R E C E N T F I N D I N G S R E C E N T F I N D I N G S This report was produced by the Poverty Global Practice, Africa Region, under the supervision of Pablo Fajnzylber. The Task Team Leaders were Theresa Osborne (Senior Economist GPV01) and Nadia Belhaj Hassine Belghith (Senior Economist GPV01). Regional Vice President: Makhtar Diop Country Director: Mark Lundell Country Manager: Coralie Gevers This volume is a product of the staff of the International Bank for Reconstruction and Development/ The World Bank. The findings, interpretations, and conclusions expressed in this paper do not necessarily reflect the views of the Executive Directors of The World Bank 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. The material in this publication is copyrighted. Copying and/or transmitting portions or all of this work without permission may be a violation of applicable law. The International Bank for Reconstruction and Development/ The World Bank encourages dissemination of its work and will normally grant permission to reproduce portions of the work promptly. For permission to photocopy or reprint any part of this work, please send a request with complete information to the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA, telephone 978-750-8400, fax 978-750-4470, http://www.copyright.com/. All other queries on rights and licenses, including subsidiary rights, should be addressed to the Office of the Publisher, The World Bank, 1818 H Street NW, Washington, DC 20433, USA, fax 202-522-2422, e-mail pubrights@worldbank.org.  iii Contents Acronyms.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v INTRODUCTION  Poverty and Employment in Madagascar 2001–2012: A Synthesis of Recent Findings. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 CHAPTER 1  Madagascar Poverty and Inequality Update: Recent Trends in Welfare, Employment, and Vulnerability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 Updated Poverty Statistics.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 The Distribution of Growth and Changes in Inequality. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 Trends in Agriculture and Employment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 Risks and Vulnerability. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 Annex 1A. Sampling, Comparability between 2010 and 2012 Data and Weights Issues. . . . . . . . . . 26 Annex 1B. Poverty Estimation Methodological Notes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 Annex 1C. Detailed Data and Results Tables.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 CHAPTER 2  Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010). . . . . . . . . . . . . . . . . . . . . . . . . . 39 Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 Introduction.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 Overview of Poverty in Madagascar. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 Methodology. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 Determinants of Urban-Rural Inequality in 2010.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 Determinants of Changes in Consumption and Inequality between 2005 and 2010, National Sample. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 Determinants of Changes in Consumption and Inequality between 2005 and 2010, Rural Households Only. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 Explaining Changing Patterns in Returns .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 Annex 2A.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 Annex 2B. Methodology.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 Annex 2C. Determinants of Urban-Rural Inequality in 2010. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 Annex 2D. Determinants of Inequality between 2005 and 2010. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 Annex 2E. Determinants of Rural Inequality between 2005 and 2010. . . . . . . . . . . . . . . . . . . . . . . . . . . 79 CHAPTER 3  Flexible Poverty Profiling and Welfare Prediction in Madagascar . . . . . . . . . 85 Introduction and Key Findings. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 Data and Methods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 Results.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 Discussion and Conclusion.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 Annex 3A. Explanation of Methods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 CHAPTER 4  Labor Demand Estimation in Rural Madagascar: Shadow Wages and Allocative Inefficiency.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 Introduction.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 Background and Data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 Data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 Madagascar’s Agricultural Sector. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 NFEs in Rural Madagascar. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 Empirical Strategy.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116 Results.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 Discussion and Conclusion.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 Annex 4A. Tables.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125 Annex 4B. Summary Statistics by Gender.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 CHAPTER 5  Transactions Costs, Poverty, and Low Productivity Traps: Evidence from Madagascar’s Informal Microenterprise Sector . . . . . . . . . . . . . . . . . . . . . . . . . 139 Summary of Results and Policy Implications. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 Introduction.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141 Firm Size Productivity Relationships. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 Data and Characteristics of Madagascar’s OOME Sector. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144 The Impact of Scale: Estimation Method and Results.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146 Labor Market Frictions.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 Profit Elasticities and Low-Productivity Traps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150 The Returns to Microentrepreneurs’ Labor. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150 Toward a Unified Theory of OOMEs: Asymmetric Information and Incomplete Markets. . . . . . . . . 152 Madagascar’s Weak Enforcement and Monitoring Infrastructure. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155 Formal Registration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155 Conclusions and Agenda for Further Research. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157 Annex 5A. Tables.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159 Photos Cover: Gudkov Andrey/Shutterstock; p. vi: Danm12/Shutterstock; p. 3: Dennis van de Water/Shutterstock; p. 7: Damien Ryszawy/Shutterstock; p. 8: Agota Kadar/Shutterstock; p. 12: Dudarev Mikhail/Shutterstock; p. 38: Stock photo/Shutterstock; p. 58: Anton Ivanov/Shutterstock; p. 61: Dudarev Mikhail/Shutterstock; p. 65: Elisabeth/Flickr © All rights reserved; p.69: Dennis van de Water/Shutterstock; p. 84: Milosk50/Shutterstock; p. 98: Dennis van de Water/Shutterstock; p. 100: Artush; p. 106: Olivier S/Shutterstock; p. 123: Damien Ryszawy/Shutterstock; p. 138: Pierre Jean Durieu/Shutterstock; p. 140: Danm12/Shutterstock; p. 158: Anton Ivanov/Shutterstock; back cover: Dietmar Temps/Shutterstock  v Acronyms AIF allocative inefficiency factor CART classification and regression tree ENEMPSI Enquête nationale sur l’emploi et le secteur informel (National employment and informal sector survey) ENSOMD Enquête Nationale sur les Objectifs Millenaire du Développement (National Survey on the Millennium Development Goals) EPM Enquête Périodique auprès des Ménages (Periodic Household Survey) FAOSTAT Food and Agriculture Statistics HH Household INSTAT Institut nationale de la statistique. (National Institute of Statistics of Madagascar) LFS labor force survey MRP marginal revenue product MSE mean-squared error NFE nonfarm enterprise NPK Nitrogen, Phosphorus, Potassium (fertilizer) OLS ordinary least squares OOME owner-operated microenterprises PPP purchasing power parity RF random forest RT regression tree SSA Sub-Saharan Africa TLU tropical livestock unit WDI World Development Indicators  1 INTRODUCTION Poverty and Employment in Madagascar 2001–2012: A Synthesis of Recent Findings Theresa Osborne June 2016 AFR Poverty Practice, World Bank M adagascar remains among the poorest Madagascar’s economy faces an array of challenges countries in the world, and has shown little in reducing poverty, including an unfavorable invest- improvement in indicators of the well-being ment climate, severe infrastructure deficits, and political of its population over recent years. Despite its unique instability (World Bank 2015). In addition, from 2001 biodiversity and abundant mineral, water,1 and labor to 2012, Madagascar experienced two political crises (in resources, it ranks among the relatively few countries in 2002 and 2009); the loss of valuable trade preferences, the world with real per capita gross domestic product with the 2005 end of the multifiber agreement, and (GDP) in 2010 lower than it was in 1960. Only the the 2009 revocation of African Growth and Opportu- Democratic Republic of the Congo and Liberia, two nity Act (AGOA) preferences;3 and a number of severe countries which have undergone periods of civil war, droughts, cyclones, and other natural shocks. have experienced a greater decline (figure I.1). As a result, Madagascar rates as the poorest country in Sub- This report synthesizes the insights obtained from a Saharan Africa (SSA) (and the world) where internation- series of five papers on poverty, inequality, labor mar- ally comparable data are available (figure I.2 maps kets, and returns to agricultural and nonfarm enter- poverty rates in SSA). This poverty is associated with low prises in Madagascar over the period 2001–12. These and declining labor productivity. By 2012, Madagascar’s papers draw on a combination of empirical techniques, GDP per employed worker had fallen to the lowest in household living standards data, and firm-level data the world with the exception of the Democratic Republic to elucidate key dynamics and structural issues driving of the Congo (figure I.3).2 poverty and welfare (in all cases measured as per capita FIGURE I.1: The Countries of the World (with Data FIGURE I.2: Headcount Poverty Rates, SSA Available) Showing Lower Real GDP per Capita in 2010 than 1970 0% Percent total decline in real GDP Zambia –10% Central Africa Republic per capita 1960–2010 Venezuela Togo Madagascar Sierra Leone Mauritania –20% Burundi Congo, Dem. Rep. –30% Cote d’Ivoire Nicaragua –40% Zimbabwe Kiribati Niger –50% Liberia –60% –70% –80% Yellow-orange = higher rates of extreme poverty. Source: World Development Indicators (WDI). Source: PovcalNet. 2 Republic of Madagascar Employment and Poverty Analysis FIGURE I.3: GDP per Employed Person (Constant $US, 1990 PPP) 4,500 4,000 3,500 3,000 2,500 2,000 1,500 1,000 500 0 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 Congo, Dem. Rep. Madagascar Niger Ethiopia Tanzania Cameroon Kenya Uganda Zambia Burkina Faso Source: WDI 2016. Note: PPP = purchasing power parity. consumption) over this dozen-year period. First, “Mada- the prediction model and isolate the most important gascar Poverty and Inequality Update: Recent Trends in predictive variables. The paper “Labor Demand Estima- Welfare, Employment, and Vulnerability” (Belghith, Ran- tion in Rural Madagascar: Shadow Wages and Allocative driankolona, and Osborne 2016) updates recent poverty Inefficiency” (Jodlowski 2016) uses multistage econo- and inequality trends since the World Bank (2014) pub- metric estimation to analyze the determinants of labor lication Face of Poverty in Madagascar: Poverty, Gender demand by rural households, both in their agricultural and Inequality Assessment. It also documents trends in and off-farm activities. Finally, in “Transactions Costs, key outcomes in both agriculture and labor markets. Poverty, and Low Productivity Traps: Evidence from Second, “Isolation, Crisis, and Vulnerability: A Decom- Madagascar’s Informal Micro-Enterprise Sector,” Bi and position Analysis of Inequality and Deepening Poverty in Osborne (2016) analyze the performance of urban- Madagascar (2005–2010)” (Thiebaud, Osborne, and Bel- based, informal owner-operated microenterprises with ghith 2016) uses re-centered influence function estima- respect to productivity and employment creation, using tion to decompose Madagascar’s rural-urban inequality econometric methods that account for possible selection into disparities in household and community attributes, bias. The main insights of these papers and their policy circumstances, and assets; as well as differential returns implications are collected here. to these assets. Using the same technique, it then decom- poses changes in per capita consumption between 2005 Although conditions have been extremely unfavorable for and 2010 into explanatory factors by quintile of the poverty reduction—with real per capita GDP declining consumption distribution. In “Flexible Poverty Profiling between 2001 and 2012—the poverty headcount rate has and Prediction of the Severity of Poverty in Madagascar” stabilized at approximately its 2001 level.4 Households (McBride and Osborne 2016), the authors use “machine were buffeted by a variety of climatic and economic learning” algorithms to profile and predict levels of wel- shocks, but the poor have adopted flexible strategies to fare in a manner that allows the data to iteratively shape return their living standards to previous levels. Poverty and Employment in Madagascar 2001–2012: A Synthesis of Recent Findings 3 Given the nature of macroeconomic and political events TABLE I.1: Trends in Poverty and Inequality over the period, urban poverty rates fluctuated more (National Basic Needs Poverty Rates) widely than (and sometimes in opposite directions to) rural ones. Between 2001 and 2010, the national poverty 2001 2005 2010 2012 rate moved in line with the urban headcount poverty rate. Poverty gap index (mean percentage shortfall of poverty line) Between 2001 and 2005, both rose, but then they fell again Urban 11.8 13.6 8.9 11.8 in 2010, even as the rural poverty rate rose (Table I.1). Rural 40.5 34.8 36.7 36.4 This pattern no longer obtained between 2010 and 2012, Total 35.9 31.3 32.0 32.2 however, when reductions in rural poverty offset increased Headcount poverty rate (percentage of the population) poverty in urban areas to produce a slight decline in the Urban 34.1 40.8 29.8 35.5 national poverty rate to its 2001 level. Rural 77.7 79.6 80.1 77.9 With approximately 78 to 80 percent of the rural popu- Total 70.8 73.2 71.7 70.7 lation remaining poor throughout the period, perhaps a Inequality indicators more meaningful indicator of rural poverty is the pov- Gini 46.9 38.9 42.7 41.0 erty gap index, which measures the severity of poverty. Coefficient Over the years 2001 to 2012, this index moved in oppo- P90/P10 8.13 4.96 6.01 6.32 site directions in rural and urban areas. The national Source: Belghith, Randriankolona, and Osborne 2016, using Enquête poverty gap index finished lower in 2012: on average, Périodique auprès des Ménages (EPM) and Enquête Nationale sur les the poor lived on 32.2 percent less than the poverty line, Objectifs Millenaire du Développement (ENSOMD). relative to 35.9 percent less in 2001 (Belghith, Randri- ankolona, and Osborne 2016). agriculture fell, labor shifted into nonfarm enterprises, Over the period 2001–12, the population responded and primary and secondary employment in services in to fluctuations in returns to their income-generating particular rose (Belghith, Randriankolona, and Osborne activities by shifting effort and resources into and out 2016). The poor accumulated assets, including more of agriculture, services, and manufacturing sectors. As education and transportation assets (Thiebaud, Osborne, the returns in urban-based sectors fell in 2005, employ- and Belghith 2016). However, these strategies could not ment shifted into agriculture; when in 2010 returns in fully offset the weak demand for labor. In 2010, those 4 Republic of Madagascar Employment and Poverty Analysis seeking but not finding work increased, even as second- FIGURE I.4: Incidence of Consumption Growth (Total) ary employment rose. Wages increased just slightly, and 2005 to 2010 only for male workers in 2010, but then returned to 10 their former (2005) levels (Belghith, Randriankolona, 8 and Osborne 2016). 6 Inequality in Madagascar fluctuates significantly over 4 Annual growth rate, % time in response to climatic and price shocks. Particu- 2 larly severe weather shocks in 2010 resulted in a decline in well-being for those at the bottom of the consump- 0 tion distribution in that year vis-à-vis 2005. In com- –2 bination with increasing returns to urban-based work –4 relative to those in 2005, this led to a regressive growth pattern and increased inequality (Figure I.3). Although –6 this implies significant consumption risks and welfare –8 losses, one cannot assess persistent inequality (inequality –10 in households’ lifetime living standards) without track- 0 20 40 60 80 100 ing households over time (with panel data.) Mada- Expenditure percentile gascar’s level of inequality as measured by the ratio of consumption for the top decile to that of the bottom Source: Calculations using EPM 2005, 2010, reported in Belghith, Randriankolona, and Osborne 2016. decile (P90/P10) ranged from 5 to 8 over the period—a low level relative to the 13.4 average for low-income countries. 2005–10, despite having accumulated more assets, the An important explanation for Madagascar’s persistent rural population was unable to completely offset a poverty is its lack of progress in generating remunerative decline in the returns to agriculture through entry into employment in the nonagricultural and urban sectors. As off-farm work. Returns to land fell by 6 percent—further shown in Thiebaud, Osborne, and Belghith (2016), the for the poor—and health shocks compounded the toll returns to education and work are higher in urban areas. (figure I.6).5 In addition, key attributes of urban communities—in particular greater access to markets, health centers, The key factors reducing agricultural incomes in 2010 and other services—increase welfare (all else equal). In were domestic rice policies and deteriorating transport addition, rural households have been more adversely conditions, which weakened internal market integration. In impacted by climatic risks. Between 2005 and 2010, the the face of rising world prices for rice, Madagascar’s over- returns to economic activities in rural areas fell signifi- whelmingly dominant crop and staple food, in 2007 the cantly for all quintiles of the consumption distribution. government removed tariffs on rice imports and decreased This, plus climatic shocks and to a lesser extent health ad valorem taxes, then in July of 2008 removed the value- shocks, explain the decline in welfare in the bottom two added tax on rice imports completely. Anticipating drought quintiles over these two years captured in figure I.4. and further increases in the world price, the government of Figure I.5 shows the main determinants of consump- Madagascar preordered rice imports (50,000 metric tons tion changes by quintile and their direction of influence. of Indian rice) and banned rice exports. These measures As households responded by seeking greater off-farm kept the price of rice relatively stable for consumers, yet employment, as shown, male-headed households were producers were unable to benefit from rising world prices. more successful than female-headed ones in offsetting In addition, rising transport costs reduced rural earnings. these losses. This was because they were better able to Between 2005 and 2010, the average real price to trans- secure employment in services sectors with apparently port a 50 kilogram bag of rice rose 42 percent—from higher profitability, whereas females were more likely $US1.40 in 2005 to $US2.00 (2005 dollars), and to a to find such employment in the primary and indus- higher level—$2.20 for the lowest consumption quintile trial (light manufacturing) sectors. Yet, over the period of the population, using Enquête Périodique auprès des Poverty and Employment in Madagascar 2001–2012: A Synthesis of Recent Findings 5 FIGURE I.5: Main Determinants of Change in Consumption (2005–10) 20% 15% Counterfactual change 10% in consumption 5% 0% –5% –10% –15% –20% –25% 20 40 60 80 Quintiles Returns to rural area Effects of climate shocks Effects of health shocks Returns to male household head Source: Thiebaud, Osborne, and Belghith 2016. Note: Effects smaller than 2% or not significant for the bottom quintile are not pictured. FIGURE I.6: Main Determinants of Change in Consumption (Rural Households, 2005–10) 20% 15% Counterfactual change in consumption 10% 5% 0% –3.8% –5% –6.4% –6.1% –6.3% –10% 20 40 60 80 Quintiles Effects of climate shocks Returns to cultivated land Effects of health shocks Returns to male household head Source: Thiebaud, Osborne, and Belghith 2016. Ménages (EPM) 2005, 2010. As a result, between 2005 of more severe poverty are the following, in order of and 2010, the ratio of paddy prices to fertilizer dropped importance:6 precipitously and was more closely correlated with con- sumption than before (figure I.7). 1. Living in a community with levels of electrification less than 27 percent of households These findings are supported further by a flexible profile of the severity of poverty (McBride and Osborne 2. Having a non-university-educated household head 2016). Of the many available household, regional, and (Having a university education makes it very likely community-level variables which one might expect to be to have higher incomes in urban areas, as would be correlated with welfare, those that are most predictive expected.) 6 Republic of Madagascar Employment and Poverty Analysis FIGURE I.7: Average Nominal Price Received for Rice Paddy (by Consumption Quintile) 8 7 6 5 4 3 2 1 0 Poorest Second Third Fourth Richest 2005 NPK/Paddy 2010 NPK/Paddy 2005 Urea/Paddy 2010 Urea/Paddy Sources: EPM 2005 and EPM 2010. 3. Having an illiterate head of household (Other inter-temporal decomposition of changes in consump- distinctions in educational attainment have little tion (Thiebaud, Osborne, and Belghith 2016) shows that predictive power.) declines in the returns to land are strongly associated with more severe poverty, and in 2010 the households 4. Living in greater remoteness from the nearest major facing lower rice prices have lower consumption. Thus, urban center (This variable predicts welfare better the benefits to poor net consumers are more than offset than other measures of access to services.) by the decline in the incomes of poor rice producers. 5. Receiving lower prices for paddy rice Both papers show that this is the case even at the bottom of the distribution. 6. Having lower livestock holdings Our findings on the role of electricity (McBride and For agricultural households analyzed separately, the Osborne 2016; Thiebaud, Osborne, and Belghith 2016) key predictive variables in order of importance are the provide indicative but not conclusive evidence on its following: importance for alleviating poverty. Electrification may simply proxy for community-level wealth or economic 1. Lower cultivated land activity, the effects of which are difficult to disentangle from those of electricity provision. Nonetheless, the com- 2. Remoteness from the nearest major urban center bined results are suggestive of a positive causal effect on 3. Living in a community with lower levels of incomes. First, we consider the possibility that electricity is electrification merely a proxy for other urban attributes. We examine the correlation between the level of a community’s electrifica- 4. Receiving a higher percentage of revenues from tion and an indicator for urban area, regional indicators, agriculture and all other indicators of remoteness which would be 5. Receiving a lower price for paddy rice expected to be correlated with urban agglomeration. We find that even controlling for these variables, the degree of The combined results of Thiebaud, Osborne, and Bel- electrification varies substantially, within and across rural ghith (2016) and McBride and Osborne (2016) provide and urban areas of varying degrees of remoteness and evidence that, while the effects of rice prices are always access to services (McBride and Osborne 2016).7 Electrifi- somewhat heterogeneous within a population, low rice cation was a more powerful predictor of welfare than any prices in Madagascar increase poverty in rural areas. An other of the available indicators of spatial advantage or Poverty and Employment in Madagascar 2001–2012: A Synthesis of Recent Findings 7 economic density. Moreover, Thiebaud, Osborne, and Bel- FIGURE I.8: Percent of Households with Electricity ghith (2016) show that between 2005 and 2010, increased in Community (by Consumption Quintiles, access to electricity was associated with a small but posi- 2005 and 2010) tive and statistically significant change in consumption for 60% parts of the consumption distribution. Over the period the 2005 2010 percentage of households with electricity increased only 50% slightly, from 15 percent to 17 percent, but the correla- tion between electrification and consumption increased. 40% Electrification reached 0.5 percent more households in the third quintile of consumption in 2010 consumption than 30% it did in 2005, 3.2 percent more of the fourth quintile, and 8.5 percent of the top quintile, and less of the bottom 20% two (figure I.8). Since the ability to shift into off-farm work was a key strategy for coping with poor returns to 10% agriculture in 2010, communities with better access to 0% electricity were likely better able to support more produc- Poorest Second Third Fourth Richest tive nonfarm enterprises (NFEs). Jodlowski’s (2016) find- ing that greater electrification was significantly correlated Note: Data are not representative at the community level. Sources: EPM 2005, 2010. with NFE revenues in 2010 (but not in other years) sup- ports this hypothesis. Electricity can raise incomes where there is potential for NFEs and a certain level of demand, but there is no evidence available that it can do so in the transport costs, worsening terms of trade in agriculture, remotest and poorest areas.8 and declining consumption (Belghith, Randriankolona, and Osborne 2016; Thiebaud, Osborne, and Belghith In addition, our findings underscore the importance of 2016). Moreover, the time to reach urban centers and reducing transport costs for poverty reduction. First, health centers is not correlated only with poverty but in 2001 and 2005, higher transport costs were associ- is highly predictive of severe poverty (McBride and ated with lower levels of rural NFE revenue (Jodlowski Osborne 2016). These results are unlikely to simply 2016). Although similar effects are no longer evidenced reflect a migration of poorer people to more remote in 2010, there is a close association between higher areas. Rather, the political crisis of 2009 reduced the 8 Republic of Madagascar Employment and Poverty Analysis availability of funding for road maintenance at a time presented in Jodlowski (2016), rural labor markets, when oil prices were high relative to 2005.9 where households are the primary employers, are subject to considerable frictions. This results in low demand More productive and remunerative off-farm employ- and a low willingness to pay a (shadow) wage, whether ment is the primary route out of poverty, in Madagascar to employ labor on the farm or in NFEs. Although it as well as in other poor agricultural economies, but it is not possible with the available data to identify the requires that constraints to larger, more efficient enter- precise source of the frictions, they likely represent some prises be alleviated. An examination of rural labor mar- combination of the risks and (nonfinancial) costs of kets (Jodlowski 2016) as well as a study of urban-based identifying, training, supervising, and releasing workers informal enterprises (Bi and Osborne 2016) suggest that (Jodlowski 2016). Thus, the source and magnitude of the informal microenterprise structures prevailing in these costs may differ by community and by household Madagascar result in a major misallocation of resources attributes. They may also vary between farm and off- in the economy. While such enterprises provide a means farm enterprises for the same household. On average, to generate a livelihood, the productivity and income Jodlowski finds the estimated size of the friction to be losses resulting from this structure are substantial. As greater in the on-farm sector than in NFEs.10 At the same they do in many other poor countries where informal time, the demand for on-farm labor is relatively respon- microenterprises have been studied, they do not tend sive to the shadow wage, whereas NFE labor is unre- to scale up and employ more workers over time. As sponsive. This suggests that in agriculture labor input Jodlowski (2016) and Bi and Osborne (2016) show, they is easier to adjust as needed, as profitability conditions underemploy workers. Therefore, as the primary source change. Agricultural laborers may also be more easily of off-farm employment, this configuration of produc- substituted for each other, as the tasks performed are tion reduces employment and wages, making the rural- less complex or specialized. For NFEs, however, there urban transition much more difficult. appear to be greater rigidities involved in adjusting labor inputs, which deters these enterprises from hiring more Despite the flexible coping strategies Madagascar’s pop- workers. To the extent that labor input is adjusted, the ulation exhibited over the years covered by this report main flexibility is in the hours of existing workers rather (2001–12), as currently configured, the potential for than the number of different employees hired. NFE labor Madagascar’s rural labor markets to generate more pro- may require more effort to find, train, and supervise. ductive employment remains low. According to estimates More skilled or specialized workers may also expect a Poverty and Employment in Madagascar 2001–2012: A Synthesis of Recent Findings 9 more regular contract, even if this contract is informal, the mean wage. Owners of multiworker OOMEs, however, and this implies greater risk to the employer. Thus, rural earn a “wage” premium of approximately 68 percent of nonfarm entrepreneurs prefer to accept lower expected the mean wage, controlling for individual characteristics profits than to incur these costs. The primary vehicle of (excluding returns to capital). employment by household-owned and -operated NFEs, therefore, is through self-employment, with minimal The persistent prevalence of microenterprises that are potential for generating employment for other workers. too small to be productive can be explained by market failures—high transactions costs and risks—that are par- The potential for remedying labor market frictions is ticularly difficult to overcome in a poor economy. OOMEs unclear. The primary source of these frictions appears to (correctly) perceive a lack of demand for their products to be market failures related to risks and the challenge of be their most immediate constraint, but it would be more incentivizing workers to be productive, honest employees efficient for there to be fewer, larger enterprises serving (“agents”) acting largely in accordance with the objec- the same level of demand than numerous OOMEs. Given tives as the owner of the enterprise (the “principle”). increasing returns to capital, OOMEs could in principle These principle-agent problems increase the transactions increase their profitability by expanding a little at a time, costs of employing workers and cause market failures reinvesting growing profits and growing to a more efficient that household enterprises overcome by employing family and profitable scale. To explain their lack of growth, there- members, first and foremost, and, in the vast majority of fore, requires a combination of conditions. cases, by staying small. Because the benefits of incurring these transactions costs must exceed their costs in order One issue is the lack of entry by larger, more efficient, for enterprises to hire more workers, where the marginal typically formal firms, which would provoke a restruc- returns to labor are low, transaction costs may loom too turing of the market and draw workers into more remu- large as a percentage of these gains. Thus, in principle, nerative work. In addition, an economy characterized by actions that would improve the profitability of enterprises, poor entrepreneurs (the OOMEs) inhibits their growth. such as investments in public infrastructure, would trans- The marginal utility of consumption for entrepreneurs’ late into jobs. At present, however, there is no compelling poor families is high, while returns to investing small evidence (from Jodlowski) that improving the profitability incremental amounts in tiny enterprises are low. In the of rural NFEs in this manner would result in a significant presence of increasing returns, firms must be created increase in either wages or employment by NFEs. larger or grow rapidly to enjoy those returns. Thus, constrained to consume all of their income, entrepre- Similarly, the potential for urban-based, microenterprises neurs lacking sufficient external financing cannot grow to generate greater employment and economic gains is their businesses. Breaking out of this low-productivity limited. Using detailed 2012 data on informal, owner- equilibrium (or “trap”) would require a more substantial operated microenterprises (OOMEs) in Madagascar’s cities increase in scale than poor households can afford. At the and towns, Bi and Osborne (2016) assess the potential of same time, transactions costs associated with external these enterprises to achieve higher incomes for their own- financing are high. Due to the difficulties associated ers and offer remunerative employment to workers. For with monitoring the use of firm resources, whether by both single-worker OOMEs—those which employ only creditors or potential partners (another principle-agent their owner—and multiworker OOMEs, which employ problem), the transactions costs of credit, partnership family, other unpaid, as well as paid labor, the activities arrangements, and share capital are high. Microcredit, pursued range from logging and mining to household if available, must come with interest rates adequate to services, transport services, and light manufacturing. Sev- cover the costs of screening and enforcement of repay- enty percent of OOMEs are the single-worker variety, and ment, and these costs are higher on a per-dollar (or this variety has significantly lower returns to capital and ariary) amount for small loans. Similarly, partnerships do to the owner’s labor due to unexploited “profit” econo- not form for the purposes of expansion precisely because mies of scale.11 The wage penalty of owning and operating entrepreneurs’ level of investible capital is low relative to a single-worker OOME rather than working for others these transactions costs. Finally, the frictions associated (controlling for worker ability, characteristics, sector, and with employing and incentivizing workers further hinder location of employment) is approximately 60 percent of firms’ profitability and growth. 10 Republic of Madagascar Employment and Poverty Analysis FIGURE I.9: Credit to the Private Sector as a Percent them created them through some type of loan, most of of GDP, Comparison Countries and Low Income which were informal. Thus, efforts to speed the develop- Average (Average 2011–14) ment of a financial sector capable of allocating financial savings to the most promising investments could be 30 effective in stimulating more dynamic change. Mada- 25 gascar’s financial markets are underdeveloped relative to other low-income and SSA countries, with a low level Percent of GDP 20 of credit to the private sector as a percentage of GDP (figure I.9). One key step would be to improve the moni- 15 toring and enforcement environment for credit, partner- 10 ships, associations, and corporate investors. In 2012, for example, credit registries and bureaus were almost non- 5 existent in the country, and the strength of legal rights to enforce repayment rated only 2 out of a 10-point scale in 0 Doing Business’s Getting Credit.12 Finally, depending on e r ia er da e ca m qu an the status of existing financial institutions, policy mak- ig an as co bi N nz Ug ag in am Ta ers could consider enhanced liquidity or risk mitigation ad w oz Lo M M measures to expand the capacity to serve larger “micro” Source: WDI. and small and medium-sized enterprises.13 This series of papers also provides insights on gender- related disparities in opportunities in Madagascar. The tentative policy implications of these findings are as Although female-headed households are not consistently follows: First, it would likely have little effect to encour- poorer than male-headed ones (McBride and Osborne age informal microenterprises to simply register without 2016), men earn significantly higher wages than women taking additional steps to improve the credibility of their (Belghith, Randriankolona, and Osborne 2016). When financial statements and integrity of firm resource uses. educational attainment, region, and urban milieu are Rather, levers to strengthen the information environment considered, men earned 37 percent more than women (through adoption and verification of accounting prac- in the labor market in 2012 (Bi and Osborne 2016). tices, credit reporting, and other means) would be needed Female entrepreneurs are less likely to own and operate to reduce the transactions costs for potential creditors a multiworker microenterprise and more likely than men and partners. In addition, while microloans to the tiniest, to own and operate their less profitable single-worker single-worker OOMEs may help to provide employment versions. Among single-worker firms, men earn higher for their poor owners, they would have little impact on profits, all else equal (Bi and Osborne 2016), and appear overall productivity, employment, and wage growth. to face fewer obstacles in undertaking certain economic activities than women, a disparity in access to opportu- Ultimately, significantly reducing the misallocation nities that widened substantially in 2010, as shown in of capital and labor in Madagascar’s economy would Thiebaud, Osborne, and Belghith (2016). require a steadily growing presence of larger, more formal firms that compete for markets (Aghion, Akcigit, These findings raise additional issues for further inves- and Howitt 2013). At the same time, a significant tigation: In particular, further investigation into the improvement in productivity could be attained through sources of labor market frictions, possible mitigating the alleviation of constraints facing OOMEs that have factors, and policy levers would be beneficial, both for already achieved a certain scale, that employ workers, formal and informal job creation. It would be worth- and that demonstrate basic entrepreneurial skills. Given while exploring further the constraints faced by female the presence of increasing returns, such firms could entrepreneurs, as well as collecting higher quality rural invest more and hire more workers if they had access farm and NFE data, perhaps detached from the EPM. to external financing. A full 92.5 percent of OOMEs Finally, as infrastructure improvements are made, it received their assets via a gift or inheritance or created would be extremely valuable to evaluate the impacts of them with their own savings, and only 1.2 percent of them on incomes and well-being in a rigorous manner. Poverty and Employment in Madagascar 2001–2012: A Synthesis of Recent Findings 11 NOTES REFERENCES 1. Madagascar ranks fifth in Sub-Saharan Africa in renewable water resources per capita, according to WDI data. Aghion, P., U. Akcigit, and P. Howitt. 2013. “What Do 2. Based only on countries with data. We Learn from Schumpeterian Growth Theory?” 3. Madagascar’s AGOA preferences were reinstated in 2014. 4. Because poverty headcount rates moved in opposite directions to Schumpter Lecture presented at the Swedish growth in two of the three subperiods (in particular, 2005–10 and Entrepreneurship Forum, Stockhim, January 2013. 2010–12), the elasticity of poverty with respect to growth would http://scholar.harvard.edu/files/aghion/files/what_do_ have the wrong sign for these and the full period, and it is not considered an informative measure of the upside potential of positive we_learn_0.pdf?m=1361377935. growth to reduce poverty in the country. Belghith, N., P. Randriankolona, and T. Osborne. 2016. 5. It is not clear to what extent adverse health shocks caused greater poverty versus the decline in incomes causing adverse health. “Madagascar Poverty and Inequality Update: Recent 6. This analysis excludes household size and composition variables, Trends in Welfare, Employment, and Vulnerability.” which overstate the adverse welfare effect of large households with more children using the per capita income welfare indicator. Washington, DC: World Bank. 7. A multivariate regression of electrification on the full set of available Bi, C., and T. Osborne. 2016. “Transactions Costs, geographic and remoteness variables leaves 42 percent of this varia- tion unexplained. Poverty, and Low Productivity Traps: Evidence from 8. The “returns” to electricity were estimated to have fallen for the Madagascar’s Informal Micro-Enterprise Sector.” bottom quintile between 2005 and 2010 (Thiebaud, Osborne, and Washington, DC: World Bank. Belghith 2016). 9. Average world oil prices were 48 percent higher in 2010 than 2005. Jodlowski, M. 2016. “Labor Demand Estimation in See World Bank, “Commodity Markets,” database, http://www Rural Madagascar: Shadow Wages and Allocative .worldbank.org/en/research/commodity-markets. 10. These differences may not be statistically significant or stable over Inefficiency.” Washington, DC: Cornell University time, as the farm-based estimates are for one year only. and the World Bank. 11. Since returns are measured in terms of profits rather than decreasing costs, they represent the combined effects of declining average costs McBride, L., and T. Osborne. 2016. “Flexible Poverty and increased market power. Profiling and Prediction of the Severity of Poverty in 12. Although these indicators had improved by 2016, they still show weak performance. Madagascar.” Washington, DC: Cornell University 13. Any interventions would ideally be designed to ensure that (i) the and the World Bank. information and monitoring environments were also improving, Thiebaud, A., T. Osborne, and N. Belghith. (ii) access to assistance was competitive and fairly distributed among individual and group enterprises, and (iii) subsidies did not undercut 2016. “Isolation, Crisis, and Vulnerability: A other developments in credit markets. Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010). Washington, DC: World Bank. World Bank. 2014. Face of Poverty in Madagascar: Poverty, Gender, and Inequality Assessment. Washington, DC: World Bank. 12  13 CHAPTER 1 Madagascar Poverty and Inequality Update: Recent Trends in Welfare, Employment, and Vulnerability Nadia Belhaj Hassine Belghith Patrick Leon Randriankolona Theresa Osborne June 2016 The authors thank Rachel Wang and Alessia Thiebaud for capable research assistance and Kristen Himelein for helpful advice. Summary F or Madagascar, the years 2001–12 lacked both in urban settings, whereas the depth of poverty for the economic growth and progress in alleviating poorest of the poor is affected to a greater extent by poverty. Political, economic, and climatic shocks conditions in rural areas. caused fluctuations in poverty, producing an increase in headcount poverty from 2001 to 2005, followed by a modest decline for several years, and a rise in 2012 when Updated Poverty Statistics the headcount poverty rate stood at 70.7 percent, using the national poverty line, and essentially the same rate as Despite adverse conditions over the past decade, trends in 2001 (70.8). From 2001 to 2013, perhaps somewhat in poverty headcount rates in Madagascar have stabi- surprisingly, increases in the headcount poverty rate lized, albeit at a high rate, approximately 71 percent.1 were accompanied by decreases in the severity of poverty Table 1.1 shows updated estimates of the national pov- and inequality and vice versa. Households living near erty rate in years with consumption data available: 2001, the poverty line—that is, those which are less poor than 2005, 2010, and 2012. As real per capita gross domestic 70 percent of the population—are buffeted by conditions product (GDP) declined from US$294 to US$267 (in TABLE 1.1: World Bank Revised Headcount Poverty Estimates (National Basic Needs Poverty Line), Earlier Estimates, and Real GDP per Capita in Years with Consumption Data Year 2001 2005 2010 2012 Official poverty estimates 69.7% 68.7% 76.5% 71.2% Percent of population in poverty, earlier estimates (World Bank 70.8% 75.0% 75.3% n.a. 2014) Total (percent of population) in absolute poverty, final revised 70.8% 73.2% 71.7% 70.7% GDP per capita in 2005 U.S. dollars $294.0 $275.5 $273.2 $267.2 Sources: Bank staff using Enquête Périodique auprès les Ménages (EPM), Enquête Nationale sur les Objectifs Millenaire du Développement (ENSOMD), and World Development Indicators (WDI). Note: Poverty line is estimated using 2010 EPM survey and adjusted for inflation in each year. 14 Republic of Madagascar Employment and Poverty Analysis 2005 U.S. dollar purchasing power parity, PPP), poverty TABLE 1.2: Poverty Rates Using International Poverty headcount rates based on the basic needs approach rose Lines, Povcalnet Method (Corrected Sampling from 70.8 percent in 2001 to 73.2 percent in 2005, then Weights, no Application of Regional Deflation) fell slightly to 71.7 percent in 2010 and to 70.7 percent in 2012—a return to their 2001 level. Year 2001 2005 2010 2012 US$1.90 2011 PPP 68.7 74.1 81.8 77.8 These estimates vary from those published earlier and US$3.10 2011 PPP 84.1 89.9 92.9 90.5 indicate that the declining trend observed in the head- US$1.25 2005 PPP 76.7 80.7 84.8 83.9 count poverty rate as of 2012 (INSTAT 2014) began US$2.00 2005 PPP 88.2 92.1 93.6 93.3 earlier than previously thought. As with any poverty Sources: EPM and ENSOMD. measurement, the precise methods used to estimate Note: PPP = purchasing power parity. the welfare aggregate (typically consumption) and the poverty line can have an impact on levels and, in some cases, trends. Headcount poverty estimation is sensitive TABLE 1.3: Trends in the Poverty Gap (Mean to small changes in the welfare indicator or the poverty Percentage Shortfall of Consumption Relative line, and the adjusted figures reported here are within to Poverty Line) the confidence interval for earlier reported estimates.2 Change Nonetheless, the revised estimates are likely to be more 2010– exact as they more accurately reflect the best available 2001 2005 2010 2012 2012 information on the geographic structure of the Malagasy Urban 11.8 13.6 8.9 11.8 2.9 population. In particular, the weights used by Madagas- Rural 40.5 34.8 36.7 36.4 –0.3 car’s National Institute of Statistics (INSTAT) to com- Total 35.9 31.3 32.0 32.2 0.2 pute population-level statistics implied a spatial partition Sources: EPM and ENSOMD. of the population, which deviated from the best esti- Note: Poverty line = Basic needs, World Bank. mates of its actual partition.3 In effect, too low a weight had been assigned to urban households and too high a weight to rural households; since urban households exhibit lower rates of poverty, this tended to overstate national poverty rates and underestimate consumption The Distribution of Growth growth between 2005 and 2010.4 (See annex 1A for and Changes in Inequality more details on this and other methodological issues.) Fluctuations in the headcount poverty rate from 2001 to Regardless of these adjustments, Madagascar’s poverty 2012 mask significant changes in the distribution of con- rates are exceedingly high, and according to internation- sumption growth. In fact, over this period, poverty head- ally comparable estimates are the highest in the world.5 count rates have risen in years when growth has been more Using the World Bank’s international poverty lines of progressive, and vice versa: poverty headcount rates have US$1.90 per capita per day (in 2011 PPP), poverty in fluctuated with the fortunes of households living close to Madagascar is 77.8 percent (table 1.2).6 the poverty line, whereas those falling lower (and higher) in the distribution have been impacted quite differently. Close to 80 percent of Madagascar’s population lives in Between 2001 and 2005, when a significant percentage of rural areas, and rural poverty rates are more than twice the nonpoor population fell into poverty, mean consump- as high as urban rates. As shown in table 1.3, although tion levels of the poor nonetheless largely improved. This rural poverty rates have stayed fairly flat— having risen is shown in the growth incidence curves (figure 1.1). slightly after 2001, then fallen back to 2001 levels in However, as shown this pattern reversed after 2005, with 2012—urban poverty rates have fluctuated much more, declines in real consumption below the 40th percentile and from 34 percent in 2001 to over 40 percent in 2005, gains at the top. Between 2010 and 2012, the pattern is 29.8 percent in 2010, and once again close to the 2001 mixed, with declines at the bottom and top and improve- level in 2012.7 ments in the middle of the distribution—where there are Madagascar Poverty and Inequality Update: Recent Trends in Welfare, Employment, and Vulnerability  15 FIGURE 1.1: Growth Incidence Curves, 2001–12 and Subperiods (Total Percentage Changes) A B 2001 to 2005 2005 to 2010 10 10 8 6 5 4 Growth rate, % Growth rate, % 2 0 0 –2 –4 –5 –6 –8 –10 –10 0 20 40 60 80 100 0 20 40 60 80 100 Expenditure percentile Expenditure percentile C D 2010 to 2012 2001 to 2012 10 10 8 8 6 6 4 4 Growth rate, % Growth rate, % 2 2 0 0 –2 –2 –4 –4 –6 –6 –8 –8 –10 –10 0 20 40 60 80 100 0 20 40 60 80 100 Expenditure percentile Expenditure percentile still many poor people. Overall, when comparing 2012 in 2005, and inched up only slightly in 2010 to 32.0 and in to 2001, the bottom range within the poor population 2012 to 32.2, still lower than in 2001, despite the decrease showed net gains in consumption: The regressive pattern in real per capita GDP over the period (table 1.3).9 of consumption growth after 2005 did not completely off- set the gains made at the bottom of the distribution from Despite negative (real) per capita GDP growth over the 2001 to 2005.8 The poverty gap, a measure of the severity period 2010–12, the headcount poverty rate dropped of poverty, correspondingly fell, from 35.9 in 2001 to 31.3 by 1 percentage point due to a favorable distribution of 16 Republic of Madagascar Employment and Poverty Analysis TABLE 1.4: Decomposition of Growth versus Inequality Contributions to Changes in Poverty Rates (2010–12) 2010 2012 Actual change Growth Redistribution Interaction Poverty line  Basic needs (World Bank) Total 71.65 70.74 –0.91 1.58 –2.77 0.28 Urban 29.82 35.52 5.71 6.83 0.38 –1.51 Rural 80.12 77.93 –2.19 0.46 –2.61 –0.04 Poverty line  Food poverty line (World Bank) Total 58.28 57.43 –0.85 2.33 –3.05 –0.13 Urban 18.35 22.67 4.32 5.09 –0.41 –0.35 Rural 66.36 64.52 –1.84 0.63 –2.50 0.03 growth near the poverty line. Negative growth would FIGURE 1.2: Inequality (Lorenz Curves) for 2001, have added 1.58 percentage points to the poverty rate 2005, 2010, and 2012 (Cumulative Share of Welfare (table 1.4) and an astonishing 6.83 percentage points Accruing to x-axis Proportion of the Population) to the urban poverty were it not for the distributional Total effects. In rural areas, the distribution of growth was 1 pro-poor overall between the two years, with the 2001 redistribution effect accounting for a reduction in the 2005 headcount poverty rate by 2.61 percentage points. 0.8 2010 2012 Whereas Madagascar’s poverty rates are exceedingly Equality Lorenz curve 0.6 high, inequality is in line with that of other low-income countries and has fallen over the period. Urban areas display more unequal distributions of real per capita 0.4 consumption than the rural zones; however, inequality in rural areas increased over the period of analysis. Fig- ure 1.2 shows the national Lorenz curves for all four sur- 0.2 vey years, again reflecting the generally equalizing trends between 2001 and 2005, which then partially reverse 0 thereafter. Lorenz curves that depict the divergence from 0 0.2 0.4 0.6 0.8 1 perfect equality at different parts of the distribution are Cumulative population proportion shown in figure 1. 3. The Gini coefficient—a summary, internationally compa- of the top decile of households to that of the bottom rable measure of inequality related to the Lorenz curve— decile (P90/P10)—helps to overcome this shortcom- also shows fluctuations, but finished the period lower.10 ing. These measures show that much of the increase in Starting at a high of 46.7 in 2001, it was 41.0 in 2012, inequality after 2005 is driven by a decline in the welfare relative to a low income average of 40 for countries share accruing to the poorest segment of the population, with available data over period 2007–11, according to which dropped from 6.97 to 5.89 percent (15 percent) the World Development Indicators (WDI). However, the between 2005 and 2012, except in the urban sectors Gini coefficient does not capture distributional changes where it declined by only 4 percent. These measures also that may occur in different parts of the welfare distribu- reveal that in the case of Madagascar inequality does not tion. It would reflect a redistribution from the middle of reflect the presence of high wealth at the top of the dis- the distribution toward the bottom in the same manner tribution: As shown in table 1.5, the average consump- as a redistribution from the top to the middle, for exam- tion ratio for the top decile to the bottom decile has been ple. Information on consumption shares by population between 5 and 8 over the period under study and thus quintiles—in particular the ratio of average consumption is consistently lower than the low-income average over Madagascar Poverty and Inequality Update: Recent Trends in Welfare, Employment, and Vulnerability  17 FIGURE 1.3: Lorenz Curves and Inequality Coefficients Madagascar Lorenz curve by area 2005 2010 2012 1 0.8 0.6 0.4 0.2 0 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 Cumulative population proportion Urban Rural Line of perfect equality 2005 2010 2012 Income shares Income shares Income shares Low. Top Low. Top Low. Top Gini p90/p10 quintile quintile Gini p90/p10 quintile quintile Gini p90/p10 quintile quintile National 38.93 4.96 6.97 46.63 42.66 6.01 6.03 49.69 41.03 6.32 5.89 47.63 Rural 35.35 4.24 7.63 43.48 37.96 4.73 6.91 45.38 37.3 5.32 6.54 44.33 Urban 39.22 5.51 6.50 45.91 38.60 5.16 6.56 45.38 38.44 5.92 6.27 45.12 Sources: ENSMOD 2005, 2010, and 2012. TABLE 1.5: Consumption (per Capita) Inequality Measures Bottom half distribution Upper half distribution Interquartile range Tails p25/p10 p50/p25 p75/p50 p90/p75 p75/p25 p90/p10 Gini 2001 1.50 1.66 1.79 1.83 2.96 8.13 46.9 2005 1.39 1.44 1.53 1.61 2.22 4.96 38.9 2010 1.48 1.51 1.59 1.69 2.40 6.01 42.7 2012 1.53 1.56 1.62 1.63 2.52 6.32 41.0 Sources: EPM 2001–2010 and ENSOMD 2012. Trends in Agriculture 2007–11 of 13.4 (per WDI). Finally, a single year’s snap- and Employment shot of the consumption distribution, particularly in a country such as Madagascar, which faces a high level of A combination of external and domestic shocks and pol- weather-related and other risk, can overstate inequality icy responses buffeted the poor over the period 2001–12. in households’ lifetime welfare. A high level of interan- Changes in the terms of trade in agriculture, consumer nual variation means that households are moving up and price inflation, weather shocks, and changing off-farm down in the distribution from year to year. labor market conditions affected households throughout the consumption distribution. Households responded 18 Republic of Madagascar Employment and Poverty Analysis by adjusting their levels of secondary employment and employment. The adverse labor market shocks in 2005 self-employment, as well as their allocation of work across combined with reasonably favorable terms of trade in sectors as returns in various sectors shifted. Overall, work- agriculture to induce a substantial movement of labor ers reporting a wage saw little to no wage growth. into agriculture. As shown in figure 1.5, one can discern a clear shift between 2001 and 2005 into agriculture for First, in 2001–2005, the political crisis of 2002 com- the top 40 percent of the distribution, with the remaining bined with cessation of most-favored-nation preferences 60 percent maintaining their extremely high (over 80 per- to produce an adverse effect on labor markets. Urban cent) rates of primary employment in the sector. One also employment in particular declined, and many people observes a decline in the percentage of household heads dependent for their incomes on the textiles sector fell employed in manufacturing from 2001 to 2005. into poverty. In addition, consumer price inflation, driven largely by local conditions, increased the cost Between 2005 and 2010, the declining profitability of of living. The consumer price index increased by over agriculture—in particular for rice cultivation—contributed 18 percent in 2005 alone, with a cumulative 54 percent to households’ seeking employment outside of the sector. increase over the four years between 2001 and 2005. By 2010, although the world price of rice had contin- Although high inflation is not unusual for the country— ued its increasing trend, the terms of trade in agriculture annualized inflation has averaged more than 11 percent shifted against producers, due in part to declining trans- since 1965—over 77 percent of households reported an port conditions and policies designed to maintain lower adverse effect of general price inflation in 2005, whereas rice prices (see Thiebaud, Osborne and Belghith 2016). only 2.9 percent did so in 2010 (and 0.8 percent in Between 2005 and 2010, the producer price of paddy 2012). Because of these and other factors, the headcount rice fell (figure 1.6), despite generally rising world food poverty rate in urban areas increased dramatically, from prices. Moreover, in contrast to 2005, in 2010, households’ 34.1 to 40.8 percent. At the same time, as figure 1.1 consumption levels were positively correlated with paddy showed, consumption at the bottom of the distribution prices across quintiles. Moreover, as transport conditions rose. Although it is unclear which conditions contributed deteriorated, for each percentage point increase in the time most to this improvement, rising rice prices after 2005 to reach input markets, the relative price of rice relative to may have improved the net incomes of rice producers (see figure 1.4). FIGURE 1.5: Sector of Main Employment Malagasy households responded to these shifts in cir- of Household Head (%) by Quintile and Year cumstances by adjusting their labor supply and sectors of 100% 90% 80% FIGURE 1.4: Rice Price Indices (2001 = 100) 70% Rice, Thai, A1.Special, $/mt, nominal $ 60% 600 Madagascar rice price 50% (100 = average 2000 price) 500 40% 30% 400 20% 10% 300 0% 2001 2005 2010 2012 2001 2005 2010 2012 2001 2005 2010 2012 2001 2005 2010 2012 2001 2005 2010 2012 200 100 Poorest Second Third Fourth Richest Services Manufacturing 0 Public administration Agriculture/primary 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 Construction Source: FAO Sources: EPM 2001–2010 and ENSOMD 2012 Madagascar Poverty and Inequality Update: Recent Trends in Welfare, Employment, and Vulnerability  19 FIGURE 1.6: Price of Rice Paddy (Producer Price FIGURE 1.7: Rice Production and Yields in Communities) by Consumption Quintile 5,000,000 35,000 900 30,000 4,000,000 800 25,000 Tonnes Kg/Ha 700 3,000,000 20,000 600 15,000 2,000,000 Price, ariary 500 10,000 1,000,000 400 5,000 300 0 0 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 200 100 Tonnes Kg/Ha 0 2005 2010 Source: FAO. Poorest Second Third Fourth Richest Community mean area cultivated per agricultural household continued its Sources: EPM 2005, 2010. downward trend (figure 1.8) before reverting in 2012 to its 2001 level.13 Over the decade agricultural produc- tivity per worker fell (figure 1.9). This, combined with fertilizer (urea), which fell overall between 2005 and 2010, demographic trends, increased logging activities and decreased by 18 percent.11 weak enforcement, especially after the political crisis of 2009, may have exacerbated Madagascar’s deforestation Although aggregate trends suggest that more land was problem, already under way (figure 1.10). brought under rice cultivation after 2008, the average productivity of this land fell. As shown in figure 1.7, Households once again responded to circumstances in rice yields flattened after the world food price spike of 2010 by shifting their employment patterns. As the terms 2008, but production continued to increase through an of trade in agriculture deteriorated, a slightly lower per- expansion in the land under rice cultivation, in part due centage of households in the 3rd and 4th quintiles had to continued high population growth.12 Yet the average heads primarily employed in agriculture, and a much FIGURE 1.8: Area of Economically Exploited Land per Agricultural Household, by Year and Consumption Quintile (Hectares) 2 1.5 1 0.5 0 2001 2005 2010 2012 Poorest Second Third Fourth Richest Total Sources: EPM and ENSOMD. 20 Republic of Madagascar Employment and Poverty Analysis FIGURE 1.9: Agricultural Value Added Per Worker the greatest percentage of all household heads in 2010. And although services were an important source of sec- 250 ondary employment in all years, there was an especially dramatic increase in employment of household heads in 200 this sector in 2010 (figure 1.11). 150 Trends in the main sectors of employment once again reversed between 2010 and 2012. In 2012 those in the 100 bottom quintiles were more likely to be employed in agriculture than in 2010, but those in the top were more 50 likely to be employed off-farm. The declining trend in manufacturing employment, which continued through 2010, began to reverse in 2012, when 10.2 percent of 0 people in the top quintile had a household head primar- 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 ily employed in manufacturing (versus only 2.2 percent in the bottom quintile). The entry into services in 2010 Source: WDI. also partially reversed in 2012 (figure 1.11) lower percentage in the top quintile (figure 1.5). One Households also responded to events by seeking a observes an even more significant decline in agriculture second job and self-employment off-farm. As shown in as a sector of secondary employment between 2005 figure 1.12, as agricultural profitability declined, the and 2010, in all quintiles. Among those with secondary proportion of both males and females looking for work employment, after agriculture, service sectors employed increased significantly between 2005 and 2010, espe- cially for females over the age of 10. In 2012, however, those looking for work fell again for both genders. At the same time, the proportion of both males and females FIGURE 1.10: Forest Cover and Changes over Time with a second job increased from 2005 to 2010, and FIGURE 1.11: Sector of Secondary Employment of Household Head (%) 100% 90% 80% 70% 60% 50% 40% 30% Forest 20% Deforestation 2005–2010 10% Deforestation 2000–2005 0% Deforestation 1990–2000 2001 2005 2010 2012 2001 2005 2010 2012 2001 2005 2010 2012 2001 2005 2010 2012 2001 2005 2010 2012 Non forest Water Plantation Poorest Second Third Fourth Richest Clouds Services Manufacturing Mangroves Public administration Agriculture/primary Tapla Construction Madagascar Poverty and Inequality Update: Recent Trends in Welfare, Employment, and Vulnerability  21 then dropped in 2012 to levels below those observed in increased between 2005 and 2010, and remained higher 2005. Women were much less likely to be self-employed in 2012 for most age ranges. The proportion of men in agriculture than men. Yet the percentage of both male self-employed in agriculture also declined in 2012 after and female workers self-employed outside of agriculture staying relatively constant for all ages between 2005 FIGURE 1.12: Labor Market Outcomes, 2005, 2010, and 2012 (Kernel-Weighted Local Polynomial Smoothed Age-Outcome Profiles Males in 2005 Males in 2010 Males in 2012 Females in 2005 Females in 2010 Females in 2012 Proportion looking for work Proportion with second job 0.10 0.8 0.08 0.6 0.06 0.4 0.04 0.2 0.02 0 0 0 10 20 30 40 0 10 20 30 40 50 60 70 80 Age Age Proportion of self-employed in agriculture Proportion of self-employed in nonfarm sector 1.0 0.20 0.18 0.8 0.16 0.14 0.6 0.12 0.10 0.4 0.08 0.06 0.2 0.04 0.02 0 0 0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80 Age Age Source: Calculated using EPM and ENSOMD. Note: y axis represents polynomial smoothed proportion (from 0 to 1) of individuals engaging in labor market behavior noted. 22 Republic of Madagascar Employment and Poverty Analysis and 2010. A shift out of agricultural work may signal Many communities also cited the lack of adequate land improving opportunities in off-farm labor markets area as a key obstacle. Some 22.5 percent of communi- and small informal enterprise; yet overall wages have ties ranked this as the number one constraint—the most not kept pace with inflation (see figure 1.13) and, as frequent among the top ranked constraints cited—and discussed in Bi and Osborne (2016), employment in the a lack of land was ranked third in frequency among smallest of such enterprises is typically less productive households’ top three constraints. This likely reflects and remunerative than in other jobs. Thus, labor pro- the high price and low profitability of inputs, which ductivity remains too low to make a greater dent in the would incentivize extensive over intensive agriculture. country’s poverty rate. Moreover, women have had more A tally of responses shows that insecurity of land tenure difficulty securing employment off-farm, and the dispar- and conflict over land were mentioned in only a few ity in wages between females and males of prime work- communities.15 ing age increased in 2010 vis-à-vis 2005 (see disparities at age 40 in figure 1.13, as indicated by the arrow). Following these top issues, many communities (34 per- cent) ranked the condition of irrigation infrastructure, Community informant surveys are broadly consistent then the condition of roads (25 percent) as among their with the trends and indicators observed in agriculture. top three problems for agricultural development. In The EPM 2010 surveyed key informant community addition, insecurity was a major issue. In 22 percent of members in each of 623 communities and obtained the communities theft of cattle and in 15 percent of com- groups’ list of the top five development problems in munities theft of crops were listed among the top three agriculture. A count of the frequency of responses for the problems, and combined they pose a greater issue than top constraints, as well as inclusion in the top three con- the condition of roads. In addition, distance to product straints, is shown in figure 1.14. Although these data are markets was cited in 19 percent of communities.16 based on perceptions rather than quantitative analysis, a couple of themes clearly emerge. First is the impor- Notably, issues related to access to credit did not rank tance of problems in input markets. The most frequently high on communities’ lists of priority problems. As ranked issue among the top three constraints is related shown, relatively few communities cited either the dis- to either the high cost or lack of access to inputs such tance to credit institutions, credit security requirements, as seeds and fertilizer.14 This was also the fourth most or high interest rates as among the top three agricul- frequently cited among communities’ number one issues. tural development problems. According to community FIGURE 1.13: Wage Trends from 2005 to 2012 Log of wage rate 11.5 11.0 Males in 2005 Males in 2010 Males in 2012 10.5 10.0 Females in 2005 Females in 2010 Females in 2012 lpoly smooth: (mean) per_day_inc 9.5 9.0 8.5 8.0 7.5 7.0 6.5 6.0 5.5 5.0 4.5 4.0 3.5 3.0 0 10 20 30 40 50 60 70 80 Age Madagascar Poverty and Inequality Update: Recent Trends in Welfare, Employment, and Vulnerability  23 FIGURE 1.14: Frequency of Community Group Rankings of Constraints to Agriculture Lack or high cost of seeds, fertilizer Selling price too low Inadequate land area Inadequate irrigation infrastructure Condition of roads Theft of oxen Markets for products too far Theft of crops Lack of animal traction Credit institutions too far away Credit security requirements Interest rate too high 0% 5% 10% 15% 20% 25% 30% 35% 40% 45% 50% Three most severe constraints Most severe constraint Source: Calculated using EPM 2010. member perceptions, therefore, issues related to greater to respond significantly to short-run shocks—as would profitability—access to inputs and markets—rather inequality and poverty measures—without these changes than a lack of access to credit by producers are the key necessarily being persistent (or indeed permanent).20 For obstacles to agricultural development. Similarly, season- instance, when there is greater spatial variance of weather ality of labor demand, inadequate access to road and shocks in a given year, measured inequality in that period irrigation infrastructure, and weak market access appear will appear higher, without this necessarily representing a among the country important constraints, based on stud- permanent condition. At the same time, when such shocks ies conducted on Madagascar over the past 15 years.17 are large and significant assets are lost, households will have difficulty recovering economically and may be forced to sacrifice long-run investments in education and health Risks and Vulnerability as part of their coping strategy. Households in Madagascar are subject to an extreme Natural conditions that result in such huge intermittent amount of weather-related risk, which can push them losses combined with the absence of adequate mechanisms deeper into poverty at any time, and these risks were most to shield against them not only have devastating short-run clearly manifested in 2010.18 Although people tend to pre- effects on consumption but also make it necessary to hedge fer a relatively even level of consumption over time even risks in a way that persistently reduces incomes.21 For as income fluctuates, they cannot perfectly smooth short- example, farmers must avoid specialization and depen- term income fluctuations arising from weather, price, or dence on food markets and rather must operate in relative temporary health shocks by borrowing, saving, and insur- autarky: the percentage of crop production for the market ing against risk (see, for example, Friedman 1957).19 With- is low, and one sees even urban households engaged in out further study, it is unclear to what extent informal agriculture for their own production. Moreover, when risk-mitigation instruments are available in rural Mada- combined with poorly performing input markets—the gascar, but most households report giving and/or receiving inability to access inputs at the right times and at advanta- remittances. Even so, the available strategies are unlikely geous prices—these issues reduce profitability substantially. to adequately address spatially correlated risks such as The returns to using fertilizer, for example, in these circum- cyclones or drought. Thus, consumption levels are likely stances can be nil (see Livingston et al. 2011).22 24 Republic of Madagascar Employment and Poverty Analysis FIGURE 1.15: Number of Negative Shocks Reported households experienced was lower in 2012 than in any (2005–12) of the other prior survey years. Apart from 2005, when a general price increase was the most frequently reported 80 2005 shock, climatic shocks are the most frequently reported 70 Percentage of households 2010 type. As shown in figure 1.16, more households reported 60 being affected by a cyclone, flood, or late rains in 2005 2012 50 than in later years. Plant and animal disease also affects 40 a significant percentage of households. 30 20 Moreover, although the type of adverse shock changes 10 from year to year and affects different households, the frequency of adverse climatic shocks is generally corre- 0 0 1 2 3 4 5 lated with poverty, as shown in figure 1.17. As shown in Number of shocks figure 1.18 2005 was also a bad year for health shocks relative to the subsequent survey years, as it was for Source: EPM. security shocks (figure 1.19). Nonetheless, the costs of these shocks appear to have been greatest for the poorest households particularly in 2010, as shown by Thiebaud, As shown in figure 1.15, the number of shocks—whether Osborne and Belghith (2016). The full statistics on the climatic, health, security, or economic shocks—that frequency of shocks is reported in annex 1C. FIGURE 1.16: Climatic, Natural, and Related Shocks 14 Percentage of households affected 2005 12 2010 10 2012 8 6 4 2 0 e an se e d n t ad er in re an se n s oc e oc e h op on as i sio sh du sh du oo ra ra ug th Fi pl ea pl a ro ) ) im at s ks ts ts se d ise k cr cl Fl O va te rly te ion e s ve is ro d Cy at ive di d ke d d in La ke Ea D te t t e t im a st an oc an oc ttl c cl igr cl rel cu fe Pl (st Pl Bl Ca a to m (li af Lo to ed ed ka t os va rc H Fo La Source: EPM. Madagascar Poverty and Inequality Update: Recent Trends in Welfare, Employment, and Vulnerability  25 FIGURE 1.17: Percentage of Households Having Stated Shock (Top Three Reported Shocks, 2012) 30% #1 Shock #2 Shock #3 Shock Drought 25% 20% Cyclone Cyclone Drought 15% Drought Cyclone household member household member Cyclone 10% Drought Illness of adult Illness of adult Cyclone Drought Illness or loss of livestock Late rain Late rain 5% 0% Poorest quintile Second quintile Third quintile Fourth quintile Richest quintile Source: ENSOMD 2012 FIGURE 1.18: Frequency of Health Shocks Reported Percentage of households affected 7 6 2005 2010 5 2012 4 3 2 1 0 Illness of adult Illness of other Death of other Death of active Other health issues of household member household member household member household member household members Source: EPM 2005, 2012, ENSOMD 2012. FIGURE 1.19: Frequency of Reported Security Shocks 9 Percentage of households affected 8 2005 2010 7 2012 6 5 4 3 2 1 0 Theft of Cattle Other Theft of Theft of Theft/loss Theft of Land Violence against standing rustling consumption production of cash stock conflict household crops goods inputs member Source: EPM 2005, 2010, ENSOMD 2012. 26 Republic of Madagascar Employment and Poverty Analysis Annex 1A. Sampling, Comparability between 2010 and 2012 Data and Weights Issues The 2012 poverty statistics were calculated from TABLE 1A.1: Sampling Objectives for EPM Enquête Nationale sur les Objectifs Millenaire du and ENSOMD Développement (ENSOMD), a household (HH) sur- vey very similar in format to the country’s previous Survey Sampling objective Living Standards Measurement Surveys (LSMS), the EPM 2010 To obtain a total HHs sample representative at the national level and at regions cross Enquête Périodique auprès des Ménages (EPM). Because urban-rural. So, that sample is supposed to be the objectives of the two surveys differed, so did the representative for each of the 44 strata of the sampling strategy. The ENSOMD was designed to track first stage sampling as there are 22 regions for Madagascar. outcomes related to the Millennium Development Goals, whereas EPM surveys are designed to capture a greater ENSOMD To draw a HH sample that is representative 2012 for the following domains: the national level, range of socioeconomic variables. For the EPM 2010, the capital, the urban and rural areas, and the sample size and structure are similar to previous versions 22 regions. The sample is not supposed to be of the EPM, whereas for the ENSOMD 2012, sample representative for an urban-rural division within a region. size and sample structure are more similar to the Mada- gascar Demographic and Health Survey (DHS) survey. In Sources: INSTAT, EPM 2010, and ENSOMD 2012 reports; Discussion with technical staffs of the surveys. particular, given the need for more detail on health out- comes and other indicators, ENSOMD did not include a community questionnaire. In order to capture DHS indicators such as mortality rates, the total HH sample regional population structure and the need to capture the size of the ENSOMD 2012 had to be very large. diversity of socioeconomic life between rural and urban areas and within urban areas. Thus, the sample of EPM For the EPM 2010, the sample structure for urban- 2010 is almost equally distributed by urban-rural strata. rural was driven generally by two criteria: geographic/ For the ENSOMD 2012, the sample structure is mainly TABLE 1A.2: Sampling Methodology for EPM 2010 and ENSOMD 2012 Rubric EPM 2010 ENSOMD 2012 Observations Sampling methodology Two stages, EAs in first Two stages, EAs in first stage and HH in second stage and HH in second stage stage Sampling frame for first stage EAs of the census EAs of the census mapping of 2008 mapping of 2008 Stratification in the first stage Regions cross urban-rural Regions cross urban-rural The EPM 2010 retained the old areas areas definition of urban-rural, while the ENSOMD 2012 used the new definition. Sampling method in the first stage Probability proportional Probability proportional to size to size Segmentation during the enumeration stepa No segmentation Segmention is used Segmentation is used only in for identified big EAs the 2012 survey selected from the first stage Sampling method at second stage Systematic sampling Systematic sampling Sources: INSTAT, EPM 2010, and ENSOMD 2012 reports; Discussion with technical staffs of the surveys. a For a selected EA in the first stage of sampling, segmentation denotes an action during the enumeration in the field; the field team divides the entire EA into two or more almost equal-size subdivisions. Afterward, the survey will be done in one subdivision randomly selected among all subdivisions of the EA. Segmentation is applied for EAs identified as large during the enumeration step by the survey team. Madagascar Poverty and Inequality Update: Recent Trends in Welfare, Employment, and Vulnerability  27 derived from the actual geographic partition of popula- in 2012. The old definition of area of residence was the tion. Only about 25 percent of the sample is drawn from definition used since the last population census in 1993, the urban area in the 2012 survey, a level more similar whereas when producing the new database of EAs in 2008, to the actual population by area of residence. Moreover, Madagascar’s National Institute of Statistics (INSTAT) the survey for 2012 is not representative of each region’s used a new definition of urban-rural. For the most part, urban and rural areas separately, as the 2010 survey is. the reclassification of EAs led some urban areas in the old Rather it is representative only by region and by rural definition to be redefined as rural. Whereas the sampling and urban areas at the national level. Nonetheless, the for EPM 2010 still relied on the old definition, both old core modules of the questionnaires, including consump- and new definitions could be captured in order to ease tion modules, were essentially identical across the two comparability. The ENSOMD 2012, however, used the new surveys.23 definition in the sampling frame. Table 1A.2 summarizes the sampling methodology used for the two surveys. Table The two surveys use a two-stage sampling procedure 1A.3 show sthe resulting sample details, and 1A.4 shows wherein the first stage, a sample of Enumeration Areas the spatial pattern of the 2010 and 2012 samples. (EAs), is randomly drawn from an EAs sampling frame and, in the second stage, a sample of HHs is drawn from a list of households obtained by an enumera- tion activity in each selected EA. In the first stage TABLE1A.3: EPM 2010 and ENSOMD 2012 of sampling, the two surveys use the same sampling Sample Comparison frame—the national list of EAs from the census map- EPM 2010 ENSOMD 2012 ping of 2008.24 Although the sampling strategy differed, Initial sample size in principle as long as the sampling weights reflect the Sample of enumeration 623 615 best available estimates of the population’s structure, areas consumption aggregates and poverty numbers should Sample intake of HH by EA 20 32 be comparable at levels for which samples are represen- Total HH sample 12,460 19,680 tative. The sampling objectives for the two surveys are Final sample size summarized in table 1A.1. Sample of EAs 623 609 An additional complicating factor is that a new official Total HH sample 12,460 16,920 definition of urban versus rural was applied beginning Source: INSTAT. TABLE 1A.4: Partition of Sample of EAs by Region and Urban-Rural Area for Each Survey EPM 2010 ENSOMD 2012 Region Urban Rural Total Urban Rural Total Analamanga 30 24 54 50 25 75 Vakinankaratra 15 15 30 6 19 25 Itasy 12 13 25 3 22 25 Bongolava 12 13 25 4 21 25 Matsiatra Ambony 14 13 27 7 18 25 Amoron’i Mania 13 13 26 3 22 25 Vatovavy Fitovinany 14 14 28 3 23 26 Ihorombe 12 12 24 4 21 25 Atsimo Atsinanana 12 13 25 4 21 25 Atsinanana 19 14 33 7 19 26 Analanjirofo 13 14 27 4 22 26 Alaotra Mangoro 13 13 26 5 20 25 (continued) 28 Republic of Madagascar Employment and Poverty Analysis TABLE 1A.4: Partition of Sample of EAs by Region and Urban-Rural Area for Each Survey (continued) Boeny 17 14 31 8 17 25 Sofia 14 17 31 4 21 25 Betsiboka 12 12 24 3 22 25 Melaky 12 12 24 3 23 26 Atsimo Andrefana 14 18 32 6 20 26 Androy 12 12 24 26 26 Anosy 14 12 26 5 22 27 Menabe 14 12 26 5 21 26 DIANA 14 12 26 10 15 25 SAVA 14 15 29 5 20 25 Total 316 307 623 149 460 609 % of urban-rural 50.7% 49.3% 100.0% 24.5% 75.5% 100.0% Sources: INSTAT, EPM 2010, and ENSOMD 2012 reports and databases; Discussion with technical staffs of the surveys. TABLE 1A.5: Weight Construction Procedure and Components Component EPM 2010 ENSOMD 2012 Observations Design weight for each EA (Pop total in strata/pop in the (Pop total in strata/pop in the EA) x Sample size of EAs in EA) x Sample size of EAs in strata strata Segmentation 1/(proportion of the Segmentation is not applied segmentation) for 2010 Design weight for each HH at (Number of enumerated HH in (Number of enumerated HH in EA level the EA)/20 the EA)/32 Nonresponse adjustment for (Sample size of EAs in strata)/ Nonresponse adjustment is not EAs (Sample size of EAs surveyed applied for 2010 in strata) Nonresponse adjustment for — (Number of identified HHs as Nonresponse adjustment is not HHs sample in strata)/(Number of applied for 2010 HHs with completed interview in the strata) Post-stratification to take into Nothing done here The structure of the population The only post-stratification account geographical structure in the 2008 census mapping adjustment done was on the of the population was used to calculate geographical repartition of adjustment factor (Wi) population and it was done only for the 2012 survey. Others The total of HH was adjusted for some EAs for which the total of HHs enumerated was too low or too high, compared to the size of the EA in the sample framea Final HHs weight Multiplication of each above Multiplication of each above component component Source: INSTAT, EPM 2010, and ENSOMD 2012 weight construction templates files; Discussion with technical staffs of the surveys. a For some EAs, the number of HHs effectively enumerated by the field work team was judged by the survey analyst team to be too low or too high given the initial size of these EAs as already reported the 2008 EAs database. To correct, the initial size in the EAs database was taken into consideration, but this correction was done for a just few number of EAs (45 EAs among the total of 623 EAs). The resulting sample structure is quite different for the all components included in the final HH weight used two surveys as shown in table 1A.4. Based on the tem- for data analysis are described in table 1A.5 for each plates files of weight construction of the two surveys, survey. Madagascar Poverty and Inequality Update: Recent Trends in Welfare, Employment, and Vulnerability  29 From table 1A.5, one can say that the main components 2012 surveys, INSTAT utilized the slightly changed of the weight adjustments represent corrections for each structure obtained from the 2008 census mapping survey following the corresponding sampling method. exercise. INSTAT has not, therefore, modelled trends in Nonetheless, there was a significant difference between population changes or urban-rural migration since that the two surveys in that a poststratification to adjust time for the purpose of altering the assured structure. the regional structure of population was done for the These assumptions may be updated after a new census ENSOMD 2012, whereas this was not done for the EPM is completed, and there are no clear indications that the 2010. This resulted in an implicit population structure rural-urban structure has altered appreciably since 2008. which differed from the best available information on Therefore, the best approach appears to be to hold the the geographic allocation of the population. structure constant in calculating sampling weights and the corresponding statistics from the 2010 and 2012 Because standards of living vary importantly by area surveys. of residence and by region, we estimate the population structure by area of residence and by region in order to Table 1A.6 exhibits results obtained for the popula- check the consistency of the actual weight of each of the tion’s structure by urban-rural areas. To avoid confusion, 2010 and 2012 surveys, taking the structure from the results are shown separately for the old and the new 2008 EAs database as the definitive reference. As the definition of area of residence. last effective population census was done in 1993, this database is the most recent and best estimate available of the geographic structure of the Malagasy population. TABLE 1A.6: Structure of Population by Urban-Rural Figure 1A.1 compares the structure of the population EPM 2010 and ENSOMD 2012 by area of residence (using the old definition) from the 2008 census mapping with results from previous EPM Old definition New definition surveys. EA EA Area database 2010 2012 database 2010 2012 Urban 22.4 20.3 24.5 16.7 10.6 20.1 When deciding on the assumed structure of the popu- lation for the purposes of the 2010 survey, INSTAT Rural 77.6 79.7 75.5 83.3 89.5 80.0 conducted a statistical test of differences between 1993 Total 100.0 100.0 100.0 100.0 100.0 100.0 and 2008 and was not able to reject that the structure Sources: Calculated from EPM 2010, ENSOMD 2012, and census remained the same. Nonetheless, for both 2010 and mapping of 2008 databases. FIGURE 1A.1: Structure of Population by Urban-Rural from Previous EPM and the 2008 Census Mapping (Old Definition Rural-Urban) 80 70 60 Rural 50 Urban 40 30 20 10 0 EPM 1993 EPM 1997 EPM 1999 EPM 2001 EPM 2004 EPM 2005 Census mapping 2008 Sources: INSTAT, EPM 1993, 1997, 1999, 2001, 2004, and 2005–2008 census mapping database of EAs. 30 Republic of Madagascar Employment and Poverty Analysis The table reveals a major discrepancy in the percent- TABLE 1A7: Structure of Population by Region, age of rural and urban populations between 2010 and Census Map Reference versus EPM 2010 2012, and in the case of the new definition of urban- and ENSOMD 2012 rural, a discrepancy in 2010 between the percentage of rural and urban populations of approximately Region EA database 2010 2012 6 percentage points. The partition of the population Analamanga 15.3 11.6 15.4 by region also shows some differences, as shown in Vakinankaratra 8.3 8.3 8.3 table 1A.7. It appears clear from this that the structure Itasy 3.4 3.7 3.4 of population provided by the EPM 2010 is problem- Bongolava 2.1 2.1 2.1 atic, while those provided by the ENSOMD 2012 seem Matsiatra Ambony 5.5 6.0 5.5 reasonable, given that post-stratification adjustments Amoron’i Mania 3.3 3.4 3.3 were made for that survey. The importance of the dis- Vatovavy Fitovinany 6.5 6.9 6.5 crepancies is exemplified by the proportion of popula- Ihorombe 1.4 1.2 1.4 tion in the large region of Analamanga, which contains Atsimo Atsinanana 4.1 4.4 4.1 the capital city. Atsinanana 5.8 6.0 5.8 Analanjirofo 4.7 4.6 4.6 Alaotra Mangoro 4.7 4.6 4.7 SUMMARY AND ADJUSTMENTS MADE Boeny 3.7 3.4 3.7 The main conclusion of the previous section is that Sofia 5.7 5.6 5.7 weights applied in both the 2010 and 2012 surveys Betsiboka 1.3 1.9 1.3 provide an inaccurate structure of population by urban- Melaky 1.3 1.4 1.3 rural area of residence, whatever the definition used (old Atsimo Andrefana 6.0 6.6 6.0 or new). In addition, the EPM 2010 actual weight does Androy 3.4 4.0 3.3 not provide a representative repartition of population by region. These issues need to be addressed as households’ Anosy 3.1 3.1 3.0 living standards vary by the geographical location of the Menabe 2.7 3.0 2.9 household. DIANA 3.2 2.8 3.3 SAVA 4.5 5.6 4.3 To address these issues, the World Bank poverty team Total 100.0 100.0 100.0 has computed and applied new HH weights for both Sources: Calculated from EPM 2010, ENSOMD 2012, and Census surveys. In addition to core design weights, the following mapping of 2008 databases. adjustments were introduced: 1. A first post-stratification procedure for the two sur- veys for EA weight. In fact, the EA weight must repro- 3. A post-stratification component to correct at the duce the structure and the size of the sampling frame same time the structure of population by urban-rural and it must be checked and corrected if not met. and by region. This took account of the old defini- 2. A recomputed nonresponse adjustment at the EA tion of area of residence, the new definition of area level and at the HH level by EA. of residence, and region. Madagascar Poverty and Inequality Update: Recent Trends in Welfare, Employment, and Vulnerability  31 Annex 1B. Poverty Estimation Methodological Notes As it is typically the case in Sub-Saharan Africa, the across geographic regions. The price indices are esti- available HH surveys in Madagascar use consumption mated using unit values from the surveys. as the key welfare measure to analyze poverty. This consumption aggregate comprises food consumption, The poverty lines are based on the cost-of-basic-needs including food produced by households themselves, as approach. The food poverty line is based on the cost of a well as expenditures on a range of nonfood goods and food basket that delivers 2,133 calories per capita (given services (for example, clothing, utilities, transporta- consumption patterns in a reference population) (see tion, communication, health, education, housing-related World Bank 2014). The basic needs poverty line adds expenditures, and imputed rent). However, the consump- an allowance for basic nonfood necessities to the food tion aggregate does not include expenditures on larger poverty line. The poverty lines have been reestimated for consumer durable items (such as cars, TVs, computers, each survey year 2001, 2005, and 2010. This reestima- and so forth), nor does it include expenditures on cer- tion is done because there was a socioeconomic crisis emonies (marriage, funerals, and the like). To the extent that occurred between these years, which may affect that better-off households devote a larger proportion of the structure of consumption. Moreover, there is a rule their total consumption to durable goods, this omission of thumb according to which poverty lines need to be creates certain biases and underestimates “true” con- reestimated at least every five years. The poverty line for sumption among wealthier families. This matters less for 2012 was estimated using the 2010 poverty line adjusted poverty analysis, where the focus lies on the bottom-end by the national consumer price index. of the distribution, but it can have a significant impact on estimated inequality. The basic needs headcount poverty rate (or, as used in the text, “poverty rate”) measures the proportion of The HH surveys collect consumption data at the the population whose monthly (price-adjusted) total household level. For the purpose of poverty and welfare household consumption per capita is below the basic analysis, total HH consumption needs to be adjusted needs poverty line, and the extreme headcount poverty for differences in household size and composition, rate (used in the text as “extreme poverty rate”) mea- which imply different consumption expenditure levels sures the proportion of the population whose monthly to achieve the same utility. There is a “public” good (price-adjusted) total household consumption per capita aspect to some categories of consumption: for example, is below the food poverty line. The annual consumption for housing and utilities, and different ages may require poverty lines for each year covered in this report is as different nutritional intake. However, the approach shown in the table 1B.1. Further technical details can followed here consists of computing consumption per be found in the Madagascar Poverty Assessment (World capita, implicitly assuming that all members of the Bank 2014). household require the same level of consumption, as this is the metric used by INSTAT as well as entities in many The national poverty line(s) reflect(s) Madagascar’s spe- other SSA countries. Paasche price indices are used to cific costs of basic consumption needs, but they are diffi- adjust consumption per capita for differences in prices cult to compare with other countries’ poverty thresholds. TABLE 1B.1: Poverty Lines Used (Annual Consumption per Capita) Year 2001 2005 2010 2012 Currency MGA* MGA MGA MGA Food poverty line 734,320 227,085 294,690 341,840 Complete poverty line (nominal values) 963,554 289,169 381,791 442,877 Temporal deflator** 1 1.501 1.32 1.16 * 1 MGA (ariary) = 5 FMG (Malagasy franc, the former national currency replaced by the MGA from 2005 onward). ** Current survey compared to the previous survey year for all years. 32 Republic of Madagascar Employment and Poverty Analysis To overcome this issue, the international poverty line of Poverty Assessment due to the adjustment in the popula- US$1.9 per capita per day (in 2011 PPP exchange rate) tion weights as described above. The World Bank (2014) is often used to evaluate a country’s poverty record vis-á- poverty figures in the report can be obtained exactly vis other developing countries or regions. using the variables in the data and the old weights. The correction of the weight variable using the same The poverty estimates for 2001, 2005, and 2010 in this post-stratification procedure described above has been paper differ from the poverty rates in the Madagascar applied to 2012 data. Annex 1C. Detailed Data and Results Tables TABLE 1C.1: Poverty Headcount and Distribution of the Poor by Region Poverty headcount rate Distribution of the poor 2001 2005 2010 2012 Change 2001 2005 2010 2012 Change Poverty line = Poverty line World Bank Urban 34.1 40.8 29.8 35.5 5.7 7.7 9.2 7.0 8.5 1.5 Rural 77.7 79.6 80.1 77.9 –2.2 92.3 90.8 93.0 91.5 –1.5 Region Analamanga 47.1 39.1 41.5 2.4 9.8 8.4 9.1 0.7 Vakinankaratra 83.3 77.6 87.7 10.1 9.6 8.8 9.9 1.1 Itasy 77.9 83.7 75.0 –8.6 3.7 3.9 3.4 –0.4 Bongolava 75.7 73.9 76.1 2.2 2.1 2.2 2.3 0.1 Matsiatra Ambony 84.5 81.0 71.9 –9.1 6.5 6.1 5.4 –0.7 Amoron’I Mania 86.1 85.9 81.7 –4.2 4.0 3.8 3.6 –0.2 Vatovavy Fitovinany 83.6 88.9 79.4 –9.5 7.3 8.1 7.4 –0.7 Ihorombe 84.4 79.1 76.6 –2.5 1.6 1.6 1.7 0.0 Atsimo Atsinanana 87.8 94.3 93.6 –0.7 4.8 5.5 5.7 0.2 Atsinanana 70.0 72.9 67.0 –5.9 5.7 5.8 5.4 –0.5 Analanjirofo 82.4 80.1 77.1 –3.0 5.4 5.3 5.1 –0.2 Alaotra Mangoro 66.6 72.3 62.8 –9.5 4.3 4.7 4.1 –0.6 Boeny 49.1 57.8 57.3 –0.5 2.4 3.0 3.1 0.1 Sofia 90.0 79.4 82.4 3.0 7.0 6.3 6.7 0.3 Betsiboka 76.9 81.9 78.9 –3.0 1.4 1.5 1.5 0.0 Melaky 81.1 79.0 81.6 2.6 1.4 1.5 1.6 0.1 Atsimo Andrefana 76.7 76.5 79.7 3.2 6.3 6.4 6.8 0.3 Androy 89.9 92.6 96.8 4.2 4.0 4.4 4.8 0.4 Anosy 76.0 78.5 85.7 7.2 3.2 3.3 3.7 0.3 Menabe 70.5 68.5 67.4 –1.1 2.5 2.6 2.7 0.0 Diana 51.8 46.2 36.4 –9.8 2.2 2.1 1.7 –0.4 Sava 76.0 71.2 71.9 0.7 4.7 4.4 4.5 0.0 Total 70.8 73.2 71.7 70.7 –0.9 100.0 100.0 100.0 100.0 0.0 (continued) Madagascar Poverty and Inequality Update: Recent Trends in Welfare, Employment, and Vulnerability  33 Poverty headcount rate Distribution of the poor 2001 2005 2010 2012 Change 2001 2005 2010 2012 Change Poverty line = Food poverty line World Bank Urban 22.4 28.3 18.3 22.7 4.3 5.9 7.8 5.3 6.7 1.4 Rural 67.7 66.0 66.4 64.5 –1.8 94.1 92.2 94.7 93.3 –1.4 Region Analamanga 33.7 23.3 29.1 5.8 8.5 6.2 7.9 1.7 Vakinankaratra 68.7 61.2 78.0 16.7 9.7 8.5 10.8 2.3 Itasy 63.0 71.0 49.8 –21.1 3.6 4.0 2.8 –1.2 Bongolava 57.1 56.7 56.2 –0.5 2.0 2.1 2.1 0.0 Matsiatra Ambony 66.4 70.4 53.9 –16.5 6.3 6.5 4.9 –1.6 Amoron’I Mania 73.0 72.7 63.1 –9.5 4.1 4.0 3.4 –0.6 Vatovavy Fitovinany 72.3 76.2 66.6 –9.6 7.8 8.5 7.6 –0.9 Ihorombe 73.0 65.9 66.0 0.1 1.6 1.7 1.8 0.1 Atsimo Atsinanana 79.6 88.7 88.8 0.1 5.3 6.4 6.7 0.2 Atsinanana 59.5 61.1 53.4 –7.7 5.9 6.0 5.3 –0.7 Analanjirofo 72.5 68.9 60.5 –8.4 5.8 5.6 4.9 –0.6 Alaotra Mangoro 48.5 58.1 38.5 –19.6 3.9 4.6 3.1 –1.6 Boeny 37.1 40.6 46.2 5.6 2.2 2.6 3.1 0.5 Sofia 78.9 63.6 72.3 8.8 7.5 6.2 7.2 1.0 Betsiboka 58.6 69.1 57.3 –11.8 1.3 1.6 1.3 –0.3 Melaky 62.9 62.4 68.1 5.6 1.3 1.5 1.7 0.2 Atsimo Andrefana 66.1 65.3 72.8 7.5 6.7 6.7 7.6 0.9 Androy 81.0 84.9 92.1 7.2 4.4 5.0 5.6 0.6 Anosy 59.1 70.7 73.6 2.9 3.1 3.7 3.9 0.2 Menabe 51.8 51.6 52.3 0.7 2.3 2.4 2.6 0.1 DIANA 34.3 29.1 23.4 –5.7 1.8 1.6 1.3 –0.3 SAVA 63.4 58.5 56.7 –1.8 4.8 4.5 4.3 –0.1 Total 60.5 59.8 58.3 57.4 –0.9 100.0 100.0 100.0 100.0 0.0 34 Sensitivity of headcount poverty rate with respect to the choice of poverty line 2001 2005 2010 2012 Poverty Change from Poverty Change from Poverty Change from Poverty Change from headcount rate actual (%) headcount rate actual (%) headcount rate actual (%) headcount rate actual (%) Poverty line = Poverty line World Bank Actual 70.8 0.0 73.2 0.0 71.7 0.0 70.7 0.0 +5% 73.0 3.1 75.5 3.1 73.6 2.8 73.0 3.2 +10% 74.8 5.7 77.4 5.8 75.8 5.7 74.9 5.8 +20% 77.7 9.8 80.9 10.5 79.3 10.6 78.4 10.9 –5% 68.9 –2.7 70.8 –3.3 69.4 –3.1 68.3 –3.4 –10% 67.1 –5.2 68.0 –7.2 66.9 –6.6 65.5 –7.4 –20% 62.4 –11.8 60.9 –16.9 60.4 –15.7 59.2 –16.3 Poverty line = Food poverty line World Bank Actual 60.5 0.0 59.8 0.0 58.3 0.0 57.4 0.0 +5% 62.4 3.2 62.9 5.2 61.2 5.0 59.9 4.4 +10% 64.4 6.5 65.5 9.6 63.5 8.9 62.5 8.8 +20% 67.4 11.4 70.3 17.6 68.4 17.4 67.1 16.9 –5% 58.3 –3.6 56.6 –5.4 55.3 –5.1 54.4 –5.2 –10% 55.1 –8.9 53.2 –11.1 51.7 –11.3 51.0 –11.2 –20% 49.0 –19.0 44.8 –25.1 43.8 –24.8 44.4 –22.7 Republic of Madagascar Employment and Poverty Analysis Madagascar Poverty and Inequality Update: Recent Trends in Welfare, Employment, and Vulnerability  35 TABLE 1C.2: Percent of Households Reporting Stated Shock (Ordered by Most Frequently Reported in 2005) Climatic shocks 2005 2010 2012 Cyclone 11.59 6.33 6.82 Plant disease (live plants) 11.17 4.28 1.46 Cattle disease 11.04 4.05 2.71 Flood 10.84 6.84 1.76 Late rain 9.09 4.16 2.48 Drought 6.88 9.52 6.19 Blocked road 5.55 0.2 0.03 Other 4.04 0.78 0.72 Early rain 1.68 0.45 0.21 Fire 1.42 0.16 0.05 Plant disease (stocked plants) 1.14 0.74 0.23 Locust invasion 0.59 2.31 1.46 Lavaka-affected crops 0.14 0.01 0.01 Forced migration due to climate shocks 0 0 Hosted relatives due to climate shocks 0.02 0.01 Security shocks 2005 2010 2012 Theft of standing crops 8.4 1.08 0.35 Cattle rustling 4.59 2.23 2.36 Other 1.68 0.79 0.51 Theft of consumption goods 1.34 0.63 0.75 Theft of production inputs 1.19 0.17 0.21 Theft/loss of cash 0.86 0.56 0.44 Theft of stock 0.59 0.11 0.08 Land conflict 0.41 0.21 0.06 Violence against household member 0.17 0.12 0.09 Health shocks 2005 2010 2012 Illness of adult household member 6.54 2.47 2.35 Illness of other household member 4.37 1.41 1.16 Death of other household member 1.58 0.9 1.74 Death of active household member 1.25 0.64 0.87 Other health issues of household members 0.9 0.36 0.36 Economic shocks 2005 2010 2012 General consumer price increase 77.7 2.91 0.81 Increase in product prices 12.87 1.84 0.39 Increase in input prices 7.25 2.42 0.25 Death of person in community 2.35 0.53 0.09 Difficulty finding buyers of agricultural products 2.12 0.33 0.29 Other 1.82 0.55 0.19 Difficulty finding buyers of nonagricultural products 0.98 0.38 0.24 Loss of job of household member 0.76 0.85 0.28 Loss of animal used for traction 0.08 0.04 Farmgate prices too low 0.99 0.09 2005 2010 2012 Other shocks 1.3 0.61 2.15 36 Republic of Madagascar Employment and Poverty Analysis NOTES 19. Some theories predict consumption growth with income growth. See, for example, Carroll (1997). Demographic issues and lifecycle 1. The 2012 poverty statistics were calculated from Enquête Nationale saving are not included in this analysis, as the timeframe for analysis sur les Objectifs Millenaire du Développement (ENSOMD), a house- is relatively short, and reliable data needed to study this aspect of hold survey very similar in format to the country’s previous Living saving are not available. Standards Measurement Surveys (LSMS), the Enquêtes Auprès les 20. Dollar, Kleineberg, and Kraay (2016), for example, find that inequal- Ménages (EPM). ity tends to show mean reversion in cross country data. 2. The 95 percent confidence interval was calculated for changes from 21. See, for example, Christiansen and Dercon (2011), Osborne (2006), 2000 to 2010 in the World Bank’s extreme and absolute poverty and Zimmerman and Carter (2003), which underscore the impor- headcount ratios, and this interval is fairly wide for the extreme tance of risk in farmers’ decisions to utilize lower-effort, lower-return poverty rate. This suggests an even wider confidence interval for the technologies. national poverty line, given that the margin of error increases as the 22. In a recent trial, on-time fertilizer applications registered value/cost poverty line approaches the mode of the distribution. ratios of greater than two in eight of the 21 cases, compared to a 3. A variety of data treatment issues were addressed in the process ratio of zero among those who received fertilizer late, and according of verifying poverty estimates for 2012, but we highlight the main to common rule of thumb value/cost ratios of greater than two are factor here. needed for farmers to adopt fertilizer into their production systems 4. The assumed population structure is based upon a 1993 census, (Livingston et al. 2011). updated by a 2008 census mapping of households. However, the 23. In addition, the sampling strategy differed; and as shown in reliability of Madagascar’s statistics is compromised by the lack of a table 1A.2, sampling was not done with replacement—so that it more recent census. is possible that there is a greater problem of selection bias in the 5. PovcalNet 2012 data. This statement refers to countries with poverty sample. In fact, selection bias is a potential problem in both surveys data only. if households that were either not included or replaced were system- 6. Although PovcalNet does not use spatial price deflators, one can atically different from those that were included. estimate the poverty rate using such deflators at the international 24. The census mapping of 2008 was done in preparation of the national poverty line, and one obtains a rate of 78.4 percent of the population population census that was supposed to take place in 2009 but was in extreme poverty and 91.6 percent poor (living under US$3.10 not undertaken due to the 2009 crisis. 2011 PPP) in 2012. 7. Although the survey instruments available do not allow us to update the geographic distribution of the population on a frequent basis, fluctuations in urban and rural poverty rates can be partially the REFERENCES result of migration of poor households to and from urban areas. 8. GDP growth estimates and poverty and consumption estimates are derived from different sources of data—the former from the national Bi, C. and T. Osborne. 2016. “Transactions Costs, accounts of a country and the latter from household surveys—and Poverty, and Low Productivity Traps: Evidence from very often estimated income and consumption diverge between these two sources. Madagascar’s Informal Micro-Enterprise Sector.” 9. Based on extreme poverty line (World Bank 2014). Washington, DC: World Bank. 10. The Gini coefficient is equal to the area between the Lorenz curve Carroll, C. 1997. “Buffer Stock Saving and the Life and the 45-degree line divided by the sum of this area and the area under this curve, and is expressed as G  – m 1  0 F(y)(1F(y))dy, Cycle-Permanent Income Hypothesis.” Quarterly where  is mean income/consumption and F(x) is the distribution of Journal of Economics 112 (1). income/consumption. 11. Staff calculations (bivariate regression) using EPM 2010. Christiansen, L. and S. Dercon. 2011 Consumption risk, 12. Madagascar’s high population growth rate, estimated at 2.78 percent technology adoption and poverty traps: Evidence (relative to a SSA mean of 2.71). 13. Since the underlying data is not a panel, one cannot conclude that from Ethiopia. Journal of Development Economics poor households lost and then regained access to land over time. 96: 159–173. Average cultivated area in 2012 was 1.68 hectares versus 1.61 in 2001, but this is not a statistically significant difference. David-Benz, Hélène CIRAD. 2011. “A Madagascar: les 14. This was computed as an aggregation of possible responses: lack of prix du riz flambent, sans rapport avec le marché seeds, lack of improved seeds, lack of fertilizer, high cost of inputs, international.” Paris: Cirad, UMR Moisa. high cost of seed, and so forth. 15. The phrasing of the questionnaire referred specifically to limitations Deaton, A. 1992. Understanding Consumption. to land area, and did not ask directly about insecurity of tenure or Clarendon Lectures in Economics. Oxford, UK: conflicts over land. It is possible that respondents blurred the issues of access and tenure security. Oxford University Press. 16. In addition, other issues, including farmers’ knowledge or support Dollar, D., K. Kleineberg, and A. Kraay 2016. “Growth is for introducing new technologies, weather or climatic issues, and soil fertility were mentioned, but not in sufficient frequency to be still good for the poor. European Economic Review included in the top constraints communities mentioned. 81: 68–85. 17. Market integration is in turn important for improving producer prices and seasonal price smoothing. Other constraints relate to Friedman, M. 1957. A Theory of the Consumption seasonality of labor inputs Moser and Barrett (2003) find that a Function. Princeton, NJ: Princeton University Press. promising system of rice intensification (SRI), while requiring low INSTAT. 2013. Enquête nationale sur l’emploi et le external inputs, is difficult for most farmers to practice because the method requires significant additional labor input at a time of the secteur informel—ENEMPSI 2012. Etude nationale. year when liquidity is low and labor effort is already high. Vice Primature chargée de l’économie et de 18. The World Bank conducted a comprehensive vulnerability assess- ment in 2012 using data through 2010 (World Bank 2012). l’industrie. Madagascar Poverty and Inequality Update: Recent Trends in Welfare, Employment, and Vulnerability  37 INSTAT. 2014. Enquête national sur le suivi des objectifs Decomposition Analysis of Inequality and Deepening du millénaire pour le développement à Madagascar. Poverty in Madagascar (2005–2010).” In this Etude nationale. Vice Primature chargée de volume. l’économie et de l’industrie. World Bank. 2012. Madagascar Three Years into the Livingston, G., S. Schonberger, and S. Delaney (2011). Crisis: An Assessment of Vulnerability and Social “Sub-Saharan Africa: The State of Smallholders in Policies and Prospects for the Future. Washington, agriculture.” IFAD. Conference on New Directions DC: World Bank AFR Social Protection Unit. for Smallholder Agriculture, 2011. World Bank. 2014. Face of Poverty in Madagascar: Minten, B. 1999. “Infrastructure, Market Access, and Poverty, Gender, and Inequality Assessment. Agricultural Prices: Evidence from Madagascar,” Washington, DC: World Bank, PREM Africa. Markets and Structural Studies Division (MSSD) Zimmerman, F. J., and M. R. Carter. 2003. “Asset Discussion Paper 26, IFPRI, Washington, DC. Smoothing, Consumption Smoothing, and the Osborne, Theresa. 2006. “Credit and Risk in Rural Reproduction of Inequality under Risk and Developing Economies. Journal of Economic Subsistence Constraints.” Journal of Development Dynamics & Control 30: 541–68. Economics 71 (2): 233–60. Thiebaud, A., T. Osborne and N. Belghith. 2016. “Isolation, Crisis, and Vulnerability: A 38  39 CHAPTER 2 Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010) Alessia Thiebaud Theresa Osborne Nadia Belhaj Hassine Belghith June 2016 The authors acknowledge the significant contributions to this paper by Patrick Leon Randriankolona. Summary B etween 2005 and 2010, Madagascar experienced urban and rural households in the top quintile, about a moderate decrease in its headcount poverty rate. half of the consumption gap (49 percent) is explained However, over the same period, the poorest of the by differences in endowments, and the other half by poor fell deeper into poverty, particularly in rural areas, differences in returns (51 percent). The key structural and inequality increased. Using an unconditional quan- correlates with consumption “explaining” disparities tile regression method proposed by Firpo, Fortin, and between rural and urban areas are remoteness from Lemieux (2009), differences in consumption between urban areas and the level of education of the household groups of interest (urban and rural households) and head. While more investments in transport connectivity changes in consumption over time (between 2005 and and education in rural areas would have a positive effect 2010) are decomposed to identify the main drivers of on consumption and reduce urban-rural inequality, to deepening poverty and increasing inequality, particularly fully realize the potential returns to these investments in rural settings. Urban-rural inequalities in 2010 are would require greater opportunities for urban migration mostly explained by a disparity in household endow- and employment, in addition to economic integration ments, which include some household assets, characteris- with urban areas. tics, shocks, and community-level variables. Differences in such endowments explain 78 percent of the total In addition, we decompose changes in consumption consumption difference between the poorest quintiles between 2005 and 2010 by quintile. We find that the in urban versus rural areas, while differences in returns increased severity of weather shocks, which dispropor- explain the remaining 22 percent, but the role of returns tionately affected rural households and those in the increases and of endowments diminishes for the higher poorest quintiles in 2010, is identified as a key driver of consumption quintiles. Among households in the bottom the observed changes. Decreasing returns to cultivated quintile, over three-fourths of the difference in consump- land and to being located in rural areas are also identi- tion levels between urban and rural households were fied as fundamental drivers. We find that households in attributable to differences in household size and compo- the poorest quintiles experienced the largest consump- sition,1 human capital, climate shocks, and distances to tion losses between 2005 and 2010. Losses were par- food markets. Toward the upper end of the distribution, ticularly large for the rural poorest and were explained the role of returns becomes more prominent. Among the primarily by an increased severity of climate shocks and 40 Republic of Madagascar Employment and Poverty Analysis by falling returns to agriculture, the latter of which is FIGURE 2.1: Incidence of Consumption Growth (Total) associated with a deterioration in the producer price 2005 to 2010 relative to input costs and deteriorating transport condi- 10 tions. In particular, climate shocks explain a –5.3 percent 8 average change in consumption over the period among households in the poorest quintile (–7.0 percent in rural 6 areas). Decreasing returns to agriculture explain a con- 4 Annual growth rate, % sumption change of –5.7 percent for households in the 2 poorest quintile (–6.4 percent in rural areas). Thus, these two factors overexplain the actual change. As with the 0 rural-urban analysis, the issues of remoteness and diffi- –2 culties accessing markets emerge as a key explanation for –4 the decline in rural incomes between the two years. As transport conditions deteriorated and rice policies acted –6 to suppress increases in rice prices, the terms of trade in –8 agriculture plummeted. –10 0 20 40 60 80 100 Offsetting these adverse effects on agriculture was a Expenditure percentile large increase in consumption unrelated to assets (except Source: Calculated using Enquêtes Auprès les Ménages (EPM) 2005, 2010. gender) for male-headed households relative to female headed ones, which are associated with a 13.8 percent increase in consumption for households in the bottom quintile (18.1 percent in rural areas). The net effect on it was the most rural provinces where average poverty consumption of households in the bottom quintile was rates increased the most. Whereas for the country as a a –3.1 percent change between 2005 and 2010 at the whole the increase in inequality was mostly driven by national level, and a –6.0 percent change in rural areas. higher consumption in the top quintile, in rural areas the We provide suggestive evidence that males were able to increase in inequality was mostly due to a deterioration shift secondary work effort into services and other activi- in consumption for the poorest households (figure 2.3). ties, whereas females faced more obstacles in doing so. The objective of this paper is to provide a deeper empiri- cal understanding of why consumption increased for Introduction some groups and not for others between 2005 and 2010, including why the largest consumption decline over Between 2005 and 2010, despite a modest decrease in the period occurred at the bottom of the distribution. Madagascar’s national headcount poverty rate (from Using recentered influence function (RIF) analysis, we 73.2 percent in 2005 to 71.7 percent in 2010), inequality uncover the main drivers of the increase in inequality— increased (see Belghith, Osborne, and Randriankolona and changes in consumption levels—both over time and 2016). The Gini coefficient rose from 38.9 to 42.7 and between urban and rural populations, for each quintile overall the incidence of growth was not favorable to the poor (figure 2.1). On a provincial level, poverty increased in 12 out of 22 provinces in Madagascar. TABLE 2.1: Trends in the Poverty Gap Moreover, there is an important rural-urban dimension (Mean Percentage Shortfall of Consumption to both persistent inequality and changes in consump- Relative to Poverty Line) tion patterns over this period. The poverty gap in rural 2005 2010 Change areas increased whereas that in urban areas decreased Urban 13.6 8.9 –4.7 (see table 2.1). And as is typically the case in poor Rural 34.8 36.7 +1.9 countries, poverty rates tend to be significantly higher Total 31.3 32.0 +0.7 in more rural provinces of Madagascar (see figure 2.2). Moreover, over the period of our study, we find that Source: Belghith, Randriankolona, and Osborne 2016. Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010) 41 FIGURE 2.2: Proportion of Rural Households and Headcount Poverty Rate (by Province) 100% Atsimo Atsinana Androy 90% Vatovavy 80% Fitovinany Amoron’i mania Betsiboka Headcount poverty rate, 2010 Itasy Haute matsiatra Anosy Sofia Atsimo-Andrefana Ihorombe 70% Melaky Analanjirofo Alaotra-Mangoro Bongolava Atsinanana Menabe 60% Sava 50% Boeny 40% Diana 30% Analamanga 20% 50% 60% 70% 80% 90% 100% Proportion of rural households, 2010 Source: EPM 2005, 2010. FIGURE 2.3: Cumulative Density Function of Log Consumption Expenditure a. All households b. Rural only 1 1 0.8 0.8 Cumulative frequency Cumulative frequency 0.6 0.6 0.4 0.4 0.2 0.2 0 0 10 11 12 13 14 10 11 12 13 14 Log consumption expenditure Log consumption expenditure 2005 2010 2005 2010 Source: EPM 2005, 2010. Note: Cutoff points for bottom and top quintiles are indicated with vertical lines. of the distribution. Because we utilize repeated cross- statements with respect to quintiles in different years sectional data, the households falling into a given quintile relate to the respective quintiles for that year only. will have shifted over time, and we cannot trace the persistence for given households of consumption or the We find that households in the poorest quintiles experi- effects of any influence variables on consumption. Rather, enced the largest consumption losses between 2005 and 42 Republic of Madagascar Employment and Poverty Analysis 2010. Losses were particularly large for the rural poorest Overview of Poverty and were explained primarily by an increased severity of climate shocks and by falling returns to agriculture. In in Madagascar particular, climate shocks explain a –5.3 percent average Poverty rates in Madagascar remain exceedingly high, change in consumption over the period among house- particularly in rural areas, and progress toward poverty holds in the poorest quintile (–7.0 percent in rural areas). reduction has been slow. According to internation- Decreasing returns to agriculture explain a consumption ally comparable estimates, Madagascar’s poverty rates change of –5.7 percent for households in the poorest are the highest in the world (Belghith, Osborne, and quintile (–6.4 percent in rural areas). Thus, these two Randriankolona 2016). While events in urban areas are factors overexplain the actual change. Offsetting these an important factor determining the headcount poverty was a large increase in consumption unrelated to assets rate, the protracted lack of progress in reducing extreme (except gender) for male-headed households relative poverty in Madagascar is largely due to a failure to to female-headed ones, which are associated with a improve the lives of the rural poor, a vast majority of 13.8 percent increase in consumption for households in whom work in agriculture or the informal sector (usually the bottom quintile (18.1 percent in rural areas). The both) (World Bank, 2015). net effect on consumption of households in the bottom quintile was a –3.1 percent change between 2005 and Between 2005 and 2010, despite a slight decrease in the 2010 at the national level, and a change of –6.0 percent overall poverty rate, the poorest fell deeper into poverty in rural areas. We suggest that males were able to shift and inequality between the bottom and the top quintiles secondary work effort into services and other activities, increased. Over the period of interest for this analysis, whereas females faced more obstacles in doing so. a modest decrease in the national headcount poverty rate was observed. The headcount poverty rate fell from Having identified the observed drivers of poverty and 73.2 percent in 2005 to 71.7 percent in 2010. However, inequality, we attempt to relate them to the broader at the same time, inequality increased in Madagascar context, events, and policies. We find that a deep urban- (see Belghith, Osborne, and Randriankolona 2016). The rural divide continues to exist in Madagascar and is Gini coefficient rose from 38.9 to 42.7 and overall the explained for the most part by differences in household incidence of growth was not favorable to the poor. On a endowments and characteristics, such as education level, provincial level, poverty increased in 12 out of 22 prov- distance to markets, and exposure to climate shocks. We inces in Madagascar (figure 2.4) find that sharp decreases in returns in rural areas and to cultivated land between 2005 and 2010, together with Poverty decreased in urban areas but increased in rural devastating effects of climate shocks, account for the areas. The urban poverty rate fell from 40.8 to 29.8, majority of the drop in consumption experienced by the and the poverty gap fell by 4.7 percentage points, from poorest households. We relate the decline in returns to 13.6 percent in 2005 to 8.9 percent in 2010. On the rural areas and cultivated land to a context of low trans- other hand, rural poverty increased by from 79.6 to 80.1 mission of international food prices to poor Malagasy while the poverty gap increased 1.9 percentage points producers, increased transport costs, rising agricultural (rising from 34.9 percent in 2005 to 36.7 percent in input costs, and deteriorating access to markets. We also 2010 (Belghith, Osborne, and Randriankolona 2016). highlight the role of climate shocks, which were more Poverty increased the most in the most rural provinces severe in 2010 than in 2005 and disproportionately (figure 2.5). affected the rural poor, contributing to significantly eroding their consumption levels. We identify increased The rural poor and urban poor exhibit significant differ- participation in informal activities, particularly those ences in household composition. As is generally the case in pursued disproportionately by male entrepreneurs, as developing countries, the rural poor of Madagascar tend to a primary means the poorest households used to offset live in households with more members, and with a higher these effects and avoid falling even deeper into poverty. proportion of children. They also have slightly younger household heads than their urban counterparts, on aver- age. Poor rural households are also more likely to be headed by a male, and the household head is more likely to Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010) 43 FIGURE 2.4: Headcount Poverty Rates and Percentage Point Change (between 2005 and 2010, by Province) a. Headcount poverty rates, 2005 b. Headcount poverty rates, 2010 c. Percentage change, 2005–2010 Source: EPM 2005, 2010. FIGURE 2.5: Proportion of Rural Households and Poverty Increase (2005–2010) 9 Atsimo-Atsinana 8 Anosy 7 Increase in headcount poverty (PPs) Betsiboka Vatovavy Itasy 6 Fitovinany 5 4 Androy Atsimo-Andrefana Boeny Alaotra-Mangoro 3 Menabe 2 1 0 Atsinanana 50% 60% 70% 80% 90% 100% Proportion of rural households, 2010 Source: EPM 2005, 2010. PP = percentage points. 44 Republic of Madagascar Employment and Poverty Analysis TABLE 2.2: Summary Statistics for Urban and Rural Households (2010) Marital status Household size Age structure Gender of head of head (Average number (Average members (Households with Age of head (Household heads of members) under 14) male head) (Average years) with spouse) Quintile Urban Rural Urban Rural Urban Rural Urban Rural Urban Rural Bottom 5.6 6.2 46% 54% 72% 78% 42 41 72% 76% Second 4.7 5.6 37% 49% 78% 82% 43 41 75% 80% Middle 4.0 4.9 33% 44% 79% 82% 41 42 72% 79% Fourth 3.6 4.3 25% 38% 76% 84% 41 42 69% 79% Top 2.9 3.4 14% 24% 75% 80% 43 43 59% 70% Education level Distance to market Security level (Avg highest level Health shocks Climate shocks (Households 1+ (Avg security: completed by head/ (Households that had (Households that had hours away from 1 = very poor, spouse, 1–4) 1+ health shocks) 1+ climate shocks) food market) 4 = very good) Quintile Urban Rural Urban Rural Urban Rural Urban Rural Urban Rural Bottom 1.9 1.4 9% 7% 23% 55% 18% 62% 3.1 3.1 Second 2.4 1.6 9% 4% 13% 41% 9% 53% 2.8 3.0 Middle 2.5 1.7 8% 4% 7% 37% 8% 51% 2.7 3.0 Fourth 2.9 1.8 6% 5% 7% 35% 7% 47% 2.7 3.0 Top 3.1 2.2 4% 6% 5% 27% 4% 33% 2.7 2.9 Source: EPM 2010. have a spouse. Similar urban-rural differences are observed poorest quintile were affected by a climate shock, about among richer households as well, as shown in table 2.2. 55 percent of their rural counterparts were, as shown in table 2.2. Poor rural and urban households also differ in the level of human capital and connectivity. Household heads in rural areas tend to be less educated than their urban Methodology counterparts, and this is observed for all quintiles. Rural households are also considerably more isolated than To uncover the proximate determinants of Madagascar’s urban ones: while only 18 percent of urban households urban-rural inequality and changes in consumption in the poorest quintile live one hour or more away from between 2005 and 2010, we utilize the unconditional the food market, 62 percent of rural households in the quantile regression method (based on the approach poorest quintile do (table 2.2). developed by Firpo, Fortin, and Lemieux 2009). This method can be used to identify the determinants of Poor rural and urban households have similar levels of disparities in consumption expenditure both between health shocks and security, but poor rural households are socioeconomic groups and over time, and can be applied significantly more affected by climate shocks. In 2010, to each quantile of the distribution. This allows one 9 percent of urban households in the poorest quintile to “explain” the distribution of consumption expen- and 7 percent of rural households in the poorest quintile diture by a set of factors observed in household- and were affected by at least one health shock. Also, rural community-level data that vary systematically with and urban households in the bottom quintile reported socioeconomic status or that have varied over time.2 The the same average level of security (a 3.1 score on a scale gap between groups or over time is decomposed into two from 1 to 4, with 1 corresponding to “very poor” security parts: one due to group differences in the magnitudes of and 4 corresponding to “very good” security). However, the variables associated with consumption levels (“deter- a sharp difference was observed in terms of climate minants”) and another due to group differences in the shocks: while only 23 percent of urban households in the effects (“returns”) of these determinants. This method Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010) 45 allows us to identify the contributions of: (1) differences setting, distance to the closest food market, security level in household and community characteristics (endowment (self-perception measure, on a scale from 1 to 5), land effects) and (2) disparities in returns to these charac- area under cultivation, and regional effects. teristics (returns effect) to differences in consumption between groups and to changes in consumption over The choice of variables used in the unconditional quan- time at different quantiles. A more detailed discussion of tile regression was made with the objective of mitigat- the methodology is presented in annex 2B. ing concerns over simultaneity bias while attempting to explain as much of the differences in consumption We carry out three separate decompositions. First, we as possible. Since the underlying model of consump- decompose the differences in consumption between tion relies on permanent and temporary influences on urban and rural households in 2010. Second, we decom- households’ real income and we seek to make causal pose changes in consumption between 2005 and 2010 (policy-relevant) inferences, particular attention was for all households. Third, we repeat this latter decompo- given to excluding variables that are less likely to be sition for rural households only, given the disproportion- exogenous or predetermined. In particular, we wish to ate deterioration in their living standards. exclude variables that could themselves be affected by differences in returns (for example, sector of employ- ment) or unobserved heterogeneity in ability or wealth DATA AND VARIABLES (for example, ownership of assets that are primarily for The Enquête Périodique auprès des Ménages (EPM), consumption purposes), as these would bias the coef- collected in 2005 and 2010, is used in this analysis. ficients and make it impossible to infer causal “effects.” The EPM is a nationally representative household-level For productive assets that are also likely correlated with survey conducted by Madagascar’s National Institute of wealth (for example, the specific type of transportation Statistics (INSTAT). It provides extensive information on asset, household use of electricity), new variables were the demographic structure, education, health, employ- created to mitigate this problem. For example, a variable ment, access to infrastructure, and consumption patterns on the proportion of households within the community of Malagasy households in both urban and rural settings. (excluding the household in question) that have elec- The EPM has been collected in 1993, 1997, 1999, 2001, tricity was preferred to a variable on the availability of 2002, 2004, 2005, and 2010. A national survey to moni- electricity within each individual household. Cellphone tor progress against the Millennium Development Goals use was excluded. Although having improved commu- (ENSOMD, in French) was also conducted in 2012 and nication technologies can increase household incomes is similar in its approach to collecting consumption data through a variety of channels, because more well-off (see, for example Belghith, Osborne, and Randriankol- households are more likely to adopt cell phones (and ona 2016). However, it was not suitable for this analysis expenditures on utilities are likely easier to capture than as it lacks several key community-level variables. other household expenditures), we suspect that the effect of unobserved wealth on household cellphone ownership In this exercise, household “endowments” are broadly would introduce bias on all coefficients. Similarly, the defined. They include (1) household characteristics, such “effect” of having a car (versus other forms of trans- as household size, proportion of children in the house- portation) would likely capture the effect of unobserved hold, gender of the household head, age of the household household wealth on consumption, in addition to the head, and marital status of the household head (whether income gains possible from owning a car. We therefore in a couple or not); (2) human capital, as measured by include an indicator variable for whether or not the education level of household head or spouse (whichever household owns any transportation asset, rather than is higher); (3) shocks such as weather shocks and health variables to differentiate which type of asset this is. We shocks (dummy variables which are equal to one if the acknowledge, however, that arguably all variables could household has experienced at least one shock); (4) access be econometrically endogenous; previous educational, to productive assets, including availability of electricity migration, and fertility decisions may be related to in the community (measured as the proportion of house- unobserved heterogeneity of the household. The exact holds in the community which have electricity, exclud- specification used for each decomposition is shown in ing the household itself) and availability of means of table 2.3. For all decompositions, full results are pre- transportation; and (5) location, such as urban or rural sented in the annexes. 46 Republic of Madagascar Employment and Poverty Analysis TABLE 2.3: Decomposition Specifications (1) (2) (3) Urban-rural inequality Inequality over time, Rural inequality over time, Variables 2010 2005–2010 2005–2010 Household size ✓ ✓ ✓ (number of members) Family composition ✓ ✓ ✓ (% of children under 14) Gender ✓ ✓ ✓ (male household head) Age of household head ✓ ✓ ✓ (years) Marital status of household ✓ ✓ ✓ head (in a couple or not) Education (highest level of ✓ ✓ ✓ household head/spouse) Climate shock ✓ ✓ ✓ (at least one climate shock) Health shock ✓ ✓ ✓ (at least one health shock) Location Not included because defines ✓ Not included because sample (urban or rural) decomposition groups restricted to rural households Isolation (time to food market is ✓ N/A* N/A* one hour or more) Security level (self-perception ✓ ✓ ✓ score, 1–5 scale) Electricity (% of households with Not included because closely ✓ ✓ electricity in community) proxies for urban location Transportation (at least one Not included because “time to ✓ ✓ means of transport) food market” is included Crop area Not included because proxies Not included because “urban/ ✓ (land area under cultivation) for rural location rural” is included Regional effects ✓ ✓ ✓ (location by province) *  It was not possible to include the variable “time to food market” in the decompositions over time because of the large number of missing values in EPM 2005. Determinants of Urban-Rural are the counterfactual differences due to disparities in endowments versus disparities in returns, holding the Inequality in 2010 other constant: If the bottom quintile of households in rural areas had had the same endowments as those in In 2010, urban households had a significantly higher urban areas, urban consumption in the bottom quintile average consumption expenditure than rural households, would have been only 19.8 percent higher than rural and this was true across all quintiles. As illustrated in consumption in the bottom quintile, due purely to the table 2.4, on average, the consumption of the poor- effect of changes on the returns to such endowments, all est urban households was 125.5 percent higher than else being equal, instead of 125.5 percent.3 Similarly, if the consumption of the poorest rural households, due the bottom quintile of households in rural areas had had to both differences in endowments and differences in the same returns to their endowments as those in urban returns. The divergence between rural and urban con- areas, urban consumption would have been 88.1 percent sumption levels is greater for the higher quintiles and higher than rural consumption. Taken together, the com- reaches 157 percent for the top quintile. Also shown bination of the better endowments and better returns Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010) 47 TABLE 2.4: Differences in Consumption Expenditure between Urban and Rural Households, and Broad Decomposition into Endowment versus Returns (2010) Percentiles 20 40 60 80 Overall (Log) Consumption Urban 12.112*** 12.527*** 12.867*** 13.257*** (Log) Consumption Rural 11.298*** 11.646*** 11.936*** 12.311*** Consumption Difference 125.5%*** 141.3%*** 153.5%*** 157.5%*** Endowment Component 88.1%*** 71.4%*** 64.5%*** 58.7%*** Returns Component 19.8%*** 40.8%*** 54.0%*** 62.3%*** Source: Calculated using EPM 2010. Note: Percentages indicate counterfactual differences (holding other factors constant). ***Significance at the 1 percent level. **Significance at the 5 percent level. *Significance at the 10% level. Robust standard errors are used. Full results can be found in annex 2C. enjoyed by urban households renders urban consump- in endowments, and the other half by differences in tion in the bottom quintile over twice as high as rural returns (51 percent). consumption in the bottom quintile. Translated into shares of the underlying disparities, figure 2.6 shows the Comparing households in the bottom quintiles across share explained by the endowment versus returns effects milieu (rural versus urban), almost three quarters of the as estimated. Differences in endowments explain 78 per- consumption difference is attributable to differences in cent of total consumption difference between the bottom household size,4 household composition, human capital, quintile of urban and rural populations 2010, while climate shocks, and distances to food markets, as shown differences in returns explain the remaining 22 percent. in table 2.5. The impacts of these differences dimin- Toward the upper end of the distribution, the role of ish for the higher quintiles, however. Across the board, returns becomes more prominent. Among the urban and urban households had higher per capita consumption rural households in the top quintile, about half of the due in part to their size and composition; however, by consumption gap (49 percent) is explained by differences construction the welfare indicator used—per capita FIGURE 2.6: Endowment and Return Effects between Urban and Rural Households (Percentage of Total Explained Divergence) 0.8 Difference in log real per capita total expenditures 0.6 Confidence interval/ endowment effect Confidence interval/ 0.4 returns effect Endowment effect Returns effect 0.2 0 0.2 0.4 0.6 0.8 Quintiles Source: EPM 2010. 48 Republic of Madagascar Employment and Poverty Analysis TABLE 2.5: Contribution to Urban-Rural Consumption Differences, percentage of Total Difference, Selected Endowments (Largest Significant Contributions for Bottom Quintile) Quintiles 20 40 60 80 Household size 13.4% 10.9% 6.2% 5.5% Percentage of children under 14 7.4% 7.5% 10.5% 11.2% Education level of household head 31.5% 26.9% 25.4% 23.3% At least one climate shock 8.7% 6.5% 2.2% 1.0% Time to food market is one hour or more 12.9% 5.7% 4.8% 4.4% Total contribution to consumption difference 73.8% 57.4% 49.1% 45.3% Source: Calculations using EPM 2010. consumption—overstates differences in actual welfare, as together, these characteristics contributed to explaining it does not adjust for adult equivalence (within house- almost three-fourths of the total difference in consump- hold shared resources and differential needs by age.) tion between the urban and rural bottom quintiles. Apart from these demographic variables, differences were explained by urban households’ having better- Climate shocks also played a key role in explaining the educated household heads. The educational attainment urban-rural consumption gap. After remoteness, climate of the household head accounted for between 23 percent shocks were the next most important factor explaining and 32 percent of total urban-rural inequality for all differences in consumption between the poorest rural quintiles. Urban households were also subject to fewer households and the poorest urban households. While climate shocks and enjoyed shorter distances to food 55 percent of the bottom quintile of rural households markets than rural households did in 2010, as illus- experienced at least one climate shock in 2010, only trated in table 2.6. For the bottom quintile, apart from 23 percent of their urban counterparts did, and this household size the next most important correlate with difference explained about 8.7 percent of the total consumption disparities was the time it takes to reach an consumption difference between the two groups in 2010 urban center. While only 18 percent of urban households (table 2.5). are located one hour or more away from the closest food market, about 62 percent of rural households are. The analysis also allows us to estimate counterfactual This difference explained 12.9 percent of urban-rural differences in consumption levels—that is, the degree consumption inequality among the poorest quintile, to which rural consumption would approach urban but significantly lower shares for richer quintiles. All consumption, if a given endowment or returns to that TABLE 2.6: Summary Statistics for Selected Determinants of Urban-Rural Inequality (Largest Contributors to Inequality) Education level Distance to market Household size Age structure (Avg highest level Climate shocks (Households 1h + (Average number of (Average members completed by head/ (Households that had away from food members) under 14) spouse)* 1+ climate shocks) market) Quintile Urban Rural Urban Rural Urban Rural Urban Rural Urban Rural Bottom 5.6 6.2 46% 54% 1.9 1.4 23% 55% 18% 62% Second 4.7 5.6 37% 49% 2.4 1.6 13% 41% 9% 53% Middle 4.0 4.9 33% 44% 2.5 1.7 7% 37% 8% 51% Fourth 3.6 4.3 25% 38% 2.9 1.8 7% 35% 7% 47% Top 2.9 3.4 14% 24% 3.1 2.2 5% 27% 4% 33% Source: EPM 2010. *1= No schooling; 2 = primary schooling; 3 = secondary schooling; 4 = higher level.  Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010) 49 endowment were equal across the groups. The results are incidence of climate shocks among urban households presented in tables 2.7 and 2.8. If returns and all other rendered consumption for the poorest urban people factors were held the same for urban and rural popula- 7.4 percent higher than for their rural counterparts—a tions, better endowments in education for the urban small but significant share of the total disparity. population in the bottom quintile would raise their consumption only 29.2 percent higher than that of the These findings suggest that among the attributes of rural bottom quintile (table 2.7), instead of the actual Madagascar’s rural and urban economies, educational 125 percent. However, urban households also have attainment and remoteness are key structural corre- better returns to education, and doubles the contribu- lates with long-run consumption levels and urban-rural tion of education to the urban-rural consumption gap: if inequality, and that climatic shocks are a major short urban and rural households all had the same education determinant. Lower returns to education in rural areas levels, higher returns to education in urban areas would may induce lower investment in schooling, and the lower still render urban consumption in the bottom quintile economic viability of connecting remote areas to markets 28.7 percent higher than the rural level (table 2.8). could induce some “program placement” bias in our estimations. However, since such investments were made Closer access to markets and reduced exposure to in the relatively distant past and most likely not with climate shocks made the bottom urban quintile sig- perfect foresight on returns in 2010, we conclude that nificantly better off than the rural poorest in 2010. A more of these investments in rural areas would have a greater proximity to food markets (assuming the same positive effect on consumption and reduce urban-rural returns for all households) also raised urban consump- inequality, but that to fully appreciate the returns to tion in the bottom quintile 11.1 percent above that of these investments, more migration, employment, and the bottom rural quintile. Similarly, the more limited integration with urban areas would be needed. TABLE 2.7: Counterfactual Percentage Differences in Consumption Expenditure between Urban and Rural Households, Endowments Percentiles 20 40 60 80 Endowment Household size 11.5%*** 10.1%*** 6.0%*** 5.3%*** Component Percentage of children under 14 6.2%*** 6.8%*** 10.3%*** 11.2%*** Education level of head/spouse 29.2%*** 26.7%*** 26.6%*** 24.6%*** At least one climate shock 7.4%*** 5.9%*** 2.0% 0.9% Time to food market is one hour or more 11.1%*** 5.1%*** 4.6%** 4.3%* Security level 1.8%*** 1.2%*** 0.4% –0.1% Source: Calculated using EPM 2005, 2010. Note: Percentages indicate counterfactual differences. ***Significance at the 1% level. **Significance at the 5% level. *Significance at the 10% level. Robust standard errors are used. Only determinants significant for the bottom quintile are presented. Full results can be found in annex 2C. TABLE 2.8: Counterfactual Percentage Differences in Consumption Expenditure between Urban and Rural Households, Returns Percentiles 20 40 60 80 Returns Household size –24.0%*** –16.9%*** 0.4% –2.6% Component Education level of head/spouse 28.7%*** 18.3%*** 11.5%*** –5.9% At least one climate shock –6.8%** –6.9%*** –3.1% –0.8% Security level –14.4%** –11.8%** –2.3% 4.6% Source: Calculated using EPM 2005, EPM 2010. Note: Percentages indicate counterfactual changes. ***Significance at the 1% level. **Significance at the 5% level. *Significance at the 10% level. Robust standard errors are used. Only determinants significant for the bottom quintile are presented. Full results can be found in the annex 2C. 50 Republic of Madagascar Employment and Poverty Analysis Determinants of Changes their endowments (assets and circumstances) and a more severe experience of shocks (figures 2.7 and 2.8). Even as in Consumption and Inequality endowments increased for all quintiles over the period, between 2005 and 2010, shifts in “returns” to these factors reduced consump- tion for the bottom quintile and increased it for the National Sample top quintile. For the other two quintiles, there was no significant change. If households in the bottom quintile From 2005 to 2010, the real consumption of the had had the same endowments in 2005 and 2010, their Malagasy households in the bottom quintile fell by consumption would have fallen by 6.9 percent over the 3.1 percent. Conversely, the consumption of households period due purely to the decline in returns. While some in the top quintile grew by 10.1 percent. Thus, inequal- moderate improvements in endowment levels were ity increased over the period, with the consumption observed, these were not sufficient to offset the dete- gap between the bottom quintile and the top quintile rioration in returns. In fact, consumption would have growing by 17 percent. The results of the decomposition increased by 4.0 percent in the bottom quintile if returns of changes in consumption for all households between had remained constant over the period. The net effect of 2005 and 2010 are presented for each quintile in a slight improvement in endowments and a considerable tables 2.9 and 2.10. Across the distribution, endowments deterioration of returns was the observed net decline in themselves improved. Of those significantly related to consumption for the poorest (3.1 percent) (table 2.9). consumption levels in the RIF regressions, the endow- The primary negative shift was in returns (table 2.10). ments that helped boost consumption (or offset con- sumption losses) for some segments of the distribution in A large portion (33.3 percent) of the (modest) improve- 2010 relative to 2005 were (1) an increase in education ment in household endowments was due to a reduced of the household head, (2) expanded access to electricity frequency of climate shocks.6 While in 2005 over (in the top three quintiles),5 and (3) greater ownership of 59 percent of households in the bottom quintile had transportation assets (table 2.11). been affected by at least one climate shock, this propor- tion had dropped slightly to 52 percent by 2010. If the The net decrease in consumption for households in the severity of climate shocks had been the same in 2005 bottom quintile was caused by a large drop in returns to and 2010, consumption in the bottom quintile would TABLE 2.9: Counterfactual Percentage Changes in Consumption Expenditure between 2005 and 2010 Percentiles 20 40 60 80 Overall (Log) 2010 consumption 11.390*** 11.764*** 12.100*** 12.587*** (Log) 2005 consumption 11.422*** 11.752*** 12.072*** 12.491*** (Log) cons. difference –3.1%** 1.1% 2.9%** 10.1%*** Endowment component 4.0%*** 2.8%*** 3.5%*** 4.8%*** Returns component –6.9%*** –1.7% –0.5% 5.0%*** Source: Calculated using EPM 2005, 2010. Note: Percentages indicate counterfactual changes. ***Significance at the 1% level. **Significance at the 5% level. *Significance at the 10% level. Robust standard errors are used. Full results can be found in Annex D. TABLE 2.10: Endowment and Returns Components as a Percentage of Total Change in Consumption (by Quintile, All Households) Percentiles 20 40 60 80 Endowment component 122% 255% 117% 49% Returns component –222% –155% –17% 51% Source: Calculated using EPM 2005, 2010. Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010) 51 TABLE 2.11: Counterfactual Percentage Changes in Consumption Expenditure between 2005 and 2010, Changes in Endowments (Holding Returns Constant) Percentiles 20 40 60 80 Located in rural area 0.2%** 0.2%** 0.2%** 0.3%** Age of household head –0.1%** –0.1%** –0.1%* 0.0% Education level of head/spouse 0.6%*** 0.7%*** 0.9%*** 1.4%*** At least one climate shock 1.3%*** 0.4% 0.4% –0.2% Security level 0.7%*** 0.5%*** 0.5%** 0.6%** Access to electricity in community 0.3%*** 0.7%*** 1.5%*** 2.6%*** Means of transport 0.5%*** 0.5%*** 0.6%*** 0.6%*** Source: Calculated using EPM, 2010. Note: percentages indicate counterfactual changes. ***Significance at the 1% level. **Significance at the 5% level. *Significance at the 10% level. Robust standard errors are used. Only determinants significant for the bottom quintile are presented. Full results can be found in annex 2D. FIGURE 2.7: Endowment and Return Effects from 2005 to 2010 (Share of Explained Divergence, by Consumption Quintile) 0.8 Difference in log real per capita total expenditures 0.6 Confidence interval/ endowment effect Confidence interval/ 0.4 returns effect Endowment effect Returns effect 0.2 0 0.2 0.4 0.6 0.8 Quintiles Source: EPM 2005, 2010. FIGURE 2.8: Changes in Consumption, Endowments, and Returns Components (2005–2010, by Quintile) 12% 10% Change in consumption 8% Net consumption change (2005–2010) 6% 4% Endowment component 2% 0% Returns component –2% –4% –6% –8% 20 40 60 80 Quintiles Source: EPM 2005, 2010. 52 Republic of Madagascar Employment and Poverty Analysis have actually increased by 1.3 percent holding all other constant over the period, the poorest quintile would factors constant (table 2.11). have experienced an increase in consumption of about 13.8 percent to due the differential benefits of having Despite their slightly lower frequency, climate shocks a male household head, all else equal. They would be had a stronger negative impact on consumption for the higher for the top, by 14.2 percent, and somewhat less poorest quintile of households in 2010 than they did in disparate in the middle (ranging from 7.4 to 8.5 percent) 2005, as shown in table 2.12. If they had been hit by (table 2.12). As discussed below, this appears to indicate the same number of shocks as in 2005, the consump- that female-headed households were less able to offset tion of the poorest households would have declined by declining returns through secondary off-farm employ- as much as 5.3 percent due only to the greater severity ment, and when employed, they received a lower wage. of the 2010 shocks. Thus, compared to 2005, signifi- cantly more devastating climate shocks affected just a For the bottom quintile of the distribution, the combined slightly lower proportion of households in 2010, and decrease in consumption caused by the deterioration in this explained a large share of the observed net drop in returns to rural activities and to more severe climate and consumption among the poorest households.7 health shocks were larger than the improvement brought about by a moderate improvement in endowments and In addition, a large portion of the drop in consumption by higher returns to male-headed households. Together expenditure was explained by decreasing returns to eco- these influences yielded a net 3.1 percent decrease in nomic activities in rural areas. Holding endowments con- consumption expenditure between 2005 and 2010 for stant, lower returns to assets and circumstances facing the the bottom quintile. For other quintiles, returns to rural rural population would account for a 5.7 percent drop in location decreased even more than for the top, but the consumption for the poorest quintile between 2005 and effects of climate shocks and health shocks were less 2010. More severe effects of health shocks also explain a severe (table 2.12). Coupled with significant improve- significant, but smaller share of the decline in consump- ments in endowments (particularly education and tion (table 2.12). However, returns do not change over access to electricity), this resulted in consumption levels the period for basic household characteristics—in par- that were either not statistically different or that were ticular, household size, percentage of household members higher than the 2005 ones for the other quintiles in our that are children, the age of the household head, marital analysis. status of the household head, and his or her educational attainment level. Moreover, we do not find evidence of a Tables 2.13 and 2.14 show the percentage of the total change in the returns to security, households’ ownership changes in consumption over the period associated of a means of transportation, or community-level access with each significant shift in “endowments” versus the to electricity. “effects” or returns to those endowments. As shown, in order of importance for shifts in consumption of the Over this period, however, the returns to opportunities bottom 40 percent of the distribution are the change in for male-headed households relative to female-headed incomes for male-headed households (with a positive ones diverged for all quintiles. Holding endowments effect, noted by “+”), followed by the change in incomes TABLE 2.12: Counterfactual Percentage Changes in Consumption Expenditure between 2005 and 2010, Returns (Holding Endowments Constant) Percentiles 20 40 60 80 Located in rural area –5.7%** –9.2%*** –15.5%*** –21.1%*** Male household head 13.8%*** 8.5%** 7.4%* 14.2%** At least one climate shock –5.3%*** –2.3%* –0.5% 1.1% At least one health shock –1.7%*** –1.2%*** 0.0% –1.0% Source: Calculated using EPM 2005, EPM 2010. Note: Percentages indicate counterfactual changes. ***Significance at the 1% level. **Significance at the 5% level. *Significance at the 10% level. Robust standard errors are used. Only determinants significant for the bottom quintile are presented. Full results can be found in the annex 2D. Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010) 53 TABLE 2.13: Contribution and Offsetting Factors to 2005–2010 Change in Consumption by Quintile, Percentage of Total Change, Selected Returns (Largest Significant Contributions for Bottom Quintile) Percentiles 20 40 60 80 Change in returns to being located in rural area –184% –882% –583% –247% Change in effects of climate shocks –169% –209% –17% 11% Change in effects of health shocks –53% –109% 0% –10% Change in returns to male household head 403% 745% 245% 139% Source: Calculated using EPM 2010, EPM 2005. TABLE 2.14: Contributions and Offsetting Factors 2005–2010 Change in Consumption, Percentage of (Absolute Value of) Total Change, Selected Endowments (Significant Effects Only) Percentiles 20 40 60 80 Located in rural area 6% 18% 7% 3% Age of household head –3% –9% –3% 0% Education level of household head/spouse 19% 64% 31% 15% At least one climate shock 41% 36% 14% –2% Security level 22% 45% 17% 6% Percentage of households with electricity in community 9% 64% 52% 27% Ownership of means of transport 16% 45% 21% 6% Source: Calculated using EPM 2010, 2005. in rural areas (–), the effect of climate shocks (–), effects From 2005 to 2010, the consumption of rural households of health shocks (–), the education of the household in the bottom quintile of the distribution fell by 6.0 per- head (+), the reduced frequency of climate shocks (+), cent, almost twice as much as the drop experienced by the level of community electrification (+), the security the poorest rural and urban households together, 3.1 per- level (+), and ownership of some means of transport cent (table 2.15). In contrast, the consumption of the top (+). Figure 2.9 shows the effects of the factors that were rural quintile did not change significantly. Results of the significant for the bottom quintiles—climate and health decomposition of changes in consumption between 2005 shocks, returns to rural economic activities, and differen- and 2010 among rural households are presented for each tial gains by male-headed households. quintile in tables 2.16 and 2.17. As expected given the results of the foregoing sec- tion “Determinants of Change in Consumption and Determinants of Changes Inequality between 2005 and 2010, National Sample,” the drop in consumption for the bottom rural quintile in Consumption and Inequality was explained for the most part by a fall in returns between 2005 and 2010, Rural (figure 2.10 and figure 2.11). Modest improvements in endowments were not sufficient to offset the decrease Households Only in consumption brought about by these falling returns. Holding returns constant, rural households would have To better understand the reasons for declines in returns had a 2.8 percent higher in consumption in 2010 than in to rural economic activity and analyze factors that may 2005. However, holding endowments constant, returns affect rural households differently, we next decompose would have caused consumption to drop by 8.6 percent changes over time for rural households only. over the same period (table 2.15). 54 Republic of Madagascar Employment and Poverty Analysis FIGURE 2.9: Main Determinants of Change in Consumption (2005–2010) Counterfactual change in consumption 20% 15% Returns to rural area 10% Effects of health shocks 5% Effects of climate shocks 0% Returns to male household head –5% –10% –15% –20% –25% 20 40 60 80 Quintiles Source: EPM 2005, 2010. Note: Effects smaller than 2% or not significant for the bottom quintile not pictured. The increased severity of climate shocks in 2010 was the droughts, which had a devastating impact on harvests. main determinant of the decline in consumption for the During the same year, the cyclone Hubert brought bottom rural consumption quintile. Madagascar is highly considerable damage to eastern coastal provinces and susceptible to cyclones, floods, droughts, locust infes- generated severe floods, which destroyed large quantities tations, and animal and plant diseases, which expose of agricultural production. As a result, food insecurity the population to considerable risks. Lacking adequate issues affected over 80 percent of the Malagasy popula- mechanisms for ex ante or ex post risk mitigation, the tion in 2010 (FAO 2010). population of Madagascar is particularly vulnerable to climate risks, which can cause a great deal of physical Whereas the percentage of rural households in the destruction and erode the livelihoods of the rural popu- bottom quintile experiencing at least one climatic lation, in particular (Auffret 2014). shock declined slightly, from over 60 percent in 2005 to 55 percent 2010, the adverse effects of climatic Between 2005 and 2010, Madagascar was hit by a series shocks in 2010 were greater than in 2005. The effect of of particularly severe climate shocks, which caused the fall in the frequency of experiencing these shocks extensive physical damage and led to widespread food would have caused a small increase in consumption insecurity. In 2008, three consecutive cyclones hit the for the poorest rural quintile, had the severity (or country, affecting 17 out of 22 regions (Auffret 2014). In “returns”) held constant with those in 2005 (+1.8 per- 2010, the southern regions were affected by prolonged cent) (table 2.16). However, consumption would have TABLE 2.15: Influences on Changes in Consumption Expenditure between 2010 and 2005, Endowments versus Returns Component (Percentage of 2005 Consumption, Rural Only) Percentiles 20 40 60 80 Overall (Log) rural 2010 consumption 11.303*** 11.652*** 11.945*** 12.323*** (Log) rural 2005 consumption 11.365*** 11.677*** 11.964*** 12.319*** Consumption change –6.0%*** –2.5%** –1.9% 0.4% Endowment component 2.8%*** 1.3%* 0.5% 1.3% Returns component –8.6%*** –3.7%*** –2.3%* –0.9% Source: Calculated using EPM 2005 and EPM 2010. Note: Percentages indicate counterfactual changes. ***Significance at the 1% level. **Significance at the 5% level. *Significance at the 10% level. Robust standard errors are used. Full results can be found in Annex E. Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010) 55 TABLE 2.16: Contributing and Offsetting Influences on Net Changes in Consumption between 2005 and 2010, Rural Only: Endowments (Percentage Changes, Significant Effects Only) Percentiles 20 40 60 80 Endowments Household size –0.6%* –0.6%* –0.6%* –0.5%* Percentage of children under 14 –0.3%** –0.4%*** –0.6% –0.9% Education level of head/spouse 0.4%*** 0.5%*** 0.5% 0.7% At least one climate shock 1.8%*** 0.5% –0.1% 0.0% Security level 0.6%*** 0.2% 0.2% 0.4% Access to electricity in community 0.4%*** 1.1%*** 1.6% 2.9% Means of transportation 0.3%*** 0.4%*** 0.4% 0.5% Cultivated land –0.5%*** –0.5%*** –0.5% –0.6% Source: Calculated using EPM 2005, 2010. Note: Percentages indicate counterfactual changes. ***Significance at the 1% level. **Significance at the 5% level. *Significance at the 10% level. Robust standard errors are used. Only determinants significant for the bottom quintile are presented. Full results can be found in annex 2C. TABLE 2.17: Contributing and Offsetting Influences on Net Changes in Consumption between 2005 and 2010, Rural Only, Returns (Percentage Changes, Significant Effects Only) Percentiles 20 40 60 80 Returns Male household head 18.1%*** 15.4%*** 9.5%** 7.1% At least one climate shock –7.0%*** –2.3% 0.9% 0.3% At least one health shock –2.0%*** –1.4%** 0.2% 0.9% Access to electricity in community –0.4%* 0.1% –0.2% 0.0% Cultivated land –6.4%*** –6.1%*** –6.3%*** –3.8%** Source: Calculated using EPM 2005, 2010. Note: Percentages indicate counterfactual changes. ***Significance at the 1% level. **Significance at the 5% level. *Significance at the 10% level. Robust standard errors are used. Only determinants significant for the bottom quintile are presented. Full results can be found in annex 2E. FIGURE 2.10: Endowment and Return Effects from 2005 to 2010 (Rural Households) 0.05 Difference in log real per capita total expenditures 0 Confidence interval/ endowment effect Confidence interval/ –0.05 returns effect Endowment effect Returns effect –0.1 –0.15 0.2 0.4 0.6 0.8 Quintiles Source: EPM 2005, 2010. 56 Republic of Madagascar Employment and Poverty Analysis FIGURE 2.11: Changes in Consumption, Rural Households, Endowments and Returns Components (2005–2010) 4% 2% Net consumption Change in consumption change (2005–2010) 0% Endowment –2% component –4% Returns component –6% –8% –10% 20 40 60 80 Quintiles Source: EPM 2005, 2010. FIGURE 2.12: Main Determinants of Change in Consumption (Rural Households, 2005–2010) Counterfactual change in consumption 20% 15% 10% Effects of climate shocks Returns to cultivated land 5% Effects of health shocks 0% Returns to male household head –5% –10% –6.4% –6.1% –6.3% –3.8% 20 40 60 80 Quintiles Source: EPM 2005, 2010. Note: Effects smaller than 2% or not significant for the bottom quintile not pictured. decreased by 7.0 percent due to the stronger effects of Figure 2.13 shows the proportion of households affected climate shocks alone (holding endowments constant) in by at least one climate shock by province in 2005 and 2010 (table 2.17, figure 2.11). The negative effects of 2010, and figure 2.14 shows the top three shocks for climate shocks on the bottom rural quintile were sig- 2010. Although not perfectly correlated, one of the nificantly more pronounced than those observed for all southern regions with a high percentage of households households combined (–5.3 percent change, as reported affected in 2005 is also highly affected in 2010, and for the national sample in the foregoing section). Rural shocks are more frequent in the south and the west in households depend for their livelihoods on subsistence both years. Droughts, cyclones, floods, late rains, plant agriculture and are more vulnerable to climatic events, diseases, locust invasions, and cattle diseases were fre- and therefore such shocks explain a greater portion of quently reported shocks in all consumption quintiles, but the change in consumption over time for rural house- they affected a significantly greater share of the poorest holds than for the national sample. households than households in other quintiles. Droughts, cyclones, and floods were the three most-cited shocks Although different households experience shocks in in the bottom three quintiles (more so than any other different years, there is some spatial correlation in the type of economic, health, or security shock). Almost one frequency of reported climatic shocks. in five households in the bottom quintile (18.6 percent) Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010) 57 FIGURE 2.13: Proportion of Households Hit by at Least One Climate Shock (by Province) a. 2005 b. 2010 Source: Calculated using EPM 2005, 2010. Note: Numbers in brackets represent the proportion of households, with 1 = 100 percent. FIGURE 2.14: Top Three Shocks by Quintile (2010) 30 Affected households (%) 25 Consumer price increase 20 15 Cattle disease Plant disease 10 Drought Drought Drought Drought Drought Cyclone Cyclone Cyclone 5 Flood Flood Flood Flood 0 First Second Third Fourth Fifth Quintiles Source: EPM 2010. reported being hit by drought in 2010, while only quintile, 6.0 percent in the third quintile, 3.7 percent in 10.3 percent in the second quintile did so, 8.5 percent in the fourth quintile, and 2.2 percent in the top quintile. the third quintile, 6.3 percent in the fourth quintile, and Finally, floods were experienced by 10.9 percent of 3.8 percent in the top quintile. The same pattern was households in the bottom quintile, 8.7 percent in the true for cyclones: 11.5 percent of the poorest households second quintile, 7.2 percent in the third quintile, 5.4 per- were hit by cyclones, but only 8.2 percent in the second cent in the fourth quintile, and only 2.0 percent in the 58 Republic of Madagascar Employment and Poverty Analysis top quintile. This, combined with the RIF decomposition asset accumulation. A significant decline in returns to results, shows the key role that climatic shocks played in cultivated land explains a large part of the decrease in deepening poverty in 2010. As both the frequency and consumption for rural households in the bottom quintile severity of these shocks are expected to increase due to of the distribution between 2005 and 2010 (table 2.17). climate change, the vulnerability of the poor popula- If endowments had not varied over the period, the poor- tion to extreme weather events is also likely to increase. est rural households would have experienced a drop Over the next 50 to 100 years, average temperatures of 6.4 percent in their consumption due to a decline in Madagascar are expected to rise by 2.5 degrees. As in returns to the land they cultivate. Middle quintiles a consequence, average annual rainfall in Madagascar experienced drops in returns to cultivated land of a is forecast to decrease, while at the same time sharp similar magnitude, but for the top quintile the drop was increases in precipitation will occur during the rainy considerably smaller. season (Auffret 2014). As was the case for the national sample, increased Changes in household and community assets also played returns to male-headed households relative to female- a role. Households in the bottom two quintiles had headed households prevented average consumption in greater means of transportation, greater community-level the bottom quintile from dropping even further. Holding access to electricity, and greater education relative to endowments constant, male-headed households would the bottom two quintiles in 2005. Offsetting this was a have experienced an 18.1 percent increase in consump- decline in the area of land cultivated (table 2.16). tion between 2005 and 2010, which exceeds the effect in the national sample (rural and urban households As with the national sample, a decline in returns to these combined of 13.8 percent) just discussed, rather than factors explains shifts between the two years more than 6 percent. Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010) 59 Explaining Changing Patterns FIGURE 2.15: Food Supply by Type of Aliment (kcal/capita/day) in Returns 100% In this section, we provide additional analysis of the underlying reasons for falling returns to land and to 80% living in rural areas between 2005 and 2010, as well as of the differential returns to male-headed households over time. The key factors identified are policies in rice 60% markets and deteriorating transport conditions, which worsened the already problematic degree of integra- 40% tion of Madagascar’s rice markets (Moser, Barrett, and Minten 2009). In addition, outcomes in labor markets 20% suggest that preferences for males in the off-farm labor market in a year of low agricultural productivity made a more significant difference in households’ ability to 0% offset agricultural losses through off-farm work. r l ali e a a a ga ca ric ric ric qu M as ne Af Af Af bi ag Se am rn rn ad ste he oz M ut M Ea So RICE MARKETS AND POLICIES Other ailments Other cereals Rice Most Malagasy households are not only highly dependent on agriculture for their livelihoods but are Source: FAOSTAT 2016. especially dependent on rice. Rice constitutes a more important proportion of their consumption than it does for households in other countries in Sub-Saharan Africa (figure 2.15), and the majority of them are both consum- Richer households tend to be slightly more concen- ers and producers of the grain. Rice paddy is produced trated in rice production, with almost 60 percent of by 87 percent of agricultural producers and is the main their total production being rice paddy relative to crop across all quintiles of the income distribution. 40 percent for the bottom quintile (table 2.18. Given TABLE 2.18: Share of Each Product in Total Production by Consumption Quintile (2005) Quintile of consumption Group of products 1 (poorest) 2 3 4 5 (richest) Paddy 39.9 46.0 49.3 54.6 59.9 Maize and other 2.2 2.4 2.7 2.2 2.1 Cassava 28.4 26.7 22.4 21.2 17.3 Sweet potatoes 6.4 7.1 6.3 5.5 4.1 Other tubers 2.6 4.1 5.4 4.0 4.0 Leguminous 2.1 2.2 2.3 2.2 2.0 Vegetable 3.0 1.9 1.9 2.2 1.8 Fruit 7.8 5.2 5.2 4.2 4.5 Industrial culture 6.1 3.5 3.6 3.2 3.3 Cash crops 1.4 1.0 0.9 0.8 0.9 Total 100.0 100.0 100.0 100.0 100.0 Source: EPM 2005. 60 Republic of Madagascar Employment and Poverty Analysis FIGURE 2.16: Location of Community Granaries FIGURE 2.17: Movements in World and Local Rice Prices 600 500 400 300 200 100 0 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 Rice, Thai, A1.special, $/mt, nominal$ Madagascar rice price (100 = average 2000 price) Source: World Bank Commodity Price databank and INSTAT. Source: Randrianarisoa 2003. 68 percent of households that produce rice, more than two-thirds need to purchase rice at some point during the year (Auffret 2014). Many also experience severe the scarcity of adequate storage facilities (figure 2.16), seasonal food shortfalls. In 2005, higher rice prices most Malagasy households are not able to economi- coincided with relatively good rice yields to improve cally store sufficient rice for self-consumption over the incomes of rice producers, particularly net rice extended periods of time. They tend to alternate sellers. In fact, poorer farmers opted to sell more of between being net sellers and net buyers of rice in their rice crop than richer farmers, who may have different seasons of the year. Large shares of the rural preferred to consume more of it (see table 2.19). The population become net rice consumers: Among the international price of rice rose significantly after 2005, TABLE 2.19: Share of Crop Sold by Consumption Quintile (2005) Quintile of consumption Group of products 1 (poorest) 2 3 4 5 (richest) Paddy 30.8 28.4 27.8 28.1 23.7 Maize and other 37.3 37.6 30.6 38.0 36.6 Cassava 29.2 31.7 30.2 28.2 32.6 Sweet potatoes 27.6 22.0 20.2 21.8 28.6 Other tubers 35.7 36.3 35.7 42.8 36.9 Leguminous 47.5 49.6 48.6 52.3 50.4 Vegetable 73.3 54.4 58.1 53.1 56.1 Fruit 57.6 58.2 58.7 57.1 52.7 Industrial culture 25.7 41.8 46.9 48.3 44.9 Cash crops 81.2 87.5 83.5 75.5 79.5 Source: EPM 2005. Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010) 61 FIGURE 2.18: International and Domestic Rice Prices, Selected Countries (2007–2009) 500 500 2000 400 400 1500 300 300 1000 200 200 500 100 100 Senegal Mali Madagascar 0 0 0 janv.–07 juil.–7 janv.–08 juil.–8 janv.–09 juil.–09 01/07 07/07 01/08 07/08 01/09 07/09 janv.–07 juil.–7 janv.–08 juil.–8 janv.–09 juil.–09 Riz CAF Riz importé conso. Riz local conso. Riz producteur Source: David-Benz and Lancon 2013. spiking in 2008 and 2012; however, this did not fully removing value-added tax on rice imports completely translate into increased revenues for Malagasy rice in July of 2008. Anticipating drought and further producers. The government of Madagascar sought suc- increases in the international rice price the government cessfully to prevent a more rapid escalation of domes- preordered rice imports (50,000 metric tons of Indian tic rice prices in the face of rising world prices, and the rice) and banned rice exports (David-Benz 2011). As domestic price of rice was kept relatively stable (fig- a result, Madagascar’s consumer rice price increases ure 2.17). In 2007, the government removed tariffs on were milder than those of several other Sub-Saharan rice imports and decreased ad valorem taxes on rice, rice importer countries (figure 2.18). 62 Republic of Madagascar Employment and Poverty Analysis FIGURE 2.19: Median Nominal Price Received Responding to the worsening terms of trade, across all for Rice Paddy (by Consumption Quintile) quintiles, there was a shift out of rice relative to 2005 (table 2.20), and poorer households sold a smaller share 900 of their crop in 2010, whereas richer households with 800 more advantageous market access increased this share 700 (table 2.21). 600 Price, ariary 500 Although lower prices improve the welfare of net 400 consumers, including those in urban areas, they are detrimental to producers. In a variety of Asian coun- 300 tries, for example, high food prices have been found to 200 push up rural wages and reduce poverty rates—in the 100 long term and in some cases in the short term, despite 0 the adverse impact on some net food buyers (Ivanic and 2005 2010 Martin 2014). Indeed, in a period of rising rice prices Poorest Second in Madagascar, between 2001 and 2005, real consump- Third Fourth tion among the bottom 40 percent grew (see Belghith, Richest Community mean Osborne, and Randriankolona 2016). Source: EPM 2005, 2010. On net, relative to 2005, producers in 2010 experienced a decline in the terms of trade in agriculture in rural In fact, prices paid to rice producers were in many cases areas, especially those in the bottom three quintiles significantly lower in 2010 than they had been in 2005; (figure 2.20). While local producer prices generally fell, the poor in particular received a much lower price. the price of agricultural inputs increased considerably, Figure 2.19 shows that the median producer price of rice and the correlation between adverse terms of trade in was higher in 2005 at 546 ariary per kilogram than in agriculture and poverty is pronounced in 2010 in a 2010, at 533 ariary, despite continued increases in both way that it was not in 2005. These factors are likely the world price and the domestic market price in urban to partly explain the decrease in returns to cultivated centers.8 This nominal decline coincided with a cumula- land and to being located in a rural area for the bottom tive increase in the consumer price index of 58 percent. quintile. FIGURE 2.20: Relative Prices of Fertilizer versus Rice (2005 and 2010 by Household Consumption Quintile) 8 7 6 5 4 3 2 1 0 Poorest Second Third Fourth Richest 2005 NPK/Paddy 2010 NPK/Paddy 2005 Urea/Paddy 2010 Urea/Paddy Source: EPM 2005, 2010. NPK = Nitrogen, Phosphorus, Potassium. Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010) 63 TABLE 2.20: Share of Each Production in Total Production by Consumption Quintile (2010) Quintile of consumption Group of products 1 2 3 4 5 Paddy 35.8 40.0 44.9 45.0 47.2 Maize and others 3.8 3.8 3.7 3.8 5.4 Cassava 33.6 24.6 23.1 18.5 16.7 Sweet potatoes 8.0 6.7 4.5 5.9 3.8 Other tubers 1.4 2.2 2.0 2.7 1.9 Leguminous 2.2 2.3 2.9 2.6 2.8 Vegetable 1.9 3.1 2.9 4.5 4.2 Fruit 7.0 10.4 8.5 10.1 9.1 Industrial culture 4.5 5.5 6.3 5.6 7.5 Cash crops 1.8 1.3 1.2 1.1 1.3 Total 100.0 100.0 100.0 100.0 100.0 Source: EPM 2010. TABLE 2.21: Share of Crop Sold by Consumption Quintile (2010) Quintile of consumption Group of products 1 2 3 4 5 Paddy 24.8 24.0 26.5 27.2 33.2 Maize and others 38.0 42.4 42.5 47.8 62.3 Cassava 31.5 36.6 42.2 39.6 47.8 Sweet potatoes 25.2 23.0 18.5 24.8 32.3 Other tubers 17.1 26.6 33.9 38.1 37.0 Leguminous 53.8 57.9 61.0 59.1 65.3 Vegetable 60.4 75.8 69.5 75.3 84.1 Fruit 60.6 76.5 67.8 64.3 73.4 Industrial culture 31.9 39.8 37.3 40.4 62.8 Cash crops 82.2 86.8 83.8 83.6 84.5 Source: EPM 2010. THE ROLE OF INFRASTRUCTURE quintile. This signals a deterioration of road connectiv- ity as a consequence of the deep political crisis of 2009 In addition to rice policies, sharp rises in transport costs (after which revenues for road maintenance diminished). due to deteriorating physical infrastructure appear to For example, for the poorest quintile in 2005, the clos- have affected market integration and the returns and est primary school was 44 kilometers away on average, endowments of rural households.9 Between 2005 and whereas this doubled to 90 kilometers by 2010. The time 2010, while the terms of trade worsened in agriculture, needed to reach a food market rose for the bottom quin- particularly for the poor, the poor also became signifi- tile from 1.9 to 2.4 hours, and the time to reach a main cantly more isolated (figure 2.21). The time required urban center from less than 6 to almost 12 hours. Also to reach various markets and services—already high in striking from the figures is that in 2010 consumption is 2005—increased dramatically between 2005 and 2010. closely and inversely related to these distances and times A sharp increase in transport costs, affecting all quintiles, to reach schools, food markets, and urban centers, much is observed between 2005 and 2010, with the highest more so than in 2005. For example, whereas it took over transport costs observed for households in the bottom 90 minutes for the richest quintile to reach food markets, 64 Republic of Madagascar Employment and Poverty Analysis FIGURE 2.21: Average Time to Food Market Access to electricity also appears to play a role in stimu- by Quintile (Hours) lating a modest increase in local incomes. For the average rural household in the poorest quintile, in 2005 less than 1 percent of other households in the community were Richest connected to the electric grid. By 2010, this proportion had not changed significantly. In contrast, rates of access to electricity increased considerably for other groups: Fourth the proportion of community households with electric- ity went from about 12 percent to about 18 percent for the top rural quintile, and from 40 percent to 48 percent Third for the top urban quintile. As illustrated in figure 2.22, little progress had been made by 2010 in reaching areas Second outside of the province of the capital city with electricity, and this may have arguably contributed to deepening the divide between urban and rural areas and between top Poorest and bottom quintiles. At the same time, over the period 2005 to 2010, it was in communities with the poorest households (in the bottom two quintiles) where the rate 0.0 0.5 1.0 1.5 2.0 2.5 of community-level electricity access was associated with 2005 2010 higher consumption. This pattern does not fit the stan- Source: EPM 2005, 2010. dard story: that income determines electrification. While the decline in returns to electricity for the bottom quintile completely offsets the gain in the electrification rate, for the second quintile, this is not the case. On the margin, it took over 144 minutes for the poorest quintile.10 In greater access to electricity appears to help increase eco- addition, the cost of transporting goods to the nearest nomic opportunities, albeit only modestly. main urban center increased dramatically between 2005 and 2010. As shown in table 2.22, the average costs by region increased by between 36 and 80 percent in the RETURNS BY GENDER rainy season over this timeframe, and up to 99 percent for Antsirinana in the dry season. Because it is unlikely All in all, between 2005 and 2010, the poorest Malagasy that the deterioration in connectivity could have been households were severely affected by a combination closely correlated with poverty rates, particularly in a of severe climate shocks, lower producer rice prices, time of postcrisis scarcity in fiscal and other resources, increasing agricultural input costs, rising transport costs it is likely that this deterioration had a causal effect on and deteriorating connections to markets. This caused poverty. By increasing the cost of accessing inputs and their consumption to decline significantly, often lead- reducing the producer price for agricultural households, ing to food insecurity, particularly in rural areas. In this while also making access to health services more dif- analysis, this is evidenced by strong decreases in returns ficult, deteriorating transport conditions reduced the real to being located in a rural area, to cultivated land, and incomes and consumption levels of those most affected. by large negative effects of climate shocks. TABLE 2.22: Transport Cost for a 50-Kilogram Bag of Rice, Dry Season and Rainy Season Average (in 2005 U.S. Dollars, by Quintile) Quintile Bottom Second Middle Fourth Top All 2005 $1.65 $1.57 $1.41 $1.32 $1.00 $1.40 2010 $2.19 $2.08 $2.07 $1.96 $1.66 $1.99 Source: EPM 2005, 2010. Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010) 65 FIGURE 2.22: Average percentage of Households with Electricity in a Community (Percentage Ranges by Province) Source: EPM 2005, 2010. 66 Republic of Madagascar Employment and Poverty Analysis However, the analysis also shows sharp increases in among urban-based owner-occupied microenterprises, returns to male-headed households relative to female- male entrepreneurs have higher returns than female ones, headed ones, with the divergence greater for poorer even controlling for other factors. Sectors such as trans- quintiles. While these increases have not been sufficiently port or logging tend to be disproportionately pursued by high to completely offset the negative impacts of falling males, and these sectors may have experienced increased returns to cultivated land and agriculture, nor the devas- profitability relative to those females pursued dispropor- tating effects of climate shocks, they have prevented the tionately, such as textiles production. poorest households headed by males from falling much deeper into poverty, particularly in rural areas. Narrowing the focus to the group that suffered the largest consumption decline over the period of interest A possible explanation is that men were better able (i.e., households in the poorest rural quintile), we find to find employment outside of the agricultural sector further support for this hypothesis. Agriculture remained between 2005 and 2010, while women were unemployed the main sector of employment for both male and female or stayed in agriculture, experiencing falling returns. As household heads in the poorest rural quintile both in shown in figure 2.23, between 2005 and 2010, the pro- 2005 and 2010. As shown in figure 2.25: Employment of portion of individuals looking for employment increased Household Heads in the Bottom Quintile in Rural Areas sharply but was considerably higher among females. (by Sector), the vast majority of male household heads Men were significantly better able to find nonfarm (92.4 percent) and female household heads (79.6 percent) employment and/or to find a second job than women were employed in agriculture in 2005. By 2010, this over the period, as shown. Moreover, the gap in wages proportion had decreased only slightly: to 90.8 percent between males and females appears higher in 2010 than for males and to 76.4 percent for females, respectively. for other years, for at least some prime working ages Despite falling returns to agriculture and severe climate (approximately age 40), as shown in figure 2.24. shocks, it appears that the overwhelming majority of household heads in the poorest rural quintile did not shift Another possible contributing explanation is that the their main source of livelihood away from the primary returns to activities in which males are more likely to sector, arguably due to a lack of alternative opportunities. engage increased relative to those for females. For the average 40-year-old worker, the wage disparity in 2010 An examination of employment patterns for household between females and males was higher than in other heads in the bottom quintile of the distribution of rural years (figure 2.24). Bi and Osborne (2016) also find that households by gender can shed light on the increase in FIGURE 2.23: Labor Market Outcomes, 2005, 2010, and 2012 (Kernel-Weighted Local Polynomial Smoothed Age-Wage Profiles 0.10 Percentage looking for work by age lpoly smooth: (mean) looking_work Males in 2005 0.08 Males in 2010 Males in 2012 0.06 Females in 2005 Females in 2010 Females in 2012 0.04 0.02 0.00 0 10 20 30 40 Age Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010) 67 FIGURE 2.23: Labor Market Outcomes, 2005, 2010, and 2012 (Kernel-Weighted Local Polynomial Smoothed Age-Wage Profiles (continued) 0.20 Proportion of 0.18 self-employed lpoly smooth: (mean) self_nonagri in non-farm sector 0.16 Males in 2005 0.14 Males in 2010 0.12 Males in 2012 0.10 Females in 2005 Females in 2010 0.08 Females in 2012 0.06 0.04 0.02 0.00 0 10 20 30 40 50 60 70 80 Age 0.8 Proportion with second job lpoly smooth: (mean) second_job Males in 2005 0.6 Males in 2010 Males in 2012 Females in 2005 0.4 Females in 2010 Females in 2012 0.2 0.0 0 10 20 30 40 50 60 70 80 Age FIGURE 2.24: Log of Wage, Male and Female Workers by Age (2005, 2010, and 2012) 11.5 Log of wage rate 11.0 Males in 2005 lpoly smooth: (mean) sper_day_inc 10.5 10.0 Males in 2010 9.5 Males in 2012 9.0 8.5 Females in 2005 8.0 Females in 2010 7.5 Females in 2012 7.0 6.5 6.0 5.5 5.0 4.5 4.0 3.5 3.0 0 10 20 30 40 50 60 70 80 Age Sources EPM 2005, 2010, and ENSOMD 2012. Notes: x axis = age of worker. y axis represents the smoothed polynomial of log wage per day. 68 Republic of Madagascar Employment and Poverty Analysis FIGURE 2.25: Employment of Household Heads in the Bottom Quintile in Rural Areas (by Sector) a. Male primary employment, by sector (rural bottom quintile) Proportion of male household 100 92.4 90.8 2005 90 2010 80 70 heads (%) 60 50 40 30 20 10 2.8 2.0 1.8 2.9 1.0 2.8 1.3 1.3 0.7 0.3 0 Not Primary Industry Trade Public Other employed sector sector private sector b. Female primary employment, by sector (rural bottom quintile) Proportion of female household 100 2005 90 79.6 76.4 2010 80 70 heads (%) 60 50 40 30 20 10.8 7.2 10 1.1 5.0 4.0 3.4 4.3 7.2 0.1 0.8 0 Not Primary Industry Trade Public Other employed sector sector private sector c. Male secondary employment, by sector (rural bottom quintile) Proportion of male household 100 2005 90 2010 80 70 heads (%) 60 57.2 50 40 34.3 34.7 34.5 30 25.4 20 10 2.5 2.4 2.2 3.4 2.9 0.1 0.0 0 Not Primary Industry Trade Public Other employed sector sector private sector d. Female secondary employment, by sector (rural bottom quintile) Proportion of female household 100 2005 90 80 2010 70 64.7 heads (%) 60 50 40 40.3 30 20.8 24.5 22.8 20 10 6.2 6.0 6.2 5.6 1.2 0.8 0.0 0 Not Primary Industry Trade Public Other employed sector sector private sector Source: EPM 2005, 2010. Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010) 69 gender-based differences in consumption over the period. segment, however, only 17.2 percent were able to find A large proportion of these household heads (both male secondary employment in the “other private” sector, with and female) took on secondary activities between 2005 the other 7.2 percent finding secondary work in the pri- and 2010. Whereas only 42.8 percent of male household mary sector and industry. It is likely that males engaged heads in the bottom quintile in rural areas had second- in more remunerative activities than their female counter- ary employment in 2005, by 2010 this proportion had parts, such as transport and other nontrade services, and reached 65.7 percent (figure 2.25). Among their female this made them better able to cope with falling returns counterparts, a similar trend was observed: in 2005, to agriculture and climate shocks. Because between 2005 whereas only 35.3 percent of female household heads and 2010 most poor rural household heads remained had secondary employment, this rose to 59.7 percent in primarily employed in agriculture, a large share of both 2010. Thus, the increases were similar—at 22.9 percent male and female household heads sought and found for males and 24.4 percent for females. Many poor rural secondary employment. Because secondary activities in household heads took on secondary activities alongside the private sector accounted for most of the new employ- their primary employment in order to cope with falling ment for both genders, there are likely to be additional returns to agriculture (their main source of livelihood) explanations for the sharp increase in returns to male and the devastating effects of climate shocks. household heads observed over the period. The increase in returns to male household heads in the bottom quintile The main gender-based differences revealed are that for could be due to the fact that the opportunities available these males almost all new secondary employment was in mostly to males were more remunerative than those the “other private sector,” which absorbed 22.5 percent available mostly to females, on average. of them. For female household heads in this population 70 Republic of Madagascar Employment and Poverty Analysis Annex 2A. MAP 2A.1: Administrative Divisions of Madagascar a. Old division b. New divisions as of 2009 Source: Wikipedia. Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010) 71 Annex 2B. Methodology The first part of this analysis investigated sources of permit the analysis of the entire distribution. Neverthe- inequality between rural and urban households. The less, they all share the same shortcoming in that they second part explored the factors influencing the evolu- involve several assumptions and computational difficul- tion of inequality in consumption over time. In this ties (Fortin, Lemieux, and Firpo 2010). section, the methodology behind these decompositions is presented in greater detail. The RIF-regression method proposed by Firpo, Fortin, and Lemieux (2009) addresses these shortcomings and Inequality is measured using real monthly household per allows one to evaluate the impact of changes in the capita consumption expenditures adjusted for spatial distribution of the explanatory variables on quantiles and temporal variations in the cost of living. The con- of the unconditional (marginal) distribution of the sumption aggregate includes expenditures on both food outcome variable. To distinguish this approach from and nonfood items and excludes both rental housing and other conditional quantile regressions (Koenker and durable goods expenses. The aggregate is constructed Bassett 1978; Koenker 2005), this method is referred to following the methodology suggested by Deaton and as unconditional quantile regression. It allows one to Zaidi (2002). The adjustment for price variations across decompose the welfare gaps at various quantiles of the regions and over time is done using the Fisher index of unconditional distribution into differences in house- unit values from the surveys. holds’ endowment characteristics such as education, age, employment status, and so forth, and differences in the To analyze the sources of inequality between rural and marginal (conditional) correlations between consump- urban groups and over time, the unconditional quantile tion and these characteristics. These components are regression method is applied. This method allows to then further decomposed to identify the specific attri- understand how the differences in the distributions of butes which contribute to the widening welfare gap. observed household characteristics between groups or over time contribute to the welfare gap. It also identifies The procedure is carried out in two stages. The first stage how the marginal “effects” of these characteristics vary consists of estimating unconditional quantile regressions across the entire distribution. Overall, the method allows on log real per capita consumption for group 1 and to distinguish between the contributions of: (a) dif- group 2 households, then constructing a counterfactual ferences in household characteristics (“endowment” distribution that would prevail if group 1 households effects); and, (b) disparities in returns to these character- had received the returns that pertained to group 2. The istics (“returns” effect) to inequality. comparison of the counterfactual and empirical distribu- tions allows us to estimate the part of the welfare gap The development of decomposition methods has been attributable to households’ characteristic differentials, a fertile area of research over the last few decades. the endowment effect, and the part explained by differ- Building on the seminal work of Oaxaca (1973) and ences in returns to those characteristics, the return effect. Blinder (1973), several procedures that allow one to go The second stage involves dividing the endowment and beyond the mean have been put forward and have been return components into the contribution of each specific used widely in the literature. Popular approaches used characteristic variable. in the decomposition of distributional statistics and the analysis of the sources of inequality include the standard The method can be easily implemented as a standard Oaxaca–Blinder decomposition method, the reweight- linear regression, and an ordinary least squares (OLS) ing procedure of DiNardo, Fortin, and Lemieux (1996) regression of the following form can be estimated: and the quantile-based decomposition approach of RIF(y, Qq ) = Xb + ε, (1) Machado and Mata (2005). The main drawback of the Oaxaca–Blinder technique is that it applies the decompo- where y is log real per capita monthly household con- sition only to the mean welfare differences between two sumption and RIF(y,Qq) is the RIF of the qth quantile population subgroups and yields an incomplete represen- of y estimated by computing the sample quantile Qq and tation of the inequality sources. The other conventional estimating the density of y at that point by the kernel methods extend the decomposition beyond the mean and method: 72 Republic of Madagascar Employment and Poverty Analysis ( { RIF(y, Qq ) = Qq + q − I y ≤ Qq }) f y (Qq ), distributions of household characteristics to inequality at the qth unconditional quantile, denoted endowment where fy is the marginal density function of y and I is an effect. The second term of the right-hand side of the indicator function. RIF can be estimated by replacing equation represents the inequality due to differences (or Qq by qth sample quantile and estimating fy by kernel discrimination) in returns to the household characteris- density. X is the regressors matrix including the inter- tics at the qth unconditional quantile. The endowment cept, b is the regression coefficient vector, and ε is the and return effects can be further decomposed into the error term. contribution of individual specific household characteris- tics (or group of some characteristics) as follows: We estimate the model for each decile from the 10th to 90th quantiles and use the unconditional quantile regres- ˆi −Q Qq ˆ* = q ∑ (X k i k ) i′ ˆ i − Xk b q ,k ˆ* − Q and Qq ˆ i′ q sion estimates to decompose the rural-urban inequality, (3) as well as the changes in consumption between 2005 and = ∑X i′ k (b ˆ i q ,k ˆ i ′ k:1 K , −b q ,k) 2010 into a component attributable to differences in the  k distribution of characteristics and a component due to where k designates the individual specific household differences in the distribution of returns. This is done as characteristics. follows: Q ˆ i is the counterfactual quantile of the uncon- ˆ * = X i ′b { } { } ( q ˆi −Q Qq ˆ i′ = Q q ˆi −Q q ˆ* + Q q ˆ* − Q ˆ i′ = X i − X i′ b q ˆi q q ) ditional counterfactual distribution which represents the (2) distribution of welfare that would have prevailed for (ˆi − b +X i′ b q ˆ i′ , q ) group i’ (rural/2005 households) if they have received group i (urban/2010 households) returns to their char- where Q ˆ is the qth unconditional quantile of log real acteristics. The decomposition results may vary with the q per capita monthly household consumption, X rep- choice of the counterfactual distribution. For example, resents the vector of covariate averages and b ˆ the if the counterfactual used is the distribution that would q estimate of the unconditional quantile partial effect. have prevailed for group i if they have received group i’ Superscripts i, i’, and * designate, respectively, the returns we would obtain different results. The choice of urban (or 2010), rural (or 2005), and counterfactual the counterfactual in this analysis is motivated by the values. The first term on the right-hand side of equa- aim of emphasizing household groups living in disadvan- tion (2) represents the contribution of the differences in taged areas. Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010) 73 Annex 2C. Determinants of Urban-Rural Inequality in 2010 TABLE 2C.1: Summary Statistics for Urban and Rural Households (2010) Marital status Household size Age structure Gender of head of head (Average number (Average members (Households with Age of head (Household heads of members) under 14) male head) (Average years) with spouse) Quintile Urban Rural Urban Rural Urban Rural Urban Rural Urban Rural Bottom 5.6 6.2 46% 54% 72% 78% 42 41 72% 76% Second 4.7 5.6 37% 49% 78% 82% 43 41 75% 80% Middle 4.0 4.9 33% 44% 79% 82% 41 42 72% 79% Fourth 3.6 4.3 25% 38% 76% 84% 41 42 69% 79% Top 2.9 3.4 14% 24% 75% 80% 43 43 59% 70% Education level Distance to market Security level (Avg highest level Health shocks Climate shocks (Households 1+ (Avg security: completed by head/ (Households that had (Households that had hours away from 1 = very poor, spouse, 1–4) 1+ health shocks) 1+ climate shocks) food market) 4 = very good) Quintile Urban Rural Urban Rural Urban Rural Urban Rural Urban Rural Bottom 1.9 1.4 9% 7% 23% 55% 18% 62% 3.1 3.1 Second 2.4 1.6 9% 4% 13% 41% 9% 53% 2.8 3.0 Middle 2.5 1.7 8% 4% 7% 37% 8% 51% 2.7 3.0 Fourth 2.9 1.8 6% 5% 7% 35% 7% 47% 2.7 3.0 Top 3.1 2.2 4% 6% 5% 27% 4% 33% 2.7 2.9 Source: EPM 2010. TABLE 2C.2: Decomposition of Log Consumption Expenditure, Urban and Rural Households (2010) Percentiles 20 40 60 80 Overall (Log) Urban consumption 12.112 12.527 12.867 13.257 (0.022)*** (0.020)*** (0.019)*** (0.024)*** (Log) Rural consumption 11.298 11.646 11.936 12.311 (0.011)*** (0.010)*** (0.010)*** (0.012)*** Difference 0.813 0.881 0.930 0.946 (0.025)*** (0.022)*** (0.021)*** (0.027)*** Endowment component 0.632 0.539 0.498 0.462 (0.037)*** (0.027)*** (0.025)*** (0.034)*** Returns component 0.181 0.342 0.432 0.484 (0.047)*** (0.032)*** (0.027)*** (0.031)*** N 11,820 11,820 11,820 11,820 Source: Calculated using EPM 2010. Note: Robust standard errors in parentheses. ***Significance at the 1% level. **Significance at the 5% level. *Significance at the 10% level. Controls include province dummies. 74 Republic of Madagascar Employment and Poverty Analysis TABLE 2C.3: Decomposition of Log Consumption Expenditure, Urban and Rural Households (2010, Endowments) Percentiles 20 40 60 80 Endowments component Household size 0.109 0.096 0.058 0.052 (0.013)*** (0.010)*** (0.008)*** (0.009)*** Percentage of children under 14 0.060 0.066 0.098 0.106 (0.013)*** (0.011)*** (0.010)*** (0.013)*** Male household head –0.005 –0.002 0.001 –0.006 (0.004) (0.003) (0.003) (0.005) Age of household head 0.001 0.001 0.000 0.001 (0.002) (0.001) (0.000) (0.001) Household head has a spouse –0.007 –0.002 0.002 0.014 (0.005) (0.005) (0.005) (0.007)* Education level of head/spouse 0.256 0.237 0.236 0.220 (0.019)*** (0.017)*** (0.016)*** (0.022)*** At least one health shock –0.001 –0.002 –0.003 –0.003 (0.002) (0.002) (0.002)* (0.002)** At least one climate shock 0.071 0.057 0.020 0.009 (0.022)*** (0.016)*** (0.013) (0.014) Time to food market is one hour 0.105 0.050 0.045 0.042 or more (0.031)*** (0.023)** (0.022)** (0.025)* Security level 0.018 0.012 0.004 –0.001 (0.005)*** (0.004)*** (0.004) (0.006) N 11,820 11,820 11,820 11,820 Source: Calculated using EPM 2010. Note: Robust standard errors in parentheses. ***Significance at the 1% level. **Significance at the 5% level. *Significance at the 10% level. Controls include province dummies. Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010) 75 TABLE 2C.4: Decomposition of Log Consumption Expenditure, Urban and Rural Households (2010, Returns) Percentiles 20 40 60 80 Returns component Household size –0.274 –0.185 0.004 –0.026 (0.084)*** (0.062)*** (0.051) (0.061) Percentage of children under 14 –0.081 –0.036 –0.088 0.035 (0.053) (0.044) (0.041)** (0.054) Male household head –0.054 –0.089 –0.130 –0.036 (0.065) (0.057) (0.058)** (0.088) Age of household head 0.072 –0.007 –0.052 0.087 (0.075) (0.065) (0.063) (0.088) Household head has a spouse 0.028 0.019 0.023 0.004 (0.057) (0.052) (0.052) (0.081) Education level of head/spouse 0.252 0.168 0.109 –0.061 (0.042)*** (0.036)*** (0.035)*** (0.052) At least one health shock –0.000 –0.008 –0.014 –0.018 (0.005) (0.004)* (0.004)*** (0.005)*** At least one climate shock –0.070 –0.072 –0.032 –0.008 (0.032)** (0.023)*** (0.019) (0.022) Time to food market is one hour –0.066 –0.002 0.013 0.042 or more (0.040) (0.030) (0.028) (0.033) Security level –0.155 –0.126 –0.023 0.045 (0.066)** (0.060)** (0.060) (0.082) Constant 0.003 0.002 0.016 0.017 (0.015) (0.011) (0.010) (0.012) N 11,820 11,820 11,820 11,820 Source: Calculated using EPM 2010. Note: Robust standard errors in parentheses. ***Significance at the 1% level. **Significance at the 5% level. *Significance at the 10% level. Controls include province dummies. Source: 76 Republic of Madagascar Employment and Poverty Analysis Annex 2D. Determinants of Inequality between 2005 and 2010 TABLE 2D.1: Summary Statistics for All Households (2005 and 2010) Location Household size Age structure Gender of head (Households in rural (Average number (Average members (Households with Age of head area) of members) under 14) male head) (Average years) Quintile 2005 2010 2005 2010 2005 2010 2005 2010 2005 2010 Bottom 95% 98% 6.1 6.1 52% 53% 84% 78% 42 42 Second 92% 94% 5.3 5.4 46% 47% 80% 81% 42 41 Middle 88% 89% 4.9 4.8 42% 42% 80% 82% 43 42 Fourth 79% 76% 4.1 4.1 34% 34% 79% 82% 43 42 Top 56% 47% 3.5 3.3 25% 22% 79% 78% 43 42 Marital status of Education level Electricity head (Highest level Climate shocks (Avg households (Households heads completed by (Households that had with electricity in with spouse) head/spouse, 1–4) 1+ climate shocks) community) Quintile 2005 2010 2005 2010 2005 2010 2005 2010 Bottom 80% 77% 1.5 1.4 59% 52% 4% 2% Second 78% 79% 1.6 1.6 57% 39% 6% 5% Middle 77% 79% 1.7 1.8 57% 35% 9% 9% Fourth 72% 76% 1.9 2.0 48% 27% 18% 21% Top 71% 67% 2.4 2.6 35% 14% 40% 48% Source: EPM 2010. TABLE 2D.2: Decomposition of Log Consumption Expenditure (between 2010 and 2005) Percentiles 20 40 60 80 Overall (Log) 2010 consumption 11.390 11.764 12.100 12.587 (0.010)*** (0.009)*** (0.010)*** (0.014)*** (Log) 2005 consumption 11.422 11.752 12.072 12.491 (0.010)*** (0.009)*** (0.009)*** (0.013)*** Difference –0.032 0.011 0.029 0.096 (0.014)** (0.012) (0.014)** (0.019)*** Endowment component 0.039 0.028 0.034 0.047 (0.007)*** (0.007)*** (0.009)*** (0.012)*** Returns component –0.071 –0.017 –0.005 0.049 (0.014)*** (0.012) (0.013) (0.018)*** N 24,157 24,157 24,157 24,157 Source: Calculated using EPM 2010. Note: Robust standard errors in parentheses. ***Significance at the 1% level. **Significance at the 5% level. *Significance at the 10% level. Controls include province dummies. Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010) 77 TABLE 2D.3: Decomposition of Log Consumption Expenditure between 2010 and 2005, Endowments Percentiles 20 40 60 80 Endowments component Located in rural area 0.002 0.002 0.002 0.003 (0.001)** (0.001)** (0.001)** (0.001)** Household size –0.000 –0.000 –0.000 –0.000 (0.003) (0.003) (0.003) (0.002) Percentage of children under 14 –0.001 –0.002 –0.002 –0.003 (0.001) (0.001) (0.002) (0.003) Male household head –0.000 –0.000 –0.000 –0.000 (0.000) (0.000) (0.001) (0.001) Age of household head –0.001 –0.001 –0.001 –0.000 (0.001)** (0.001)** (0.000)* (0.000) Household head has a spouse –0.000 –0.000 0.000 0.000 (0.000) (0.000) (0.000) (0.001) Education level of head/spouse 0.006 0.007 0.009 0.014 (0.001)*** (0.002)*** (0.002)*** (0.003)*** At least one climate shock 0.013 0.004 0.004 –0.002 (0.004)*** (0.003) (0.004) (0.004) At least one health shock 0.002 –0.003 –0.007 –0.003 (0.003) (0.002) (0.002)*** (0.004) Security level 0.007 0.005 0.005 0.006 (0.002)*** (0.002)*** (0.002)** (0.003)** Access to electricity 0.003 0.007 0.015 0.026 in community (0.001)*** (0.002)*** (0.003)*** (0.005)*** Means of transport 0.005 0.005 0.006 0.006 (0.001)*** (0.001)*** (0.001)*** (0.001)*** N 24,157 24,157 24,157 24,157 Source: Calculated using EPM 2005, 2010. Note: Robust standard errors in parentheses. ***Significance at the 1% level. **Significance at the 5% level. *Significance at the 10% level. Controls include province dummies. 78 Republic of Madagascar Employment and Poverty Analysis TABLE 2D.4: Decomposition of Log Consumption Expenditure (between 2010 and 2005, Returns) Percentiles 20 40 60 80 Returns component Located in rural area –0.059 –0.097 –0.169 –0.237 (0.027)** (0.027)*** (0.036)*** (0.069)*** Household size 0.014 –0.020 0.016 0.008 (0.039) (0.030) (0.031) (0.040) Percentage of children under 14 –0.018 –0.022 –0.055 –0.097 (0.026) (0.022) (0.026)** (0.036)*** Male household head 0.129 0.082 0.071 0.133 (0.038)*** (0.032)** (0.038)* (0.059)** Age of household head 0.031 0.016 0.017 0.008 (0.042) (0.036) (0.041) (0.057) Household head has a spouse –0.023 –0.010 –0.026 –0.152 (0.033) (0.029) (0.034) (0.053)*** Education level of head/spouse 0.019 0.020 0.044 0.098 (0.027) (0.024) (0.028) (0.044)** At least one climate shock –0.054 –0.023 –0.005 0.011 (0.016)*** (0.014)* (0.015) (0.018) At least one health shock –0.017 –0.012 –0.000 –0.010 (0.006)*** (0.005)*** (0.005) (0.008) Security level –0.022 –0.018 –0.004 –0.045 (0.039) (0.033) (0.036) (0.049) Access to electricity –0.008 –0.001 0.006 0.018 in community (0.007) (0.007) (0.010) (0.018) Means of transport –0.002 –0.003 –0.012 –0.017 (0.005) (0.005) (0.006)* (0.010)* Constant –0.029 0.057 0.102 0.322 (0.089) (0.078) (0.092) (0.143)** N 24,157 24,157 24,157 24,157 Source: Calculated using EPM 2005, 2010. Note: Robust standard errors in parentheses. ***Significance at the 1% level. **Significance at the 5% level. *Significance at the 10% level. Controls include province dummies. Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010) 79 Annex 2E. Determinants of Rural Inequality between 2005 and 2010 TABLE 2E.1: Summary Statistics for Rural Households (2005 and 2010) Marital status Household size Age structure Gender of head of head (Avg number of (Avg members (Households with Age of head (Households heads members) under 14) male head) (Average years) with spouse) Quintile 2005 2010 2005 2010 2005 2010 2005 2010 2005 2010 Bottom 6.2 6.2 52% 54% 84% 84% 42 41 80% 76% Second 5.4 5.6 47% 49% 82% 82% 42 41 78% 80% Middle 4.9 4.9 42% 44% 82% 82% 42 42 80% 79% Fourth 4.2 4.3 36% 38% 80% 80% 43 42 74% 79% Top 3.5 3.4 27% 24% 79% 79% 43 43 69% 70% Education level Electricity (Avg highest level Climate shocks (Avg households Transportation Land completed by head/ (Households that had with electricity in (Households with 1+ (Average acres of spouse, 1–4) 1+ climate shocks) community) means of transport) exploited land) Quintile 2005 2010 2005 2010 2005 2010 2005 2010 2005 2010 Bottom 1.5 1.4 60% 55% 0.9% 1.1% 9% 7% 119 93 Second 1.6 1.6 61% 41% 1.2% 2.0% 14% 14% 127 112 Middle 1.6 1.7 60% 37% 2.0% 3.8% 16% 20% 138 121 Fourth 1.7 1.8 57% 35% 3.8% 7.0% 22% 25% 138 129 Top 2.0 2.2 51% 27% 12.2% 17.5% 33% 35% 132 149 Source: EPM 2005, 2010. TABLE 2E.2: Decomposition of Log Consumption Expenditure, Rural Households (2005 and 2010) Percentiles 20 40 60 80 Overall (Log) 2010 rural consumption 11.303 11.652 11.945 12.323 (0.011)*** (0.009)*** (0.009)*** (0.012)*** (Log) 2005 rural consumption 11.365 11.677 11.964 12.319 (0.010)*** (0.009)*** (0.010)*** (0.012)*** Difference –0.062 –0.025 –0.019 0.004 (0.015)*** (0.013)** (0.013) (0.017) Endowment component 0.028 0.013 0.005 0.013 (0.008)*** (0.008)* (0.008) (0.010) Returns component –0.090 –0.038 –0.023 –0.009 (0.015)*** (0.013)*** (0.013)* (0.017) N 17,755 17,755 17,755 17,755 Source: Calculated using EPM 2005, 2010. Note: Robust standard errors in parentheses. ***Significance at the 1% level. **Significance at the 5% level. *Significance at the 10% level. Controls include province dummies. 80 Republic of Madagascar Employment and Poverty Analysis TABLE 2E.3: Decomposition of Log Consumption Expenditure, Rural Households (2005 and 2010, Endowments) Percentiles 20 40 60 80 Endowments component Household size –0.006 –0.006 –0.006 –0.005 (0.003)* (0.004)* (0.003)* (0.003)* Percentage of children under 14 –0.003 –0.004 –0.006 –0.009 (0.001)** (0.002)*** (0.002)*** (0.003)*** Male household head 0.000 0.000 0.000 0.000 (0.001) (0.001) (0.001) (0.001) Age of household head –0.001 –0.001 –0.001 –0.000 (0.001) (0.001) (0.001) (0.000) Household head has a spouse 0.000 –0.000 –0.000 –0.001 (0.000) (0.000) (0.000) (0.001) Education level of head/spouse 0.004 0.005 0.005 0.007 (0.002)*** (0.002)*** (0.002)*** (0.003)*** At least one climate shock 0.018 0.005 –0.001 –0.000 (0.004)*** (0.004) (0.004) (0.004) At least one health shock 0.004 –0.001 –0.007 –0.011 (0.003) (0.003) (0.003)*** (0.004)*** Security level 0.006 0.002 0.002 0.004 (0.002)*** (0.002) (0.002) (0.002)* Access to electricity 0.004 0.011 0.016 0.029 in community (0.001)*** (0.002)*** (0.002)*** (0.004)*** Means of transportation 0.003 0.004 0.004 0.005 (0.001)*** (0.001)*** (0.001)*** (0.002)*** Land –0.005 –0.005 –0.005 –0.006 (0.001)*** (0.001)*** (0.001)*** (0.002)*** N 17,755 17,755 17,755 17,755 Source: Calculated using EPM 2005, 2010. Note: Robust standard errors in parentheses. ***Significance at the 1% level. **Significance at the 5% level. *Significance at the 10% level. Controls include province dummies. Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010) 81 TABLE 2E.4: Decomposition of Log Consumption Expenditure, Rural Households (2005 and 2010, Returns) Percentiles 20 40 60 80 Returns component Household size –0.000 –0.049 0.039 0.030 (0.043) (0.032) (0.031) (0.037) Percentage of children under 14 –0.010 –0.022 –0.038 –0.139 (0.030) (0.025) (0.027) (0.035)*** Male household head 0.166 0.143 0.091 0.069 (0.043)*** (0.036)*** (0.040)** (0.057) Age of household head 0.054 0.034 0.000 –0.036 (0.045) (0.038) (0.041) (0.053) Household head has a spouse –0.017 –0.042 –0.004 –0.052 (0.038) (0.032) (0.035) (0.051) Education level of head/spouse 0.037 0.040 0.038 0.071 (0.029) (0.025) (0.028) (0.038)* At least one climate shock –0.073 –0.023 0.009 0.003 (0.018)*** (0.015) (0.016) (0.019) At least one health shock –0.020 –0.014 0.002 0.009 (0.006)*** (0.005)** (0.005) (0.008) Security level –0.057 –0.029 –0.006 –0.085 (0.042) (0.034) (0.036) (0.045)* Access to electricity –0.004 0.001 –0.002 –0.000 in community (0.002)* (0.002) (0.003) (0.005) Means of transportation –0.001 0.007 –0.013 –0.012 (0.006) (0.006) (0.006)* (0.009) Land –0.066 –0.063 –0.065 –0.039 (0.016)*** (0.013)*** (0.014)*** (0.018)** Constant –0.086 –0.077 –0.110 0.076 (0.088) (0.076) (0.081) (0.105) N 17,755 17,755 17,755 17,755 Source: Calculated using EPM 2005, 2010. Note: Robust standard errors in parentheses. ***Significance at the 1% level. **Significance at the 5% level. *Significance at the 10% level. Controls include province dummies. 82 Republic of Madagascar Employment and Poverty Analysis NOTES Blinder, Alan S. 1973. “Wage Discrimination: Reduced Form and Structural Estimates.” Journal of Human 1. Madagascar uses per capita consumption as its welfare indicator for measuring poverty (as is consistent with the World Bank’s official Resources 8 (4): 436–55. international poverty measure). Because of economies of scale and Dang, H., P. Lanjouw, J. Luoto, and D. McKenzie. differential consumption needs within the household, inequalities due to differences in household size and percentage of children are likely 2014. “Using Repeated Cross-Sections to Explore to be overstated. Movements into and out of Poverty.” Journal of 2. A full explanation of all causes of a distribution or its changes would also require an examination of broader policy, institutional, and Development Economics. 107: 112–28. contextual factors, which are not observable at this level of data. David-Benz, Hélène. 2011. “A Madagascar: les prix 3. All percentages indicate counterfactual changes in relative consump- du riz flambent, sans rapport avec le marché tion levels between the two groups, meaning that all other factors are held equal to their values at baseline (rural group or year 2005). international.” Paris: Cirad, UMR Moisa. 4. Madagascar uses per capita consumption as its welfare indicator for David-Benz, Hélène, and Frederic Lancon. 2013. measuring poverty (as is consistent with the World Bank’s official international poverty measure). Because of economies of scale and “Transmission des prix internationaux du riz sur differential consumption needs within the household, inequalities due les marches africains: le long terme, la crise de to differences in household size and percentage of children are likely to be overstated. 2008 . . . et maintenant?” Presented at Third Africa 5. Access to electricity declined in the bottom two quintiles, so the Rice Congress, Cirad, Paris. positive contribution to the endowments effects is only for the top three quintiles. Deaton and Zaidi. 2002. “Guidelines for Constructing 6. Endowments in terms of household size, percentage of children, gen- Consumption Aggregates for Welfare Analysis,” der of the household head, age of the household head, marital status, unpublished. and health shocks do not explain significant changes in consumption expenditure between 2005 and 2010. This is not surprising given DiNardo, Fortin, and Lemieux. 1996. “Unconditional that levels of these endowments have changed very little or not at all Quantile Regressions,” Econometrica 77 (3) : 953–973. over the period. 7. The data does not allow one to discern whether households EPM (Enquête Périodique auprès des Ménages). 2005, remained in the same quintile in 2005 and 2010 or not. However, 2010. INSTAT, Government of Madagascar. the results on climate shocks indicate that either (1) households that were not in the bottom quintile in 2005 had fallen into it by 2010 FAO (Food and Agriculture Organization of the United partly as a result of the severity of climate shocks they experienced, Nations). 2010. Mission d’évaluation de la securité and/or (2) households that were already in the bottom quintile in 2005 saw their consumption decrease further by 2010 partly for the alimentaire a Madagascar. Rome: FAO. same reason. Firpo, S., N. Fortin, and T. Lemieux. 2009. 8. Unfortunately, comparable data on prices at the community level “Unconditional Quantile Regressions.” Econometrica were not collected in 2001 and 2012. 9. A study by Minten (1999) demonstrates the importance of distance 77 (3): 953–73. to a road and “soft” infrastructure to improve competition among Fortin, N., T. Lemieux, and S. Firpo. 2010. traders over the quality of a road for improving market integration within Madagascar, but the study is dated. Decomposition Methods in Economics. Vancouver: 10. Unfortunately, there was no community survey in 2012 to assess University of British Columbia. http://econ.sites.olt whether conditions had changed. .ubc.ca/files/2013/05/pdf_paper_thomas-lemieux- decomposition-methods-economics.pdf. INSTAT (National Institute of Statistics, Madagascar). REFERENCES 2014. Enquête national sur le suivi des objectifs du millénaire pour le développement (ENSOMD) à Auffret, Philippe. 2014. “Madagascar: Three Years into Madagascar. Antananarivo. Crisis.” Social Protection and Labor Discussion Ivanic, M. and W. Martin. 2014. http://documents Paper 1416, World Bank, Washington, DC. .worldbank.org/curated/en/106581468325435880/ Belghith, N., T. Osborne, and P. Randriankolona. 2016. Short-and-long-run-impacts-of-food-price-changes- “Madagascar Poverty and Inequality Update: Recent on-poverty. Trends in Welfare, Employment, and Vulnerability.” Koenker. 2005. Quantile Regression. Cambridge Washington, DC: World Bank. University Press, Cambridge, UK. Bi, C., and T. Osborne. 2016. “Transactions Costs, Koenker and Bassett. 1978. “Regression Quantiles,” Poverty, and Low Productivity Traps: Evidence from Econometrica. 46(1): 33–50. Microenterprises in Madagascar.” Washington, DC: Machado and Mata. 2005. “Counterfactual World Bank. decomposition of changes in wage distributions Isolation, Crisis, and Vulnerability: A Decomposition Analysis of Inequality and Deepening Poverty in Madagascar (2005–2010) 83 using quantile regression,” Applied Econometrics Madagascar. Washington, DC: USAID, Cornell 20(4): 445–465. University, INSTAT, and FOFIFA (Malagasy for Moser, C., C. Barrett, and B. Minten. 2009. “Spatial Centre Nationale pour la Recherche Appliquée Integration at Multiple Scales: Rice Markets in au Développent Rurale, edited by B. Minten, J.-C. Madagascar.” Agricultural Economics 40: 281–94. Randrianarisoa, and L. Randrianarison. Ithaca, NY: Oaxaca, Ronald L. 1973. “Male-Female Wage Cornell Food and Nutrition Policy Program. Differentials in Urban Labor Markets.” International World Bank. 2014. Face of Poverty in Madagascar: Economic Review 14 (3): 693–709. Poverty, Gender, and Inequality Assessment. Randrianarisoa, J.-C. 2003. “Analyse spatiale de la Washington, DC: World Bank, PREM Africa. production rizicole malgache” in Agriculture, World Bank. 2015. Madagascar: Systematic Country pauvreté rurale et politiques économiques à Diagnostic. Washington DC: World Bank. 84  85 CHAPTER 3 Flexible Poverty Profiling and Welfare Prediction in Madagascar Linden McBride* Theresa Osborne June 2016 *Cornell University This paper was funded by the World Bank and carried out under the direction and supervision of the Africa Poverty Global Practice. Christopher Barrett (Cornell University) helped supervise the work. The authors are grateful to Kathleen Beegle and Dominique van de Walle for peer review comments. Introduction and Key Findings I n Madagascar, where 70.7 percent of the population the analysis provides only correlations and is thus an is below the national poverty line, an understanding insufficient basis for predicting the impacts of policies; of the characteristics and conditions that best predict however, the ranking of the most predictive variables differing levels of deprivation cannot only provide is nonetheless suggestive of both further analysis and a profile of the poorest groups in the country but it policy priorities. can also be one fundamental step in the targeting of interventions and the design of effective policy responses. This analysis utilizes the nationally representative 2010 This paper applies regression tree and regression forest Madagascar Enquête Périodique auprès des Ménages analysis—flexible, data-driven methods for the construc- (EPM), the latest available household consumption tion of prediction models—to identify the observable survey for which community-level variables are avail- household and community-level characteristics that are able. The application of regression tree (RT) and random the most powerful predictors of households’ consump- forest (RF) analyses to these data allows us to identify tion levels.1 These “machine learning” approaches several clusters of variables that are highly predictive take a different approach to conventional poverty of household per capita consumption and use these to profiling, which describes data according to arbitrarily group households by similar characteristics and con- chosen differentiating factors or inflexible (parametric) sumption levels. The two methods give similar but not econometric estimates. By using out-of-sample predictive identical results. Combining insights of each, of the performance as the criterion for choosing best estimates, many available household, regional, and community- the approach used here develops conditional, data-driven level variables which one might expect to be correlated profiles of who is poor and how poor they are. This with poverty, those that are most predictive of increasing results in a nested set of the most important factors that severity of poverty are the following, in order of impor- predict welfare, as represented by per capita consump- tance: (1) living in a community with levels of electrifica- tion expenditures, thus providing a more nuanced guide tion at less than 27 percent of households; (2) having a for targeting of programs.2 Like other poverty profiles, non-university-educated head of household; (3) having 86 Republic of Madagascar Employment and Poverty Analysis an illiterate head of household; (4) living in greater assumptions. In addition, they offer higher out-of- remoteness from the nearest major urban center, a vari- sample predictive accuracy than traditional methods able which predicts welfare better than other measures (Breiman 2001). This feature is important because the of urban attributes or access to services; (5) achieving population in which we are interested is rarely the lower prices for paddy rice, and/or other indicators of precise one observed in the available data. In the case of performance of rice markets; and (6) having lower live- Madagascar, the most recent available data (with a com- stock holdings. plete set of variables) are a sample from 2010, whereas the sample of interest is the population of all households These results indicate that having a university educa- in 2016 or later. Under the assumption that the same tion makes it very likely to have higher incomes in data-generating process underlies both the sample we urban areas, as would be expected, but apart from this have available and the population in which we may be distinction and illiteracy, differing levels of educational interested, we seek a method that is most accurate for attainment are not important predictors of welfare. For that population and not just accurate in the sample we agricultural households, analyzed separately, the key pre- happen to have available. Therefore, the “out-of-sam- dictive variables in order of importance are the follow- ple” predictive ability of this approach is both temporal ing: (1) less cultivated land, (2) greater remoteness from and spatial. Because we want to use past survey data the nearest major urban center, (3) having low or modest to try to profile the poor in a larger population and in levels of electrification in the community, (4) getting subsequent periods, such as for the purposes of target- higher percentage of one’s revenues from agriculture, and ing interventions, out-of-sample predictive performance (5) having a lower price of paddy rice. matters a great deal.3 While the importance of certain variables identified The RT version of CART operates by recursively parti- may not be surprising, using these methods one is able tioning the data into consumption groups by variables, to sort through a long list of variables that might have their threshold values, and consumption thresholds that otherwise been expected to play as meaningful a role provide the best prediction, and then into subgroups, in predicting levels of poverty (or consumption). For which successively reduce error in predicting consump- example, variables such as the household’s being female tion levels. It essentially runs a horse race among all of headed, living in the capital, and the various regional the potentially predictive variables to rank their explana- variables do not play an important role in explaining the tory and predictive power, while splitting the population variation in consumption in the 2010 EPM. Thus, the into more-poor or less-poor groups by characteristics specific ranking of the most predictive variables would and context, allowing the data to tell the analyst which be difficult to ascertain a priori. variables are more or less important in profiling and targeting the poor. Data and Methods The random forest (RF) version of the algorithm pro- vides several innovations over and above the RT: instead The methods used for the analysis include a recursive of building just one regression tree, it builds hundreds partitioning algorithm known as classification and (a forest), over randomly selected subsets of the data, regression tree (CART), as well as an extension of this and then averages across all of these trees in what is algorithm, called random forests (RF). (See annex 3A called “bootstrap aggregation” (or “bagging”) for a final for details.) These techniques permit a fully flexible, prediction (Breiman 2001).4 With this innovation, which data-driven approach to determining the profiling allows for averaging across low bias, de-correlated trees, parameters more useful for sorting households into RFs produce low-bias, low-variance predictions that increasingly homogeneous groups in terms of their are highly accurate out of sample (Breiman 2001). In characteristics, circumstances, and consumption levels. addition to identifying key variables that allow for more These methods offer several advantages over traditional nuanced profiling of the poor than is feasible using tra- approaches to poverty profiling. In particular, where ditional descriptive and econometric methods, this tech- the true underlying model is unknown, these methods nique allows one to rank their importance and observe are more flexible and can be more informative than how their partial correlations vary and co-vary with the parametric models that require several possibly faulty outcome variable of interest. How important a role a Flexible Poverty Profiling and Welfare Prediction in Madagascar  87 given variable plays in predicting consumption levels is to the nearest urban center and the cost of transporting assessed by the extent to which its exclusion from the 50 kilograms of rice to the nearest urban center during predictive model increases the out-of-sample (squared) the wet and dry seasons. Summary statistics for all of the prediction error.5 In addition, the relationship between above variables are provided in annex table 3A.1. a single variable and consumption can be plotted by incrementally changing the variable over its range (while holding all other variables at their means) to see how Results the predicted response changes (Hastie, Tibshirani, and Friedman 2009). Such plots showing large jumps in the CART RESULTS predicted response due to an incremental change in the The distribution of log per capita consumption (in 2001 variable of interest could suggest areas where thresholds deflated ariary per person) is presented in figure 3.1, or other nonlinearities may lie. This can reveal consump- where we can see that the majority of households are tion patterns that bifurcate around values for particular consuming below the poverty line (the vertical red household or community characteristics. While RFs offer line), the mean log per capita consumption is 12.03, a more robust and lower variance prediction than RTs, the median is 11.97, and the poverty line is 12.17. their results are somewhat more difficult to interpret Figures 3.2 through 3.4 present the results of the RT visually. analyses. Figure 3.2 includes the full sample; figure 3.3 includes the full sample but excludes from the search A comprehensive set of observable household character- algorithm the demographic variables of household size istics, assets and community level data are included in and the dependency ratio for reasons that are detailed the analyses that follow. First are household characteris- below; figure 3.4 includes only agricultural households tics, circumstances, and assets, including (1) the house- and excludes demographic variables. hold size and dependency ratio; (2) the age, education level, sex, employment status, and marital status of the In the RT presented in figure 3.2, the oblong circle at household head; (3) the number of primary-school- each node contains the conditional mean logged per cap- educated individuals in the household; (4) the ownership ita consumption and the percentage of the total sample of productive agricultural assets such as plows, carts, that meets the condition(s) displayed at this and any harrows, and manual agricultural equipment; (5) tropical preceding nodes. In the case of the very top of the tree, livestock units (TLUs) owned;6 (6) the amount of land the conditional mean is the sample mean, a logged per cultivated in ares by the household,7 including ownership capita consumption of 12.0; 100 percent of the sample is of a nonagricultural enterprise; (7) the percent of house- found there. Displayed on each branch descending from hold revenue from various sources (fishing, nonagricul- the first node is the first most predictive logical condition tural enterprise, agriculture, and livestock); (8) whether the household is an agricultural household; (9) whether FIGURE 3.1: Log per Capita Consumption the household is a net consumer or net producer of Distribution paddy rice and of dehulled rice;8 (10) whether the household lives in a rural or urban area; and (11) what Kernel density estimate climate, economic, health, security, and other shocks the 0.6 household has been exposed to in the past year. Also included are a variety of community-level variables, such as (a) the level of local security; (b) the mean and 0.4 Density standard deviation of the price and availability of white, imported, and paddy rice and inorganic fertilizer inputs (urea and nitrogen, phosphorus, potassium fertilizer 0.2 (NPK) over the seasons in the local community;9 and (c) regional dummy variables. Finally, to capture the 0 remoteness of each community, several variables are 5 10 15 20 included: distance in hours to the nearest market, health center, school, public transportation, and location to Log per capita consumption and poverty line kernel = epanechnikov, bandwidth = 0.0971 purchase agricultural inputs; and distance in kilometers 88 Republic of Madagascar Employment and Poverty Analysis FIGURE 3.2: Regression Tree of Log per Capita Household Consumption on Household and Community Level Variables, 2010 EPM (n = 12,460) In(pcexp)=12 100% Percent of households with electricity <0.275 >=0.275 In(pcexp)=11.8 In(pcexp)=12.7 76.7% 23.3% Household size >=4.5 Dependency ratio<=0.118 <4.5 <0.118 In(pcexp)=11.6 In(pcexp)=12 In(pcexp)=12.6 In(pcexp)=13.2 39.5% 37.2% 17.6% 5.7% Household head literate=0 Dependency ratio>=0.125 Household head has university degree=0 =1 <0.125 =1 In(pcexp)=11.3 In(pcexp)=11.7 In(pcexp)=11.9 In(pcexp)=12.4 In(pcexp)=12.4 In(pcexp)=13.1 13.1% 26.4% 28.2% 9.0% 14.5% 3.1% FIGURE 3.3: Regression Tree of Log per Capita Household Consumption on Household and Community Level Variables (Demographic Variables Excluded), 2010 EPM (n = 12,460) In(pcexp)=12 100% Percent of households with electricity <0.275 >=0.275 In(pcexp)=11.8 In(pcexp)=12.7 76.7% 23.3% Household head has university degree=0 Household head literate=0 =1 =1 In(pcexp)=11.6 In(pcexp)=11.9 In(pcexp)=12.6 In(pcexp)=13.2 25.6% 51.1% 18.8% 4.5% Imported rice available Household head has university degree=0 in the community=0 =1 =1 In(pcexp)=11.2 In(pcexp)=11.7 In(pcexp)=11.9 In(pcexp)=12.5 5.6% 20.0% 49.7% 1.4% Flexible Poverty Profiling and Welfare Prediction in Madagascar  89 FIGURE 3.4: Regression Tree of Logged per Capita Household Expenditures on Household and Community Level Variables, 2010 EPM Agricultural Households (n = 8,145) In(pcexp)=11.8 100% Percent of households with electricity <0.75 >=0.75 In(pcexp)=11.7 In(pcexp)=12.2 90.7% 9.3% Imported rice available in the community Percent of households with electricity =0 =1 <0.525 >=0.525 In(pcexp)=11.4 In(pcexp)=11.8 In(pcexp)=12.1 In(pcexp)=12.5 11.8% 78.9% 6.5% 2.8% KM to nearest urban center Land cultivated (acres) >=117 <117 <176 >=176 In(pcexp)=11.2 In(pcexp)=11.6 In(pcexp)=11.7 In(pcexp)=12 6.4% 5.5% 59.1% 19.7% Hours to heath center Lives in Diana Hours to heath center >=7 <7 =0 =1 >=0.852 <0.852 In(pcexp)=10.5 In(pcexp)=11.3 In(pcexp)=11.7 In(pcexp)=12.2 In(pcexp)=11.9 In(pcexp)=12.2 0.9% 5.5% 56.9% 2.3% 13.0% 6.7% Mean price imported rice Land cultivated (acres) >=1048 <1048 <601 >=601 In(pcexp)=11.6 In(pcexp)=11.8 In(pcexp)=12.1 In(pcexp)=12.8 19.2% 37.7% 6.0% 0.7% St dev. price of paddy rice <150 >=150 In(pcexp)=11.5 In(pcexp)=11.8 15.5% 3.7% identified by the algorithm. If a given household’s char- and therefore those that appear in the RT explain more acteristics meet the logical condition presented in the left variation in consumption than do those that do not branch (in the case of figure 3.2, this first condition is appear in the tree. whether the household is located in a community where fewer than 27.5 percent of households have electricity), From the CART analysis shown in figure 3.2, we see then one continues down the left branch of this first that households living in communities without much node. If the household meets the condition in the oppos- electricity, households with larger household size and/or ing branch (or equivalently, fails to meet the condition larger dependency ratios, households living in communi- in the left branch), then one continues down the right ties where imported rice is not available, and heads of branch. As one moves down the tree, the conditional household with low education (illiterate) are at the lower mean of consumption and the percent of households that end of the per capita consumption distribution, whereas remain in that branch (meeting the preceding conditions) households living in areas with more electricity, with are each presented at each node. Likewise, the terminal lower dependency ratios, and headed by individuals with node for each branch is the mean predicted per capita university-level education have higher consumption. expenditure for households meeting all the conditions in each of the preceding nodes in that branch of the The electricity variable may to some extent capture tree. All variables listed in annex table 3A.1 are utilized, differences between larger and smaller urban and 90 Republic of Madagascar Employment and Poverty Analysis increasingly rural households that are not sufficiently consumption needs and household-level economies of absorbed by the urban-rural variable and regional scale through an adult equivalence scale (see, for exam- dummies, which are also included. In fact, the average ple, Deaton and Zaidi 1999; Deaton 1997). Because rural household lives in a community where 4 percent household composition has large effects on per capita of households have electricity, while the average urban consumption by construction, for the rest of the analysis, household lives in a community where 35 percent of these demographic variables—household size and the households have electricity. Moreover, the average house- dependency ratio—are omitted. hold in the capital lives in a community where 77 per- cent of households have electricity, while the average Figure 3.3 displays the results with basic demographic household in the rest of the country lives in a community variables omitted. Absent household size and the where only 18 percent of households have electricity. dependency ratio, access to electricity remains a key One cannot infer causality in the sense that access to explanatory variable, and education emerges as even electricity directly raises incomes and therefore consump- more important—in particular, whether the household tion levels. This result could be due to the increased head is illiterate or has a university education. Moreover, economic activity or wealth in the community. In this the availability of imported rice in the community and case, these community attributes would cause greater whether or not the household is in the Diana region— income-generating opportunities, but the inference that the northern-most region of Madagascar where many electricity itself causes higher income could be spurious. households rely on fishing, forest products, and agri- Alternatively, community-level electrification may proxy culture for their livelihoods—bifurcates the households for the households’ unobserved level of wealth.10 We having among the lowest consumption levels. Among examine the likelihood of these interpretations in some households with slightly greater consumption levels (just of the following material. below the consumption poverty line but above the most destitute households in the sample), the remoteness vari- The variables that appear in the tree in figure 3.2 allow able, “hours to nearest health center” bifurcates house- the algorithm to capture more variance in the dependent holds, placing those traveling for longer than 0.85 hours variable, consumption, than do the variables that do to a health center in a lower consumption branch. Other not appear. Any variable that did not appear in this tree indicators of access to services and remoteness are less failed to improve the sum of squared prediction error of important predictors than this one. the model by a minimum of 0.007. In this respect, it may provide just as much insight to consider the variables Next, we perform the same analysis over the subset that the algorithm does not select as those that it does, of 8,145 households in which the head of household as we build a differentiated profile of poor households in reported agricultural work as his or her primary income- Madagascar. In figure 3.2 we do not see several vari- generating activity (with demographic variables again ables we might expect, including female-headed house- excluded). The resulting regression tree is presented in holds, land ownership, whether the household lives in a figure 3.4, where we can see that electricity again serves rural or urban environment, or any information about as the first splitting factor as well as the variable that households’ livelihoods. We will explore the correlates of bifurcates the households with the highest consumption several of these variables further to understand whether, levels (log per capita consumption of 12.5) from those for example, the rural-urban and livelihoods informa- with slightly lower consumption (log per capita con- tion is being picked up by the electricity and education sumption of 12.1). variables. Following the leftmost branch of figure 3.4 from the Interpretation of the results in figure 3.2 is complicated root node, we see that the agricultural households with by the fact that household size and dependency ratio the lowest consumption levels (log per capita consump- variables capture both welfare and measurement issues. tion of 10.5) are found in communities that lack avail- Although there is a real relationship between dependency ability of imported rice, are far from the nearest urban and welfare, whenever one uses per capita consumption centers, and are far from the nearest health center: the as the welfare indicator, one will overstate the rela- lowest-consumption agricultural households are those tive impact of household size on household members’ living in the remotest areas. Among households in the welfare, as this indicator fails to adjust for age-specific middle of the expenditure distribution, having larger Flexible Poverty Profiling and Welfare Prediction in Madagascar  91 land holdings and living in less remote areas are associ- each type of rice based on whether they bought or pro- ated with slightly higher consumption levels. Meanwhile, duced more by weight for each type of rice. As shown in having smaller land holdings, household residence in any table 3.2, 63.7 percent of households are net producers region besides Diana, and facing higher imported rice of paddy rice, that is, they produce more paddy rice than prices and lower standard deviation of paddy rice prices they buy. However, only 0.4 percent are net producers of are associated with slightly lower consumption. Note dehulled rice. Meanwhile, only 2.0 percent of households that, in comparison to the full sample, education vari- are net consumers of paddy rice while 72.4 percent are ables (such as literacy and having completed university) net consumers of dehulled rice.11 Overall, 4.3 percent are not differentiating factors for the consumption levels of the population is involved in neither production nor among agricultural households. Education primarily dif- consumption of either type of rice. The fact that the local ferentiates consumption levels between agricultural and nonavailability of imported rice bifurcates those house- nonagricultural households. holds at the lower end of the consumption distribution suggests that the variable may proxy for local commu- While these results provide clear profiling information, nity purchasing power, that is, the availability of this rice the implications for food and policy and public invest- is lower where households are less able to afford it. Or ments in infrastructure are less clear. To aid our interpre- it may proxy for market integration and/or the prefer- tation, table 3.1 shows that the price, standard deviation, ences of rice consumers in some remote areas. Moreover, and the local availability of rice differ significantly by where it is available, but at a higher price than 1,048 type. White and imported rice are both more expensive ariary per kilogram, only households with extremely low and more available than is paddy rice in local markets. landholdings are poorer. Such households would tend to Although imported rice is slightly more expensive than be net consumers of rice, and a lower imported rice price “white” rice, it has the lowest standard deviation of all would tend to push down other local prices (figure 3.4). rice prices across the seasons, and thus could be less Further insight on these questions is gained from the RF expensive in certain seasons. Imported rice is available in analysis that follows. 80 percent of communities, but is not quite as ubiqui- tous as the slightly less expensive white rice, available in RANDOM FOREST RESULTS 98 percent. To pin down the importance ranking of predictive vari- One would expect rice prices to affect rice producers and ables, we next ran both the full and agricultural- net rice consumers differently. Households in 2010 gen- household-only datasets through the RF algorithm. erally sold paddy rice and purchased dehulled rice. We Because RFs are more robust than single-regression trees, classify households as net consumers or net producers of they can help confirm and extend several of the insights TABLE 3.1: Mean and Standard Deviation of Seasonal Prices (per Kilogram) and Local Availability of Rice by Type, at the Community Level Mean price SD of prices Availability of rice type Type across seasons across seasons in local community White rice 974.98 169.73 97.5% Paddy rice 730.00 168.39 49.8% Imported rice 997.60  99.27 80.3% TABLE 3.2: Consumer or Producer Status by Type of Rice (Population Weighted) Neither net producer Net producer Net consumer nor net consumer Paddy rice 63.7%  2.0% 34.3% Dehulled rice  0.4% 72.4% 27.2% 92 Republic of Madagascar Employment and Poverty Analysis from the RTs. However, RFs are more difficult to In the RF variable importance plot for the full sample (fig- interpret than RTs as they cannot offer a single branch- ure 3.5) we see, as anticipated by the single RT analysis ing figure displaying the conditional relationships and reported in figure 3.3, that variables such as the percent of interactions among the variables. Therefore, we report households with electricity in the community and whether two types of output from the RF algorithm: variable the head of household has a university degree or is literate importance plots (figures 3.5 and 3.7) and partial depen- play a large role in reducing the out-of-sample prediction dence plots (figures 3.6 and 3.8). The variable impor- error. In addition, the remoteness variable, “kilometers to tance measures how great a role a given variable plays in the nearest urban center,” the mean price of paddy rice in reducing the error of the out-of-sample prediction across the community, the TLU holdings of the household, and the forest, while partial dependence plots display the the number of primary school educated members in the effects of variables of interest on the forest’s prediction household play a more substantial role than the other of consumption. variables in reducing the out-of-sample prediction error. FIGURE 3.5: Variable Importance Plot, 2010 EPM (n = 12,460) pct_hh_electricty_community hh_head_university_ed hh-head_literate km_nrst_urban mean_price_paddy_rice tlu nbr_w_primary_ed sd_price_white_rice sd_price_paddy_rice mean_price_white_rice hh_head_secondary_ed mean_price_npk transp_50kg_wetseas sd_price_imprted_rice land_cultivated_ares transp_50kg_dryseas pct_revenue_non_ag mean_price_urea hours_to_health_center imprted_rice_avail_community mean_price_imprtd_rice net_cons_dehulled sd_price_npk hours_to_aginputs pct_revenue_livestock hh_head_agriculture age_of_hh_head hours_to_public_transport hh_head_married pct_revenue_ag 25 30 35 40 45 50 55 60 %IncMSE Note: The x-axis of this dot plot is percent increase in mean-squared error (MSE). %IncMSE is the percent that MSE of predicted logged per capita income increases due to the perturbance of this variable. pct_hh_electricity_community is the percent of households in the community with electricity; hh_head_university is a binary variable indicating whether or not the household head has completed university; hh_head_literate is a binary variable indicating whether or not the head of household is literate; sd_price_white_rice is the standard deviation in the price of white rice across the seasons as available in the local community; nbr_w_primary_ed is the number of members of the household with a primary school education. EPM = Enquête Périodique auprès des Ménages. FIGURE 3.6: Partial Dependence Plots, 2010 Log per Capita Consumption (y Axis) on Key Variables (n = 12,460) a. Part. dep. on pct_hh-electricity_community b. Part. dep. on hh_head_university_ed c. Part. dep. on hh_head_literate 12.6 12.6 12.6 12.4 12.4 12.4 12.2 12.2 12.2 12.0 12.0 12.0 11.8 11.8 11.8 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 pct_hh_electricity_community hh_head_university_ed hh_head_literate d. Part. dep. on km_nrst_urban e. Part. dep. on mean_price_paddy_rice f. Part. dep. on tlu 12.6 12.6 12.6 Flexible Poverty Profiling and Welfare Prediction in Madagascar  12.4 12.4 12.4 12.2 12.2 12.2 12.0 12.0 12.0 11.8 11.8 11.8 0.0 0.2 0.4 0.6 0.8 1.0 0 5,000 10,000 15,000 0 50 100 150 200 250 300 350 km_nrst_urban mean_price_paddy_rice tlu Note: pct_hh_electricity_community is the percent of households in the community with electricity; hh_head_university is a binary indicating whether or not the household head has completed university; sd_ price_white_rice is the standard deviation in the price of white rice across the seasons as available in the local community; mean_price_white_rice is the mean price of white rice across the seasons as available in the local community; hh_head_literate is a binary indicating whether or not the head of household is literate; nbr_w_primary_ed is the number of members of the household with a primary school education. Part_dep=partial dependence. 93 94 Republic of Madagascar Employment and Poverty Analysis FIGURE 3.7: Variable Importance Plot, 2010 EPM Agricultural Households (n = 8145) land_cultivated_ares km_nrst_urban pct_hh_electricity_community pct_revenu_ag mean_price_paddy_rice sd_price_white_rice tlu mean_price_white_rice mean_price_npk sd_price_paddy_rice pct_revenue_livestock sd_price_imprtd_rice hours_to_public_transport mean_price_urea pct_revenue_non_ag mean_price_imprtd_rice transp_50kg_wetseas hours_to_aginputs imprtd_rice_avil_community transp_50kg_dryseas hours_to_health_center hours_to_market sd_price_npk age_of_hh_head nbr_w_primary_ed hh_head_literate hh_head_married net_cons_dehulled Diana rural 20 25 30 35 40 45 50 %IncMSE Note: The x-axis of this dot plot is percent increase in MSE. %IncMSE is the percent that MSE of predicted logged per capita income increases due to the perturbance of this variable. land_cultivated_ares is the land area cultivated by the household in the local measurement unit of ares; pct_hh_ electricity_community is the percent of households in the community with electricity; mean_price_paddy_rice is the mean price of paddy rice across the seasons as available in the local community; remoteness is an index capturing how remote the community is in terms of access to services; sd_ price_white_rice is the standard deviation in the price of white rice across the seasons as available in the local community. EPM = Enquête Périodique auprès des Ménages. In considering variable importance for prediction of price of paddy rice, and TLU holdings are important per capita expenditures among agricultural households predictors in correctly predicting where an agricultural only (figure 3.7) we see a somewhat different ranking of household will lie on the expenditure distribution. variables by predictive importance. Notably, land area cultivated (ares) and the percentage of revenue from The RF results for both the full and agricultural house- agricultural activities emerge as more important in this hold samples yield a different ranking of key variables subset of the data. Consumption is increasing with the related to outcomes in local rice markets and thus a dif- area of cultivated land, as would be expected in a con- ferent interpretation of results. Both analyses underscore text with such small farm sizes, and is decreasing in the the performance of markets for the country’s staple food percentage of revenues from agriculture. In addition, as and dominant crop for poverty reduction. Because the with the full sample, we see that distance to the nearest different indicators in rice markets are related and may urban center, electrification in the community, the mean have nonlinear effects, however, the ranking is sensitive FIGURE 3.8: Partial Dependence Plots, 2010 Log per Capita Consumption (y Axis) on Key Variables, Agricultural Households (n = 8,145) (Relevant Range Circled) a. Part. dep. on land_cultivated_ares b. Part. dep. on km_nrst_urban c. Part. dep. on pct_hh_electricity-community 12.4 12.4 12.4 12.2 12.2 12.2 12.0 12.0 12.0 11.8 11.8 11.8 11.6 11.6 11.6 0 1,000 2,000 3,000 4,000 5,000 6,000 7,000 0 1,000 2,000 3,000 4,000 0.0 0.2 0.4 0.6 0.8 1.0 land_cultivated_ares km_nrst_urban pct_hh_electricity-community d. Part. dep. on pct_revenue_ag e. Part. dep. on mean_price_paddy_rice f. Part. dep. on sd_price_white_rice Flexible Poverty Profiling and Welfare Prediction in Madagascar  12.4 12.4 12.4 12.2 12.2 12.2 12.0 12.0 12.0 11.8 11.8 11.8 11.6 11.6 11.6 0.0 0.2 0.4 0.6 0.8 1.0 0 5,000 10,000 15,000 0 1,000 2,000 3,000 4,000 pct_revenue_ag mean_price_paddy_rice sd_price_white_rice (continued) 95 96 FIGURE 3.8: Continued g. Part. dep. on land_cultivated_ares h. Part. dep. on km_nrst_urban i. Part. dep. on pct_hh_electricity-community 12.4 12.4 12.4 12.2 12.2 12.2 12.0 12.0 12.0 11.8 11.8 11.8 11.6 11.6 11.6 0 1,000 2,000 3,000 4,000 5,000 6,000 7,000 0 1,000 2,000 3,000 4,000 0.0 0.2 0.4 0.6 0.8 1.0 land_cultivated_ares km_nrst_urban pct_hh_electricity-community j. Part. dep. on pct_revenue_ag k. Part. dep. on mean_price_paddy_rice l. Part. dep. on sd_price_white_rice 12.4 12.4 12.4 12.2 12.2 12.2 12.0 12.0 12.0 11.8 11.8 11.8 11.6 11.6 11.6 0.0 0.2 0.4 0.6 0.8 1.0 0 5,000 10,000 15,000 0 1,000 2,000 3,000 4,000 pct_revenue_ag mean_price_paddy_rice sd_price_white_rice (continued) Republic of Madagascar Employment and Poverty Analysis FIGURE 3.8: Continued m. Part. dep. on tlu n. Part. dep. on mean_price_white_rice o. Part. dep. on mean_price_npk 12.0 12.0 12.0 11.9 11.9 11.9 11.8 11.8 11.8 11.7 11.7 11.7 11.6 11.6 11.6 0 50 100 150 200 250 300 350 500 1,000 1,500 2,000 2,500 3,000 3,500 0e+00 1e+05 2e+05 3e+05 4e+05 5e+05 6e+05 tlu mean_price_white_rice mean_price_npk p. Part. dep. on sd_price_paddy_rice q. Part. dep. on pct_revenue_livestock r. Part. dep. on sd_price_imprtd_rice 12.0 12.0 12.0 Flexible Poverty Profiling and Welfare Prediction in Madagascar  11.9 11.9 11.9 11.8 11.8 11.8 11.7 11.7 11.7 11.6 11.6 11.6 0 5,000 10,000 15,000 20,000 0.0 0.2 0.4 0.6 0.8 1.0 0 200 400 600 800 1,000 sd_price_paddy_rice pct_revenue_livestock sd_price_imprtd_rice Note: part. Dep. = partial dependence. 97 98 Republic of Madagascar Employment and Poverty Analysis to the precise methodology used. Given that RF provides other variables, the marginal effects (represented by the a more accurate out-of-sample prediction and better slopes of the curves) are not as pronounced. Rather, for illustrates possible nonlinearities, we derive our interpre- the binary education variables (household head has a tation of results on rice markets from this methodology. university degree and household head is literate) we see The RF results suggest that the availability and price of slopes indicative of a positive marginal effect across the imported rice is not as important a determinant of wel- mean range of the household consumption distribution. fare as in the RT results. Rather, the mean price of paddy For continuous variables, such as kilometers to the near- rice and standard deviations of the price for white rice est urban center, the mean price of paddy rice, and TLU and paddy rice ranking higher in importance. The avail- holdings, we see clear slopes where the bulk of the data ability of imported rice falls to 20th place in the ranking lie, as indicated by the blue circles. of variables, below the price of transport, the price of urea fertilizer, and hours to the nearest health center. Figure 3.8 reports partial dependence plots for the most important variables (those reported in figure 3.7) for To better observe the role of each of these important agricultural households only. The marginal effect of land predictors, figures 3.6 and 3.8 display their partial area cultivated appears to be significant at low levels of dependence plots, that is, the relationships between the cultivated land, where the observations are most dense. variables on the horizontal axis and predicted log per The curve flattens out at approximately 1.5 hectares capita consumption (on the vertical axis), holding all (1,500 ares), but as shown the cultivated areas per other variables at their mean values. The “rug plots” at household tend to be much lower. TLU holdings follow the bottom of each plot indicate the data density, with a similar though much more gradual trajectory as that circles to highlight the relevant ranges. In figure 3.6, seen for land. The distance to the nearest urban center where partial dependence plots are reported for the has a negative correlation with consumption in the full sample analysis, one observes a steep relation- relevant range, as expected. Finally, the mean price of ship between the percentage number of households paddy rice has a positive relationship with consumption with electricity in the community and welfare. For the in the range where the data are available. TABLE 3.3: Multivariate Correlates of Electrification Variable (N = 12,460, Region Dummies Included) Dependent variable: % HH with electricity (1) (2) (3) (4) (5) (6) (7) Rural (0/1) –0.445*** –0.439*** –0.409*** –0.406*** –0.407*** –0.386*** –0.361*** (0.00473) (0.00472) (0.00474) (0.00473) (0.00472) (0.00480) (0.00443) Owns nonagricultural enterprise 0.0570*** –0.0392*** –0.0372*** –0.0368*** –0.0370*** –0.0293*** (0.00431) (0.00559) (0.00558) (0.00556) (0.00547) (0.00496) Percentage of revenue nonagriculture 0.195*** 0.189*** 0.188*** 0.177*** 0.136*** (0.00747) (0.00746) (0.00745) (0.00736) (0.00672) Cost of transporting 50 Kg rice, wet season –0.00000113 –0.00000133* –0.00000155** –0.00000213*** (0.000000721) (0.000000719) (0.000000714) (0.000000688) Cost of transporting 50 Kg rice, dry season –0.00000258*** –0.00000177** –0.000000216 0.00000148* (0.000000891) (0.000000894) (0.000000893) (0.000000883) Distance to nearest urban center –0.000106*** –0.000111*** –0.0000304** (0.0000132) (0.0000130) (0.0000132) Flexible Poverty Profiling and Welfare Prediction in Madagascar  Hours to reach nearest public transport –0.00355*** –0.00146*** (0.000221) (0.000215) Hours to reach nearest school 0.000480 0.00136* (0.000803) (0.000731) Hours to reach agricultural input supplier –0.000672*** –0.000685*** (0.000219) (0.000213) Hours to nearest health center –0.00136*** –0.00101*** (0.000421) (0.000380) Hours to nearest food market 0.00248*** 0.00127*** (0.000363) (0.000357) _cons 0.503*** 0.478*** 0.438*** 0.452*** 0.460*** 0.475*** 0.650*** (0.00409) (0.00446) (0.00462) (0.00482) (0.00491) (0.00495) (0.00564) R-squared 0.415 0.423 0.453 0.457 0.460 0.478 0.582 Note: Standard errors in parentheses; * p < 0.10, ** p < 0.05, *** p < 0.01. 99 100 Republic of Madagascar Employment and Poverty Analysis To ascertain why the electrification variable plays such a Further investigation shows that while the costs of large role throughout these analyses and to better under- transporting 50 kilograms of rice during the dry and stand what other things this variable may be capturing, a wet seasons (column 4) are not significantly correlated closer look at the correlates of this variable via multivar- with electrification, the distance of the household from iate regression are provided in table 3.3. Both livelihood the nearest urban center is significant and negatively variables and remoteness variables are considered in this correlated with electrification (column 5), as we might examination of correlates. In table 3.3’s first regression expect. In column 6, several proxies for remoteness, used (column 1), we see that whether the household is located throughout the regression tree and forest analyses, are in a rural environment or not is statistically significant included. All but hours to the nearest school are sig- and explains 42 percent (R-squared) of the variation in nificant. Of the significant coefficients, all are negative the electrification variable. Subsequent regressions in except hours to the nearest market. In the final column, this table, in which livelihood, remoteness, and regional regional dummy variables are included (a breakdown of dummy variables are progressively added, decrease electrification by region is shown in annex table 3A.2); the magnitude of the rural variable slightly, but do not although the coefficients on these regional dummies are decrease its significance. In regressions 2 and 3, liveli- suppressed here, all are statistically significant. With hood variables such as ownership of and percent revenue all variables included, the final regression “explains” from nonagricultural enterprise are added. When percent 58 percent (adjusted R-squared) of the variation in revenue from nonagricultural enterprise is added to electrification across communities. While this is a high the regression in column 3, we see that the ownership adjusted R-squared, it still leaves much of the variation coefficient switches signs, suggesting that when the in electrification unexplained, suggesting that things amount of income derived from nonagricultural income we cannot observe in the data—perhaps differences in is accounted for, nonagricultural-enterprise-owning opportunity, population density and market size, the households are less likely to reside in electrified areas, costs of delivering electricity, or political connectedness but those that do live in such areas derive more of their of some communities—are also driving this variable. income from nonagricultural endeavors. Flexible Poverty Profiling and Welfare Prediction in Madagascar  101 Discussion and Conclusion Yet among agricultural households, educational attain- ment does not appear to be an important predictor Several clusters of variables emerge from the analyses of expenditures. Rather, other productive assets such as highly predictive of a household’s falling along the as land area and livestock holdings (below a certain high or low end of the expected per capita consumption threshold), the market prices of farm outputs and inputs, distribution. In the full sample, the poorest households community-level electrification, and several proxies for are those found in communities where fewer households remoteness (distance from nearest urban center, distance have electricity, a variable that is correlated with remote- from nearest health center) appear to play a larger role. ness, livelihood strategies, and regions but may also be The poorest households live in the remotest areas and correlated with unobserved factors such as market and have low land and livestock holdings. On the list of educational opportunities that we cannot observe. In the variables that are not as predictive of consumption per full sample, household-level features such as having a lit- capita are ownership of agricultural equipment, regional erate or university-educated head of household also play indicators, gender of household head, and marital status. a large role in separating higher and lower consumption households. The poorest households have an illiterate These findings provide implications for targeting and head of household, while the wealthiest households guideposts for key policy areas. In particular, they imply have a head with a university degree. Although heads of the targeting of interventions to reach households with household with university degrees are not more likely to especially low land holdings, in communities with be employed (91 percent employed) than those without productive potential but lacking electricity, and to better (95 percent employed), they are much more likely to be connect those in more remote areas. Although the issues living in an urban environment than a rural one as well, associated with rice policies are complex and merit and are more likely to be living in the capital than not. further investigation, these results indicate that on the Literacy follows the same employment and environment whole higher producer prices aid poverty reduction. pattern. 102 Republic of Madagascar Employment and Poverty Analysis Annex 3A. Explanation of Methods Because the data available for the analysis of poverty in highly accurate out-of-sample predictions with minimal Madagascar are not ideal for obtaining sound identifica- variance (Kleinberg, Mullainathan, and Obermeyer tion for the purposes of inference, this paper takes the 2015). OLS, as the best linear unbiased estimator, does approach of predictive analytics. Predictive analytics not allow for such trade-offs. Third, these methods allow differ from traditional regression analysis in several fun- for nonparametric analysis with unlimited interactions damental ways, and therefore offer several advantages. and without a predefined functional form; instead, the First, these methods target prediction of an outcome over data define the form. In this paper, the regression tree and above parameterization of a model. Second, and and regression forest analyses are implemented in R relatedly, these methods make a bias for variance trade- using packages developed by Therneau, Atkinson, and off such that they do not produce unbiased coefficient Ripley (2015) and Liaw and Wiener (2002), respectively. estimates in the manner of OLS. Rather they produce TABLE 3A.1: Summary Statistics, EPM 2010 (N = 12,460, Household Survey Weights Applied) Mean (household Linearized Variable Variable name weighted) std. err. 95% conf Interval Household size hh_size 4.76 0.03 4.70 4.81 Age of head of household age_of_hh_head 41.96 0.16 41.63 42.28 Head of household literate (y/n) hh_head_literate 0.73 0.00 0.72 0.74 Head of household has completed primary hh_head_primar~d 0.30 0.01 0.29 0.31 school Head of household has completed secondary hh_head_second~d 0.15 0.00 0.14 0.15 school Head of household has completed university hh_head_univer~d 0.06 0.00 0.05 0.06 Number of households in the ea with pct_hh_electrity_ 0.17 0.00 0.16 0.18 electricity community Percent of revenue from fishing pct_revenue_fish 0.03 0.00 0.02 0.03 Percent of revenue from nonagricultural pct_revenue_no~g 0.26 0.00 0.26 0.27 enterprise Percent of revenue from agriculture pct_revenue_ag 0.51 0.00 0.51 0.52 Percent of revenue from livestock pct_revenue_livestock 0.11 0.00 0.10 0.11 Own agricultural cart (y/n) owns_ag_cart 0.08 0.00 0.08 0.09 Own plow (y/n) owns_ag_plow 0.10 0.00 0.10 0.11 Own harrow (y/n) owns_ag_harrow 0.08 0.00 0.07 0.08 Owns agricultural equipment (y/n) owns_ag_equip 0.77 0.00 0.76 0.78 Own nonagricultural enterprise (y/n) owns_non_ag_enterprise 0.35 0.01 0.34 0.36 Number of household members with a nbr_w_primary_ed 1.11 0.01 1.08 1.14 primary school education Tropical livestock units owned by household tlu 1.79 0.07 1.65 1.92 Dependency ratio depr 0.43 0.00 0.43 0.44 Head of household is female (y/n) hh_head_female 0.20 0.00 0.19 0.21 Head of household married (y/n) hh_head_married 0.75 0.00 0.74 0.76 Head of household divorced or separated hh_head_div_sep 0.10 0.00 0.09 0.11 (y/n) Head of household is widowed (y/n) hh_head_widowed 0.09 0.00 0.08 0.09 Head of household is employed (y/n) hh_head_employed 0.95 0.00 0.95 0.96 (continued) Flexible Poverty Profiling and Welfare Prediction in Madagascar  103 TABLE 3A.1: Continued Mean (household Linearized Variable Variable name weighted) std. err. 95% conf Interval Household is agricultural household hh_head_agriculture 0.68 0.01 0.67 0.69 Household lives in community with bad or bad_security 0.33 0.01 0.32 0.34 very bad security conditions Household living in community with average ok_security 0.31 0.01 0.30 0.32 security conditions Household lives in rural area (y/n) rural 0.75 0.00 0.74 0.76 Mean community price of white rice across mean_price_white_rice 974.98 2.85 969.40 980.56 seasons Standard deviation of community price of sd_price_white_rice 169.73 4.83 160.26 179.20 white rice across seasons Mean community price of imported rice mean_price_imprtd_rice 997.60 4.47 988.84 1006.35 across seasons Standard deviation of community price of sd_price_imprtd_rice 99.28 3.87 91.70 106.85 imported rice across seasons Mean community price of paddy rice across mean_price_paddy_rice 730.00 8.70 712.95 747.05 seasons Standard deviation of community price of sd_price_paddy_rice 168.39 4.81 158.95 177.83 paddy rice across seasons Mean community price of npk across seasons mean_price_npk 5720.18 422.66 4891.70 6548.65 Standard deviation of community price of npk sd_price_npk 176.40 11.74 153.38 199.41 across seasons Mean community price of urea across seasons mean_price_urea 1564.03 8.01 1548.33 1579.73 Standard deviation of community price of sd_price_urea 62.22 1.82 58.65 65.80 urea across seasons White rice available in community (y/n) white_rice_avail 0.97 0.00 0.97 0.98 Paddy rice available in community (y/n) paddy_rice_avail 0.50 0.01 0.49 0.51 Imported rice available in community (y/n) imprtd_rice_avail 0.80 0.00 0.79 0.81 Npk available in community (y/n) npk_avail_community 0.39 0.01 0.38 0.40 Urea available in community (y/n) urea_avail_comunity 0.36 0.01 0.35 0.37 Land cultivated (ares) land_cultivated_ares 101.94 2.08 97.85 106.02 Net producer of paddy rice (y/n) net_prod_paddy 0.64 0.01 0.63 0.65 Net consumer of paddy rice (y/n) net_cons_paddy 0.02 0.00 0.02 0.02 Net producer of dehulled rice (y/n) net_prod_dehulled 0.00 0.00 0.00 0.01 Net consumer of dehulled rice (y/n) net_cons_dehulled 0.72 0.01 0.71 0.73 Household lives in capital (y/n) capital 0.07 0.00 0.06 0.08 Household experienced a climate shock (y/n) climate_shock 0.34 0.01 0.33 0.35 Household experienced an economic shock economic_shock 0.10 0.00 0.09 0.11 (y/n) Household experienced a health shock (y/n) health_shock 0.06 0.00 0.05 0.06 Household experienced a security shock (y/n) security_shock 0.06 0.00 0.05 0.06 Household experienced other shock (y/n) other_shock 0.01 0.00 0.00 0.01 Hours from community to nearest market hours_to_market 3.62 0.11 3.41 3.82 Hours from community to nearest health hours_to_health_center 2.75 0.08 2.59 2.91 center (continued) 104 Republic of Madagascar Employment and Poverty Analysis TABLE 3A.1: Continued Mean (household Linearized Variable Variable name weighted) std. err. 95% conf Interval Hours from community to location where ag hours_to_aginputs 9.51 0.15 9.22 9.80 inputs can be purchased Hours from community to nearest school hours_to_school 0.94 0.02 0.89 0.99 Hours from community to nearest public hours_to_public_transp 9.14 0.14 8.88 9.41 transportation Kilometers from community to nearest urban km_nrst_urban 92.60 1.39 89.88 95.31 center Cost of transporting 50kg of rice to nearest transp_50kg_wetseas 4210.15 70.77 4071.43 4348.86 urban center, wet season Cost of transporting 50kg of rice to nearest transp_50kg_dryseas 3873.09 57.93 3759.54 3986.64 urban center, dry season Household located in Analamanga (Y/N) Analamanga 0.17 0.01 0.16 0.18 Household located in Vakinankaratra (Y/N) Vakinankaratra 0.08 0.00 0.07 0.08 Household located in Itasy (Y/N) Itasy 0.03 0.00 0.02 0.03 Household located in Bongolava (Y/N) Bongolava 0.02 0.00 0.02 0.02 Household located in MatsiatraAmbony (Y/N) MatsiatraAmbony 0.05 0.00 0.04 0.05 Household located in AmoroniMania (Y/N) AmoroniMania 0.03 0.00 0.03 0.03 Household located in VatovavyFitovi~y (Y/N) VatovavyFitovi~y 0.06 0.00 0.05 0.06 Household located in Ihorombe (Y/N) Ihorombe 0.02 0.00 0.01 0.02 Household located in AtsimoAtsinanana (Y/N) AtsimoAtsinanana 0.03 0.00 0.03 0.04 Household located in Atsinanana (Y/N) Atsinanana 0.06 0.00 0.06 0.07 Household located in Analanjirofo (Y/N) Analanjirofo 0.05 0.00 0.05 0.06 Household located in AlaotraMangoro (Y/N) AlaotraMangoro 0.05 0.00 0.04 0.05 Household located in Boeny (Y/N) Boeny 0.04 0.00 0.03 0.04 Household located in Sofia (Y/N) Sofia 0.06 0.00 0.05 0.06 Household located in Betsiboka (Y/N) Betsiboka 0.01 0.00 0.01 0.01 Household located in Melaky (Y/N) Melaky 0.01 0.00 0.01 0.01 Household located in AtsimoAndrefana (Y/N) AtsimoAndrefana 0.06 0.00 0.06 0.07 Household located in Androy (Y/N) Androy 0.03 0.00 0.03 0.03 Household located in Anosy (Y/N) Anosy 0.03 0.00 0.03 0.03 Household located in Menabe (Y/N) Menabe 0.03 0.00 0.02 0.03 Household located in Diana (Y/N) Diana 0.04 0.00 0.04 0.05 Std. err = standard error  TABLE 3A.2: Electrification by Region 6. Following guidance from Harvest Choice, TLU were calculated as follows: tlu = 0.7*ox + 0.7*cow + 0.1*sheep + 0.1*goat + 0.2*pig + 0.01*chicken + 0.01*turkey + 0.01*duck + 0.01*goose + Region Electrification .001*rabbit. Analamanga 48.5% 7. The are is a local unit of area measurement; 1 are equals 100 square meters or 0.01 hectares. Vakinankaratra 11.0% 8. The 2010 EPM household survey has modules on both production Itasy 13.6% and consumption of multiple commodities, including rice. Each module includes kilograms produced and consumed of paddy rice Bongolava 4.4% and dehulled rice. Net sellers/buyers of paddy rice and net sellers/ buyers of dehulled rice are identified separately by calculating the Matsiatra Ambony 10.1% marketable surplus (marketable surplus = production – consump- Amoron’i Mania 4.9% tion) of these two commodities for each household. 9. Availability of each of these commodities is a binary variable Vatovavy Fitovinany 4.2% indicating whether the commodity price was reported for a given Ihorombe 5.2% community in the community survey. Where no price was reported for a given commodity, this variable is zero for that commodity. Atsimo Atsinanana 2.5% Where any price was reported for a given commodity, this variable is Atsinanana 22.5% one for that commodity. 10. Causal inference of the effect of expanded electricity would require Analanjirofo 10.0% additional empirical methods, which would require either more integrated or experimental data. Alaotra Mangoro 10.4% 11. Unfortunately, it is not possible to observe in the data whether the Boeny 23.7% dehulled rice category is composed of imported or white rice variet- ies or both (and in what proportion). Sofia 7.3% Betsiboka 5.0% Melaky 4.8% REFERENCES Atsimo Andrefana 12.6% Breiman, L. 2001. “Random Forests.” Machine Learning Androy 0.3% 45: 5–32. Anosy 9.6% Deaton, A. S. 1997. The Analysis of Household Surveys: Menabe 10.2% A Microeconomic Approach to Development Policy. Diana 20.3% Baltimore, MD: Johns Hopkins University Press. Sava 7.1% Deaton, A. S., and S. Zaidi. 1999. “Guidelines for Constructing Consumption = Aggregates for Source: EPM 2010 Welfare Analysis.” Mimeo. Princeton, NJ: Princeton University. Hastie, T., R. J. Tibshirani, and J. Friedman. 2009. The NOTES Elements of Statistical Learning: Data Mining, 1. We use per capita consumption as the welfare indicator in this Inference, and Prediction. 2nd ed. New York: analysis. Springer. 2. While the dependent variable in this analysis is household per capita consumption expenditures, throughout the analysis the less cumber- Kleinberg, J., S. Mullainathan, and Z. Obermeyer. 2015. some terms consumption or household consumption are used. “Prediction Policy Problems.” American Economic 3. We acknowledge the assumption that the same-data generating process may be violated in future periods, but because we include Review: Papers and Proceedings, 105(5): 491–95. variables such as experience of a climatic or health shock and key Liaw, A., and M. Wiener. 2002. “Classification and prices, which fluctuate over time, this assumption is not as strong as Regression by Random Forest.” R News 2:18–22. it may at first appear. 4. The advantage of the random selection of subsets of data in this Therneau, T., B. Atkinson, and B. Ripley. 2015. algorithm is that it de-correlates the trees from one another and rpart: Recursive Partitioning and Regression also reserves a subset of the data, not used to build a given tree, for unbiased testing of the accuracy of the prediction. This out-of-sample Trees. R package version 4.1-10. Phoenix, AZ: testing error is known as the out-of-bag error. Mayo Foundation. http://CRAN.R-project.org/ 5. In particular, the mean-squared error (MSE) measure of each vari- able’s importance in a regression forest is measured by randomly package=rpart. perturbing the variable of interest and recording the extent to which the out-of-bag error differs from that found with the unperturbed data (Hastie, Tibshirani, and Friedman 2009). The differences are averaged across all trees and then divided by the standard deviation of the differences to produce a normalized measure of the increase in MSE (%IncMSE), comparable across all variables. 106  107 CHAPTER 4 Labor Demand Estimation in Rural Madagascar: Shadow Wages and Allocative Inefficiency Margaret Jodlowski* June 2016 *Cornell University This research was conducted as part of a World Bank-supported analytical program under the task leadership of Theresa Osborne, who provided comments and guidance on earlier drafts, as well as the supervision of Christopher B. Barrett of Cornell University. Introduction U nderstanding the factors and circumstances that FIGURE 4.1: Agriculture Value Added per Worker influence rural labor demand in Madagascar is (Average 2011–14) of central importance for informing pro-poor 400 growth policies. As with many poor countries, most of Madagascar’s rural labor force is concentrated in 350 agriculture and lives in poverty. Agricultural productiv- ity is among the lowest in the world (see figure 4.1). 300 Agricultural wage laborers, who tend to be among the poorest of the poor, are typically underemployed and 250 paid very little. Many poor people work in both farm US$ 2005 and rural nonfarm enterprises (NFEs), which have been 200 shown to reduce rural poverty in poor countries (Barrett, Reardon, and Webb 2001), and yet these workers remain 150 poor. A more expansive and efficient labor market is of key importance, not only for job creation and wages, but 100 also for the productivity of farm and nonfarm enter- prises. Thus, understanding the determinants of labor 50 demand, including of any inefficiencies in these markets, is an important priority for designing pro-poor policies. 0 ar da ia so ia a a bi ny op an c Fa an as m Ke nz hi Ug Za ag a The major employer of labor in rural Madagascar is the Et Ta in ad rk Bu M household—households employ labor (household mem- bers and hired laborers) on farm plots and in household- Source: World Development Indicators (WDI). operated NFEs, and they are also the suppliers of labor. Agricultural workers typically work on household farms, 108 Republic of Madagascar Employment and Poverty Analysis and the vast majority of those who earn income off-farm that the observed wage rate for laborers is equivalent do so in the informal economy—working in an NFE, for those who are self-employed is similarly invalid, as often receiving in-kind compensation, or being self- those who work to earn a wage and those who work on employed. More than 85 percent of Malagasy workers the household farm are likely to differ in both observ- are employed in nonwage activities, and in 2005 only able and unobservable ways. Thus, Jacoby developed a 11 percent of rural adults were employed as a nonfamily method to estimate structural time allocation models for worker in an NFE (Stifel, Rakotomanana, and Celada households in the absence of observed market wages. 2007). Between 2001 and 2010, the percentage of house- Barrett, Sherlund, and Adesina (2008) generalized holds operating an NFE increased from 26.3 percent to Jacoby’s approach to accommodate risk, search, and 43.9 percent, while the percentage of these households transactions costs, as well as occupational and location that employed hired labor in their NFE declined from preferences. 30.8 percent in 2001 to 14.3 percent in 2005 and then stayed relatively constant between at 16 percent in 2010. This paper attempts to build an empirical understanding Thus, a movement into NFEs was not accompanied of the functioning of Madagascar’s rural labor markets, by a greater willingness to hire nonhousehold labor. If while also deriving insights into the factors affecting the this trend occurred despite the higher profit potential revenues of rural households. In particular, following from hiring such workers, it would suggest a friction on Randrianarisoa, Barrett, and Stifel (2009), we estimate the demand side of these labor markets, which reduces the proximate drivers of demand for labor by rural both labor and enterprise incomes in rural areas. Thus, households over the decade 2001 to 2010 using the understanding these and other outcomes requires an Enquête Périodique auprès des Ménages (EPM) for the understanding of the factors influencing a household’s years 2001, 2005, and 2010. Because farm and non- demand for labor. farm labor demand may differ, we analyze each sector separately, using only 2001 for the on-farm sector due Constraints to raising labor demand can arise through to data limitations in the subsequent surveys. We adapt the effects on the profitability (marginal revenue prod- the methods developed to study labor supply by Jacoby uct, MRP) of labor, or through frictions in the labor (1993) and Barrett, Sherlund, and Adesina (2008) to market, and thereby the level of employment relative to the problem of labor demand, using their approach to the efficient (profit-maximizing) level. Further, there is address the issue of unobserved wages. We also relax the strong evidence from across the continent that agri- assumption that the wage is equal to the marginal rev- cultural factor markets, especially the land and labor enue product of labor. We examine the shadow wage— markets, do not function competitively and are subject the wage firms (households in our case) would be willing to market failure. These failures are of potentially diverse to pay labor. origin, and include poor infrastructure and labor super- vision problems. Barrett and Dillon (2016) reject the Shadow wages are composed of two elements, the hypothesis of a well-functioning, complete, and competi- marginal revenue product of labor and an allocative tive labor market in five Sub-Saharan African countries.1 inefficiency factor (AIF) that captures the effects of the nonwage costs (or benefits) which firms see when Despite the widespread presumption that labor markets employing workers. If the nonwage costs exceed the in poor rural economies are inefficient, there is relatively benefits, this adjustment pushes down the wage the little research on the determinants of labor demand in employer is willing to pay. If there are nonwage benefits such settings (see Hammermesh 1996). Jacoby (1993) to employing workers, such as retaining high-quality was one of the first papers to structurally estimate workers, future training benefits, or employment as a shadow wages as the marginal revenue product of labor, means of sharing resources with workers, the AIF will in the presence of an informal (or nonexistent) wage result in a willingness to pay more than the marginal economy where shadow wages are determined within revenue product of labor. In our sample, we find that the household. He observed that traditional methods for there are relatively few cases of the latter. Given these analyzing the labor supply decisions of households, and two elements, shadow wages can be impacted by a vast other household-based time allocation models, are inap- array of factors. Affecting the marginal revenue prod- propriate for contexts where self-employment is ubiqui- uct of labor are technology; relative output and factor tous and wage rates are not observable. The assumption price movements (macroeconomic variables and levels Labor Demand Estimation in Rural Madagascar: Shadow Wages and Allocative Inefficiency 109 of market integration); and the cost and availability The availability and use of other inputs affects labor of other inputs, including infrastructure services, all of demand as well. We find that on-farm labor demand which affect the marginal revenue product of labor. In in 2001 is positively related to the land area cultivated addition there are the risks—costs of hiring, training, and livestock holdings of the household, as might be supervising, and letting go of workers—that affect the expected.2 For NFEs, a variety of factors significantly AIF. These can be affected by institutional arrange- affect labor demand, with different ones emerging as ments and household characteristics affecting the important over time. First, the number of working- ability to reduce these costs. We estimate the observable aged men and women in the household increased determinants of job creation (the extensive margin of demand in 2001 and 2010, a likely result of the lower labor demand growth), as well as the increase in hours labor market frictions involved in employing fam- worked for the same number of jobs (the intensive mar- ily labor. In 2001, having more education increased gin), in addition to the AIF in rural labor markets. In labor demand, but in 2010 it reduced it. Our results addition, we estimate the responsiveness (or elasticity) highlight the importance of physical infrastructure for of households’ total demand for labor, both paid and increasing the NFE revenue. In 2001 and 2005, higher unpaid, with respect to shifts in the supply of labor (or transport costs are associated with lower levels of NFE other nondemand-side drivers of wages), as well as the revenue. In 2010, the availability of electricity and efficiency of these labor markets. irrigation networks had a positive and significant rela- tionship with revenue for NFEs, but they did not affect Based on the estimated divergence between the marginal labor demand in any year studied, suggesting that revenue product and wages paid for households paying labor market frictions are not helped by these services. wages, we find evidence of significant allocative ineffi- Own investment in these small enterprises, measured ciency in rural labor markets. A wedge equal to approxi- by the value of equipment, also rises throughout the mately half the marginal revenue product in effect halves decade. the households’ willingness to pay for labor. The wedge also appears to be higher in the NFE sector than on- We also find that the demand for farm labor is wage farm: at the most extreme, wages for NFE workers in elastic, while NFE labor demand is inelastic and becomes 2001 are only about 10 percent of the marginal revenue more inelastic over time. Based on the (Hicks-Marshall) product of labor, while standard economic theory would theory of derived demand, elastic labor demand indicates equate them. The finding of such a large divergence that units of labor are easy to adjust with circumstances between the observed wage and the marginal revenue and workers are easily substituted, perhaps because the product of labor is not necessarily evidence of miscalcu- tasks performed by farm labor are neither highly special- lation on the part of households. This wedge is related ized nor complex. For NFEs, however, this may not be to nonwage costs and risks of hiring workers, but it is an the case. Rather than identify, hire, train, and supervise important factor affecting both the potential to generate paid workers, NFEs prefer to utilize less labor and accept labor income and the profitability of household farm lower profits, given these nonwage costs. As a result, and nonfarm enterprises. Further, the allocative ineffi- they tend to generate employment only of the household ciency we estimate is almost always negative: a negative members and are not currently promising candidates for value for the AIF indicates that labor is underdemanded providing wage labor in rural areas. by these household enterprises. This implies that there are barriers, only some which we can observe, to labor demand. We estimate the factors that are related to labor Background and Data being over or under demanded and find that house- hold enterprises for which the household head is well As a country with a poverty rate of 77.8 percent, the educated significantly over demand labor for both farms first decade of the millennium was not kind to Madagas- and NFEs in 2001. Also, the value of equipment was car.3 The country experienced two political crises during significantly related to an increased likelihood of labor the period covered by this study, first in 2002 and again being over demanded in 2001, but significantly related to in 2009, as well as a fiscal crisis in 2007. Labor markets an increased likelihood of labor being under demanded were at least somewhat flexible in absorbing workers in 2010: capital investments into these small enterprises into different sectors as macro-conditions and policy may be outpacing labor demand. responses changed. 110 Republic of Madagascar Employment and Poverty Analysis FIGURE 4.2: Sector of Primary Employment 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 2001 2005 2010 2012 2001 2005 2010 2012 2001 2005 2010 2012 2001 2005 2010 2012 2001 2005 2010 2012 Poorest Second Third Fourth Richest Services Public administration Construction Manufacturing Agriculture/primary The decade saw people shift employment out of 2005 levels for all but the richest quintile, which saw a manufacturing and services and into agriculture (in more modest decline (see figure 4.3). These shifts sug- 2005), and there was some evidence of urban-to-rural gest that the profitability and opportunities in different migration. Employment in agriculture by the coun- sectors were subject to a variety of economic and other try’s relatively non-poor increased 24.8 percent from shocks (see, for example, Belghith, Randriankolona, 2001 to 2005, while employment in manufacturing and Osborne 2016; and Thiebaud, Osborne, and Bel- and services declined by 8.9 percent and 14.9 percent, ghith 2016). respectively for the same group (Stifel, Rakotomanana, and Celada 2007). For those in the richest two income quintiles, primary employment in agriculture peaked in Data 2005, with the richest quintile making the largest jump, increasing from 32 percent employed in agriculture in The data used in this paper are from the three most 2001 to nearly 50 percent in 2005. However, by 2012, recent waves of the EPM (2001, 2005, and 2010). employment in agriculture had fallen relative to 2005, While the core modules for NFEs stayed the same nonetheless remaining higher for the middle three quin- throughout the three waves, funding and time con- tiles than its 2001 level (figure 4.2). Across all income straints prevented the fielding of a detailed agricultural quintiles, the number of people whose secondary module in the years after 2001. Consequently, this anal- employment is in services increased markedly in 2010, ysis uses the estimates from the NFE data to describe with the largest increase coming from the poorest (from the dynamics of labor-demand elasticities over time, about 10 percent to nearly 50 percent), but this second- while providing a static estimation of the shadow wage ary service sector growth was not sustained. By 2012, elasticity of demand for agricultural labor in 2001. secondary employment in services had fallen back to The EPM surveys followed a two-stage stratification Labor Demand Estimation in Rural Madagascar: Shadow Wages and Allocative Inefficiency 111 FIGURE 4.3: Sector of Secondary Employment 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 2001 2005 2010 2012 2001 2005 2010 2012 2001 2005 2010 2012 2001 2005 2010 2012 2001 2005 2010 2012 Poorest Second Third Fourth Richest Services Public administration Construction Manufacturing Agriculture/primary procedure, with the first stratification being at the As a measure of remoteness, we utilize the transportation faritany or province level, while the second divided cost of shipping a 50 kilogram bag of rice to the nearest areas within these provinces into urban areas and rural main urban center during the rainy season. This proxies areas. For further distinction, major cities (grand centres for the expense both of sourcing inputs to production, urbain) were differentiated from smaller urban centers as well as the cost of marketing any agricultural surplus, (centres urbain secondaire). The agriculture section of or, in some cases, the final product of the NFE. Any sub- the survey from the 2001 survey asks respondents about stitutes for labor in the production function, especially each plot they cultivate separately, including informa- modern ones like machinery or chemical pesticides or tion about family labor, wage labor, and animal labor fertilizers, will be imported via the nearest urban center. for each crop they report having cultivated. In each sur- The other community-level controls include an indicator vey, households reported the number of workers (both of whether the community is considered part of a zone family members and nonfamily members), equipment rouge, indicating a high level of crime and insecurity; values, and various expenses for each NFE operated by whether the community has national television and a member of the household. These data, supplemented radio coverage; and an indicator of whether the com- by household demographic information, form the basis munity has access to rural financial services. Especially of our analysis. In addition to the household survey data in this tumultuous decade, physical security was likely from the EPM, there are community-level data for each to be a major driver of labor supply movements, with round, providing information on many community- people avoiding more violent areas. Access to finance level factors. We chose those that could influence the and broadcast media indicates the extent to which a demand for labor as well as the cost of searching for commune is able to invest in their enterprises and keep and hiring labor. abreast of market conditions. 112 Republic of Madagascar Employment and Poverty Analysis Madagascar’s Agricultural be driven by complementarity of the animal traction and NPK fertilizer inputs, the interaction of which is positive Sector and significant in the farm gross revenue estimation. It seems, therefore, that while female-headed households Madagascar’s agricultural sector is characterized by the face differential levels of access to agricultural inputs dominant production of rice, the country’s staple grain, that are complements or substitutes to labor, they do not and of several nonrice food crops, grown both for home employ significantly different quantities of labor, either consumption and market sale. In addition, export crop from their own family or hired.4 production is concentrated in coastal areas and includes commodities such as vanilla, coffee, cocoa, and spices. Agricultural producers choose both the extensive and Rice production uses more labor than any other kind of intensive margin of input use from a diverse set of input production, with an average of 54.08 person-days per choices. For tractability, we chose six inputs out of this hectare; the next most labor intensive crop types are set that saw the most widespread use across the sample; export crops, which use 33.15 person-days on average. these are summarized in table 4.1, disaggregated by the Because of its labor intensity, 32 percent of plots grow- type of crop.5 Rice, the most commonly grown crop, uses ing rice used hired labor, compared to only 5 percent of significantly more labor (both family and hired) as well plots growing export crops. However, there is no signifi- as more tractor-hours than either non-rice food crops cant difference between the wages paid to hired workers or export crops. NPK fertilizer use is not common and based on the crop type. Even for the most labor-intensive does not differ across crop types. Labor, especially family production, most of the labor comes from the family, labor, is the most commonly used input, regardless of the from 84.6 percent for rice to 95.5 percent for export crop type, apart from land. crops. Across all growing types, plot ownership rates are around 90 percent. However, some factors do seem to be correlated with an increase in the amount of hired labor NFEs in Rural Madagascar used, especially education. Farm operators who have not completed secondary school—those with no education The NFEs described in this survey are small-scale busi- or some primary education (78.8 percent of operators in nesses, largely operating in the informal sector and the sample) or who completed primary but not farther characterized by their small size. Each employs only (14.7 percent)—hire 4.35 person-days of nonfamily 1.5 people on average (whether hired worker or house- labor; whereas those who have completed secondary hold member) and earn an annual revenue of MGA school (5 percent), or have post-secondary education 8.6 million, or US$1,448. There is, understandably, a (1.5 percent), hire almost exactly double that amount, good deal of heterogeneity in the operation and structure 8.68 person-days, on average, a statistically significant of these firms over the four sectors (agriculture, manu- difference. facturing, services, and trade). Over 500 unique business types are recorded over the three EPM survey rounds, Summary statistics comparing plots operated by with weavers and seamstresses being the most common. female- and male-headed households can be found Grocers and other vendors are the next most frequent. below in annex 4B. There is no significant difference in Agricultural enterprises, which grew from 4 percent of the amount of hired labor employed by female-headed enterprises in 2001 to nearly 25 percent in 2010, are the households, which constitutes 16.1 percent of the most informal (only 6 percent registered with the gov- sample. There is also no significant difference in plot ernment, compared to 25.5 percent of trade enterprises ownership rates, or even in the average area cultivated and 27.8 percent of those in services) and operate fewer by households headed by either gender. However, male- months out of the year (8.5 months) than the other headed households use significantly more rented tractor- types of enterprises (10.75 months for services, 10.64 hours and animal traction, both household-owned and for trade, and 9.31 for manufacturing). Thus, these are rented. Additionally, male-headed households purchase especially small and informal operations in an economy significantly more pesticide and apply significantly more dominated by informality. organic fertilizer, which could be a direct result of the higher amount of animal traction. Farm revenue func- Over the decade, the percentage of surveyed households tion estimates (see table 4A.4) show that this seems to operating an NFE increased, while the percentage of Labor Demand Estimation in Rural Madagascar: Shadow Wages and Allocative Inefficiency 113 TABLE 4.1: Agricultural Inputs by Crop Type (Means, 2001) Rice Nonrice food crops Export crops Area 68.89 46.30*** 66.15  (in ares) (114.6) (101.8) (89.87) Family labor 51.56 25.76*** 43.94**  (days) (79.56) (49.57) (70.40) Hired labor 15.04 7.895*** 2.986***  (days) (46.67) (40.77) (7.745) Animal, own 39.24 17.24*** 0.0485***  (hours) (403.6) (216.0) (0.547) Tractor, own 24.75 11.46 0.0544***  (hours) (436.4) (279.3) (0.515) NPK fertilizer 2.831 1.928 0***  (kg) (46.29) (66.28) (0) TLU 2.439 3.599*** 1.054*** (6.945) (10.80) (2.221) Equipment value, log 3.362 3.272*** 3.316 (1.318) (1.347) (1.170) N 3,479 2,814 423 Note: Standard deviation in parentheses. TLU = tropical livestock unit. The are is a local unit of area measurement: 1 are = .01 hectare. ***Significantly different from rice at 1%. **Significantly different from rice at 5%. *Significantly different from rice at 10% levels. these households that employed hired labor in their NFE TABLE 4.2: Household Composition Changes declined from 2001 to 2005, and then stayed relatively constant between 2005 and 2010. During the same 2001 2005 2010 time, households’ investment in NFEs, measured by the Men in household 1.078 1.107 1.174a,b (0.024) (0.014) (0.011) value of their equipment, also increased. In 2001, of the Women in household 1.018 1.297a 1.259a,b 5,080 households surveyed, 1,334 operated a NFE (26.3 (0.022) (0.012) (0.009) percent), and of those enterprises in operation, only 411 Note: Standard errors in parentheses. hired any nonfamily labor (30.8 percent of households  Significantly different from 2001. a operating a NFE). This percentage increased slightly to  Significantly different from 2005. b 30.3 percent of households operating a NFE in 2005, although fewer enterprises use hired labor in this year (14.3 percent of households operating a NFE). Finally, government. Although the number of NFEs increases in 2010, 43.9 percent of households operated a NFE, from year to year, the average number operated by a sin- although a similar percentage (16.0 percent) used hired gle household declines between 2001, 2005, and 2010, labor. The drop in the percentage of households hiring and of those, a decreasing percent respond yes to the labor could be a result of increasing household size, as question “Does the enterprise still have actual activity?” shown in table 4.2. As households in rural and second- This could indicate that more households had started ary urban areas grow, these family members can replace operating NFEs between 2005 and 2010, but that by the hired workers, and, given the increase in the number of time they were surveyed in the last round, these opera- NFEs in the sample, may start operating small enter- tions had ceased their activities. Over the same period, prises of their own. the average amount of hourly wages paid to both family and hired workers increased from MGA 8.75 and MGA Table 4.3 shows that the composition and structure 52.51 in 2005 to MGA 10.15 and 59.54 in 2010, respec- of NFEs change over time. Since NFEs were identified tively. In 2001, NFE operators paid an average wage of through household sampling, the analysis is not repre- MGA 30.00 between hired workers and family work- sentative of larger firms. Only 10 percent of these NFEs ers, whose wages were not reported separately in the in the sample, across all years, are registered with the survey. The value of equipment owned by NFEs increases 114 Republic of Madagascar Employment and Poverty Analysis TABLE 4.3: NFE Summary Statistics (by Year) 2001 2005 2010 mean mean mean Enterprise has had actual activity in the last year 0.979 0.961a 0.953a,b (1 = yes)* (0.143) (0.209) (0.212) Wage paid to household members (MGA/day) — 8.751 10.15 (161.2) (97.81) Wage paid to hired workers (MGA/day) — 52.51 59.54 (463.3) (484.9) Received financial aid 0.011 0.013 0.018a,b (0.002) (0.002) (0.002) Number of household employees 1.231 1.486a 1.634a,b (0.91) (0.931) (1.114) Number of hired employees 0.356 0.345 0.451a (1.645) (1.668) (1.442) Value of equipment (10,000 MGA) 111.8 187.5a 264.8a (6965) (2,905) (2,945) Years in operation 6.166 6.369 8.285a,b (9.495) (7.468) (8.832) Number of enterprises operated by a household 1.321 1.211a 1.118a,b (0.536) (0.448) (0.368) Agriculture enterprise (1 = yes) 0.0523 0.042 0.28a,b (0.223) (0.201) (0.449) Manufacturing enterprise (1 = yes) 0.154 0.024a 0.0422a (0.361) (0.153) (0.201) Trade enterprise (1 = yes) 0.174 0.0723a 0.342a,b (0.379) (0.259) (0.474) Services enterprise (1 = yes) 0.481 0.52a 0.336a,b (0.500) (0.500) (0.472) Monthly wages paid (MGA) 30.00 — — (189.75) N 1,568 3,333 5,783 Note: Standard deviation in parentheses.  a Significantly different from 2001. bSignificantly different from 2005. markedly as well: from MGA 1.118 million in 2001 to also face a lack of access to finance: less than 10 per- MGA 1.875 million in 2005 and MGA 2.648 million in cent of communities have access to some sort of rural 2010.6 Within a given year, there are more family work- financial institution. Access to broadcast media is more ers on average in each NFE than hired workers, and they, common, but by no means universal: more than half of unsurprisingly, receive less in wages than their nonfamily farms in the sample are in communities with no access to counterparts. national television or radio. Table 4.4 shows the summary statistics for these Traditional labor supply models make assumptions that community-level characteristics in 2001, as well as the are empirically intractable, especially in places with summary statistics for some plot-level characteristics limited formal labor markets, such as rural Madagas- that impact production but are not inputs chosen by car. As a result, Barrett, Sherlund, and Adesina (2008) the operator. Most plots in the sample face some sort of developed an extension of Jacoby (1993) in light of the disadvantage: erosion is the most common, with nearly implausibility of one of the original model’s assumptions. three-fourths of plots considered eroded. Communities As with many other early labor supply models, Jacoby Labor Demand Estimation in Rural Madagascar: Shadow Wages and Allocative Inefficiency 115 TABLE 4.4: Plot and Community Characteristics There are complications to empirically analyzing deter- for Farms (2001) minants of rural labor demand. The first is related to the structure of the data sets themselves, and the others Mean are related to the data that are not observable. First, as Plot characteristics discussed, the EPM data set is repeated cross-sections, Hillside .217 rather than panel data. This complicates matters sig-   (1 = plot is on the hillside) (.412) nificantly, as there are distinct benefits to panel data, Hilltop .116 specifically in the ability to control for unobserved   (1 = plot is on the top of a hill) (.32) heterogeneity. Eroded .732   (1 = plot is eroded) (.443) Another factor complicating demand estimation in this Sandy .146   (1 = plot soil is sandy) (.353) context is that the majority of labor demand in rural Pest .372 Madagascar comes from family enterprises, in which  (1 = plot experienced a pest attack in the last (.483) workers are not compensated with a wage but rather via year) a share of the profits or other in-kind remuneration. As a Weather .516 result, researchers either must make very strong assump-  (1 = plot experienced a weather shock in the (.5) tions, such as that those working on a family farm would last year) be paid the same as an observationally similar individual Community characteristics with a recorded wage rate. Alternatively, researchers Transport cost 10366 must impute a wage rate for these individuals. (14,015) Zone rouge (1 = yes) .154 Wages, when observed and recorded, often represent (.361) only the recorded cost of hiring a worker, although there Access to broadcast media (1 = yes) .431 are many other costs associated with the hiring and (.495) maintenance of staff. The researcher does not observe Access to finance (1 = yes) .0704 searching, hiring, monitoring, or supervision costs, or the (.256) costs of firing workers, even when wages are recorded. N 7,671 Yet employers factor these costs into hiring and wage Note: Standard deviation in parentheses. Transport costs reflect cost of decisions. In this context, such costs are likely to exhibit transporting 50 kilograms of rice to nearest urban center in the rainy season. a good deal of heterogeneity related to whether these costs are borne by an agricultural enterprise or a NFE, as well as related to location and attributes of the employer assumed the textbook equilibrium condition of MRPL = and enterprise. For example, Otsuka and Yamano (2006) w, that the market wage is equal to the marginal rev- point out, “The cost of monitoring the work efforts enue product of labor. Empirically, there are numerous of [agricultural] laborers in ecologically diverse farm reasons why this condition will be violated, including environments is exceedingly high.” As a result, labor risk, search, and enforcement costs, among many others; is often only demanded for tasks that do not require statistically, papers that use Jacoby’s method to structur- much skill or are easy to monitor, when, in the absence ally estimate labor supply routinely reject the hypothesis of these high monitoring costs, demanding additional of equality: for example, Jacoby (1993) in Peru; Bar- labor for more specialized tasks could be revenue- and rett, Sherlund, and Adesina (2008) in Côte d’Ivoire; and welfare-improving. In order to estimate labor demand Skoufias (1994) in India. The deviation between MRPL parameters consistently, therefore, we must control for and w is defined as naïve allocative inefficiency (AI). this systematic variation in the “true” shadow cost of Here, naïve reflects that this inefficiency is relative to a employing labor. naïve model where such a deviation does not exist. The existence of AI is not necessarily an indication of error Although in this context labor may move without on the part of hiring households. In light of this, Barrett, restriction between on- and off-farm employment, the Sherlund, and Adesina (2008) propose a method that two sectors may differ appreciably in terms of whether takes into account the nonobservability of wages and of they are affected by seasonality and weather shocks. allocative inefficiencies for most households. As a result, there are likely to be structural differences 116 Republic of Madagascar Employment and Poverty Analysis between in terms of labor demand patterns. To address where m* i is the equilibrium amount of labor employed this issue, we estimate labor demand in the two sectors by household i, mi is the observed level of labor separately. Because of changes in the survey, we are only employed, w* i is the shadow wage rate, which itself able to accomplish this for the 2001 round of the EPM, must be estimated, and Ai is a vector of household and as data on agricultural inputs, including labor, were not enterprise characteristics, including the characteristics collected in later survey rounds. of the community in which the household resides. We assume that ηi, the error term, is normally distributed with mean zero, as required for Tobit maximum likeli- hood estimation. Empirical Strategy The four steps required to estimate this final model are The empirical strategy we implement is designed to as follows: parameterize the household-level conditional-factor demand functions for labor in farms and NFEs, while 1. First, one estimates the enterprise production func- addressing the theoretical and data challenges described. tion and recovers the implied marginal revenue Estimation is theoretically motivated by an enterprise- product of labor. based household model, in which households choose 2. For the subsample of enterprises that pay workers a consumption of home-produced and market goods; wage, one estimates the enterprise-specific divergence labor allocation among leisure, home production, and between the observed wage rate paid to workers and wage employment for each household member; and the estimated MRPL from step 1, as a function of whether or not to hire nonfamily labor to supplement enterprise, employer, and community attributes. or replace family labor. The enterprise the household operates can be either a farm or a NFE. The household 3. One calculates the shadow wage, w*, for all enter- makes these decisions in order to maximize household prises by adjusting the estimated MRPL for the utility, subject to a budget constraint. estimated AIF from step 2. 4. One estimates the labor demand function, A household’s hiring decisions are nonseparable from equation (1). their consumption and their labor market participation decisions. This nonseparability arises because family and nonfamily labor are not perfect substitutes, due to STEP 1: ESTIMATING MRPL supervision and search costs of hired labor, risk premia, and liquidity constraints. Therefore, household demo- First, we use the entire sample to estimate stochastic rev- graphics, in addition to standard firm characteristics that enue functions for both farm and nonfarm production.7 might affect these sources of friction, must be considered The dependent variable in the first step is the annual rev- when estimating labor demand. enue (gross revenue, minus expenses and salaries paid, for each NFE operated by a household and gross agricul- Estimation follows a four-step procedure, as outlined tural revenue per plot for farm production)8 in order to in Barrett, Sherlund, and Adesina (2008). The final step aggregate across the wide variety of products produced contains the primary model of interest: household-level by households in the sample and ensure comparabil- demand for labor in rural Madagascar. Because labor ity between farms and NFEs. The regressors for farm employed is censored at zero, the main empirical model production include the quantities of the main inputs to we estimate is a censored Tobit regression: production: total labor, animal traction, tractor usage, NPK fertilizer, and land. To capture land quality, we also m* = b0 + b1w* + b 2 Ai + ηi include controls for the plot-level characteristics listed in i i table 4.1. There are fewer observed inputs for the NFEs: such that: those we use are total labor, the value of equipment, (1) m* if m* > 0 and the amount of financial aid received.9 Controls for  the number of years the NFE has been in operation, the mi =  i i   0 if m* ≤ 0, sector it operates in, and whether it is reported to have i Labor Demand Estimation in Rural Madagascar: Shadow Wages and Allocative Inefficiency 117 had “actual” activity in the past 12 months are also STEP 2: ESTIMATING THE ALLOCATIVE included as controls.10 This approach is not without its INEFFICIENCY FACTOR drawbacks, namely, the likely simultaneity of the input The naïve allocative AIF is estimated in the second step application rates, as revenue and input application rates for each household-enterprise that hires wage labor may be affected by unobserved factors that violate the ˆ L from step 1 and the in our sample.12 Using the MRP orthogonality condition for OLS estimates to be unbi- observed wage w from the data set, AIF is defined within ased. Even when using household-level fixed effects, the the subsample of household enterprises that paid work- data we have on plot-level characteristics do not capture ers by the following relationship: the diversity of agronomic conditions that farm opera- tors observe. Farmers use these conditions, such as soil  w  AIF = ln   . (4) composition and quality, drainage, slope, and location  MRP L    on the farm, when making decisions about applica- tion rates of other inputs, including labor. Also, inputs’ Since our hypothesis is that w ≤ MRPL, we expect this application responds to unobserved shocks in the error expression will be negative. A negative AIF indicates term, η, (such as pests), also violating the orthogonality that, relative to labor for market wages, on-farm (or on- condition. enterprise) labor is under demanded; the opposite holds for a positive AIF value, which indicates labor is being We estimate the production function using a generalized over demanded. Of course, an AIF value of zero means Leontief second-order flexible functional form which that the wage equals the marginal revenue product of allows for the flexible identification of complementari- labor and so there is no naïve inefficiency. Because the ties between inputs. While any second-order flexible MRPL may deviate systematically from w across the dif- function form provides an exact second-order approxi- ferent enterprises we observe based on their characteris- mation of the true, unknown function at the sample tics, especially between those engaged in agriculture and means, the generalized Leontief specification additionally those not engaged in agriculture, we attempt to identify allows for input values to be zero, as is often the case in the characteristics correlated with AIF by regressing it on this context, in contrast to the translog functional form H, a set of enterprise and operator characteristics. This (Chambers 1988). The generalized Leontief specification regression takes the following form: is as follows: AIF = a0 + a1H + m, (5) m 1 m m 1/ 2 TRk = γ 0 + ∑ γ ik x1/ ik 2 + ∑ ∑ γ ij x1/ ik x jk + ΓZk +  k , (2) 2 1/ 2 where m is a mean zero iid error term. This set of char- i =1 2 i =1 j =1 acteristics H includes demographic variables, such as the number of working-age adults of each gender and the where TR is total revenue for household-enterprise k; number of children in the household; characteristics of xi is the quantity used for each of the m inputs; Z is a the household head, such as age, education, and migrant vector of plot, enterprise, and community character- status; and community characteristics such as province- istics that directly affect production; and  is a mean level dummies, access to transportation or financial zero independent and identically distributed (iid) error services, and physical insecurity. These results provide term.11 correlates of AI, which can help us understand patterns of allocative inefficiency within subsets of the popula- Denoting labor by subscript L, the  estimated marginal tion. More crucially for this analysis, estimation of equa- ˆ revenue product of hired labor, MRP L , for each house-  tion (5) yields predictions of AI for the households that hold-enterprise can be estimated by taking the partial we do not observe hiring nonfamily labor. derivative of (2) with respect to labor, which we will call xL, where L  [1, m], using the parameter estimates, γ ˆk, of the γ0 to γm terms above: STEP 3: IMPUTING SHADOW WAGES   ∂TR TR  m  MRP L = = γˆ L + ∑ x1/ xj γ 2 1/ 2 ˆ Lj  . (3) The third step combines the estimated allocative inef-  ∂xL xL  L   j =1 ficiency AI from step 2 and the estimated MRPL from 118 Republic of Madagascar Employment and Poverty Analysis step 1 to impute shadow wages, w ˆ *, for all households positive for farms and NFEs in 2001 and NFEs in 2010. by rearranging equation (4) to estimate: However, additional labor was associated with a signifi-   ˆ ˆ cant decrease in net revenue in NFEs in 2005, a year of w* = e ∗ MRP L (6) ˆ AI serious disruption in urban labor markets. This may be This shadow wage constitutes a sufficient statistic to the result of rigidities in the level of labor employed at a address the issue of nonseparability of household pro- time of falling profitability. Other inputs that positively duction and consumption decisions (Jacoby 1993). contributed to farm revenue include NPK fertilizer usage and household plot ownership versus other forms of land rights. This indicates that the use of modern inputs STEP 4: ESTIMATING LABOR DEMAND increases farm revenue, and that land tenure and security does as well. Households may be more likely to invest in Taking the imputed shadow wage w ˆ *, we estimate labor the long-term productivity and health of their plot if they demand as in equation (1). In step 4, the dependent vari- own it. able is a household enterprise’s latent demand for labor. We bootstrap the standard errors (with 500 replications) Our analysis of NFEs shows that increases in years in of the Tobit regressions in order to mitigate the problems operation has a positive and significant relationship with produced by sequential, multistep regressions estima- NFE revenue in 2005 and 2010. A dummy variable that tions such as this.13 indicates that a household enterprise received financial aid is positively associated with revenue in 2010, indicat- The Tobit model gives only one point estimate for each ing the importance of outside sources of capital for these coefficient, and so to isolate the change in the probabil- small businesses. Such sources of financial aid include ity of using labor (the extensive margin) from the change microfinance institutions, government grants, and, most in the amount of labor used (the intensive margin), we frequently, financial support from friends and family, follow McDonald and Moffitt (1980), who were the including remittances. There are, therefore, likely impor- first to propose this decomposition as an extension of tant network effects that help determine whether an NFE the Tobit model.14 We therefore report three separate will have financial success in a given year. There are also marginal effects: important benefits, in terms of increased revenue, from ∂E (d i ) better infrastructure: increasing transport costs have a a. : change in the unconditional expectation of ∂xi negative relationship with NFE revenue (significant in all latent demand for labor years except 2010). In 2010, access to electricity and irri- gation systems (both measured at the community level, ∂E (d i |d i > 0) b. : the change in the expected level of rather than the household level) improve NFE revenue as ∂xi well. Finally, there is a significantly negative association observed use conditional on the household actually with physical insecurity and revenue in 2005 for NFEs. using labor, or the intensive margin Unsurprisingly perhaps, physical insecurity, theft, and ∂P (d i > 0) violence are bad for business. c. : the change in the probability of labor ∂xi being used, or the extensive margin The elasticities and effects on the estimated marginal revenue products of labor are shown in table 4.5. The MRP elasticity for each input (total labor for both farms and NFEs, and also land and NPK for farms) was cal- Results culated by computing the elasticity for each household and then taking the average across these values, rather PRODUCTION FUNCTION AND MARGINAL than computing the elasticity at the mean of each input REVENUE PRODUCT OF LABOR ESTIMATES variable. For farms, a 1 percent increase in labor used The full estimation results of the stochastic revenue increases revenue by 0.30 percent; for NFEs in the same function for agriculture (in 2001) appear in table 4A.4 year a 1 percent increase in labor used increases net and the results for NFEs’ net revenue for 2001, 2005, revenue by 0.18 percent. The elasticity of revenue with and 2010 appear in table 4A.5 The main input variable respect to labor use, however, declines over the next two of interest, total labor, is statistically significant and sampled years: a 1 percent increase in labor increases Labor Demand Estimation in Rural Madagascar: Shadow Wages and Allocative Inefficiency 119 TABLE 4.5: Estimated Elasticities of Revenue and Marginal Revenue Product of Labor NFE Farm Elasticities 2001 2005 2010 2001 Total labor 0.018 0.0081 a 0.009 a 0.301 (0.006) (0.002) (0.0015) (0.801) Land area — — — –0.072 (0.292) NPK — — — 0.001 (0.0039) NFE Equipment –0.078 –0.016 –0.0284 (0.011) (0.0121) (0.0041) Marginal revenue product Total labor (MGA) 60,647.2 58,321.6 27,231.7a 7,689.131 (4001.2) (1147.7) (80876.5) (1302.234) Land (MGA) — — — –30,797.18 (432,491.5) NPK (MGA) — — — 828.488 (1987.6) NFE equipment –19,651.2 –14,002.6a –7454.57a (7332.12) (9768.81) (2618.7) AIF, means –2.331 –2.567a –1.918a –1.578 (.226) (.291) (.147) (.070) AIF, medians –1.502 –2.065a –1.872a –1.573 Note: Standard deviations in parentheses. NFE = nonfarm enterprise. NPK = fertilizer. AIF = allocative inefficiency factor.  s Significantly different from 2001 value. revenue by 0.08 percent in 2005 and 0.09 percent in nonfarm enterprises in that year. But it is also signifi- 2010. In the latter half of the decade, following the two cantly higher than the average wage paid to employees political crises, the elasticity of revenue with respect to in the sample, which are lowest on average for NFE labor declines significantly. In 2001, the estimated mar- employees in 2005 and 2010 (about MGA 2,000 per ginal revenue product of labor for farms is significantly day) and highest on average (MGA 7,000 per day) for lower than for NFEs: the marginal revenue product of NFE workers in 2001, with farmworkers in that year labor for farms in 2001 is about MGA 7,000: nearly ten earning around MGA 5,000 per day. (See table 4.6 for times less than the marginal revenue product of labor for the wage data). TABLE 4.6: Estimated Shadow Wages and Observed Wages (MGA per Day) Farm, 2001 NFE, 2001 NFE, 2005 NFE, 2010 Shadow wages mean mean mean mean Nonhiring enterprise 7434.72 7478.84 3233.421a 3869.321a (2864.544) (25316.15) (5429.388) (9386.16) Hiring enterprise 7512.422** 10287.15*** 6945.05a,** 8008.15a,** (955.843) (16112.61) (12015.4) (1817.816) Observed wages 5652.614††† 6867.779††† 2667.61a,††† 2049.57a,††† (927.656) (32226.84) (5445.34) (14927.18) Note: Standard deviation in parentheses. NFE = nonfarm enterprise.  ***, **, *Significantly different from hiring at 1%, 5%, 10% levels, respectively.  †††,††,† Significantly different from MRPL at 1%, 5%, and 10% levels, respectively. aSignificantly different from 2001. 120 Republic of Madagascar Employment and Poverty Analysis ALLOCATIVE INEFFICIENCY FACTOR were more likely to have underutilized labor in this year. ESTIMATES In 2005, the opposite is true: labor is underdemanded in service sector NFEs. As for farms in 2001, labor is It is possible to test whether the textbook condition under demanded on plots growing rice and non-rice food assumed by Jacoby (1993)—that the market wage equals crops. Increased total household landholding and, once the marginal revenue product of labor—is indeed vio- again, education of the household head are associated lated in this case  by comparing the observation-specific with an increased likelihood of over demanding labor. ˆ values for MRP L estimated in step 1 with the observed Households with well-educated heads may over demand wage for enterprises that hire workers (and, importantly, labor because their educational advancement allows also record a wage).  The test is a bivariate regression of them, or perhaps even obligates them, to serve as an log wages on log MRP ˆ L , with the null hypothesis that ˆ = 1): employment safety net for their families and communi- there is no allocative inefficiency (i.e., a ˆ = 0 and b  ties. It is important to remember, however, that these ˆ ˆ L ) + e. (7) results capture associations rather than causal relation- ln(w) = a + b ln(MRP ships, and that the existence of allocative inefficiency, We reject this null for all years. This reaffirms the finding especially in such a context where the labor market is so of Barrett, Sherlund, and Adesina (2008), among others, informal, does not necessarily reflect operator error or and demonstrates the need to estimate the divergence misallocation. Instead, it reflects market-wide frictions between the observed wage and the MRP ˆ L systemati- that affect all households, with certain household char- cally. The marginal revenue product of labor must be acteristics being correlated with the increased likelihood adjusted for the unobserved costs of hiring workers to of labor being over- or undersupplied. accurately estimate demand. To make this adjustment, we regress the AIF, as calculated from equation (4), on a set of household characteristics in order to recover corre- SHADOW WAGE AND DEMAND lates to use for more accurate estimation of the shadow FOR NONFAMILY LABOR ESTIMATES wage. The full results of this estimation are found in table 4A.7. Across all four specifications, the AIF was Following equation (6), we estimate shadow wages—the negative 92.1 percent of the time, implying that the willingness to pay a wage for hiring labor from the firm’s wages paid are consistently lower than the MRP of labor perspective—for the whole sample, both  hiring and non- and that there are important nonwage costs of hiring hiring households, using the estimated MRPˆ L from step 1  labor for household enterprises. Further, a negative value and the estimated AIF from step 2. Calculated shadow of AIF indicates that labor is being underutilized relative wages, for both hiring and nonhiring farm plots, are to market wage work on these plots. found in table 4.6, in MGA per day. For all four specifi- cations, the shadow wage for the nonhiring enterprises Table 4A.7 shows that the determinants of allocative is significantly lower than for the hiring enterprises, inefficiency change from year to year for the NFEs. as expected. Among the hiring enterprises, the average In 2001, for example, labor is more likely to be over annual mean of the shadow wages ranges from MGA demanded in NFE operations where the household head 3,233 for an NFE in 2005 to MGA 7,479 for an NFE has at least some secondary school education, with in 2001. The shadow wage falls significantly for NFEs no effect in later years. The likelihood of labor being from its 2001 high in both 2005 and 2010. This could over demanded relative to its MRP also increases with reflect the existence of excess labor supply in rural areas, increased equipment in 2001 as well. In 2010, how- prompted by increased urban-to-rural migration. This ever, increased equipment value becomes significantly is supported in part by data which show the number of correlated with the likelihood that labor is being under working-age adults in the household increasing through- demanded. NFE operators may not be able to hire at a out the decade (see table 4.2). rate that keeps pace with their increasing capital invest- ments. In 2010, the year in which secondary employment Based on these reported wages and the national poverty in services increased dramatically across the income lines for each year, wages in 2005 were actually highest, distribution, enterprises in the service industry were as measured by the number of days a person would have significantly more likely to have over demanded labor, to work to meet the national poverty line. In 2001, the as were manufacturing enterprises. Trade enterprises average NFE worker would have to work 144 days; in Labor Demand Estimation in Rural Madagascar: Shadow Wages and Allocative Inefficiency 121 2005, that number decreases to 114.5 days. It peaks in employing workers were to become more costly, firms 2010, when workers would have to work 228.7 days to would respond by reducing the intensity at which their meet the national poverty line for that year. The average current employees work, rather than reducing the num- NFE hires less than half of a worker and operates nine ber of employees they have. In 2001, on the other hand, months out of the year, in a given year. Labor usage by the extensive margin dominates, meaning that changes in household NFEs or farms is therefore not high enough to demand shifters were more likely to change the probabil- be a viable route out of poverty. Labor is underutilized ity that additional labor was used, rather than increas- from a firm perspective as well, based on estimates of the ing the intensity of work for currently employed labor. allocative inefficiency. Based on the changes in the mar- Because the MRPL and the allocative efficiency both ginal revenue product of labor, these additional workers contribute to determining the shadow wage and factors benefit the firms that employ them, though at a falling that influence one over the other, it is not possible to rate over the decade: the MRPL for NFEs is less than half estimate their effects on the labor demand separately its 2001 value in 2010. It could be that these household from each other. firms first absorb the labor offered by family mem- bers, as the number of family employees increases over For farms, the most economically significant demand time. The number of hired, nonfamily employees is not shifter is the education level of the household head significantly different between 2001 and 2005, but the (table 4.7). Specifically, a household head having some 2010 level is significantly greater. Together, these trends high school increases the amount of labor demanded by indicate that labor utilization increases as excess supply 34 percent. The effect is not seen for higher levels of edu- pushes wages down. For the sake of comparison, Barrett, cation, but that could be a result of very few household Sherlund, and Adesina (2008) found observed wages for heads having completed high school or postsecondary workers on Ivorian rice plots to be 56.7 percent of the education. Labor demand also increases in the number marginal revenue product of labor there; in our case we of TLUs, which indicates that either farm animals, such find observed farmworker wages to be 73.4 percent of as larger draft animals, are complementary to labor, the MRPL and NFE worker wages to be only 11.4 per- rather than a substitute for it, or that larger herds of cent, in 2001. Thus, we see greater evidence of market smaller animals or chickens require more labor to care failures in the NFE sector, as wages there are much far- for them. Similarly, increased landholdings also increase ther from estimates of the marginal revenue product of the amount of labor demanded. Of the community labor. This suggests greater difficulty in finding appropri- characteristics, being in a zone rouge area increases labor ate labor for NFEs, compared to farms, where the work demand. This could indicate that although the cost of is likely to be less specialized. traveling to a job is higher, those households in areas with greater levels of physical insecurity may hire more Table 4A.9 presents the estimated marginal effects from labor to serve as guards for crops and livestock. the Tobit regression of demand for labor and the boot- strapped standard errors for the farms; these results for The results for NFEs show important changes in NFEs can be found in table 4A.10. Column 1 presents terms of what affects labor demand over time. In both the estimate for the unconditional demand for labor, 2001 and 2010, the number of working-age men and while columns 2 and 3 present the intensive marginal working-age women in a household is related to a and extensive marginal labor demand effect, respectively, significant increase in the amount of labor demanded, following a McDonald and Moffitt (1980) decomposi- with an insignificant coefficient in 2005. This was tion. We find that, controlling for community and house- the year in which household size was the largest and hold characteristics, the quantity of labor demanded falls allocative inefficiency was the most negative on aver- as shadow wages rise, as expected. We find that labor age, which indicates labor is being undersupplied. demand for farm work is elastic, but that the opposite The number of family members employed in this year holds for nonfarm work: NFE labor demand is inelastic, is significantly greater than in 2001: households are and becomes more inelastic over time. The intensive growing and these members are being put to work, margin dominates the extensive one for all cases except but labor remains undersupplied, meaning that labor NFEs in 2001, implying that the elasticity of labor market frictions remain costly. Because most of the demand is being driven by the intensive margin. If, there- increase in household size comes from additional adult fore, shadow wages were to suddenly to increase, and women and children, however, the new household 122 Republic of Madagascar Employment and Poverty Analysis TABLE 4.7: Farms: Estimated Demand for Labor (Select Results) Farm 2001 (2) (3) (1) Conditional Probability   Unconditional (intensive margin) (extensive margin) Shadow wage (ln MGA) –1.599*** –1.324*** –0.1024*** (0.107) (0.085) (0.01463) Some high school 0.340* 0.3230* 0.0182** (0.198) (0.1923) (0.0078) Zone rouge 0.657*** 0.6276*** 0.0325*** (0.121) (0.1198) (0.0068) TLU 0.0196*** 0.1358*** 0.0013*** (0.00709) (0.0295) (0.0005) Total land (in Ares) 0.162*** –0.444*** 0.0105*** (0.0339) (0.0785) (0.0023) Rice –0.531*** –0.808*** –0.0434*** (0.0933) (0.1301) (0.01061) Export crop –1.213*** 0.2814 –0.0374 (0.175) (0.24) (0.03169) District (fixed) effects Yes Yes Yes Constant 20.69*** (1.484) s 1.032*** (0.0334) n 2,390 2,390 2,390 Note: Standard errors in parentheses. Unconditional refers to the unconditional expectation of the observed dependent variable. TLU = tropical livestock unit. ln=natural logarithm. ***p < 0.01, **p < 0.05, *p < 0.1  members could be ill suited for the work in the Interestingly, the role played by education is not con- household’s NFEs. The additional burden of supply- sistent. In 2001, the impact of education was positive. ing for a growing household might drive many NFEs Completing primary school was associated with a out of business, reducing their quantity demanded 30.7 percent increase in the quantity of labor demanded, for labor in this year. Additionally, the number of and having completed secondary school with a 24.8 per- children becomes significantly related to quantity of cent increase, making educational attainment the most labor demanded in 2010, indicating that household economically significant demand shifter in this year. size and composition becomes increasingly important However, while the relationship was still positive in throughout the decade, perhaps as rural households 2005, it was not significant, and the relationship actually grow as a result of increased urban-to-rural migration. became negative in 2010, although the magnitude is Also different between 2005 and the other two years is small compared to 2001 and significant only for some the relationship between quantity of labor demanded high school and completed high school. This could indi- and the gender of the household head. In 2001, male cate that while educational attainment is important, it is household heads demanded less labor, while in 2005 not enough to increase the quantity of labor demanded they demanded significantly more. Here, the household in the adverse economic conditions created by the 2009 having a male head is associated with a 27 percent political crisis. In light of such instability, few household increase in the quantity of labor demanded. characteristics are important. Labor Demand Estimation in Rural Madagascar: Shadow Wages and Allocative Inefficiency 123 Another potential impact of the 2009 crisis can be seen demand is not consistent year to year, indicating that no in the coefficients for transport cost. For 2010 it is actu- one sector experienced consistent growth or decline of ally positive and significant. Thus, being in an area that labor demanded. For example, in 2001, manufacturing is more remote is actually associated with an increase enterprises demanded significantly more labor, while in the quantity of labor demanded. These remote areas in 2010 they demanded significantly less. In fact, in also were those least likely to feel the direct impact 2010, all sectors (agriculture, trade, and manufacturing) of the crisis, as they are farther from the country’s demanded significantly less labor relative to services, urban areas, where political violence was concentrated. which is consistent with the service sector growth shown The zone rouge indicator is insignificant in 2005 and in figures 4.2 and 4.3 for the latter half of the decade. 2010, showing that physical violence and insecurity may have started to be associated with lower levels of labor demanded. As with farms, physical insecurity, Discussion and Conclusion measured by this indicator, is significant and positive in 2001. Finally, the relationship of equipment value to This paper estimated which factors increased (or labor demanded changed markedly over time as well. decreased) rural labor demand for both on-farm and While increased value of equipment was associated with off-farm work by households in Madagascar, during increased quantities of labor demanded in 2001 and a politically and economically turbulent decade. This 2005, indicating that equipment is complementary to the estimation took into account the inherent nonobservabil- activities of labor or that the equipment requires work- ity of wages for most informal sector or self-employed ers to maintain and service it, the relationship becomes workers, as well as the unobserved nonwage costs of significantly negative in 2010, which is also when the hiring and employing labor, which we found to be about stock of equipment owned is also the highest. This sug- half (51.8 percent) of the total cost of demanding labor gests that by the end of the decade, capital has started in this context. These costs, which represent a variety of to become a substitute for labor in these NFEs. The unobserved labor market frictions, are significant and relationship between sector-level indicators and labor make it more difficult for households to employ workers. 124 Republic of Madagascar Employment and Poverty Analysis Indeed, the results of our analysis show that labor is For the informal enterprise sector, capital investments undersupplied on more than 90 percent of household (measured by the value of equipment) have a changing plots or in household NFEs. For all of these households, relationship with labor demand over time. While the the observed wage is consistently and significantly lower two demonstrate complementarity in the first two years than the marginal revenue product of labor. In order studied (2001 and 2005), with the value of equipment for labor to be efficiently and effectively allocated for associated with increased levels of labor demand, this both profit maximization and poverty reduction, the relationship reverses in 2010. Interestingly, more remote labor market frictions that inflate the cost of hiring and areas are associated with higher labor demand in 2010 maintaining workers for employer households should be as well, and there is no longer a detectable positive exter- reduced. Because these frictions are unobserved almost nality from physical insecurity. One resource not covered by nature, specific prescriptions for their removal are dif- in the survey, perhaps because it is unlikely to exist, is ficult to make. Nonetheless, lessons can be applied from access to business development services. Such services places where we observe well-functioning labor markets can better equip microentrepreneurs to build their capa- and work that specifically tests for the presence of labor bilities in all areas of business management, including market failures. Complete and competitive labor markets labor relations. Technical and vocational skills develop- rely on infrastructure that facilitates easy job searches, ment is another overlooked but potentially important have processes in place to write and enforce contracts, way of improving engagement between the demands of and have unrestricted worker mobility. Worker mobil- the nonfarm labor market and the skills of the popula- ity is, at least in part, a function of a stable and peaceful tion in rural areas (IFAD 2011). society, where travel and relocation are not limited by fears of violence or unrest along the route. Especially Finally, our results imply that exogenous wage growth in the specific context of Madagascar, where cultural (or wage growth due to large urban-center-based labor ties to specific places and plots of land are strong, a demand) would have markedly different effects on farm lack of worker mobility may be especially widespread versus nonfarm employment. Based on the estimated (Stifel, Fafchamps, and Minten 2011). Systems that help elasticities, wage increases would have only a small mitigate output risk, especially in agriculture, would also effect in reducing the quantity of labor demanded in help labor markets function: concerns about crop fail- the more rural NFE sector, but the opposite is true in ures or price shocks keep farmers from using the optimal the farm sector. This higher wage responsiveness could level of all inputs, including labor. reflect the thin margins typically observed in agricul- ture, whereby small increases in costs would result in Nonetheless, despite the high cost of employing work- a greater adjustment of the relevant input. However, in ers for households, we do find some evidence of which both the farm and nonfarm sectors, significant frictions factors are related to an increased quantity demanded exist. Wages paid in this environment are routinely of labor. We find that, especially for farms, educational significantly less than the MRP of labor. By identifying attainment of household heads can stimulate rural labor and reducing these labor-market frictions, which include demand, a positive externality of educational attainment. search and monitoring costs, among other unobserved Indeed, education has benefits on the supply side of the costs, the demand for labor by an individual household labor market as well, as it allows workers to access bet- will increase, especially for the more shadow wage ter opportunities and attenuate their exposure to labor responsive farm households. This hiring, in turn, is asso- market risks (IFAD 2011). The positive relationship ciated with higher levels of revenue both for farms and between land holdings and livestock holdings on quan- for NFEs, while increased employment opportunities tity of labor demanded indicates that asset accumulation, at either the extensive or intensive margin benefit wage especially productive assets like these, are beneficial for earners, especially the landless poor. Therefore, reducing the local labor market, as these results suggest the two the barriers to a more efficient labor market should be are complements rather than substitutes: having more poverty reducing for both demanders and suppliers of land or more livestock elicits a positive labor demand rural labor. response from farm households.15 Labor Demand Estimation in Rural Madagascar: Shadow Wages and Allocative Inefficiency 125 Annex 4A. Tables TABLE 4A.1: Plot and Community Characteristics for Farms (2001) Mean Plot characteristics Hillside .217   (1 = plot is on the hillside) (.412) Hilltop .116   (1 = plot is on the top of a hill) (.32) Eroded .732   (1 = plot is eroded) (.443) Sandy .146   (1 = plot soil is sandy) (.353) Pest .372  (1 = plot experienced a pest attack in the last year) (.483) Weather .516  (1 = plot experienced a weather shock in the last year) (.5) Community characteristics Transport cost 10366 (14,015) Zone rouge .154  (1 = yes) (.361) Access to broadcast media .431   (1 = yes) (.495) Access to finance .0704   (1 = yes) (.256) N 7,671 Note: Standard deviation in parentheses. Transport costs reflect cost of transporting 50 kilograms of rice to nearest urban center in the rainy season. TABLE 4A.2: Agricultural Inputs by Crop Type (Means, 2001) Rice Non-rice food crops Export crops Area 68.89 46.30*** 66.15  (ares = .01 ha) (114.6) (101.8) (89.87) Family labor 51.56 25.76*** 43.94**  (days) (79.56) (49.57) (70.40) Hired labor 15.04 7.895*** 2.986***  (days) (46.67) (40.77) (7.745) Animal, own 39.24 17.24*** 0.0485***  (hours) (403.6) (216.0) (0.547) Tractor, own 24.75 11.46 0.0544***  (hours) (436.4) (279.3) (0.515) NPK fertilizer 2.831 1.928 0***  (kg) (46.29) (66.28) (0) TLU 2.439 3.599*** 1.054*** (6.945) (10.80) (2.221) Equipment value, log 3.362 3.272*** 3.316 (1.318) (1.347) (1.170) N 3,479 2,814 423 Note: Standard deviation in parentheses. TLU = tropical livestock unit. ***, **, *Significantly different from rice at 1%, 5%, 10% levels respectively  126 Republic of Madagascar Employment and Poverty Analysis TABLE 4A.3: NFE Summary Statistics 2001 2005 2010 mean mean mean Enterprise has had actual activity in the last year 0.979 0.961a 0.953a,b   (1 = yes) (0.143) (0.209) (0.212) Wage paid to household members (MGA/day) — 8.751 10.15 (161.2) (97.81) Wage paid to hired workers (MGA/day) — 52.51 59.54 (463.3) (484.9) Received financial aid 0.011 0.013 0.018a,b (0.002) (0.002) (0.002) Number of household employees 1.231 1.486a 1.634a,b (0.91) (0.931) (1.114) Number of hired employees 0.356 0.345 0.451a (1.645) (1.668) (1.442) Value of equipment (10,000 MGA) 111.8 187.5a 264.8a (6965) (2,905) (2,945) Years in operation 6.166 6.369 8.285a,b (9.495) (7.468) (8.832) Number of enterprises operated by a household 1.321 1.211a 1.118a,b (0.536) (0.448) (0.368) Agriculture enterprise (1 = yes) 0.0523 0.042 0.28a,b (0.223) (0.201) (0.449) Manufacturing enterprise (1 = yes) 0.154 0.024a 0.0422a (0.361) (0.153) (0.201) Trade enterprise (1 = yes) 0.174 0.0723a 0.342a,b (0.379) (0.259) (0.474) Services enterprise (1 = yes) 0.481 0.52a 0.336a,b (0.500) (0.500) (0.472) Monthly wages paid3 30.00 — — (189.75) N 1,568 3,333 5,783 Note: Standard deviation in parentheses. Financial aid is the amount, in MGA, that the enterprise “benefitted from” over the past 12 months. Possible sources (from codes on the survey)  include microfinance institutions, help from parents or friends, or government grants.  Wages are not differentiated between household and nonhousehold labor in 2001.  Significantly different from 2001. bSignificantly different from 2005. a TABLE 4A.4: Farm Production Function Estimates (2001) Gross revenue Gross revenue Inputs Total labor 2 –0.0004*** Total labor (days) 0.0190*** (0.0002) (0.0031) Total labor x area 0.0002 Area (ares) –0.0525*** (0.0002) (0.0024) Total labor x NPK –0.0019 NPK (kilograms) 0.0230** (0.0013) (0.0095)5) Total labor x tractor 0.0002 Tractor, own (hours) 0.0007 (0.0011) (0.0052) Total labor x animal traction –0.0006* Animal traction, own (hours) 0.0032 (0.0004) (0.0033)5) (continued) Labor Demand Estimation in Rural Madagascar: Shadow Wages and Allocative Inefficiency 127 TABLE 4A.4: Farm Production Function Estimates (2001) (continued) Gross revenue Gross revenue Area 2 0.0013*** Plot characteristics (1 = yes) (0.00008) Plot is eroded 0.0082 Area x NPK 0.0008 (0.0100) (0.0009) Plot is sandy –0.0109 Area x tractor 0.00005 (0.0104) (0.0005) Pest attack –0.0257*** Area x animal traction 0.0005 (0.0087) (0.0003) Weather shock –0.0077 NPK2 –0.0003 (0.0085) (0.0002) Household owns plot 0.0543*** NPK x tractor –0.0032 (0.0142) (0.0103) Hillside plot –0.0041 NPK x animal traction 0.0031* (0.0107) (0.0016) Hilltop –0.0020 Tractor2 –0.0001 (0.0138) (0.0001) Community characteristics Tractor x animal traction 0.0003 (0.0002) Transport cost -0.0000004 (0.0000004) Animal traction 2 -0.000002 (0.00003) Access to broadcast media –0.0180 (0.0144) Equipment value 0.000003 (0.00001) Access to finance 0.0206 (0.0215) TLU –0.0004 (0.0005) Zone rouge –0.0092 (0.0135) Household head education (none/some primary omitted) Electricity available –0.0260* Completed primary school 0.0463* (0.0144) (0.0262) Irrigation available 0.0152 Some high school 0.0085 (0.0107) (0.0159) Household (Fixed) Effect No Completed high school 0.0161 (0.0455) Constant 3.548*** (0.0389) Post high school –0.0208* (0.0113) N 3,133 Note: Standard errors in parentheses. ***p<0.01, **p<0.05, *p<0.1. TABLE 4A.5: NFE Net Revenue Function Estimates (2001, 2005, and 2010) 2001 2005 2010   Revenue1 Revenue1 Revenue1 Inputs Total labor (person-months) 0.0584*** –0.206** 0.0805*** (0.0145) (0.0977) (0.0259) Equipment value, ln (MGA) –0.137*** -0.0537 –0.0999*** (0.0160) (0.0340) (0.0137) Total labor2 0.0127*** 0.257*** 0.0458** (0.00395) (0.0573) (0.0225) Total labor x equipment value –0.0136*** –0.0380 -0.00400 (0.00418) (0.0313) (0.00692) Equipment value2 0.0480*** 0.0662*** 0.0731*** (0.00446) (0.0111) (0.00603) (continued) 128 Republic of Madagascar Employment and Poverty Analysis TABLE 4A.5: NFE Net Revenue Function Estimates (2001, 2005, and 2010) (continued) 2001 2005 2010   Revenue1 Revenue1 Revenue1 NFE characteristics Years in operation 0.0002 0.00169* 0.002*** –0.0004 (0.000895) (0.0004) Received financial aid (1 = yes) –0.0236 0.0170 0.0760*** (0.0206) (0.0517) (0.0269) Agriculture –0.0152 0.117*** 0.00436 (0.0237) (0.0296) (0.00315) Manufacturing –0.0302** 0.0796* 0.0163* (0.0145) (0.0434) (0.00851) Trade –0.0307** 0.116*** 0.0276*** (0.0125) (0.0272) (0.00432) Services 0.00993 0.106*** — (0.0107) (0.0151) TLU 0.0006 -0.00796 — –0.0004 (0.00914) Land area –0.0001** –0.0159 — (0.00005) (0.0295) Household head education (none/some primary omitted) Completed primary 0.0283 0.00914 0.0950*** (0.0178) (0.0158) (0.0192) Some high school 0.0381*** 0.0766*** 0.108*** (0.0088) (0.0184) (0.0101) Completed high school 0.0621*** 0.0430 0.107*** (0.0129) (0.0341) (0.0234) Post high school 0.00627 — –0.0172* (0.0096) (0.00905) Community characteristics Access to finance 0.0109 0.0710*** –0.0280* (0.0281) (0.0219) (0.0165) Access to broadcast media –0.0229 0.0720 –0.00819 (0.0148) (0.0461) (0.0207) Zone rouge 0.00325 –0.077*** –0.0224 (0.0195) (0.0224) (0.0292) Transport cost –0.0001*** –0.000004* –0.000005 (0.00005) (0.000002) (0.000003) Electricity available — 0.00133 0.0444*** (0.0207) (0.0150) Irrigation available — –0.00320 0.0277** (0.0187) (0.0135) Household (fixed) Effect No No No Constant 4.078*** 2.386*** 3.145*** (0.221) (0.0756) (0.0607) Observations 1,510 1,382 5,373 Note: NFE = nonfarm enterprise. Standard errors in parentheses.***p<0.01, **p<0.05, *p<0.1 Labor Demand Estimation in Rural Madagascar: Shadow Wages and Allocative Inefficiency 129 TABLE 4A.6: Estimated Elasticities of Revenue and Marginal Revenue Product of Labor NFE Farm Elasticities 2001 2005 2010 2001 Total labor 0.018 0.0081a 0.009a 0.301 (0.006) (0.002) (0.0015) (0.801) Land area — — — –0.072 (0.292) NPK — — — 0.001 (0.0039) NFE equipment –0.078 –0.016 –0.0284 (0.011) (0.0121) (0.0041) Marginal revenue product Total labor (MGA) 60,647.2 58,321.6 27,231.7a 7,689.131 (4001.2) (1147.7) (80876.5) (1302.234) Land (MGA) — — — –30,797.18 (432,491.5) NPK (MGA) — — — 828.488 (1987.6) NFE equipment –19,651.2 –14,002.6a –7454.57a (7332.12) (9768.81) (2618.7) Allocative inefficiency, means –2.331 –2.567a –1.918a –1.578 (.226) (.291) (.147) (.070) Allocative inefficiency, medians –1.502 –2.065a –1.872a –1.573 Note: Standard deviations in parentheses. NPK = fertilizer. NFE = nonfarm enterprise. a Significantly different from 2001 value. TABLE 4A.7: Estimation of Naïve Allocative Inefficiency Factor (AIF) in Labor Hiring Decision Farms 2001 NFEs 2001 NFEs 2005 NFEs 2010   AIF AIF AIF AIF Number of working-aged men (ln) 0.0709 0.789 –0.231 0.155 (0.0702) (0.515) (0.279) (0.170) Number of working-aged women (ln) –0.0255 –0.294 –0.0845 –0.190 (0.0741) (0.495) (0.324) (0.199) Number of children (ln) –0.113 –0.433 –0.248 0.142 (0.0782) (1.147) (0.445) (0.307) Age of head (ln) 0.00474 –0.458 –0.450 0.114 (0.104) (0.584) (0.404) (0.256) Head is male –0.0238 0.236 –0.350 –0.287 (0.0855) (0.519) (0.345) (0.226) Head is a migrant 0.0514 –0.146 –0.138 — (0.125) (0.393) (0.211) Household head education (some primary/none omitted) Completed primary 0.251* 0.495 0.483 –0.276 (0.148) (0.822) (0.377) (0.357) Some high school 0.239* 0.947** 0.461 –0.0623 (0.128) (0.406) (0.369) (0.177) Completed high school 0.139 1.048** — 0.118 (0.516) (0.500) (0.283) (continued) 130 Republic of Madagascar Employment and Poverty Analysis TABLE 4A.7: Estimation of Naïve Allocative Inefficiency Factor (AIF) in Labor Hiring Decision (continued) Farms 2001 NFEs 2001 NFEs 2005 NFEs 2010   AIF AIF AIF AIF Post high school –0.0813 0.782 0.590 –0.117 (0.0643) (0.496) (0.413) (0.207) Equipment value, ln 0.0194 0.103*** –0.0134 –0.0555** (0.0207) (0.0304) (0.0370) (0.0244) TLUs 0.00390 (0.00399) Total land (ln) 0.0846*** (0.0267) Distance from plot to village (minutes walking) 0.000455 (0.00130) Rice –0.238*** (0.0621) Export crop –0.472*** (0.122) Manufacturing enterprise 0.235 0.116 0.440* (0.720) (0.508) (0.245) Trade enterprise –0.186 0.0454 –0.386** (0.448) (0.325) (0.177) Services enterprise 0.141 –0.655*** 0.672*** (0.392) (0.234) (0.205) Number of enterprises, per household 0.145 0.433** –0.264 (0.343) (0.193) (0.170) NFE has actual activity1 –0.150 0.764 –0.130 (0.849) (0.637) (0.385) Years in operation (ln) 0.202 –0.0306 0.0500 (0.157) (0.110) (0.0682) Constant 15.53*** 12.15*** –4.713*** –1.379 (0.645) (2.593) (1.724) (1.011) District Fixed Effects Yes Yes Yes Yes n 432 204 202 427 R-squared 0.644 0.220 0.119 0.111 Note: Standard errors in parentheses. NFE = nonfarm enterprise. ln=natural logarithm. *** p<0.01, ** p<0.05, * p<0.1. TABLE 4A.8: Estimated Shadow Wages and Observed Wages (MGA per Day) Farm 2001 NFE 2001 NFE 2005 NFE 2010 mean mean mean mean Shadow wages Nonhiring enterprise 7434.72 7478.84 3233.421a 3869.321a (2864.544) (25316.15) (5429.388) (9386.16) Hiring enterprise 7512.422** 10287.15*** 6945.05a,** 8008.15a,** (955.843) (16112.61) (12015.4) (1817.816) Observed wages 5652.614††† 6867.779††† 2667.61a,††† 02049.57a,††† (927.656) (32226.84) (5445.34) (14927.18) Note: Standard deviation in parentheses. NFE = nonfarm enterprise. ***, **, *Significantly different from hiring at 1%, 5%, 10% levels respectively. , , Significantly different from MRPL at 1%, 5%, and 10% levels respectively. aSignificantly different from 2001. ††† †† † Labor Demand Estimation in Rural Madagascar: Shadow Wages and Allocative Inefficiency 131 TABLE 4A.9: Farms: Estimated Demand for Labor Farm 2001 (2) (3) (1) Conditional Probability   Unconditional (intensive margin) (extensive margin) Shadow wage (log (ln) MGA) –1.599*** –1.324*** –0.1024*** (0.107) (0.085) (0.01463) Number of working-age men (ln) –0.0138 0.0283 0.0022 (0.104) (0.0997) (0.00773) Number of working-age women (ln) 0.0910 0.0818 0.0063 (0.0906) (0.0791) (0.00626) Number of children (ln) –0.136 –0.117 –0.0091 (0.107) (0.092) (0.00749) Age of head (ln) –0.164 –0.107 –0.0083 (0.157) (0.1361) (0.01077) Head is male –0.0560 –0.0561 –0.0041 (0.118) (0.1001) (0.00705) Household head education (some primary/none omitted) Completed primary 0.376 0.2814 0.01602 (0.263) (0.231) (0.0100) Some high school 0.340* 0.3230* 0.0182** (0.198) (0.1923) (0.0078) Completed high school –1.312 –0.7562 –0.1561 (0.895) (0.0877) (0.2531) Post-high school –0.157 0 –0.0103 (0.0960) (0) (0.0073) Transport cost –1.83e-06 0.1385 0 (4.13e-06) (0.0896) (0) Access to broadcast media 0.227** 0.1028 0.0106 (0.108) (0.1317) (0.0070) Access to finance 0.143 0.111 0.0072 (0.147) (0.1198) (0.0084) Zone rouge 0.657*** 0.6276*** 0.0325*** (0.121) (0.1198) (0.0068) Equipment value (ln) –0.0139 –0.0044 –0.0003 (0.0274) (0.0262) (0.0020) TLU 0.0196*** 0.1358*** 0.0013*** (0.00709) (0.0295) (0.0005) Total land (ln) 0.162*** –0.444*** 0.0105*** (0.0339) (0.0785) (0.0023) Rice –0.531*** –0.808*** –0.0434*** (0.0933) (0.1301) (0.01061) Export crop –1.213*** 0.2814 –0.0374 (0.175) (0.24) (0.03169) Province FE Yes Yes Yes Constant 20.69*** (1.484) σ 1.032*** (0.0334) n 2,390 2,390 2,390 Note: Standard errors in parentheses. Unconditional refers to the unconditional expectation of the observed dependent variable *** p<0.01, ** p<0.05, * p<0.1. 132 TABLE 4A.10: NFEs—Estimated Demand for Labor 2001 2005 2010 (2) (3) (2) (3) (2) (3) Conditional Probability Conditional Probability Conditional Probability (1) (intensive (extensive (1) (intensive (extensive (1) (intensive (extensive   Unconditional margin) margin) Unconditional margin) margin) Unconditional margin) margin) Shadow wage (n) MGA) –0.270*** –0.1040*** –0.129*** –0.0555*** –0.0267 –0.0148 –0.0190** –0.0147* –0.0039* (0.0229) (0.01758) (0.0222) (0.0149) (0.1481) (0.1748) (0.00948) (0.0078) (0.0021) Men (ln) 0.207*** 0.2140*** 0.265*** 0.407*** –0.0686 –0.0374 0.282*** 0.2320*** 0.0613*** (0.022) (0.0311) (0.0393) (0.155) (0.3770) (0.4460) (0.0265) (0.0214) (0.0071) Women (ln) 0.151** 0.1330*** 0.165*** 0.620*** 0.0400 0.0218 0.165*** 0.1361*** 0.0360*** (0.085) (0.0349) (0.0432) (0.144) (0.2333) (0.2653) (0.0286) (0.0235) (0.0069) Children (ln) 0.458 –0.0866 –0.1074 0.271 –0.1681 –0.0917 0.263*** 0.2165*** 0.0572*** (0.347) (0.0612) (0.0758) (0.201) (1.1316) (1.144) (0.0384) (0.0309) (0.0089) Age of head (ln) 0.341* –0.0454 –0.0563 0.148 –0.0069 –0.0038 0.000360 0.0003 0.0001 (0.186) (0.0317) (0.0395) (0.121) (0.0893) (0.0656) (0.0342) (0.0284) (0.0075) Head is male –0.556** –0.0792*** –0.098*** 0.270*** 0.1395 0.0705 –0.0291 –0.0237 –0.0060 (0.143) (0.0267) (0.0331) (0.101) (0.7534) (0.8994) (0.0269) (0.0220) (0.0053) Household head education Primary school 0.307*** 0.2036*** 0.219*** 0.0733 0.0500 0.0273 –0.0216 –0.0179 –0.0050 (0.054) (0.0668) (0.0607) (0.0919) (0.2843) (0.3290) (0.0577) (0.0472) (0.0137) Some high school 0.270* 0.1306*** 0.155*** 0.178 0.0398 0.0221 –0.0635** –0.0519** –0.0151** (0.131) (0.0332) (0.0377) (0.123) (0.2305) (0.2617) (0.0267) 0.0208 (0.0068) Completed high school 0.248* 0.2849*** 0.293*** 0.282 0.0850 0.0496 –0.138* –0.1086** –0.0388 (0.139) (0.0605) (0.0486) (0.248) (0.4967) (0.5488) (0.0737) (0.0548) (0.0260) High school+ 0.115 — — — — — 0.00800 — — (0.144) (0.0202) Transport cost, ln 0.000003 –0.0008 –0.0011 –0.00002 0.0672 0.0366 0.0143** 0.0118** 0.0031** (0.000006) (0.0072) (0.0089) (0.00002) (0.3714) (0.4343) (0.00609) (0.0049) (0.0013) Broadcast media –0.181 0.0485 0.0587 0.558*** 0.6683 0.3644 –0.0595 –0.0497 –0.0116 (0.152) (0.0378) (0.0445) (0.210) (3.8151) (4.392) (0.0497) (0.0440) (0.0091) Access to finance –0.539** –0.1746*** –0.226*** –0.0598 –0.0046 –0.0024 0.0171 0.0142 0.0036 (0.154) (0.0535) (0.0678) (0.441) (0.1227) (0.0715) (0.0374) (0.0317) (0.0078) Access to — — — 0.0331 0.0083 .0074 0.0332 0.0271 0.0074 electricity (0.160) (0.113) (0.0786) (0.0332) (0.0277) (0.0079) (continued) Republic of Madagascar Employment and Poverty Analysis TABLE 4A.10: NFEs—Estimated Demand for Labor (continued) 2001 2005 2010 (2) (3) (2) (3) (2) (3) Conditional Probability Conditional Probability Conditional Probability (1) (intensive (extensive (1) (intensive (extensive (1) (intensive (extensive   Unconditional margin) margin) Unconditional margin) margin) Unconditional margin) margin) Access to — — — 0.268* 0.0418 0.0374 0.00600 0.0049 0.0013 irrigation (0.138) (0.540) (0.3517) (0.0319) (0.0241) (0.0064) Zone rouge 0.240*** 0.1165** 0.1346** –0.0958 –0.0557 –0.0304 –0.0281 –0.0229 –0.0064 (0.022) (0.0574) (0.0615) (0.160) (0.3235) (0.3669) (0.0878) (0.0738) (0.0220) Equipment value, ln 0.080*** 0.0163*** 0.020*** 0.118*** 0.0522 0.0285 –0.0167*** –0.0139 –0.0037*** (0.008) (0.0020) (0.0026) (0.0262) (0.2862) (0.3403) (0.00501) (0.0041) (0.0011) TLU 0.142** 0.0050*** 0.006*** –0.0249 0.0118 0.0064 –0.00137 –0.0011 –0.0003 (0.001) (0.0015) 0.0018 (0.0518) (6.520) (3.578) (0.00203) (0.0017) (0.0005) Land area –0.0001 –0.0003 –0.0004 –0.0495 –0.1275 –0.0695 0.0001 0.00001 0.0000 (0.001) (0.0002) (0.0003) (0.101) (48.818) (26.184) (0.000959) (0.00001) (0.000) Agriculture 0.118 –0.0103 –0.0128 –0.184 — — –0.0778*** –0.063*** –0.0182*** (0.240) (0.0630) (0.0790) (0.224) (0.0232) (0.0185) 0.0060 Manufacturing –0.0810 0.0786* 0.0940* 0.894** — — –0.466*** –0.321*** –0.2130*** (0.194) (0.0446) (0.0510) (0.371) (0.0731) (0.0388) (0.0523) Trade –0.223 –0.0304 –0.0380 1.331*** — — –0.0486** –0.040** –0.0110** (0.168) (0.0357) (0.0452) (0.319) (0.0229) (0.0188) (0.0055) Services 0.056 0.0269 0.0333 0.409*** — — — — — (0.143) (0.0325) (0.0402) (0.0947) Years in operation –0.003 0.0021* 0.0026* 0.00838 0.0025 0.0014 0.00382*** 0.0031*** 0.0008*** Labor Demand Estimation in Rural Madagascar: Shadow Wages and Allocative Inefficiency (0.003) (0.0011) (0.0014) (0.00525) (0.0142) (0.0165) (0.00119) (0.0009) (0.0003) Province fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Constant 3.199*** 0.537 0.784*** (0.761) (0.429) (0.141) Sigma 0.997*** 0.413*** 0.431*** (0.0312) (0.0201) (0.0103) n 1369 1369 1369 1589 1589 1589 2,462 2,462 2,462 Note: Standard errors in parentheses. NFE = nonfarm enterprise. Unconditional refers to the unconditional expectation of the observed dependent variable. TLU = tropical livestock unit. Ln=natural logarithm. ***p<0.01, **p<0.05, *p<0.1. 133 134 Republic of Madagascar Employment and Poverty Analysis Annex 4B. Summary Statistics by Gender TABLE 4B.1: Agriculture: Plot and Community Characteristics: Male- and Female-Headed Households Male Female mean n mean n Plot Characteristics Hillside 0.173*** 8210 0.132 1821 (0.378) (0.339) Hilltop 0.0895 8210 0.0846 1821 (0.286) (0.278) Eroded 0.576*** 8210 0.487 1821 (0.494) (0.5) Sandy 0.115** 8210 0.0956 1821 (0.32) (0.294) Pest attack 0.285*** 8210 0.235 1821 (0.452) (0.424) Weather shock 0.401*** 8210 0.344 1821 (0.49) (0.475) Community Characteristics Average transport cost 11125*** 5576 12835 993 (12916) (13366) ”Zone rouge” indicator 0.178 5530 0.197 985 (0.383) (0.398) Access to broadcast media 0.518*** 5530 0.446 985 (0.5) (0.497) Access to finance 0.073 6205 0.0757 1150 (0.26) (0.265) Note: Standard deviation in parenthesis; ***, **, *Significantly different from female at 1%, 5%, 10% respectively. As shown in table 4B.1, male-headed households are experience both weather and pest shocks more fre- more likely to operate hillside plots, eroded plots, or quently than female-headed households. It does seem, sandy plots. Rather than this indicating some sort of however, that women operate plots in communities that preferential treatment for female-headed households, are more remote: the transport costs are significantly this may reflect the greater likelihood that a male- lower for male-headed households, and they are also headed household will operate any plot, regardless of its more likely to have access to broadcast media. quality or position. Similarly, male-headed households Labor Demand Estimation in Rural Madagascar: Shadow Wages and Allocative Inefficiency 135 TABLE 4B.2: Agriculture—Summary Statistics: Male- and Female-Headed Households Male Female mean n mean n Household owns plot 91.40% 8210 90.20% 1821 (0.28) (0.298) Area (hectares) 59.83 5642 56.18 1039 (107.3) (115.1) Family labor (days) 41.15 5673 35.42 1043 (71.36) (54.91) Hired labor (days) 11.49 5674 10.17 1043 (45.5) (24.78) Animal traction, own (hours) 31.61** 5674 5.48 1043 (349.9) (58.01) Animal traction, rented (hours) 299.4** 5673 635.00 1043 (4440) (6541) Tractor, own (hours) 20.83* 5674 0.16 1043 (394.2) (0.867) Tractor, rented (hours) 523.1** 5674 405.50 1043 (3726) (2717) Times weeding 1.77 5674 1.79 1043 (6.867) (4.372) NPK (kilogram) 2.58 5674 0.64 1043 (59.07) (3.987) Urea (kilogram) 6.17 5674 7.02 1043 (106.1) (70.61) Organic fertilizer (MGA) 9065*** 1211 4,831.00 172 (15919) (9840) Pesticide (MGA) 2030* 5674 864.20 1043 (21387) (8763) Agricultural equipment (MGA) 432,807.00 7698 218,879.00 1659 (8423000) (3744000) Revenue per acre 529,484.00 3635 132,433.00 684 (11760000) (484610) Revenue per month of labor 11,930,000.00 3560 2,028,000.00 681 (279000000) (5372000) Revenue per value of equipment 118.1*** 3309 1,518.00 639 (1638) (19630) Note: Standard deviation in parentheses. ***, **, *Significantly different from female at 1%, 5%, 10%, respectively. 136 Republic of Madagascar Employment and Poverty Analysis TABLE 4B.3: NFEs—Summary Statistics: Male- and Female-Headed Households 2001 2005 2010 male mean female mean male mean female mean male mean female mean Enterprise has had actual activity in the 0.979 0.981 0.956** 0.974 0.952 0.954 last year (0.145) (0.136) (0.223) (0.161) (0.213) (0.209) Salary paid to household members — — 10.74* 2.933 10.04 10.58 (184.5) (48.45) (97.16) (100.4) Daily salary paid to hired workers — — 63.58*** 19.97 67.18*** 28.66 (100s MGA) (512.4) (269.7) (526.4) (254.8) Number of household employees 1.263** 1.129 1.514*** 1.405*** 1.664 1.511 (0.922) (0.866) (0.924) (0.945) (1.145) (0.972) Number of hired employees 0.408*** 0.194 0.405*** 0.165*** 0.496 0.265 (1.788) (1.071) (1.878) (0.739) (4.907) (1.272) Value of equipment 1353000*** 376537 229.5*** 63.78 312.2*** 73.11 (7943000) (1403000) (2,411) (389.5) (3,271) (679.3) Years in operation 6.153 6.206 6.153*** 7.004 8.035*** 9.301 (9.783) (8.544) (7.094) (8.447) (8.512) (9.966) Number of enterprises operated 1.35*** 1.233 1.229*** 1.159 1.126*** 1.086 by a household (0.546) (0.493) (0.461) (0.403) (0.377) (0.324) Agriculture enterprise 0.0504 0.0582 0.0486*** 0.0225 0.285** 0.257 (0.219) (0.234) (0.215) (0.149) (0.452) (0.437) Manufacturing enterprise 0.161 0.132 0.0309*** 0.00356 0.0504*** 0.00876 (0.368) (0.339) (0.173) (0.0596) (0.219) (0.0932) Trade enterprise 0.176 0.169 0.0851*** 0.0344 0.326*** 0.406 (0.381) (0.376) (0.279) (0.182) (0.469) (0.491) Services enterprise 0.468* 0.521 0.507** 0.556 0.338 0.328 (0.499) (0.500) (0.500) (0.497) (0.473) (0.47) Monthly wages paid 400528 232474 — — — — (2147000) (2644000) N 1,190 378 2,490 843 4,641 1,142 Note: Standard deviation in parentheses. ***, **, *Significantly different from female at 1%, 5%, 10%, respectively. NOTES 5. In the production function estimation, hired labor and family labor are aggregated together to overcome issues of dimensionality in later 1. Further, the failures they find are widespread and structural in nature steps. They are presented here separately to highlight differences in and unrelated to household characteristics, including gender of the the amount used of each. household head and remoteness. 6. Summary statistics are calculated using population weights. 2. Community-level access to services including irrigation and electric- 7. We use the primal approach to estimation, as the dual approach, ity were not significant in our estimations of the marginal revenue which requires input price data and (ex ante expected) output prices, product of labor on-farm in that year, and therefore we found no is infeasible. Of course, the problem of unobserved wages (that is, the evidence of an effect on labor demand of these types of infrastructure price of labor) is exactly the problem this paper’s method is designed in these data. However, applying labor and fertilizer (NPK) both con- to overcome. tributed positively to farm plot-level revenue. The effect of individual 8. Gross revenue is not used as the dependent variable for the farm and (within-household) plot size on revenues was not significant, possibly nonfarm production function because gross revenue is not recorded because there is an inverse relationship between land quality and plot for NFEs on the survey instrument. Instead, the survey asks for size in Madagascar. revenue but defines it as total revenue minus certain expenses. Due 3. This poverty rate is for 2012 based on the World Bank’s extreme to the difficulty for respondents of performing that calculation, it is poverty line of $1.90 per day per person (2011 U.S. purchasing possible respondents actually recorded gross revenue for this ques- power parity dollars). tion. Summary statistics, which show no negative values in response 4. Unfortunately, the data do not report the activities of the laborers, to this question, support this possibility. whether they are family or not. While there is no significant differ- 9. Financial aid is the amount, in MGA, that the enterprise “benefit- ence between the quantities of labor, it is entirely possible that the ted from” over the past 12 months. Possible sources (from codes on level of expenditure on hired labor does differ, and that men and the survey) include microfinance institutions, help from parents or women hire labor to perform very different tasks. friends, or government grants. Labor Demand Estimation in Rural Madagascar: Shadow Wages and Allocative Inefficiency 137 10. This is an indicator variable where 1 indicates that the respondent Chambers, R. G. 1988. Applied Production Analysis: answered “yes” to the question “Did this enterprise have actual A Dual Approach. Cambridge, UK: Cambridge business activity in the last 12 months?” 11. Note that we generalize this functional form further by allowing University Press. inputs to enter the revenue function as polynomials (with a squared Hammermesh, D. S. 1996. Labor Demand. Princeton, term). 12. It is labeled”naïve” because the observed inefficiency does not imply NJ: Princeton University Press. an error, or waste, by the decision maker; it merely indicates diver- Horowitz, J. L. 2001. “The Bootstrap.” Handbook of gence between w and MRP ˆ . L 13. Bootstrapping can also be sued where the distribution of error terms Econometrics, 5, 3159–3228. is not known. See Horowitz (2001). IFAD (International Fund for Agricultural Development). 14. It is possible to calculate the decomposition manually, but is most easily and efficiently achieved using the Stata command “margins,” 2011. Rural Poverty Report. www.ifad.org/rpr2011. which is what we used here. Jacoby, H. G. 1993. “Shadow Wages and Peasant Family 15. There are also positive externalities from a labor demand perspective Labour Supply: An Econometric Application to the for the physical insecurity indicator, indicating that not all increased labor demand reflects improvements in overall well-being. Peruvian Sierra.” Review of Economic Studies 60(4): 903–21. McDonald, J. F., and R. A. Moffitt. 1980. “The Uses of REFERENCES Tobit Analysis.” Review of Economics and Statistics 62(2): 318–21. Barrett and Dillon. 2016. “Agricultural factor markets Otsuka, K., and T. Yamano. 2006. “The Role of Rural in Sub-Saharan Africa: An updated view with formal Labor Markets in Poverty Reduction: Evidence from tests for market failure.” Food Policy. Article in Press. Asia and East Africa.” Background paper for World Barrett, C. B., T. Reardon, and P. Webb. 2001. “Nonfarm Bank. 2008. World Development Report 2008: Income Diversification and Household Livelihood Agriculture for Development. Washington, DC: 2008. Strategies in Rural Africa: Concepts, Dynamics, and Randrianarisoa, J. -C., C. B. Barrett, and D. C. Stifel. Policy Implications.” Food Policy 26 (4): 315–31. 2009. The Demand for Hired Labor in Rural Barrett, C. B., S. M. Sherlund, and A. A. Adesina. 2008. Madagascar. Washington, DC: USAID. http://www “Shadow Wages, Allocative Inefficiency, and Labor .cfnpp.cornell.edu/images/wp234.pdf. Supply in Smallholder Agriculture.” Agricultural Skoufias, E. 1994. “Using Shadow Wages to Estimate Economics, 38 (1): 21–34. Labor Supply of Agricultural Households.” American Behrman, J. R. 1999. “Labor Markets in Developing Journal of Agricultural Economics 76 (2): 215–27. Countries.” O. Ashenfelter, D. Card (Eds.), Stifel, D., M. Fafchamps, and B. Minten. 2011. “Taboos, Handbook of Labor Economics, vol. 3B: Agriculture and Poverty.” Journal of Development 2859–2939. Studies 47 (10): 1455–1481. Belghith, N., P. Randriankolona, and T. Osborne. 2016. Stifel, D., F. H. Rakotomanana, and E. Celada. 2007. “Madagascar Poverty and Inequality Update: Recent Assessing Labor Market Conditions in Madagascar, Trends in Welfare, Employment, and Vulnerability.” 2001–2005. Washington, DC: World Bank. Washington, DC: World Bank. Thiebaud, A., T. Osborne, and N. Belghith. Benjamin, D. 1992. “Household Composition, Labor 2016. “Isolation, Crisis, and Vulnerability: A Markets, and Labor Demand: Testing for Separation Decomposition Analysis of Inequality and Deepening in Agricultural Household Models.” Econometrica Poverty in Madagascar (2005–2010). Washington, 60(2): 287-322. DC: World Bank. 138  139 CHAPTER 5 Transactions Costs, Poverty, and Low Productivity Traps: Evidence from Madagascar’s Informal Microenterprise Sector Chuqiao Bi and Theresa Osborne* June 2016 *AFR Poverty Global Practice, World Bank. The authors thank Frank Vella for valuable suggestions and guidance. Summary of Results and Policy Implications T his paper utilizes 2012 data on informal, and are just one-third those of multi-worker OOMEs. owner-occupied microenterprises (OOMEs) These returns fall below prevailing lending rates in the in Madagascar to assess the potential of these country at the time. The wage penalty, estimated as the enterprises to achieve higher incomes for their owners average returns to the owner’s time, is approximately and offer remunerative employment to workers. We 60 percent of the mean wage in the broader labor mar- note first that there is significant overlap in activities ket. Owners of multi-worker OOMEs, however, earn a for single-worker OOMEs—which employ only their premium of approximately 68 percent of the mean wage, owner—versus others, which employ family and other controlling for individual characteristics. unpaid as well as paid labor. The sectors range from logging and mining to household, transport services, and To derive policy implications requires that we attempt manufacturing. We first estimate the returns to capital to explain these findings. OOMEs perceive a lack of and labor at varying scales of operation, taking account demand for as their most immediate constraint, and the of individual differences in ability or opportunity that size of the market certainly plays a major role in reduc- may affect the decision to own such an enterprise rather ing enterprise profits. Nonetheless, it would still be more than obtaining other means of employment in Madagas- efficient for there to be fewer, larger enterprises serving car’s labor markets. the same level of demand. OOMEs could in principle increase their profitability by expanding a small amount The results show that OOMEs that have only the owner at a time, reinvesting growing profits and growing to working in them—with no family, other unpaid, or paid a more efficient and profitable scale. Such a market workers—have significantly lower returns to capital restructuring would expand incomes and thereby mar- and to the owner’s labor, controlling for worker char- kets, given the importance of the OOME sector to the acteristics. Moreover, we find evidence of profitability overall economy. Yet this process does not take place. increasing with scale, but these potential gains are not exploited. Returns to capital are increasing, and there is We propose a simple theoretical model to explain the significant underemployment of workers relative to the persistence of low-productivity OOMEs. In a dynamic, profit-maximizing level. The returns to capital for single- general equilibrium framework, a combination of worker OOMEs, which comprise 70 percent of OOMEs, conditions is needed. One of these is a lack of entry average approximately 12 percent per annum (nominal) by larger, more efficient, typically formal firms, which 140 Republic of Madagascar Employment and Poverty Analysis would provoke a restructuring of the market and draw appear to be primarily to be able to serve larger cus- workers into more remunerative work. In addition, tomers, but this also depends on entrepreneurial ability, in a poor economy like Madagascar’s, the very nature and only so many enterprises can serve these limited of OOMEs inhibits their growth: First, the marginal markets. The second implication is that while very small utility of consumption for poor entrepreneurs is high, loans to the tiniest single-worker OOMEs may help to and in the presence of increasing returns, the returns to provide employment for the poor, they would have little small incremental investments for the smallest OOMEs impact on overall productivity, employment, and wage are low. Thus, entrepreneurs typically consume all the growth. firm’s income. Breaking out of this low-productivity equilibrium (or “trap”) would require a more substan- Ultimately, significantly reducing the misallocation tial increase in scale than poor households can afford. of capital and labor in Madagascar’s economy would However, because of the difficulties associated with require a steadily growing presence of larger, more monitoring the use of firm resources, whether by credi- formal firms that compete for markets. These firms tors or potential partners, the transaction costs of expan- are essential to creating more productive employment, sion through external financing are high. Credit costs stimulating greater demand, integrating Madagascar’s more to cover the higher enforcement risks and small internal markets, and better accessing export markets. loan sizes, and partnerships are much less likely to form Achieving such a transformation will require allevia- precisely because entrepreneurs’ level of investible capital tion of the constraints that such firms face to entering, is low relative to these transactions costs. Similar issues growing, and competing on an even playing field, while of information asymmetries and incentives (“transac- earning acceptable (risk-adjusted) returns. Therefore, an tions costs”) likely hinder more optimal levels of labor obvious policy implication is to attempt to alleviate the employment as well. constraints to formal firm investment. At the same time, some measure of investment and productivity growth The policy implications of this paper are tentative but appears feasible through the development of OOMEs are as follows: First, it would likely have little effect to that have already achieved a certain scale, or for those encourage firms to simply register without taking addi- with demonstrable entrepreneurial skills. Given the pres- tional steps to improve the credibility of their financial ence of increasing returns, such firms could invest more statements. This and other means to strengthen the and hire more workers if they had access to financing. information environment would be needed to reduce Therefore, a fourth policy implication is to intensify the transactions costs for potential creditors and part- efforts to improve the monitoring and enforcement ners. In the current context, the benefits of registration environment for creditors, partners, and associations. In Transactions Costs, Poverty, and Low Productivity Traps: Evidence from Madagascar’s Informal Microenterprise Sector  141 2012, credit registries and bureaus were almost nonex- (GDP) per person employed—a broad measure of labor istent in Madagascar, and the strength of legal rights to productivity—has fallen in real terms and is now the enforce repayment rated only 2 out of a 10-point scale in second lowest in the world (among countries with data) Doing Business’s Getting Credit, although these indica- after the Democratic Republic of Congo (see figure 5.1). tors improved by 2016. Finally, there is an argument for This is partly because Madagascar’s labor force is more stimulating more rapid evolution of credit markets to concentrated in agriculture, a sector which exhibits serve the higher potential OOMEs. In order to sustain particularly poor low labor productivity: 73.2 percent of impacts of any such program, it would be important household heads claim agriculture as their main sector of to ensure that (1) the monitoring environment was employment, and 83 percent in the bottom 80 percent of also improving, (2) that access to such resources was the consumption distribution.1 competitive and fairly distributed between and among individual and group enterprises, and (3) that subsidies For the 26 percent of the population not primarily did not undercut other developments in credit markets. employed in agriculture, the vast majority are employed Without a broader expansion of demand for goods and in informal microenterprises. In 2012, over 87 percent services, as some firms grew, others may be edged out of of employed workers worked in enterprises with five the market, and thus it is difficult to predict the direction or fewer workers (ENEMPSI 2012), and 80 percent of change in employment levels without better under- of nonfarm employees did so (figure 5.2). In contrast, standing the labor market frictions identified in this only 4.5 percent of nonagricultural workers worked in paper. Therefore, further investigation into the sources establishments with 100 or more workers. In addition, and possible means to address these frictions is needed, approximately 75 percent of nonfarm jobs in the country as is robust evaluation of the impacts of any credit sup- were informal (INSTAT 2013), as were approximately port programs. 90 percent of new jobs created in 2010. The formal private sector accounts for very little employment, signaling severe constraints on the growth of this sec- Introduction tor: Only an estimated 3.9 percent of nonfarm workers were employed by the formal private sector, and only Madagascar’s high rates of poverty, reaching over 11 percent of employed workers receive a wage (INSTAT 70 percent of the population, are closely associated with 2013).2 Although Madagascar is not alone in its preva- its inability to create and sustain productive employment lence of small, unproductive firms, it appears to be an for its workforce. Since 2001, gross domestic product extreme case. FIGURE 5.1: GDP per Person Employed, Madagascar and Comparators, 2001–2014 (1990 PPP U.S. Dollars) 4000 3500 3000 2500 2000 1500 1000 500 0 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 Cameroon Burkina Faso Madagascar Tanzania Congo, Dem. Rep. Kenya Zambia Uganda Niger Ethiopia Source: World Development Indicators (WDI). Note: PPP = purchasing power parity. 142 Republic of Madagascar Employment and Poverty Analysis FIGURE 5.2: Types of Workers and Share of Employment by Firm Size 100% 35% 30% 80% Share of employment (%) 25% 60% 20% 15% 40% 10% 20% 5% 0% 0% Agriculture Nonagriculture 1 2 3–5 6–20 21–100 >100 Self-employed Non-qualified worker High/middle manager Firm size Family aid Qualified worker Source: Sy and Diatta (2014) using data from the LFS 2012. Note: The panel b is restricted to the nonagricultural sector and excludes public administration and public companies. Although employment in OOMEs is a potential pathway TABLE 5.1: Reasons for Operating an OOME out of poverty for some, a high concentration of labor (Percentage by Category) in such enterprises is also associated with the economy’s low productivity.3 Owner operation and microscale are Single worker Multiworker All also strongly associated with informality—the partial or Did not find paid work 6.9 4.8 6.3 at large company limited use of formal accounting practices, registration, Did not find paid work 17.9 12.6 16.4 and compliance with regulatory and tax requirements. at small business Employment in these enterprises pays little relative to To get a better income 44.4 49.9 46.0 that in larger, more formal businesses: according to Mad- To be independent 17.0 20.3 17.9 agascar’s National Institute of Statistics (INSTAT 2013), Family tradition 13.8 12.4 13.4 84,000 versus 184,000 ariary per month. As a result, 8 out of 10 unemployed workers seek wage employment Source: ENEMPSI 2012. rather than trying to create a business, and only 8 percent prefer to work for themselves (12 percent are indifferent). Although some existing microentrepreneurs choose to profits of key characteristics of the entrepreneur or firm, work for themselves or to carry on a traditional family as well as indicators of levels of competition. We then business, according to the 2012 labor force survey (LFS, estimate the effects of this mode of employment on the conducted by INSTAT), 22.7 percent chose their activity owner’s labor earnings. In all cases, we take into account because they were unable to find a sufficiently remunera- the possible estimation bias associated with the sorting tive job. Nonetheless, 46 percent claimed to do so as a or selection of workers into microentrepreneurship. Such way to increase their incomes (table 5.1). individuals may be more capable entrepreneurs and/or less employable in the formal sector, may have greater This paper explores the productivity and income impli- familiarity with business, or may have differential oppor- cations of there being such a high level of employment of tunities that are not observed in the data—all sources of the country’s economic resources (capital and labor) in “unobserved heterogeneity” which can bias estimates. OOMEs. To develop an empirical understanding of the issues, we estimate the contributions to firm profits of We find that the vast majority of OOMEs operate capital and labor, as well as how returns change with the substantially below efficient scale, causing a sizeable scale of operation. In addition, we estimate the effects on inefficiency in the allocation of both capital and labor Transactions Costs, Poverty, and Low Productivity Traps: Evidence from Madagascar’s Informal Microenterprise Sector  143 in the economy. First, our estimates show a significant more likely to own a single-worker one, and in turn, in difference in returns to both capital and labor between the case of single-worker OOMEs, earn lower profits OOMEs that utilize only the owner’s labor (henceforth than their male counterparts. “single-worker” OOMEs) and those that also have other workers (“multi-worker” OOMEs), whether those work- We offer a simple theoretical explanation for the persis- ers are paid or not. Returns to capital by single-worker tent productivity losses observed as a result of subop- OOMEs are below the cost of capital in the economy timal scale. Our explanation relies on four conditions: and below those of multi-worker OOMEs, which in con- First, an important general equilibrium condition outside trast earn acceptable returns. Returns to the enterprise of our model must prevent the entry of more efficient head’s labor effort in single-worker OOMEs are lower (typically larger) firms that would compete aggressively than labor incomes earned elsewhere in the economy. with OOMEs and spur Schumpeterian growth, as in this Accounting for observed and unobserved differences case OOMEs could not compete and would eventu- in workers, this mode of employment reduces labor ally exit (Aghion, Akcigit, and Howitt 2013). We take earnings by an average of 60,000 ariary (approximately this condition as given in Madagascar. There is a low US$30) per month, relative to a mean monthly employ- reported level of formal firm activity, and explaining ment income in urban areas of 101,000 ariary. Yet the this, however critical, is beyond the scope of this paper. returns to the labor of owners of multi-worker OOMEs Rather, our model focuses on the partial equilibrium is on average 68,000 ariary more than it would be in the constraints to the growth of OOMEs, incorporating the labor market for equivalent workers. remaining conditions. The second condition is that the initial wealth of entrepreneurs must be low. Third, credit Second, we find evidence of increasing returns to capital markets must be inadequate to serve a significant frac- within the population of OOMEs and of increasing returns tion of (the most able) entrepreneurs. Third, and relat- to scale in much of the relevant range for most OOMEs. edly, the costs of mitigating information asymmetries and As measured, these returns to scale can accrue either of enforcing agreements among potential investors must through declining average costs or through improved be high. Although there are no data available to provide pricing. Those single-worker OOMEs which register their a formal test of the theory, we provide some descriptive businesses experience a significant increase in profitability, evidence and suggest further avenues of research. but there is no statistically significant effect of registration on the profits of multi-worker OOMEs, once firm and owner characteristics are taken into account. Firm Size Productivity Relationships In addition, we find evidence of that family history or wealth and gender are important determinants of The advantages of greater firm size could differ across firm entry and size. Enterprises’ current asset levels countries at any time by the countries’ areas of compara- are significantly related to having a traditional (prior) tive advantage, technological readiness, institutional family business, to the mother’s level of education, and arrangements, policy context, and market access. None- to the father’s employment history—factors that are theless, a priori, the productivity losses for Madagascar likely correlated with the availability of start-up capital. of the prevalence of OOMEs appear likely to be large. In addition, although the level of enterprise assets is not associated with the gender of the head in the full Internationally, countries with a higher GDP per capita sample, male ownership increases the profitability of tend to have larger firms, fewer microenterprises, single-worker firms (although there is no evidence that and more individuals employed in large firms. See, this is the case for multi-worker firms), all else equal. for example, Poschke (2014). Based upon the most Having more education reduces the likelihood of owning recent comprehensive data on manufacturing firms in and operating any OOME but positively affects returns 124 countries, average establishment size is strongly once one does. However, these returns are lower than correlated with GDP per capita—with an elasticity of the returns to schooling in the labor market generally. 0.26 (Bento and Restuccia 2014). Similarly, using data Finally, there is evidence of disadvantageous gender- from 47 developing countries, Ayyagari, Demirgüç-Kunt, based sorting into less profitable activities: females are and Maksimovic (2011) find that large firms are more less likely than men to own a multi-worker OOME and innovative and productive than small firms.4 144 Republic of Madagascar Employment and Poverty Analysis Moreover, studies of micro- and small enterprises sug- The reasons for the persistent prevalence in poor coun- gest that informal microenterprises possess little growth tries of informal microenterprises (which do not grow potential. In a study of seven Sub-Saharan African (SSA) into more productive small, medium, or large firms) are countries, Van Biesebroeck (2005) finds that it is rare not fully understood, and they may be country specific. for micro- and small firms in these economies to reach Grimm, Kruger, and Lay (2011), for example, find that medium or large scale, and that, as in countries outside returns to capital at low levels of operation are high in the region, large firms are the most productive. These a sample of seven West African countries (with some firms do not tend to grow and employ more workers over exceptions), and therefore microentrepreneurs could time (La Porta and Shleifer 2014). Thus in contrast to use internally generated resources to grow, as McKenzie the case of more developed economies such as the United and Woodruff (2006) point out. A full explanation must States, firms in SSA which start out small are unlikely to therefore include impediments to the entry of larger, contribute substantially to long run productivity growth. more productive firms as well as to the growth of micro and small ones. If they were more efficient, large firms The predominance of OOMEs—or the lack of larger, would typically drive a significant share of their smaller, more formal firms—also adversely affects wages and less-efficient counterparts out of business while also job creation. Although in many countries it is believed absorbing the labor released as those firms exit. Barri- that smaller firms create more jobs, in some countries ers to this process can be related to many factors, which they are also disproportionately responsible for even require a country-specific diagnosis in order to assess the more job destruction. (For the case in the United States, contribution of causes, such as a lack of infrastructure, see Neumark, Wall, and Zhang 2009) and Li and Rama an unfavorable investment climate, trade barriers, or (2015).5 Moreover, a significant body of literature finds difficulties in accessing key inputs such as financial and that laborers in informal enterprises earn significantly human capital. Similarly, the constraints that are most lower wages than those employed in formal enterprises, binding for OOME growth may be country specific.7 To after controlling for worker characteristics (Montenegro shed light on these questions in the case of Madagascar, and Patrinos 2014). In Madagascar, only 30 percent of we therefore focus on the returns to the microentrepre- OOMEs employ workers besides their owners, and only neur, to his or her decision to hire (more) workers, to 8 percent employ paid laborers. On average, they employ invest in the enterprise, and to apply his or her own time 1.4 workers: 1.0 is the owner, 0.3 are unpaid workers, and effort to the enterprise. and 0.1 are paid (authors’ calculations using ENEMPSI 2012). Given the observed patterns, there are serious doubts regarding the potential of OOMEs to alleviate Data and Characteristics of poverty and improve productivity. Madagascar’s OOME Sector Nonetheless, a priori it is not clear what the economic We utilize data from the Enquête nationale sur l’emploi losses are, what blocks OOME growth, and what the et le secteur informel (ENEMPSI) that was collected policy levers are for poverty reduction in countries where in 2012 in two phases, consisting in the first phase of they predominate. Microenterprises may be an efficient an LFS, and in the second phase a survey of informal outcome for some activities and markets. The products enterprises identified through the LFS. For the LFS, a and services they produce may differ from those produced stratified random sample was drawn of over 11,000 by larger firms in ways that are more in line with market households from which 41,000 individuals ages five demand.6 Larger firms selling close substitutes may serve and over were surveyed. From this sample, a listing of a segment of the market, while OOMEs serve another. individuals reporting that they owned and operated an Entrepreneurs may be more productive given their prefer- informal enterprise was produced, and a representative ences and the incentive advantages of working for them- sample was drawn of such enterprises for the second selves. In larger organizations, the costs of monitoring phase. For the purposes of the survey, informality was workers may be high. For a variety of reasons, a disper- defined as not having a statistical number or, in the sion of firm sizes would be expected, and some microen- case of people working “on their own account,” as not terprises may be efficient in some contexts. Yet if any of keeping financial accounts (INSTAT 2013). Only urban- these circumstances make OOMEs an optimal outcome, based enterprises were included, whether from large or then they should show similar productivity as larger firms. secondary cities of the country. Transactions Costs, Poverty, and Low Productivity Traps: Evidence from Madagascar’s Informal Microenterprise Sector  145 The resulting sample was of 5,692 enterprises, of which to wireless telecommunications, veterinary services, over 3,968 are owned and operated by a single person footwear manufacturing, real estate, photography, and with no other workers (single-worker/owner); and 1,724 automobile rental. of which had additional employees, whether unpaid, paid, or both (multi-worker). The activities performed There is a significant gender dimension to the types of by these enterprises can be classified into the following activities pursued. Female-owned businesses are more broad sectors: the primary sector (primarily forestry and likely to be engaged in retail, spinning, and textile forest-related products); industry (manufacturing, con- activities, and male-owned businesses are more likely to struction, mining and quarrying); trade and commerce be engaged in construction, transport, lumbering, and (wholesale and retail); and all other services (trans- metallurgy. Although the gender of the entrepreneur portation, hotels and food service, household services, does not change the probability of being employed in information and communication, real estate, professional industry, being male increases the probability of heading services, and others). Table 5.2 shows the distribution by an OOME in the primary sector by 86 percent relative to gender of single- versus multi-worker OOMEs. Table 5.3 the probability for females, and the other services sector provides breakdown of the sample by type and sector. by 12 percent, but decreases the probability of being in the trade sector by 26 percent.8 The activities represented in the sample do not appear at first glance to be those often efficiently done by individu- In Madagascar, the majority of OOMEs operate at a als working on their own with little capital. Table 5.4 very small scale. The level of capital invested is low, but shows the most frequent specific activities performed, by rarely zero: only 1.49 percent of informal firms claim to gender and OOME type. As shown, there is significant have zero assets—1.6 percent of single-worker OOMEs overlap in activities performed by both single- and multi- and 1.1 percent of multi-worker OOMEs. Mean capital worker firms, with both engaged in various types of invested in each firm was approximately 736,000 ariary, retail, household services, and mining. In addition to the or US$335.9 The means and median levels of assets are top 11 most frequent activities shown, OOMEs engage as shown in table 5.5. Assets are lower for single-worker in activities ranging from the sale of motor vehicles OOMEs on average, as well as for female-owned ones, but there is overlap in the distributions. In addition to low levels of capital, the level of employ- TABLE 5.2: Gender-type Distribution of OOME’s ment generated per OOME is very low. Over 70 percent of OOMEs employ only their owners and no other Of which: Male Female workers, paid or unpaid. Almost all (99.5 percent) Single worker 63.6 76.0 employ five or fewer employees, and these proportions Multi-worker 36.4 24.0 are very similar to what they were when those firms All 44.9 55.1 started their businesses: 99.5 percent had fewer than five Source: ENEMPSI 2012. employees, and 76.0 percent claim to have started with TABLE 5.3: Sample of Informal Microenterprises ENEMPSI 2012 (Not Population Weighted) Single-worker/owner Multiple worker Sectors represented Registered* Not registered Registered Not registered Total All 353 3,615 382 1,342 5,692 Industry 26 1,444 58 648 2,176 Primary sector 3 106 7 62 178 Services (except trade) 144 792 91 197 1,224 Trade/commerce 179 1,261 226 438 2,104  The enterprise counts as registered if it is registered with the commerce department, has a license, has a carte professionnelle, or is registered with * the social security administration. 146 Republic of Madagascar Employment and Poverty Analysis TABLE 5.4: Most Frequent Activities by OOME Type and Gender (Percent) Male-owned single worker Female-owned single worker Male-owned single worker Female-owned multi-worker Building construction (10.7) Spinning, weaving, finishing Mining of uranium and thorium Retail sale in nonspecialized textiles (31.9) ores (15.1) stores (23.8) Other land transport (7.9) Retail sales via stalls and Building construction (12.1) Spinning, weaving, and finishing markets (11.1) textiles (12.9) Mining of uranium and thorium Retail sale of food, beverages, Retail sale in nonspecialized Restaurants and mobile food ores (7.8) tobacco (9.3) stores (8.5) services (10.7) Retail sale of food, beverages Retail sale in nonspecialized Lumbering (6.2) Retail sale of food, beverages, and tobacco (6.8) stores (8.6) and tobacco (8.5) Retail sale via stalls and Household services (7.2) Retail sale via stalls and Retail sale via stalls and markets (6.8) markets (5.9) markets (7.5) Retail trade not in stores, stalls, Manufacture of food Other land transport (5.4) Manufacture of other food or markets (3.5) products (4.1) products (6.4) Manufacturing of nonmetallic Manufacturing except fur Retail sale of food, beverages, Mining of uranium and thorium mineral products (3.0) apparel (3.5) and tobacco (5.3) ores (4.4) Supporting transport Retail of other goods in Retail sale of other goods in Manufacturing except fur activities (2.8) specialized stores (3.1) specialized shops (4.1) apparel (2.9) Other personal services (2.6) Restaurants and mobile food Manufacture of nonmetallic Household services (2.9) services (2.9) mineral products (3.9) Retail sale of other goods in Other personal services (2.6) Manufacture of other fabricated Retail sale of other goods in specialized shops (2.6) metal products (2.6) specialized shops (2.1) Wholesale of agricultural raw Mining of uranium and thorium Extraction of sand stones and Wholesale of agricultural raw materials (2.5) ores (2.5) clay (2.4) materials (1.6) Source: ENEMPSI 2012. TABLE 5.5: Asset Values by Single- and Multi-worker OOMEs and Gender of Owner (1,000 Ariary) Mean Median Gender of owner Male Female All Male Female All Single worker 608 481 532 50 43 46 Multi-worker 1,344 1,027 1,200 103 113 109 just the one owner/worker. Although 8.0 percent of busi- The Impact of Scale: Estimation nesses were created in the past year, most (53.6 percent) were created at least five years earlier, and 34.3 percent Method and Results at least 10 years earlier. Some OOMEs employ unpaid To understand the economic effects of the OOMEs’ workers, who are typically family members. The level scale of operation, we estimate the relationship between of employment of unpaid and workers averaged across firms’ monthly cash profits, denoted p, and their use of OOMEs was 35 hours and 24 hours per month, respec- inputs.10 We estimate these relationships for all OOMEs, tively. Multi-worker firms utilized on average 118 hours as well as separately for multi-worker firms and single- per month of unpaid labor and 81 hours per month worker firms, given that single- and multi-worker firms of paid labor. Yet average hours per paid worker were may be fundamentally different in some way. Indexing close to or exceeded full time (calculated as 40 hours per firms by i, we estimate the following equation: week, or 172 hours per month): Paid workers worked on average 1660 per month, and unpaid workers only ˆ iν + ε i , p i = a + bf (ki ) + γ (g(hi )) + q(Xi ) + ξη (1) 22.6 hours per month. Transactions Costs, Poverty, and Low Productivity Traps: Evidence from Madagascar’s Informal Microenterprise Sector  147 Where f(k)is a polynomial function of k, capital invested (i) owning and operating a single-worker OOME, (ii) a (or assets), and g(h)represents a function of labor hours multi-worker OOME, or no OOME, using multinomial used. We distinguish between hours of work by the probit estimation as shown: owner-operator, or “head,” by unpaid workers, and by paid workers, which combined are represented by the pr(y j = ν) = Φ(z jν Γ). (2) vector hi. We estimate the functional forms for f(k) and g(h) by testing the significance levels of higher order Φ( ) represents the standard normal cumulative distribu- terms and dropping those which are not statistically tion function and the error term, η, is distributed accord- significantly different from zero.11 This approach has ing to the standard normal. The vector z includes a set the advantage of allowing some flexibility of functional of individual and spatial characteristics: the individual’s form, including nonconstant elasticities of profits with level of education, gender, region, size of city (large or respect to inputs, and in contrast to a log specification, secondary), and a set of family history variables: the of allowing some variables to be less than or equal to years of education of the mother, years of education of zero. We also include Xi, a set of conditioning variables the father, the father’s job status, and the father’s previ- which could affect profitability, including indicator vari- ous sector of employment. In all cases, there is a subset ables for the region of the country, size of urban location ˜ j, ∈ zj among the family history variables which are z (major urban versus secondary urban), and the gender of excluded from the main equation (1). We then estimate the owner-operator. We estimate the relationships with equation (1) using an approximation of η, the general- this parsimonious set of Xi variables, and with a more ized residual from equation (2) following Das, Newey, comprehensive set of firm characteristics, which includes and Vella (2003). In practice, we use a polynomial in the years of schooling of the head of the enterprise, the the predicted probabilities of each outcome, and we use reason the firm gives for pursuing the particular activity, cross validation techniques to select the preferred specifi- its broad sector (industry, trade, services), the age of cation.13 In cases where there is no appreciable selection the firm, and the age of the owner-operator. In addi- bias in comparison with the ordinary least squares (OLS) tion, since equation (1) is not a production function, but estimates, and where the polynomial approximation rather captures the combined effects of scale through terms were not jointly significant, we prefer OLS for both average costs and revenues or prices charged on efficiency reasons.14 profits, we attempt to control for demand-side and pric- ing issues by also including the main type of competitor The results from equation (2), which provides an (large versus small, commercial versus noncommercial).12 estimate of the main factors affecting the probability of owning an OOME are summarized in table 5.6. We find The term ξη ˆ iν represents an estimate of unobserved that having a father who was a salaried manual laborer determinants of individual i’s decision to own and or who worked in a nonagricultural sector is predic- operate an OOME, which may be correlated with tive of owning some type of OOME; that having more included regressors and pi. The remaining errors term ei education makes owning such an enterprise less likely is independent and identically distributed. If laborers in and being in a large urban center more likely. Moreover, the economy sort into sectors, firm types, and types of being male is associated with owning a multi-worker employment by educational attainment and ability, more OOME, whereas being female is associated with owning (or less) capable workers may sort into OOMEs rather a single-worker OOME. than other means of employment. Alternatively, less capable workers may sort into owning OOMEs because The estimates of equation (1), shown in table 5.7, pro- they have more difficulty retaining salaried employment. vide the following key insights into the issue of (profit- In either case, unobserved heterogeneity in ability, oppor- ability) returns to scale. First, the structure of returns tunities, or other characteristics of the entrepreneur may differs substantially between single-worker and multi- be correlated with thus biasing the coefficient on the worker OOMEs.15 The returns to both capital and the owners’ time, assets, or other variables. We attempt to head’s labor are significantly higher for multiple-worker address this (selection bias) issue through a two-step esti- firms, even when firm characteristics such as sector, edu- mation strategy which incorporates a multinomial choice cational attainment, location, and unobservable factors model in the first step. In particular, we use the LFS to are taken into account. Moreover, returns to capital are estimate the probability of an individual j’s choice, v, of low and diminishing for single-worker OOMEs, on an 148 Republic of Madagascar Employment and Poverty Analysis TABLE 5.6: Significant Influences on Likelihood of Owning and Operating an OOME Single worker OOME Multi-worker OOME Father salaried manual laborer 0.447* 0.253 (2.34) (1.07) Father worked in agriculture –0.385*** –0.279*** (–6.30) (–3.68) Father worked in trade 0.124 0.269** (1.78) (3.19) Male –0.157*** 0.192*** “= 1 male, = 0, female” (–4.97) (4.97) Age 0.0180*** 0.0208*** (16.53) (15.81) Years of education –0.0471*** –0.0336*** (–8.76) (–5.16) Large urban 0.153*** 0.188*** (4.26) (4.27) N 20,065  p-value = .10; **p-value = .05; ***p-value=.01; t-statistics in parentheses. * Note: Includes regional indicators and other variables. See annex table 5A.1 for all details. TABLE 5.7: Estimated Monthly Returns to Firm Inputs No firm characteristics With firm characteristics Single worker Multiple worker Single worker Multiple worker All Assets 0.0130*** 0.0330*** 0.0104** 0.0384*** 0.0191*** Assets2 –1.37xe–10** ~ –.000000107* ~ .000000134*** head’s hours 1.026*** 1.868** 0.743*** 1.893** 1.057*** head’s hours2 –0.00120* –0.00355* ~ –0.00363* –0.00169** Unpaid hours 0.985*** 0.325** 0.356*** Unpaid hours2 –0.000848* ~ ~ Paid hours 0.494*** 0.376*** 0.490*** Paid hours2 –.000266*** –0.000289*** –0.000275*** Male = 1 58.64*** ~ 83.07*** ~ 63.72*** Years schooling 11.49*** ~ 10.69*** Returns to additional year of head’s schooling 6.84% ~ 4.90% N 2,332 1,210 2,775 1,355 4,125  p-value = .10; ** p-value = .05; *** p-value=.01. ~ = not statistically significant. * Note: Includes regional indicators and other variables. Details are reported in annex table 5A.2. annualized basis averaging 4.5 percent (controlling for These estimates indicate that the returns to capital for firm characteristics), but not for multi-worker OOMEs. single-worker OOMEs fall well below the opportunity Similarly, in the pooled sample, returns to capital are cost of capital in the economy. First, they are lower increasing, as shown by the positive coefficient on assets than returns in slightly larger OOMEs, which have an squared, in part due to the increased returns associated annualized return above 36 percent. The average (bank) with moving from a single- to a multi-worker firm.16 lending interest rate at the time was 18 percent, and Transactions Costs, Poverty, and Low Productivity Traps: Evidence from Madagascar’s Informal Microenterprise Sector  149 interest rates on microcredit were 36 percent. Thus, there other than do large commercial entities. Large compa- is a substantial economic loss associated with the market nies would require a higher return on capital in line with structures observed, where 70 percent of OOMEs are the the cost of capital in the economy; they must also pay single-worker type. the costs of formality. Finally, because they enjoy greater market power, their presence may permit OOMEs in the Perhaps surprisingly, having more education does not same sector to set higher prices as well. dramatically improve profitability of OOMEs. This may explain why education has a negative selection effect on The structure of returns also varies, as one would entry of these entrepreneurs. Returns to schooling con- expect, by sector. Estimates for industry and trading ditioning on owning an OOME and on other included (sectors with sufficient observations to estimate equa- variables are positive and significant for single-worker tion (1) separately) are shown in table 5.8. In industry, OOMEs, but are not statistically significant for multi- returns to capital and the owners’ time are again higher worker ones. Moreover, these returns are low relative to for multi-worker firms, although these differences are international benchmarks of approximately 10 percent less pronounced in trading. Male entrepreneurs and (see Montenegro and Patrinos 2014). In the broader LFS, those with more education have higher profits in some the estimated returns range from 10.4 to 11.7 percent, subsets of the sample, and returns to capital are increas- depending on the set of conditioning variables, a level ing for industry and mildly decreasing in trading. close to international benchmarks.17 However, given that more education is predictive of owning a larger OOME, conditioning on size and other characteristics may Labor Market Frictions understate the actual returns to education. The estimates reported imply that firms not only employ In addition, in the estimations with firm characteristics, too little capital, but also too little labor, conditioning on we find that the type of competition faced from the their level of assets. First, expanding from a single-worker firm’s main competitors has an impact on the firm’s prof- firm to a multi-worker firm improves the returns to the itability. Of the four types of competition firms could owner’s time and capital appreciably. Second, conditioning report—large and small commercial operators and large on assets, although the returns to paid labor are diminish- and small noncommercial operators—competing primar- ing, they are still positive for most firms.18 The estimated ily with a large commercial operator increased profits profit-maximizing level of paid worker hours per month significantly (see table 5A.3). Smaller entities appear to would be 760 hours, which translates to 4.4 full time- provide more direct or intense competition with each equivalent workers. This is substantially higher than TABLE 5.8: Estimates of Returns to Productive Factors by Broad Sector Industry Trading Single worker Multiple worker Single worker Multiple worker Assets .0199*** .261** .105~ 0.122 ** Assets2 ~ –.0000826** –0.0000425~ –0.0000183 ** Assets3 ~ 53 10–9** ~ Head hours .56*** 1.041** 0.340~ .58~ Head hours2 ~ Unpaid hours ~ .475 ** Unpaid hours2 ~ Paid hours ~ –.09~ Paid hours squared ~ Male = 1 79.8*** ~ 103* 290.7*** Schooling of head 8.431*** ~ ~ 32.9*** N 1,269 285 731 384 150 Republic of Madagascar Employment and Poverty Analysis TABLE 5.9: Marginal Profit of Paid Labor increase their profits simply by scaling up, using the and Wages Paid available technologies.20 As shown in table 5.10, with the exception of paid worker hours and owner hours at or Percentile of observations above the 90th percentile of the factor’s use, all elastici- 25th 50th 75th ties are positive. Moreover, the profit elasticity with Marginal (profit) product .263 .357 .414 respect to capital assets is generally increasing as one Survey estimates average wage .213 .400 .714 moves up the distribution. A 1 percent increase in the Friction/wage 1.23 .89 .58 firm’s asset base (or capital) will contribute only slightly to profits at the low end of the scale, but as the firm’s capital expands, it increasingly adds to profitability, current full-time employment of on average 81 paid hours holding other factors costs constant. However, because and .47 paid workers for multi-worker firms. these gains are not offset by the cost of borrowing, single-worker firms would have to achieve a certain scale The gap between actual and optimal labor employment to make borrowing for investment purposes worthwhile. suggests an important labor market friction, typically With respect to labor inputs, to maximize profits would due to the costs of finding, monitoring, managing, and entail hiring paid labor at approximately the 75th per- dismissing workers (Rogerson, Shimer, and Wright centile level for multi-worker firms, and above the 2005). The relationship between the unexploited mar- 90th percentile of all OOMEs. ginal profit and wages suggests that the friction dimin- ishes as a percentage of average wage as wages (and These positive elasticities, combined with the coefficients labor productivity) increase (see table 5.9). To assess on returns reported in table 5.7, provide evidence of whether such costs vary with firm characteristics, we a major allocative inefficiency and a low-productivity examine correlations of these with the number of hours “trap.” Too much capital and labor are employed in of paid labor used. All else equal, male-owned firms hire single-worker OOMEs, which have low returns to both 330 more paid labor hours per month. Firms headed by factors, and among multi-worker OOMEs there remain an individual whose father was a “boss” hire 250 more unexploited economies of scale. These results contrast labor hours, whereas those whose father was employed with those found in several West African countries. in agriculture hire 120 hours less. Being located in a Grimm, Kruger, and Lay (2011) find that returns to large urban area is associated with hiring 146 hours capital at low levels of operation are as high as 70 per- more of labor. Finally, for each year of completed educa- cent per month in all urban areas but one (Lomé) and tion, heads of firms hired 29 more hours of paid labor fall with the level of capital. Therefore, their results do per month. While the difference between single-worker not support a hypothesis of a low-productivity trap: for OOMEs, which have 4.5 years of education on aver- this to occur, the returns to capital at a very small scale age, and multi-worker OOMEs, at 4.8 years, is not very would be too low to enable poor microentrepreneurs large, informal firms with paid workers have an average to grow their businesses through internally generated of 6.0 years of schooling. Based on these correlations, resources. Yet in Madagascar, where returns are lower at it appears to be more costly for female owners to hire the lowest levels of investment and generally rising with paid workers, and having more education and training scale, this hypothesis appears to have validity. appears to reduce the costs of hiring paid workers. See annex table 5A.6 for more details.19 The Returns to Profit Elasticities and Low- Microentrepreneurs’ Labor Productivity Traps Many OOME heads responded to the survey by saying that they chose to operate a business to improve their To assess whether there are increasing (profit) returns to incomes. (See the “Introduction” to this chapter.) To see scale, we compute the elasticities of profits with respect what the effects on labor income of this mode of (self-) to each factor of production and for all factors added employment are, we compare average labor income per together. If these elasticities are positive, then firms could month among different categories of workers as captured Transactions Costs, Poverty, and Low Productivity Traps: Evidence from Madagascar’s Informal Microenterprise Sector  151 TABLE 5.10: Elasticity of Profits with Respect to Factors of Production by Percentile of Inputs Used (Unweighted) 10th 25th 50th 75th 90th 95th Assets All 0.0005 0.0025 0.0070 0.0307 0.1347 0.4635 Single worker 0.0002 0.0012 0.0032 0.0132 0.0681 0.2119 Multi-worker 0.0026 0.0074 0.0153 0.0705 0.3778 0.7901 Paid worker hours All 0.1007 0.2414 0.1797 0.1667 0.1492 –0.2277 Multi-worker 0.0744 0.1713 0.1183 0.0808 –0.0500 –0.4821 Unpaid worker hours All 0.0710 0.1248 0.2133 0.3002 0.2344 0.3875 Multi-worker 0.0648 0.1139 0.1947 0.2740 0.2139 0.3536 Owner hours All 0.3687 0.4154 0.3839 0.2124 0.0787 –0.0722 Single-worker 0.4294 0.5122 0.6752 0.6815 0.8842 0.7680 Multi-worker 0.4020 0.3972 0.3413 0.0744 –0.2803 –0.4699 All factors, total and using no negative factors All 0.5409 0.7841 0.7839 0.7100 0.5971 0.5510 All (no neg) 0.5409 0.7841 0.7839 0.7100 0.5971 0.8510 Single-worker 0.4297 0.5133 0.6784 0.6947 0.9523 0.9799 Multi-worker 0.5438 0.6899 0.6696 0.4996 0.2613 0.1917 M-w (no neg) 0.5438 0.6899 0.6696 0.4996 0.5917 1.1437 Note: Elasticity estimates are computed using the coefficients estimated and reported above and locally smoothed values of the independent variables and associated profits in the sample within 5 percent on either size of the indicated percentiles. in the LFS. Because wages are not observed for owners positive wage. The excluded category of workers are of OOMEs, we estimated a shadow wage equal to the therefore those workers most like the OOME head, but average estimated contribution to profits of the OOME who work for a wage in the private sector. The results heads’ labor using the coefficients estimated and the level are summarized in table 5.11. All else equal, heads of of OOME labor applied for each firm. We then estimate single-worker OOMEs are paid significantly less than a linear wage equation (3) in two steps, as follows: their counterparts who work in the private sector. The loss in income of 60,000 ariary per month is substantial w j = a + bX j + mη ˆ j + εj, (3) relative to the mean earnings of urban-based workers where wj represents the monthly labor income of indi- of 101,000 ariary (INSTAT 2013). When one estimates viduals in the LFS, indexed by j. Xj includes the age, level these relationships using OLS, however, the effect of of education and training, region, size of urban center, single-OOME ownership is not significant, implying that sector of employment, and an array of family history their unobserved ability increases their actual earnings variables for individual j. η ˆ j is a polynomial approxima- relative to comparator workers. In contrast, heads of tion to the generalized residual from equation (2), which multi-worker OOMEs make on average 68,000 ari- in this case captures unobserved heterogeneity affecting ary per month more than their counterparts in regular labor income—in the case of OOME heads, unobserv- employment. Their business assets and scale appear able ability or opportunity. We include binary indicator complementary to their labor effort, even when one con- (“dummy”) variables for those owning and operating ditions on family history and unobservables. Although an OOME, for workers in the public administration, heading a single-worker OOME represents a major inef- workers in agriculture, industry, trade, and services, and ficiency, operating a larger multi-worker OOME can be for paid employees (not heads) of OOMEs receiving a a gainful choice for those willing and able to do so. 152 Republic of Madagascar Employment and Poverty Analysis TABLE 5.11: Wage/Average Returns to Labor economies of scale. Thus, the apparent persistence of (1000 Ariary per Month) low-productivity OOMEs requires a more multifaceted explanation of the failure of OOMEs to grow despite the Bias corrected OLS possibility of financing growth through credit, internal Single-worker OOME –59.57** –3.90 resources, or mergers and partnerships. (–2.85) (–1.06) Multi-worker OOME 68.47*** 53.43*** We propose a hypothesis that relies upon a combination (10.86) (11.58) of market failures to produce a persistent (“steady”) Employee of OOME –12.00*** –11.80*** state of low productivity for OOMEs. First, there (–3.40) (–3.51) must be a constraint to the competitive pressures that Received vocational training 76.56*** 76.53*** OOMEs face from larger, more competitive firms, both (8.41) (14.62) in serving the market and in retaining labor. Second, OOME owners must be poor—that is, they must have Male 31.59*** 29.87*** low levels of wealth relative to the range where returns (13.83) (12.49) are highly increasing. Third, financial markets must fail Years of education 9.042*** 8.819*** to serve a significant portion of entrepreneurs. Finally, (17.53) (21.94) and relatedly, due to information asymmetries, it must Note: t-stats (bootstrapped as appropriate) in parentheses be costly for entrepreneurs to merge their resources through partnerships (or other similar financial or ownership arrangements) to take advantage of scale Toward a Unified Theory economies. of OOMEs: Asymmetric To illustrate in a dynamic setting how these circum- Information and Incomplete stances could produce a low-productivity trap, we consider a simple model of the entrepreneur’s decision of Markets how much to invest and how many entrepreneurs with which to partner. We assume that competitive pressures The pattern identified in returns to capital and labor esti- from larger entities are limited for reasons outside our mated from the sample of Malagasy OOMEs presents a model. puzzle. That is, in the presence of increasing returns, even if firms start small, over time they could invest a small Entrepreneurs have available wealth in period t denoted amount more in each period, employing more labor xt. The value of this wealth can be captured in a value as assets expand, and thus achieve higher profitability function V, as follows: over time.21 This is true even if entrepreneurs are risk V (xt )  Et ∑ t = 1 m(ct ) = m(ct ) + EtV (xt + 1), T averse.22 Those that achieved greater scale could employ (4) more workers and drive the less productive firms out of business, while offering higher wages as well. Thus, where µ(ct) denotes contemporaneous utility. In each even if many entrepreneurs face borrowing constraints, period t there are two decisions: first the entrepreneur if the most able can access financing and there were decides the number of owners, B, to have (re-)invest in otherwise relatively free entry and exit, resources (capital the enterprise, knowing that each will invest an optimal and labor) would flow to the most productive enter- level for them, kt* in the second stage. Solving the prob- prises able to grow more quickly. Credit constraints can lem backward, each entrepreneur must decide how much constitute a barrier to entry if these constraints apply to of available resources to consume (ct), how much to almost all entrepreneurs. However, they do not provide a save in liquid assets (at), and how much to invest in the complete explanation for the inefficient market structure enterprise, (kt), which can include paid labor, raw materi- observed in Madagascar. With economies of scale, saving als, or capital assets. To simplify the model, assume that small amounts of retained earnings each period would all owners invest and intend to split profits evenly, have result in firm and productivity growth. Moreover, even equal wealth, face the same constraints, and have the microenterprises that cannot borrow could combine same utility functions. Total capital invested in t will thus their capital to create larger firms and achieve greater equal Kt ≡ ktBt. Transactions Costs, Poverty, and Low Productivity Traps: Evidence from Madagascar’s Informal Microenterprise Sector  153 Because there is no credit market, in all periods, entre- FIGURE 5.3: Shape of Profit and Value Functions preneurs must have nonnegative assets for all t: i.e., xt − kt − ct = at ≥ 0. If the entrepreneur is borrowing- constrained in period t in the sense that he would be better off borrowing if he could in period t, then his V(x) household consumes and invests all resources, that is, ␲(␹) ct = xt – kt, and he sets savings (at) to zero.23 In this case, in period t the entrepreneur sets the marginal utility of consumption equal to the expected value in the subse- quent period of resources earned through the enterprise ␹ (in t+1). These resources are thus determined by the marginal value product of kt. In particular, for all t, and optimum number of enterprise owners B* and market demand d, the entrepreneur sets consumption ct such x that  dV (p  ) dp  (kt B*, d)  Solving now for the first stage of the entrepreneur’s m ′(ct ) = Et  t + 1 t + 1 ε t +1  , (5) dx dk problem, taking as given that any potential partners   t +1 t  will set kt = k*(xt) in the second stage, the entrepreneur chooses B, according to the following optimization prob- where εt+1 is a random shock to profits centered at 1. We lem to maximize his individual benefit p : , assume the standard properties of the utility function − m ′(c) > 0, and m ′′(c) < 0. That is, the marginal utility  p(k*B; d)   (k*, B) =  max p − m(B − 1)  of consumption is declining with the level of consump- B  B   tion and is high when consumption is low, as is typically for B ≥ 1. assumed. Under standard assumptions, with concave There is a monitoring cost of m, which increases linearly profit (that is, p ′′(K) < 0), V (xt )) is also concave. In this with the number of owners, incurred to ensure that other case, however, this is not guaranteed. Profits, p  , are an owners do not withdraw excessive resources from the increasing function of total investments made, ktB*. firm. If the expression in brackets is concave, the solu- In these circumstances, there is a range of x for which tion to the first order condition gives optimal B as: E[V(xt+1)] is concave, and a range for which it may be convex.24 More formally, if there is a range of total p ′k* + 2 (p ′k*)2 − 4mp firm capital, kB for which p ′′ > 0, then it is possible B*(x, m, d) = . dV dp(kt B*) 2m for φ(xt )  t + 1 to be increasing in xt—i.e., dxt + 1 dkt For a solution with B > 1 to exist, (p ′k*)2 − 4mp ≥ 0 convex. Under certain functional forms, including if —that is, the incremental profits from an additional V ′′′ = p ′′′ = 0, there is a value of x, denoted x , such that investor must be high relative to current profits and φ ′ < 0 for x < x , and φ ′ > 0 for x > x. This implies that monitoring costs. Otherwise, B* = 1. the expected value function is an upside-down S-shape, as shown in figure 5.3. The implication of the concav- Differentiating the first order condition further, one dB* ity of V(x) for x < x despite increasing returns to K is obtains that < 0. As expected, the optimal number dm that since the marginal utility of c is high, entrepreneurs of partners is lower the higher is the cost of monitor- choose a level of Kt such that there are unexploited dB* d 2p increasing returns to scale (in expectation) and for which ing them. Moreover, > 0 if > 0 in the relevant dk* dk*2 p ′(K) is low. How probable it is that the entrepreneur range. That is, the number of partners is increasing in the adopts a low-productivity strategy, with low returns level of capital all partners are willing to invest. Moni- to capital, therefore, depends on his level of wealth, x, toring carries a fixed cost, and if potential entrepreneurs relative to x. ,25 The poorer the entrepreneur is, the more are too poor, they tend to work alone. Each entrepreneur likely she is to set k below the range where returns are knows that other possible partners are, like himself, high (See figure 5.3). drawing their household consumption resources from 154 Republic of Madagascar Employment and Poverty Analysis TABLE 5.12: Client Types and Firm Profits are limited liability companies. For similar reasons, offer- ing credit—which from the lender’s perspective looks Average value similar to becoming a partner as it requires the assump- Share of added by main Main Clients firms client type tion of risk and monitoring costs—will typically not be Public and semi-public sector 0.08 322.6 worthwhile. Large private company 1.38 305.7 (commercial) Finally, there is an important market-level effect of the Small business (noncommercial) 3.78 269.5 general lack of demand, which could be introduced in Large private company 1.08 220.3 a general equilibrium extension of this (partial equilib- (noncommercial) rium) model. In a situation where incomes are low (that Small business (commercial) 16.52 182.3 is, almost all businesses are too small), and entrepre- Households 77.01 164.8 neurial wealth and labor earnings are low, equilibrium market demand for goods and services is reduced, as are opportunities to offer more differentiated goods and services (Greenwald and Stiglitz 1990). As shown the revenues of the firm, and that given entrepreneurs’ in table 5.12, for 77 percent of OOMEs the main client poverty, there is a high marginal utility of consumption. types are households, followed by other small businesses This may only increase the costs of monitoring other at 16.5 percent. The type of client matters, as shown in owners’ actions and firm resources. Therefore, partner- table 5.12. The average value added for OOMEs is high- ships are unlikely to form. Among informal enterprises est among those able to serve the public sector and large in Madagascar, as is generally observed elsewhere, private companies. Perhaps this explains why, from each shared ownership is rare: less than 1 percent of OOMEs entrepreneur’s perspective, rather than a lack of credit, FIGURE 5.4: Most Important Obstacles to OOMEs’ Business 35 Single worker 30 Multi-worker Percent of firms citing obstacle 25 20 15 10 5 0 t it) s n er n tit ion tit s t it) er e s en en an rial tie er io om io ac ed ed th s) n) y) rk at st ct pe ct m m sp l io qu te cu O cr cr wo cu du m du ax ip ge a iffi e ng h u an w m ,t of pro co ro bl a uc eq d ld an ns ita tti p fie m ica lit ra m d nd io ck of uc of ge su ali o at n ua of n a (la ow m w (to qu of lty si ch ul h y ry o lo (q ly eg tie Fl cu ck Te g e ne (to F pp nc in iffi La yr l hi cu nd Su na ac (d an ffi Fi Fi m Di m e nc of o na To ck Fi La Obstacle Transactions Costs, Poverty, and Low Productivity Traps: Evidence from Madagascar’s Informal Microenterprise Sector  155 OOMEs claim that their main constraints are a lack of reported, most were informal: 48 percent were from demand. As shown in figure 5.4, the most important family or friends, 20.6 percent from a microcredit obstacle to OOMEs that entrepreneurs most frequently institution, 15 percent from suppliers, 9.2 percent cite is the lack of customers, followed by too much from customers, 3.5 percent from money lenders, and competition. 2.0 percent from banks. Moreover, in a regression of enterprise assets on other characteristics, the education of the entrepreneur’s mother and father increased his Madagascar’s Weak or her enterprise assets, as did being male (conditioning on sector and region). The age of the establishment was Enforcement and Monitoring not significant, suggesting that firms’ assets are typi- Infrastructure cally not accumulated over time. Madagascar’s financial markets are relatively undevel- Although OOMEs rate demand-side issues as their oped in part because they lack the supporting infor- most serious obstacles, in the hypothetical situation mation and enforcement infrastructure required to that they would be granted access to finance, their efficiently screen borrowers and enforce repayment. In priority would be to invest in capital. As shown in 2012, credit registries covered over 0.1 percent of the figure 5.5, the most frequent response entrepreneurs adult population; credit bureaus covered zero percent gave to the question “What would be your priority if of the population, and the strength of legal rights to you could benefit from credit for your activity?” was enforce repayment rated only 2 on a 10-point scale in to improve the quality of machinery and tools, fol- Doing Business’s Getting Credit. Credit was also costly lowed by increasing the stock of raw materials and in 2012, with bank-lending interest rates averaging opening another establishment. A very small percentage approximately 18 percent, and real interest rates of responded that they would hire more workers. These 11.6 percent (according to central bank data).26 At the responses are consistent with the fact that returns to same time, microcredit had been growing. Launched capital are increasing, whereas the estimated marginal in 1990, Madagascar’s microfinance sector had about product of hiring workers is diminishing and involves 31 players in 2012, which included state, foreign inves- important transactions costs. tor, and donor-supported initiatives, operating under a legal framework and regulated by Madagascar’s Cen- tral Bank. The average lending rate was 36 percent— Formal Registration a rate almost equivalent to the estimated return on capital in the OOME sector. Given these realities, for The debate regarding the potential of microenterprises single-worker OOMEs, borrowing small amounts is closely intertwined with questions regarding the costs would not be profitable. Moreover, formal creditors and benefits of firm formality. Thus, a firm characteristic would be unlikely to lend to OOMEs without their of policy interest is the decision to register one’s firm adopting more transparent accounting standards and/or with the authorities. Because the intention in the survey offering high collateral relative to loan values in order was to sample informal firms, these data provide only a to reduce the monitoring costs m relative to loan partial perspective on the issue. values. Nonetheless, some firms in the survey were more formal As a result, use of credit by Madagascar’s OOME sec- than others. Surveyed firms were asked whether they tor is extremely low. A full 92.5 percent of OOMEs had any of the following types of registration: registra- received their assets through a gift, inheritance, or own tion with Ministry of Commerce or the social security savings and only 1.2 percent of them through some fund or possession of a license or professional identifi- type of loan. Moreover, only 3.6 percent of single- cation card.27 If we consider a firm with at least one of worker OOMEs and 3.8 percent of multi-worker these forms of registration, only 9 percent of OOMEs OOMEs utilized some type of credit to finance the were registered in 2012: 5.7 percent of single-worker operations of their enterprise. Of all credit transactions OOMEs and 9.3 percent of multi-worker OOMEs. The 156 Republic of Madagascar Employment and Poverty Analysis FIGURE 5.5: Frequency of Response to “What would equation (2) with the outcomes v set either to the deci- be your priority if you could benefit from credit sion to operate an unregistered OOME of the appropri- for your activity?” ate type (single- or multi-worker) or a registered OOME of that type and include the variable capturing registra- 20 tion. Again, we use the method of Das, Newey, and Vella 18 (2003) to address possible bias in this coefficient. 16 14 In the first stage (equation 2), we find that the sector 12 and type of work of the individual’s father, the father’s 10 schooling, and the age and gender of the head of the OOME are significant determinants of the decision to 8 register. In particular, having a father who was more 6 educated makes it more likely that the enterprise will be 4 registered, as will having a head who is older or male. 2 (See annex table 5A.6) 0 Improve machines, furniture, tools Increase stock of raw materials Open another establishment in another activity Open another establishment in this activity Improve the site Spend more outside of the establishment Hire workers The effect of registration is found to be positive and statistically significant for single-worker OOMEs, but is not for multi-worker firms, with other significant results summarized in table 5.14. The bias correction terms are significant, and the coefficients on the variable registered differ significantly when these terms are included (vis-à- vis OLS), suggesting that the effects of registration are attributable to some extent to unobserved heterogeneity in either ability or opportunity. Source: ENEMPSI 2012 Whereas the act of registering does not improve profit- TABLE 5.13: Population of Informal Enterprises ability on its own, it appears to be associated with access by Sector and Type to services and markets that do. Registered firms are more likely to have electricity: 28.3 percent versus only Of which: 6.4 percent of unregistered firms. Also, 31 percent have Percentage single Of which: a telephone as opposed to only 6.5 percent of unreg- of total worker registered istered firms. They are more likely to have a computer Industry 43.1% 71.8% 1.7% (4 percent) or Internet connection (2 percent) than are Primary 4.2% 65.9% 4.1% unregistered firms (less than 1 percent in both cases), Other services 17.4% 74.2% 16.2% although the level of use of these technologies is very Trade 35.3% 67.9% 15.8% Source: ENEMPSI 2012. TABLE 5.14: Statistically Significant Firm Characteristics for Profits breakdown by broad sector is shown in table 5.13. Those Variables Multi-worker Single-worker that did so also achieved a higher level of profitability Registration ~ 100.5** on average. However, this may not be due to the effect of registering: those firms which register may be run by Has electricity 37.3* 40.47** the most capable individuals, with more opportunities Fixed locale ~ ~ for which registration is a requirement, and thus the Exporter 245.2** 249.7** coefficient on registration from an OLS regression could Note: Regional and city-size dummies, other characteristics, and family be biased. To address this issue, we estimate equation history variables included. *p=.10; **p=.05; *** p=.01. ~ = not statistically significantly different  (1), this time including a binary indicator variable for from 0. In all cases, cross-validation was used to select specification of whether or not the firm is registered. We first estimate bias-corrected estimation. Transactions Costs, Poverty, and Low Productivity Traps: Evidence from Madagascar’s Informal Microenterprise Sector  157 TABLE 5.15: Percentage of Firms Using Different Types of Locations, Registered versus Unregistered Registered Unregistered Travelling, improvised on road or public market 10.3 21.7 Stationary on the road 6.5 6.7 Vehicle 14.9 1.2 At home (either fixed or unfixed installation) 14.9 40.2 Fixed locale in public market, workshop, shop/boutique, or restaurant 52.1 25.7 TABLE 5.16: Correlation of Firm Registration and Main Client Type misallocation of both capital and labor. The returns to invested capital for the tiniest of these enterprises are Probit estimation below the opportunity cost of capital in the economy, coefficient and returns to the entrepreneur’s labor are significantly Public sector clients 0.736* lower than wages of similar individuals in the labor (1.73) market. Large private commercial client 0.332* (1.77) Although owner-operated microenterprises are moti- Assets 3.39e-08*** vated to invest more to raise their incomes, they are (8.67) constrained by competition among similar enterprises, Large urban 0.539*** by low demand and investor capacity in a context of (10.95) widespread poverty, difficult access to export markets, Direct exporter 0.797* and high transactions costs in capital and labor markets. (1.76) Although Madagascar’s financial markets are undevel- N 5338 oped, the very nature of owner-operated microenter- Note: t statistics in parentheses. prises increases the importance of high transaction costs and limits the markets’ potential to grow. In particular, the high cost of monitoring prospective partners or bor- rowers relative to profit levels, exacerbated by the lack low in general. They are also more likely to have a fixed of formal separation between household and enterprise locale (table 5.15). But perhaps the greatest benefit to finances and the risk aversion of poor entrepreneurs, registration, however, is improved ability to serve higher makes this problem difficult to solve using conventional value customers. As shown in table 5.16, there is a cor- credit markets. Moreover, the costs of microfinance relation between registration and serving the highest exceed the returns to capital at the smallest scale of value clients—public entities, large private companies, operation. Monitoring and control technologies can and export markets. be expensive, and they are typically more difficult to implement for informal firms precisely because they are Conclusions and Agenda informal; these firms lack the practice of keeping clear and precise accounts which can be checked and audited. for Further Research Because firms with the scale and technologies needed to compete against the market of low-productivity informal Analysis of the patterns of returns, investment, and firms also face barriers to entry, an equilibrium market employment on the part of owner-operated microenter- configuration persists, with many low-productivity firms prises in Madagascar suggests both potential and limits competing for the same limited markets. to their ability to contribute to poverty reduction in the country. The current structure of markets, character- While there appears some potential to foster the growth ized by the prevalence of a large number of unproduc- of OOMEs that have already achieved a certain scale, tive informal microenterprises, results in a substantial to significantly reduce the misallocation of capital and 158 Republic of Madagascar Employment and Poverty Analysis labor misdirected to single-worker OOMEs in particular poor state of its infrastructure, poor access to external would require an alleviation of constraints inhibiting markets, and a difficult investment climate. These bar- investment by larger, more sophisticated firms, which riers would need to be lowered for investments in high- can offer more productive employment and stimulate potential OOMEs to bear fruit. Moreover, unless the greater demand and better access export markets. At the labor market functions efficiently, investments may not same time, given these constraints, the absence of larger, bring the jobs required to lift substantial numbers of the more efficient firms that could stimulate a restructuring population out of poverty. Thus, a greater understanding of markets and increased labor productivity, incomes, of the sources of high-transaction costs is needed, so that and demand helps to perpetuate the low productivity policies and institutional innovations can be explored “trap” indicated by our results. Madagascar’s investment to reduce the frictions inherent in informal and formal climate suffers from political instability, but also from a labor markets.28 Transactions Costs, Poverty, and Low Productivity Traps: Evidence from Madagascar’s Informal Microenterprise Sector  159 Annex 5A. Tables TABLE 5A.1: First Step Estimates: Single- versus Multi-worker OOME Single worker OOME Multi-worker OOME Father’s level is upper, engineer and similar (salaried) –0.095 0.134 (–0.40) (0.48) Father’s level is middle management, foreman (salaried) 0.111 0.192 (0.55) (0.78) Father’s level is skilled worker (salaried) 0.192 0.256 (1.01) (1.10) Father semi-skilled worker 0.192 0.140 (0.97) (0.57) Father salaried manual laborer 0.447* 0.253 (2.34) (1.07) Father was unsalaried head 0.308 0.437 (1.60) (1.88) Father independent worker 0.302 0.297 (1.69) (1.35) Father apprentice –0.708 –0.0170 (–1.00) (–0.02) Father worked in agriculture –0.385*** –0.279*** (–6.30) (–3.68) Father worked in industry 0.0491 0.139 (0.72) (1.66) Father worked in trade 0.124 0.269** (1.78) (3.19) Father’s years of education –0.0106 –0.00673 (–1.58) (–0.82) Mother’s years of education –0.00807 –0.00175 (–1.12) (–0.20) Male –0.157*** 0.192*** “= 1 male, = 0, female” (–4.97) (4.97) Age 0.0180*** 0.0208*** (16.53) (15.81) Years of education –0.0471*** –0.0336*** (–8.76) (–5.16) Large urban 0.153*** 0.188*** (4.26) (4.27) Includes region dummies (not shown) N 20065 Note: t statistics in parentheses. ***p = .01; **p = .05; *p = .10. 160 Republic of Madagascar Employment and Poverty Analysis TABLE 5A.2: Value Added Estimation without Firm Characteristics Single Worker Multi-worker 2-step OLS 2-step OLS Assets 0.0130*** 0.0149*** 0.0330*** 0.0319*** (2.87) (3.77) (5.30) (5.85) Assets squared –0.000000137** –0.000000154*** 5.74e–08 –1.16e–09 (–2.23) (–2.75) (0.98) (–0.02) Hours head 1.030*** 1.059*** 1.868** 1.315* (4.02) (4.86) (2.14) (1.82) Hours head squared –0.00120* –0.00145** –0.00355* –0.00208 Squared head monthly working hours (–1.78) (–2.52) (–1.70) (–1.21) Paid hours 0.494*** 0.552*** (3.05) (4.18) Paid hours squared –0.000266*** –0.000320*** (–4.32) (–6.04) Unpaid hours 0.985*** 0.405* (2.78) (1.91) Unpaid hours squared –0.000848* –0.0000426 (–1.94) (–0.25) = 1 if large urban 91.32*** 67.71*** 73.69 84.80** (5.79) (5.39) (1.38) (2.07) ‘1 = male 58.62*** 88.19*** 11.35 26.21 (3.77) (7.11) (0.21) (0.67) Includes region dummies (not shown) N 2332 3347 1210 1592 *p<0.10; **p<0.05; ***p<0.01. Note: t-statistics in parentheses.  TABLE 5A.3: Estimates of Determinants of Profits, with Firm Characteristics Single worker Multi-worker All Total assets (1000 ariary) 0.0104** 0.0384*** 0.0191*** (2.54) (14.75) (5.94) Total assets squared –0.000000107* 0.000000134*** (–1.90) (4.16) Boss working hours 0.743*** 1.893** 1.057*** (3.25) (2.50) (3.66) Boss working hours squared –0.000707 –0.00363** –0.00169** (–1.18) (–1.97) (–2.29) Paid worker hours 0.376*** 0.490*** (2.62) (6.21) Paid worker hours squared –0.000289*** –0.000275*** (–5.05) (–7.76) Unpaid worker hours 0.325** 0.356*** (2.32) (4.92) (continued) Transactions Costs, Poverty, and Low Productivity Traps: Evidence from Madagascar’s Informal Microenterprise Sector  161 Single worker Multi-worker All Years of schooling, head 11.49*** 7.046 10.69*** (6.13) (1.14) (4.48) Exporter 264.3** 77.15 189.0 (2.55) (0.27) (1.53) Competitors small commercial –26.38 –419.0*** –191.1*** (–0.56) (–2.93) (–3.27) Competitors large noncommercial –65.47 –549.9*** –265.2*** (–1.03) (–3.04) (–3.45) Competitors small noncommercial –53.39 –386.7*** –200.2*** (–1.10) (–2.60) (–3.32) Product selected for family tradition 10.20 208.7** 69.40** (0.38) (2.32) (2.04) Product selected business know 12.31 115.9 38.86 (0.52) (1.42) (1.28) Product chosen for profit 83.58*** 304.9*** 147.4*** (3.34) (3.69) (4.64) Product chosen for stable revenues 34.13 147.1 69.91* (1.03) (1.50) (1.74) Belongs to organization of producers 53.38 369.8*** 158.2*** (1.35) (3.19) (3.27) Industry 73.11* 4.998 46.00 (1.90) (0.04) (0.98) Service 78.62* –57.84 45.15 (1.95) (–0.46) (0.90) Trade 111.8*** 113.0 111.9** (2.88) (0.97) (2.34) Large urban 33.94** 45.38 32.60* (2.37) (0.96) (1.79) Male 83.07*** 25.76 63.72*** (6.09) (0.59) (3.75) Father boss (no salary) –3.304 178.9** 81.74** (–0.10) (2.12) (2.18) Father upper salaried –3.376 777.0*** 307.4*** (–0.05) (3.85) (3.71) Father worked in agriculture –33.75** –92.75* –57.24*** (–2.28) (–1.94) (–3.05) Constant –45.12 595.6** 169.0* (–0.57) (2.26) (1.68) N 2775 1355 4125 Note: t statistics in parentheses. *** p = .01; **p = .05; *p = .10. 162 Republic of Madagascar Employment and Poverty Analysis TABLE 5A.4: Multinomial Probit Estimates of Determinants of Registration Whole sample Single-worker OOME Multi-worker OOME Choice 1: Choice 2: Choice 1: Choice 2: Choice 1: Choice 2: Own, not Own, Own, not Own, Own, not Own, Dependent variables registered registered registered registered registered registered Father upper level salaried 0.0450 –0.263 –0.0173 –0.718 0.116 0.342 position (0.20) (–0.74) (–0.07) (–1.64) (0.37) (0.61) Father was middle manager/ 0.223 –0.186 0.165 –0.432 0.232 0.209 foreman (1.16) (–0.57) (0.77) (–1.16) (0.87) (0.39) Father skilled salaried 0.257 0.0368 0.231 –0.295 0.171 0.467 (1.43) (0.12) (1.15) (–0.84) (0.68) (0.89) Father semi-skilled salaried 0.346* –0.734** 0.322 –0.848** 0.246 –0.449 (1.84) (–2.10) (1.54) (–2.12) (0.94) (–0.77) Father salaried manual 0.494*** –0.0154 0.539*** –0.396 0.0988 0.411 laborer (2.72) (–0.05) (2.67) (–1.07) (0.39) (0.77) Father was head (no salary) 0.399** 0.238 0.297 –0.0612 0.352 0.596 (2.19) (0.76) (1.46) (–0.17) (1.41) (1.13) Father worked on own 0.358** 0.107 0.337* –0.187 0.187 0.486 (unsalaried) (2.11) (0.36) (1.77) (–0.56) (0.80) (0.95) Father was unsalaried –0.769 0.0595 –0.573 –10.43 –10.17 1.033 apprentice (–1.07) (0.08) (–0.79) (–0.00) (–0.00) (1.19) Father worked in agriculture –0.358*** –0.367*** –0.350*** –0.351*** –0.157* –0.251** (–6.09) (–3.87) (–5.51) (–2.88) (–1.82) (–2.06) Father worked in industry 0.0941 0.0335 0.0242 0.0640 0.197** –0.0478 (1.43) (0.32) (0.34) (0.49) (2.08) (–0.34) Father worked in trading 0.146** 0.289*** 0.0822 0.0774 0.170* 0.376*** (2.16) (2.84) (1.13) (0.58) (1.73) (2.97) Father’s years of education –0.0133** 0.00258 –0.0120* 0.00250 –0.00873 0.00479 (–2.05) (0.25) (–1.70) (0.19) (–0.93) (0.37) Mother’s years of education –0.00562 –0.00854 –0.00868 –0.00411 0.00374 –0.00876 (–0.81) (–0.80) (–1.15) (–0.29) (0.37) (–0.65) Male = 1 –0.1000*** 0.256*** –0.246*** 0.226*** 0.224*** 0.275*** (–3.31) (5.01) (–7.44) (3.35) (5.14) (4.23) Age 0.0192*** 0.0206*** 0.0150*** 0.0162*** 0.0163*** 0.0174*** (18.44) (11.86) (13.29) (7.12) (11.13) (8.01) Years of education –0.0622*** 0.0298*** –0.0543*** 0.0235** –0.0490*** 0.0411*** (–11.94) (3.76) (–9.57) (2.28) (–6.46) (4.09) Large or secondary urban 0.127*** 0.446*** 0.0896** 0.467*** 0.107** 0.342*** center (3.72) (7.13) (–7.92) (–8.49) (2.20) (4.33) N 20065 20065 20065 Note: t statistics in parentheses. Includes region dummies. *p<0.10; **p<0.05; ***p< .01. Transactions Costs, Poverty, and Low Productivity Traps: Evidence from Madagascar’s Informal Microenterprise Sector  163 TABLE 5A.5: Wage and Average Revenue Product of Labor Regression Corrected OLS Single worker OOME head –59.57** –3.900 (–2.85) (–1.06) Multi-worker OOME head 68.47*** 53.43*** (10.86) (11.58) Received training for main job 76.56*** 76.53*** (8.41) (14.62) Age 1.318*** 1.364*** (12.02) (15.26) Male = 1 31.59*** 29.87*** (13.83) (12.49) Years of completed education 9.042*** 8.819*** (17.53) (21.94) Paid worker in OOME –12.00*** –11.80*** (–3.40) (–3.51) Father skilled salaried worker –26.04*** –24.94*** (–2.76) (–4.49) Father semi-skilled salaried worker –28.95*** –27.30*** (–2.64) (–3.68) Father salaried manual laborer –38.35*** –35.65*** (–3.89) (–5.23) Father head of own company –29.70** –30.05*** (–2.17) (–4.05) Father worked on own (no salary) –33.69*** –33.18*** (–3.60) (–6.06) Father apprentice –93.77*** –96.43** (–3.14) (–2.26) Father’s years of education 1.175** 1.161** (1.93) (2.42) Mother’s years of education 1.780*** 1.727*** (2.81) (3.25) Main job in public administration 113.6*** 114.5*** (14.16) (21.96) Agricultural laborer –29.01*** –25.71*** (–6.30) (–4.44) Large urban center 12.39*** 12.85*** (5.20) (4.75) Primary job in industry –6.280 –3.569 (–1.48) (–0.65) Primary job in services –46.03*** –43.07*** (–10.14) (–7.68) Primary job in trade 6.140 8.878 (1.55) (1.62) _cons 14.58 11.17 (1.30) (1.00) N 12455 12455 Note: t statistics in parentheses. Region dummies also included. *p<0.10; **p<0.05; ***p< .01. 164 Republic of Madagascar Employment and Poverty Analysis TABLE 5A.6: Multivariate Tobit Estimation of Correlates of Paid Hours of Labor Employed (Last Month) Coefficient t-stat Owner received vocational training for main job –88.7* –1.93 Male 330.7*** 7.68 Years of education 29.2*** 5.13 Father head/boss 249.7 3.28 Father worked in agriculture –114.7** –2.54 Large urban center 146.4*** 3.14 Industry –259.3** –2.55 Services –423.8*** –3.82 Trade –441.4*** –4.18 Constant –1193.4*** –4.67 N 3722 Note: Estimation includes region dummies and reasons for operating OOME. TABLE 5A.7: Significant Determinants of OOME Registration Single-worker Multiworker Unregistered Registered Unregistered Registered Father semi-skilled salaried 0.322 –0.848** 0.346* –0.734** (1.54) (–2.12) –1.84 (–2.10) Father salaried manual laborer 0.539*** –0.396 0.494*** –0.0154 (2.67) (–1.07) –2.72 (–0.05) Father was head (no salary) 0.297 –0.0612 0.399** 0.238 (1.46) (–0.17) –2.19 –0.76 Father worked on own (unsalaried) 0.337* –0.187 0.358** 0.107 (1.77) (–0.56) –2.11 –0.36 Father worked in agriculture –0.350*** –0.351*** –0.358*** –0.367*** (–5.51) (–2.88) (–6.09) (–3.87) Father worked in trading 0.0822 0.0774 0.146** 0.289*** (1.13) (0.58) –2.16 –2.84 Father’s years of education –0.0120* 0.00250 –0.0133** 0.00258 (–1.70) (0.19) (–2.05) –0.25 Male = 1 –0.246*** 0.226*** –0.1000*** 0.256*** (–7.44) (3.35) (–3.31) –5.01 Age of head 0.0150*** 0.0162*** 0.0192*** 0.0206*** (13.29) (7.12) –18.44 –11.86 Years of education –0.0543*** 0.0235** –0.0622*** 0.0298*** (–9.57) (2.28) (–11.94) –3.76 Large or secondary urban center 0.0896** 0.467*** 0.127*** 0.446*** (–7.92) (–8.49) (–8.87) (–10.32) Note: t statistics in parentheses. Region dummies also included. *p<0.10; **p<0.05; ***p< .01. Transactions Costs, Poverty, and Low Productivity Traps: Evidence from Madagascar’s Informal Microenterprise Sector  165 NOTES 18. The potential for increasing profits by adjusting the head’s and unpaid labor is typically not great for OOMEs. For single-worker 1. These figures were calculated using survey information from firms, if one attributes a reservation wage to the head equal to the Enquête Nationale sur les Objectifs Millenaire du Développement difference in the value marginal revenue product of paid and unpaid (ENSOMD) 2012. workers in multi-worker firms—wage of approximately 400 ariary 2. The percentage of workers thus employed appears to have fallen, per hour—the optimal level of the head’s labor input in single- based on 1-2-3 surveys (nested household employment/labor worker firms would be about 1.5 heads working full time. This is force, and microenterprise, and poverty/consumption surveys), higher than the observed mean of 146 hours. However, since owners which are not comparable, between 1995 and 2010 (Nordman, cannot bring in another owner without sharing profits, their labor Rakotomanana, and Rouboud 2012). allocation decision is not flexible enough to expand incrementally to 3. A key feature of these enterprises is that they are owner operated. 1.5 owners. Moreover, while using unpaid labor also raises profits, Indeed, microenterprises operated by someone other than the owner the firm’s accounts do not fully capture the benefit flows to unpaid are practically nonexistent in Madagascar. laborers out of profits, and typically the application of unpaid labor 4. They also find that firms that are privately owned, those that are is constrained by the availability of family labor—not something the incorporated, and those with more access to external financing are enterprise can optimize freely. more innovative. 19. There is no correlation with the age of the firm, again suggesting that 5. The United States presents a different case, however. There, once firm OOMEs generally decide on their scale of operation based on a set age is taken into account, firm size plays no significant role in job of constraints and opportunities at entry. creation (Haltiwanger, Jarmin, and Miranda 2010). 20. To identify technical increasing returns (decreasing average total 6. For example, consumers in developing countries may prefer to costs), the elasticities of output with respect to inputs would be purchase goods from a series of small merchants rather than from greater than one. Here, we have only measures of profits, rather a supermarket with refrigeration, higher electricity, and advertising than average total costs (ATC). Assuming that prices are constant, a costs but offering greater convenience and quality or variety. positive elasticity implies that ATC are decreasing at that point of the 7. In some SSA countries, for example, microcredit is widely available factor use distribution. to micro-entrepreneurs. In Togo, for example, there are over 140 21. If assets are lumpy, they could save up for them over a number of microfinance institutions, which serve millions of clients, whereas in periods. Eventually, firm size would be limited either by market size Madagascar, microcredit is more limited. or by exhausting increasing returns to scale. 8. Unconditional probit estimation of heading an OOME in the respec- 22. If returns to enterprise inputs are decreasing, even if entrepreneurs tive sectors. In all cases except in industry, the coefficient on “male” are borrowing-constrained, without risk aversion they will invest was statistically significant at the 1 percent level. and achieve profit maximization in a steady state. If entrepreneurs 9. This is nonetheless higher than the level of capital invested among are risk averse, they will invest less, but will not be borrowing- urban SSA microenterprises that Grimm, Kruger, and Lay (2011) constrained in a steady state (Osborne 2006). If there are increasing found (of 80 international dollars.) returns, however, this explanation is no longer adequate to explain 10. We attempted to deduct depreciation of fixed assets to compute prof- underinvestment. its but were unable to obtain sensible results, and therefore utilize 23. Even in this model, in some periods the entrepreneur will save positive cash flow instead. assets, and at least one steady state will result with positive savings. 11. The returns to capital and labor are shown as separable here after 24. The precise ranges will depend upon parameters of the profit and testing whether cross-product terms were significant. They were sig- utility functions, as well as the distribution of et+1. –, 25. If there is a positive probability of a high draw of et+1 such as xt+1 > x nificant only in the pooled sample, and for convenience we therefore assume separability. then entrepreneurs may escape the poverty trap. 12. Given that the survey lacked measures of physical units of output, 26. World Development Indicators reports interest rates of approxi- as well as output prices and some input prices, including for unpaid mately 60 percent of that year, with 52 percent of this constituting labor, we could not estimate a production function per se. We there- the risk spread. However, we report World Bank staff estimates fore interpret our findings as the joint effects of average costs and using central bank data. different degrees of market power. 27. In French, these are Registre du Commerce, Caisse Nationale de 13. This procedure minimizes the squared sum of out-of-sample predic- Protection Sociale, Patente, and Carte professionnelle. tion error. Identification of the second stage equation is provided by 28. For example, as is indicated by Doing Business Employing Workers the exclusion of several family history variables in that equation, as data, Madagascar has the 10th highest minimum wage in the well by the nonlinearity of the relationships from the first stage. world as a percent of average labor productivity, at 0.9. Countries 14. In general, we lose up to 400 observations when using the correc- with a higher ratio are Burkina Faso, Haiti, Honduras, Kenya, tion step, due to the lack of family history variables. When OLS Mozambique, Senegal, Sierra Leone, Togo, and Zimbabwe. estimates are more precise and have a lower mean squared error, we report only these. 15. Tests of the significance of the difference in the coefficient on assets rejected the null of no difference at the 1 percent level. 16. The return to the head’s time is higher than for other “unpaid” labor, REFERENCES and the return to unpaid labor is higher than that for paid labor, as one would expect, given the definition of returns as net cash flow within the period. Since there is typically some kind of noncash Aghion, P., U. Akcigit, and P. Howitt. 2013. “What Do compensation for “unpaid” workers that is not captured in the firms’ We Learn from Schumpeterian Growth Theory?” accounts, this does not mean that firms can expand their unpaid workforce without bearing any cost. Schumpter Lecture presented at the Swedish 17. This is based upon a Mincer-style regression, with region, male, and Entrepreneurship Forum, Stockholm, January 2013. urban indicator variables also variously included. Men are paid http://scholar.harvard.edu/files/aghion/files/what_do_ 37 percent more than females conditioning on these variables and years of schooling. we_learn_0.pdf?m=1361377935. 166 Republic of Madagascar Employment and Poverty Analysis Ayyagari, M., A. Demirgüç-Kunt, and V. Maksimovic. McKenzie, C. J., and C. Woodruff. 2006. “Do Entry 2011. “Firm Innovation in Emerging Markets: The Barriers Provide an Empirical Basis for Poverty Traps? Role of Finance, Governance, and Competition.” Evidence from Mexican Microenterprises.” Economic Journal of Financial and Quantitative Analysis 46 Development and Cultural Change 55(1): 3–42. (6): 1545–1580. Montenegro, C., and H. Patrinos. 2014. “Comparable Bento and Restuccia. 2014. ”Misallocation, Estimates of Returns to Schooling around the Establishment Size, and Productivity.” University of World.” Policy Research Working Paper 7020, World Toronto Department of Economics Working Paper Bank, Washington, DC. No. 517. July. Neumark, D., B. Wall, and J. Zhang. 2009. “Do Small Das, M., W. Newey, and F. Vella. 2003. “Nonparametric Businesses Create More Jobs? New Evidence for the Estimation of Sample Selection Models.” Review of United States from the National Establishment Time Economic Studies. 70: 33–58. Series.” Review of Economics and Statistics 93 (1): Greenwald, B., and J. Stiglitz. 1990. “Asymmetric 16–29. Information and the New Theory of the Firm: Nordman, C., F. Rakotomanana, and F. Rouboud. 2012. Financial Constraints and Risk Behavior.” American “Informal versus Formal: A Panel Data Analysis of Economic Review 80 (2): 160–5. Earnings Gaps in Madagascar.” DIAL Document de Grimm, M., J. Kruger, and J. Lay. 2011. “Barriers to Travail, Antananarivo. Entry and Returns to Capital in Informal Activities: Osborne, T. 2006. “Credit and Risk in Rural Developing Evidence from Sub-Saharan Africa.” Review of Economies.” Journal of Economic Dynamics and Income and Wealth 57 (special issue, May). Control 30: 541–68. Haltiwanger, J., R. Jarmin, and J. Miranda. 2010. “Who Poschke, M. 2014. “The Firm Size Distribution Across Creates Jobs? Small versus Large versus Young?” Countries and Skill Biased Technical Change in Working Paper 16300, National Bureau of Economic Entrepreneurial Technology.” IZA Discussion Research, Cambridge, MA. Paper 7991, Institute for the Study of Labor, Bonn. INSTAT (National Institute of Statistics, Madagascar). Rogerson, R., R. Shimer, and R. Wright. 2005. “Search- 2013. Enquête National sur l’Emploi et le Secteur Theoretic Models of the Labor Market: A Survey.” Informel: ENEMPSI 2012. Antananarivo: INSTAT. Journal of Economic Literature 43 (December): Irinnews.org. 2013. “Microcredit helps small businesses 959–88. buck the system in Madagascr.” May 16. http://www Sy and Diatta. 2014. ”Jobs in Madagascar.” Unpublished .irinnews.org/news/2013/05/16/microcredit-helps- working paper. World Bank. small-businesses-buck-system-madagascar. Van Biesebroeck, J. 2005. “Firm Size Matters: Growth LaPorta, R., and A. Shleifer. 2014. “Informality and and Productivity Growth in African Manufacturing.” Development.” Working Paper No. 20205. National Economic Development and Cultural Change 53 (3): Bureau of Economic Research, Cambridge, MA. 545–583. Li, Y., and M. Rama. 2015. “Firm Dynamics, Productivity World Bank. 2015a. Madagascar: Diagnostic Trade Growth, and Job Creation in Developing Countries: Integration Study—Levelling the Playing Field for The Role of Micro- and Small Enterprises.” World Renewed and Inclusive Growth. Internal draft for Bank Research Observer 30 (1): 3–38. review. Washington, DC: World Bank. World Bank. 2015b. Madagascar: A Systematic Country Diagnostic. Washington, DC: World Bank.