Volume 27 • Number 3 • 2013 ISSN 0258-6770 (PRINT) ISSN 1564-698X (ONLINE) THE WORLD BANK ECONOMIC REVIEW Volume 27 • 2013 • Number 3 THE WORLD BANK ECONOMIC REVIEW Water Nationalization and Service Quality Fernando Borraz, Nicolás González Pampillón, and Marcelo Olarreaga Mass Media and Public Policy: Global Evidence from Agricultural Policies Alessandro Olper and Johan Swinnen Liability Structure in Small-Scale Finance: Evidence from a Natural Experiment Fenella Carpena, Shawn Cole, Jeremy Shapiro, and Bilal Zia SYMPOSIUM ON FINANCIAL STRUCTURE AND DEVELOPMENT GUEST ¸, and Justin Yifu Lin EDITORS: Robert Cull, Asli Demirgüc Financial Structure and Economic Development: A Reassessment ¸-Kunt, and Justin Yifu Lin Robert Cull, Asli Demirgüc The Evolving Importance of Banks and Securities Markets Asli Demirgüc¸-Kunt, Erik Feyen, and Ross Levine Notes on Financial System Development and Political Intervention Fenghua Song and Anjan Thakor Financial Development: Structure and Dynamics Pages 389–572 Augusto de la Torre, Erik Feyen, and Alain Ize Job Growth and Finance: Are Some Financial Institutions Better Suited to the Early Stages of Development than Others? Robert Cull and L. Colin Xu www.wber.oxfordjournals.org 2 THE WORLD BANK ECONOMIC REVIEW editor Andrew Foster, Brown University co-editors Peter Lanjouw and Luis Servén, World Bank The editorial team would like to thank former editors Elisabeth Sadoulet and Alain de Janvry for overseeing the review process of the articles in this issue. assistant to the editor Marja Kuiper editorial board Harold H. Alderman, World Bank (retired) Caroline Freund, World Bank Chong-En Bai, Tsinghua University, China Paul Glewwe, University of Minnesota, USA Pranab K. Bardhan, University of California, Philip E. Keefer, World Bank Berkeley, USA Norman V. Loayza, World Bank Kaushik Basu, World Bank Jeremy Magruder, University of California, Thorsten Beck, Tilburg University, Berkeley, USA Netherlands William F. Maloney, World Bank Johannes van Biesebroeck, K.U. Leuven, David J. McKenzie, World Bank Belgium Jaime de Melo, University of Geneva, Maureen Cropper, University of Maryland, Switzerland USA Ugo Panizza, UNCTAD Asli Demirgüç-Kunt, World Bank Nina Pavcnik, Dartmouth College, USA Jean-Jacques Dethier, World Bank Vijayendra Rao, World Bank Quy-Toan Do, World Bank Martin Ravallion, Georgetown University, Frédéric Docquier, Catholic University of USA Louvain, Belgium Jaime Saavedra-Chanduvi, World Bank Eliana La Ferrara, Università Bocconi, Italy Claudia Sepúlveda, World Bank Francisco H. G. Ferreira, World Bank Dominique Van De Walle, World Bank Augustin Kwasi Fosu, United Nations Christopher M. Woodruff, University of University, WIDER, Finland California, San Diego, USA The World Bank Economic Review is a professional journal used for the dissemination of research in development economics broadly relevant to the development profession and to the World Bank in pursuing its development mandate. It is directed to an international readership among economists and social scientists in government, business, international agencies, universities, and development research institutions. The Review seeks to provide the most current and best research in the field of quantita- tive development policy analysis, emphasizing policy relevance and operational aspects of economics, rather than primarily theoretical and methodological issues. Consistency with World Bank policy plays no role in the selection of articles. The Review is managed by one or two independent editors selected for their academic excellence in the field of development economics and policy. The editors are assisted by an editorial board composed in equal parts of scholars internal and external to the World Bank. World Bank staff and outside researchers are equally invited to submit their research papers to the Review. For more information, please visit the Web sites of the Economic Review at Oxford University Press at www.wber.oxfordjournals.org and at the World Bank at www.worldbank.org/research/journals. Instructions for authors wishing to submit articles are available online at www.wber.oxfordjournals.org. Please direct all editorial correspondence to the Editor at wber@worldbank.org. SUBSCRIPTIONS:A subscription to The World Bank Economic Review (ISSN 0258-6770) comprises 3 issues. Prices include postage; for subscribers outside the Americas, issues are sent air freight. Annual Subscription Rate (Volume 27, 3 Issues, 2013): Institutions—Print edition and site-wide online access: £184/ $277/E277, Print edition only: £169/$255/E254, Site-wide online access only: £154/$232/E231; Corporate—Print edition and site-wide online access: £277/$415/E415, Print edition only: £254/$381/E380, Site-wide online access only: £230/$346/E347; Personal—Print edition and individual online access: £45/$67/E67. US$ rate applies to US & Canada, EurosE applies to Europe, UK£ applies to UK and Rest of World. There may be other subscription rates available; for a complete listing, please visit www.wber.oxfordjournals.org/subscriptions. Readers with mailing addresses in non-OECD countries and in socialist economies in transition are eligible to receive complimentary subscriptions on request by writing to the UK address below. Full prepayment in the correct currency is required for all orders. Orders are regarded as firm, and payments are not refundable. Subscriptions are accepted and entered on a complete volume basis. Claims cannot be con- sidered more than four months after publication or date of order, whichever is later. All subscriptions in Canada are subject to GST. Subscriptions in the EU may be subject to European VAT. If registered, please supply details to avoid unnecessary charges. For subscriptions that include online versions, a proportion of the subscription price may be subject to UK VAT. Personal rates are applicable only when a subscription is for individual use and are not available if delivery is made to a corporate address. The current year and two previous years’ issues are available from Oxford University Press. Previous BACK ISSUES: volumes can be obtained from the Periodicals Service Company, 11 Main Street, Germantown, NY 12526, USA. E-mail: psc@periodicals.com. Tel: (518) 537-4700. Fax: (518) 537-5899. CONTACT INFORMATION: Journals Customer Service Department, Oxford University Press, Great Clarendon Street, OxfordOX2 6DP, UK. E-mail: jnls.cust.serv@oup.com. Tel: þ44 (0)1865 353907. Fax: þ 44 (0)1865 353485. In the Americas, please contact: Journals Customer Service Department, Oxford University Press, 2001 Evans Road, Cary, NC 27513, USA. E-mail: jnlorders@oup.com. Tel: (800) 852-7323 (toll-free in USA/Canada) or (919) 677-0977. Fax: (919) 677-1714. In Japan, please contact: Journals Customer Service Department, Oxford University Press, Tokyo, 4-5-10-8F Shiba, Minato-ku, Tokyo, 108-8386, Japan. E-mail: custserv.jp@oup.com. Tel: þ 81 3 5444 5858. Fax: þ 81 3 3454 2929. POSTAL INFORMATION: The World Bank Economic Review (ISSN 0258-6770) is published three times a year, in February, June, and October, by Oxford University Press for the International Bank for Reconstruction and Development/THE WORLD BANK. Send address changes to The World Bank Economic Review, Journals Customer Service Department, Oxford University Press, 2001 Evans Road, Cary, NC 27513-2009. Periodicals postage paid at Cary, NC and at additional mailing offices. Communications regarding original articles and editorial management should be addressed to The Editor, The World Bank Economic Review, The World Bank, 3, Chemin Louis Dunant, CP66 1211 Geneva 20, Switzerland. ENVIRONMENTAL AND ETHICAL POLICIES: Oxford Journals, a division of Oxford University Press, is committed to working with the global community to bring the highest quality research to the widest possible audience. Oxford Journals will protect the environment by implementing environmentally friendly policies and practices wherever possible. Please see http://www.oxfordjournals.org/ethicalpolicies.html for further information on environmental and ethical policies. DIGITAL OBJECT IDENTIFIERS: For information on dois and to resolve them, please visit www.doi.org. PERMISSIONS: For information on how to request permissions to reproduce articles or information from this journal, please visit www.oxfordjournals.org/jnls/permissions. ADVERTISING: Advertising, inserts, and artwork enquiries should be addressed to Advertising and Special Sales, Oxford Journals, Oxford University Press, Great Clarendon Street, Oxford, OX2 6DP, UK. Tel: þ 44 (0)1865 354767; Fax: þ 44(0)1865 353774; E-mail: jnlsadvertising@oup.com. DISCLAIMER: Statements of fact and opinion in the articles in The World Bank Economic Review are those of the respective authors and contributors and not of the International Bank for Reconstruction and Development/THE WORLD BANK or Oxford University Press. Neither Oxford University Press nor the International Bank for Reconstruction and Development/THE WORLD BANK make any representation, express or implied, in respect of the accuracy of the material in this journal and cannot accept any legal responsibility or liability for any errors or omissions that may be made. The reader should make her or his own evaluation as to the appropriateness or otherwise of any experimental technique described. The World Bank Economic Review is printed on acid-free paper that meets the minimum require- PAPER USED: ments of ANSI Standard Z39.48-1984 (Permanence of Paper). INDEXING AND ABSTRACTING: The World Bank Economic Review is indexed and/or abstracted by CAB Abstracts, Current Contents/Social and Behavioral Sciences, Journal of Economic Literature/EconLit, PAIS International, RePEc (Research in Economic Papers), and Social Services Citation Index. COPYRIGHT # 2013 The International Bank for Reconstruction and Development/THE WORLD BANK All rights reserved; no part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise without prior written permission of the publisher or a license permitting restricted copying issued in the UK by the Copyright Licensing Agency Ltd, 90 Tottenham Court Road, London W1P 9HE, or in the USA by the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923. Typeset by Techset Composition Limited, Chennai, India; Printed by The Sheridan Press. THE WORLD BANK ECONOMIC REVIEW Volume 27 † 2013 † Number 3 Water Nationalization and Service Quality 389 Fernando Borraz, Nicola ´ s Gonza´ lez Pampillo ´ n, and Marcelo Olarreaga Mass Media and Public Policy: Global Evidence from Agricultural Policies 413 Alessandro Olper and Johan Swinnen Liability Structure in Small-Scale Finance: Evidence from a Natural Experiment 437 Fenella Carpena, Shawn Cole, Jeremy Shapiro, and Bilal Zia SYMPOSIUM ON FINANCIAL STRUCTURE AND DEVELOPMENT ¨c GUEST EDITORS: Robert Cull, Asli Demirgu ¸ , and Justin Yifu Lin Financial Structure and Economic Development: A Reassessment 470 Robert Cull, Asli Demirgu ¸ -Kunt, and Justin Yifu Lin ¨c The Evolving Importance of Banks and Securities Markets 476 Asli Demirgu ¸ -Kunt, Erik Feyen, and Ross Levine ¨c Notes on Financial System Development and Political Intervention 491 Fenghua Song and Anjan Thakor Financial Development: Structure and Dynamics 514 Augusto de la Torre, Erik Feyen, and Alain Ize Job Growth and Finance: Are Some Financial Institutions Better Suited to the Early Stages of Development than Others? 542 Robert Cull and L. Colin Xu Water Nationalization and Service Quality ´ lez Pampillo ´ s Gonza Fernando Borraz, Nicola ´ n, and Marcelo Olarreaga The objective of this paper is to explore the impact of Uruguay’s privatization and subsequent nationalization of water services on network access and water quality. The results suggest that although the early privatization of water services had little impact on access to the sanitation network, the subsequent nationalization led to an increase in network access at the bottom of the income distribution as well as an improvement in water quality. JEL codes: D60, H51, I10, I30, L33, O12 In the 1970s, the government’s role as the provider of basic services in sectors with a natural monopoly component, such as the water supply and sanitation, was rarely questioned. Indeed, it was thought that private firms were likely to abuse their monopoly power in this type of market because they would concen- trate their supply on rich households, leaving poor households without access to basic services. However, public companies would have incentives to ensure access to the maximum number of potential voters (at least in democracies). By the late 1980s, the weak economic performance and low productivity of many public companies around the world changed this view (World Bank 2004). Poor management practices due to political agendas rather than profit- oriented motives shed light on the substantial inefficiencies and poor service quality provided by these public companies. In the 1990s, the privatization of water services was perceived by some as a potential solution to the poor perfor- mance of publicly owned water monopolies that left more than one billion people in developing countries without access to clean and safe water and 40 percent of the world’s population without access to safe and clean sanitation services (Segeredt 2005). In its Human Development Report of 2006, the UNDP notes, “not having access to water and sanitation is a polite euphemism for a form of deprivation that threatens life.” Additionally, Galiani et al. Fernando Borraz (corresponding author) is a researcher at the Banco Central del Uruguay and a professor at the Deparmento de Economı ´a FCS-UDELAR; his email address is fborraz@bcu.gub.uy. Nicola´ s Gonza ´ n is graduate student at CEMFI, Spain; his email address is ´ lez Pampillo ngonzalez@correo.um.edu.uy. Marcelo Olarreaga is a professor at the University of Geneva and a researcher at CEPR; his email address is marcelo.olarreaga@unige.ch. THE WORLD BANK ECONOMIC REVIEW, VOL. 27, NO. 3, pp. 389 –412 doi:10.1093/wber/lht001 Advance Access Publication January 29, 2013 # The Author 2013. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 389 390 THE WORLD BANK ECONOMIC REVIEW (2005) provide empirical evidence that a transition toward privatized water services in Argentina in the 1990s led to a rapid decline in child mortality. France was an early example of a country with privately provided water services. Throughout the 1990s, many countries privatized water services, beginning with England in 1989 and followed by Eastern European and Latin American countries. A few Asian and African countries followed in the mid- and late 1990s (Hall, Lobina, and Corral 2010). Nevertheless, the share of public water companies is still very large.1 This share has increased over the last 10 years as negative reactions to privatization occurred in many countries. Uruguay is a recent example of this reversal. Until 1993, all water services in Uruguay were publicly owned, with the exception of a few small, community- based private companies that operated in the Department of Canelones since the 1940s in areas that the public company did not serve.2 In 1993, the initial wave of privatization was implemented in the Department of Maldonado, affecting approximately 3,000 customers. A second, larger wave of privatiza- tion followed in the same department, affecting over 20,000 customers, includ- ing some in the capital of Maldonado. The privatization was reversed in 2004 when an amendment to the Uruguayan constitution was passed declaring water “part of the public domain.” The private provision of water was made illegal.3 The apparent reasons for Uruguay’s renationalization of water companies were no different from those observed in other Latin American countries in the last decade (e.g., Bolivia, Argentina, Brazil). The privatization of water services did not achieve the promised results.4 Private companies became deeply unpop- ular due to perceptions of low or declining water quality as well as the high prices charged by private companies. A series of highly publicized episodes of low quality water supplied by Uragua and Aguas de la Costa (subsidiaries of the Spanish water companies Aguas de Barcelona and Aguas de Bilbao) led (as early as 2003, in the middle of a financial crisis) to the well-publicized request by the then-Minister of Economics and Finance, Alejandro Atchugarry, for Uragua to leave the country. Whether public or private water provision leads to better access and quality is an empirical question. The objective of this paper is to explore the impact of the privatization and subsequent nationalization of water services on water 1. According to Hall, Lobina, and Corral (2010), water services are owned and run by the public sector in 90 percent of the 400 largest cities in the world. This figure should be compared to the share of formal employment in public companies across all sectors, which is 5 percent, on average, according to Kikeri (1999). 2. For an assessment of the performance of community-driven water providers in developing countries, see Whittington et al. (2009). 3. Water privatization was also made illegal in the Netherlands. 4. The reason for the failure of privatization is not necessarily inherent to privatization but may also be explained by poorly designed contracts (in terms of required investments) or inadequate regulatory bodies. These are often associated with problems due to corruption (see Chong and Lopez-de-Silanes 2005). ´ lez Pampillo Borraz, Gonza ´ n, and Olarreaga 391 quality (microbiological and inorganic tests) and access to sanitation networks (the percentage of households with water-sealed toilets connected to sewer lines) in Uruguay. Studying Uruguay’s water services is interesting because household access to Uruguay’s piped sewerage network is particularly low compared to countries at similar levels of development. With an access rate below 50 percent, Uruguay performs poorly relative to other Latin American countries, such as Chile, Colombia, Mexico, and some comparable Brazilian states, in terms of income per capita. Additionally, as in many other Latin American countries, issues re- garding the quality of water provided by private companies were part of the reason for nationalization. The existing empirical evidence on the impact of privatization on water quality and access is relatively small and tends to suggest that privatization either has a positive impact or no impact. Barrera-Osorio, Olivera, and Ospino (2009) use a difference-in-difference methodology to assess the impact of water privatization in Colombia on several outcomes, such as coverage (the percent- age of households connected to water services) and water quality (the frequen- cy of the service and characteristics of the water, such as its color). They find that in urban areas, water access increases and water quality improves as a result of privatization, whereas negative effects on access are detected in rural areas. This finding is consistent with the notion that as water services are pri- vatized, the poorest consumers may be left behind. In a developed country context, Wallsten and Kosec (2005) analyze the effects of water ownership on water quality, measured as the number of viola- tions of the Safe Drinking Water Act in the United States between 1997 and 2003. Using panel data at the community level and controlling for community fixed effects, they find that ownership is irrelevant with respect to compliance with the Safe Drinking Water Act. This result may have been observed be because the study was conducted in a developed country with high income levels and a strong demand for high quality water.5 Moreover, an important body of literature examines the impact of the pri- vatization of water services on child mortality in developing countries. Child mortality is an indirect measure of water quality, but increases in water access and quality have been demonstrated to be negatively associated with child mor- tality (Lee, Rosenzweig, and Pitt 1997; Shi 2000; Galdo and Bertha 2005). Using a panel data framework, Galiani, Gertler and Schargrodsky (2005) provide convincing evidence that in Argentinean municipalities where water services were privatized, the incidence of child mortality from water-related diseases declined significantly (whereas the incidence of child mortality for 5. In their review of the literature, Nauges and Whittington (2009) suggest that income elasticities of demand oscillate between 0.1 and 0.4. 392 THE WORLD BANK ECONOMIC REVIEW other reasons remained stagnant). They therefore provide indirect evidence of improvements in water quality and access.6 The empirical methodology that we followed is similar to that employed in Galiani, Gertler, and Schargrodosky (2005). Using panel data for the periods surrounding the episodes of privatization and nationalization, we identify differences in sanitation rates between regions that first privatized their water suppliers and later nationalized them and those that did not using a difference-in-difference estimator.7 Our study has at least two important differ- ences from Galiani, Gertler and Schargrodosky (2005). First, as explained above, we focus on the nationalization of water services, not only on their pri- vatization. Second, our dependent variables are direct measures of the quality of service provided by water companies (access to and quality of water). Health outcomes are important, particularly child mortality, but they tend to be determined by many external factors other than water services. In a middle-income country such as Uruguay, child mortality is rarer than in low- income countries. Therefore, water quality may improve dramatically without any observable change in child mortality.8 Direct measures of the quality of water services seem more appropriate in such a context. Our results suggest that although the earlier privatization period had little impact in terms of access to the sanitation network, the nationalization of water services had a positive and statistically significant impact on access, par- ticularly among households in the bottom 25 percent of the income distribu- tion. Nationalization also seems to have led to improved water quality. Indeed, the impact of nationalization on the detection of abnormal levels in microbio- logical and inorganic water tests is consistently negative and has a relatively large coefficient. It is important to note that although it may be tempting to conclude from our results that the public sector can provide water services as well as or better than the private sector, this conclusion cannot be supported on the basis of our empirical evidence. Indeed, the control group in our difference-in-difference strategy includes cities that were consistently served by public companies, making it impossible to answer such a question.9 What the results suggest is that the privatization of water companies had little impact in terms of network 6. Note that the same critiques of private water suppliers that were present in Uruguay in the late 1990s and early 2000s were also made in the Argentinean press at the time (high prices, water provided by private companies being unfit for human consumption, or the fact that these private companies only honor half of their investment commitments). 7. Because the data on water quality only span the nationalization period, we cannot measure the impact of the privatization of water services. 8. Over the last decade, the average number of child deaths in Uruguay was 10 per jurisdiction. Of those, less than 10 percent were water related. See Borraz and Olarreaga (2011). The reason the effect of privatization on child mortality was estimated as being relatively large in Argentina, which is also a middle-income country, is that although the two countries have similar levels of income per capita, Uruguay is much more homogenous than Argentina in terms of both income and race. 9. We are grateful to a referee for this clarification. ´ lez Pampillo Borraz, Gonza ´ n, and Olarreaga 393 access, confirming the public opinion that privatization did not deliver the promised results. However, the subsequent nationalization of the water compa- nies delivered progress in terms of both network access and water quality rela- tive to those companies that were consistently publicly owned. This finding contradicts most of the existing evidence for developing countries, which gener- ally shows that water privatization leads to improved service quality. The reminder of the paper is organized as follows. The next section de- scribes the water system in Uruguay. Section II discusses the empirical method- ology, and section III presents the data and some descriptive statistics. Section IV presents the results, and section V concludes. I . T H E WA T E R S Y S T E M IN UR UGUAY Until the early 1990s, water and sewage services in most of the country were exclusively provided by a publicly owned company, called Obras Sanitarias del Estado (OSE), except for a series of small, community-based providers that were created in the 1940s by residents with the objective of fostering land sales in the area of El Pinar in the Department of Canelones. The largest of these community-based providers was Aguas Corrientes El Pinar, which served fewer than 1,000 clients in the 1990s and only had 12 full-time staff. In 1993, the first privatization of water services previously provided by OSE occurred in the Department of Maldonado. The new company, Aguas de la Costa, supplied water and sewage in the wealthier areas of La Barra, Manantiales, and Jose ´ Ignacio, which are to the east of the internationally known resort of Punta del Este. This was a joint venture between a local company, S.T.A. Ingenieros and Benecio S.A., which had approximately 10 percent ownership, and the Spanish company Aguas de Barcelona, a subsidiary of Suez Lyonaisse des Eaux, which owned the rest of the company. Aguas de la Costa signed a 25-year concession in 1993. The company had approximately 3,100 customers. A second water service privatization occurred in 2000. Uragua began provid- ing water and sewage services in urban and suburban areas of Maldonado. Maldonado has 150,000 of the 3,300,000 inhabitants of Uruguay. Uragua served the area west of the Maldonado stream, with the exception of the city of Aigua ´ . Specifically, Uragua served the capital of Maldonado (50,417 inhabi- tants), San Carlos (23,878), Pan de Azu ´ polis (7,579), Cerros ´ car (6,969), Piria Azules, Nueva Carrara, Pueblo Gerona, West of River Solı ´s, Silver River, Highway 9 North, and Punta del Este (7,298 inhabitants).10 Uragua was owned by Aguas de Bilbao, a Spanish water provider. Thus, by 2000, only the city of Aigua ´ in Maldonado was served by OSE, and all other jurisdictions in Maldonado were served by two private 10. We consider all year-round residents and not the tourist population, which can reach hundreds of thousands during the summer. 394 THE WORLD BANK ECONOMIC REVIEW companies, Uragua and Aguas de la Costa. OSE had an exclusive monopoly in the rest of the country, except for the area of El Pinar, which was served by small, community-based providers. OSE accounted for over 90 percent of all connections. In 2004, Uragua began litigation with OSE because of a breach of contract following some well-publicized episodes of colored water in Maldonado. This led to a referendum, which declared water to be part of the public domain, and an amendment to the Uruguayan constitution. Uragua and the Uruguayan government reached an agreement, and all of Uragua’s assets were transferred to OSE by the end of 2005. After the agreement with the government, the company left the country in 2005. Aguas de la Costa’s assets were transferred to OSE in 2005. Aguas Corrientes el Pinar was nationalized in December 2006, and its assets were also transferred to OSE. Maldonado was particularly affected by the return to nationalization. At the turn of the century, the only city in Maldonado served by the publicly owned OSE was Aigua ´ , with 2,676 inhabitants. By 2006, OSE was the only provider of water services in Maldonado. Our empirical analysis will therefore focus on the change of ownership in the Department of Maldonado because we do not have data at the city level for the small, community-based providers in the Department of Canelones. As discussed in the data section, when examining the impact of the privatization and subsequent nationalization of water services on network access, our treated group is restricted to three cities due to data availability, Maldonado, Pan de Azu ´ car, and San Carlos. We use 32 cities as a control group, including the 19 departmental capitals (except Maldonado) and other large cities. In our examination of the impact on water quality, we use six treated cities in the Department of Maldonado: Maldonado, Pan de Azu ´ car, Piria´ polis, Punta Ballena, Punta del Este, and San Carlos. The control group has 26 cities, including all departments’ capitals except Maldonado. Table 1 presents the evolution of some indicators for the public company OSE before and after nationalization. Note that the number of employees did not change substantially following nationalization. This is important because we do not wish to attribute any change in the performance of the water com- panies to a change in the composition of the workforce due to layoffs at the time of nationalization. The absence of layoffs was confirmed through inter- views with the managers of the water company, which indicated that only the top managers of the nationalized companies were removed by OSE. As expect- ed, the volume of water produced increased following nationalization because the water service coverage increased. Moreover, the population served with sewerage rose from 531,300 in 2004 to 729,100 in 2006. Total gross fixed assets (including work in progress) increased sharply (by approximately 38 percent) between 2004 and 2006. This increase was partly due to the acquisi- tions of the private companies and partly the result of new infrastructure in- vestments made through a sanitation project supported by the World Bank T A B L E 1 . Indicators of the Public Company: OSE Units 2002 2003 2004 2005 2006 Service area Total population - water supply ’000 inhabitants 3,158.4 3,178.2 3,100.7 3,245.0 3,253.7 Total population - wastewater ’000 inhabitants 1,791.6 1,805.9 1,541.2 1,919.0 1,935.7 Total number of staff # FTE 4.816 4.508 4.362 4.174 4.280 Water service Population served - water ’000 inhabitants 3,041.6 3,063.8 2,833.5 2,980.9 3,055.3 Water connections year end ’000 connections 735.1 732.9 737.7 799.9 822.5 Volume of water produced Million m3/year 282.7 275.1 288.1 309.4 320.0 Total volume of water sold Million m3/year 135.3 131.0 132.3 142.0 147.1 Sewerage Service Population served - sewerage ’000 inhabitants 467.8 479.7 531.3 633.8 729.1 Borraz, Gonza Sewerage connections year end ’000 connections 162.2 166.1 169.5 195.6 205.9 Length of sewers Km 1,683.0 1,759.6 1,872.5 2,410.8 2,494.5 Financial information Total operating revenues Million URY of 2010 3,951.3 6,015.2 6,036.5 627.3 6,848.2 Total billings to residential customers Million URY of 2010 2,252.1 3,417.8 3,537.0 3,718.0 4,070.0 ´ lez Pampillo Total billings to industrial customers Million URY of 2010 840.9 12,19.0 1,269.5 1,340.1 1,499.5 Total water and wastewater operational expenses Million URY of 2010 2,499.8 3,736.6 3,733.9 3,689.6 4,010.7 Labor costs Million URY of 2010 1,342.1 1,851.9 1,842.9 1,688.4 1,848.8 Total gross fixed assets including work in progress Million URY of 2010 17,907.5 24,051.4 2,356.2 24,330.8 2,4604.8 ´ n, and Olarreaga Tariff information Fixed charge per month for water and wastewater URY of 2010 per month 86.4 140.7 144.5 145.7 145.0 services for residential customers Connection charges - water URY of 2010 1,414.1 1,814.4 1,803.5 1,891.0 1,957.6 Connection charges - sewers URY of 2010 567.3 726.7 722.3 757.1 784.3 Source: OSE. 395 396 THE WORLD BANK ECONOMIC REVIEW following nationalization. Finally, OSE’s prices increased sharply in 2003 during the privatization period and then remained stable following nationaliza- tion. Note that these are the prices of the public company, but during the pri- vatization period, there were no major differences between the prices of public or private companies because the prices were controlled by the Uruguay Energy and Water Services regulator (URSEA). II. ME T H O D O LO GY The identification strategy we followed is similar to that employed by Galiani et al. (2005), who searched for systematic differences in changes in child mor- tality rates between regions that have privatized and those that have not changed the ownership structure of water companies using a difference-in- difference approach.11 We followed their approach but used a double treatment that included the privatization of water providers in some cities and their sub- sequent renationalization. Selection bias due to cherry picking at the moment of privatization is an issue in this framework. Governments may decide to privatize companies in cities that are more profitable and have better prospects to maximize the short- run financial benefits of privatization. If profitability positively affects perfor- mance over time (i.e., performance is serially correlated), we will observe that nationalization will lead to a better-performing water company, but this effect will simply be due to the trend in the performance of water companies in that region. In this context, even if nationalization adversely affected performance, the estimation might not identify this effect because the treatment group in- cludes a disproportionate number of utilities that perform well. We address this issue, as we discuss in the data section below, using parallel trend tests for the outcome variables (access to sewage network and water quality) before privatization. In the absence of any difference in trends before privatization, there would be little evidence of cherry picking along these di- mensions, and time fixed effects in our difference-in-difference specification can therefore control for these common trends. Our econometric model is given by yu ¼ fNit þ gPit þ x0it b þ at þ ai þ uit ð1Þ where yit is the outcome variable in city i in year t. We consider two different outcome variables: sanitation rates and water quality. The units of observation i are cities. The parameter Nit is a dummy variable that is equal to one after 11. Lee, Rosenzweig, and Pitt (1997) show that this reduced form approach may downward bias the effects of the treatment because households adjust their behavior to the new environment and provide an alternative structural approach to estimate the impact of improved quality on health outcomes. ´ lez Pampillo Borraz, Gonza ´ n, and Olarreaga 397 the renationalization in the three cities that were previously privatized. The parameter Pit is a dummy variable indicating that the water company in city i in period t is privately owned; x’it is a vector of control variables; b is the corre- sponding vector of coefficients, at is a year effect; ai is a city-specific fixed effect; and uit is a city time-varying error (distributed independently across departments and time). The parameters of interest are f and g, which measure the impact of the dual treatment of the privatization and the subsequent nationalization of water services. When the sanitation rate is the outcome variable, a positive value for f or g indicates that the nationalization or privatization of private companies led to higher sanitation rates relative to companies that were consistently publicly owned. The water quality sample does not cover the preprivatization period, as dis- cussed in the next section. Thus, when considering water quality tests (abnor- mal levels of microbiological and organoleptic elements) as the dependent variable, we only have one treatment, Nit. A positive coefficient implies that na- tionalization led to a higher number of tests with abnormal results relative to companies that were consistently publicly owned. An important issue in panel data models is that observations tend to be cor- related across time within individual cities. One possible solution to the serial correlation problem is to use robust standard errors clustered at the city level. In this context, asymptotic statistical inference depends on the number of clus- ters and time periods. A small number of clusters may result in biased (clus- tered) standard errors, tending to underestimate inference precision. Bertrand, Duflo, and Mullainthan (2004) analyze the performance of the following alternative solutions to the serial correlation problem: 1) using para- metric methods, that is, specifying an autocorrelation structure; 2) using the block bootstrap; 3) ignoring time series information, that is, averaging the data before and after treatment; and 4) using an empirical variance-covariance matrix. Bertrand, Duflo and Mullainthan (2004) find that the empirical variance-covariance matrix outperforms the others, but it does not work prop- erly with small samples. Cameron, Gelbach, and Miller (2008) state that the block bootstrap works properly with small numbers of groups.12 In our case, bootstrapped, clustered standard errors are similar to clustered standard errors. In most cases, the latter method reports greater standard errors, and in some cases, conventional standard errors are larger than the latter method. As a conservative criterion, we decided to report the method with the largest standard errors in each case. Depending on the nature of the outcome variable, continuous, count, or fractional, we use different estimators that will be discussed in the results section. 12. An alternative solution is to use the method proposed by Bell and McCaffrey (2002), “bias correction of clustered standard errors,” but this approach unfortunately cannot be applied in a difference-in-difference framework. 398 THE WORLD BANK ECONOMIC REVIEW I I I . D ATA S O U R C E S AND VA R I A B L E C O N S T R U C T I O N We begin by describing the data sources and variable construction for the anal- ysis of the impact of nationalization on access to water sanitation networks. We then turn to the data sources and variable construction for the analysis of the impact of nationalization on water quality. Access to Sanitation Networks The data regarding the percentage of households with water-sealed toilets con- nected to sewer lines are obtained from the annual Uruguayan national house- hold survey, Encuesta Continua de Hogares (ECH), conducted by the National Statistical Office of Uruguay, Instituto Nacional de Estadı ´stica. The ECH is the main source of socioeconomic information on Uruguayan households and their members at the national level. The surveys were conducted throughout the year with the objective of generating a description of the socioeconomic situation of the entire population. The ECH also includes questions about household living conditions. In par- ticular, the survey asks whether water-sealed toilets are connected to sewer lines. Therefore, we generate a dummy variable that takes the value one if the household is connected to sewer lines and zero otherwise. Then, we aggregate the data by city to obtain the percentage of households with sanitation access in each city. We therefore consider panel data by city from 1986 to 2009. The time span includes the preprivatization period in the case of Maldonado because the two privatizations in Maldonado occurred in 1993 and 2000, as discussed above. The ECH survey is only representative at the department level or at the city level for the largest cities in terms of population. Therefore, and to ensure a representative sample, we only retain the capital cities of the different depart- ments and other large cities in our sample. We have a total of 35 cities, three in the treatment group (Maldonado, Pan de Azu ´ car, and San Carlos) and 32 in the control group (Artigas, Bella Unio ´ n, Canelones, Carmelo, Colonia, Dolores, Durazno, Florida, Fray Bentos, Lascano, Libertad, Melo, Mercedes, Minas, Montevideo, Paso de los Toros, Paysandu ´ , Periferia Canelones, Rivera, Rocha, Rosario, Rı ´o Branco, Salto, San Jose ´ de Mayo, San Ramo ´ n, Santa Lucı´a, Sarandı´ del Yı ´, Sarandı ´ Grande, Tacuarembo ´ , Tranqueras, Treinta y Tres, Trinidad, and Young).13 Because some of the selected cities were not surveyed in certain years ( pri- marily in the older edition of the ECH), we have an unbalanced panel that may imply panel attrition bias. For instance, of the 24 time periods in the treatment group, Pan de Azu ´ car appears 15 times, and San Carlos appears 18 times. In the control group, we have observations for Lascano in 10 of the 24 time periods; for Bella Unio ´ n, Libertad, and Rosario in 12 of the years; for Santa 13. Of these 35 cities, only 19 are capital cities. ´ lez Pampillo Borraz, Gonza ´ n, and Olarreaga 399 Lucı ´a in 15 of the years; for Carmelo, Dolores, Paso de los Toros, Rı ´o Branco, San Ramo ´ del Yı ´ n, Sarandı ´, and Sarandı ´ Grande in 17 of the years; and for Young in 18 of the years. Therefore, we have a total of 735 observations.14 Some cities began to appear in the ECH because of their rapid population growth; therefore, we checked the robustness of results to a smaller subsample in terms of the time span from 1993 to 2009. Two observations must be made. First, there are far more control cities than treatment cities; hence, we also estimate our model restricting the number of controls to capital cities alone. The sanitation data for each capital city are available in every year of the full period; hence, in this instance, the panel attri- tion problems may be solved. Second, because we lose observations as we drop cities, small sample bias is a potential issue. Thus, a tradeoff exists between small sample bias and the potential for panel attrition bias, which provides some robustness checks for our results. The top panel of table 2 provides descriptive statistics by treatment and control groups for the network access sample before and after privatization and nationalization. Overall, the treatment and control groups present similar characteristics with relatively small differences, on average, although they tend to be statistically significant. It will therefore be important to control for these characteristics in our econometric analysis.15 More important, the network access rates are not significantly different between treated and control cities before privatization and before nationalization. However, treated cities have a statistically larger network access rate following nationalization. Whether this larger network access rate can be attributed to a causal effect will be addressed using the difference-in-difference method described in the previous section. As discussed above, a concern one may have with our methodology is that although, on average, network access rates in control and treated cities did not differ before privatization (and nationalization), they may have had different trends, which would bias our estimates of the impacts of privatization and na- tionalization. To address this issue, we performed a test of parallel trends for the period before the privatization. Thus, we introduce a time trend in the spec- ification of equation (1) and check for its statistical significance. The results of the different specifications and samples are reported in table 3. In all columns except D, the coefficient on the time trend for treated cities is statistically insig- nificant, suggesting that treated and control cities had common trends before and after privatization.16 In column D, where we only use departmental capi- tals as treated and control cities, the coefficient is negative and statistically sig- nificant. This result would be a concern if we were to find that the 14. There are 35 cities. We have 21 cities times 24 years (1986–2009 period), yielding 504 observations. In addition, we have 1  10 þ 3  12 þ 2  15 þ 7  17 þ 2  18 ¼ 231 observations. 15. These differences reflect that only a few cities’ water services were privatized and subsequently nationalized. 16. We obtain qualitatively identical results from a parallel test in the trends for the period before the nationalization, which are available upon request. 400 T A B L E 2 . Descriptive Statistics for Treatment and Control Groups Publicly owned period (1986– 2000) Privately owned period (2001 – 2005) Publicly owned period (2006 – 2009) THE WORLD BANK ECONOMIC REVIEW Control Difference Difference Difference Treated (1) (2) (1) 2 (2) Treated (1) Control (2) (1) 2 (2) Treated (1) Control (2) (1) 2 (2) Access to sewage network sample Network evacuation rate 0.47 (0.10) 0.44 (0.19) 0.03 0.53 (0.07) 0.49 (0.21) 0.04 0.69 (0.10) 0.54 (0.22) 0.15** Education (head 6.78 (0.42) 6.58 (0.77) 0.20* 7.22 (0.55) 7.18 (0.73) 0.04 7.82 (0.38) 7.77 (0.75) 0.05 of household) (log) Household per capita 7.90 (0.10) 7.58 (0.25) 0.32*** 7.52 (0.14) 7.43 (0.22) 0.09* 7.83 (0.18) 7.63 (0.21) 0.20*** income (log) Accumulated 6.95 (0.17) 7.11 (0.23) 2 0.16*** 7.09 (0.20) 7.24 (0.27) 2 0.15** 6.97 (0.17) 7.07 (0.31) 2 0.10 precipitations Observations 31 385 15 165 11 128 Water quality sample Fecal coliforms NA NA NA 0.00 (0.00) 0.00 (0.00) 0.00 0.00 (0.00) 0.65 (6.19) 2 0.65 Pseudomonas aeroginosa NA NA NA 0.00 (0.00) 0.04 (0.19) 2 0.04 0.00 (0.00) 0.14 (0.35) 2 0.14** pH NA NA NA 7.11 (0.48) 7.35 (0.47) 2 0.24* 6.94 (0.39) 7.31 (0.48) 2 0.37*** Cloudiness NA NA NA 1.57 (1.86) 1.73 (1.93) 2 0.16 0.64 (0.63) 1.37 (1.74) 2 0.73** Count of noncompliance NA NA NA 0.33 (0.49) 0.46 (0.79) 2 0.13 0.05 (0.21) 0.45 (0.78) 2 0.40*** tests Minimum NA NA NA 9.02 (0.30) 4.70 (1.11) 4.32*** 8.62 (0.44) 4.86 (1.15) 3.76*** temperature (8C) Average temperature (8C) NA NA NA 16.96 (0.06) 17.50 (0.94) 2 0.54** 16.74 (0.15) 17.55 (0.99) 2 0.80*** (log) Accumulated NA NA NA 4.18 (0.05) 4.33 (0.19) 2 0.15*** 4.20 (0.24) 4.29 (0.27) 2 0.09* precipitations Observations 12 54 22 94 Note: Figures in parentheses are standard deviations. *, **, and *** indicate statistical significance at the 10 percent level, 5 percent level, and 1 percent level, respectively. Source: URSEA, ECH, and the National Meteorology Office. T A B L E 3 . Test of Parallel Trends before Privatization in the Network Access Sample Aa Ba Cb Dc Variables (1) (2) (1) (2) (1) (2) (1) (2) Time trend 0.002 2 0.003 0.010** 0.008* 0.003 2 0.002 0.006** 0.001 (0.002) (0.003) (0.004) (0.005) (0.002) (0.003) (0.002) (0.004) Time trend by treatment status 0.004 0.004 0.004 0.004 2 0.000 2 0.005 2 0.007** 2 0.010** (0.004) (0.006) (0.004) (0.006) (0.003) (0.005) (0.003) (0.005) Education 0.030 0.038 0.037 0.041*** (0.027) (0.033) (0.027) (0.015) (log) Household per capita 0.205* 0.234** 0.185* 0.098 income (0.105) (0.107) (0.106) (0.104) (log) Accumulated precipitation 0.052** 0.050 0.046** 0.053*** (0.021) (0.043) (0.021) (0.020) Borraz, Gonza Year effects No Yes No Yes No Yes No Yes Observations 416 416 261 261 400 400 270 270 Log likelihood 2 199.18 2 184.35 2 126.12 2 116.78 2 191.76 2 176.84 2 124.75 2 122.75 Sample period 1986– 2000 1993– 2000 1986– 2000 1986– 2000 ´ lez Pampillo Note: Estimates are obtained using a Papke & Wooldridge Fractional Logit Model. Marginal effects are reported, and robust standard errors clustered at the city level are reported in parentheses. All models include city fixed effects. Source: Authors’ estimation using ECH and National Meteorology Office data. *, **, and *** indicate statistical significance at the 10 percent level, 5 percent level, and 1 percent level, respectively. ´ n, and Olarreaga a Treatment Units (three cities): Maldonado, San Carlos, and Pan de Azu ´ car; Control Units (32 cities): Artigas, Bella Unio ´ n, Canelones, Carmelo, Colonia, Dolores, Durazno, Florida, Fray Bentos, Lascano, Libertad, Melo, Mercedes, Minas, Montevideo, Paso de los Toros, Paysandu ´ , Periferia Canelones, Rivera, Rocha, Rosario, Rı ´o Branco, Salto, San Jose´ de Mayo, San Ramo ´ n, Santa Lucı´a, Sarandı´ del Yı ´, Sarandı´ Grande, Tacuarembo ´, Tranqueras, Treinta y Tres, Trinidad, and Young. b Treatment Unit (one city): Maldonado; same Control Units as in A. c Treatment Unit (one city): Maldonado; Control Units (17 cities): Artigas, Canelones, Colonia, Durazno, Florida, Fray Bentos, Melo, Mercedes, Minas, Paysandu ´ de Mayo, Tacuarembo ´ , Rivera, Rocha, Salto, San Jose ´ , Treinta y Tres, and Trinidad. 401 402 THE WORLD BANK ECONOMIC REVIEW nationalization or privatization of water services led to a decline in network access, which could be explained by its trend before the change of ownership. However, as discussed in the results section, we find that network access in- creased before nationalization; therefore, the differences in trends could only downward bias the estimated positive impact of the nationalization. Water Quality Tests The data on water quality come from URSEA. In 2004, URSEA created a water quality unit, in partnership with the chemistry department of University of La Repu ´ blica, which was charged with monitoring the water supply system nationwide according to the guidelines of the World Health Organization. This unit conducts a number of water tests (in the distribution network) to measure the quality of the water provided to consumers. Our units of observation are cities from Uruguay’s 19 departments. We use data for the 2004–2009 period, but some observations are missing for some of the cities. As treatment cities, we use Maldonado, Pan de Azu ´ polis, ´ car, Piria Punta Ballena, Punta del Este, and San Carlos. We use the following control cities: Artigas, Canelones, Colonia, Dolores, Durazno, Florida, Fray Bentos, La Paloma, La Paz, Las Piedras, Melo, Mercedes, Minas, Montevideo, Pando, Paysandu ´ , Progreso, Rivera, Rocha, Salto, San Jose ´ ntida, Tacuarembo ´ , Atla ´, Toledo, Treinta y Tres, and Trinidad.17 We use data for two different microbiological tests and two organoleptic tests that we will use as outcome variables because of their importance in terms of direct and indirect negative effects on health. The two microbiological tests are for fecal coliforms and pseudomonas aeruginosa. The two organolep- tic tests are pH and cloudiness tests. The first test indicates whether the fecal coliforms exceed the accepted maximum value. The second test indicates whether pseudomonas aeruginosa is present. In the case of organoleptic tests, the water quality unit reports the ob- served value. In the case of pH tests, the upper limit is 8 mg/L, and higher levels are considered abnormal. The upper limit in the cloudiness test is five; hence, a result greater than or equal to this value represents a high level of cloudiness in the water, which could indirectly affect health via a higher likeli- hood of bacteria formation and a reduction in the quantity of water consumed. The outcome variable that captures abnormal levels of microbiological or organoleptic substances is constructed as follows. For each test, we generate a dummy variable that takes the value one if the test is above the accepted limit and zero otherwise. We then sum these four binary variables to create a count variable that measures the number of tests that showed abnormal levels in each city. 17. Note that we have six treated cities in the water sample instead of the three in the network access sample. This is still a relatively small number of treated cities, which suggests that the results should be interpreted with caution. ´ lez Pampillo Borraz, Gonza ´ n, and Olarreaga 403 A potential drawback of these data is that the tests have improved and become more precise over time, making it possible to detect abnormal levels of substances more frequently. Thus, our estimates may be biased because we would observe a deterioration of the water quality throughout the period as a result of the improved testing techniques rather than the poor water quality. This issue is addressed in our econometric framework through the use of year fixed effects as control variables. In the bottom panel of table 2, we provide descriptive statistics by treatment and control groups. It is important to note that there was no statistically signifi- cant difference in the count of abnormal results between treated and control cities before nationalization, whereas after nationalization, the count of abnor- mal results was significantly lower in treated cities.18 This result seems to be driven by a lower count of abnormal levels of cloudiness and pseudomonas aeruginosa. Note that we cannot perform a test of parallel trends because we only have two periods (2004 and 2005) before nationalization in the water quality sample. The nationalization dummy therefore takes the value one in treated cities between 2006 and 2009 and zero otherwise. Note also that we do not have data for the preprivatization period. We therefore cannot control for privatization in treated cities because the privatization and nationalization of water services would be perfectly collinear with the city fixed effects. I V. R E S U L T S We begin by discussing the results of the estimation of equation (1) when access to sanitation rates is the outcome variable. We then turn to the estimates obtained when water quality is the outcome variable. Access to Sanitation Networks We estimate equation (1) for different subsamples, with and without control variables. Because the left-hand variable is a fractional variable ( percentage of households with water-sealed toilets connected to sewer lines), we use a Papke and Wooldridge (2008) estimator. Control variables include average completed years of education of the head of the household and average real per capita household income at the city level as well as accumulated precipitation at the department level. We expect the three control variables to be positively corre- lated with the sanitation rate. Table 4 presents the results. There is a positive and statistically significant impact of nationalization on sanitation rates in all subsamples and specifica- tions. A positive effect means that cities in which water services were national- ized experienced an increase in access to sanitation networks. The coefficient 18. Note that some of the control variables present relatively large differences between treated and control cities, which again reflects that only a few cities’ water companies were privatized and subsequently nationalized. 404 THE WORLD BANK ECONOMIC REVIEW that captures the causal impact is 0.15, on average, which means that nationali- zation led to a 15 percent increase in access to sanitation networks. However, the impact of privatization is never statistically significant, except in column D, where the coefficient is negative and statistically significant. This result suggests that in cities where sewer services were privatized, there was no increase in san- itation access rates relative to the preprivatization period.19 Note that these results need to be interpreted with caution because we only have three treated cities in columns A and B and only one treated city in columns C and D because of the data constraints discussed in section III. However, the low number of treated cities makes it more difficult to identify a statistically signifi- cant coefficient, as discussed in McKenzie (2012), which is not the case for the impact of nationalization in the results reported in table 4.20 To determine whether the increase in access to sanitation networks occurred where we expect (i.e., among poor households), we aggregate the data at the city level using only the lower 25th household income percentile in one case and the higher 25th in the other.21 We then append these data and introduce a dummy variable that indicates that the observation corresponds to the bottom 25 percent of the income distribution as well as interaction variables between this dummy and the nationalization and privatization dummies. A positive co- efficient of the interaction of the bottom 25 percent income dummy and the na- tionalization dummy would indicate that poor households experienced a larger increase in their network access rate after nationalization. Table 5 presents the results for the four different samples with and without using the education level of the head of the household, household per capita income, and accumu- lated precipitation. Without these controls, the interaction of the bottom 25 percent income dummy with nationalization is always positive and statistically significant, but after introducing the control variables, the interaction is only significant for sample D. The interaction of the bottom 25 percent dummy with the privatization dummy is negative and statistically significant when using the control variables, but not in the specifications without them. If we combine these results, across all samples we find that the bottom 25 percent has greater access to the network during the nationalization period than during the privatization period.22 19. We also tested whether network access was higher during periods with the public provision of water services (i.e., the preprivatization period and postnationalization period) and found a statistically significant coefficient for most specifications (available upon request). However, the coefficients are 50 percent smaller when using this alternative definition, which confirms that most of the positive impact is due to the nationalization of water services and not to the preprivatization period. 20. This may explain the statistically insignificant results for privatization. 21. Unfortunately, the ECH data do not follow households over time. Therefore, we cannot implement our difference-in-difference methodology at the household level. 22. We also run these regressions for households above and below the poverty line. The results, which are available upon request, suggest that nationalization increased the network access rate of poor households. T A B L E 4 . Impact of Privatization & Nationalization on Network Evacuation Rate in Maldonado Aa Ba Cb Dc Variables (1) (2) (1) (2) (1) (2) (1) (2) Private provision 0.037 0.055 0.037 0.053 2 0.003 2 0.035 2 0.086*** 2 0.081 (0.042) (0.066) (0.042) (0.067) (0.036) (0.054) (0.031) (0.049) Renationalized 0.163*** 0.143** 0.163*** 0.140** 0.244*** 0.173*** 0.154*** 0.139** provision (0.053) (0.058) (0.053) (0.061) (0.036) (0.056) (0.037) (0.055) Education 0.041 0.049 0.045 0.025** (0.028) (0.031) (0.029) (0.012) (log) Household per 0.210*** 0.221*** 0.191*** 0.169*** capita income (0.063) (0.085) (0.062) (0.060) (log) Accumulated 0.075** 0.072* 0.072** 0.063*** Borraz, Gonza precipitation (0.030) (0.043) (0.030) (0.022) Observations 735 735 580 580 702 702 432 432 Log likelihood 2 355.21 2 331.55 2 282.16 2 263.89 2 340.11 2 316.54 2 198.54 2 195.19 Sample period 1986– 2009 1993– 2009 1986– 2009 1986– 2009 ´ lez Pampillo Note: Estimates are obtained using a Papke & Wooldridge Fractional Logit Model. Marginal effects are reported, and robust standard errors clustered at the city level are reported in parentheses. All models include city fixed effects. Source: Authors’ estimation using ECH and National Meteorology Office data. *, **, and *** indicate statistical significance at the 10 percent level, 5 percent level, and 1 percent level, respectively. ´ n, and Olarreaga a Treatment Units (three cities): Maldonado, San Carlos, and Pan de Azu ´ car; Control Units (32 cities): Artigas, Bella Unio ´ n, Canelones, Carmelo, Colonia, Dolores, Durazno, Florida, Fray Bentos, Lascano, Libertad, Melo, Mercedes, Minas, Montevideo, Paso de los Toros, Paysandu ´ , Periferia Canelones, Rivera, Rocha, Rosario, Rı ´o Branco, Salto, San Jose´ de Mayo, San Ramo ´ n, Santa Lucı´a, Sarandı´ del Yı ´, Sarandı´ Grande, Tacuarembo ´, Tranqueras, Treinta y Tres, Trinidad, and Young. b Treatment Unit (one city): Maldonado; same Control Units as in A. c Treatment Unit (one city): Maldonado; Control Units (17 cities): Artigas, Canelones, Colonia, Durazno, Florida, Fray Bentos, Melo, Mercedes, Minas, Paysandu ´ de Mayo, Tacuarembo ´ , Rivera, Rocha, Salto, San Jose ´ , Treinta y Tres, and Trinidad. 405 406 T A B L E 5 . Impact of Nationalization on Access to Sewage Network in Maldonado at the Bottom and Top 25 percent of the Income Distribution THE WORLD BANK ECONOMIC REVIEW Aa Ba Cb Dc (1) (2) (1) (2) (1) (2) (1) (2) Private provision 0.013 0.083 0.022 0.084 2 0.030 2 0.005 2 0.152*** 2 0.119** (0.046) (0.089) (0.047) (0.091) (0.043) (0.053) (0.030) (0.048) Renationalized provision 0.093* 0.143** 0.105* 0.146** 0.158*** 0.164*** 0.016 0.044 (0.056) (0.063) (0.056) (0.066) (0.045) (0.053) (0.035) (0.052) Private provision by Bottom 25 percent 0.032 2 0.134* 0.013 2 0.139* 0.025 2 0.112*** 0.101*** 0.013 (0.024) (0.070) (0.024) (0.072) (0.023) (0.033) (0.015) (0.043) Renationalized provision by Bottom 25 0.109*** 2 0.047 0.091** 2 0.049 0.108*** 2 0.028 0.177*** 0.094** percent (0.036) (0.041) (0.037) (0.043) (0.022) (0.033) (0.013) (0.042) Bottom 25 percent 2 0.362*** 0.111 2 0.348*** 0.103 2 0.364*** 0.128 2 0.415*** 2 0.185* (0.018) (0.086) (0.018) (0.093) (0.018) (0.089) (0.009) (0.110) Controls No Yes No Yes No Yes No Yes Observations 1470 1470 1160 1160 1404 1404 864 864 Log likelihood 2 653.93 2 597.64 2 524.11 2 482.46 2 625.26 2 568.37 2 341.32 2 338.75 Sample period 1986– 2009 1993– 2009 1986– 2009 1986– 2009 Note: Estimates are obtained using a Papke & Wooldridge Fractional Logit Model. Marginal effects are reported, and robust standard errors clustered at the city level are reported in parentheses. All models include city fixed effects. Source: Authors’ estimation using ECH and National Meteorology Office data. *, **, and *** indicate statistical significance at the 10 percent level, 5 percent level, and 1 percent level, respectively. a Treatment Units (three cities): Maldonado, San Carlos, and Pan de Azu ´ car; Control Units (32 cities): Artigas, Bella Unio ´ n, Canelones, Carmelo, Colonia, Dolores, Durazno, Florida, Fray Bentos, Lascano, Libertad, Melo, Mercedes, Minas, Montevideo, Paso de los Toros, Paysandu ´ , Periferia Canelones, Rivera, Rocha, Rosario, Rı ´o Branco, Salto, San Jose´ de Mayo, San Ramo ´ n, Santa Lucı´a, Sarandı´ del Yı ´, Sarandı´ Grande, Tacuarembo ´, Tranqueras, Treinta y Tres, Trinidad, and Young. b Treatment Unit (one city): Maldonado; same Control Units as in A. c Treatment Unit (one city): Maldonado; Control Units (17 cities): Artigas, Canelones, Colonia, Durazno, Florida, Fray Bentos, Melo, Mercedes, Minas, Paysandu ´ de Mayo, Tacuarembo ´ , Rivera, Rocha, Salto, San Jose ´ , Treinta y Tres, and Trinidad. ´ lez Pampillo Borraz, Gonza ´ n, and Olarreaga 407 Because serial correlation may be an issue in this difference-in-difference framework, we parametrically model the serial correlation of the error term as a first-order autoregressive process. We use different estimation methods that imply different assumptions. The results are reported in table 6. In the first column, we report feasible generalized least squares estimates, allowing the error term to be correlated across cities (i.e., we allow for error correlation across panels, or what is typically termed spatial correlation). In the second column, the correlation of errors across individuals is assumed to be identically and independently distributed. In the third column, we use a within estimator with city fixed effects, and in the last column, we provide estimates using the Driscoll and Kraay (1998) method to obtain Newey-West standard errors, which allow error autocorrelation of the general form. We use the subsample, which only includes capital cities other than Montevideo. Again, in all cases, nationalization has a positive impact on access to the sanitation network. This increase in sanitation access is consistent with the fact that a USD 50 million loan was obtained from the World Bank for an OSE project on sanita- tion and residual treatment after the nationalization of private companies. However, the external funding the World Bank provided to Uruguay’s water company was not unusual. Three loans have been granted for the improvement of water services in Uruguay since 1999 that covered the preprivatization period, the privatization period, and the nationalization period.23 Thus, there is no systematic bias toward the nationalization period. More important, we could not find any indication in the official documents that the last loan should be used primarily for improvements in treated cities (those that were national- ized). Thus, one should not expect that this last loan is the reason for the im- provement in water services in treated cities. We obviously cannot exclude the possibility that the loan contributed to improved access and water quality in cities where nationalization took place, but this seems to have been driven by something deeper than the simple availability of funds. On its website, OSE reports that following nationalization, work related to sanitation improve- ments, which had ceased in 2002, was restarted. Sanitation projects in Ciudad de la Costa, Punta del Este, and Maldonado, where private companies were located, apparently became a priority in OSE’s agenda regardless of the avail- ability of funding from external sources. 23. The World Bank provided Uruguay with financial support to develop water and sanitation services through three investment loans at different time periods: 1) Water Supply Rehabilitation project (1988–1999), USD 22.3 million; 2) OSE Modernization and Systems Rehabilitation project, APL-1 (2000–2007), USD 27 million; 3) OSE Modernization and Systems Rehabilitation project, APL-2 (ongoing since 2007), USD 50 million. Other loans focused on technical support, such as the Public Services Modernization Technical Assistance project (2001– 2008) for USD six million. The objective of these loans was to help Uruguay make investments in the water infrastructure, improving the efficiency and coverage of the water supply and sanitation services. See The World Bank (2010) for additional details. 408 THE WORLD BANK ECONOMIC REVIEW T A B L E 6 . Impact of Nationalization on Network Access Controlling for First-order Autoregressive Disturbances Pooled Within Within Variables FGLS Pooled OLS Estimator Estimator D-K Private provision 2 0.003 2 0.019 2 0.007 2 0.024 (0.019) (0.037) (0.048) (0.038) Renationalized provision 0.163*** 0.159*** 0.168*** 0.168*** (0.026) (0.048) (0.066) (0.059) Education 0.024*** 0.023*** 0.019** 0.032** (0.003) (0.008) (0.008) (0.011) (log) Household per capita 0.159*** 0.157*** 0.164*** 0.137*** income (0.013) (0.035) (0.035) (0.037) (log) Accumulated precipitation 0.017*** 0.029 0.023 0.042* (0.007) (0.018) (0.016) (0.021) Observations 432 432 414 432 AR(1) 0.329 0.270 0.209 0.06 Sample period 1986– 2009 Notes: All columns report estimates using panel data estimators controlling for first-order autoregressive (AR(1)) disturbances. Marginal effects are reported, and robust standard errors clustered at the city level are reported in parentheses. All models include individual effects, year effects, and time trends by city. Treatment Unit (one city): Maldonado; Control Units (17 cities): Artigas, Canelones, Colonia, Durazno, Florida, Fray Bentos, Melo, Mercedes, Minas, Paysandu ´, Rivera, Rocha, Salto, San Jose ´ de Mayo, Tacuarembo ´ , Treinta y Tres, and Trinidad. Source: Authors’ estimation using ECH and National Meteorology Office data. *, **, and *** indicate statistical significance at the 10 percent level, 5 percent level, and 1 percent level, respectively. Water Quality Tests The outcome variable is the number of tests that reported an abnormal level of microbiological or organoleptic substances. This type of right-hand variable requires an appropriate estimator. We will use Poisson and negative binomial estimators to account for overdispersion in the water tests (a nonconstant ratio of variance over a conditional mean). Because there is a large number of zeros in the data (see table 2), we also performed a Vuong test, which indicated that the zero-inflated Poisson was the appropriate model.24 For control variables, we used accumulated precipitation and the minimum and average temperatures at the department level. We expect precipitation to be positively correlated with the number of abnormal quality tests because a high level of precipitation is likely to negatively affect the functioning of the water network, making water tests more likely to detect higher levels of 24. When testing the negative binomial estimator, the Vuong test takes the value 5.79 in the specification without controls and 6.45 in the specification with controls. When testing the Poisson estimator, the Vuong test takes the value 1.81 without controls and the value 1.87 with controls. They all reject that the ordinary Poisson or negative binomial models should be preferred to the zero-inflated estimators at the 5 percent level, which is unsurprising given the large number of zeroes. ´ lez Pampillo Borraz, Gonza ´ n, and Olarreaga 409 undesired substances. Low temperatures may increase the likelihood of failures in the distribution network, and a high average temperature may contribute to the reproduction of bacteria, such as coliforms. Hence, we expect a negative coefficient on the former and a positive coefficient on the latter. Table 7 reports the estimates with and without control variables. The first two columns present the zero-inflated Poisson estimates, and the last two columns present the zero-inflated negative binomial estimates. The control vari- ables have the expected signs in both specifications, but none of the variables are statistically significant. The nationalization of water services is always nega- tively associated with abnormal levels of undesirable substances in water quality tests, and the impact is statistically significant at the 10 percent level. It is also very large, with a reduction of 0.7 tests per city exhibiting abnormal levels after nationalization. Thus, the results suggest that water quality im- proved with nationalization. As discussed above, note that because of the small number of treated cities, the results should be interpreted with caution even T A B L E 7 . Impact of Nationalization on Water Quality (count of abnormal tests of levels of microbiological elements) Poisson Negative Binomial Variables (1) (2) (3) (4) Renationalized provision 2 0.651* 2 0.642* 2 0.650* 2 0.665* (0.357) (0.371) (0.355) (0.390) Education 0.367 0.466 (0.247) (0.306) (log) Household per capita income 2 0.016 2 0.145 (0.865) (0.848) Precipitation 0.423 0.611 (0.330) (0.385) Minimum temperature 2 0.146 2 0.235 (0.118) (0.165) Average temperature 0.014 0.002 (0.146) (0.149) Individual effects Yes Yes Yes Yes Year effects Yes Yes Yes Yes Observations 182 182 182 182 Log likelihood 2 101 2 100 2 101 2 100 Sample 2004– 2009 Notes: All columns report estimates using a zero inflated model. Marginal effects are reported, and robust standard errors clustered at the city level are reported in parentheses. All models include city and year fixed effects. Vuong tests indicate that the zero inflated models are appropri- ate. Treatment Units (six cities): Maldonado, Pan de Azu ´ polis, Punta del Este, Punta ´ car, Piria Ballena, and San Carlos. Control units (26 cities): Artigas, Canelones, Colonia, Dolores, Durazno, Florida, Fray Bentos, La Paloma, La Paz, Las Piedras, Melo, Mercedes, Minas, Montevideo, Pando, Paysandu ´ , Progreso, Rivera, Rocha, Salto, San Jose ´ ntida, Tacuarembo ´ , Atla ´, Toledo, Treinta y Tres, and Trinidad. Source: Authors’ estimation using URSEA, ECH and National Meteorology Office data. *indicates statistical significance at the 10 percent level. 410 THE WORLD BANK ECONOMIC REVIEW though we have a larger number of treated cities in the water quality sample than in the network access sample. Moreover, because the water quality sample does not cover the preprivatization period, we cannot control for the determinants of privatization, which leaves us with a weaker identification strategy than for the network access sample.25 V. C O N C L U S I O N S The question of market versus government failures in the provision of water services is complex and unlikely to be answered without empirical evidence. In this paper, we examine the impacts of the privatization and nationalization of water services on service quality. Thus, in contrast to most of the existing liter- ature, we identify the impact not only through the privatization of public firms but also through the nationalization of private firms. Another important aspect of the study is the focus on direct measures of service quality (access to the network and water quality) rather than on indirect measures, such as health outcomes. Using difference-in-difference estimators, we find that Uruguay’s privatiza- tion of water companies in 1993 and 2000 yielded little progress in terms of access to sanitation networks. However, the nationalization of all private com- panies in 2006 led to an improvement in access to the sewage network as well as an improvement in water quality. The improvement in access following the nationalization of water services tended to favor of the poor because greater in- creases in access were observed for poor households. These results are in contrast to existing evidence on the privatization of water services in other Latin American countries, which finds that privatization led to a decline in child mortality and an increase in water access and quality. Future research should attempt to disentangle the determinants of these two outcomes to improve understandings of why privatization and nationalization had different impacts in Uruguay. Private and public companies have different objectives: private firms tend to maximize profits, whereas the objectives of public companies are more varied and may include motives ranging from polit- ical consideration to corruption or social objectives. These differences in objec- tives frequently motivate calls for both privatization and nationalization (see Chong and Lopes-de-Silanes 2005) and may explain part of our empirical find- ings for Uruguay. However, other potential explanations may also be impor- tant. For example, the type of regulations at the time of privatization or nationalization such as required investment or the requirement to provide uni- versal access, the functioning of regulatory bodies, or poorly designed contracts and bidding processes may help to reconcile the results found here and in the 25. Note that when using a public provider dummy that takes a value of one when the city is served by a publicly owned water company in lieu of a nationalization dummy, we obtain qualitatively and quantitatively similar results to those reported in table 7. ´ lez Pampillo Borraz, Gonza ´ n, and Olarreaga 411 rest of the literature. A detailed examination of these differences will improve understandings of what works and what does not with respect to water privati- zation. Additionally, differences in the functioning of public companies (e.g., external funding, the composition of the board of directors) may help to explain differences at the time of the privatization or nationalization of water services. Finally, the results of this paper suggest that the focus on private versus public ownership of natural monopolies such as water providers may be misleading. The institutional environment within which the natural monopoly operates may be much more important. AC K N OW L E D G M E N T S We are grateful to the editor, Elisabeth Sadoulet, Dale Whittington, Guido Porto, Juan Robalino, Roxana Barrantes, Guillermo Valles, Mina Mashayekh, Cristian Rodriguez, three anonymous referees, seminar participants at the 2011 Latin American and Caribbean Environmental Economics Program workshop in San Jose ´ de Costa Rica, the 2011 Bolivian Conference on Development Economics in La Paz, the 2011 Latin American and Caribbean Economic Association annual meeting in Santiago de Chile, and the Geneva Trade and Development Workshop for their very constructive comments and suggestions. We are also grateful to Alvaro Capandeguy from URSEA for providing us with access to data on water quality and to Abelardo Gianola, also from URSEA, for providing us with information regarding price regulation before and after nationalization. The views expressed here are those of the authors and do not necessarily reflect the views of the Banco Central del Uruguay. This work was conducted with the aid of a grant from the Latin American and Caribbean Environmental Economics Program. All errors are our own. REFERENCES Barrera-Osorio, F., M. Olivera, and C. Ospino. 2009. “Does Society Win or Lose as a Result of Privatization? The Case of Water Sector Privatization in Colombia.” Economica 76 (304): 649– 74. Bertrand, M., E. Duflo, and S. Mullainthan. 2004. “How Much Should We Trust Differences-in-Differences Estimates?” Quarterly Journal of Economics 119 (1): 249– 75. Borraz, F., and M. Olarreaga. 2011. “Services Ownership and Access to Safe Water: Water Nationalization in Uruguay.” In M. Mashayekhi, M. Olarreaga, and G. Porto, eds., Services, Trade and Development, 265–82. United Nations Conference on Trade and Development. Cameron, A., J.B. Gelbach, and D.L. Miller. 2008. “Bootstrap-Base Improvements for Inference with Clustered Errors.” The Review of Economics and Statistics 90 (3): 414–27. Chong, A., and F. Lopez-de-Silanes. 2005. Privatization in Latin America: Myths and Reality. Palo Alto, CA: Stanford University Press. Driscoll, J., and A. Kraay. 1998. “Consistent Covariance Matrix Estimation With Spatially Dependent Panel Data.” Review of Economics and Statistics 80 (4): 549– 60. 412 THE WORLD BANK ECONOMIC REVIEW Galdo, G., and B. Bertha. 2005. “Evaluating the Impact on Child Mortality of a Water Supply and Sewerage Expansion in Quito: Is Water Enough?” OVE Working Papers 0105. Inter-American Development Bank, Office of Evaluation and Oversight (OVE), Washington, DC. Galiani, S., P. Gertler, and E. Schargrodsky. 2005. “Water for Life: The Impact of the Privatization of Water Services on Child Mortality.” Journal of Political Economy 113 (1): 83 –120. Galiani, S., P. Gertler, E. Schargrodsky, and F. Sturzenegger. 2005. “The Benefits and Costs of Privatization in Argentina: A Microeconomic Analysis.” In A. Chong, and F. Lopez-de-Silanes, eds., Privatization in Latin America: Myths and Realities, 67– 116. Palo Alto, CA: Stanford University Press. Hall, D., E. Lobina, and V. Corral. 2010. “Replacing Failed Private Water Contracts.” PSIRU report. Public Services International Research Unit, University of Greenwich, London, UK. Kikeri, S.. 1999. “Privatization and Labor: What Happens to Workers when Governments Divest?” World Bank Technical Paper No. 396. World Bank, Washington, DC. Lee, L., M. R. Rosenzweig, and M. M. Pitt. 1997. “The Effects of Improved Nutrition, Sanitation, and Water Quality on Child Health in High-Mortality Populations.” Journal of Econometrics 77: 209 –35. McKenzie, D.. 2012. “Beyond Baseline and Follow-up: The Case for More T in Experiments.” Journal of Development Economics 99 (2): 210– 21. Nauges, C., and D. Whittington. 2009. “Estimation of Water Demand in Developing Countries: An Overview.” World Bank Research Observer 25 (2): 263– 94. Papke, L., and J.M. Wooldridge. 2008. “Panel Data Methods for Fractional Response Variables with an Application to Test Pass Rates.” Journal of Econometrics 145 (1-2): 121–33. Segerfeldt, F.. 2005. Water for Sale: How Business and The Market Can Resolve the World’s Water Crisis. Washington, DC: Cato Institute. Shi, A.. 2000. “How Access to Urban Potable Water and Sewerage Connection Affects Child Mortality.” The World Bank, Policy Research Working Paper 2274. World Bank, Washington, DC. Wallsten, S., and K. Kosec. 2005. “Public or Private Drinking Water? The Effects of Ownership and Benchmark Competition on U.S. Water System Regulatory Compliance and Household Water Expenditures.” AEI-Brookings Joint Center Working Paper No. 05-05. AEI-Brookings Joint Center, Washington, DC. Whittington, D., J. Davis, L. Prokopy, K. Komives, R. Thorsten, H. Lukacs, W. Wakeman, and A. Bakalian. 2009. “How Well is the Demand-Driven, Community Management Model for Rural Water Supply Systems Doing? Evidence from Bolivia, Peru, and Ghana.” Water Policy 11 (6): 696 –718. World Bank. 2004. “Reforming Infrastructure: Privatization, Regulation and Competition.” Washington, DC: World Bank and Oxford University Press. ———. 2010. “Modernizing Public Water Services in Uruguay through a long-term Partnership.” Working Paper. World Bank, Washington, DC. Mass Media and Public Policy: Global Evidence from Agricultural Policies Alessandro Olper and Johan Swinnen Mass media play a crucial role in information distribution and in the political market and public policy making. Theory predicts that information provided by the mass media reflects the media’s incentives to provide news to different groups in society and affects these groups’ influence in policy making. We use data on agricultural policy from 69 countries spanning a wide range of development stages and media markets to test these predictions. Our empirical results are consistent with theoretical hypotheses that public support for agriculture is affected by the mass media. In particular, an in- crease in media (television) diffusion is associated with policies that benefit the majori- ty to a greater extent and is correlated with a reduction in agriculture taxation in poor countries and a reduction in the subsidization of agriculture in rich countries, ceteris paribus. The empirical results are consistent with the hypothesis that increased compe- tition in commercial media reduces transfers to special interest groups and contributes to more efficient public policies. JEL Codes: D72, D83, Q18 There is a rapidly growing body of literature on the economics of the mass media. This literature has led to a series of important new hypotheses and in- sights in an area that has been long neglected by economists.1 An important strand of this literature concerns the role of mass media in political markets and its effect on public policy making. Most of this literature on the relation- ship between the mass media and public policy is theoretical. A few empirical studies have attempted to assess the effect of media on policy outcomes. Some key findings from this literature suggest that access to mass media empowers Alessandro Olper is an associate professor at the University of Milano, Italy, and a research fellow at Centre for Institution and Economic Performance; his email address is alessandro.olper@unimi.it. Jo Swinnen is a professor of economics and director of Centre for Institution and Economic Performance at the University of Leuven, Belgium, and visiting professor at Stanford University, United States. His email address is Jo.Swinnen@econ.kuleuven.be. This work was supported by the KU Leuven Research Council (Methusalem project). The authors would like to thank the participants in several conferences and meetings, the editor, and three anonymous referees for the World Bank Economic Review for useful comments and suggestions. A supplemental appendix to this article is available at http://wber.oxfordjournals.org/. 1. See McCluskey and Swinnen (2010) and Pratt and Stro ¨ mberg (2011) for reviews. THE WORLD BANK ECONOMIC REVIEW, VOL. 27, NO. 3, pp. 413 –436 doi:10.1093/wber/lht008 Advance Access Publication March 26, 2013 # The Author 2013. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 413 414 THE WORLD BANK ECONOMIC REVIEW people politically and, as such, increases the benefits they receive from govern- ment programs (Stro ¨ mberg and Snyder 2008). This influence has been found for different types of government programs and different countries, such as un- employment relief in the United States (Stro ¨ mberg 2004b), public food provi- sion and disaster relief in India (Besley and Burgess 2001, 2002), and educational spending in Uganda and Madagascar (Reinikka and Svensson 2005; Francken et al. 2009). All of these studies measure the effect within a single country, which has the benefit of holding many other factors fixed but has the potential disadvantage of limited variation in policy and media. Our paper seeks to contribute to this empirical literature by analyzing the impact of mass media on policy making for a specific type of policy across a wide variety of countries and years. We use a new dataset from the World Bank that includes measures of agricultural subsidization and taxation for a much wider set of countries and a longer period of time than has been previ- ously available (Anderson and Valenzuela 2008). We use these data as the de- pendent variable. Agricultural policy (subsidization or taxation) is an excellent policy instrument to study the impact of media on policy choice across a wide variation of coun- tries for both empirical and theoretical reasons. Empirically, agricultural policy is an important policy for governments in both rich and poor countries. In poor countries where agriculture is a very important share of the economy and food is a major consumption item, the importance of agriculture as a public policy issue is obvious. However, in rich countries, agricultural policy remains disproportion- ately important compared to the relatively small share of agriculture in terms of economic output. For example in the European Union, the Common Agricultural Policy continues to absorb 40 percent of the entire European Union budget. Another symptom of the continued importance of agricultural policy for rich countries is the impasse in the current WTO negotiations, where disagreements over agricultural policies threaten to undermine the entire WTO agreement. Another empirical factor is the substantial ad hoc and case study evidence that mass media can play an important role in influencing agricultural policy. Several studies have highlighted the important role of mass media in influenc- ing voters and government policy on key recent agricultural and food policies, such as the use of genetically modified organisms (Curtiss et al. 2006; Marks, et al. 2003; Vigani and Olper 2012), policy reactions to food safety crises (Swinnen et al. 2005; Verbeke et al. 2000), and trade disputes (Kuzyk and McCluskey 2006; Swinnen and Francken 2006). Agricultural policy is also an interesting case from a theoretical perspective. The literature on the political economy of agricultural policy identifies group size (the number of farmers versus the number of food consumers in the economy) as an important causal factor. Group size is argued to play an im- portant role because it affects collective action costs (based on Olson 1965) and affects the per capita costs and benefits of agricultural policy, which, in turn, affects political outcomes in the presence of voter information costs Olper and Swinnen 415 (based on Downs 1957) or when political activities are proportional to the size of potential policy costs and benefits (Swinnen 1994). Recent papers in the media economics literature claim that the mass media can play an important role in public policy by altering these political economy mechanisms (Stromberg 2001, 2004a; Kuzyk and McCluskey, 2006). Oberholzer-Gee and Waldfogel (2005) argue that the link between group size and political mobili- zation depends on the structure of media markets. In a series of influential papers, Stro ¨ mberg (2001, 2004a) has shown that competition among mass media outlets leads to the provision of more news/information to large groups such as taxpayers and dispersed consumer interests, altering the trade-off in po- litical competition and thus influencing public policy. He refers to this outcome as “mass media-competition-induced political bias.” The purpose of our paper is to evaluate whether the mass media have an impact on the political economy of agricultural policies by exploiting taxation and subsidization data from 69 countries that were observed from 1960 to 2004. The paper also contributes to an emerging body of literature analyzing whether the diffusion of free and independent media are key ingredients in more efficient public policies. Besley and Burgess (2001, 2002) use a political agency model to demonstrate that a more informed and politically active elec- torate increases the incentives for a government to be responsive. Prat and Stro¨ mberg (2005) find for Sweden that people who begin watching commercial television (TV) news programs increase their political knowledge and participa- tion. Overall, this and other evidence (such as Besley and Burgess 2001, 2002; Francken et al. 2009, 2012; Reinikka and Svensson 2005; Stro ¨ mberg and Snyder 2008) support the notion that the mass media reduce the power of special interest lobbies relative to unorganized interests. The paper also contributes to the literature on the political economy of agri- cultural policies. There is extensive literature, both theoretical and empirical, on the determinants of agricultural policy making (see de Gorter and Swinnen 2002; Swinnen 2010; Anderson et al. 2013 for surveys), but no previous study has ex- amined the role of the media in this process. Our paper is the first to do so. Our analysis, which exploits both cross-country and time-series variation in the data, indicates that the mass media may have a substantive impact on public agricultural policy. In the developing world, agricultural taxation is reduced by the presence of mass media outlets, whereas agricultural support is reduced in developed countries. Our results thus suggest that competition in the media market is associated with a reduction of policy distortions in agricul- tural markets. I. CONCEPTUAL FRAMEWORK ¨ mberg’s In this section, we first present a conceptual framework based on Stro (2004a) theory of mass media and political competition. Then, we discuss the main implications of the model in light of the worldwide characteristics and 416 THE WORLD BANK ECONOMIC REVIEW regularities of agricultural policies, and we identify testable hypotheses regard- ing the effect of mass media competition on agricultural policy outcomes. The Basic Theory In Stro¨ mberg’s (2004a) model, mass media outlets affect public policy because they provide the channel through which politicians convey campaign promises to the electorate. Political parties make binding announcements regarding the amount of public money they plan to spend on various government programs. The two parties propose public spending with the objective of maximizing their vote shares within a given budget constraint. Commercial media2 are the only channel through which the parties’ plat- forms are announced to the voters. Media firms allocate a certain quantity of information on political platforms and proposed spending levels with the ob- jective of maximizing their audiences (number of readers/viewers), all of which are voters. Voters purchase media products (e.g., newspapers) and adjust their expectations of how much the parties will spend on the basis of information provided by the media. They then vote for one of the parties. The party that wins the election implements the promised spending plan. Voters value news related to political platforms in the media because it allows them to maximize the benefits they receive from government programs. It is assumed that readers (voters) use the news they receive from the media to decide on private action, which affects the value they realize from a govern- ment program. More precise news about future policies makes it more likely that readers will take the correct private actions. For example, early news about changes in agricultural subsidies help farmers to produce the correct crops to realize the full value of these subsidies. Voters purchase the media product that provides them with the most infor- mation on the government programs they value, conditional on other (exoge- nous) characteristics, such as ideological preferences. The media maximize expected profits. They have two types of costs: “first copy costs,” or the cost of producing one unit of news space, and “reproduc- tion and distribution costs,” or the average cost of reproducing and delivering the media product to a certain audience. This cost function is consistent with the notion that there are roughly constant long-run marginal costs in distribut- ing TV and radio news (or in printing and delivering newspapers). The revenue of the media includes the price they can charge for their product plus the price per reader/viewer paid by advertisers. This structure of the media industry implies that different groups receive different media coverage. News items of interest to small groups and groups with limited attractiveness for advertisers receive less coverage. 2. In his theoretical derivations, Stromberg (2004a) refers to the media as “newspapers” but explains that for the purposes of the main points of his analysis, the cost and revenue structures of TV and radio are similar in the relevant aspects. Olper and Swinnen 417 This bias in the mass media’s news coverage translates into policy bias. As media coverage of different issues changes, the efficiency with which politicians can reach different groups with campaign promises also changes. If a party promises a group that receives very little news coverage that it will raise spend- ing, only a small fraction of the voters who would benefit become aware of the promise. Therefore, a spending promise to this group will not win many votes for the party. Consequently, this group of voters will not attract many favor- able policies. Instead, promising to raise spending for groups that attract sub- stantial media coverage (for example, because they are large groups or because the groups are more valuable to advertisers) will lead to a stronger voter re- sponse and thus to policies that are more favorable to these groups. Implications for Agricultural Policy What does this theory imply for the impact of mass media on agricultural policy?3 The most important stylized fact about agricultural policy is the so-called development paradox, the policy switch from the taxation to the sub- sidization of agriculture associated with economic development (Anderson and Hayami, 1986; Anderson, 2009). The classic interpretation of this pattern is that when a country becomes richer, farm groups, compared to consumer and taxpayer groups, become more effective in collective action situations as the number of farmers declines and development reduces communication and transportation costs. Both factors reduce organizational costs and free rider problems in collective action situations (Olson, 1965). Moreover, because the per capita cost experienced by the rest of society falls with fewer farmers, the opposition of taxpayers and consumers to (agricultural) subsidies decreases as the number of farmers decreases as a result of economic development (Becker 1983; Swinnen 1994; Anderson 1995). The model developed here suggests that the relationship between agricultural policy and economic development will be affected by the introduction of media competition in the political market. Voter preferences and government policies will be affected by how the media industry provides information to citizens. One key prediction of the model is that, ceteris paribus, government trans- fers such as agricultural protection should be biased toward large groups as an effect of media competition. Because the agricultural group (the number of farmers) is relatively large in poor countries and relatively small in richer ones, an important implication of the model is that, ceteris paribus, the effect of media competition on agricultural policy should differ in poor versus rich countries. More specifically, we expect that the impact on agricultural protec- tion induced by mass media competition should be positive in poor countries 3. Stromberg’s (2004a) basic model has a fixed budget constraint. This assumption is later relaxed (see section 3.2 and footnote 16 of his paper) to allow for government programs, such as taxes and subsidies, which influence the budget itself. Endogenizing the government budget in this extended model does not affect the main results (or our hypotheses). 418 THE WORLD BANK ECONOMIC REVIEW and negative in rich countries. Thus, we can formulate the following empirical prediction. Hypothesis 1 (Group size effect): Mass media competition-induced political bias will reduce agricultural protection in rich (developed) countries but will increase agricultural protection in poor (developing) countries, ceteris paribus. Another prediction of the model is that, ceteris paribus, government trans- fers will be biased toward groups that are more attractive to advertisers. Stromberg (2004a) refers to the case of the show Gunsmoke, which was can- celled despite its high ratings because its audience was perceived as “too old and too rural” to be of much interest to advertisers. The implication for agricultural protection is not obvious because “being at- tractive to advertisers” may apply to many things, including age (young people are more easily influenced by advertisements than older people), income (richer people have more money to spend), and so forth. The latter argument would imply that mass media competition would induce government transfers to be biased toward relatively richer groups who have more income to spend and are therefore more attractive to advertisers. Nevertheless, the implication for agricultural protection is not trivial. The relationship between economic development and the rural-urban income gap is the subject of some debate. Recent papers have argued that this relationship is nonlinear, with a relatively small gap in very poor countries (as incomes in both urban and rural areas are very low), urban incomes rising relative to rural incomes when countries grow, and the income gap narrowing again at high income levels (see Hayami 2007; McMillan and Rodrik, 2011). This nonlinear relationship would also follow if the spread of TV were uneven between urban and rural areas, reflecting the differences in income (demand for TVs). There would be a relatively small gap in very poor countries (as TV distribution in both urban and rural areas is very low) and a rise in urban TV distribution rela- tive to rural TV distribution when countries grow, with the gap in TV distribu- tion narrowing again at high income levels. In either case, this channel would add a nonlinear relationship between average incomes and media effects, with a prourban media bias effect that would be strongest at medium income levels. However, a problem with using observed (sectoral) income to document this nonlinear relationship between the rural-urban income gap and overall income is that the observed sectoral income levels are obviously affected by the policies themselves. With occasionally very large subsidies or taxes, these policy trans- fers clearly affect the relative income measures. With transfers going from rural to urban areas in poor countries (and vice versa in rich countries), we would expect the pretransfer rural-urban income ratio to be higher (lower) than that observed in poor (rich) countries. In medium income countries, the transfers are relatively lower; thus, the observed income ratio is closer to the pretransfer ratio. This situation would imply that the expected “advertiser-value effect” of media competition on agricultural protection should be as follows: small in Olper and Swinnen 419 (very) poor countries, because pretransfer rural and urban incomes are similar, and low and negative in rich and emerging countries, where pretransfer urban incomes are much higher than rural incomes. Thus, we can formulate the fol- lowing empirical prediction (conditional on an observed nonlinear relationship between the rural-urban income ratio and economic development). Hypothesis 2 (Advertiser value effect): Mass media competition-induced po- litical bias will reduce agricultural protection in rich (developed) and emerging countries and will have little effect in poor (developing) countries, ceteris paribus. In combination, these two effects imply that the total media effect will in- crease agricultural protection (or reduce agricultural taxation) in poor coun- tries owing to the group size effect, reduce agricultural protection in emerging countries owing to the advertiser value effect, and more strongly reduce agri- cultural protection in rich countries owing to the reinforcement of group size and advertiser value effects. The next sections present the data and empirical strategy used to test the hypotheses. I I . D ATA AND E M P I R I CA L S P E C I F I CAT I O N We test our predictions on annual observations from a sample of approximate- ly 70 developing and developed countries from all continents. Overall, we employ a panel of more than 2,000 observations, but the panel structure is un- balanced. Specifically, for a few developed countries, the starting year is ap- proximately 1960; for the majority of the sample, the starting year is approximately 1970; and for transition countries, it is approximately 1992. The last year of observations is 2004. Table S.1 (in the supplemental appendix, available at http://wber.oxfordjournals.org/) reports the full list of countries with data and the 1970 and 2002 values of key policy and media variables. Dependent Variables Our dependent variables are measures of agricultural protection. As an indica- tor of agricultural taxation and subsidization, we use the relative rate of assis- tance (RRA) to agriculture, taken from the Agricultural Distortions Database from the World Bank (see Anderson and Valenzuela 2008 for details). The RRA index is calculated as the ratio between the agricultural and nonagricul- tural nominal rates of assistance: ÂÀ Á À Á à RRA ¼ 1 þ NRAag = 1 þ NRAnonag À 1 where NRAag is the nominal rate of assistance to agriculture and NRAnonag is the nominal assistance to nonagricultural sectors. The NRAag measures total transfers to agriculture as a percentage of the undistorted unit value. It is 420 THE WORLD BANK ECONOMIC REVIEW positive when agriculture is subsidized, negative when it is taxed, and zero when net transfers are zero. The NRAag at the agricultural level is obtained as the weighted average of assistance at the product level using the undistorted value of production as a weight. It includes a wide range of policies, such as the assistance provided by all tariff and nontariff trade measures applied to ag- ricultural products and any domestic price-distorting measures. In addition, the price equivalent of any direct interventions on inputs is included. One advantage of using RRA (instead of NRA) as the dependent variable is that, especially in developing countries, an important indirect source of agricul- tural taxation is trade protection for the manufacturing sector as a component of import-substitution policies. Thus, RRA is a more useful indicator in an in- ternational comparison of the extent to which a country’s policy regime has an anti- or proagricultural bias. However, as a robustness check and to assess whether changes in agricultural policy or industrial protection are important el- ements in the media-protection relationship, we ran a series of additional re- gressions using the NRA as the dependent variable. Mass Media Variables To test our hypotheses, we use the penetration of TV sets as an indicator. More specifically, our variable is the natural logarithm of TV sets per 100 in- habitants, based on data from the Arthur S. Banks Cross National Time-Series Data Archive, supplemented by UNESCO Statistics on TV and data from the International Telecommunication Union (2010).4 The rationale for using this proxy is that, although the share of informed voters is not observed, we can observe the share of media users. Because both move together, it is sufficient to examine the levels and changes in the share of ¨ mberg 2004b). Moreover, media users to test the effect of media bias (see Stro in our specific context, another justification for the use of these indicators is derived from Stro ¨ mberg’s (2004a, p. 266) argument that “the emergence of broadcast media increased the proportion of rural and low-education media consumers as it became less expensive to distribute radio waves than newspa- pers to remote areas, and as these groups preferred audible and visual enter- tainment to reading. As politicians could reach rural and low-education voters more efficiently, the model predicts an expansion in programs that benefits these voters.” 4. Both the Arthur Banks and the International Telecommunication Union data are based on data originally collected by UNESCO. These data are collected annually from 1970 onward. From 1960 to 1970, the data were collected every five years. Thus, for that period, we only use data for 1960, 1965, and 1970 without any interpolation. A potential limitation of these data lies in the fact that some countries have a licensing scheme whereby TV sets (or radios) must be registered. Because households may have more than one TV receiver or may not register, the number of licensed receivers may understate the true number of TVs and radios. Our identification strategy exploits the within-country variation in the data (see section 5). As long as the number of licenses and of TVs/radios display similar growth paths, this limitation of the data should not pose a major problem. Olper and Swinnen 421 Income and Group Size Variables The hypotheses advanced in section II imply that the relationship between the media variable and RRA (NRA) is conditional on the level of development (income), partially due to group size effects. As an indicator of development (income), we use real per capita GDP in purchasing power parity ( gdppc) taken from the Penn World Tables.5 The most direct indicator of (relative) group size is the share of agricultural employment, emps, based on FAO data. However, as is well known, both are strongly correlated because the agricultur- al employment share decreases with economic development. Therefore, gdppc is itself an indicator of relative group size. In our basic specifications, we use gdppc as the primary indicator for the conditional effect of media on agricultural protection because the employment share data are of poor quality, which precludes a consistent comparison across countries and over time, especially for developing countries (see Timmer and de Vries 2007 for a discussion). One reason for the poor quality is the differ- ences in national definitions of “agricultural labor.” Another reason is that the yearly “observations” in the FAO labor statistics are linear interpolations between census data collected once every decade. We understand that there are also problems with the measurement of gdppc because national accounts data are noisy over short time horizons (see Deaton 2005), but we believe the data problems are less important for gdppc than for emps. However, we perform a series of robustness checks using emps as indicator. Both variables (the level of development and employment share) are also in- cluded as control variables because both have been identified as major determi- nants of agricultural protection outside the media effect. Other Control Variables In addition to the variables discussed above, in the empirical specifications, we include controls that are likely to affect the level of agricultural protection, as suggested by previous studies. Standard control variables in studies on the po- litical economy of agricultural policies are indicators of comparative advan- tage, trade status, terms of trade effects, and political institutions (see Olper 2007; Swinnen 2010; Olper and Raimondi 2012). To control for comparative advantage, we include agricultural land per capita, landpc, and the agricultural export share, exps, measured as net exports over production. These two vari- ables are based on data from FAO and the World Bank’s Agricultural Distortions Database. Because of the possibility that governments set agricul- tural protection to exploit terms of trade effects, we also control for country size using the log of population, logpop. We proxy for political institutions by 5. Specifically, we use the variable rgdpch from the Penn World Tables, version 6.3. 422 THE WORLD BANK ECONOMIC REVIEW T A B L E 1 . Descriptive Statistics Std. Mean Dev. Min. Max. Obs. Countries RRA 12.71 64.63 2 94.62 404.87 2,020 69 NRA 20.11 67.19 2 93.11 432.72 2,231 69 Log TVs (  100 inhabitants) 1.89 2.10 2 6.91 4.60 2,231 69 GDP per capita (in purchasing power 9,808 10,313 259 45,947 2,231 69 parity) Agricultural employment share 0.38 0.29 0.01 0.92 2,231 69 Land per capita 1.72 3.97 0.04 41.51 2,231 69 Net export share 0.01 0.36 2 1.73 1.28 2,087 69 Log population 9.97 1.29 7.21 14.07 2,231 69 Democracy index (Polity2) 3.26 7.13 2 9.00 10.00 2,231 69 Government consumption 17.13 9.18 1.38 85.37 2,230 69 Trade to GDP ratio 51.66 31.17 5.00 622.63 2,224 69 Sach-Warner trade policy index 0.61 0.49 0.00 1.00 2,178 69 Economic crisis 0.26 0.44 0.00 1.00 2,231 69 Source: Own calculations based on the data described in the text. adding the Polity2 index of democracy taken from the Polity IV database (Marshall and Jaeggers 2007).6 In addition to these standard control variables, we use a series of covariates to check the robustness of our findings (Olper et al. 2013). They include two different indicators of (aggregated) openness: the ratio of trade to GDP from the Penn World Tables and the Sachs-Warner index of openness as defined by Wacziarg and Welch (2008).7 We use government consumption to GDP from the Penn World Tables as a proxy for government size. Finally, because eco- nomic crises may trigger policy reforms, we add two lags of a crisis variable measured with a dummy equal to one for every year that the real GDP per capita growth rate is negative (zero otherwise). Table 1 presents summary sta- tistics for the variables described above. I I I . D E S C R I P T I V E S TAT I S T I C S AND A N A LY S I S Figures 1 and 2 present trends in the agricultural policy indicators (RRA and NRA) and the media variable (number of TVs per 100 inhabitants) for the period from 1970 to 2004 for selected countries. Table S.1 in the online appen- dix presents the figures for 1970 and 2002 for all countries in the dataset. The 6. The Polity2 index assigns a value ranging from 2 10 (autocracy) to þ 10 (democracy) to each country and year, with higher values associated with stronger democracies. 7. Wacziarg and Welch (2008) updated the Sachs and Warner index and exploited its time dimension. The index is equal to one when a country is considered open and zero otherwise. Thus, it captures reforms in the overall trade policy. Olper and Swinnen 423 F I G U R E 1. Indicators of Agricultural Policies (RRA, NRA) in Selected Countries (1970–2004) Note: Three-year moving averages of both RRA and NRA. Source: Own calculations based on data from the Agridistortions databases (World Bank). F I G U R E 2. Indicators of Mass Media: TVs per 100 Inhabitants in Selected Countries (1970–2004) Source: Own calculations based on the data described in the text. figures reveal that there is substantive variation in all indicators both over time and between countries. Figure 1 illustrates the stylized fact that RRA and NRA are higher in rich countries (Canada and Germany) than in poor countries (India and Kenya). 424 THE WORLD BANK ECONOMIC REVIEW Support for farmers was particularly high in Germany, with an RRA between 50 percent and 100 percent for much of the period. The level of taxation faced by farmers was particularly severe in India before 1980, with RRA figures of approximately 50 percent. Figure 1 also shows that the variation in NRA and RRA between countries has declined over time, with agricultural taxation de- clining (RRA increasing) in poor countries and agricultural subsidies for agri- culture falling (RRA decreasing) in rich countries. On average, the RRA for the full sample increased from approximately 0 percent in the 1970s to approx- imately 20 percent in the 1980s, subsequently declining to approximately 10 percent in 2000 (the NRA values differ, but the patterns are similar). Figure 2 illustrates the differences in TV penetration across countries and over time. The number of TVs increased everywhere over time, but, not sur- prisingly, the numbers are much higher in rich countries than in poor ones. The number of TVs increased from approximately 35 per 100 inhabitants in 1970 to approximately 60 in Germany and 75 in Canada by 2000. In the poor countries, the number of TVs was nearly zero until the mid-1980s. It remains very low in Kenya. The increase is somewhat more rapid in India, reaching ap- proximately 10 per 100 inhabitants in 2000. A comparison of figures 1 and 2 suggests a different relationship between ag- ricultural policies and the spread of mass media in rich and poor countries: for poor countries (such as Kenya), the RRA increased over the 1970–2004 period, and mass media also grew. In contrast, for rich countries (such as Germany), the RRA declined, whereas mass media continued to grow. To further analyze this relationship between mass media and agricultural policies, table 2 reports simple pair-wise correlations between the media vari- able and the agricultural protection indicators for different levels of develop- ment (organized as percentiles of per capita GDP). The pair-wise correlations T A B L E 2 . Correlation Coefficients between TV Penetration and Agricultural Protection Indicators by Percentile of per Capita GDP TV vs. Agricultural protection Percentiles of gdppc RRA NRA , 5% 0.228 0.245 , 10% 0.292 0.249 , 25% 0.290 0.205 , 50% 0.406 0.360 . 50% 0.102 0.044 . 75% 2 0.352 2 0.391 . 90% 2 0.624 2 0.612 . 95% 2 0.656 2 0.629 Note: The percentiles of gdppc refer to two different samples due to data availability for RRA and NRA. The samples have the following median values of gdppc: USD 5,949.5 for the RRA sample and USD 5,356.6 for the NRA sample. Source: Own calculations based on the data described in the text. Olper and Swinnen 425 are consistent with the suggestions from figures 1 and 2. The correlation coeffi- cient is approximately þ 0.30 for percentiles below the median ( , 50 percent). For per capita GDP values above the 75th percentile, the correlation becomes negative, and increasingly so, reaching a value below 2 0.60 at the 95th percen- tile. These correlations are consistent with hypotheses 1 and 2: there is a differ- ent relationship between agricultural policies and the spread of mass media in rich and poor countries. For poor countries, the spread of mass media is associ- ated with an increase in RRA (NRA), whereas in rich countries, the growth of mass media is associated with a decline in RRA (NRA). Of course, these are merely descriptive statistics. For a more formal analysis, we now turn to an econometric analysis. I V. E C O N O M E T R I C S T R A T E G Y AND I D E N T I F I CAT I O N I S S U E S Our hypotheses suggest that in countries with low gdppc, the media variables, and RRA should be positively related, and when gdppc is high, there should be an inverse relationship between these variables. A priori, we do not know at what level of gdppc the relationship changes sign. By using a general specifica- tion, we can derive the gdppc value at the turning point, if any, from the esti- mated coefficients: RRAit ¼ a0 þ a1 mitÀ1 þ a2 ðmitÀ1  gdppcitÀ1 Þ þ a3 gdppcitÀ1 þ bxitÀ1 þ 1it ð1Þ where RRAit measures the relative rate of assistance in country i and year t, mit-1 refers to the one-year lagged media variable, and xit-1 is a vector of addi- tional controls. Taking the partial derivative of RRA with respect to the media variable, we obtain @ RRA ¼ a1 þ a2 gdppc: @m Given our hypotheses, we expect that a1 . 0 and a2 , 0, such that a1 þ a2 gdppc is positive (negative) as gdppc is higher (lower) than gdppc*, with gdppc* ¼ a1/– a2 the level of development at which our media-protection rela- tionship changes sign. We refer to this as the “turning point.” Note that a key requirement for the predictions to hold is that gdppc* should lie within the range of gdppc values in the dataset. Regarding identification, our main concern is omitted variable bias. If the media variables are correlated with unobserved determinants of the protection level, our estimates will be inconsistent. Note that, a priori, the direction of the bias is not predictable. Therefore, our basic specification always includes a set of country (hi) and year fixed effects (qt): RRAit ¼ a1 mitÀ1 þ a2 ðmitÀ1  gdppcitÀ1 Þ þ a3 gdppcitÀ1 þ bxitÀ1 þ hi þ qt þ 1it : ð2Þ 426 THE WORLD BANK ECONOMIC REVIEW However, this approach does not allow us to properly isolate the causal effect of increased mass media consumption for two main reasons. First, country fixed effects do not control for unobservable, country-specific, time-varying factors correlated with both the media and agricultural protection. Second, po- tential measurement errors in the gdppc indicator may introduce further endo- geneity problems (Deaton 2005). Indeed, the correlation between gdppc and the media variables may bias the estimated media coefficients (and the interac- tion term between the media and gdppc), a problem exacerbated by the fixed effects transformation (Wooldridge 2002). Thus, to make our identification as- sumption more credible, we adopt two additional strategies. First, given our specific concern regarding (omitted) time-varying factors cor- related with both media variables and protection, in addition to the time dummies, we include continent-year interaction effects to control for changes over time that affect countries within a region similarly.8 In addition, as discussed in the data section, beyond the traditional determinants of agricultural protection that have been found relevant in previous studies, we include several other covari- ates, such as indicators of trade openness, trade policy reforms, government con- sumption, and economic crises, to increase the similarity between the countries. Second, to directly address measurement error problems and other forms of endogeneity, we also employ dynamic panel methods. Specifically, we use the system generalized methods of moment (SYS-GMM), developed by Arellano and Bover (1995) and Blundell and Bond (1998). By estimating a system of equations in first differences and levels and employing instruments, this ap- proach should allow for consistent estimations even in the presence of measure- ment errors and other forms of endogeneity (see Bond et al. 2001). Our SYS-GMM dynamic panel model has the following specification: DRRAit ¼ a1 DRRAitÀ1 þ a2 DmitÀ1 þ a3 DðmitÀ1  gdppcitÀ1 Þ þ a4 DgdppcitÀ1 þ þ b0 DxitÀ1 þ qt þ yit ð3aÞ RRAit ¼ a0 þ a1 RRAitÀ1 þ a2 mitÀ1 þ a3 ðmitÀ1  gdppcitÀ1 Þ þ a4 gdppcitÀ1 þ þ b0 xitÀ1 þ qt þ yit ð3bÞ where D denotes first differences, that is, Dyit ¼ yit – yit-1, RRAit-1 is the lagged dependent variable and y it is a disturbance term. In estimating the system of equations (3a)–(3b), the (endogenous) lagged dependent variable is instrument- ed by its t – 2, t – 3, and longer lags, using the lagged levels for the first- 8. These interaction effects capture any regional differences in the agricultural protection dynamic. We also tested a second specification in which we included continent-specific polynomial terms over time, and the results were qualitatively and quantitatively similar. Olper and Swinnen 427 differences equation (3a) and the lagged differences for the level equation (3b). Similarly, to address endogeneity in other explanatory variables (such as the media variable and its interaction with gdppc), they can be instrumented by their respective t – 2, t – 3, and longer lags. The validity of a particular assump- tion can then be tested using standard generalized methods of moment tests of overidentifying restrictions. In summary, the SYS-GMM specification should allow for greater flexibility, improved control for omitted time-varying factors through the lagged dependent variable and, finally, greater consistency even in the presence of endogenous regressors. However, it is important to stress that this estimator does not resolve endogeneity problems due to omitted variables with persistent effects, such as when the TV trend is correlated with the RRA (NRA) trend, as a result of an omitted (possibly time-invariant) variable. V. R E G R E S S I O N R E SU LT S This section presents the results of our econometric analyses. We present first the results of the static model and afterward the dynamic panel results. We evaluate the robustness of the results by testing whether the results are sensitive to country and time coverage and to the use of different indicators for key variables. Static Model Table 3 reports the static fixed effects results of different specifications based on equation (2), with columns (1)–(4) using RRA and (5)–(6) using NRA as the dependent variable. In every regression, the standard errors are corrected for heteroskedasticity and autocorrelation of unknown form and are clustered within countries. The results in column 1 show that the simple fixed effects specification, without controls apart from gdppc, yields statistically significant coefficients ( p-value , 0.01) for both the linear effect and the interaction effect of TV pen- etration with gdppc. The positive sign for the linear term and the negative sign for the interaction effect are consistent with hypotheses 1 and 2. The penetra- tion of TV is associated with a higher RRA at low levels of development but with a lower RRA at higher levels. In regression (1), the turning point for the relationship is a per capita GDP level of USD 6,013. This number is virtually identical to the sample median value, which is equal to USD 5,949. Columns (2) and (3) report regressions that control for the standard agricul- tural protection covariates (agricultural employment share, comparative advan- tage, country size, and the quality of democracy) and additional variables, such as trade openness, government consumption to GDP, and crises indicators. In column (4), we add a set of continent-year interaction effects to control for differences in regional protection dynamics. The different specifications 428 THE WORLD BANK ECONOMIC REVIEW T A B L E 3 . Effect of TV Penetration on Agricultural Protection Dependent variable RRA RRA RRA RRA NRA NRA Variables (1) (2) (3) (4) (5) (6) Log TV 8.839 6.912 8.358 8.131 9.433 8.025 (0.005) (0.012) (0.003) (0.004) (0.004) (0.014) Log TV * GDP per capita 2 0.147 2 0.122 2 0.138 2 0.125 2 0.177 2 0.148 (  100) (0.006) (0.006) (0.004) (0.003) (0.001) (0.001) GDP per capita 0.007 0.006 0.008 0.008 0.010 0.009 (0.018) (0.024) (0.008) (0.010) (0.009) (0.013) Employment share 2 1.200 2 1.054 2 0.909 2 1.426 2 1.247 (0.050) (0.081) (0.163) (0.015) (0.037) Land per capita 2 1.622 2 2.033 2 1.712 2 3.076 2 2.905 (0.137) (0.069) (0.223) (0.047) (0.085) Export share 2 10.885 2 9.133 2 8.974 2 15.993 2 12.964 (0.264) (0.354) (0.476) (0.134) (0.335) Log population 2 0.071 2 0.132 0.159 2 0.096 2 0.059 (0.765) (0.574) (0.629) (0.733) (0.848) Polity2 (democracy index) 0.951 0.926 0.831 1.314 1.224 (0.002) (0.001) (0.004) (0.001) (0.001) Government consumption 0.537 0.630 0.176 0.320 (0.217) (0.137) (0.726) (0.544) Trade to GDP 0.014 2 0.038 2 0.007 2 0.029 (0.772) (0.510) (0.884) (0.607) Sachs-Warner trade policy index 20.975 17.553 18.127 16.444 (0.000) (0.001) (0.000) (0.001) Lagged_1 crisis 2 0.090 0.847 0.382 1.185 (0.956) (0.601) (0.815) (0.463) Lagged_2 crisis 1.007 2.072 0.780 1.871 (0.427) (0.121) (0.536) (0.166) Time fixed effects Yes Yes Yes Yes Yes Yes Country fixed effects Yes Yes Yes Yes Yes Yes Continental-years interaction effects No No No Yes No Yes Observations 2,025 1,996 1,935 1,935 2,001 2,001 Countries 69 69 67 67 69 69 Adj. R 2 0.857 0.865 0.871 0.880 0.858 0.866 Critical GDP per capita 6,013 5,666 6,057 6,504 5,330 5,422 Note: OLS regressions; p-values based on robust standard errors clustered by countries in pa- rentheses; all controls entered with one year lagged; continental (Asia, Africa, Latin America) and year interaction effects included as indicated (see text). Source: Own calculations based on the data described in the text. yield consistent results both in terms of coefficients and the significance of the media variables. Columns (5) and (6) report the results of regressions analogous to columns (3) and (4) but using the nominal rate of assistance (NRA) as the dependent variable instead of RRA. The estimated media coefficients are very similar, and the results are thus robust to using different indicators of agricultural support. Olper and Swinnen 429 To put the estimates into perspective, we illustrate the size of the media effects using the Philippines and Taiwan as examples. These two countries have average per capita GDP values in the period covered by the analysis of USD 1,854 and USD 9,987, respectively, which are significantly lower and higher than the critical turning point of the estimated relationship. Using the estimated coefficients of the full model (column (4)), a 10 percent increase in the share of households with TVs would be associated with a 4.8 percent in- crease in agricultural protection in the Philippines, but the same increase would reduce the Taiwan’ agricultural protection by 6.4 percent.9 This finding sug- gests that if there is a causal affect, its magnitude could be substantial. In a working paper version of this article (Olper and Swinnen 2012), we pre- sented similar regressions using radio penetration as media variable. When using radio penetration as an indicator, the patterns of the results are similar to those obtained using TV, but the significance levels of the media variables are lower and less robust. More specifically, including additional controls, the radio penetration linear term is positively and significantly correlated with pro- tection, but the interaction term with per capita GDP, while still negative, is no longer statistically significant. A possible interpretation of these results is that radio is a more important news source in poor countries, whereas TV matters most in emerging and rich countries. Dynamic Panel Model As discussed in section IV, the results obtained from the (static) fixed effects model may still suffer from endogeneity bias, particularly as a result of mea- surement errors. To account for this problem, table 4 reports the results of dynamic panel estimates that control for persistency in agricultural protection. To avoid problems resulting from the use of an excessive number of instru- ments in the SYS-GMM estimator, the specification only controls for the stan- dard covariates (as in column (2) of table 3).10 Columns (1) and (2) report the results using ordinary least squares (OLS) and least squares with dummy variables. The OLS and least squares with dummy variables results serve as benchmarks for the evaluation of the SYS-GMM specification and should be upward and downward biased com- pared to the SYS-GMM, respectively. In the SYS-GMM estimates, the media variables and per-capita GDP are treated as endogenous variables and instru- mented with their t 2 2 and higher lagged values. The SYS-GMM results for RRA and NRA are presented in columns (3) and (4). As expected, the magni- tude of the coefficient on lagged RRA is above the estimated least squares with dummy variables value and below the estimated OLS value (and similar for 9. Both elasticities are evaluated at the mean value of the RRA distribution, equal to 12.7 percent. 10. Note that this does not substantially affect our results because the autoregressive term largely absorbs these omitted terms (see Roodman, 2009). 430 THE WORLD BANK ECONOMIC REVIEW T A B L E 4 . TV Penetration and Agricultural Protection: Dynamic Panel Model Dependent variable RRA RRA RRA NRA Estimation method OLS LSDV SYS-GMM SYS-GMM Variables (1) (2) (3) (4) Lagged RRA (NRA) 0.9104 0.7407 0.8309 0.8183 (0.000) (0.000) (0.000) (0.000) Log TV 0.838 2.058 2.956 3.218 (0.006) (0.020) (0.010) (0.017) Log TV * GDP per capita (  100) 2 0.028 2 0.037 2 0.078 2 0.092 (0.002) (0.006) (0.000) (0.000) GDP per capita 0.0014 0.0018 0.0037 0.0043 (0.002) (0.028) (0.000) (0.000) Employment share 0.024 2 0.195 0.145 0.181 (0.396) (0.230) (0.090) (0.066) Land per capita 2 0.248 2 0.481 2 0.459 2 0.518 (0.001) (0.133) (0.000) (0.002) Export share 2 3.124 2 1.153 2 5.394 2 7.618 (0.001) (0.799) (0.001) (0.000) Log population 2 0.022 0.261 2 0.040 2 0.049 (0.411) (0.715) (0.476) (0.456) Polity2 (democracy index) 0.123 0.337 0.190 0.236 (0.023) (0.001) (0.023) (0.008) Time fixed effects Yes Yes Yes Yes AR2 test ( p-value) 0.03 0.08 AR3 test ( p-value) 0.25 0.34 Hansen ( p-value) 0.39 0.65 Diff-in-Hansen ( p-value) 0.65 0.43 Instruments 82 82 Observations 1,984 1,984 1,984 2,058 Countries 69 69 69 69 R2 0.94 0.94 Critical GDP per capita 2,992 5,563 3,790 3,498 Note: p-values based on robust standard errors clustered by countries in parentheses; SYS-GMM based on xtabond2 in Stata, with instruments structured with lag (3) for RRA (NRA), and lag (2) for the media variables and gdppc; Additional instruments used for the level equation are the t – 3 first difference of the RRA (NRA) and the t – 2 first difference for media variables and gdppc; the collapse option is also used to control for instrument proliferation. LSDV represents least squares with dummy variables. Source: Own calculations based on the data described in the text. NRA). Moreover, neither the basic Hansen test of overidentifying restrictions nor the Difference-in-Hansen test, related to the additional instruments used by the level equation, detects any problem with instrumental validity. These obser- vations all suggest that our instruments are valid and informative and the SYS-GMM estimator is consistent. The estimated coefficients of TV penetration presented in columns (3) and (4) are significant and consistent with hypotheses 1 and 2 in the SYS-GMM re- gressions. These estimated coefficients measure short-term correlations. To compare them with the static results, one should use the long-run correlations, Olper and Swinnen 431 which can be obtained by dividing the estimated coefficients by one minus the autoregressive coefficient. We obtain values equal to 17.46 for the linear term and 2 0.0046 for the interaction with income level and similar values for the NRA specification. Thus, the magnitudes of the estimated (long-run) media correlations in the dynamic SYS-GMM model are (in absolute value) approxi- mately two times higher than those of the static model (see columns (3) and (5) of table 3). This result is consistent with the presence of attenuation bias due to measurement errors in the variables. This problem is exacerbated in the fixed effect specification but is efficiently accounted for in the generalized methods of moment approach (see Wooldridge, 2002, p. 313).11 Further Robustness Tests We performed a series of additional robustness tests to further check our results, testing whether the results are sensitive to country and time coverage and to the use of different indicators for the group size effect and development. Columns (1)–(4) of table 5 report sensitivity analyses for the media- protection relationship with different ranges of countries and periods. One problem with our results may be that both the structural adjustment programs of the 1980s and the beginning of the GATT Uruguay Round in the mid-1980s caused an effect that interfered with our media-protection relationship: a re- duction in agricultural taxation in developing countries and a reduction in agri- cultural protection in developed countries. Columns (1) and (2) examine this possibility by running the model using only observations before and after 1985, respectively. Although the magnitude of the estimated relationship is dif- ferent in both periods (approximately three times larger in magnitude after 1985),12 the relationship is also estimated with high precision for years before 1985, suggesting that the abovementioned confounding effects do not drive the results. Another possibility is that our nonlinear media-protection relationship is driven by some group of sensitive observations related to a particular group of poor or rich countries. To check this possibility, in columns (3) and (4), we ex- cluded from the regressions observations for countries with a gdppc lower than USD 1,000 and higher than USD 25,000, respectively.13 The results remain consistent and significant, although the sizes of the coefficients change some- what. Dropping observations for the poorest countries increases the magnitude 11. Note that when comparing the SYS-GMM and the static fixed effects results, at least in terms of the magnitude of the media effect, these may also differ because the two models differ not only in terms of the estimator used but also in terms of the covariates included. 12. One likely reason for this difference is that the expansion of TVs in developing and emerging countries did not begin before the early 1980s. 13. Note that the results are fairly robust to the use of other gdppc thresholds. For example, by excluding observations with gdppc values below USD 5,000 or higher than USD 20,000 and thus working with approximately half of the sample, the media-protection relationship is still statistically significant. 432 T A B L E 5 . Robustness Checks: SYS-GMM Regressions over Different Samples THE WORLD BANK ECONOMIC REVIEW Dependent variable RRA RRA RRA RRA RRA RRA NRA Interaction with gdppc gdppc gdppc gdppc gdppc emp emp Variables (1) (2) (3) (4) (5) (6) (7) Lagged RRA (NRA) 0.893 0.796 0.838 0.877 0.799 0.909 0.895 (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) Log TV 1.478 6.313 3.604 1.966 2.073 2 2.306 2 3.094 (0.025) (0.011) (0.050) (0.004) (0.016) (0.104) (0.028) Log TV * GDP per capita (or emps) 0.029 0.093 0.088 0.038 0.084 3.996 5.277 (0.001) (0.005) (0.000) (0.002) (0.001) (0.011) (0.001) GDP per capita 0.0014 0.0044 0.0041 0.0018 0.0039 0.0022 0.0032 (0.002) (0.004) (0.000) (0.005) (0.001) (0.027) (0.002) Other controls Yes Yes Yes Yes Yes Yes Yes Time dummies Yes Yes Yes Yes Yes Yes Yes AR2 test ( p-value) 0.35 0.01 0.04 0.04 0.07 0.03 0.08 AR3 test ( p-value) 0.10 0.27 0.25 0.25 0.61 0.37 0.52 Hansen ( p-value) 0.96 0.22 0.92 0.30 0.91 0.10 0.02 Instruments 69 70 76 79 84 69 69 Observations 834 1,206 1,720 1,698 1,785 1,984 2,058 Countries 55 68 63 69 63 69 69 Sample , 1985 . 1985 gdppc . USD 1,000 gdppc , USD 25,000 Richer/poorest excluded All All Critical GDP per capita (emps) 5,097 6,788 4,096 5,173 2,468 (0.58) (0.59) Notes: Each column reports a regression for a specific countries/years sample with characteristics indicated at the bottom of the table: , ( . ) 1985 means that the regression is run for data before or after 1985; gdppc . 1,000 ( , 25,000) means that the regression considered only countries/years data with gdppc higher (lower) than 1,000 (25,000) international U.S. dollars; richer/poorest excluded are excluding the three richest and poorest countries from the regression (namely, the United States, Norway, and Switzerland and Mozambique, Ethiopia, and Zimbabwe). The p-values in parentheses are based on robust standard errors clustered by countries; SYS-GMM is based on xtabond2 in Stata, with instruments structured with lag (3) for RRA and lag (2) for media variables and gdppc (emps); additional instruments used for the level equation are the t – 3 first difference of the RRA and the t – 2 first difference for media variables and gdppc; the collapse option is used to control for instrument proliferation. Additional controls included in every re- gression are emps, landpc, exps, log( pop), and Polity2, all lagged one year, and year fixed effects. Source: Own calculations based on the data described in the text. Olper and Swinnen 433 of the estimated relationship; dropping observations for the richest countries reduces the estimated effects compared to the benchmark SYS-GMM regression in Column (3) of table 4. These results are consistent with hypothesis 2 (adver- tiser value effect), suggesting that farmers in the poorest developing countries may be too poor to attract advertisers and media coverage. The findings may also be consistent with an unbalanced increase in media diffusion between rural and urban areas. If rural areas have some lag in media penetration com- pared to urban areas, the media protection relationship will be weaker in the poorest countries. In column (5), we simultaneously exclude observations for the three richest (the United States, Norway, and Switzerland) and the three poorest countries (Mozambique, Ethiopia, and Zimbabwe) from the regression. Once again, the results are very robust. Columns (6) and (7) of table 5 present an additional robustness check using the agricultural employment share (emps) to interact with the media effect. For both RRA and NRA, all key variables have signs consistent with our hypothe- ses. Although the regression results with emps are less stable than those with gdppc (likely as a result of severe measurement errors in the emps variable), the key conclusions are robust to a change in the structural variable.14 Finally, as discussed in the conceptual framework section, there may be ad- ditional nonlinearity in the relationship between media coverage and policies resulting from a nonlinear relationship between economic growth and the rural-urban income gap and the relative urban-rural gap in TV distribution, which would affect the value of both groups to advertisers and thus the channel through which the mass media affect agricultural protection. To test for this, we added an additional interaction effect between TV and the square of gdppc to the model specifications. In some specifications, this additional media interaction term is significant and negative, consistent with the hypothe- sis. However, the results are less robust than those without the additional non- linear term. These results are presented in Olper and Swinnen (2012). VI. CONCLUSIONS This paper provides evidence of the relationship between mass media competi- tion and agricultural protection for a large group of countries. Stro¨ mberg’s (2004a) theory predicts that information provided by the mass media, reflecting 14. There are well-known problems with the agricultural employment data, which are generally linear interpolations between a few observations (often one per decade) and suffer from serious measurement errors (see Timmer and de Vries, 2007). The model specification tests reported at the bottom of the table indicate a well-specified SYS-GMM model for the RRA regression but not for the NRA regression, where the Hansen test rejects the null of the validity of the additional overidentifying restrictions. However, the autocorrelation tests indicate that the model is well specified. Considering the strong measurement errors in the employment data and that Hansen tests have weak power, such results are unsurprising. 434 THE WORLD BANK ECONOMIC REVIEW the media’s incentives to provide news to different groups in society, affects government policy making and who benefits from government policies. The theory predicts that mass media competition will induce a policy bias toward large groups and groups that are more valuable to advertisers; these groups are more informed because the mass media target them. We apply this theory to agricultural policy. This results in the hypotheses that (a) given the changing role of the agricultural sector due to economic de- velopment, the impact of mass media competition on agricultural policy will differ between poor and rich countries, ceteris paribus, and (b) this effect is contrary to the so-called development paradox of agricultural policies. Thus, the traditional change in agricultural policy from taxation to subsidization that is associated with economic development will be smoothed in the presence of mass media competition. We hypothesize that this is due to a combination of the group size effect, with larger groups being more attractive to the media, and the advertiser value effect, with richer groups being more attractive audi- ences for the media. We use data on agricultural policy from 69 countries spanning a wide range of development stages and media markets to test these predictions. Our empiri- cal results are consistent with the theoretical hypotheses. We find a significant and robust correlation between public support for agriculture and TV penetration, which is conditional on the structure of the economy. In particular, an increase in media penetration is correlated with policies that benefit the majority to a greater extent; it is correlated with a reduction in agricultural taxation in poor countries and a reduction in agricultural subsidies in rich countries, ceteris paribus. These results are robust to the use of different indicators of agricultural poli- cies, different media variables and different control variables and estimation techniques. REFERENCES Anderson, K. 1995. “Lobbying Incentives and the Pattern of Protection in Rich and Poor Countries.” Economic Development and Cultural Change 43(2): 401 –23. ——— 2009. “Five Decades of Distortions to Agricultural Incentives.” In K. Anderson, ed., Distortions to Agricultural Incentives, 3 – 64. Washington, DC: The World Bank. Anderson, K., and Y. Hayami. 1986. The Political Economy of Agricultural Protection. Sydney, Australia: Allen and Unwin. Anderson, K., and E. Valenzuela. 2008. “Estimates of Distortions to Agricultural Incentives, 1955 to 2007.” Washington, DC: The World Bank. Anderson, K., G. Rausser, and J. Swinnen. Forthcoming. “Political Economy of Public Policies: Insights from Distortions to Agricultural and Food Markets.” Journal of Economic Literature. Arellano, M., and O. Bover. 1995. “Another Look at the Instrumental Variable Estimation of Error-Components Models.” Journal of Econometrics 68: 29–51. Becker, G. S. 1983. “A Theory of Competition among Pressure Groups for Political Influence.” Quarterly Journal of Economics 98: 371 –400. Olper and Swinnen 435 Besley, T., and R. Burgess. 2001. “Political Agency, Government Responsiveness and the Role of the Media.” European Economic Review 45: 629 –40. ———. 2002. “The Political Economy of Government Responsiveness: Theory and Evidence from India.” The Quarterly Journal of Economics 117(4): 1415– 51. Blundell, R. W., and S. R. Bond. 1998. “Initial Conditions and Moment Restrictions in Dynamic Panel Data Model.” Journal of Econometrics 87: 115–43. Bond, S., A. Hoeffler, and J. Temple. 2001. “GMM Estimation of Empirical Growth Model.” CEPR Discussion Papers 3048. Center for Economic and Policy Research, Washington, DC. Curtis, K., J.J. McCluskey, and J. Swinnen. 2008. “Differences in Global Risk Perceptions of Biotechnology and the Political Economy of the Media.” International Journal of Global Environmental Issues 8 (1/2): 77 –89. Deaton, A. 2005. “Measuring Poverty in a Growing World (or Measuring Growth in a Poor World).”Review of Economics and Statistics 87 (1):1–19. De Gorter, H., and J. Swinnen. 2002. “Political Economy of Agricultural Policy.” In B. Gardner, and G. Rausser, eds., Handbook of Agricultural Economics, Vol. 2B, 1893– 1943. Amsterdam, The Netherlands: Elsevier. Downs, A. 1957. An Economic Theory of Democracy. New York: Harper and Row. Francken, N., B. Minten, and J. Swinnen. 2009. “Media, Monitoring, and Capture of Public Funds: Evidence from Madagascar.” World Development 37 (1): 242– 55. ———. 2012. “The Political Economy of Relief Aid Allocation: Evidence from Madagascar.” World Development 40 (3): 486 –500. Hayami, Y. 2007. “An Emerging Agricultural Problem in High-Performing Asian Economies.” Policy Research Working Paper 4312. The World Bank, Washington, DC. International Telecommunication Union (ITU). 2010. World Telecommunication Indicators 2009. Geneva, Switzerland: ITU. http://www.itu.int/. Kuzyk, P., and J.J. McCluskey. 2006. “The Political Economy of the Media: Coverage of the U.S.-Canadian Lumber Trade Dispute.” World Economy 29 (5): 637–54. McMillan, M.S., and D. Rodrik. 2011. Globalization, Structural Change and Productivity Growth. NBER Working Paper 17143. National Bureau of Economic Research, Cambridge, MA. Marks, L. A., N. Kalaitzandonakes, and S. S. Vickner. 2003. “Evaluating Consumer Response to GM Foods: Some Methodological Considerations.” Current Agriculture, Food & Resource Issues 4: 80– 94. Marshall, M. G., and K. Jaggers. 2007. Polity IV Project: Dataset Users’ Manual. Arlington, VA: Polity IV Project. McCluskey, J., and J. Swinnen. 2010. “Media Economics and the Political Economy of Information.” In D. Coen, W. Grant, and G. Wilson, eds., The Oxford Handbook of Business and Government, 643–62. Oxford: Oxford University Press. Oberholzer-Gee, R., and J. Waldfogel. 2005. “Strengh in Numbers: Group Size and Political Mobilization.” Journal of Law and Economics 48: 73–91. Olper, A. 2007. “Land Inequality, Government Ideology and Agricultural Protection.” Food Policy 32: 67– 83. Olper, A., J. Falkowski, and J. Swinnen. 2013. “Political Reforms and Public Policies: Evidence from Agricultural Protection.” World Bank Economic Review. doi:10.1093/wber/lht003. Olper, A., and V. Raimondi. Forthcoming. “Electoral Rules, Forms of Government and Redistributive Policy: Evidence from Agriculture and Food Policies.” Journal of Comparative Economics. doi:10.1016/j.jce.2012.03.002. Olper, A., and J. Swinnen. 2012. “Mass Media and Public Policy: Global Evidence from Agricultural Policies.” LICOS Discussion Paper 320. LICOS Centre for Institutions and Economic Performance, Leuven, Belgium. 436 THE WORLD BANK ECONOMIC REVIEW Olson, M. Jr. 1965. The Logic of Collective Action. Public Goods and the Theory of Groups. Cambridge, MA: Harvard University Press. ¨ mberg. 2005. “Commercial Television and Voter Information.” CEPR Discussion Prat, A., and D. Stro Papers 4989. Center for Economic and Policy Research, Washington, DC. ———. 2011. “The Political Economy of Mass Media.” CEPR Discussion Papers 8246. Center for Economic and Policy Research, Washington, DC. Reinikka, R., and J. Svensson. 2005. “Fighting Corruption to Improve Schooling: Evidence from a Newspaper Campaign in Uganda.” European Economic Review 3 (2– 3): 259– 67. Roodman, D. 2009. “A Note on the Theme of Too Many Instruments.” Oxford Bulletin of Economics and Statistics 71 (1): 135–58. ¨ mberg, D. 2001. “Mass Media and Public Policy.” European Economic Review 45: 652– 63. Stro ——— 2004a. “Mass Media Competition, Political Competition, and Public Policy.” Review of Economic Studies 71: 265– 84. ——— 2004b. “Radio’s Impact on Public Spending.” Quarterly Journal of Economics 119: 189–221. ¨ mberg, D., and J.M. Snyder. 2008. “The Media’s Influence on Public Policy Decisions.” In Stro R. Islam, ed., Information and Public Choice. From Media Markets to Policy Making, 17– 31. Washington, DC: The World Bank. Swinnen, J. 1994. “A Positive Theory of Agricultural Protection.” American Journal of Agricultural Economics 76 (1): 1– 14. ——— 2010. “The Political Economy of Agricultural Distortions: The Literature to Date.” In K. Anderson, ed., Political Economy of Distortions to Agricultural Incentives, 81– 104. Cambridge, UK: Cambridge University Press. Swinnen, J., and N. Francken. 2006. “Summits, Riots, and Media Attention: The Political Economy of Information on Trade and Globalization.” The World Economy 29 (5): 637– 54. Swinnen, J., J. McCluskey, and N. Francken. 2005. “Food Safety, the Media, and the Information Market.” Agricultural Economics 32 (1): 175 –88. Timmer, M. P., and G. J. de Vries. 2007. A Cross-Country Database for Sectoral Employment and Productivity in Asia and Latin America, 1950-2005. Groningen, The Netherlands: Groningen Growth and Development Centre Research Memorandum, GD-98, University of Groningen. Verbeke, W., R.W. Ward, and J. Viaene. 2000. “Probit Analysis of Fresh Meat Consumption in Belgium: Exploring BSE and Television Communication Impact.” Agribusiness 16: 215–34. Vigani, M., and A. Olper. 2012. “GMO Standards, Endogenous Policy and the Market for Information.” LICOS Discussion Paper No. 306/2012. LICOS Centre for Institutions and Economic Performance, Leuven, Belgium. Wacziarg, R., and K.H. Welch. 2008. “Trade Liberalization and Growth: New Evidence.” The World Bank Economic Review 22 (2): 187 –231. Wooldridge, J. M. 2002. Econometric Analysis of Cross Section and Panel Data. Cambridge, MA: MIT Press. Liability Structure in Small-Scale Finance: Evidence from a Natural Experiment Fenella Carpena, Shawn Cole, Jeremy Shapiro, and Bilal Zia Microfinance, the provision of small individual and business loans, has experienced dramatic growth, reaching over 150 million borrowers worldwide. Much of the success of microfinance has been attributed to attempts to overcome the challenges of information asymmetries in uncollateralized lending. However, very little is known about the optimal contract structure of these loans, and there is substantial variation across lenders, even within a particular setting. This paper exploits a plausibly exogenous change in the liability structure offered by a microfinance program in India, which shifted from individual to group liability lending. We find evidence that the lending model matters: for the same borrower, the required monthly loan installments are 11 percent less likely to be missed under the group liability setting in comparison with individual liability. In addition, compulsory savings deposits are 20 percent less likely to be missed under group liability contracts. JEL codes: D14, O12, O16, O17 Theory and evidence highlight financial market imperfections as a central cause of poverty and a key impediment to growth (Banerjee and Newman, 1993; Rajan and Zingales, 1998). In theories of capital accumulation, for example, financial market imperfections influence the ability of the poor to borrow for investments in education and physical capital. Additionally, in models that explain entrepreneurship, information asymmetries and transaction costs prevent the poor from undertaking profitable entrepreneurial activities because they often have no collateral. The lack of access to financial services Fenella Carpena is a PhD student in the Department of Economics at the University of California, Berkeley; her email address is fenella@econ.berkeley.edu. Shawn Cole is an Associate Professor of Business Administration at Harvard Business School; his email address is scole@hbs.edu. Jeremy Shapiro is a Consultant at McKinsey & Company; his email address is jeremypshapiro@gmail.com. Bilal Zia (corresponding author) is an Economist in the Development Research Group at the World Bank; his email address is bzia@worldbank.org. This project is a collaborative research effort with Saath Microfinance. The authors thank Saath, Xavier Gine ´ , David McKenzie, Petia Topalova, and workshop participants at the World Bank for helpful comments and suggestions. Stuti Tripathi and the Center for Microfinance at IFMR provided excellent research assistance. This work was supported by the World Bank Gender Action Plan, the HBS Division of Faculty Research and Development to S.C., and the NSF Graduate Research Fellowship Program to F.C. THE WORLD BANK ECONOMIC REVIEW, VOL. 27, NO. 3, pp. 437 –469 doi:10.1093/wber/lhs031 Advance Access Publication December 4, 2012 # The Author 2012. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 437 438 THE WORLD BANK ECONOMIC REVIEW may thus leave many productive opportunities for the poor untapped and may generate persistent income inequality and low growth (Beck et al., 2007).1 Microfinance, which is the provision of credit, savings, and other financial services to low-income households and entrepreneurs, has exploded in popular- ity and coverage in recent years, particularly in its ability to meet the large unmet demand for finance (Morduch, 1999; Armenda ´ riz de Aghion and Morduch, 2010). Both emerging markets and developed economies, including the United States, now provide microfinance services through a variety of public and private channels. The growth of microfinance has been unprecedent- ed: between 2004 and 2008, the sector’s average annual asset growth rate was 39 percent, reaching US$60 billion in total assets by the end of 2008 (Chen et al., 2010). A careful evaluation of microfinance in Banerjee et al. (2009) reveals that microcredit has important effects on business outcomes and the composition of household expenditures. The rapid growth of microfinance and its potential for promoting development have attracted the interest not only of governments, donors, and socially oriented investors but also of mainstream commercial banks. Perhaps the most celebrated feature of microfinance is the group liability contract, which is a lending methodology pioneered by the Grameen Bank in Bangladesh. Under this contract, loans without collateral are extended to a group of borrowers whose members are jointly liable for each other’s repay- ment. Because groups form voluntarily and group members are responsible for paying off each other’s debts, borrowers have the incentive to screen for risky clients, monitor their peers, and enforce repayment. The success of this model with the Grameen Bank led to its replication in many countries around the world, reaching more than 150 million individuals by the end of 2007 (Daley-Harris, 2009). This model is particularly important because small firms suffer the most from institutional weakness (Beck et al., 2005) and because the structure of the banking sector may have important distributional impacts on growth (Cetorelli and Gambera, 2001). Although most microfinance organizations use group liability, not all do so. On the one hand, group liability may solve information asymmetry problems by leveraging social ties and the borrower’s knowledge about the community, and it may reduce monitoring costs to the lender by motivating borrowers to monitor each other. On the other hand, social sanctions may be limited, bad clients may free-ride on good clients, and borrower groups may collude against the lender. In addition to group liability lending, many microfinance programs employ a variety of approaches to maintain high repayment rates. For example, some programs implement frequent repayment schedules and progressive lending, or they require collateral substitutes. However, very little is known about the efficiency of these designs in ensuring repayment. 1. See World Bank (2008) for a literature summary. Carpena, Cole, Shapiro, and Zia 439 The question of an optimal loan contract structure remains largely unan- swered in both the theoretical and the empirical microfinance literature. Theoretical studies have primarily focused on explaining how and why the group liability mechanism works and offering competing predictions on its benefits, whereas the empirical literature lags behind the theory. Two impor- tant exceptions are Gine ´ and Karlan (2009) and Attanasio et al. (2011). Gine ´ and Karlan (2009) report on a field experiment in the Philippines to test the effect of individual versus group liability lending. Their analysis focuses on the importance of peer monitoring and finds no significant difference in default among individual and group borrowers. Gine ´ and Karlan identify the effects of peer monitoring, but they do not focus on the effects of joint liability. In con- trast, our paper examines the effect of the contract structure on the group of borrowers who are willing to borrow with either individual or group liability. Attanasio et al. (2011) conducted a field experiment in Mongolia in which vil- lages were randomly assigned to obtain access to group loans, individual loans, or no loans. The main objective of Attanasio et al. (2011) is to measure the impact of both types of microcredit on different poverty measures. The authors find a positive impact for group liability loans on food consumption and entre- preneurship, with no difference in repayment rates between individual and group liability. The identification of the impact of group liability on outcomes such as the default rate is complicated by the standard problems of selection and omitted variable bias. Individuals with different financial habits might choose one form of contract but not the other. Alternatively, lenders with different levels of sophistication may attract different client mixes and offer different contracts. One cannot simply compare clients across lending contracts because self- selection or other aspects of the program may be the cause of any observed differences. In this paper, we use a natural experiment to compare loan repayment and savings discipline between individual and group lending models.2 In this setting, group lending differs from individual lending in both the liability struc- ture and the repayment practices. In group lending, borrowers are liable for the scheduled payments of group members, and the loan officer interacts primarily with the group leader, who collects payments from the other group members. Under individual lending, the borrowers are personally liable and interact directly with the loan officer. Our empirical strategy takes advantage of a change in the lending policies of Saath, a non-government organization provid- ing microfinance services in India. Saath switched from individual to group lending. This transition was governed by a strict policy rule: after a particular date, all of the borrowers completing an individual liability cycle were offered a group liability loan in their next loan cycle. The individual liability loan com- pletion dates were distributed relatively uniformly throughout the year, offering 2. Throughout this paper, we use the terms “group liability” and “joint liability” interchangeably. 440 THE WORLD BANK ECONOMIC REVIEW a natural variation in the timing of the loan contract transitions. Thus, in July, for example, individual liability borrowers completing a loan cycle would switch to group liability in the following loan, whereas those whose loan cycle ended after July would remain under an individual contract setting until the end of their cycle. This plausibly exogenous change, which was phased in over time, generates natural control groups and allows us to credibly identify the causal impact of the group liability structure in what amounts to a repeated difference-in-difference framework. At any particular point in time, our “treat- ment” group consists of clients who have fully repaid their individual liability loan and who currently have a group liability loan, whereas our “control” group consists of individual liability loan clients who will eventually convert to a group liability loan. Our primary analysis focuses on loan performance and estimates the effect of group liability on these outcomes. We find that the group liability structure significantly improves repayment rates. In particular, clients are approximately 11 percent less likely to miss a monthly repayment in the group liability setting relative to individual liability; this effect holds even with individual fixed effects. We also find that there is greater discipline regarding the monthly com- pulsory savings deposits when clients have a group liability loan. Specifically, compulsory deposits are approximately 20 percent less likely to be missed in the group liability setting. Our results provide the first credible evidence that group liability contracts improve upon individual liability, particularly in en- suring repayment and increasing savings discipline among clients. These results, however, are subject to some important caveats. First, the transition to group liability lending was accompanied by other changes to the lending structure, particularly an increase in the loan size, which may have raised the continuation value of borrowing. Nevertheless, we argue that our es- timates of the impact of these changes in the lending model should be in the lower bounds. Second, because our empirical strategy focuses on clients who chose to borrow under both individual and group liability settings, the external validity of our results may be limited. Third, limitations in the data availability preclude us from examining loan outcomes such as delinquency or prepayment. Finally, our empirical strategy does not allow us to test for the specific mecha- nisms by which group lending improves repayment. Based on our discussions with Saath, interviews of field officers, and our reading of the evidence, we speculate that “peer pressure” is the mechanism at work. We discuss these caveats further in Section IV. From a practical and policy perspective, our results are quite timely. Microlenders worldwide are increasingly weakening joint liability in their lending approaches (Armenda ´ riz de Aghion and Morduch, 2010). BancoSol in Bolivia has shifted a significant proportion of its lending portfolio from group to individual lending, and even the Grameen Bank has moderated its joint lia- bility clause, allowing defaulters to get back on track without invoking group pressure. Therefore, our results suggest a cautionary tale for microfinance. Carpena, Cole, Shapiro, and Zia 441 Many MFIs are moving away from joint liability to individual liability, but this transition is not supported by strong empirical evidence. This finding is impor- tant because, to our knowledge, only two other papers examine the relative merits of joint and individual liability contracts. Our paper underscores the fact that more research is required to provide better policy guidance for MFI practitioners worldwide. The rest of this paper is organized as follows. Section I reviews the existing literature on the liability structure in microfinance. Section II provides the back- ground for the microfinance program that we study and explains the change in the liability structure of its loan products. In Section III, we provide a descrip- tion of the data and summary statistics. We discuss our empirical strategy and results in Section IV. Finally, Section V concludes the paper. I. PREDICTIONS OF GROUP LIABILITY A wealth of theoretical literature in microfinance explores the mechanisms behind group liability contracts, particularly in terms of how they mitigate in- formation asymmetries and enforcement problems. Stiglitz (1990) shows that the group liability structure overcomes ex ante moral hazard because it creates incentives for group members to monitor each other’s loans. Similarly, Banerjee et al. (1994) study credit cooperatives and underscore the role of peer monitoring. These authors describe a model in which higher monitoring results in higher borrower effort and, hence, a higher probability of project success. Even if a project succeeds, however, borrowers may refuse to repay or may claim that the project failed so that they can avoid repayment. This type of strategic default is captured in several theoretical studies on group liability. For example, Besley and Coate (1995) provide a model demonstrating that joint lia- bility may harness social capital to increase a borrower’s willingness to repay. Likewise, Armenda ´ riz de Aghion (1999) demonstrates that joint liability agree- ments may reduce the incidence of strategic default because borrowers may impose social sanctions on the defaulter. In addition to examining moral hazard, the theoretical literature investigates how joint liability mitigates adverse selection. Ghatak (2000) describes a model in a scenario in which borrowers have ex-ante information about the riskiness of other borrowers’ investment projects, but lenders do not. Joint lia- bility thus acts as a screening device that induces “assortative matching.” Specifically, borrowers with safe investments partner with other safe borrowers, leaving risky borrowers to form groups among themselves. These theoretical models, among others, have shown that group liability may improve repayment rates by alleviating imperfections in the credit market. However, whether group liability outperforms other contract structures remains an open question in the microfinance literature. For example, Besley and Coate (1995) point out in their model that if borrowers cannot repay as a group, then some group members will not find it worthwhile to contribute 442 THE WORLD BANK ECONOMIC REVIEW their share of repayment even though they would have repaid under individual lending. Inconclusive empirical evidence accompanies these ambiguous theoretical predictions. Some empirical studies support the theoretical advantages of group liability. For instance, in Bangladesh, Sharma and Zeller (1997) show that groups that were formed through self-selection had better repayment rates; however, this study may suffer from omitted variable biases. Other studies provide little empirical support for this theory. For example, Ahlin and Townsend (2007) use Thai data to show that repayment rates are negatively as- sociated with social ties. Only a handful of studies examine the merits of group liability relative to other contract structures. Fischer (2010) conducts a series of lab experiments with actual microfinance clients and provides evidence that the contract struc- ture affects project selection. Specifically, he finds that group liability increases risk-taking relative to individual liability contracts because borrowers free-ride on the insurance provided by their partners. In a randomized experiment in India in which borrowers were assigned to either weekly or monthly repayment meetings, Feigenberg et al. (2010) find that more frequent repayment meetings build social capital among borrowers, which, in turn, leads to reduced default. The most relevant studies on repayment rates under different loan liability structures are Gine ´ and Karlan (2009) and Attanasio et al. (2011). Gine ´ and Karlan (2009) report evidence from two field experiments in the Philippines. In the first experiment, borrowers who had signed up under a group liability structure were converted to individual liability. Because both the joint and indi- vidual liability groups previously underwent the same screening, the authors can independently identify the peer monitoring effect under group liability. However, they cannot identify or rule out any impact of screening with this methodology. In addition, the group repayment and monitoring mechanisms may be entrenched and difficult to undo, even with an individual liability struc- ture. Their second experiment randomly introduced either group or individual liability lending to new borrowers. However, the experiment was conducted at the loan center level, and take-up was quite uneven between the group and in- dividual loan centers, resulting in potential statistical power concerns. In both instances, the authors find that the default rates are invariant to the contract structure. Attanasio et al. (2011) conduct a field experiment in Mongolia where villag- es were randomly assigned to group loans, individual loans, or no loans. The authors seek to measure the impact of individual and group loans on reducing poverty. In particular, the study finds that clients who received group loans had higher food consumption and were more likely to operate a business than clients in the control villages. Clients in the individual-lending villages had no significant increases on any of these measures. For the repayment outcomes, the authors find no significant differences in the repayment rates between indi- vidual and group liability. Carpena, Cole, Shapiro, and Zia 443 Although loan default and repayment are the primary outcomes of interest when examining group liability contracts, the economics literature on rotating savings and credit organizations (Roscas) suggests that group liability may have positive effects on savings. Bouman (1995) argues that participating in credit and savings groups allows individuals to avoid demands for financial support from their relatives because contributions to a Rosca are generally recognized by society as senior claims. In a theoretical model, Ambec and Treich (2007) show that Roscas can serve as a commitment device that helps people to over- come self-control problems. Gugerty (2007) provides support for this model, reporting that many Rosca participants in rural Kenya cite “you can’t save alone” or “sitting with other members helps you to save” as their primary mo- tivation for participating in a Rosca. Our paper complements Gine ´ and Karlan (2009) by examining the optimal contract structure in an alternative setting. Although the original experiment in Gine´ and Karlan (2009) focuses on moving from group to individual liability contracts, we explore the reverse; that is, we explore the shift from individual to group liability. The following section describes the setting and our empirical strategy in more detail. II. EMPIRICAL SETTING Our partner institution, Saath, is a non-government organization based in Ahmedabad, India. Founded in 1989, Saath implements development initiatives in slum communities, including health, infrastructure improvement, and liveli- hood training programs. Additionally, Saath provides credit and savings servic- es to the urban poor through its microfinance unit. In 2009, Saath Microfinance had over 6,400 active clients in 4 branches with a savings portfo- lio of INR 18 million (USD 390,000) and a loan portfolio of INR 19 million (USD 410,000).3 Although Saath has provided mentoring support to community-based credit and savings groups since the mid-1990s, its microfinance unit was not formally established until 2002. In that year, Saath integrated these credit and savings groups into its organization and registered them as cooperative societies with the Indian government. Saath also began managing these credit and savings co- operatives at this time, evolving into the Saath Microfinance Unit. Today, Saath Microfinance provides various financial services to slum communities, in- cluding voluntary savings accounts, compulsory savings accounts, and group li- ability loans. Savings Products Since its inception in 2002, Saath Microfinance has offeredvoluntary savings accounts to its clients. These voluntary savings accounts earn an interest rate of 3. Based on Saath’s 2008–2009 Annual Report. 444 THE WORLD BANK ECONOMIC REVIEW 6 percent per year and do not require a minimum balance. As the name sug- gests, members are not obliged to make regular deposits into voluntary savings accounts. Any amount can be deposited, but only six withdrawals can be made per year. In November 2007, Saath Microfinance initiated compulsory savings ac- counts for its members. Specifically, members are required to deposit INR 100 (USD 2) every month into compulsory savings accounts for the duration of their membership with Saath Microfinance. Clients may withdraw any amount from their compulsory savings at any time as long as a minimum balance of INR 3,500 (USD 70) is maintained. Similar to voluntary savings, compulsory savings earn an interest of 6 percent per year. Any amount that the client de- posits over the compulsory savings of INR 100 is deposited into the client’s voluntary savings account. The goal of the compulsory savings account is to allow clients to build a financial buffer stock against adverse shocks and to provide low-cost capital to Saath. These compulsory deposits were mandated for all borrowers independent of the switch to group liability loans. Hence, all outstanding loans under both individual and group liability were required to make compulsory deposits after November 2007. In Section IV, we compare adherence to these compulsory deposits for the same person as she moves from individual to group liability. Loan Products In addition to savings products, Saath Microfinance provides loans for asset creation (e.g., house repairs), production (e.g., business working capital), and consumption (e.g., health, social functions). From Saath’s beginnings in 2002 until November 2007, it provided credit through individual liability loans. Beginning in November 2007, Saath discontinued its individual liability loans, instead offering group liability loans to members applying for credit. Under the individual liability loan model, a client was required to have been a member of Saath for at least six months with a savings account to be eligible for a loan. Members could borrow up to three times their savings account balance at an interest rate of 18 percent per year.4 These individual-liability loans generally require no collateral; however, each loan applicant must meet two requirements. First, the loan applicant must have two “guarantors” who also have savings accounts with Saath. Second, the combined savings balances of the loan applicant and the two guarantors must be greater than or equal to the loan amount applied for. Although the guarantors are, in principle, re- quired to maintain these savings balances throughout the duration of the loan, in practice, this rule is not strictly enforced. The guarantors are not eligible for 4. Microfinance organizations typically quote interest rates in one of two forms: “declining,” the standard used in developed markets, where the amount of interest due each period is calculated based on the interest rate and the remaining principal, and “flat,” where the interest payments are calculated using the original principal amount. Thus a 10 percent “flat” rate is significantly higher than a 10 percent “declining” rate. Saath quotes rates using the standard declining balance approach. Carpena, Cole, Shapiro, and Zia 445 a loan until the loan that they guaranteed has been fully repaid, but loan repay- ment is the sole responsibility of the borrower. The borrowers are required to make monthly installments that cover principal and interest. The monthly prin- cipal installment is a fixed amount, and because the interest rate reflects a de- clining balance, the total installment amount ( principal plus interest) varies every month. If the borrower defaults, Saath reserves the right to seize the bor- rower’s savings. If the savings are not sufficient to cover the loan, Saath re- serves the right to take the guarantor’s savings as well. However, in practice, as an NGO whose mission is to empower the poor, Saath has never seized any of its individual borrowers’ or guarantors’ savings. With the group liability model, however, Saath extends credit to groups of individuals at an interest rate of 24 percent per year. Four loan size categories are available to clients: (1) Rs. 3,000–5,000, (2) Rs. 6,000–10,000, (3) Rs. 11,000–20,000, and (4) Rs. 21,000–30,000. These groups are formed primarily through self-selection with joint applications submitted to Saath. The groups are composed of three to six individuals, all of whom must be Saath Microfinance members. Within each group, several criteria must be fulfilled. First, at least 50 percent of the group must have been Saath Microfinance members for at least 6 months and must have at least asavings account. Second, at least 50 percent of the group must be female. Third, relatives or in- dividuals from the same household are not allowed in the same group. Finally, the loan terms must be homogenous across group members; that is, the number of installments and the monthly installment due dates must be the same, and the loan amount must not vary widely within each group. As in the individual liability model, group liability borrowers are required to make monthly installment payments for both principal and interest, although in this setting, the total installment amounts ( principal plus interest) are equal every month. (In the individual liability model, the monthly principal installment re- payment is fixed, but the interest and, therefore, the installment size vary each month.) Before any loans are disbursed, the group members are also required to sign a “mutual agreement form” stating that they are liable to pay each other’s debts in the event of default or delinquency. Borrower groups who have defaulted or are delinquent are not eligible to receive another loan from Saath. The Shift from Individual to Group Liability Saath’s decision to shift from offering individual liability to group liability loans in November 2007 was driven by a change in the management’s priori- ties. Saath wanted to lend to more people, provide larger loan amounts, and expand its microfinance operations geographically, but its lending activities had become stagnant under the individual liability model. In particular, the “guarantors” requirement for individual liability loans restricted credit eligibili- ty because Saath had already reached a point where almost all of its members were either borrowers or guarantors. Additionally, savings clients were reluc- tant to stand as guarantors for other clients’ loans, and the loan amounts were 446 THE WORLD BANK ECONOMIC REVIEW limited to 3 times the total savings account balance of the borrower. Saath’s management thus shifted to group liability loans to overcome these restrictions in its individual liability model. In terms of the models discussed above, the limited ability of Saath members to pledge savings as collateral prevented Saath from expanding, and Saath saw group liability as a way to solve this problem. In the year following this change, Saath gained almost 800 new clients and in- creased its reach from 11 to 20 wards. The transition from individual liability to group liability loans was imple- mented using the following rule. Beginning in November 2007, all new loans disbursed were group liability loans; Saath would no longer disburse individual liability loans. However, existing loans whose terms lasted beyond November 2007 were unaffected. For example, individual liability loan clients who com- pleted their loan in February 2008 continued under the individual liability con- tract until then. After February 2008, if they chose to borrow again, they received a group liability loan. The date to switch from individual to group lia- bility was therefore determined by the individual liability loan completion dates. These completion dates and the subsequent conversion to group liability loans were distributed relatively uniformly throughout the year. Although Saath’s loan product moved from individual to group liability be- ginning in late 2007, the location for repayments, the frequency of loan repay- ment collection, and the salaries for the field officers remained similar across time in our dataset. An empirical concern for the subsequent analysis is whether Saath’s policy shift from individual to joint liability was accompanied by a shift in its loan collection techniques. In June 2010, we conducted short interviews with Saath field officers regarding repayment collection. We were able to interview 10 out of the 12 Saath field officers who collected repayment among the sample of clients we study; the remaining field officers are no longer with Saath. The results from this survey confirm that the location and frequen- cy of loan repayments remained the same across the two settings. Specifically, for both individual and group liability loans, we find that field officers collected repayment at the client’s household every month 100 percent of the time. Additionally, 90 percent of the time under both regimes, the lender turned down future loan requests from defaulting clients. For group liability borrowers, Saath does not require a group to designate one of its members as the “group leader.” However, in practice, all groups have a leader who is in charge of collecting repayments from the other members. Hence, for individual borrowers, the field officers visit each bor- rower, whereas for group borrowers, the field officers typically visit only the group leader’s household. If any member of a particular borrower group fails to make a scheduled payment, the field officer assembles all of the group members together and collects the installment amount from the other members, as stipulated in the group liability contract. Both individual liability borrowers and joint liability borrower groups in default are not granted future Carpena, Cole, Shapiro, and Zia 447 loans. In addition, over the period that we study, there was no change in wages among the field officers, who continued to receive a fixed monthly sum.5 In summary, borrowing conditions changed in the following ways: (1) bor- rowers were obliged to enter joint liability lending groups rather than borrow- ing independently from Saath; (2) the loan officers collected repayment from the group leader rather than from each individual borrower; (3) monthly in- stallments became fixed rather than varying; and (4) the interest rate increased from 18 percent to 24 percent. Given that multiple dimensions of the contract structure were changing, we discuss the implications of each of these changes on our analysis in Section IV. I I I . D ATA CO L L E C T I O N AND S U M M A RY S TAT I S T I C S In this study, we use data from Saath Microfinance’s administrative software systems. The loan data cover outstanding loans every month from April 2005 through March 2009. Because the change in the type of loan contract occurred in November 2007, the data contain more than two years of monthly data on individual liability loans and more than one year on group liability loans. Data are available electronically from only 2 of the 4 Saath Microfinance branches, Behrampura and Vasna; we focus on these branches. These two branches are the largest and the oldest, accounting for the vast majority of Saath’s clients. The data were maintained for accounting purposes to record cash flowing in and out of each branch. The data are therefore of very high quality. The data do not, however, contain information on the terms of each loan, such as matu- rity dates, installment amounts, and amounts outstanding. These data were re- corded by the loan officers in the client passbooks and administrative ledgers. As a result, we are unable to examine the overdue amounts, prepayments, and similar measures.6 The data on loans cover April 2005 to March 2009, but savings data were only available from January 2008 to March 2009. The savings data include the monthly aggregate deposit and withdrawal amounts for the compulsory savings accounts. 5. After March 2009, field officers received 1 percent of the loan interest they collect. Surveys of field officers indicate that they were not aware of this change in compensation structure before it occurred. Nevertheless, to isolate the focus of our study on contract structure, we exclude months after March 2009 in the analysis. 6. The loan data come from three software systems that Saath Microfinance has used at different points in time. Each of the two branches in our dataset used a separate system until early 2008, when the current system was introduced in both branches. Because client identifiers were not carried over from one software system to another, we had to rely on using client names to track individuals over time. These names were unique because they included first, middle, and last names. In identifying clients across systems, 80 percent of the clients had exact name matches, while 14 percent had to be matched by hand due to name spelling errors. The remaining 6 percent, however, could not be matched to the current software system. It is likely that these clients have withdrawn their membership with Saath Microfinance and therefore have not borrowed under the group liability setting because Saath migrated information from the previous to the current software system only for existing members. 448 THE WORLD BANK ECONOMIC REVIEW As previously described, in our main analysis, we study borrowers who re- ceived both individual and group liability loans to overcome the selection problem. Hence, in our dataset, these clients begin with an individual liability loan and, after November 2007, receive a group liability loan. In Saath’s Behrampura and Vasna locations, we identified a sample of 276 such clients, representing 22 percent of the loan client base in these two branches as of March 2009. Table 1 provides the summary statistics for our sample. Collectively, these clients received a total of 748 loans from Saath, of which 450 were individual liability loans and 298 were group liability loans. The average individual liabil- ity loan amount was approximately INR 10,000 (USD 220), and the group lia- bility loan average was approximately INR18,000 (USD 390). Figure 1 plots the number of group liability loans disbursed over time. As the figure shows, the borrowers in our sample, all of whom received individual liability loans, switched to group liability loans in varying months. Our empirical strategy takes advantage of this staggered timing to compare individual liability loan clients who received group liability loans to future recipients to identify the impact of group liability on loan repayment behavior and savings discipline. I V. E M P I R I C A L S T R A T E G Y AND AN A LY S I S Empirical Strategy To study the effect of contract structure on lending outcomes, we exploit the natural experiment provided by Saath’s change in policy. The presence of an exogenous policy change is important. Without exogenous variation, it would be very difficult to determine whether differences in outcomes were attributable to contract structure or to any number of other unobservable characteristics of either the borrowers or the lending institutions. Indeed, theory predicts that dif- ferent contracts are optimal for different types of borrowers. To overcome the selection problem, we focus our attention on the Saath bor- rowers who received both individual and group liability loans. We exploit the T A B L E 1 . Summary Statistics Individual Liability Group Liability Total No. of Total No. of No. of Ave. Loan Amt No. of Ave. Loan Amt Branch Clients Loans Loans (Rs.) Loans (Rs.) Behrampura 198 512 303 9981.1 209 19081.34 Vasna 78 236 147 9927.211 89 16764.04 Full Sample 276 748 450 9963.014 298 18389.26 Notes: This table reports the summary statistics for the borrowers in our sample. These bor- rowers received both individual liability and group liability loans. Source: Authors’ analysis based on data sources discussed in the text. Carpena, Cole, Shapiro, and Zia 449 F I G U R E 1. Group Liability Loans Disbursement in Sample Notes: This figure plots the number of group liability loans disbursed over time in our sample. All of the borrowers were individual liability borrowers who subsequently received a group liability loan after the policy change. Source: Authors’ analysis based on the data sources discussed in the text. natural phasing-in of group liability in what amounts to a repeated difference-in-difference framework. At any point in time, our “treatment” group thus consists of clients who have fully repaid their individual liability loan and who currently have a group liability loan, whereas our “control” group consists of individual liability loan clients who will eventually convert to a group liability loan. Specifically, we estimate the following equation: yilt ¼ a þ bTil þ gi þ dt þ 1ilt ð1Þ where the subscript i refers to individuals, l refers to loans, and t refers to months. T is a dummy variable equal to 1 if loan l of client i is a group liability loan and 0 if it is an individual liability loan. yilt is a measure of loan repay- ment or savings discipline. The estimate of b then provides the effect of switch- ing an individual who is already borrowing to the group liability loan. We include time effects dt because the conversion to group liability loans was stag- gered across individuals, and the individual fixed effects gi absorb the time- invariant characteristics of each borrower. Limitations In what follows, we note some features of our setting that may limit the gener- alizability of the empirical results. Concurrent Changes in Loan and Savings Products. As discussed in Section II, the shift from individual to group liability lending contracts was 450 THE WORLD BANK ECONOMIC REVIEW accompanied by other changes in the contract features. Specifically, in the group liability setting, the loan officers collected repayment from the group leader rather than from each individual borrower, monthly installments became fixed rather than varying, and the interest rate increased from 18 percent to 24 percent. Furthermore, simultaneous with the change to joint lia- bility loan products, the savings rules shifted because Saath began requiring all of its members to maintain compulsory savings accounts. In our view, an ideal experimental evaluation would include (1) requiring borrowers to enter joint liability lending groups rather than borrowing on their own and (2) loan officers collecting repayment from the group leader rather than from each individual borrower, leaving the installment sizes and the inter- est rate fixed. The group lending contracts offered by the majority of MFIs in India collect repayment either from one person (an assigned leader) or from every borrower in the group simultaneously. Hence, we believe that the change in the mode of payment is a feature of the group liability contract. Although the interest rate change and the change in monthly installments are not typical, our regression coefficients, which estimate the impact of the contract change, are likely to be lower bounds. In theory, the increase in the in- terest rate could have several effects: a price effect might reduce demand, whereas a higher interest rate could increase the repayment burden and induce default. Most evidence suggests that microfinance borrowers are not very price elastic, so we are not overly concerned about the demand effects. The increase in the interest rate should bias us against a finding that joint liability lending reduces default. Furthermore, our empirical analysis only considers individuals who borrowed under both the individual and group liability regimes, thus ac- counting for any self-screening among clients based on the increase in interest rates between the two loan contracts. The change in monthly repayment installments and the mode of payment also warrants further discussion. The repayment schedule for individual liabili- ty loans required fixed principal repayments along with interest. Hence, the nominal size of the monthly payments declined over the cycle of the loan. In contrast, the group liability repayment structure is based on a fixed monthly re- payment throughout the term of the loan. Because our analysis focuses on the shift from individual to joint liability loans for the same person, for the same loan amount, we pick up the effect of a lower payment under individual liabili- ty (because the borrower is at the end of her loan cycle) versus a relatively higher fixed payment under joint liability. This shift should bias us against finding a reduction in default. This bias effect is likely even greater in our case because the average loan size and the corresponding repayment installment size are higher under joint liability. The mode of payment also shifted under the individual and group liability settings. In the former, the field officer visited each individual liability borrower to collect repayment, whereas in the latter, the field officer only visited the group leader, who was in charge of collecting repayment from the other Carpena, Cole, Shapiro, and Zia 451 members. Nevertheless, this change in the mode of payment is a feature of group lending because in the setting that we study, group liability is a lending contract that involves both joint repayment to a group leader and joint liability. Thus, the “group” features, such as repayment to a group leader, may lower default, whereas the higher interest rates may increase default, so the effect that we capture may well be a lower bound. Saath’s savings products also changed during our study period, as discussed in Section II. However, we note first that in our main analysis of loan repay- ment and savings discipline, we focus only on the individuals who converted to group liability loans, exploiting the timing of their switch. This focus allows us to control for any changes that occurred at the microfinance institution level under the two loan contract regimes. Thus, in the context of the MFI-wide change in savings requirements, we are comparing a shift from individual to joint liability for the same person (when we include individual fixed effects) who faces mandatory savings under both liability structures. Because our em- pirical strategy rests on the continuous, rolling changeover from individual to joint liability after the announcement, our sample consists of borrowers who are opening and maintaining mandatory savings accounts prior to shifting to a joint liability loan. Although the introduction of a mandatory savings account may have influenced the composition of borrowers, the internal validity of our results remains unaffected because our analysis considers only those individuals who chose to renew their loans with Saath. Theoretically, the imposition of mandatory savings may have two opposing effects: (a) it may discourage borrowers from renewing their loans because the real cost of borrowing has increased through the imposition of a mandatory savings plan; or (b) it may encourage borrowers to renew their loans because individuals appreciate the saving discipline provided by compulsory accounts. This latter point is not trivial. Individuals may fail to save enough because they consistently put off saving for their future (Laibson, 1997), they may be tempted to spend on immediate consumption (Banerjee and Mullainathan, 2010), or they may face intra-household constraints (Ashraf et al., 2010). In a recent field experiment, Atkinson et al. (2010) find that prompting individuals to save at the time of loan repayment doubles the amount of savings. The overall effect of mandatory savings on borrower selection is therefore ambiguous. Unfortunately, we lack any household-level data that would allow us to empirically differentiate these effects. In addition, we only have basic socioeconomic data from Saath records for the clients who eventually joined the joint liability groups. Nevertheless, we run a simple regression of renewal on the percentage payments missed and find a strong statistically significant negative coefficient. Clearly, there is screening based on past loan performance. However, we also find that this screening occurs even for previous individual-to-individual renewals and therefore cannot be considered an effect of either joint liability or mandatory savings. 452 THE WORLD BANK ECONOMIC REVIEW EXTERNAL VALIDITY. Our study sample consists of individual borrowers who have repaid their individual loans and choose to borrow under joint liability. We therefore estimate the effect of joint liability on improving repayment rates among those who choose to borrow under joint liability. Although restricting our analysis to this sample may compromise external validity, we believe that this sample is highly relevant; measuring the effect on those who decline to borrow under joint liability would have little relevance for the outside world. Of course, the sample is also selected based on individuals who chose to borrow in an individual liability setting. However, given that most theory sug- gests that joint liability leads to stricter screening and stricter monitoring, this additional screen may not be that restrictive. The setting that we study may be anomalous in that the typical transition in the microfinance industry is presently the reverse, but we believe that careful study of the effects of shifts in liability in either direction is informative and valuable. Indeed, the recent collapse of Banco del Exito (BANEX), one of the largest microlenders in Nicaragua, highlights the importance of examining the relative merits of group liability and individual liability contracts. Furthermore, the main question that we ask—whether conditional on borrowing, liability rules impact repayment performance—has important policy and theoretical im- plications because we consider how joint liability lending may improve upon individual liability lending. It is difficult to imagine how any single study could capture both the compositional effects and the effect of contract structure on those who have borrowed in individual and joint liability settings. By focusing on clients who borrowed under both individual and joint liabili- ty, our study further highlights the importance of examining the impact of lending contracts on financial inclusion. In the setting that we study, almost 80 percent of the borrowers who completed their individual liability loan and could have borrowed under joint liability did not do so. Although this high percentage indicates that it is possible for the shift in joint liability to have neg- ative consequences on financial inclusion, we do not have data on why people chose to borrow or not to borrowwith Saath. Many new clients (who had not previously borrowed under individual liability) joined Saath after joint liability and may have chosen to stop borrowing for a number of reasons (e.g., no further project/investment needs, shifting business or employment status). Without further data, we cannot determine the impact of joint liability on fi- nancial inclusion. Finally, although we find strong evidence that joint liability improves upon individual liability lending in terms of repayment behavior, we cannot be certain that the treatment effect is similar for other MFIs. The MFI that we study, Saath, has operations that are fairly typical of small MFIs around the world. However,we evaluate a particular joint liability lending program in which borrower groups have a group leader who collects repayment for the group as opposed to public repayment (e.g., at the village center), a collection Carpena, Cole, Shapiro, and Zia 453 method that may bemore common among other MFIs. In this sense, the exter- nal validity of our findings may be limited. DATA LIMITATIONS. Because the data that are available were used primarily for accounting purposes, the dataset does not contain information on the terms of each loan, such as maturity dates, installment amounts, and outstanding loan amounts, all of which were recorded by loan officers in paper ledgers. Thus, our analysis is limited to observing whether a client made a loan repayment or a compulsory savings deposit for a particular month, and we are unable to con- sider outcomes on overdue loan amounts, prepayments, and other measures. Furthermore, aside from data on gender and the client’s neighborhood, the data do not contain information on other demographic or household character- istics of the clients. Because of data limitations, our ability to understand why borrowing amounts increase is limited. Under individual liability, borrowers were limited to loan amounts that were a fixed proportion of their savings and their guaran- tors’ savings. These restrictions were removed under joint liability and replaced with a strict appraisal process for group members. Ideally, if we had deposit data under individual liability, we could determine whether these borrowing constraints were binding under individual liability. Unfortunately, Saath did not keep good historical records of these data; hence, it is not possible to statis- tically distinguish demand and supply effects on loan size. Effect of Lending Structure on Loan Repayment We now turn to the critical question of loan repayment. We note that the joint liability structure will, in theory, induce not only better screening but also greater monitoring efforts. Our empirical design does not distinguish between the two potential causes of improved repayment but rather estimates the com- bined causal effect. Table 2 presents the OLS estimates of Equation 1, where the outcome of in- terest is a dummy variable for a missed payment. This dummy variable indi- cates whether the client failed to make a repayment for a particular month. Saath Microfinance clients are required to make monthly repayments until the principal balance is paid in full, beginning 30 days after disbursement. Hence, the dependent variable takes the value of 1 for a particular month if the total amount repaid by the borrower for that month is nil and 0 otherwise. We use this dummy variable as our first measure of monthly loan repayment discipline. In our sample, almost 20 percent of the individual liability but only 0.1 percent of the group liability monthly loan repayments were not made. Our main results are presented in Table 2. Column (1) presents the regression results with no fixed effects, controlling only for which branch the borrower uses. Taken at face value, group lending reduces the probability of missing a payment by 17.5 percentage points. In columns (2)-(4), we add individual fixed effects, calendar month fixed effects, and both sets of fixed effects, respectively. 454 T A B L E 2 . Dependent Variable: Missed Payment Dummy THE WORLD BANK ECONOMIC REVIEW (1) (2) (3) (4) (5) (6) Group Liability Loan Dummy 2 0.175*** 2 0.163*** 2 0.152*** 2 0.112*** 2 0.159*** 2 0.081** (0.014) (0.012) (0.042) (0.026) (0.015) (0.033) Behrampura Branch 0.067*** 0.072*** 0.067*** (0.016) (0.015) (0.016) Number of Previous Loans 2 0.008 0.045*** (0.007) (0.017) Loan Age in Months 2 0.028*** 2 0.014** (0.007) (0.007) Loan Age Squared 0.003*** 0.002*** (0.001) (0.001) Loan Age Cubed 2 0.000*** 2 0.000*** (0.000) (0.000) Constant 0.128*** 0.091*** 0.217** 0.181*** 0.167*** 0.272*** (0.013) (0.004) (0.087) (0.086) (0.020) (0.086) Control for Calendar Month No No Yes Yes No Yes Individual FEs No Yes No Yes No Yes R-squared 0.082 0.229 0.098 0.246 0.095 0.253 N 6055 6055 6055 6055 6055 6055 Mean of Dep Var 0.105 0.105 0.105 0.105 0.105 0.105 *p , 0.10, **p , 0.05, ***p , 0.01. Notes: This table reports the results from the OLS regressions using panel data from April 2005 to March 2009, where the dependent variable is a dummy for missing a payment. This dummy takes the value of 1 for a particular month if the total amount repaid by the borrower for that month is nil and 0 otherwise. The observations are at the loan-month level. The sample includes the loans of clients who received both an individual and a group lia- bility loan. The variable Group Liability Loan Dummy is a dummy variable equal to 1 if the loan is a group liability loan and 0 if it is an individual lia- bility loan. Behrampura Branch is a dummy for one of the two MFI branches in the sample. Number of Previous Loans is the number of loans that the client received before the current loan. Loan Age in Months is the difference between the month of the observation and the month when the loan was dis- bursed, plus 1. Standard errors, clustered at the individual level, are given in parentheses beneath each point estimate. Source: Authors’ analysis based on data sources discussed in the text. Carpena, Cole, Shapiro, and Zia 455 Finally, in columns (5) and (6), we add controls for the age of the loan: repay- ments may be higher early in the cycle, when borrowers are flush with cash, or later in the cycle, when borrowers seek to repay a loan to obtain a new one. The coefficient drops, although only coefficients (1) and (6) have 95 percent confidence intervals that (barely) do not overlap. Because only overdue loans last over 12 months, the loan age coefficients may “soak up” some of the treat- ment effect, particularly when individual and month fixed effects are present.7 Our preferred point estimate is column (4), which indicates that group lending reduces the probability of a missed payment by 11.2 percentage points. This effect is large and meaningful and may have significant implications for the profitability of a lender. INTERNAL VALIDITY. We conduct a number of robustness tests. First, we report a direct “falsification” test of our analysis by using data from our clients’ previ- ous individual-to-individual loan renewals. Specifically, we focus on the clients in our sample who had at least two individual liability loans. The sample is reduced because many clients did not have multiple loans in the past. Among these clients, we study whether the borrowing experience with the microlender is related to loan repayment; in other words, does having a second loan lead borrowers to repay more reliably? Table 5 presents these results and shows no significant effect on missed payments.8 A second concern regarding the internal validity of our analysis is the fact that joint liability contracts are a completely new contractual arrangement. In particular, because the arrangement is new, clients may be in a “honeymoon” period.9 During this period, clients may be on their best repayment behavior while they are learning the rules of the game. As they gain a betterunderstanding of the consequences of missing a payment, they may begin to behave more stra- tegically. To test for this “honeymoon” effect, we compare the default rates of new clients (i.e., first-time Saath borrowers) under the individual and joint liabil- ity regimes over our sample period. Specifically, we find that first-time Saath bor- rowers make late payments 49 percent of the time in the individual liability setting and 2 percent of the time under joint liability. Because we are comparing clients who are borrowing for the first time in either setting, this result suggests that the “honeymoon” effect does not drive clients’ repayment behavior. 7. A simple way to address the relationship between loan age and repayment status is to to restrict the sample to the first twelve months of repayment data: doing so with the same specifications as reported in Table 2 yields point estimates ranging from 2 .074 to 2 .153, statistically indistinguishable from each other, but all statistically different from 0 at the one percent level. 8. The specific date (November 2007) for the falsification test was not chosen arbitrarily (rather, it was precisely one year prior to the actual change in date), but we have conducted the analysis for all months at least one year before the policy change and find that our effect is dramatically larger than at any other date. Specifically, we re-ran our specification with each of the previous twelve months as our placebo date and found only two cases to be significant, but of much lower magnitude. 9. We thank an anonymous referee for raising this point. 456 THE WORLD BANK ECONOMIC REVIEW A third concern is that a client’s propensity to repay may be correlated with the time in the loan cycle. Specifically, clients may be more likely to make re- payments toward the end of their last individual liability cycle to ensure their eligibility for a group liability loan in the future. We note that this situation would bias estimates against finding that group liability improves borrower per- formance. Nevertheless, we investigate this possibility using an event-time re- gression with the dependent variable for missed payment as previously described, where the event is the conversion from an individual liability loan to a group liability loan. Figure 2 plots the coefficients for each event-time dummy. The first month of repayment in the group liability setting is time ¼ 0, the final repayment month for the individual liability loan is time ¼ 2 1, the second-to-last individ- ual liability loan repayment month is time ¼ 2 2, and so on. Thus, the figure describes loan repayment behavior under the individual liability contract before switching to group liability. Saath requires its borrowers to pay their current loan in full before they are given their next loan. Therefore, by defini- tion, all clients in our sample made a repayment at time ¼ 2 1. Examining the periods in which time 7 2 2 reveals no pattern to support the idea that clients strategically repaid their individual liability loans so that they could borrow under the group liability setting. Alternatively, clients may be more likely to make repayments early in the loan cycle because they may be flush with cash from a recent loan disbursal. We examine this possibility, again using an event-time regression, as shown in Figure 3. We estimate how repayment rates change around loan renewal times when a client pays off an individual liability loan and renews another individu- al liability loan (blue line) as well as cases in which a client pays off a group li- ability loan and renews a second group liability loan (red line). Note that the first month of repayment in the second loan cycle is time ¼ 0, and the final re- payment in the first loan cycle is time ¼ 2 1. Similar to Figure 2, at time ¼ 2 1, all of the clients made a repayment by definition, so the missed payment dummy must mechanically equal zero. Figure 4 shows that the prior missed payments are uncorrelated with the number of months since loan origination. Our study sample consists of individual liability clients who chose to renew their borrowing under the group liability setting. These clients may be better at repayment than borrowers who did not want to enter into a group liability loan contract. However, our analysis focuses exclusively on those who renew and includes individual fixed effects. Hence, an interpretation of our results is that even “good” clients exhibited higher repayment discipline under the group liability setting in comparison with the individual liability setting. However, we acknowledge that by focusing only on those clients who borrowed under both types of contracts, we limit the external validity of our results. The outcome that we have considered thus far, whether a client missed a loan installment for a particular month, is a rough measure because repay- ments may be partial. That is, a client may have repaid an amount greater than Carpena, Cole, Shapiro, and Zia 457 F I G U R E 2. Event Time Regression: Missed Payment on Event Time Dummies of Switching from Individual to Group Liability Loan Notes: This figure plots coefficients for the event-time dummies where the event is the conversion from an individual liability to a group liability loan. The dependent variable is a dummy for missing a monthly repayment, which takes the value 1 for a particular month if the total amount repaid by the borrower for that month is nil and 0 otherwise. The first month of repayment in the group liability setting is time ¼ 0; the final repayment month in the individual liability setting is time ¼ 2 1. The dashed lines indicate the 95 percent confidence interval. Source: Authors’ analysis based on the data sources discussed in the text. zero but less than the required installment amount. Thus, another measure of repayment discipline is the standard deviation of the principal amount repaid for individual liability loans and the total amount repaid for group liability loans. As described in Section 2, only the principal installment amount is fixed in the individual liability setting, whereas in group liability, the required total installment amount ( principal plus interest) is equal every month. If the 458 THE WORLD BANK ECONOMIC REVIEW F I G U R E 3. Event Time Regression: Missed Payment on the Event Time Dummies for Switching from the First to the Second Loan Cycle Notes: This figure plots the coefficients for the event-time dummies where the event is the shift from the client’s first loan cycle to the second loan cycle for each of the individual and group liability loans. The dependent variable is a dummy for missing a monthly repayment, which takes the value 1 for a particular month if the total amount repaid by the borrower for that month is nil and 0 otherwise. The first month of repayment in the second loan cycle is time ¼ 0, and the final repayment month in the first loan cycle is time ¼ 2 1. The dashed lines indicate the 95 percent confidence interval. Source: Authors’ analysis based on the data sources discussed in the text. required amount is repaid each month, then the standard deviation would be zero. However, if there are many months in which people pay less or more than the required amount, then the standard deviation would be higher. Table 6 provides OLS estimates in which the dependent variable is the standard deviation of repayment. Again, our estimates show that there is greater loan Carpena, Cole, Shapiro, and Zia 459 F I G U R E 4. Calendar Month of Loan Origination and Missed Payments Notes: This figure plots the percentage of monthly repayments that were missed using the first three repayments from the loan disbursement of the client’s most recent individual liability loan. The calendar month of loan origination refers to the calendar month when the loan was disbursed. The size of the bubbles represents frequencies. The red line represents the best-fit line. The sample includes clients who received both an individual and a group liability loan. Source: Authors’ analysis based on the data sources discussed in the text. repayment discipline in the group liability setting relative to individual liability, although this effect is not statistically significant. Finally, we note that the average loan size increased from 10,000 INR under individual liability to 18,000 INR under joint liability (see Table 1). This in- crease is large and warrants further discussion. In some ways, the increase can be thought of as an effect of the group lending model because Saath was willing to extend larger loans to borrowers on the strength of social collateral. Indeed, Saath management told us that they were willing to give larger loans precisely because of the joint liability framework. However, one may reasonably wonder whether the increase in loan size itself affects repayment rates. For example, if borrowers invest in convex pro- duction technologies, higher credit limits could increase repayment. We test for this possibility in two ways. First, we split the sample into four quartiles based on the percentage increase in the credit limit from which a borrower benefited when she or he converted to joint liability lending. We do not observe any sys- tematic variation in the treatment effect estimate along this dimension (results not reported). Second, because our dataset includes the reported purpose of each group liability loan, we can conduct separate analyses for loans taken for the purpose of consumption, productive activities, and asset creation. Again, we find no evidence that the treatment effect varies across these three categories (results not reported). 460 THE WORLD BANK ECONOMIC REVIEW DISCUSSION. Given these results, a natural question that arises is why group lia- bility outperforms individual liability for clients who are already borrowing. Although the guarantors requirement in individual liability contracts provided incentives for guarantors to monitor loans and enforce repayment, in practice, these incentives were quite weak; the microlender rarely seized the savings of the guarantors of defaulting clients and did not strictly enforce that the guaran- tors maintain their savings account balance. The microlender collected repay- ment from the guarantors only if all other options (e.g., seizing the defaulting borrower’s savings, revisiting the defaulting borrower, threatening to charge penalties, rescheduling the loan, and having the branch manager intervene) had been exhausted. In contrast, the group liability structure strengthens coopera- tion and trust among the group members, as indicated by the fact that almost all of the joint liability borrower groups designate a group leader, even though it is not required. It is possible that having a group leader increases incentives for monitoring and enforcing repayment. For example, having a group leader may create a person of authority who can enforce repayment or impose sanc- tions in the event of default. We also note that the borrowing requirements for individual liability loans suggest that the effects that we find would likely be smaller in magnitude if these requirements were not in place. Because individual liability loan borrow- ers must have two guarantors and because Saath may use the guarantors’ savings in the event of default, it is possible that these requirements encourage the screening of potential individual liability loan borrowers. We are not able to control for such effects. Savings Discipline We now turn our attention to the compulsory savings deposits required by Saath. As discussed in Section I, we might expect savings to be higher in the joint liability setting because participating in a borrower group may allow indi- viduals to avoid financial demands from their families or to overcome self- control problems. Furthermore, compulsory savings are required to continue as a member in good standing with Saath; those who do not meet the compulsory savings requirements are not allowed to borrow until these requirements have been met. Thus, the same form of peer pressure that applies to loan repayment may apply to compulsory savings. Although Saath initiated both a shift to group liability lending and compul- sory deposits in November 2007, we can separately identify the effect of group liability on compulsory savings by exploiting the time-series variation in loan renewals. Specifically, although the compulsory savings were mandated across the board for all borrowers after November 2007, the shift from individual to group liability was staggered depending on when each individual loan term expired. As explained previously, these renewals were distributed relatively uni- formly throughout the year, resulting in variations in loan contracts at a time when compulsory savings were uniform. Hence, we can study adherence to Carpena, Cole, Shapiro, and Zia 461 compulsory savings for the same person who borrowed under an individual lia- bility contract after November 2007 and who eventually converted to a group liability contract.10 Table 3 presents the OLS estimates in which our dependent variable is a dummy for missing a compulsory savings deposit. The dependent variable takes the value of 1 if the client deposited less then INR 100 and 0 otherwise. The point estimate in column (4), which includes month and individual fixed effects, indicates that the same borrower is 20.5 percentage points less likely to miss a compulsory deposit in a group lending arrangement than when borrow- ing individually. This finding suggests that one possible mechanism through which group lia- bility reduces loan delinquency may be increased savings; a greater savings balance may provide a buffer for borrowers hit with liquidity shocks. Heterogeneous Effects We test for heterogeneous effects along two dimensions. As before, our depen- dent variable is a dummy for whether a client missed a loan repayment for a particular month. Columns (1)-(3) indicate a significantly larger impact of group liability in reducing missed payments among men, although the effect for women remains negative and significant. It is important to note, however, that the control group means for females are also significantly lower, with missed payment rates at 16 percent for females and 23 percent for males.11 The results suggest that group liability effectively neutralizes this gender differ- ential in missed payments. In columns (4)-(6), we examine whether group lending improves repayment behavior more for clients who initially exhibited poor repayment discipline under individual liability. We define ‘borrower quality’ as the percentage of missed payments in the client’s first individual liability loan in the data, and we divide the sample in two along this measure. Note that because our definition of borrower quality makes use of a client’s repayment behavior in her first indi- vidual loan, the regression columns (4)-(6) are restricted to the subsample of clients who had at least two individual loans and include repayment data only from the client’s second individual loan onwards. In terms of past missed payments, we find that group liability has a larger impact on borrowers who, at the outset, were of poor quality. Clients who missed 10 percent of their first individual loan monthly payments were 2.4 percent less likely to miss repayments under the group liability regime. Hence, similar to the gender results, the introduction of group liability is effective in reducing missed payments among those with inconsistent payment records. 10. because we only have savings data from January 2008 onwards, we cannot study the effect of compulsory savings under the individual liability setting, as we have no pre-period data (i.e. savings data pre-November 2007). 11. A difference in means test is significant at the 1 percent level. 462 T A B L E 3 . Dependent Variable: Missed Compulsory Deposit THE WORLD BANK ECONOMIC REVIEW (1) (2) (3) (4) (5) (6) Group Liability Loan Dummy 2 0.234*** 2 0.338*** 2 0.074* 2 0.205** 2 0.230*** 2 0.198*** (0.035) (0.032) (0.041) (0.045) (0.038) (0.069) Behrampura Branch 0.003 2 0.008 2 0.000 (0.036) (0.034) (0.035) Number of Previous Loans 2 0.001 0.124 (0.017) (0.098) Loan Age in Months 2 0.040** 0.050*** (0.019) (0.019) Loan Age Squared 0.004 2 0.006** (0.003) (0.003) Loan Age Cubed 2 0.000 0.000** (0.000) (0.000) Constant 0.389*** 0.698*** 0.607*** 0.807*** 0.461*** 0.623*** (0.039) (0.027) (0.045) (0.040) (0.049) (0.123) Control for Calendar Month No No Yes Yes No Yes Individual FEs No Yes No Yes No Yes R-squared 0.054 0.489 0.103 0.511 0.059 0.514 N 2929 2929 2929 2929 2929 2929 Mean of Dep Var 0.205 0.205 0.205 0.205 0.205 0.205 *p , 0.10, **p , 0.05, ***p , 0.01. Notes: This table reports the results from the OLS regressions with panel data from January 2008 to March 2009, where the dependent variable is a dummy for missing a compulsory deposit. This dummy takes the value of 1 for a particular month if the client deposited less than the required amount of Rs. 100 and 0 otherwise. The observations are at the individual-month level. The sample includes clients who received both an individual and a group lia- bility loan. The variable Group Liability Loan Dummy is a dummy variable equal to 1 if the client had a group liability loan and 0 if it is an individual li- ability loan. Behrampura Branch is a dummy for one of the two MFI branches in the sample. Number of Previous Loans is the number of loans that the client received before the current loan. Loan Age in Months is the difference between the month of the observation and the month when the loan was dis- bursed, plus 1. Standard errors, clustered at the individual level, are given in parentheses beneath each point estimate. Source: Authors’ analysis based on data sources discussed in the text. T A B L E 4 . Heterogeneous Effects by Gender and Borrower Quality (1) (2) (3) (4) (5) (6) Group Liability Loan Dummy 2 0.224*** 2 0.212*** 2 0.132*** 2 0.137*** 2 0.122*** 2 0.075 (0.034) (0.032) (0.042) (0.028) (0.025) (0.048) Female 2 0.041 (0.039) Group Liability Loan* 0.063* 0.065* 0.070** Female (0.037) (0.035) (0.032) Borrower Quality 0.223** (0.111) Group Liability Loan* 2 0.234* 2 0.240* 2 0.248** Borrower Quality (0.119) (0.123) (0.121) Behrampura Branch 0.061*** 0.079*** (0.017) (0.024) Number of Previous Loans 0.045*** 0.025 (0.017) (0.023) Carpena, Cole, Shapiro, and Zia Loan Age in Months 2 0.014* 0.012 (0.007) (0.008) Loan Age Squared 0.002*** 2 0.001 (0.001) (0.001) Loan Age Cubed 2 0.000*** 0.000 (0.000) (0.000) Constant 0.165*** 0.108*** 0.296*** 0.088*** 0.075*** 0.147 (0.038) (0.011) (0.087) (0.022) (0.010) (0.144) Individual FEs No Yes Yes No Yes Yes Month FEs No No Yes No No Yes R-squared 0.084 0.231 0.255 0.095 0.239 0.268 N 6055 6055 6055 2579 2579 2579 463 (Continued ) 464 THE WORLD BANK ECONOMIC REVIEW TABLE 4. Continued (1) (2) (3) (4) (5) (6) Mean of Dep Var in Control (Males) 0.226 0.226 0.226 Mean of Dep Var in Control (Females) 0.157 0.157 0.157 *p , 0.10, ** p , 0.05, ***p , 0.01. Notes: This table reports the results from the OLS regressions where the dependent variable is a dummy for missing a payment. This dummy takes the value of 1 for a particular month if the total amount repaid by the borrower for that month is nil and 0 otherwise. The regressions in columns 1 to 3 provide the heterogeneous effects by gender and include the full sample of clients. The regressions in columns 4 to 6 provide the heterogeneous effects by borrower quality, defined by the percentage of missed payments in the client’s first individual liability loan in the data. The regressions in columns 4 to 6 include only the subsample of clients who had at least two individual liability loans and only loans from the 2nd loan onwards. All of the observations are at the loan-month level. The variable Group Liability Loan Dummy is a dummy variable equal to 1 if the loan is a group liability loan and 0 if it is an individual liability loan. Borrower Quality is the percentage of missed payments in the client’s first individual liability loan in the data. Behrampura Branch is a dummy for one of the two MFI branches in the sample. Number of Previous Loans is the number of loans that the client received before the current loan. Loan Age in Months is the difference between the month of the observation and the month when the loan was disbursed, plus 1. Standard errors, clustered at the individual level, are given in parentheses beneath each point estimate. Source: Authors’ analysis based on data sources discussed in the text. TA B L E 5 . Falsification Test (1) (2) (3) (4) (5) (6) False Treatment 2 0.004 2 0.011 0.049 0.002 0.021 0.058 (0.023) (0.026) (0.041) (0.025) (0.025) (0.050) Behrampura Branch 0.076*** 0.089*** 0.066*** (0.027) (0.027) (0.029) Number of Previous Loans 2 0.028** 0.046 (0.014) (0.035) Loan Age in Months 2 0.025 2 0.014 (0.019) (0.022) Loan Age Squared 0.003 0.004 (0.003) (0.003) Loan Age Cubed 2 0.000 2 0.000 (0.000) (0.000) Constant 0.095*** 0.058*** 0.222 0.139*** 0.158*** 0.319** (0.025) (0.013) (0.142) (0.020) (0.043) (0.129) Control for Calendar Month No No Yes No No Yes Individual FEs No Yes No No No Yes Carpena, Cole, Shapiro, and Zia R-squared 0.011 0.140 0.042 0.000 0.016 0.175 N 1584 1584 1584 1584 1584 1584 Mean of Dep Var 0.140 0.140 0.140 0.140 0.140 0.140 *p , 0.10, **p , 0.05, ***p , 0.01. Notes: This table reports the results from a falsification test, where the dependent variable is a dummy for missed payment. This dummy variable takes on the value of 1 for a particular month if the total amount repaid by the borrower for that month is nil and 0 otherwise. The regressions corre- spond with those in Table 2. The sample contains the observations of the individual liability loans among clients who received both an individual and a group liability loan and who obtained at least 2 individual liability loans, the most recent of which was disbursed after November 2006. Observations are at the loan-month level. False Treatment is a dummy variable equal to 1 if the loan was disbursed after November 2006 and 0 otherwise. Behrampura Branch is a dummy for one of the MFI branches in the sample. Number of Previous Loans is the number of loans that the client received before the current loan. Loan Age in Months is the difference between the month of the observation and the month when the loan was disbursed, plus 1. Standard errors, clustered at the individual level, are given in parentheses beneath each point estimate. Source: Authors’ analysis based on data sources discussed in the text. 465 466 THE WORLD BANK ECONOMIC REVIEW T A B L E 6 . Dependent Variable: Standard Deviation of Repayment (1) (2) (3) Group Liability Loan Dummy 2 288.666 2 306.367* 2 220.794 (185.204) (185.391) (251.901) Behrampura Branch 839.101*** (149.442) Constant 1352.427*** 798.894*** 281.277*** (100.308) (120.412) (83.967) Individual FEs No No Yes R-squared 0.005 0.041 0.472 N 689 689 689 Mean of Dep Var 1243.496 1243.496 1243.496 *p , 0.10, **p , 0.05, ***p , 0.01. Notes: This table reports results from the OLS regressions using cross-sectional data, where the dependent variable is the standard deviation of monthly repayment. If the required amount is repaid every month, then the standard deviation would be zero. If there are many months where people pay less or more than the required amount, then the standard deviation would be higher. The observations are at the loan level. The sample includes the loans of clients who received both an individual and a group liability loan. The variable Group Liability Loan Dummy is a dummy variable equal to 1 if the loan is a group liability loan and 0 if it is an individual liability loan. Behrampura Branch is a dummy for one of the two MFI branches in the sample. Standard errors, clustered at the individual level, are given in parentheses beneath each point estimate. Source: Authors’ analysis based on data sources discussed in the text. V. D I S C U S S I O N AND CONCLUSION Microfinance has reached over 150 million borrowers worldwide and is growing at a cumulative 40 percent average growth rate. Recent initial public offerings (IPOs), which valued the Mexican microfinance institution Compartamos at $2 billion and SKS in India at $1.5 billion, have attracted the attention of the global financial markets. However, there have also been spec- tacular failures, such as the collapse of Banco del Exito (BANEX), which was recently the largest micro, small, and medium enterprise lender in Nicaragua, with a $125 million dollar loan portfolio. Suffering from a 45 percent delin- quency rate, BANEX was ordered into liquidation.12 As many microlenders around the world weaken their group liability ap- proach and shift toward individual lending, understanding the role of group li- ability in enhancing performance has become critical in microfinance programs. However, the empirical literature provides little guidance for policy makers and microfinance practitioners because few empirical studies have com- pared group liability contracts with other lending strategies. In this paper, we exploit an exogenous change in the liability structure in an Indian microfinance program, in which the program shifted from an individual liability structure to a group liability structure. We find evidence that for the 12. See: http://financialaccess.org/node/3547 Carpena, Cole, Shapiro, and Zia 467 same borrower, the shift to group liability reduces default rates and improves savings discipline. Under the group liability setting, the required monthly loan installments are 11 percent less likely to be missed, and compulsory savings de- posits are approximately 20 percent less likely to be missed relative to individu- al liability. Thus, our findings indicate that group lending outperforms individual lending in loan repayment and savings discipline. We see our study as an important piece of evidence rather than a definitive answer to the question of the optimal lending structure for microfinance. The microlender that we study, Saath, has operations that are fairly typical of MFIs that lend in urban areas. Moreover, the management, infrastructure, and stated goals of Saath are not markedly different from other lenders throughout India or in other low-income settings. Saath’s most remarkable characteristic is prob- ably its small size; the microfinance industry includes an important right tail of very large lenders. However, small institutions such as Saath (an NGO with $410,000 dollars in its total loan portfolio as of 2009) make up a non-trivial portion of the industry. Second, as with any natural experiment, we caution that there are limita- tions to our study. Our sample consists of only those who elected to continue from the individual to the group lending model and thus may not be represen- tative of the entire population that would be effected by changes in lending models. We are unable to clearly identify the mechanisms through which group lending improves repayment, and our data are not sufficient to allow us to pre- cisely calculate the effect of the lending structure change on lender profitability. An ideal experiment to answer these questions might have randomly assigned individuals to a range of different lending models, such as group lending with self-selected group members, group lending with randomly assigned group members, and group lending without group liability. Nevertheless, it may be useful to discuss our view of the mechanisms at work based on our reading of the evidence and on numerous conversations with Saath clients, staff, and management. We believe that peer pressure is im- portant. The Saath operations manual states, “The concept of peer pressure must be executed properly in favor of the organization and concept of micro finance,” arguing that the group leader should take responsibility for ensuring that the members repay and stressing the importance of joint liability. In the case of repeated missed payments, the manual instructs staff to ensure that “the members take the responsibility of closing the loan amount of” a delin- quent borrower. We also interviewed 10 Saath field officers regarding their views of changes to the loan collection process with group lending. All ten field officers mentioned that collection under joint liability was easier because clients were more disciplined about paying on time, and several others men- tioned the importance of “collective responsibility” within the group. These qualitative reports are consistent with a recent paper on microfinance in Andhra Pradesh, which finds that peer pressure is an important determinant of loan repayment (see Breza (2012)). 468 THE WORLD BANK ECONOMIC REVIEW Taken as a whole, we believe that our results provide a cautionary tale for policy-makers and microfinance institutions that are eager to convert from group to individual lending models. Although most microfinance organizations around the world have reported repayment rates that are impressively high, the industry has also witnessed both idiosyncratic failure and widespread collapse, such as the recent crisis in Andhra Pradesh. Our results highlight the impor- tance of the group lending structure in facilitating the sustainable provision of credit to the poor. Our findings suggest a need for further research to identify the precise mechanisms of how joint liability affects borrower behavior, the role of savings as a mechanism for improving repayment, and the relative merits of peer screening versus peer monitoring under group liability contracts. Finally, it would be useful to understand whether the group lending structure itself could be improved upon, perhaps by introducing formal group or sub- group level repayment incentives or insurance mechanisms. REFERENCES Ahlin, C., and R. M. Townsend. 2007. “Using Repayment Data to Test Across Models of Joint Liability Lending.” The Economic Journal 117 (517): F11–F51. Ambec, S., and N. Treich. 2007. “Roscas as Financial Agreements to Cope with Self-Control Problems.” Journal of Development Economics 82 (1): 120 –37. ´ riz de Aghion, B. 1999. “On the Design of a Credit Agreement with Peer Monitoring.” Armenda Journal of Development Economics 60 (1): 79– 104. ´ riz de Aghion, B., and J. Morduch. 2010. The Economics of Microfinance. Cambridge, MA: Armenda MIT Press. Ashraf, N., D. Karlan, and W. Yin. 2010. “Female Empowerment: Further Evidence from a Commitment Savings Product in the Philippines.” World Development 38 (3): 333– 44. Atkinson, J., A. de Janvry, C. McIntosh, and E. Sadoulet. 2010. “Prompting Microfinance Borrowers to Save: A Field Experiment from Guatemala.” Economic Development and Cultural Change (forthcoming). Attanasio, O., B. Augsburg, R. D. Haas, E. Fitzimons, and H. Harmgart. 2011. “Group Lending or Individual Lending? Evidence from a Randomized Field Experiment in Mongolia.” Working Paper W11/20. Institute for Financial Studies, London, UK. Banerjee, A., T. Besley, and T. Guinnane. 1994. “Thy Neighbor’s Keeper: The Design of a Credit Cooperative with Theory and a Test.” Quarterly Journal of Economics 109 (2): 491–515. Banerjee, A., E. Duflo, R. Glennerster, and C. Kinnan. 2009. “The Miracle of Microfinance? Evidence from a Randomized Evaluation.” Working Paper. Massachusetts Institute of Technology, Department of Economics, Cambridge, MA. Banerjee, A., and S. Mullainathan. 2010. The Shape of Temptation: Implications for the Economic Lives of the Poor. Working Paper. Massachusetts Institute of Technology, Department of Economics, Cambridge, MA. Banerjee, A. V., and A. F. Newman. 1993. “Occupational Choice and the Process of Development.” Journal of Political Economy 101 (2): 274– 98. ¨c Beck, T., A. Demirgu ¸ -Kunt, and V. Maksimovic. 2005. “Financial and Legal Constraints to Growth: Does Firm Size Matter?” The Journal of Finance 60 (1): 137– 77. Beck, T., A. Demirgu ¸ -Kunt, and M. S. Martinez-Peria. 2007. “Reaching Out: Access to and Use of ¨c Banking Services across Countries.” Journal of Financial Economics 85 (1): 234 –66. Carpena, Cole, Shapiro, and Zia 469 Besley, T., and S. Coate. 1995. “Group Lending, Repayment Incentives and Social Collateral.” Journal of Development Economics 46 (1): 1 – 18. Bouman, F. J. A. 1995. “Rotating and Accumulating Savings and Credit Associations: A Development Perspective.” World Development 23: 371– 84. Breza, E. 2012. “Peer Effects and Loan Repayment: Evidence from the Krishna Default Crisis.” Working Paper. Massachusetts Institute of Technology, Department of Economics, Cambridge, MA. Cetorelli, N., and M. Gambera. 2001. “Banking Market Structure, Financial Dependence and Growth: International Evidence from Industry Data.” The Journal of Finance 56 (2): 617 –48. Chen, G., S. Rasmussen, and X. Reille. 2010. “Growth and Vulnerabilities in Microfinance.” CGAP Focus Note No. 61. Daley-Harris, S. 2009. State of the Microcredit Summit Campaign Report 2009. Washington, DC: Microcredit Summit Campaign. Feigenberg, B., E. Field, and R. Pande. 2010. “Building Social Capital through Microfinance.” Working Paper 16018. National Bureau of Economic Research, Cambridge, MA. Fischer, G. 2010. “Contract Structure, Risk Sharing and Investment Choice.” Working Paper EOPP/ 2011/23. London School of Economics, London, UK. Ghatak, M. 2000. “Screening by the Company You Keep: Joint Liability Lending and the Peer Selection Effect.” Economic Journal 110 (465): 601– 31. ´ , X., and D. Karlan. 2009. “Group versus Individual Liability: Long Term Evidence from Gine Philippines Microcredit Lending Groups.” Policy Research Working Paper 4008. World Bank, Policy Research Department, Washington, DC. Gugerty, M. K. 2007. “You Can’t Save Alone: Commitment in Rotating Savings and Credit Associations in Kenya.” Economic Development and Cultural Change 55 (2): 251– 82. Laibson, D., 1997. “Golden Eggs and Hyperbolic Discounting.” Quarterly Journal of Economics 112, 443–477. Morduch, J. 1999. “The Microfinance Promise.” Journal of Economic Literature XXXVII (4): 1569–1614. Rajan, R., and L. Zingales. 1998. “Financial Dependence and Growth.” American Economic Review 88 (3): 559–86. Sharma, M., and M. Zeller. 1997. “Repayment Performance in Group-Based Credit Programs in Bangladesh: An Empirical Analysis.” World Development 25 (10): 1731–42. Stiglitz, J. E. 1990. “Peer Monitoring and Credit Markets.” World Bank Economic Review 4 (3): 351–66. World Bank. 2008. Finance for All? Policies and Pitfalls in Expanding Access. Washington, DC: World Bank. SYMPOSIUM ON FINANCIAL STRUCTURE AND DEVELOPMENT Financial Structure and Economic Development: A Reassessment’ ¨c Robert Cull, Asli Demirgu ¸ -Kunt, and Justin Yifu Lin While the literature on the effects of financial development is large, relatively few studies have examined whether and how financial structure—the mix of fi- nancial institutions and the services that they offer—matters for economic growth and inequality. Moreover, the literature has been largely silent about whether the relationships between financial structure and firm outcomes ( per- formance and access to finance) change as a country develops. The group of papers published in this special section helps to fill those gaps in the literature. The four that appear here were part of a larger conference on “Financial Structure and Economic Development” that took place at the World Bank on June 16, 2011. Further information on the issue of financial structure in devel- opment can be found in other papers presented at the conference and available in World Bank Working Papers (Lin, Sun, and Jiang, 2009; Beck, Demirgu ¸ -Kunt, and Singer, 2011; Kpodar and Singh, 2011).1 ¨c The first of the papers in this issue by Asli Demirgu ¸ -Kunt, Erik Feyen, and ¨c Ross Levine, uses quantile regressions to assess the relationship between eco- nomic and financial development at each percentile of the distribution of economic development. Thus, the quantile regressions provide information on how the associations between economic development and both bank and secu- rities market development change as countries grow richer. ’ Robert Cull (corresponding author; rcull@worldbank.org) is a Lead Economist in the Development Research Group of the World Bank. Asli Demirgu ¸ -Kunt (ademirguckunt@worldbank. ¨c org) is Director, Development Policy, Development Economics Vice Presidency, and Chief Economist, Finance and Private Sector Development Network at the World Bank. Justin Yifu Lin ( justinlin@ccer. pku.edu.cn) is Professor and Honorary Dean of the National School of Development at Peking University and former Chief Economist and Senior Vice President of the World Bank. 1. Lin, Sun, and Jiang study these issues through the lens of economic theory. Beck, Demirgu ¸ -Kunt, and Singer offer new data and analysis that pushes beyond measures of financial ¨c structure based only on banks and stock markets to include credit unions, building societies, community banks, microfinance institutions, finance companies, and factoring companies. Kpodar and Singh examine the effects of financial structure on poverty and inequality. THE WORLD BANK ECONOMIC REVIEW, VOL. 27, NO. 3, pp. 470– 475 doi:10.1093/wber/lht006 Advance Access Publication February 18, 2013 # The Author 2013. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 470 ¨c Cull, Demirgu ¸ -Kunt, and Lin 471 The main results are that as economies develop (1) both banks and markets become larger relative to the size of the overall economy; (2) the association between an increase in bank development and an increase in economic output becomes smaller; and (3) the association between an increase in securities market development and an increase in economic output becomes larger. The quantile regressions suggest that financial structure changes—becoming more market-oriented—as economies develop. This is consistent with theoreti- cal arguments that economic development increases the demand for the services provided by securities markets relative to services provided by banks (Allen and Gale, 2000; Boyd and Smith, 1998). Hence these findings support theoretical predictions that suggest securities markets become more important for econom- ic activity and that banks become less important as countries develop economi- cally. This finding is also of significant policy relevance. If the mixture of banks and markets should change as indicated as economies develop, then policy and institutional impediments to the evolution of the financial system can have significant costs for economic development. In the next paper, Fenghua Song and Anjan Thakor provide a theoretical analysis to explain how banks and capital markets compete with each other to attract borrowers, but also complement each other and co-evolve over time. In their model, banks are superior to markets in credit analysis because they can more accurately certify borrowers that are worthy of credit. In contrast, capital markets hold the advantage in aggregating information. By providing a trading venue, informed investors’ private information about project payoff enhance- ment opportunities is impounded into security prices, thereby reducing the prob- ability that valuable investment opportunities are passed up. Banks and markets compete with each other to attract borrowers based on their relative advantages. Their model also incorporates two novel features absent from other theories of financial system architecture: securitization and risk-based bank capital. Securitization creates a natural symbiosis by involving the bank in the origina- tion and screening of loans and capital markets in the provision of financing. This, in turn, creates a positive feedback loop from bank development to market development. In addition, improvements in banks’ screening that derive from banking sector evolution increase the confidence that capital market in- vestors have in the quality of securitized borrowers, which stimulates greater informed trading in the capital market and thus capital market evolution. Capital market evolution also affects bank lending through its effects on bank capital. As informed trading through markets increases, the costs of equity capital declines for firms, including banks. By raising additional capital at reduced cost, banks can extend more credit, lending to riskier borrowers who were previously excluded from credit markets. Expansion into riskier market segments provides an incentive for banks to improve the precision of their screening efforts. Thus, capital markets improvements spur banking sector development through bank capital. The analysis can therefore account not only for competition and comple- mentarity between bank and market financing, but also their co-evolution. 472 THE WORLD BANK ECONOMIC REVIEW Their theoretical framework can also account for political intervention in fi- nancial sector development where the political goal is to expand credit avail- ability. By giving politicians a choice in how to intervene in the financial system, the authors endogenize the manner of political intervention at different stages of financial development. There are two methods of intervention: equity capital subsidies in exchange for government ownership of banks (as we see in many emerging economies), and direct-lending regulations that force banks to increase their lending to low-quality borrowers that banks would not extend credit to in the absence of regulation. By contrast, regulations do not subsidize banks, and in fact impose a cost on them. In the early stages of development, capital is very expensive due to less devel- oped financial markets. Banks therefore find the capital subsidy attractive, and are willing to increase size by lending to low-quality borrowers in order to obtain it. Direct-lending regulations do not work well because banks are making low profits in the early development stage, so they will not incur losses by lending to more borrowers (of lower quality) without some form of compensa- tion. At the intermediate development stage, capital becomes sufficiently cheap that the capital subsidy no longer works well as an inducement to expand lending to lower-quality borrowers. At the same time, directed-lending regula- tions also do not work well, because bank profits are still not sufficiently high. Thus, there is no political intervention to promote expanded credit availability. At the advanced stage of financial development, capital becomes even cheaper, so capital subsidies continue to not to be effective in inducing banks to expand lending. What is left is for politicians to directly push/force the banks to expand lending to low-quality borrowers, which imposes costs on the banks. Regulation becomes more effective in the advanced stage because banks are making enough profits to cover these additional costs. Banks will obey the regulations, because failing to do so would result in revocation of the banking license and the loss of large future profit streams. In summary, the model pre- dicts a U-shaped pattern of political intervention: It is highest in the early and advanced stages of financial development, though in different forms, but lowest in the intermediate stage. The main reason for this stage-dependent intervention is that the value of capital subsidies depends on the cost of equity capital for banks, which in turn depends on the level of development of the capital market. While direct empiri- cal tests of these hypotheses are not pursued, the paper provides insights as to how the incentives of financial services providers, borrowers, and politicians shape the structure of the financial sector, and how that changes over time. In the third paper, Augusto de la Torre, Erik Feyen, and Alain Ize use a battery of sixteen indicators of the size, depth, and efficiency of financial sectors to describe the path of development across countries over the past 35 years, and to benchmark that development using regressions that control for each country’s stage of economic development and other arguably exogenous factors (such as population size and density). The authors then view these ¨c Cull, Demirgu ¸ -Kunt, and Lin 473 paths of financial development through the lens of the frictions that hindered fi- nancial contracting. They define two broad categories of frictions. The first set restricts agents’ capacity to establish and enforce bilateral contracts (so-called agency frictions), while the second impedes agents’ capacity to participate and coordinate their financial activities in collectively desirable ways (collective fric- tions). The authors go on to argue that the arc of financial development reflects countries’ efforts to find the path of least resistance around those frictions. Data on financial structure support the notion that the sequencing of finan- cial services broadly conformed with what one would expect based on the gradual grinding down of the frictions along the paths of least resistance. For example, agency frictions associated with the costs of information and contract enforcement meant that bank deposits preceded bank credit, credit to govern- ments developed before credit to private actors, and the full development of credit markets lagged that for bank credit. With regard to collective action fric- tions and network effects, the cross-country patterns show that external funding of the government preceded domestic funding, and wholesale (non- deposit) funding lagged retail deposit funding, but often took off rapidly once minimum thresholds had been reached. Finally, the development of capital markets showed large returns to scale reflecting network effects, and intercon- nectedness within the financial system and globalization exploded as financial systems matured. Countries often deviated from their development paths as predicted by the benchmark regressions and, though the regressions summarize strong central tendencies, lower income countries generally have not retraced the past steps taken by high income countries. The authors therefore speculate that across-the-board innovations (that lifted all countries, regardless of their stage of development) and path dependencies reflecting dynamic interactions between financial and economic development both have factored in the finan- cial development experiences of individual countries. The paper concludes with an empirical analysis of factors that could poten- tially account for some countries’ large deviations from benchmark development paths. They show that deviations were linked to policy-related variables that af- fected the enabling environment for financial contracting such as enforcement costs, creditor and property rights, and the quality of credit information, though contractual rather than informational frictions explain a larger share of the policy-induced development differences across countries. The authors also point out that contractual frictions are likely to be more difficult to resolve because they tend to require reform of local institutions, whereas many informational frictions could conceivably be eased through technological innovations. Not sur- prisingly, deviations from benchmark development paths were also strongly linked to financial crashes. And the associated lags in financial development were long-lived and evident across a large number of indicators. However, because the empirical analysis is based on cross-country regressions, it does not lend itself to identifying the best policies to avert financial crashes. 474 THE WORLD BANK ECONOMIC REVIEW In the final paper of the collection, Robert Cull and Colin Xu use firm-level data from 89 countries to test whether financial structure affects labor growth rates. One of the predictions of New Structural Economics is that, because labor is more abundant than capital in poor countries, labor-intensive indus- tries should characterize the early stages of development. Businesses in labor-intensive industries tend to start out small, and small, local banks are likely to be better positioned than large banks and stock markets to collect the soft information and undertake the sustained monitoring that enables financial institutions to lend to small businesses.2 The authors find that labor growth is in fact swifter in low-income countries that have a higher level of private credit/GDP, consistent with the predictions from new structural economics. There is also evidence from a variety of instrumental variables regressions that the relationship is causal. In high-income countries, labor growth rates are in- creasing in the level of stock market capitalization, also consistent with predic- tions from new structural economics, though the authors are unable to provide evidence that the association is causal. The authors find no evidence that small- scale firms in low-income countries benefit most from private credit market de- velopment. Rather, the labor growth rates of large firms increase more with the level of private credit market development, a finding consistent with the history-based political economy view that banking systems in low-income countries serve the interests of the elite, rather than providing broad-based access to financial services (see Calomiris and Haber, 2011). A limitation of the study is that the measure of banking sector development, private credit/GDP, does not provide information on the size distribution of banks, making it impossible to test fully all of the predictions from new struc- tural economics regarding the suitability of financial structures at different stages of development. Still, the results do suggest that larger firms are captur- ing a disproportionately large share of the credit in poor countries with rela- tively well-developed banks. Though the methods vary, all four papers in this special section provide evi- dence or arguments that financial structure should vary with the stage of eco- nomic development, a point which has been under-emphasized in the literature to date. Implicit or explicit in those analyses is the notion that bank-based structures are likely to be better at promoting growth during the early stages of development, and that they gradually give way to capital markets as economies develop. Another theme from these papers that has been underplayed in the lit- erature is the dynamic interplay between banks and capital markets that leads to their co-evolution. While the papers are successful in describing the develop- ment of financial systems from both theoretical and empirical perspectives, more research is needed to identify the triggers and catalysts that ignite the in- terplay between banks and markets. The paper by de la Torre, Feyen and Ize provides important clues about the types of financial market frictions have 2. See Lin, Sun, and Wu (2012) for a more detailed development of these themes. ¨c Cull, Demirgu ¸ -Kunt, and Lin 475 been easiest to resolve at different stages of economic development and Song and Thakor offer a theory that shows how political interventions shape finan- cial structure at different development stages. But these are only a beginning. Identifying specific policy priorities for different stages of development and ex- plaining why some countries pursue suitable paths while other do not remains a challenge. Our guess is that many of the reasons why countries adopt seem- ingly inappropriate financial sector policies are rooted in political economy, both at the national and international levels, but much more work is needed to establish that conjecture in a meaningful, actionable way. Our overall hope is that these papers provide a fresh perspective on an under-researched, but po- tentially important issue to a wide audience of readers. REFERENCES Allen, F., and D. Gale. 2000. Comparative Financial Systems. Cambridge, MA: MIT Press. Beck, T., A. Demirgu ¸ -Kunt, and D. Singer. 2011. “Is Small Beautiful? Financial Structure, Size, and ¨c Access to Finance.” World Bank Policy Research Working Paper 5806. Washington, DC. Boyd, J.H., and B.D. Smith. 1998. “The Evolution of Debt and Equity Markets in Economic Development.” Economic Theory 12(3): 519–60. Calomiris, C., and S. Haber. 2011. “Fragile Banks, Durable Bargains: Why Banking is All About Politics and Always Has Been.” Stanford University, mimeo, Stanford, CA. Kpodar, K., and R. Singh. 2011. “Does Financial Structure Matter for Poverty? Evidence from Developing Countries.” World Bank Policy Research Working Paper 5915. Washington, DC. Lin, J. Y., X. Sun, and Y. Jiang. 2009. “Toward a Theory of Optimal Financial Structure.” World Bank Policy Research Working Paper 5038. Washington, DC. Lin, J.Y., X. Sun, and H.X. Wu. 2012. “Banking Structure, Labor Intensity, and Industrial Growth: Evidence from China.” World Bank, mimeo, Washington, DC. The Evolving Importance of Banks and Securities Markets ¨c Asli Demirgu ¸ -Kunt, Erik Feyen, and Ross Levine The roles of banks and securities markets evolve during the process of economic development. As countries develop economically, (1) the size of both banks and securities markets increases relative to the size of the economy, (2) the association between an increase in economic output and an increase in bank development becomes smaller, and (3) the association between an increase in economic output and an increase in securities market development becomes larger. These findings are consistent with theories predicting that as economies develop, the services provided by securities markets become more important for economic activity, whereas those provided by banks become less important. JEL codes: O16, G1, G2, O43 Several economic theories stress that banks provide services to the economy that differ from those provided by securities markets, predicting that both the operation of banks and the functioning of securities markets have independent influences on economic development. For example, Acemoglu and Zilibotti (1997), Allen and Gale (1997, 1999), Boot and Thakor (1997, 2000), Dewatripont and Maskin (1995), Holmstrom and Tirole (1993), and Rajan (1992) argue that banks have a comparative advantage in reducing the market frictions associated with financing standardized, shorter-term, lower- risk, well-collateralized endeavors, whereas decentralized markets are relatively more effective in custom designing arrangements to finance more novel, longer- run, higher-risk projects that rely on intangible inputs. Consistent with these Asli Demirgu ¨c¸ -Kunt (ademirguckunt@worldbank.org) is the Director of Development Policy in the World Bank’s Development Economics Vice Presidency (DEC) and Chief Economist of the Financial and Private Sector Network (FPD). Erik Feyen (corresponding author; efeijen@worldbank.org) is a Senior Financial Economist in the Financial Systems Practice of the World Bank. Ross Levine (ross_levine@haas.berkeley.edu) is the Willis H. Booth Professor of Banking and Finance at the Haas School of Business at the University of California, Berkeley. We received very helpful comments from Thorsten Beck; Norman Loayza; Yona Rubinstein; seminar participants at the World Bank’s conference, “Financial Structure and Economic Development,” on June 16, 2011; and three anonymous referees. Mauricio Pinzon Latorre provided outstanding research assistance. A supplemental appendix to this article is available at http://wber.oxfordjournals.org/. THE WORLD BANK ECONOMIC REVIEW, VOL. 27, NO. 3, pp. 476– 490 doi:10.1093/wber/lhs022 Advance Access Publication August 24, 2012 # The Author 2012. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 476 ¨c Demirgu ¸ -Kunt, Feyen, and Levine 477 theories, Demirgu ¨c¸ -Kunt and Maksimovic (1998), Levine and Zervos (1998), and Beck and Levine (2004) provide evidence that better-functioning banks and securities markets exert robust, independent, and positive effects on economic activity. A substantial body of economic theory emphasizes that the comparative importance of banks and markets for economic activity changes during the process of economic development, with markets becoming relatively more important for economic activity. For example, the concepts articulated in Goldsmith (1969), Allen and Gale (1995, 2000), Boot and Thakor (1997, 2000), Boyd and Smith (1998), Weinstein and Yafeh (1998), Morck and Nakamura (1999), and Song and Thakor (forthcoming) suggest that (1) banks and markets provide different, though sometimes complementary, financial ser- vices and that (2) the services provided by markets become comparatively more important for promoting economic activity as countries develop economically. In particular, these theories suggest that as economies develop, a greater number of projects require customized financial arrangements rather than stan- dardized contracts and rely on intangible assets rather than easily collateralized capital inputs. Because these models also suggest that banks have a compara- tive advantage in financing standardized, well-collateralized endeavors, whereas securities markets are better at custom designing arrangements to finance novel projects that rely on intangible inputs, these theories imply that the services provided by securities markets have a greater impact on economic activity as economies grow, whereas those provided by banks become less important. Empirical research has been largely unsuccessful at clarifying the evolving importance of banks and markets during the process of economic development, as exemplified by Beck and Levine (2002), Demirgu ¸ -Kunt and Maksimovic ¨c (2002), and Levine (2002). Demirgu ¸ -Kunt and Levine (2001) show that banks ¨c and securities markets tend to become more developed as economies grow and that securities markets tend to develop more rapidly than banks. Thus, finan- cial systems generally become more market-based during the process of economic development. However, this pattern could simply reflect reverse causality. Economic progress may boost the development of securities markets more than it boosts the development of banks. The observation that financial systems tend to become more market-based as economies develop does not nec- essarily imply that securities markets exert a larger impact on economic activity in more economically advanced economies. In this paper, the changing importance of banks and securities markets with the development of economies is evaluated in two steps. First, using a newly developed database, the evolution of banks and securities markets during the process of economic development is reassessed. That is, as countries develop economically, what happens to the size of banks and securities markets relative to the size of the overall economy? Second, the changes to the associations between economic activity and bank and stock market development as 478 THE WORLD BANK ECONOMIC REVIEW countries develop economically are examined. This investigation involves regressing economic activity on both bank and securities market development and assessing how the estimated coefficients change as countries develop economically. This analysis provides information on how the associations between economic activity and the different components of the financial system evolve during economic development. The primary methodological contribution of this paper is the use of quantile regressions to assess how the associations between economic activity and bank and securities market development evolve as countries grow (Koenker and Basset 1978). Ordinary least squares (OLS) regressions provide information on the association between, for example, economic development and bank devel- opment for the average country, defined as a country at the average level of economic development. However, quantile regressions provide information on the relationship between economic activity and bank development at each per- centile of the distribution of economic development. As emphasized throughout the paper, these quantile analyses do not yield a sharp causal interpretation. Rather, they show how the estimated coefficients of the financial development indicators vary at different levels of economic development. In this way, the analyses illustrate how the associations between economic development and both bank and securities market development change during the process of economic development. New data contribute to these analyses of finance and development with the construction of a database that covers 72 countries over the period from 1980 through 2008. For the analyses, the data are aggregated into five-year averages (data permitting), so there are a maximum of six observations per country. In addition to using standard indicators, such as bank credit to the private sector as a share of gross domestic product (GDP), stock market capitalization rela- tive to GDP, and the value of stock market transactions relative to GDP, the analyses employ data on the capitalization of private domestic bond markets relative to GDP (for country averages, please see table S1 in the supplemental appendix, available at http://wber.oxfordjournals.org/). Both banks and securities markets become larger relative to the size of the overall economy as countries develop economically, confirming the results in Demirgu ¸ -Kunt and Levine (2001). These findings hold across various mea- ¨c sures of bank and securities market development, including measures incorpo- rating private domestic bond markets. It is important to note that the measures of bank and securities market development are scaled by GDP. Thus, the find- ings show that the growth of marketable securities and bank loans outpaces the growth of economic activity as countries develop economically. The analyses also indicate that (1) the association between economic activity and bank development decreases with economic development, but (2) the asso- ciation between economic activity and securities market development increases as countries grow. Put differently, as economies develop, the marginal increase in economic activity associated with an increase in bank development falls, ¨c Demirgu ¸ -Kunt, Feyen, and Levine 479 whereas the marginal increase in economic activity associated with an increase in securities market development rises. Although instrumental variables are not employed to identify a causal effect, these results are consistent with the predic- tions emerging from the large body of theoretical research discussed above: as economies develop economically, the services provided by securities markets will become more important for future economic development, whereas those provided by banks will become less important. This research is policy relevant. First, if the optimal mixture of banks and markets changes as an economy develops, such a relationship is an indication of the costs of policy and institutional impediments to the evolution of the financial system. This is the first paper to show that the association between economic activity and stock market development increases as economies grow, whereas the association between economic activity and bank development decreases. Furthermore, this work suggests that the associations between economic activity and both bank and securities market development change with economic development. This change implies that the estimated elasticities from previous research regarding the impact of changes in bank or stock market development on economic development will yield misleading informa- tion about countries with incomes that are far from the sample average. Previous studies do not account for the evolving importance of banks and markets during the process of economic development. The new data, methods, and analyses contribute to a better understanding of the dynamic relationships among economic development, financial institu- tions, and securities markets, but these contributions come with qualifications and limitations. First, although the analyses are policy relevant, direct policy instruments and levers are not studied. Thus, the results suggest that impedi- ments to the evolution of financial systems will hinder economic activity, but they do not provide guidance on exactly which types of policies foster the healthy development of financial systems. Second, although the analyses reduce concerns about reverse causality, they do not specify a particular causal mechanism, nor do they rule out reverse causality or omitted variable bias. A substantial body of theory predicts that as economies develop, financial systems will become more market based and the marginal impact of securities markets on economic activity will increase, whereas that of banks will decrease. Our findings are consistent with these pre- dictions and inconsistent with simple reverse causality scenarios. Although a simple reverse causality scenario might predict that economic development increases the size of banks and securities markets relative to the size of the overall economy and that securities markets grow faster than banks, a simple reverse causality scenario does not yield predictions about the differential change in the association between economic activity and bank and securities market development as economies grow. That is, a simple reverse causality sce- nario does not predict that the association between economic activity and bank development diminishes in magnitude while the association between economic 480 THE WORLD BANK ECONOMIC REVIEW activity and securities market development increases in magnitude as countries develop economically. Although these differential effects might be accounted for by sophisticated reverse causality scenarios, potential omitted variable biases, or as yet unformalized theories of finance and development, this paper provides the first empirical evidence that is consistent with an influential theo- retical body of literature predicting that securities markets become more impor- tant for economic activity and that banks become less important as countries develop economically. DATA AND S U M M A RY S TAT I S T I C S Several measures of bank and stock market development are constructed and used to analyze the relationship between economic activity and the structure of the financial system. Economic theory highlights the advantages of developing indicators reflecting the degree to which banks and markets ameliorate market frictions and thereby (1) improve ex ante information about possible invest- ments; (2) enhance the monitoring of investments after financing occurs; (3) facilitate the trading, diversification, and management of risk; (4) ease the mobilization and pooling of savings; and (5) foster the exchange of goods, services, and financial claims. However, such empirical proxies do not exist for a broad cross-section of countries over the last few decades. Instead, measures of the size and activity of banks and securities markets are compiled and employed to examine the relationship between financial systems and economic development. These measures are constructed over the period from 1980 to 2008, and the primary sources of these indicators are provided in table 1. Private credit is used to measure bank development and equals deposit money bank credit to the domestic private sector as a share of GDP. Private credit isolates the credit issued to the private sector and excludes the credit issued to governments, government agencies, and public enterprises. Private credit also excludes credits issued by central banks. Not surprisingly, there is enormous cross-country variation in private credit. For example, averaging the 1980–2008 period, private credit was less than 10 percent of GDP in Angola, Cambodia, and Yemen, and it was greater than 85 percent of GDP in Austria, China, and the United Kingdom. Table 2 indicates that the annual average value of private credit across countries was 39 percent, with a standard deviation of 36 percent. Stock value traded is used to measure market development and equals the value of the stock market transactions as a share of GDP. This market develop- ment indicator incorporates information on the size and activity of the stock market, not simply the value of the listed shares. Previous work by Levine and Zervos (1998) indicates that the trading of ownership claims on firms in an economy is closely tied to the rate of economic development. There is substan- tial variation across counties. As shown in table 2, although the mean value of the stock value traded is approximately 29 percent of GDP, the standard ¨c Demirgu ¸ -Kunt, Feyen, and Levine 481 T A B L E 1 . Variable Definitions and Sources Name Source Definition Dependent variable and baseline financial sector controls Log real GDP per World Development Logarithm of real GDP per capita capita Indicators (constant 2000 USD). Private credit International Financial Deposit money bank credit to the private Statistics sector as a percentage of GDP. Stock value traded Standard & Poor’s Value of stock market transactions as a percentage of GDP. Stock market Standard & Poor’s The value of listed shares on a country’s capitalization stock exchange as a percentage of GDP. Securities market Standard & Poor’s; Bank Stock market capitalization plus domestic capitalization of International private bond market capitalization as a Settlements percentage of GDP. Standard controls Log initial GDP per World Development Log initial real GDP per capita capita Indicators (constant 2000 USD). Log average years of Barro and Lee (2010) Log (1 þ average years of schooling). schooling Log openness to trade World Development Log sum exports and imports of goods Indicators and services as a percentage of GDP. Log government size World Development Log general government consumption as Indicators a percentage of GDP. Log inflation rate International Financial Log (1 þ annual change of CPI). Statistics T A B L E 2 . Descriptive Statistics Variable Mean Standard deviation Maximum Minimum Dependent variable and baseline controls Log real GDP per capita 7.58 1.57 10.94 4.13 (constant 2000 USD) Private credit 39.28 35.90 319.71 0.00 Stock value traded 28.80 57.44 632.34 0.00 Stock market capitalization 47.70 58.39 561.44 0.00 Securities market capitalization 59.08 71.20 588.27 0.00 Standard controls Log average years of schooling 1.86 0.50 2.65 0.03 Log openness to trade 4.26 0.61 6.12 2 1.18 Log inflation rate 0.15 0.37 5.48 2 0.52 Log government size 2.72 0.43 4.42 0.32 Note: Descriptive statistics are calculated on all available annual data for the 1980– 2008 period. deviation is approximately double this value. In Armenia, Tanzania, and Uruguay, the stock value traded annually averaged less than 0.23 percent over the 1980–2008 sample (10th percentile). In contrast, the stock value traded 482 THE WORLD BANK ECONOMIC REVIEW averaged over 75 percent in Hong Kong, Saudi Arabia, Switzerland, and the United States (90th percentile). This paper’s results are robust to using other market development indicators, such as stock market capitalization, which simply measures the value of the listed shares on a country’s stock exchanges as a share of GDP, and securities market capitalization, which equals the capi- talization of the stock market plus the capitalization of the private domestic bond markets divided by GDP. Log real GDP per capita is used to measure economic activity and equals the logarithm of GDP per capita in constant 2000 US dollars. Consistent with theories guiding the empirical analyses (which are discussed in the introduc- tion), log real GDP per capita is examined rather than GDP per capita growth to obtain estimates of the association between economic activity and both bank and securities market development. With the current specification, the estimated coefficients provide information on how log real GDP per capita changes when, for example, securities market development changes. To assess the independent link between finance and economic development, the regressions control for many country characteristics. Some specifications include standard controls, such as years of schooling, openness to trade, infla- tion, government size, initial GDP per capita of the economy in 1980, and dummy variables for the five-year periods of analysis. The correlations in table 3 highlight the key features of the financial system and economic development. First, both bank development and securities market development are positively correlated with economic development. Second, bank development and securities market development are positively correlated with each other, suggesting that financial development involves both larger banks and larger markets. Although these are simple correlations, these basic patterns hold when controlling for many other national traits. THE RELATION SHI PS A MON G BANKS, MARK ET S, A N D E CO N O M I C DE V E LO PM E N T To assess how the relationships between economic activity and both bank development and stock market development evolve with economic develop- ment, quantile regressions are used with data averaged over non-overlapping five-year periods. OLS provides information on the relationship between log real GDP per capita and financial development for a country at an average level of economic development. However, OLS does not provide information on how the relationship between economic activity and financial development differs for countries at different levels of economic activity. The quantile regressions model the relationship between log real GDP per capita and financial development at the specific percentiles (or quantiles) of the log real GDP per capita. Thus, in a quantile regression of log real GDP per capita on private credit, the procedure is able to yield a different estimated coefficient of private credit for each percentile (or quantile) of log real GDP per T A B L E 3 . Correlations Log real GDP Private Stock value Log average Log openness Log inflation Log government Correlations per capita credit traded years of schooling to trade rate size Private credit 0.67*** 1 Stock value traded 0.41*** 0.51*** 1 Log average years of schooling 0.71*** 0.49*** 0.26*** 1 Log openness to trade 0.25*** 0.21*** 0.08*** 0.31*** 1 Log inflation rate 2 0.15*** 2 0.16*** 2 0.12*** 2 0.03** 2 0.13*** 1 Demirgu Log government size 0.28*** 0.21*** 0.04** 0.25*** 0.28*** 2 0.08*** 1 Source: Author’s analysis based on data described in the text. ¨c¸ -Kunt, Feyen, and Levine Note: Correlations are calculated on all available annual data for the 1980– 2008 period. *, **, and *** denote the significance level of correlation at the 10, 5, and 1 percent levels, respectively. 483 484 THE WORLD BANK ECONOMIC REVIEW capita. For example, the estimated coefficient at the 50th percentile is a median regression, yielding the estimated relationship between the log real GDP per capita and private credit at the median level of economic activity. By comput- ing the quantile regression for each of the 5th to the 95th quantiles, the analyses provide information on how the relationship between economic activity and financial development differs across distinct levels of log real GDP per capita. Neither the OLS nor the quantile regressions identify the causal impact of bank and securities market development on economic development. Rather, the goal is to explore whether and how the relationship between changes in economic activity and changes in both bank and market development varies with the level of economic development. Illustrating the Quantile Regression Results In panel A of figure 1, the graph on the upper left side plots the coefficients from 91 separate quantile regressions for each percentile from the 5th through the 95th percentiles of log real GDP per capita, where the dependent variable is log real GDP per capita, and the main regressor is private credit. The regres- sions control for stock value traded. A circle represents each coefficient esti- mate produced by the quantile regression associated with the corresponding percentile. The left axis provides information on the values of the coefficient es- timates. Thus, the estimated coefficient depicts the sensitivity of log real GDP per capita associated with a change in private credit at each percentile of eco- nomic development. These estimates are statistically significant. Additional in- formation on the sensitivity of these estimates is provided below. The graph also plots the actual values of private credit at each percentile, which are desig- nated with a triangle. The scale of the values of private credit is provided on the right axis. The graphs in the remainder of panel A of figure 1 provide similar information on the relationship between economic activity and stock market development. The upper-right graph provides information for stock value traded. The lower graphs confirm the increasingly relevant role of securi- ties markets by documenting similar upward trends for both securities market capitalization and stock market capitalization. Panel B of figure 1 provides the same types of quantile analyses while con- trolling for other characteristics of the national economies. The standard con- trols are log real GDP per capita in 1980, government size, openness to trade, inflation, average years of schooling, and period fixed effects. Each of the eight graphs in panels A and B of figure 1 provides two addition- al pieces of information. First, the horizontal dotted line is the OLS estimate of the coefficient of the financial development indicator. Thus, in the graph on the upper left side of panel A in figure 1, this line is simply the coefficient of private credit from an OLS regression of log real GDP per capita on private credit for the full sample of country-year observations, controlling for stock value traded. When moving away for the mean log real GDP per capita, the quantile estimates become statistically different from the OLS estimates. The ¨c Demirgu ¸ -Kunt, Feyen, and Levine 485 nature of these deviations is explored below. Second, the solid line shows the estimated linear relationship between each estimated coefficient of the financial development indicator (i.e., the circles) and the GDP per capita percentile asso- ciated with the coefficient. As a specific example, consider the graph in the upper-right quadrant of panel B in figure 1. The estimated coefficients of stock value traded after conditioning on the standard controls and period fixed Figure 1. Quantile Coefficients for Private Credit and Securities Market Activity Source: Author’s analysis based on the data described in the text. Note: The dependent variable is log real GDP per capita. The figure depicts the coefficients of the quantile regressions of private credit, stock value traded, securities market capitalization, and stock market capitalization as independent variables for each of the 5th to 95th percentiles of the GDP per capita distribution on the left axis. Private credit is defined as deposit money bank credit to the private sector as a percentage of GDP. Stock value traded is the value of stock value transactions as a percentage of GDP. Stock market capitalization is the value of listed shares on a country’s stock exchanges as a percentage of GDP. Securities market capitalization is defined as stock market capitalization plus sdmestic private bond market capitalization as a percentage of GDP. Percentile values are reported on the right axis. Data are five-year non-overlapping country averages. Panel A does not control for additional variables. Panel B controls for standard controls: initial GDP per capita, government size, openness to trade, inflation, average years of schooling, and time-fixed effects. The horizontal dotted line depicts the OLS estimate. The solid lines represent the linear fit. 486 THE WORLD BANK ECONOMIC REVIEW Figure 1. Continued effects are collected. Then, these estimated coefficients are regressed on the GDP per capita percentile associated with the estimates. Panel A in table 4, column (4) provides the results of this regression. The estimated coefficient for each GDP per capita percentile provides the trend line graphed in figure 1. Discussing the Quantile Regression Results In terms of bank development, figure 1 shows that as the log real GDP per capita rises, (1) private credit rises (triangles) and (2) the marginal increase in the log real GDP per capita associated with an increase in private credit dimin- ishes (circles). Put differently, the level of bank development increases, but its association with economic activity diminishes. In panel A of table 4, the signifi- cance of this association is tested formally. In this panel, the dependent vari- able is the estimated linear association between economic activity and either bank development or securities market development at each percentile of the distribution of GDP per capita underlying figure 1. Regressions (1) and (2) show that this relationship is statistically significant: as economic activity T A B L E 4 . Robust Regression Results of the Linear Regression Fits from Figure 1 Panel A: Linear model Dep. Var.: Percentile regression coefficient Private credit Dep. Var.: Percentile regression coefficient Stock value traded 1 (No controls) 2 (With all controls) 3 (No controls) 4 (With all controls) Percentile 2 1.24E 2 04*** 2 1.02E 2 05*** 4.18E 2 05*** 3.79E 2 05*** [0.00] [0.00] [0.00] [0.00] Constant 2.51E 2 02*** 4.45E 2 03*** 2.05E 2 03*** 2 1.34E 2 03 [0.00] [0.00] [0.00] [0.00] Controls No Yes No Yes Observations 91 91 91 91 Panel B: Quadratic model Percentile 2 1.09E 2 04*** 2.67E 2 05*** 2.13E 2 04*** 4.88E 2 05*** [0.00] [0.00] [0.00] [0.00] Square of percentile 2 1.37E 2 07 2 3.69E 2 07*** 2 1.63E 2 06*** 2 1.02E 2 07** [0.00] [0.00] [0.00] [0.00] Constant 2.48E 2 02*** 3.77E 2 03*** 2 1.44E 2 03*** 2 1.58E 2 03 Demirgu [0.00] [0.00] [0.00] [0.00] Controls No Yes No Yes Observations 91 91 91 91 ¨c¸ -Kunt, Feyen, and Levine Source: Author’s analysis based on data described in the text. Note: The table displays the robust regressions results of the linear fits depicted in figure 1. The dependent variables are coefficients of the quantile regres- sions of private credit and stock value traded for each of the 5th to 95th percentiles of the GDP per capita distribution, respectively, on five-year non- overlapping country averages. Panel A reports a linear model where the regressors are a constant and the income percentile associated with the coefficient. Panel B shows the results for the quadratic model using the same independent variables in the linear model plus the square of the percentile associated with the coefficient. Columns 1 and 3 use the coefficients of quantile regressions without additional controls ( panel A of figure 1). Columns 2 and 4 use the coeffi- cients of quantile regressions that include standard controls: Initial GDP per capita, Government size, Openness to trade, Inflation, Average years of school- ing, and time-fixed effects ( panel B of figure 1). The p-values in brackets are based on robust country-level clustered standard errors. *, **, and *** denote the significance on the 10, 5, and 1 percent levels, respectively. 487 488 THE WORLD BANK ECONOMIC REVIEW increases, there is a significant reduction in the estimated coefficient of private credit. The results are the opposite for securities market development. As log real GDP per capita rises, (1) stock value traded rises, and (2) the marginal increase in log real GDP per capita associated with an increase in stock value traded rises. That is, as countries develop economically, securities market development increases, and its association with economic activity increases. Regressions (3) and (4) in panel A of table 4 show that this effect is statistically significant: the association between economic activity and stock value traded increases as the log real GDP per capita rises. These results suggest that the relationship between bank development and economic activity differs from the relationship between securities market development and economic activity. Figure 1 suggests that there might be a nonlinear relationship (1) between economic activity and bank development and (2) between economic activity and securities market development. To assess the sensitivity of our findings and provide more information on the nature of the relationship, we examine these relationships more rigorously in panel B of table 4, which includes a quadratic term to allow for a potential nonlinear, parabolic relationship. This potential relationship makes it possible to estimate the level of economic activity at which the associations between financial and economic development start to decrease as the economy develops further. Consistent with figure 1, the regression results in panel B of table 4 suggest that there are a nonlinear associations between economic activity and both bank and securities market development. At very low levels of economic devel- opment, the association between economic activity and bank development is increasing in economic development, but the slope quickly becomes negative. In particular, regressions (1) and (2) of panel B indicate that the slope of the association between economic activity and bank development becomes negative after real GDP per capita reaches $1,032 in 2000 US dollars (36th percentile). For securities market development, panel B of table 4 indicates that the associa- tion between economic activity and securities is always increasing in the 5th to 95th percentile interval of economic development, but at a decreasing rate (regressions (3) and (4)). In other words, only the upward sloping part of the estimated parabola is relevant. For example, the regression (4) estimates suggest that as economies grow and move to the next percentile, the coefficient increases by more than 5 percent for countries below the 20th percentile. In contrast, the coefficient increases by just 0.7–1.0 percent for each additional percentile from the 78th percentile upward. Broader Implications of Quantile Analyses The results of the above analysis are consistent with several lines of theoretical research on the evolving importance of banks and financial markets during the process of economic development. As noted in the introduction of this paper, Allen and Gale (2000), Boot and Thakor (1997, 2000), Boyd and Smith ¨c Demirgu ¸ -Kunt, Feyen, and Levine 489 (1998), Song and Thakor (forthcoming), and others stress that at higher levels of economic development, economies require the types of custom-designed fi- nancial arrangements that ease the financing of novel, longer-term investments that often employ more intangible inputs than the types of projects that domi- nate economic activity at lower levels of economic development. These theories predict that securities markets are comparatively better than banks at financing these activities. Thus, influential lines of theoretical analysis predict that the services provided by securities markets will become more important for foster- ing economic activity as economies grow, whereas those provided by banks will tend to become less important. The quantile regression results are consis- tent with these predictions. However, the quantile regression results are incon- sistent with the view that economic development is simply associated with an increase in bank and stock market development with no differential effect on their association with economic activity. That is, as countries develop economi- cally, the association between economic activity and bank development tends to weaken, whereas the association with securities market development tends to strengthen. CONCLUSIONS Banks and markets evolve during the process of economic development. As economies grow, both the banking system and financial markets become more developed, but the association between economic activity and bank develop- ment tends to fall, and the association between economic activity and securities market development tends to increase. For the first time, the quantile analyses employed in this paper directly assess the predictions emerging from an influen- tial line of theoretical work on financial structure. The findings are consistent with the view that (a) financial institutions provide different financial services from those provided by financial markets, and (b) as economies grow, the services provided by securities markets become more important for promoting economic activity, whereas those provided by banks become less important. As such, this research suggests that policies and institutions that impede the evolu- tion of the structure of financial systems as economies grow can have detrimen- tal ramifications for economic development. REFERENCES Acemoglu, D., and F. Zilibotti. 1997. “Was Prometheus Unbound by Chance? Risk, Diversification, and Growth.” Journal of Political Economy 105: 709–775. Allen, F., and D. Gale. 1995. “A Welfare Comparison of the German and U.S. Financial Systems.” European Economic Review 39: 179–209. ———. 1997. “Financial Markets, Intermediaries, and Intertemporal Smoothing.” Journal of Political Economy 105: 523– 546. 490 THE WORLD BANK ECONOMIC REVIEW ———. 1999. “Diversity of Opinion and Financing of New Technologies.” European Economic Review 39: 179–209. ———. 2000. Comparing Financial Systems. Cambridge, MA: MIT Press. Beck, T., and R. Levine. 2002. “Industry Growth and Capital Accumulation: Does Having a Market- or Bank-based System Matter?” Journal of Financial Economics 64: 147–180. ———. 2004. “Stock Markets, Banks, and Growth: Panel Evidence.” Journal of Banking and Finance 28: 423–442. Boot, A. W. A., and A. Thakor. 1997. “Financial System Architecture.” Review of Financial Studies 10: 693 –733. ———. 2000. “Can Relationship Banking Survive Competition?” Journal of Finance 55: 679– 713. Boyd, J. H., R. Levine, and B. D. Smith. 2001. “The Impact of Inflation on Financial Sector Performance.” Journal of Monetary Economics 47: 221–248. Boyd, J. H., and B. D. Smith. 1998. “The Evolution of Debt and Equity Markets in Economic Development.” Economic Theory 12: 519– 560. ¨c Demirgu ¸ -Kunt, A., and R. Levine. 2001. Financial Structures and Economic Growth: A Cross-Country Comparison of Banks, Markets, and Development. Cambridge, MA: MIT Press. ¨c Demirgu ¸ -Kunt, A., and V. Maksimovic. 1998. “Law, Finance, and Firm Growth.” Journal of Finance 53: 2107– 2137. ———. 2002. “Funding Growth in Bank-based and Market-based Financial Systems: Evidence from Firm Level Data.” Journal of Financial Economics 65: 337–363. Dewatripont, M., and E. Maskin. 1995. “Credit Efficiency in Centralized and Decentralized Economies.” Review of Economic Studies 62: 541 –555. Goldsmith, R. W. 1969. Financial Structure and Development. New Haven, CT: Yale University Press. Holmstrom, B., and J. Tirole. 1993. “Market Liquidity and Performance Monitoring,” Journal of Political Economy 101: 678– 709. Koenker, T., and G. Basset. 1978. “Quantile Regressions.” Econometrica 46(1): 33 –50. Levine, R. 2002. “Bank-based or Market-based Financial Systems: Which Is Better?” Journal of Financial Intermediation 11: 398– 428. Levine, R., and S. Zervos. 1998. “Stock Markets, Banks, and Economic Growth.” American Economic Review 88: 537– 558. Morck, R., and M. Nakamura. 1999. “Banks and Corporate Control in Japan.” Journal of Finance 54: 319 –340. Rajan, R. G. 1992. “Insiders and Outsiders: The Choice between Informed and Arm’s Length Debt.” Journal of Finance 47: 1367– 1400. Song, F., and A.V. Thakor. Forthcoming. “Financial System Architecture and the Co-evolution of Bank and Capital Markets.” Economic Journal. Weinstein, D. E., and Y. Yafeh. 1998. “On the Costs of a Bank-centered Financial System: Evidence from the Changing Main Bank Relations in Japan.” Journal of Finance 53: 635 –672. Notes on Financial System Development and Political Intervention* Fenghua Song and Anjan Thakor We study the impact of political intervention on a financial system that consists of banks and financial markets and develops over time. In this financial system, banks and markets exhibit three forms of interaction: they compete, they complement each other, and they co-evolve. Co-evolution is generated by two new ingredients of finan- cial system architecture relative to the existing theories: securitization and risk-sensi- tive bank capital. We show that securitization propagates banking advances to the financial market, permitting market evolution to be driven by bank evolution, and market advances are transmitted to banks through bank capital. We then examine how politicians determine the nature of political intervention designed to expand credit availability. We find that political intervention in banking exhibits a U-shaped pattern, where it is most notable in the early stage of financial system development (through bank capital subsidy in exchange for state ownership of banks) and in the advanced stage (through direct lending regulation). Despite expanding credit access, political intervention results in an increase in financial system risk and does not contribute to financial system evolution. Numerous policy implications are drawn out. JEL codes: G21, G28 There is strong evidence that the development of the financial system – consist- ing of banks and financial markets – positively affects real-sector growth through the efficient mobilization and allocation of capital. This point was recognized at least as far back as Gurley and Shaw’s (1955, 1956) discussions of the relationship between financial system development and growth in the real sector. Gurley and Shaw (1955) note, “Conventional theories of income, interest, and money have given insufficient attention to important reciprocal relationships * Fenghua Song (song@psu.edu) is Assistant Professor of Finance at Smeal College of Business, Pennsylvania State University, University Park, PA 16802. Anjan Thakor (thakor@wustl.edu; corresponding author) is Research Associate with European Corporate Governance Institute (ECGI) and John E. Simon Professor of Finance at Olin Business School, Washington University in St. Louis, Campus Box 1133, One Brookings Drive, St. Louis, MO 63130. We thank Justin Yifu Lin, Claudio Raddatz, participants at the World Bank Conference on Financial Structure and Economic Development (June 2011), and particularly two anonymous referees for their comments, and to the World Bank for financial support. We alone are responsible for any remaining errors. THE WORLD BANK ECONOMIC REVIEW, VOL. 27, NO. 3, pp. 491 –513 doi:10.1093/wber/lhs011 Advance Access Publication April 27, 2012 # The Author 2012. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 491 492 THE WORLD BANK ECONOMIC REVIEW between real development and financial development . . . An additional complex- ity is that development of financial institutions, including nonbank intermediar- ies, is both a determined and a determining variable in the growth process.” More recently, this point has been made by various authors. For example, Gertler (1988) points out the financial system matters because it impinges on the resolution of informational asymmetries that have real effects, and that financial contracts and institutions are determined simultaneously with real variables. This view is well exemplified in the following statement by Levine (2010) in a debate on the role of the financial system: “Finance is powerful . . . the last few centuries demonstrate that financial innovation is crucial, indeed indispensable, for sustained economic growth and prosperity.” This recognition of the role of the financial system in economic development notwithstanding, it is less clear whether the effect of the financial system can be attributed to the development of capital markets or banks and other finan- cial institutions, or even whether the architecture of the financial system (i.e., the configuration/mix of markets and institutions within a financial system) matters for the real-sector growth. Complicating this picture is the fact that political intervention in credit markets seems to be ubiquitous, and often has telling consequences (Brown and Dinc ¸ 2005; Calomiris and Wallison 2009). This raises interesting questions about the manner in which financial systems evolve, how they affect economic growth, and the endogenous nature of polit- ical intervention at different stages of financial system development. The purpose of this paper is to theoretically address these questions. To put our research in perspective, it is useful to begin by briefly reviewing the existing literature on financial system architecture, which deals with the mix of bank and market financing in the overall allocation of credit. Financial systems are typically classified as bank-based or market-based, depending on the relative shares of financing activities performed through banks and other financial intermediaries as compared to those conducted through capital markets.1 An important issue in this regard has to do with which system maxi- mizes real-sector economic growth, but on this the empirical evidence is mixed. For example, Beck and Levine (2002), Demirgu ¨c¸ -Kunt and Levine (2001), and Levine (2002) show that the positive impact of financial system development on economic growth is unaffected by whether the evolution of the financial system is due to bank or financial market development, and the correlation between a country’s economic growth and the architecture of its financial system is rather low. Deidda and Fattouh (2008) find, however, that a change from a bank-dominated system to one with both banks and markets can actually hurt economic growth.2 1. Typical examples of market-based systems are those in the U.S. and U.K., and Japan and Eurozone countries like Germany are traditionally viewed as having bank-based systems. 2. See also the growing body of other empirical research on financial system architecture as well as its impact on growth (Levine and Zervos 1998; Tadesse 2002). Song and Thakor 493 The theoretical literature on financial system architecture has been largely focused on examining how banks and markets compete for borrowers and the comparative capital allocational efficiencies of bank-based systems versus market-based systems, in terms of providing risk sharing (Allen and Gale 1997), generating information feedback from security prices (Boot and Thakor 1997a; Subrahmanyam and Titman 1999), creating financial innovation incen- tives (Boot and Thakor 1997b), resolving asset-substitution moral hazard (Boot and Thakor 1997a), and protecting borrower proprietary information and strengthening R&D incentives (Bhattacharya and Chiesa 1995; Yosha 1995).3 This “banks versus markets” distinction, based on the dominant view that banks and markets compete, clearly has powerful policy implications. It means that regulators must make choices of focus to seek the right balance between the two sectors when attempting to influence the trajectory of financial system development. However, this neat distinction between banks and markets, rooted in historical patterns of financial system evolution from many decades ago, seems to be rapidly losing relevance in an era in which securitization and widespread use of capital market financing by banks have largely eroded the boundaries between banks and markets, if not dissolved them completely. Mindful of these contemporary developments, we pose the following specific questions about financial system development: (1) What are the various forms of interaction that banks and markets engage in at different stages of development of the financial system; and (2) What forms of endogenously determined political interventions in the financial system should we expect in these different stages and what are the potential consequences of such intervention? To address these questions, we develop a theoretical framework in which we do not restrict the interaction between banks and markets to competition, but permit them to interact in a richer manner. Our analysis reveals that banks and markets exhibit three forms of interaction. They compete, they complement each other, and they co-evolve. Strong institutions like banks are necessary for markets to perform well, and well-functioning markets are essential for banks to be induced to be sufficiently well capitalized to expand credit availability to borrowers without exposing themselves to excessive risk. Thus, the notion that one part of the financial system takes precedence over the other at a particular stage in terms of the development focus of policymakers is rendered obsolete. Our analysis pivots on two key frictions that impede a borrower’s ability to obtain external financing and influence the cost of this financing. One is the “certification friction,” which arises from the fact that there is imperfect infor- mation about borrower credit quality. As a result, a creditworthy borrower may be erroneously denied credit were it not be “certified” as worthy of credit. 3. See also Allen and Gale (1999), Dewatripont and Maskin (1995), and von Thadden (1995). 494 THE WORLD BANK ECONOMIC REVIEW The other is the “financing friction,” which arises due to uncertainty about the availability of opportunities for the borrower to enhance its project payoff. This means that good investment opportunities may be passed up, which increases a borrower’s cost of obtaining funds from the capital market. We show that banks are better at diminishing the certification friction due to their credit-analysis expertise, whereas capital markets are superior at resolving the financing friction by providing a trading venue through which informed investors’ private information about the project payoff enhancement opportun- ity is (noisily) impounded into security prices, thereby lowering the probability that valuable investment opportunities are passed up. Banks and markets compete to attract borrowers based on their relative advantages in resolving the two frictions. Relative to the existing theories of financial system architecture, the recipe for our analysis includes two new ingredients: securitization and (risk-based) bank capital. Securitization blends the roles of banks and markets by involving the bank in the origination and screening of the loan and the market in the provision of financing. This, in turn, generates a positive feedback loop from bank development to market development. Improvements in the bank’s screen- ing technology, as a result of bank evolution, increase the confidence that capital market investors have in a securitized borrower’s credit quality, which stimulates greater informed trading in the capital market. This lowers the finan- cing friction and spurs capital market evolution. Bank capital connects the two sectors in a different way. As the capital market evolves, informed trading in the market increases, which reduces the fi- nancing friction for all seeking market financing, including banks which now confront a lower cost of equity capital. This enables the bank to raise addition- al equity from the capital market to extend riskier loans that were previously eschewed, and expands the banking sector’s lending scope. Moreover, as banks lend to riskier borrowers with lower prior credit qualities, they find it privately optimal to improve the precision of their screening to better distinguish credit- worthy borrowers from those who are not creditworthy. This stimulates evolu- tion of the banking sector. That is, advances in the capital market are propagated to the banking sector through bank capital. The analysis thus uncovers not only competition and complementarity between bank and market financing, but also co-evolution. Bank evolution spurs capital market evolution, and capital market evolution in turn spurs bank evolution. That is, banks and capital markets co-evolve with each other. The limitation of the standard approach that views banks and capital markets as competitive tubs on their own bottoms in the analysis of financial system de- velopment becomes starkly apparent. Securitization plays a key role in the co-evolution of banks and markets in our analysis. Thus, the shadow banking system, where securitization-related activities are featured prominently, is of pivotal importance for the issues we examine. In a sense, our analysis highlights the “bright side” of the interaction between banks Song and Thakor 495 and markets through shadow banking. But the shadow banking system also played a key role in the recent financial crisis (see, for example, Stein 2010). Gennaioli, Shleifer, and Vishny (2011) present a formal model of the link between securitization and trading of asset-backed securities in the shadow banking sector and show that securitization forges market-based connections between banks, leading to a reduction in idiosyncratic risk but an increase in systemic risk. This may be viewed as the “dark side” of shadow banking. In addition, our framework also provides a setting in which we can investi- gate the role of political intervention in financial system development, some- thing with potential ramifications for financial crises and stability. This is virgin territory in the theory of financial system architecture. Nonetheless, pol- itical intervention in the financial system is not only commonplace but appears to be on the rise. Credit markets are hard for politicians to resist intervening in because expanded credit availability for potential voters can generate large per- ceived political benefits. It has been empirically documented that in emerging markets, cozy ties between bankers and politicians as well as state ownership of banks allow for a great deal of political influence in credit markets (e.g., Brown and Dinc ¸ 2005; Dinc ¸ 2005; La Porta, Lopez-de-Silanes, and Shleifer 2002). And the recent financial crisis has shown that even advanced financial systems, such as the one in the U.S., are not immune to political influence. See Calomiris and Wallison (2009), for example, for a discussion of how political influences on securitization agencies like Fannie Mae and Freddie Mac, as well as on credit rating agencies and banks, gave shape to some of the initial forces that influenced the course of the 2007-09 financial crisis. Giving politicians a choice of how to intervene in the financial system, we endogenize the manner of political intervention at different stages of financial system development, where the political goal is to expand credit availability. We find that in the early stage of financial development, when both the certi- fication friction and the financing friction are large, political intervention to expand credit availability takes the form of government ownership of banks in exchange for direct capital subsidies to banks. In the intermediate stage, there is little political intervention. In the advanced stage, political intervention returns, but this time takes the form of direct lending regulation that mandates that banks invest in low-quality, high-risk (excluded) borrowers, even if doing so imposes a loss on the banking sector. The reason for this stage-dependent intervention is that the value of capital subsidies depends on the cost of equity capital for banks, which in turn depends on the stage of development of the capital market. Moreover, bank profitability is also dependent on the stage of financial market development, so the feasibility of enacting regulations that impose losses on banks is also stage dependent. In the intermediate stage, bank capital is neither sufficiently costly to make an equity subsidy attractive to banks, nor are the banking sector’s profits sufficiently high to absorb the losses imposed by direct lending regulations. Thus, political intervention in banking exhibits a U-shaped pattern: it is most notable in underdeveloped and highly 496 THE WORLD BANK ECONOMIC REVIEW developed financial systems, but it takes different forms. Differences in form notwithstanding, the effect of political intervention is the same – an increase in financial system risk. The rest of the paper is organized as follows. In Section I, we describe the theoretical framework. Section II analyzes the three forms of interaction between banks and markets. Section III examines political intervention at dif- ferent stages of financial system development. Section VI concludes with a dis- cussion of policy implications of our analysis. I. THE FRAMEWORK In this section, we describe the economic framework, with a minimum of mathematical details. The reader is referred to Song and Thakor (2010, 2011) for more formal treatments of these ideas, including proofs of results. The focus here is on explaining the logical structure that leads to these results and the economic intuition underlying them. Agents and Economic Environment Consider an economy where each borrower needs outside finance to invest in a project. The success of the investment depends on the borrower’s type. The borrower may be either authentic or a crook, and only an authentic borrower will invest in the project, whereas a crook will just abscond with the funds, leaving nothing for the financiers. Each borrower’s type is its private informa- tion. The credit quality of a borrower, q [ [0,1], is simply the common prior assessment about the probability with which the borrower is authentic, with a higher q representing a better credit quality. An authentic borrower will have an opportunity to enhance its project payoff, given a particular macroeconomic state; this will be discussed in more detail shortly. We study a financial system in which there are two frictions affecting an au- thentic borrower’s ability to obtain external finance and the cost of this financing. (1) The first is a certification friction, which arises due to imperfect informa- tion about borrower type. As a result, even a creditworthy (i.e., authen- tic) borrower may be (erroneously) denied credit. (2) The second friction, financing friction, arises as follows. Each borrower’s project value depends on whether the borrower invests in costly payoff enhancement. The investment in payoff enhancement achieves its goal only if the macroeconomic condition turns out to be favorable (e.g., a strong consumer demand for the borrower’s product, and a viable supply of labor force). Absent information about the macroeconomic condition, an authentic borrower finds it privately unprofitable to invest in payoff enhancement, since the expected return from such an invest- ment is less than its cost, where the expectation is calculated based on the borrower’s prior assessment about the macroeconomic condition. Song and Thakor 497 There are traders in the capital market who find it personally profitable to invest in acquiring information about the macroeconomic condition that is useful to authentic borrowers for making their payoff-enhancement investment decisions. The motivation for the informed traders is that ac- quisition of this information can be used by these traders to take profit- able long positions in the security issued by the borrower when the traders learn the macroeconomic condition is conducive to making the project-payoff enhancement decision. In addition to informed traders, there are also liquidity traders who are uninformed about the macroeco- nomic condition and trade for pure liquidity needs in the capital market. Trading by liquidity traders introduces noise in the price formation process and masks informed trades, but the informed traders’ information will still be (noisily) transmitted to borrowers through the equilibrium prices of securities traded in the capital market. To see this more clearly, suppose the informed traders first observe a favorable macroeconomic condition that they conjecture the authentic borrower will exploit to enhance the project payoff, and then they take long positions in the bor- rower’s security that (noisily) reveal their conjecture. By observing capital market trading and the market-clearing price for its security, the authentic borrower infers the macroeconomic state and decides to make the payoff-enhancement investment, thereby rationalizing the informed traders’ initial conjecture. That is, the capital market provides borrowers valuable information through the feedback effect of security prices. The larger the number of informed traders, the greater the impact of their trading on the price of the security, so security prices become more informative about the macroeconomic condition when there are more informed traders present in the capital market. Hence, the authentic bor- rowers’ investment in project-payoff enhancement becomes more efficient when there are more informed traders. However, to become an informed trader requires personal investment in information acquisition by the trader, which is costly. This cost will be greater the less developed the fi- nancial system. As the financial system develops, an increasing amount of payoff-relevant information is available in the public domain for inves- tors, so it costs less to acquire additional information. This induces more investors to become informed in more advanced financial systems, making borrowers’ payoff-enhancement decisions more informative. Think of the larger number of informed and sophisticated traders in the U.S. compared to say Zambia. Banks and Capital Markets Banks are better than markets at diminishing the certification friction because of their expertise in credit screening that may help to attenuate asymmetric-information problems (e.g., Allen 1990; Boyd and Prescott 1986; 498 THE WORLD BANK ECONOMIC REVIEW Coval and Thakor 2005; Ramakrishnan and Thakor 1984). Specifically, bank screening yields an informative signal that noisily reveals the borrower’s type, where the precision of the signal is increasing in the bank’s investment in a screening technology, which is privately costly to the bank. Banks may also arise to reduce state-verification costs that endogenously give rise to debt con- tracts (Townsend 1979). Nonetheless, not all information asymmetries may be eliminated by banks, so credit market distortions associated with practices like credit rationing, that do not resemble standard Walrasian market-clearing out- comes, may persist (Stiglitz and Weiss 1981). This may create a role for capital market financing in different forms to provide amelioration that is unattainable with bank-intermediated contracting. Capital markets are better than banks at resolving the financing friction by providing a setting in which security trading leads to the (noisy) impounding of informed traders’ private information about the macroeconomic condition. As explained earlier, this friction is dissipated to a greater extent in more devel- oped financial systems, where borrowers learn more accurately about the macroeconomic condition and make more efficient investment decisions. We have modeled the financing friction resolved by the capital market purely in terms of the feedback effect of market prices as, for example, in Boot and Thakor (1997a). However, the literature has highlighted various other eco- nomic functions served by markets. For example, Allen and Gale (1999) show how markets can aggregate divergent opinions about new technologies.4 This can enhance synergies between financial system development and real-sector development, as shown by Laeven, Levine, and Michalopoulos (2012). Such in- formation aggregation may be particularly important when equity rather than debt is used for financing, since opinions may diverge more about the upside potential of a new technology. Thus, markets may rely on equity to a greater extent than banks do, even if banks were allowed to take equity positions in borrowers.5 Viewed in this light, the resolution of the financing friction by the market may be viewed more broadly as the resolution of an information aggre- gation problem. Nonetheless, our specification of the role of markets as com- municating valuable payoff-relevant information to firms is entirely consistent with this alternative aggregation-of-diverse-opinions viewpoint, since it is pre- cisely this form of aggregation that paves the way for the market to generate in- formation that may not be readily available even to the firm’s manager. Financing Sources Each borrower has three sources from which it can choose to finance its project: (1) borrow directly from the capital market by issuing a debt security; 4. Other papers have also used heterogeneous prior beliefs to explore issues in banking, e.g., Song and Thakor (2007). 5. Since credit rationing in models like Stiglitz and Weiss (1981) results from the use of debt contracts, the market can eliminate it by relying on equity. Song and Thakor 499 (2) let a bank screen and certify its credit quality first and then borrow from the capital market via securitization; or (3) take a loan from a bank. With direct capital market financing, the bank is out of the picture, so there is no “bank certification” provided to the borrower. With securitization, the bank screens the borrower’s credit quality first, and then seeks financing for the borrower from the capital market via securitized debt only if the screening yields a favorable outcome. The bank incurs a fixed cost to set up a trust for securitization. With a bank loan too, the bank screens the borrower’s credit quality first, and then extends the loan only if the screening outcome turns out to be favor- able. Of course, since the screening is noisy, the bank’s credit decision is subject to both Type-1 and Type-2 errors. Thus, with both securitization and bank financing, the borrower’s credit quality is “certified” by the bank, albeit with the recognition that the certification is not perfect. A key difference between a direct bank loan and securitization is that, with the former, the bank has to raise funding for the loan, whereas with the latter, the funding comes from the capital market. For the bank loan, there is a risk-sensitive capital requirement that is increasing in borrower risk. The bank funds the rest of the loan from deposits. Deposit gathering is costly due to the cost of setting up branches, employing tellers, etc., and the total cost of deposit gathering is increasing in the amount of deposits raised by the bank. As we will discuss in more depth later, bank certification will influence the credit terms that the borrower can obtain from the market with securitization. Moreover, bank certification also impinges on how many traders choose to become informed in the capital market and hence the accuracy of the payoff-relevant information that the borrower can infer from security prices. I I. R E SU LTS In this section, we discuss two main results arising from our analysis of the model described in the previous section. The first result concerns the bor- rower’s choice of financing source. The second result characterizes the inter- action between banks and capital markets: how does the evolution of one sector of the financial system affect the other sector, and how does a financial system evolve as a whole? Borrower’s Choice of Financing Source We first analyze a borrower’s choice of financing source. We focus on authentic borrowers, as crooks will make the same financing choices as authentic bor- rowers in equilibrium to avoid being identified. First, consider borrowers with the lowest credit quality, i.e., borrowers in cohorts with the highest probability of having crooks. It is clear that as a borrower’s credit quality (q) declines, a bank’s expected payoff from lending to the borrower declines as well due to the increasingly high likelihood of lending money to a crook; this is because 500 THE WORLD BANK ECONOMIC REVIEW the bank’s screening is noisy. Thus, there exists a credit quality cutoff, call it ql . 0, below which a borrower cannot obtain financing. This cutoff deter- mines the banking sector’s lending scope, with a lower cutoff corresponding to a broader lending scope. To examine the borrower’s choice of financing source beyond the credit ex- clusion cutoff of ql, it is useful to introduce a little notation. The benefit to the borrower of having its loan originated by the bank is that the bank provides screening and hence “certifies” an affirmatively-screened borrower as credit- worthy, which reduces the borrower’s cost of credit. Let VS represent this screening benefit. It is clear that @ VS/@ q , 0, that is, the certification benefit declines as the borrower’s observable quality improves. This screening benefit is available only with a bank loan and securitization. Bank certification produces another benefit for the borrower if it chooses se- curitization. Because a certified borrower is more likely to be authentic, the project-payoff enhancement conjectured by informed traders who observe a fa- vorable macroeconomic state is more likely to be realized with a certified bor- rower; note that a crook will not invest in payoff enhancement even if the macroeconomic condition is favorable. This means that when a borrower is bank-certified and has securitized its debt, a larger number of capital market investors will acquire information about the payoff-enhancement-relevant macroeconomic condition than if the borrower finances in the capital market without bank certification. Thus, bank certification enables an authentic borrow- er to make a more informed decision about its investment in payoff enhance- m ment. Let VPE be the value of payoff enhancement to the borrower with direct sec market financing, and VPE be the value of payoff enhancement with securitiza- sec m sec tion. Our arguments above indicate that VPE ðqÞ . VPE ðqÞ, @ VPE ðqÞ=@ q . 0, m and @ VPE ðqÞ=@ q . 0. Moreover, since the borrower’s bank-certification benefit sec m will get smaller as q increases, we expect @ ½VPE ðqÞ À VPE ðqފ=@ q , 0. Recall that bank financing is also associated with a deposit-gathering cost that is increasing in the amount of deposits gathered. Denote this cost by D(q). Since the bank faces a risk-sensitive (equity) capital requirement that is decreas- ing in q, it follows that @ D(q)/@ q . 0. With these preliminaries, we can write down the incremental value of bank financing relative to capital market financing as m Dbank ¼ VS ðqÞ À DðqÞ À VPE ðqÞ; ð1Þ and the incremental value of securitization relative to capital market financing as sec m Dsec ¼ VS ðqÞ À F þ VPE ðqÞ À VPE ðqÞ; ð2Þ where F is the fixed cost of setting up the securitization trust. Song and Thakor 501 Let us now focus on borrowers who are not subject to credit exclusion, that is, q ! ql . For these borrowers, if q is relatively low, then VS(q) is relatively high. Moreover, for low q, banks face high capital requirements, so D(q) will be m sec relatively low. The low q also implies that VPE ðqÞ will be relatively low; VPE ðqÞ m will be low too but higher than VPE ðqÞ. Hence, for low q, both Dbank and Dsec will be strictly positive. But if F is large enough, we will have Dbank . Dsec. Now, as q increases, we see that m @ Dbank @ VS ðqÞ @ VPE ðqÞ @ DðqÞ ¼ À À , 0; ð3Þ @q @q @q @q m sec @ Dsec @ VS ðqÞ @ VPE ðqÞ @ VPE ðqÞ ¼ À þ , 0: ð4Þ @q @q @q @q sec Since @ D(q)/@ q . 0 and @ VPE ðqÞ=@ q . 0, we see that j@ Dbank =@ qj .j@ Dsec =@ qj. Thus, as q increases, the benefit of bank financing relative to market financing falls faster than does the benefit of securitization relative to market financing. This is intuitive. As observable borrower quality improves, banks need to keep less capital and therefore gather more deposits, which is costly and diminishes the advantage of bank financing relative to market financing. Moreover, the certification benefit of bank origination also declines as observ- able borrower quality goes up; this decrease in the relative advantage of bank financing over market financing also occurs with securitization. However, since the fixed cost of setting up securitization trust (F) is invariant to borrower quality, the advantage of securitization relative to market financing does not decline as fast as the relative advantage of bank financing. Figure 1 depicts these pictorially. Summarizing, we have our first result: Result 1. An authentic borrower chooses its financing source in equilibrium as follows: (1) There exists a low quality cutoff, ql . 0, such that an authentic borrow- er with q , ql cannot obtain financing. (2) There exists a high quality cutoff, qh . ql, such that an authentic bor- rower with q ! qh directly borrows from the capital market and the one with q [ [ql, qh) approaches a bank. (3) There exists a medium quality cutoff, qm [ (ql, qh), such that an authen- tic borrower with q [ [ql, qm) prefers a bank loan and the one with q [ [qm, qh) prefers securitization. In equilibrium, every crook within a prior quality cohort q chooses the same financing source as the authentic borrower in that cohort. Note that the equilibrium above is pooling in the sense that the observation- ally indistinguishable crooks within any observable quality cohort always 502 THE WORLD BANK ECONOMIC REVIEW F I G U R E 1. Borrower’s Choice of Financing Source choose the same financing source as the authentic borrowers in that cohort. This is a sequential equilibrium that satisfies the universal divinity refinement of Banks and Sobel (1987), as proved in Song and Thakor (2010). The basic idea is that the set of pricing responses that would induce defection of the crooks from the equilibrium has a larger measure than the set of responses that would induce defection of the authentic borrowers from the equilibrium. Thus, financiers would view a defector to be more likely to be a crook, and this means the defector would receive a worse price by defecting than staying with the equilibrium. Thus, “hiding in the crowd” is an optimal strategy for the crooks. As for the financing-source-choice dichotomy, the intuition can be summar- ized as follows. Both bank screening and capital market information revelation give rise to noisy signals, and the noise-to-information ratio increases as the borrower’s observable credit quality declines. Thus, when this credit quality (q) is sufficiently low, the borrower is deemed insufficiently creditworthy to receive credit from any source. Borrowers of very high quality go directly to the capital market. These are borrowers who may be viewed as having sufficiently high credit reputations that they do not need bank certification. For example, we know that the highest-credit-quality firms are able to avail of commercial paper as a source of direct market financing, but this is not available to lower- quality firms. The intuition here is when observable credit quality is high, not only is the marginal value of bank certification low, but the incentives for Song and Thakor 503 investors to become informed in the capital market are high, so market prices are very informative to the managers of firms in making real decisions. This quality-cutoff division between bank and market financing also appears in pre- vious research (Bolton and Freixas 2000; Boot and Thakor 1997a). However, our analysis goes a step further to argue that borrowers of intermediate quality prefer to take bank loans that are then securitized. This allows them to take ad- vantage of bank screening and certification, which stimulates information pro- duction in the capital market and hence elevates price informativeness, while at the same time avoiding bank funding which embeds the bank’s cost of comply- ing with capital requirements. One might wonder if the result that it is the borrowers of intermediate quality who are securitized is consistent with the observed securitization of highly-risky subprime mortgages preceding the recent financial crisis. There are two points to note on this. First, the securitization of subprime mortgages is a relatively recent phenomenon, and for much of the time period over which mortgage securitization has existed, only qualifying prime mortgages were securitized, whereas riskier mortgages were not. Moreover, government subsid- ies and various other forms of political intervention played a role in the emer- gence of subprime mortgage securitization, an issue that we explicitly discuss later in our analysis. Second, the term “intermediate quality” is a relative term. These could still be fairly risky borrowers. The point is that there are even riskier borrowers who are served by bank loans that stay on the books of banks. Co-evolution of Banks and Markets Our previous analysis highlights competition between banks and markets. In this section, we discuss a richer form of bank-market interaction and examine collaboration and co-dependence between banks and markets, and how a fi- nancial system evolves as a whole. Our framework shows that markets and banks are linked in the manner of co-dependence through two channels: securitization and risk-sensitive bank capital requirements. Let us explain the intuition. With securitization, the bank certifies a borrower’s credit quality via credit analysis and the capital market finances the borrower, so each sector of the financial system operates at its best – banks focus on credit analysis and markets focus on providing financing. Improvements in the bank’s credit analysis technology, as a result of develop- ment in the banking sector, enhance the capital market investors’ confidence in a securitized borrower’s credit quality. This, in turn, encourages greater informed trading in the capital market, since payoff enhancement is more prob- able with an authentic borrower, which then improves the information conveyed by the market price of securitized debt regarding the payoff- enhancement-relevant macroeconomic condition. The consequence is a lower- ing of the financing friction and a spurring of capital market evolution. That 504 THE WORLD BANK ECONOMIC REVIEW is, securitization propagates banking advances to the capital market, permitting market evolution to be driven by bank evolution. We next examine how development in the capital market affects banks. Note that banks need to raise costly equity capital from the market to support their lending, which forges a natural link between banks and markets. When the capital market develops, there is an improvement in the infrastructure for information acquisition and processing, so it costs less for investors to become informed, leading to an increase in the number of informed traders. With greater informed trading, the informativeness of the price of traded claims increases, which enables the authentic borrower to extract more precise infor- mation about the payoff-enhancement opportunity. Better investment decisions by the borrower result in a lowering of the cost of equity capital for the bank lending to the borrower. This allows the bank to raise additional equity capital (that is mandated by the risk-based capital requirement) to extend riskier loans that were previously eschewed. That is, advances in the capital market enable banks to expand their lending scope (i.e., ql decreases) and lend to riskier bor- rowers that were previously excluded from credit availability. Moreover, as banks lend to riskier borrowers, they find it privately optimal to improve their credit analysis precision to distinguish more accurately between authentic borrowers and crooks, even in the absence of a general im- provement in information technology. That is, development in the capital market feeds the development of the banking sector through risk-sensitive bank capital. This means that it is through (risk-sensitive) bank capital that market advances that diminish the financing friction end up being transmitted to banks, permitting banks to more effectively resolve the certification friction for borrowers and expand their lending scope. These feedback loops generate a virtuous cycle in which each sector of the fi- nancial system benefits from the development of the other, so banks and markets not only act as competitors as Result 1 shows, but also as collabora- tors that complement each other. To summarize, we have: Result 2. Bank evolution spurs capital market evolution, and capital market evolution spurs bank evolution: banks and markets co-evolve with each other. The key to this co-evolution result is two-fold. On the one hand, banks need equity capital to support on-balance-sheet loans, and the capital market is the source. Advances in the capital market make this equity capital cheaper, bene- fiting banks and permitting an expansion in lending scope. On the other hand, banks act as “filters” for the capital market, screening out egregiously bad credit risks and thereby increasing the marginal return to information produc- tion for investors, leading to more efficient security prices and capital market evolution. In this sense, our analysis complements the important observation by Laeven, Levine, and Michalopoulos (2012) about the synergies between finan- cial development and real-sector development. In their analysis, real-sector de- velopment makes financial-sector screening technologies obsolete. So profitable Song and Thakor 505 investments are not financed, unless screening technology innovations allow the financial sector to catch up. The contrast is that we focus on the interaction between banks and markets within the financial system, whereas they focus on the interaction between the financial system as a whole and the real sector. III. POLITICAL INTERVENTION In this section, we examine how political intervention can play a role at various stages of the financial system development. Why Do Politicians Intervene? Objective of Political Intervention Why do politicians want to intervene in the financial system? One reason is that politicians seek to increase their wealth and power.6 This is achieved by increasing the likelihood of getting re-elected, as a longer tenure in office allows for greater accumulation of power and influence and better opportun- ities for wealth accumulation both while holding office and after. In turn, this may be facilitated by increasing the number of borrowers receiving credit from the financial system, which then creates an incentive for politicians to intervene in the financial system to expand credit availability. The idea is as follows. Being able to tout expanded credit availability as a benefit generated for potential voters provides politicians with a way to curry favor with their voter base and increase the odds of being re-elected. Even apart from such private benefits, politicians may be philosophically aligned with the idea that having credit available to a broader set of people is a good social goal. As an example, the idea of universal homeownership for Americans has been appealing to many U.S. politicians, and this may induce them to push banks to expand credit availability to more Americans to enable a higher number of home purchases. Specifically, in our framework the objective of political inter- vention is to expand the lending scope of the financial system, that is, to de- crease the credit exclusion cutoff, ql: recall only borrowers with credit qualities q [ [ql, 1] are able to access the financial system for credit (either from the banking sector or the capital market), so a lower ql corresponds to a wider lending scope. Having said this, we wish to emphasize that there are plausible circumstances in which such a goal is consonant with the goal of maximizing power and wealth. How Do Politicians Intervene? What kind of intervention do politicians engage in? In terms of specific strat- egies for political intervention, we consider two sets of strategies. First, politi- cians may expand the lending scope of the banking sector by providing capital subsidies to banks in exchange for government ownership. The capital infusion 6. A direct, albeit crude, way to do this would be to accept bribes from certain groups and engage in other illegal activities. While this is a realistic alternative specification, we rule it out in our analysis. 506 THE WORLD BANK ECONOMIC REVIEW is subsidized in the sense that the ownership taken by the government is smaller than what its equity infusion would justify in a competitive capital market. Because banks receiving subsidized equity capital have more capital to satisfy risk-based capital requirements, they are willing to lend to riskier (lower-quality) borrowers that were previously excluded. That way, politicians can influence the financial system to expand its lending scope. Alternatively, politicians may enact direct lending legislation that requires banks to expand lending scope and extend credit to riskier borrowers, even if doing so imposes losses on the banking sector. Banks failing to obey the regula- tions face revocation of their licenses to operate. As long as banks are suffi- ciently profitable, the licenses will be valuable and banks will comply.7 To analyze political intervention at different stages of financial system devel- opment, we use the cost of acquiring information about the payoff- enhancement-relevant macroeconomic condition as a measure of the stage of fi- nancial system development, with the cost declining as the financial system develops from a primitive stage to an advance stage. The idea is that this cost is high in the early stage of financial system development because the necessary infrastructure for obtaining information about firms’ investment condition is still primitive and processing information can be quite costly. By contrast, in an advanced financial system, there is considerable information that can be acquired and processed at a relatively low cost. Political Intervention at the Early Stage of Financial System Development At the early stage of financial system development, the high cost of becoming an informed investor leads to a relatively small number of informed traders. With low informed trading, market prices are not very informative, so the bor- rower is more likely to pass up value-enhancing investment opportunities, which in turn drives up the cost of equity capital for the bank lending to the borrower. This relatively high cost of bank equity reduces the bank’s profit, which leads to denial of credit to high-risk (low-quality) borrowers, that is, the bank’s lending scope is narrow at this stage (ql being relatively high). This narrow lending scope prompts politicians to intervene in the banking sector to expand its lending scope. What form of intervention will politicians find optimal? First, consider direct lending regulation in which politicians enact regula- tions that force banks to expand the lending scope without any subsidy. This strategy of intervention is unlikely to work out at the early stage, because banks’ profits from lending are too low to absorb the losses such risky lending imposes on banks. If politicians force banks to lend to low-quality borrowers, 7. Clearly, the intervention strategy space of politicians is much richer in practice than the two strategies we have chosen to analyze. However, we believe those are two commonly deployed strategies, and analyzing a more complex set of strategies becomes analytically intractable. Song and Thakor 507 the loss to the banks may lead them to voluntarily surrender their banking licenses and quit the banking industry. A second approach for politicians is to induce each bank to increase its lending scope to lend to low-quality borrowers who were initially denied credit access (i.e., those with q , ql ) by “compensating” the banks. We consider gov- ernment ownership in exchange for subsidized equity as a tool to achieve this. Specifically, the subsidy manifests itself through a government ownership of the bank that is smaller than what is justified by the amount of equity capital sub- sidized, if the subsidized capital were to be raised from a competitive capital market. Note that equity capital is more costly at the early stage due to less informed trading in the financial market, so each bank will value subsidized equity capital relatively highly. Clearly, a larger capital subsidy corresponds to a larger increase in the banks’ lending scope, which increases the private bene- fits that politicians obtain from broader credit access. What are the consequences of an expanded bank lending scope with a gov- ernment equity subsidy? Will it also induce banks to increase their screening precision and hence spurs capital market evolution as it does without political intervention (see Section II)? To analyze this, note that, without a government equity subsidy, a bank finds it privately optimal to increase its screening preci- sion as it lends to risky, previously excluded borrowers, as the bank bears the entire cost of not being able to screen crooks accurately. With an equity subsidy, the government has part ownership of the bank and hence shares the cost of imprecise screening, so a bank’s incentive to improve its screening preci- sion is weakened. As a result, we do not get the same positive propagation of development from banks to capital markets, and the financial system does not experience the same evolutionary boost, despite its lending scope being expanded. In fact, the larger the government equity subsidy, the less likely it is that the banking sector will improve its screening technology. Thus, political intervention at the early stage of a financial system may retard its development, despite a broader lending scope. Moreover, because the expansion of the banking sector’s lending scope is not accompanied by a concomitant increase in bank screening precision, systemic risk in the financial system increases as well. Political Intervention at the Intermediate Stage of Financial System Development The cost of becoming informed is lower at the intermediate stage, so there are more informed traders in the capital market than in the early stage of develop- ment. Thus, equity capital is less expensive for banks in the intermediate stage than in the early stage. As our previous analysis shows, the financial system’s lending scope is broader (lower ql ) in the intermediate stage. Politicians, however, may want to expand the scope even further. But an equity subsidy is now less attractive to a bank, and for each unit of equity subsidized, the bank responds by increasing its lending scope less in the intermediate stage than in 508 THE WORLD BANK ECONOMIC REVIEW the early stage. An equity subsidy is thus less efficient for politicians in the intermediate stage than in the early stage. Direct regulation may not work either because banks are not yet profitable enough to absorb the accompanying losses. Thus, at the intermediate stage, we see that political intervention is muted compared to the early stage. It is even possible that when equity capital is suffi- ciently less expensive for banks, an equity subsidy from the government may disappear altogether at the intermediate stage. At the same time, if banks’ profits are still not high enough to absorb the loss associated with direct lending regulation, direct lending regulation that mandates lending to high-risk bor- rowers will not be feasible either. That is, there may be no political intervention at all at some specific intermediate stage of financial system development. Political Intervention at the Advanced Stage of Financial System Development At an advanced stage of financial system development, information acquisition in the capital market is the least costly, so the number of informed traders in the capital market is the largest. Thus, politicians view a capital subsidy as even less attractive now than at the intermediate stage. The relatively low cost of bank equity generates relatively high profits for banks in this stage of financial system development. Direct lending regulations that mandate that banks lend to risky, excluded borrowers now become feas- ible. Banks comply because their high profits make their banking licenses valuable. What are the consequences of the expansion of the banking sector’s lending scope at this stage of financial system development? Note that although banks bear the full cost of imprecise screening as they lend to risky, previously- excluded borrowers under the lending regulation, they will not increase their screening precision by investing more in the screening technology. To see why, note that if banks were to improve their screening precision when they are compelled to expand their lending scope, they would have done so voluntarily even without such regulatory pressure. The reason why banks had chosen not to lend to those highly risky ( previously excluded) borrowers in the first place is because the marginal benefit of improved screening is outweighed by its mar- ginal cost. Direct lending regulations compel banks to lend but do not funda- mentally alter this benefit-cost tradeoff. As a result, political invention at the advanced stage of financial system development also increases financial system risk, as it does at the early stage. We can summarize this discussion as follows: Result 3. Political intervention in the financial system is stage dependent: (1) At an early stage of financial system development, when information ac- quisition and processing in the capital market are sufficiently costly so informed trading is sufficiently low and bank equity cost is sufficiently high, politicians intervene by providing subsidized capital to banks. Song and Thakor 509 (2) At an intermediate stage of financial system development, when informa- tion acquisition and processing cost in the capital market increases but banks’ profit is still low, politicians avoid intervening with either equity subsidy or direct lending regulations. (3) At an advanced stage of financial system development, when informa- tion acquisition and processing cost becomes even lower and banks’ profits become higher, politicians intervene by enacting direct lending regulations to force banks to expand lending scope to low-quality bor- rowers that create losses for banks. To capture the intuition for this result in a summary form, note that the choice of political intervention rests on the tradeoff between the cost to the taxpayers/ voters of the chosen intervention and the effectiveness of the intervention in in- fluencing banks to expand lending scope. At an early stage of financial develop- ment, bank lending scope is constrained by the high cost of raising the equity capital needed to support risky lending. Because capital is a costly resource, politicians are able to significantly influence banks by providing it at a subsidy via government ownership of banks. So that is the chosen mode of interven- tion. At an advanced stage, capital is cheap, so offering that as a carrot to banks to expand lending scope is far less effective. But banks are also more profitable at this stage, so a license to operate has high value, and politicians can simply enact regulations requiring banks to engage in risky lending (even at an expected loss) because there is a sufficient profit cushion for each bank to satisfy its participation constraint. It is in the middle – the intermediate stage of development – that capital is not costly enough and bank profits are not high enough to make either intervention strategy viable. I V. P O L I C Y I M P L I C A T I O N S AND CONCLUSION In this paper we have discussed the theoretical underpinnings of financial system evolution. The main result is that banks and capital markets within the financial system exhibit three forms of interaction: competition, complementar- ity and co-evolution. This is in contrast to the standard view that banks and markets only compete with each other. This new perspective is rich in implications about development policy, but as our analysis indicates, we should be careful to distinguish between policy interventions aimed at correcting market failures and those driven by political economy factors. In the past, development agencies have sometimes adopted the view that emerging capitalistic economies, like say Romania and other former Soviet-bloc countries, needed to develop capital markets first in order to have the financial systems needed to support vigorous economic growth (e.g., see Meyendorff and Thakor 2002). In many instances, this led to consid- erable resources being devoted to developing stock exchanges and over-the-counter markets like RASDAQ, the Romanian version of NASDAQ, 510 THE WORLD BANK ECONOMIC REVIEW well before robust financial institutions were developed. What our analysis shows, consistent with the policy conclusions in Meyendorff and Thakor (2002), is that such a policy prescription is misguided. Banks and markets have to be developed together; focusing first on markets and not on banks actually runs the risk of retarding capital market growth itself. With a policy not focused on co-evolution, one may sacrifice the efficacy of both market and banking system development. Thus, in the interest of correcting potential market failure, development policy should focus on the simultaneous develop- ment of banks and markets. What is interesting is that this can also deal effect- ively with political economy forces, since it can make political intervention less likely as politicians would be unable to use the perceived failure of the financial system as a politically convenient reason to intervene. This brings us to our examination of endogenous political intervention in fi- nancial system development. Such intervention can be broadly viewed as being one of two types – either motivated by the desire to correct a market failure or motivated by the selfish interests of politicians. An example of the former mo- tivation is the liquidity-provision intervention by the Federal Reserve during the recent crisis when the private provision of liquidity in the shadow banking system dried up. Such intervention is ex post efficient in the sense that it mini- mizes market-failure-related distortions. This is not the kind of political inter- vention that our analysis focuses on. Specifically, the intervention in our analysis is motivated by the selfish interests of politicians – the desire to (ineffi- ciently) expand credit availability to enhance the re-electability prospects of politicians. We showed that political intervention is just as likely in well- developed financial systems as it is in emerging, underdeveloped systems. In emerging financial systems, politicians provide subsidized equity in exchange for ownership of banks. In well-developed financial system, politicians enact regulations that mandate that banks lend to previously excluded high-risk bor- rowers. In both cases, the effect is the same – systemic risk of the financial system is increased. Thus, if the goal is to improve financial system stability, the discussion cannot proceed without explicitly considering political incentives to undertake initiatives that increase financial system risk. In case this increased systemic risk leads to a financial crisis, there may be an opportunity for polit- ical intervention that is aimed at correcting a market failure. There can thus be an interaction between the two motivations for policy intervention by the gov- ernment. Purely politically motivated intervention can increase the likelihood of intervention aimed at correcting market failures. Also interesting is the potential interaction between policy mis-steps due to the failure to recognize that banks and markets must co-evolve on the one hand and the likelihood and nature of political intervention in the financial system on the other hand, an issue touched upon briefly earlier. This takes us beyond the issues analyzed earlier because the nature of political intervention now becomes less predictable, since it is likely to be targeted at whatever is per- ceived to be a failure in the financial system. This introduces a form of Song and Thakor 511 endogenous uncertainty into the financial system and further elevates systemic risk. A key point emerging from our analysis is that securitization is one of the elements that is responsible for the virtuous loop that positively connects the co-evolution of banks and markets. Thus, if there is any government interven- tion that our analysis supports, it is in the form of encouraging the develop- ment of securitization markets, and not any direct intervention to expand bank lending either through government ownership of banks or direct lending legis- lation that compels banks to lend to otherwise-excluded high-risk borrowers.8 The reason is that the encouragement of securitization helps to improve eco- nomic efficiency by reducing the impact of frictions related to asymmetric in- formation that can lead to market failures (e.g., credit rationing or other forms of credit exclusion due to a limited scope of direct bank lending), whereas direct intervention to expand credit availability is motivated solely by political considerations and comes at the expense of economic efficiency. In other words, politicians need to distinguish between market failures arising from co- ordination failures among profit-maximizing participants – that may be subject to attenuation through selective ex post government intervention – and those that are linked to the endogenous uncertainty created by government mis-steps. We have not considered the impact of the judicial system on the interactions we have examined, so a few remarks on that are appropriate. As La Porta, Lopez-de-Silanes, Shleifer, and Vishny (1998) have shown, legal systems and the implied property rights have important consequences for finance and growth. There are at least two major issues here. One is the impact of the judi- cial system on how much information about potential borrowers is available in the public domain. Laws differ across countries when it comes to how much credit information lenders, credit reporting bureaus and others can share about borrowers. Our analysis suggests that the more the judicial system permits the release and dissemination of such information, the smaller will be the (screen- ing) advantage of banks over markets. Hence, capital markets will experience a greater share of total credit extension in economies with judicial systems of this sort. The other issue is creditor versus borrower (shareholder) rights. Again, legal systems differ in the extent to which they favor one group over the other (La Porta, Lopez-de-Silanes, Shleifer, and Vishny 1998). Since the economic forces that propel banks and markets to develop further originate from cred- itors in our model, the analysis hints at the superiority of legal systems that favor creditor rights, at least in terms of correcting market failures and facilitat- ing bank-market co-evolution. However, this is no more than a hint, so it would be an interesting one to examine in future research. 8. Clearly, our model does not argue that ex post efficient government interventions during financial crises are not worthwhile. 512 THE WORLD BANK ECONOMIC REVIEW REFERENCES Allen, Franklin. 1990. “The Market for Information and the Origin of Financial Intermediation.” Journal of Financial Intermediation 1(1): 3– 30. Allen, Franklin, and Douglas Gale. 1997. “Financial Markets, Intermediaries, and Intertemporal Smoothing.” Journal of Political Economy 105(3): 523 –46. ———. 1999. “Diversity of Opinion and Financing of New Technologies.” Journal of Financial Intermediation 8(1 –2): 68–89. Banks, Jeffrey, and Joel Sobel. 1987. “Equilibrium Selection in Signaling Games.” Econometrica 55(3): 647 –61. Beck, Thorsten, and Ross Levine. 2002. “Industry Growth and Capital Allocation: Does Having a Market– or Bank–Based System Matter?” Journal of Financial Economics 64(2): 147–80. Bhattacharya, Sudipto, and Gabriella Chiesa. 1995. “Proprietary Information, Financial Intermediation, and Research Incentives.” Journal of Financial Intermediation 4(4): 328 –57. Bolton, Patrick, and Xavier Freixas. 2000. “Equity, Bonds, and Bank Debt: Capital Structure and Financial Market Equilibrium under Asymmetric Information.” Journal of Political Economy 108(2): 324–51. Boot, Arnoud, and Anjan Thakor. 1997a. “Financial System Architecture.” Review of Financial Studies 10(3): 693– 733. ———. 1997b. “Banking Scope and Financial Innovation.” Review of Financial Studies 10(4): 1099– 131. Boyd, John, and Edward Prescott. 1986. “Financial Intermediary–Coalitions.” Journal of Economic Theory 38(2): 211– 32. Brown, Craig, and Serdar Dinc ¸ . 2005. “The Politics of Bank Failures: Evidence from Emerging Markets.” Quarterly Journal of Economics 120(4): 1413–44. Calomiris, Charles, and Peter Wallison. 2009. “The Last Trillion–Dollar Commitment: The Destruction of Fannie Mae and Freddie Mac.” Journal of Structured Finance 15(1): 71– 80. Coval, Joshua, and Anjan Thakor. 2005. “Financial Intermediation as a Beliefs–Bridge between Optimists and Pessimists.” Journal of Financial Economics 75(3): 535–69. Deidda, Luca, and Bassam Fattouh. 2008. “Banks, Financial Markets and Growth.” Journal of Financial Intermediation 17(1): 6– 36. ¨c Demirgu ¸ -Kunt, Asli, and Ross Levine. 2001. Financial Structure and Economic Growth: A Cross– Country Comparison of Banks, Markets, and Development. Cambridge, MA: The MIT Press. Dewatripont, Mathias, and Eric Maskin. 1995. “Credit and Efficiency in Centralized and Decentralized Economies.” Review of Economic Studies 62(4): 541–55. ¸ , Serdar. 2005. “Politicians and Banks: Political Influences on Government–Owned Banks in Dinc Emerging Markets.” Journal of Financial Economics 77(2): 453– 79. Gennaioli, Nicola, Andrei Shleifer, and Robert Vishny. 2011. “A Model of Shadow Banking.” Working Paper, Harvard University, Cambridge, MA. Gertler, Mark. 1988. “Financial Structure and Aggregate Economic Activity: An Overview.” Journal of Money, Credit and Banking 20(3): 559– 88. Gurley, John, and Edward Shaw. 1955. “Financial Aspects of Economic Development.” American Economic Review 45(4): 515– 38. ———. 1956. “Financial Intermediaries and the Saving– Investment Process.” Journal of Finance 11(2): 257 –76. La Porta, Rafael, Florencio Lopez-de-Silanes, and Andrei Shleifer. 2002. “Government Ownership of Banks.” Journal of Finance 57(1): 265–301. La Porta, Rafael, Florencio Lopez-de-Silanes, Andrei Shleifer, and Robert Vishny. 1998. “Law and Finance.” Journal of Political Economy 106(6): 1113– 55. Song and Thakor 513 Laeven, Luc, Ross Levine, and Stelios Michalopoulos. 2012. “Financial Innovation and Endogenous Growth.” Working Paper, Brown University, Providence, RI. Levine, Ross. 2002. “Bank– Based or Market–Based Financial Systems: Which Is Better?” Journal of Financial Intermediation 11(4): 398– 428. ———. 2010. “Financial Innovation Boosts Economic Growth: A Debate with Joseph Stiglitz.” The Economist (February – March). London: The Economist Newspaper Ltd. Levine, Ross, and Sara Zervos. 1998. “Stock Markets, Banks, and Economic Growth.” American Economic Review 88(3): 537–58. Meyendorff, Anna, and Anjan Thakor. 2002. Designing Financial Systems in Transition Economies: Strategies for Reform in Central and Eastern Europe. Cambridge, MA: The MIT Press. Ramakrishnan, Ram, and Anjan Thakor. 1984. “Information Reliability and a Theory of Financial Intermediation.” Review of Economic Studies 51(3): 415– 32. Song, Fenghua, and Anjan Thakor. 2007. “Relationship Banking, Fragility, and the Asset– Liability Matching Problem.” Review of Financial Studies 20(6): 2129–77. ———. 2010. “Financial System Architecture and the Co–evolution of Banks and Capital Markets.” The Economic Journal 120(547): 1021–55. ———. 2011. “Financial Markets, Banks, and Politicians.” Working Paper, Pennsylvania State University, University Park, PA. Stein, Jeremy. 2010. “Securitization, Shadow Banking, and Financial Fragility.” Daedalus 139(4): 41– 51. Stiglitz, Joseph, and Andrew Weiss. 1981. “Credit Rationing in Markets with Imperfect Information.” American Economic Review 71(3): 393– 410. Subrahmanyam, Avanidhar, and Sheridan Titman. 1999. “The Going–Public Decision and the Development of Financial Markets.” Journal of Finance 54(3): 1045– 82. Tadesse, Solomon. 2002. “Financial Architecture and Economic Performance: International Evidence.” Journal of Financial Intermediation 11(4): 429–54. Townsend, Robert. 1979. “Optimal Contracts and Competitive Markets with Costly State Verification.” Journal of Economic Theory 21(2): 265–93. von Thadden, Ernst-Ludwig. 1995. “Long– Term Contracts, Short– Term Investment and Monitoring.” Review of Economic Studies 62(4): 557– 75. Yosha, Oved. 1995. “Information Disclosure Costs and the Choice of Financing Source.” Journal of Financial Intermediation 4(1): 3 – 20. Financial Development: Structure and Dynamics Augusto de la Torre, Erik Feyen, and Alain Ize1 This paper analyzes the process of financial development over the last three to four decades from the perspective of the fundamental frictions (agency and collective) to which economic agents were exposed. A comprehensive statistical benchmarking anal- ysis showed that financial development followed regular dynamics that can be largely explained by the underlying frictions. In particular, the sequencing, returns to scale, and shape of the developmental paths for various types of financial activities—including public debt, banking, insurance, asset management, and capital markets—broadly matched benchmark predictions. Reflecting financial innovation and the dynamic in- teraction between financial and economic development, financial development paths were also found to be strongly dependent on initial conditions. At the same time, policy differences, including the failure to improve the quality of the enabling environ- ment and prevent financial crashes (the dark side of finance), were found to explain a sizable share of the deviations of individual country paths from the benchmarks. JEL codes: G2, G38, O16, O54 What has shaped the process of financial development (FD), and how regular has it been? Has this process followed a single or multiple paths? What were the sequences and shapes of the paths followed by different financial services and activities (as measured by FD indicators) as economies developed? How did policy—whether aimed at strengthening the enabling environment for fi- nancial contracting or ensuring the sustainability of FD—affect these paths? Remarkably, the literature ( particularly the empirical literature) that at- tempts to explain how economic development and financial sector policies jointly affect FD is still nascent.2 The proposition that financial structure is shaped by the efforts of market participants to circumvent and reduce the 1. The authors work for the World Bank as Chief Economist for Latin America and the Caribbean (adelatorre@worldbank.org), Senior Financial Specialist (efeijen@worldbank.org), and Senior Consultant (aize@worldbank.org). The paper benefited from valuable comments by Eduardo Fernandez-Arias and other participants in the December 2010 LAC Flagship Authors’ Workshop. We thank Magali Pinat for excellent research assistance. The views in this paper are entirely those of the authors and do not necessarily represent the views of the World Bank, its executive directors and the countries they represent. A supplemental appendix to this article is available at http://wber. oxfordjournals.org/. 2. See Huang (2005) for a comprehensive review of the literature on the determinants of FD. THE WORLD BANK ECONOMIC REVIEW, VOL. 27, NO. 3, pp. 514– 541 doi:10.1093/wber/lht005 Advance Access Publication February 26, 2013 # The Author 2013. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 514 de la Torre, Feyen, and Ize 515 frictions that hinder financial contracting is of course a familiar one.3 However, few papers have attempted to analyze these frictions in terms of the way they interact and what this interaction may imply for the dynamics of FD.4 At the same time, FD has typically been understood as a relatively smooth and predictable march from “relationship-based finance” to “arm’s- length finance” involving a systematic process of market completion driven by a gradual reduction of agency frictions.5 However, the global financial crisis showed that FD has a “dark side” associated with volatile boom-bust cycles. Thus, what may appear as progress in market completion can, in fact, exacer- bate market failures, thereby undermining the sustainability of FD. This paper begins by exploring and explaining the patterns of FD based on a simple typology of the frictions that hinder financial contracting. Following work by de la Torre and Ize (2010, 2011), the paper separates the frictions into agency frictions, which restrict the scope of bilateral contracting, and col- lective frictions, which restrict the scope of multilateral interaction and partici- pation.6 The structure of the financial system and its evolution over time (hence FD) reflect efforts to find the path of least resistance around these fric- tions. This hypothesis leads to broad predictions regarding the sequencing of fi- nancial activities, the shape of their development paths, the importance of scale effects, and the role of public policy. The hypothesis also suggests that the same easing of frictions that underlies the “bright side” of FD (innovation, market completion, improved allocation of funds and risks) can breed the ten- sions and fault lines of the “dark side” (systemic fragility, excessive fluctua- tions, crises). 3. See, for example, Merton and Bodie (2004). 4. Interestingly, more work has been done on measuring the impact of FD on economic growth than on exploring how FD is affected by policy and economic growth (see, for example, Beck and Levine, 2004). When discussing the impact of financial structure on growth, the literature, at least until very recently, generally concluded that function matters more than form. See Demirgu ¸ -Kunt and Levine ¨c (1999) or Allen and Gale (2000). More recent papers (such as Demirgu ¨c¸ -Kunt, Feyen, and Levine, 2011) have come closer to recognizing that there might be such a thing as an “optimal” financial structure, that is, that form might also matter. 5. An earlier strand of thought viewed FD as driven by the steady mitigation of asymmetric information failures, such as moral hazard and adverse selection (see, for instance, Akerlof [1970], Spence [1973], and Stiglitz and Weiss [1981]). A more recent strand has emphasized enforcement costs and lack of collateral leading to problems of limited pledgeability (see Holmstrom and Tirole [1996], and Geanakoplos, [2009]). Rajan and Zingales (2003) present a more complete narrative rooted in the same basic threads. 6. Failures to reduce agency frictions continue to dominate the FD literature. Such failures may occur at the level of the investor (reflecting an inability to monitor or a lack of interest in doing so), at the level of the borrower (reflecting problems of governance), or somewhere in between (reflecting problems of incentives and “skin in the game” at some level of the contractual chain or the monitoring pyramid). Failures to resolve multilateral-participation frictions, although less frequently discussed, are equally important. These failures are a routine occurrence in less-developed financial systems and justify much of the state’s catalytic and financial infrastructure-building role. However, such failures may also occur in well-developed systems, particularly in the process of spreading risk (see Anginer, de la Torre, and Ize [2011]). 516 THE WORLD BANK ECONOMIC REVIEW To explore such issues, we use a battery of 16 financial indicators that relate to different dimensions (depth-size or liquidity-efficiency) and channels (markets or institutions) of FD. To create a uniform set of dynamic bench- marks that facilitates comparisons across countries, we use a comprehensive set of country-level controls consisting of variables that reflect the stage of eco- nomic development (initial and current per capita income) as well as structural variables that are arguably exogenous to financial policy, including population size and density. We find that FD, as measured by the benchmarks, followed regular dynamics. However, development paths depended on the initial level of per capita income. Thus, the lower-income countries did not generally retrace the steps followed in the past by the current higher-income countries. This finding suggests that across-the-board innovations (that lift all boats at the same time) and path dependencies (that reflect dynamic interactions between fi- nancial and economic development) are both relevant for a country’s FD trajec- tory. The fact that the data we use cover only a small slice of the world’s FD history (35 years at most) compounds the importance of such initial income effects. Reflecting the differential impact of frictions on different financial activities, paths differed widely across FD indicators. Indeed, the sequencing of financial services broadly conformed to what one would expect based on the gradual smoothing of the frictions along the paths of least resistance. For example, re- flecting differences in the difficulty of coping with bilateral agency frictions (in- formation or contract enforcement costs) across financial services, credit to governments developed before credit to private participants, bank deposits pre- ceded bank credit, and the development of capital markets and associated insti- tutions, such as mutual funds and other forms of asset managers, followed the development of bank credit. At the same time, reflecting collective frictions and network effects, external funding of the government preceded domestic funding, retail funding preceded wholesale funding, casualty insurance preced- ed life insurance, the development of capital markets lagged behind and was highly convex (a manifestation of large scale and network effects), and inter- connectedness increased significantly as financial systems matured. We further explore the dynamics of FD based on an analysis of how and why countries deviated from their benchmark paths. For this exploration, we use as additional controls a number of policy-related variables proxied by various enabling-environment indicators. We also include a financial crash vari- able to capture the dark side of FD. We find that enabling-environment factors, such as enforcement costs, creditor or property rights, and credit information, played an important role in promoting FD. However, we find that contractual frictions (rather than informational frictions) explained the bulk of policy- induced developmental differences across countries. Although informational frictions could be mostly overcome through technological innovations that could be more easily imported, contractual frictions mostly reflected weakness- es in local (nonimportable) institutions that were more difficult to resolve. de la Torre, Feyen, and Ize 517 Regarding booms and busts, we find that financial crashes accounted for large and lasting lags in FD that cut across a surprisingly large number of indicators and dimensions of FD. The lasting impact of financial crashes clearly puts a premium on central bankers’ and supervisors’ capacity to anticipate and control unsustainable booms through monetary or macroprudential policies. The rest of the paper is organized as follows. Section 2 briefly presents the conceptual framework for financial frictions that underpins FD (the bright side of finance) and links it with financial instability (the dark side of finance). Section 3 describes the data and procedures that are used to estimate the FD paths. Section 4 presents the dynamics of FD as given by the benchmark paths. Section 5 explores the role of policy in explaining deviations from the bench- marks. Section 6 concludes. I . F I N A N C I A L F R I C T I O N S , F I N A N C I A L D E V E LO P M E N T, AND F I N A N C I A L S TA B I L I T Y The financial services industry emerges to help market participants find ways to reduce or circumvent the two classes of frictions that hinder financial contracting: bilateral agency frictions and multilateral collective frictions. Each class can be sub- divided into two categories: informational frictions (which relate to agents’ limited and often asymmetric capacity to understand information and to the costs of ob- taining information) and relational frictions (which hinder agents’ capacity to agree, act upon, and enforce bilaterally or collectively beneficial financial contracts). This simple dichotomy underpins four paradigms, of which two (asymmetric infor- mation and costly enforcement) are associated with agency frictions and two (col- lective action and collective cognition) are associated with collective frictions.7 Asymmetric information frictions hinder FD because they lead to a misalign- ment of incentives between the principal and the agent. This misalignment, in turn, can trigger the commonly known market failures of adverse selection, risk shifting, shirking, and false reporting. Thus, information asymmetry frictions limit financial contracts to those contracts in which the agent has sufficient re- sources of its own at risk (“skin in the game”) and/or where the principal can adequately screen and monitor the agent. Enforcement frictions also lead to a misalignment of incentives between the principal and the agent and, in this case, because of imperfect pledgeability, to a situation in which the agent is unable to credibly commit to honor the contract. Imperfect pledgeability thus restricts financial contracts to those contracts that can be effectively collateralized. Collective frictions hinder FD by constraining collective (rather than bilat- eral) participation. Much of the gains from financial activity relate to the re- duction in transaction costs and the increase in liquidity and risk diversification benefits that result from multilateral arrangements in which many players 7. For more details on the paradigms, see de la Torre and Ize (2010, 2011). 518 THE WORLD BANK ECONOMIC REVIEW participate. Such arrangements can either take place in markets, where transac- tions can be conducted around a trading platform, or through financial institu- tions that offer services whose benefits are pooled across a large number of customers. The higher the number of participants is, the higher the benefits of participation will be. However, although participation creates positive external- ities for society at large, it may be hindered by coordination failures. For example, because of first-mover disadvantages, an investor may abstain from buying a long-term security for which there is no secondary market and, hence, no liquidity. In this case, there is a multilateral (individual versus social) misalignment of incentives. Participation may also be hindered by collective (even symmetric) cognition frictions; one does not participate in an activity that is not well understood. Therefore, market participants who wish to engage in financial contracting must find the path of least resistance around these frictions and the associated market failures. Once a decision is made to participate, private responses to coping with frictions can be divided into two subsets: responses aimed at less- ening the frictions themselves (acquiring information, using collateral, delegat- ing) and responses aimed at lessening the exposure to these frictions (diversifying and pooling risk, buying insurance and hedges, staying liquid). In turn, the state facilitates these private responses through a set of progressively more intrusive public interventions: (i) the provision of a basic contractual and informational infrastructure that facilitates contracting; (ii) coordination ar- rangements that facilitate participation (catalytic involvement to promote market development, systemic lending of last resort, government guarantees); (iii) the regulation and taxation needed to internalize externalities or protect consumers; and (iv) the direct provision by the state of financial services. Financial structure is a snapshot, at a given point in time (hence, for a given technological and state of market development), of the composition of financial services aimed at coping with financial frictions. FD reflects the evolution of fi- nancial structure over time. At lower stages of FD, financial institutions resolve agency frictions by relying on nontradable and immovable collateral and relationship-based transactions. However, as the informational and contractual environment improves, private information becomes public, other types of col- lateral become available and tradable, and monitoring can increasingly rely on third parties, statistical methods, and accounting and disclosure standards.8 8. Different components of the financial system help address frictions in different ways. Consider information frictions: Capital markets provide price signals and motivate the supply of hard, public information by borrowing firms; banks generate proprietary information on clients; fund managers contribute to information gathering by monitoring marketable assets; and market facilitators (auditors, rating agencies, credit bureaus) contribute by selling specialized information and analysis. Consider risk management: Capital markets allow investors to diversify risk by buying assets with different risk profiles, and banks, insurance companies, and asset managers provide diversification through pooling. Finally, consider liquidity: Capital markets provide the liquidity that allows participants to unwind assets at limited cost, and banks offer deposits that can be redeemed on demand and at par. de la Torre, Feyen, and Ize 519 The gradual easing of agency frictions thus helps boost participation, which unleashes positive network and scale externalities (e.g., liquidity, learning spill- overs, efficiency) and sets in motion a virtuous development cycle. In turn, rising participation gradually increases the degree to which financial institu- tions and capital markets complement each other. The entire process is acceler- ated by financial innovation.9 The gradual easing of finance frictions provides a few predictable regularities that can be empirically verified. The first regularity concerns the sequence in which various financial activities are likely to develop, the second regularity concerns the shape of the paths they are likely to follow once they begin, and the third concerns the volatility of FD, reflecting the seeds of financial instabili- ty that germinate while FD takes place and that may eventually develop into a financial crisis. The order of development of financial activities should reflect the intensi- ty of the frictions to which they are exposed. The activities that are the least constrained should develop first. However, those financial services that are strongly inhibited by collective frictions should develop only after a criti- cal mass of participation has been reached that is sufficient to trigger the positive network externalities needed to sustain their development. We would thus expect participation-intensive financial activities to be those that have the most rapid development paths (the most “buoyant” or convex) once they have passed this initial threshold. Because activities that exhibit the highest returns to scale are expected to be those exposed to the highest collective frictions, scale effects should correlate with the order in which financial activities develop and, subsequently, the buoyancy of their development.10 9. The history of FD is marked by major waves of innovation. Consider, for instance, the role in the exponential ascent of finance in the Western world stemming from the invention of Italian banking (based originally on trade-related bills of exchange) by the Medici in the late 14th century; the introduction of payment systems based on checking accounts, fractional reserve banking, and central banking during the 17th century; the development of the government bond market, its seeds already visible in the late middle ages; the invention of the joint-stock, limited liability company in the early 17th century and the associated mushrooming of stock exchanges; the emergence of marine insurance and life insurance in the second half of the 17th century; or, in the latter part of the 20th century, the development of securitization and derivative products. For an insightful and entertaining rendition of the history of finance in the Western world, see Ferguson (2008). For a recently updated review of the roots and dynamics of financial innovation, see Lerner and Tufano (2011). For the role of competition and deregulation in FD, see Rajan and Zingales (2003). Examples of theoretical and methodological breakthroughs that have dramatically influenced FD include double-entry bookkeeping, probability theory, life expectancy tables and actuarial science, and the Black-Scholes option theory. 10. The pattern can be broadly corroborated through comparative historical studies. The literature on the history of finance in the Western world is vast. See, for instance, Ferguson (2008) and Rajan and Zingales (2003). 520 THE WORLD BANK ECONOMIC REVIEW However, the same frictions that underlie FD (the bright side of finance) are also likely to be at the root of its dark side (financial instability). We define the “dark side” as the type of financial malady in which the actual success in the process of smoothing frictions can endogenously lead to systemic instability.11 This type of adverse dynamic may ensue if the easing of frictions boosts partici- pation in such a way that it unleashes new and possibly more severe forms of frictions. FD may thus advance along a fragile or self-destructive trajectory that may lead to large and lasting reversals in developmental indices. The dark side may arise at the interface between agency and collective action frictions. The smoothing of agency frictions facilitates a switch from private to public information, which, in turn, promotes participation but leads to a socially insufficient supply of screening and monitoring. Investors free ride on the increased availability of public information and choose to remain short and rely on market liquidity and public information to exit at the first sign of possible trouble.12 The boost in participation can also promote risk origination without sufficient “skin in the game.” As observed in the global crisis, this sit- uation can give rise to a complex and opaque chain of transactions, which is ultimately unstable.13 This situation may also arise from dynamics associated with the intensifica- tion of collective action problems. The positive externalities of increased market participation in good times may turn into crippling negative externalities in bad times. Thus, market withdrawal in times of stress may be individually optimal but socially harmful because it can trigger self-fulfilling liquidity collapses.14 Financial institutions may become too large (from a social point of view) because participants may not internalize the negative externalities (e.g., domino effects and contagion) associated with the failure of systemically important fi- nancial institutions and the too-big-to-fail or too-interconnected-to-fail syn- dromes that are associated with these failures. The dark side may also be associated with the swelling of collective cogni- tion problems. The successful easing of frictions (including through innovation) may lead to problems of collective cognition that, in a world of collective un- certainty, may result in wide mood swings. The bonanza associated with en- hanced participation feeds a collective mood of optimism that unleashes bouts 11. There are other types of FD maladies. For instance, the lack of success in smoothing the frictions that make some forms of financial contracting impossible can itself be considered a malady. The success in smoothing frictions that leads to developmental inefficiencies can also be considered a malady. For example, financial innovations that are designed only to evade taxes or regulations may be beneficial to their creators and users, but not to society at large. Indeed, much of the increase in funds mobilization and allocation prior to the subprime crisis was arguably socially wasteful, even if it did not have adverse systemic stability consequences (Haldane, 2010). We are indebted to one of the referees for noting this point. 12. Huang and Ratnovski (2010) show that the dark side of bank wholesale funding dominates if bank assets are at arm’s length and tradable. 13. See Ashcraft and Schuerman (2008) and Gorton and Metrick (2010). 14. See Shleifer and Vishny (2011). de la Torre, Feyen, and Ize 521 of exuberance. The euphoric mood initially accentuates the upswing; however, once unexpected realities emerge, euphoria can easily turn into despair, wors- ening the collapse.15 I I . D ATA AND ESTI MATION PROC ED URES In this section, we conduct a simple empirical analysis of FD indicators. Our aim is to illustrate the above conceptual framework by showing that the se- quence of development of various financial activities and the shape of the path that they follow as they develop is consistent with (and thus validates, at least broadly) the broad footprints described above. We measure domestic FD based on 14 depth-size indicators that include the following:16 † Key components of commercial banks’ operations (retail and wholesale funding, credit to the private sector, credit to government, and claims on other domestic financial institutions) † Insurance company premia (life and casualty) † Mutual fund and pension fund assets † Public and private debt securities (domestic and international capitalization) † Equity (total stock market capitalization) We complement these indicators with two indicators of efficiency-liquidity for which there is sufficient cross-country data: † Banks’ net interest margin † Equity market turnover To make the data as comparable as possible across countries, we create FD benchmark paths for each FD indicator by controlling for structural factors that can be considered policy exogenous (at least in the short term), including population (size and density) and three other country-specific characteristics that, for a given level of economic development, have a sizable impact ( positive or negative) on FD (fuel exporter, offshore financial center, transition 15. The South Sea bubble and panic of 1720, which materialized in the wake of the financial revolution of the late 18th century, is an early example of bright side leading to dark side. The importance of mood swings for financial bubbles and panics has its roots in Keynes’s animal spirits and Hyman Minsky’s writings on financial crises (see Minsky 1975). More recently, the importance of mood swings was popularized by Kindleberger (1989) and Shiller (2006). 16. The data are from FinStats, a worldwide financial database compiled by the World Bank, which covers 40 key financial indicators for the 1980– 2010 period (coverage quality varies between variables). The data come from a variety of sources, including the IMF’s International Financial Statistics, BIS, WDI, S&P, Bankscope, Axco, and national sources. 522 THE WORLD BANK ECONOMIC REVIEW country).17 We also control for economic development, as proxied by per capita GDP. Although the latter is clearly not exogenous to FD, this fact is not a significant concern for what we want to achieve here.18 Reflecting that FD can be expected to be generally path dependent, we also include as controls the initial level of GDP per capita (as measured for the ear- liest date for which financial statistics are available) and its interaction with current GDP per capita. By introducing a level effect, including initial GDP per capita allows countries at different initial levels of economic development to follow their own dynamic FD path. The interaction term, for its part, allows the slope of this dynamic FD path to vary with the level of initial economic development. As shown in the supplemental appendix S.II, this specification is key to properly representing the dynamics of FD, whether they reflect the impact of fi- nancial innovations or the interactions between economic and FD. Financial in- novations are typically transferable across countries and thus tend to be introduced across the board (in low- as well as high-income countries)—that is, financial innovations can cause a boost in financial activity that “lifts all boats at the same time,” with dynamic paths suddenly surging above the cross- section (benchmark) line at all per capita income levels in a parallel fashion.19 In addition to allowing for such parallel surges, including the initial income level in the estimating regressions also helps to capture path dependence. Initial conditions matter because today’s FD depends on today’s output, which in turn depends on yesterday’s FD. Path dependence thus implies that better initial institutions can become self-reinforcing.20 To facilitate the interpretation of the coefficients, we decompose the income effects into an initial income effect, yi i i 0, and an income-growth effect, yt – y0 (all 21 in logs). The estimating equation is as follows: 17. The controls were selected iteratively, based on individual statistical significance and collective explanatory power. 18. Detecting FD patterns does not require identifying and isolating the ways in which the dynamics of financial and economic development interact. In addition, as long as FD affects economic development with a longer lag than the other way around, financial sector policies will have at least a temporary impact on FD that is not fully captured by economic development. If so, generating a benchmark path for each FD indicator and comparing countries against the benchmark is informative in terms of the quality of FD policies. Thus, deviations from benchmarks can be at least partially interpreted as reflecting differences in policies and (policy-driven) institutions. See supplemental appendix S.I (available at http://wber.oxfordjournals.org/) for a formalization of this argument. 19. Consider, for example, the cases of credit card services and e-banking. These services are now found in most developing countries, and although they cover a smaller fraction of the adult population, they work with comparable functionality and quality as in rich countries. In both cases, developing countries have been able to leapfrog because the associated technology is relatively easy to import and adapt and the services do not depend significantly on local contractual institutions. 20. See North (1990). 21. To better capture the underlying FD patterns, we employ quantile (median) regressions, which are less influenced by outliers. de la Torre, Feyen, and Ize 523 À ÁÀ Á ð1Þ FDit;j ¼ aj0 þ aj1 yi0 þ aj2 þ aj3 yi0 yit À yi0 þ aj4 sit þ Vj0 Xit þ Vj1 Zit þ 1jt where FDi,j i t is the (log of ) the indicator j of FD for country i at time t, st is the i (log of ) the country’s population size, Xt is a vector of other country-specific structural characteristics, and Zi t is a vector of policy variables. The a1 coefficient measures the elasticity of FD with respect to the country’s initial per capita income. The higher this elasticity is, the more dependent the development of the financial indicator (or financial activity) is on the country’s initial level of economic development (per capita income). Thus, this coeffi- cient can be viewed as a proxy for sequencing: The activities that are more income elastic develop “later” than those activities that are less income elastic. Similarly, the a2+a3y0 coefficient measures the elasticity of FD with respect to the country’s per capita income growth.22 The higher this elasticity is, the faster the activity develops with the country’s economic development. Thus, growth elasticity characterizes the “buoyancy” of the financial-activity path. Finally, the a4 coefficient measures returns to scale (elasticity relative to popu- lation size). Activities with higher returns to scale prosper in larger countries. We run this regression in two stages. First, to fully capture the longer-run dynamics of FD and thus obtain a more meaningful benchmark, we estimate the regression over the whole panel but without policy and institutional vari- ables. The results of this first stage are discussed in section 3. In the second stage, we reestimate the equation with policy-related institutional variables to capture their specific impact on FD. This second stage, whose results are exam- ined in section 4, is conducted over a narrower sample because of limitations in the data on enabling-environment indicators. I I I . F D G ROW T H PAT T E R N S Table 1 reports the results of the first-stage regressions over the full sample. Note first the impact of the structural controls. Reflecting returns to scale and network effects, FD tended to lag in countries with smaller or more dispersed populations. In contrast, reflecting uneven growth across sectors, FD in oil- exporting economies lagged behind FD in other economies at similar levels of income. Similarly, FD lagged behind in transition economies, in this case re- flecting a pretransition economic system that did not favor FD. Inversely, and for obvious reasons, FD in offshore centers generally led FD in other countries. Table 2 shows the variance decomposition of these estimates, expressed by groups of variables, in terms of simple averages (of absolute values) over all in- dicators as well as ranges (maxima and minima). The initial level of economic development ( per capita income) accounts for the bulk (nearly half ) of the 22. The coefficient is measured for, y0 ¼  y0 is the median initial income level for the entire y0 where  country sample. 524 T A B L E 1 . Basic Benchmark Regressions Bank Bank Claims Bank Credit Bank Bank Insurance Insurance Private Net Interest On Dom. To Domestic Non-Deposit Premiums Premiums THE WORLD BANK ECONOMIC REVIEW Credit Margin Fin. Sector Government Deposits Funding (Life) (Non-Life) 1 2 3 4 5 6 7 8 Panel A Log Initial GDPPC 0.372*** 2 0.261*** 0.822*** 0.285*** 0.288*** 0.380*** 0.619*** 0.267*** Log GDPPC minus Log 0.840*** 0.120 2 0.286 1.634*** 1.535*** 2 0.271 0.745** 2 0.155 Initial GDPPC Interaction 7.95e205 2 0.0837*** 0.223*** 2 0.183*** 2 0.0964*** 0.146*** 0.133*** 0.0686*** Log Population 0.0721*** 2 0.0660*** 0.243*** 0.0940*** 0.0367*** 0.0717*** 0.0424** 2 0.0496*** Log Population density 0.0193*** 2 0.0293*** 0.339*** 0.200*** 0.0870*** 0.0452*** 0.0999*** 2 0.0403*** Fuel dummy 2 0.272*** 0.00729 2 0.256*** 2 0.262*** 2 0.163*** 2 0.0551 2 0.687*** 2 0.202*** Offshore dummy 0.331*** 0.105** 2 0.634*** 0.166*** 0.333*** 0.428*** 2 0.130 0.107** Transition dummy 2 0.0350 0.187*** 2 0.102 2 0.0864 2 0.170*** 0.220*** 2 0.779*** 2 0.0863* Constant 0.285*** 3.709*** 2 8.413*** 2 1.285*** 0.815*** 2 0.441*** 2 6.126*** 2 1.708*** Observations 4,075 1,785 1,643 4,003 4,097 3,983 2,138 2,308 Pseudo R2 0.388 0.294 0.247 0.141 0.401 0.285 0.384 0.357 Panel B Log initial GDPPC 0.317*** 0.734*** 0.672*** 0.415*** 1.010*** 0.159*** 1.030*** 2 0.134*** Log GDPPC minus 2 3.501** 2 1.423*** 1.853*** 0.253 2.332*** 0.0639 2 0.426 2 2.144*** Log initial GDPPC Interaction 0.566*** 0.472*** 2 0.0452 0.0900 2 0.0578 2 0.0207 0.239*** 0.167** Log population 2 0.0994 0.135*** 0.462*** 0.118*** 0.112*** 0.0973*** 0.122*** 2 0.243*** Log population density 2 0.152*** 0.00934 0.0661*** 0.0756*** 2 0.131*** 0.0571*** 2 0.0115 2 0.253*** Fuel dummy 0.360** 2 0.224** 2 0.0575 0.0716 2 0.785*** 2 0.357*** 0.0507 2 0.00290 Offshore dummy 2 0.278 0.960*** 2 0.592*** 0.391*** 2 0.0158 2 0.345*** 0.150 2 0.0280 Transition dummy 2 1.834*** 2 1.421*** 0.722*** 2 0.669*** 2 0.504** 2 0.118 2 0.499*** 2 0.474*** Constant 0.247 2 5.554*** 2 4.359*** 2 0.975*** 2 6.488*** 1.579*** 2 7.983*** 4.601*** Observations 568 613 1,682 1,818 889 978 985 1,198 Pseudo R2 0.169 0.383 0.375 0.274 0.353 0.0808 0.382 0.138 Notes: This table displays the median regression results of equation (1) using a panel of country-year data for the 1980– 2010 period. GDPPC stands for gross domestic product per capita. Source: Authors’ analysis based on data described in the text. T A B L E 2 . Variance Decomposition of the Benchmark Regressions Current Private Strength of Credit Strength of % of total Log initial adjusted Population Other structural credit legal rights information investor Enforcement variance GDPPC GDPPC size characteristics crash index index protection index costs Residual Without policy variables Average 22 15 4 7 .. .. .. .. .. 52 Min. 1 4 0 1 .. .. .. .. .. 32 Max. 44 33 24 17 .. .. .. .. .. 80 With policy variables Average 18.4 17.9 4.1 6.8 4.4 3.3 2 1.7 1.3 40.2 Min. 0.5 5.5 0 1 0.5 0 0.1 0.4 0 23.5 Max. 42 32.8 25.7 21.2 18.7 18.2 4.3 4.8 3.3 63.9 de la Torre, Feyen, and Ize Notes: This table provides a variance decomposition (using the regression coefficients from tables 1 and 4), calculated for indicator F as VAR(F) ¼ SbCOV(F, X ) þ COV(F,1). Other structural characteristics refer to the joint variance component of population density, fuel exporters, offshore centers, and the transition dummy. GDPPC stands for gross domestic product per capita. Source: Authors’ analysis based on data described in the text. 525 526 THE WORLD BANK ECONOMIC REVIEW explained cross-country variations in FD, which, in turn, account for approxi- mately half of the total variance. Thus, much of the cross-country FD pattern is simply explained by initial conditions, which, in an important sense, summa- rize the entire history of the economic and FD of a given country prior to the first observation in the sample period. However, countries’ economic growth thereafter and over the sample period explains a substantial portion (approxi- mately 30 percent) of the explained portion of FD paths, with population size and other structural controls accounting for the rest. Figure 1 orders financial indicators in accordance with their elasticities with respect to initial per capita income and contrasts these elasticities with the growth and returns-to-scale elasticities. Panel C of figure 1 shows the matrix of cross-correlations across indicators. As a whole, figure 1 broadly validates one of the three main predictions of section 2: Sequencing, buoyancy, and returns to scale are all substantially correlated. The correlation between sequencing and buoyancy is particularly high; the later an activity develops, the faster it develops once it takes off. This result is expected once strong collective ( partic- ipation) frictions are overcome. The positive correlation between sequencing and returns to scale also confirms that returns to scale relate to collective fric- tions. Larger returns to scale require higher critical mass for an activity to develop, and thus, it is more difficult for this activity to develop in the presence of substantial collective frictions. To further test the hypothesis that sequencing, buoyancy, and returns to scale are interrelated, we group the indicators according to their initial income, income growth, and size elasticities (table 3). Regarding income, the high corre- lation between initial income and income growth elasticities simplifies this task because all FD indicators can be placed in three groups (column headings of table 3): early developers with low buoyancy ða1 , 0; a2 þ a3  y0 , À0:1Þ, middle developers with average buoyancy ða1 [ ½0; 0:5Š; a2 þ a3 y0 [ ½À0:1; 1ŠÞ,  and late developers with high buoyancy ða1 . :5; a2 þ a3  y0 . 1Þ. Financial in- dicators in the first group developed early but declined in importance as income grew. In contrast, the third group appeared late but grew very quickly. The middle group was somewhere in between. Regarding size, we divide activities into those with negative returns to scale (a4 , 0), low returns to scale ða4 [ ½0; 0:1ŠÞ, and high returns to scale (a4 . .1) (row headings of table 3). Most financial activities (12 out of 16) cluster in the diagonal cells of the table, fully matching the predictions of section 3. That is, the later the initial level of economic development at which a financial activity developed is, the faster it developed as income grew and as it was increasingly subject to increasing returns to scale. The financial services above the diagonal or below the diagonal do not fully match the predictions of section 3. Above the diagonal are financial services that, given their low returns to scale, de la Torre, Feyen, and Ize 527 F I G U R E 1. Sequencing, Buoyancy, and Returns to Scale Notes: On the basis of equation (1), this figure present elasticities of initial income (a1) and y0), where  population size (a4) – panel A-, and “buoyancy” (a2 þ a3 y0 is the median of the initial income distribution – panel B-. The buoyancy elasticity, which depends on initial income, measures the financial development response to GDP per capita growth. Panel C contains the pairwise correlations between these three series. 528 THE WORLD BANK ECONOMIC REVIEW T A B L E 3 . Typology of Financial Activities Based on Sequencing and Returns to Scale Early developers Middle developers Late developers y0 , À0:1 a1 , 0; a2 þ a3  y0 [ ½À0:1; 1Š a1 [ ½0; 0:5Š; a2 þ a3  y0 . 1 a1 . 0:5; a2 þ a3  Negative returns to scale (a4 , 0) Foreign public debt Casualty insurance Pension fund assets* Low returns to scale ða4 [ ½0:0:1ŠÞ Domestic public debt Life insurance Bank credit to government Bank private credit Bank retail funding Bank wholesale funding Net interest margins High returns to scale (a4 . 0.1) Stock market capitalization Domestic private debt Stock market turnover Bank claims on financial sector Mutual fund assets Foreign private debt Notes: This table presents a taxonomy of financial activities grouped along two dimensions: (1) sensitivity to population size (returns to scale) and (2) sensitivity to GDP per capita (both initial income and income growth effects). The returns to scale dimension is based on elasticity a4 from equation (1). y0 ), where  Sensitivity to GDP per capita is based on the initial income elasticity (a1) and the buoyancy term (a2 þ a3  y0 is the median of the initial income distribution. The regression coefficients are taken from table 2. The Net interest margin coefficients were multiplied by 2 1 to ensure that an increasing value signifies a financial development improvement. * indicates not significantly different from zero. de la Torre, Feyen, and Ize 529 developed later than expected (casualty insurance) or more buoyantly than ex- pected (life insurance).23 Below the diagonal is one financial indicator (stock market capitalization) that, considering its high returns to scale, developed earlier than expected. However, these three FD indicators that deviate from predictions are only mildly deviant (there are no activities in the top-right cell or in the bottom-left cell of table 3), and the reasons underlying their deviancy can be easily identified, as discussed below. To better comprehend the process of FD, we examine each financial activity by grouping them into four categories (government borrowing, banking servic- es, capital markets, and institutional investors) that are broadly ordered accord- ing to the sequencing of their development. To illustrate the dynamics, we chart the predicted growth trajectories followed by some of the indicators for individual countries as their income grew over their initial income level (figures 2 and 3). We plot the initial and current income on the horizontal axis and the value of the indicator on the vertical axis. The continuous lines corre- spond to the initial cross-country benchmarks (i.e., the projected values of the indicators when initial income varies across countries while current income remains equal to initial income). The dotted lines correspond to the expected country-specific FD trajectories, given its initial income (i.e., the values of an FD indicator projected for a country as its per capita income grows above its initial income).24 Government Borrowing Somewhat surprisingly, public-sector borrowing was the financial activity that developed the earliest. Furthermore, public-sector borrowing developed in in- ternational markets before it did at home (figure 1). These features are easy to explain based on frictions. Public-sector borrowing developed early because sovereigns are well known; hence, agency frictions were comparatively easier to overcome.25 However, public-sector borrowing developed abroad first because well-established international markets, by definition, have overcome the rele- vant collective frictions. Hence, it is not surprising that government borrowing initially took place abroad and in foreign currency.26 As frictions eroded and countries’ FD deepened, governments were able to substitute external 23. The indicator for pension fund assets also appears above the diagonal; however, its size elasticity is not significantly different from zero. 24. In both cases, to better reveal the income dynamics, the values of all structural controls are kept equal to their median values over the whole sample. 25. Nonetheless, governments’ capacity to issue debt may be subject to establishing minimum credibility as a debtor. See, for example, Dickson (1967) for a discussion of the measures taken by the English monarchy to bolster the credibility of its debt in the late 18th century. 26. This is, of course, the basic premise of the original sin literature, which focuses on the inability of emerging economy sovereigns and corporates to issue long-term debt denominated in local currency (Calvo and Reinhart [2002]; Eichengreen, Hausmann, and Panizza [2003]). The fact that a better foreign institutional framework facilitates enforcement (especially postdefault value recovery) is an important driver of the “original sin” story (De la Torre and Schmukler [2004]). 530 THE WORLD BANK ECONOMIC REVIEW F I G U R E 2. Estimated Development paths for Government Funding Notes: The figure presents the development paths of bank credit to government and state- owned entities (A) and international public debt (B) (both as a percentage of GDP), based on equation (1). The initial income path line is plotted against log initial income and is constructed by multiplying initial income by its estimated coefficient and adding a constant term to preserve scale. This constant is calculated as the sum of the sample median values of all right-hand variables in equation (1), with the exception of the estimated growth term, multiplied by the associated estimated coefficients and the estimated constant. The complete path is plotted against log GDP per capita and is composed of the predicted country-year values from equation (1). borrowing with domestic borrowing. However, reflecting collective- participation frictions, governments typically started by borrowing from local banks instead of markets (figure 2). Thus, domestic public debt developed later than domestic bank credit to government or foreign public debt. de la Torre, Feyen, and Ize 531 F I G U R E 3. Estimated Development Paths for Bank Intermediation Note: This figure uses the same derivations as figure 2 as applied to bank credit to the private, nonfinancial sector (A) and bank claims on other financial institutions (B). Banking Services At the initial stages of FD, the lack of public information, a sound contractual framework, and tradable collateral restricts lending to mostly relationship lending based on private information rather than at arm’s length (based on public information). Hence, banks would be expected to develop relatively 532 THE WORLD BANK ECONOMIC REVIEW early, ahead of capital markets or asset managers such as mutual funds.27 The fact that banks can also solve collective frictions efficiently through liquidity pooling and the provision of basic payment services further enhances their early attractiveness. Indeed, this early attractiveness is exactly what one ob- serves. Banking services were the next financial activity to develop (figure 1). However, because attracting deposits is easier than overcoming agency frictions that limit lending, bank deposit taking should precede bank lending. Moreover, banks initially lent to governments rather than to private agents. Indeed, both features are corroborated by figure 1. At the same time, bank retail funding (deposits) took off before wholesale funding. As frictions eased, retail investors were increasingly able to shift into higher yielding, market- based instruments or to have their funds managed by asset managers or institu- tional investors rather than banks.28 Reflecting participation frictions, bank claims on other financial institutions developed later but grew much faster once they developed (figure 3).29 The gradual overcoming of collective frictions allowed banks to increasingly participate in an interinstitutional market that became denser as the number of players rose and the latter became more inter- connected. Finally, note in figure 1 that improvements in bank efficiency (a re- duction in net interest margins) began occurring very early, presumably as soon as banks started to operate. Capital Markets In view of the stronger agency frictions associated with arm’s-length transac- tions, capital markets should only develop once public information has improved sufficiently. This hypothesis is largely consistent with the evidence. In view of high returns to scale, stock markets developed after banking (figure 1). Nonetheless, they developed ahead of prediction, which may be explained by the fact that equity issuance is an essential component of corporate finance (and gov- ernance) for larger firms. However, reflecting size thresholds and collective fric- tions that limit participation, the development of primary markets (market capitalization) should precede the development of secondary markets (market turnover), where trading liquidity is essential. As corroborated by figure 1, the secondary market developed after the primary market. At the same time, equity markets are likely to develop ahead of corporate bond markets because their 27. In particular, unlike markets, banks can develop even without a good legal framework (Rajan and Zingales [2001]). 28. Although our sample indicates that retail funding preceded wholesale funding, this sequence need not apply at all times and to all financial systems. For example, much of the precrisis banking growth in the Eastern European countries, which started from relatively limited levels of FD, originated from wholesale funding (mostly external) rather than domestic funding. Moreover, there appears to have been a shift (at least temporary) in many advanced economies back to retail funding following the global financial crisis. 29. The late but very rapid growth of interbank lending can be viewed as growth analogous to the rapid rise of bank interconnectedness that preceded the subprime crisis in the United States (Shin [2010]). de la Torre, Feyen, and Ize 533 unlimited upsides can better compensate for the downsides associated with agency frictions. Again, figure 1 shows that this sequencing held: Private debt se- curities followed equity and developed abroad before developing at home. Institutional Investors The development of asset managers is likely to be constrained by the develop- ment of capital markets. In particular, the growth of mutual and pension funds should reflect the fact that marketable, liquid assets—which the funds need to invest in—appear relatively late. Furthermore, pension funds developed earlier than mutual funds (figure 1) because the collective frictions that needed to be overcome for their development were largely solved by an act of government (e.g., the creation of privately administered, fully funded, individual retirement savings accounts, often of a mandatory nature). In contrast, the frictions hamper- ing the development of mutual funds (large returns to scale) had to be resolved by market forces. In both cases, however, their growth was extremely buoyant, reflecting the fact that marketable, liquid assets in which these funds invested grew very rapidly once they began to develop. Regarding insurance, life insurance companies are likely to develop relatively late because, in addition to being exposed to collective frictions (they also need to invest in market-based assets), they are exposed to larger agency frictions that reflect their longer investment ho- rizons. This fact may explain why life insurance developed later than predicted. In contrast, casualty insurance is likely to develop earlier, both because it is partly influenced by policy (as in the case of mandatory insurance for motor vehi- cles) and because the risks it faces are normally distributed and can be more easily addressed through risk pooling. Furthermore, their negative returns to scale, which make this activity somewhat of an outlier, can be explained by the predominance of foreign trade insurance in smaller open economies.30 I V. T H E R O L E OF POLICY In this section, we briefly explore the role of policy and institutions in explain- ing the deviations from benchmark paths. We use three variables (creditor rights, investor protection, and enforcement costs) that reflect the quality of the contractual environment and one variable (creditor information) that reflects the quality of the informational environment.31 To measure the quality of 30. The latter accounts for a disproportionally high share of total casualty insurance, reflecting the importance of foreign trade for the smaller economies (see Feyen, Lester and Rocha [2011]). 31. The investor protection index, creditor rights and creditor information indices are taken from the World Bank’s Doing Business database. Although the creditor information quality variable only measures one dimension of the quality of the informational environment (namely, the quality and coverage of credit bureaus), we used only this variable because it had the best coverage and was generally highly correlated with other information indices. The contract enforcement index is the first principal component of the following indicators (also from Doing Business): contract enforcement costs, number of days to enforce a contract (in logs), and number of procedures to enforce a contract. 534 THE WORLD BANK ECONOMIC REVIEW macroprudential management, we construct a credit crash dummy that mea- sures severe annual drops in the ratio of private credit to GDP.32 Table 4 pre- sents the results of the regressions similar to table 1 but with the policy variables included, estimated over the more limited period for which the data are available. The variance decomposition for these new estimates, which is similar to the decomposition shown for the regressions without policy vari- ables, appears in table 2. Together, the enabling-environment indicators account for a significant (albeit limited) fraction of the total explained variance of the financial indica- tors, which rises from 48 to 60 percent when the policy variables are incorpo- rated.33 As expected, better creditor rights promoted banking activity (not only private credit but also, and apparently even more, bank claims on other finan- cial institutions) as well as capital markets (stock market capitalization and, even more strongly, private bond market capitalization) and life insurance. Similarly, lower enforcement costs facilitated bank funding (both retail and wholesale) and bank private lending while contributing to lowering intermedia- tion margins. Unsurprisingly, the quality of investor protection had a particu- larly large impact on stock market activity (both in the primary and secondary markets). Furthermore, as expected, the quality of the informational environ- ment was important for bank private lending and, even more so, for private (corporate) debt capitalization (particularly at home).34 It is noteworthy, however, that the contractual variables (particularly credi- tor rights) accounted for the lion’s share of the cross-country differences, with information accounting for only a relatively small fraction.35 This result likely reflects the fact that improvements in the informational environment, which are more technology dependent, could be more simply introduced and imported from abroad.36 In contrast, improvements in the contractual environment re- quired a strengthening of local institutions that was harder to deliver. This 32. The “crash variable” for a particular country is the fraction of years of the entire period in which annual private credit dropped by 20 percent or more. 33. Many of the enabling-environment indicators face measurement problems and only cover limited dimensions of policy. Thus, the share of variation across countries explained by policy may be, in practice, substantially higher than identified here. 34. Somewhat surprisingly, however, the credit information variable appears to discourage bank funding and results in higher bank intermediation margins. The reduction in bank funding may reflect the emergence of alternative (market-based) channels of financial intermediation. Furthermore, as the more competitive segments of the borrowing market ( particularly corporates) migrate to the capital markets, bank loans become increasingly concentrated in households and smaller enterprises, where margins are higher owing to lower competition and higher risk. 35. The importance of the legal and institutional environment for financial and economic development is consistent with the evidence discussed in Beck and Levine (2005). 36. This conclusion deserves an important qualification because of the already noted limitations of our index, which only measures one dimension of a broader, multifaceted reality. To the extent that informational frictions reflect both information gathering and information processing costs (i.e., problems of bounded understanding and rationality in an increasingly complex environment), it could be argued that such frictions are unlikely to vanish any time soon. T A B L E 4 . Extended Benchmark Regressions Panel A Bank Claims Bank Insurance Insurance Bank Private Net Interest On Dom. Fin. Bank Credit To Bank Domestic Non-Deposit Premiums Premiums Credit Margin Sector Government Deposits Funding (Life) (Non-Life) 1 2 3 4 5 6 7 8 Log initial GDPPC 0.266*** 2 0.260*** 0.664*** 0.415*** 0.269*** 0.411*** 0.508*** 0.199*** Log GDPPC minus Log 0.456*** 0.524*** 2 0.817 2.065*** 1.049*** 2 0.378* 0.391 2 1.114*** initial GDPPC Interaction 2 0.00235 2 0.134*** 0.283*** 2 0.253*** 2 0.0817*** 0.115*** 0.0998** 0.187*** Log population 0.0406*** 2 0.112*** 0.294*** 0.204*** 0.0576*** 0.0754*** 0.0626*** 2 0.0520*** Log population density 0.0465*** 2 0.0167 0.348*** 0.175*** 0.0623*** 0.0152 0.139*** 2 0.0284*** Fuel dummy 2 0.233*** 2 0.0135 0.289* 2 0.464*** 2 0.227*** 2 0.145*** 2 0.519*** 2 0.183*** Offshore dummy 0.271*** 0.00753 2 0.767*** 0.157* 0.362*** 0.675*** 2 0.195* 0.0437 Transition dummy 2 0.373*** 0.152** 2 1.572*** 2 0.146 2 0.319*** 2 0.118 2 1.645*** 2 0.272*** Private credit crash 2 5.963*** 2.945*** 2 3.188** 2 1.724*** 2 3.329*** 2 5.281*** 2 1.782*** 2 0.409 Strength of legal rights 0.0288*** 2 0.00336 0.242*** 2 0.0454*** 0.00687 0.0178* 0.277*** 0.0561*** index Credit information 0.0425*** 0.0857*** 2 0.0264 2 0.210*** 2 0.0449*** 2 0.0560*** 0.0546*** 0.0180** index Strength of investor 0.0167 0.0103 2 0.0933* 0.155*** 0.0570*** 2 0.0390** 2 0.0250 2 0.0414*** de la Torre, Feyen, and Ize protection index Enforcement costs 2 0.00326*** 0.00178** 2 0.00486* 2 0.00268** 2 0.00238*** 2 0.00378*** 0.00668*** 0.000972 Observations 2,148 1,731 1,056 2,140 2,160 2,094 1,805 1,857 R2 0.710 0.479 0.395 0.317 0.662 0.604 0.633 0.537 (Continued ) 535 TABLE 4. Continued 536 THE WORLD BANK ECONOMIC REVIEW Panel B Pension Fund Mutual Stock Market Stock Market Domestic Domestic Public Foreign Foreign Public Assets Fund Assets Turnover Capitalization Private Debt Debt Securities Private Debt Debt Securities Securities Securities 1 2 3 4 5 6 7 8 Log initial GDPPC 2 0.0629 0.713*** 0.593*** 0.481*** 1.173*** 0.346*** 1.006*** 2 0.224*** Log GDPPC minusLog 2 2.166 2 2.300*** 1.268*** 0.644* 5.861*** 1.311** 2 0.894 2 1.431** initial GDPPC Interaction 0.285 0.476*** 2 0.0795 2 0.000579 2 0.487*** 2 0.224*** 0.327*** 0.133 Log population 2 0.0525 0.287*** 0.619*** 0.155*** 2 0.0301 0.154*** 0.123*** 2 0.467*** Log population density 2 0.142*** 2 0.119*** 0.00520 0.0397** 2 0.0740* 0.129*** 2 0.0179 2 0.184*** Fuel dummy 0.182 2 0.335** 2 0.183** 0.0137 2 0.380*** 2 0.00603 0.0845 2 0.0817 Offshore dummy 0.121 1.230*** 2 0.334*** 0.0959 2 0.567*** 2 0.509*** 2 0.243* 2 0.478*** Transition dummy 2 3.052*** 2 1.387*** 0.635*** 2 0.926*** 2 0.744** 0.359* 2 0.673*** 2 0.460*** Private credit crash 2 5.985*** 2 7.414*** 2 6.495*** 2 4.187*** 4.262** 2 2.988** 2 0.742 8.200*** Strength of legal rights 0.189*** 0.000925 0.00960 0.0372** 0.176*** 2 0.0530** 0.0421* 2 0.0906*** index Credit information 0.275*** 2 0.324*** 2 0.0881*** 2 0.132*** 0.181*** 0.108*** 2 0.111*** 0.115*** index Strength of investor 0.0476 0.0617 0.157*** 0.121*** 0.0480 0.0952*** 2 0.0470 2 0.0532 protection index Enforcement costs 2 0.00971 2 0.00576 2 0.0127*** 0.00268 0.00390 2 0.00294 0.00623** 2 0.00164 Observations 565 567 1,292 1,344 645 707 883 1,073 R2 0.378 0.669 0.598 0.490 0.567 0.277 0.617 0.330 Notes: This table extends table 1 by adding the following additional policy variables: Private credit crash (which assumes a value of 1 if private credit to GDP drops by over 20 percent for a particular country-year) and a set of variables taken from the World Bank Doing Business Database, including the Strength of legal rights index (the extent to which creditors are legally protected), the Credit information index (the quality of credit information), the Investor protection index (the extent to which investors are protected by law), and Enforcement costs (the cost to enforce a contract). The contract en- forcement index is the first principal component of the following indicators (also from Doing Business): contract enforcement costs, number of days to enforce a contract (in logs), and number of procedures to enforce a contract. GDPPC stands for gross domestic product per capita. ***, **, and * indicate p , .01, p , .05, and p , .1 Source: Authors’ analysis based on data described in the text. de la Torre, Feyen, and Ize 537 interpretation is bolstered by the results of a regression of creditor rights and credit information against country income (incorporated through both a linear term and a quadratic term). The creditor rights index is convex, whereas the credit information index is concave.37 This result suggests that improvements in the informational environment are achieved early on, whereas improvements in creditor rights take much longer to materialize. Thus, one should observe fewer cross-country differences regarding the quality of information compared to the quality of contracts. Remarkably, the credit crash dummy accounts for a sizable share of total ex- plained variance, particularly in the case of private credit (where it accounts for nearly 30 percent) but also for other banking indicators, such as bank funding and bank margins, and even for financial activity outside banking. For example, the credit crash dummy had a sizable negative impact on stock trading. It is also remarkable that banking systems took such a long time to recover from such impacts. This finding can be inferred from the fact that the credit cash dummy, which is estimated over a 30-year time span, remains highly significant when regressed against the pure cross-section of last period indicators rather than the full panel.38 This result is a noteworthy reminder that FD (the bright side) and financial instability (the dark side) strongly inter- act in lasting, complex ways. The interaction between bright and dark sides is illustrated further in figure 4, which plots the predicted and actual dynamic paths of private bank credit for all countries in the database, separated into three groups according to their initial income levels. Even at such an aggregate level of analysis, a cursory inspection suggests that, in most cases, busts occurred following booms in which the actual development paths went over their predicted values. Thus, financial unsustainability (the dark side of FD) appears to be related to abnor- mal deviations from the regular development paths inferred from a cross- country benchmarking analysis (the bright side of FD). Of course, this hypoth- esis must be validated by a detailed statistical analysis at the country level, a topic left for further research. V. F I N A L T H O U G H T S The broad cross-country benchmarking approach we have developed here can help to measure the quality of policies at any point in time and to identify the forces and frictions shaping FD. This approach began from the premise that both the evolution of financial structure (FD) and its sustainability (financial stability) are shaped by four fundamental types of frictions and their associated market failures. Two such frictions restrict agents’ capacity to establish and enforce bilateral contracts (agency frictions), and the other two frictions restrict 37. The underlying regression is available from the authors upon request. 38. The underlying regression is available from the authors upon request. 538 THE WORLD BANK ECONOMIC REVIEW F I G U R E 4. Estimated Private Credit Development Paths by Region Notes: This figure presents the actual and predicted paths of bank credit to the private sector (as a percentage of GDP) for high-income (A), middle-income (B), and low-income (C) countries. Predicted paths are derived from equation (1) using group medians. The samples are restricted to the countries for which data are available back to 1980. de la Torre, Feyen, and Ize 539 agents’ capacity to participate and coordinate their financial activities in ways that are collectively desirable (collective frictions). The predictions derived from this approach regarding the order of development, returns to scale, and the shape of the developmental paths of various financial activities were broadly satisfied. In particular, where strong participation frictions were re- sponsible for hindering the growth of financial activities, development only began after some threshold. However, once this threshold was passed, it was followed by buoyant dynamics because rising participation and interconnected- ness generated positive externalities that promoted further participation and interconnectedness. This paper has also argued that the same frictions that feed the development forces of the bright side feed the forces of instability from the dark side, making them interact in complex and unexpected ways and leading to booms followed by busts.39 Indeed, the paper has shown that FD was substantially vo- latile and that countries paid a heavy and lasting price for financial collapses. Avoiding such collapses requires the early identification of the buildup of sys- temic stress. By inferring a predicted development path from a broad universe of observations, our benchmarking approach may help to discriminate between sustainable and unsustainable trajectories. A key topic for further research is to investigate whether deviations from benchmarks can help to predict crashes on a country-by-country basis. An important caveat applies, however. Although our analysis suggests the presence of developmental regularities, it does not ensure predictability, partic- ularly as one goes farther into the future. On the one hand, financial innova- tions may introduce inflexion points that are undetectable from the existing historical data. If so, apparent short-term upward deviations from the bench- mark paths could, in fact, become sustainable because the benchmark paths themselves bend upward. On the other hand, it is possible for the benchmark paths to bend downward, at least in the longer term. For example, some of the indicators that have exhibited rapidly growing paths may eventually slow down, in accordance with logistic (S-shaped) trajectories. 39. The evidence presented on our measure of “buoyancy” is arguably consistent with the empirical suggestion that there can be “too much finance” (see Arcand, Berkes and Panizza [2012]). In effect, high buoyancy implies a decreasing impact of FD on economic growth. In this sense, finance may resemble a luxury good; its use increases significantly as income rises, yet its benefits (whether in welfare or growth) exhibit falling marginal returns. To be sure, buoyancy by itself does not necessarily imply that there can be “too much finance.” If one reasonably assumes nonsatiation, more finance should always be better. However, once the “dark side” is factored in, it is no longer clear that more finance is necessarily better because the marginal costs of financial instability may eventually come to dominate the marginal benefits of more FD. Ultimately, however, the balance between marginal costs and marginal benefits depends on the policy response, which puts a premium on keeping the forces of the dark side at bay as FD deepens. 540 THE WORLD BANK ECONOMIC REVIEW REFERENCES Akerlof, G. 1970. “The Market for Lemons: Quality Uncertainty and the Market Mechanism.” Quarterly Journal of Economics 84 (3): 488– 500. Anginer, D., A. de la Torre, and A. Ize. 2011. “Risk-Absorption by the State: When is it Good Public Policy?” World Bank Policy Working Paper 5893. Washington, DC. Arcand, J.L., E. Berkes, and U. Panizza. 2012. “Too Much Finance?” IMF Working Paper WP/12/161. International Monetary Fund, Washington, DC. Ashcraft, A., and T. Schuermann. 2008. “7 Deadly Frictions in Subprime Mortgage Securitization.” PRMIA Credit Forum. Professional Risk Managers’ International Association, Northfield, MN. Allen, F., and D. Gale. 2000. Comparing Financial Systems. Cambridge, MA: MIT Press. Beck, T., and R. Levine. 2004. “Stock Markets, Banks and Growth: Panel Evidence.” Journal of Banking and Finance 28 (3): 423– 42. ´ nard, and M. M. Shirley, eds., ———. 2005. “Legal Institutions and Financial Development.” In C. Me Handbook of New Institutional Economics, 251–78. The Netherlands: Springer. Calvo, G., and C. Reinhart. 2002. “Fear of Floating.” Quarterly Journal of Economics 117 (2): 379 –408. Claessens, S., and L. Laeven. 2003. “Financial Development, Property Rights, and Growth.” Journal of Finance 58 (6): 2285– 812. De la Torre, A., and A. Ize. 2010. “Regulatory Reform: Integrating Paradigms.” International Finance 13 (1): 109 –39. ———. 2011. “Containing Systemic Risk: Paradigm-Based Perspectives on Regulatory Reform.” Economia 11(1): 25 –64. De la Torre, A., and S. Schmukler. 2004. “Coping with Risk through Mismatches: Domestic and International Financial Contracts for Emerging Economies.” International Finance 7: 349– 90. ¨c Demirgu ¸ -Kunt, A., and R. Levine. 1999. “Bank-Based and Market-Based Financial Systems: Cross-Country Comparisons.” World Bank Policy Working Paper 2143. Washington, DC. ¨c Demirgu ¸ -Kunt, A., E. Feyen, and R. Levine. 2011. “The Evolving Importance of Banks and Securities Markets.” World Bank Policy Research Working Paper 5805. Washington, DC. Dickson, P. 1967. The Financial Revolution in England. London: Macmillan Eichengreen, B., R. Hausmann, and U. Panizza. 2003. “Currency Mismatches, Debt Intolerance and Original Sin: Why They are Not the Same and Why it Matters.” NBER Working Paper No. 10036. National Bureau of Economic Research, Cambridge, MA. Ferguson, Niall. 2008. The Ascent of Money: A Financial History of the World. New York: The Penguin Press. Feyen, E., R. Lester, and R. Rocha. 2011. “What Drives the Development of the Insurance Sector: An Empirical Analysis Based on a Panel of Developed and Developing economies.” World Bank Policy Research Working Paper 5572. Washington, DC. Geanakoplos, J. 2009. “The Leverage Cycle.” Cowles Foundation Discussion Paper 1715. Cowles Foundation for Research in Economics, Yale University, New Haven, CT. Gorton, G., and A. Metrick. 2010. “Securitized Banking and the Run on Repo.” Yale ICF Working Paper 09-14. International Center for Finance, Yale University, New Haven, CT. Haldane, A., S. Brennan, and V. Madouros. 2010. “What is the Contribution of the Financial System: Miracle or Mirage?” In The Future of Finance: The LSE Report, 87-120, London: London School of Economics. Holmstrom, B., and J. Tirole. 1996. “Private and Public Supply of Liquidity.” NBER Working Paper 5817. National Bureau of Economic Research, Cambridge, MA. Huang, R., and L. Ratnovski. 2010. “The Dark Side of Bank Wholesale Funding.” Working Paper 1223. European Central Bank, Frankfurt, Germany. de la Torre, Feyen, and Ize 541 Huang, Y. 2005. “What Determines Financial Development?” Discussion Paper 05/580. University of Bristol, Bristol, UK. Kindleberger, C.P. 1989. Manias, Panics and Crashes: A History of Financial Crises. Wiley Investment Classics. Hoboken, NJ: Wiley. Lerner, J., and P. Tufano. 2011. “The Consequences of Financial Innovation: A Counterfactual Research Agenda.” NBER Working Paper 16780. National Bureau of Economic Research, Cambridge, MA. Merton, R., and Z. Bodie. 2004. “The Design of Financial Systems: Towards A Synthesis of Function and Structure.” NBER Working Paper 10620. National Bureau of Economic Research, Cambridge, MA. Minsky, H. 1975. John Maynard Keynes. New York: Columbia University Press. North, D.C. 1990. Institutions, Institutional Change and Economic Performance. Cambridge, UK: Cambridge University Press. Rajan, R., and L. Zingales. 2001. “Financial Systems, Industrial Structure and Growth.” Oxford Review of Economic Policy 17: 467– 82. ———. 2003. Saving Capitalism from the Capitalists: Unleashing the Power of Financial Markets to Create Wealth and Spread Opportunity. New York: Crown Business Division of Random House. Shiller, R. 2006. Irrational Exuberance. New York: Random House. Shin, H.S. 2010. “Securitization and the Financial System.” In Risk and Liquidity, Clarendon Lectures in Finance, 152– 70. Oxford: Oxford University Press. Shleifer, A., and R. Vishny. 2011. “Fire Sales in Finance and Macroeconomics.” Journal of Economic Perspectives 25 (1): 29– 48. Spence, M. 1973. “Job Market Signaling” Quarterly Journal of Economics 87 (3): 355–74. Stiglitz, J., and A. Weiss. 1981. “Credit Rationing in Markets with Imperfect Information.” American Economic Review 71 (3): 393– 410. Job Growth and Finance: Are Some Financial Institutions Better Suited to the Early Stages of Development than Others?1 Robert Cull and L. Colin Xu Evidence based on firm-level data from 89 countries with updated country-level data on financial structure suggests that in low-income countries, labor growth is more rapid in countries with a higher level of private credit/GDP. This positive re- lationship with private credit is especially pronounced in industries that depend heavily on external finance. The results, which are robust to multiple estimation ap- proaches, are consistent with the predictions of new structural economics. In high- income countries, labor growth rates increase with the level of stock market capi- talization, consistent with predictions from new structural economics. However, the association disappears when stock market development is treated as an endogenous explanatory variable using instrumental variable regressions. There is no evidence that small-scale firms in low-income countries benefit the most from the develop- ment of the private credit market. Rather, the labor growth rates of larger firms increase to a greater extent than others with the level of private credit market development, a finding consistent with the perspective from historical political economy that banking systems in low-income countries serve the interests of the elite rather than providing broad-based access to financial services. JEL codes: G2, O1 Although the relative advantages of bank- and market-based systems have been clearly presented in the literature, the available empirical evidence does not 1. Robert Cull (corresponding author) is a lead economist at the Development Research Group of the World Bank; his email address is rcull@worldbank.org. L. Colin Xu is also employed as a lead economist at that institution; his email address is lxu1@worldbank.org. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations or those of the Executive Directors of the World Bank or the governments they represent. The authors thank Stijn Claessens, three anonymous referees, and participants at the World Bank Conference on Financial Structure and Economic Development (June 16, 2011) for many helpful comments and suggestions. THE WORLD BANK ECONOMIC REVIEW, VOL. 27, NO. 3, pp. 542– 572 doi:10.1093/wber/lhs050 Advance Access Publication January 24, 2013 # The Author 2013. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 542 Cull and Xu 543 indicate that either type is more effective in promoting growth.2 Cross-country evidence has demonstrated that the overall level of financial development, rather than its institutional composition, is robustly linked to economic growth (Demirgu ¸ -Kunt and Levine 2001; Levine 2002). Although cross-country indi- ¨c cators of financial development have proliferated and been refined over the past decade, they necessarily entail limitations when describing the nuances of the financial structure of a given country. Therefore, it is possible, and even likely, that in a given country or at a specific point in time, productive activities would be better supported by banks or markets. This paper examines whether firms from countries in the early stages of de- velopment grow more rapidly under a bank-based system that harnesses local information than under a market-based system that does not. Our approach relies on a measure of firm growth from the WBES database, the percentage increase in the number of workers over the two years prior to the survey. Labor growth is a topic of intense interest, but there are also practical reasons for focusing on this measure in this analysis.3 Other measures of firm growth (such as sales growth) are available, but our measure of labor growth is avail- able for a larger set of firms and countries. Testing the hypothesis that financial structure has different effects on firm growth depending on a country’s level of economic development requires the widest possible sample of countries because the financial structure indicators are measured at the country level. Our most expansive regression models employ information from over 50,000 firms in 89 countries.4 Relying on labor (rather than sales) growth has an additional advantage: labor is likely to be measured with less error than sales for both accounting and tax reasons, and such measurement errors may differ systemati- cally by the level of development. 2. The relative advantages of different types of financial systems have been long debated, most notably in comparisons between market-based and bank-based financial structures (Gershenkron 1962; Demirgu ¸ -Kunt and Levine 2001). Banks, for example, are able to exploit economies in processing ¨c information regarding the creditworthiness of prospective borrowers and often form long-run relationships with firms that reduce information asymmetries and permit the effective monitoring of firms’ activities. Securities markets provide an incentive to gather information about firms, provide a liquid platform for investors to buy and sell shares in those firms, and may improve the corporate governance of firms by facilitating takeovers. More generally, markets are likely to do a better job of aggregating and transmitting information signals to investors than banks, which could improve the allocation of financial resources and thus promote growth. See Levine (2002) for an extensive review of the literature on the relative merits of banks versus markets. 3. For example, job growth is the topic of the upcoming World Development Report for 2013 and a key concern of the current U.S. administration. 4. The firm-level survey information as well as a steady expansion in the number of countries that have financial structure data available enables us to undertake this analysis. For comparison, the original financial structure indicators presented in Levine (2002) were available for only 48 countries. 544 THE WORLD BANK ECONOMIC REVIEW These themes were recently developed in the “New Structural Economics” approach to studying economic development (Lin 2010). With respect to finan- cial sector development, Lin, Sun, and Wu (2012) note, In a poor country, where the endowment structure is characterized by labor abundance and capital scarcity, labor-intensive industries are consistent with the comparative advantages deter- mined by its endowment structure. Since labor-intensive businesses are usually smaller, they are more informationally opaque and require less amount of external finance than firms in capital- intensive industries. Lending to smaller businesses often requires banks to collect soft information on borrowers. Due to organizational complexity and the corresponding difficulty in communicat- ing soft information, it is more difficult for large banks to effectively collect and utilize soft infor- mation about borrowers when making lending decisions. But smaller banks have advantages in monitoring small firms and satisfying their financial needs. Therefore, poor countries . . . should have smaller local banks play a dominant role in their banking sector. As a result, proponents of this approach hold that there is an endogenously de- termined optimal financial structure at each stage of an economy’s develop- ment. Measurement problems are a key challenge to empirically testing this approach to financial development. Case studies of individual countries lack sufficient variation in financial structure over time to permit formal hypothesis testing, and although cross-country financial indicators have come a long way in a short time, they remain relatively crude. Determining a country’s optimal industrial structure and assessing the suitability of its financial structure using cross-country regressions alone would be a daunting task. To foreshadow our main result, we find that firms grow more rapidly in countries with low levels of per capita income when the banking system is relatively well developed. We find no such results for other measures of the financial structure, including measures of stock market development. Moreover, we find no strong relationships between the financial structure and firm labor growth of countries with higher levels of per capita income, in line with previous findings in the literature (Levine 2002; Demirgu ¨c¸ -Kunt and Levine 2001). Our results are robust to the use of the instrumental variables method to address the potential endogeneity of our financial struc- ture variables. Previous studies have found that credit from suppliers (trade credit) can be an alternative source of funding for firms with limited access to bank loans. For example, small firms in the United States that lack an established relation- ship with a bank hold significantly higher levels of accounts payable than other firms (Petersen and Rajan 1997). More generally, using the regression method- ology pioneered by Rajan and Zingales (1998), Fisman and Love (2003) show that firms in industries that are more dependent on trade credit financing grow relatively more rapidly in countries with relatively underdeveloped financial sectors. Because our focus is on developing and transitional economies, and as Cull and Xu 545 a result of data constraints, we do not include trade credit in our analysis.5 We note, however, that if trade credit naturally arises as a substitute for bank credit in less developed financial sectors, we would be less likely to find a sig- nificant relationship between banking sector development and firm growth. Because that relationship is only significant for the low-income subsample in our analysis, trade credit may not be an adequate substitute for bank credit in those environments. By contrast, trade credit could be a more effective substi- tute for bank credit in more advanced countries, which could help to explain why we do not find a significant relationship between banking sector develop- ment and firm labor growth for the sample of higher-income countries. We also apply the Rajan and Zingales regression methodology to test whether firms in industries that rely heavily on external financing have higher labor growth rates in countries with relatively well-developed financial sectors. We find evidence consistent with that proposition, but only for firms in low- income countries with relatively well-developed banking sectors. Although the instrumental variables approach focuses on between-country differences, the Rajan and Zingales approach captures within-country, between-industry differ- ences in labor growth rates. The fact that both approaches yield similar results supports the plausibility of our findings. Our confidence regarding the benefi- cial role of banks in poor countries is further bolstered by the finding that a more developed banking system is associated with higher investment rates, more employee training, and larger firm sizes only in poor countries, which suggests that banks spur both physical and human capital investments in addi- tion to increasing job growth in these countries. The WBES also contains a substantial amount of information on firm char- acteristics (size, industry, ownership structure, and legal status) that enables us to better identify the types of firms that benefit from a relatively well-developed banking system in countries that are at an early stage of development. Although our results seem to indicate that banks are the financial institutions that are best suited to serve firms in low-income countries, they do not appear to disproportionately serve the small-scale firms that characterize the early stages of economic development. In low-income countries, although labor growth increases with the size of the banking sector, large firms are the primary beneficiaries of this increase in labor growth. The remainder of the paper is organized as follows. Section I describes the enterprise survey and financial structure data in greater detail and presents summary statistics. Section II describes the variables that we use as instruments 5. Fisman and Love (2003) use data from 37 industries in 43 countries to conduct their analysis. Industry-level trade credit financing in the United States is used to summarize the natural reliance on that form of financing in a frictionless financial sector. Most of those countries have higher per capita income levels than the countries that compose our sample (see table 1). Thus, the assumption that U.S. trade credit usage is a reasonable benchmark may be more applicable to higher-income countries than to those in our sample. In any event, we were unable to include a consistent measure of trade credit usage for a sufficiently wide set of countries in our OLS and instrumental variables regressions. 546 THE WORLD BANK ECONOMIC REVIEW for our indicators of financial development in the firm growth regressions. Section III explains our estimation approach based on instrumental variables and the approach based on the Rajan and Zingales methodology and presents our main regression results. Section IV examines the types of firms that are most affected by the financial structure in high- and low-income countries. Section V provides a series of robustness checks for our main findings that in- corporate measures of banking sector concentration and efficiency and the quality of the business environment. Section VI examines whether other firm characteristics and performance measures are related to financial development to better understand the potential mechanisms by which banking sector devel- opment fosters firm growth in low-income countries. Section VII concludes. I . D ATA Sampling from the universe of registered businesses and following a uniform stratified random sampling methodology, the core WBES employs a standard- ized survey instrument to benchmark the investment climate of individual econ- omies across the world.6 Table 1 lists the countries surveyed and the year and number of observations for each survey. The country list demonstrates that our focus is on developing and transitional economies. In the empirical analysis that follows, we often split the sample in half and, for ease of exposition, refer to these economies as the “poor” and “rich” subsamples. The split-sample tests roughly provide comparisons of firms in low- and middle-income coun- tries. The survey contains sufficient information to allow for firm performance analyses and reports detailed information on firm employment, age, industry, ownership, legal status, and the number of establishments. Table 2 provides the definitions and sources of the key variables used in the analysis. As noted above, our analysis is designed to explain the variation in firm labor growth. We use the percentage growth in the number of full-time em- ployees over the two years prior to the survey as the dependent variable in our regressions because that information was asked of firms in a wider sample of countries than, for example, information about sales growth.7 In addition, the information provided by firm owners regarding their number 6. A detailed description of the sample design and sample frame can be found at http://www. enterprisesurveys.org/documents/Sampling_Note.pdf. 7. We did, however, run regressions using both sales growth and a measure of total factor productivity as the dependent variable. We generally did not find strong links between those variables and our measures of financial sector structure. This is likely because the number of available observations was small for those variables. For example, the maximum number of observations for the sales growth variable was 41 countries in the rich subsample and 23 for the poor subsample. For total factor productivity, there were 36 countries in the rich subsample and 23 for the poor subsample. In comparison, there were 50 available rich countries and 43 poor ones for the labor growth variable. In the sales growth regressions for the low-income sample, the coefficient for private credit/GDP was positive in the OLS, GMM, and Rajan and Zingales regressions (described below), but it achieved significance only in the Rajan and Zingales regressions. T A B L E 1 . List of Countries and Years of Each Survey with Sample Size Number Number Number Number Number Country, year of firms Country, year of firms Country, year of firms Country, year of firms Country, year of firms Albania, 2002 163 Colombia, 2006 890 Hungary, 2005 583 Moldova, 2002 167 Slovenia, 2002 183 Albania, 2005 199 Costa Rica, 2005 316 India, 2002 919 Moldova, 2003 98 Slovenia, 2005 220 Angola, 2005 346 Croatia, 2002 176 Indonesia, 2003 67 Moldova, 2005 331 South Africa, 2006 831 Angola, 2006 342 Croatia, 2005 226 Ireland, 2005 492 Mongolia, 2004 160 Spain, 2005 597 Argentina, 2006 923 Czech Republic, 2002 256 Kazakhstan, 2002 244 Morocco, 2004 756 Sri Lanka, 2004 386 Armenia, 2002 164 Czech Republic, 2005 315 Kazakhstan, 2005 562 Mozambique, 2002 109 Swaziland, 2005 242 Armenia, 2005 347 Dominican Republic, 2005 123 Kenya, 2003 201 Namibia, 2006 293 Swaziland, 2006 240 Azerbaijan, 2002 154 Democratic Republic of 340 Kenya, 2006 650 Namibia, 2005 –2006 300 Syrian Arab 466 Congo, 2005 Republic, 2003 Azerbaijan, 2005 339 Ecuador, 2003 392 Korea, Rep., 2005 573 Nicaragua, 2003 413 Tajikistan, 2002 171 Belarus, 2002 243 Egypt, Arab Rep., 2004 911 Kyrgyz Republic, 2002 160 Oman, 2003 317 Tajikistan, 2003 102 Belarus, 2005 313 Egypt, Arab Rep., 2006 943 Kyrgyz Republic, 2003 97 Panama, 2006 527 Tajikistan, 2005 189 Benin, 2004 170 El Salvador, 2003 415 Kyrgyz Republic, 2005 192 Paraguay, 2006 543 Tanzania, 2003 216 Bolivia, 2006 529 Estonia, 2002 159 Lao People’s Dem. 224 Peru, 2002 105 Tanzania, 2005 401 Rep., 2005 Bosnia and 172 Estonia, 2005 207 Latvia, 2002 169 Peru, 2006 574 Tanzania, 2006 395 Herzegovina, 2002 Bosnia and 196 Ethiopia, 2002 351 Latvia, 2005 196 Philippines, 2003 607 Thailand, 2004 1,374 Herzegovina, 2005 Botswana, 2005 311 Gambia, 2005 194 Lebanon, 2006 331 Poland, 2002 477 Uganda, 2003 240 Botswana, 2006 303 Gambia, 2006 192 Lesotho, 2003 42 Poland, 2003 96 Uganda, 2005 573 Brazil, 2003 1,547 Georgia, 2002 165 Lithuania, 2002 192 Poland, 2005 931 Uganda, 2006 570 Bulgaria, 2002 235 Georgia, 2005 194 Lithuania, 2004 206 Portugal, 2005 498 Zambia, 2006 491 Bulgaria, 2005 278 Germany, 2005 1,190 Lithuania, 2005 192 Romania, 2002 245 Burkina Faso, 2006 45 Greece, 2005 537 Macedonia, Former 162 Romania, 2005 559 Yugoslav Rep., 2002 Cull and Xu Burundi, 2006 289 Guatemala, 2003 435 Macedonia, Former 191 Russian Federation, 2002 481 Yugoslav Rep., 2005 Burundi, 2005 –2006 292 Guinea-Bissau, 2005 179 Madagascar, 2005 226 Russian Federation, 2005 581 Cambodia, 2003 24 Guinea-Bissau, 2006 175 Malawi, 2005 134 Rwanda, 2005 209 Cameroon, 2006 113 Guinea, 2005 –2006 244 Malaysia, 2002 839 Saudi Arabia, 2005 536 Cape Verde, 2006 46 Guyana, 2004 144 Mali, 2003 125 Senegal, 2003 206 Chile, 2004 903 Honduras, 2003 352 Mauritius, 2005 158 Slovak Republic, 2002 161 547 China, 2002 2,374 Hungary, 2002 237 Mexico, 2006 1,251 Slovak Republic, 2005 205 548 THE WORLD BANK ECONOMIC REVIEW T A B L E 2 . Variable Definitions and Sources Variable name Definitions Lgrow (labor growth) Firm-level employment growth rate, from Investment Climate Assessment data. Private credit/GDP Private Credit/GDP, from World Bank Financial Development and Structure Database (also available from WDI). Stock market/GDP Stock market capitalization/GDP, World Bank Financial Development and Structure Database. Share of state banks Share of banking sector assets held by government-owned banks, from Micco, Panizza, and Yanez (2007). Share of foreign banks Share of banking sector assets held by foreign-owned banks, from Claessens and Van Horen (2012). Ln(Lt-2) Log(firm size in terms of the number of employees, twice-lagged), from Investment Climate Assessments. Ln(firm age) Log(firm age), from Investment Climate Assessments. Ln(GDP per capitat-2) Log(GDP per capita, twice-lagged), country-level, WDI data. of employees is likely to be more accurate than information regarding their sales, especially for smaller firms that either do not maintain quality ac- counting records or are reluctant to fully report their sales (e.g., because of potential tax consequences). For the 89 countries represented by the firms considered in our regressions, the WBES were conducted at different points in time between 2002 and 2006. We use three firm characteristics as controls in our firm growth regressions: age, size as measured by the number of employees two years prior to the survey, and the percentage of shares held by foreign owners. Both firm age and the number of employees enter the regressions in log form. We do not have strong priors regarding how these characteristics affect firms’ labor growth, al- though if older firms are more likely to have reached their equilibrium size, they may have slower growth rates than younger firms. Percentage growth in employees may be greater for smaller firms because they begin from a low base of employees, but larger firms may have advantages, especially in terms of access to finance, that make hiring additional workers easier. It is also unclear whether firms with high shares of foreign ownership expand employment more or less quickly than firms with higher shares of domestic ownership. Better access to finance by foreign firms would allow for more rapid expansion, but a shorter-term focus on the part of those firms may limit firm expansion. The key explanatory variables in our analysis are financial structure indica- tors drawn from the World Bank Database on Financial Development and Structure, which was updated in November 2010.8 We rely on two primary variables that are intended to capture different aspects of each country’s 8. More detailed descriptions of the dataset can be found in Beck, Demirgu ¸ -Kunt, and Levine ¨c (2000, 2010). The permanent URL for accessing the dataset is http://go.worldbank.org/X23UD9QUX0. Cull and Xu 549 financial structure: the ratio of private credit to GDP and the ratio of stock market capitalization to GDP. From previous research on the finance-growth nexus, we would expect growth in employment to be positively linked to both of these measures. However, new structural economics predicts that the associ- ation between firm growth and private credit relative to GDP, which is a bank- based measure of financial development, should be stronger for countries in early stages of development. We acknowledge, however, that this variable does not provide as direct a test of the predictions of new structural economics as we would like because it does not summarize the size distribution of banks, and small banks are the institutions that are hypothesized to best serve small- scale manufacturers in the early stages of development (as described above). By contrast, the relationship between employment growth and stock market capi- talization (relative to GDP) is likely to be stronger for countries in more ad- vanced stages of economic development. Another aspect of financial structure that may affect firm growth is the nature of ownership—the extent to which the banking sector is state-owned or foreign. However, neither state nor foreign ownership of banks explains varia- tion in labor growth in our sample (results not reported). The ratios of private credit to GDP and stock market capitalization to GDP are therefore the focus of the empirical analysis that follows. Perhaps the first thing to notice from table 3, which provides summary sta- tistics for our four indicators of financial development, is that country coverage is broader for the private credit measure, with 46 countries at or below the sample median for per capita income and 45 countries above the median. Broader coverage of countries is another reason that private credit is the primary financial indicator in our analysis. For stock market capitalization/ GDP and the shares of banking sector assets held by state and foreign-owned banks, the sample is skewed in favor of high-income countries (37–41 country observations) as opposed to lower-income countries (19 –22 observations). In part, this pattern of data availability may reflect the fact that stock markets and large shares of foreign bank participation are more likely to be features of the financial systems of advanced countries than of developing countries. However, this is not the case for the share of sector assets held by state-owned banks. The prevalence of observations from wealthier countries for that vari- able likely indicates the relative difficulty of collecting data that summarize fi- nancial development in low-income countries. In any event, for indicators other than private credit/GDP, the relatively small set of lower-income coun- tries makes it more difficult to test hypotheses regarding the optimal financial structure based on predictions from new structural economics.9 9. The reason for a significantly higher likelihood of missing observations for our indicator of stock market development may be that when this variable is unobserved for poor countries, its true value is zero. Below, we examine how this consideration affects our results. 550 THE WORLD BANK ECONOMIC REVIEW T A B L E 3 . Variations in Poor and Rich Countries For the poor region N Mean SD p10 p25 p50 p75 p90 CV p75/p25 p90/p10 Labor growth 46.000 0.493 0.430 0.126 0.233 0.378 0.654 0.749 0.872 2.800 5.947 Private credit/GDPt-2 46.000 0.195 0.196 0.047 0.071 0.142 0.263 0.427 1.007 3.729 9.129 Stock market cap./GDPt-2 19.000 0.144 0.140 0.007 0.033 0.089 0.215 0.381 0.972 6.524 57.142 Investment rate 41.000 0.045 0.031 0.015 0.024 0.043 0.059 0.072 0.699 2.430 4.723 Training dummy 46.000 0.333 0.169 0.126 0.204 0.318 0.438 0.512 0.509 2.144 4.063 Ln(workerst-2) 46.000 3.002 0.811 1.714 2.617 3.031 3.447 4.133 0.270 1.317 2.411 State bank ownership share 22.000 0.204 0.252 0.000 0.000 0.083 0.315 0.688 1.238 . . Foreign bank ownership share 22.000 0.363 0.275 0.057 0.134 0.330 0.487 0.757 0.758 3.634 13.279 For the rich region N Mean SD p10 p25 p50 p75 p90 CV p75/p25 p90/p10 Labor growth 45.000 0.286 0.182 0.120 0.171 0.239 0.327 0.576 0.638 1.914 4.801 Private credit/GDPt-2 45.000 0.515 0.405 0.169 0.195 0.391 0.765 1.153 0.786 3.917 6.823 Stock market cap./GDPt-2 41.000 0.309 0.338 0.065 0.116 0.217 0.355 0.543 1.093 3.063 8.396 Investment rate 24.000 0.038 0.026 0.012 0.018 0.035 0.048 0.077 0.688 2.621 6.363 Training dummy 45.000 0.507 0.182 0.296 0.398 0.466 0.660 0.745 0.358 1.660 2.517 Ln(workerst-2) 45.000 3.120 0.548 2.602 2.808 3.082 3.283 3.913 0.176 1.169 1.504 State bank ownership share 37.000 0.162 0.199 0.000 0.010 0.096 0.203 0.452 1.230 20.300 . Foreign bank ownership share 37.000 0.460 0.326 0.065 0.181 0.435 0.730 0.917 0.708 4.033 14.043 Notes: The poor and the rich samples each account for 50 percent of the (collapsed) country sample that deletes firms that have missing observations for labor growth and our base control variables. N represents the number of countries; SD represents the standard deviation; CV represents the coefficient of variation; p10, p25, p50, p75, and p90 represent the 10th, 25th, 50th, 75th, and 90th percentiles, respectively. Cull and Xu 551 As expected, higher-income countries have more developed financial sectors in terms of their average private credit/GDP, stock market capitalization/GDP, and the share of assets held by foreign-owned banks (table 3).10 The average share of sector assets held by state-owned banks is slightly skewed in favor of lower-income countries (0.20 vs. 0.16 for higher-income countries), although the median state-owned bank asset share is almost identical for the two groups. At least two other features of the summary statistics for the financial indica- tors stand out. First, although higher-income countries tend to have more devel- oped financial sectors, there is an overlap in the distributions of the financial indicators between higher- and lower-income countries. For example, among lower-income countries, those in the top 10th percentile have private credit/GDP ratios of 42.7 percent or higher, which ranks them above the median for higher- income countries. A second feature is that within the group of low- (or high-) income countries, there is substantial variation in all financial development indi- cators despite the relatively small sample sizes. For example, the ratio of stock market capitalization to GDP runs from 1 percent at the 10th percentile to 38.1 percent at the 90th percentile for countries in the lower-income sample. On the one hand, this pattern suggests that if there is an optimal financial structure that varies with the stage of development as reflected in per capita income levels, a sizable fraction of the countries in our sample are not achieving it. On the other hand, and perhaps more practically, the substantial variation in financial struc- tures within both the high- and low-income samples enables us to test whether certain structures yield better outcomes in terms of employment growth and whether those structures differ depending on a country’s income level. II. INSTRUMENTS Endogeneity poses the primary challenge to identifying a causal relationship between financial structure and firm growth at different stages of economic de- velopment. Therefore, we use instrumental variable estimation techniques. A brief description of the specific variables that we consider and the motivation for their inclusion as instruments are presented below. Commodities, Natural Resources Engerman and Sokoloff (1997) argue that the land endowments of Latin America were amenable to commodities that featured economies of scale in production (sugar cane, rice, silver) and thus the use of a large share of slave and indigenous labor. Power was historically concentrated in the hands of the plantation and mining elite, and the institutional structure that arose and per- sisted offered economic opportunity only to a small group. In contrast, the 10. The difference is statistically significant for private credit/GDP and stock market capitalization/ GDP at conventional levels but less so for the share of assets held by foreign-owned banks (with a p value of 12 percent). 552 THE WORLD BANK ECONOMIC REVIEW endowments of North America lent themselves to commodities grown on family farms (particularly wheat and maize), which fostered the growth of a relatively large middle class in which power was widely distributed. The insti- tutions that arose and persisted came to reflect this relative equality of econom- ic opportunity. These factors help to account for the disparity in per capita income between North America and Latin America. Therefore, we include dummy variables to indicate whether any of a given commodity is grown in each country as potential instruments. Corn, maize, and wheat growth are our proxies for the egalitarian institutions emphasized by Engerman and Sokoloff, which we expect to be positively linked to financial development. A related body of literature examines whether natural resource endowments, particularly oil, are associated with lower levels of economic development. The existence of a so-called resource curse has been extensively analyzed and debated (Sachs and Warner 1995, 2001; Lederman and Maloney 2008). We therefore include a measure of an economy’s dependence on oil, the net exports of petroleum per worker, in the set of potential instruments. We expect that variable to be negatively associated with the financial development indicators.11 Settler Mortality Another influential strand of the literature on the effects of endowments on institutional development and growth trajectories focuses on rates of settler mortality (Acemoglu, Johnson, and Robinson 2001, 2002). In climates and en- vironments that were inhospitable to development, as reflected by high rates of settler mortality, Europeans created states and institutions that enabled elites to extract wealth from the colonies, usually in the form of minerals and cash crops. By contrast, in more hospitable environments, settler colonies emerged in which institutions were created to protect property rights and to foster more broad-based economic opportunity. Acemoglu, Johnson, and Robinson (2001) find robust evidence that settler mortality has a pervasive influence on a series of key property rights institutions, such as protection against expropriation by the government, constraints on the executive, and the establishment of democ- racy. Because relatively widespread access to financial services is likely to be reflective of the more egalitarian institutions that emerged in settler colonies, we include the logarithm of annualized deaths per thousand European soldiers in the set of potential instruments, which we expect to be negatively associated with the financial development indicators. Fractionalization and Trust Trust, which arises as a manifestation of the social and institutional environ- ment, may help to facilitate financial intermediation. A higher level of trust, for 11. A key advantage of using this measure is that it is available for most countries. Moreover, Lederman and Maloney (2008) show that measures based on net exports are more closely linked to actual natural resource reserves than other trade-based endowment measures. Cull and Xu 553 instance, may reduce the information asymmetry between lenders and borrow- ers; it may also facilitate the social norm of not defaulting on debts. For example, Guiso, Sapienza, and Zingales (2004) find that in Italy, social capital development has a positive effect on the use of formal financial services, in- cluding credit. Therefore, we include the average level of trust in a country from 1981 to 2006 among the potential instruments, and we expect it to be positively linked to the financial development indicators.12 We also include measures of ethnic and linguistic fractionalization among the set of potential instruments. We expect them to be negatively linked to the financial development indicators. Legal Origin La Porta, Lopez-de-Silanes, Shleifer, and Vishny (1998) first documented that the legal protection of outside investors was stronger in common law countries (of English legal origin) than in civil law countries rooted in Roman legal tradi- tions, particularly those of French legal origin. Their interpretation was that English common law traditions evolved to protect private property and support private market outcomes, whereas French civil law tended to unify the legal system and cement state control of the judicial system, often to the detriment of private market development. Because the legal traditions in most countries were introduced by outside colonizers, LaPorta et al. argued that legal origin could be treated as exogenous in regressions explaining financial development and growth. Using legal origin as an instrument for investor protection, they demonstrated a strong link between those protections and measures of financial development. We therefore include dummy variables for English and French legal origins in the set of potential instruments. We expect English (French) legal origin to be positively (negatively) linked to financial development. We ran a series of exploratory regressions (unreported) of financial develop- ment on firm characteristics and potential instruments, each added separately to the regression. The strongest relationships were the negative ones between settler mortality and both key measures of financial development (private credit/GDP; stock market capitalization/GDP), a positive correlation between private credit/GDP and interpersonal trust, and a negative correlation between net petroleum exports per capita and private credit/GDP. Surprisingly, the legal origins, the commodities-based endowment measures, and the ethnic 12. As described in Knack and Keefer (1997) and Zak and Knack (2001), our measure of trust is based on data from the World Values Survey first conducted in 1981. Specifically, the measure is the percentage of respondents in each country agreeing that “most people can be trusted” against the alternative that “you can’t be too careful in dealing with people.” We use the average from 1981 to 2006 (rather than from the actual year of the WBES survey) because data on trust are available for distinct years for different countries, and relying on the survey year would lead to too many missing values for trust for our countries. Averaging also has the benefit of smoothing the measurement errors for the trust variable. 554 THE WORLD BANK ECONOMIC REVIEW fractionalization measures were not significantly associated with our measures of financial development in these regressions. Settler mortality does not appear in the first stage results in table 6 despite its strong association with financial development. This is because the instrument set does not pass the Hansen’s J test of overidentifying restrictions when settler mor- tality is included (results unreported). For the exclusion restrictions to be satisfied, the instruments should affect labor growth only through their effects on financial development. Because settler mortality has been shown to be strongly linked to broad measures of institutional and economic development (Acemoglu, Johnson, and Robinson 2001, 2002)—for instance, we can conceivably see how settler mortality may directly affect local infrastructure and labor regulations, both of which may directly affect firm growth (Xu 2011)—it is not surprising that its in- clusion results in the failure of the overidentification restriction test (i.e., Hansen’s J test). We note, however, that its inclusion does not change the relationships between financial structure and labor growth highlighted below.13 Because trust may have a strong influence on contractual arrangements, we find it more plausible that its effects on labor growth may work largely through its positive impact on private credit levels. Similarly, the strong negative relationships between the measures of countries’ reliance on oil exports and various measures of financial development recently documented in Beck (2011) indicate that wide- spread access to financial services, particularly credit, is rare in resource-based economies. Therefore, the notion that net oil exports primarily affect labor growth through their effects on financial development and structure also seems plausible. In the Generalized Method of Moments (GMM) regressions that follow, we will show that the test of overidentifying restrictions is passed when trust and net oil exports are used as instruments in most cases, and we will offer additional estimation approaches in the few cases in which the test is not passed. III. REGRESSIONS We present our results in several steps, first ignoring the potential endogeneity of financial structure and then addressing this potential endogeneity with either the instrumental variables or the Rajan and Zingales approach. Ordinary Least Squares We estimate the following base regression: DLij ¼ a þ bFIRMij þ gFINj þ 1ij ð1Þ where DL is the percentage change in the number of workers employed by firm i in country j over the two years prior to the enterprise survey, as described above. FIRM represents the three firm characteristics that we use as controls: 13. Results are not reported here but are available from the authors. Cull and Xu 555 age, size measured in (the logarithm of ) total workers, and the share of foreign ownership.14 As described in section II, we expect that labor growth would be slower for older, better established firms. We have no strong priors regarding how firm size or the share of foreign ownership affect labor growth. FIN repre- sents the indicators of financial development and structure that are the focus of the analysis. Because the dependent variable measures labor growth over the two years prior to each survey, the financial structure variables and firm size are also measured two years prior to the survey (denoted by the subscript “t-2” in the tables). We begin with ordinary least squares (OLS) regressions describing the asso- ciations between the financial development indicators and labor growth. We present results for the full sample of countries and for high- and low-income subsamples. We divide the countries into high- and low-income categories based on the median per capita income level for the 91 countries for which we have private credit/GDP figures, as we did for the summary statistics in table 3. Standard errors are clustered at the country level in all models to avoid exag- gerating the precision in firm-level regressions with country-level variables (Moulton 1990). The OLS regressions suggest relationships between the financial indicators and firm labor growth that are consistent with predictions of new structural economics (table 4). For example, private credit/GDP is only significantly posi- tively associated with labor growth in the sample of low-income countries, whereas stock market development/GDP is positively associated with labor growth among the high-income group. The magnitudes of those coefficients are also large. For example, a one-standard-deviation increase in private credit/ GDP in the low-income group yields an approximately 20 percentage-point in- crease in labor growth compared to a mean labor growth rate of 49.3 percent (37.8 percent median) for that sample.15 A one-standard-deviation increase in stock market capitalization/GDP in the high-income group is associated with a 7.3 percentage point increase in labor growth compared to the 29 percent mean (24 percent median) growth rate for that sample. Coefficients for the stock market capitalization and private credit variables are similar to those from the respective subsamples when we interact them with the dummy vari- ables for poor and rich regions (see model 4).16 14. In unreported specifications, we include industry dummy variables. The qualitative results are similar to those presented here. 15. Bear in mind that all labor growth rates are based on a two-year period. 16. Because the number of firms differs across countries in table 1, one might be concerned that countries with more firms are weighted too heavily in the regressions. We therefore reran our regressions weighting each firm by 1/(number of firms surveyed in that country). We also ran regressions weighted by the ratio of the nonagricultural population to the average firm size in each country (larger firm size in a country decreases the weight given to its firms; greater nonagricultural population increases those weights). Because the results for the financial structure variables are quite similar in the weighted regressions to those from our unweighted specifications, we do not present the weighted regressions in the paper. 556 T A B L E 4 . OLS Regressions of Labor Growth on Financial Structure: Pooled, Poor, and Rich Samples THE WORLD BANK ECONOMIC REVIEW Pooled Poor Rich (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Ln(workerst-2 ) 2 0.060*** 2 0.050*** 2 0.054*** 2 0.065*** 2 0.074*** 2 0.024 2 0.056*** 2 0.067*** 2 0.068*** 2 0.070*** ( 2 4.742) ( 2 2.919) ( 2 4.475) ( 2 7.688) ( 2 5.095) ( 2 0.791) ( 2 3.409) ( 2 7.739) ( 2 7.557) ( 2 8.137) Ln(firm age) 2 0.119*** 2 0.120*** 2 0.110*** 2 0.108*** 2 0.112*** 2 0.119*** 2 0.113*** 2 0.106*** 2 0.111*** 2 0.102*** ( 2 8.333) ( 2 10.278) ( 2 8.853) ( 2 11.221) ( 2 9.249) ( 2 5.643) ( 2 8.038) ( 2 10.058) ( 2 9.267) ( 2 8.842) Foreign ownership 0.088* 0.062 0.070* 0.115*** 0.144*** 0.065* 0.142*** 0.119*** 0.116*** 0.120*** (1.913) (1.103) (1.700) (5.863) (3.634) (1.777) (4.568) (5.918) (5.243) (6.005) Private creditt-2 0.030 0.230 0.798*** 1.256*** 2 0.101** 2 0.218*** (0.223) (1.018) (6.875) (10.486) ( 2 1.972) ( 2 2.758) Stock mkt. cap./GDPt-2 2 0.036 2 0.119 0.513 2 0.915*** 0.013 0.216*** ( 2 0.421) ( 2 0.617) (1.160) ( 2 3.795) (0.268) (4.142) Private creditt-2  Poor 1.100*** (8.414) Private creditt-2  Rich 2 0.165* ( 2 1.943) Stock mkt. cap./GDPt-2  Poor 2 1.151*** ( 2 4.216) Stock mkt. cap./GDPt-2  Rich 0.220*** (4.086) Ln(GDP per capitat-2) 2 0.105** 2 0.040 2 0.117*** 2 0.026 ( 2 2.514) ( 2 1.491) ( 2 2.922) ( 2 0.625) Intercept 1.018*** 1.042*** 1.761*** 1.299*** 0.898*** 0.952*** 1.668*** 1.062*** 1.028*** 1.233*** (15.253) (14.504) (6.703) (8.264) (14.347) (14.527) (7.167) (10.289) (9.740) (4.268) Year dummies Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Number of observations 50,114 40,443 40,443 40,443 21,581 12,542 12,542 28,533 27,901 27,901 Adjusted R 2 0.079 0.089 0.112 0.170 0.147 0.114 0.243 0.102 0.100 0.110 Note: Standard errors clustered at the country level. *, **, and *** represent significance at the 10, 5, and 1 percent levels; t statistics are in parentheses. Cull and Xu 557 In the low-income sample, the coefficient for stock market capitalization is positive in one specification (model 6) but does not achieve statistical sig- nificance. In model 7, that coefficient is negative and significant.17 In the high-income sample, the coefficient for private credit/GDP is negative and significant in model 10, but this result may be due to the high correlation between the private credit and stock market capitalization indicators. Among the group of 61 countries in our regressions that have values for both of those indicators (which is biased toward higher-income countries because many low-income countries lack information on stock market capi- talization), the correlation is 0.66. However, when the stock market capitali- zation variable is dropped, the private credit variable remains significant in the high-income sample (model 8), although it is much smaller (in absolute value). Overall, the associations between private credit and labor growth are strong in the low- income sample, whereas stock market capitalization is not positively related to labor growth. The reverse is true among the high- income countries. Among the control variables, there is a strong negative relationship between firm age and labor growth across both high- and low-income coun- tries. There is also a significant negative association between firm size and labor growth. The share of foreign ownership in a firm is positively associat- ed with labor growth in each sample. Because we split the sample by the level of per capita income, there is less of a need to include that variable in the labor growth regressions. However, in the regressions that pool all coun- tries, that variable is negative and significant, indicating that labor growth is slower in countries with high incomes (table 4, model 3). For the subsample regressions, the coefficient for per capita income is also negative, and it is significant in the low-income sample. Perhaps more important, its inclusion does not qualitatively change the results obtained for the other variables (table 4, models 7 and 10). Instrumental Variable Regressions Because labor growth is as likely to affect financial structure (or omitted factors may drive both high labor growth and financial development), it is important to investigate whether the associations between financial structure and firm labor growth in table 4 are plausibly causal. To address concerns 17. The reason for a significantly higher likelihood of missing observations for our indicator for stock market development may be that its true value is zero when it is unobserved for poor countries. To consider this possibility, in a sensitivity check (not reported in the tables), we set the stock market indicator to zero when it is missing for the low-income sample. The results for the low-income sample remain similar: private credit/GDP remains positive and significant even when stock market development is controlled for; stock market capitalization is insignificant when it is included alone or with private credit/GDP. 558 THE WORLD BANK ECONOMIC REVIEW regarding the endogeneity of the financial structure variables, we present GMM estimates in table 5.18 As described above, we use the average level of interpersonal trust from 1981 to 2006 and net petroleum exports per worker as instruments. Both are measured at the country level. The first-stage F statis- tics for the excluded instruments and Shea’s adjusted partial R 2 statistics indi- cate that the instruments are strong predictors of financial development in the poor subsample (see also the first-stage regressions in table 6). They also perform reasonably well for the full sample and for the rich subsample when private credit/GDP is the endogenous variable. They do not perform well when stock market capitalization is treated as endogenous for the full and rich- country samples. The instruments’ inability to explain stock market capitaliza- tion is not surprising because, as mentioned above, trust levels and net petro- leum exports are not strongly correlated with stock market capitalization. The main finding from the GMM regressions in table 5 is that private credit/ GDP remains positively associated with labor growth in the sample of low- income countries (see models 5 and 7), and the coefficient is similar to those from the OLS regressions in table 5. We view this finding as evidence consistent with predictions from new structural economics and Gerschenkron (1962): in the early stages of economic development, banks are better able to foster firm growth than are stock markets. The p values for Hansen’s J test of overidentify- ing restrictions are far larger than the critical values for those regressions, pro- viding additional support for the validity of our instruments. Stock market capitalization/GDP is not significantly linked to labor growth in the sample of high-income countries, in contrast to the OLS regressions. The first-stage F statistics for the excluded instruments are small whenever stock market capitalization is the endogenous financial structure variable in the GMM regressions for that subsample. Moreover, recall that our sample of “high- income” countries is closer to a sample of middle-income countries. Were we to include firms from industrialized countries, results for the market capitalization variable might differ because those countries tend to have the most advanced securities markets in the world. Finally, the instruments do not pass Hansen’s J test of overidentifying restrictions when stock market capitali- zation appears in the GMM regressions for the low-income subsample, indicat- ing that trust and net petroleum exports are not likely to affect labor growth through their effects on market capitalization alone.19 Overall, the GMM 18. We choose GMM over two-stage least squares because GMM can account for the heteroskedasticity of an unknown form (Wooldridge 2002). In practice, the two-stage least squares results were qualitatively similar to the GMM results presented here. 19. We derive similar results when we use Limited Information Maximum Likelihood estimation, which is more amenable to weak instruments. For the private credit variable, the results are very similar across the GMM and Limited Information Maximum Likelihood estimations. The Limited Information Maximum Likelihood regressions are available upon request from the authors. T A B L E 5 . GMM Estimates of Determinants of Labor Growth: The Pooled, Poor, and Rich Samples Pooled sample Poor subsample Rich subsample (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Ln(GDP per 2 0.104 2 0.458 2 0.041 2 0.040 0.104 2 0.311 capitat-2) ( 2 1.494) ( 2 0.241) ( 2 0.961) ( 2 0.508) (0.773) ( 2 0.748) Ln(workerst-2) 2 0.076*** 2 0.080 2 0.069*** 2 0.099 2 0.077*** 2 0.072*** 2 0.075*** 2 0.060*** 2 0.066*** 2 0.084*** 2 0.061*** 2 0.097** ( 2 7.396) ( 2 1.461) ( 2 7.537) ( 2 0.462) ( 2 5.106) ( 2 2.739) ( 2 5.222) ( 2 2.578) ( 2 6.967) ( 2 4.258) ( 2 4.512) ( 2 2.551) Ln(firm age) 2 0.129*** 2 0.165 2 0.109*** 2 0.106** 2 0.113*** 2 0.094*** 2 0.112*** 2 0.090*** 2 0.089*** 2 0.143*** 2 0.098*** 2 0.102*** ( 2 4.994) ( 2 1.575) ( 2 7.885) ( 2 2.454) ( 2 9.534) ( 2 5.338) ( 2 9.275) ( 2 5.461) ( 2 4.515) ( 2 3.216) ( 2 6.702) ( 2 4.179) Foreign 0.122*** 0.070 0.113*** 0.091 0.150*** 0.109*** 0.149*** 0.103*** 0.140*** 0.050 0.138*** 0.068 ownership (4.215) (0.543) (3.557) (0.444) (4.015) (2.867) (4.004) (2.693) (4.161) (0.477) (4.114) (0.659) Private creditt-2 0.342 0.363 0.871*** 0.889*** 2 0.471 2 0.603 (0.750) (0.886) (5.801) (5.537) ( 2 1.290) ( 2 1.253) Stock mkt. cap./ 2.635 3.834 1.381* 0.216 1.591 2.103 GDPt-2 (0.533) (0.212) (1.703) (0.275) (0.737) (0.628) Intercept 0.980*** 0.370 1.662*** 3.312 0.898*** 0.912*** 1.139*** 1.254*** 1.179*** 0.524 0.414 2.737 (9.407) (0.307) (4.137) (0.360) (14.928) (9.795) (3.978) (2.851) (8.031) (0.611) (0.455) (1.363) Number of 49,910 40,285 49,910 40,285 21,535 12,542 21,535 12,542 28,375 27,743 28,375 27,743 observations Adjusted R 2 0.047 . 0.087 . 0.148 0.088 0.149 0.091 0.051 . 0.020 . First-stage 2.388 0.177 3.468 0.017 8.295 9.631 8.053 12.835 0.963 0.137 3.094 0.170 F-statistics for excluded instruments Shea’s adj. 0.151 0.003 0.144 0.0002 0.473 0.219 0.444 0.266 0.050 0.001 0.021 0.010 partial R 2 Cull and Xu p value, 0.225 0.843 0.134 0.599 0.942 0.057 0.760 0.020 0.722 0.872 0.923 0.418 Hansen’s J Note: Standard errors are clustered at the country level. *, **, and *** represent significance at the 10, 5, and 1 percent levels; t statistics are in parentheses. The two financial variables ( private credit/GDP and stock market capitalization/GDP) are considered endogenous. The instruments are petroleum net 559 exports per capita (lagged) and average trust 1981– 2006. 560 THE WORLD BANK ECONOMIC REVIEW T A B L E 6 . First-Stage Results from GMM Estimates in Table 5 Pooled Poor Rich Stock Stock Stock Stock Stock Stock Private market/ Private market/ Private market/ Private market/ Private market/ Private market/ creditt-2 GDPt-2 creditt-2 GDPt-2 creditt-2 GDPt-2 creditt-2 GDPt-2 creditt-2 GDPt-2 creditt-2 GDPt-2 coef./t coef./t coef./t coef./t coef./t coef./t coef./t coef./t coef./t coef./t coef./t coef./t Ln(firm age) 0.044*** 0.019* 0.016* 0.002 0.005 0.001 0.004 2 0.002 0.048** 0.018 0.011 2 0.003 (3.028) (1.755) (1.677) (0.232) (0.545) (0.196) (0.446) ( 2 0.556) (2.350) (1.320) (0.826) ( 2 0.296) Foreign 2 0.005 0.016 0.008 0.011 2 0.034** 2 0.013 2 0.033** 2 0.020 0.048 0.039 0.032 0.031 ownership ( 2 0.131) (0.523) (0.230) (0.402) ( 2 2.492) ( 2 1.401) ( 2 2.045) ( 2 1.349) (0.900) (1.036) (0.674) (0.943) Ln(workerst-2) 0.017* 0.010 0.010 0.012 0.019*** 0.007* 0.016*** 0.006* 0.003 0.008 0.015 0.014 (1.736) (1.151) (1.367) (1.530) (2.650) (1.904) (2.586) (1.839) (0.230) (0.719) (1.121) (1.264) Trust 0.015** 0.001 0.013* 2 0.000 0.018*** 0.001 0.017*** 2 0.002 0.012 2 0.001 0.005 2 0.005 (2.162) (0.252) (1.818) ( 2 0.023) (3.513) (0.336) (3.294) ( 2 0.558) (1.056) ( 2 0.095) (0.480) ( 2 0.515) Petroleum net 2 0.076* 0.007 2 0.080*** 0.005 2 0.303** 2 0.387*** 2 0.339* 2 0.458*** 2 0.069 0.015 2 0.055 0.025 export per worker ( 2 1.720) (0.205) ( 2 2.608) (0.157) ( 2 1.985) ( 2 2.811) ( 2 1.937) ( 2 3.284) ( 2 1.357) (0.359) ( 2 1.295) (0.580) Ln(GDP per 0.142*** 0.105*** 0.055 0.072 0.243*** 0.144** capitat-2) (4.289) (3.298) (1.410) (1.528) (3.273) (2.561) Intercept 2 0.121 0.206* 2 1.013*** 2 0.510*** 2 0.210** 0.071 2 0.518** 2 0.338 0.077 0.360** 2 1.706*** 2 0.696* ( 2 1.030) (1.657) ( 2 6.042) ( 2 3.298) ( 2 2.162) (1.479) ( 2 2.239) ( 2 1.248) (0.357) (1.977) ( 2 3.729) ( 2 1.716) Year dummies Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Number of 49,910 40,285 49,910 40,285 21,535 12,542 21,535 12,542 28,375 27,743 28,375 27,743 observations Adjusted R 2 0.250 0.042 0.430 0.162 0.675 0.597 0.687 0.643 0.135 0.050 0.278 0.118 Cull and Xu 561 regressions do not provide support for a causal link between stock market capitalization and labor growth among firms in high-income countries.20 Rajan and Zingales Methodology Another widely used method for identifying a causal link between financial sector development and growth comes from Rajan and Zingales (1998), who focus on how industry-level growth rates vary with both dependence on exter- nal sources of financing and the level of financial sector development in a country.21 The key identifying assumption is that capital markets in the United States are relatively frictionless; thus, an industry’s reliance on external finance in that country is a reliable indication of its true technological demand for ex- ternal financing. Industry-level dependence on external finance computed from U.S. data is therefore the key exogenous source of variation in their analysis. The equation that we estimate is as follows: DLijk ¼ a þ bj þ gk þ dEXTDEPj þ uEXTDEPj à FINk þ 1ijk ð2Þ where DL is the percentage change in the number of workers employed by firm i in industry j in country k over the two years prior to the enterprise survey, as described above. Dummy variables for each industry are represented by bj, and gk represents dummies for each country. EXTDEP is the industry-level depen- dence on external finance based on U.S. data, which is equal to external finance (capital expenditures minus cash flow from operations) divided by capital expenditures, as defined in Rajan and Zingales (1998). We interact that variable with FIN, which represents our indicators of financial development ( private credit/GDP and stock market capitalization/GDP) in each country. A positive, significant coefficient for that interaction term is supportive of the notion that the labor growth of firms in externally dependent industries is faster than other industries in countries with relatively well-developed financial sectors. Because our aggregate variable now varies at the country-industry level, we cluster our standard errors at that level. Because this approach focuses on within-industry, between-country differenc- es in labor growth, it is complementary to our instrumental variable regressions (which focus on cross-country differences). Moreover, the inclusion of country and industry dummy variables makes this approach less susceptible to criticisms regarding omitted variable bias and model misspecification than more traditional 20. When we replace the stock market capitalization variable with the value of shares traded in the stock market (divided by GDP), we do not find a significant positive link with labor growth in the high-income sample. We do, however, continue to find a positive relationship between private credit/ GDP and labor growth in the low-income sample when the stock market turnover variable replaces the market capitalization variable in our regressions (results not reported). 21. See, for example, Fisman and Love (2007) and Kroszner, Laeven, and Klingebiel (2007) for applications of this method and Xu (2011) for a summary of the use of this method in identifying the effects of the business environment in general. 562 THE WORLD BANK ECONOMIC REVIEW cross-country regressions. Rajan and Zingales relied on industry-level data because these were the most disaggregated comprehensive data on growth at that time. In contrast, our surveys enable us to examine firm-level growth in a consis- tent manner across countries. Finer disaggregation should improve the precision of our estimates relative to industry-level approaches. Our main finding is that the interaction between industry-level dependence on external finance and private credit/GDP is significantly positively linked to labor growth in the sample of low-income countries (table 7). By contrast, there is no significant relationship between labor growth and the interaction between stock market development and dependence on external finance in the sample of low-income countries. This same pattern is obtained in our instru- mental variable regressions. Our finding indicates that when evaluated at the median industry level of external finance (0.24), a country at the 90th percen- tile of banking development would have a labor growth rate that is higher by 11.9 percentage points than a country at the 10th percentile of banking devel- opment.22 This increase in the labor growth rate is equivalent to 28 percent of the standard deviation of labor growth for the poor country sample. As in our instrumental variable regressions, we find no significant relation- ship between labor growth and the interaction between industry-level depen- dence on external finance and indicators of financial development within the sample of high-income countries. Although neither the instrumental variables nor the Rajan and Zingales approaches can definitively establish a causal link between banking sector development and labor growth in the early stages of development, both approaches yield similar results and are similar to the baseline OLS results. At a minimum, we consider the persistent contrasts in the correlations between the high- and low-income samples across estimation methods to be noteworthy. I V. I N T E R A C T I O N S B E T W E E N F I R M C H A R A C T E R I S T I C S AND FINANCIAL STRUCTURE In the previous section, we presented evidence consistent with a positive rela- tionship between private credit/GDP and firm growth in poorer countries. To obtain a better idea of what types of firms are likely to benefit from bank finance in these poorer countries,23 we interact the financial structure variables with firm characteristics using the following equation: DLij ¼ a þ bFIRMij þ gFINj þ dFIRMij FINj þ 1ij : ð3Þ The notation is the same as in equation (1). We focus first on the relationship between firm size, financial structure, and firm growth. Small firms are defined 22. That is, 1.302  0.24  (0.427 – 0.047). 23. We do not find strong and consistent patterns for the rich subsample. T A B L E 7 . Rajan and Zingales Approach: The Effect of Financial Development on Labor Growth by Dependence on External Finance Poor Rich (1) (2) (3) (4) (5) (6) (7) (8) Ln(Lt-2) 2 0.070*** 2 0.024 2 0.074*** 2 0.044*** 2 0.080*** 2 0.081*** 2 0.080*** 2 0.081*** ( 2 3.126) ( 2 0.708) ( 2 7.172) ( 2 3.444) ( 2 7.664) ( 2 7.535) ( 2 16.897) ( 2 16.722) Ln(firm age) 2 0.115*** 2 0.127*** 2 0.117*** 2 0.122*** 2 0.111*** 2 0.111*** 2 0.111*** 2 0.111*** ( 2 10.236) ( 2 7.602) ( 2 13.317) ( 2 10.889) ( 2 8.526) ( 2 8.368) ( 2 14.904) ( 2 14.565) Foreign ownership 0.145** 0.063 0.139*** 0.045 0.154*** 0.150*** 0.134*** 0.134*** (2.508) (1.326) (4.844) (1.412) (5.714) (5.438) (6.781) (6.693) Ln(GDP per capitat-2) 0.011 0.161 0.005 0.508*** 2 0.030 2 0.032 2 0.085*** 2 0.088*** (0.203) (1.263) (0.052) (3.871) ( 2 0.730) ( 2 0.753) ( 2 5.467) ( 2 5.622) External dependence 2 0.379*** 2 0.075 2 0.007 2 0.070** ( 2 3.534) ( 2 0.687) ( 2 0.202) ( 2 2.568) Private credit* t-2Ext. Dep. 1.199*** 1.302*** 2 0.066 2 0.037 (5.623) (5.604) ( 2 0.982) ( 2 1.140) Stock market cap.* t-2Ext. 0.367 2 0.926 0.038 2 0.006 Dep. (0.701) ( 2 1.489) (0.677) ( 2 0.195) Country dummies No No Yes Yes No No Yes Yes Industry dummies No No Yes Yes No No Yes Yes Number of observations 14,910 9,153 14,910 9,153 19,124 18,758 19,124 18,758 Adjusted R 2 0.130 0.129 0.194 0.229 0.109 0.110 0.147 0.149 Notes: The table shows OLS regressions; standard errors are clustered at the country-industry level. *, **, and *** represent significance at the 10, 5, and 1 percent levels; t statistics are in parentheses. External dependence is the industry-level dependence on external finance based on U.S. data, which is equal to external finance (capital expenditures minus cash flow from operations) divided by capital expenditures, as defined in Rajan and Zingales (1998). The industry classifications in our dataset (and the correspondence with the industry classifications in the Rajan and Zingales paper) are as follows: 1, tex- Cull and Xu tiles (same); 2, leather (same); 3, garments (same); 5, food (same); 7, metals and machinery (metal products); 8, electronics (electric machinery, which contains electronic components); 9, chemicals and pharmaceutics (average of “drugs” and “other chemicals”); 11, wood and furniture (same); 12, nonme- tallic and plastic materials ( plastic products); 15, auto and auto components (motor vehicles); 16, other manufacturing (other industries); 23, construction or transportation (transportation equipment). The following industries (mostly services) in our data are coded as missing because external financial dependence measures are not available for them in Rajan and Zingales (1998): other unclassified, telecommunications, accounting and finance, advertising and marketing, other services, retail and whole- 563 sale trade, hotels and restaurants, real estate and rental services, and mining and quarrying. 564 THE WORLD BANK ECONOMIC REVIEW as those with 10 or fewer workers; medium-sized firms have from 11 to 50 workers; and large firms have more than 50 workers. Note that our definitions of size cutoffs are somewhat smaller than those in many studies of mature economies, reflecting the fact that our sample is composed of developing and transitional countries, which generally feature more firms than mature economies. In models in which private credit/GDP is the financial structure variable, the coefficients for medium- and large-size are negative and significant, indicating that the growth rates of small firms are higher than for others.24 The negative coefficients for medium- and large-size are offset, however, by the positive co- efficients on the interactions between firm size and private credit/GDP (table 8).25 Those coefficients imply that the labor growth rates for large firms in the low-income sample only begin to exceed those for small firms when private credit reaches 30 –38 percent of GDP. Within the low-income sample, only approximately one-fifth of the countries meets or exceeds that standard. This finding suggests that the positive relationships between private credit/ GDP and labor growth for the poorer countries reported in the previous section were driven by large firms in a small set of countries with well-developed credit markets. This pattern would seem to be at odds with the predictions of new structural economics in which banks are hypothesized to be well positioned to aid small-scale entrepreneurs in the earliest stages of development.26 We next divide the firms by age, with old, middle-age, and young correspond- ing to the top, middle, and bottom thirds of the age distribution in our sample (table 8). In the low-income sample, labor growth declines with firm age, indi- cating that better-established firms are less likely to be in an expansionary phase that requires the hiring of new workers.27 Moreover, because we compute labor growth on a percentage basis as incremental gains in employment affect small firms to a much greater extent than they do large ones. Within the low-income sample, the beneficial effects of private credit development on the labor growth of firms appear to be largely independent of firm age. The coefficients on all of the interactions between firm age and private credit in both the OLS and GMM regressions are positive, significant, and of almost identical magnitude. A similar pattern emerges when we divide the firms into high-, mid-, and low-capital intensity, corresponding to the top, middle, and bottom thirds of the sample as measured by the ratio of capital to labor at the industry level. Within the sample 24. These coefficients are not shown in table 8 but are discussed in the table notes. 25. In the GMM estimation, the interaction terms of financial structure variables with firm characteristics are considered endogenous. The instruments consist of the interaction of our old instrumental variables with firm characteristics. 26. We acknowledge that well-developed banks might have enabled these firms to grow from small to large over time. 27. See Table 8 notes for the coefficients of the old and middle-aged dummy variables. The same pattern is also evident for the high-income sample, though results for that sample are not shown in table 8. Cull and Xu 565 T A B L E 8 . Private Credit and Labor Growth by Firm Characteristics, Poor Subsample OLS GMM Private credit by firm size Priv.  Small ( 10 workers) 0.533*** 0.557 (2.823) (1.511) Priv.  Medium (11– 50 workers) 0.795*** 0.918*** (5.857) (6.097) Priv.  Large ( . 50 workers) 0.942*** 1.041*** (10.173) (8.592) Private credit by firm age Priv.  Young 0.830*** 0.812*** (7.113) (4.878) Priv.  Middle-aged 0.850*** 0.900*** (6.870) (5.531) Priv.  Old 0.797*** 0.870*** (6.074) (5.919) Private credit by firm capital intensity Priv.  Low 0.808*** 0.827*** (5.621) (4.961) Priv.  Medium 0.875*** 0.876*** (7.195) (7.090) Priv.  High 0.810*** 0.761*** (7.324) (3.814) Notes: This table shows regression coefficients for the interaction between private credit/GDP (twice-lagged) and various firm characteristics (twice-lagged size, age, and capital intensity). The interaction terms for each characteristic are included in separate regressions. In addition to the firm characteristics, the regressions also control for the share of foreign ownership in the firm and the logarithm of GDP per capita in each country. Standard errors are clustered at the country level. *, **, and *** represent significance at the 10, 5, and 1 percent levels. Firm characteristics enter the regressions not only as part of an interaction with private credit but also on their own. For reference, the coefficients for the medium and large firm dummies for the OLS regression described in the top third of the table are 2 0.244*** and 2 0.359***, respectively. The coeffi- cients for the middle-aged and old dummies for the OLS regression described in the middle third of the table are 2 0.107*** and 2 0.184***, respectively. The coefficients for the medium and high capital intensity dummies for the OLS regression described in the bottom third of the table are 2 0.018 and 2 0.106**, respectively. of lower-income countries, neither young nor less capital-intensive firms appear to benefit from banking sector development to a greater extent than others in terms of labor growth. To summarize, contrary to predictions from new structural economics that banks should disproportionately benefit small-scale entrepreneurs and manu- facturers in the early stages of economic development, our results indicate that the labor growth rates of larger firms are increasing in the level of private credit to GDP. Although banks appear to be more important than stock markets for generating firm growth in low-income economies, they do not seem to target the clients that new structural economics predicts would be the most natural. However, we cannot rule out the possibility that banks enabled a 566 THE WORLD BANK ECONOMIC REVIEW subset of small firms to grow over time. Of course, it remains an open question why only a subset of firms would have benefited from bank finance. V. A D D I T I O N A L R O B U S T N E S S C H E C K S In table 9, we present OLS regressions that control for additional country-level variables to assess the robustness of our findings regarding financial develop- ment and firm labor growth. Because job growth likely depends on the quality of the institutional environment and the general level of stability, we include two variables: an index of adherence to the rule of law and another measuring political stability. Certain industries may have experienced more rapid job growth during the period in which our surveys were conducted; thus, we also include variables that measure the shares of GDP attributable to agriculture, manufacturing, and services (the omitted category) in our analysis. Finally, we control for the stringency of labor regulations by including a measure of the costs of firing an employee. Higher costs would likely be associated with slower labor growth. Although these control variables are occasionally signifi- cant, their inclusion does not change our main finding: private credit/GDP remains significantly positively linked to labor growth within the low-income sample. Although stock market capitalization/GDP remains positively linked to labor growth in the high-income sample, the inclusion of the additional con- trols renders its coefficient insignificant. Banking sector structure, competition, and efficiency may also help to deter- mine the labor growth rates of firms. For example, it may be that labor growth in low-income countries is more closely related to the levels of competition and concentration in the banking sector than to banking sector depth (as measured by private credit/GDP). The inclusion of the banking sector concentration vari- able may also provide a more direct test of the predictions of new structural economics because that variable provides some information on the size distri- bution of banks. To examine these possibilities, we include two variables: the share of banking sector assets held by the three largest banks as a measure of concentration and the ratio of banking sector costs to income as a proxy for efficiency. Neither of these variables is significant in the regression, and the private credit variable remains significant and of similar magnitude as in our base regressions. To conserve space, we do not present these results here. VI. MECHANISMS THROUGH WHICH BANK FINANCE SPURS F I R M D E V E LO P M E N T As a final empirical exercise and to further assess the plausibility of our main finding, we examine the relationships between our indicators of financial devel- opment and other firm-level performance variables (table 10). Within the sample of low-income countries, we find that private credit/GDP is positively and significantly associated with firms’ investment rates, which indicates that T A B L E 9 . Sensitivity Checks, Industry Controls, Rule of Law, Political Stability, and Firing Costs Poor Rich coef./t coef./t coef./t coef./t coef./t coef./t coef./t coef./t coef./t coef./t coef./t coef./t Ln(Workerst-2) 2 0.073*** 2 0.036 2 0.078*** 2 0.022 2 0.076*** 2 0.028 2 0.067*** 2 0.068*** 2 0.072*** 2 0.073*** 2 0.062*** 2 0.062*** ( 2 5.235) ( 2 1.504) ( 2 4.774) ( 2 0.806) ( 2 5.062) ( 2 0.948) ( 2 7.035) ( 2 6.745) ( 2 8.429) ( 2 8.359) ( 2 7.435) ( 2 7.445) Ln(firm age) 2 0.113*** 2 0.106*** 2 0.118*** 2 0.110*** 2 0.114*** 2 0.130*** 2 0.103*** 2 0.108*** 2 0.103*** 2 0.107*** 2 0.113*** 2 0.119*** ( 2 9.493) ( 2 7.541) ( 2 9.818) ( 2 6.676) ( 2 9.428) ( 2 5.593) ( 2 10.391) ( 2 9.821) ( 2 9.681) ( 2 9.249) ( 2 10.714) ( 2 9.991) Foreign 0.142*** 0.024 0.173*** 0.075** 0.143*** 0.063* 0.127*** 0.123*** 0.139*** 0.138*** 0.100*** 0.097*** ownership (3.524) (0.397) (3.759) (2.258) (3.568) (1.686) (6.059) (5.522) (6.324) (5.921) (5.026) (4.623) Private creditt-2 0.825*** 0.859*** 0.798*** 2 0.089 2 0.095* 2 0.103* (6.949) (7.683) (6.135) ( 2 1.457) ( 2 1.789) ( 2 1.807) Stock/GDPt-2 0.107 0.614** 0.329 0.051 0.007 0.015 (0.189) (1.962) (0.785) (0.866) (0.112) (0.265) Agri./GDPt-2 0.347* 0.860* 0.410 0.779 (1.658) (1.734) (0.712) (1.274) Manufacture/ 0.002 0.026 0.001 2 0.001 GDPt-2 (0.446) (1.568) (0.421) ( 2 0.312) Rule of lawt-2 0.060 0.231 2 0.043 2 0.027 (1.101) (1.623) ( 2 1.573) ( 2 0.818) Political 2 0.082 2 0.014 2 0.063* 2 0.079** stabilityt-2 ( 2 1.357) ( 2 0.145) ( 2 1.885) ( 2 2.457) Firing costst-2 0.000 0.002* 0.001 0.001 (0.001) (1.710) (1.543) (1.007) Intercept 0.785*** 0.432 0.868*** 0.793*** 0.912*** 0.845*** 0.992*** 0.959*** 1.053*** 1.004*** 1.025*** 1.008*** (10.101) (1.644) (11.580) (7.082) (10.717) (7.874) (9.634) (9.561) (9.622) (9.341) (9.160) (8.858) Cull and Xu Year dummies Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Number of 21,490 12,542 17,859 11,417 20,729 12,542 28,052 27,420 27,826 27,390 25,740 25,108 observations Adjusted R 2 0.152 0.158 0.175 0.144 0.148 0.127 0.104 0.104 0.108 0.106 0.110 0.107 Note: The table reports OLS regressions of labor growth on indicators of financial development, firm characteristics, and country-level controls. Standard errors are clustered at the country level. *, **, and *** represent significance at the 10, 5, and 1 percent levels; t statistics are in parentheses. 567 568 THE WORLD BANK ECONOMIC REVIEW T A B L E 1 0 . Potential Mechanisms through which Financial Development Spurs Firm Development Dependent variable Explanatory variable Investment rate Training Firm Size Private credit Poor subsample 0.048*** 0.417*** 1.035*** (3.698) (4.099) (3.959) Rich subsample 0.001 2 0.108* 0.238 (0.063) ( 2 1.691) (1.488) Stock market capitalization Poor subsample 2 0.073* 2 0.502 0.495 ( 2 1.813) ( 2 1.403) (0.390) Rich subsample 2 0.008 2 0.088 0.316 ( 2 1.380) ( 2 1.030) (1.586) Notes: This table shows coefficients for indicators of financial development from OLS regres- sions of firm outcomes (investment rate, training, and firm size) on financial indicators and other controls. Each coefficient in the table comes from a separate regression. Those regressions also control for firm characteristics (twice-lagged number of workers, age, and share of foreign owner- ship in the firm) and per capita income in each country; in the case of firm size as the dependent variable, we do not control for lagged firm size. Standard errors are clustered at the country level. *, **, and *** represent significance at the 10, 5, and 1 percent levels. Training is a dummy vari- able indicating whether the firm offers training to its employees. Firm size is measured by (the log of ) total workers. the labor growth associated with banking sector development is accompanied by capital investment. Similarly, private credit/GDP is positively associated with a dummy variable indicating that the firm provides training to its employ- ees. This result suggests that bank finance is associated not only with an in- creasing number of workers at firms in low-income countries but also with the development of their human capital, presumably to achieve productivity gains. Finally, in low-income countries, private credit is positively associated with firm size measured by the number of employees. One concern was that our measure of labor growth focused on a short time horizon (two years). The pos- itive and significant relationship between private credit and firm size is consis- tent with the notion that bank finance sustains firms over longer periods, enabling them to expand over time. In contrast, neither private credit nor stock market capitalization is signifi- cantly associated with investment rates, firm size, or employee training within the sample of high-income countries. Furthermore, there is not a strong relationship between stock market capitalization and the other firm outcome variables within the sample of low-income countries. These patterns mimic those found for labor growth and thus are consistent with the beneficial effects of banking sector development in lower-income countries. Cull and Xu 569 VII. CONCLUSIONS We have combined data from detailed firm surveys with indicators of financial development and structure that cover a broader set of countries than those in previous studies to analyze the relationships between firm growth, financial structure, and the stage of economic development. In low-income countries, we find evidence that labor growth is swifter in countries with a higher level of private credit/GDP, consistent with the predictions of new structural econom- ics. This conclusion holds under two well-established estimation approaches: instrumental variable regressions and the Rajan and Zingales regression meth- odology. The plausibility of the finding of the beneficial role of banks in poorer countries is further supported by our finding that a more developed banking system is associated with higher investment rates, more employee training, and larger firm sizes only in poor countries, suggesting that banks spur both physi- cal and human capital investments in addition to increasing job growth in such countries. In high-income countries, we find that labor growth rates increase with the level of stock market capitalization using OLS regressions, which may be seen as consistent with predictions from new structural economics. However, we are unable to reproduce those results using the two additional estimation methods described above. Moreover, when we examine the types of firms that appear to benefit from well-developed private credit markets in low-income countries, we find no evidence that the small-scale firms benefit most. Rather, labor growth rates increase with the level of private credit market development for larger firms. This pattern is consistent with arguments rooted in political economy that banking systems in low-income countries serve the interests of the elite rather than providing broad-based access to financial services. We acknowledge that our approach comes with caveats and highlights the gaps in our knowledge regarding the effects of financial structure. For example, our inability to find a suitable instrument for stock market development makes it difficult to study whether market capitalization enables firms to grow more quickly in high-income countries. Again, our results using the Rajan and Zingales methodology fail to reveal any evidence of a positive link between stock market development and the labor growth of firms in industries that depend heavily on external finance. We also acknowledge that our indicators of financial structure do not provide as direct a test of the predictions of new structural economics as we would like because they do not summarize the size distribution of banks, and small banks are the institutions that are hypothe- sized to best serve small-scale manufacturers in the early stages of development. It is possible, therefore, that alternative financial structure variables would yield different results regarding labor growth. Deriving policy implications from our findings is not straightforward. For example, we are reluctant to recommend that governments actively attempt to shape the size composition of banks, with the exception of allowing sufficient 570 THE WORLD BANK ECONOMIC REVIEW entry to facilitate competition in the provision of financial services (see Barth, Caprio, and Levine 2006 on the downside of rigid entry restrictions). Moreover, we are mindful of the fact that careful empirical studies of bank branching in the United States show that it led not only to larger banks, inten- sified competition, and improved bank performance but also to substantial in- creases in the incomes of the less affluent by increasing the relative wage rates and working hours of unskilled workers.28 Although these were indirect bene- fits to the less well off, it is not clear whether those benefits were smaller than those that would have been generated by relationship lending to small business- es via smaller banks. Rather than focusing on policy advice regarding the size distribution of financial institutions in a country, a better route might be to focus on the features of the environment that make the provision of financial services possi- ble (such as adherence to the rule of law, property rights protection, effective dispute resolution mechanisms, and credit and collateral registries). Multiple studies have demonstrated that the entry rate and growth of new firms are posi- tively associated with financial development. These effects are especially strong for small firms, which report being more constrained by financial obstacles than others.29 A focus on the enabling environment might therefore be the most effective way to support the growth and entry of small firms. Finally, although the enterprise surveys that we use offer rich information across a wide range of countries, our labor growth rates can only be computed over a two-year window. We therefore rely on a snapshot of firm growth to draw inferences. Case studies of firm financing patterns and financial develop- ment in low-income countries over longer periods are likely to offer a more complete picture of how firms grow. To the extent that some low-income coun- tries have been better able to achieve financial inclusion, these case studies may shed light on how banks can foster growth across a broader spectrum of firms during the early stages of development. However, to the extent that the political economic incentives that limit the inclusiveness of financial systems in low-income countries cannot be overcome, it would be wise to explore how alternative policies and intermediaries could better achieve that aim. REFERENCES Acemoglu, D., S. Johnson, and J.A. Robinson. 2001. “The Colonial Origins of Comparative Development: An Empirical Investigation.” American Economic Review 91 (5): 1369– 1401. ———. 2002. “Reversal of Fortunes: Geography and Institutions in the Making of the Modern World Income Distribution.” Quarterly Journal of Economics 117 (4): 1231– 94. 28. See Jayaratne and Strahan (1998) and Beck, Levine, and Levkov (2010). 29. See Beck, Demirgu ¨c¸ -Kunt, and Maksimovic (2005); Beck, Demirgu ¸ -Kunt, Laeven, and ¨c Maksimovic (2006); Klapper, Laeven, and Rajan (2006); and Laeven (2003). Cull and Xu 571 Barth, J.R., G. Caprio Jr., and R. Levine. 2006. Rethinking Bank Regulation: Till Angels Govern. New York: Cambridge University Press. Beck, T. 2011. “Finance and Oil: Is There s Resource Curse in Financial Development?” European Banking Center Discussion Paper 2011-004. Tilburg University, Tilburg, the Netherlands. Beck, T., A. Demirgu ¸ -Kunt, L. Laeven, and V. Maksimovic. 2006. “The Determinants of Financing ¨c Obstacles.” Journal of International Money and Finance 25 (6): 932–52. Beck, T., A. Demirgu ¸ -Kunt, and R. Levine. 2000. “A New Database on Financial Development and ¨c Structure.” World Bank Economic Review 14 (3): 597–605. ———. 2010. “Financial Institutions and Markets Across Countries and over Time: The Updated Financial Development and Structure Database.” World Bank Economic Review 24 (1): 77–92. ¨c Beck, T., A. Demirgu ¸ -Kunt, and V. Maksimovic. 2005. “Financial and Legal Constraints to Firm Growth: Does Firm Size Matter?” Journal of Finance 60 (1): 137– 77. Beck, T., R. Levine, and A. Levkov. 2010. “Big Bad Banks? The Winners and Losers from Bank Deregulation in the United States.” Journal of Finance 65 (5): 1637–67. Claessens, S., and N. van Horen. 2012. “Being a Foreigner among Domestic Banks: Asset or Liability?” Journal of Banking & Finance 36 (5): 1276– 1290. ¨c Demirgu ¸ -Kunt, A., E. Feyen, and R. Levine. 2011. “Optimal Financial Structures and Development: The Evolving Importance of Banks and Markets.” World Bank, mimeo. ¨c Demirgu ¸ -Kunt, A., and R. Levine. 2001. Financial Structure and Economic Growth: A Cross-Country Comparison of Banks, Markets, and Development. Cambridge, MA: MIT Press. Engerman, S., and K. Sokoloff. 1997. “Factor Endowments, Institutions, and Differential Paths of Growth among New World Economies: A View from Economic Historians of the United States.” In Stephen H. Haber, ed., How Latin America Fell Behind, 260– 304. Stanford, CA: Stanford University Press. Fisman, R., and I. Love. 2007. “Financial Development and Growth Revisited.” Journal of the European Economic Association 5 (2 –3): 470–79. ———. 2003. “Trade Credit, Financial Intermediary Development, and Industry Growth.” Journal of Finance 58 (1): 353– 74. Gerschenkron, A. 1962. Economic Backwardness in Historical Perspective, a Book of Essays. Cambridge, MA: Harvard University Press. Guiso, L., P. Sapienzia, and L. Zinglales. 2004. “The Role of Social Capital in Financial Development.” American Economic Review 94 (3): 526–56. Klapper, L., L. Laeven, and R. Rajan. 2006. “Entry Regulation as a Barrier to Entrepreneurship.” Journal of Financial Economics 82 (3): 591– 629. Knack, S., and P. Keefer. 1997. “Does Social Capital Have an economic Payoff? A Cross-Country Investigation.” Quarterly Journal of Economics 112 (4): 1252–88. Kroszner, R.S., L. Laeven, and D. Klingebiel. 2007. “Banking Crises, Financial Dependence, and Growth.” Journal of Financial Economics 84 (1): 187– 228. Jayaratne, J., and P.E. Strahan. 1998. “Entry Restrictions, Industry Evolution and Dynamic Efficiency: Evidence from Commercial Banking.” Journal of Law and Economics 41 (1): 239–74. Laeven, L. 2003. “Does Financial Liberalization Reduce Financing Constraints?” Financial Management 2003 (Spring): 5 – 34. La Porta, R., F. Lopez-de-Silanes, A. Shleifer, and R.W. Vishny. 1998. “Law and Finance.” Journal of Political Economy 106 (6): 1131–50. ´a 9 (1): Lederman, D., and W.F. Maloney. 2008. “In Search of the Missing Resource Curse.” Economı 1–58. Levine, R. 2002. “Bank-Based or Market-Based Financial Systems: Which is Better?” Journal of Financial Intermediation 11 (4): 398 –428. Lin, J. 2010. “New Structural Economics: A Framework for Rethinking Development.” World Bank Policy Research Working Paper 5197. Washington, DC. 572 THE WORLD BANK ECONOMIC REVIEW Lin, J.Y., X. Sun, and H.X. Wu. 2012. “Banking Structure, Labor Intensity, and Industrial Growth: Evidence from China.” World Bank, mimeo. Micco, A., U. Panizza, and M. Yanez. 2007. “Bank Ownership and Performance. Does Politics Matter?” Journal of Banking and Finance 31 (1): 219–41. Moulton, B. 1990. “An Illustration of a Pitfall in Estimating the Effects of Aggregate Variables on Micro Units.” Review of Economics and Statistics 72 (3): 334–38. Petersen, M.A., and R.G. Rajan. 1997. “Trade Credit: Theories and Evidence.” Review of Financial Studies 10 (3): 661– 91. Rajan, R.G., and L. Zingales. 1998. “Financial Dependence and Growth.” American Economic Review 88 (3): 559 –86. Sachs, J.D., and A. Warner. 1995. “Natural Resource Abundance and Economic Growth.” National Bureau of Economic Research Working Paper 5398. National Bureau of Economic Research, Cambridge, MA. ———. 2001. “The Curse of Natural Resources.” European Economic Review 45 (4– 6): 827– 38. Xu, L.C. 2011. “The Effects of Business Environments on Development: A Survey of New Firm-Level Evidence,” World Bank Research Observer 26 (2): 310– 40. Wooldridge, J.M. 2002. Econometric Analysis of Cross Section and Panel Data. Cambridge, MA: The MIT Press. Zak, P.J., and S. Knack. 2001. “Trust and Growth.” Economic Journal 111 (April): 295–321. Forthcoming papers in THE WORLD BANK ECONOMIC REVIEW • How Can Safety Nets Contribute to Economic Growth? Harold Alderman and Ruslan Yemtsov • Political Reforms and Public Policy: Evidence from Agricultural and Food Policies Alessandro Olper, Jan Fałkowski, and Johan Swinnen • Food Security and Storage in the Middle East and North Africa Donald F. Larson, Julian Lampietti, Christophe Gouel, Carlo Cafiero, and John Roberts • Preferential Market Access Design: Evidence and Lessons from African Apparel Exports to the United States and the European Union Jaime de Melo and Alberto Portugal-Perez • Network Proximity and Business Practices in African Manufacturing Marcel Fafchamps and Måns Söderbom • The Impact of Financial Literacy Training for Migrants John Gibson, David McKenzie, and Bilal Zia • Policy Barriers to International Trade in Services: Evidence from a New Database Ingo Borchert, Batshur Gootiiz, and Aaditya Mattoo THE WORLD BANK 1818 H Street, NW Washington, DC 20433, USA World Wide Web: http://www.worldbank.org/ E-mail: wber@worldbank.org