using both the cluster method and tions of LAC presents no consensus on how to the ­AI. ­ group the ­ countries. We elected to include 9. These results correspond to the use of the clus- 1 Mexico as part of South America, given its sim- ter method in combination with LandScan ilarities (such as population, area, and density) gridded population data for ­ 2012. ­ with other countries in this ­ subregion. ­ 0. This finding is consistent with results reported 2 1 3. Calculations based on WUP, July 2 ­017. by Uchida and Nelson (2010) and the World Growth is calculated as the compound annual Bank (2008), using the original version of growth rate between year x and year x+5 for the ­AI. urban population (WUP estimates are avail- 21. The estimated elasticity between GDP per able at five-year ­ intervals). Over the same capita and the urban share is 4 ­ 012. ­ .2 for 2 period, growth in the rural population passed The complete set of mechanisms linking from ­ 1.18 percent per year between 1960 and urbanization to economic development is not 1961 to a contraction of ­ 0.26 percent per year entirely ­ understood. We are grateful to Gilles between 2014 and ­ 2015. The rural population Duranton for this ­ observation. in the region has been shrinking since ­ 1994. 22. We used the cluster algorithm and LandScan 14. We find that across the world there is a nega- 2012 population ­ grid. As shown by Roberts tive and statistically significant relationship al. (2017), similar results are found using et ­ ­(p<0.01) between the urban share and the the ­ AI. ­ urban population growth ­ rate. ­ 23. We conducted a similar analysis to the one 1 5. For all countries (except Austria), urban crite- in table 1.1, but with dummies for Sub- ria fall into at least one of these four broadly Saharan Africa countries, and found that they defined ­ categories. However, every country’s are systematically underperforming on official definition is slightly different and may include and consistent ­measures. ­ particularities that are not fully reflected 24. On the one hand are static productivity gains among these ­ four. For example, the definition (gains in the level of productivity) when peo- for Honduras was counted in the “population ple move from traditionally less-productive size” and “urban services or characteristics” sectors, such as agriculture, to more modern categories, but it also has elements that do not productive sectors, such as manufacturing and fit neatly in either (such as “communication by other industrial s­ ectors. On the other hand are land [road or train] or regular air or maritime dynamic productivity gains (gains in the growth ­ service”). Austria’s definition, according to the of productivity) resulting from changes in tech- definition in the WUP, is based on commuting nology within each ­ sector. ­ patterns into an urban core and does not fall 5. GGDC has data for 1950–2010, but our anal- 2 into any of these ­ categories. ysis covers 1960–2009 to ensure a balanced 16. The mean difference between Latin America ­panel. ­ and the Caribbean and the rest of the world is 26. The GGDC dataset has been used by Timmer, also statistically significant at the 10 percent de Vries, and de Vries (2015) and Rodrik level in a simple one-sided, two-sample, t-test ­ (2015). Data from the World Development where the alternative hypothesis is that the Indicators do not cover such a long period, nor mean difference is negative (that is, the mean are they consistent across ­ years. LAC-9 con- for LAC countries is less than that for non- sists of Argentina, Bolivia, Brazil, Chile, LAC ­ countries). Likewise, a Mann–Whitney U Colombia, Costa Rica, Mexico, Peru, and test—which may be more appropriate given the República Bolivariana de Venezuela. Asia con- large size difference between the LAC and non- sists of India; Indonesia; Japan; Republic of LAC samples and the absence of normality— Korea; Malaysia; Philippines; Singapore; rejects the null hypothesis that the LAC and Taiwan, China; and T ­ hailand. North America non-LAC samples come from the same under- consists of the United S ­ tates. Africa consists of lying population (Roberts et ­ al. ­ 2017). Botswana, Ethiopia, Ghana, Kenya, Malawi, U rbani z ati o n , E c o n o m i c D e v e l o p m ent , and S tru c tura l T rans f o r m ati o n   45 Mauritius, Nigeria, Senegal, South Africa, 31. A recent paper by Gaubert (2017), which Tanzania, and Z ­ambia. Europe consists of studies the location choices of firms in a Denmark, Germany, France, Great Britain, Italy, range of sectors across cities, also suggests Netherlands, Spain, and ­ Sweden. Middle East that, in a general equilibrium model, there and North Africa consists of the Arab Republic are productivity benefits from reducing con- of Egypt and ­Morocco. gestion costs (increasing housing supply) 27. We conducted a regression analysis to test the because they allow for more efficient spatial relationship between the share of people organization of production in differentiated employed in services and industry on the one goods ­sectors. hand and aggregate labor productivity (gross 32. As outlined in a recent World Bank book, value added per employee) on the ­ other. We Africa Cities: Opening the Doors to the World found a significant and positive relationship (Henderson, Venables, and Lall 2017), institu- between the shares of employment absorbed tional and regulatory constraints can lead to a by the services and industrial sectors and a misallocation of land and labor, fragmented country’s labor ­ productivity. We also found physical development, and limited productivity that the Latin America and the Caribbean gains. All these can hold down the emergence ­ region is delivering significantly lower produc- of the tradable sector and trap cities into pro- tivity gains—because of labor reallocation ducing only locally traded goods and ­ services. (from agriculture) to services and industry— than that observed in the rest of the ­world. ­ 28. The primary limitation is that this data set References covers only a small subset of countries in the Ades, Alberto, and Edward Glaeser. 1994. “Trade world. In addition, the statistical foundations ­ and Circuses: Explaining Urban Giants.” of GDP and employment estimates in many Quarterly Journal of Economics 110 (1): (developing) countries are subject to substan- 195–227. tial measurement error (see, for example, Alvarez-Cuadrado, Francisco, and Markus Devarajan 2013; Jerven 2013), which can Poschke. 2011. “Structural Change Out of skew productivity ­ estimates. ­ Agriculture: Labor Push versus Labor Pull.” 9. Using the World Bank Enterprise Surveys, 2 American Economic Journal: Macroeconomics Francis, Saliola, and Seker (2013) estimate per- 3 (3): 127–58. formance using comparable firm-level data for Arthur, W. Brian. 1994. Increasing Returns and 31 countries in the ­ region. Their study finds Path Dependence in the Economy. Ann Arbor: that the annual growth rate in real labor pro- University of Michigan Press. ductivity is declining in the manufacturing and Bairoch, Paul, and Christopher Braider. 1988. services sectors in Latin America and the Cities and Economic Development from the Caribbean, while concurrently those sectors Dawn of History to the Present. Chicago: add ­jobs. These trends show that LAC region’s University of Chicago. businesses are expanding their workforce but Bonomi Barufi, Ana Maria, Eduardo A. Haddad, that revenue gains are ­ lagging. ­ and Peter Nijkamp. 2016. “Industrial Scope of 30. There is some growing evidence (Kim and Agglomeration Economies in Brazil.” Annals Zangerling 2016) that lower productivity of Regional Science 56 (3): 707–55. gains in the service sector in Latin America and Chen, M., H. Zhang, W. Liu, and W. Zhang. the Caribbean might be linked to the concen- 2014. “The Global Pattern of Urbanization tration of labor in nontradable and low-value- and Economic Growth: Evidence from the added services, which appear to have limited Last Three Decades.” PLOS One 9 (8): productivity gains from ­ agglomeration. A e103799. paper by Bonomi Barufi, Haddad, and Chenery, Hollis B., and Lance Taylor. 1968. Nijkamp (2016) finds that high- and low-tech “Development Patterns: Among Countries and manufacturing benefit the most from agglom- over Time.” Review of Economics and Statistics eration economies in Brazil, ­ followed by ser- 50 (4): 391–416. vices associated with higher knowledge Comin, Diego, William Easterly, and Erick Gong. intensity. Low-skilled services and medi- ­ 2010. “Was the Wealth of Nations Determined um-tech manufacturing have the lowest coeffi- in 1000 BC?” American Economic Journal: cients of agglomeration ­ economies. Macroeconomics 2 (3): 65–97. 46   RAISING THE BAR Davis, Donald R., and David E. Weinstein. 2002. Herrendorf, Berthold, Richard Rogerson, and “Bones, Bombs, and Break Points: The Akos Valentinyi. 2013. “Growth and Structural Geography of Economic Activity.” American Transformation.” NBER Working Paper 18996, Economic Review 92 (5): 1269–89. National Bureau of Economic Research, Davis, James C., and J. V. Henderson. 2003. Cambridge, MA. “Evidence on the Political Economy of the Jerven, M. 2013. Poor Numbers: How We Are Urbanization Process.” Journal of Urban Misled by African Development Statistics and Economics 53 (1): 98–125. What to Do about It. Ithaca, NY: Cornell de Rojas, Jose Luis. 2012. Tenochtitlán: Capital University Press. of the Aztec Empire. Gainesville, FL: University Kim, Yoonhee, and Bontje Zangerling. 2016. of Florida Press. Mexico Urbanization Review: Managing Devarajan, Shantayanan. 2013. “Africa’s Statistical Spatial Growth for Productive and Livable Tragedy.” Review of Income and Wealth 59 (2): C i t i e s i n M e x i c o . Wa s h i n g t o n , D C : 1–7. World Bank. Diamond, Jared M. 1997. Guns, Germs and Steel: Krugman, Paul. 1991a. Geography and Trade, The Fate of Human Societies . New York: Cambridge, Mass: MIT Press. W. W. Norton. ———. 1991b. “Increasing Returns and Economic Dijkstra, L ., and H . Poelman. 2014. “A Geography.” Journal of Political Economy 99 Harmonised Definition of Cities and Rural (3): 483–99. Areas: The New Degree of Urbanization.” ———. 2007. “The ‘New’ Economic Geography: Regional and Urban Policy Working Paper Where Are We?” In Regional Integration in 01/2014, European Commission, Brussels. East Asia, edited by Masahisa Fujita, 23–34. Duranton, Gilles. 2015. “A Proposal to Delineate New York: Palgrave Macmillan. Metropolitan Areas in Colombia.” Desarrollo Krugman, Paul, and Raul Livas Elizondo. 1996. y Sociedad 15: 223–64. “Trade Policy and the Third World Metropolis.” The Economist. 2014. “The Conquest of Mexico: Journal of Development Economics 49 (1): On the Trail of Hernán Cortés.” December 17. 137–150. Ellis, P., and M. Roberts. 2015. Leveraging Kuznets, Simon. 1973. “Modern Economic Urbanization in South Asia: Managing Spatial Growth: Findings and Reflections.” American Transformation for Prosperity and Livability. Economic Review 63 (3): 247–58. Washington, DC: World Bank. Maloney, William F., and Felipe Valencia Caicedo. Francis, David C., Federica Saliola, and Murat 2016. “The Persistence of (Subnational) Seker. 2013. “Measuring Firm Performance in Fortune.” The Economic Journal 126 (598): Latin America and the Caribbean.” Brief, 2363–2401. World Bank, Washington, DC. Michaels, Guy, and Ferdinand Rauch. 2013. Gaubert, Cecile. 2017. “Firm Sorting and “Resetting the Urban Network: 117–2012.” Agglomeration.” Working Paper, University of Economic Series Working Paper 684, University California, Berkeley. of Oxford, UK. Haas, Jonathan, Shelia Pozorski, and Thomas Olsson, Ola, and Douglas A. Hibbs Jr. 2005. Pozorski, eds. 1987. The O rigin s an d “Biogeography and Long-Run Economic Development of the Andean State. Cambridge: Development.” European Economic Review Cambridge University Press. 49: 909–38. Henderson, J. V. 2003. “The Urbanization Organisation for Economic Co-operation and Process and Economic Growth: The So-What Development. 2012. Redefining “Urban”: A Question.” Journal of Economic Growth 8 (1): New Way to Measure Metropolitan Areas. 47–71. Paris: OECD Publishing. Henderson, J. V., Adam Storeygard, Tim L. Pages, Carmen, ed. 2010. The Age of Productivity. Squires, and David N. Weil. 2016. “The Global Washington, DC: Inter-American Development Spatial Distribution of Economic Activity: Bank. Nature, History, and the Role of Trade.” NBER Roberts, Mark, Brian Blankespoor, Chandan Working Paper 22145, National Bureau of Deuskar, and B enjamin Stewar t. 2017. Economic Research, Cambridge, MA. “Urbanization and Development: Is Latin Henderson, J. V., Anthony J. Venables, and Somik America and the Caribbean Different from the Vinay Lall. 2017. Africa’s Cities: Opening Doors Rest of the World?” Policy Research Working to the World. Washington, DC: World Bank. Paper 8019, World Bank, Washington, DC. U rbani z ati o n , E c o n o m i c D e v e l o p m ent , and S tru c tura l T rans f o r m ati o n   47 Rodrik, Dani. 2015. “Premature Deindus­ D e velopm e nt?” Jo u r n a l of E c o n o m i c trialization.” Journal of Economic Growth Literature 51 (2): 325–69. 21 (1): 1–33. Timmer, M. P., G. J. de Vries, and K. de Vries. Rozenfeld, Hernán D., Diego Rybski, Xavier 2015. “Patterns of Structural Change in Gabaix, and Hernan A. Makse. 2011. “The D evelopi ng C ou nt r ie s .” I n Ro u tl e dge Area and Population of Cities: New Insights Handbook of Industry and Development, from a Different Perspective on Cities.” edited by J. Weiss and M. Tribe, 65–83 . American Economic Review 101 (5): 2205–25. Abingdon, UK: Routledge. Saavedra-Chanduvi, Jaime, and Ethel Sennehauser. Uchida, H irotsugu, and A nd rew Nelson. 2009. Reshaping Economic Geography in Latin 2008.  Agglomeration Index: Towards a America and the Caribbean: A Companion New Measure of Urban Concentration . Volume to the 2009 World Development Washington, DC: World Bank. Report. Washington, DC: World Bank. Wahl, Fabian. 2016. “Does Medieval Trade Spence, M ichael, Pat ricia Cla rke A n nez , Still Matter? Historical Trade Centers, and Robert M. Buckley. 2009. Urbanization Agglomeration and Contemporary Economic and Growth: Commission on Growth and Development.” Regional Science and Urban D e v e l o p m e n t . Wa s h i n g t o n , D C : Economics 60 (C): 50–60. World Bank. World Bank. 2008. World Development Report, Spolaore, Enrico, and Romain Wacziarg. 2013. 20 09: Reshaping Economic Geography. “How Deep Are the Roots of Economic Washington, DC: World Bank. The Many Dimensions of Urbanization and the 2 Productivity of Cities in Latin America and the Caribbean Mark ­­Roberts Introduction cross-country differences in gross domestic The urban share of a country’s population, as product (GDP) per ­­ capita. measured in chapter 1, is a useful indicator, In addition, this chapter examines produc- but it captures only one aspect of a country’s tivity outcomes in individual LAC urban urbanization ­­ process. The character of areas, benchmarking these against urban 2 urbanization may thus differ fundamentally ­­ areas elsewhere in the world. Because no set- among countries despite similar ­­ shares. tlement exists in isolation, we analyze the Compare, for example, the notoriously productivity dispersion across urban areas in “sprawling” nature of urbanization in the LAC countries, again comparing them with United States, complete with its sometimes those in non-LAC countries and ­­ regions. The seemingly never-ending suburbs, with the better integrated are a country’s urban areas more “compact” urban development more through flows of goods, services, labor, capi- typical of Western European ­­countries. tal, and ideas, the more productivity at the This chapter aims to go beyond the previ- margin might be equalized between them, ous chapter’s analysis and provide a more and the greater their contribution to produc- in-depth examination of patterns of urban- tivity and growth at national ­­ level.3 ization in the Latin A merica and the To facilitate the comparison of urbaniza- Caribbean (LAC) region and to see how these tion patterns in the LAC region with those compare with those in the rest of the world elsewhere, we build on the methods for the along several dimensions of urbanization: globally consistent definition of urban areas density of urban areas, prevalence of agglom- introduced in chapter 1, using the cluster algo- erations that consist of multiple “cities,” and rithm of Dijkstra and Poelman (2014) to con- rates of urban ­­primacy.1 We also analyze struct a global dataset of almost 64,000 urban whether these dimensions are related to ­­ areas in 192 countries for 2012. With this The author thanks Angelica Maria Sanchez Diaz and Jane Park for excellent research assistance for this chapter, as well as inputs. Benjamin Stewart for invaluable Geographic Information System ­­ 49 50   RAISING THE BAR dataset, we analyze differences in population North America and Western Europe but size and density across urban ­­areas. We also ­­ in common with the rest of the world. For identify the presence of multicity agglomera- South America, there is some evidence tions (MCAs), urban areas that consist of two of a positive relationship, but this is far cities. We analyze differences in the or more ­­ weaker than that for North America and prevalence and characteristics of these MCAs Western ­­Europe. Although alternative between LAC and the rest of the ­­ world. By explanations exist, these results are combining these data with high-resolution consistent with the hypothesis that the nighttime lights data, we also construct a difficulties associated with governing measure of productivity at the individual large metropolitan areas, which arise from ­­ urban area level. This facilitates analysis of the “fragmentation” of infrastructure the productivity performances of urban areas and basic service provision across in the LAC region against those in the rest of multiple local governments, can stifle the world, as well as of productivity disper- the contribution that such areas make sion across urban areas in LAC ­­countries. to national productivity unless effective The main findings of the chapter are as coordination mechanisms are in place. ­­ follows: Whereas North American and Western European countries have succeeded in •  Urban areas in the LAC region stand out advanced. this, LAC countries are less ­­ internationally for their high population • A significant number of LAC countries densities, particularly in comparison exhibit unusually high urban primacy, with urban areas in Europe and Central benchmarked against comparator Asia (ECA) and North America, where world. These countries in the rest of the ­­ urban areas, on average, have similar countries include Barbados and Dominica populations but cover larger geographic in the Caribbean, Costa Rica and Panama areas. Most individual LAC countries ­­ in Central America, and Argentina, also exhibit significantly higher urban Paraguay, and Uruguay in South ­­ America. densities than comparator ­­countries.4 However, despite concerns frequently • The high urban densities in LAC expressed about the negative repercussions countries may be exerting a negative of high primacy on national economies, we effect on national levels of GDP per find no evidence of a negative relationship capita. This suggests that LAC cities may ­­ with GDP per ­­capita. lack the “enabling environment” in both • Urban areas in South America and policy choices and infrastructure levels, Mexico have relatively high productivity, required to mitigate productivity-sapping globally ­­benchmarked. However, they congestion costs and prevent them from lag the global frontier of productivity overpowering the productivity benefits ­­ performance. Urban areas in the rest of of agglomeration ­­economies. LAC tend to exhibit average productivity • Among world regions, Latin America given their population ­­sizes. and the Caribbean has the second most • LAC countries show relatively high ­­ MCAs in the world. A large share of the productivity dispersion across urban areas, LAC region’s urban population lives in higher than in more developed countries. these agglomerations and, as economic This high dispersion is associated with development continues, we can expect relatively low average national road further. this share to increase ­­ ­­ density. This is consistent with relatively • There is no positive significant poorly integrated internal markets, and relationship, for the subregions of the is suggestive of a spatial misallocation of Caribbean or Central America, between resources across urban areas that may be the share of a country’s population living undermining their aggregate contribution in MCAs and its GDP per capita, unlike to national productivity. T he M an y D imensi o ns o f U rbani z ati o n and the P r o du c ti v it y o f Cities in l a c    51 Defining a Global Data Set of provision is less efficient to jurisdictions Urban Areas where it is more ­­ efficient. Against this, how- ever, many analysts argue that, because of A byproduct of the application of the cluster interjurisdictional spillovers, local govern- algorithm introduced in chapter 1 (in the ment fragmentation—absent effective mecha- “Urbanization in the LAC Region and the nisms for coordination between local Rest of the World” section) is a global data governments—undermines the efficient pro- set, for 2012, of 63,629 urban areas in vision of infrastructure and basic services in 192 countries, including 7,197 urban areas in the wider urban ­­ area. Given that governance 34 LAC ­­countries.5 Each of these areas meets challenges often increase as the size of an the criteria for an urban area specified by the urban area rises, discussion tends to focus on algorithm, a ­ spatially contiguous area for large metropolitan areas (Muzzini et ­­ al. which population density (measured at a reso- 2016, for Argentina; Kim and Zangerling lution of 1 km 2) is at least 300 people per 2016, for ­­Mexico).8 square kilometer throughout the entire area Unfortunately, there is no global data set 5,000. It and the overall population is at least ­­ of local government administrative boundar- is important to note that urban areas thus ies that would allow us to quantify the num- defined do not necessarily conform to official ber of local government units within each administrative boundaries of towns and urban ­­area.9 We are, however, able to iden- cities. Rather, they correspond to a wide array ­­ tify urban areas that consist of multiple of places, ranging from settlements that just “­ c ities,” where a “city” in this context is meet the criteria6 to extended urban agglom- defined by its administrative ­­ boundaries. In erations that cover hundreds of square kilo- some cases, such a “city” may represent a dis- meters and include tens of millions of ­­ people. tinct center or subcenter of an urban area, At the extreme upper end of the distribution but, in others, it may amount to little more are nine urban areas—Delhi in India; Dhaka than a suburb of another “city” in the same and Rajshahi in Bangladesh; Jakarta and area, even though it is administratively Surabaya in Indonesia; Lahore in Pakistan; distinct. Following CIESIN (2013); Zhou, ­­ and Beijing, Chongqing, and Shanghai in Hubacek, and Roberts (2015); and Ellis and China—all of which have implausibly large Roberts (2016); we refer to such urban areas estimated populations of more than 45 mil- as multicity agglomerations (MCAs). lion and which we, therefore, drop from fur- We identify “cities” in urban areas, and ther analysis, leaving us with a final global thus MCAs, by using Geographic Information sample of 63,620 urban ­­ areas.7 System (GIS) techniques to overlay a global In this global sample, we would also ide- layer of individual settlement points on a ally like to quantify the number of local gov- global map of our urban ­­ areas.10 The results ernment units in each area, which would that we report in the main text focus on allow us to compare local government frag- MCAs defined based on cities that have a mentation in urban areas in the LAC region minimum population of ­­ 100,000 each. This ­­ with the equivalent in the rest of the world. is because the global layer of settlement points Whether such fragmentation is good or bad more rel iably ident i f ie s such cit ­­ ie s. for infrastructure and service delivery, and Importantly, however, our main regression ­­ thus productivity, is open to debate. By draw- results (reported in the “Implications for ing an analogy with competitive markets for National Productivity” section later in this private goods, Tiebout (1956) hypothesized chapter) are robust to redefining an MCA as that competition between different political an urban area containing two or more settle- jurisdictions within an urban area may lead ments, as taken from the global settlement ­­ to efficiency in the local public sector. This is point layer, of any population ­­size.11 because residents will vote with their feet and In total, the number of MCAs with two or move from jurisdictions where local service more cities, each with a minimum population 52   RAISING THE BAR of 100,000, identified is small—only 295 region with those in the rest of the world globally, including 54 in the LAC region along dimensions that go beyond the simple (0.46 percent of all urban areas ­­ ­­ worldwide). comparison of urban ­­ shares. The feature of LAC urbanization that most stands out from such a comparison is the higher average pop- Urban Areas in the LAC Region areas. As table ulation density in its urban ­­ Are More Densely Populated ­­ 2.1 shows, at 2,360 people per square kilo- Than Those Elsewhere meter, the mean population density for LAC Using our global sample of urban areas, we 1.54 times the global ­­ urban areas is ­­ figure. compare patterns of urbanization in the LAC Similarly, the median population density for TABLE ­­2.1  Summary Statistics for Global Sample of Urban Areas 25th 75th 99th Mean percentile Median percentile percentile Maximum World (N = 63,620; total urban population ≈ ­­3.73 billion) Population 58,642 6,632 10,154 22,417 732,556 43,790,629 Area (km2) ­­36.8 ­­5.5 ­­10.9 ­­22.8 ­­389.4 22,321 2 Population density (per km ) 1,529 723 1,181 1,901 6,488 36,186 No. of cities in urban area ­­0.057 0 0 0 1 66 Latin America and the Caribbean (N = 7,197; total urban population ≈ ­­432.9 million) Population 60,151 7,259 11,762 24,928 832,365 20,588,698 Area (km2) ­­18.4 ­­4.4 ­­7.2 ­­13.1 ­­209.6 3,404 2 Population density (per km ) 2,360 1,304 1,961 2,948 8,180 19,232 No. of cities in urban area ­­0.080 0 0 0 1 30 Caribbean (N = 473; total urban population ≈ ­­23.6 million) Population 49,826 6,790 11,419 23,001 668,129 3,431,292 Area (km2) ­­18.3 ­­4.7 ­­8.3 ­­14.7 ­­176.9 569 2 Population density (per km ) 2,007 1,045 1,568 2,495 6,944 9,956 No. of cities in urban area ­­0.101 0 0 0 1 20 Central America (N = 1,778; total urban population ≈ 110 million) Population 61,860 6,941 11,043 24,047 916,871 19,782,701 Area (km2) ­­20.3 ­­5.4 ­­8.3 ­­14.1 ­­212.4 2,651 2 Population density (per km ) 1,930 1,052 1,542 2,383 6,778 10,232 No. of cities in urban area ­­0.069 0 0 0 1 16 South America (N = 4,946; total urban population ≈ ­­299.4 million) Population 60,524 7,427 12,182 25,368 819,726 20,588,698 Area (km2) ­­17.7 ­­3.7 ­­6.6 ­­12.5 ­­207.0 3,404 2 Population density (per km ) 2,549 1,479 2,152 3,117 8,757 19,232 No. of cities in urban area ­­0.082 0 0 0 1 30 Source: Calculations based on analysis of urban areas defined using the cluster algorithm of Dijkstra and Poelman (2014), as applied to LandScan 2012 gridded population ­­data. ­­ rea. Note: “No. of cities in urban area” refers to the number of cities with a population exceeding 100,000 whose settlement points intersect with an urban a An urban area that intersects with two or more such settlement points is defined as a multicity agglomeration. T he M an y D imensi o ns o f U rbani z ati o n and the P r o du c ti v it y o f Cities in l a c    53 FIGURE ­­2.1  Percentage of Urban Areas with Population Densities Higher Than the Global Median, by Region 90 80 70 60 Percent 50 40 30 20 10 0 South Caribbean Central Middle East Sub- South East Asia Europe North America America and North Saharan Asia and and Central America Africa Africa Pacific Asia Source: Calculations based on analysis of urban areas defined using the cluster algorithm of Dijkstra and Poelman (2014), as applied to LandScan 2012 gridded population ­­data. Note: An urban area is classified as dense if its mean population density exceeds the global median of 1,180 people per square kilometer. LAC urban areas, which is almost 2,000 peo- of urban areas classified as “dense” is at least ple per square kilometer, exceeds the median 50 percent for all but four of the 34 LAC 66 percent. for all urban areas globally by ­­ ­­ countries in the global sample (figure 2.3). In Likewise, if we define urban areas as six of these countries—Aruba, Belize, Brazil, either “dense” or “not dense” depending on Dominica, Peru, and República Bolivariana whether their mean population densities de Venezuela—that proportion exceeds exceed or fall below the median of almost 90 percent and in a further seven is 80–90 1,200 people per square kilometer for all ­­ percent. Three countries that break the areas globally, South America, Central p at tern are A rgentina, Grenada, and ­ America, and the Caribbean lead the way Barbados, with dense proportions of roughly globally on the relative prevalence of dense 36 percent, 25 percent, and 20 percent 2 .1). The contrast is par- urban areas (figure ­­ ­­ respectively. In several, mainly Caribbean, ticularly marked against the more developed countries (A ntigua and Barbuda, the regions of EC A and Nor th A ­­ merica. Bahamas, Guyana, Jamaica, and ­­ St. Kitts Whereas the proportion of urban areas clas- and Nevis) the split between “dense” and sified as dense exceeds 65 percent in each of “not dense” is roughly 50:50, which, by con- the three LAC subregions, in ECA the figure distribution. struction, mirrors the global ­­ is just less than 14 percent and in North The finding of high urban population den- America a little over 2 ­­ percent. The higher sities in most LAC countries also carries over urban densities in the LAC region are the when, instead of comparing them with the result not so much of differences in the pop- rest of the world, we compare each LAC ulations of urban areas but in the land areas country with a corresponding set of three 2.2). LAC urban areas that they cover (figure ­­ comparator countries selected using the tend to be geographically much smaller than methodology set out in box ­­ 2.1. As annex 2B those in ECA and North ­­ America.12 shows, 20 LAC countries exhibit mean High population densities across LAC urban population densities significantly urban areas are, moreover, attributable not greater than those in the corresponding just to a few large countries: the proportion ­­comparators.13 54   RAISING THE BAR FIGURE ­­2.2  Distribution of Area Size and Population across Urban Areas, Selected Regions a. Size of urban area b. Population of urban area 0.8 0.10 0.08 0.6 0.06 0.4 0.04 0.2 0.02 0 0 0 20 40 60 80 100 8 10 12 14 16 18 Area (km2) Population (log) caribbean central America South America Europe and central Asia North America Source: Calculations based on analysis of urban areas defined using the cluster algorithm of Dijkstra and Poelman (2014), as applied to LandScan 2012 gridded population d ­­ ata. Note: Panels a and b show, for different regions, the distribution of area (in square kilometers) and population (in natural logs), respectively, of urban areas using an Epanechnikov ­­kernel. For expositional purposes, the distributions of area are trimmed at 100 ­­km2. FIGURE ­­2.3  Percentage of Urban Areas with Population Densities Higher Than the Global Median, by LAC Country 100 80 Percent 60 40 20 0 ez P zil ue eru Co uad B lo or ra ia ug y Su Ch y rin ile Ar uya e Do Aru a in a ep ba Pa oliv a ge na ica id G t. lic an nad cia To es go . K a rb i s a s, a N e m s Ba ena a rb da Co Be s sta lize Ho nama Ni ndu a El ara ras Gu lva a em r M ala ico St Bah Ba Hait at do Ja evi o Ur gua ua B bi in b itt ma ud Gr aic Pa Ric Sa gu G am nd Th Ec a, R d in a in e S ub ad n R Cu ba ad re Lu ex nt m Br m l c d n an ica Ve Tr d th ua in m tig an Do An nt ce in .V St South America Caribbean Central America Source: Calculations based on analysis of urban areas defined using the cluster algorithm of Dijkstra and Poelman (2014), as applied to LandScan 2012 gridded population ­­data. Note: Countries are sorted in descending order within each of three subregions (South America, Caribbean, and Central ­­America). The black dashed line is included to facilitate comparison of the distribution of “dense” versus “not dense” urban areas in any given LAC country to the distribution of such areas ­­ dense”). LAC = Latin America and the Caribbean. globally (by definition, 50 percent of urban areas globally are classified as “ T he M an y D imensi o ns o f U rbani z ati o n and the P r o du c ti v it y o f Cities in l a c    55 BOX ­­2.1  Comparing Apples with Apples: Selecting Comparators for LAC Countries Throughout this chapter, we benchmark individual landlocked, or the rest and, for each LAC country, Latin American and Caribbean countries against a searched for countries in the rest of the world fall- corresponding matched set of comparator countries category. For each LAC country, ing into the same ­­ drawn from the rest of the ­­world. These comparisons comparators. In this gives a long list of potential ­­ complement the more straightforward regional com- the second stage, we then whittled down the list parisons, that is, the comparisons of Latin America to a final set of three comparators by selecting and the Caribbean (LAC) against other World Bank– the “nearest neighbors” on population, land area, defined regions, presented in this ­­ chapter. Such and overall mean population ­­ density. In doing so, regional comparisons suffer from the problem that we imposed the restriction that the set of com- the differences they reveal may be driven by differ- parators must include at least one country from ences in the composition of the “types” of countries, each of the East Asia and Pacific and Europe and for example, in the proportion of small island nations Central Asia regions, which are commonly used that make up each ­­ region. The individual-country comparator regions for the LAC region (Ferreyra benchmarking is intended to provide a cleaner com- al. ­­ et ­­ 2017). We selected the third comparator parison, because a LAC country can, in key respects, world. country unrestrictedly from the rest of the ­­ be considered more like its corresponding set of This helps to avoid all comparator countries being comparator countries than countries outside that ­­ set. drawn from, for example, Sub-Saharan Africa, and In selecting comparator countries for a given LAC helps to ensure geographic diversity among the country, there is a temptation to use the development ­­comparators. level (gross domestic product per capita) as one of the One might ask, “Why restrict the number of criteria. It would seem natural to compare a middle-­ ­­ comparators for each LAC country to three?” The income LAC country against middle-income countries choice was based on experimentation with the ­­data. in the rest of the ­­ world. However, we avoid this temp- For any given LAC country, it was found that, as tation because gross domestic product per capita is the number of countries in its comparison set was too closely related to the outcome of ­­ productivity. expanded, the quality of the “match” with the added Instead, we adopted a two-stage procedure for marginal country decreased, undermining the qual- selecting comparator ­­ countries. In the first stage, ity of the ­­ comparison. Three was found to be the we classified all countries globally as island, group. optimal average size of the comparison ­­ H ig h u rba n den sit ie s repre sent a 2016). High urban densities can also Roberts ­­ double-edged ­­sword. On the one hand, they work to propagate infectious disease vectors can help to stimulate powerful agglomeration and act as a stimulus for crime and ­­ violence. economies that spur productivity through a Glaeser (2011) has dubbed these conges- variety of mechanisms, including through the tion forces the “demons of density,” and there spillover of knowledge between firms and can come a tipping point for any urban area workers, the growth of a large local base of where the positive externalities and spillovers intermediate input suppliers, and better match- of density come to be outweighed by the neg- ing of workers with jobs (Marshall 1890; ative effects of congestion, such that the ­­ Duranton and Puga 2004). However, high effects of increased density on productivity densities also give rise to adverse congestion negative. The exact urban density of this are ­­ forces, which can undermine productivity “tipping point” is dependent on the “enabling ­­ within cities. These include not only traffic environment” for positive net agglomeration congestion externalities but also costs that effects that cities offer, where this enabling arise more generally from the pressure of environment is itself a function of policy urban population on the supply of basic urban choices that affect the management of cities services and infrastructure, land and housing a nd l e ve l s of u rb a n i n f r a s t r u c t u r e markets, and the environment (Ellis and development. For example, investments in ­­ 56   RAISING THE BAR urban infrastructure and increases in the sup- point”—for example, improvements in the ply of affordable housing will tend to allevi- technology for fighting and deterring crime ate congestion costs at any given urban may also push the point farther away. ­­ density and push the “tipping point” further Comprehensive data on the types of conges- ­­ away. Technologies and the mix of industries tion forces and the costs in terms of produc- that characterize an urban area may also tivity and welfare are lacking; box ­­ 2.2 exert an important influence on the “tipping available. discusses the limited information ­­ BOX ­­2.2  Congestion Forces in LAC Urban Areas How strong are congestion forces in Latin American equal to the entire population of the ­­ Philippines. and Caribbean urban areas, and how large are Furthermore, it took the LAC region just under 25 years the costs that they impose on firms and workers? to reduce the share of its urban population in slums Unfortunately, because of the absence of compre- ­­ from about 35 percent in 1990 to its 2014 figure. If hensive data, we can provide no direct answer to we extrapolate forward this pace of reduction, slums this quesion. More generally, within the field of come. will remain a feature of LAC cities for decades to ­­ urban economics, there is a surprising dearth of rigorous empirical research on the quantifica- Traffic Congestion tion of congestion costs, their relationship to urban Although regional rates of motorization in the LAC density, and the influence that policy can have on region (about 100–300 vehicles per 1,000 people) mitigating their effects. There are, however, three are a fraction of existing rates in developed nations areas where we can provide some basic descriptive (roughly 500–700 vehicles in Europe and the United ­­i nformation. States), they are nonetheless associated with traf- fic congestion that is among the worst in the world Congestion in Housing Markets (Barbero 2012; see also chapter ­­ 4). According to Absent a sufficiently elastic supply of formal hous- 2016 TomTom traffic index data, which cover 390 ing, high urban densities can generate strong upward cities in 41 countries, Mexico City holds the dubi- pressure on rents and house ­­prices. The consequent ous honor of being the world’s most congested city pricing-out of households from the formal hous- because travel time in the city is, on average, 66 per- ing market can, in turn, cause the proliferation of cent higher during the day than it would be in a free- informal housing (often, slums), which acts as the flow traffic ­­situation.b outward manifestation of “excessive” congestion A further eight LAC cities, out of the 12 LAC markets. forces in land and housing ­­ cities in the data, feature in TomTom’s list of the In the Latin America and the Caribbean (LAC) 100 most congested ­­ cities. Even in the least con- region, congestion in these markets is evident in the gested of these cities, Belo Horizonte in Brazil, existence of the infamous favelas of Rio de Janeiro travel time within the city is, on average, 27 percent and the villas of Gran Buenos Aires. ­­ More generally, higher during the day than it would be if the roads although countries such as Mexico have made inroads were ­­uncongested. To put these numbers into into expanding their housing stock and the access of perspective, travel times in London are, on average, low-income households to mortgage finance (Kim 44 percent higher than in the free-flow situation, and Zangerling 2016), slums remain a notable feature and in New York 35 percent higher.­­ More generally landscape.a According to UN-­ of the region’s urban ­­ B2.2.1), traffic congestion rises much more (figure ­­ Habitat data, one in every five urban residents in the rapidly with population density for the LAC cities LAC region was living in a slum in 2014, implying an than for the non-LAC cities for which TomTom overall urban slum population of 104 million, roughly ­provides ­­data. (continued) T he M an y D imensi o ns o f U rbani z ati o n and the P r o du c ti v it y o f Cities in l a c    57 BOX ­­2.2  Congestion Forces in LAC Urban Areas (continued) FIGURE ­­B2.2.1  Relationship between Traffic Congestion and Population Density, LAC Cities versus Non-LAC Cities 80 Mexico City 60 Congestion level (%) Buenos Rio de Janeiro Aires Santiago de Chile 40 Salvador Recife Fortaleza São Paulo Belo Horizonte 20 Brasília Porto Alegre Curitiba 0 6 7 8 9 10 Population density (log) Non-LAC LAC Source: Calculations based on TomTom traffic index data (­­https://www.tomtom.com/en_gb/ ­­trafficindex/). Population density calculations based on urban areas defined using the cluster algorithm of Dijkstra and Poelman (2014), as applied to LandScan 2012 gridded population d ­­ ata. Note: Congestion is measured as the percentage of extra travel time for trips by road in a city compared with the free-flow traffic ­­situation. The data cover 390 cities globally, including 12 in three LAC countries (Brazil, Chile, and ­­Mexico). LAC = Latin America and the Caribbean. Air Pollution that, although for non-LAC developing country This is an area where LAC cities perform much cities there is a significant positive correlation better, as seen in figure B2.2.2, panel a, which between a city’s population density and its PM 2.5 shows box plots of concentration in ambient air of concentration, no such correlation exists for 2.5 μm or less ­­ fine particulate matter of ­­ (PM 2.5) LAC ­­cities. based on data covering about 3,000 cities glob- This is not to say, however, that the air in LAC ally, from the Global Urban Ambient Air Pollution cities is safe to breathe; it is far from it: Only 11 database of the World Health Organ i zation out of the 128 LAC cities in the WHO database ­­ (WHO). Median air pollution across the 128 LAC have PM 2.5 levels less than what the WHO guide- cities in the database is notably less than that in lines stipulate as representing a significant health other developing world ­­ regions. Panel b shows ­­t hreat. (continued) 58   RAISING THE BAR BOX ­­2.2  Congestion Forces in LAC Urban Areas (continued) FIGURE B2.2.2  Air Pollution in Cities in Latin America and the Caribbean and Other Regions a. Concentration of PM2.5, by region b. Relationship between air pollution and population density in developing country cities South Asia 5 Middle East and North Africa East Asia and PM 2.5 concentration (log) Paci c 4 Sub-Saharan Africa Europe and Central Asia Latin America and 3 the Caribbean Europe and Central Asia (high income) North America 2 0 50 100 150 200 0 2 4 6 8 10 PM2.5 concentration Population density (log) lAc Non-lAc Source: Calculations based on data on levels of PM2.5 taken from the World Health Organization’s Global Urban Ambient Air Pollution Database ­­(http://www​ .who.int/phe/health​_topics/outdoorair/databases/ ­­cities/en/). Note: Panel a shows, for each region, a box plot of the mean annual PM2.5 measured at the city ­­level. Data cover measures of PM2.5 mostly in 2013 and 2014, ­­ ountries. Regions are sorted descending by the regional ­­average. The left and right caps are the minimum and maximum for almost 3,000 cities in 101 c value, excluding outliers. To identify outliers, we calculate the interquartile range; values outside the range defined by (25th Percentile – 1.5 × Interquartile Range, 75th Percentile + 1.5 × Interquartile Range) are considered outliers. Panel b shows a scatterplot for the relation between the natural log of a city’s population density and its (natural log) PM2.5 ­­concentration. The sample for panel b covers 384 cities in 43 developing ­­countries. LAC = Latin America and the Caribbean; PM2.5 = annual concentration of fine particulate matter of 2.5 μm or less. a. Although Mexico has made inroads into expanding its housing stock, the policies that have made this expansion possible have been criticized for, among other things, contributing to uncoordinated urban ­­growth. Hence, much of the low-cost housing has been constructed on the outskirts of municipalities with poor access to employment opportunities and an absence of links with urban planning and infrastructure ­­provision. This has contributed to much of the housing being left ­­vacant. For a full discussion of these issues, see Kim and Zangerling ­­(2016). b. Although the TomTom data measure traffic congestion by comparing to the free-flow traffic situation, this does not provide an accurate measure of the true deadweight loss of ­­congestion. The cost imposed on other road users by the marginal road user is given by the marginal cost of travel minus the average cost of ­­travel. The deadweight loss of traffic congestion is then equal to the sum of these costs where the sum is over the road users who would not travel in the presence of optimal congestion ­­pricing. See Akbar and Duranton (2017) for a further discussion and an attempt to empirically estimate the true deadweight loss of traffic congestion for Bogotá. T he M an y D imensi o ns o f U rbani z ati o n and the P r o du c ti v it y o f Cities in l a c    59 A Significant Share of Latin America more cities, however: 30 against São and the Caribbean’s Urban 23. These MCAs exhibit high average Paulo’s ­­ Population Lives in Large MCAs population densities, exceeding the global Besides high population densities, another median for all urban areas of just under notable feature of LAC urbanization is the 1,200 people per square kilometer. Again, presence of ­­MCAs. We define an MCA as an this provides a major contrast with ECA, urban area that consists of two or more where 24 out of 40 MCAs are dense, and “cities,” each of which is defined based on its North America, where only 9 out of 33 administrative ­­ boundaries. While the num- MCAs are ­­dense. ber of these urban areas—295 globally—is Except for EAP, agglomerations in the very small set against the total number of Caribbean and South America also generally urban areas (see the “Defining a Global Data consist of more cities than agglomerations in Set of Urban Areas” section), their size and other regions ­­do. The mean number of cities economic significance make them of special per agglomeration in the Caribbean is 6.75,­­ interest. Although MCAs represent only ­­ and for South America, just less than ­­ 5. The ­­ 0.46 percent of the urban areas in our global mean number of cities per agglomeration in sample, they are home to an estimated ­­ 1.27 Central America is ­­4.42, which is less than in billion people—about one-third of the North America, but more than in ECA, Sub- world’s urban ­­ population. For a full list of Saharan Africa, the Middle East and North the LAC region’s MCAs, see annex ­­ 2C. Africa, and South Asia. However, the distri- Such urban areas are also of interest because bution of the number of cities per agglomera- they tend to represent large metropolitan areas tion shows a heavy positive skew in all that are typically fragmented into multiple regions. The modal number of cities per ­­ local government ­­ jurisdictions. In São Paulo, agglomeration for all regions, including LAC for example, the urban area, defined using the and its subregions, is ­­ two. cluster algorithm, encompasses 34 municipali- The share of overall urban population liv- ties; Mexico City, 57; and Santo Domingo in ing in MCAs ranges from just over 37 per- 14 the Dominican Republic, 19 (map 2.1). ­­ cent in Central America to 41 percent in Although, as discussed earlier in this chapter, it South America, similar to EAP and South has been argued, following Tiebout (1956), 2.4). Had we included the nine Asia (figure ­­ that the fragmentation of metropolitan areas excluded large urban areas in the sample, this into multiple jurisdictions can improve the effi- share would have been far higher in EAP and ciency of local service delivery by promoting South Asia than in the LAC ­­ subregions.16 At competition between these jurisdictions, it can just under 45 percent, the share of North also create difficulties in coordinating infra- America’s urban population living in MCAs structure provision and service delivery at the is also higher than in the LAC ­­subregions. ­­ level of the metropolitan area. Without mecha- By distribution across countries, one half nisms for metropolitan coordination, these (27 out of 54) of the LAC region’s MCAs are difficulties can, in turn, have negative reper- in only two countries: Brazil (19 MCAs) and cussions for the metro area’s productivity Mexico (8 MCAs) (figure 2.5, panel a). (Ahrend et ­­al. 2014; see also chapter ­­6). Argentina, Chile, and Peru each has three, With 54 MCAs, LAC is second only to República Bolivariana de Venezuela and the East Asia and Pacific (EAP) in World Bank Dominican Republic two ­­ each. A further regions (table ­­2.2). Most of these MCAs (38) nine countries in the region possess a single are in South America, with 12 in Central agglomeration, and the other 17 have no America and 4 in the ­­ Caribbean.15 By popu- MCAs. Among LAC countries with at least ­­ lation, the largest of these MCAs is São Paolo, one MCA, there is quite marked heterogene- with an estimated 20.6 ­­ ­­ million inhabitants. ity in the share of urban population living in The urban area of Buenos Aires encompasses those areas (figure 2.5, panel b). In Venezuela, 60   RAISING THE BAR MAP ­­2.1  Examples of Multicity Agglomerations in Latin America and the Caribbean That Span Multiple Municipalities a. São Paulo, Brazil São Paulo cluster Other urban clusters 0 10 20 km Municipalities intersecting São Paulo cluster b. Mexico City, Mexico c. Santo Domingo, Dominican Republic Caribbean Sea Mexico City cluster Other urban clusters Santo Domingo cluster Mexico City Other urban clusters 0 10 20 km 0 10 20 km Municipalities intersecting Mexico City cluster Municipalities intersecting Santo Domingo cluster Note: The red areas correspond to urban areas defined using the cluster algorithm of Dijkstra and Poelman (2014), as applied to LandScan 2012 gridded population d ­­ ata. The yellow ­­ efined. The dark blue lines represent the boundaries of lines represent subnational administrative boundaries at the Admin-2 (municipality) level that belong to a city as officially d Admin-2 areas that intersect with the urban area but that do not belong to the officially defined ­­city. T he M an y D imensi o ns o f U rbani z ati o n and the P r o du c ti v it y o f Cities in l a c    61 TABLE ­­2.2  Number of Multicity Agglomerations, by Region All multicity agglomerations Dense multicity agglomerations % of global Mean no. of % of global Mean ­­no. of Total total cities Total total cities World Bank regions East Asia and Pacific 90 ­­30.51 ­­5.40 83 ­­34.3 ­­5.67 Latin America and 54 ­­18.31 ­­4.98 54 ­­22.31 ­­4.98 the Caribbean Europe and 40 ­­13.56 ­­3.90 24 ­­9.92 ­­3.58 Central Asia North America 33 ­­11.19 ­­4.45 9 ­­3.72 ­­7.67 South Asia 33 ­­11.19 ­­4.15 28 ­­11.57 ­­4.50 Sub-Saharan Africa 24 ­­8.14 ­­3.54 24 ­­9.92 ­­3.54 Middle East and 21 ­­7.12 ­­3.62 20 ­­8.26 ­­3.70 North Africa Total 295 100 ­­4.59 242 100 ­­4.88 LAC subregions ­­South America 38 ­­12.88 ­­4.97 38 ­­15.7 ­­4.97 ­­Central America 12 ­­4.07 ­­4.42 12 ­­4.96 ­­4.42 Caribbean 4 ­­1.36 ­­6.75 4 ­­1.65 ­­6.75 Source: Calculations based on analysis of urban areas defined using the cluster algorithm of Dijkstra and Poelman (2014), as applied to LandScan 2012 gridded population ­­data. Note: A multicity agglomeration is defined as an urban area with two or more cities, each of which has a population of at least 100,000; an agglomeration is classified as dense if its mean population density exceeds the sample median for all urban areas ­­globally. “Mean no. of cities” refers to those with a population of at least 100,000 per multicity agglomeration. FIGURE ­­2.4  Percentage of Urban Population Living in Multicity Agglomerations, by Region 50 45 40 35 30 Percent 25 20 15 10 5 0 North South South East Asia Caribbean Central Middle East Sub- Europe America America Asia and Pacific America and Saharan and North Africa Africa Central Asia Source: Calculations based on analysis of urban areas defined using the cluster algorithm of Dijkstra and Poelman (2014), as applied to LandScan 2012 gridded population ­­data. Note: A multicity agglomeration is defined as an urban area with two or more cities, each of which has a population of at least ­­100,000. 62   RAISING THE BAR FIGURE ­­2.5  Multicity Agglomerations, by LAC Country a. Number of multicity agglomerations b. Percentage of urban population living 20 80 in a multicity agglomeration 18 70 16 60 Population in MCAs (%) 14 50 12 10 40 8 30 6 20 4 10 2 0 0 e ica ra lic Ch y B ba ne ua ia ile ge ru Co nama El lom a lva ia Ha r Br iti Gu Me azil em o Cu la ela r B M zil lo o a ile in ge ru Ve Rep ina zu lic sta B ra ca Pa uay Sa a Gu H r em i Cu a Bo ba Ec ivia r do zu do a Pa tin at ait at xic ,R do do Ve Ec oliv Sa b a bi El nam al Co exic Co la, R Ar Pe Pa pub gu m Ar Pe a nR aR ne ub Pa Ri Ch t m n Br lva ua l g ica n ica st e in Co m Do Do ­­ ata. Source: Calculations based on analysis of urban areas defined using the cluster algorithm of Dijkstra and Poelman (2014), as applied to LandScan 2012 gridded population d Note: A multicity agglomeration is defined as an urban area with two or more cities, each of which has a population of at least ­­100,000. the share is only some 17 percent, but in primacy in many of the region’s countries, Costa Rica it is more than 70 percent (in the where primacy refers to the share of a coun- on ly ag g lomeration of San ­­ Jose). I n try’s urban population residing in its largest Argentina, Brazil, Colombia, and Mexico, city. Primacy is considered excessive when it ­­ the shares are 40–50 ­­ percent. acts as a drag on overall national productiv- There is a strong positive correlation, glob- ity and on economic ­­ growth. It is caused by ally, between a country’s urban share—as esti- overcongestion in the largest city, which itself mated using the cluster algorithm—and the results from policy distortions that bias the share of its national population living in allocation of resources toward that city at the ­­ MCAs (figure 2.6). This implies that, as LAC expense of other cities or rural ­­areas. High countries continue to develop and urbanize, primacy rates in major Latin American coun- we can expect the potential governance tries have been linked to the widespread trade challenges of having multiple jurisdictions policy distortions of the import substitution within large metropolitan areas to mount, par- industrialization era of the 1960s and 1970s, ticularly for relatively populous LAC countries and to the concurrent high rates of political such as Ecuador, Guatemala, and Peru, which centralization in many LAC countries (Ades share.17 are still at an intermediate urban ­­ and Glaeser 1995; Davis and Henderson 2003; Krugman and Elizondo ­­ 1996).18 Using A Third of LAC Countries a framework in which the effects of primacy Analyzed Suffer from Potentially on long-run economic growth can vary non- Excessive Primacy linearly with a country’s development level One of the most debated characteristics of and its overall size, Henderson (2000) identi- urbanization in the LAC region is excessive fies 24 countries worldwide as suffering from T he M an y D imensi o ns o f U rbani z ati o n and the P r o du c ti v it y o f Cities in l a c    63 FIGURE ­­2.6  Cross-Country Relationship between Urban Share and Share of National Population Living in Multicity Agglomerations 100 Population living in multicity agglomerations (%) 80 60 CRI DOM CHL 40 ARG PRY PER PAN COL BRA SLV MEX HTI GTM 20 CUB BOL ECU VEN 0 HND BLZ NIC LCA GUY SUR JAM URY BHS TTO BRB 20 40 60 80 100 Urban share (%) Source: Calculations based on analysis of urban areas defined using the cluster algorithm of Dijkstra and Poelman (2014), as applied to LandScan 2012 gridded population ­­data. Note: A multicity agglomeration is defined as an urban area with two or more cities, each of which has a population of at least 100,000; urban share is a country’s urban share of the population as measured on the basis of the cluster a ­­ lgorithm. The figure illustrates 176 countries covered by the global data set of urban ­­areas. We prefer to fit a nonlinear relationship in figure 2.6 rather than a linear relationship because this avoids a negative estimated intercept. Logically, the share of a country’s population living in multicity agglomerations must be zero bound. For a list of country abbreviations, see annex 2A. excessive primacy in ­­1990.19 Out of these 24, FIGURE ­­2.7  Urban Primacy, by Region 11 were in the LAC region.20 70 Yet, the strong persistence often inher- Share of urban population living in ent in urban systems still sees high urban 60 largest urban area (%) primacy rates in the region even if the 50 original factors have ­­ dissipated. 21 When 40 we look at current rates of urban primacy (measured using our global data set of 30 consistently defined urban areas), we see a 20 marked difference between the Caribbean 10 and Central and South A merica 0 (figure ­­2 .7). an c sia rth rica ica a ica ica sia The above simple comparisons may, how- cifi ric be hA lA er er er Af Af Pa Am m m rib ra ut an rth ever, be ­­misleading. It is natural to expect hA lA d nt Ca So an ar No Ce ra ut ah sia nt No that Caribbean countries, given their small So d nd -S Ce an tA b ta Su s pe sizes, will tend to exhibit higher primacy, and Ea as ro eE Eu dl Henderson (2000) also reports that the id M “optimal” primacy, when the rate of long-run Source: Calculations based on analysis of urban areas defined using the cluster algorithm of Dijkstra economic growth is maximized, is decreasing and Poelman (2014), as applied to LandScan 2012 gridded population ­­data. with a country’s population ­­ size.22 It is more Note: For each region, the figure shows the unweighted mean urban primacy rate across ­­countries. Urban primacy is defined as the share of a country’s urban population living in its largest urban ­­area. relevant to compare Caribbean countries with The figure is based on the nontrimmed global sample of urban ­­areas. North America comprises comparators. Figure ­­ their ­­ 2.8, panel a, shows Bermuda, Canada, and the United ­­States. 64   RAISING THE BAR FIGURE ­­2.8  Urban Primacy, LAC Countries and Comparators a. Caribbean 100 Share of urban population living 90 in largest urban area (%) 80 70 60 50 40 30 20 10 0 ica a he os a vis s ia go lic iti a ba a ne ud ub ad aic uc Ha ub ad Cu ,T ba Ne in di en rb Ar m .L m as rb ep To na nd Ba Ja Gr St Do m Ba nR re d sa ha d an eG an ica itt Ba ad th ua in .K id m d St tig in an Do An Tr nt ce in .V St b. Central America 80 Share of urban population living 70 in largest urban area (%) 60 50 40 30 20 10 0 Costa Rica Panama El Salvador Nicaragua Belize Guatemala Honduras Mexico c. South America 100 90 Share of urban population living 80 in largest urban area (%) 70 60 50 40 30 20 10 0 e y y a ru a ile a r a RB zil do ua a n in ivi bi am Pe a gu Ch ya t m , Br ua l ug ela en Bo rin Gu ra lo Ec Ur g zu Pa Co Su Ar ne Ve LAC country Average comparators LAC World Development Indicators Source: Calculations based on analysis of urban areas defined using the cluster algorithm of Dijkstra and Poelman (2014), as applied to LandScan 2012 gridded population data, and World Development Indicators data (­­http://data.worldbank.org/data-catalog/world-development​-indicators). Note: The comparators for each LAC country were selected as described in box ­­2.1. “Average comparators” refers to the unweighted mean urban primacy rate in the comparator countries, and “LAC World Development Indicators” refers to a LAC country’s urban primacy rate for 2012 as reported in the World Development Indicators. LAC = Latin America and the Caribbean. T he M an y D imensi o ns o f U rbani z ati o n and the P r o du c ti v it y o f Cities in l a c    65 that on such a comparison 7 out of 14 there is any relationship between the three Caribbean countries exhibit urban primacy key dimensions of urbanization examined— notably above the average for their corre- density, MCAs, and urban primacy—and ­­ sponding sets of comparator countries. For national GDP per capita ­­ levels. For 2012, the other seven, there is either no notable dif- we examine the relationship between a ference or the urban primacy rate is less than country’s (natural log) GDP per capita and the average for the comparator ­­ countries. (i) two alternative measures of urban den- Benchmarking urban primacy rates in sity; (ii) the share of its overall population LAC countries against those in their compar- living in MCAs; and (iii) its urban primacy ator countries also reveals that, although rate. On (i), the two measures of urban den- ­­ average urban primacy rates may not appear sity that we explore are, first, the share of a particularly high in either Central or South country’s overall population living in dense America, in some of the countries in these urban areas, and, second, the (natural log) subregions high primacy may still be a poten- mean density of a country’s urban areas tial problem for overall national productivity weighted by the share of each urban area in and economic growth (figure 2.8, panels b a country’s overall urban population (“Log and c). In Central America, Costa Rica and (Weighted Density)” in table 2.3). Panama have urban primacy rates that nota- Because the relationships presented are bly exceed the averages for their comparator correlations, the following results, even if countries. In South America, Argentina, ­­ consistent with theories outlined previ- Chile, Paraguay, Peru, Suriname, and ously, 24 cannot be regarded as providing Uruguay all have urban primacy rates that causal ­­evidence.25 appear high against their ­­ comparators. Table ­­2.3 presents the results of several Several of these countries appear among the regressions, all estimated using a single list that Henderson (2000) identifies as suf- global cross-section of 169 countries, of a fering from excessive primacy in ­­ 1990. More country’s (natural log) GDP per capita on generally, relative to their comparison groups, dimensions of ­­ urbanization. Throughout 15 out of 35 LAC countries exhibit high, and columns 1–3 a country’s GDP per capita potentially excessive, urban ­­primacy. continues to be positively and significantly For comparison, figure ­­ 2.8 also reports correlated with the share of its overall pop- urban primacy rates for 2012 from the ulation living in urban areas (its urban World Bank’s World Development Indicators share), as measured using the cluster algo- (WDI). Urbanization metrics reported in ­­ rithm, even after accounting for other WDI, including urban primacy, are based dimensions of ­­u rbanization. 26 The results on national definitions of urban ­­ areas. 23 As in columns 1a and 1b show that, at any can be seen, levels of urban primacy calcu- given urban share, a country’s GDP per lated using our global data set conform well capita is negatively related with the two with those reported in W DI for both alternative measures of urban ­­ density. Caribbean and South American ­­ countries. Although this negative relationship is insig- Large differences are, however, apparent for nificant when the measure of densit y Costa Rica, El Salvador, Nicaragua, and is “Percentage of population in dense,” it is ­­Panama. highly significant when the measure is “Log (Weighted Density)”—compare col- umn 1b with column 1a. In the latter case, Implications for National a 10 percent increase in weighted urban Productivity: Density and density is associated with about a 5 percent MCAs Matter, but Urban Primacy drop in GDP per ­­ capita. In columns 1a Does Not and 1b, the share of a country’s overall In this section, we analyze whether, con- population living in MCAs bears no rela- trolling for a country’s overall urban share, tionship to GDP per ­­ capita. 66   RAISING THE BAR TABLE ­­2.3  Cross-Country Regression of Log(GDP per Capita) on Different Dimensions of Urbanization (1a) (1b) (2a) (2b) (3a) (3b) *** *** *** *** *** Urban share (%) ­­0.047 ­­0.036 ­­0.044 ­­0.036 ­­0.043 ­­0.034*** ­­(0.008) ­­(0.005) ­­(0.009) ­­(0.005) ­­(0.008) ­­(0.005) Percentage of population in dense ­­− 0.010 ­­− 0.006 ­­− 0.006 ­­(0.007) ­­(0.008) ­­(0.008) Log(Weighted Density) ­­− 0.477*** ­­− 0.415** ­­− 0.432*** ­­(0.140) ­­(0.163) ­­(0.163) Percentage of population in MCAs ­­− 0.002 ­­0.002 ­­− 0.004 ­­− 0.001 ­­− 0.003 ­­0.001 ­­(0.006) ­­(0.006) ­­(0.006) ­­(0.007) ­­(0.006) ­­(0.007) (North America) × (Percentage of ­­ ­­0.032*** ­­0.025*** ­­0.026** ­­0.019** Population in MCAs) ­­ 0.011) ( ­­(0.008) ­­ 0.011) ( ­­(0.008) (Western Europe) × (Percentage of ­­ ­­0.032*** ­­0.024*** ­­0.030*** ­­0.022** Population in MCAs) ­­(0.012) ­­(0.009) ­­ 0.011) ( ­­(0.009) ­­ (South America) × (Percentage of ­­0.006 ­­0.010* ­­0.005 ­­0.011* Population in MCAs) ­­(0.005) ­­(0.006) ­­(0.005) ­­(0.006) ­­ (Central America) × (Percentage ­­0.006 ­­0.008 ­­0.007 ­­0.009 of Population in MCAs) ­­(0.006) ­­(0.007) ­­(0.006) ­­(0.007) (Caribbean) × (Percentage of ­­− 0.013 ­­− 0.011 ­­− 0.012 ­­− 0.009 Population in MCAs) ­­(0.017) ­­(0.017) ­­(0.017) ­­(0.016) Urban primacy (%) ­­− 0.014 ­­− 0.020 ­­(0.013) ­­(0.013) [Urban primacy (%)]2 ­­0.000 ­­0.000 ­­(0.000) ­­(0.000) Constant ­­6.716*** ­­10.614*** ­­6.738*** ­­10.115*** ­­7.123*** ­­10.780*** ­­(0.256) ­­(1.165) ­­ 0.257) ( ( ­­ 1.339) ­­(0.416) ­­ 1.264) ( No. of countries 169 169 169 169 169 169 Adjusted R2 ­­0.341 ­­0.388 ­­0.349 ­­0.389 ­­0.346 ­­0.388 Source: Calculations based on analysis of global data set of urban areas as constructed using the cluster algorithm of Dijkstra and Poelman (2014) and the World Development Indicators data (­­http://data.worldbank.org/data−catalog/world-development-indicators). Note: The dependent variable is the natural log of GDP per capita in 2012 international dollars (purchasing power parity exchange rates); robust standard ­­errors. “Urban share (%)” denotes the percentage share of a country’s overall population living in urban areas; “Percentage of population in dense” denotes the percentage share of a country’s overall population living in dense urban areas, where a dense urban area is one that has a mean population density that exceeds the global median for all urban areas; “weighted density” denotes the mean density of urban areas within a country weighted by the share of each urban area in a country’s overall urban population; “Percentage of population in MCAs” denotes the percentage share of a country’s overall population living in MCAs, where an MCA is defined as an urban area with two or more cities, each of which has a population of at least ­­100,000. GDP = gross domestic product; MCA = multicity agglomeration. * p < ­­0.1. **p < ­­0.05. ***p < ­­0.01. Columns 2a and 2b then introduce inter- large metropolitan areas, in which case an actions between the share of a country’s pop- increase in the share of the population living ulation living in MCAs and the country’s in such areas may be expected to lead to a net ­­ region. The intention is to explore whether positive increase in GDP per ­­ capita. there are heterogeneous effects across Consistent with this hypothesis, we see ­­ regions. For example, we might expect that that, for both North America and Western more developed regions have been more suc- Europe, there are positive and statistically sig- cessful in designing and implementing insti- nificant interaction effects with “Percentage of tutions to overcome the coordination population in MCAs,” and that this is the case challenges associated with the governance of irrespective of the measure of urban density T he M an y D imensi o ns o f U rbani z ati o n and the P r o du c ti v it y o f Cities in l a c    67 ­­ used in the regression specification. Where the ­­settlements. Here, the estimated coefficient measure of density is “Log (Weighted on “Percentage of population in MCAs” is Density),” a 1 percentage point increase in the not only negative but also statistically signifi- share of the overall population living in MCAs cant in several of the specifications is associated with a ­­2.5 percent increase in 2D). In North American and Western (annex ­­ GDP per capita for North American countries European countries, however, this negative and an almost identical ­­ 2.4 percent increase ­­ effect is more than overturned. By contrast, for Western European countries—see column for LAC countries, the effect of an increase in ­­ 2b. This is consistent with the idea that coun- “Percentage of population in MCAs” on tries in these regions have succeeded in devel- GDP per capita remains ­­negative. oping and implementing institutions that, although perhaps not completely solving the governance challenges of large metropolitan International Benchmarking of areas, address them sufficiently to allow for a LAC Urban Areas’ Productivity: net positive effect on productivity. ­­ For South Better Than Average, but America, there is also a (marginally) signifi- Lagging the Global Frontier cant positive interac tion effec t with Whereas the analysis in the chapter has “Percentage of population in MCAs” when focused thus far on different dimensions of the measure of density is “Log (Weighted urbanization and their relationship to GDP Density),” but not when it is “Percentage of per capita at the national level, it now turns to population in dense.” For South America, a an analysis of productivity measured at the 1 percentage point increase in “Percentage of level of individual urban areas, again bench- population in MCAs” is associated only with marking against the rest of the ­­ world. Ideally, 1.0 percent increase in GDP per ­­ a ­­ capita. we would like to be able to use subnational Finally, for both the Caribbean and Central economic accounts data to construct measures America, there are no significant interactions of labor and total factor productivity (TFP) at with “Percentage of population in MCAs,” level. However, most the individual urban area ­­ regardless of the measure of ­­ density. countries lack such data and, where they are These results suggest that LAC countries available, as, for example, for Brazil, India, have yet to reach the point of institutional and the European Union, they relate to subna- maturity in the governance of large metro- tional administrative areas whose boundaries politan areas where they can fully leverage may only crudely approximate those of “true” gain.27 This is these areas for net productivity ­­ urban areas (see chapters 1 and ­­ 6). especially true for Caribbean and Central To overcome this challenge, we turn to American ­­countries.28 nighttime lights data, which have the Columns 3a and 3b of table 2.3 ­­ further advantage of being globally available at a investigate the effects of urban primacy on fine spatial resolution, and so allow for the GDP per capita conditional on other dimen- construction of a consistent proxy measure sions of ­­urbanization. In these columns, of economic activity for our full global although the estimated coefficient on urban sample of urban ­­ areas. We proxy each ­ p rimacy is negative, it is statistically urban area’s economic activity by the total ­­ insignificant. Hence, unlike Henderson light emitted from the area at night—a (2000), we find no evidence that a country’s measure that, following, for example, urban primacy has a significant negative Addison and Stewart (2015), we refer to as effect on its GDP per ­­capita. the “sum of ­­ lights.” This proxy measure, The aforementioned results on MCAs are calculated for 2015 using data averaged strengthened when we redefine an MCA as over all cloud-free nights, then acts as the an urban area with two or more settlements, basis for constructing a measure of produc- irrespective of the populations of those tivity (box ­­2 .3). 68   RAISING THE BAR BOX ­­2.3  VIIRS Nighttime Lights Data The use of nighttime lights data to proxy for eco- Evidence of the suitability of the new VIIRS nomic activity has, since the seminal work of data to proxy for economic activity is provided Henderson, Storeygard, and Weil (2011, 2012), ­­ in table B2.3.1. Column 1 reports the results, for become a well-established ­­ practice. The 2015 2015, from a regression of GDP levels on a mea- annual composite product that we use in this chap- sure of economic activity (the “sum of lights”) ter, however, differs from that used in most previ- derived from the VIIRS data for a global sample ous ­­r esearch. a Most previous research relied on of 181 countries, whereas columns 2 and 3 report nighttime lights data derived from sensors on board corresponding results for a sample of 31 LAC meteorological satellites that were part of a program respectively. countries and Brazil’s municipalities, ­­ 1960s.b Instead, we use night- that originated in the ­­ In all three cases, the VIIRS data are very strongly time lights data from a new satellite instrument, the positively correlated with GDP (R 2 values range Visible Infrared Imaging Radiometer Suite (VIIRS), from ­­ 0 .96). And, as shown by columns 4 0.77 to ­­ launched in 2011, which provides much higher res- through 6, these correlations remain significant olution ­­data. The new data also overcome several even after controlling for population, which sug- problems with the “old” data, such as “overglow” gests that the VIIRS data are picking up varia- or “blooming,” which cause light to spill over the tion in productivity, in addition to variation in emitted.c,d geographic area from which it is ­­ ­­p opulation. TABLE B2.3.1  Regression of Log(GDP) on VIIRS Nighttime Lights Data, 2015 (3) (6) (1) (2) Brazilian (4) (5) Brazilian Global LAC municipalities Global LAC municipalities Log(NTL) 0.780*** 0.920*** 0.837*** 0.513*** 0.549*** 0.454*** (0.032) (0.032) (0.006) (0.040) (0.060) (0.009) Log(Population) 0.447*** 0.421*** 0.584*** (0.047) (0.061) (0.012) Constant 15.75*** 14.02*** 7.25*** 11.88*** 11.87*** 3.94*** (0.384) (0.381) (0.037) (0.428) (0.370) (0.078) No. of 181 31 5,418 181 31 5,418 observations R2 0.826 0.956 0.769 0.890 0.981 0.835 Source: Analysis based on nighttime lights data from the VIIRS 2015 annual composite product (https://ngdc.noaa.gov/eog/viirs/download_dnb_composites.html); World Bank World Development Indicators data (http://data.worldbank.org/data-catalog/world-development-indicators); and Instituto Brasileiro de Geografia e Estatística data. Note: For the global and LAC country samples, GDP is measured in constant international dollars (2005 PPP exchange rates); for Brazilian municipalities, GDP is measured in current local currency units. log(NTL) denotes the natural logarithm of an area’s “sum of lights” in 2015 as calculated using the VIIRS data; log(Pop) denotes the natural logarithm of an area’s population in 2015. Robust standard errors are in parentheses. GDP = gross domestic product; LAC = Latin America and the Caribbean; NTL = nighttime lights; PPP = purchasing power parity; VIIRS = Visible Infrared Imaging Radiometer Suite. *p < 0.1. **p < 0.05. ***p < 0.01. a. All nighttime lights data products, including the 2015 annual composite product that we use in this chapter, are produced by the National Oceanic and Atmospheric Administration of the ­­U.S. government (­­https://ngdc.noaa.gov/eog). U.S. Department of Defense’s Defense Meteorological Satellite Program. b. The satellites were part of the ­­ c. For example, the Pacific Ocean can be lit up as far as 50 km from the coastline near Los Angeles (Pinkovskiy ­­ 2013). d. Chapters 4 and 6 use an algorithm attributable to Abrahams, Lozano-Gracia, and Oram (2017) to address the problem of “overglow” in the old ­­ data. ­­ nalysis. Despite the superiority of the new data, the old data are preferred in these chapters because they require time series data for their a T he M an y D imensi o ns o f U rbani z ati o n and the P r o du c ti v it y o f Cities in l a c    69 TABLE ­­2.4  The 15 Urban Areas in the LAC Region with the Highest Estimated Economic Activity, as Measured by Nighttime Lights Data, 2015 Relative ­­Population No. of cities in ­­ Rank Country Urban area sum of lights Population density urban area 1 Argentina Buenos Aires 388.0 14,183,924 4,167 30 2 Brazil São Paulo 284.7 20,588,698 6,455 23 3 Mexico Mexico City 218.9 19,782,701 7,462 16 4 Brazil Rio de Janeiro 161.9 9,932,480 5,730 7 5 Chile Santiago 105.6 5,837,310 5,238 3 6 Peru Lima 96.4 9,056,851 8,931 22 7 Brazil Brasilia 71.5 2,019,961 4,126 1 8 Brazil Porto Alegre 66.5 3,453,232 3,299 9 9 Brazil Belo Horizonte 55.2 4,181,234 4,937 6 10 Mexico Monterrey 53.3 3,870,579 4,373 8 11 Mexico Guadalajara 50.7 4,219,190 5,822 4 12 Colombia Bogotá 49.7 7,861,739 13,445 2 13 Brazil Campinas 49.5 2,304,343 2,609 4 14 Brazil Curitiba 49.0 2,773,894 3,003 4 15 Paraguay Asuncion 48.7 2,172,047 2,886 5 Source: Analysis of nighttime lights data from the 2015 VIIRS annual composite product (­­https://ngdc.noaa.gov​/­eog/viirs/download_dnb_composites​.html). Note: The relative sum of lights is the ratio of an urban area’s sum of lights to the unweighted mean sum of lights for all urban areas in the LAC region; the relative sums of lights, population, and population density are for the urban areas as derived using the cluster algorithm of Dijkstra and Poelman (2014); both population and population density are calculated using LandScan 2012 gridded population ­­data. “ ­­ No. of cities in urban area” refers to the number of cities with a population of at least 100,000 whose settlement points intersect the urban ­­area. LAC = Latin America and the Caribbean; VIIRS = Visible Infrared Imaging Radiometer Suite. On the basis of the nighttime lights data, a significant and positive mean difference in 2.4 shows the 15 urban areas in the table ­­ economic activity for urban areas in the LAC LAC region with the highest estimated region against those in the rest of the world absolute economic ­­activity. 29 These areas (column ­­2). LAC urban areas tend to have lev- correspond with some of the largest urban els of economic activity significantly greater areas in the region with all but one, than that predicted by their populations—­ Brasilia, an ­­ MCA. However, an urban relative to the rest of the world they are, on area’s economic activity rank does not nec- average, more ­­productive. Because we do not essarily coincide with its population rank: control for an urban area’s capital stock given Buenos Aires, with an estimated popula- data limitations, this higher productivity may 14.2 million, beats both São Paolo tion of ­­ be attributable to a higher capital–labor ratio, and Mexico City, with estimated popula- higher TFP, or both—in other words, the ­­ tions of about 20 million. This suggests higher level of productivity most accurately that Buenos Aires is the more productive, at productivity. represents higher labor ­­ least from a labor productivity ­­standpoint. As seen in column 3, and in figure 2.9, ­­ Notwithstanding this observation on this better than average performance is driven Buenos Aires, column 1 in table ­­ 2.5 shows ­­ by urban areas in South America and Mexico. that, globally, there is a significant positive 2.9 that However, it is also clear from figure ­­ relationship between an urban area’s popula- even these areas fall short of the global “fron- tion and its economic activity, as measured by tier” of productivity performance (marked by ­­ its sum of lights, in 2015. There is, however, ­­ the outer envelope of data points in the figure). 70   RAISING THE BAR TABLE 2.5  Relationship between Log(Nighttime Lights) and Compared with South America and Mexico, Log(Population), All Urban Areas Globally urban areas in the Caribbean and the rest of (1) (2) (3) Central America are, on average, less Log(Population) 1.281*** 1.271*** 1.270*** productive. The relationships between popu- ­­ (0.041) (0.040) (0.040) lation and estimated economic activity for these two subregions are statistically indistin- Latin America and the Caribbean 1.164*** guishable from the global ­­ relationship. Hence, (0.361) from a global perspective, the productivity South America (and Mexico) 1.411*** performance of urban areas in the Caribbean (0.324) and Central America (excluding Mexico) may Central America (except Mexico) –0.123 be judged ­­“average.”30 (0.401) We can also use the residual from the Caribbean –0.307 regression in column 1 of table ­­ 2.5 as a mea- (0.940) sure of an urban area’s productivity relative Constant –7.637*** –7.674*** –7.660*** to its population-based predicted level, where a positive (negative) value implies higher (0.628) (0.600) (0.601) (lower) than predicted ­­ productivity. Again, No. of observations 63089 63089 63089 given that this regression does not control for Adjusted R2 0.351 0.375 0.382 an urban area’s capital stock, this measure Source: Calculations based on nighttime lights data from the 2015 VIIRS annual composite product can best be thought of as a measure of (rela- (https://ngdc.noaa.gov/eog/viirs/download_dnb​_composites.html). Note: The dependent variable is the natural log of an urban area’s sum of lights, where urban areas tive) labor ­­productivity. have been identified by applying the cluster algorithm of Dijkstra and Poelman (2014) to LandScan On the basis of this measure, figure ­­ 2.10 2012 gridded population data; standard errors are clustered at the country level. VIIRS = Visible Infrared Imaging Radiometer Suite. benchmarks, for each LAC country, the *p < 0.1. **p < 0.05. ***p < 0.01. mean productivity in its urban areas against FIGURE ­­2.9  Relationship between Log(Nighttime Lights) and Log(Population), All Urban Areas Globally 15 10 Log(Nighttime Lights) 5 0 −5 8 12 16 20 Log(Population) Quadratic fit for non-LAC countries Non-LAC country Quadratic fit for South American countries and Mexico South America, with Mexico Quadratic fit for Central American countries, without Mexico Central America, without Mexico Quadratic fit for Caribbean countries Caribbean Source: Calculations based on nighttime lights data from the 2015 VIIRS annual composite product (­­https://ngdc​.noaa.gov/eog/viirs/download_dnb​ _composites.html). Note: Log(Nighttime Lights) denotes the natural logarithm of an urban area’s sum of lights for 2015, where urban areas have been identified by applying the cluster algorithm of Dijkstra and Poelman (2014) to LandScan 2012 gridded population d ­­ ata. LAC = Latin America and the Caribbean; VIIRS = Visible Infrared Imaging Radiometer Suite. T he M an y D imensi o ns o f U rbani z ati o n and the P r o du c ti v it y o f Cities in l a c    71 FIGURE ­­2.10  Mean Urban Area Productivity in LAC Countries Benchmarked against International Comparators a. Caribbean 3 2 1 0 –1 –2 –3 –4 go he a vis a os ia a a ica ba lic s iti ne ud ub aic ad uc Ha ub ad Cu ,T ba Ne in di en rb Ar m .L m as rb ep To na nd Ba Ja Gr St Do m Ba nR re d sa ha d an eG an ica itt Ba ad th ua in .K id m d St tig in an Do An Tr nt ce in .V St b. Central America 3 2 1 0 –1 –2 –3 –4 Belize Mexico Costa Rica Panama Honduras El Salvador Nicaragua Guatemala c. South America 5 4 3 2 1 0 –1 –2 –3 –4 a y ay ile e zil BR a r a u na do ua tin ivi bi am r Pe a u Ch ya m la, Br ua l ug ag en Bo rin Gu lo ue Ec r Ur g Pa Co Su Ar z ne Ve LAC country Comparator 1 Comparator 2 Comparator 3 Source: Based on nighttime lights data from the 2015 VIIRS annual composite product (­­https://ngdc.noaa.gov​/­eog/viirs/download_dnb_composites.html). Note: Productivity is measured using the residuals from the regression in column 1 of table ­­2.5. A country’s mean urban area productivity is given by the mean of these residuals across its urban ­­areas. LAC = Latin America and the Caribbean; VIIRS = Visible Infrared Imaging Radiometer Suite. 72   RAISING THE BAR the corresponding means for each of its generally, factors—from moving to the comparator countries, where the compara- urban areas in which they earn the highest tor countries are again selected following returns, which is where they will be most the methodology in box ­­ 2.1. In most cases, ­­ productive. Hence, it may be expensive, in mean productivity in urban areas is lower monetary or nonmonetary terms, for a than in at least one of the comparator coun- ­­ worker to move between urban areas. One tries, but in only two cases—Haiti and reason could be a lack of domestic market Jamaica—does mean productivity lag all integration, which could include an inade- comparators. T his matches the th ree ­­ quately developed national transport net- i mpression g iven by fig u re ­­ 2 .9 that , work, another could be that housing is much although LAC urban areas are by no means more ­­expensive. the worst performers on productivity, they These two explanations of persistent can certainly improve a ­­ lot. The only LAC differences in productivity both imply that countries for which mean productivity the contribution of urban areas to aggre- across urban areas is higher than in all three gate GDP will fall below potential, but the comparator countries are Trinidad and two differ on their welfare implications: Tobago and the Bahamas in the Caribbean, when persistent productivity differences and Argentina, Ecuador and, more surpris- are driven by differences in amenities ingly, República Bolivariana de Venezuela across urban areas and there is perfect fac- in South ­­A merica. tor mobility, the only reason why a worker would choose not to relocate from a less to a more productive area is because he or she Productivity is Highly Dispersed is already happy there. ­­ By contrast, when across LAC Urban Areas persistent productivity differences are Dispersion of productivity across a coun- driven by barriers to migration, workers try’s urban areas provides direct informa- may be “trapped,” in which case a reduc- tion on inequalities in performance and tion in barriers would increase the aggre- can yield indirect clues on their aggregate gate contribution from urban areas to contribution to GDP and productivity national GDP and productivity, and would ­­ nationally. This is because, in a world of be welfare ­­ enhancing. However, even perfect factor mobility, we expect produc- where productivity differences are driven tivity across urban areas to tend to equal- by amenities and everyone is happy where ize at the margin as people and firms they are, the possibility of an improvement gravitate to the places where they will earn in both the aggregate contribution of urban the highest returns, which is where they areas to national GDP and to welfare will be most ­­productive. To the extent that remains. Hence, under free mobility, a ­­ productivity differences persist under such worker may be “enticed” to move to a conditions, we would expect these to be more productive urban area if the things the result of differences in, for example, that he or she values become more readily urban amenities that workers value and available in that area or, equivalently, the that, all else constant, may persuade them things that he or she dislikes become less to accept a lower wage in one urban ­­ prevalent. A reduction in pollution or a re a t h a n i n a not her ( Ro s e n 19 79; crime in the more productive area might, Roback ­­1982). for example, act as an ­­enticement. In this An alternative explanation of persistent case, relocation will benefit aggregate GDP differences in productivity across urban and ­­welfare. areas at the margin is that they are due to Productivity dispersion across urban barriers that prevent workers—and, more areas in each of the three LAC subregions T he M an y D imensi o ns o f U rbani z ati o n and the P r o du c ti v it y o f Cities in l a c    73 is notably higher than in North America FIGURE ­­2.11  Distribution of Productivity across Urban Areas, (figure ­­2 .11). 31 The distributions of pro- Selected Regions ductivity in urban areas in the three LAC subregions contain notable overlaps with 0.8 not only the corresponding distribution for North America but also the corresponding 0.6 distribution for ­­ SSA. Whereas the most productive urban areas in the LAC region Density r iva l m a ny N o r t h A m e r i c a n u rb a n 0.4 areas, the least productive trail the best performers in ­­S SA. 32 0.2 Of course, the higher productivity disper- sion across urban areas in the three LAC subregions could be largely attributable to 0 between-country productivity differences −10 −5 0 5 10 rather than within-country productivity dif- Productivity ferences in each (sub)region. ­­ This would be Caribbean Central America, without Mexico consistent with, for example, a story of a South America and Mexico Sub-Saharan Africa relative lack of regional integration in the North America LAC region compared with North ­­ America. However, we also observe high productivity Source: Calculations based on nighttime lights data from the 2015 VIIRS annual composite product dispersion across urban areas within indi- ­­(https://ngdc.noaa.gov/eog/viirs/download_dnb​_composites.html). Note: The figure shows density plots of the residuals from the regression in column 1 of vidual LAC countries against their compar- table ­­2.5. These residuals may be interpreted as measuring productivity across urban areas ator countries (figure 2.12).­­ Dispersion where urban areas have been identified by applying the cluster algorithm of Dijkstra and Poelman (2014) to LandScan 2012 gridded population d ­­ ata. VIIRS = Visible Infrared Imaging within LAC countries is higher than in all Radiometer Suite. three comparators in virtually all cases, Trinidad and Tobago and Uruguay ­ aside. In this instance, we restrict the selection of dispersion of productivity across urban areas comparator countries to high-income coun- within a country and the (natural) log of the tries, without imposing any restrictions on country’s density of roads (length of roads the region from which the comparators are per 100 km2 of land area), which we take as a drawn, but otherwise follow the methodol- prox y m e a su re of dom e s t ic m a rke t ogy for selecting comparators set out in ­­i ntegration. 33 As discussed in detail in box ­­ 2.1. The rationale is that we expect chapter 4, national road densities in LAC such countries to exhibit high domestic countries lag those in the most developed market ­­ integration. Therefore, if we observe ­­countries. high dispersion in LAC countries against The above evidence matches the idea their comparators, this suggests that the that high dispersion of productivity across high dispersion may, at least in part, be LAC urban areas is driven by barriers asso- driven by a relative lack of domestic market ciated with weak domestic market integra- ­­integration. ­­ tion that prevents workers from moving. Further evidence that a lack of domestic Easing these barriers would have beneficial market integration may be contributing to effects not only for aggregate GDP and h ig h pro duc t iv it y d i sp er sion ac ro s s productivity but also for aggregate welfare. LAC country urban areas is provided by A not her, potential ly complement a r y, 2.13, which shows, for a global sample figure ­­ explanation is grounded in persistent dif- of 112 countries, a highly statistically signifi- ferences in amenities across urban areas cant negative relationship between the (box ­­2 .4). 74   RAISING THE BAR FIGURE ­­2.12  Productivity Dispersion (Measured by the Coefficient of Variation) across Urban Areas in LAC Countries Benchmarked against High-Income International Comparators a. Caribbean 180 160 140 120 100 80 60 40 20 0 Dominican Cuba Jamaica Haiti Trinidad and Republic Tobago b. Central America 800 150 700 600 100 500 400 300 50 200 100 0 0 Honduras El Salvador Nicaragua Guatemala Panama Mexico Costa Rica c. South America 600 100 500 80 400 60 300 40 200 100 20 0 0 Peru Guyana Colombia Bolivia Ecuador Venezuela, Brazil Chile Paraguay Uruguay Argentina RB LAC country Comparator 1 Comparator 2 Comparator 3 Source: Calculations based on nighttime lights data from the 2015 VIIRS annual composite product ( ­­ https://​ngdc​.noaa.gov/eog/viirs/download_dnb​ _composites.html). Note: Productivity is measured using the residuals from the regression in column 1 of table ­­2.5. Productivity dispersion across a country’s urban areas is measured by the coefficient of ­­variation. Comparators for each LAC country are restricted to high-income countries, but with no restrictions as to which regions the comparators are drawn from. The method for selecting comparators is as otherwise described in box ­­2.1. LAC = Latin America and the Caribbean; VIIRS = Visible Infrared Imaging Radiometer Suite. T he M an y D imensi o ns o f U rbani z ati o n and the P r o du c ti v it y o f Cities in l a c    75 ­­Conclusions FIGURE ­­2.13  Productivity Dispersion across Urban Areas in a Country Is Negatively Correlated with National Road Density, Three main initial points emerge: urban areas 112 Countries in most LAC countries have high population density; the LAC region is home to many 3 Within-country dispersion of city productivity MCAs that present complicated governance challenges for delivering infrastructure and basic urban services;34 and many countries 2 still exhibit relatively high urban primacy, even though the era of import substitution PER GTM PAN BOL COL industrialization has long passed and SLV CHL CRI LAC countries have seen political and fiscal 1 PRY MEX JAM decentralization reforms over the last three ARG NIC BRA ­­decades. On national GDP per capita, we have also DMA seen evidence to suggest that, in the absence 0 of an adequate enabling policy environment, −2 0 2 4 6 8 the high density of the LAC region’s urban Road density (log) areas may be having a negative effect, and LAC countries Non-LAC countries that governance challenges associated with the region’s MCAs may be constraining their Source: Calculations based on nighttime lights data from the 2015 VIIRS annual composite product ­­(https:// it. By contrast, there is no evi- contribution to ­­ ngdc.noaa.gov/eog/viirs/download_dnb​_composites.html) and road density data from the World Development Indicators database (http://data.worldbank.org/data-catalog/world-development​-indicators). dence to suggest that high urban primacy Note: Productivity is measured using the residuals from the regression in column 1 of table ­­2.5. Productivity ­­ rates are acting as such a drag. dispersion across a country’s urban areas is measured by the interquartile range of the distribution of ­­productivity. Road density is the ratio of the length of the country’s total road network to the country’s land Although urban areas in South America area and is measured in kilometers per 100 ­­km2 of land ­­area. LAC = Latin America and the Caribbean; and Mexico are relatively productive by VIIRS = Visible Infrared Imaging Radiometer Suite. For a list of country abbreviations, see annex 2A. BOX ­­2.4  Cities and Aggregate Growth: United States and Brazil In a recently published paper, Hsieh and Moretti through reducing barriers to migration or making (2017) present a model of spatial equilibrium the productive cities relatively more attractive, will that facilitates empirical analysis of the contribution improve both aggregate GDP and aggregate ­­ welfare. that cities make to aggregate gross domestic product Applying their framework to the United States (GDP). In the model, aggregate GDP is increasing in using data for 220 metropolitan areas, Hsieh and the total factor productivity (TFP) of each city in a Moretti find that, after conditioning on observ- country, but decreasing in the dispersion of nominal able worker characteristics, nominal wage disper- wages across cities in a ­­country. A high dispersion sion increased by a factor of two across ­­ U.S. cities of nominal wages reflects large marginal productiv- between 1964 and 2009, reflecting a worsening spa- ity differences across cities—a sure sign that not all ­­ tial distribution of workers. They calculate that, if it workers are living and working in the cities where U.S. GDP were not for this deteriorating distribution, ­­ they will be most ­­ productive. Policies that reduce the in 2009 would have been ­­ 13.5 percent higher than it dispersion of nominal wages across cities by facilitat- was. They attribute most of the increased nominal ­­ ing the movement of workers to cities in which they wage dispersion across ­­ U.S. cities to constraints on will be more productive—the high TFP cities—either housing supply, arising from, for example, tight land (continued) 76   RAISING THE BAR BOX ­­2.4  Cities and Aggregate Growth: United States and Brazil (continued) use restrictions in high TFP cities such as New York; declined in the metro areas in the period, with the San Francisco; and San Jose, California. ­­ Hsieh and largest declines in the highest productivity ­­ cities. Moretti contend that, by contributing to a dearth of Still, Brazil remains a long way from an efficient affordable housing, these restrictions have deterred allocation of workers, and Bastos reports that wage workers from moving there. ­­ They calculate that, dispersion across Brazilian metro areas remains if land use restrictions in these three cities were ­­ higher than across their U.S. counterparts. This sug- ­­ brought into line with those in the median ­­U.S. city, gests that the potential contribution to aggregate GDP ­­ U.S. GDP would increase by nearly 10 ­­ percent. of Brazil’s most productive metro areas has still to be In his background paper for this book, Bastos leveraged. As with the United States, a shortage fully ­­ (2017) takes the Hsieh and Moretti framework and of affordable housing in the most productive metro applies it to the Latin America and the Caribbean areas is a ­­ culprit. High-wage metro areas in Brazil region’s biggest economy and most populous coun- experienced a larger increase in their formal housing try, ­­ Brazil. Unlike Hsieh and Moretti’s findings deficits between 2000 and 2010 than low-wage metro for the United States, he finds that the dispersion areas. Although informal housing presumably filled ­­ of nominal wages (conditioned on worker charac- the gap for some migrants to high-wage metro areas teristics) across 36 Brazilian metropolitan areas ­­ (see box 2.2), the poor quality of such housing may declined in ­­1999–2014. Employment and popu- have deterred would-be migrants, “trapping” them in lation growth over this period were fastest in the less productive ­­cities. metro areas with the highest (conditional) nomi- Despite the appeal of both the Hsieh and nal wages in ­­1999. This all points to an improving Moretti framework and Bastos’ empirical applica- ­­ spatial allocation of workers across metro areas. tion of it to Brazil, some caveats are in ­­ order. The Bastos presents evidence that this improved spa- most notable is, perhaps, that because of data avail- tial allocation may be attributable to a relative ability Bastos’ analysis is necessarily confined to improvement in living conditions in the most pro- the formal sector, which employs only a minority of ductive metro ­­ areas. For example, homicide rates ­­workers in Brazil. global standards, they lag the global frontier within LAC countries is also high relative to of productivity ­­performance. By contrast, high-income comparator countries, and urban areas in the Caribbean and Central appears to be at least partly driven by weak America (outside Mexico) appear average in internal market ­­ integ ration. Keeping terms. The impression of untapped global ­­ workers “trapped” in relatively unproduc- urban productivity gains continues to hold tive urban areas, this weakness may be con- when we compare LAC countries against straining the aggregate contribution of comparator countries that are similar both urban areas to national productivity and from a geographic viewpoint (whether an ­­welfare. island, non-island, or landlocked), and in In part II we turn to a more rigorous density. terms of their size and population ­­ empirical analysis that picks up on several of Behind the averages lies considerable pro- this chapter’s themes as they relate to the ductivity variation: whereas the most pro- determinants of cities’ productivity: internal ductive LAC urban areas are on a par with market integration (chapters 3 and 4), human many in North America, the least produc- capital (chapter 5), and fragmented urban tive are on a par with many in ­­ Africa. governance in large metropolitan areas Productivity dispersion across urban areas (chapter ­­6). T he M an y D imensi o ns o f U rbani z ati o n and the P r o du c ti v it y o f Cities in l a c    77 Annex 2A: List of Comparator Countries for Each LAC Country Global comparators High-income comparators     Code Country Comparator 1 Comparator 2 Comparator 3 Comparator 1 Comparator 2 Comparator 3 ATG Antigua and Micronesia, Cyprus Seychelles New Caledonia Iceland Bermuda Barbuda ­­Fed. ­­Sts. ARG Argentina Cambodia Ukraine Algeria Saudi Arabia Netherlands Poland ABW Aruba Micronesia, Cyprus Seychelles Bermuda New Caledonia Iceland ­­Fed. ­­Sts. BHS Bahamas, The New Caledonia Cyprus Cabo Verde Iceland New Caledonia Cyprus BRB Barbados Vanuatu Cyprus São Tomé and New Caledonia Iceland Malta Príncipe BLZ Belize Brunei Montenegro Djibouti Brunei Equatorial Estonia Darussalam Darussalam Guinea BOL Bolivia Moldova Turkmenistan Chad Austria Slovak Republic Czech Republic BRA Brazil China Turkey United States United States Canada Saudi Arabia CHL Chile Cambodia Sweden Cameroon Greece Portugal Belgium COL Colombia Myanmar Ukraine South Africa Spain Poland France CRI Costa Rica Cambodia Croatia Eritrea Croatia Finland Norway CUB Cuba Papua New Ireland Sri Lanka New Zealand Ireland Cyprus Guinea DMA Dominica Kiribati Cyprus Seychelles New Caledonia Iceland Bermuda DOM Dominican Papua New Ireland Sri Lanka New Zealand Ireland Cyprus Republic Guinea ECU Ecuador Cambodia Romania Senegal Greece Belgium Portugal SLV El Salvador Cambodia Belgium Togo Denmark Israel United Arab Emirates GRD Grenada Micronesia, Cyprus Seychelles New Caledonia Iceland Bermuda ­­Fed. ­­Sts. GTM Guatemala Cambodia Portugal Côte d’Ivoire Netherlands Belgium Greece GUY Guyana Timor-Leste Estonia Montenegro Equatorial Brunei Estonia Guinea Darussalam HTI Haiti Papua New Cyprus Sri Lanka New Zealand Ireland Cyprus Guinea HND Honduras Cambodia Bulgaria Benin United Arab Sweden Israel Emirates JAM Jamaica New Zealand Cyprus Ireland Ireland New Zealand Cyprus MEX Mexico Indonesia Turkey Iran, Islamic France Germany Italy ­­Rep. NIC Nicaragua Cambodia Bulgaria Denmark Denmark Finland Croatia PAN Panama Timor-Leste Georgia Bosnia and Kuwait Lithuania Croatia Herzegovina PRY Paraguay Serbia Turkmenistan Lao PDR Austria Slovak Republic Luxembourg PER Peru Cambodia Ukraine Angola Saudi Arabia Netherlands Poland (continued) 78   RAISING THE BAR ANNEX 2A  List of Comparator Countries for Each LAC Country (continued) Global comparators High-income comparators Code Country Comparator 1 Comparator 2 Comparator 3 Comparator 1 Comparator 2 Comparator 3 KNA ­­ t. Kitts and S Micronesia, Cyprus Seychelles Bermuda New Caledonia Iceland Nevis ­­Fed. ­­Sts. LCA ­­St. Lucia Tonga Cyprus São Tomé and New Caledonia Iceland Malta Príncipe VCT ­­St. Vincent and Micronesia, Cyprus Seychelles New Caledonia Iceland Bermuda the Grenadines ­­Fed. ­­Sts. SUR Suriname Brunei Montenegro Djibouti Brunei Equatorial Estonia Darussalam Darussalam Guinea TTO Trinidad and Fiji Cyprus Mauritius New Zealand Cyprus Ireland Tobago URY Uruguay Timor-Leste Norway Lithuania Kuwait Lithuania Oman VEN Venezuela, RB Malaysia Ukraine Mozambique Poland Netherlands Spain Annex 2B: Statistical Tests of Differences in Population, Area, and Population Density between LAC Countries and Their Comparators Log population Log area Log population density ( p values) ( p values) ( p values) Country > ­­comp.a t-testb KS testc > ­­comp.a t-testb KS testc > ­­comp.a t-testb KS testc Caribbean Cuba + ­­0.621 ­­0.913 − ­­0.000 ­­0.000 + ­­0.000 ­­0.000 Dominican Republic + ­­0.304 ­­0.233 − ­­0.000 ­­0.000 + ­­0.000 ­­0.000 Haiti − ­­0.085 ­­0.161 − ­­0.006 ­­0.027 + ­­0.031 ­­0.162 Jamaica − ­­0.562 ­­0.443 − ­­0.000 ­­0.000 + ­­0.000 ­­0.000 Central America Belize − ­­0.325 ­­0.662 − ­­0.093 ­­0.079 + ­­0.397 ­­0.007 Costa Rica + ­­0.025 ­­0.040 + ­­0.656 ­­0.014 + ­­0.012 ­­0.195 El Salvador + ­­0.000 ­­0.000 + ­­0.101 ­­0.000 + ­­0.047 ­­0.011 Guatemala + ­­0.000 ­­0.000 + ­­0.002 ­­0.000 + ­­0.235 ­­0.349 Honduras + ­­0.453 ­­0.421 − ­­0.562 ­­0.014 + ­­0.018 ­­0.056 Mexico − ­­0.000 ­­0.000 − ­­0.000 ­­0.000 + ­­0.001 ­­0.000 Nicaragua + ­­0.001 ­­0.004 − ­­0.853 ­­0.216 + ­­0.000 ­­0.000 Panama + ­­0.688 ­­0.049 − ­­0.000 ­­0.000 + ­­0.000 ­­0.000 South America Argentina − ­­0.738 ­­0.426 − ­­0.264 ­­0.000 + ­­0.240 ­­0.000 Bolivia + ­­0.108 ­­0.544 − ­­0.000 ­­0.000 + ­­0.000 ­­0.000 Brazil + ­­0.000 ­­0.000 − ­­0.000 ­­0.000 + ­­0.000 ­­0.000 Chile + ­­0.000 ­­0.000 + ­­0.085 ­­0.000 + ­­0.000 ­­0.000 (continued) T he M an y D imensi o ns o f U rbani z ati o n and the P r o du c ti v it y o f Cities in l a c    79 ANNEX 2B  Statistical Tests of Differences in Population, Area, and Population Density between LAC Countries and Their Comparators (continued) Log population Log area Log population density ( p values) ( p values) ( p values) Country > ­­comp.a t-testb KS testc > ­­comp.a t-testb KS testc > ­­comp.a t-testb KS testc Colombia + ­­0.007 ­­0.003 − ­­0.000 ­­0.000 + ­­0.000 ­­0.000 Ecuador + ­­0.000 ­­0.000 + ­­0.001 ­­0.000 + ­­0.000 ­­0.000 Guyana + ­­0.227 ­­0.080 − ­­0.065 ­­0.270 + ­­0.000 ­­0.000 Paraguay + ­­0.370 ­­0.883 − ­­0.000 ­­0.000 + ­­0.000 ­­0.000 Peru + ­­0.000 ­­0.001 − ­­0.000 ­­0.000 + ­­0.000 ­­0.000 Uruguay + ­­0.056 ­­0.009 − ­­0.000 ­­0.000 + ­­0.000 ­­0.000 Venezuela, RB + ­­0.000 ­­0.000 − ­­0.000 ­­0.000 + ­­0.000 ­­0.000 ­­ ata. Source: Calculations based on analysis of urban areas defined using the cluster algorithm of Dijkstra and Poelman (2014), as applied to LandScan 2012 gridded population d Note: The table shows results only for countries where one of the six hypothesis tests conducted shows a significant difference between a LAC country and its corresponding set of ­­ A. KS test = Kolmogorov-Smirnov test; LAC = Latin America and the Caribbean. comparator ­­countries. For a full list of comparator countries, see annex 2 a. This column indicates whether the mean of each variable across urban areas is greater (+) or less (–) than that for the pooled set of comparator country urban areas. b. This column represents a two-tailed and two-sample t-test of the difference in means between a country and its comparators. c. This column compares a country’s distribution with that of its comparators where the null hypothesis is that the distributions are identical. Annex 2C: List of Multicity Agglomerations in the LAC Region Relative sum of Population No. of cities in ­­ Rank Country Urban area lights Population density urban area 1 Brazil São Paulo 285 20,588,698 6,455 23 2 Mexico Mexico City 219 19,782,701 7,462 16 3 Argentina Buenos Aires 388 14,183,924 4,167 30 4 Brazil Rio de Janeiro 162 9,932,480 5,730 7 5 Peru Lima 96 9,056,851 8,931 22 6 Colombia Bogotá 50 7,861,739 13,445 2 7 Chile Santiago 106 5,837,310 5,238 3 8 Mexico Guadalajara 51 4,219,190 5,822 4 9 Brazil Belo Horizonte 55 4,181,234 4,937 6 10 Mexico Monterrey 53 3,870,579 4,373 8 11 Brazil Recife 42 3,465,982 6,461 5 12 Brazil Porto Alegre 67 3,453,232 3,299 9 13 Colombia Medellín 16 3,450,578 15,399 3 14 Dominican Republic Santo Domingo 25 3,431,292 6,027 2 15 Venezuela, RB Caracas 29 3,325,327 8,862 4 16 Brazil Fortaleza 40 3,272,611 6,260 3 17 Guatemala Ciudad de 30 3,061,338 4,992 4 Guatemala 18 Brazil Salvador 32 2,797,798 7,551 2 19 Brazil Curitiba 49 2,773,894 3,003 4 20 Ecuador Guayaquil 17 2,600,395 10,200 2 (continued) 80   RAISING THE BAR ANNEX 2C  List of Multicity Agglomerations in LAC (continued) Relative sum of Population No. of cities in ­­ Rank Country Urban area lights Population density urban area 21 Haiti Port-au-Prince 5 2,497,164 6,121 3 22 Brazil Campinas 50 2,304,343 2,609 4 23 Costa Rica San Jose 26 2,272,653 4,040 2 24 Paraguay Asuncion 49 2,172,047 2,886 5 25 Cuba La Habana 14 2,054,052 5,205 20 26 Mexico Toluca 25 2,021,447 2,405 2 27 Brazil Belem 21 2,005,080 8,325 2 28 Brazil Goiania 34 1,867,097 3,057 3 29 Colombia Barranquilla 21 1,859,324 8,875 2 30 El Salvador San Salvador 10 1,825,864 4,435 6 31 Bolivia La Paz 18 1,806,596 8,055 2 32 Brazil São Goncalo 25 1,786,076 4,011 3 33 Brazil Santos 34 1,465,472 4,120 5 34 Brazil Itaquari 34 1,404,090 3,672 5 35 Dominican Republic Santiago de los 11 1,256,166 2,622 2 Caballeros 36 Brazil Natal 16 1,138,317 4,381 3 37 Panama Panama 13 1,111,798 5,499 2 38 Colombia Bucaramanga 8 1,052,221 4,864 2 39 Mexico San Luis Potosi 13 1,029,379 4,885 2 40 Brazil Teresina 14 936,407 4,653 2 41 Mexico Cuernavaca 12 934,042 2,480 2 42 Argentina Greater Mendoza 37 927,595 2,316 4 43 Chile Viña del Mar 17 832,365 3,350 3 44 Peru Trujillo 6 821,578 8,326 2 45 Brazil São José dos 18 819,726 3,192 2 Campos 46 Brazil Coxipo da Ponte 19 803,896 2,976 3 47 Chile Talcahuano 16 759,765 3,269 2 48 Mexico Tampico 14 715,843 3,221 2 49 Mexico Heroica Veracruz 10 688,503 5,109 3 50 Colombia Pereira 3 631,139 7,821 2 51 Argentina San Juan 16 481,644 2,898 2 52 Brazil Volta Redonda 9 475,645 2,809 2 53 Peru El Tambo 2 431,053 12,497 2 54 Venezuela, RB Guarenas 5 399,325 6,136 2 Source: Calculations based on analysis of urban areas defined using the cluster algorithm of Dijkstra and Poelman (2014), as applied to LandScan (2012) gridded population data, and nighttime lights data from the 2015 VIIRS annual composite product (­­https://ngdc.noaa.gov/eog/viirs/ d ­­ ownload_dnb_composites.html). Note: Multicity agglomerations are identified by using Geographic Information Systems techniques to overlay a global layer of individual settlement points on a global map of urban areas, as derived using the cluster ­­algorithm. Each multicity agglomeration is named after the most populous settlement point that falls within its ­­area. Where an urban area intersects with two or more settlement points, each of which had an estimated population of 100,000 or more in 2000, we identify this as an multicity agglomeration. The global settlement point layer that we use is the Center for International Earth Science Information Network’s Global Rural-Urban Mapping Project Settlement Point Layer v 1.1 (­­http://sedac​ .ciesin.columbia.edu/data/set/grump​-v1-settlement-points-rev01). Relative sum of lights is the ratio of an urban area’s sum of lights to the unweighted mean sum of lights for all urban areas in Latin America and the Caribbean. Both population and population density are calculated using LandScan (2012) gridded population ­­data. “ ­­ No. of cities in urban area” refers to the number of cities with a population of at least 100,000 whose settlement points intersect the urban a­­ rea. VIIIRS = Visible Infrared Imaging Radiometer Suite. T he M an y D imensi o ns o f U rbani z ati o n and the P r o du c ti v it y o f Cities in l a c    81 Annex 2D: Cross-Country Regression of Log(GDP per Capita) on Different Dimensions of Urbanization: Alternative Definition for a Multicity Agglomeration (1a) (1b) (2a) (2b) (3a) (3b) Urban share ­­ .054*** 0 ­­ .044*** 0 ­­ .042*** 0 ­­ .048*** 0 ­­ .041*** 0 ­­ .046*** 0 ­­(0.007) ­­ 0.004) ( ­­(0.006) ( ­­ 0.004) ­­(0.007) ­­ 0.004) ( Percentage of population in dense ­­− 0.011* ­­0.007 ­­0.007 ­­(0.006) ­­(0.006) ­­(0.006) Log(Weighted Density) ­­− 0.469*** ­­− 0.125 ­­− 0.151 ­­(0.145) ­­(0.173) ­­(0.171) Percentage of Population in MCAs ­­− 0.008* ­­− 0.007 ­­− 0.017*** ­­− 0.016*** ­­− 0.017*** ­­− 0.017*** ­­ 0.004) ( ­­(0.004) ­­(0.005) ­­(0.005) ­­(0.005) ­­(0.005) (North America) × (Percentage of ­­ .035*** 0 ­­ .028*** 0 ­­ .034*** 0 ­­ .027*** 0 Population in MCAs) ­­(0.005) ­­(0.005) ­­(0.006) ­­(0.005) ­­ (Western Europe) × (Percentage of ­­ .041*** 0 ­­ .033*** 0 ­­ .042*** 0 ­­ .034*** 0 Population in MCAs) ­­(0.006) ­­(0.006) ­­(0.006) ­­(0.006) (South ­­ America) × (Percentage of ­­ .008** 0 ­­ .011*** 0 ­­ .008** 0 ­­ .011*** 0 Population in MCAs) ­­ 0.004) ( ­­ 0.004) ( ­­(0.003) ­­(0.003) ­­ (Central America) × (Percentage of ­­0.005 ­­0.007 ­­0.006 ­­0.008 Population in MCAs) ­­(0.005) ­­(0.006) ­­(0.005) ­­(0.005) (Caribbean) × (Percentage of Population ­­0.005 ­­0.005 ­­0.006 ­­0.006 in MCAs) ­­(0.007) ­­(0.007) ­­(0.007) ­­(0.007) Urban primacy (%) ­­− 0.014 ­­− 0.012 ­­(0.012) ­­(0.012) [Urban primacy (%)]2 ­­0.000 ­­0.000 ­­(0.000) ­­(0.000) Constant ­­ .539*** 6 ­­10.304*** ­­ .434*** 6 ­­ .402*** 7 ­­ .754*** 6 ­­ .863*** 7 ­­(0.235) ­­(1.199) ­­(0.210) ­­(1.417) ­­(0.400) ­­ 1.396) ( No. of countries 169 169 169 169 169 169 2 Adjusted R ­­0.353 ­­0.399 ­­0.513 ­­0.509 ­­0.511 ­­0.509 Source: Calculations based on analysis of global data set of urban areas as constructed using the cluster algorithm of Dijkstra and Poelman (2014) and World Development Indicators data (­­http://data.worldbank​.org/data-catalog/world-development-indicators). Note: The dependent variable is the natural log of GDP per capita in 2012 international dollars (PPP exchange rates); robust standard e­­ rrors. “Urban share” denotes the percentage share of a country’s overall population living in urban areas; “Percentage of population in dense” denotes the share of a country’s overall population living in dense urban areas, where a dense urban area is one that has a mean population density that exceeds the global median for all urban areas; “weighted density” denotes the mean density of urban areas within a country weighted by the share of each urban area in a country’s overall urban population; “Percentage of populations in MCAs” denotes the share of a country’s overall population living in MCAs, where an MCA is defined as an urban area that contains two or more cities of any population ­­size. GDP = gross domestic product; MCA = multicity agglomeration; PPP = purchasing power parity. *p < 0.1. **p < 0.05. ***p < 0.01. Notes urban settlements that might usefully be cities). referred to as towns (rather than ­­ 1. Urban primacy is defined as the share of a 3. In settings with high factor mobility, it has been country’s urban population living in its most traditional to argue that capital and labor will populous urban ­­area. ­­ move until a spatial equilibrium is reached. In 2. The preference in this chapter is to refer to this equilibrium, utility levels across homoge- areas.” This is because “urban areas” “urban ­­ neous agents will be equalized (Rosen 1979; provides a more apt description for small Roback 1982). All else equal, this will tend to 82   RAISING THE BAR make for the spatial equality of wages and /­data/set/grump-v1-settlement-points-rev01). profits, not to mention the spatial equality of This data set provides geographic coordinates productivity levels at the margin (Glaeser et al. for 70,629 individual settlements, as well as 1992; Glaeser 2000). Even in spatial equilib- associated estimates of population for 1990, rium, however, differences in productivity 1995, and­­2000. For a complete description, (and in wages) at the margin will remain, to see CIESIN ­­ (2017). Throughout the chapter, the extent that there are differences in ameni- we follow the convention of naming an urban ties (for example, differences in climate) that area after the largest settlement point that it households value across ­­ areas. See the intersects ­­with. “Productivity Is Highly Dispersed across LAC 11. See annex ­­ 2D. Settlements in this case can Urban Areas” section in this chapter. include places that are officially classified as 4. A LAC country’s comparators are its “nearest rural even though they intersect with an urban neighbors” on population, land area, and area as defined by the cluster ­­ algorithm. average population ­­ density. Comparator 12. As table ­­ 2.1 shows, LAC urban areas have a countries are also countries that are similarly median area of ­­ 7.2 km2, compared with geographically located, that is, an island or those in ECA ­­ (22.0 km2) and North America non–island nation, and landlocked or 2 ­­(21.4 ­­km ). The corresponding mean areas ­­nonlandlocked. are ­­18.4 km2 (LAC), 51.2­­ km2 (ECA), and 5. This is based on the application of the cluster 91 km2 (North America). ­­ Our findings on the algorithm to LandScan 2012 globally gridded high density of LAC urban areas relative to population ­­ data. In their background paper, those in North America echo earlier research Roberts et al.­­ (2017) also apply the cluster findings by Ingram and Carroll (1981) who, algorithm to GHS-Pop and WorldPop gridded for a small sample of 24 large Latin American population ­­ data. They find that the resultant cities, found that average population densities maps of urban areas for LAC show a high in the 1950s–1970s resembled those of “old” level of agreement with the map produced North American cities such as New York, using LandScan 2012 ­­ data. Chicago, Philadelphia, Washington, DC, and 6. These settlements may be more aptly described Boston in the north and east, but were consid- as towns rather than ­­ cities. erably higher than those of newer North 7. These nine urban areas represent less than American cities such as Houston, San Diego, 0.23 percent of the original sample of 63,629 ­­ San Jose, and Phoenix in the south and ­­ west. urban ­­areas. 13. This assessment is based on a series of simple 8. The final section in chapter 6, “Institutional two-sample ­­t-tests. In performing these t-tests, Fragmentation, Metropolitan Coordination, we pool urban areas in the three comparator and Productivity,” identifies three main lines countries and test whether the mean across of thought regarding whether fragmentation urban areas for the LAC country is signifi- is good or bad for the economic performance cantly different from the corresponding mean of an urban ­­ area. Like Tiebout, the “polycen- ­­ for this pooled set of urban areas. Alternative trist” school argues that fragmentation is results based on performing t-tests against good, whereas the “centrist” school argues each individual comparator country are, over- bad. In between these two extremes, that it is ­­ all, consistent with those based on the pooling the “regionalist” view recognizes the benefits of urban areas in the comparator countries, of having multiple local governments while especially when comparing levels of urban highlighting the importance of metropolitan population ­­ density. We prefer to report results coordination, defined as the efforts of govern- based on pooling primarily for reasons of mental institutions to manage and solve prob- ­­ space. Similar comments apply to the results lems in common between ­­ jurisdictions. of the Kolmogorov-Smirnov test that are 9. Chapter 6 develops such a data set, but only reported in annex 2B, as well as to all subse- for a subsample of LAC metropolitan ­­ areas. quent analysis in the chapter relating to the 10. The global settlement point layer that we use benchmarking of individual LAC countries is the Center for International Earth Science against their ­­comparators. Information Network’s Global Rural–Urban 14. In Mexico City, only 16 of the 57 municipali- Mapping Project (GRUMP) Settlement Point ties belong to the city as defined by its official Layer v ­­1.1 ­­(http://sedac.ciesin.columbia.edu​ administrative ­­boundaries. T he M an y D imensi o ns o f U rbani z ati o n and the P r o du c ti v it y o f Cities in l a c    83 15. Out of the 12 MCAs in Central America, eight overwhelm agglomeration economies; high are in ­­Mexico. urban primacy may negatively affect national 16. Including the nine largest urban areas in the GDP per capita for similar reasons; and a global sample, the share of EAP’s overall large share of a country’s population living in urban population living in MCAs rises to ­­ 56.4 MCAs may also adversely affect national percent, and the share for South Asia rises to GDP per capita if the costs of fragmentation ­­72.3 ­­percent. outweigh the benefits, and there is a lack of 17. See the section in this chapter on “Implications metropolitan ­­coordination. for National Productivity” and chapter 6. 25. Potential sources of bias that would need to be 18. High urban primacy has also been linked to investigated before drawing causal inferences an absence of well-developed national trans- include both omitted variables and reverse port networks (Ades and Glaeser 1995; Davis ­­ causality. For example, in the relationship and Henderson 2003). between GDP per capita and urban primacy, 19. More specifically, Henderson (2000) uses an both could be partly driven by the level of augmented Solow-Swan growth framework, a development of national transport networks neoclassical growth model in which capital and a country’s openness to international ­­ trade. accumulation is subject to diminishing mar- 2.3, “urban share (%),” “Percentage 26. In table ­­ ginal returns and growth is ultimately driven of population in dense,” and “Percentage of by exogenous technological ­­ progress. For an population in MCAs” are all measured as overview of this model see, for example, Barro shares of a country’s population; all measures and Sala-i-Martin ­­(2003). are on a scale of ­­ 0–100. 20. The 11 LAC countries that Henderson (2000) 27. Consistent with this, CAF (2017) reports that, identifies as suffering from excessive primacy against Organisation for Economic in 1990 are Argentina, Chile, Costa Rica, Co-operation and Development countries, Dominican Republic, El Salvador, Guatemala, Latin America has a low prevalence of metro- Nicaragua, Panama, Paraguay, Peru, and politan governance bodies, informal or ­­ Uruguay. However, Henderson calculates ­­ formal. Half of Latin American metropolitan urban primacy using official national defini- areas have no coordination mechanisms what- tions of urban ­­ areas. soever, and only one in five cities has some 21. By the late 2000s, all but two LAC countries form of formal ­­framework. had directly elected local mayors, and the 28. In Mexico, for example, Kim and Zangerling average share of subnational spending in total (2016) document how comprehensive expenditures had reached 31.4 ­­ percent. This ­­ national efforts to designate and coordinate contrasts with the early 1980s, when only six metropolitan areas have only begun in the LAC countries had directly elected mayors past decade, particularly ­­box  3.3 ­­(pp. ­­50–51). and the equivalent share was 13.1 ­­ percent 29. In the analysis that follows, we drop from our (Chona, ­­n.d.). global sample urban areas that either (i) have 22. In table A2 of his paper, Henderson (2000) a zero or negative sum of lights or (ii) fall in reports that, for a country with a GDP per the top percentile of urban areas on the global capita of $17,200 (1987 constant interna- distribution of sum of lights, but that have a tional dollars), the estimated “optimal” pri- population of less than ­­ 200,000. On (ii), this macy rate declines from 23 percent at a leads to the exclusion of, for example, small national population of 8 million to 18 percent urban areas centered on oil refining that at 22 ­­ million. At a population of 100 million, ­­ appear very bright at night because of flaring. the optimal rate is 10 ­­ percent. A similar rate of Excluding areas on the basis of (i) and (ii) decline in the optimal rate is reported at lower leads to a final sample of 63,089 urban areas GDP per ­­capita. section. for the analysis in this ­­ 23. WDI follows the United Nations’ World 30. The finding that LAC urban areas are, on Urbanization Prospects database in adopting average, more productive than those in the rest national definitions of urban ­­ areas. of the world may seem to contradict the find- 24. These theories are that high levels of urban ing in table 1.2 of chapter 1 that LAC coun- density may have a negative effect on national tries have levels of GDP per capita that fall GDP per capita if they give rise to excessive close to the fitted line for the global relation- congestion forces (“demons of density”) that ship between a country’s development level 84   RAISING THE BAR and its urban share (as measured on the basis Akbar, ­­P. ­­A., and ­­G. ­­Duranton. ­­2017. “Measuring of the cluster algorithm). ­­ However, the units of the Cost of Congestion in a Highly Congested analysis that underpin these two findings are City: Bogotá.” Working Paper 04/2017, very different, individual urban areas in this Development Bank of Latin America, ­­ Caracas. chapter versus countries in chapter ­­1. B a rb ero, ­­J . ­­2 012 . Inf ra s t r u c t u re i n t he 31. Figure ­­2.11 groups Mexico with South America Development of Latin ­­ America . Caracas: rather than Central America, because Mexico’s Development Bank of Latin ­­ America. productivity distribution more closely resem- Barro, ­­R. ­­J., and ­­X. ­­Sala-i-Martin. ­­2003. Economic bles those of South American countries than of ­­Growth. Cambridge, MA: MIT ­­ Press. other Central American ­­ countries. Bastos, ­­ 2017. “Spatial Misallocation of Labor P. ­­ 32. For the Caribbean, productivity levels across in Brazil.” Background paper for this book, urban areas exhibit an interesting bi-modal World Bank, Washington, DC. distribution, driven by the three largest coun- CAF (Development Bank of Latin ­­ America). tries in the subregion, Haiti, Cuba, and ­­2 017. U r b a n G r o w t h a n d A c c e s s t o Dominican Republic, which are home to 387 O ppor tunities: A C halle nge for L atin of the Caribbean’s 473 urban ­­ areas. Hence, ­­A m e r i c a . 2 017 R e p or t o n E c o nom i c the lower mode corresponds to urban areas in Development ­­(RED). ­­Caracas. Haiti and the upper mode to urban areas CIESIN (Center for International Earth Science in Cuba and the Dominican ­­ Republic. Information Network). ­­ 2013. “Report for 33. Although based on a smaller sample of 91 Phase I: Mapping, Quantification and Analysis countries, a similarly strong negative correla- of Evolution of Patterns of Urban Physical tion is evident between the dispersion of pro- Extent and Morphology in South Asian Cities, ductivity across urban areas within a country 1 9 9 9 – 2 010 .” B a c k g r o u n d p a p e r f o r ­­ and the (natural) log level of the country’s Leveraging Urbanization in South Asia: density of paved ­­roads. M an ag ing S pati al Tran sfor m ation for 34. For detailed descriptions of these challenges in Prosperity and Livability , World Bank, the cases of Argentina, Central America, and Washington, ­­DC. Mexico see Muzzini et ­­ al. (2016), World Bank ———. ­­ 2017. Documentation for the Global (2016), and Kim and Zangerling (2016) Rural-Urban Mapping Project, Version 1 ­­respectively. (GRUMPv1): Settlement Points, Revision ­­ 01. New ­­York. Chona, G. n.d. Intergovernmental Fisc al References Relations and Decentralization in Latin Abrahams, ­­A., ­­N. Lozano-Gracia, and ­­C . ­­Oram. America and the Caribbean: Challenges and 2017. “Deblurring DMSP Nighttime Lights: ­­ Policy ­­Q uestions. New York: Inter-American A New Method Using Gaussian Filters and Development ­­B ank. ­­h ttp://decentralisatie​ Frequencies of Illumination.” Unpublished .org/?wpfb_dl=385. ­­manuscript. Davis, ­­J., and ­­J. ­­V. ­­Henderson. ­­2003. “Evidence Addison, ­­D., and ­­B . ­­Stewart. ­­2015. “Nighttime on the Political Economy of the Urbanization Lights Revisited: The Use of Nighttime Lights Process.” Journal of Urban Economics 53 (1): Data as a Proxy for Economic Variables.” ­­ ­­98–125. Policy Research Working Paper 7496, World Dijkst ra, ­­L ., and ­­H . ­­P oel man. ­­2 014. “A Bank, Washington, ­­DC. Harmonised Definition of Cities and Rural Ades, ­­A. ­­F., and ­­E . ­­L . ­­Glaeser. ­­1995. “Trade and Areas: The New Degree of Urbanization.” Circuses: Explaining Urban ­­ Giants.” Quarterly Regional Working Paper, Directorate-General Journal of Economics 110 (1): ­­ 195–227. for Regional and Urban Policy, European Ahrend, ­­R ., ­­E . Farchy, ­­I . Kaplanis, and ­­A . ­­C . Commission, ­­Brussels. Lembcke. ­­ ­­ 2014. “What Makes Cities More Duranton, ­­G ., and ­­D. ­­P uga. ­­2 004. “Micro- Productive? Evidence on the Role of Urban Fou nd at ion s of Urba n A g g lomerat ion Governance from Five OECD ­­ Countries.” ­­E conomies.” In Handbook of Regional and Reg ional Development Work i ng Paper Urban Economics , Volume 4: Cities and 2 014 / 05, O r g a n i s at ion for E c onom ic Geography, edited by ­­ V. Henderson and J. ­­ Co-operation and Development, ­­ Paris. ­­J.-F.  Thisse, ­­2063–2117. Amsterdam: ­­Elsevier. T he M an y D imensi o ns o f U rbani z ati o n and the P r o du c ti v it y o f Cities in l a c    85 Ellis, ­­P., and ­­M . ­­R oberts. ­­2 016. Leveraging Kim, ­­Y., and ­­B . ­­Z angerling. ­­2 016. Mexico Urbanization in South Asia: Managing Spatial Urbanization Review: Managing Spatial Transformation for Prosperity and ­­ Livability. Growth for Productive and Livable Cities in Washington, DC: World ­­ Bank. ­­Mexico. Washington, DC: World ­­ Bank. Ferreyra, ­­M. ­­M., ­­C . Avitabile, ­­J. Botero Álvarez, Krugman, ­­P., and ­­R. ­­L . ­­Elizondo. ­­1996. “Trade ­­F. Haimovich Paz, and ­­S . ­­Urzúa. ­­2017. At a Policy and the Third World Metropolis.” Journal Crossroads: Higher Education in Latin of Development Economics 49 (1): ­­137–50. America and the ­­ Caribbean. Washington, DC: Marshall, ­­A . ­­1890. Principles of ­­E conomics. World ­­Bank. London: Macmillan and ­­ Co. Glaeser, ­­ E. ­­ 2000. “New Economics of Urban and Muzzini, ­­E ., ­­B . Eraso Puig, ­­S . Anapolsky, Regional Growth.” In The Oxford Handbook ­­T.  Lonnberg, and ­­V. ­­Mora. ­­2016. Leveraging the of Economic Geography, edited by ­­G . ­­L . Potential of Argentine Cities: A Framework for Clark, ­­M . ­­S . Gertler, ­­M . ­­P. Feldman, and Policy ­­Action. Washington, DC: World ­­ Bank. ­­ K. Williams, ­­ 83 –98. New York: Oxford P i n k o v s k i y, ­­M . ­­L . ­­2 0 1 3 . “ E c o n o m i c University ­­Press. Discontinuities at Borders: Evidence from ———. ­­2 011. Triumph of the City: How Our Satellite Data on Lights at ­­ Night.” Working Greatest Invention Makes Us Richer, Smarter, paper, MIT Economics, Cambridge, ­­ MA. Greener, Healthier, and ­­ Happier. New York: Roback, ­­ ­­ J. 1982. “Wages, Rents, and the Quality Penguin ­­Press. of ­­L ife.” Journal of Political Economy 90: Glaeser, ­­E ., ­­H . ­­D. Kallal, ­­J . Scheinkman, and ­­1257–78. ­­A . ­­Shleifer. ­­1 992 . “Grow th in ­­C ities.” Roberts, ­­M ., ­­B . Blankespoor, ­­C . Deuskar, and Jour n al of Politic al E conomy 10 0 (6): ­­B . ­­S tewa r t . ­­2 017. “ Urba n i z at ion a nd ­­1126–52. ­­ Development. Is Latin America and the Henderson, ­­ J. ­­ V. ­­ 2000. “The Effects of Urban Caribbean Different from the Rest of the Concentration on E conom ic Grow ­­ th.” World?” Policy Research Working Paper 8019, Working Paper 7503, National Bureau of World Bank, Washington, ­­ DC. Economic Research, Cambridge, ­­ MA. Rosen, ­­ 1979. “Wages-based Indexes of Urban S. ­­ Henderson, ­­J. ­­V., ­­A. Storeygard, and ­­D. ­­N. ­­Weil. Quality of Life.” ­­ In Current Issues in Urban 2011. “A Bright Idea for Measuring Economic ­­ Economics, edited by ­­ P. Mieszkowski and ­­ M. ­­G rowth.” American Economic Review 101 Straszheim, ­­ 74–104. Baltimore: John Hopkins (3): ­­194–99. University Press ———. ­­ 2012. “Measuring Economic Growth Tiebout, ­­ C. ­­ 1956. “A Pure Theory of Local M. ­­ from Outer ­­S pace.” American Economic ­­Expenditures.” Journal of Political Economy Review 102 (2): ­­ 994–1028. 64 (5): ­­416–24. Hsieh, ­­C . ­­T., and ­­E . ­­Moretti. ­­2 017. “Housing Wo r l d ­­B a n k . ­­2 0 16 . C e n t r a l A m e r i c a Constraints and Spatial ­­ Misallocation.” Urbanization ­­ Review. Making Cities Work Working Paper 21154, National Bureau of for C entral ­­A meric a . Washington, DC: Economic Research, Cambridge, ­­ MA. World ­­B ank. I n g r a m , ­­G . ­­K . , a n d ­­A . ­­C a r r o l l . ­­1 9 81. Zhou, ­­N ., ­­K . Hubacek, and ­­M . ­­Roberts. ­­2015. “Sy mposiu m on Urban isation and “A nalysis of Spatial Pat terns of Urban Development: The Spatial Structure of Latin Growth across South Asia Using DMSP-OLS A m e r i c a n C it i e s .” Jo u r n a l o f U rb a n Nighttime Lights Data.” Applied Geography Economics 9 (2): ­­ 257–73. 63: ­­292–303. PART The Determinants of City Productivity in Latin America and II the Caribbean Cities in Latin America and the Caribbean density have somewhat weaker r ­ oles. This (LAC) need to raise the bar to reach the suggests an absence of wider positive global “frontier” of productivity perfor- agglomeration effects beyond those associ- mance and further contribute to the region’s ated with skill, which may be linked to the economic ­ development. To help understand lack of an adequate “enabling environment” what is required to achieve this, part II of for generating these ­effects. Chapters 4 and the book takes a more in-depth and rigorous 5 then provide deeper analysis of the roles look at the key determinants of city produc- of market access and skill, r ­ espectively. tivity in the r­ egion. Chapter 3 takes a rela- Chapter 6 goes beyond the concept of den- tively broad look at these determinants sity to analyze the role of a city’s spatial using a rich data set of household surveys form more generally in determining its pro- for 16 LAC countries, focusing on market ductivity, as well as the role of a city’s frag- access, skill, and ­ density. Although skill is a mentation into different administrative strong predictor of productivity differences jurisdictions and the mechanisms for metro- across cities, market access and, especially, politan ­coordination. The Empirical Determinants of City Productivity Mark ­Roberts 3 Introduction implies that cities are indeed “special places,” which, through their environment, can help One city may be more productive than workers and firms become more productive another for two basic ­ reasons. First, a city than they might otherwise b ­ e. Aligned with may be home to workers and firms whose this second explanation, urban economics characteristics make them more productive: it has identified three closely interrelated, and, may have an unusually talented workforce, to a significant degree, overlapping, theories whose members would be equally productive of “urban s ­ uccess.” These theories, all of no matter where they l ­ived.1 From a produc- which focus on different types of positive tivity viewpoint, such a city is the sum of its agglomeration effects, aim to explain differ- ­parts. Second, a city may have attributes ences in productivity across cities beyond associated with its environment that, because those associated with ­ sorting. of positive externalities and spillovers, Agglomeration ­economies. The first the- enhance the productivity of workers and ory is that cities can generate higher produc- firms beyond that expected on the basis of tivity than rural areas because of the positive their individual ­characteristics. Such a city externalities, known as agglomeration econo- becomes more than the sum of its ­ parts. mies, that their large population sizes or den- The first implies that, at least from a pro- create. Agglomeration economies can sities ­ ductivity perspective, there is nothing special arise through several ­ m echanisms. 2 The about cities and the concentrations of peo- “thick” labor markets that characterize cities ple and firms that are their defining can help generate better matches between characteristic. Differences in productivity ­ workers and firms, so that each person is across cities are entirely attributable to com- more likely to find his or her “perfect” ­ job. positional differences associated with the Cities can also provide the conditions for the “sorting” of workers and firms into different growth of a large and diversified array of spe- cities (for example, the tendency of more cialized suppliers of goods and services, skilled and able workers to move to certain which provide the intermediate inputs that cities or for more inherently productive firms help fuel the growth of the local e ­ conomy. to gravitate toward certain c­ ities). The second The work discussed in this chapter is based primarily on background papers by Quintero and Roberts (2017) and Reyes, (2017). The author thanks Jane Park for her excellent research assistance with the ­ Roberts, and Xu ­ chapter. 89 90   RAISING THE BAR The geographic proximity of people and emphasizing connectivity to the markets of firms in cities can give rise to the, often unin- other surrounding areas and ­ cities.5 tended, spillover of ideas as workers learn Although the three theories have been well from each other through observation and studied for developed countries and a handful ­interaction.3 of developing countries, little rigorous empiri- Human capital externalities ( ­ HCEs). The cal evidence exists on their relevance for most second theory is that cities can generate developing countries, including for countries higher productivity not so much because of in Latin America and the Caribbean (LAC) their size or density but because they tend to (Overman and Venables 2005; Henderson have higher overall human capital or skill, 2010; Duranton 2015).6 The main aim of this which helps generate positive ­ H CEs. In chapter, therefore, is to shed empirical light on many ways, this theory can be considered a the relative importance of these three theories special case of the first t­ heory. Whereas the or, to put it another way, the channels through first theory emphasizes several channels which positive agglomeration effects ­ arise. It through which agglomeration impacts posi- also distinguishes the extent to which varia- tively on a city’s productivity, HCE theory tions in productivity across cities, and between focuses on just one of these channels— urban and rural areas, are attributable to com- namely, the spillover of ideas between positional differences in the workforce—­ p eople. Furthermore, in doing so, it also ­ “sor t i ng”— ver su s t he u nderly i ng hypothesizes that the spillover of ideas is environment. The chapter draws on a data set ­ more likely to come from higher- than of harmonized household survey and sample lower-­ skilled workers, leading to the predic- census microdata for 16 LAC c ­ ountries. These tion that a worker’s individual productivity data have been matched with data that will be increasing with the average human describe differences in cities’ ­ environments. capital of the city in which she or he lives Complementing this analysis, which views (Rauch 1993; Moretti ­ 2004).4 productivity through the lens of workers, the Market ­a ccess. The third theory is that chapter also reports analysis based on firm- cities can generate higher productivity level World Bank Enterprise Survey (WBES) because they also tend to benefit from higher data for a global sample of c ­ ities. This global levels of access to large consumer markets perspective allows for a comparison on and to supplier markets of intermediate whether LAC differs from the rest of the ­ inputs. This superior access stems from both world on its strength of city-level determi- a city’s own “internal” market and its con- nants of firm productivity, controlling for the nectivity to other surrounding areas and characteristics of individual ­ f irms. It goes cities. Higher consumer and supplier market ­ beyond the three theories of urban success to access make it easier for firms to cover the highlight the characteristics of a city’s busi- fixed costs of setting up a new plant, which ness environment that are important for helps stimulate increases in profits and determining ­productivity. productivity (Krugman 1991a, 1991b; The chapter’s main findings are as follows: K rugman and Venables 1995; Fujita, Krugman, and Venables 1999). Again, this •  Nominal wages are, on average, higher theory is closely related to the first theory in urban than rural areas through- insofar as it focuses attention on a specific out LAC, reflecting higher average (sub)set of mechanisms through which posi- productivity. Higher productivity is ­ tive agglomeration effects may ­ a rise. It also typically seen in and around larger shares with agglomeration economies the- and more densely populated ­ cities. ory the hypothesis that a larger “internal” •  Much of this productivity variation market aids city productivity by stimulating stems from observable workforce com- the growth of a large and diversified array positional differences associated with of specialized suppliers of intermediate the sorting of workers between cities ­ goods. But it then goes beyond this by also and ­a reas. Notably, more productive T he E m p iri c a l D eter m inants o f Cit y Pr o d u c ti v it y    91 areas tend to be populated by better Cities Are More Productive educated ­workers. Than Rural Areas •  An important component of subnational productivity remains, whose variation Average Productivity in Cities Exceeds cannot be explained by workforce com- That in Rural Areas positional ­ d ifferences. This suggests Cities (and, more generally, urban areas) offer that, from a productivity viewpoint, potential productivity advantages over rural sorting is not the entire story and that areas, largely explained by the three theories cities are more than the sum of their of urban s ­ uccess. In line with these theories, parts. This is consistent with the three ­ countries worldwide have higher average theories of urban ­ success. nominal urban than rural w ­ ages. Urban •  Positive agglomeration effects are pres- firms can generally afford to pay higher ent in LAC ­ countries. These effects are wages than rural firms because their employ- driven, however, mainly by HCEs with ees are more ­ productive.7 a lesser role for market ­ access. Once Figure 3 ­ .1 shows the presence of large an area’s average level of human capi- urban–rural wage ratios in a sample of 15 tal (or skill level) and its market access LAC countries for which we have data for have been controlled for, population both types of area (seven each South density exerts no positive influence on American and Central American, and one productivity. This suggests that other ­ ­Caribbean).8 The figure’s data sources under- channels for positive agglomeration lie much of the rest of the analysis in the effects—for example, positive external- “ L a r g e S u b n a t i o n a l Va r i a t i o n s i n ities associated with labor market pool- Productivity” and “Explaining Underlying ing or spillovers of knowledge beyond Variations in Productivity” sections (box ­ 3.1). those emphasized by HCE theory— may not be ­ operative. Compositional Differences Associated •  One potential explanation is that the with Sorting Explain a Lot, but Not enabling environment for these other Everything types of positive agglomeration effects may not be present in LAC ­ cities. Cur- Given the large urban–rural productivity dis- rent levels of infrastructure and exist- parities, it might be thought that, despite the ing policies may not be adequate to already high urbanization rates in the region, support the high population densities there might be large unexploited productivity that characterize LAC cities, resulting gains to be had from rural–urban ­ migration. in excessively strong congestion forces However, it is also possible that the differ- that offset these other positive ­ effects. ences in average wages between urban and •  The finding of a lack of significant rural areas are attributable to compositional agglomeration effects beyond HCEs differences in the workforce between the two and market access is confirmed when types of area, rather than to anything special one analyzes global W BE S ­ d ata. about c ­ ities. These compositional differences These data also highlight obstacles to can arise from the “sorting” of workers into hiring skilled labor (which are worse different areas based on observable character- in the LAC region than elsewhere) as istics (such as educational attainment) and a major constraint on firm productiv- not so easily observable characteristics (such cities. Other elements of a city’s ity in ­ as ability and motivation) (Combes and business—and, therefore, also wider Gobillon ­2015). enabling environ ment, including Table ­ 3.1 shows important differences in modern infrastructure, basic protec- key observable characteristics of workers tion from crime, and access to formal between urban and rural areas in the banking finance—are also critical for 15-country ­ data set. Most important, in all ­productivity. 15 countries, workers who live in cities are, 92   RAISING THE BAR FIGURE ­3.1  Ratio of Nominal Mean Urban to Nominal Mean Rural Wage in 15 LAC Countries, 2000–14 3.0 2.6 2.5 2.3 2.2 2.2 2.1 2.0 2.0 1.9 2.0 1.7 1.6 1.6 1.4 1.4 1.4 Ratio 1.5 1.2 1.0 0.5 0 ia ru il a r ile y as a ico ala r ca a lic do do ua az bi m gu liv Pe ur ub Ri Ch em ex m na Br ua lva ug ra Bo nd sta ep lo M Pa ca Ec at Ur Sa Ho Co nR Co Ni Gu El ica in m Do South America Central America and the Caribbean Source: Calculations based on household survey microdata from SEDLAC (http://sedlac. econo.unlp.edu.ar/eng/) for all countries except ­Brazil. For Brazil, the calculation is based on IPUMS International (https://international.ipums.org/international/) population census sample ­microdata. Note: The ratio is the mean hourly nominal wage for urban relative to rural residents in each country calculated using pooled data for 2000–2014, where the mean hourly wage has been detrended using survey-year fixed effects. The figure is organized in descending order of urban–rural nominal wage ratio in each ­subregion. Argentina is excluded because its household survey (the Encuesta Permanente de Hogares) covers urban areas ­only. LAC = Latin America and the Caribbean; SEDLAC = Socio-Economic Database for Latin America and the Caribbean. ­BOX 3.1  SEDLAC: A Treasure Trove of Harmonized Data Our analysis in this chapter draws on successive The version of SEDLAC used for this book cov- rounds of household survey microdata for 16 LAC ers different survey years for different countries— countries that, apart from Brazil, come from the for example, 1974–2014 for Argentina, 1987–2013 Socio-Economic Database for Latin America and for Chile, and 2001–14 for ­ C olombia. To ensure the Caribbean ( ­ SEDLAC).a This database has been consistency across the LAC countries of analysis, constructed by the Center for Distributive, Labor we use SEDLAC data only from 2000 onward. To and Social Studies (CEDLAS) at the Universidad allay potential concerns over a lack of represen- National de La Plata and the World Bank’s Pov- tativeness of the survey data at the level at which erty Group for the LAC ­ region. The raw microdata we analyze it, we pool successive cross-sections from household surveys are not uniform across of ­d ata.b This has the effect of greatly increasing LAC countries, but the beauty of SEDLAC is that it ­ s ample sizes for subnational areas, and therefore also provides harmonized survey m ­ icrodata. Hence, also increasing the statistical precision of our the team behind SEDLAC makes strenuous efforts ­e stimates. to ensure that the data are comparable across coun- SEDLAC provides microdata on household tries and over time “by using similar definitions of members working in the formal and informal variables in each country/year, and by applying con- s ectors. To avoid potential selection bias, our ­ sistent methods of processing the data” (CEDLAS analysis focuses on a broad sample of workers and World Bank ­ 2014). covering formal and informal sectors irrespective (continued) T he E m p iri c a l D eter m inants o f Cit y Pr o d u c ti v it y    93 ­BOX 3.1  SEDLAC: A Treasure Trove of Harmonized Data (continued) of job ­ c haracteristics. However, we chose to For our empirical investigation of agglomeration restrict our samples only to wage workers, exclud- effects, in the “Explaining Underlying Variations ing self-employed workers whose reported income in Productivity” section, we further match the har- levels may not be comparable across countries monized survey data from SEDLAC with data from (Duranton ­ 2 016). Likewise, our samples exclude a LAC geospatial database that was constructed workers who report zero income (mostly family for this book in collaboration with the University ­ rade). Our final helpers in agriculture and retail t of Southampton’s GeoData Center (Branson et ­ a l. samples comprise all employed wage workers ­2016).d This database aligns with the identifiers for age ­14– 65 years. A worker’s wage is taken to be SEDLAC. It is this matching that subnational areas in ­ the nominal hourly wage earned in the primary also enables the mapping of subnational variations in ­o ccupation. c mean hourly wages shown in maps 3.1 and 3.2. a. The “Cities Are More Productive Than Rural Areas” section focuses on only 15 of the 16 countries because Argentina’s household survey covers only urban ­areas. For Brazil, we instead take microdata on workers from the population census sample for 2000 provided by IPUMS International ­(https://international.ipums. org/international/). We perform our own harmonization of the IPUMS International data for Brazil with the SEDLAC data for the other 15 countries in our s ­ ample. SEDLAC covers 24 ­countries. However, besides Brazil, eight countries were dropped either because changes in administrative units and their coding over time prevented SEDLAC from providing reliable geographic identifiers or because technical difficulties prevented the loading of the microdata from ­SEDLAC. The eight dropped countries are Paraguay, Suriname, and República Bolivariana de Venezuela in South America, and the Bahamas, Belize, Guyana, Haiti, and Jamaica in the ­Caribbean. b. To account for this, all regressions in the first four sections of this chapter include survey-year fixed e ­ ffects. c. Wages are measured at 2005 purchasing power parity exchange r ­ ates. In SEDLAC, rural wages are also increased by 15 percent to capture differences in rural– urban prices (CEDLAS and World Bank 2014, ­23). In our analysis, we undo this by multiplying the mean hourly wage of a rural worker by a factor of 0 ­ .8695. Although the results are not reported in this chapter, to test the robustness of results based on the broad sample, our background work also considered a narrower sample of workers that is restricted to “prime age” men, age 20–55 years, working in the private ­sector. We generally find very similar results for our broad and narrow samples (see Quintero and Roberts 2017). d. For more information, visit http://www.geodata.soton.ac.uk/geodata/. TABLE ­3.1  Differences in Characteristics between Urban and Rural Workers in 15 LAC Countries Years of Workers Age in years schooling with higher Region Country Area (mean) (mean) education (%) Male (%) Married (%) South Bolivia Urban ­36.5 ­10.4 ­17.6 ­57.6 ­65.1 America Rural ­40.3 ­6.0 ­4.7 ­72.9 ­75.5 Brazil Urban ­33.0 ­8.1 ­9.0 ­56.3 ­56.0 Rural ­32.4 ­4.5 ­1.2 ­70.9 ­61.4 Chile Urban ­39.5 ­11.6 ­20.5 ­60.0 ­61.4 Rural ­39.9 ­8.5 ­5.8 ­73.9 ­63.0 Colombia Urban ­37.2 ­9.6 ­16.1 ­56.6 ­57.3 Rural ­37.3 ­5.2 ­2.1 ­74.9 ­64.3 Ecuador Urban ­38.4 ­10.6 ­17.3 ­60.5 ­53.0 Rural ­38.7 ­6.5 ­3.4 ­70.7 ­56.2 Peru Urban ­37.4 ­10.6 ­24.4 ­56.2 ­56.5 Rural ­38.3 ­6.3 ­4.5 ­68.6 ­66.2 Uruguay Urban ­39.6 ­10.1 ­11.2 ­55.7 ­62.2 Rural ­41.2 ­7.5 ­3.9 ­67.6 ­68.8 Seven Urban ­37.2 ­10.1 ­18.5 ­57.0 ­57.6 countries Rural ­38.1 ­6.0 ­3.5 ­71.7 ­64.9 (continued) 94   RAISING THE BAR TABLE ­3.1  Differences in Characteristics between Urban and Rural Workers in 15 LAC Countries (continued) Years of Workers Age in years schooling with higher Region Country Area (mean) (mean) education (%) Male (%) Married (%) Central Costa Rica Urban ­37.0 ­9.9 ­17.5 ­60.0 ­54.6 America Rural ­36.0 ­7.2 ­6.4 ­71.4 ­61.4 and the Caribbean Dominican Urban ­36.5 ­9.8 ­16.6 ­61.0 ­56.7 Republic Rural ­37.3 ­6.6 ­4.7 ­73.4 ­59.6 El Salvador Urban ­36.9 ­8.9 ­11.2 ­53.3 ­57.1 Rural ­35.4 ­4.9 ­1.3 ­66.7 ­58.4 Guatemala Urban ­34.6 ­7.4 ­6.7 ­58.9 ­59.9 Rural ­34.2 ­3.6 ­0.5 ­71.9 ­67.4 Honduras Urban ­35.3 ­8.1 ­7.9 ­56.5 ­56.1 Rural ­35.7 ­4.3 ­0.7 ­72.6 ­62.4 Mexico Urban ­36.6 ­9.8 ­15.2 ­61.2 ­61.4 Rural ­37.1 ­6.2 ­3.1 ­69.6 ­68.6 Nicaragua Urban ­35.7 ­8.3 ­13.1 ­56.2 ­57.7 Rural ­35.2 ­4.3 ­1.7 ­74.1 ­64.4 Panama Urban ­37.7 ­11.6 ­16.1 ­58.6 ­60.5 Rural ­38.2 ­7.5 ­4.1 ­73.7 ­64.9 Eight Urban ­36.5 ­9.6 ­14.5 ­60.1 ­60.0 countries Rural ­36.4 ­5.7 ­2.8 ­70.8 ­65.0 LAC Fifteen Urban ­36.9 ­9.9 ­16.7 ­58.4 ­58.7 countries Rural ­37.2 ­5.8 ­3.2 ­71.2 ­65.0 Source: Calculations based on household survey microdata from SEDLAC (http://sedlac.econo.unlp.edu.ar/eng/), for all countries except ­Brazil. For Brazil, the calculation is based on IPUMS International (https://international.ipums.org/international/) population census sample ­microdata. Note: The table is sorted alphabetically by country name in each ­subregion. Descriptive statistics are based on wage/salary employees age 14–65 ­years. The values reported for LAC overall, and for the subregions (South America, and Central America and the Caribbean), are for the pooled sample of workers across all component c ­ ountries. All differences ­ est. IPUMS = Integrated Public Use Microdata Series; in means and in proportions between urban and rural areas are statistically significant at the 1 percent level in a two-tailed t LAC = Latin America and the Caribbean; SEDLAC = Socio-Economic Database for Latin America and the Caribbean. on average, significantly better educated rural areas once we control for these differ- than their rural ­ c ounterparts. This is the ences and, hence, ­ sorting. In other words, case regardless of whether we measure edu- does a city-dwelling worker earn signifi- cation by number of years of schooling or by cantly more than an “equivalent” worker completion of higher e ­ ducation. Apart from who lives in the countryside? If so, this Brazil, Costa Rica, El Salvador, Guatemala, would suggest the existence of an urban pro- and Nicaragua, urban workers are also, on ductivity premium that may, at least partly, average, slightly younger than rural w­ orkers. be explained by the three theories of urban Workers in urban areas are also more likely ­success. to be female—which could be taken as an To disentangle the degree to which differ- indication that urbanization promotes ences in nominal wages, and hence ­productivity, female labor force participation—and less between urban and rural areas are attribut- likely to be ­married. able to observable compositional differences Given the differences in worker charac- in the workforce versus other factors associ- teristics, the question arises of whether there ated with cities, we ran a series of augmented remains a significant difference between Mincerian wage regressions using microdata average nominal wages between urban and for workers (Mincer 1974), where these data T he E m p iri c a l D eter m inants o f Cit y Pr o d u c ti v it y    95 are again taken from the Socio-Economic explained by the differences in the observed Database for Latin A merica and the characteristics of workers between the type of Caribbean (SEDLAC) or ­ I PUMS. In these ­ area. We term this quantity the “worker pre- regressions, we include number of years of mium” to distinguish it from the estimated schooling, age and its square, and a worker’s coefficient on the urban dummy in our gender and marital status as key observable regressions, which reflects the existence of an ­characteristics.9 We also include a dummy “urban ­premium.” variable that takes the value one if a worker Figure ­ 3 .2 shows the results of this lives in an urban area and zero ­otherwise. If exercise. For all 15 LAC countries, differ- ­ the estimated coefficient on this urban ences in observable worker characteristics dummy is positive and significant, then this are important in explaining why nominal suggests a positive urban productivity pre- wages, and thus productivity, tend to be mium that allows an urban worker to earn areas. On average higher in cities than rural ­ more, in nominal terms, than an observation- across the countries, the average rural nom- ally equivalent rural ­ worker. From these inal wage would be 38 percent higher if regressions, we can also calculate the percent- the characteristics of rural workers were age difference in the average nominal wage changed to be the same as those of urban between urban and rural areas that can be workers—57 percent higher in the most FIGURE ­3.2  Urban and Worker Premiums in 15 LAC Countries 1.0 0.91 0.8 0.64 0.6 0.57 0.55 0.53 Premium 0.50 0.49 0.49 0.47 0.46 0.41 0.41 0.38 0.37 0.4 0.35 0.36 0.32 0.30 0.29 0.27 0.23 0.24 0.24 0.19 0.20 0.2 0.16 0.17 0.13 0.12 0.00 0 a u il a r ile y as a ico ala r ic ua ca do do a az i bi m r bl liv Pe ur Ri gu Ch g em ex m na Br ua a u ra Bo nd alv sta ep u lo M Pa ca Ec at Ur Ho Co nR S Co Ni Gu El ca i in m Do South America Central America and the Caribbean Urban premium Worker premium Source: Calculations based on household survey microdata from SEDLAC (http://sedlac.econo.unlp.edu.ar/eng/) for all countries except ­Brazil. For Brazil, the calculation is based on IPUMS International (https://international.ipums.org/international/) population census sample ­microdata. ˆ Note: Urban premium is calculated as [exp( α)−1], where αˆ is the estimated coefficient on a dummy variable, DU, which takes the value one (zero) when a worker lives in an urban (rural) area from a regression of the (natural) log of the nominal wage on DU and a set of observable worker characteristics (age, age squared, number of years of schooling, gender, ˆ and marital ­status). Worker premium is calculated as [exp(δ)−1], ˆ is the difference in the fitted natural log wage between urban and rural areas based on the difference in where δ mean values of each of the worker characteristics between these ­areas. When multiplied by 100, the values of both the urban and worker premiums in the chart give, all else equal, the percentage difference in the mean hourly wage between urban and rural a ­ reas. The figure is sorted in descending order of urban premium in each ­subregion. LAC = Latin America and the Caribbean; SEDLAC = Socio-Economic Database for Latin America and the Caribbean. 96   RAISING THE BAR extreme case, C ­ olombia. But we also productivity—tend to be seen in subna- observe that, in 14 of the 15 countries, there tional areas that correspond to major remains a statistically significant urban ­cities.12 As with our analysis of urban–rural productivity premium even after controlling wage ratios, the question arises as to the for the differences in worker ­characteristics. extent to which these differences are the The size of this premium ranges from product of sorting ­ (that is, differences in 12 percent in Costa Rica to 91 percent in workforce composition) versus differences Bolivia, with an average of 36 p ­ ercent. in the underlying productivity of ­ areas. Uruguay is the exception: the small differ- To answer this question, we again esti- ence in average wages between urban and mate a series of augmented Mincerian wage rural areas (figure ­3.1) can be attributed to regressions, one for each LAC country, that compositional differences in worker charac- control for key observable characteristics of teristics (sorting) rather than to urban ­workers.13 This time, instead of just includ- ­success. ing a simple binary urban dummy, we Although the significant urban produc- include a dummy variable for each subna- tivity premium in all but one LAC country tional area in a c ­ ountry.14 The estimated is consistent with the three theories, differ- coefficient on the dummy for a given subna- ences in the unobservable characteristics of tional area can be interpreted as an estimate workers could also explain that premium— of its (natural) log of underlying productiv- for example, more able or motivated people ity (or “location premium”), having con- (characteristics that are not easily captured trolled for the characteristics of its workforce by data) may be sorting into urban a ­ reas. and, hence, ­ sorting. Even then, it is likely that these unobserved As map ­ 3.1b and map ­ 3.2b show, when differences are correlated with observed we map these estimates of underlying pro- differences in number of years of school- ductivity, we see much less variation across ing; and, to the degree that this is true, we subnational areas than for average nominal can expect our estimates of the urban wages. This indicates that compositional ­ productivity premium to be relatively differences in the workforce associated with ­u nbiased.10 the sorting of workers across places is a major factor that drives productivity differ- ences between cities and, more generally, Large Subnational Variations subnational ­areas. in Productivity, Explained But sorting does not tell the full story of Partly by Sorting productivity differences between locations The evidence in the previous section shows because, even after controlling for composi- that LAC cities are generally more produc- tional differences in the workforce, some tive than rural areas and that this produc- variation in nominal wages remains across t iv it y d i f ferenc e rem a i n s even a f ter subnational areas in countries ( ­figure 3.3 controlling for observable differences in shows box plots of estimated location premi- workforce composition associated with ums for the sample of 16 c ­ ountries). The ­ sorting. However, it provides no informa- variation in underlying productivity is partic- tion on the wider geographic variations in ularly pronounced in Costa Rica, Ecuador, productivity in countries, which, as again Honduras, and Peru. There are also import- reflected by variations in average nominal ant differences in the size of the median wages, are large — maps 3.1 and 3. 2 , location premium across countries in, for panel a. The subnational areas depicted in example, Costa Rica and Ecuador, indicating the figures typically correspond to level 2 that more of a residual effect of location is administrative units or municipios, in the left over after controlling for sorting in 16 countries for which we have ­ d ata.11 these two countries than for, say, Honduras T he h ig hest nom i na l wages — a nd so and ­Uruguay. T he E m p iri c a l D eter m inants o f Cit y Pr o d u c ti v it y    97 MAP ­3.1  Subnational Variations in Nominal Wages in South America a. Without controlling for differences in worker characteristics b. Controlling for differences in worker characteristics N N Venezuela, RB Venezuela, RB Guyana Guyana Suriname Suriname Bogotá, Colombia Bogotá, Colombia Brasilia, Brazil Brasilia, Brazil Quito, Ecuador Quito, Ecuador Lima, Peru Lima, Peru Santa Cruz, Bolivia Santa Cruz, Bolivia Paraguay Paraguay Rio de Janeiro, Brazil Rio de Janeiro, Brazil São Paulo, Brazil São Paulo, Brazil Santiago, Chile Santiago, Chile Montevideo, Uruguay Montevideo, Uruguay Location premium Location premium (before sorting) (worker sorted) 6.0–8.0 6.0–8.0 4.0–6.0 Location premium 4.0–6.0 3.0–4.0 Greater Buenos Aires and (before sorting): 3.0–4.0 Greater Buenos Aires and 2.0–3.0 Greater La Plata, Argentina Argentina 2.0–3.0 Greater La Plata, Argentina 1.0–2.0 1.0–2.0 Location premium 6.0–8.0 0.5–1.0 0.5–1.0 (worker sorted): 4.0–6.0 Argentina 0.0–0.5 0.0–0.5 No data 3.0–4.0 No data 0.5–1.0 0 250 500 1,000 km 0 250 500 1,000 km Nonsample or 2.0–3.0 Nonsample or 0.0–0.5 Central America Central America Source: Quintero and Roberts 2017. Note: We use a separate point layer for Argentina to retain all locations for which SEDLAC allows us to estimate location p ­ remiums. Unlike other countries in our sample, these correspond to major cities/urban agglomerations, such as City of Buenos Aires and Greater La Plata, for which we lack a Geographic Information System shapefile of administrative ­boundaries. Location premium in the maps is calculated as exp(ˆ α) and expressed in 2005 purchasing power parity exchange rates, where, α ˆ is the estimated coefficient from a series of country-specific regressions on a location dummy, Li,l(i),t, which takes the value one when a worker i lives in a location l in the year t and zero ­otherwise. These regressions also include survey-year fixed ­effects. In panel a, the location premium is estimated without controlling for observable worker ­characteristics. In panel b, the location premium is estimated controlling for observable worker characteristics (age, age squared, number of years of schooling, gender, marital status) SEDLAC = Socio-Economic Database for Latin America and the Caribbean. MAP ­3.2  Subnational Variations in Nominal Wages in Central America and the Caribbean a. Without controlling for di erences in b. Controlling for di erences in worker characteristics worker characteristics N N Santo Domingo, Santo Domingo, Dominican Republic Dominican Republic Mexico City, Mexico Mexico City, Mexico Location premium Location premium Guatemala City, Guatemala (worker sorted) Guatemala City, Guatemala Tegucigalpa, Honduras (before sorting) Tegucigalpa, Honduras San Salvador, El Salvador San Salvador, El Salvador 6.0–9.7 Managua, Nicaragua Panama City, Panama 6.0–8.0 Panama City, Panama 4.0–6.0 4.0–6.0 Managua, Nicaragua 3.0–4.0 San Jose, Costa Rica 3.0–4.0 San Jose, Costa Rica 2.0–3.0 2.0–3.0 1.0–2.0 Venezuela, RB 1.0–2.0 Venezuela, RB 0.5–1.0 0.5–1.0 0.0–0.5 0.0–0.5 No data 0 250 500 1,000 km Colombia No data 0 250 500 1,000 km Colombia Nonsample or Nonsample or South America South America Source: Quintero and Roberts 2017. Note: Location premium in the maps is calculated as exp(α) ˆ and expressed in 2005 purchasing power parity exchange rates, where α ˆ is the estimated coefficient from a series of country-specific regressions on a location dummy, Li,l(i),t, which takes the value one when a worker i lives in a location l in the year t and zero ­otherwise. These regressions also include survey-year fixed ­effects. In panel a, the location premium is estimated without controlling for observable worker ­characteristics. In panel b, the location premium is estimated controlling for observable worker characteristics (age, age squared, number of years of schooling, gender, marital status). 98   RAISING THE BAR FIGURE ­3.3  Subnational Variations in Underlying Productivity in 16 LAC Countries 1.2 1.0 Estimated location premium 0.8 0.6 0.4 0.2 0 r ile a ru ia a il y ca r a lic a ico ala as do do ua az in bi gu m liv Pe ur ub Ri Ch em ex nt m na Br ua lva ug ra Bo nd sta ep lo ge M Pa ca Ec at Ur Sa Ho Co nR Co Ar Ni Gu El ica in m Do South America Central America and the Caribbean Source: Calculations based on household survey microdata from SEDLAC (http://sedlac.econo.unlp.edu.ar/eng/) for all countries except Brazil. For Brazil, the calculation is based on IPUMS International (https://international.ipums.org/international/) population census sample ­microdata. Note: This figure is organized in descending order of the median estimated location premium in each ­subregion. Estimated location premiums measure subnational variations in underlying productivity after controlling for observable worker characteristics within the broad sample (all wage/salary employees age 14–65 ­years) and survey-year fixed effects. The upper and lower caps, respectively, indicate the maximum and the minimum estimated location premiums for each ­country. The bottom of the box, the border of two colors, and the top of the box, respectively, depict the first quartile, the median, and the third quartile of the estimated location premiums in each ­country. IPUMS = Integrated Public Use Microdata Series; LAC = Latin America and the Caribbean; SEDLAC = Socio-Economic Database for Latin America and the Caribbean. Explaining Underlying reported in the urban economics literature Variations in Productivity: for developed countries, but lower than esti- The Three Theories mates reported for ­ China. Using comparable regression specifications, Chauvin et ­ a l. Consistent with the three theories of urban (2017) report an elasticity of nominal wages success, we see that estimated underlying for population density of ­ 4.6 percent for ­U.S. productivity levels (estimated location premi- metropolitan statistical areas, and 1 ­ 9.2 per- ums) across our sample of 16 LAC countries cent for a sample of Chinese provincial and are positively and significantly correlated prefectural ­cities. For our 16-country LAC with population density, average number of sample, the estimated elasticities of underly- years of schooling among the working-age ing productivity for skill and market access population (that is, skill),15 and a measure of are 62 percent and 4 ­ respectively. p ercent, ­ market access (figure ­ 3.4).16 For population Using a similar, but not identical specifica- density, which provides our measure of tion, Hering and Poncet (2010) report an agglomeration, the estimated elasticity of estimated elasticity of underlying productiv- underlying productivity is 9 ­ p ercent.17 This ity with respect to market access of 8 p ­ ercent is higher than corresponding estimates for Chinese cities. So, as with the estimated T he E m p iri c a l D eter m inants o f Cit y Pr o d u c ti v it y    99 FIGURE ­3.4  Correlation between Underlying Productivity and Population Density, Average Number of Years of Schooling, and Market Access a. population density b. Average number of years of schooling 2 2 location premium (worker sorted) location premium (worker sorted) 1 1 0 0 –1 –1 –2 –2 y = 0.0912X – 2.1592 y = 0.6234X – 3.3960 –3 R2 = 0.7619 –3 R2 = 0.8086 –5 0 5 –1.5 –1.0 –0.5 0 0.5 log(population Density) log(Average years of Schooling) c. market access 2 location premium (worker sorted) 1 0 –1 –2 y = 0.0435X – 1.6268 –3 R2 = 0.7695 –10 0 10 20 log(market Access) Source: Calculations based on Quintero and Roberts 2017. Note: Scatterplots show the correlation between the estimated location premiums (expressed in natural logs) from Quintero and Roberts (2017) and the natural logs of population density, average number of years of schooling, and market access controlling for country fixed e ­ ffects. Hence, subnational administrative areas are the units of observation and the correlations are estimated on the basis of the within-country variation in the ­data. Market access is measured as MAi = ∑i≠j(Pj/t2ij) where MAi is subnational area i ’s market access, Pj is the population of subnational area j, and tij is the estimated travel time (by road) between subnational areas i and j­. elasticity for population density, this is lower with our measure of underlying productivity in the LAC region than what has been found variables. and our main explanatory ­ for ­China.18 Yet it is also true that population density, Strong Human Capital Externalities, skill, and market access are all positively cor- Some Role for Market Access, but related with one another:19 more densely popu- Little Evidence of Wider Positive lated areas tend to exhibit higher skill and Agglomeration Effects greater market ­ access. To disentangle the rela- tive importance of the three variables, columns In column 1, we see that the estimated elastic- 1 through 3 of table 3.2 report regression ity of underlying productivity with respect to results, where we also control for physical geo- population density declines to ­ 4.9 percent graphic conditions that could be correlated once we control for geographic c ­ onditions. 100   RAISING THE BAR TABLE ­3.2  Results of Regressions on the Determinants of Underlying Productivity Variations across Subnational Areas Dependent variable: Location premium (ln) (1) (2) (3) (4) (5) Population density (ln) ­0.049*** ­0.013* ­0.005 ­0.023* ­0.002 Average number of years of ­0.576*** ­0.574*** schooling (ln) Percentage of working-age ­0.021*** ­0.020*** population with higher education Market access (ln) ­0.015*** ­0.027*** Mean air temperature (ln) ­0.030 ­0.044 ­0.051 ­0.036 ­0.045 Terrain ruggedness (ln) ­− 0.031** ­− 0.024*** ­− 0.017 ­− 0.026* ­− 0.024 Total precipitation (ln) ­− 0.028 ­− 0.008 ­− 0.010 ­− 0.003 ­− 0.001 Constant ­− 0.99*** ­−2.37*** ­−2.70*** ­−1.28*** ­−1.82*** No. of observations 5,750 5,750 5,050 5,750 5,050 2 R ­0.757 ­0.814 ­0.831 ­0.785 ­0.804 Adjusted R2 ­0.756 ­0.813 ­0.830 ­0.785 ­0.803 Sources: Quintero and Roberts 2017; population data: Gridded Population of the World, v4. Note: In all columns, country effects have been controlled for and standard errors have been clustered by ­country. In all columns, the dependent variable is the estimated location premium (measured in natural logs) from a series of country-level first-stage regressions after controlling for observable worker characteristics in the broad sample (all wage/salary employees age 14–65 years) and survey-year fixed ­effects. Worker characteristics include age, age squared, marital status, gender, and number of years of ­schooling. ***p < 0.01. **p < 0.05. *p < 0.1. Although this estimate remains statistically extent—market ­ access. By contrast, other significant, it is less than the estimated elastic- types of positive agglomeration effects associ- ity of 9 percent that we reported above when ated with the (more general) theory of not controlling for g ­ eography. Once we also agglomeration economies seem to be ­ absent. introduce skill, as measured by average num- This includes positive effects stemming from, ber of years of schooling, in column 2, how- for example, labor market pooling or more ever, both the estimated size and statistical general spillovers of knowledge beyond those significance of the elasticity of underlying pro- emphasized by HCE ­ theory. ductivity with respect to population density dramatically. Including market access in fall ­ Absence of Wider Positive column 3, then leads population density to Agglomeration Effects May Be lose its significance ­ completely. Although skill Linked to an Inadequate Enabling and market access are significant, the former Environment has a larger effect on underlying ­ productivity. Although an increase in the average number of The absence of wider positive agglomeration years of schooling from the 25th to the 75th effects could be linked to the high popula- percentile in our sample implies an estimated tion densities of many LAC cities (see chap- productivity increase of ­ 23.4 percent, moving 2). One hypothesis is that these high ter ­ from the 25th to the 75th percentile for mar- densities are leading to excessive congestion ket access implies a productivity increase of forces that are negating many of the positive only ­4.1 ­percent. externalities normally associated with urban It thus seems that, although LAC countries ­ density. In this context, it is not necessarily do experience positive agglomeration effects, the high densities per se that matter, because these are not as strong, overall, as have been LAC cities may lack the enabling environ- reported for ­ C hina. Furthermore, these ment to foster these wider positive agglomer- agglomeration effects are mainly those associ- ation ­effects. Therefore, current policies and ated with the theories of HCEs and—to a levels of infrastructure in LAC cities may not much lesser, but still statistically significant be sufficient to prevent these densities T he E m p iri c a l D eter m inants o f Cit y Pr o d u c ti v it y    101 creating excessive congestion forces. lea r n i ng on ly f rom h ig h ly educ ated Chapter 2 discussed issues of metropolitan colleagues. This alternative measure is also a ­ coordination in the context of multicity highly statistically significant predictor of agglomerations, a topic that chapter 6 looks variations in underlying productivity across at in more ­ d epth. It also mentioned the subnational ­ areas. Comparing the results in continued widespread existence of slums, columns 4–5 with those in columns 2–3, how- which pose challenges for infrastructure ever, we can also see that our regressions fit provision—for two ­ reasons. One, as high- better when using average number of years of lighted by Fay et ­ a l. (2017), is the slums’ schooling rather than the share of the work- location, which is often in flood-prone or ing-age population with higher ­ education.22 environmentally protected ­ areas. The other is that their dense and disorderly develop- Effects Are Heterogeneous ment hinders work on access roads or water, Across Countries sewerage, and ­ d rainage. Fay et a ­ l. (2017) also highlight a lack of access to sanitation Beyond the above average results for all as a serious infrastructure issue, even for the 16 LAC countries, we now investigate the middle classes, in Latin American ­ cities. considerable heterogeneity in estimated effects However, further work is required to fully of population density, skill, and market access substantiate the above ­ hypothesis. This is on underlying productivity across individual because it may be that, relative to our mea- countries (figure ­ 3.5).23 Contrary to the over- sures of human capital and market access, all average results, we find that the estimated population density is a relatively poor measure elasticity of underlying productivity with of agglomeration, in which case its estimated respect to population density is positive and coefficient may be biased ­ d ownward. 20 statistically significant for Brazil, Dominican Indeed, in previous chapters, we emphasized Republic, Ecuador, and Peru: the estimated that the boundaries of subnational administra- 95 percent confidence intervals, shown in tive units often conform only poorly with the panel a of the f ­igure, do not encompass the “true” boundaries of ­ cities. To the extent that value zero (indicated by the dashed ­ line). By agglomeration is more poorly measured, this contrast, we estimate a significant negative could provide another explanation of why we elasticity for Chile and ­ Nicaragua. find no evidence of positive agglomeration On market access, the overall positive effects beyond those associated with HCEs average (and statistically significant) influ- and market a ­ ccess. It is also somewhat of a ence is driven mainly by four countries conundrum as to why excessive congestion (Brazil, Costa Rica, Ecuador, and ­ Nicaragua). forces associated with high densities combined For the remaining countries in the figure, the with an inadequate enabling environment effect of market access is not significantly might be thwarting certain types of positive different from ­ zero.24 agglomeration effects, but not ­ HCEs.21 With skill, the effect of average number of In addition to reporting average number of years of schooling is statistically significant years of schooling, table ­ 3.2 also reports for all ­countries. Even here, the estimated results using an alternative measure of skill: strength of HCEs varies dramatically across the share of the working-age population who c ountries. We estimate extremely strong ­ have completed higher education—columns HCEs in Bolivia but comparatively weak 4 and 5. This is closer to the measure of skill externalities in the Dominican Republic, preferred in the academic literature on HCEs, El Salvador, and Nicaragua.25 much of which argues that raising the top of In addition to heterogeneity across coun- the human capital distribution will generate tries, it is also fruitful to analyze, following learning spillovers but that raising the bot- the example of Du ranton (2016) for tom will not (Glaeser ­ 1999). This amounts to Colombia, heterogeneity across different sub- arguing that workers experience significant groups of workers (box ­ 3.2). 102   RAISING THE BAR FIGURE ­3.5  Cross-Country Heterogeneity in Estimated Elasticities of Underlying Productivity with Respect to Population Density, Average Number of Years of Schooling, and Market Access a. Population density 0.15 0.10 0.05 Elasticity 0 –0.05 –0.10 ru il r a ile as lic ico r ca ala a do do az bi gu Pe ur ub Ri Ch em ex m Br ua lva ra nd sta ep lo M ca Ec at Sa Ho Co nR Co Ni Gu El ica in m Do South America Central America and the Caribbean b. Average years of schooling 1.8 1.6 1.4 1.2 Elasticity 1.0 0.8 0.6 0.4 0.2 0 a ile ru il or a as ico ala ca r a lic do az ivi bi gu ad Pe ur ub Ri Ch m ex m Br a l ra Bo nd u alv sta ep e lo M ca Ec at Ho Co nR S Co Ni Gu El ica in m Do South America Central America and the Caribbean (continued) T he E m p iri c a l D eter m inants o f Cit y Pr o d u c ti v it y    103 FIGURE ­3.5  Cross-Country Heterogeneity in Estimated Elasticities of Underlying Productivity with Respect to Population Density, Average Number of Years of Schooling, and Market Access (continued) c. Market access 0.08 0.06 0.04 0.02 Elasticity 0 –0.02 –0.04 –0.06 –0.08 r il ru a ile a ca lic ico ala r as do do az bi gu Pe ur ub Ri Ch em ex m Br ua lva ra nd sta ep lo M ca Ec at Sa Ho Co nR Co Ni Gu El ica in m Do South America Central America and the Caribbean Source: Quintero and Roberts 2017. Note: Figures show the estimated elasticities for each country derived from regressing—in individual country-level regressions—estimated subnational ­ .2. The squares represent the point estimates, whereas the underlying productivity (measured in natural logs) on the variables shown in column 3 of table 3 upper and lower caps indicate the upper and lower bounds of the 95 percent confidence i­ntervals. ­BOX 3.2  Which Groups of Workers Benefit More? To explore heterogeneous effects of population den- age of workers in all 16 ­countries. Given that our sity, skill as measured by average number of years of sample already excludes all self-employed workers schooling, and market access on estimated underly- and workers who report zero income (see box 3 ­ .1), ing productivity variations across different types of we defined informal workers as those who work workers, we performed a series of regressions for the for firms with five or fewer ­employees. a All other following four dimensions, drawn from our broad workers are assumed to be ­ f ormal. Our regres- sample, which consists of all employed wage workers sions for private vs. public sector and formal vs. age 14–65 years: informal workers exclude Brazil, because, for that country, no public sector workers were left in our 1. Young vs. old original broad sample after data cleaning, and 2. Male vs. female IPUMS International does not provide data that 3. Private vs. public sector allow us to distinguish between formal and infor- 4. Formal vs. informal mal workers in a manner akin to that for other We used 35 years as the dividing line between countries, for which the data instead come from young and old because this is roughly the mean ­SEDLAC. (continued) 104   RAISING THE BAR ­BOX 3.2  Which Groups of Workers Benefit More? (continued) Table ­B3.2.1 summarizes the ­ results. The most ­ stimated for young and for male workers, respec- e striking differences in estimated effects come in tively, which are statistically significant at the 10 per- private vs. public sector workers and formal vs. cent level or ­better. Although they are smaller than informal ­ w orkers. In both cases, the estimated for private vs. public sector and formal vs. informal elasticities of underlying productivity with respect workers, differences can also be observed in the to skill and market access are higher for private estimated elasticity of underlying productivity with and formal workers than for public and informal respect to skill for young vs. old and male vs. female ­ workers. For market access, the estimated elastici- ­ workers. Old workers appear to benefit more from ties are close to zero for public and informal work- stronger HCEs than young workers, and male work- insignificant. ers, and statistically ­ ers than female ­ workers. Market access also has an insignificant effect for Population density exerts a negligible and statis- old and for female ­ workers. However, the ­ estimated tically insignificant effect on underlying productiv- elasticities in both cases are much closer to those subgroups. ity for all ­ TABLE B3.2.1  Heterogeneous Effects of Determinants on Underlying Productivity across Worker Subgroups (1a) (1b) (2a) (2b) (3a) (3b) (4a) (4b) Young Old Male Female Private Public Formal Informal (ln) Population density ­ -0.001 0.004 0.006 −0.001 0.002 0.012 0.004 −0.002 Average number of years 0.466*** 0.580*** 0.541*** 0.495*** 0.548*** 0.172* 0.559*** 0.401*** (ln) of schooling ­ Market access ­ (ln) 0.014*** 0.010 0.011* 0.008 0.018*** −0.001 0.019*** 0.004 *** *** *** *** *** *** *** Constant −2.554 −2.392 −2.356 −2.281 −2.586 −0.804 −2.349 −1.565*** No. of observations 3,756 3,757 3,758 3,754 3,758 3,440 3,758 3,732 2 Adjusted R 0.790 0.744 0.689 0.675 0.717 0.680 0.668 0.687 Source: Quintero and Roberts 2017; population data: Gridded Population of the World, v4. Note: In all columns, country effects have been controlled for and standard errors have been clustered by ­country. In all columns, the dependent variable is the estimated location premium (measured in natural logs) from a series of country-level first-stage regressions after controlling for observable worker characteristics in the broad sample (all wage/salary employees age 14–65 years) and survey-year fixed ­effects. Worker characteristics include age, age squared, marital status, gender, and number of years of ­schooling. a. The Socio-Economic Database for Latin America and the Caribbean provides two indicators for whether a worker is considered informal (CEDLAS and World Bank 2014). The first, based on a “productive” definition of informality, identifies a worker as informal if “(s)he belongs to any of the following categories: (i) unskilled self-employed, (ii) salaried worker in a small private firm, (iii) zero-income ­workers.” The second, based on a “legalistic” or “social protection” notion of informality, identifies a salaried worker as informal if “s(he) does not have the right to a pension linked to employment when r ­ etired.” We rely on the “productive” indicator because this suffers from fewer missing observations; however, because our sample already excludes both self-employed and zero-income workers, this amounts to equating informal employment with employment by small ­firms. ***p < 0.01. **p < 0.05. *p < 0.1. What about Firms? insights into the drivers of urban s ­ uccess. Evidence from World Bank And analyzing firm data allows for a more Enterprise Surveys direct measurement of productivity because it allows us to build measures of labor produc- So far in the chapter, we have followed the ­ TFP). tivity and of total factor productivity ( dominant approach of the academic urban Unfortunately, there are no equivalents of economics literature of using microdata on SECLAC or IPUMS for firm microdata. The individual ­workers. It is also possible, how- background paper by Reyes, Roberts, and Xu ever, to analyze productivity differences, and (2017) for this book does, however, take their determinants, from the perspective of advantage of what we suggest is the next best firms, which can yield complementary t h i ng — h a r mon i z e d W B E S d at a for T he E m p iri c a l D eter m inants o f Cit y Pr o d u c ti v it y    105 FIGURE ­3.6  Different Dimensions of a City’s Porta et a­ l. 1998; Beck, Demirgüç-Kunt, and Business Environment Maksimovic 2005; Bloom et ­ a l. ­2 010). Refined BE includes the entry and exit barri- ers that exist for firms in a city, as well as a • Basic protection Basic business • Infrastructure city’s labor regulations and tax e ­ nvironment. environment • Human capital The agglomeration environment encapsu- • Access to finance lates (i) whether a firm is based in a large city, and therefore has the potential to benefit from agglomeration economies; 27 (ii) “capacity Refined business • Barriers to entry/exit agglomeration,” defined as the concentration • Labor regulations of firms in a city that possess high capacity environment • Tax environment either in technology, management, or ability to adapt to a changing competitive environment; and (iii) informal competition in a city, mea- • In a big city sured by the share of firms in the city that Agglomeration environment • Capacity agglomeration self-report as competing with informal f ­ irms.28 • Informal competition Capacity agglomeration is proxied by the share of firms in a city that employ more than 50 ­workers. The use of this proxy is consistent Source: Reyes, Roberts, and Xu 2017. with evidence that shows that large firms are more productive and export more than 2 ­ 006–2015. These data cover almost 49,000 smaller firms (Bernard et ­ al. 2007; Melitz and firms in up to 529 cities, drawn from 110 coun- Ottaviano 2008); are more innovative (Cohen globally. They include 66 cities in 23 LAC tries ­ and Levin 1989); and conduct research and countries. The fact that LAC firms and cities ­ development (R&D) more efficiently (Cohen form part of a much broader global sample and Klepper ­ 1996). Perhaps because the R&D means that the data can be used to compare centers of large firms provide key spillovers the strength of city-level determinants of firm for small firms (Acs, Audretsch, and Feldman productivity, as measured by labor productiv- 1994), large firms are associated with higher ity and TFP, in the LAC region with those in industrial agglomeration (Barrios, Bertinelli, world.26 The analysis conducted the rest of the ­ and Strobl 2006; Holmes and Stevens 2 ­ 002). by Reyes, Roberts, and Xu (2017) focuses on For the United States, the exogenous reloca- the relationship between firm productivity and tion of large firms has been found to positively different elements of a city’s business environ- affect incumbent firms’ TFP (Greenstone, ment (BE), which itself can be considered part Hornbeck, and Moretti 2010), and firms are of a city’s wider enabling environment for pro- more likely to become large when colocated ductivity ­enhancement. The BE in this context with other large firms (Li, Lu, and Wu ­ 2012). is broadly defined as comprising three main Using the same proxy for capacity agglomera- elements: basic BE, refined BE, and the agglom- tion, Li, Long, and Xu (2017) find that this eration environment (figure 3.6). measure helps explain China’s productivity Basic BE refers to those aspects of a city’s advantage over India in a quantitatively BE that many analysts view as fundamental important way; Clarke, Qiang, and Xu (2015) for ­development. It includes the basic func- find that it has predictive power for firm-level tions of government protection, including job growth using WBES ­ data. containing corruption and providing basic Like the analysis of worker data in previ- protection from ­ crime. It also encompasses a ous sections, however, it is also possible that good supply of human capital and infra- productivity differences across cities may be structure, as well as access to finance—a key driven by compositional differences associ- element for many researchers (Levine 1997; ated with the sorting of firms across ­ cities. Demirgüç-Kunt and Maksimovic 1998; La Different cities can be expected to be home to 106   RAISING THE BAR firms with different mixes of ­ characteristics. and TFP, where LAC is a dummy variable To control for this, Reyes, Roberts, and Xu equal to one for LAC cities and zero for cities (2017) regress firm productivity not only on in the rest of the w ­ orld. The regressions variables that capture the three different ele- include only interactions for LAC where these ments of a city’s BE (basic and refined BE, were found to be statistically significant in and the agglomeration environment) but also one or more of the regressions estimated by on key characteristics of firms that can be Reyes, Roberts, and Xu ( ­ 2017).30 As seen, the observed in the WBES ­ data.29 Given that the number of these interactions is relatively regression analysis also includes country small. It follows that the BE in LAC cities ­ fixed effects, in effect it examines the deter- influences firm productivity in much the minants of firm productivity in countries at same way as it does in the rest of the w ­ orld. the city-industry level, assessing how much Therefore, and perhaps somewhat surpris- the strength of these determinants varies ingly, the refined BE that cities offer is not a between LAC and the rest of the ­ world. significant determinant of firm productivity, Table ­ 3 .3 shows the main regression whether labor productivity or TFP, either in results for (natural log) labor productivity the LAC region or the rest of the ­ world.31 TABLE ­3.3  The Effects of a City’s Business Environment on Firm Productivity (1) Log(Labor Productivity) (2) Log(TFP) BE element BE variable Coefficient SE Coefficient SE Basic BE Corruption obstacle 0.079 ­0.106 ­0.223 ­0.127 Security cost ­−2.903** ­0.830 ­−3.540** ­0.758 Outage ­0.049 ­0.114 ­0.014 ­0.102 Web intensity ­0.464** ­0.119 ­0.333** ­0.108 Web intensity × LAC ­− 0.001 ­0.210 ­− 0.016 ­0.194 Skilled labor obstacle ­− 0.237* ­0.095 ­− 0.138 ­0.114 Overdraft facility ­0.327** ­0.112 ­0.165 ­0.112 Trade credit ­0.285 ­0.149 ­− 0.079 ­0.129 Refined BE Land access obstacle ­− 0.076 ­0.108 ­− 0.081 ­0.105 Tax rate obstacle ­0.197 ­0.104 ­− 0.238* ­0.111 Labor regulation obstacle ­0.091 ­0.121 ­0.181 ­0.119 Agglomeration BigCity ­0.102* ­0.040 ­0.076 ­0.048 environment BigCity × LAC ­− 0.205* ­0.086 ­− 0.019 ­0.072 Capacity agglomeration ­0.258 ­0.230 ­0.505* ­0.199 Inf competition ­− 0.043 ­0.114 ­0.002 ­0.102 Inf competition × LAC ­0.186 ­0.204 ­0.076 ­0.192 No. of observations 48,614 19,603 Adjusted R2 ­0.370 ­0.278 Source: Calculations based on Reyes, Roberts, and Xu 2017. Note: Labor productivity is measured as sales divided by the number of permanent ­employees. TFP (total factor productivity) is estimated as the residual from industry-specific production functions with log value added as the dependent variable and log capital and log labor as the independent ­variables. Capital is the replacement cost of land and ­machinery. Labor is the number of permanent employees plus ­0.5 times the number of temporary ­employees. Regressions include a full set of country and industry fixed effects, as well the following controls: Foreign (share of foreign ownership of the firm), OwnLargest (ownership share of the largest owner), L0 20–100 (= 1 if firm’s number of employees three years ago was 20–100; = 0 otherwise), L0 100+ (= 1 if firm’s number of employees three years ago exceeded 100; = 0 otherwise), Age 6–10 (= 1 if firm’s age is between 6 and 10 years; = 0 otherwise), Age 10+ (= 1 if firm’s age is 10 years or greater; = 0 otherwise), and exporter (= 1 if firm exports; = 0 o ­ therwise). BE = basic environment; SE = standard error. **p < 0.01. *p < 0.05. Heteroscedasticity-corrected SEs were clustered at the city level. T he E m p iri c a l D eter m inants o f Cit y Pr o d u c ti v it y    107 By contrast, several different components of LAC cities have long been notorious for their basic BE have significant effects on productivity high crime ­ rates. According to data from the in LAC and non-LAC cities ­ alike. A lack of basic Brazilian think tank Igarapé Institute, LAC protection—as proxied by a high city-­ industry was, excluding war zones, home to 43 of the average spending by firms (as a proportion of 50 most murderous cities in the world in sales) on security (by a high value of the variable 2016, with San Salvador, the capital city of El “security cost”)—has a large and significant Salvador, holding the dubious distinction of negative effect on labor productivity and on being the global “murder capital” with 137 TFP. The presence of modern infrastructure in a ­ homicides per 100,000 ­ inhabitants.33 As fig- city (in fact, of the internet as captured by the ure 3­ .7 shows, average spending by firms on variable “web intensity”) has a significant posi- security across LAC cities (as a share of sales) tive effect on labor productivity and T­ FP. Access also exceeds that in all other regions of the to formal finance, measured by the share of world (including average spending by firms in firms with an overdraft facility in a city (“over- East Asia and Pacific cities), except for Europe draft facility”), likewise has a significant posi- and Central Asia and Sub-Saharan Africa. tive effect on firm productivity, although this Our findings using the worker data (see effect is confined to labor ­ productivity.32 A “Explaining underlying variations in produc- shortage of skilled labor in a city (“skilled labor tivity”), highlight the central role of skill in obstacle”) has a significant negative effect on determining urban success in the ­ region. That firm labor productivity, but not on ­ TFP. skilled labor obstacles have significant nega- The findings that a lack of basic protection tive effects on firm productivity corroborates from crime in cities and the existence of this ­finding. skilled labor obstacles have significant nega- Although the effects of the basic BE and tive effects on productivity are particularly refined BE on firm productivity are similar ­ important even if their effects are the same for in the LAC region to those in the rest of the LAC cities as for those in the rest of the world. world , t he ­ s it u at ion d i f fer s for t he FIGURE ­3.7  Security Costs Incurred by Firms in Cities, Latin America and the Caribbean and Other Regions 0.030 0.025 0.020 Mean security cost 0.015 0.010 0.005 0 East Asia and South Asia Middle East and Latin America Europe and Sub-Saharan the Pacific North Africa and the Caribbean Central Asia Africa Source: Analysis of World Bank Enterprise Survey data from Reyes, Roberts, and Xu ­2017. Note: The y axis shows the mean spending by firms on security across cities (as a share of sales) in a ­region. For each city, security cost is the city-industry average of the share of a firm’s sales paid for ­security. Europe and Central Asia covers 27 countries, only one of which (Sweden) is a Western European country. LAC = Latin America and the Caribbean. 108   RAISING THE BAR agglomeration e ­ nvironment. Outside the the results for TFP also suggest that the ben- LAC region, there is a “big city” labor pro- efits of agglomeration might not be as strong ductivity premium of 10 percent, which is ­ orld. in the LAC region as in the rest of the w level. statistically significant at the 5 percent ­ Besides being in a big city, capacity agglom- For LAC, however, the premium is ­ negative. eration is also found to be a statistically sig- Conditional on firm characteristics and nificant determinant of TFP, although, in other elements of the BE, labor productivity this case its effect in the LAC region is not is some 10 percent lower for a LAC firm in a found to differ significantly from that in the big city than for one in a smaller ­ city. The world. The importance of a city’s rest of the ­ existence of this negative effect is consistent BE and agglomeration environment for the and, therefore, serves to reinforce the find- productivity of its firms is further investi- ing of a lack of wider agglomeration effects, gated in box ­ 3.3 for Colombian manufac- beyond HCEs, that was found using worker turing firms, where particularly rich data data. Although not statistically significant, exist for firms and ­cities. BOX 3.3  The Determinants of Manufacturing Firm Productivity across Colombian Municipalities With nearly 50 million inhabitants and an estimated data, the authors estimate the total factor productiv- urban population of 35 million, Colombia is one of ity (TFP) of each f­ irm.d The second data set is a panel the Latin America and the Caribbean (LAC) region’s of municipal characteristics, which covers more than most populous and highly urbanized c ­ ountries. a 1,300 variables for ­ 1993–2014. e By joining these Although the areas east of the Andes account for two data sets, Balat and Casas study the effect of a nearly 60 percent of Colombia’s total land area, most municipality’s characteristics on the productivity of economic activity is concentrated in smaller areas to firms. Key variables studied include those for the its ­ the west: the Andean region, with the capital city of municipality relating to the nature of manufacturing Bogotá, and the Caribbean coast ­ region. agglomeration, fiscal performance, quality of educa- In their background paper, Balat and Casas tion, and rates of conflict and ­ violence. (2017), show that more than 70 percent of Colom- Descriptive analysis shows that Bogotá not only bian manufacturing firms are concentrated in seven forms the largest manufacturing cluster with the high- major cities—Bogotá, Medellín, Cali, Bucara- est absolute concentration of firms in every industry manga, Manizales, Cartagena, and Barranquilla— but also exhibits the highest average manufacturing in these two r­ egions. Balat and Casas seek to explore firm ­productivity. But Bogotá does not top productiv- whether this clustering benefits the productivity of ity charts in every i­ndustry. Cali, for example, Colom- Colombian manufacturing firms, while also analyz- bia’s third most populous municipality, is the most ing the broader characteristics of Colombian munic- productive in apparel and metal products; Medellín ipalities that may either promote or hinder manufac- in wood products and computing machinery; Barran- turing firm ­productivity. quilla in textiles and printing; Manizales in machin- Balat and Casas use two distinct, and extremely ery; Bucaramanga in rubber and plastic products; and data sets. The first is a firm-level input-output rich, ­ Cartagena in food and paper ­ products. panel that, for each firm, provides annual data on Further analysis reveals that an industry’s local- the revenue generated from each product sold, use ization in a municipality, as measured by its size rel- of raw materials, investment, number of employees, ative to the size of the manufacturing sector there, overall wage bill, and, crucially, the municipality in is a significant and robust determinant of firm which the firm is ­ located.b The data set covers 2005– ­productivity. f By contrast, a diversified manufac- 2013 and 22 manufacturing industries, for a total of turing base and a high level of competition within almost 27,000 firm-year o ­ bservations.c Using these a municipality appear to be damaging for firm pro- (continued) T he E m p iri c a l D eter m inants o f Cit y Pr o d u c ti v it y    109 BOX 3.3  The Determinants of Manufacturing Firm Productivity across Colombian Municipalities (continued) ductivity, although these results are less ­ robust. The are also found to be more productive in municipali- overall scale of manufacturing activity within a ties that provide higher-quality schools, as reflected, municipality is likewise estimated to be unimport- for example, in lower student–teacher ratios and ant for firm ­productivity. higher scores in mandatory high school exit e ­ xams.g Beyond the nature of the agglomeration environ- High rates of crime and violence exert statistically ment, Balat and Casas also find other municipali- and economically significant negative effects on ty-level characteristics (many of which have been T FP. Thus, a one-standard-­ ­ deviation increase in the highlighted as of more general importance in the theft rate or in the number of terrorist attacks that main text of this chapter) to be important for firm occur in a municipality is associated with a decrease T FP. Increased municipal expenditure on trans- ­ in firm TFP of up to 5 ­ p ercent. Likewise, average port infrastructure is beneficial for TFP, whereas productivity losses stemming from the presence an undesirable business environment ­ generated by of paramilitary and drug-­ t rafficking groups in a increased taxes on firms adversely affects i ­t. Firms municipality are estimated to be up to ­ percent. 3.2 ­ Source: Based on Balat and Casas 2017. a. For consistency with chapters 1 and 2, the estimate of Colombia’s urban population is derived by applying the cluster algorithm of Dijkstra and Poelman (2014) to LandScan 2012 gridded population ­data. b. The source of this panel data is the Superintendencia de Sociedades (Superintendence of Corporations) ­database. c. Balat and Casas (2017) exclude firms engaged in the manufacture of coke, refined petroleum products, nuclear fuel, and basic metals (including metals such as gold, silver, platinum, and nickel) from their ­analysis. This is because these firms are commodity producers, and therefore their dynamics are different from those of other manufacturing ­firms. d. Balat and Casas (2017) estimate firm-level TFP using an extended Cobb-Douglas production function with industry-specific ­coefficients. Their approach helps them to address major methodological challenges that arise from the inability to observe productivity shocks that might affect firms’ input choices (endogeneity) and entry–exit decisions ( ­ selection). e. This data set is maintained by the Center of Economic Development Studies at the Universidad de los A ­ ndes. f. In constructing measures of localization and diversity, Balat and Casas (2017) experiment with several measures of the size of an industry, including number of firms in that industry, its employment, size of capital stock, and ­production. g. Such as Saber 11, a standardized test similar to the SAT in the United S ­ tates. that is conducive to the fostering of wider Conclusions positive agglomeration ­ effects. Hence, cur- Although compositional differences in the rent policies and levels of urban infrastruc- workforce associated with the sorting of ture may not be sufficient to prevent workers across locations have a major role negative congestion effects offsetting more in explaining productivity variations, urban general positive agglomeration ­ effects. success in the LAC region is also crucially Results based on WBES firm-level data dependent on the existence of a strong over- show that additional elements in a city’s all stock of human capital ­ (that is, a high environment matter for productivity, such as level of city skill) and, at least for certain prevention of crime and theft (which are countries, good access to large consumer notoriously high in LAC cities) targeted at and supplier markets through transport firms, provision of modern infrastructure ­ networks. By contrast, other types of posi- (Internet access), and access to formal bank- tive agglomeration effects, such as those, finance. The findings on the importance ing ­ for example, that we expect to arise from of skill and crime are further reinforced by a labor pooling or more general knowledge case study of the determinants of firm pro- spillovers, seem to be weak to nonexistent ductivity in Colombian ­ municipalities. in the region’s ­ cities. A possible hypothesis Given the importance of the theories of that may explain this is that, in the context market access and of HCEs in explaining of high urban population densities, LAC cit- urban success in the LAC region, the next ies may not have an enabling environment two chapters examine them in more ­ depth. 110   RAISING THE BAR Annex 3A: Results of ­ pproach. Nevertheless, the overall qualita- a Regressions on the tive picture remains the ­ s ame. Hence, the Determinants of Underlying estimated elasticity of a worker’s nominal Productivity Variations Based wage with respect to the population density on the Single-Stage Approach of the area in which she or he lives drops As an alternative to the two-stage approach drastically as we introduce, first, an area’s to analyzing the determinants of productiv- overall average number of years of school- ity across subnational areas, many papers ing, then its market access—­ columns 2 and have followed a single-stage approach 3 of table 3A.1. Its ­ s tatistical significance (Duranton 2016; Chauvin et ­ a l. ­ 2 017). declines. Whether the variable becomes also ­ Instead of first estimating location premiums statistically insignificant at all conventional and then regressing these premiums against levels (up to the 10 percent level) depends potential area-level determinants of produc- crucially on how we cluster the standard tivity, the single-stage approach simply errors. When they are clustered at the area ­ includes the potential area-level determi- level, population density remains signifi- nants directly in a regression of a worker’s cant at the 5 ­ level. It may be argued, percent ­ (natural log) wage on her or his observable however, that this is too restrictive because characteristics (directly in a Mincerian-style it rules out correlation between errors for wage ­regression).34 workers who live in, for example, neigh- As table 3A.1 shows, when we adopt this boring a ­ reas. When we instead cluster approach, we obtain higher estimated coef- standard errors at national level, popula- ficients on population density than those tion density is insignificant even at the reported in table ­ 3.2 using the two-stage 10 percent ­ level. 35 TABLE ­3A.1  Results of Regressions on the Determinants of Underlying Productivity Variations Based on the Single-Stage Approach Dependent variable: Nominal hourly wage (ln) (1) (2) (3) (4) (5) Population density (ln) ­0.057*** ­0.024*** ­0.012** ­0.035*** ­0.017*** Average number of years of ­0.636*** ­0.605*** schooling (ln) Percentage of working-age popu- ­0.015*** ­0.014*** lation with higher education Market access (ln) ­0.013*** ­0.016*** Mean air temperature (ln) ­0.013 ­0.018 ­0.020 ­0.018 ­0.022 Terrain ruggedness (ln) ­− 0.000 ­0.002 ­0.006 ­− 0.005 ­− 0.000 Total precipitation (ln) ­− 0.044*** ­− 0.027*** ­− 0.025*** ­− 0.035*** ­− 0.033*** Constant ­−2.02*** ­−3.41*** ­−3.49*** ­−2.09*** ­−2.27*** No. of observations 4,000,142 4,000,142 3,766,690 4,000,142 3,766,690 R2 ­0.337 ­0.346 ­0.349 ­0.343 ­0.346 2 Adjusted R ­0.337 ­0.346 ­0.349 ­0.343 ­0.346 Source: Quintero and Roberts 2017; population data: Gridded Population of the World, v4. Note: All estimations are based on the broad sample (all wage/salary employees age 14–65 ­years). In all columns, the dependent variable is the natural log of the nominal hourly wage in the main ­occupation. All regressions include country-year fixed effects and observable characteristics of individual workers (age, age squared, marital status, gender, and number of years of schooling), both individually and interacted with a full set of country ­dummies. ***p < 0.01. **p < 0.05. *p < 0.1. Standard errors are clustered at the area level. T he E m p iri c a l D eter m inants o f Cit y Pr o d u c ti v it y    111 Notes Brazil, China, and ­ India. Duranton (2016) examines these two theories for Colombia, in 1. The basic idea here is that through skill-selec- addition to investigating empirically the role tive migration workers with different charac- of market access in driving productivity differ- teristics “sort” into different ­ cities. It is this ences across ­cities. Authors who have studied sorting that leads to compositional differences the empirical relationship between subna- workforce. But sorting is not necessarily in the ­ tional levels of productivity and market access independent of the three theories of urban without necessarily accounting for agglomera- success that we describe ­ below. For example, tion economies and HCEs in a developing the theory of human capital externalities country context include Fally, Paillacar, and (HCEs) could not explain productivity differ- Terra (2010) for Brazil and Hering and Poncet ences across cities without the compositional (2010) and Roberts et ­ China. al. (2012) for ­ differences in the workforce that arise from 7. For a detailed explanation of why nominal sorting. Similarly, if the amenities that differ- ­ rather than real wages provide the appropri- ent cities offered were the same, there would ate measure of productivity, see Combes and be no incentive for more highly educated Gobillon ­(2015). workers to sort into larger and more densely 8. Household survey microdata are available for populated cities absent higher returns to Brazil from the Socio-Economic Database for human capital in those ­ cities. Latin America and the Caribbean ( ­SEDLAC). 2. Agglomeration economies can also be sepa- Although this data can be used to calculate the rated into localization economies and urban- overall urban–rural wage ratio, it cannot ization ­ economies. Localization economies be used for the more general analysis of subna- are the positive externalities associated with tional productivity variations that we perform the clustering together within a city of firms in the “Large Subnational Variations in from the same industry (Marshall 1890). Productivity” and “Explaining Underlying Urbanization economies are the positive exter- Variations in Productivity” sections of this chap- nalities associated with the geographic con- ter. For consistency between sections, we there- centration of a set of different industries fore prefer to use IPUMS International data for within a given city (Jacobs ­ 1969). throughout. Although SEDLAC does not Brazil ­ 3. Duranton and Puga (2004) identify three classify areas as urban or rural, it identifies mechanisms underpinning agglomeration whether households are urban and ­ rural. We economies, which they term “sharing, match- follow the SEDLAC documentation (CEDLAS ing, and ­learning.” and World Bank 2014) in assuming that urban 4. If unskilled and skilled workers are comple- (rural) households live in urban (rural) ­ areas. mentary inputs in the production processes 9. In these regressions, age proxies a worker’s within a city’s firms, then this can also gener- ­experience. ate observationally similar effects to human 10. Controlling for (time-invariant) unobservable capital ­externalities. This issue is explored in characteristics of workers requires panel data detail in chapter ­5. (see, for example, Combes, Duranton, and 5. This chapter focuses on domestic market Gobillon 2008; D’Costa and Overman, 2014), access—access to markets within the same which are unavailable for a large sample of country. However, as is discussed in chapter 4, ­ LAC ­countries. access to international markets through both 11. We use Admin-2 level data for 9 of the 16 ports and airports is also likely to be an import- countries in our sample: Bolivia, Brazil, ant determinant of a city’s ­ productivity. To Colombia, Dominican Republic, El Salvador, help distinguish between agglomeration econ- Honduras, Mexico, Nicaragua, and ­ Uruguay. omies and market access, the measure of mar- For the other seven countries; we use Admin-1 ket access used in this chapter excludes from its level data for Argentina, Guatemala, and calculation an area’s own p ­opulation. The Panama; and Admin-3 level data for Chile, exclusion of an area’s own population also Costa Rica, Ecuador, and ­ Peru. Admin-2 areas helps to mitigate reverse causality ­ concerns. correspond to municipios in Brazil, Colombia, 6. Chauvin et ­ al. (2017) examine empirically the Dominican Republic, El Salvador, Honduras, roles of agglomeration economies and HCEs Mexico, Nicaragua, and Uruguay, and to pro- in driving urban productivity differences in Bolivia. vincias in ­ 112   RAISING THE BAR 12. Such as Bogotá, Buenos Aires, Lima, Mexico 19. The Pearson correlation coefficient for pop- City, Panama City, Santa Cruz, Santiago, and ulation density and average number of years São ­Paulo. of schooling is ­ 0.30, whereas that for popu- 13. Again, the observable worker characteristics lation density and market access is ­ 0.67. For that we control for are number of years of average number of years of schooling and schooling, age and its square, gender, and market access, the correlation coefficient is marital ­status. 0.32. All estimated correlation coefficients ­ 14. As with the regressions in the previous section, are significant at the 5 percent ­ level. we include survey-year fixed ­ effects. We also 20. As noted, in our regressions we are able to estimate the regressions without a constant, control only for the sorting of workers based allowing us to include a full set of subnational on their observable ­ characteristics. However, area ­dummies. sorting may also be taking place on the basis 15. We use average number of years of schooling of such unobservable characteristics of among the working-age population rather workers as their ability and ­ motivation. To than only employed workers as our measure the extent that these unobservable character- of skill because, in principle, there is no reason istics of workers are not correlated with why knowledge may not spill over from an their observable characteristics, our estimate unemployed to an employed member of the of the coefficient on population density will ­workforce. be ­biased. If more able and motivated work- 16. Our specification of market access follows a ers sort toward denser areas, then we would classic Harris-style formulation (Harris 1954), expect the direction of this bias to be in which the market access of an area is the upward, thereby strengthening our result of travel-time discounted sum of populations in all an estimated absence of agglomeration other subnational areas within the same ­economies. ­ country. In calculating market access, we 21. As an alternative to the two-stage approach to exclude an area’s own population, which helps analyzing the determinants of subnational distinguish the market access variable from productivity that we have adopted in this population density, while mitigating endogene- chapter, many papers have followed a single-​ ity problems associated with reverse ­ causation. stage approach (Duranton 2016; Chauvin We view market access as capturing the benefi- al. ­ et ­ 2017). Annex 3A shows that, when we cial effects, for final goods producers, of access apply this approach to our LAC sample, we to consumer markets and access to suppliers of obtain higher estimated coefficients on popu- intermediate ­ inputs. Although these two types lation density than those reported in table ­ 3.2 of access (to consumers and suppliers) are, in using the two-stage ­ approach. Nevertheless, principle, two separate concepts, empirical the overall qualitative picture regarding weak work has found it hard to separate them positive agglomeration effects beyond HCEs because of their extremely high correlation (see, remains the ­ same. A further empirical concern for example, Redding and Venables ­ 2004). with the regressions in table 3 ­ .2, and not dis- 17. We prefer population density to overall popula- cussed in the main text, is that there may be tion as our measure of agglomeration because endogeneity stemming from either reverse the subnational administrative areas that we causation from an area’s location premium to use in our analysis only provide approxima- its levels of population density, human capital, tions of “true” cities (see chapters 1 and ­ 2). and market access or omitted variables that 18. Like the analysis in annex 3A, Hering and are correlated with both the left- and right- Poncet (2010) regress individual wages hand side v ­ ariables. Addressing these endoge- directly on market access controlling for indi- neity concerns would require the additional vidual observable c ­ haracteristics. Rather than use of instrumental variables ­ estimation. We base their market access variable on popula- would require a minimum of three instru- tion, they derive the variable through a two- ments, one for each of our key independent step procedure that first involves the variables. Finding plausible instruments is, ­ estimation of a gravity trade ­ equation. We however, ­ tricky. Although data on precolonial were unable to apply such a procedure owing population densities represent a possible to an absence of bilateral trade flow data for instrument for population density, this would subnational ­areas. still leave us two instruments ­ short. T he E m p iri c a l D eter m inants o f Cit y Pr o d u c ti v it y    113 22. The corresponding single-stage approach 1 million ­ residents. Because it is a simple results are in columns 4 and 5 in table 3A.1, binary dummy variable for whether a firm is annex ­ 3A. As can be seen, in this case, popula- in a national capital or big city, the measure tion density remains significantly positive at the of agglomeration used by Reyes, Roberts, 1 percent level, even after controlling for the and Xu is somewhat cruder—given limita- share of working-age population who possess tions of the WBES data—than the continu- higher education and for market ­ access. ous measure of population density used in 23. Figure ­ 3.5 does not show results for all 16 the earlier worker-based ­ analysis. countries in our overall ­ sample. Where results 28. This variable is denoted “Inf competition” in are not shown, this is either because of 3.3. Informal firms tend to be relatively table ­ extremely wide confidence intervals due to unproductive compared with formal firms, in small numbers of subnational areas or, as with part because they possess less managerial Panama, a lack of sufficient observations to capital along with little organizational com- permit ­estimation. plexity and related know-how (La Porta and 24. In contrast to figure ­3.5c, chapter 4 reports a Shleifer ­2014). Given this, Reyes, Roberts, significant positive effect of market access on and Xu hypothesize that firms that face productivity for Mexican subnational ­ areas. higher informal competition are likely to The difference in results may be partly attrib- benefit less from positive spillover effects utable to differences in data and partly due to within their ­ industries. differences in m ­ ethods. For example, chapter 4 29. This includes controlling for the industry bases its measure of productivity on nighttime that a firm belongs to, as well as the extent lights data rather than nominal wages, while it to which a firm is foreign owned, the owner- also merges some subnational areas to form ship share of its largest owner, whether a larger metropolitan ­areas. Chapter 4 also relies firm is an exporter, a firm’s size as measured on panel rather than cross-sectional data, but by the number of workers it employs, and a does not control for sorting based on the firm’s ­age. observable characteristics of individual 30. Besides the results in table 3.3, Reyes, workers. This does not rule out the impor- ­ Roberts, and Xu also estimate regressions for tance for productivity of improving access three-year growth rates of labor productivity to international markets—both through and TFP, and a firm’s export ­ share. We do improved road and rail access to ports and air- not discuss these results because our focus is ports and improvements in both port and air- more on the long-run determinants of city port infrastructure—even where the coefficient productivity. Interactions for LAC with both ­ on market access is statistically ­ insignificant. “Web intensity” and “Inf competition” are 25. Estimated variations in the strength of HCEs included in table 3.3 because these interac- may also be attributable to differences in the tions are significant in one or more of these quality of education across countries, such supplemental regressions estimated by Reyes, that one additional year of schooling in, for Roberts, and Xu. example, Nicaragua does not accomplish the 31. The exception is tax rate obstacle, which has a same amount of learning as an additional year significant negative effect on TFP at the 5 per- of schooling in Bolivia. The extent to which cent level. the estimated variations in the strength of 32. This suggests that access to finance boosts HCEs are attributable to educational quality labor productivity through facilitating an differences is an important area of future increase in the capital–labor ratio rather than ­research. necessarily fostering ­ innovation. 26. Labor productivity is measured as sales 33. See https://www.economist.com/blogs/graphic​ divided by the number of permanent 2017/03/daily-chart-23. However, the detail/​ ­ employees. TFP is estimated as the residual LAC region’s high homicide rates are primar- from industry-specific production functions ily related to gang warfare associated with with log value added as the dependent vari- drug ­trafficking. It is not clear how much this able and log capital and log labor as the inde- violent crime affects ­ firms. pendent ­variables. 34. Our preference for the two-stage approach is 27. In this case, a large city is defined as either a partly because, unlike the single-stage national capital or a city with more than approach, it allows us to obtain estimates of 114   RAISING THE BAR levels of underlying productivity for subna- L atin A me ric a and the C ­ aribbean. tional ­areas. The main concern of the two- Washington, DC: World ­ Bank. stage approach is that it might not provide Chauvin, ­ J. ­ E . Glaeser, ­ P., ­ Y. Ma, and ­ Tobio. K. ­ reliable standard errors in the second stage, 2017. “What Is Different about Urbanization ­ because the location premiums that act as the in Rich and Poor Countries? Cities in Brazil, dependent variable in the second-stage regres- China, India and the United ­ States.” Journal of sion are themselves estimates, and thus con- Urban Economics 98 (C): ­ 17–49. tain an estimation error that is likely greater Clarke, ­ G., ­ C. ­Z. Qiang, and ­ L. ­ C. ­Xu. ­2015. “The for smaller subnational ­ areas. However, given Internet as a General-Purpose T ­ echnology.” that the number of observations (about 4 mil- Economics Letters 135 (C): ­ 24–27. lion) that underlies our estimation of the loca- Cohen, ­ W. ­M., and ­ R. ­ Levin. ­ C. ­ 1989. “Empirical tion premiums is large, and that 9 ­ 9.4 percent Studies of Innovation and Market S ­ tructure.” of estimated location premiums from our first- In Handbook of Industrial Organization, stage regressions are significant at the 5 per- Volume 2, edited by ­ R. Schmalansee and ­ R. ­D. cent level (only 38 of 5,872 location premiums Willing, ­ 1059–1107. Amsterdam: ­ Elsevier. are insignificant at the 5 percent level), this Cohen, ­ W. ­M., and ­ K lepper. ­ S. ­ 1996. “Firm Size concern seems to be relatively ­ minor. a nd t he Nat u re of I n novat ion w it h i n 35. The “ideal” level of clustering for the stan- Industries: The Case of Process and Product dard errors likely lies between the area and ­R&D.” Review of Economics and Statistics national ­levels. 78 (2): ­232–43. Combes, ­ P., ­G. Duranton, and ­ L. ­ Gobillon. ­ 2008. “Spatial Wage Disparities: Sorting Matters!” References Journal of Urban Economics 63 (2): ­ 723–42. Acs, Z ­ .­ J., ­D. B­ . Audretsch, and M ­ .­ ­ eldman. P. F Combes, ­ P., and ­ L. ­ G obillon. ­ 2 015. “The 1994. “R&D Spillovers and Recipient Firm ­ Empirics of Agglomeration E ­ conomies.” In ­Size.” Review of Economics and Statistics 76 Handbook of Regional and Urban Economics, (2): ­336–40. Volume 5 , edited by ­ G . Duranton, ­ J . V. Balat, J ­ ., and ­ C. ­Casas. 2 ­ 017. “Firm Productivity Henderson, and ­ W. Strange, ­ 2 47–348. and Cities: The Case of ­ Colombia.” Background Amsterdam: ­Elsevier. paper for this book Washington, ­ DC. D’Costa, ­ S ., and ­ H. ­ O verman. ­ G. ­ 2 014. “The Barrios, ­ L . Bertinelli, and ­ S ., ­ E. ­S trobl. ­2006. Urban Wage Growth Premium: Sorting or “Geographic Concentration and Establishment Learning?” Regional Science and Urban Scale: An Extension Using Panel ­ D ata.” Economics 48 (C): ­ 168–79. Journal of Regional Science 46 (4): 7 ­ 33–46. Demirgüç-Kunt, A ­ ., and ­ V. M­ aksimovic. 1 ­ 998. Beck, ­ T., ­A. Demirgüç-Kunt, and ­ Maksimovic. V. ­ “Law, Finance, and Firm G ­ rowth.” Journal of 2 005. “Financial and Legal Constraints to ­ Finance 53 (6): ­ 2107–37. Growth: Does Firm Size Matter?” Journal of Dijkst ra, ­ L ., and H ­ . ­ P oel man. ­ 2 014. “A Finance 60 (1): ­ 137–77. Harmonised Definition of Cities and Rural Bernard, ­ A. ­B ., ­ J. ­B . Jensen, ­ S. ­ J. Redding, and Areas: The New Degree of Urbanization.” ­ P. ­K. ­ S chott. ­ 2 007. “Firms in International Regional Working Paper, Directorate-General ­Trade.” Journal of Economic Perspectives 21 for Regional and Urban Policy, European (3): ­105–30. Commission, ­Brussels. Bloom, ­ N ., ­A . Mahajan, D ­ . McKenzie, and J ­. Duranton, G. 2015. “Growing through Cities in Roberts. ­ ­ 2010. “Why Do Firms in Developing Developing Countries.” World Bank Research Countries Have Low Productivity?” American Observer 30(1): 39-73. Economic Review 100 (2): ­ 619–23. ———. ­ 2 016. “Ag g lomerat ion E f fec t s i n Branson, ­ J., ­A. Campbell-Sutton, ­ G. ­M. Hornby, ­C olombia.” Journal of Regional Science 56 ­ D. ­ D . Hornby, and ­ C. ­ H ill. ­ 2 016. “A (2): ­210–38. Geospatial Database for Latin America and Duranton, ­ G ., and ­ D. ­ P uga. ­ 2 004. “Micro- the Caribbean: G ­ eodata.” Southampton, U.K.: Fou nd at ion s of Urba n A g g lomerat ion University of S ­ outhampton. ­E conomies.” In Handbook of Regional and CEDLAS (Center for Distributive, Labor and Urban Economics , Volume 4: Cities and Social Studies) and World B ­ ank. 2­ 014. A Guide Geography, edited by ­ V. Henderson and ­ J. ­ J.-F. to SEDLAC—Socio-Economic Database for Thisse, ­ 2063–2117. Amsterdam: ­ Elsevier. T he E m p iri c a l D eter m inants o f Cit y Pr o d u c ti v it y    115 Fally, ­ T., ­R . Paillacar, and ­ C. ­ Terra. ­ 2 010. Levine, ­ R. 1 ­ 997. “Financial Development and “Economic Geography and Wages in Brazil: Economic Growth: Views and A ­ genda.” Journal Evidence from Micro-data.” Journal of of Economic Literature 35 (2): ­ 688–726. Development Economics 91 (1): ­ 155–68. Li, ­ D., ­Y. Lu, and ­ M. ­ Wu. ­ 2 012. “Industrial Fay, ­ M ., ­L. ­ A . Andres, ­ C. ­ J. ­E . Fox, ­ U. ­G. Agglomeration and Firm Size: Evidence from Narloch, ­ S . Straub, and M ­ .­ A. ­ Slawson. 2 ­ 017. ­C h i na.” R eg io n al S c i e n c e a n d Urba n Rethinking Infrastructure in Latin America Economics 42 (1–2): ­ 135–43. and the C aribbean; Spending Better to Li, ­ W., ­ X. C ­ . Long, and ­ L. ­ C. ­ 2 017. X u. ­ Achieve More. Washington, DC: World ­ Bank. “Regulation, Agglomeration, and the Reversal Fujita, ­ M ., ­ P. Krugman, and ­ A. ­ J. ­Venables. of Fortune bet ween China and I ­ndia.” ­1999. The Spatial Economy: Cities, Regions Working paper, World Bank, Washington, ­ DC. and International ­ Trade. Cambridge, MA: Marshall, ­A . ­1890. Principles of ­ E conomics. MIT ­Press. London: Macmillan and ­ Co. Glaeser, ­ E. ­ 1999. “Learning in C ­ ities.” Journal of Melitz, ­ M. ­ J., and ­ G. ­ Ottaviano. ­ 2008. “Market Urban Economics 46 (2): ­ 254–77. Size, Trade and ­ P roductivity.” Review of Greenstone, ­ M ., ­R . Hornbeck, and ­ E. ­Moretti. Economic Studies 75 (1): ­ 295–316. 2010. “Identifying Agglomeration Spillovers: ­ Mincer, ­J . ­1974. Schooling, Experience, and Evidence from Winners and Losers of Large ­E arnings . New York: National Bureau of Plant ­Openings.” Journal of Political Economy Economic ­Research. 118 (3): ­536–98. Moretti, ­ E. ­2004. “Human Capital Externalities Harris, C ­ .D­ .1 ­ 954. “The Market as a Factor in in ­C ities.” In Handbook of Regional and the Localization of Industry in the United Urban Economics, Volume 4: Cities and ­States.” Annals of the Association of American Geography, edited by ­ V. Henderson and J J. ­ ­ .-F. Geographers 44 (4): ­ 315–48. Thisse, ­ 2243–91. Amsterdam: ­ Elsevier. Henderson, ­ J. ­ V. ­2010. “Cities and Development.” Overman, ­ H. ­ G ., and ­ A. J ­. V­ enables. ­ 2 005. Journal of Regional Science 50 (1): ­ 515–40. “Cities in the Developing World.” Center for Hering, ­ L ., and ­ S. ­ Poncet. ­ 2010. “Market Access Economic Performance Discussion Paper 695, and Individual Wages: Evidence from ­ China.” London School of Economics and Political Review of Economics and Statistics 92 (1): Science, ­London. ­145–59. Quintero, ­ L ., and M ­ .­ Roberts. 2 ­ 017. “Explaining Holmes, ­ T. J ­ ., and J­.J­.S ­ tevens. 2 ­ 002. “Geographic Spatial Variations in Productivity: Evidence Concentration and Establishment ­ S cale.” from 16 LAC ­ Countries.” Background paper Review of Economics and Statistics 84 (4): for this book World Bank, Washington, ­ DC. ­682–90. Rauch, ­ J. ­E. ­1993. “Productivity Gains from Jacobs, ­J . ­1969. The Economy of C ­ ities. New Geographic Concentration of Human Capital: York: Random ­ House. Evidence from the C ­ ities.” Journal of Urban Krugman, ­P. ­1 991a. Geography and ­ Trade . Economics 34 (3): ­ 380–400. Cambridge, MA: MIT ­ Press. Redding, ­ S., and ­ A. ­ Venables. ­ J. ­ 2004. “Economic ———. ­ 1991b. “Increasing Returns and Economic Geography and International ­ I nequality.” ­Geography.” Journal of Political Economy 99 Journal of International Economics 62 (1): (3): ­483–99. ­53–82. K r ug man, ­ P., a nd ­ A. ­J. ­Venables. ­ 1 995. Reyes, ­ M. Roberts, and ­ J., ­ L. ­ C. ­ Xu. ­2017. “The “Globalization and the Inequality of N ­ ations.” Heterogeneous Growth Effects of the Business Quarterly Journal of Economics 110 (4): Environment: Firm-Level Evidence for a Global ­857–80. Sample of C ­ ities.” Policy Research Working La Porta, ­ R., and ­ A. ­Shleifer. ­2014. “Informality Paper 8114, World Bank, Washington, ­ DC. and ­D evelopment.” Journal of Economic Roberts, ­ M ., ­ U. Deichmann, B ­ . Fingleton, and Perspectives 28 (3): ­ 109–26. T. ­ ­ S hi. ­ 2 012. “Evaluating China’s Road to La Porta, ­ R., F ­ . Lopez-de-Silanes, A ­ . Shleifer, and Prosperity: A New Economic Geography R. ­ ­ Vishny. ­ 1998. “Law and ­ Finance.” Journal Approach.” Regional Science and Urban of Political Economy 106 (6): ­ 1113–55. Economics 42 (4): ­ 580–94. Transport Infrastructure and Agglomeration in Cities Harris Selod and Souleymane ­ Soumahoro 4 Introduction through the lens of investment in transport, and its implications for agglomeration effects Transport investment can contribute to cities’ in cities if such investment were increased— productivity. Improved transport systems, for ­ important issues for two main ­ reasons. First, example, may lower production costs in an transport investment can be critical to industrial cluster and generate efficiency make cities function more efficiently and gains through localization ­e conomies . become more s ­ ustainable. Second, there is Similarly, they may generate, through urban- ample empirical evidence suggesting that ization economies, positive externalities to improved transport systems increase produc- all firms in large urban ­ centers.1 There are tivity in cities by facilitating the spatial con- also potentially wider economic benefits, centration of firms (Ghani, Goswami, and which may themselves directly or indirectly Kerr 2016), by increasing firm birth (Holl induce sizable effects on productivity, includ- 2004) and employment (Mesquita Moreira ing increased employment and market oppor- al. 2013), and by improving firm efficiency et ­ tunities and enhanced human capital (Datta ­2012). (Similarly, chapter 3 documents externalities (HCEs) in education and health that access to markets is a significant determi- due to improved access to ­ transport. nant of productivity for a sample of 16 LAC Yet, despite increasing recognition of the countries, even with cross-country heteroge- importance of infrastructure for growth, the neity in the estimated ­ elasticities.) stock of physical capital in Latin America and The main findings are as follows: the Caribbean (LAC) is thought to be low for the region’s development ­ level. Recent data •  Despite recent years’ growing policy suggest that paved road density in the LAC enthusiasm for infrastructure invest- region is only marginally higher than in Sub- ment, LAC continues to exhibit low road Saharan Africa (SSA) and about one-quarter density and poor road quality, which that of the Middle East and North Africa likely translates into deficient access to (MENA), the next least-performing region transport infrastructure around cities as (Dulac ­2013). well as high congestion in ­cities. This chapter explores the issue of low •  The prevalence of high physical and physical capital stock in the LAC region nonphysical transport costs in the LAC 117 118   RAISING THE BAR region is a major constraint to domes- also shows t hat reduced congest ion , tic and international ­ trade. It is also the rather than the increased stock of roads, is source of negative externalities (conges- the relevant ­ mechanism. tion and pollution) that challenge the In terms of the productivity effects of productivity and sustainability of LAC transport capital in the urban space, various ­cities. papers examine the role of transport invest- °° Physical investment can be accompa- ment in the movement of inputs within cities, nied by policy reforms that encour- especially labor, an issue that has been exten- age competition in the transport sively investigated in the context of high- and industry and improve logistics and middle-income ­ countries. Evidence from ­ U.S. customs ­procedures. cities suggests, for example, that roads °° Inefficient regulations to reduce con- enhance productivity in cities through their gestion and related negative external- stimulating effects on employment (Duranton ities could be complemented with, or and Turner 2012) and on domestic trade incentives.2 replaced by, price ­ flows (Duranton, Morrow, and Turner 2014). •  In Mexico (a case study), investment in In Mexico, Gonzalez-Navarro and Quintana- roads is generally associated with local Domeque (2016) exploit a random allocation job growth, increased manufactur- of public funding to street paving to provide ing specialization, and local economic evidence of the beneficial effects of road ­development. upgrading. Their findings suggest that road ­ paving increased household-level acquisition of durable goods (vehicles, appliances, and Transport, Agglomeration, and home improvements) through its positive Productivity: A Brief Review effect on property values and access to c ­ redit. Transport investment can in theory have Another effect associated with transport significant effects on productivity by exoge- improvement includes the decentralization of nously lowering the costs of labor and inter- production and population from core cities to mediate goods for firms (Venables 2007; peripheral areas (Baum-Snow 2007; Baum- Graham ­ 2 007). However, in reviewing Snow et ­ al. 2012), where land is cheaper and recent estimates of the relevant elasticities, the intercity transport network more easily t he literat u re ack nowledges nua nced ­ accessible. There is also evidence from the ­ f indings (see Gramlich 1994; Deng 2013; United States (Duranton and Turner 2011) and Redding and Turner 2014; Straub 2011; Japan (Hsu and Zhang 2012) that additional Trebilcock and Rosenstock 2015; Berg et al. roads may incentivize intracity (noncommer- 2017). For example, from a macroeconomic ­ ongestion. cial) driving, therefore not relieving c perspective, several studies at the country The literature also examines the economic or state level find positive returns to trans- consequences of improved intercity transport port capital (Aschauer 1993; Calderón and infrastructure, in developed and developing Servén 2004a), whereas others identify countries. In the United States, for example, ­ insignificant effects (Holtz-Eakin 1994; highway connection is found to increase Garcia-Mila, McGuire, and Porter 1996). 3 earnings in services (Chandra and Thompson Holtz-Eakin (1994) argues, for example, 2000), boost the wage of skilled relative to that the impact of infrastructure capital on low-skilled workers (Michaels 2008), and aggregate productivity is unlikely to be stimulate city-level specialization in heavy robust once the simultaneity bias linking goods through its effects on the weight of city the two is accounted ­ for. Fernald (1999) exports (Duranton, Morrow, and Turner challenges this view and provides evidence 2 014). In Brazil, Bird and Straub (2014) ­ of the causal productivity gains associated exploit the creation of Brasilia and subse- with road investment in vehicle-intensive quent infrastructure investment as a “natural industries in the United ­ States. The author experiment,” and find that access to roads T rans p o rt I n f rastr u c t u re an d A gg l o m erati o n in Cities   119 reduced the inequality in the spatial distribu- movement of agricultural products (cash tion of economic activities among Brazilian crops, timber, and livestock), minerals (cop- ­ regions. In India, Donaldson (forthcoming) per, silver, and coal), and people, the newly examines the welfare effects of colonial rail- developed transport network accelerated roads and finds that those railroads stimu- market integration and created opportunities lated trade, reduced trade costs and 2006). for specialization (Summerhill ­ interregional price gaps, and ultimately increased real agricultural ­ income. Railroads as an engine of growth in the 19th Similar positive productivity effects are and early 20th centuries.  Railroads are found near the “Golden Quadrilateral,” a described as the most attractive growth-­ major highway improvement project involving enhancing infrastructure in 19th-century Latin 5,846 km of roads in India (Ghani, Goswami, America (Coatsworth 1979; Summerhill ­2006). and Kerr ­2016). In China, roads and railroads The financing opportunities brought by the collectively contributed to an increase in first wave of globalization, coupled with county-level gross domestic product (GDP) improved political stability, generated more per capita (Banerjee, Duflo, and Qian ­ 2012). enthusiastic investment trends between 1870 Still in China, the effects of the country’s and ­1914.4 In 1900, nearly 55,000 km of rail- national expressway network exhibited signif- road track were in operation in the LAC region, icant subnational heterogeneity (Roberts et ­al. 75 ­ percent of which was in Argentina, Brazil, 2012), partially echoing the findings of Faber and ­ Mexico. Between 1900 and 1930, rail net- (2013) who documents a depressing effect of works continued to expand at an annual rate of the highway network on the income of periph- ­ 3.6 percent, before leveling off after World War eral regions due to shrunk industrial ­output. Depression. The development of I and the Great ­ In SSA, Storeygard (2016) identifies a causal the rail network in the LAC region appeared to link between low transport costs and city- be completed in the first half of the 20th cen- level growth as measured by nighttime ­ lights. tury and only to have accompanied the early In Ghana and Kenya, access to rail is found to urbanization. The share of the popu- stages of ­ stimulate local economic development in the lation living in urban areas, for example, short and long run (Jedwab, Kerby, and increased from 2 ­ 7.6 percent in 1930 to 7 ­ 5.4 Moradi 2015; Jedwab and Moradi ­ 2016). percent in 2000, but the aggregate length of railroads in service actually declined over the 4.1). This finding comes as no period (figure ­ Transport in Latin America and surprise given that the main function of rail- the Caribbean: History, roads was to link primary resources (agricul- Current State, and Challenges tural and mineral) to markets, including through ports, which became less important in History of Transport in the LAC Region subsequent phases of economic ­ development. Most historians agree that the economic The economic benefits of early investments modernization of Latin America in the 19th in railroads are well documented and include century coincided with the development of increased social savings and improved export physical infrastructure (Bulmer-Thomas e ­ t al. ­performance.5 Social savings—or the poten- ­ 2006). Innovations in transport helped many tial output gains from the reduction of trans- countries overcome the challenge of rugged port costs as a result of the more efficient terrain and hard-to-navigate w ­ aterways. The allocation of labor and capital—increased synergy of modern railroads with improved substantially with investment in ­ railroads. ports, and to a lesser extent roads, reduced In Argentina and Brazil, for example, social transportation costs, facilitated the allocation savings generated from railroad freight were of productive resources, and is believed to estimated to be about 26 and 22 percent of have spurred economic growth. By providing GDP in 1913, respectively (Summerhill for low-cost national and international ­ 2006). In Mexico, investment in railroads 120   RAISING THE BAR FIGURE ­4.1  Length of Railroad Track in Service and Urban Share in Latin America and the Caribbean, 1900–2007 145 80 135 70 125 Railway track in service (1,000 km) 60 Urban population share (%) 115 50 105 40 95 30 85 20 75 65 10 55 0 1900 1920 1940 1960 1980 2000 2020 Year Urban share Linear t of the track in service Track in service Source: Calculations based on the MOxLAD database (http://www.lac.ox.ac.uk/moxlad-database). Note: The MOxLAD database groups sources, including Mitchell International Historical Statistics, World Development Indicators, Financial Statistics of the International Monetary Fund, and those of the Economic Commission for Latin America and the C ­ aribbean. FIGURE ­4.2  Export Density and Railroad Density in Selected Latin were below 9 percent of GDP in 1 ­ 890. In the American Countries, 1900–30 United Kingdom and Germany, they were respectively 11 percent and 5 percent in 8 ­the  1890s. CUB 7 Investments in railroads, combined with Export density (log) SLV port improvements, also stimulated exports HTI 6 DOM URY and so accelerated the integration of Latin CRI ARG American economies into the world ­ market. 5 CHL HND GTM MEX Railroad coverage and export performance 4 VEN ECU NIC were highly correlated in 1900 –1930 PAN PER BRA COL (figure 4.2). The cross-country elasticity of ­ BOL 3 export density with respect to railroad den- 0 1 2 3 4 ­ eriod.6 This suggests sity was about 1 in that p Railway density (log) that investment in railroads boosted primary commodity exports or conversely that the Source: Calculations based on the MOxLAD database (http://www.lac.ox.ac.uk/moxlad-database). profitability of exports was important during Note: Linear regression linking the natural log of export density (export value per square kilometer of land area) to the natural log of railroad density (kilometers of railroads per 1,000 km2 of land area) in the first wave of globalization, leading to selected Latin American countries (using averages during ­1900–30). For a list of country abbreviations, more investment in ­ railroads. In both cases, see annex 2A. y = 1.001x + 3.3568; R2 = 0.7596. higher spending on railroad infrastructure was key to the economic success of Latin engendered a social rate of return that America during the Belle ­ Époque.7 exceeded 38 percent of GDP in ­ 1910. To put these figures into perspective, social savings The rise of roads from indigenous trails to from railroad freight in the United States modern motorways.  Before the railroad T rans p o rt I n f rastr u c t u re an d A gg l o m erati o n in Cities   121 revolution, early modern roads in Latin America public investments in roads exceeded were built upon precolonial indigenous 30 percent in 1937, up from ­ 7.3 percent in ­ routes. In the aftermath of World War I and 1925 (Cardenas ­ 1987). the Great Recession, a period that coincided Box 4­ .1 presents, as an illustration, the with stagnating railroad services, innovations history of road development in Mexico, in combustion engine technology accelerated which today has one of the most developed the acquisition of motor vehicles in Latin road networks in ­ the LAC region. America, generating new opportunities for Map ­ 4.1 shows the growth of the road profitable road ­ investment. The bulk of such network over the past three decades, for investment was, however, concentrated in a which we have reconstructed panel georefer- few ­ countries. In 1930, for example, enced information on the road extent and Argentina, Brazil, and Mexico—the three road type (see Blankespoor, Bougna et al. largest economies—collectively owned nearly 2017 for more ­detail). The maps show the rel- 80 percent of passenger vehicles and about atively recent investments in larger capacity 74 percent of commercial vehicles in Latin roads over the past two ­decades. ­America.8 These statistics could reflect a In contemporary LAC, roads have grad- bidirectional causal link between the growth ually emerged as a cost-efficient alternative in vehicle ownership and the development of to costly r ­ailroads. Figure 4 ­ .3 shows roads. In Mexico, for example, the share of ­ growth in the total length of road built in BOX ­4.1  History of Road Development in Mexico Investment in roads started with the Spanish Colony (Bess ­2014). In the 1960s, roads were built to respond (1521–1810), when roads were focused on transport- to the needs of private firms and to serve the national ing natural resources, especially silver and gold, to and state governments’ objective to build strategic the port of Veracruz to ship them to ­ Spain. Whereas relationships with rural communities (Bess ­ ­ 2017). A the focus in the late 19th and early 20th centuries government-owned company was set up to build more was on improving railroads throughout the country, Mexico. than 1,000 km of toll roads in the center of ­ roads started to receive more attention in the 1920s In the first half of the 1990s, a very ambitious pro- and ­ 1930s. Subsequently, the focus shifted again gram of road construction was launched, which led to reconstructing and enlarging the road network, to the construction of 5,800 km of privately financed which had been damaged during the Mexican Revo- highways (Foote ­ 1997).b Road-building policies were lution (Bess 2016a, 2016b). In the 1940s and 1950s, pursued in the following decades in continuity with with great hope in the promises of industrialization past policies, including a program for basic infra- and a generalized drive toward economic modern- structure in the 2000s and the construction and ization, Mexico perceived road construction as renovation of more than 23,000 km of roads in the necessary to allow market growth and improve the second half of the decade as part of a program to accessibility of subnational regions (Bess ­ 2014). This address rural ­ poverty. In 2014, the federal govern- conception led to the building of new roads in Mex- ment launched the 2014–2018 National Infrastruc- ico by state road-building agencies, which mobilized ture Program, c projecting a more than 20 percent large public spending and private domestic—and increase in the average annual investment in this sec- ­foreign—investment.a tor relative to the previous 20 years (Pérez-Cervantes In the early 1950s, the first freeway, from Mexico and Sandoval-Hernández 2 ­ 017). The most ambitious City to Acapulco, was opened and became a model part of this new program focuses on the south of the freeways. During this period, road for building future ­ country—no large project is planned for the ­ north. building played a key role in modernizing the Mexican The center of Mexico—more populated and richer— economy and in developing major commercial industries receives smaller ­ projects. a. The United States, for example, invested millions of dollars directly in Mexican transport industry and ­infrastructure. b. The extremely high tolls that made this investment possible, however, ended up preventing trucks from using these new roads, later forcing the government to restructure the highway network and implement a toll road rescue plan to help state-owned Mexican banks finance roads on nonmarket terms (Foote ­1997). c. Programa Nacional de ­Infraestructura. 122   RAISING THE BAR MAP ­4.1  The Evolution of the Road Network in Mexico, 1985–2016 a. 1985 b. 1999 N W E S c. 2008 d. 2016 Road Type Multilane divided Two lanes or divided Pavement 0 250 500 1,000 Kilometers Gravel or earth road Source: Blankespoor, Bougna et al. 2017. FIGURE ­4.3  Length of Roads and Urban Population Share in Latin America and the Caribbean, 1950–2000 3,500 80 3,000 70 60 2,500 Urban population share (%) Road length (1,000 km) 50 2,000 40 1,500 30 1,000 20 500 10 0 0 1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 Road length Urban share Source: Calculations based on the MOxLAD database (­http://www.lac.ox.ac.uk/moxlad-database). T rans p o rt I n f rastr u c t u re an d A gg l o m erati o n in Cities   123 the LAC region as well as the rate of urban- was far from enabling the LAC region to ization between 1950 and ­ 2000.9 Put into catch up with Western Europe and the perspective with figure 4 ­ .1, figure 4­ .3 United ­ States. In short, LAC and South Asia reveals two notable ­ p atterns. First, road were not too different in road density in 1961 investment began to take off the moment but became far apart in ­ 2000. Road density railroads in operation stagnated, corrobo- in the LAC region in 2000 was only margin- rating the idea that roads substituted for ally greater than that in South Asia in ­ 1961. declining rail services (Summerhill ­ 2006). This may reflect the recent high growth Second, unlike railroads, the development experience in South Asia, which translated of road net works in the L AC reg ion into much greater infrastructure ­ investment. occurred concurrently with urbanization Yet, LAC countries continue to rely (figure ­4.3). This matches recent empirical increasingly on roads as the main mode of findings from research documenting the passenger transport and surface ­ freight. LAC causal effect of road improvements on has the highest road occupancy rate in the urban population growth in Africa (Jedwab world with more than 800,000 vehicle-km to and Storeygard 2017a, 2017b; Blankespoor, paved lane–km (Dulac 2013)—nearly four Mesplé-Somps et al. 2017).10 times the rate in China and about twice that The upward trend in the aggregate Africa. Similarly, ownership of motorized in ­ road length between 1950 and 2000 (see vehicles is also on the rise and is expected to f igure 4.3) hides some interesting features ­ accelerate because of the increasing size of of the cu rrent state of transpor t the middle classes in Latin America (Fay i nfrastructure. On road density, the infra- ­ et ­ 2017). In comparison to rail, road is a l. ­ structure stock in the LAC region has consis- today the dominant mode of land transport, tently remained below that of other regions, with some 80 percent of the 1,920 ­ billion except for SSA and M ­ ENA. In 1961, for ton-km of total surface freight ­ (table 4.2). example, average road density per 100 km 2 However, the relative importance of roads in of land was about 4 ­ .2 km in the LAC region, surface freight varies widely across LAC 1 .7 km in SSA, and 1 ­ ­ .2 km in MENA ­c ountries.11 For example, although roads (table ­4.1). Between 1961 and 2000, road account for nearly all surface freight density grew at an average annual rate of in Argentina, Costa Rica, and Uruguay, rail- about 6­ .5 percent in the LAC region, which, roads remain an important complement in relative terms, was below the annual mode of transport to roads in Brazil, growth rate in SSA ­ (9.2 percent), MENA C olombia, a nd Mex ico, w it h nea rly ­ (15.4 percent), and South Asia ( percent). ­16.0 ­ 20 ­ p ercent of average surface freight Unlike South Asia, this rate of investment (figure ­4.4). TABLE ­4.1  Density of All Roads (Paved and Nonpaved) in Regions of the World, 1961–2000 Road density (kilometers per 100 km2 of land) Annual growth rate (%) Region 1961 2000 1961–2000 Latin America ­4.22 ­14.97 ­6.53 East Asia and Pacific ­12.04 ­19.98 ­1.69 Western Europe ­84.55 ­135.38 ­1.54 United States ­62.85 ­69.51 ­0.27 Middle East and ­1.17 ­8.18 ­15.42 North Africa South Asia ­12.85 ­93.06 ­16.00 Sub-Saharan Africa ­1.74 ­7.97 ­9.17 Source: Calculations based on Mitchell 2007. 124   RAISING THE BAR TABLE ­4.2  Modal Share of Surface Freight, by Current State of Transport: Growing Region, 2015 Investment Region Road (%) Rail (%) Recent data for 2000–2013 suggest that the Africa ­80.74 ­19.26 LAC economies recently witnessed a vigorous Asia Pacific ­56.77 ­43.23 pickup in total infrastructure investment after Europe ­55.96 ­44.04 two decades of stagnation (Fay et ­ 2017). In al. ­ Latin America ­78.28 ­21.72 2008–2013, annual average infrastructure Middle East ­97.23 ­2.77 spending accounted for ­ 2 .7 percent of LAC GDP, suggesting a regain in policy interest for North America ­62.41 ­37.59 ­ infrastructure. Similarly, transport investment Source: International Transport Forum ­2017. in particular exhibited an increasing trend in Note: The modal share of surface freight indicates the split of goods transported by road or rail, excluding waterways, which were not available 2000–2013 (figure ­ 4.5): for example, trans- for the ­analysis. port spending as a share of regional GDP FIGURE ­4.4  Modal Split of Surface Freight in Latin America and went up threefold, from a mere 0 ­ .4 percent in the Caribbean, 2012 2003 to about ­ 1.3 percent in 2009, led by public ­ i nve s t m e n t . T h i s i s a l l t h e 100 more remarkable given the region’s past per- formance, characterized by a decline in Share of surface freight (%) 80 transport investment at an annual ­ 0.69 per- cent in 1980–2002 (Calderón and Servén 60 2004b). Two factors might explain this recent ­ trend: the increasing interest of the public 40 sector and the growth momentum of 2003– 2013, sometimes referred to as the “Latin 20 American decade,” which may have contrib- uted to fiscal space for infrastructure 0 Brazil Argentina Mexico Colombia Uruguay Costa Rica ­development.12 The upward regional trend in transport Rail Road spending hides wide differences among ­ c ountries and subregional groups (fig- Source: Calculations based on Freight Transport and Logistics Yearbook ­2015 (https://publications. iadb.org/handle/11319/6885). ure ­4.6). In most subregions, no trend is dis- cernible in spending except for the Andean FIGURE ­4.5  Investment in Transport Infrastructure in Latin countries, where transport investment more America and the Caribbean, 2000–13 than doubled in ­ 2008–15. Specifically, such investment increased by the equivalent of 1.5 ­ 1.30 percent of GDP in Bolivia, ­ 2 .16 percent Transport investment in Latin America in Colombia, and ­ 2 .27 percent in ­ Peru. and the Caribbean (% of GDP) Transport investment declined, however, in 1.0 all countries of the Southern Cone, except for Paraguay and Uruguay, which saw increases equivalent to about 0 ­ .6 percent and 0.5 ­ 0.7 percent of their respective ­ GDPs. Similar to recent trends in road building (see previous section), spending in total trans- 0 port infrastructure (road, rail, waterways, air) in most LAC countries was largely devoted to 11 00 01 02 03 08 09 10 12 13 04 05 06 07 20 20 20 20 20 20 20 20 20 20 20 20 20 20 roads in 2008–2015 (table 4 ­ .3). Panama is an Total Private Public exception, given that the largest share of its infrastructure investment went to water Source: Fay et ­al. 2017 based on the INFRATALAM database (­w ww.infralatam.info). T rans p o rt I n f rastr u c t u re an d A gg l o m erati o n in Cities   125 FIGURE ­4.6  Change in Transport Investment as a Share of GDP, 2008–15 2.5 Change in transport investement (%) 2.0 1.5 1.0 0.5 0 –0.5 –1.0 go ala as a e ico ca a ia a ile a na lic r ru ay y il do gu m bi in liz ua liv az ur Ri ba Ch Pe ub ya gu em ex nt m na Be lva ra ug Bo Br nd sta Gu To lo ep ge ra M Pa ca at Ur Sa Ho Co Pa Co nR Ar d Ni Gu an El ica ad in id m in Do Tr Mexico and Central America Andean States Southern Cone Caribbean Source: Calculations based on the INFRALATAM database, ­w ww.infralatam.info. Note: Total transport investment includes private and public ­spending. The baseline and end-line years are 2008 and 2015 for most countries, except for Chile (end-line year is 2014), Dominican Republic (baseline year is 2009), and Uruguay (end-line year is ­2012). Growth is calculated as percentage change over the period ­considered. TABLE ­4.3  Transport Infrastructure Investments, by Sector, 2008–15 % of GDP Subregion Country Roads Railroads Air Waterways Total Andean States Bolivia ­3.41 ­0.10 ­0.10 ­0.02 ­3.84 Columbia ­2.40 ­0.01 ­0.06 ­0.09 ­2.56 Peru ­2.03 ­0.60 ­0.11 ­0.13 ­2.88 Southern Cone Argentina ­0.07 — ­0.00 ­0.00 ­0.75 Brazil ­0.70 ­0.20 ­0.07 ­0.08 ­1.06 Chile ­1.26 ­0.28 ­0.05 ­0.09 ­1.69 Paraguay ­1.73 — ­0.05 ­0.03 ­1.80 Uruguay ­0.30 ­0.01 ­0.00 ­0.06 ­0.45 Central America Belize ­0.94 — — ­0.09 ­0.97 and the Caribbean Costa Rica ­0.99 ­0.01 ­0.07 ­0.18 ­1.25 El Salvador ­0.87 — ­0.03 ­0.04 ­0.93 Guatemala ­1.15 ­0.00 ­0.02 ­0.07 ­1.23 Guyana ­0.90 — — — ­0.90 Honduras ­1.67 — ­0.06 ­0.48 ­2.21 Mexico ­0.68 ­0.05 ­0.01 ­0.04 ­0.77 Nicaragua ­1.95 — ­0.03 ­0.03 ­1.99 Panama ­1.43 — ­0.07 ­2.19 ­3.68 Dominican Republic ­0.37 — — ­0.12 ­1.32 Trinidad and Tobago ­0.59 — — — ­0.59 Source: Calculations based on the INFRALATAM database (www.infralatam.info). ­ DP. — = missing data. Note: These figures are calculated as the average of total spending (private and public) as a share of G 126   RAISING THE BAR FIGURE ­4.7  Evolution of Paved Road Density, Selected Regions, roads between LAC and EAP and MENA 1961–2000 widened until ­2000. The LAC region’s increase in paved road density in 1960–2000 was only 10 Paved road density (km per 100 km2) marginally higher than ­SSA’s. 9 8 7 Cities are poorly connected.  Poor road 6 networks can have large effects on local 5 4 development and productivity, as they often 3 obstruct access to markets and economic 2 opportunities. LAC cities, for example, ­ 1 appear to have limited access to improved 0 road infrastructure, as reflected in the average 1961 1965 1970 1975 1980 1985 1990 1995 2000 cumulative length of primary roads within a Sub-Saharan Africa Middle East and North Africa 100 km radius of cities with at least one Latin America and East Asia and Pacific million ­i nhabitants.14 This indicator of the Caribbean connectivity to surrounding markets (figure ­ 4.8, blue circles), is slightly less Source: Calculations based on Mitchell 2007. than 500 km in LAC cities, and varies in other regions from 530 km (South Asia) to 1,439 km transport, the bulk of which served to main- (North A ­ merica). In short, except for SSA tain and expand the Panama C ­ anal. The where the length of primary roads around Andean states, including Bolivia, Columbia, large cities is about 213 km, LAC cities fall and Peru, emerged as the largest investors in short on their surrounding road ­infrastructure. roads with an average subregional investment The ranking of LAC and SSA based on of about ­2.61 percent of ­ GDP. this indicator does not change if city popula- The breakdown of spending by sector tion is considered, as shown by the kilome- also showcases the increasing recognition of ters of roads per 1,000 people (see the orange the economic benefits of integrated networks dots in figure ­4.8). Within LAC, Peru has the (Leal and Pérez, 2012), combining national least road infrastructure around large cities and transcontinental road systems with rail- (200 km), Puerto Rico, Honduras, and Chile roads and ­ ports. In Brazil, Chile, Honduras, the most (768 km, 734 km, and 730 ­ k m). Panama, and Peru, higher investments in Surprisingly (figure ­4.9), Argentina, Mexico, road coexist with nonroad spending in rail, and Brazil, among the economic leaders of air, and water ­ t ransport. Prominent exam- the region, each has a surrounding road indi- ples of road-integration achievements include cator below the regional average of 499 km the Trans-Amazonian Highway in Brazil (428 km, 397 km, and 348 ­ km).15 and the Pan-American Highway corridor connecting Latin America to North ­ America. Transport costs are too high in the LAC region.  Stylized facts abound on the prevalence of exorbitant transport costs in the Current Challenges LAC region, reflecting physical and nonphysical Road quality has improved, but not trade. In most factors and acting as barriers to ­ enough.  Despite its prominence in recent LAC countries, transport costs more than infrastructure policies, road development in export tariffs (Mesquita Moreira, Volpe the LAC region has not translated into better Martincus, and Blyde 2 ­ 008). This can be seen networks. In the 1960s, for example, LAC ­ ­ .10, where data on intraregional in figure 4 had similar paved road densities to regions exports among LAC countries (orange dots) like East Asia and Pacific (EAP), MENA, and and exports to the United States (blue dots) ­ .7).13 However, the gap in paved SSA (figure 4 suggest that freight expenditures exceed tariff T rans p o rt I n f rastr u c t u re an d A gg l o m erati o n in Cities   127 FIGURE ­4.8  Average and Per Capita Road Length in a 100-Kilometer Radius around Cities with at Least 1 Million Inhabitants 1,600 0.50 1,400 0.45 0.40 Average length of roads (km) Roads per 1,000 people (km) 1,200 0.35 1,000 0.30 800 0.25 600 0.20 0.15 400 0.10 200 0.05 0 0 Sub-Saharan Latin America South Asia East Asia and Middle East Europe and North Africa and the Paci c and North Central Asia America Caribbean Africa Average road length (left axis) Per capita road length (right axis) Source: Calculations based on DeLorme 2015 and Blankespoor, Khan, and Selod 2017. Note: Road length within 100 km radius of a city’s center does not measure density because effective land area may vary, for example, for coastal cities in the presence of water s­ urface. FIGURE ­4.9  Average Road Length in a 100-Kilometer Radius around Cities with at Least 1 Million Inhabitants Puerto Rico Honduras Chile Guatemala Costa Rica Dominican Republic Ecuador Paraguay LAC 19 Haiti Cuba Colombia Argentina Venezuela, RB El Salvador Mexico Brazil Bolivia Panama Peru 0 100 200 300 400 500 600 700 800 900 Average road length (km) Source: Calculations based on DeLorme 2015 and Blankespoor, Khan, and Selod 2017. Note: The y-axis shows the 19 individual countries and their collective label “LAC 19.” The x-axis shows the average road length in kilometers within 100 kilometers radius of a city’s center. Road length within a 100-kilometer radius of a city’s center does not measure density because effective land area may vary, for example, in the presence of water s ­ urface. 128   RAISING THE BAR FIGURE ­4.10  Ad Valorem Freight and Real Tariffs for Intraregional costs in most LAC countries, except for Exports and Exports to the United States, 2005 Uruguay (exports to the United States) and Ecuador (intraregional e ­xports).16 LAC also 25 tends to exhibit higher transport costs than developed countries because of poor infrastructure, weak competition in the trucking Paraguay industry and dysfunctional c ­ustoms. Freight 20 expenditures are about 7­ .2 percent of regional import value in the LAC region, about twice the 3.7 percent in the United States (Mesquita ­ Ratio of freight expenditures to trade Moreira, Volpe Martincus, and Blyde ­ 2008). 15 The costs of urban congestion.  In LAC Bolivia cities, transport is associated with high Chile congestion. Although regional rates of ­ motorization in LAC (about 100–300 vehicles 10 Ecuador Colombia per 1,000 people) fall short of existing rates Argentina in developed nations (about 500–700 in Brazil Argentina Uruguay Canada, Europe, and the United States), they Venezuela Peru Paraguay are nonetheless linked to congestion, accidents, Chile 5 Brazil and pollution that are among the highest in Bolivia Ecuador Uruguay México the world (Barbero ­2012). On congestion, the Peru Venezuela 2016 TomTom traffic index shows that Colombia México Mexico City, with an extra travel time of 0 66 percent against the noncongested situation, 0 5 10 15 20 25 is the most congested city in the ­world. Eight Ratio of tariff revenue to trade other LAC cities are among the world’s top United States Latin America and the Caribbean 45-degree line 100 congested places (table ­4.4). High congestion and related social and Source: Mesquita Moreira, Volpe Martincus, and Blyde 2008. environmental challenges can be extremely TABLE ­4.4  LAC Cities Are among the Top 100 Congested Places in the World World rank City Country Congestion Morning peak Evening peak 1 Mexico City Mexico 66 96 101 8 Rio de Janeiro Brazil 47 63 81 17 Santiago de Chile Chile 43 73 88 19 Buenos Aires Argentina 42 64 68 28 Salvador Brazil 40 63 70 43 Recife Brazil 37 60 65 47 Fortaleza Brazil 35 56 57 71 São Paulo Brazil 30 42 53 99 Belo Horizonte Brazil 27 42 59 Source: TomTom traffic index (­w ww.tomtom.com). ­ ongestion. Note: The TomTom index for congestion measures the percentage of extra travel time (relative to a free-flow situation) as a result of traffic c The TomTom data cover 48 countries and 390 cities ­worldwide. T rans p o rt I n f rastr u c t u re an d A gg l o m erati o n in Cities   129 expensive for the economy of the LAC The geographic distribution of economic region. Bull and Thomson (2002), for exam- activities is highly uneven, with a large con- ple, estimate that the costs of negative exter- centration in the Mexico City Metropolitan nalities linked to traffic congestion in large Area, which in 2010 contributed a quarter of cities are nearly ­ 3 .5 percent of the LAC national gross value added, although it cov- region’s aggregate G ­ DP. Policy efforts to ered less than 0 ­ .3 percent of national reduce such negative externalities have often territory. This spatial concentration, espe- ­ favored investments in large public trans- cially in the center of the country and the por t systems and hard-to - enforce periphery of Mexico City, can be seen in for- regulations—and the latter, such as restric- mal establishments (map 4.2, panel a). tions on vehicle use, have seen mixed out- Manufacturing firms are even more concen- comes in many L AC c ­ ities. Gasoline trated (panel b), consistent with theories of emissions, for example, were reduced by agglomeration (Redding and Turner 2 ­ 014). 9–11 percent during peak hours and by Most of the clustering is in central Mexico 6 percent during the day after vehicle use where the transport network is denser (facili- restrictions in Quito, Ecuador (Carrillo, tating the shipment of goods) and where Malik, and Yoo 2016). Although similar major agglomeration centers are located regulations induced a shift to public trans- (potentially supplying ­ labor).19 As expected, port in Santiago de Chile (De Grange and firms in other sectors (mainly in commerce Troncoso 2011), they failed to reduce emis- and services) are less concentrated than firms sions or vehicle use in Mexico (Eskeland and in manufacturing (maps not ­ shown). Feyzioglu 1997; Davis ­ 2 008). Combining A small number of studies have docu- these regulation-based policies with price mented rising geographic concentration of incentives may be more efficient at reducing industries in Mexico (Unger ­ 2003). A look at congestion and related externalities (Berg the Ellison and Glaeser index at the 6-digit et ­al. ­2017). industry classification level (Ellison and Glaeser 1997; see box 4.2 for details) reveals patterns akin to these ­ findings. First, indus- Roads and Agglomeration tries in Mexico have become more geograph- Economies: Evidence from ically concentrated over the past decade, as Mexico reflected in the increase of the mean value of Bridging the transport infrastructure gap the index (across all industries), from ­ 0.381 in Latin America could generate large eco- in 2004 to ­ 0.430 in 2014 (a significant ­ 10.3 nomic ­gains. To illustrate them, this section percent ­ increase). Second, the mean value of presents a case study on M­ exico. As already the Ellison and Glaeser index for manufac- discussed in box 4 ­ .1, Mexico has invested turing industries in Mexico is, on average, heavily in roads over recent decades, and 12–15 percent higher than the mean value of may therefore offer a learning opportunity the same index calculated for all industries, for similar investments in other L AC confirming that manufacturing industries are ­countries.17 on average more concentrated than other ­ i ndustries. Finally, comparison with com- monly agreed-on thresholds in the literature Geographic Patterns of Economic (see box ­ 4.2) suggests that about 98–100 per- Activity cent of industries (for overall industries and Industries are increasingly concentrated. for manufacturing only) are concentrated Mexico’s 2014 Economic Census indicates (with an Ellison and Glaeser index greater that most formal employment is in services than ­ 0 .05). The fraction of concentrated (44.2 percent), commerce ­ ­ (29.6 percent), manufacturing industries (98 percent) in manufacturing ( ­23.5 percent), agriculture Mexico is greater than the one reported for ­ (0.9 percent), and mining (­0.8 ­p ercent).18 Canada (75 percent) in Behrens and Bougna 130   RAISING THE BAR MAP ­4.2  Spatial Distributions of Formal Establishments and Manufacturing Firms in Mexico, Overlaid on the Road Network, 2014 a. Spatial distribution of formal establishments b. Spatial distribution of manufacturing firms Pavement Gravel or earth road Two lanes or divided Multilane divided Source: DeLorme, AAA, DENUE, and INEGI. Reproduced from Blankespoor, Bougna et al. 2017. ­ 014. Firm locations (dots) are overlaid on 2016 r Note: Universe of all formal firms (panel a) and of manufacturing firms only (panel b) for 2 ­ oads. AAA = American Automobile Association; DENUE = Directorio Estadístico Nacional de Unidades Económicas; INEGI = Instituto Nacional de Estadísticas y Geografía. BOX ­4.2  Measuring Industrial Concentration: The Ellison and Glaeser Index The Ellison and Glaeser index defines concentration where Gi = ∑r (Sri −xr)2 is the spatial Gini coefficient of as agglomeration above what would be observed if industry i, S ri is the share of employment of locality plants simply chose locations ­ randomly. This mea- r in industry i, xr is the share of total employment sure provides an unbiased estimate of agglomerative in each locality r, Hi = ∑ j Z 2 Herfindahl index ji is the ­ source. It can be inter- forces independently of their ­ of the plant size distribution of industry i, and Zji preted as the probability that a firm choosing its ­ represents the employment share of a particular firm location follows the prior firm rather than locating i. j in industry ­ randomly, and is given by the following: Following Ellison and Glaeser (1997), an indus- try is said to be strongly concentrated if the index γi = ( 2 Gi − 1 − ∑ r xr Hi ) is greater than 0 ­ .05, weakly concentrated if the index is 0­ –0.05, and not concentrated if the index (1 − ∑ x ) (1 − H ) r 2 r i is ­negative. Source: Ellison and Glaeser 1997; Blankespoor, Bougna et al. 2017. (2015) and about the same as the one (Pérez and Palacio 2009; Kim and Zangerling reported for other developed countries ­ 2016). To contribute to this line of inquiry, the (Ellison and Glaeser 1997; Duranton and present study relies on the Krugman Overman ­2008). Specialization index at city level, which provides a tractable way to quantify Cities are increasingly specialized.  The specialization and potential changes in other important stylized fact, along with firm specialization patterns over time (box ­ 4.3).20 concentration, is the increasing specialization The computed index indicates a fall in the of Mexican ­ localities. Using different spatial proportion of weakly specialized localities scales, some studies indicate an important (with an index lower than ­ 0.35) in Mexico increase in local specialization in recent years from ­90.7 percent in 2004 to ­ 84.7 percent T rans p o rt I n f rastr u c t u re an d A gg l o m erati o n in Cities   131 BOX ­4.3  Measuring Municipality Specialization: The Krugman Specialization ­Index The Krugman Specialization index is widely u ­ sed. of industry i in the total output or employment It measures deviation of industry shares by comput- across all localities (in ­ Mexico). ing the share of employment that would have to be The index can take values between zero and t ­ wo. relocated to achieve an industry structure equivalent If the relative specialization measure is zero, the eco- g roup. It is to the average structure of the reference ­ nomic structure of a locality is identical to the eco- given by the following formula: nomic structure of the overall ­ economy. The higher the index, the more the economic structure of the M locality deviates from the overall economy (refer- KSIi = ∑S i =1 i m − Si ence group) and the more that locality is s ­ pecialized. Localities for which the specialization index is where S mi is the output or employment share of higher than ­ 0.75 are considered highly specialized, industry i in locality m, and S i is the average share whereas localities for which the specialization index is below ­0.35 are considered weakly ­ specialized. Source: Blankespoor, Bougna et al. 2017. Note: We calculate specialization indexes for a reconstructed sample of localities by merging municipalities that belong to the same metropolitan area, leading to a universe of 316 reconstructed metropolitan areas and 1,832 standalone ­municipalities. in ­2 014. As for highly specialized locali- measured by a reduction in the time and cost ties, their percentage increased from 6 to travel. For example, Blankespoor, Bougna of ­ ­ 9.2 ­percent over the p ­ eriod. Together, these et al. (2017) report that average travel time to findings provide evidence of increased spe- the nearest port of entry to the United States cialization in Mexican l ­ocalities. Map ­4.3 (among 44 entry ports) from any of the 2,094 shows the spatial distribution of the Krugman Mexican localities decreased by more than Specialization index for all localities in an hour and half between 1986 and 2 ­ 014. Mexico in 2014, revealing clusters of special- Similarly, travel time to the nearest port ized localities near one another (the red dot (among six major Mexican ports) decreased showing high ­ specialization). There is some by more than 40 minutes over the same indication from zooming on the map that ­period. 21 These are nonnegligible decreases specialized localities are often near larger in the average minimum travel time to the roads (two-lane or multi-lane roads), consis- United States border (about 8 percent) and to tent with the empirical findings presented ports (about 6 ­percent). 4.6. later in table ­ Improvements in access to national mar- kets can be measured using a market access i ndicator. By design, for a particular area, ­ Roads and Market Access this indicator reflects a discounted sum of As roads link cities and cities to ports, access population or income of the surrounding to international and domestic markets is a areas, where the discount factor is propor- function of local road ­ i nfrastructure. The tional to travel time (box 4 ­ .4). 22 The ratio- growth of road networks (see box 4 ­ .1), and nale behind this indicator is to numerically the current state of the roads network gauge the size of surrounding markets for depicted on the map of firm concentration locally produced goods, accounting for the (see map ­4.2) show recent road investments ease with which such goods can be trans- in ­Mexico. ported to these adjacent domestic m ­ arkets. Subsequent improvements in access to A higher value indicates greater market international markets can be directly ­ access. From 1994 to 2014, the indicator 132   RAISING THE BAR MAP ­4.3  Output Locality Specialization, Overlaid on the Road Network, 2014 N W E S KSI, output (2014) KSI ≤ 0.35 0.35 < KSI ≤ 0.75 KSI > 0.75 Road type (2016) Multilane divided Two lanes or divided Pavement 0 250 500 1,000 Kilometers Gravel or dirt road Source: DENUE ­(INEGI). Note: KSI is the Krugman Specialization index described in box ­4.4. DENUE = Directorio Estadístico Nacional de Unidades Económicas; INEGI =Instituto Nacional de Estadísticas y Geografía. BOX ­4.4  Market Access Market access is defined in each locality as follows: a measure of trade ­ elasticity. Travel times t ijt are cal- culated on the countrywide road network assuming MAit ∑P τ i≠j −θ jt ijt that speed is a function of road type (Blankespoor, Bougna et al. 2017). For the trade elasticity param- eter, lacking a specific study for Mexico, we use the where Pjt is the population of locality j at time t same value suggested by Donaldson (forthcoming) (which proxies for the size of the local market in j), for India ­(q = 3.8). The same market access indica- tijt is the time required to travel between locality i and tor is used in different empirical works (Jedwab and j given the state of the road network at time t, and q  is ­Storeygard ­2017a, 2017b). Source: Blankespoor, Bougna et al. 2017. ­ alculated. Note: The measure of market access excludes the population of the locality for which it is being c T rans p o rt I n f rastr u c t u re an d A gg l o m erati o n in Cities   133 MAP ­4.4  Changes in Market Access in Mexico, 1986–2014 a. 1986 b. 2014 Roads 2016 Multilane divided Roads 1985 Two lanes or divided Two lanes or divided Pavement Pavement Gravel or Earth Road Gravel or Earth Road Market Access 1986 Market Access 2014 Lowest quartila - 1986 Lowest quartila - 1988 Highest quartila - 1986 Highest quartila - 1988 Source: DeLorme, American Automobile Association, and Economic Censuses ­(Instituto Nacional de Estadísticas y Geografía). ­ ccess. Note: On both panels, the colors describe the quartiles of the 1986 distribution of the index, with darker shades indicating higher degrees of market a more than doubled among the 2 ,094 period) interacted with the measure of mar- Mexican localities analyzed in Blankespoor, ket access, as well as time and location fixed Bougna et al. (2017). 23 The change between ­effects. 1986 and 2014 is shown in map 4 ­ .4. Despite a rich set of controls and fixed Overall, as suggested by the spatial distri- effects, two factors undermine causal infer- bution of the index, most improvements ence: the bias arising from nonrandom place- in market access occurred predominantly in ment of roads and the recursion problem Mexico. the center of ­ inherent to using a market access indicator when the explained variable is a function of population (Baum-Snow et ­ 2 017). We a l. ­ The Economic Impacts of Road add ress t hese issues by resor ti ng to Improvement ­instrumentation.24 This section investigates the effects of improved market access from road invest- Employment.  Table ­ 4.5 summarizes the ment on employment, specialization, and ordinary least square (OLS) and instrumental local productivity (as measured by nighttime variable (IV) results for the specification, lights) in Mexican ­ cities. A balanced panel with employment the dependent v ­ariable. data set of 2,094 localities is analyzed over The effect of market access on employment is different periods, depending on data positive and statistically ­ s ignificant. ­ constraints. The specification with employ- Specifically, the IV result suggests that a ment as the explained variable covers six 10 percent increase in market access results periods of five-year intervals from 1986 to in a ­1.6 percent increase in ­ employment. ­ 2014. The specification with specialization as Other interesting findings include the the explained variable covers 2004, 2009, interaction between market access and and ­2014. The specification with nighttime relevant indicators of population and human lights comprises the years 1996, 2000, and capital. For example, more urbanized areas ­ 2010. All specifications include time-varying ­ (the metropolitan center and populated locality characteristics (education, popula- localities) and areas with less than the average tion, oil-reserves, and the pre–/post–North level of education are likely to benefit more American Free Trade Agreement [NAFTA] from increased market ­ access. The latter may 134   RAISING THE BAR TABLE ­4.5  The Effects of Market Access on Employment Total employment Variables OLS IV (Road count) Market access ­0.149*** ­0.163*** ­(0.0111) ­(0.0202) Market Access × Population Dummy ­0.0717*** ­0.0629*** ­(0.0117) ­(0.0154) Market Access × Education Dummy ­− 0.0468*** ­−0.0474*** ­(0.00334) ­(0.00329) Market Access × NAFTA Dummy ­− 0.114*** ­−0.115*** ­(0.00233) ­(0.00213) Market Access × Capital City ­− 0.0225 ­−0.0228 ­(0.0226) ­(0.0250) Market Access × Oil Dummy ­0.839** ­0.832** ­(0.385) ­(0.399) Education dummy ­0.155*** ­0.155*** ­(0.0581) ­(0.0522) Population dummy ­0.851*** ­0.805*** No. of observations 11,379 11,251 Adjusted R2 ­ .423 0 ­0.424 Source: Blankespoor, Bougna et al. 2017. Note: Standard errors are clustered at the locality level to adjust for ­heteroscedasticity. The road count is the number of roads intersecting a circle with a 10 km ­radius. Education and population controls are measured at the initial date ( ­ 1986). Constant is not ­shown. IV = instrumental variable; OLS = ordinary least square. ***p < 0.01. **p < 0.05. *p < 0.1. suggest that, with roads, cities may attract Local productivity.  As shown in table ­ 4.7, low-skilled ­labor. The estimates also provide the elasticity of local productivity, as measured evidence that the positive effect of improved by nighttime lights, with respect to market domestic market access on employment is significant.27 access is positive and statistically ­ partially attenuated after NAFTA came Specifically, the result from the IV regression force. Finally, cities in oil-producing into ­ indicates that a 10 percent increase in market regions seem to benefit more from improved access increases nighttime luminosity by access to domestic markets than cities in ­ 0.9 percent, controlling for the population of other ­regions. the locality, suggesting that market access is an productivity. This important driver of city-level ­ Specialization.  Table ­ 4.6 presents the finding is consistent with previous studies results of the regression of city-level documenting a sizable impact of transport specialization on market ­ access. It suggests a infrastructure on productivity in Mexico positive and statistically significant response (Becerril -Torres, Álvarez-Ayuso, and del of specialization to improved market ­ access.25 Moral-Barrera 2010; Brock and German-Soto A 10 percent increase in market access, the 2013; Duran-Fernandez and Santos 2014a, findings suggest, translates into a 7 percent 2014b). Relevant channels through which these ­ increase in output specialization, as reflected effects materialize include spatial concentration in the IV estimation ­result.26 and specialization (Dávila 2008; Monge ­ 2012). T rans p o rt I n f rastr u c t u re an d A gg l o m erati o n in Cities   135 TABLE ­4.6  The Effects of Market Access on Local reduce congestion and pollution have had Specialization only mixed success in the LAC region, it will Krugman Specialization index be important to come up with alternative (output) cost-efficient ways to reduce these externali- Variable OLS IV (doughnut) instruments. ties, possibly using price ­ Market access ­0.455*** ­0.704*** ­(0.176) ­(0.231) Notes Observations 4,303 3,628 1. Localization economies are the positive exter- Adjusted R2 ­0.0234 ­0.0233 nalities associated with the clustering within a Source: Blankespoor, Bougna et al. 2017. city of firms from the same industry (Marshall Note: The reported results are for the instrumentation with the “doughnut,” ­ 1890). Urbanization economies are the posi- which is calculated as the market access when excluding all localities tive externalities associated with the geo- within a 25 km circle (for details, see Blankespoor, Bougna et al. 2017). Education and population variables are measured at the initial date ­(1986). graphic concentration of different industries Constant not ­shown. IV = instrumental variable; OLS = ordinary least square. within a given city (Jacobs ­ 1969). See Graham Standard errors are clustered at the locality level to adjust for (2005) and Redding and Turner (2014) for a ­heteroscedasticity. ***p < 0.01. **p < 0.05. *p < 0.1. detailed discussion of transport and agglomer- ation ­economies. TABLE ­4.7  The Effects of Market Access on 2. The pricing of externalities, however, may be Nighttime Lights resisted by those facing higher transport Nighttime lights ­costs. 3. See Graham (2005) for a review of the rele- Variable OLS IV (road count) vant ­papers. Market access ­0.044** ­0.086** 4. The period 1870–1914 also coincided with ­(0.022) ­(0.044) the world’s first wave of globalization in trade and ­finance. No. of observations 5,144 4,965 5. From the work of Fogel (1962), the concept of Adjusted R2 ­0.084 ­0.081 “social savings” refers to a growth accounting Source: Blankespoor, Bougna et al. 2017. approach to assess the historical implications of Note: The road count is the number of roads intersecting a circle with a new technology on economic g ­ rowth. In prac- a 10 km ­radius. Education and population variables are measured at tice, the rate of social savings is derived from the initial date ­(1986). Constant not ­shown. IV = instrumental variable; OLS = ordinary least square. estimating the cost savings induced by the new Standard errors are clustered as the locality level to adjust for technology (in this case, railroads) relative to ­heteroscedasticity. the next best a ­lternative. The social savings ***p < 0.01. **p < 0.05. *p < 0.1. approach has been extensively used to analyze the impact of innovations in transport, espe- Conclusions cially ­railroads. A few exceptions include Von Tunzelmann (1978), who focuses on the effects This chapter confirms that LAC suffers from of steam power in the United Kingdom, and a significant transport infrastructure ­gap. Bogart (2009) who focuses on the impacts of Bridging it may have local effects on the turnpike trusts set up to levy road tolls in growth of jobs, the specialization of cities, ­England. and economic development, as shown by the 6. Export density is defined as the average export Mexico case ­ study. Complementary policies value per square kilometer of land area, and could also have an impact, such as policies railroad density measures the average number that encourage competition, improve customs of km of railroads per 1,000 km2 of land a ­ rea. 7. The Belle Époque (from the end of the Franco– procedures, and improve efficiency via price- Prussian War in 1871 to the eruption of World based ­regulations. War I in 1914) coincided with a significant The extremely high congestion is likely to flow of foreign capital into Latin America and exert a toll on city productivity, because of the economic modernization of most coun- the time and money lost in intracity trans- region. tries in the ­ port, and on environmental and health costs 8. These figures come from Mitchell (2007) and affecting w­ orkers. Because regulations to are based on 18 Latin American ­ countries. 136   RAISING THE BAR 9. The urban share in figure ­ 4.3 is from World them here (see Blankespoor, Mesplé-Somps Urbanization Prospects data and may overes- et al. 2017 for a detailed presentation of local timate the actual rate (see chapters 1 and ­ 2). specialization in ­Mexico). 10. To our knowledge, no similar study exists for 21. These figures are theoretical measures based Latin ­ America. The georeferenced and historic on network extent and road type, but they do data that would be needed to explore how not account for ­ congestion. roads influenced urbanization in the LAC 22. See Blankespoor, Mesplé-Somps et al. (2017) region has not yet been ­compiled. for the use of an alternative measure, market 11. Data on airborne and maritime freight trans- potential, which they define as a discounted port were not available for this ­ analysis. sum of surrounding incomes, where the dis- 12. External debt, net of foreign exchange reserves, count factor is a function of transport ­ costs. decreased from ­28.6 percent of GDP in 1998– All the results in this chapter are robust to the 2002 to ­ 5.7 percent in 2008 (Ocampo ­ 2015). use of either the market access or market 13. Paved road density is the length of roads in potential ­indicator. km per 100 km2 of land a ­ rea. 23. Contrary to the measures of access to interna- 14. Cities refer here to urban a ­gglomerations. tional markets, measures of access to domestic The data come from World Urbanization markets also change with concurrent changes Prospects, georeferenced by Blankespoor, in the population d ­ istribution. The doubling Khan, and Selod ­ (2017). of the market access indicator thus reflects not 15. This is a local indicator of infrastructure avail- only road improvements but also population ability that does not measure overall connec- ­increases. tion to other cities or positions of the transport 24. To address the recursion problem, market ­ network. It also fails to account for road access is instrumented with the number of ­quality or ­congestion. roads intersecting a circle of 10 km radius 16. Freight expenditures do not include insurance around each ­ locality. Sources of changes in and are defined as the costs of transporting accessibility are then only due to variation in goods to the international port of the country ­ roads. To address the endogeneity of road of origin and of delivering them to the port of placement in the specification regressing the country of ­destination. specialization, a “doughnut” market access, 17. This section draws largely on Blankespoor, excluding all localities within a 25 km radius, Mesplé-Somps et al. (2017), commissioned as instrumented. See Baum-Snow et ­ is ­ al. (2017) a background p ­ aper. It deals only with the and Blankespoor, Mesplé-Somps et al. (2017) agglomeration impacts of improved transport for a detailed discussion about these issues accessibility, and does not focus on measuring and the strategies to overcome ­ them. other ­impacts. 25. For this result on specialization and the next 18. Figures for informal employment are on local productivity, only the estimated elas- ­ unavailable. Looking at formal employment ticities of the variables of interest with data only, as seen in previous chapters, the respect to market access are ­ presented. See notable trend over the past two decades has Blankespoor, Mesplé-Somps et al. (2017) for been the decrease in the share of formal man- other estimated ­ coefficients. ufacturing employment, showing that Mexico 26. For employment specialization, Blankespoor, is following a trend of tertiarization similar to Mesplé-Somps et al. (2017) report a smaller that experienced by developed economies in elasticity of about 3 ­ percent. previous ­decades. 27. Nighttime lights are defined and measured as 19. Unger and Chico (2004) note that the cluster- in chapter 6; see box 6.1. ing of firms in Mexico often occurs in places where the required labor skills can be found, highlighting labor market pooling References (the “matching” argument put forward in Aschauer, ­ D. ­ 1 993. “Genuine Economic A. ­ economic geography) as a mechanism of Returns to Infrastructure ­ I nvestment.” Policy ­agglomeration. Studies Journal 21 (2): ­ 380–90. 20. Patterns of specialization measured with the A ., ­ Banerjee, ­ N. ­ E . Duflo, and ­ 2012. “On Qian. ­ Krugman Specialization index for employ- the Road: Transportation Infrastructure and ment are very similar and we do not show Economic Growth in ­ China.” Working Paper T rans p o rt I n f rastr u c t u re an d A gg l o m erati o n in Cities   137 N o. 17897, National Bureau of Economic ­ Blankespoor, ­ A . Khan, and ­ B ., ­ H. ­ 2017. S elod. ­ Research, Cambridge, ­ M A. A Consolidated Dataset of Global Urban Baum-Snow, ­ N. ­ 2 007. “Did Highways Cause Populations: ­1 969–2015. Technical ­ n ote. Suburbanization?” The Quarterly Journal of World Bank, Washington, ­ DC. Economics 122 (2): ­ 775–805. Blankespoor, ­ B., ­S. Mesplé-Somps, ­ H. Selod, and Baum-Snow, ­ N ., L­ . Brandt, J ­. ­ V. Henderson, G. ­ ­ Spielvogel. ­ 2017. “Roads and Structural ­ M . ­ A . Turner, and ­ Q. ­ Zhang. ­ 2012. “Roads, Transformation in Mali.” Unpublished Railroads and Decentralization of Chinese ­manuscript. ­Cities.” Review of Economics and Statistics Blankespoor, ­ B., ­T. Bougna, ­ R. Garduno-Rivera, 99 (3): ­ 435–48. and ­ Selod. ­ H. ­ 2017. “Roads and the Geography Baum-Snow, ­ N., ­ V. Henderson, ­ J. ­ M. ­A. Turner, of Economic Activities in M ­ exico.” Policy ­ Q . Zhang, and ­ L. ­ B randt. ­ 2 017. “Does Research Working Paper WPS 8226, World Investment in National Highways Help or Bank, Washington, ­ DC. Hurt Hinterland City Growth?” Unpublished Bogart, ­ D. ­2009. “Turnpike Trusts and Property ­manuscript. Income: New Evidence on the Effects of B a rb ero, ­J . ­2 012 . Inf ra s t r u c t u re i n t he Transport Improvements and Legislation in Development of Latin ­ A merica . Caracas: Eighteenth-Century ­England.” The Economic Development Bank of Latin ­ A merica. History Review 62 (1): 1 ­ 28–52. Becerril-Torres, ­ O., ­I . Álvarez-Ayuso, and ­ L . del Brock, ­ G., and ­ B. ­German-Soto. ­ 2013. “Regional Moral-Barrera. 2 ­ 010. “Do Infrastructures Industrial Growth in Mexico: Do Human Influence the Convergence of Efficiency in Capital and Infrastructure Matter?” Journal of México?” Journal of Policy Modeling 32 (1): Policy Modeling 35 (2): ­ 228–42. ­120–37. Bull, ­ A ., and ­ I. ­T homson. ­ 2002. “Urban Traffic Behrens, ­ K., and ­T. ­ Bougna. ­ 2015. “An Anatomy Congestion: Its Economic and Social Causes of the G eog raph ical Concent ration of and ­Consequences.” Cepal Review 76: ­ 105–116. C a nad ia n M a nu fac t u r i ng ­ I ndu st r ie s.” Bu l mer-T homas , ­ V., ­ J . C oat swor t h , a nd Regional Science and Urban Economics ­R.  Cortes-Conde, ­eds. ­2006. The Cambridge 51 (C): ­ 47–69. Economic History of Latin America, Volume Berg, ­ C. N ­ ., U ­ . Deichmann, Y ­ . Liu, and H­ .­Selod. ­I I. Cambridge, U.K.: Cambridge University ­ 2017. “Transport Policies and ­ Development.” ­Press. The Journal of Development Studies 53 (4): Calderón, ­ C ., and ­ L. ­Servén. ­2004a. The Effects ­465–80. of Infrastructure Development on Growth Bess, ­ M. ­ 2 014. “Routes of Conflict: Building and Income Distribution. Washington, DC: Roads and Shaping the Nation of Mexico, World ­Bank. ­1 941–1952 .” The Journal of Transport —— —. ­2004b. “Trends in Infrastructure in Latin History 35 (1): ­ 78–96. A merica, ­ 1 980 –2001.” Policy Research —— —. ­2016a. “‘Neither Motorists nor Pedestrians Wo r k i n g P a p e r 3 4 0 1 , Wo r l d B a n k , Obey the Rules: Transit Law, Public Safety, Washington, ­DC. and the Policing of Northern Mexico’s Roads, Cardenas, ­ 1987. “La Industrialización Mexicana E. ­ ­1 920s-1950s.” The Journal of Transport durante la Grande ­ Depresión.” Working Paper, History 37 (2): ­ 155–74. El Colegio de México, Mexico ­ City. ———. ­ 2 016b. “Revolutionary Paths: Motor Carrillo, ­ P. ­E ., ­A. ­S . Malik, and ­ Y. ­ Yoo. ­ 2016. Roads, Economic Development, and National “Driving Restrictions that Work? Quito’s Pico Sovereignty in 1920s and 1930s ­ M exico.” y Placa ­ P rogram.” Canadian Journal of Mexican Studies/Estudios Mexicanos 32 (1): Economics 49 (4): ­ 1536–68. ­56–82. Chandra, ­ A ., and ­ E. ­ T hompson. ­ 2000. “Does —— —. ­2017. Routes of Compromise: Building Public Infrastructure Affect Economic Activity? Roads and Shaping the Nation of Mexico, 1917– Evidence from the Rural Interstate Highway 1952. Lincoln, NE: University of Nebraska ­ Press. ­S ystem.” Region al Scie nc e an d Urban Bird, ­ J ., and ­ S. ­S traub. 2014. “The Brasilia Economics 30 (4): ­ 457–90. Experiment: Road Access and the Spatial Coatsworth, ­ H. ­ J. ­ 1979. “Indispensable Railroads Pattern of Long-Term Local Development in a Backward Economy: the Case of M ­ exico.” in ­ B razil.” Policy Research Working Paper The Journal of Economic History 39 (4): 6964, World Bank, Washington, ­ DC. ­939–60. 138   RAISING THE BAR Datta, S ­. ­2 012. “The Impact of Improved Ellison, ­ G ., and ­ E. ­Glaeser. ­ 1997. “Geographic Highways on Indian ­ F irms.” Journal of C oncent rat ion i n ­ U . S . M a nu fac t u ri ng Development Economics 99 (1): ­ 46–57. Industries: A Dartboard ­ Approach.” Journal Dávila, ­ A. ­2008. “Los Clústeres Industriales del of Political Economy 105 (5): ­ 889–927. noreste de Mèxico ­ (1993–2003). Perspectivas Eskeland, ­ G., and ­T. ­Feyzioglu. 1 ­ 997. “Rationing de desarrollo en el marco de una mayor inte- Can Backfire: The ‘Day without a Car’ in gración económica con Texas?” Región y Mexico ­City.” World Bank Economic Review Sociedad 20 (41): 5 ­ 7–88. 11 (3): 3­ 83–408. Dav is , L ­. ­ 2 0 08. “T he E f fec t of Dr iv i ng Faber, B ­.­ 2013. “Trade Integration, Market Size, Restrictions on Air Quality in Mexico C ­ ity.” and Industrialization: Evidence from China’s Journal of Political Economy 116 (11): National Trunk Highway ­ System.” Working ­38–81. Paper, University of California, ­ Berkeley. De Grange, ­ L ., and ­ R. ­Troncoso. ­ 2011. “Impacts Fay, ­ M ., ­ L ., ­ A . Andres, ­ C . Fox, ­ U . Narloch, of Vehicle Restrictions on Urban Transport S . Straub, and ­ ­ M. ­ Slawson. ­ 2017. Rethinking Flows: The Case of Santiago, ­ Chile.” Transport Infrastructure in Latin America and the Policy 18 (6): ­ 862–69. ­C aribbean. Washington, DC: World ­ Bank. DeLorme. ­ 2 015. “Digital Atlas of the Earth Fernald, ­ ­ . 1 J. G ­ 999. “Roads to Prosperity? Database.” Yarmouth, M ­ E. Assessing the Link between Public Capital and Deng, ­ T. ­ 2 013. “ I mp ac t s of Tra n sp or t ­P roductivity.” American Economic Review Infrastructure on Productivity and Economic 89 (3): ­ 619–38. Growth: Recent Advances and Research Fogel, R ­ .­1962. “A Quantitative Approach to the ­C hallenges.” Transport Reviews 33 (6): Study of Railroads in American Economic ­686–99. Growth: A Report of Some Preliminary Donaldson, ­ D. ­Forthcoming. “Railroads of the ­F indings.” Journal of Economic History Raj: Estimating the Impact of Transportation 22 (2): ­163–97. ­Infrastructure.” American Economic ­ Review. Foote, ­W. ­1997. Mexico’s Troubled Toll ­ Roads. Dulac, ­J . ­2 013. Infrastructure Requirements: ICWA Letters, February ­ 3 . Hanover, NH: Estimating Road and Railway Infrastructure Institute of Current World ­ Affairs. C a pa c i t y a n d C o s t s t o ­ 2 0 5 0 . Pa r i s : Garcia-Mila, ­ T. ­ T., ­ J. McGuire, and ­ R. ­ Porter. H. ­ International Energy ­ Agency. 1996. “The Effect of Public Capital in State-Level ­ Duran-Fernandez, R ­ ., and G ­ . S ­ antos. 2 ­ 014a. Production Functions ­ Reconsidered.” The Review “Road Infrastructure Spillovers on the of Economics and Statistics 78 (1): ­ 177–80. Manufacturing Sector in M ­ exico.” Research in Ghani, ­ E ., ­ A. ­G. Goswami, and ­ W. ­ R. ­ Kerr. ­2016. Transportation Economics 46 (C): ­ 17–29. “Highway to Success: The Impact of the —— —. ­ 2 014b. “Regional Convergence, Road Golden Quadrilateral Project for the Location Infrastructure, and Industrial Diversity in and Performance of Indian ­ M anufacturing.” ­M ex ico.” R e se a rc h i n Tra n s po r t atio n The Economic Journal 126 (591): ­ 317–57. Economics 46: 1 ­ 03–10. Gonzalez-Navarro, ­ M ., and ­ C . Quintana- D u ra nton , ­ G ., a nd ­ H. ­ O ver m a n . 20 0 8. Domeque. ­ 2016. “Paving Streets for the Poor: “Exploring the Detailed Location Patterns of Experimental Analysis of Infrastructure U.K. Manufacturing Industries Using Micro- ­ ­Effects.” Review of Economics and Statistics Geographic ­D ata.” Journal of Regional 98 (2): ­ 254–67. Science 48 (1): ­ 213–43. Graham, ­ D. ­ 2 005. “Transport Investment, J. ­ Duranton, G ­ ., and M ­ . A­ . T­ urner. 2 ­ 011. “The Agglomeration and Urban ­ P roductivity.” Fundamental Law of Road Congestion: Unpublished ­manuscript. Evidence from US Cities.” American Economic —— —. ­ 2007. “Agglomeration, Productivity and Review 101 (6): ­ 2616–52. Transport ­I nvestment.” Journal of Transport —— —. 2­ 012. “Urban Growth and Transportation.” Economics and Policy 41 (3): ­ 317–43. Review of Economic Studies 79 (4): ­ 1407–40. G ra m l ich , E ­. ­ M. ­ 1 99 4. “I n f ra s t r uc t u re Duranton, ­ G., ­ ­ . Morrow, and M P. M ­ .­ A. ­Turner. Investment: A Review ­ E ssay.” Journal of ­ 2014. “Roads and Trade: Evidence from the Economic Literature 32 (3): ­ 1176–96. ­U S.” Review of Economic Studies 81 (2): Holl, ­ A. ­ 2 004. “Transport Infrastructure, ­681–724. A g g lom e r at io n E c o nom i e s , a nd F i r m T rans p o rt I n f rastr u c t u re an d A gg l o m erati o n in Cities   139 Birth: Empirical Evidence from ­ Portugal.” ­C aribbean. Washington, DC: Inter-American Journal of Regional Science 44 (4): ­ 693–712. Development ­Bank. Holtz-Eakin, ­ D. 1994 “Public-Sector Capital and Michaels, ­ G. 2 ­ 008. “The Effect of Trade on the Productivity Puzzle.” Review of Economics t he D em a nd for Sk i l l — Ev idence f rom and Statistics 76 (1): ­ 12–21. the I nterstate H ighway ­ S y s t e m .” H su , ­ W. T, a nd ­ H. ­ Z h a ng. ­ 2 012 . “T he Review of Economics and Statistics 90 (4): Fundamental Law of Highway Congestion: ­683–701. Evidence from National Expressways in Mitchell, ­B . ­2 007. International Historical J apan.” Working paper, Department of ­ Statistics 1750–2005: ­ A mericas. New York: E conomics, The National University of Palgrave ­Macmillan. ­Singapore. Monge, ­ M. ­ 2012. “Analisis de la Cadena Productiva International Transport Federation. ­ 2 017. de Tequila: El caso de J ­alisco.” License Capacity to Grow: Transport Infrastructure Dissertation, Departamento de Economía, Needs for Future Trade ­ G rowth. Paris: Universidad Autónoma Metropolitana- Organisation for Economic Co-operation and Azcapotzalco, Mexico ­ City. ­Development. Ocampo, ­ A. ­ J. ­ 2015. “Uncertain ­ Times.” Finance Jacobs J. 1969. The Economy of Cities. New York: and Development 52 (3): ­ 6 –11. Random House. Pages-Serra, ­C . ­2010. The Age of Productivity: Jedwab, ­ E . Kerby, and A R ., ­ ­ . ­ M oradi. ­ 2 015. Transforming Economies from the Bottom “History, Path Dependence and Development: ­U p . Wash i ng ton , DC : I nter-A mer ic a n Evidence from Colonial Railroads, Settlers and Development ­Bank. Cities in ­ K enya.” The Economic Journal Pérez, ­ S ., and ­ M. ­ Palacio. ­2 009. “Desarrollo 127 (603): ­ 1467–94. Regional y Concentración Industrial: Impacto Jedwab, ­ R ., and ­ A. ­ M oradi. ­ 2 016. “T he en el Empleo ­ (1994–2004).” Observatorio de Per m a ne nt E f fe c t s of Tra n sp or t at ion la Economía Latinoamericana ­ 117. No. ­ Revolutions in Poor Countries: Evidence from Pérez-Cervantes, F., and A. Sandoval-Hernandez. ­A frica.” Review of Economics and Statistics 2017. “Short-Run Market Access and the 98 (2): ­ 268–84. Construction of B et ter Transpor tation J e dwab, R . , a nd A . S tore y g a rd . 2 017a . Infrastructure in Mexico.” Economía 18 (1): “ E c o n o m i c a n d Po l i t i c a l F a c t o r s i n 225–50. Infrastructure Investment: Evidence from Redd i ng , S ­. J ­ . a nd M­ . ­A. ­ Tu r ner. 2 ­ 014. Railroads and Roads in Africa 1960-2015.” “Transpor tation Costs and the Spatial Unpublished manuscript. O r g a n i z at ion of E c onom i c A ­ c t iv it y.” Jedwab, ­ R ., and ­A. ­S toreygard. 2017b. “The Work i n g Pap e r ­ N o. 2 0 2 35, N at io n a l Average and Heterogeneous Ef fec ts of Bureau of Economic Research, Cambridge, Transportation Investments: Evidence from ­M A. Sub-Saharan Africa ­ 1 9 6 0 – 2 0 1 0 .” Roberts, ­ M ., U­ . Deichmann, ­ B . Fingleton, and Unpublished ­manuscript. ­ T. ­S hi. ­2 012. “Evaluating China’s Road to Kim, ­ Y., and ­ B. ­ Z angerling. ­ 2 016. Mexico Prosperity: A New Economic Geography Urbanization ­R eview. Washington, ­ D C: ­A pproach.” Regional Science and Urban World Bank. Economics 42 (4): ­ 580–94. Leal, ­ E ., and ­ G. ­P érez. ­2 012 . “Port-Raill Storeygard, A ­ .­2016. “Farther on Down the Road: Integration: Challenges and Opportunities for Transport Costs, Trade and Urban Growth in Latin ­A merica.” FAL Bulletin 310 ­ (7). Sub-Saharan ­A frica.” Review of Economic Mesquita Moreira, M ­ ., ­C . Volpe Martincus, and Studies 83 (3): ­ 1263–95. ­J. ­Blyde. ­2008. Unclogging the Arteries: the Straub, ­ S. ­2011. “Infrastructure and Development: Impact of Transport Costs on Latin American A Critical Appraisal of the Macro-level and Caribbean ­ Trade. Washington, DC: Inter- L iterature.” Journal of Development Studies ­ American Development ­ Bank. 47 (5): ­ 683–708. Mesquita Moreira, ­ J. Blyde, ­ M ., ­ C . Volpe, and Summerhill, ­ W. ­ 2006. “The Development of R. ­ ­D . ­M olina. ­2 013. Too Far to E x por t: ­Infrastructure.” In The Cambridge Economic Domestic Transport Costs and Regional History of Latin America, Volume II, edited Export Disparities in Latin America and the by ­ V. Bulmer Thomas, ­ H. Coatsworth, and J. ­ 140   RAISING THE BAR ­ . Cortes Conde, ­ R 293–326. Cambridge, U.K.: Unger, ­ R. ­ K ., and ­ C hico. ­ 2 004. “La Industria Cambridge University P ­ ress. Automotriz en tres Regiones de ­ M éxico. Trebilcock, ­ M ., and ­ M. ­ 2 015. R osenstock. ­ Un análisis de ­ c lùsteres.” El Trimestre “Infrastructure Public-Private Partnerships in Económico 71 (284): ­ 909–41. the Developing World: Lessons from Recent Venables , A ­ . J. 20 07. “Evaluating Urban ­Experience.” Journal of Development Studies Transpor t I mprovements: Cost– B enefit 51 (4): ­ 335–54. Analysis in the Presence of Agglomeration Unger, ­K. ­ 2003. “Los Clústeres Industriales en and Income T ­ a x a t i o n .” J o u r n a l o f Mèxico: Especializaciones Regionales y la Transport Economics and Policy 41 (2): Política ­ i ndustrial. ” Documento de trabajo ­173–88. número 278, División de Economía, Centro Von Tunzelmann, ­ N. ­ G. ­ 1978. Steam Power and de Investigación y Docencia Económicas, British Industrialization to 1860. Oxford, Mexico ­City. U.K.: Oxford University ­ Press. Human Capital in Cities María Marta Ferreyra 5 Introduction rural areas even after taking this sorting into account. Moreover, a strong determinant Cities benefit individuals, and individuals of city productivity is aggregate human capi- benefit cities. By bringing individuals and tal—stronger, in fact, than population den- firms together, cities make individuals more sity and market access (see chapter 3). productive. Yet, it is not just the number of Cities with larger stocks of human capital individuals that makes a city productive; it is might be more productive for two main rea- also their “quality”—their human capital. sons. First, a greater share of skilled workers In other words, although all individuals con- should raise the productivity of unskilled tribute to city productivity, they do not all workers, to the extent that skilled and contribute equally. unskilled workers are complementary. For In this chapter, we investigate the role of example, when a construction company hires aggregate human capital in cities’ productivity. college-educated managers, these new hires When choosing where to live in a country, might streamline and speed up the produc- individuals compare locations on the basis of tion process, raising the productivity (and multiple attributes, including job opportuni- wages) of the unskilled construction workers. ties, housing values, and amenities such as cul- These complementarity effects are usually tural attractions and neighbors’ demographic reflected in skilled workers’ wages. This characteristics. Individuals thus sort across would be the case, in our example, if the locations, and skilled individuals are more college-educated managers were compen- likely to sort into cities than into rural areas.1 sated for their contribution to the productiv- Because skilled individuals are more pro- ity of their less-skilled colleagues. ductive than their unskilled counterparts, The second reason is the effects of human their sorting into cities could, in principle, capital externalities (HCEs). In a city with a lead to higher productivity there than in rural higher share of skilled workers, all workers areas. Yet cities in Latin America and the have greater opportunity to learn from skilled Caribbean (LAC) are more productive than workers—for example, by exchanging ideas, The author gratefully acknowledges the excellent research assistance of Angelica Sanchez Diaz. 141 142   RAISING THE BAR knowledge, and information even if they do income and college share. Because we find that not belong to the same firm. Because skilled college-educated workers gravitate toward workers are usually not paid for contributing areas with high college shares, the question to the productivity of others outside their firm, arises as to how areas with low college shares their presence yields positive externalities. can attract college-educated workers. One Given the importance of aggregate human possibility is to implement policies that raise capital in city productivity, this chapter first the demand for such workers. We run a simu- examines the distribution of human capital lation of such a policy for one area. across cities (more specifically, areas) in the The chapter’s main findings are as follows: region.2 It then presents estimates of the pro- ductivity gains due to aggregate human capi- •  Relative to small areas, large areas tal (henceforth, “returns to aggregate human have higher shares of skilled individ- capital”), and of their variation across coun- uals, and their income distributions tries. It considers two measures of aggregate are more unequal. Migrants to large human capital: the share of college-educated areas are more skilled than migrants workers (henceforth, “college share”), and to small areas. In order to acquire average years of schooling. skilled human capital, small areas rely Because the estimated returns might not more than large areas on “import- necessarily reflect externalities, this chapter ing” it. Most individuals in large then investigates possible channels of produc- areas work in services, but skilled and tivity gains from aggregate human capital. unskilled individuals work in different For unskilled workers, a college share type of services. Urban shares of pop- increase is expected to raise productivity ulation across countries differ mostly because of both complementarities and exter- depending on the urban share of their nalities. For skilled workers, in contrast, a unskilled population. college share increase raises their aggregate •  On average in the LAC region, an addi- supply. This, by itself, would lead to a wage tional year of average education raises decline for skilled workers. As a result, if a nominal wages (henceforth, “wages”) college share increase raises wages for skilled by 9.2 percent, and an additional per- workers, it must be due to externalities.3 We centage point in the college share raises therefore explore whether the response to a wages by 2 percent. These returns are college share increase varies among workers commensurate with private returns, of different educational attainments because although they are larger for average this can provide us with evidence about the years of schooling than for college existence of HCEs. share. Returns to aggregate human When these externalities exist, the policy capital are heterogeneous across coun- maker may want to subsidize policies that tries. They are U-shaped with respect raise an area’s aggregate human capital, either to the country’s average aggregate by forming it locally, or by attracting it from human capital.4 other areas. Yet, to attract or retain skilled •  In the LAC region, wages for college-ed- workers, the area must offer the attributes ucated workers rise with an increase sought by these individuals. For this reason, in college share. This supports the we investigate individuals’ valuation of loca- existence of HCEs in the LAC region. tional attributes. Crucially, these include the Returns to college share are U-shaped college share because college-­ educated indi- relative to a person’s educational attain- viduals may contribute to others not only by ment and are highest for the least raising their productivity but also by enrich- educated. Returns to average years of ing social interactions, becoming civically schooling are also U-shaped relative to involved, and contributing to lower crime. a person’s number of years of schooling, We examine the case of Brazil, a country yet they are highest for workers with the with large regional disparities in average highest number of years of schooling. H u m an C a p ita l in C ities   143 •  In Brazil, workers of all educational indicated otherwise. Income and wages are attainments value college-educated measured in nominal terms. neighbors and intercity connectivity. They value density only to the extent Stylized Fact 1. Human Capital Is educated that it raises the share of college-­ Unequally Distributed across and neighbors. within Countries •  In Brazil, simulation results for a policy that raises demand for college-educated In LAC areas, the average share of skilled workers show that such a policy would adult population is equal to 13 percent. benefit not only these workers but also Behind this average lies substantial variation others, who would gain in welfare and across and within countries: the median productivity through the greater pres- share of skilled adult population in Argentina ence of college-educated workers. is 32 percent, against 3.7 percent in Honduras (figure 5.1, panel a). The variation across areas is lowest in the countries with the low- Some Stylized Facts est and the highest average share of skilled adult population, yet is high in countries The unit of observation in this section (as in with intermediate levels of average share of chapter 3 and in the section below titled skilled adult population. Similar patterns “Returns to Aggregate Human Capital”) is hold for average yea rs of school i ng an administrative unit, except in a few cases (figure 5.1, panel b), as we would expect in which administrative units have been from the high area-level correlation (equal to merged.5 For most countries, the administra- 0.87) between the share of skilled adult pop- tive unit is a municipality; we present infor- ulation and average years of schooling. mation for the most recent year with data Human capital is more unequally distributed (see annex 5A for further details). For sim- within LAC countries than in comparator plicity, we use the term “area” for the unit of countries such as Poland or Turkey. observation. We measure area size by popula- Even among the largest areas in the region, tion; we thus consider an area larger than the share of the skilled adult population varies another when it has a greater population. We widely (figure 5.2). Ciudad de Buenos Aires consider an area to be small, medium, or tops the distribution, with a 60 percent share large when its population is below the coun- of skilled population. Its surrounding area, try’s median, between the country’s median Gran Buenos Aires, reaches only 23 percent. A and 75th percentile, and above the country’s country’s largest areas are not necessarily the 75th percentile, respectively.6 In figures that most educated. For example, the largest areas compare urban and rural areas, Argentina is in Brazil—Rio de Janeiro and São Paulo— not included because that country’s house- have around the same share of skilled adult hold surveys do not cover rural areas. population (about 30 percent) as Campinas, a We define the adult population as individu- medium-​ sized area that hosts multiple higher als age 25–64 years. In this section, “skilled” education institutions. Moretti (2004b) simi- individuals are those who have some higher larly notes that, in the United States, small education, regardless of whether they have and medium-sized areas that host large higher completed it; thus “unskilled” individuals are education institutions have disproportionately those who have at most completed high large shares of skilled individuals. school and have not started higher education. We classify individuals by their educational attainment into those with elementary, sec- Stylized Fact 2. Skilled Individuals Are ondary, and higher education when they have More Likely Than Others Are to Live in completed elementary, secondary, or higher Urban Areas education at most, respectively, and have not started the following level. “Average” refers Skilled individuals have a greater tendency to a simple (unweighted) average, unless than unskilled individuals to live in urban 144   RAISING THE BAR FIGURE 5.1  Distribution of Human Capital at the Area Level, circa 2014 a. percentage of adult population with some higher education b. Average years of schooling Argentina Argentina Chile Chile peru uruguay Dominican Republic Dominican Republic Colombia mexico Bolivia Ecuador uruguay Colombia mexico peru Costa Rica Costa Rica Ecuador Bolivia Brazil paraguay paraguay El Salvador El Salvador Brazil Nicaragua Honduras Guatemala Nicaragua Honduras Guatemala 0 20 40 60 0 5 10 15 percent Years Source: Calculations based on SEDLAC (for countries other than Brazil) and IPUMS (for Brazil). Note: Panel a shows a box plot per country for the distribution of the percentage of population with some higher education at the area level. Panel b shows a box plot per country for the distribution of the average years of schooling at the area level. Indicators are calculated for the population age 25–64 years. See annex 5A for the years used for each country. Countries are sorted by median values for each indicator. Outliers are indicated by dots. When there are no outliers, the left and right caps show the minimum and maximum value, respectively, and the box indicates the 25th and 75th percentiles (with the median indicated inside the box). In the cases of outliers, the left and right caps are the minimum and maximum values excluding the outliers. To identify outliers, we calculate the interquartile range; values outside the range defined by (25th percentile – 1.5 * interquartile range, 75th percentile + 1.5 * interquartile range) are considered outliers. areas (figure 5.3). Of the skilled adult popula- propensity to live in urban areas. Absent tion, 92 percent lives in urban areas, against other changes, countries with low shares of only 67 percent of the unskilled adult urban population will become more urban to population. the extent that their unskilled population For each country, the red diamond shows moves to urban areas. the overall propensity of the adult population to live in urban areas, and the vertical differ- Stylized Fact 3. The Higher the Area ence between the orange and blue bars below Population, the Greater the Share of shows the gap in the propensity to live in Skilled Individuals urban areas bet ween the skilled and unskilled. Although the fraction of skilled As shown in figure 5.4, a greater share of the population living in urban areas varies little population is skilled in large areas than in across countries, the fraction of unskilled small or medium-sized areas (Argentina is population varies more, driving the differ- the exception). On average, a 1 percent ences across countries in the overall increase in population is associated with H u m an C a p ita l in C ities   145 FIGURE 5.2  Population and Human Capital in the Largest Areas, circa 2014 60 Ciudad de Buenos Aires Adult population with some higher education (%) Arequipa Gran La Plata Mar del Plata Cercado Gran Cordoba Lima 40 Trujillo Porto Alegre Murillo Salta Valparaiso Gran Mendoza Quito Gran Rosario Curitiba Bogota, D.C. Gran Santa Fe Concepcion Brasilia Andres Ibañez Belo Horizonte Santiago Piura Chiclayo Montevideo Puebla Cartagena Medellin Rio De Janeiro Campinas São Paulo Santiago De Bucaramanga San Luis Potosi Alajuela Guayaquil Barranquilla Guadalajara Los Caballeros Quere Taro San Recife Monterrey Managua Torreon Salvador Distritos Asuncion Mexico D.F. Cachapoal Cautin Goiânia Gran Buenos Aires Belem Salvador 20 Distrito Central São Luis Cali Tijuana Fortaleza Pereira Cucuta Guarulhos Canoas Manaus Juarez Toluca San Pedro Sula Porto Velho Guatemala Nova Iguaçu San Marcos Altamira Alta Verapaz Huehuetenango 0 0 5 10 15 20 Population (millions) Source: Calculations based on SEDLAC (for all countries other than for Brazil) and IPUMS (for Brazil). Note: For each of the largest areas, the figure shows population (in millions) and percentage of adult population with some higher education. The figure shows selected areas in the region where population is greater than a given threshold (equal to 1 million for Brazil, Guatemala, and Mexico, and 500,000 for the remaining countries). The horizontal axis is in (natural) logarithmic scale. IPUMS = Integrated Public Use Microdata Series; SEDLAC = Socio-Economic Database for Latin America and the Caribbean. FIGURE 5.3  Percentage of the Adult Population Living in Urban Areas, circa 2014 100 90 80 70 60 percent 50 40 30 20 10 0 ca a ico bl n ua ala a a ile ru as y r r ay il do tin do bi i pu ica ua liv az Ri Ec c pe ur Ch ag gu ex em m i ua lva en ug Br Bo Re min nd sta r lo m ra ca at g ur Sa Co Ho pa Co Ar Ni Do Gu El unskilled adult population (age 25–64 years) living in urban areas Skilled adult population (age 25-64 years) Adult population living in urban areas living in urban areas Source: Calculations based on SEDLAC (for countries other than Brazil) and IPUMS (for Brazil). Note: Skilled population has at least some higher education. Urban areas are defined by national statistics offices. IPUMS = Integrated Public Use Microdata Series; SEDLAC = Socio-Economic Database for Latin America and the Caribbean. 146   RAISING THE BAR FIGURE 5.4  Percentage of Skilled Population, by Area Size, circa 2014 45 40 35 30 25 Percent 20 15 10 5 0 b n ico ca a ala a ia a ile ru as Do ay y r r il do bi in do gu pu ica ua liv az Co lic Ri Pe ur Ch gu ex em m nt ua lva ra ug Re min Bo Br nd sta lo M ge ra ca at Ec Ur Sa Co Ho Pa Ar Ni Gu El Small Medium Large Source: Calculations based on SEDLAC (for countries other than Brazil) and IPUMS (for Brazil). Note: The figure shows the average percentage of the adult population (age 25–64 years) with some higher education, by area size. Area size classification follows country-specific population thresholds, as explained in annex 5A. IPUMS = Integrated Public Use Microdata Series; SEDLAC = Socio-Economic Database for Latin America and the Caribbean. a 0.29 percent increase in the share of skilled On average, the elasticity of the Gini coef- adult population. 7 The corresponding ficient with respect to area population is increase is lower (equal to 0.12 percent) in equal to 0.03.8 In other words, on average a the United States (Behrens and Robert- 1 percent increase in population is associated Nicoud 2015). with a 0.03 percent increase in inequality (as measured by the Gini coefficient). The elas- ticity is lower for the United States (equal to Stylized Fact 4. Larger Areas Have 0.012), indicating a stronger tendency toward Greater Income Inequality income inequality in large LAC cities. We measure income inequality for each area When controlling for the share of skilled through the Gini coefficient (whose value population, the LAC elasticity of the Gini ranges between 0 and 1; higher values indi- coefficient with respect to area population cate greater income inequality). As shown in falls on average from 0.03 to 0.017, and from figure 5.5, in most countries larger areas tend 0.012 to 0.009 for the United States. This to have greater income inequality. The pat- implies that the share of skilled population tern does not hold, or holds less strongly, for accounts for 43 percent of the association Argentina, Bolivia, Ecuador, Guatemala, between city population and income inequal- Paraguay, and Peru. ity in the LAC region yet for only 25 percent As it turns out, large cities are more of this association in the United States. So unequal largely because they have greater education is more strongly associated with shares of skilled population. We arrive at this income inequality in large LAC cities than in conclusion by estimating the elasticity of the the United States. Gini coefficient with respect to area popula- The fact that city population, skills, and tion, with and without controlling for the inequality are more strongly associated in the share of skilled population. LAC region than in the United States reflects H u m an C a p ita l in C ities   147 FIGURE 5.5  Average Gini Coefficient, by Area Size, circa 2014 0.60 0.55 0.50 Gini coefficient 0.45 0.40 0.35 0.30 a bl n ia a a ico ca ile ala ru ay as y r r il do gu bi in do pu ica ua liv az Pe ic Ri ur Ch gu ex em nt m lva ra ua ug Br Bo Re min nd sta lo ra ge M ca at Ec Ur Sa Co Pa Ho Co Ar Ni Do Gu El Small Medium Large Source: Calculations based on SEDLAC (for countries other than Brazil) and IPUMS (for Brazil). Note: The figure shows the average Gini coefficient by area size (weighted average; weight is area size). Area size classification follows country-specific population thresholds, as explained in annex 5A. IPUMS = Integrated Public Use Microdata Series; SEDLAC = Socio-Economic Database for Latin America and the Caribbean. the LAC region’s relative skill scarcity. In (ECA), 11 percent of the population living in LAC countries, the share of skilled popula- urban areas is foreign born.10 Hence, here tion is lower than in the United States. For we focus on domestic migrants. example, the share of individuals with some On average, 7.16 percent of household higher education in the average LAC country heads age 25–35 years have migrated within (18.4 percent) is roughly one-third of that in their countries in the last five years.11 In most the United States (59 percent). Second, countries, migrants to large areas are more returns to higher education are higher in the likely to be skilled than migrants to medium LAC region than in the United States. For or small areas (except in Ecuador, El example, returns to complete higher educa- Salvador, and Paraguay) (figure 5.6). In other tion are equal to 104 percent for the average words, large areas benefit from the inflow of LAC country, more than twice as high as in skilled migrants at a higher rate than medium the United States.9 or small areas. This, in turn, reinforces large areas’ advantage in human capital.12 Stylized Fact 5. Migrants to Large Areas Are More Skilled Than Migrants Are to Stylized Fact 6. To Acquire Skilled Small Areas Human Capital, Small Areas Rely More Than Large Areas Are on Migration The migration that fuels areas’ growth in the LAC region is of domestic rather than In most countries, the share of skilled popu- foreign origin. Only 2.8 percent of these lation that arrived via migration is larger in areas’ adult population is foreign born. smaller areas (figure 5.7)—smaller areas are Argentina and Costa Rica are the countries more likely to “import” their skilled popula- whose areas attract most international tion. This might be because larger areas are migration (see annex 5D). In contrast, in more likely to host higher education institu- comparator countries in East Asia and tions and so develop their own skilled popu- Pacific (EAP) and Europe and Central Asia lation, which then stays in the area.13 148   RAISING THE BAR FIGURE 5.6  Percentage of Skilled Migrants, by Area Size, circa 2014 80 70 60 50 Percent 40 30 20 10 0 ala a ia a a ico ile as y r ay r il do do in bi gu ua liv az ur Ch gu em ex nt m ua lva ra ug Bo Br nd lo ge ra M ca at Ec Ur Sa Co Ho Pa Ar Ni Gu El Small Medium Large Source: Calculations using IPUMS (for Brazil, Colombia, El Salvador, and Mexico) and SEDLAC (for all other countries). Note: The figure shows, among household heads age 25–35 years who have migrated within the past five years, the percentage who are skilled (who have at least some higher education) by size of their destination area. Area size classification follows country-specific population thresholds; as explained in annex 5A. IPUMS = Integrated Public Use Microdata Series; SEDLAC = Socio-Economic Database for Latin America and the Caribbean. FIGURE 5.7  Percentage of Skilled Individuals Who Are Migrants, by Area Size, circa 2014 50 45 40 35 30 Percent 25 20 15 10 5 0 ico le a a a ala a y y r r il do do i bi in gu ua ua liv i az Ch ex em m nt ua lva ag ug ra Bo Br lo M ge ca at Ec r Ur Sa Co Pa Ar Ni Gu El Small Medium Large Source: Calculations using IPUMS (for Brazil, Colombia, El Salvador, and Mexico) and SEDLAC (for all other countries). Note: The figure shows, for each area size, the percentage of skilled individuals who are migrants. A migrant is someone who has moved within the past five years from a different department (Admin-1 unit) except in the case of Bolivia, where migrations are from any other place in the country. The sample consists of household age 25–35 years. Area size classification follows country-specific population thresholds, as explained in annex 5A. IPUMS = Integrated Public Use Microdata Series; SEDLAC = Socio-Economic Database for Latin America and the Caribbean. Stylized Fact 7. The Larger the Area, areas host larger shares of skilled individu- the Greater the Employment als (stylized fact 3), who are more likely Share in Services than their unskilled counterparts to work in As figure 5.8 shows, in larger areas, a services regardless of area size (annex 5B).14 greater share of the adult population is Second, in larger areas, individuals of all employed in services (except in Argentina). skill levels are more likely to work in ser- This is due to two reasons. First, larger vices (annex 5C). H u m an C a p ita l in C ities   149 FIGURE 5.8  Percentage of Employment in Services, by Area Size, circa 2014 100 90 80 70 Percent 60 50 40 30 20 ca a a bl n ico ile ala a ia y as ay ru r r il do bi gu do in pu ica ua az liv Ri ic ur Pe Ch gu ex em m nt lva ra ua ug Br Re in Bo nd sta lo ra M ge ca m at Ec Ur Sa Co Ho Pa Co Ar Ni Do Gu El Small Medium Large Source: Calculations based on SEDLAC (for countries other than Brazil) and IPUMS (for Brazil). Note: The figure shows the employment share (in percent) of the service sector, by area size. Area size classification follows country-specific population thresholds; as explained in annex 5A. The sample consists of workers age 25–64 years. IPUMS = Integrated Public Use Microdata Series; SEDLAC = Socio- Economic Database for Latin America and the Caribbean. Nonetheless, the specific service sector take place in cities, labor productivity is where individuals work in large areas lowest in wholesale, retail, hotels, and depends on their skill level (figure 5.9). restaurants—namely, in the service sec- Skilled individuals are most likely to work in tors where unskilled individuals are most public administration, education, health, likely to work.15 social work, financial intermediation, and real estate; unskilled individuals in whole- Summary sale, retail, hotels, restaurants, transport and communications. I n services in large In the LAC region, skilled individuals tend areas, skilled individuals are also more likely to sort into large areas, and migrants to than unskilled individuals to work in the large areas are more skilled than migrants public sector, mainly in public administration to small areas. Large areas are also more and education. likely than small areas to develop their own Structural transformation in the LAC human capital than to “import” it from region seems to have shifted an increasing other areas. share of workers into low-productivity Larger areas have more unequal income service sectors since 1960 (see chapter 1). distributions, largely because they have A lt houg h bot h sk i l led a nd u nsk i l led greater shares of skilled individuals. Put dif- i ndividuals may be vulnerable to this ­ ferently, the greater inequality of larger areas trend, unskilled individuals seem to be is a consequence of their ability to attract more vulnerable given the service sectors high-earning, “successful” individuals. where they tend to work. Data from the Importantly, the sorting of skilled individuals Gron ingen Grow th and Development into large areas does not necessarily mean Center (GGDC), also used in chapter 1, that these individuals have a taste for size (or show that, among activities that usually population density) in itself; rather, it may 150   RAISING THE BAR FIGURE 5.9  Percentage of Service Workers, by Sector in Large Areas and by Skill Level, circa 2014 100 90 80 70 60 Percent 50 40 30 20 10 0 Unskilled Skilled Public administration, education, health, and social work Financial intermediation and real estate Transport and communications Wholesale, retail, hotels, and restaurants Construction Other activities Source: Calculations based on SEDLAC (for countries other than Brazil) and IPUMS (for Brazil). Note: For each skill level, the figure shows the distribution of employees across subsectors in the service sector in Latin America and the Caribbean. Unskilled workers are those who have completed high school at most, and skilled workers have at least some higher education. The sample consists of workers age 25–64 years who work in the service sector and live in large areas in Latin America and the Caribbean. Large areas are determined following country-specific population thresholds; see annex 5A for further information. IPUMS = Integrated Public Use Microdata Series; SEDLAC = Socio-Economic Database for Latin America and the Caribbean. reflect a preference for locational attributes than in the United States. For example, it is that they more typically find in large areas, possible that only a few areas offer good job including job opportunities, cultural ameni- opportunities to college graduates. ties, and a high college share. It may also Across LAC countries, differences in the reflects their greater ability to pay for hous- share of urban population are mostly driven ing because housing prices are usually higher by the unskilled. It is possible that, as a coun- in large areas. We revisit these issues later in try’s share of skilled population grows the chapter. (because, perhaps, of the expansion of educa- A lt houg h t he posit ive asso ciat ion tion in the country), the share of urban popu- between area size, education, and inequality lat ion g rows si mply bec ause sk i l led has been documented for the United States population sort into large areas. But this as well, we find that it is stronger for LAC. share may also rise as the unskilled move to This may be because a smaller share of the urban areas. Because the services in which the population in the LAC region is skilled, and unskilled typically work in large areas (whole- returns to higher education are substantially sale, retail, hotels, and restaurants) are of low higher. Yet, it is also possible that locational productivity, urbanization may thus continue attributes may be less evenly distributed the trend of shifting workers into low-­ across a country’s areas in the LAC region productivity service sectors (see chapter 1). H u m an C a p ita l in C ities   151 Returns to Aggregate entirely due to these complementarities, Human Capital without any role for HCEs. A critical distinction between comple- Through education, workers become more mentarities and HCEs is that the former are productive and earn concomitantly higher likely internalized in skilled workers’ wages, salaries. Using the same sample of LAC work- but the latter are not. For example, if a ers as in Quintero and Roberts (2017) and as skilled worker raises the productivity of in chapter 3, we find that a worker’s addi- unskilled workers inside his or her firm by tional year of schooling increases her or his sharing her or his knowledge and skills, that salary, on average, by 8.92 percent.16 This is worker is likely to be paid for it. In contrast, on a par with estimates from other researchers. if the skilled worker raises the productivity For example, when estimating returns to of a worker in another firm, most likely that education for nations all over the world, skilled worker is not paid for it. Montenegro and Patrinos (2014) find an aver- The distinction between complementarities age return of 10 percent for high-income and HCEs is important for policy. Because economies, and of 9.2 percent for LAC. HCEs are a market failure, their correction Education benefits society as well. For requires policy intervention—for example, example, more-educated individuals may through government subsidies to the forma- contribute to the generation and dissemina- tion of human capital. Complementarities, in tion of ideas, knowledge, and products. They turn, do not require policy intervention. may be more informed and engaged citizens, A form of complementarity arises when a and may advocate for greater quality (and greater share of college-educated individuals perhaps variety) of public goods. They may raises demand for the services provided be less likely to engage in criminal activities. by less skilled individuals, that is, when In this section, we focus on a specific kind college-educated individuals take more cab of social benefit from education, namely the rides, or hire more house cleaners and nan- returns that a person’s human capital has on nies, than less-educated individuals. Because the productivity (and hence wages) of others salaries for the unskilled adjust to reflect this in her or his area. Although a person’s human greater demand, this phenomenon does not capital could also benefit others in larger geo- constitute a market failure. graphic units (such as the state and the coun- In an area, HCEs can arise through inter- try), we focus on benefits in the area because actions among individuals. For example, they this is the geographic unit in which most of a can arise when workers from different compa- person’s interactions are likely to take place. nies interact in formal settings such as confer- To estimate these social benefits, research- ences, public presentations, and joint projects ers have often compared wages for workers among companies. They can also arise when who live in areas with different levels of workers from different companies interact in aggregate human capital, controlling for informal settings such as school meetings, other area-level characteristics and for work- civic associations, or even the neighborhood. ers’ characteristics. We use the term “return All these interactions provide learning oppor- to aggregate human capital” to refer to these tunities for workers, who thus exchange ideas wage differences. Returns to aggregate and share relevant knowledge and skills. human capital, however, may reflect not only Verbal exchanges may not even be required HCEs but also complementarities between for learning because a worker can learn from skilled and unskilled human capital. When another merely by observing that person. these complementarities are present, higher HCEs can also arise from the actions and aggregate human capital raises productivity behaviors of skilled individuals that benefit (and hence wages) for the unskilled and, all individuals in the area. For example, perhaps, for the skilled as well. Moreover, because skilled individuals are less likely to the return to aggregate human capital may be engage in crime than unskilled individuals, 152   RAISING THE BAR an area with more aggregate human capital for individual-level characteristics and for is likely safer. To the extent that safety con- other area characteristics, including popula- tributes to productivity, this might be tion density and market access. reflected in wages.17 Similarly, college-edu- Two caveats are in order. The first is that it cated individuals are more likely to be civi- is possible that area density, aggregate human cally engaged and to demand better public capital, and market access are endogenous, services from local authorities. To the extent that is, they are correlated with some unob- that these improve productivity, they might served area or individual characteristic (see be reflected in wages as well. chapter 3). Lacking instruments for the whole region, we proceed as if endogeneity were not a concern (with the same caveats made in Estimated Returns to Aggregate that earlier chapter), and use the term Human Capital “returns to aggregate human capital” to When measuring returns to aggregate human denote the coefficient on aggregate human capital, researchers have used two measures: capital. The second caveat is that, as indi- average years of schooling of those living in cated above, the returns to aggregate human an area, and the share of individuals with capital may not only (if at all) reflect HCEs. some higher education (whether they com- We return to this point in the next section. pleted it or not). In our investigation of the As column 1 shows, an additional year of determinants of area-level productivity in average schooling in our sample is associated chapter 3, we used (log) average years of with a salary increase of 9.2 percent, which is schooling, and share of individuals in the slightly larger than the estimated average pri- working-age population (WAP) with com- vate return to education in the LAC region pleted higher education.18 (recall that this is equal to 8.92 percent in our In table 5.1, columns 1 and 2 show the sample). This finding is consistent with other coefficients on two alternative measures of studies, both for the developed and developing aggregate human capital, namely average world, which have estimated returns to aggre- years of schooling and share of individuals in gate human capital in the range of 50–100 the WAP with completed higher education percent of private returns (Duranton 2014). (or college share). Both regressions control The estimated return for the LAC region is large: if all individuals acquired an extra year TABLE 5.1  Returns to Aggregate Human Capital, 2000–14 of schooling, each of them would reap a salary (1) (2) increase of about 18 percent, in roughly equal Average no. of years of schooling 0.092 *** parts from own and aggregate human capital. (0.012) In our sample, a 1 percentage point increase in the college share is associated with a 2 per- % WAP with completed higher education 0.020*** cent average salary increase (column 2). As (0.003) with the average number of years of school- Adjusted R 2 0.832 0.803 ing, the estimated return to college share is in No. of observations 5,050 5,050 line with the private returns to higher educa- Source: Calculations using SEDLAC for all countries except for Brazil, and IPUMS for Brazil. Sample is tion: on average in the LAC region, a higher the same as that used by Quintero and Roberts 2017. Column 2 reports the same results as those of education graduate earns 104 percent more column 5 in table 3.2. Note: This table regresses estimated area-level productivity on aggregate human capital. The than a high school graduate, controlling for coefficients represent returns to aggregate human capital; when multiplied by 100, returns are observed characteristics (Ferreyra et al. 2017). expressed in percent. In these regressions, a unit of observation is an area; all Latin American and Caribbean areas are pooled in the regressions. Both regressions control for population density, Thus, a 1 percentage point increase in college market access, air temperature, terrain ruggedness, and precipitation; both include country share will raise average wages by about 1 per- fixed effects. Area-level productivities are estimated by regressing, for each country, log wages on individual-level characteristics (age, age squared, years of schooling, gender, and marital cent, which is commensurate with our esti- status) and year fixed effects. Average years of schooling is calculated for individuals in the WAP. mated return to the college share. IPUMS = Integrated Public Use Microdata Series; SEDLAC = Socio-Economic Database for Latin America and the Caribbean; WAP = working-age population (individuals age 14–65 years). Our estimated return to college share in *** p < 0.01, ** p < 0.05, * p < 0.1. Standard errors are clustered by country. the LAC region is somewhat larger than that H u m an C a p ita l in C ities   153 in Moretti (2004a) for the United States area has an average number of years of school- (which is in a range of 0.6–1.2 percent), yet of ing of 7.35, which in most countries is equiva- similar magnitude. The estimate might be lent to just having finished elementary larger for LAC because, in the average LAC education. With such low educational attain- area, only 5 percent of the WAP has com- ment, an additional year of average education pleted higher education, against 23 percent in might have high returns in wages but might U.S. cities (Moretti 2004b).19 not affect the college share.20 It is reassuring, To provide context for the estimated however, that country-level estimates of return to college share, an area would, on returns to human capital using average years average, need to raise its college share by of schooling and college share are highly and 4.6 percentage points to attain the same positively correlated (correlation = 0.75).21 social benefit derived from an extra year of Returns to aggregate human capital are average schooling. Relative to the average heterogeneous across countries (see chapter 3). college share (equal to 5 percent), this is a In figure 5.10, we investigate whether this het- very sizable increase. It is approximately erogeneity is related to countries’ aggregate equal to the increase in college share in the human capital, measured as the average LAC region between 2002 and 2012, during (across areas) of aggregate human capital. For the region’s remarkably large and fast higher both measures of aggregate human capital, education expansion (Ferreyra et al. 2017). there seems to be a U-shaped association Hence, returns to aggregate human capital between a country’s average aggregate human appear large when aggregate human capital is capital and its return to aggregate human cap- measured by average years of schooling, but ital. In other words, as aggregate human not as large when measured by college share. capital rises, returns to aggregate human cap- ­ This might be because the average educational ital first fall, and then rise. One possible attainment is low in the region: the average explanation for this pattern is that aggregate FIGURE 5.10  Returns to Aggregate Human Capital, 2000–14 a. Returns to average years of schooling b. Returns to share of higher education graduates 0.20 BOL 0.10 HND HND 0.08 GTM 0.15 0.06 GTM Returns Returns 0.10 MEX CHL BOL PER 0.04 BRA COL BRA MEX 0.05 ECU CRI PER SLV 0.02 ECU CHL NIC SLV NIC COL CRI DOM DOM 0.00 0.00 5 6 7 8 9 10 0 2 4 6 8 Average years of schooling Average share of higher education graduates (%) Source: Calculations using SEDLAC for all countries except for Brazil, and IPUMS for Brazil. Panel a sample and panel b returns are from Quintero and Roberts 2017. Note: The vertical axis shows, for each country, the estimated returns to aggregate human capital. The horizontal axis shows, for each country, the average of the corresponding variable; the average is calculated over the country’s areas. Average years of schooling, and share of higher education graduates, correspond to individuals age 14–65 years. Returns can be expressed in percent if multiplied by 100. To obtain these returns, for each country we regress area-level productivity on the corresponding measure of aggregate human capital; these regressions control for area density, market access, air temperature, terrain ruggedness, and precipitation. Area-level productivities are estimated by regressing, for each country, log wages on individual-level characteristics (age, age squared, years of schooling, gender, and marital status) and year fixed effects. We do not run these regressions for Argentina, Panama, or Uruguay because they have few areas. Coefficients from the quadratic specification in panel b are significantly different from zero. Coefficients from the quadratic specification in panel a are not significantly different from zero (if a linear specification is fitted to the data in panel a, the corresponding coefficient is not significantly different from zero either). IPUMS = Integrated Public Use Microdata Series; SEDLAC = Socio-Economic Database for Latin America and the Caribbean. For a list of country abbreviations, see annex 2A. 154   RAISING THE BAR human capital has decreasing returns when it workers benefits unskilled workers because is low, but increasing returns once it surpasses of complementarities and HCEs. Because the a certain threshold—perhaps indicating the two effects work in the same direction, need for a critical share of skilled workers (or a return to the share of skilled workers for for workers with a minimum number of years the unskilled does not provide evidence of schooling) who can benefit from the pres- of HCEs because it could be entirely due to ence of other skilled workers. complementarities. In contrast, an increase in the share of skilled workers depresses the wages of other skilled Complementarities Versus Human workers because it raises their relative supply, Capital Externalities yet also raises their wages via HCEs. For skilled Although the evidence above indicates posi- workers, the net effect is positive only if HCEs tive returns to aggregate human capital, are sufficiently large.22 Thus, a positive return we recall that this association might not be to college share for skilled workers provides evi- due—at least not solely—to HCEs. Consider, dence of the presence of HCEs. for example, the return to college share. This When he investigates whether the return return could be positive not only because of to college share varies by educational attain- HCEs, but also because of workers’ comple- ment in the United States, Moretti (2004a) mentarities with college-educated workers. finds that this return is positive for college-­ To distinguish between the two factors, educated individuals, thus confirming the it is useful to investigate whether the return presence of HCEs. Furthermore, he finds that to college share varies by own education the return declines with a worker’s educa- (Moretti 2004a). On the assumption that tional attainment. workers of different skills are complemen- For LAC, we also find that returns to col- tary, an increase in the share of skilled lege share are positive for college-educated FIGURE 5.11  Returns to Aggregate Human Capital, by Individual’s Own Education, 2000–14 a. Returns to share of college graduates b. Returns to average years of schooling 0.030 0.18 0.16 0.025 0.14 0.020 0.12 Returns 0.10 Returns 0.015 0.08 0.010 0.06 0.04 0.005 0.02 0 0 2 4 6 8 10 12 14 16 18 ar e ar e da e e at er uc igh te im m im et on Som nd et uc igh ed h ple y y ry So ary n at er pr pl co pl pr So io Years of education m se om m n ed e h io Co Co m C c se Source: Calculations based on Socio-Economic Database for Latin America and the Caribbean for all countries except for Brazil, and IPUMS for Brazil. Sample is the same as used by Quintero and Roberts 2017. Note: Panel a shows, for each educational attainment, the return to the share (in percent) of college graduates. In both panels, returns can be expressed in percent if multiplied by 100. To construct panel a, we pool data from all countries and regress log wages on individual characteristics (age, age squared, indicators of educational attainment, gender, and marital status) interacted with country dummies, area-level characteristics (density, share of college graduates, market access, air temperature, terrain ruggedness, and precipitation), country-year fixed effects, and the interaction between indicators of individual educational attainment and the area share of college graduates. Individuals with completed primary (secondary) have not started secondary (higher) education. Panel b shows, for each value of own years of schooling, the return to average years of schooling. To construct panel b, we pool data from all countries and regress log wages on individual characteristics (age, age squared, years of schooling, years of schooling squared, gender, and marital status) interacted with country dummies, area-level characteristics (density, average years of schooling, market access, air temperature, terrain ruggedness, and precipitation), country-year fixed effects, the interaction between own years of schooling and average years of schooling, and the interaction between own years of schooling squared and average years of schooling. All relevant coefficients for these panels are significantly different from zero. H u m an C a p ita l in C ities   155 workers, thus providing evidence of the worker in this area unskilled. Thus, an presence of HCEs (figure 5.11, panel a). additional year of average education does Unlike Moretti (2004a), we find that the not alter the average college share in the returns to college share are U-shaped relative LAC region, but it changes the average skill to own educational attainment: they are high- of the unskilled. For example, a person with est for the least educated and decline with edu- five years of schooling may not benefit if cational attainment, as in Moretti (2004a), but average years of schooling rises from seven rise again for college-​ educated workers. The to eight years because this would make fact that, in the LAC region, returns are higher other unskilled workers more educated for college-educated workers than for workers (and hence employable). This would explain with complete secondary education, or some the descending portion of figure 5.11, panel higher education, suggests that human capital b. However, the same situation may benefit externalities might be higher for college-­ a person with 16 years of schooling because educated workers than for workers with those it may allow her to specialize in complex other attainments. But, more important, activities and leave easier ones to the aver- returns to college share are highest for the least age worker, who is now more educated. educated—either because a higher college ­ Thus, the returns to average years of school- share implies greater demand for their services ing may capture more complementarity (for example, as restaurant workers or cab effects than the returns to college share drivers), because it allows them to work in the because of the low number of average years same firms as skilled college educated workers, of schooling in the region. In contrast, the or because it allows them to learn from college returns to college share may capture more educated workers outside the firm. HCEs because it is plausible that college-­ This U-shaped pattern also holds for the educated workers would generate more pos- alternative measure of aggregate human itive externalities than workers who have capital, namely average years of schooling only finished elementary school. 23 (figure 5.11, panel b). In particular, an individ- The U-shape of the return to aggregate ual can reap increasing returns from average human capital for own education (see figure years of schooling once she or he completes at 5.11, panels a and b) is reminiscent of least seven years of schooling (roughly equiva- the U-shape of the returns for a country’s lent to finishing elementary school). With 9.36 average aggregate human capital (see figure years of schooling, the average individual has 5.10). In other words, aggregate human already surpassed this threshold. capital has high returns for an individual (or Although the returns to both measures of average area) with low education; these aggregate human capital are U-shaped in returns fall as the individual (or a country’s own human capital, who enjoys the highest average area) acquires more education, and returns varies depending on the measure: finally rise once the individual (or average people with the lowest educational attain- area) has acquired sufficient education. ment benefit the most from college share To summarize, the evidence suggests (figure 5.11, panel a), yet people with the that an area’s aggregate human capital highest number of years of schooling benefit raises average productivity and that at least the most from average years of schooling part of this return can be attributed to (figure 5.11, panel b). These differences may HCEs. The least skilled individuals benefit be related to the region’s level of educational the most from an increase in college share, attainment and with the fact that workers and individuals with the highest years of might not be perfect substitutes within a schooling benefit the most from an increase given skill level. in average years of schooling. By enhancing To see this, recall that the average years workers’ productivity, areas with high of schooling equals 7.35 years (roughly aggregate human capital are thus attractive equal to elementary school) in the average to all individuals, holding other locational area in the region, which makes the average attributes constant. 156   RAISING THE BAR Attracting Skilled Individuals different skill levels in Brazil. We also study to Cities the productivity spillovers of college gradu- ates onto other workers and consider the Given the evidence that skilled human capi- effects of a hypothetical program that tal raises aggregate productivity, it seems as attracts college graduates to an area. 24 though local leaders would be interested in We draw largely on Fan and Timmins (2017), attracting such individuals to local commu- a background paper for this book. Box 5.1 nities. These efforts, however, can be describes their model. successful only to the extent that the com- Fan and Timmins (2017) use data from a munities offer the locational attributes 5 percent sample of the 2010 Population sought by skilled individuals. This issue may Census in Brazil. They focus on the loca- be of particular importance to medium-sized tional choices of household heads age 25–35 and small areas, which rely more than large years, who choose among almost 1,400 areas on migration to raise college share municipalities in 27 states. 25 Lacking better (stylized fact 6). data, they consider a person as having moved In this section, we study the determinants if she or he resides in a municipality outside of location choice among individuals of her or his birth state. BOX 5.1  An Equilibrium Model of Household Sorting for Brazil Fan and Timmins (2017) study how individuals challenge in the estimation of preferences for local choose their municipality (or locality) of residence in attributes. By using state-of-the-art methods, the Brazil. They develop an equilibrium model of house- authors overcome this challenge and recover prefer- hold locational choice and local labor markets. They ences over these attributes. estimate the parameters of individuals’ utility func- One local attribute considered by people is their tions and of local labor markets’ productivity, and use expected income in that location. The authors model the parameter estimates to perform counterfactuals. it as a function of individual characteristics and a In the model, when choosing among locations, indi- local productivity term that varies by educational viduals consider natural attributes such as elevation and attainment (for example, the composition of eco- climate, and others such as job opportunities (prox- nomic activity of a given municipality may be par- ied by their expected income in the specific location), ticularly fitting for college graduates). Expected housing values, population density, density of college income is also an outcome of individuals’ sorting. graduates, connectivity with other locations, and other For example, local productivity for high school attributes not observed by the researcher.a Importantly, graduates depends on the local density of high preferences over these attributes are allowed to vary school graduates (because an increase in this density by educational attainment—for example, the density would render them less scarce and hence less valu- of college graduates may be more valuable to college able) and on the local density of college graduates graduates than to less-educated individuals. The model (because an increase in this density could make them accounts for the fact that moving is costly, which in more productive, as we saw in the previous section). turn makes people more likely to remain in a location, The model accounts for the fact that changes in even when it is not their preferred one. locational attributes (including labor market con- Some local attributes, such as elevation and ditions) may lead people to change their locational precipitation, do not depend on people’s collec- choices, which in turn changes the local attributes tive decisions, yet others do. Such is the case for a resulting from individuals’ sorting and leads to fur- municipality’s population density, the density of col- ther re-sorting. The model, then, can be used to lege graduates, and housing prices. In other words, evaluate the equilibrium impact of specific policies some local attributes are the outcome of individu- on population density, skill composition of the pop- als’ sorting and are thus endogenous. This poses a ulation, housing prices, incomes, and utilities. a. A municipality’s density of college graduates is defined as the municipality’s number of college graduates divided by the municipality’s area. Thus population density is equal to the density of college graduates plus the density of less-educated individuals. H u m an C a p ita l in C ities   157 Brazil is a country with large regional dis- is positively and significantly correlated with parities. Incomes for the average individual in the number of museums, theaters, and restau- the midwest or southeast (the richest regions) rants. Even if less-educated individuals can- are almost twice as high as in the north- not afford all these amenities, they may still east (the poorest region). In the sample, enjoy their presence. Finally, preferences for 16.9 percent of individuals have moved. college-educated individuals might also Moving is most likely among college-­ reflect the preference for other, unmeasured educated individuals and among individuals amenities that are correlated with the pres- born in the northeast region. The southeast ence of such individuals, as is the case of attracts the highest share of migrants. lower crime. Individuals who completed at least high school prefer locations with additional uni- Preferences for Locational Attributes versity buildings. This might reflect the An important finding from the model’s esti- value they attach to institutions that their mation is that moving costs are steep. These child ren mig ht at tend because about costs may not only be pecuniary but also 80 percent of higher education students in reflect other considerations such as the diffi- the LAC region live with their parents while culty of separating from family members and attending higher education (Ferreyra et al. friends. Fan and Timmins (2017) also 2017). It might also reflect university spill- uncover a pattern of individuals’ preferences overs in the community, for example, in locational attributes. through extension activities with the com- In principle, individuals might like dense munity at large, or through research and places, perhaps because greater density facili- innovations that benefit the community. tates social interaction. The authors find that Most individuals value additional road whether density is liked or not depends on and rail density, reflecting the value they the composition of the additional population, place on the connectivity of the location with and on whether it leads to a greater or lower others. All individuals value the presence of a share of college-educated individuals. People shoreline. Although they like higher tempera- like population growth, or changes in popu- tures and more abundant winter rain, they lation composition, so long as they do not dislike summer rain. lower the college share. Thus, people weigh multiple locational Importantly, individuals of all skill levels attributes when choosing where to live. enjoy living in locations with a higher share Critical attributes, however, are the presence (and density) of college graduates. Preference of college-educated neighbors and intercity for college graduates captures not only an connectivity. This suggests that areas seeking intrinsic preference for more educated neigh- to attract such individuals can resort to poli- bors but also a preference for local attributes cies that raise the local demand for col- associated with the presence of such neigh- lege-educated workers, improve intercity bors in the community. As Fan and Timmins connectivity, or expand higher education. (2017) document, in Brazil the density of col- Moreover, that college share is highly and lege graduates is positively and significantly positively correlated with the presence of correlated with the provision of trash collec- other urban amenities, such as public services tion, sewage, and water, perhaps because and cultural attractions, indicates that governments are more likely to provide these expanding the provision of these amenities services to areas with a relatively large share may also raise the college share. of middle- and high-income individuals who To the extent that household preferences in exert political pressure to receive these ser- Brazil are similar to those in other LAC coun- vices and can pay for them, or because these tries, the estimates suggest that the sorting of individuals can afford housing in areas with skilled individuals into large areas observed in high service provision. Similarly, the authors the LAC region (see the stylized facts earlier in the document that the share of college graduates chapter) does not reflect their taste for area size; 158   RAISING THE BAR rather, it mainly reflects their taste for the loca- for which reason Fan and Timmins (2017) tional attributes more usually found in large explore the effects of a hypothetical pro- areas, including a high college share. gram that expands the demand for higher education graduates in Feira de Santana, Labor Demand a mid-sized municipality in Brazil’s relatively poor northeast. The program raises wages Fan and Timmins (2017) also estimate labor offered to higher education graduates by demand for individuals in local labor mar- 50 percent. Hiring college instructors for a kets. How much employers are willing to pay local college, or hiring physicians and to those of a given educational attainment researchers for a local hospital, would exem- depends on local labor market conditions. plify this type of program. Table 5.2 shows that, as expected, employers Table 5.3 shows average program effects are willing to pay less to workers with less for individuals by educational attainment in than completed higher education when the the municipality. The program increases the local labor market displays a greater density number of higher education graduates in of those skills (see the first three coefficients Feira de Santana by 17.2 percent. Further, in row 1). Employers are willing to pay it raises the density and share of college slightly more to college graduates when the graduates. The inflow of higher education density of college graduates is higher, graduates attracts less-educated individuals although this effect is not significantly differ- as well, who arrive in the municipality to ent from zero. enjoy the presence of a greater share of higher The table further shows that employers are education graduates and to benefit from the willing to pay more to workers of any skill greater labor demand induced by their pres- level when the local labor market displays a ence. Most immigrants are from neighboring greater density of higher education graduates. towns. The table shows average effects for The positive spillovers of college graduates on individuals in Feira de Santana, which, after less-skilled workers might be due to both the program, also includes immigrants from complementarities and HCEs (as seen). As in other municipalities. Moretti (2004a), these spillovers are larger As intended, average income rises by than those on college-educated workers. about 50 percent for college graduates yet also rises for less-educated individuals. For higher education graduates, the increase in The Overall Effects of Raising the Share income is the total outcome of the positive of Higher Education Graduates effect of the program itself, which raises Because the presence of higher education labor demand for college graduates, and the graduates raises the attractiveness and pro- positive (albeit small) effect of the greater ductivity of a community, local leaders might presence of college graduates. For less-­ be interested in attracting such individuals, educated individuals, two competing forces TABLE 5.2  Local Labor Demand, Brazil, 2010 Individual’s educational attainment Percent salary increase that employers are Less than Completed Completed Completed willing to pay in locations with: primary primary high school higher education An additional 1 percent density of workers with the −0.717 −0.795 −0.291 0.013a same educational level An additional 1 percent density of college graduates 0.613 0.831 0.376 n.a. Source: Calculations based on Fan and Timmins 2017. Data are from IPUMS. Note: a. Preference coefficients for the corresponding attribute that are not significantly different from zero. Results are for household heads age 25–35 years. H u m an C a p ita l in C ities   159 TABLE 5.3  The Effects of Raising Labor Demand for Higher Education Graduates in Feira de Santana, Brazil, 2010 Individual’s educational attainment Completed Less than Completed Completed higher Effects in Feira de Santana primary primary high school education Percent increase in population 13.43 11.43 9.69 17.20 Percent increase in income 0.67 4.83 3.36 50.20  Percent increase in income due to change in density of own −9.19 −9.39 −2.81 0.20 type of worker  Percent increase in income due to change in density of 9.86 14.22 6.17 n.a. higher education graduates Change in quality of life (expressed as a percent of income) 19.14 17.71 54.31 29.06 a  Due to change in population density −15.95 −12.85 −36.96 −22.26  Due to change in share of higher education graduatesa 59.58 49.95 147.18 85.40 b Total welfare change (expressed as a percent of income) 19.81 22.54 57.67 79.26 Source: Calculations based on Fan and Timmins 2017. Data are from IPUMS. Note: Estimates are for household heads age 25–35 years. Quality of life is the component of utility from a particular location that is common to all individuals of a given educational attainment. It is a function of location characteristics, net of housing prices. a. This represent the contribution of change in population density and change in share of higher education to the total change in quality of life. b. This represents the sum of percent increase in income and change in quality of life. are at play: the supply increase of individuals Feira de Santana as well as to those who of their skill level (which lowers their wages), move in (mainly from neighboring munici- and the positive spillovers from higher educa- palities), regardless of their educational tion graduates (which raises their wages). attainment. In contrast, individuals who stay The latter effect prevails. in the neighboring municipalities experience Moreover, thanks to the program, all indi- (net) negative effects. On the one hand, these viduals experience greater satisfaction with “stayers” benefit from lower density and their location (alternatively, they gain “qual- lower housing prices; but, on the other hand, ity of life”). Although Feira de Santana they lose quality of life and labor demand becomes denser and housing prices rise, it spillovers from the college share decline. also becomes more attractive given the Although the net effect of these forces is neg- greater presence of higher education gradu- ative, on average each of these municipalities ates, and the latter effect prevails. loses relatively little because individuals from The total welfare effect of the program is numerous neighboring municipalities move the net effect of income growth and of change to Feira de Santana. in quality of life. Because income grows and Through this simulation, the authors quality of life rises for all individuals, the illustrate how an area can benefit from pol- program raises welfare for all individuals. icies that raise demand for highly educated Furthermore, Fan and Timmins (2017) show individuals. An important message is that that if individuals did not value the presence increasing college share has the potential of of college-educated neighbors, quality of life raising both quality of life and incomes. would actually fall for all individuals. This It can raise quality of life because individu- would in turn result in lower population als of all skill levels enjoy having skilled growth because some of this growth is due to neighbors, either because they enjoy inter- the increased college share. acting with them or because their presence Thus, raising the demand for higher edu- is associated with a greater volume of urban cation graduates in Feira de Santana benefits amenities. It can raise incomes because not only those workers but also others. skilled individuals raise the productivity of Benefits accrue to the original residents of all others. 160   RAISING THE BAR Several caveats are in order. First, to con- higher education are substantially higher. Yet duct a full cost-benefit analysis of such a it is also possible that locational attributes simulation, one would need additional infor- may be less evenly distributed across a coun- mation, such as the fiscal cost of the pro- try’s areas in the LAC region than in the gram. Second, in principle the total effect of United States. For example, it is possible that this type of program (taking into account only a few areas offer good job opportunities both the positive effects on Feira de Santana to college graduates. and the negative effects on the neighboring Across LAC countries, differences in the municipalities) can be either positive or nega- share of urban population are mostly driven tive. This depends on whether the gains of by the unskilled. When living in large areas, college graduates to Feira de Santana out- the unskilled tend to work in low-productivity weigh the losses to other municipalities, services, such as wholesale, retail, hotels, and which in turn depends on the initial distribu- restaurants. To the extent that urban popula- tion of population and amenities (including tion shares continue to grow, fueled by the college share) across municipalities. Third, migration of unskilled workers to urban one must exercise caution if designing policies areas, the concern is that they will shift into favoring a specific location because these pol- low-productivity sectors. This will continue icies have a mixed track record. 26 Fourth, if the trend (chapter 1) of shifting workers into all locations implemented this type of policy low-productivity sectors. without increasing the country’s aggregate Our estimates show that individuals of all human capital, the country as a whole might skill levels prefer to live in areas with greater not gain. In particular, raising the share of shares of skilled people. The latter contrib- higher education graduates in each location ute not only to quality of life in an area but would ultimately require a nationwide also to workers’ productivity. On average, increase in college share. returns to aggregate human capital in the LAC region are large, of about the same size as private returns. HCEs account for at least Conclusions part of the returns to aggregate human capi- As in other regions of the world, in the LAC tal. The least-skilled individuals benefit the region, larger geographic areas attract more- most from an increase in college share, and skilled individuals. These individuals may individuals with the highest years of school- not be interested in area size (or density) in ing benefit the most from an increase in itself but rather in the amenities, jobs, and average years of schooling. college share typically found in large areas Because at least part of the estimated (as is the case in Brazil). Yet, by virtue of returns to aggregate human capital are due to attracting more skilled individuals, larger HCEs, it is efficient to enact policies that cor- areas are places with greater inequality. rect the market failure—for example, by sub- Migration patterns are part of this picture sidizing the formation or acquisition of because migrants to large areas are more skilled human capital in an area. This seems likely to be skilled than migrants to small particularly true for small areas, which tend areas. Large areas are also more likely than to have lower shares of skilled population small areas to develop their own human cap- and largely import their skilled human capi- ital than to “import it” from other areas. tal. Further research, however, is required to Although the positive association between quantify the exact size of HCEs and of the area size, education, and inequality has been optimal subsidy. documented for the United States as well If returns to aggregate human capital were (Behrens and Robert-Nicoud, 2015), we find solely due to complementarities between that it is stronger for LAC. This may be skilled and unskilled human capital, policy because a smaller share of the population intervention would not be required to correct in the LAC region is skilled, and returns to a market failure. Nonetheless, policy makers H u m an C a p ita l in C ities   161 might still want to enact policies to raise attract skilled human capital in the short aggregate human capital because of the posi- run by improving connectivity with other tive impact of skilled workers on the produc- areas, increasing the supply of amenities tivity and welfare of all workers. desired by skilled individuals, and raising Attracting skilled individuals to areas demand for skilled human capital. In the with low college shares may seem difficult, medium and long run, however, their best given that these individuals have a prefer- strategy might be to develop human capital ence for areas with high college shares. locally. As seen in this chapter, even keeping Yet, the evidence in this chapter indicates students in school for an extra year can that such areas (as well as others) can yield large returns. Annex 5A: Areas Used in the Stylized Facts Administrative Median 75th ­percentile Country Year unit Merged administrative units (thousands) (thousands) SEDLAC         Argentina 2014 Urban agglomerate n.a. 315 2,263 Bolivia 2011 Province Cercado and Quillacollo; Andres Ibañez and Warnes; 40 290 Ingavi and Murillo. Chile 2013 Province Some districts in Maramarga and Valparaiso; Santiago 154 1,159 Metropolitan area and some districts of Cahapoal. Colombia 2010 Municipalities Pereira, La Virginia and Dosquebradas; Cucuta, Los 26 637 Patios, El Zulia, and Villa del Rosario; Giron, Piedecuesta, Bucaramanga and Floridablanca; Soledad Barraquilla and Malambo; Palmira, Yumbo, and Cali; Valle del Aburra metropolitan area; Bogotá, Sibate, and Mosquera. Costa Rica 2010 Canton San José city covering selected districts from the follow- 27 316 ing provinces: Alajuela, Cartago, Heredia, and San José. Dominican 2014 Municipalities La Calena, Santiago de los Caballeros and Pedro García; 12 225 Republic National Districto and selected municipalities of San Cristobal and Santo Domingo. Ecuador 2012 Canton Guayaquil and Duran. 23 256 El Salvador 2014 Municipalities San Salvador Metropolitan area and Cuscatlan. 10 110 Guatemala 2014 Department n.a. 455 1,084 Honduras 2012 Municipalities San Pedro Sula and La Lima. 11 93 Mexico 2014 Municipalities Eighteen groupings in total. Examples: Tuxtla Gutierrez, 34 835 Berriozabal, and Chiapa de Corzo; Morelia and Tarimbaro; Ramos Arizpe and Saltillo; Distrito Federal and selected municipalities of Mexico and Hidalgo. Nicaragua 2005 Municipalities n.a. 23 125 Paraguay 2008 Municipalities Districts of Asuncion and additional municipalities from 13 202 Central Area (such as Limpio, Villa Elisa, Luque) Peru 2013 Province Metro area of Lima and Callao. 50 705 (continued) 162   RAISING THE BAR ANNEX 5A  Areas Used in the Stylized Facts (continued) Population Population Administrative threshold 1 threshold 2 Country Year unit Merged administrative units (thousands) (thousands) Uruguay 2011 Aggregated n.a. 12 159 city (“localidad agregada”) IPUMS         Brazil 2010 Municipalities n.a. 44 448 >20,000 inhabitants Colombia 2005 Municipalities n.a. 34 458 >20,000 inhabitants El Salvador 2007 Municipalities n.a. 36 95 >20,000 inhabitants Mexico 2010 Municipalities n.a. 13 145 Note: n.a. = not applicable, indicating countries in which there was no merge of administrative units. IPUMS = Integrated Public Use Microdata Series; SEDLAC = Socio-Economic Database for Latin America and the Caribbean. Annex 5B: Percentage of Employment in Services, by Educational Attainment 100 90 80 70 60 Percent 50 40 30 20 10 0 a lic zil y r ile ca ico a y ala r ua s a u do do ra ua a in bi i r liv Pe a ub Ri gu Ch g u em ex nt m Br lva ua ug ra Bo nd sta ep ra lo ge M ca Ec at Ur Sa Ho Pa Co nR Co Ar Ni Gu El ica in m Do Primary education Secondary education Tertiary (higher) education Source: Calculations based on Socio-Economic Database for Latin America and the Caribbean (for countries other than Brazil) and IPUMS (for Brazil). Note: Figure refers to workers in the adult population (age 25–64 years). The figure shows the percentage of individuals of each educational attainment who are employed in services. For the definition of educational attainments, see the “Some Stylized Facts” section earlier in this chapter. H u m an C a p ita l in C ities   163 Annex 5C: Probability of Working in the Service Sector for Skilled and Unskilled Workers, by Area Size 100 90 80 70 60 Percent 50 40 30 20 10 0 Unskilled Skilled Unskilled Skilled Unskilled Skilled Small Medium Large Source: Calculations based on SEDLAC (for countries other than Brazil) and IPUMS (for Brazil). Note: For areas of a given size, the figure shows the percentage of workers employed in services for skilled and unskilled workers. The figure refers to workers in the adult population (age 25–64 years). Skilled workers have at least some higher education. IPUMS = Integrated Public Use Microdata Series; SEDLAC = Socio-Economic Database for Latin America and the Caribbean. Annex 5D: Percentage of Urban Population Born Abroad 14 12 10 8 Percent 6 4 2 0 ca a a ay lic y r ia ile a r ua ico ru il a iti do do ua az am tin aic bi liv Ha Pe ub Ri gu Ch ag ex m Br ua lva ug en m Bo n sta ep ra r lo M Pa ca Ec Ja Ur Sa g Pa Co nR Co Ar Ni El ica in m Do Source: Calculations using IPUMS. Note: The figure shows, for each country, the fraction of individuals who were born abroad among those who are classified as urban by national statistics offices. Because households are not classified as urban or rural in Argentina or Uruguay, figures for Argentina and Uruguay are for total population. IPUMS = Integrated Public Use Microdata Series. 164   RAISING THE BAR Notes 9. Percent of skilled population is calculated relative to the population age 25-65 years in 1. In this chapter, “skilled” individuals are those each country. Sources for LAC: SEDLAC for with postsecondary education. More specifi- all countries other than Brazil; IPUMS for cally, the term comprises individuals with Brazil. Source for the United States: U.S. completed or unfinished higher education as Census Bureau, Current Population Survey discussed in the section on stylized facts and 2010. Returns to higher education in the individuals with completed higher education LAC region are from Ferreyra et al. (2017). as discussed in the subsequent two sections. Returns to higher education in the United 2. The definition of a “city” is in the “Some States are based on Card (2001) and Stylized Facts” section. Heckman et al. (2006). 3. Of course, this positive effect could also arise 1 0. Comparator countries include Indonesia, if skilled workers were complements among Kyrgyz Republic, Malaysia, Slovenia, themselves. Following the literature (Moretti Thailand, and Ukraine. Source: IPUMS. In 2004a; Guo, Roys, and Seshadri 2016; selecting comparators, we apply a different Ciccone and Peri 2006), we assume that criterion from that in chapter 2. Following workers of different skill levels are comple- other World Bank studies (such as Ferreyra ments and that workers of the same skill level et al. 2017), our comparators for LAC are are substitutes. developing countries from EAP and ECA with 4. For each country, we calculate the average information on international migration at the (over cities) of cities’ average years of area level in IPUMS. schooling, and cities’ share of college-­ educated 11. In the data, recent migration can be measured by workers. whether an individual currently resides in a dif- 5. Specifically, we use level-2 administrative ferent place from five years ago, in which case units. We follow the same criteria as in she or he has moved some time during the past chapter 6 on the merging of administrative ­ five years. Following related work (Bayer, units (see table in annex 5A for details). Kehoe, and Timmins 2009; Lall, Timmins, and 6. We adopt these cut-offs because the distribu- Yu 2009), we focus on the 25–35-year-old group tion of area size is highly skewed to the right. to capture own migration decisions (as opposed In other words, most areas are small, and a to one’s parents), during the years in which indi- few areas are large. With these thresholds, viduals are most mobile (because they are less the group of “small” areas comprises a large likely to migrate once they start a family). number of areas, and the group of “large” 12. Migrants might actually become skilled at areas comprises a small number of areas. The their destination. Although we have no data groups are not equally sized, but they are rel- to assess this possibility, Ferreyra et al. (2017) atively homogeneous on area size. To facili- document that 80 percent of higher education tate comparisons with the United States students live with their parents during college, (Moretti 2004a; Behrens and Robert-Nicoud in which case they most likely do not move 2015), we define groups by population for college. instead of by population density. Annex 5A 13. For example, in Colombia all large areas have lists the population thresholds used to build at least one higher education institution, only the area groups. 30 percent of medium-sized areas have one, 7. This value is the coefficient of the regression and virtually no small areas have one. of log area share of skilled individuals on log 14. The share of individuals employed in services area population, pooling data for all areas and averages 55 percent, 74 percent, and 87 percent countries. When country fixed effects are among those with primary, secondary, and included, the coefficient is equal to 0.28. higher education. Both coefficients are significantly different 15. We measure labor productivity as the ratio from zero. between value added and employment, using 8. This is the average of country-specific elastici- the GGDC 10 sector database for 2011 for the ties, estimated separately by country. When LAC countries with available data (Argentina, pooling data for all countries, the estimated Brazil, Costa Rica, and Peru). We consider elasticity is 0.029; if country fixed effects are agriculture and mining as nonurban sectors. included in this regression, the estimate is 16. These estimates arise from the first stage of 0.042. the estimation of the determinants of city H u m an C a p ita l in C ities   165 productivity (see chapter 3). Returns to increase in their share would drive down their education are estimated as the coefficient on wages, holding other things constant, by virtue years of schooling in the regression of log of increasing their relative supply. wages on years of schooling, age, age squared, 23. It is possible, however, that the high returns to gender, marital status, and an area fixed effect. an additional year of schooling reflect exter- Because a separate regression is run for each nalities arising from crime reduction. To the country, we obtain returns to schooling extent that people are more productive in for each country. These range from about safer places (see chapter 3), this might give rise 6 percent in the Dominican Republic, to HCEs and be reflected in the returns to Nicaragua, and Peru, to about 11 percent in average years of schooling. Brazil and Uruguay; across countries, their 24. The estimation of these spillovers bears simi- average is 8.92 percent. larities to that of the estimation of returns to 17. Box 3.3 documents that this is indeed the case aggregate human capital, though it is not in Colombia, based on Balat and Casas exactly the same. (2017). 25. These municipalities, with 97 percent of 18. As in Ferreyra et al. (2017), higher education Brazil’s population, are those with data. comprises both short-cycle and bachelor’s 26. See, for example, Neumark and Simpson programs, akin to associate and bachelor’s (2015). In the same spirit, the World Bank’s programs, respectively, in the United States. 2009 World Development Report argues in 19. The U.S. average is for individuals age 25 and favor of spatially blind policies and mainly older. The LAC average is for individuals against spatially targeted policies except age 14–65; it is the average of the area college when countries are fragmented for linguis- shares for the areas (and years) included in tic, political, religious, or ethnic reasons the regressions that estimate returns to aggre- (World Bank 2009). gate human capital. 20. However, returns to aggregate human capital also appear relatively low in the United States References when aggregate human capital is measured Balat, J., and C. Casas. 2017. “Firm Productivity by the share of higher education graduates. a nd C it ie s: T he C a s e of C olombi a .” Although the average city share of higher edu- Background paper for this book. World Bank, cation graduates is larger in the United States Washington, DC. (23 percent), it is still quite low. It is possible Bayer, P., N. Keohane, and C. Timmins. 2009. that returns to the share of higher education “Migration and Hedonic Valuation: The Case graduates are relatively low for the observed of Air Quality.” Journal of Environmental range of this share, but might be higher for Economics and Management 58 (1): 1–14. higher ranges. Behrens, K., and F. Robert-Nicoud. 2015. 21. To obtain country-level returns, for each “Agglomeration Theory with Heterogenous country we regress area-level productivity on Agents.” In Handbook of Regional and Urban the corresponding measure of aggregate Economics, Volume 5, edited by Gilles human capital; these regressions control for Duranton, J. Vernon Henderson, and William area density, market access, air temperature, C. Strange, 171–87. Amsterdam: Elsevier. terrain ruggedness, and precipitation. To esti- Card, D. 2001. “Estimating the Return to mate area-level productivities, for each coun- Schooling: Progress in Some Persistent try we regress log wages on individual Econometric Problems.” Econometrica 69 (5): characteristics (age, age squared, years of 1127–60. schooling, gender, and marital status) and Ciccone, A., and G. Peri. 2006. “Identifying year fixed effects. We do not run these regres- Human-Capital Externalities: Theory with sions for Argentina, Panama, or Uruguay Applications.” Review of Economic Studies because they have few areas. 73 (2): 381–412. 22. This net effect might be positive as well because Combes, P., and L. Gobillon. 2015. “The Empirics of complementarities among skilled workers. of Agglomeration Economies.” In Handbook We follow Moretti (2004a, 2004b), Ciccone of Regional and Urban Economics, Volume 5, and Peri (2006), and Combes and Gobillon ed ited by Gi l les Du ra nton , J. Ver non (2015) in assuming that skilled workers are Henderson, and William C. Strange, 247–348. substitutes among themselves and that an Amsterdam: Elsevier. 166   RAISING THE BAR Duranton, G. 2014. “Growing through Cities in Montenegro, C. E., and H. A. Patrinos. 2014. Developing Countries.” World Bank Research “Compa rable E sti mates of Ret u r ns to Observer 30 (1): 39–73. Schooling around the World.” Policy Research Fan, L., and C. Timmins. 2017. “A Sorting Model Wo r k i n g P a p e r 7 0 2 0 , Wo r l d B a n k , Approach to Valuing Urban Amenities in Brazil.” Washington, DC. Background paper for this book. World Bank, Moretti, E. 2004a. “Estimating the Social Return to Washington, DC. Higher Education: Evidence from Longitudinal Ferreyra, M. M., C . Avitabile, J. Botero, and Repeated Cross-Sectional Data.” Journal of F. Haimovich, and S. Urzua. 2017. At a Econometrics 121 (1–2): 175–212 Crossroads: Higher Education in Latin ———. 2004b. “Human Capital Externalities in America and the Caribbean. Washington, DC: Cities.” In Handbook of Regional and Urban World Bank. Economic s, Volume 4 , edited by J. V. Guo, J., N. Roys, and A. Seshadri. 2016. Henderson and J. F. T hisse, 2243 –91. “Estimating Aggregate Human Capital Amsterdam: Elsevier. Externalities.” Working Paper, University of Neumark, D., and H. Simpson. 2015. “Place- Wisconsin–Madison. Based Policies.” In Handbook of Regional Heckman, J., L. Lochner, and P. Todd. 2006. and Urban Economics, Volume 5, edited by “Earnings Functions, Rates of Return, and Gilles Duranton, J. Vernon Henderson, and Treatment Effects: The Mincer Equation and William C. Strange, 1197–1287. Amsterdam: Beyond.” In Handbook of the Economics of Elsevier. E duc ation , Volume 1, edited by E . A. Quintero, L., and M. Roberts. 2017. “Explaining Hanushek. New York: Elsevier. Spatial Variations in Productivity: Evidence Lall, S. V., C. Timmins, and S. Yu. 2009. from 16 LAC Countries.” Background paper “Connecting Lagging and Leading Regions: for this book, World Bank, Washington, DC. The Role of Labor Mobility.”  Brookings- World Bank. 2009. World Development Report: Wharton Papers on Urban Affairs 2009 (1): Reshaping Economic Geography. Washington, 151–74. DC: World Bank. Urban Form, Institutional Fragmentation, and 6 Metropolitan Coordination Nancy Lozano-Gracia and Paula Restrepo Cadavid Introduction agglomeration economies are also at the core This chapter attempts to explain the effect of of the links between the spatial aspects of urban form and institutional structure on pro- urban form and city productivity. Cities can ductivity. Urban form has multiple spatial use a given area of land and space in very dif- dimensions, such as the geometric shape of a ferent ways. Such differences are closely linked city’s urban extent; the internal structure of to the way transport systems are designed, the the city as determined, for example, by its transport modes used (private or public), com- transport network; and the land use patterns muting times, matching between workers and as reflected through the spatial distribution of firms, how firms interact with each other, and population and buildings within a city. This the type and intensity of human interaction. approach goes beyond the economic litera- In the economics literature, urban form ture’s frequent focus on a single dimension of has been linked to economic performance urban form: density. This chapter also (Parr 1979; Ciccone and Hall 1996), sustain- explores an institutional aspect of urban form, ability (Breheny 1992; De Roo and Miller namely the fragmentation of governance in 2000), quality of life (Squires 2002), commut- large metropolitan areas and concomitant ing costs (Wheeler 2001), and knowledge attempts at metropolitan coordination. spillovers through human interactions (Lynch As with chapters 2 through 5, we focus on 1981; Jaffe, Trajtenberg, and Henderson city-level productivity measures and intro- 1993; Glaeser 1998). Overall, denser cities are duce identification methods that aim to thought to improve labor productivity assess the links between urban form and a through better matching of firms and workers city’s institutional structure, on the one hand, and enhanced interactions that facilitate the and city-level productivity on the other. spread of tacit knowledge, both of which are The same channels—sharing, matching, thought to occur more easily the closer people and learning—identified by Duranton and and firms are to each other (see Ciccone and Puga (2004) to explain the emergence of Hall 1996; Rosenthal and Strange 2004; This chapter is based on background papers by Duque et al. (2017a), Duque et al. (2017b), and Duque et al. (2017c). The authors thank Grace Cineas, Jane Park, and Wilson A. Velasquez for excellent research assistance provided for the work on this chapter. 167 168   RAISING THE BAR Cervero 2001). Furthermore, recent work has • Beyond density, other spatial dimensions shown that density of employment and popu- of urban form matter for productivity. lation can lift innovation and overall metro- Smooth, rounded, compact, and politan productivity. All else equal, the internally well-connected cities tend number of inventions (measured as patents) to have higher productivity levels than per capita is about 20 percent higher in a met- rugged or elongated cities, or cities with ropolitan area that is twice as dense—with poorly connected streets. density measured as employment density—as •  Large metropolitan areas in the LAC another metropolitan area in the United region comprise, on average, just over States (Carlino, Chatterjee, and Hunt 2007). nine administrative units. Half of them Metropolitan sprawl is also associated with have a metropolitan governance body. lower average labor productivity (Fallah, The fragmentation levels observed in Partridge, and Olfert 2011). the region are detrimental for produc- However, although the links for the relation- tivity. However, unlike what Ahrend ship between density and productivity are well et al. (2014b) find for Organisation for established in the developed world, little Economic Co-operation and Develop- research has been done for developing coun- ment (OECD) countries, we find no tries, including, prior to this book, for countries evidence that the presence of a gover- in Latin America and the Caribbean (LAC) (see nance body at the metropolitan level chapter 3). Although such links have been estab- mitigates the negative effects of frag- lished between density and productivity, much mentation. This may point to ineffec- less is known about the links between other spa- tive governance arrangements or insti- tial dimensions of urban form, and productivity. tutions that do not effectively support Recent steps in this direction are presented in interjurisdictional coordination. the works of Harari (2016) and Tewari, Alder, and Roberts (2016). The former focuses on the geometry of urban extents of over 450 Indian Urban Form and Productivity cities and finds that more compact cities, with Measuring Urban Form an urban geometry conducive to shorter poten- tial within-city trips, are characterized by larger In this section, we focus on spatial aspects of populations, lower wages, and higher housing urban form; in the next section, we focus on rents. These findings suggest that a city’s resi- its institutional aspects. Hence, “urban form” dents value compactness as a consumption in this section refers to spatial urban form. amenity. By contrast, firms do not appear to be Economists have commonly focused on directly affected by city shape in their location only one dimension of urban form: population choices, and no evidence is found of a signifi- density. In the economics literature, the com- cant effect on the productivity of firms for that mon conclusion is that less dense cities face subset of Indian cities. In their analysis of urban higher commuting rates (Wheeler 2001) and development patterns of Indian cities, Tewari, have lower knowledge spillovers (Lynch 1981; Alder, and Roberts (2016) find a robust and Jaffe, Trajtenberg, and Henderson 1993; positive relationship between a city’s initial com- Glaeser 1998), negatively affecting a city’s pro- pactness and its subsequent economic growth, ductivity levels. Some authors argue that estimated using nighttime lights data. improved highways, public transit services The chapter’s main findings are as follows: (Glaeser and Khan 2004; Chatman and Noland 2014), and advances in communication •  Although the average LAC city is rounded, technologies (Partridge et al. 2009) have helped has smooth borders (perimeters), has a reduce the productivity costs of sprawling.1 dense street network, and tends to be Density alone, however, does not describe compactly built, the region’s cities show a all the multiple dimensions of urban great diversity of urban form. form. Because it is measured as an average, U rban F o r m , I nstituti o na l F rag m entati o n , an d Metr o p o l itan C o o r d inati o n   169 it does not capture how density varies over A roundness index can be calculated to space within cities as a result of the interwo- measure the degree to which the shape of an ven decisions of individuals, firms, and gov- urban area deviates from its equal-area circle. ernment on where to live, locate, or build It is calculated as the share of the total area infrastructure—often affecting, for example, of the urban extent that is inside the equal- street layout and land use patterns. area circle about its center of gravity. 3 The The urban planning and geography litera- roundness index equals 1 for a perfect circle. ture has long discussed different ways of As the index moves toward 0, the urban area characterizing urban form. For Batty and becomes more irregular and less compact (see Longley (1994), for example, it has many annex 6B).4 For example, for all forms shown dimensions because it includes all elements in the first row of table 6.1, the shape shown that form the spatial layout of cities, such as in column a will have a roundness value streets, buildings, or open spaces. closer to 1, whereas that in column c will Cities may be characterized in three key, have a value closer to zero. interrelated, dimensions of urban form: The smoothness of the perimeter provides another way to measure a city’s compactness 1. The border’s shape and perimeter (Harari 2016), resting on the fact that, among (Angel, Parent, and Civco 2010a) all shapes of a given area, the circle has the 2. The internal structure of the urban area minimal length of contact with its periphery. 3. The land use patterns observed within In a walled city, for example, looking at the city boundaries and reflecting the use smoothness of the perimeter would be a nat- of space and the distribution of popu- ural measure of its compactness, all else lation within the city (see Whyte 1968; being equal (Angel, Parent, and Civco 2010b). Batty and Longley 1994; Batty 2008; A smoothness index can be calculated as Prosperi, Moudon, and Claessens 2009) the ratio of the perimeter of the equal-area These dimensions link to the efficiency of circle and the perimeter of the shape (Angel, city transport, the cost of providing urban Parent, and Civco 2010a). A smoothness infrastructure and services, and environmen- index equal to 1 indicates a totally smooth tal sustainability. perimeter found in a perfect circle. A smooth- ness index close to 0 indicates a highly irregu- Border shape and perimeter lar perimeter, which is very common in cities Whether the border of a city is shaped as a that have grown unplanned or are in rugged circle or as a tentacle-like shape has implica- ­topography. 5 The shapes in the smoothness tions for trip lengths, as does the smoothness index row of table 6.1 provide examples of of its perimeter. varying levels of smoothness. The shape in A perfect circle—the most compact geo- column a will have a value closer to 1 com- metric shape—has geometric properties such pared with those in columns b and c. as a minimum surface area and a maximum accessibility from and to any interior point Internal structure of the city (see Thompson 1952; Angel, Parent, and The internal structure of the city affects the Civco 2010b). A circular form can reduce trip way people and products move within a city. lengths and increase accessibility compared A key element in this structure relates to two with an elongated form.2 Better accessibility aspects of its connective infrastructure: the improves matching between workers and structure given by the layout of the road net- jobs, consumers and goods, and firms and work in the city and the degree to which all output markets, affecting productivity. Cities segments in the network are interconnected. with compact and circular shapes also have Figures shown under the “Internal structure” lower costs per capita in providing basic row in table 6.1 provide an example of street infrastructure, which benefits from econo- networks with different layouts and levels or mies of density (Litman 2015). connectedness of the network segments. 170   RAISING THE BAR TABLE 6.1  Examples of Urban Areas with High, Medium, and Low Values of the Indexes That Describe Urban Form High Medium Low (a) (b) (c) Shape and Roundness Ambato (peru) Guaratinguetá (Brazil) Castries (Saint Lucia) perimeter index 0.92 0.53 0.41 Smoothness Feira de Santana (Brazil) Mérida (Venezuela, RB) León (Mexico) index 0.79 0.51 0.27 Internal Circuity Caracas (Venezuela, RB) 1.18 San Jose (Costa Rica) 1.07 Sorriso (Brazil) 1.00 structure Intersection Cap-Haitien (Haiti) 127.43 Belo Jardim (Brazil) 73.80 Guayama (Puerto Rico) 20.05 density Street Morelia 20.66921 Paysandu (Uruguay) 10037.48 Tuxtepec (Mexico) 5,011.50 (Mexico) density (continued) U rban F o r m , I nstituti o na l F rag m entati o n , an d Metr o p o l itan C o o r d inati o n   171 TABLE 6.1  Examples of Urban Areas with High, Medium, and Low Values of the Indexes That Describe Urban Form (continued) High Medium Low (a) (b) (c) Land use Sprawl Ciudad Bolívar (Venezuela, RB) Barcelona (Venezuela, RB) Kingston (Jamaica) 0.22 Fullness Kingston (Jamaica) Maturin (Venezuela, RB) Antofagasta (Chile) 0.14 index 0.92 0.93 Source: Duque et al. 2017a. Note: Roundness and smoothness images represent the urban area. Circuity, street density, and intersection density show the layout of road networks in an urban area. Fullness pixels represent density of built-up area and sprawl pixels represent population density. See annex 6C for correlation matrix between urban form indicators. Cities where the road network has grown To assess the degree of connectivity for all unplanned—such as table 6.1, third row segments, we can use measures of intersec- (“Circuity”), column a—are usually associ- tion density and street density as indicators of ated with longer commuting times and lower a city’s internal structure. Both metrics can accessibility indicators than cities where the provide information on the ease of movement road network follows a grid pattern—such as within a city because circular roads and net- table 6.1, third row, column c (see Boeing works with many dead-end streets make it 2017). Regular urban structures and harder to reach all points in a city (see Boeing high-density street networks are associated 2017). High values of these two measures are with more efficient, shorter, and cheaper associated with high walking rates and an trips, which reduce congestion costs and increased use of nonmotorized modes allow for nonmotorized modes of transport (Cervero and Kockelman 1997). (see Mills and Hamilton 1989; Bogart 1998; For this work, we calculate two indexes to Bertaud 2004; Giacomin and Levinson 2015; reflect the internal structure of cities: Huang and Levinson 2015; Cervero and Kockelman 1997). Reducing the costs of 1. Circuity of the road network . This interaction through bet ter- connected measure indicates how circular the networks and providing denser intersections street network’s layout is. It is cal- can potentially improve matching and learn- culated as the average ratio between ing in cities. the lengths of each segment and the 172   RAISING THE BAR straight-line distance between the two of the population (sprawl) and of the built-up nodes it links (Boeing 2017). The cir- areas (fullness of the form) within a city’s cuity value is equal to 1 when all the boundaries (table 6.1): streets in the network are straight and 1. Sprawl. The population distribution greater than 1 when the street network within city boundaries gives an indica- has curved roads. tion of how land is used in a city (Fallah, 2. Intersection density and street density. Partridge, and Olfert 2011). A sprawl These indexes assess the degree of index that measures the degree of even- connectivity of the road network. ness in that distribution can provide a Intersection density is calculated as the good measure of land use within cities. number of nodes divided by the area It takes a value close to 1 when the pop- the network covers, considering only the ulation is highly concentrated in a por- set of nodes with more than one street tion of the urban area. The sprawl index emanating from them and thus excluding is calculated as the normalized differ- streets with a dead end (Boeing 2017). ence between the share of areas with Street density is calculated as the sum of population density below the regional, all segments of the street network (in km) or LAC, average density and the share in the undirected representation of the of areas with population density above street network, divided by the area of the that (Fallah, Partridge, and Olfert city in square kilometers (Boeing 2017).6 2011).8 The population counts for each area within the city were retrieved from Land use patterns high-resolution population grid layers How cities distribute and organize land can for 1990, 2000, and 2015 within the affect the way firms and households in a city derived urban extents.9 interact. Cities can grow by sprawling, with 2. Fullness of the form. This indicator population locating in patches of land that measures the presence of built-up areas leapfrog through empty spaces, such as within the urban extent as a fraction Ciudad Bolivar in República Bolivariana de of the total area. A fuller city where Venezuela—table 6.1, sixth row, column a. built-up area is denser, as shown in col- Sprawling cities are inefficient in providing umn a of the last row of table 6.1, may infrastructure and public services because be conducive to more interactions; how- the per-unit cost of development increases ever, a high fullness can also suggest a with sprawl (Knaap and Nelson 1992; city with little open space, which may Knaap, Ding, and Hopkins 2001; Fallah, undermine productivity.10 A fullness Partridge, and Olfert 2011). Having less index equals 1 for a city where all land sprawl is conducive to lower commuting is fully built up; an index close to 0 is times, easier interactions, and higher pro- indicative of a city with many unbuilt ductivity (Wheeler 2001). Land use interven- areas within its boundaries. tions that contribute to the colocation of residences and jobs have the potential to increase employment accessibility (see Avner Measuring Productivity and Lall 2016; Quirós and Mehndiratta 2015). However, some separation of land One of the key challenges in studying urban uses may be desirable because allowing for form, and the relationship between such form the colocation of firms potentially leads to and productivity, is the lack of comparable agglomeration economies. Given the diffi- data across countries and over time. As seen culties in accessing information on effective in part I of this book, national definitions of land uses and the distribution of jobs for a urban areas vary, often dramatically, across broad set of cities,7 we focus on two indica- countries; within a country, cities’ adminis- tors of land use that look at the distribution trative boundaries seldom conform to U rban F o r m , I nstituti o na l F rag m entati o n , an d Metr o p o l itan C o o r d inati o n   173 the actual extent of a city. And, although and their shapes drawn using the areas out- cross-sectional analysis measured productiv- lined from NTL imagery for the three years ity at city level in chapters 2 and 3, productiv- (box 6.1). Further, similar to Tewari, Alder, ity is harder to measure over time; and that and Roberts (2016) and chapter 2 of this measurement over time is needed to tease out book, we used an aggregate measure of lumi- the links between cities’ form and their pro- nosity extracted from NTL data to calculate ductivity. Recent advances in Geographic an estimate for city output (Y ) per square I nformation S cience (GIS cience) and kilometers, which we take as a proxy mea- Computational Geometry provide some valu- sure of productivity.12 Specifically, the sum of able methods for our purposes. luminosity within the defined city boundaries In a first attempt to provide a comprehen- extracted (see box 6.1) is calculated and then sive characterization of urban form in the divided by the area in square kilometers.13 LAC region over time, we use nighttime lights (NTL) imagery, for 1996, 2000, and A Variety of Urban Forms in the LAC 2010, to identify all urban areas in the region, Region outline their borders, and extract indicators of their form.11 This effort allows us to pro- Using the above indexes we now characterize vide a standardized characterization of the urban form of LAC cities along the three key urban form of 919 LAC cities. dimensions: shape, internal structure, and As a starting point, all cities with more land use. The shape indicators for 2010 gener- than 50,000 people in 2010 were identified ally indicate that LAC cities are more rounded BOX 6.1  Outlining Urban Extents Using Nighttime Lights Two nighttime lights (NTL) products are available tion of city boundaries and later the aggregation of from the Defense Meteorological Satellite Program luminosity levels within such boundaries, several nighttime lights Operational Linescan System corrections are needed. First, the literature recog- (DMSP-OLS) for the years included in this a ­ nalysis: nizes the problem of overglow in DMSP-OLS, which the so-called “ordinary” product (the NTL) and is the effect of light spilling beyond boundaries— the radiance-calibrated product. For this work, we for example, light from coastal cities appearing up use radiance-calibrated yearly composites for 1996, to 50 km out to sea (Croft 1978; Wu et al. 2014). a 2000, and 2010 from the National Centers for To accurately allocate light intensity to a city and Environmental Information of the National Oceanic more accurately outline the form of cities, a cor- and Atmospheric Administration to delineate urban rection for overglow is necessary. For this, we con- extents. We chose the radiance-calibrated product ducted a deblurring process by restacking the light over the ordinary product because the radiance-­ on its source pixels (Abrahams, Lozano Gracia and calibrated data correct for saturation issues found Oram 2016). Yearly composites are not comparable in the NTL product, avoiding underestimation of because of the lack of the sensors’ onboard calibra- total light for the largest, and hence brightest, cit- tion, so an additional intercalibration is necessary ies (Zhang, Schaaf, and Seto 2013). Previous work for a multitemporal analysis of urban form and city also suggests that the radiance-calibrated product is productivity in the LAC region (Cao et al. 2016; Hsu a better proxy than the ordinary product for socio- et al. 2015; Pandey, Joshi, and Seto 2013; Zhang and economic variables (Hsu et al. 2015; Ma et al. 2014). Seto 2011). Following Duque et al. (2017a), we used The composites used for this work have a spatial res- a threshold approach to delineate urban extents each olution of 30 arc-seconds (about 1 km at the equator). year for LAC cities with more than 50,000 inhabi- Before the radiance-calibrated data can be used tants in 2010. Map B6.1.1 shows three examples of for analysis at the city level, including the delinea- the urban extents obtained using the NTL for 2010. (continued) 174   RAISING THE BAR BOX 6.1  Outlining Urban Extents Using Nighttime Lights (continued) MAP B6.1.1  Examples of Urban Extents over the DMSP-OLS Radiance-Calibrated 2010 Composite a. Bogotá, Colombia b. Santiago de Chile, c. San José de Costa Rica, Chile Costa Rica 0 10 Km Source: Elaboration based on Duque et al. 2017b. Note: All maps are at the same spatial scale. The yellow line shows the urban boundary. Lighter- shaded (whiter) areas are those that have greater nighttime lights intensity. Note: The radiance-calibrated product is used for the analysis, but in the text we generally refer to radiance-calibrated as nighttime lights. a. These are problems not present in the data from VIIRS sensors, but those data are constrained by their short period. than elongated, have urban perimeters that Sellers (2007), who also find that cities in are more smoothed than complex, and have Asia and LAC have very dense city structures little open space inside city boundaries. relative to cities like Dallas in the United The degree of roundness in LAC cities tends States and Sydney in Australia. In terms of to be high or close to 1, with median and mean circuity, the typical LAC city, according to values above 0.5 in the three years (table 6.2).14 our findings, looks like Bogotá or Mexico A slight decrease is observed between 1996 City. A small increasing trend is observed in and 2010, indicating a weak trend toward less circuity values from 1996 to 2010, which round and more elongated urban extents.15 might be due to recent growth of settlements Smoothness of the city’s perimeter is also in more rugged terrain.16 observed to decrease slightly over time, from The sprawl indicator suggests that LAC 0.67 to 0.64, indicating a trend toward less cities do not face high degrees of sprawl but smoothed urban perimeters, which could have grown with compact patterns and high reflect urban growth along corridors linking density overall, with the mean for 2010 being other cities. 0.575. Huang, Lu, and Sellers (2007), find Internal structure indicators suggest that that LAC cities are among the world’s least most LAC cities follow regular patterns that sprawling cities. This does not seem to have resemble a grid, with high values of intersec- changed much over time, with only a slight tion density and street density and with cir- decrease from the 1996 value of 0.598. cuity levels very close to 1. The LAC region’s Completing the picture of urban form, full- average circuity value in 2010 was 1.04 ness values are high overall with an average (a slight increase from 1996), which suggests of 0.602 in 2010, showing a small increase that street networks for most LAC cities are during the period. These results are consistent about only 4 percent longer than if they were with previous findings (Angel, Parent, and all composed of straight lines. This regularity Civco 2010a; Huang, Lu, and Sellers (2007); is in line with the findings by Huang, Lu, and Inostroza, Baur, and Csaplovics 2013).17 U rban F o r m , I nstituti o na l F rag m entati o n , an d Metr o p o l itan C o o r d inati o n   175 TABLE 6.2  Descriptive Statistics of Urban Form in LAC Cities Variable Year p25 Median p75 Mean SD Min Max Shape and perimeter Roundness 1996 0.725 0.828 0.879 0.787 0.121 0.266 0.947 2010 0.712 0.782 0.877 0.782 0.121 0.35 0.952 Smoothness 1996 0.620 0.721 0.767 0.674 0.14 0.09 0.888 2010 0.567 0.7 0.755 0.644 0.146 0.16 0.856 Internal structure Intersection density 1996 46.96 64.9 84.13 65.87 28.88 0.33 184.97 2010 36.02 51.26 66.59 51.93 23.51 0.2 148.01 Street density 1996 7,988.80 10,452.40 12,985.50 10,474.40 3,791.50 135.9 20,669.20 2010 6,428.30 8,643.80 10,773.70 8,582.40 3,249.50 87.1 19,040.00 Circuity 1996 1.019 1.028 1.046 1.037 0.034 1.002 1.419 2010 1.021 1.032 1.05 1.04 0.029 1.004 1.241 Land use Sprawl 1996 0.479 0.595 0.721 0.598 0.177 0.108 1 2010 0.475 0.583 0.677 0.575 0.148 0.074 1 Fullness 1996 0.494 0.63 0.763 0.618 0.192 0.014 0.996 2010 0.482 0.623 0.739 0.602 0.181 0.028 0.993 Source: Calculations based on Duque et al. 2017a. Note: Roundness ranges from 0 to 1. A high roundness value approximates a circle; a low roundness value approximates an irregular shape. Smoothness ranges from 0 to 1. A high smoothness value indicates a smooth perimeter; a low smoothness value indicates a highly irregular perimeter. Intersection density can take on values of 0 or greater. A high intersection density value indicates a higher concentration of intersecting streets given the city area; lower intersection density indicates a lower concentration of intersecting streets given the city area. Street density can take on values of 0 or greater. A high intersection density value indicates a higher number of streets; a low intersection density value indicates a lower number of streets. Circuity can take on values of 1 or greater. A circuity value greater than 1 indicates a street network that is not straight (curvy); a circuity value of 1 or close to 1 indicates a street network that is straight. Sprawl ranges from 0 to 1. A high sprawl value indicates a population that is evenly distributed; a low sprawl value indicates a population that is concentrated. Fullness ranges from 0 to 1. A high fullness value indicates a city that is compact and built up with minimal empty spaces; a low fullness value indicates a city that is sprawling with empty spaces. See annex 6C for correlation matrix between urban form variables. LAC = Latin America and the Caribbean; SD = standard deviation. Figure 6.1 shows three indicators, one for like São Paulo and Puebla appear to have each dimension of urban form analyzed. densely built urban forms whereas cities The three panels illustrate the large variabil- like Brasilia and Cali tend to have a higher ity in the form of LAC cities. Some cities proportion of open spaces within their like Santa Cruz (Bolivia) or Puebla (Mexico) urban areas. have high smoothness values, whereas cities Despite large regional variability in urban like Medellín and São Paolo have low form, little change was seen between 1996 smoothness values (panel a). and 2010 in roundness, smoothness, circu- The indicators for internal structure ity, fullness, and sprawl, which changed less also show variability. Cities like La Paz, than 5 percent between 1996 and 2010 (fig- Santiago de Chile, and Lima have rela- ure 6.2). Such small changes stress the per- tively high street density values whereas sistence of urban form, highlighted in the cities in Central A merica such as San literature and chapters 1 and 2. Salvador (El Salvador) and Panama City The indicators that changed most over the (Pa na ma) show relat ively low values period of analysis are intersection density (panel b). The average LAC city seems to and street density, both decreasing by nearly have high fullness, with the average full- 20 percent, suggesting that LAC cities have ness index just above 0.6 (panel c). Cities b e c om e le s s c on ne c t e d on avera ge , 176   RAISING THE BAR FIGURE 6.1  Urban Form in Latin America and the Caribbean Shows Great Variability a. Smoothness more smooth 1.0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 less smooth 0.1 0 z la i ín lo os ru eb au ell ot aC ed pu op sp nt m Sã ui Sa nl Sa median value of the indicator position in the distribution of the average city b. Street density 19,000 High street density 17,000 15,000 13,000 11,000 low street 9,000 density 7,000 5,000 3,000 1,000 az Sa a o eio r ity do lim iag p aC ac lva la nt m m Sa na n pa Sa median value of the indicator position in the distribution of the average city (continued) U rban F o r m , I nstituti o na l F rag m entati o n , an d Metr o p o l itan C o o r d inati o n   177 FIGURE 6.1  Urban Form in Latin America and the Caribbean Shows Great Variability (continued) c. Fullness more full 1.0 0.9 0.8 0.7 0.6 less full 0.5 0.4 0.3 0.2 0.1 0 lo la ty lam ilia li Ca eb au Ci as Be pu op ico Br ex Sã m median value of the indicator position in the distribution of the average city Source: Elaboration based on Duque et al. 2017a. potentially hindering the exchange of goods FIGURE 6.2  Change in Urban Form Indicators, 1996–2000 and ideas. This result supports the claim in chapter 4 that transport investments in LAC 20 15 cities have not enhanced city connectivity 10 within the region. 5 Percent In short, although the average city is 0 round and tends to have smooth perimeters, –5 –10 a dense and gridded street network, and a –15 densely built footprint, averages are deceiv- –20 ing and hide a wide diversity of urban forms ss ss s ity n ity lw es tio ne ne in the LAC region. ra rcu ns lln c Sp d th de se Fu Ci un oo er et Ro Sm t re In St Does Urban Form Matter for Productivity in the LAC Region? Source: Elaboration based on Duque et al. 2017a. Beyond describing urban form and trends in the region, we wish to establish whether vector of urban form variables including there is evidence of links between the vari- shape, internal structure, and land use met- ous urban form metrics and a city’s produc- rics and on a vector of control variables tivity. Does form ultimately matter for including geographic characteristics such productivity in LAC cities? To respond to as distance to the nearest international this question, we estimate an empirical border, temperature, precipitation, and model that regresses city productivity on a coastal location. 178   RAISING THE BAR To construct a proxy of productivity at the potential form.19 Strategy 1 uses all seven city level we used the density of NTL emitted urban form metrics, but strategy 2 is limited by the urban extent, controlling for popula- to using roundness and smoothness metrics tion density.18 We first regressed density of only because these are the only two form NTL on population density so as to isolate variables for which instruments can be con- the variation in NTL density explained by structed (see Duque et al. 2017a). population density alone. We then used the The results confirm that shape matters for residuals from this regression as a measure of productivity in LAC cities. city-level productivity, where a larger residual Table 6.3 presents four specifications of indicates that a city of given population den- the empirical model using strategy 1. sity is more productive. We used our sample Alternative specifications are used to disen- of 919 cities across 32 LAC countries (see tangle the relative importance of the different box 6.1). urban form metrics, and to avoid problems of One of the main empirical challenges in multicollinearity as some of the urban met- answering this question is to tackle the rics variables are highly correlated. 20 Six of endogeneity of urban form when estimating the seven form metrics are in the specifica- its effect on productivity. City form can, tions in table 6.3. For purposes of simplicity, in fact, be taken as the result of the interac- we do not present the variable for street den- tion between decisions taken by firms, house- sity (it is highly correlated with intersection holds, and government (within the constraints density). Similar results were found using posed by topography); and hence both urban strategy 2. form and productivity stem from the inter- Our results confirm the importance of play between agglomeration economies and considering dimensions of urban form that congestion forces. In the simplest example, go beyond population density when looking cities that grow dense can facilitate agglomer- at city productivity. Three of the seven urban ation economies by increasing proximity of metric indicators are significant at the firms, fostering productivity. But highly pro- 90 percent level or higher in all model speci- ductive cities are also more likely to have gov- fications in which they are included. ernments able to invest in city centers, invest Regarding the shape of cities, we find that in better-planned street networks, and better the coefficients for roundness and smooth- manage land use patterns to reduce sprawl. ness are positive and significant across all To tackle these endogeneity concerns, we specifications. 21 Other things being equal, adopted two alternative identification strate- a more circular urban shape and a smooth gies. Strategy 1, as in Fallah, Partridge, and perimeter are associated with higher produc- Olfert (2011), uses lagged explanatory vari- tivity. This suggests that the way a city grows, ables from an earlier year to mitigate the within its boundaries and at its periphery, direct simultaneity between the dependent can have an impact on productivity. This and independent variables. In our case, we implies that city leaders can influence the regressed city productivity in 2010 on urban productivity of their cities with policies form metrics from either 1990 or 1996, that shape the physical form they take. depending on data availability. Strategy 2 Infrastructure investments, land use, and uses an instrumental variables (IV) approach, zoning regulations not only are therefore exploiting both temporal and cross-sectional tools for planning the form a city takes but variation in city shape. For this, we used the can also—through their role in building the time variation of NTL data to build a panel shape, texture, and land use of a city—­ of time-city observations. We followed influence productivity. Harari (2016) and constructed a synthetic Changing our focus from the periphery to instrument that uses the potential shape of a the internal structure, we find that having a city as a starting point, and calculated the dense street network (higher values of street city form indicators on the basis of such density or intersection density) is associated U rban F o r m , I nstituti o na l F rag m entati o n , an d Metr o p o l itan C o o r d inati o n   179 TABLE 6.3  Regression Results for Urban Form and City Productivity with Outliers (a) (b) (c) (d) Roundness (1996) 0.459*** (0.0656) Smoothness (1996) 0.496*** 0.485*** (0.1702) (0.1584) Fullness (1990) 0.683** 0.536* 0.433 (0.3012) (0.2981) (0.3297) Fullness2 (1990) –0.717*** –0.529** –0.390 (0.2400) (0.2424) (0.2607) Circuity (1996) 0.125 –0.065 0.088 0.107 (0.7760) (0.7864) (0.7815) (0.7472) Street density (1996) 0.406*** 0.386*** 0.364*** 0.389*** (0.0684) (0.0677) (0.0728) (0.0749) Sprawl (1990) 0.085 (0.0677) Constant –0.732 –0.558 –0.634 –1.026 (0.5758) (0.8088) (0.7401) (0.7055) N 919 919 919 919 R 0.282 0.269 0.279 0.283 Source: Duque et al. 2017a. Note: The dependent variable is the residuals of the regression of nighttime lights density on population density. Street density 1996 has been rescaled (divided by 1 × 10 4). Robust standard errors clustered at the country level appear in parentheses. All models include controls for geographic characteristics as measured by natural amenities that are distances (in thousand kilometers) to international border, temperature, precipitation, and coast indicator, and include country fixed effects. ***p < 0.01. **p < 0.05. *p < 0.1. with higher productivity. 22 Our results sug- whether there is a nonlinear relation between gest that a 10 percent increase in intersec- fullness and productivity. Despite having esti- tion density would be associated with mated coefficients for both terms of the productivity levels that are about 39 percent expected sign and significance, we find the higher. Although this may seem like a large combined effect of fullness on productivity effect, increasing intersection density by not significantly different from zero for most 10 percent would require significant efforts values of fullness. Finally, contrary to the and investments. Rio de Janeiro, for exam- results presented by Fallah, Partridge, and ple, increased its street density by only Olfert (2011), the sprawl variable is not signif- 4 percentage points between 1996 and 2010. icant across all specifications, which suggests On regularity of urban structure, our results that, after controlling for the shape and inter- show no evidence in favor, at least from a nal texture characteristics of the city, there is productivity perspective, of a regular grid- no evidence that the distribution of population ded street network: the coefficient of circu- density within cities, as measured through the ity was found to be not significant across all sprawl index, affects their productivity. The specifications, after controlling for other results here suggest that building denser street measures of urban form. networks in more elongated cities could lift For land use patterns, we introduce a qua- these cities’ productivity toward that of dratic term for the fullness variable to test rounded, smoother, and more compact cities. 180   RAISING THE BAR To summarize, urban form matters for fragmentation and introduce metropolitan city productivity in the LAC region, and spe- coordination proxies. cific characteristics (such as roundness and Most of the literature follows three lines of smoothness) appear to create more conducive thought when looking at the links between urban spaces for firms and households to the governance structures of cities and their interact. We also confirmed the validity of economic performance: polycentrist, centrist, moving beyond population density and and regionalist. broadening the measurements of urban form The polycentrist view argues that institu- to include proxies of intracity connectivity tional fragmentation in cities is equivalent to and land use (such as built-up area fragmen- creating additional layers of decentralization tation). The results suggest that policy mak- that can, in fact, enhance economic growth ers have several instruments at hand to (Fischer 1980) through two mechanisms: bet- increase their cities’ productivity by influenc- ter information, which leads to more efficient ing the form their cities take. Although provision of public goods (Ostrom 2010), and rounded, smoother, and more compact cities increased competition between individual of a given population density tend to be more local governments (Stansel 2005). This is productive, the evidence suggests that a more consistent with the arguments put forward by elongated city could become more productive Charles Tiebout in 1956, suggesting that by improving connectivity, for example by competition between local governments leads building a denser street network. to efficiency gains. The centrist view argues that the presence of multiple local governments within metro- Institutional Fragmentation, politan areas may generate coordination fail- Metropolitan Coordination, and ures that reduce efficiency in providing Productivity transport infrastructure and land use plan- This section moves beyond spatial urban form ning, with negative repercussions for eco- and focuses on institutional fragmentation.23 nomic performance (Ahrend, Gamper, and In chapter 2 we saw that LAC stands out rela- Schumann 2014b). Fragmentation may also tive to other regions for its high number of reduce a metropolitan area’s ease of doing multicity agglomerations (MCAs). LAC coun- business because of the additional bureaucracy tries also stand out against those in North that it imposes on firms (Kim, Schumann, and America and Western Europe for exhibiting a Ahrend 2014) and the associated higher trans- negative and significant relationship between action costs and barriers to the diffusion of the share of their population living in MCAs growth-promoting policies (Cheshire and and their gross domestic product per capita. Gordon 1996; Feiock 2009). The centrist view These results are consistent with the hypothe- thus argues that the costs of fragmentation are sis that institutional fragmentation can have higher than the efficiency benefits it may bring. negative effects on national productivity. The regionalist view is a middle way In this section, we seek to assess whether between the two: it recognizes the benefits of the fragmentation of urban areas across dif- local governments while highlighting the ferent administrative units has an effect on importance of metropolitan coordination, productivity, and whether the existence of defined as the efforts of governmental institu- metropolitan governance bodies attenuates tions to manage and solve problems in the negative effects of fragmentation. In common between municipalities (Ríos 2015). contrast to chapter 2, we focus on productiv- ­ According to Grassmueck and Shields (2010), ity at the level of individual cities rather than more important than the presence of multiple at the national level, following the methods local governments is the way in which they used in chapter 3 to control for the sorting of interact and perceive each other. For a sample workers across different cities. We also of OECD countries, Ahrend, Gamper, and expand the indicators of institutional Schumann (2014b) found that the presence of U rban F o r m , I nstituti o na l F rag m entati o n , an d Metr o p o l itan C o o r d inati o n   181 a governance body that coordinates munici- administrative units within a city; the palities halved the productivity penalty asso- nu mb er of ad m i n i s t rat ive u n it s p er ciated with fragmentation, measured by the 100,000 inhabitants; and the share of the number of municipalities in a given metro- population living in the central city. Each politan area. Foster (1993) and Nelson and covers different aspects of fragmentation Foster (1999) also found empirical support as proposed by Hendrick and Shi (2015): for the regionalist view: they found a positive fragmentation of a given urban extent; association between income growth and the scale of institutional fragmentation; and presence of overarching decision-making dominance of the central city in the metro- mechanisms such as multijurisdictional, politan region (table 6.4). All three vari- multipurpose regional governments. Also, the ­ ables are constructed using spatial data presence of single-purpose districts associ- (box 6.2). ated with large-scale infrastructure provision In a similar way, we focus on a subset of (such as water and wastewater systems) has variables with the aim of covering the multi- been found to foster income growth. ple dimensions of metropolitan coordina- Empirical studies looking at the role of tion. Coordination can result from the institutional fragmentation and governance presence of institutions (metropolitan gover- on economic performance have focused nance bodies), coordinated planning pro- mostly on developed countries. 24 Further, cesses (for example, for land use planning their results do not consistently support one and mobility), or special-purpose entities of the three views. In this section, we extend that overlap with administrative units (usu- the interpretations to LAC, and test whether ally for providing certain public services). empirical data are supportive of one of these We use three proxies (see table 6.4) to cap- lines of thought. ture each of these dimensions: the presence of a metropolitan governance body; the per- centage of municipalities covered by an inte- Measuring Fragmentation and grated transport system; and the total Coordination in LAC Cities number of single-purpose districts for public We focus on three variables to measure service provision in the metropolitan area institutional fragmentation: the number of (see box 6.2). TABLE 6.4  Institutional Fragmentation and Metropolitan Coordination Dimension Description Institutional fragmentation I. Size of region Number of administrative units 2010 II. Political fragmentation Number of administrative units per 100,000 inhabitants 2010 III. Central city domination Central-city population share 2010, where the central city is defined as the city whose administrative area overlaps the most with the identified urban extent Metropolitan governance I. Governance Presence of a governance body II. Land use plan and mobility Percentage of municipalities covered by integrated transport systems (metro, bus) between municipalities and central city III. Coordination for SPDs Presence of an SPD for water Presence of an SPD for energy Presence of an SPD for waste collection SPD water + SPD energy + SPD waste = SPD sum Note: SPD = single-purpose district. 182   RAISING THE BAR BOX 6.2  Constructing Institutional Fragmentation and Metropolitan Coordination Variables We used data from Defense Meteorolog ical Using a combination of desk review and spatial S atellite Program-Operational Linescan System ­ data, we constructed a database to characterize the (DMSP-OLS) NTL imagery (see box 6.1) to iden- metropolitan areas in the region in terms of insti- tify urban extents in the LAC region. We considered tutional fragmentation and metropolitan coordi- that a metropolitan area existed when more than nation. We used the administrative boundaries of one municipality or equivalent administrative unit local governments that conform with the metropol- intersected a single urban extent with more than itan areas and distributed population data to cal- 500,000 people in 2010. We used the administra- culate institutional fragmentation measures using tive unit boundaries from the World Bank’s LAC geoprocessing tools in ArcGIS. Administrative Geospatial Database, constructed for this book for boundaries were obtained from OpenStreetMap a this purpose (Branson et al. 2016). Metropolitan and the World Bank’s LAC Geospatial Database. area boundaries were obtained by aggregation of all Population counts at the administrative unit and administrative units that intersected the same urban urban extent levels were estimated using the Global extent. We verified each obtained metropolitan area Human Settlement Layer distributed population with ancillary information from official sources to grids produced by the Joint Research Centre of the include those municipalities that are part of the offi- European Union (Freire and Pesaresi 2015; Pesaresi cial metropolitan area denomination but were not et al. 2015). All metropolitan coordination vari- intersected by the urban extent. Map B6.2.1 pres- ables were obtained through a desk review using ents some examples of identified metropolitan areas: official information, further detailed in Duque Mexico City, Rio de Janeiro, and Buenos Aires. et al. (2017b). MAP B6.2.1  Examples of Metropolitan Areas a. Mexico City b. Rio de Janeiro c. Buenos Aires Source: Elaboration based on Duque et al. 2017b. Note: Urban extents extracted from 2010 nighttime images (in red), over the Global Human Settlement Layer built-up layer for 2014 (Freire and Pesaresi 2015), with administrative boundaries (light purple). a. For more information, see http://www.openstreetmap.org/copyright. LAC Cities: How Fragmented? unit overlapping with its urban extent. How Coordinated? This corresponds to 110 of the 919 cities We limited our analysis of institutional frag- included in the urban form analysis (see the mentation and coordination to metropolitan second section in this chapter, “Urban Form areas with more than 500,000 inhabitants and Productivity”). This restriction in our in 2010 and more than one administrative city sample was to focus on “larger cities,” U rban F o r m , I nstituti o na l F rag m entati o n , an d Metr o p o l itan C o o r d inati o n   183 usually thought to face the most challenging conduc ted by A h rend , G a mp er, a nd coordination issues. It also allows us to Schumann (2014). 25 In table 6.5 we compare compare, albeit not perfectly, results with the 110 metropolitan areas in the LAC those from a study of institutional fragmen- region with the 225 metropolitan areas tation and coordination in OECD countries included in Ahrend, Gamper, and Schumann TABLE 6.5  Institutional Fragmentation and Metropolitan Coordination, LAC Region versus OECD LACa OECD Mean SD Mean SD Number of MAs > 500,000 110 225 Average population of MA 2,257,472 3,309,093 2,072,762 3,552,573 a Average population density of central city (inhabi- 3,339.251 2,025.94 1,877.66 1,901.62 tants per square kilometer) Average population density of MAs (inhabitantss 806.4 868.97 703.26 754.114 per square kilometer) Average number of admin units per MA 9.39 11.42 86.19 135.76 Average number of admin units per 100,000 0.55 0.60 5.64 7.27 inhabitants Share of MA population living in the central citya 38.83 72.98 Share of MA with joint governance body 48.86 67.84 Source: Elaboration based on Duque et al. 2017b and Ahrend, Gamper, and Schumann 2014a. Note: LAC = Latin America and the Caribbean; MA = metropolitan area; OECD = Organisation for Economic Co-operation and Development; SD = standard deviation. a. Differences may be a result of inconsistencies in how central cities are identified between Duque et al. (2017b) and Ahrend, Gamper, and Schumann (2014a). TABLE 6.6  Top 15 Fragmented Metropolitan Areas, LAC Region versus OECD No. of admin. MA popula- No. of admin- Population Country Central city unitsa tion (2010) Country Central city istrative unitsa (2010) Mexico Mexico City 76 21,242,585 France Paris 1,375 11,693,218 Brazil São Paulo 39 20,483,833 Korea, Rep. of Seoul 964 22,529,435 Mexico Puebla 38 2,805,693 United States Chicago 540 9,461,105 Brazil Curitiba 37 3,640,533 Czech Republic Prague 435 1,829,843 Brazil Ribeirao Preto 34 5,085,801 France Toulouse 434 1,217,316 Brazil Belo Horizonte 34 1,573,563 United States New York 356 16,539,430 Costa Rica San Jose 31 2,319,583 France Rouen 346 698,385 Brazil Porto Alegre 31 13,588,699 United States Minneapolis 329 3,348,859 Argentina Buenos Aires 31 4,103,952 France Lyon 327 1,894,945 Brazil Sorocaba 27 1,950,203 Austria Vienna 313 2,683,251 Brazil Londrina 25 1,041,624 Germany Hamburg 308 2,984,966 Brazil Brasilia 23 3,879,415 Australia Melbourne 281 4,105,857 Brazil Rio de Janeiro 21 12,104,842 Australia Sydney 279 4,555,516 Brazil Joinville 20 2,922,544 Germany Berlin 276 4,374,708 Mexico Oaxaca de Juarez 20 1,130,568 Spain Madrid 272 6,507,502 Source: Elaboration using Duque et al. 2017b and Ahrend, Gamper, and Schumann 2014a. Note: LAC = Latin America and the Caribbean; OECD = Organisation for Economic Co-operation and Development. a. “No. of administrative units” refers to number of local governments operating in a metropolitan area. 184   RAISING THE BAR (2014a) for 28 OECD countries (excluding more skilled workforce. From this first step Chile and Mexico). we obtain the productivity differentials— The average metropolitan area in the which cannot be explained by workers’ own LAC region and OECD has a very similar observable characteristics—and regress them population of about 2 million. Population on institutional fragmentation variables, met- density is, on average, higher in the LAC ropolitan coordination variables, and other region than in OECD metropolitan areas, control variables (table 6.4). We limit our particularly within central cities. However, analysis to the metropolitan areas in the pre- metropolitan areas in the LAC region are vious subsection for which productivity dif- much less fragmented than those in the ferentials (estimated by Quintero and Roberts OECD (table 6.6): 0.55 administrative 2017) are ­ available. 26 This constrains our units per 100,000 inhabitants versus 5.64. sample to 73 LAC cities, across 14 LAC coun- Administrative fragmentation is even more tries. Results are in table 6.7. marked, with the average metropolitan Columns 1 through 3 in table 6.7 repli- area in the OECD having some 86 adminis- cate three of the specifications estimated in trative units, against 9.39 in the LAC chapter 3. In all three cases, for our more ­ region. limited sample of cities, we see that popula- About half of the metropolitan areas in tion density has essentially no estimated the LAC region have a metropolitan gover- effect on productivity. This is the case even nance body, against 68 percent in the without controlling for a city’s geographical OECD. The presence of governance bodies characteristics and its stock of human capi- and other coordination mechanisms in tal (as measured by its average years of LAC metropolitan areas is indicative by schooling). 27 Consistent with chapters 3 itself of an intent by policy makers in the and 5, we also find evidence of strong region to foster coordination in large and human capital externalities (column 3). fragmented urban areas. Below we exam- Columns 4 and 5 then add the fragmenta- ine whether these efforts are having any tion and metropolitan variables that are the measurable impact. core of analysis in this chapter. In both col- umns, this results in the estimated coeffi- cient on the log of population density Do Fragmentation and Metropolitan becoming both positive and strongly statis- Coordination Matter for Productivity tically significant. On first inspection, this in the LAC Region? would seem to imply the existence of strong To estimate the relationship between institu- agglomeration economies, even having con- tional fragmentation and metropolitan area trolled for human capital, in contradiction productivity, we build on the two-stage to results presented earlier in this book. empirical approach followed by Quintero However, this would be the incorrect take- and Roberts (2017) for LAC cities in away because the regressions also control chapter 3 and follow a similar analysis to the ­ for area and include a variable (the number one conducted by Ahrend, Gamper, and of administrative units per 100,000 inhabi- Schumann (2014b) for OECD countries. tants), the definition of which includes pop- The first step consists of extracting the ulation. Given this, unlike the regressions in productivity differentials that result from columns 1–3, the coefficient on population population sorting: more skilled workers tend density cannot be interpreted as an estimate to prefer living in larger cities (see chapters 3 of the elasticity of productivity with respect and 4). This is necessary because otherwise to population density. one may confound agglomeration benefits To see this, consider the implied effect of with productivity increases linked to having a an increase in a city’s population density that U rban F o r m , I nstituti o na l F rag m entati o n , an d Metr o p o l itan C o o r d inati o n   185 TABLE 6.7  Regression of a City Productivity Premium (ln) on Institutional Fragmentation and Metropolitan Coordination Variables Variables (1) (2) (3) (4) (5) Log(Population Density) −0.002 0.005 −0.014 0.203*** 0.205*** (0.0271) (0.0327) (0.0259) (0.0579) (0.0607) Log(Average Years of Schooling 2010) 2.011*** 1.454*** 1.449*** (0.3408) (0.3239) (0.3414) Log(Area km2) 0.264*** 0.262*** (0.0625) (0.0644) Log(Number of Admin Units 2010) −0.172*** −0.173*** (0.0614) (0.0626) Number of admin units per 100,000 inhabitants 0.212*** 0.210*** 2010 (0.0700) (0.0712) Share of population living in central city 2010 −0.345*** −0.336** (0.1272) (0.1360) Governance body 0.021 (0.0482) Integrated transport system −0.000 (0.0006) Sum (single-purpose districts) 0.029 (0.0279) log(Terrain Ruggedness) −0.011 −0.016 −0.021 −0.022 (0.0284) (0.0223) (0.0191) (0.0204) log(Mean Air Temperature) −0.002 0.090 0.203*** 0.189** (0.0960) (0.0770) (0.0706) (0.0772) log(Total Precipitation) −0.007 −0.070 −0.112** −0.093 (0.0782) (0.0623) (0.0538) (0.0611) Constant 1.158*** 1.166*** −3.580*** −5.450*** −5.551*** (0.1704) (0.2938) (0.8357) (0.9404) (0.9727) Country dummies Y Y Y Y Y No. of observations 71 71 71 71 71 Adjusted R2 0.641 0.619 0.768 0.832 0.826 Source: Duque et al. 2017b. Note: City productivity premiums for metropolitan areas are estimated using Quintero and Roberts’ (2017) narrow sample (restricted to prime-age men working in the private sector). Sum (single-purpose districts) refers to the presence of a single-purpose district for water, energy, and waste added together so that the max value for the variable is 3. Standard errors are in parentheses. ***p < 0.01. **p < 0.05. *p < 0.1. results from a 1 percent decline in its area, coefficient on log (population density) and β ˆ holding everything else constant. The results is the estimated coefficient on log (area km 2). in columns 4 and 5 imply that the effect of For column 4, this value would be equal to this increase in population density would be −0.061, whereas, for column 5, it would be given by α − β, where α is the estimated equal to −0.057. Hence, in both cases, an 186   RAISING THE BAR increase in population density resulting from suggest the presence of nonlinearities in the a reduction in area is estimated to have a relationship between the fragmentation of a negative effect on productivity. This is con- metropolitan area and its productivity and chapter 2, sistent with the results presented in ­ that, starting from a certain level of fragmen- where urban areas in the LAC region are tation, the benefits of more responsive local shown to be relatively dense, not so much government or greater competition can com- because they have higher populations but pensate for costs due to coordination because their populations tend to be squeezed failures. into smaller areas.28 Figure 6.3 visualizes the relationship Fragmentation—as measured by the num- between fragmentation and productivity, and ber of administrative units and the number of sheds more light on its nonlinear characteris- administrative units per 100,000 inhabitants— tics. This representation of the results shown is found to matter for productivity. As in table 6.7, column 4, summarizes the net observed in table 6.7 (columns 4 and 5) both effect of increasing the number of adminis- variables are statistically significant, albeit trative units, holding population size constant, having different signs (the first negative, for cities having 500,000, 1 million, or 10 the second positive). Our results, combined million inhabitants. with the results from chapter 2, are consis- The main message from the figure is that tent with the hypothesis that fragmentation in LAC cities the negative effects of fragmen- may be dampening the benefits of agglomera- tation dominate, and, given current struc- tion economies in LAC cities. These results tures, only extreme fragmentation would FIGURE 6.3  What Levels of Fragmentation Are Needed to Reap the Benefits? 1.2 1.0 0.8 City productivity premium (ln) 0.6 0.4 0.2 0 –0.2 –0.4 –0.6 –0.8 –1.0 0 5 10 15 20 25 30 35 40 45 50 Level of fragmentation (number of administration units) 0.5 million inhabitants 1 million inhabitants 10 million inhabitants Source: Calculations based on Duque et al. 2017b. Note: The figure displays the net effects of different fragmentation levels for metropolitan areas of different population sizes (0.5 million, 1 million, and 10 million) using the coefficients for number of administrative units and number of administrative units per 100,000 inhabitants in 2010 (from column 3, table 6.7). U rban F o r m , I nstituti o na l F rag m entati o n , an d Metr o p o l itan C o o r d inati o n   187 lead to any benefits and only then for cities likely needs to expand the desk review that are not too populous. A metropolitan approach (used to gather the metropolitan area of 500,000 inhabitants would need at variables in this chapter) to include variables least 16 administrative units to start reaping that reflect reality on the ground, for exam- any benefits, and a metropolitan area of ple, through a detailed city survey. 1 million inhabitants at least 42 administra- tive units. Because the average metropolitan area in the LAC region has 9.39 administra- Conclusions tive units, our results suggest that, in prac- This chapter suggests that the average LAC tice, most LAC metropolitan areas are city is rounded and has smooth perimeters probably being affected negatively by their and a dense street network. Controlling for fragmentation (rather than benefitting from a city’s average population density, these it). For a city of 10 million inhabitants, the characteristics seem to be positively linked net effect of fragmentation on productivity is to a city’s economic performance, likely negative at all levels of fragmentation. These supporting the emergence of agglomeration results are similar to those in Ahrend, economies through different mechanisms. Gamper, and Schumann (2014b) and are An important outcome is that a city can consistent with the centralist view that argues grow in different shapes and still achieve for the presence of coordination failures and high productivity by guaranteeing a high their negative repercussions on economic per- rate of inner-city connectedness (equally, a formance. Finally, we find that an increase in compact but poorly connected city can the central city domination—measured as the show low productivity.) The results also share of the population in the metropolitan underscore the fact that urban form tends area living in the central city—may negatively to persist over time, requiring policy mak- affect economic performance. ers to think far ahead and ensure good We find no evidence that metropolitan accessibility within cities. This has import- coordination variables have an effect on eco- ant implications for policy makers because nomic performance (see table 6.7, col- mayors often ask what they can do to umn 5). 29 In fact, contrary to what was improve the productivity of their cities. City found by Ahrend, Gamper, and Schumann planning and land management policies are (2014b), none of these variables appears sig- not often regarded as instruments to foster nificant. There is therefore no evidence in productivity and growth in cities. The favor of the regionalist view because our results in this chapter suggest otherwise: results show that the presence of a gover- these are important tools that local govern- nance body or integrated public services ments have at hand to increase productivity does not necessarily foster increased produc- in LAC cities. tivity for LAC cities. However, these last Institutional fragmentation matters for results need to be viewed with caution productivity, and most metropolitan areas because they are conditioned on the vari- in the region are hurt by it; but there is no ables used to measure the degree of metro- evidence of metropolitan coordination miti- politan coordination. For example, it is gating these impacts. This raises doubts quite possible that a metropolitan coordina- over the effectiveness of current bodies for tion body is in place, at least on paper, but metropolitan coordination, in part stem- that it is not effective in practice in solving ming from overlapping responsibilities coordination failures. To fully understand across local governments and government the role of such bodies in reducing the perils agencies, and from these bodies’ limited of fragmentation in the LAC region, one authority. 188   RAISING THE BAR Annex 6A: Seventy-Three Cities in Institutional Fragmentation and Coordination Analysis Country Central city Country Central city Brazil Florianopolis Argentina Rosario Mexico Tuxtla Gutierrez Brazil Vitoria Colombia Pereira (Centro Occidente) Bolivia Santa Cruz Mexico Morelia Brazil Santos Peru Arequipa Argentina Cordoba (Capital) Brazil Joinville Uruguay Asuncion Argentina Salta Bolivia La Paz Brazil Ribeirao Preto San Salvador San Salvador Mexico Cancun Brazil Manaus Brazil Londrina Colombia Barranquilla Guatemala Quetzaltenango Brazil Goiania Mexico Veracruz Brazil Belem Brazil Sorocaba Brazil Brasilia Mexico Saltillo Mexico Toluca Mexico Tampico Costa Rica San Jose Colombia Cucuta Brazil Campinas Brazil Cuiaba Colombia Cali Mexico Chihuahua Ecuador Guayaquil Brazil Sao Jose Dos Campos Mexico Puebla Peru Trujillo Brazil Curitiba Brazil Aracaju Guatemala Guatemala Mexico Aguascalientes Brazil Salvador Bahia Argentina Tucuman Brazil Fortaleza Mexico Queretaro Colombia Medellín Mexico Merida Dominican Republic Santo Domingo Argentina Mendoza Mexico Cuernavaca Brazil Porto Alegre Brazil Teresina Brazil Recife Bolivia Cochabamba Mexico Monterrey Mexico Torreon Brazil Belo Horizonte Brazil Joao Pessoa Mexico Guadalajara Brazil Maceio Chile Santiago Mexico San Luis Potosi Colombia Bogotá Colombia Bucaramanga Peru Lima Panama Panama City Brazil Rio de Janeiro Brazil Natal Mexico Mexico City Brazil Sao Luis Brazil São Paulo U rban F o r m , I nstituti o na l F rag m entati o n , an d Metr o p o l itan C o o r d inati o n   189 Annex 6B: Urban Form Indicators n o d e s i t l i n k s ( B o e i n g 2 017 ). T h e unweighted circuity of an area m is calcu- Roundness index lated as follows: Draw the equal-areas circle about the proxi- mate center Cp and calculate the area of over- m DN Cu , m = lap Os of the equal-area circle and the shape. m DE The following is the formula for calculating where the Exchange Index Ix: C u,m is the average unweighted circuity in Os area m, IX =  A m D is the sum of the network distances N between all origin-destination pairs in the (Angel, Parent, and Civco 2010b). sample, and m D is the sum of the Euclidean distances E Smoothness index between all origin-destination pairs in the Find the perimeter P of the shape. The sample (Giacomin and Levinson 2015). following is the formula for calculating the ­ Exchange Index Ip: Intersection density I p = PA / P = (2 πA ) / R  Intersection density is the node density of the set of nodes with more than one street ema- (Angel, Parent, and Civco 2010b). nating from them (thus excluding dead ends) (Boeing 2017). Fullness index Calculate the radius rA of a small neighbor- Street density hood in the shape, so that πr2A = A/100, with Street density is the sum of all edges in the undirected (an undirected graph’s edges point rA = (A / 100π). mutually in both directions) representation of the graph (an abstract representation of a set Find the average fullness of the shape, F S , of elements and the connections between as the average of the fullness Fi of a small cir- them) (Boeing 2017). cle of radius rA about the center of every pixel i in the shape Sprawl index  m  Metropolitan sprawl is measured as follows: FS =  ∑ Fi  / n   i =1  Sprawl = ((L% − H%)+1)) × 0.5 (Angel, Parent, and Civco 2010b). where L% is the share of the metropolitan Circuity population living in block groups with popu- Circuity divides the sum of all edge lengths lation density below the overall metropolitan by the sum of the great-circle distances median block group. H% is the share of met- between the nodes incident (element of a ropolitan population living in block groups graph) to each edge. This is the average with density above that of the overall metro- ratio between an edge length and the politan median block g roup (Fallah, straight-line distance between the two Partridge, and Olfert 2011). 190   RAISING THE BAR Annex 6C: Correlation Matrix between Urban Form Variables (1) (2) (3) (4) (5) (6) (7) Roundness (1996) 1 Smoothness (1996) 0.7453* 1 Fullness (1990) −0.1573* −0.3100* 1 Intersection density (1996) 0.1427* −0.0309 0.3429* 1 Street density (1996) 0.1527* −0.0465 0.4361* 0.9470* 1 Circuity (1996) −0.1993* −0.1172* −0.0586 −0.3901* −0.4351* 1 Sprawl (1990) 0.0422 0.2216* −0.7122* −0.5454* −0.5939* 0.2032* 1 Note: * Denotes significance at 5 percent. Notes squares that fall within the boundaries of all 1. Sprawling is usually referred to as low-density cities in our sample. Population density levels expansion of cities, or density decline are calculated for each square (or pixel) and (Brueckner and Fansler 1983; Civco et al. hence not at the aggregate city level. 2000; Fulton et al. 2001). 9. We used the GHS-Pop layers as outlined in 2. Accessibility is defined as the number of points Pesaresi et al. (2015). An important constraint that can be reached in a predefined period, of this indicator is that it does not take into for example one hour. account the actual height of buildings, but 3. The roundness index has been referred to in the assumes that population is equally distributed geography literature as the exchange index. We in all built-up areas in an administrative unit. refer to it here as roundness for ease of interpre- 1 0. Recent work by the World Health tation (see Angel, Parent, and Civco 2010b). Organization suggests a positive relationship 4. The Shape Metrics Toolbox was used to calcu- between urban green space and health, which late the shape metrics of the urban extent may in turn affect productivity (WHO 2016). polygons. This software is intellectual prop- 11. NTL products have high correlation with erty of the Center for Land Use Education and human activities (Hsu et al. 2015), and have Research at the University of Connecticut been used previously for regional and global (http://clear.uconn.edu/tools/Shape_Metrics​ analysis of urbanization (Cheng et al. 2016; /­index.htm). Pandey, Joshi, and Seto 2013; Sutton, Cova, 5. The correlation between roundness and and Elvidge 2006; Zhang and Seto 2011; smoothness is 0.93, hence the measures are Zhou, Hubacek, and Roberts 2015; Zhou not used concurrently in the regressions. et al. 2015), population modeling (Anderson 6. The undirected representation of a street net- et al. 2010; Lo 2001), and economic perfor- work considers that all nodes of the network mance (Cao et al. 2016; Forbes 2013; point in all directions. Henderson, Storeygard, and Weil 2012; Shi 7. Ideally, one would like to build indicators that et al. 2014; Small, Elvidge, and Baugh 2013; consider the variation in actual land uses, Chen and Nordhaus 2011). NTL data are but such information is rarely available in also used in chapters 1, 2, and 4 of this book. LAC cities; and, although there is work using 12. For a discussion on the use of luminosity satellite imagery to obtain an approximation density as a proxy for economic statistics, see of land use through land cover classes, such Chen and Nordhaus (2011). efforts are time and computer intensive and 13. Although the more recent VIIRS NTL data require high resolution (day-time) satellite used in chapter 2 overcome some of the chal- imagery (see Antos et al. 2016). lenges of the DMSP-OLS data, they lack the 8. Because population is available at a spatial time dimension needed for analyzing urban resolution of 250 meters, the distribution of form over time (they are available only since reference is the distribution of 250-meter 2013). U rban F o r m , I nstituti o na l F rag m entati o n , an d Metr o p o l itan C o o r d inati o n   191 14. Because of space limitations, table 6.2 shows density are not presented in the table but are statistics for only 1996 and 2010. similar to those of street density. This is 15. Huang, Lu, and Sellers (2007) find that LAC expected because these two variables are cities are more elongated than those in Asia highly positively correlated. and Europe, which tend to be more circular; 23. Although institutional fragmentation and LAC cities are less elongated than those in the metropolitan governance are, on their own, United States. Shanghai in China and key issues to be studied and understood in Manchester in the United Kingdom represent depth, we focus here only on their effect on the average circular city in their regions city productivity, given the overall focus of the whereas Boston has a more elongated shape book. We leave the in-depth study of these and represents the average city in the United two topics for further research. States. Huang, Lu, and Sellers (2007) use sat- 24. See, for example, Ahrend, Gamper, and ellite images of 77 metro areas in Asia, Schumann 2014b; Carr and Feiock 1999; Australia, Europe, Latin America, and the Parks and Oakerson 1989; Brezzi and Veneri United States to calculate seven spatial metrics 2014. These studies focus on metropolitan that capture five dimensions of urban form. governance in the United States, OECD coun- Note that Huang, Lu, and Sellers (2007) tries, and the European Union. group Japan with Europe, not with Asia 25. Analysis by Ahrend, Gamper, and Schumann 16. Expansion in LAC cities in 2000–10 was in (2014) includes metropolitan areas with a terrains with an average slope above population of 500,000 or more, similar to the 16 ­percent (Duque et al. 2017c). population threshold in this book. However, 17. Huang, Lu, and Sellers (2007) use a measure differences in how metropolitan areas are called porosity that measures the ratio of open delineated may affect comparability of indica- space to total urban area, which can be under- tors. The metropolitan areas defined by stood as a complement to measures like full- Ahrend, Gamper, and Schumann (2014) are ness and sprawl used in this book. They find functional economic areas characterized by a that LAC has less open space than Asia, densely inhabited “city” and a commuting Europe, and the United States, and slightly zone whose labor market is highly integrated more than Australia. with the core. This analysis relies on identifying 18. Measured as the natural logarithm of metropolitan areas through use of NTL. The deblurred NTL (lumens) per km2. This is con- city cores in Ahrend, Gamper, and Schumann sistent with measures used in previous chap- (2014) are defined by the LandScan population ters; by dividing both sides of the equation by database. Polycentric cores and the hinterlands city area, this is only a rescaling exercise to of the functional areas are identified on the facilitate comparison across cities and basis of commuting data (travel from home to interpretation. work) in 2000 (census year) with the require- 19. Following Harari (2016), the identification ment that more than 15 percent of the resident relies on the fact that exogenous changes in population of any of the cores commutes to the city form over time can result from work in the other core. encountering topographic obstacles along its 26. In their study, the authors use micro data on expansion path. nominal hourly wages in the main occupation. 20. The following urban form metrics were found As independent variables, they use a vector of to be highly positively correlated: roundness observable characteristics per worker (age, and smoothness, and intersection density and age squared, gender, marital status, and years street density. There is also a high and nega- of education completed) and municipality tive correlation between fullness and sprawl. fixed effects. 21. Similar results are found using strategy 2 for 27. In contrast, chapter 3 reports the absence of a the coefficient on roundness, which is positive significant relationship between productivity and highly significant at the 5 percent level. and population density only after controlling The coefficient of smoothness was found to be for geographical characteristics, human capi- negative but not significant. tal, and market access. 22. Nonlinear relationships were tested but found 28. Similar nonlinearities determine the effect of a to be nonsignificant. Results from intersection change in population on productivity, which 192   RAISING THE BAR by definition, would be a function of the num- Land Use.” Policy Research Working Paper ber of administrative units per 100k individu- No. 7904, World Bank, Washington, DC. als and the percentage of the population in the Batty, M. 2008. “The Size, Scale, and Shape of central city. Although marginal effects for all Cities.” Science 319: 769. population levels are not presented here Batty, M., and P. Longley. 1994. Fractal Cities: A because of space constraints, the resulting Geometry of Form and Function. London: nonlinear function suggests a positive Academic Press. marginal effect that increases with population ­ Bertaud, A. 2004. The Spatial Organization of up to an inflection point that occurs at popu- Cities: Deliberate Outcome or Unforeseen lation levels of about 10 million people when Consequence? Berkeley, CA: Institute of the marginal effect starts decreasing, but Urban and Regional Development. remains positive. Boeing, G. 2017. “OSMnx: New Methods for 29. Results for institutional fragmentation and Acquiring, Constructing, Analyzing, and metropolitan coordination variables remain Visualizing Complex Street Networks.” the same after including a series of control Working Paper, University of California, variables (table 6.7, column 5). Berkeley. Bogart, W. T. 1998. The Economics of Cities and Suburbs. Upper Saddle River, NJ: Pearson Education Company. References Branson, J., A. Campbell-Sutton, G. M. Hornby, Abrahams, A., N. Lozano Gracia, and C.  Oram. D. D. Hornby, and C. Hill. 2016. “A Geospatial 2016. “Deblurring DSMP Nighttime Lights.” Dat aba s e for L at i n A mer ic a a nd t he Working paper, World Bank, Washington, DC. Caribbean,” draft version 1. University of Ahrend, R., C. Gamper, and A. Schumann. Southampton, Southampton, U.K. 2014a. “The OECD Metropolitan Governance B r e h e n y, M . J . , e d . 1 9 9 2 . S u s t a i n a b l e Survey.” Regional Development Working Development and Urban Form. London: Pion Papers , Orga n isat ion for E conom ic Limited. Co-operation and Development, Paris. Brezzi, M., and P. Veneri. 2014. “Assessing ———. 2014b. “W hat Makes Cities More Polycentric Urban Systems in the OECD: Productive? Evidence on the Role of Urban C ou nt r y, R e g ion a l a nd M e t rop ol it a n Governance from Five OECD Countries.” Perspectives.” Regional Development Working OECD Regional Development Working Paper No. 2014/01, OECD, Paris. Papers, 2014/05, OECD Publishing. Brueckner, J., and D. A. Fansler. 1983. “The Anderson, S. J., B. T. Tuttle, R. L. Powell, and Economics of Urban Sprawl: Theory and P.  C . S ut to n . 2 010 . “C h a r a c t e r i z i n g Evidence on the Spatial Sizes of Cities.” Relationships between Population Density Review of Economics and Statistics 65 (3): and Nighttime Imagery for Denver, Colorado: 479–82. I s su e s of S c a le a nd R e pre s e nt at ion .” Cao, Z., Z. Wu, Y. Kuang, N. Huang, and International Journal of Remote Sensing M. Wang. 2016. “Coupling an Intercalibration 31 (21): 5733–46. of Radiance-Calibrated Nighttime Light Angel, S., J. Parent, and D. Civco. 2010a. “The Images and Land Use/Cover Data for Modeling Fragmentation of Urban Footprints: Global and Analyzing the Distribution of GDP in Evidence of Urban Sprawl 1990 –2000.” Guangdong, China.” Sustainability 8 (2): Lincoln Institute of Land Policy Working 1–18. Paper, Cambridge, MA. Carlino, G. A., S. Chatterjee, and R. M. Hunt. ———. 2010b. “Ten Compactness Properties of 2007. “Urban Density and the Rate of Circles: Measuring Shape in Geography.” Invention.” Journal of Urban Economics Canadian Geographer 54 (4): 441–61. 61 (3): 389–419. Antos, S. E., S. V. Lall, and N. Lozano Gracia. Carr, J. B., and R. C. Feiock. 1999. “Metropolitan 2016. “The Morphology of African Cities.” Government and Economic Development.” Policy Research Working Paper 7911, World Urban Affairs Review 34 (3): 476–88. Bank, Washington, DC. Cervero, R. 2001. “Efficient Urbanisation: Av ner, Paolo, a nd S om i k V. L a l l. 2016 . Economic Performance and the Shape of the “Matchmaking in Nairobi: The Role of Metropolis.” Urban Studies 38 (10): 1651–71. U rban F o r m , I nstituti o na l F rag m entati o n , an d Metr o p o l itan C o o r d inati o n   193 Cervero, R., and K. Kockelman. 1997. “Travel NTL Imagery.” Background paper for this Demand and the 3Ds: Density, Design and book, World Bank, Washington, DC. Diversity.” Transportation Research. Part D Duranton, G., and D. Puga. 2004. “Micro- Transportation Environment 2 (3): 199–219. fou nd at ion s of Urba n A g g lom erat ion Chatman, D., and R. Noland. 2014. “Transit Economies.” In Handbook of Urban and S e r v i c e , Phy sic a l A g g lom erat ion a nd Regional Economics, Volume 4, edited by Productivity in US Metropolitan Areas.” J. V. Henderson and J.-F. Thisse, 2063–2117. Urban Studies 51 (5): 917–37. New York: North Holland. Chen, X., and W. D. Nordhaus. 2011. “Using Fallah, B., M. Partridge, and M. Olfert. 2011. Luminosity Data as a Proxy for Economic “Urban Sprawl and Productivity: Evidence Statistics.” Proceedings of the National from US Metropolitan Areas.” Papers in Academy of Sciences of the United States of Regional Science 90 (3): 451–73. America 108 (21): 8589–94. Feiock, R. C. 2009. “Metropolitan Governance Cheng, Y., L. Zhao, W. Wan, L. Li, T. Yu, and and Institutional Collective Action.” Urban X. Gu. 2016. “Extracting Urban Areas in Affairs Review 44 (3): 356–77. China Using DMSP/OLS Nighttime Light Fischer, M. M. 1980. “Regional Taxonomy: A Data Integrated with Biophysical Composition Comparison of Some Hierarchic and Non- Information.” Journ al of Geographic al Hierarchic Strategies.” Regional Science and Sciences 26 (3): 325–38. Urban Economics 10 (4): 503–37. Cheshire, P. C ., and I. R. Gordon. 1996. Forbes, D. J. 2013. “Multi-Scale Analysis of the “Territorial Competition and the Predictability Relationship between Economic Statistics and of Collective (In)action.” International Journal DMSP-OLS Night Light Images.” GIScience of Urban and Regional Research 20 (3): & Remote Sensing 4 (4): 165–71. 383–99. Foster, K. A. 1993. “Exploring the Links between Ciccone A., and R. E. Hall. 1996. “Productivity Pol it ic a l S t r uc t u re a nd M e t rop ol it a n and the Density of Economic Activity.” Growth.” Political Geography 12 (6): 523–47. American Economic Review 86 (1): 54–70. Freire, S., and M. Pesaresi. 2015. “GHS Civco, D. L., J. D. Hurd, C. L. Arnold, and Population Grid, Derived from GPW4, S. Prisloe. 2000. “Characterization of Suburban Multitemporal (1975, 1990, 2000, 2015).” Sprawl and Forest Fragmentation through European Commission, Joint Research Centre. Remote Sensing Application.” Proceedings of the Fulton, W., R. Pendall, M. Nguyen, and ASPRS Annual Convention, Washington, DC. A. Harrison. 2001. Who Sprawls the Most? Croft, T. A. 1978. “Night-Time Images of the How Growth Patterns Differ across the Earth from Space.” Scientific American 239 United States. Washington, DC: Brookings (1): 68–79. Institution. de Roo, G., and D. Miller. 2000. Compact Cities Giacomin, David J., and David M. Levinson. and Sustainable Urban Development: A 2015. “Road Network Circuity in Metropolitan Critical Assessment of Politics and Plans from Areas.” Environment and Planning B: Planning an International Perspective. Aldershot, U.K.: and Design 42 (6): 1040–53. Ashgate. Glaeser, E. L. 1998. “Are Cities Dying?” Journal Duque, J. C., N. Lozano Gracia, J. Patino, and of Economic Perspectives 12 (2): 139–60. P. Restrepo. 2017a. “Urban Form and Glaeser E. L., and M. E. Khan. 2004. “Sprawl Productivity: In W hat Shape A re Latin and Urban Growth.” Handbook of Regional American Cities?” Background paper for this and Urban Economics 4: 2481–2527. book, World Bank, Washington, DC. Grassmueck, G., and M. Shields. 2010. “Does ———. 2017b. “Institutional Fragmentation and Government Fragmentation Enhance or Hinder Metropolitan Coordination in Latin American Metropolitan Economic Growth?” Papers in Cities: What Consequences for Productivity Regional Science 89 (3): 641–57. and Growth?” Background paper for this Harari, M. 2016. “Cities in Bad Shape: Urban book, World Bank, Washington, DC. Geometry in India.” Working Paper, The Duque, J. C., N. Lozano-Gracia, J Patino, Wharton School, University of Pennsylvania, P. Restrepo, and W. A. Velasquez. 2017c. Philadelphia. “Spatio-temporal Dynamics of Urban Growth Henderson, J. V., A. Storeygard, and D. N. Weil. in Latin American Cities: An Analysis Using 2012. “Measuring Economic Growth from 194   RAISING THE BAR Outer Space.” American Economic Review Lo, C. P. 2001. “Modeling the Population of 102 (2): 994–1028. China Using DMSP Operational Linescan Hendrick, R., and Y. Shi. 2015. “Macro-level System Nighttime Data.” Photogrammetric Deter m i na nts of L o ca l G over n ment Engineering and Remote Sensing 67 (9): Interaction: How Metropolitan Regions in the 1037–47. United States Compare.”  Urban Affairs Lynch, K. 1981. A Theory of Good City Form. Review 51 (3): 414–38. Cambridge, MA: MIT Press. Hsu, F. C., K. E. Baugh, T. Ghosh, M. Zhizhin, Ma, L., J. Wu, W. Li, J. Peng, and H. Liu. 2014. and C . D. Elvidge. 2015. “DMSP- OL S “Evaluating Saturation Correction Methods Radiance Calibrated Nighttime Lights Time for DMSP/OLS Nighttime Light Data: A Case Series with Intercalibration.” Remote Sensing Study from China’s Cities.” Remote Sensing 7 (2): 1855–76. 6 (10): 9853–72. Huang, J., and D. Levinson. 2015. “Circuity in Mills, E. S., and B. W. Hamilton. 1989. Urban Urban Transit Networks.” Working Paper Economics, 4th Edition. Glenview IL: Scott, 201501, Nexus Research Group, University of Foresman, and Company. Minnesota, Minneapolis. Nelson , A . C ., a nd K . A . Foster. 1999. Huang, J., X. X. Lu, and J. M. Sellers. 2007. “Metropolitan Governance Structure and “A Global Comparative Analysis of Urban Income Growth.” Journal of Urban Affairs Form: Applying Spatial Metrics and Remote 21: 309–24. Sensing.” Landscape and Urban Planning Ostrom, E. 2010. “Beyond Markets and States: 82 (4): 184–97. Polycentric Governance of Complex Economic Inostroza L ., R. Baur, and E . Csaplovics. Systems.”  Tran sn ation al C or poration s 2013. “Urban Sprawl and Fragmentation in Review 2 (2): 1–12. Latin America: A Dynamic Quantification Pandey, B., P. K. Joshi, and K. C. Seto. 2013. and Characterization of Spatial Patterns.” “Monitoring Urbanization Dynamics in India Journal of Environmental Management 115: Using DMSP/OLS Nighttime Lights and SPOT- 87–97. VGT Data.” International Journal of Applied Jaffe, Adam, Manuel Trajtenberg, and Rebecca Earth Observation and Geoinformation 23: Henderson. 1993. “Geographic Localization 49–61. of Knowledge Spillovers as Evidenced by Patent Parks, R. B., and R. J. Oakerson. 1989. “Metropolitan Citations.” Quarterly Journal of Economics Organization and Governance.” Urban Affairs 108 (3): 577–98. Quarterly 25 (1): 18–29. Kim, S. J., A. Schumann, and R. Ahrend. 2014. Parr, J. B. 1979. “Regional Economic Change “What Governance for Metropolitan Areas?” and Reg ional Spatial St r uc t u re: S ome Regional Development Working Paper, OECD, Interrelationships.” Environment and Planning Paris. A 11 (7): 825–37. Knaap, G., and A. C. Nelson. 1992. “The Partridge, M. D., D. S. Rickman, A. Kamar, and Regulated Landscape: Lessons on State Land M. R. Olfert. 2009. “Agglomeration Spillovers Use Planning from Oregon.” Cambridge, and Wage and Housing Cost Gradients across MA: Lincoln Institute of Land Policy the Urban Hierarchy.” Journal of International Knaap, G., C. Ding, and L. D. Hopkins. 2001. Economics 78: 126–140. “Managing Urban Growth for the Efficient Pesaresi, M., D. Ehrilch, A. J. Florczyk, S. Freire, Use of Public Infrastructure: Toward a Theory A. Julea, T. Kemper, P. Soille, and V. Syrris. of Concurrency.” International Regional 2015. “GHS Built-Up Grid, Derived from Science Review 24 (3): 328–343. Landsat, Multitemporal (1975, 1990, 2000, Litman, T. 2015. “Analysis of Public Policies That (2014).” Eu ropea n C om m ission , Joi nt Unintentionally Encourage and Subsidize Research Centre (JRC) [Dataset] PID: http:// Urban Sprawl.” Victoria Transport Policy data​.europa.eu/89h/jrc-ghsl-ghs​_ built_ldsmt​ Institute. Supporting paper commissioned by _­globe​_ r2015b. LSE Cities at the London School of Economics Prosperi, D., A. V. Moudon, and F. Claessens. and Political Science, on behalf of the Global 2009. “The Question of Metropolitan Form: Commission on the Economy and Climate Introduction.” Footprint 3(2): 1–4. (www.newclimateeconomy.net) for the New Quintero, L., and M. Roberts. 2017. “Explaining Climate Economy Cities Program. Spatial Variations in Productivity: Evidence U rban F o r m , I nstituti o na l F rag m entati o n , an d Metr o p o l itan C o o r d inati o n   195 from 16 LAC Countries.” Working Paper, Development in the Post-Reform Period: An World Bank, Washington, DC. Empirical Analysis.” Working paper. Quíros, T. P., and S. R. Mehndiratta 2015. Tiebout, C. M. 1956. “A Pure Theory of Local “Accessibility Analysis of Growth Patterns in Expenditures.” Journal of Political Economy Buenos Aires, Argentina: Density, Employment 64: 416. and Spatial Form.” Transportation Research Thompson, D. W. 1952. On Growth and Form. Record: Journ al of the Transportation Second Edition. Cambridge: Cambridge Research Board 2512. University Press. Ríos, A. A. 2015. “Metropolitan Coordination in Wheeler, C. H. 2001. “Search, Sorting, and Mexico.” Current Urban Studies 3 (1): 11–17. Urban Agglomeration.”  Journal of Labor Rosenthal, S. S., and W. C. Strange. 2004. Economics 19 (4): 879–99. “Evidence on the Nature and Sources of WHO (World Health Organization). 2016. Urban Agglomeration Economies.” In Handbook of Green Spaces and Health: A Review of Urban and Regional Economics, Volume 4, Evidence. Copenhagen: WHO Regional Office edited by J. V. Henderson and J.-F. Thisse, for Europe. 2119–71. New York: North Holland. Whyte, W. 1968. The Last Landscape. Garden Shi, K., B. Yu, Y. Huang, Y. Hu, B. Yin, Z. Chen, City, NY: Doubleday. L. Chen, and J. Wu. 2014. “Evaluating the Wu, J., L. Ma, W. Li, J. Peng, and H. Liu. 2014. Ability of NPP-VIIRS Nighttime Light Data to “Dynamics of Urban Density in China: Estimate the Gross Domestic Product and the Estimations Based on DMSP/OLS Nighttime Electric Power Consumption of China at Light Data.” IEEE Journal of Selected Topics Multiple Scales: A Comparison with DMSP- in Applied Earth Observations and Remote OLS Data.” Remote Sensing 6 (2): 1705–24. Sensing 7 (10): 4266–75. Small, C., C. D. Elvidge, and K. Baugh. 2013. Zhang, Q., and K. C. Seto. 2011. “Mapping “Mapping Urban Structure and Spatial Urbanization Dynamics at Regional and Connectivity with VIIRS and OLS Night Light Global Scales Using Multi-Temporal DMSP/ Imagery.” Paper presented at the Urban Remote OLS Nighttime Light Data.” Remote Sensing Sensing Event (JURSE), São Paulo, April of Environment 115 (9): 2320–29. 21–23. Zhang, Q., C. Schaaf, and K. C. Seto. 2013. “The Squires, G. D. 2002. Sprawl: C auses and Vegetation Adjusted NTL Urban Index: A C o n se q u e n c e s a n d Pol i c y R e s po n se s . New Approach to Reduce Saturation and Washington, DC: The Urban Institute. Increase Variation in Nighttime Luminosity.” Stansel, D. 2005. “Local Decentralization and Remote Sensing of Environment 129: 34-41. Local Economic Growth: A Cross-Sectional Zhou, N., K. Hubacek, and M. Roberts. 2015. Examination of US Metropolitan “Analysis of Spatial Patterns of Urban Growth Areas.” Journal of Urban Economics 57 (1): across South Asia Using DMSP-OLS Nighttime 55–72. Lig hts Data.” Applied Geog raphy 63: Sutton, P. C., T. J. Cova, and C. D. Elvidge. 2006. 292–303. “Mapping ‘Exurbia’ in the Conterminous Zhou, Y., S. J. Smith, K. Zhao, M. Imhoff, United States Using Nighttime Satellite A. Thomson, B. Bond-Lamberty, and C. D. Imagery.” Geocarto International 21 (2): 39–45. Elvidge. 2015. “A Global Map of Urban Extent Tewari, M., S. Alder, and M. Roberts. 2016. from Nightlights.” Environmental Research “Patterns of India’s Urban and Spatial Letters 10 (5): 54011. ECO-AUDIT Environmental Benefits Statement The World Bank Group is committed to reducing its environmen- tal footprint. In support of this commitment, we leverage elec- tronic publishing options and print-on-demand technology, which is located in regional hubs worldwide. Together, these initiatives enable print runs to be lowered and shipping distances decreased, resulting in reduced paper consumption, chemical use, greenhouse gas emissions, and waste. We follow the recommended standards for paper use set by the Green Press Initiative. The majority of our books are printed on Forest Stewardship Council (FSC)–certified paper, with nearly all containing 50–100 percent recycled ­ content. The recycled fiber in our book paper is either unbleached or bleached using totally chlorine-free (TCF), processed chlorine–free (PCF), or enhanced elemental chlorine–free (EECF) processes. More information about the Bank’s environmental philosophy can be found at http://www.worldbank.org/corporateresponsibility. With more than 70 percent of its population living in cities, Latin America and the Caribbean (LAC) is among the most urbanized regions in the world. Yet, even though LAC cities are, on average, more productive than those elsewhere in the world, their productivity lags that of North American and Western European cities. Closing this gap will help LAC countries raise their living standards and be among the world’s richest countries. Raising the Bar for Productive Cities in Latin America and the Caribbean explores the productivity of LAC cities and the factors that explain it. Using original empirical research, the book documents the relatively high population density, strong concentration of human capital in the largest cities, and other features of LAC cities that distinguish them from cities in the rest of the world. This book also explores how three key factors—urban form, skills, and access to markets—determine productivity in LAC cities. Although these cities benefit strongly from human capital and skills, they fail to reap the wider benefits of agglomeration. This is, in part, due to an inadequate enabling environment, as well as excessive congestion forces associated with infrastructure deficiencies and lack of administrative coordination within metropolitan areas. Further, the poor integration of LAC cities within countries contributes to large performance differences across cities and undermines cities’ aggregate contribution to national productivity. Raising the Bar will be of interest to policy makers, researchers, and the public at large. SKU 211258