53728 THE WORLD BANK ECONOMIC REVIEW Trade Li beralization and Growth: New Evidence Romain Wacziarg and Karen Horn Welch Comprehensive Wealth and Future Consumption: Accounting fo r Population Growth Susana Ferreira, Kirk Hamilton, and Jeffrey R. Vincent Comparison of Net Benefits of Incentive-Based and Command and Control Environmental Regulations: The Case of Santiago, Chile Raul O'Ryan and Jose Miguel Slinchez Women's Power, Conditional Cash Transfers, and Schooling in Nica ragua Seth R. Gitter and Bradford L. Barham Does Aid for Education Educate Children? Evidence from Panel Data Axel Dreher, Peter Nunnenkamp, and Rainer Thiele World Bank Lending and Financial Sector Development Robert Cull and Laurie Effron HIV Pandemic, Medical Brain Drain, and Economic Development in Sub-Saharan Africa Alok Bhargava and Frederic Docquier Foreign Direct Investment, Access to Finance, and Innovation Activity in Chinese Enterprises Sourafel Girma, Yundan Gong, and Holger Gorg THE WORLD BANK ECONOMIC REVIEW EDITOR Jaime de Melo, University cfGeneva ASSISTANT TO THE EDITOR Marja Kuiper EDITORIAL BOARD Chong-En Bai, Tsinghua University, China Karla Hoft: World Bank Jean-Marie Baland, UniversityofNamur, Hanan Jacoby, World Bank Belgium Elizabeth M. King, World Bank Kaushik Basu, Cornell University, USA Aart Kraay, World Bank Alok Bhargava, Houston University, USA Justin Yifu Lin, World Bank Fram;ois Bourguignon, Paris School of Thierry Magnac, Universiti de Toulouse I, Economics, France France Kenneth Chomitz, World Bank Jonathan Morduch, New York University, Luc Christiaensen, World Bank USA Stijn Claessens, International iWonetary Fund Jacques l\1orisset, World Bank Maureen Cropper, Uni'l.JersityofMaryland, Juan-Pablo Nicolini, UniversidadTorcuato USA di Tella, Argentina Jishnu Das, World Bank Boris Pleskovic, World Bank Stefan Dercon, University of Oxford, UK Martin Rama, World Bank Shantayanan Devarajan, World Bank l\1artin Ravallion, World Bank Augustin Kwasi Fosu, World Institute for Elisabeth Sadoulet, University ofCalifornia, Development Economics Research Berkeley, USA (WIDER), Helsinki, Finland Joseph Stiglitz, Columbia Unieversity, USA Jan VVillem Gunning, Free University, Jonathan Temple, University ofBristol, UK The Netherlands Ruslan Yemtsov, World Bank The World Bank Economic Review is a professional journal for the dissemination of World Bank-sponsored and other research that may inform policy analysis and choice. 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 quantitative development policy analysis, emphasizing policy relevance and operational aspects of economics, rather than primarily theoretical and methodological issues. It is intended for readers familiar with economic theory and analysis but not necessarily proficient in advanced mathematical or econometric techniques. Articles illustrate how professional research can shed light on policy choices. Consistency with World Bank policy plays no role in the selection of articles. Articles are drawn from work conducted by \Vorld Bank staff and consultants and by outside researchers. :"Ion-Bank contributors are encouraged to submit their work. Before being accepted for publication, articles are reviewed by three reterees-one from the vVorld Bank and two from outside the institution. Articles must also be endorsed by two members of the Editorial Board before final acceptance. For more information, please visit the \Neb sites ofthe Economic Review at Oxford University Press at www.wber.oxfordjournals.org and at the \Vorld Bank at www.worldbank.org/researchljournals. Instructions for authors wishing to submit articles are available online at WW\v.wber.oxfordjoumals.org. Please direct all editorial correspondence to the Editor at wber@Worldbank.org. THE WORLD BANK ECONOMIC REVIEW Volume 22·2008· Number 2 Trade Liberalization and Growth: New Evidence 187 Romain Wacziarg and Karen Horn Welch Comprehensive Wealth and Future Consumption: Accounting for Population Growth 233 Susana Ferreira, Kirk Hamilton, and Jeffrey R. Vincent Comparison of Net Benefits of Incentive-Based and Command and Control Environmental Regulations: The Case of Santiago, Chile 249 Raul O'Ryan and Jose Miguel Sanchez Women's Power, Conditional Cash Transfers, and Schooling in Nicaragua 271 Seth R. Gitter and Bradford L. Barham Does Aid for Education Educate Children? Evidence from Panel Data 291 Axel Dreher, Peter Nunnenkamp, and Rainer Thiele World Bank Lending and Financial Sector Development 315 Robert Cull and Laurie Effron HIV Pandemic, Medical Brain Drain, and Economic Development in Sub-Saharan Africa 345 Alok Bhargava and Frederic Docquier Foreign Direct Investment, Access to Finance, and Innovation Activity in Chinese Enterprises 367 Sourafel Girma, Yundan Gong, and Holger Gorg 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 22, 3 Issues, 2008): Academic libraries-Print edition and site-wide online access: US$182/£118/182, Print edition only: US$173/£112/173, Site-wide online access only: US$173/£1121173; Corporate-Print edition and site-wide online access: US$273/£177/273, Print edition only: US$259/£168/259, Site-wide online access only: US$259/£168/259; Personal-Print edition and individual online access: US$59/£38/59. US$ rate applies to US & Canada, Euros 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.orglsubscriptions. Readers with mailing addresses in non-DE CD 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 considered 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. BACK ISSUES: The current year and two previous years' issues are available from Oxford University Press. Previous 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@oxfordjoumals.org. Tel: +44 (0)1865 353907. Fax: +44 (O)1865 353485. In the Americas, please contact: Journals Customer Service Department, Oxford University Press, 2001 Evans Road, Cary, NC 27513, USA. E-mail: jnlorders@oxfordjournals.org. Tel: l800) 852-7323 (toll-free in USA/Canada) or (919) 677-0977. Fax: (919) 677-1714. In Japan, please contact: Jouroals Customer Service Department, Oxford University Press, Tokyo, 4-5-10-8F Shiba, Minato-ku, Tokyo, 108-8386, japan. E-mail: okudaoup@po.iijnet.or.jp. Tel: +81 3 5444 5858. Fax: +81 3 3454 2929. POSTAL INFORMATION: The World Bank Economic Review (ISSN 0258-6770) is published three times a year, in Feb., jun., and OCt., by Oxford University Press for the International Bank for Reconstruction and DeveiopmentlTHE 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. Communications regarding original articles and editorial management should he addressed to The Editor, The World Bank Economic Review, The World Bank, 3, Chemin Louis Dunant, CP66 1211 Geneva 20, Switzerland. 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.orgljnls/permissions. ADVERTISING: Inquiries about advertising should be sent to Helen Pearson, Oxford Journals Advertising, PO Box 347, Abingdon OX14 1GJ, UK. E-mail: helen@oxfordads.com. Tel: +44 (0)1235201904. Fax: +44 (0)8704 296864. 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 DevelopmentlrHE 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. PAPER USED: The World Bank Economic Review is printed on acid-free paper that meets the minimum requirements of ANSI Standard Z39A8-1984 (Permanence of Paper). INDEXING AND ABSTRACllNG: The World Bank Economic Review is indexed andlor abstracted by CAB Abstracts, Current Contents/Social and Behavioral Sciences, Journal of Economic LiteraturelEconLit, PAIS International, RePEc (Research in Economic Papers), and Social Services Citation Index. COPYRIGHT 2008 The International Bank for Reconstruction and DeveiopmenthHE WORLD BANK All rights reserved; no part of this publication may be reproduced, s~ored in a retrieval system, or trans mitted 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 WIP 9HE, or in the USA by the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923. Trade Liberalization and Growth: New Evidence Romain Wacziarg and Karen Horn Welch A new data set of on openness indicators and trade liberalization dates allows the 1995 Sachs and Warner study on the relationship between trade openness and econ omic growth to be extended to the 19905. New evidence on the time paths of econ omic growth, physical capital investment, and openness around episodes of trade policy liberalization is also presented. Analysis based on the new data set suggests that over the 1950-98 period, countries that liberalized their trade regimes experienced average annual growth rates that were about 1.5 percentage points higher than before liberalization. Postliberalization investment rates rose 1.5-2.0 percentage points, confirming past findings that liberalization fosters growth in part through its effect on physical capital accumulation. Liberalization raised the average trade to GDP ratio by roughly 5 percentage points, suggesting that trade policy liberalization did indeed raise the actual level of openness of liberalizers. However, these average effects mask large differences across countries. JEL codes: Fl, F4, 04 Many developing countries have embarked on programs of external economic liberalization in recent decades. In 1960, just 22 percent of all countries, repre senting just 21 percent of the global population, had open trade policies, in the sense defined by Sachs and Warner (1995). By 2000, some 73 percent of countries, representing 46 percent of the world's population, were open to international trade (figure 1). 1 Romain Wacziarg (corresponding author) is an associate professor of economics at the Stanford Graduate School of Business; his email address is wacziarg@gsb.stanford.edu. Karen Horn Welch is director, Domestic Public Equity, at the Stanford Management Company, in Menlo Park, California; her email address is karen.welch@stanford.edu. The authors thank Jaime de Melo, John McMillan, Paul Segerstrom, and Jessica Wallack; an anonymous referee; and seminar participants at Columbia University, Harvard University, the University of Colorado at Boulder, and Stanford University for useful comments. They thank Peter Henry and Jonas Vlachos and for sharing their data. This article was written while Romain Wacziarg was the Edward Teller National Fellow at the Hoover Institution. 1. The main reason for the discrepancy between the share of countries that are open and the share of the world's population living in open countries is that as of 2000, the world's two largest countries, China and India, remained essentially closed. Sachs-Warner (1995) classify India as open as of 1994. The authors revisited this issue and could not confirm their finding. In fact, in terms of both policy indicators and trade volumes, China appears to be twice as open as India. This issue is discussed later in the article and in an appendix to the working version of this paper (Wacziarg and Welch 2003). For an in-depth comparison of the trade regimes of India and China, see Wacziarg (2003). THE WORLD BANK ECONOMIC REVIEW, VOL. 22, No.2, pp. 187-231 doi: 10.1 093/wber/lhn007 Advance Access Publication June 3, 2008 The Author 2008. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development I THE WORLD BANK. All rights reserved. For permissions, pblse e-mail: journals. permissions@oxfordjournals.org 187 188 THE WORLD BANK ECONOMIC REVIEW FIGURE 1. Openness to Trade, 1960-2000 Note: Openness is defined according to the Sachs and Warner (1995) criteria. Sample includes 141 countries. ----Percentage of countries that are open to trade 70%~::=f~~~;~~~I~~~~··~~~------~----------------~~~~- I of world population living in countries that are open to trade 1: " ~ 40% O%~-- -~------~------~----.~ ~------~----~-----~ ... ...------ ..~ 1960 1965 1970 1975 1980 1985 1990 1995 2000 Source: Authors' analysis based on data described in the text. The effect of this trend toward greater trade policy openness on per capita income growth is the topic of a large body of research. Until recently, a growing academic consensus had emerged that both trade policy openness and higher ratios of trade volumes to gross domestic product (GDP) were positively correlated with growth, even after controlling for a variety of other growth determinants. Attempts to establish a causal link also suggested a positive impact of trade. 2 In a sweeping critical survey of this literature, Rodriguez and Rodrik (2000) argue that these findings are less robust than claimed, because of difficulties in measuring openness, the statistical sensitivity of the specifications, the collinearity of protectionist policies with other bad policies, and other econometric difficulties. Further research on this import ant topic is called for in view of the doubts their study created about the linkages between trade openness and growth. 3 Taking over where Rodriguez and Rodrik (2000) left off, the article pursues three goals. The first goal is to update the Sachs-Warner classification by 2. Particularly noteworthy are the contributions of Edwards (1992), Dollar (1992), Ben-David (1993), Sachs and Warner (1995), Ades and Glaeser (1999), and Alesina, Spolaore, and Wacziarg (2000). Among studies trying to establish a causal link running from openness to growth or income levels, see Frankel and Romer (1999), who measure openness by trade volumes, and Wacziarg (2001), who captures openness by using a composite trade policy index. 3. Harrison and Hanson (1999) also criticize the Sachs-Warner classification, in a spirit similar to that of Rodriguez and Rodrik. Their criticisms are revisited in detail later in the article. Wacziarg and Horn Welch 189 presenting a comprehensive cross-country database of trade indicators (tariffs, nontariff barriers, and other measures of trade restrictions) and policy liberali zation dates for the 1990s. The second goal is to extend the Sachs-Warner empirical results on outward orientation and growth to the 1990s. The third, and most important, goal is to exploit the timing of liberalization in a within country setting to identify the changes in growth, investment rates, and open ness associated with discrete changes in trade policy. The availability of almost 50 years of data makes it possible to compare the performance of countries under liberalized and nonliberalized regimes across time. The main empirical analysis presents estimates for the within-country response of per capita income growth, the investment rate, and the ratio of imports plus exports to GDP to trade liberalization, controlling for country and time effects. New evidence is presented on the within-country path of growth in relation to the date of major trade policy changes. Evidence from the large sample is supplemented by a discussion of several developing countries' experiences with trade reform. The cross-sectional results confirm recent criticisms of the Sachs-Warner findings by showing that these were sensitive to the openness classification used in the 1970-89 period and do not hold for the 1990s. The vast majority of countries in the sample used here are classified as having been open during the 1990s; a simple dichotomous indicator of openness no longer discriminates between slow- and fast-growing countries. The findings here suggest that researchers should exercise caution when using simple dichotomous policy indicators such as the Sachs-Warner dummy variable. However, the dates of trade liberalization--collected by Sachs-Warner from a comprehensive survey of a broad country-specific case literature and updated here to the late 1990s can be used to estimate the within-country growth and investment effects of trade policy liberalization. In contrast to the cross-sectional findings presented here, the results based on within-country variation suggest that over time the effects of increased policy openness within countries are positive, economically large, and statistically significant. The article examines a subsample of developing countries for which detailed information was collected on the broader economic and political context of trade reform. It then interprets the large sample results in the context of these country case studies. This effort reveals two lessons. First, the extent to which per capita income growth changed after trade reforms varied widely across countries. While the average effect obtained in the large sample is positive, roughly half of the countries experienced zero or even negative changes in growth following liberalization. Second, generalizations about the factors that may explain these differences are difficult to draw. The institutional environ ment of countries, the extent of political turmoil, the scope and depth of econ omic reforms, and the characteristics of concurrent macroeconomic policies all seem to have a role to play, to varying degrees in different countries. While this article paints a picture that is highly favorable to outward-oriented policy 190 THE WORLD BANK ECONOMIC REVIEW reforms on average, it cautions against one-size-fits-all policies that disregard local circumstances. The article is organized as follows. Section I presents an updated data set of lib eralization dates and policy openness indicators and uses the data to replicate the Sachs-Warner growth regressions. Section II presents within-country evidence on trade liberalization, growth, investment, and trade volumes and discusses the timing of these effects. Section III examines 13 country cases of trade liberaliza tion in order to illustrate the country-specific complexities that underlie the results from the larger sample. The last section provides some concluding remarks. I. TRADE LIBERALIZATION IN THE 1990s This section updates the Sachs-Warner classification and results. It also addresses the Rodrfguez and Rodrik critique of their study. The Sachs- Warner Criteria An update of the Sachs-Warner classification is called for not only because of the problems with their classification of open and closed countries but also because the underlying data-on tariffs, nontariff barriers, exchange rate black market premia, socialist economic systems, and export marketing boards-are of independent interest. This section presents a comprehensive database of these variables for the 1990s. It also presents the results of a painstaking check of the Sachs-Warner classification of openness and updates their data on trade policy openness through 2000. Sachs-Warner constructed a dummy variable for openness based on five individual dummy variables for specific trade-related policies. A country was classified as closed if it displayed at least one of the following characteristics: (1) Average tariff rates of 40 percent of more (TAR). (2) Nontariff barriers covering 40 percent or more of trade (NTB). (3) A black market exchange rate at least 20 percent lower than the official exchange rate (BMP). (4) A state monopoly on major exports (XMB). (5) A socialist economic system (as defined by Kornai 1992) (SOC). Tariff and nontariff barriers restrict trade directly. A black market premium (BMP) on the exchange rate could have effects equivalent to formal trade restrictions. If, for example, exporters have to purchase foreign inputs using foreign currency obtained on the black market but remit their foreign exchange receipts from exports to the government at the official exchange rate, the BMP acts as a trade restriction. On the basis of Lerner symmetry between import tariffs and export taxes, Sachs-Warner also included the state monopoly on exports criterion as a trade restriction. The socialist regime dummy variable accounts for the trade-limiting aspects of centrally planned economies. Wacziarg and Horn Welch 191 It is important to distinguish the Sachs-Warner dummy variable for openness, which pertains to the 1970s and 1980s, from the Sachs-Warner liberalization dates, which extend from 1950 to 1994 and were compiled independently using a different methodology. While the Sachs-Warner dummy variable was based on the five criteria cited above, the dates of liberalization were obtained from a comprehensive survey of country case studies of liberalization. Where possible, the criteria used to construct the cross-sectional dummy variable for the 1970s and 1980s were used to establish the date of liberalization. Data limitations and lack of consistency in the definitions of the available measures of trade restric tions across time periods, however, prevented Sachs-Warner from using their five criteria to establish the dates of liberalization. 4 The Sachs-Warner methodology was followed as closely as possible in the update presented here. An Openness Dummy Variable for the 1990s The sample is based on the 118 countries included in the Sachs-Warner data set. 5 The sample also includes the new data on 23 Eastern European countries and former Soviet republics included in version 6 of the Penn World Tables (Heston, Summers, and Aten 2002). The openness dummy variable (OPEN90-99) was based on the five criteria Sachs-Warner use, in order to maintain as much consistency as possible between their data set and the data used here. Data limitations made it impossible to update their dummy variable to the 1990s based on exactly the same data, however. 6 The main differences between the two data sets include the following: (1) Because of data availability problems, unweighted tariff data were used here; Sachs-Warner used own import-weighted data. Countries that exceed the TAR threshold in the new data set based on unweighted data could conceivably not exceed the threshold based on weighted average data. This is unlikely to be a big problem, however, because the use of unweighted rather weighted tariffs does not result in countries being classified differ ently in the subsample in which both measures are available. (2) Nontariff barrier data comparable to those used by Sachs-Warner are hard to obtain. Sachs-Warner used average nontariff barrier data for 1985-88 from the Barro-Lee data set, itself based on data from the United Nations Conference on Trade and Development (UNCTAD). Their data cover only 29 countries for the period 1995-98. Where 4. As Sachs-Warner write (p. 24), "Our choice of dating is surely subject to further refinement .... We relied on a wide array of secondary sources, which sometimes contradicted each other." The appendix to their article describes how they compiled their dates of liberalization and identifies the corresponding data sources for each country in their sample. A similar appendix for the updated dates is available in the working paper version of this study (Wacziarg and Welch 2003). 5. Sachs-Warner characterized the openness status of only 111 of these countries. 6. The data sources are detailed in Wacziarg and Welch 2003. The full data set is available in electronic format at www.stanford.edul-wacziarglpapersum.html. Table I-A displays the data used to construct the updated openness indicator. 192 THE WORLD BANK ECONOMIC REVIEW comparable data on nontariff barriers were missing, the countries were classified based only on the other four Sachs-Warner criteria. The limited availability of nontariff barrier data for the 1990s based on a consistent definition required the compilation of an additional nontar iff barrier data set, which may be independently useful to researchers. In addition to the 1995 - 98 average core nontariff barrier data used in the analysis, the data set contains average core nontariff barrier data for 1989-94 and 1999 data for all nontariff barriers. 7 (3) Sachs-Warner relied on an export marketing index from a World Bank study of African countries (Husain and Faruqee 1994) as the basis for their XMB variables and on the Kornai (1992) classification of socialist countries as the basis for their SOC dummy variable. In the absence of updated indices from single sources, the same methodology could not be used with the updated data. The XMB and SOC dummy variables were therefore obtained from a comprehensive review of country case studies. The XMB criterion is no longer confined to African countries (as it was in Sachs-Warner), but applies to all countries in the updated data. The definition of an export marketing board was expanded to encompass any form of state monopoly over major exports. 8 (4) Data on the BMP from Easterly and Sewadeh (2002), the primary source for updating these data, are missing for Belarus, Tajikistan, and Uzbekistan, and only very limited data are available for Armenia, Azerbaijan, Georgia, the Kyrgyz Republic, and Moldova. All are classified as open based on the overall index drawing on limited data. Whenever BMP data were available for former Soviet republics, the data indicate that in 2001 all of these countries except Latvia and Lithuania were closed. (5) Sachs-Warner deviated in some cases from their self-imposed classifi cation rules. Some adjustments were meant to capture the fact that some countries had undergone changes in trade policy only mid-period, so that a classification based on period averages could be misleading. Other adjustments were made for others' reasons, described in their article. Lacking objective reasons to deviate from stated rules, the updated classification presented here abstains from any such adjustments. Several features of the new data are worth noting. (The underlying data used to construct the openness status dummy variable for the period 1990-99 are displayed in table A-1.) First, 46 countries that were classified as closed by Sachs-Warner in the 1970-89 period are classified as open in the 1990s. 7. The difference in the definitions reflects the 1999 change in UNCTAD's reporting. Before 1999, UNCTAD collected data on core nontariff barriers, including quotas, licensing, prohibitions, and administered pricing. In 1999, it began reporting all nontariff barriers, which also include technical measures and automatic licensing. 8. Wacziarg and Welch (2003) provide additional details and country-specific sources on export marketing boards and the political transitions from socialism. Wacziarg and Hom Welch 193 Sachs-Warner characterized nine of these countries as closed based on their dates of liberalization. Second, 30 countries were not classified in the Sachs-Warner study, including 23 Eastern European countries and former Soviet republics. 9 Ten of these countries remained closed in the 1990s. Third, of the 111 countries Sachs-Warner classify, 78 were closed and 33 were open in the 1970-89 period. In the 1990s, 32 countries were closed and 79 open. Of the 141 countries classified in the new data set, 42 were closed and 99 open during the 1990s. No country that was classified as open by Sachs-Warner in 1970-89 was classified as closed in the updated data set. An important and often overlooked drawback of the Sachs-Warner openness dummy variable is that it is based on averages of BMP data over each of two decades (1970-79 and 1980-89), averages of nontariff barriers and tariffs (TAR) over the last years of their sample period (1985-88), and end-of-period data for the export marketing board (XMB) and socialist (SOC) dummy vari ables. In the new data set, the XMB and SOC variables are based on their 1999 values rather than beginning-of-period or decade-long data, in order to maintain as much consistency as possible with the Sachs-Warner method oJogy.lO Similarly, the nontariff barrier data are available only for 1995-98; decade averages of the tariff data, which are available, are therefore used. As a result, some countries classified as closed could conceivably have become open late in the decade, and some countries classified as open could have been closed over most of the period. Decade dummy variables thus provide only a rough characterization of a country's outward orientation, especially in a decade in which many countries actively engaged in liberalization. A better approach is to rely more on liberalization dates, as is done below. Trade Liberalization Dates since 1994 In principle, the liberalization date is the date after which all of the Sachs-Warner openness criteria are continuously met (data limitations often imposed reliance on country case studies of trade policy). The choice of liberalization dates was based on primary-source data on annual tariffs, nontariff barriers, and BMPs. A variety of secondary sources was also used, particularly to identify when export market ing boards were abolished and multiparty governance systems replaced Communist Party rule. Because of data limitations, the European Bank for Reconstruction and Development (EBRD 1994) classification and standards of 9. The other seven countries are Cape Verde, Iceland, Lesotho, Liberia, Malta, Panama, and Swaziland. Because of lack of data, Sachs-Warner did not classify these and four other countries (Comoros, Fiji, Seychelles, and Suriname). The new data set did not allow for the determination of the openness status of these four countries in the 1990s. 10. Sachs-Warner's XMB indicators are based on data from 1991; the SOC indicators are based on data from 1987. Using 1999 data is thus consistent with their approach, however questionable that approach may be. Most countries that abolished export marketing boards in the 1990s did so during the first half of the decade. However, relying on end-of-period SOC data means that some Eastern European countries and former Soviet republics are classified as open. 194 THE WORLD BANK ECONOMIC REVIEW openness were used for several transition economies, just as they are in Sachs-Warner. Table A-2 presents the dates of trade liberalization. l l Despite the clear criteria stated above, Sachs-Warner's dates of liberalization could not conform to their five formal criteria for openness, because comparable data were lacking for many time periods. Hence, there is much scope for disagree ment with the Sachs-Warner classification, especially in light of new data pub lished since their study. Systematic review of the Sachs-Warner dates since 1990 raised questions about the liberalization status or dates for several countries. 12 Sixteen countries labeled as closed at the end of the Sachs-Warner sample period (1994) liberalized between 1995 and 2001 (table I)Y The dates of liberalization cited by Sachs-Warner differ in five countries (Cote d'Ivoire, the Dominican Republic, Mauritania, Niger, and Trinidad and Tobago). Thirty-five countries remained closed as of 2001, including five that were not classified in the Sachs-Warner study and four (Belarus, Croatia, Estonia, and India) for which the authors disagree with Sachs-Warner's assessment (table 2). Of 141 countries in the sample, 18 liberalized between 1995 and 2001 and 35 remained closed as of 2001. The Rodriguez and Rodrik Critique Rodriguez and Rodrik (2000) find that the BMP and XMB variables played a major role in the classification of countries as open or closed. They state that a dummy variable for openness based on the BMP and XMB criteria alone leads to the classification of countries as open or closed that is much closer to that generated by OPEN (the Sachs-Warner dummy variable) than one based on the SOC, TAR, and NTB dummy variables alone. They show that the BMP and XMB criteria generate a dummy variable that differs from the Sachs-Warner dummy variable in only six cases, while the TAR, NTB, and SOC criteria used jointly generated a dummy variable that differs from the Sachs-Warner dummy variable in 31 cases. Hence, they argue that the Sachs-Warner dummy variable for 1970-89 largely reflected the BMP and XMB criteria. Moreover, they argue that the XMB criterion affected only the African countries (many of which were classified as closed based on this criterion alone) and therefore amounted to an Africa dummy variable. 14 11. The working paper version of this study (Wacziarg and Welch 2003) provides detailed country summaries of liberalization episodes, along with an explanation of the dates chosen. 12. Wacziarg and Wallack (2004) systematically checked the Sachs-Warner liberalization dates before 1990 in a subset of their sample, uncovering little disagreement. 13. Table 1 also presents data for Cape Verde and Panama, which were not classified in the Sachs-Warner study. 14. Sachs-Warner based the XMB criterion entirely on the Husain and Faruqee's (1994) study of African countries that had been involved in a World Bank or International Monetary Fund structural adjustment program between 1987 and 1991. Rodriguez and Rodrik (2000) noted that Sachs-Warner classify all but one of the Sub-Saharan African countries as closed based on the XMB criterion, which is not applied to any other region. This study gathered and used XMB data for countries other than African ones. Wacziarg and Horn Welch 195 TABLE 1. Liberalization Dates of Countries That Differ from or Were Not Included in Sachs-Warner List Country Date of liberalization Cape Verde 1991 Dominican Repu blic 199r Trinidad and Tobago 1992" Cote d'Ivoire 1994" Niger 1994" Armenia 1995 Azerbaijan 1995 Egypt, Ara b Rep. of 1995 Mauritania 1995" Mozambique 1995 Tanzania 1995 Bangladesh 1996 Ethiopia 1996 Georgia 1996 Madagascar 1996 Panama 1996 Tajikistan 1996 Venezuela, R.B. de 1996 Burkina Faso 1998 Burundi 1999 Pakistan 2001 Serbia and Montenegro 2001 Sierra Leone 2001 "Year differs from that in Sachs and Warner (1995) (see text for explanation). Source: Authors' analysis based on data described in the text. TABLE 2. Countries that Remained Closed as of 2001 Algeria India" Russian Federation Angola Iran, Islamic Rep. of Rwanda Belarus' Iraq Senegal Central African Republic Kazakhstan Somalia Chad Lesotho b Swaziland b China Liberia b Syrian Arab Republic Congo, Dem. Rep. of Malawi Togo Congo, Rep. of Malta b Turkmenistan Croatia" Myanmar Ukraine Estonia" Nigeria Uzbekistan Gabon Papua New Guinea Zimbabwe Haiti 'Disagreement with Sachs and Warner (1995) (see text for explanation). bNot classified in Sachs and Warner (1995). Source: Authors' analysis based on data described in the text. To what extent are the updated Sachs-Warner data subject to the Rodriguez and Rodrik critique? BMP was the sole criterion on the basis of which 26 of 196 THE WORLD BANK ECONOMIC REVIEW 42 countries were classified as closed in the 1990s; XMB was the sole criterion on which nine countries were classified as closed. Three countries were classi fied as closed based on both the BMP and XMB criteria, leaving just four countries (Bangladesh, China, India, and Pakistan) classified as closed based on the other three criteria. Bangladesh was classified as closed based on both the TAR and BMP criteria. China was classified as closed based on the BMP and SOC criteria. India was classified as closed because of its tariff and nontariff barriers. Pakistan was classified as closed because of tariffs. The BMP and XMB criteria generated a dummy variable that differs from the 1990-99 updated Sachs-Warner dummy variable in only two cases, while the TAR, NTB, and SOC criteria used jointly generate a dummy variable that differs from the updated Sachs-Warner dummy variable in 38 cases.15 The openness status dummy variable for 1990-99 is thus subject to the same criti cisms Rodriguez and Rodrik lodged against the Sachs-Warner classification for the 1970-89 openness dummy variable. The Rodriguez and Rodrik critique is valid in terms of country status based on the OPEN90-99 dummy variable. It is less valid for the liberalization dates. As most countries were classified as closed based on the XMB and BMP criteria, not surprisingly, when they open up these variables change. The XMB and BMP vari ables determined the year of liberalization in many countries that opened up during the 1990s. The exceptions tend to be Eastern European countries and former Soviet republics, which opened based on the SOC criterion (general reforms related to liberalization). The TAR criterion was not a decisive factor in assigning a liberalization date for any country; NTB was the determining factor only in Panama. However, policy changes that reduced the BMP or removed XMBs were generally accompanied by changes in the levels of other types of trade barriers, such as tariff and nontariff barriers, that had initial values below the Sachs-Warner thresholds of 40 percent. Hence, liberalization dates do not simply capture changes in the BMP and XMB variables, but they also reflect broader liberalization. Given that the dates of liberalization in the new data set were cross-checked against a case study literature of outward-oriented reforms in developing countries, it is likely that they reflect important shifts in trade policy.16 Updating the Sachs- Warner Results The Sachs-Warner study attracted considerable attention in part because their estimated effect of the cross-sectional dummy variable for openness in explain ing annual growth between 1970 and 1989 was very large (about 2 percentage 15. Among the countries in which the TAR, NTB, and SOC dummy variables and the updated Sachs-Warner dummy variable disagree, 20 are in Africa and 10 are Eastern European countries or former Soviet republics. These countries were classified as closed based on either the XMB criterion or the BMP criterion, or both. 16. Wacziarg and Wallack (2004) show that the Sachs-Warner liberalization dates are good indicators of the timing of major trade policy changes by thoroughly checking these dates against the case study literature of trade liberalization in 25 developing countries. Wacziarg and Horn Welch 197 points). The updated data on trade policy openness make it possible to extend the Sachs-Warner regressions through the late 1990s. As this is not the main focus of this article, these results are reported only briefly. As a consistency check, the Sachs-Warner regression was first replicated for 1970-89 (column 1 in table 3 replicates column 7 in Sachs-Warner's table 11). The only difference is that the new calculations are based on a newer release of the Penn World Tables data (version 6 instead of version 5). The openness dummy variable for 1970-89 enters highly significantly, with a magnitude of 1.98 percentage points of annual growth. This result is consistent with the results in Sachs-Warner, who find a coefficient of 2.2. In contrast, the updated Sachs-Warner dummy variable enters insignificantly in the same specification for the 1990s (column 2 of table 3). The cross-sectional effect of openness on growth was estimated by construct ing openness indicators based on the dates of liberalization. The openness status for 1980, for example, takes on a value of 1 if a country had liberalized by 1980 and a value of 0 otherwise. Subsequent growth (after 1980) can then be regressed on this variable and other controls. Dummy variables were con structed for each decade (1970, 1980, and 1989) in this fashion. An advantage of this method over the period-specific dummy variables is that the period specific dummy variables are based partly on information from the end of the period (TAR, NTB, XMB, and SOC) and partly on period averages (BMP). Constructing openness indicators based on the dates of liberalization instead isolates only the countries that were open at the beginning of a period. The econometric specification is identical to that in Sachs-Warner; it restricts the time span of each regression to a single decade. The effect of the liberaliza tion status in the 1970s is weaker and smaller than in the 1980s but positive and significant at the 90 percent level. The Sachs-Warner results were likely driven by the strong effect of liberalization on growth in the 1980s (columns 3 and 4 of table 3). This effect is positive but statistically indistinguishable from zero in the 1990s when countries are grouped according to their liberalization status as of 1989. These results suggest that the Sachs-Warner cross-sectional findings are highly sensitive to the decade under consideration and that the updated openness indicator can no longer effectively distinguish fast-growing from slow-growing countries. 17 II. WITHIN-COUNTRY LIBERALIZATION DYNAMICS This section argues that better use can be made of data on the dates of liberali zation. With almost 50 years of data on growth and openness, it is possible to 17. Wacziarg and Welch (2003), who conduct many more replications of the initial Sachs-Warner cross-sectional findings, conclude that no matter how the liberalization dummy variable was defined, the results for the 1990s show an insignificant effect of the updated dummy variable on growth. This result is in sharp contrast with the results for the 1970-89 period. 198 THE WORLD BANK ECONOMIC REVIEW TABLE 3. Replication of Sachs-Warner Cross-sectional Regressions (1) (2) (3) (4) (5) Growth Growth Growth Growth Growth Variable 1970-89 1989-98 1970-80 1980-89 1989-98 Real GDP per capita (t) -1.5929 -1.150 -1.292 1.397 1.261 (4.89) (1.95) (2.83) (3.84) (2.13) Sachs-Warner openness 1.9845 0.136 dummy variable(1970-89 or 1990-98 periods) (3.87) (0.21) Openness status based on 1.387 2.574 0.521 liberalization dates (t) (1.86) (4.17) (0.84) Secondary-school enrollment 0.8059 4.689 0.169 1.822 4.872 rate (t) (0.68) (2.43) (0.10) (1.40) (2.52) Primary-school enrollment 1.4003 1.381 2.455 -0.139 1.616 rate (t) (1.65) (0.86) (2.01) (0.11) (0.99) Government Consumption to -0.0844 -0.063 -0.005 -0.065 -0.059 GDP ratio (t, t + X) (3.02) (1.32) (0.19) (2.51) (1.26) Number of revolutions per 0.4359 -0.986 1.238 -0.211 -1.030 year (t, t + X) (0.58) (1.08) (1.12) (0.21) (1.13) Number of assassinations 0.0296 0.483 0.276 0.188 0.473 per capita per year (t, t+X) (0.13) (1.56) (0.94) (0.54) (1.54) Deviation of the price level -0.1709 -0.734 -0.476 0.350 -0.721 of investment (t), as in Sachs-Warner (0.53) (1.24) (0.99) (0.87) (1.23) Gross domestic investment! 0.0757 0.051 0.076 0.103 0.040 real GDP (t, t + X) (2.64) (1.01) (2.02) (2.30) (0.76) Extreme political repression -0.6974 0.165 -0.907 -0.780 0.224 (from Sachs-Warner) (1.66) (0.28) (1.47) (1.51 ) (0.38) Population density (t - 10) 0.0006 0.0009 0.001 0.001 0.001 (0.90) (1.40) (0.60) (0.87) (1.49) Intercept 12.2482 7.752 9.334 10.635 8.288 (4.87) (1.81) (2.84) (3.86) (1.92) Adjusted R2 0.546 0.211 0.35 0.53 0.32 Number of observations 91 89 99 97 89 Note: Numbers in parentheses are t-statistics. The beginning date of each period (1970 in columns 1 and 3, 1980 in column 4, and 1989 in columns 2 and 5) is denoted by t. (t, t + X) denotes the average computed between dates t and t + X (X = 20 in column 1 and 10 in columns 2-5). The dependent variable is defined as the real annual per capita growth rate of GDP in the relevant period. Source: Authors' analysis based on data described in the text. Growth, income, and investment data are from Heston, Summers and Aten (2002). Wacziarg and Horn Welch 199 assess the within-country effects of discrete changes in trade policy openness. 1S This section compares the means of economic growth and other variables of interest, such as physical capital investment rates and trade volumes, under lib eralized and nonliberalized regimes. Liberalization and Growth Fixed-effects regressions of growth on a binary liberalization indicator, defined by the dates of liberalization, were run to assess the within-country effect of growth on liberalization. The regressions amount to difference regressions in growth or difference-in-difference regressions in log income: (1 ) where Yit is per capita income in country i at time t and L1Bit = 1 if t is greater than the year of liberalization and no reversals of the trade policy reforms have occurred, and 0 otherwise. The sample was not restricted to countries that underwent reforms. The residual term is modeled as Bit = Vi + l1t + !-tit and in all regressions, the Vi terms are treated as country fixed effects and l1t terms as fixed effects. Over the sample period 1950-98,31.7 percent of country-year observations occur in a liberalized regime (LIBit = 1) (table 4). The conditional mean of annual growth of per capita GDP given that a country is liberalized is 2.71 percent, while the mean is 1.18 percent in a nonliberalized regime, a difference of 1.53 percentage points of annual growth. These simple conditional means are based on both cross-sectional and within-country variation. Panel (1) of table 5 displays country and time fixed-effects regressions of growth on the liberalization indicator, in order to isolate within-country vari ation. The regression for 1950-98 indicates a within-country difference in growth between a liberalized and a nonliberalized regime of 1.42 percen tage points (column 1). This coefficient is estimated with a high level of statistical precision (the t-statistic exceeds 5).19 The estimated within-country 18. Sachs and Warner provide some within-country evidence on liberalization and growth for a sample of 37 reformers, presenting estimates for one fixed-effects regression of growth on dummy variables for three time periods around liberalization episodes. They show that average growth was depressed by 0.88 percentage points in the three years before liberalization, rose 1.09 percentage points a year in the three years following liberalization, and rose 1.33 percentage points a year thereafter relative to growth in the three years before liberalization. These limited results are of the same order of magnitude as the more detailed research presented here, which investigates the robustness of these estimates, extends them in time (the sample period spans 1950-98 rather than 1966-93) and space (the sample includes up to 133 countries rather than 37), and presents new evidence on investment and openness. 19. This effect was estimated allowing for first-order autocorrelation of the residuals, using the Baltagi-Wu fixed-effects method. The coefficient on liberalization was 1.32, with a t-statistic of 4.14, in line with the fixed-effects results reported here. The simpler fixed-effects estimates, with t-statistics based on robust standard errors, are reponed here because of concerns over the small T properties of the Baltagi-Wu estimator, particularly when the sample is restricted to specific decades. 200 THE WORLD BANK ECONOMIC REVIEW TABLE 4. Summary Statistics for Variables Used in Fixed-Effects Regressions Number of Standard Variable observations Mean deviation Minimum Maximum Liberalization 7,191 0.317 0.465 0.0 1.0 Investment rate 5,078 15.291 9.128 -3.590 52.880 (percent) Openness ratio 5,078 60.505 42.880 3.110 473.860 Growth (annual 4,936 1.784 6.153 -48.732 43.754 percent) Per capita GDP 5,072 5,739.380 5,826.636 276.000 39,129.000 (purchasing power parity US$) -- .. Source: Authors' analysis based on data described in the text. TABLE 5. Fixed-Effects Regressions of Growth, Investment, and Openness on Liberalization Status, 1950-98 Item (1) 1950-98 (2) 1950-70 (3) 1970-90 (4) 1990-98 Dependent variable: Growth Li beralization 1.417 0.611 1.787 2.547 (5.05) (1.29) (3.11) (2.39) Number of observations 4,936 1,728 2,312 1,116 Number of countries 133 108 112 133 Adjusted R2 0.05 0.03 0.04 0.04 Dependent variable: Investment rate Liberalization 1.937 2.545 1.237 0.762 (9.06) (7.57) (2.91) (2.16) Number of observations 5,078 1,844 2,321 1,140 Number of countries 136 110 117 136 Adjusted R2 0.04 0.10 0.11 0.02 Dependent variable: Openness Liberalization 5.531 2.302 4.097 -1.803 (7.42) (1.89) (3.74) (0.83) Number of observations 5,078 1,844 2,321 1,140 Number of countries 136 110 117 136 Adjusted R2 0.22 0.02 0.14 0.08 Note: Numbers in parentheses are robust t-statistics. Regressions are based on the specifica tions in equations (1)-(3).AII regressions include time and country fixed-effects (estimates not reported). Source: Authors' analysis based on data described in the text. effect increases over time, reaching its maximum in the 1990s (column 2-4). These results stand in sharp contrast to the cross-sectional results: countries that liberalized in the 1990s experienced a larger postliberalization increase in growth than countries that liberalized in any other decade. Indeed, the esti mated difference in growth in the 1990s is roughly 2.55 percentage points. Wacziarg and Horn Welch 201 Liberalization and Investment The empirical literature on trade and growth suggests that the effects of liberal ization on economic growth are mediated largely by the rate of physical capital investment. Several researchers, including Levine and Renelt (1992), Baldwin and Seghezza (1996), and Wacziarg (2001), suggest that the investment rate is an important channel linking trade and growth. This finding is based largely on cross-country findings. Fixed-effects regressions of investment rates on the liberalization indicator were run in order to investigate this issue in a within country context: (2) where lit is physical capital investment and Y it is GDP in country i at time t, and Wit captures country and year effects. Panel (2) of table 5 reports the estimates of such regressions. The within country evidence confirms past cross-country findings. For the period 1950 98, countries with liberalized regimes experienced average rates of physical capital investment that were 1.94 percentage points higher than those of countries with nonliberalized regimes. This represents 20 percent of this vari ahle's standard deviation in the pooled sample. The effect is largest in the initial period of the sample (1950-70). Fixed-effects regressions of growth on the investment rate were run in order to get a rough notion of how much of the effect of trade policy openness on growth can be attributed to the investment channel. The coefficient on invest ment in the baseline 1950-98 regression was 0.15 percentage points, with a t-statistic of 8.05.2° The effect of liberalization on investment in the corre sponding regression was 1.94 percentage points. Multiplying the two yields an estimate of the effect of liberalization on growth through investment of roughly 0.29 percentage points, about 21 percent of the total effect of liberalization on growth. The analysis provides suggestive evidence that investment constitutes an important channel through which trade-centered liberalization affects growth within countries. Liberalization and Openness Is trade policy liberalization followed by a break in the volume of trade, as measured by the ratio of imports plus exports to GDP? If this is the case, it suggests that liberalization did increase the level of openness of the economy. Determining this effect is important, because announced reforms may be poorly implemented or counteracted by alternative trade barriers. If trade liberalization is associated 20. The full results are presented in the working paper version of this study (Wacziarg and Welch 2003). 202 THE WORLD BANK ECONOMIC REVIEW with increases in trade volumes, one could be more confident that it actually raised the level of exposure of the reforming country to the world economy?l This issue is examined by running the following regression: (3) where X it denotes exports and Mit denotes imports. The results suggest that lib eralization raises openness by 5.53 percentage points of GDP for the full sample period (Panel (3) of table 5). This effect is indistinguishable from zero in the 1990-98 time period, however, perhaps because more time is needed to observe the effects of recent liberalizations on trade volumes. In most periods, however, trade liberalization is associated with sustained and large increases in the effec tive level of exposure of the typical reforming country to the world economy. Timing of Effects The simple average difference between growth in nonliberalized and liberalized regimes may mask important timing issues. It provides no information on how soon the effects occur or whether they cease to be felt a few years after reform. This subsection examines the time path of growth, investment, and openness for an average country before and after liberalization. Average annual growth rates, investment rates, and openness ratios are dis played in figures 2 through 4 for 20 years before and 20 years after liberaliza tion in a sample of 81 countries that underwent permanent liberalizations (that is, liberalizations that were not reversed as of 2000). As several countries had varying numbers of years of data before and after their liberalization, the average at each point in time is based on different samples of countries. 22 Several observations can be made about the figures. First, despite not con trolling for any fixed effects, the increase in growth following liberalization is remarkably similar to that shown in table 5: growth before trade-centered reforms averages 1.5 percent and rises to roughly 3 percent postreform (figure 2). Second, there does not seem to be a strong time pattern: the effects appear to be immediate and do not die out after a few years. Third, the few years immediately preceding liberalization are low-growth years: reforms are often preceded by downturns or crises. The investment rate seems to take off during the 10 years following liberaliza tion and remain high thereafter (figure 3). The plotted effect seems larger than that uncovered in the fixed-effects regressions. Openness follows a more or less 21. Even absent effects on actual openness, liberalization could still have effects on growth and investment, through pro-competitive effects or technological transfers, for example. 22. The figures did not look different when the sample was restricted to countries with continuously available data. The availabiliry of data forced a reduction in the time span to eight years before and after liberalizations and in the country coverage to 39 countries. These figures are available in the working paper version of this study (Wacziarg and Welch 2003). Wacziarg and Horn Welch 203 FIGURE 2. Sample Means for Growth before and after Liberalization o Annual growth - - - - - Average preliberalization - - -Average postliberalization -~-- Growth, 3-year moving average o o o 4 o o o o 0 0 o 0 il 0 0 ~ 0 0 0 -I o -2 -20 -10 o 10 20 Year T Source: Authors' analysis based on data described in the text. linear upward trend, without an apparent break at the date of liberalization (figure 4). More formal tests based on fixed effects did reveal an effect attribu table to liberalization, even after controlling for time fixed effects, however. Dummy variables for four (nonoverlapping) periods surrounding the reforms were defined in order to further examine the timing of the growth, investment, and openness responses to liberalization. Fixed-effects regressions were then run on growth, investment, and openness. The specification is as follows: where D lit = 1 if T 3::; t ::; T - 1 and zero otherwise; D2it 1 if T::; t ::; T + 2; D 3it = 1 if T + 3 ::; t::; T + 6, and D4it 1 if t> T + 6; and T denotes the date of liberalization. The coefficients on these dummy variables capture the average difference in growth between these years and the period preceding three years before liberalization (the baseline period). The corresponding speci fications for the investment rate and openness ratio were also run (table 6).23 23. Countries that experienced policy reversals or multiple liberalizations, for which definitions of the dummy variables are not straightforward, had to be dropped. Dropping these variables reduced the size of the sample for the growth regression from 133 to 118 countries. 204 THE WORLD BANK ECONOMIC REVIEW FIGURE 3. Sample Means for Investment before and after Liberalization o Investment rate - - Average preliberalization - - - - Avcrage postliberaliution ---Investment, 3-year moving average o 25 o o 0 o 0 o 20 o e g 0.. 15 0 o 0 0 0 ]0 '-;----------,----------,- - - - - - - ' 1 - - - - - - ' 1 -20 -]0 o ]0 20 Vear T Source: Authors' analysis based on data described in the text. The results are consistent with the observations made about figures 2-4. Countries that liberalize often do so following periods of economic turmoil: growth is depressed by 0.55 percentage points in the three years before liber alization relative to the preceding years. Tornell (1998) shows that 60 percent of episodes of economic reform, including trade reform, occur in the aftermath of a domestic political or economic crisis. Measuring growth differences relative to "early prereform" outcomes prevents falsely attributing to reforms growth differences that stem from depressed economic circum stances in the years immediately preceding the reforms. In the three years following liberalization, growth rises slightly (by 0.30 percentage points), but the effect is statistically indistinguishable from zero. Sustained growth differences become apparent three years after reform, with annual increases in growth of 1.44 points in period T + 3 to T + 6 and of 1.0 percentage point after that relative to the baseline period. The typical timing pattern revealed by these regressions shows growth to be slightly depressed before liberalization and to increase 1.0 -1.5 percentage points three years after reforms. A similar pattern applies to investment and openness. These esti mates reflect sample averages and may mask interesting country-specific differences, as discussed below. Wacziarg and Horn Welch 205 FIGURE 4. Sample Means for Openness before and after Liberalization o Opcrmcss - - - - - Average preliheraliution - Average postiibernlization ~~~-Openness, 3~year moving average 100 o 80 o -- o ---- o o 60 o o -------------0 o -e o 40 -20 -10 o 10 20 YearT Source: Authors' analysis based on data described in the text. TAB L E 6. Fixed-Effect Regressions: Timing of the Effects of Liberalization on Growth, Investment, and Openness Item (1) Growth (2) Investment (3) Openness DJ -0.555 -1.040 1.979 (1.14) (2.88) (1.32) D2 0.300 -0.160 0.795 (0.61) (0.41 ) (0.63) D3 1.438 1.197 3.606 (3.27) (2.98) (2.21) D4 1.015 2.129 13.371 (2.30) (5.47) (9.17) Number of observations 4,230 4,357 4,357 Number of countries 118 121 121 Adjusted R2 0.04 0.08 0.26 Note: Number in parentheses are robust-statistics. Regressions are based on the specification in equation (4). All regressions include time and country fixed-effects (estimates not reported). Definition of dummy variables, where T represents the date of liberalization, is as follows: Dl 1 if T - 3 ::; t ::; T - 1 and zero otherwise. D z 1 if T::; t ::; T + 2 and zero otherwise. D3 = 1 if T + 3::; t::; T + 6 and zero otherwise. D4= 1 if t > T + 6 and zero otherwise. Source: Authors' analysis based on data described in the text. 206 THE WORLD BANK ECONOMIC REVIEW Concurrent Policies It is difficult to attribute differences in growth purely to trade liberalization. Countries carrying out trade reforms often simultaneously adopt policies favor ing domestic deregulation, privatization, and other microeconomic reforms and macroeconomic adjustments, making it difficult to interpret the coefficient on liberalization in a within-country growth regression as the total effect of trade liberalization per se. 24 A more realistic interpretation of these estimates is that they capture the impact of trade-centered reforms more broadly. In what follows, we describe our efforts to address this important concern. SCOPE OF REFORMS. The working paper version of this study (Wacziarg and Welch 2003) distinguishes countries that carried out overall reforms from those that carried out external sector reforms in relative isolation from other dom estic reforms. Wacziarg and Wallack (2004) examine 22 episodes of trade liber alization, most of them in developing countries in the 1980s. Fourteen of these episodes were accompanied by market-oriented domestic reforms; eight occurred in relative isolation from major shifts in domestic policy. The distinc tion between pure trade reforms and overall reforms was based largely on whether the countries implemented a substantial program of privatization and deregulation at the same time as trade reforms. Isolating the sample of countries that were part of the Wacziarg and Wallack (2004) study and examining whether the within-country effects of liberalization on growth differed between trade reformers and overall reformers reveal several noteworthy findings. First, even though the sample was restricted to 22 countries, the estimates were remarkably similar to those obtained for the full sample of 133 countries. Second, the estimates of the impact of trade liberaliza tion in countries that carried out trade reforms in isolation were similar to the corresponding estimates for countries that also reformed their domestic sectors, despite the crude nature of the distinction between overall reformers and pure trade reformers. While the interpretation of these suggestive results requires caution, a plausible conclusion is that the effect of trade-centered reforms is in large part attributable to an external reform component. This issue is further addressed below in the context of individual country experiences. OrnER EXTERNAL REFORMS. Trade reforms are sometimes associated with other types of external reforms, such as capital market liberalization. To the extent such reforms are adopted simultaneously, estimates may capture the impact of these financial reforms rather than trade reforms. This argument is frequently invoked to criticize the type of estimates presented above. 24. An analogous point is often made in a cross-country context. RodrIguez and Rodrik (2000) and other observers suggest that "bad" government policies tend to go together, making it difficult to disentangle the effects of protectionist trade policy from those of poor macroeconomic management, poor governance, or poor institutions in general. Wacziarg and Horn Welch 207 This issue is investigated by looking at data on the timing of financial reforms. Bekaert, Harvey, and Lunblad (2001) examine the impact of capital market lib eralization on economic growth in a panel context, using both cross-sectional and within-country time variation. Using data from Bekaert and Harvey (2000) on the dates of official regulatory reforms pertaining to financial markets, they find robust positive effects of financial liberalization. 25 Their dates are com pared with the dates of trade liberalization in the data set used here. Bekaert and Harvey (2000) characterize the date of official financialliberali zation for 40 of the 106 countries in the sample that had liberalized by 2001.26 Of these, only two (Brazil and Turkey) have exactly the same year of official financial regulatory reform and trade liberalization. Only nine countries implemented financial sector reforms within three years before and after the date of trade liberalization, and just 17 did so within five years before and after. Many countries that enacted trade reforms never enacted financial liberal ization, so the numbers cited above overstate the extent of coincidence between financial and trade liberalization dates. There is thus little evidence that trade reform and financial market liberalization occur concurrently and that the esti mettes may confound the effects of the two types of reform. III. COUNTRY CASE STUDIES The econometric results presented above summarize the effect of trade liberali zation on growth and other variables for a sample of very diverse countries. Fixed-effects regressions allow all time-invariant country characteristics to be controlled for. The estimated coefficients on liberalization are not country specific, however; they represent average responses. The reaction of individual countries to reforms is likely to vary, especially as the depth and scope of reforms differed across countries. Much can be learned from the considerable heterogeneity in the response of growth to trade reform. This section examines specific cases of reform in countries representative of the broader sample for which enough data on growth, investment, and openness are available before and after reforms. The goal is to get a sense of the subtleties of reform in specific cases and to illustrate the economic mechanisms that give rise to the average estimated effects. The time paths of growth, investment, and openness are first examined for a subsample of 24 developing countries for which data are available for at least eight years before and after liberalization. A more detailed discussion then focuses on 13 of these countries. 25. Henry (1999, 2000) uses data on economic and political reforms for a smaller set of 18 developing countries. 26. Details of the comparison between the Bekaert and Harvey (2000) dates and the dates presented here are available on request. 208 THE WORLD BANK ECONOMIC REVIEW The average difference in growth, investment rates, and openness ratios between the pre- and postliberalization periods is shown for 24 countries (table 7). The countries were chosen from the sample of 39 countries for which at least eight years of data are available on either side of the date of liberaliza tion, restricting the sample to emerging markets, the main focus of this study. The data reveal positive growth differences in 13 of the 24 countries and negative differences in six of them; the remaining five countries exhibit an effect close to zero. Postliberalization growth effects appear large in Mauritius, Indonesia, Uruguay, Republic of Korea, Chile, Taiwan (China), and Uganda. Among countries that experienced positive differences, the magnitude of the growth increase ranged from 0.83 percentage points of per capita income growth in Poland to 3.62 points in Mauritius. The range of growth decline was of a similar magnitude. Before and after comparisons of investment rates and openness also reveal large variations across countries. The postliberalization surge in investment rates was particularly strong in the Republic of Korea, Taiwan (China), Indonesia, Jordan, and Guinea-Bissau. About half of the 24 countries exhibited zero or negative differences in investment rates. TABLE 7. Mean Growth, Investment, and Openness Changes in 24 Countries Growth Investment Openness Year of Sample Country difference difference difference liberalization period Mauritius 3.62 0.34 35.90 1968 1951-98 Indonesia 3.32 9.80 25.96 1970 1961-98 Uruguay 3.08 -1.01 11.22 1990 1951-98 Korea, Rep. of 3.02 18.44 43.40 1968 1954-98 Chile 2.80 1.12 26.33 1976 1952-98 Taiwan 2.29 9.91 55.77 1963 1952-98 Uganda 2.24 1.63 -6.60 1988 1951-98 Ghana 1.99 -3.91 9.13 1985 1956-98 Guinea 1.85 -2.74 7.28 1986 1960-98 Guyana 1.80 7.49 84.49 1988 1951-98 Benin 1.74 1.64 8.72 1990 1960-98 Mali 1.19 0.86 15.68 1988 1961-98 Poland 0.83 -4.30 3.35 1990 1971-98 Paraguay 0.42 2.01 49.71 1989 1952-98 Cyprus 0.34 -4.05 29.13 1960 1951-96 Colombia 0.18 0,48 5.91 1986 1951-98 Tunisia -0.30 -5.58 31.94 1989 1962-98 Philippines -0.40 1.03 39.54 1988 1951-98 Israel -0.96 6.10 21.42 1985 1951-98 Botswana 1.99 3.98 22.27 1979 1961-98 Mexico -2.16 -4.59 17.56 1986 1951-98 Hungary -2.41 -1.19 -4.17 1990 1971-98 Guinea-Bissau -2.95 5.59 9.89 1987 1961-98 Jordan -4.28 5.75 40.61 1965 1955-98 Source: Authors' analysis based on data described in the text. Wacziarg and Horn Welch 209 Closer examination of postliberalization changes in growth, investment, and openness for a restricted sample of developing countries thus reveals consider able heterogeneity in their experiences with reform. The following case studies develop hypotheses that could account for these differences. From the sample of 24 developing countries for which there are at least eight years of data on either side of liberalization, a subsample of 13 countries was selected to study in greater detail. A set of countries was chosen that was small enough to allow their preexisting conditions, overall policy environment, and macroeconomic circumstances to be examined while maintaining a geo graphically diverse sample reflecting the range of country-specific growth effects identified above. The goal was to uncover patterns that could explain cross-country differences in individual countries' responses to liberalization and suggest directions for future research. The subsample was selected to include a geographically diverse set of countries that experienced growth effects of liberalization in roughly the same proportions as the 24 countries discussed above. It includes 13 countries, seven of which experienced higher mean growth rates following liberalization (Indonesia, Republic of Korea, Chile, Taiwan (China), Uganda, Ghana, and Poland). The growth difference was negative in four countries (Israel, Botswana, Mexico, and Hungary). In two countries, Colombia and the Philippines, liberalization was associated with roughly zero difference in their mean growth rates. Table A-3 describes all countries' concurrent reforms, macroeconomic environment, and political context. Examination of these case studies suggests that the packaging and timing of reforms are important factors in explaining differences in postliberaliza tion growth patterns. Countries that followed through by deepening trade reforms over time did better than countries that did not. Neither active gov ernmental disengagement from industrial policy nor broad-based reforms were necessary conditions for success. Countries that counteracted short lived programs of external liberalization with domestic interventions and countries that adopted tight macroeconomic policies, faced unfavorable terms of trade shocks, or suffered from political instability did not perform as well as other countries. Sustained Reforms In the majority of countries that experienced higher growth following liberali zation, trade reforms were not strictly limited to the period of liberalization; these countries continued to deepen trade reform after liberalization. Chile, for example, which liberalized in 1976, recovered from the Latin American debt crisis and continued to grow during the late 1980s. During this period, it decreased tariffs and implemented several bilateral free trade agreements. Both Korea and Taiwan (China), which liberalized in the 1960s, continued to lower tariffs and remove nontariff barriers, particularly during the mid-1980s and 210 THE WORLD BANK ECONO~lIC REVIEW 1990s. Indonesia sustained the initial reforms of 1970 with reductions in export duties in 1976 and additional trade-centered liberalization throughout the 1980s. In Uganda, the 1988 liberalization was followed by a second wave of external reforms in 1993-94. Scope of Reforms Whether trade reforms were part of a package of other domestic reforms or occurred in relative isolation does not seem to help predict the effect on growth. Among countries that implemented broad-based reforms, and in which postliberalization growth increased, Chile and Poland stand out as prototypical success stories of reform. Both implemented broad-based domestic reforms, of which trade liberalization was only a part. In Colombia, Hungary, and Mexico, which Wacziarg and Wallack (2004) classify as broad-based refor mers, average growth following liberalization actually fell. Political instability is probably at the heart of Colombia's lack of increased growth. In Hungary, the decline may have occurred because the domestic portion of the reform program (banking sector reforms, privatizations) was in large measure delayed until 1995. To the extent that external and domestic reforms are complemen tary, the full effects for Hungary may not be apparent in the growth data, which extend only to 1998. The case of Mexico is more complex. The privatization program began before trade liberalization, in 1984, with the sale of small- and medium-sized businesses, and continued after 1986, with the sale of larger enterprises, such as the national telephone company, parts of the banking industry, and the national airline. While Mexico maintained large government oligopolies that prevented broad industrial restructuring and resource reallocation, one can hardly argue that its entry into the General Agreement on Tariffs and Trade (GATT) in 1986 and the concurrent reduction in external barriers occurred in isolation from other domestic reforms. The flip side of this coin is a country like Ghana, which, according to Wacziarg and Wallack (2004), implemented trade reforms in relative isolation (privatiza tion, for instance, did not begin until the early to mid-1990s). It experienced a 2 percentage point increase in mean growth after the 1985 liberalization. Other interesting cases are the success stories of Southeast Asia, where many economies, including Korea and Taiwan (China), implemented policies aimed at increasing foreign direct investment (FDI) at the same time or after external liberalization. Indonesia, Korea, and Taiwan (China) pursued growth strategies with widespread government involvement in the economy. In Indonesia, government involvement increased during the 19705, after external liberali zation began. Both Korea and Taiwan (China) adopted activist industrial pol icies, with the government involved in "picking winners." That the growth performance of these countries was unprecedented the 1998 Asian crisis shows that government disengagement from the economy is not a necessary condition for successful trade reforms. What all these countries shared was an Wacziarg and Horn Welch 211 outward-oriented development model in which increasing exports was a central pillar of the growth strategy. One cannot point to the breadth of reform as an unambiguous criterion explaining differences in the growth response to liberalization. The picture that emerges is far from simple. The set of economies that experienced higher growth following liberalization includes both those that maintained heavy gov ernment involvement in the economy (Indonesia, Korea, and Taiwan [China]) and those that actively reduced the role of government (Chile and Poland). The set of countries that experienced negative or zero growth differentials after lib eralization includes Colombia, Hungary, and Mexico, countries that actively disengaged the government from domestic economic activity at the time of trade reforms. Counteractive Policies Some of the 13 countries in the sample implemented policies that actively counteracted the effects of trade reform and as a result did not experience increases in growth rates. 27 In Israel, social pacts based on broad coalitions of labor, government, and industry set the patterns for prices, wages, and the exchange rate in ways that mitigated the effects of trade openness on domestic producers. In the Philippines, trade liberalization was accompanied by a large increase in the share of state-owned enterprises in the economy, including a doubling of the share in GDP of financial transfers from the government to state-owned enterprises between 1987 and 1989. Such interventions, designed partly to protect domestic producers in the face of increased import com petition, may have precluded the realization of gains from trade. Macroeconomic Factors Countries that did not experience growth increases after liberalization often suffered from mitigating circumstances, associated in particular with restrictive macroeconomic policies or terms of trade shocks. In Hungary and Mexico, two countries in which growth fell following liberalization, trade reform was followed by tight monetary policies involving high interest rates, which depressed growth. In Mexico, currency overvaluation undid the effects of trade liberalization in the late 1980s and early 1990s. In Botswana, terms of trade considerations account for the absence of a postliberalization growth surge. Volatility on world diamond markets increased shortly after Botswana implemented trade reforms in 1979. The weak diamond marked caused a recession in 1981-82 that resulted in a postliberalization growth rate that was about 2 percentage points lower than the preliberalization rate. Thus, terms of trade considerations are essential in accounting for the absence of a postliberalization growth surge in Botswana. 27. Wacziarg and Wallack (2004) discuss some of these cases in greater detail. 212 THE WORLD IIANK ECONOMIC REVIEW Political Instability Several countries suffered from severe forms of political instability, preventing realization of the gains from trade liberalization. A prime example is Colombia, where instability persisted throughout the 1990s. Other examples include Israel and the Philippines. In contrast, economies that seem to have experienced higher growth following reform also witnessed periods of relative political stability. Taiwan (China) is a case in point, as are Chile, Indonesia, and Korea, where liberalization coincided roughly with the rise to power of authoritarian regimes, resulting in a degree of lasting political stability follow ing periods of political unrest. IV. CONCLUSION This article presents an updated data set of trade policy indicators and liberali zation dates. It revisits the evidence on the cross-country effects of Sachs-Warner's simple dichotomous indicator of outward orientation on econ omic growth, confirming the pitfalls of this indicator first identified by Rodriguez and Rodrik (2000). It shows that the Sachs-Warner dichotomous indicator effectively separates fast-growing from slow-growing countries in the 1980s and to a lesser extent in the 1970s, but fails to do so in the 1990s. Simple dichotomous indicators of outward orientation are too crude to capture the complexities of trade policy. Instead, liberalization dates that capture episodes of discrete shifts in trade policy can be useful for estimating within-country growth responses The Sachs-Warner dates of liberalization were painstakingly checked and updated, based on quantitative data and a thorough review of country-specific case studies of reform. The new and robust evidence indicates that these dates of liberalization mark breaks in growth, investment, and openness within countries. Over the 1950-98 period, countries that liberalized their trade regimes experienced average annual growth rates that were about 1.5 percen tage points higher than before liberalization. Postliberalization investment rates rose 1.5-2.0 percentage points, confirming past findings that liberalization fosters growth in part through its effect on physical capital accumulation. Liberalization raised the average trade to GDP ratio by roughly 5 percentage points, after controlling for year effects, suggesting that trade policy liberaliza tion did indeed raise the actual level of openness of liberalizers. Trade-centered reforms thus have significant effects on economic growth within countries. These within-country estimates represent the average effect of liberalization on growth, investment, and openness; they mask differences in the individual responses of countries to trade liberalization. Restricting the sample to 13 countries sheds light on the sources of these differences. Countries that experi enced positive effects tended to deepen trade reforms. But active industrial policies, such as those implemented in Southeast Asia, did not preclude growth Wacziarg and Horn Welch 213 gains from trade liberalization, and broad-based reforms appear to be neither a necessary nor a sufficient condition for reaping these gains. Countries that experienced negative or no effects on growth tended to have suffered from political instability, adopted contractionary macroeconomic policies in the aftermath of reforms, or undertaken efforts to counteract trade reform by shielding domestic sectors from necessary adjustments. Future research should seek to clarify the factors accounting for heterogeneity in the growth effects of trade reform. ApPENDIX TABLE A-I. Trade Policy Variables for Economies in Sample, 1990s Core nontariff Export Average barrier coverage Average BMP, Marketing Board Socialist OPEN90-99 tariff, 1990- rate (percent), 1990-99 (1 = country has (l = country Economy (1 = open)' 99 (percent)b 1995 _98' (percent)d exporting board)' is socialist)' AIl:'ania 1 15.90 7.53 0 0 Algeria 0 23.97 177.91 0 0 Angola 0 23.62 0 0 Argentina 12.54 2.1 9.30 0 0 Armenia 0 0 0 Australia 7.91 0 0 0 Austria 6.91 0 0 0 Azerbaijan 1 0 0 0 Bangladesh 0 43.70 83.27 0 0 Barbado, 1 15.58 2.31 0 0 Belarus 0 12.63 1 0 Belgium 6.91 0 0 0 Benin 28.61 1.0 1.93 0 0 Bolivia 10.34 1.49 0 0 Bot,wan,} 20.55 7.82 0 0 Bra7.il 17.32 21.6 13.76 0 0 Bulgaria 17.37 7.44 0 0 Burkina Faso 1 29.13 1.98 0 0 Burundi 0 7.40 29.55 0 0 Cameroon 18.43 1.98 0 0 Canada 6.81 0 0 0 Cape Verde 1 22.05 0 0 0 Central African 0 12.80 1.55 0 Republic Chad 0 15.80 1.98 1 0 Chile 1 11.33 5.2 9.84 0 0 China 0 31.06 35.89 0 1 Colombia 1 14.30 10.3 8.87 0 0 Congo, Oem. 0 25.47 34.67 0 Rep. of Congo, Rep. of 0 17.97 1.98 1 0 Costa Rica 1 10.60 6.20 5.37 0 0 Cote d'Ivoire 1 22.00 30.90 1.98 0 0 Croatia 0 37.76 0 0 Cyprus 10.64 21.60 2.16 0 0 (Continued) 214 THE WORLD BANK ECONOMIC REVIEW TABLE A-I. Continued Core nontariff Export Average barrier coverage Average BMP, Marketing Board Socialist OPEN90-99 tariff, 1990 rate (percent), 1990-99 (I country has (I country Economy (1 = open)' 99 (percent)b 1995-98< (percent)d exporting board)e is socialist)' Czech Republic 6.08 0.22 0 0 Denmark 6.91 0 0 0 Dominican 16.70 6.20 16.31 0 0 Republic Ecuador 11.29 9.34 0 0 Egypt, Arab 30.23 12.45 0 0 Rep. of EI Salvador 1 9.38 5.20 13.59 0 0 Estonia 0 1.12 25.09 0 0 Ethiopia 0 22.55 111.43 0 0 Finland 1 6.91 0 0 0 France 1 6.91 0 0 0 Gabon 0 19.87 1.98 1 0 Gambia, The 13.55 4.69 0 0 Georgia 0 0 0 Germany 6.91 0 0 0 Ghana 14.93 2.96 0 0 Greece 6.91 1.24 0 0 Guatemala to.27 6.03 0 0 Guinea 3.99 0 0 Guinea-Bissau 1 0 0 0 Guyana 0 13.70 28.23 0 0 Haiti 0 10.00 81.12 0 0 Honduras 8.90 9.21 0 0 Hong Kong 2.10 -0.02 0 0 (China) Hungary 1 12.11 5.40 0 0 Iceland 1 3.98 1.24 0 0 India 0 48.65 93.80 7.45 0 0 Indonesia 1 16.27 31.30 7. to 0 0 Iran, Islamic 0 1,199.31 0 0 Rep. Iraq 0 138,935.90 0 0 Ireland 3.98 2.50 0 0 Israel 7.80 2.09 0 0 Italy 6.91 0 0 0 Jamaica 14.68 15.46 0 0 Japan 1 5.98 -0.35 0 0 Jordan 1 15.83 3.37 0 0 Kazakhstan 0 55.34 0 0 Kenya 27.47 15.94 0 0 Korea, Rep. of 11.28 25.00 0.Q3 0 0 Kyrgyz 0 0 Republic Latvia 1 5.73 7.29 0 0 Lesotho 1 17.40 3.49 0 0 Liberia 0 2,306.86 0 0 Lithuania 4.33 7.45 0 0 Luxembourg 6.91 0.38 0 0 Macedonia, 18.45 0 0 l:"YR Madagascar 7.13 5.93 0 0 Malawi 0 19.80 28.83 0 0 Malaysia 1 11.70 19.60 1.35 0 0 (Continued) Wacziarg and Horn Welch 215 TABLE A-I. Continued Core nontariff Export Average barrier coverage Average BMP, Marketing Board Socialist OPEN90-99 tariff, 1990 rate (percent), 1990-99 (1 country has (1 country Economy (1 open)' 99 (percent)b 1995-98' (percem)d exporting board)' is socialist)' Mali 15.66 1.98 0 0 Malta 7.23 1.20 0 0 Mauritania 28.23 1.55 0 0 Mauritius 27.00 16.70 5.25 0 0 Mexico 12.53 13.40 2.24 0 0 Moldova 0 0 0 Morocco 23.75 13.40 3.54 0 0 Mozamblque 16.25 6.87 0 0 Myanmar 0 5.70 2,280.77 0 0 Nepal 0 15.28 24.23 0 0 Netherlands 6.91 0 0 0 New Zealand 1 6.35 2.50 0 0 Nicaragua 1 9.90 9.98 0 0 Niger 1 18.30 1.87 0 0 Nig~ria 0 29.74 11.50 151.32 0 0 Norway 4.87 0 0 0 Pakistan 0 54.73 9.74 0 0 Pan.1ma 10.67 0 0 0 Papua New 0 ]6.67 16.57 1 0 (,uinea Paraguay 10.91 0.00 11.83 0 0 Pen, 16.80 8.75 0 0 Philppines 19.09 4.36 0 0 Poland 12.46 2.42 0 0 Portugal 1 6.91 2.04 0 0 Romania 0 13.50 104.30 0 0 Russian 0 11.24 50,979.69 0 Federation Rwanda 0 38.40 50.78 0 0 Senegal 0 13.05 1.98 0 Sierra Leone 0 30.25 61.47 0 0 Singapore 0.32 2.10 0.80 0 0 Slovak 7.35 5.34 0 0 Republic Slovenia 1 10.60 10.06 0 0 Somalia 0 246.55 0 0 South Africa 9.05 8.30 3.46 0 0 Spain 6.91 1.71 0 0 Sri Lanka 24.34 22.70 7.84 0 0 Swaziland 15.10 7.62 0 0 Sweden 6.91 0.00 0 0 Switzerland 1.38 0.00 0 0 Syrian Arab 0 16.00 279.97 0 0 Republic Taiwan 9.85 0.95 0 0 (China) Tajikistan 0 0 Tanzania 0 25.12 22.17 0 0 Thailand 29.54 17.50 1.80 0 0 Togo 0 15.25 1.98 I 0 Trinidad and 14.86 13.22 0 0 Tobago Tunisia 28.25 3.67 0 0 Turkey 15.28 19.80 1.15 0 0 (Continued) 216 THE WORLD BANK ECONOMIC REVIEW TABLE A-I. Continued Core nontariff Export Average barrier coverage Average BMP, Marketing Board Socialist OPEN90-99 tariff, 1990 rate (percent), 1990-99 (1 country has (1 = country Economy (1 = open)" 99 (percent) b 1995-98< (percent)d exporting board)e is socialistIC Turkmenistan o 42.86 1 0 Uganda 1 14.37 3.10 19.33 0 0 Ukraine o 9.73 9.02 I 0 United 6.91 0.00 0 0 Kingdom United States 5.96 0.00 0 0 Uruguay 1 14.00 0.00 9.88 0 0 Uzbekistan o Dual 0 0 exchange rate Venezuela 14.31 17.70 4.13 0 0 Yemen, Re. of 1 20.00 8.34 0 0 Serbia and o 106.44 0 0 Montenegro Zambia o 18.43 1.00 62.55 0 0 Zimbabwe o 20A3 132.81 0 0 - not available. aBased on application of Sachs and Warner (1995) criteria; see Wacziarg and Welch (2003) for details. bUnweighted average tariff, 1990-99, based on data from UNCTAD (2001), World Bank (2000), and WTO (various years). and it includes adjustments for time-varying population growth rates in the discount rates in PVL1Cit , PV(L1'Yit Wit) and the wealth-dilution term. Both equations were estimated. Given that population growth rates varied over time in the countries in the sample, the expectation was that the results for equation (8) would be stronger than those for equation (7) in ways defined below. Strictly interpreted, equation (6) implies the joint hypotheses f30 0 and f3I = 1: there is a one-to-one relation between genuine savings and consump tion changes. A weaker hypothesis is simply f3I > 0: genuine savings and con sumption changes are positively correlated. The theory refers to a situation in which genuine savings include all changes in wealth. Any empirical estimates of genuine savings will inevitably be incomplete, which could bias the estimate of f3I away from 1. This bias was expected to be smaller, and the estimates of f3I therefore closer to 1, for more comprehensive savings measures. Equations (7) and (8) were estimated sequentially to test this hypothesis. Initially, g was set equal to gross national savings. It was then adjusted sequentially, for the depreciation of produced capital, the depletion of natural capital, and the dilution of produced and natural capital. As discussed in the next section, these Ferreira, Hamilton, and Vincent 239 adjustments are crude. The resulting measurement error could weaken the convergence of the estimates of 131 toward 1. One might expect country fixed effects to be added to equations (7) and (8) (changing the intercepts from 130 to 13oi) to control for omitted wealth com ponents that are more or less constant over time. In fact, in the presence of time-invariant total omitted wealth X, equation (3) becomes (9) where g IS true (unobserved) genuine savings. 6 As long as population N is changing over time, an ordinary fixed effect will not solve the problem of the omission of X'YIN from the genuine savings variable in equations (7) and (8). The problem occurs even if the population growth rate is constant (and nonzero). To solve this problem one must include the ratio of the population growth rate to the total population as an additional explanatory variable. Equations (7) and (8) thus become (10) and (11) PVLlCit + PV(Ll'YitWit) = 130 + 131git + 132iN'Yit + Bit· It From equation (9) one would expect 132i to be negative if significant wealth were omitted and zero otherwise. The absolute value of this coefficient pro vides an estimate of Xi, the total omitted wealth for country i. These two equations and, for comparison, fixed-effects versions of equations (7) and (8) are estimated. 7 III. DATA AND ESTIMATION ISSUES This section describes the data sources, the construction of key variables, and estimation issues. Data Data for constructing the variables in equations (7), (8), (10), and (11) were obtained from the World Development Indicators (World Bank various years). 6. The authors are indebted to a referee for pointing this out. 7. Fixed effects were used instead of random effects, because the sample is not random: it includes the population of developing countries. An F-test was used to test the null hypothesis that the fixed effects were equal to zero. The hypothesis was rejected in all models. 240 THE WORLD BANK ECONOMIC REVIEW All monetary variables are expressed in constant 2000 U.S. dollars. 8 The final sample includes 64 non-OEeD countries. Complete series for 1970-2003 were available for most countries, with missing values occurring mainly in 2003. Hence the panel is unbalanced but reasonably complete. The time horizon, T, was set equal to 20 years in constructing the present value of future changes in per capita consumption. Although Ferreira and Vincent (2005) use 10 years in their benchmark model, they show that the econometric relation between genuine savings and their consumption measure improves when they extend the time horizon to 20 years. This improvement is not surprising: "green" accounting theory refers to an infinite time horizon. The 20-year time horizon reduced the sample in the econometric analysis to 1970-82: a test was run to determine whether genuine savings in year twas positively correlated with the actual present value of consumption changes during the period t + 1 to t + 21. Each equation was estimated sequentially for four comprehensive savings measures. In increasing order of completeness, the four measures are as follows: 1. Gross savings. This measure implicitly includes both gross investment in produced capital within the country's borders and the current change in holdings of foreign assets. 2. Net savings. This measure equals gross savings minus depreciation of pro duced capital. 3. Green savings. This measure was constructed by subtracting estimates of the current depletion of subsoil assets and forest resources from net savings. 4. Population-adjusted savings. This measure equals green savings minus the wealth-dilution term. Summary statistics for the 64 countries reveal that the means of PVAC and PVAC + PV(Ayw) are the same order of magnitude as the savings measures, suggesting that an empirical relation between current savings and future con sumption changes is plausible (table 1).9 The means of the savings measures decrease sharply with the progressive adjustments for depreciation of produced capital (net savings), depletion of natural capital (green savings), and wealth dilution (population-adjusted savings). In line with Dasgupta's (2003) findings, the mean of population-adjusted savings is negative, suggesting that on average 8. A reviewer pointed out that exchange rates in 2000 may not be representative, because countries were still recovering from the effects of the Asian 6nancial crisis. Choosing a different base year would alter the relative size of the different variables in the model across countries (because of different base year exchange rares). With country fixed effects. however, it is within-country variation over time rather than variation across countries that determines the econometric results. The results are therefore invariant to the choice of base year. 9. Supplemental appendix S.2 provides detail on data soun;es and the procedures followed in constructing the variables. Supplemental appendix S.3 provides detail on individual countries. Ferreira, Hamilton, and Vincent 241 TABLE 1. Descriptive Statistics for Key Variables Number of Standard Variable Observations Mean Deviation Minimum Maximum PV~C 799 111.1 356.9 -2,201.3 1,414.9 PV~C + PV(~')'W) 790 50.2 379.3 -2,067.4 1,284.1 Gross savings 794 240.8 363.5 -86.8 2,516.3 Net savings 794 144.7 254.1 -221.3 2,124.0 Green savings 794 53.6 198.2 -1,591.3 1,489.1 Population-adjusted 793 -148.5 289.4 -2,321.1 1,028.2 savings Population growth 858 2.37 0.81 -0.13 4.37 rate (percent) Total population 858 41.5 136 0.120 1,010.0 (millions) Note: All variables except population growth rate and total population are expressed in per capita terms in 2000 U.S. dollars. PV~C and population-adjusted savings are computed using country-specific constant interest and population growth rates, as in equation (7); PV~C + PV(~')'W) is computed using country-specific constant interest rates but time-varying population growth rates, as in equation (8). The sample of 64 countries covers the period 1970-82. Source: Authors' analysis based on data from World Bank (various years). the countries in the sample dissaved after accounting for the spreading of exist ing wealth across growing populations. The adjustment for wealth dilution also increases the variability of the savings measure, as indicated by the larger standard deviation for population-adjusted savings than for green savings. Estimation Issues There is a risk of endogeneity when population-adjusted savings is the savings measure. The estimates of natural capital in the wealth-dilution term are con structed using data on future resource rents. It is possible that a positive con sumption shock in period t + s, which is reflected in PVACit on the left side of the regression equations, might induce a country to extract more resources to pay for the additional consumption. Because resource rents would also increase in period t + s, the dependent and explanatory variables would be simul taneously determined. This risk is greater when the adjustments for time varying population growth rates are included, because future resource rents that appear in the current wealth-dilution term also appear in the PV(AYitU'it) term on the left side of equations (8) and (11). Although this risk is reduced by the facts that current wealth was not used in constructing PV(AYitU'it) (see equation (5)) and that future per capita wealth in PV(AYitWit) is weighted by the discount rate and the change in the population growth rate, it does not necessarily become negligible. Equations that involved population-adjusted savings using instrumental variables were estimated using the generalized two-stage least squares fixed-effects estimator of Balestra and Varadharajan-Krishnakumar (1987) in 242 THE WORLD BANK ECONOMIC REVIEW order to reduce the bias that could result. The set of instruments included lagged values of green savings, produced capital, the percentage of the popu lation of working age, and the population growth rate, and a time trend. These variables were correlated with contemporaneous savings in the first-stage regressions (and were thus relevant instruments); lagging them ensured that they were exogenous. Standard errors were corrected in all models for serial correlation, and time dummy variables were included to control for unobserved global factors that affected consumption-savings decisions across all countries in a given year. The risk of spurious correlation must be considered given that time-series data were used. Formal testing for stationarity and cointegration is not appro priate for the data, because time-series tests have little power in samples that cover as short a period as those used here (Banerjee 1999). There is ample evi dence that consumption is nonstationary but cointegrated with income, so that savings is stationary (see, for example, Davidson and others 1978; Hamilton 1994). Although consumption is nonstationary, the dependent variable is con structed as the present value of future changes in per capita consumption. Examination of the autocorrelograms for these series confirms their stationar ity. Ferreira and Vincent (2005) examine autocorrelograms for longer time series of similar variables, because the time horizon in their study was only 10 years. They, too, find that the series are stationary. IV. RESULTS Estimates of /31 were derived for the four model specifications and the four savings measures. Comparison of estimates in the last row (population adjusted) of table 2 with those in the second from last row (green) indicates the impact of accounting for wealth dilution; comparison of estimates in column 3 with those in column 1 (and column 4 with column 2) indicates the impact of accounting for changes in the population growth rate through adjustment to the dependent variable; and comparison of the estimates in column 2 with those in column 1 (and column 4 with column 3) indicates the impact of con trolling for time-invariant omitted wealth by including the ratio of the popu lation growth rate to total population as an additional explanatory variable. For reference, column 5 shows results from a model with the dependent vari able used by Ferreira and Vincent (2005) (that is, the difference between average future consumption and current consumption). That model includes no adjustments for population other than the variables expressed in per capita terms. A 20-year time horizon and country-specific (but time-invariant) interest rates were used instead of the 10-year horizon and the fixed 3.5 percent rate used by Ferreira and Vincent to make the,results more directly comparable to those in the other columns. Consider first the results in columns 1-4. The most striking result is the adjustment for natural resource depletion. The hypothesis /31 > 0 is supported TAB LE 2. Estimates of f31 for Four Model Specifications and Four Savings Measures Variable 1 Equation (7) 2 Equation (10) 3 Equation (8) 4 Equation (11) 5 Ferreira and Vincent (2005) Dependent variable pv~c PV~c PV~c + PV(~')'W) PV~c + PV(~')'W) "C-C Time dummy variables? Yes Yes Yes Yes Yes Country fixed effects? Yes No Yes No Yes Control for omitted wealth? No Yes No Yes No Savings measure Gross -0.642* -0.084 0.764* -0.106 -0.597.... (0.365) (0.255) (0.415) (0.258) (0.268) Net -0.610* -0.200 -0.729* -0.234 -0.533 H (0.364) (0.316) (0.412) (0.324) (0.274) Green 0.425** 0.405** 0.558 H 0.504H 0.801 ** (0.203) (0.178) (0.274) (0.197) (0.362) Population-adjusted 0.413** 0.392** 0.496*** 0.788*" ~ 0.560" (0.163) (0.165) (0.213) (0.182) (0.287) ~ j' ... * Significant at the 1 percent level; Hsignificant at the 5 percent level; *significant at the 10 percent level. ~ Note: Numbers in parentheses are robust standard errors corrected for serial correlation. Two-stage least squares estimates are shown for ;! population-adjusted savings. Fixed-effects estimates are shown for equations (7) and (8). Pooled ordinary least squares estimates are shown for equations (10) and (11). The sample of 64 countries and 788 observations covers the period 1970-82. f Source: Authors' analysis based on data from World Bank (various years). ... !i S1 ~ ... N .j:> w 244 THE WORLD BANK ECONOMIC REVIEW only for the two savings measures that include this adjustment, green savings and population-adjusted savings. The estimates for gross savings and net savings are negative; although none is significantly different from zero at the 5 percent significance level, the estimates for equations (7) and (8) are significant at the 10 percent level and have very large negative values. The expected posi tive correlation between current savings and future consumption changes occurs only when the savings measure is expanded to include natural capital. The sign change between net and green savings is consistent with the expec tation that estimates of f31 should be closer to 1 for savings measures that are more comprehensive. The estimates on green and population-adjusted savings remain significantly below 1 in all four equations, however. In addition, the estimates for population-adjusted savings are virtually identical to those for green savings rather than being closer to 1. One would expect the coefficient on population-adjusted savings to be biased upward if the instrumental vari ables did not completely purge that saving measure of endogeneity. In fact, the adjustment for wealth dilution does not substantially improve the empirical relation between current savings and future consumption changes. The adjustment to the dependent variable increases the absolute value of the estimates, with the estimates for green and population-adjusted savings in columns 3 and 4 about one-fifth to one-third larger than the corresponding esti mates in columns 1 and 2 and reaching values of 0.5 or more. Although these increases are not statistically significant, they are nonetheless substantial. There is thus some evidence that the adjustment moves the estimates closer to 1. The control for omitted wealth affects only the coefficients on gross and net savings, which rise toward zero and become less significant (compare column 2 with column 1 and column 4 with column 3). It makes sense that omitted vari ables bias should be greater for these measures than for the more comprehen sive ones. Although the coefficients on green and population-adjusted savings barely change, the estimates of f32i, the country-specific coefficients on the ratio of population growth rate to total population, are significantly different from zero for many countries when they are the savings measures. Consider the results for population-adjusted savings in equation (11), the model that includes all three population adjustments. Estimation of this model generated estimates of f32i for 62 countries (estimates for China and India were not possible, because of the low estimates of the omitted-wealth control vari ables for these countries).lO Fifty-one of the estimates-four-fifths of the total-were nonpositive (either significantly negative or not significantly differ ent from zero), as theory predicts they should be. (Recall that a negative f32i indicates a positive amount of omitted wealth.) Forty-five were negative, with half of those (22) significantly different from zero. There is thus econometric 10. For China and India, figures in hundredths (population growth) are divided by figures in billions (population). The resulting variables are dose enough to zero to make the matrix of explanatory variables singular when they are included in it. Ferreira, Hamilton, and Vincent 245 evidence that the estimates of produced and natural capital failed to account for significant amounts of time-invariant national wealth in about half of the countries. The plausibility of the 51 nonpositive estimates of {32i can be gauged by comparing them with the difference between the present value of future con sumption flows, capitalized over a 20-year time horizon, and the sum of the values of produced and natural capital. The present value of consumption is a broad measure of a country's total wealth (Hamilton and Hartwick 2005; World Bank 2006); this procedure identifies the residual amount of the con sumption stream that must be generated by some form of capital other than produced and natural capital. These variables need to be in total, not per capita, terms to be compared properly, because -f32i provides an estimate of total, not per capita, omitted wealth. A simple ordinary least squares regression was run with the 51 nonpositive estimates as the dependent varia ble and the corresponding residual amounts (means for 1970-82) as the explanatory variable. Although the fit was not tight (R 2 0.101), the slope parameter (-0.151) was significantly different from zero (P = 0.023). The magnitude of the slope parameter implies that the omitted wealth components determined by the estimation of equation (11) accounted for about one-sixth of the wealth omitted from the estimates of produced and natural capital. Comparison of the results in column 5 with those in column 1 indicates that differences between the definition of the dependent variable in the study by Ferreira and Vincent (2005) and this study have the greatest impact when the savings measures incorporate adjustments for natural capital. The more restric tive theoretical basis of the model used by Ferreira and Vincent, which draws on Weitzman (1976) rather than Dasgupta (2001), causes it to exaggerate the correlation between current savings and future consumption: the coefficients on green and population-adjusted savings in column 5 are not significantly different from 1, whereas those in column 1 are. V. DISCUSSION The econometric results indicate that there is a positive correlation in develop ing countries between current per capita savings and the present discounted value of changes in future per capita consumption when the measure of savings is expanded to incorporate natural resource wealth. This result holds when additional adjustments for wealth dilution linked to population growth and the effects of changing population growth rates are taken into account. Conventional savings measures are negatively related with the present value of changes in consumption; adding adjustments for natural capital reverses the relation and makes it positive, as theory predicts it should be. The improved performance of savings measures after making this adjustment is consistent with the results from the more restrictive model used by Ferreira and Vincent (2005). The results presented here imply that policymakers and economists 246 THE WORLD BANK ECONOMIC REVIEW should interpret the net national savings rates for developing countries pub lished in the World Development Indicators (World Bank various years) as signals of future consumption paths if and only if the rates include this adjust ment for natural capital. The three population-related adjustments evaluated lead to only minor improvements in the relation between current saving and changes in future con sumption. The coefficients on green and population-adjusted savings increase by a third when the dependent variable is adjusted for changes in the popu lation growth rate over time, but the increases are not statistically significant. Controlling for time-invariant omitted wealth by including the ratio of popu lation growth rate to total population improves the relation in the sense that the coefficients on gross and net savings are no longer significantly less than zero; it does not affect the coefficients on green and population-adjusted savings. This suggests that the most important unobserved time-invariant com ponents of total wealth are related to natural wealth and that the estimates of natural wealth, crude though they are, account for them surprisingly well. The lack of significant impact of the adjustment for wealth dilution is surprising. The adjustment has a substantial impact on the savings estimates, as shown in table 1 and supplemental appendix 5.3. Measurement error may well be to blame. The estimates of the stocks of produced and natural capital are crude, especially for produced capital (Pritchett 2000). Measurement error could also explain the lack of significance of the increases in the coefficients on green and population-adjusted savings when the dependent variable is adjusted for changes in the population growth rate over time, as this adjustment involves the capital stock estimates. Adjusting genuine savings for wealth dilution is justified theoretically, and the esti mates presented here and those of Dasgupta (2003) and Hamilton and Atkinson (2006) indicate that it has a potentially large impact on estimates of genuine savings. Better estimates of capital stocks are needed, however, before it can confi dently be stated that this adjustment significantly improves the performance of genuine savings as an indicator of future consumption changes. The data provided in national accounts in developing countries are of ques tionable quality. The analysis suggests three priorities for producing better data: strengthening basic national accounts data, including data on gross savings and depreciation; updating and refining estimates of natural resource extraction and harvest costs, as well as constructing time series of resource life times; and extending the coverage of natural resource data, particularly for agricultural soils, fisheries, and diamonds. REFERENCES Arrow, Kenneth J., Partha S. Dasgupta, and Karl-Goran Majer. 2003. "The Genuine Savings Criterion and the Value of Population." Economic Theory 21(2):217-25. Asheim, Geir B. 2004. "Green National Accounting with a Changing Population." Economic Theory 23(3):601-19. Ferreira, Hamilton, and Vincent 247 Balestra, Pietro, and Jayalakshmi Varadharajan-Krishnakumar. 1987. "Full Information Estimations of a System of Simultaneous Equations with Error Component Structure." Econometric Theory 3(2):223-46. Banerjee, Anindya. 1999. "Panel Data, Unit Roots, and Cointegration: An Overview" Oxford Bulletin of Economics and Statistics 6(0):607-29. Burnside, Craig, and David Dollar. 2000. "Aid, Policies, and Growth." American Economic Review 90(4):47-68. Dasgupta, Partha S. 2001. Human Well-being and the Natural Environment. New York: Oxford University Press. ---.2003. "Population, Poverty, and the Human Environment." In Karl-Goran Maler, and Jeffrey R. Vincent eds., Environmental Degradation and Institutional Responses, vol. 1 of Handbook of Environmental Economics. Amsterdam: North-Holland. Dasgupta, Partha S., and Geoffrey M. Heal. 1979. Economic Theory and Exhaustible Resources. Cambridge: Cambridge University Press. Dasgupta, Partha S., and Karl-Goran Maler. 2001. "Wealth as a Criterion for Sustainable Development." Discussion Paper 139. Beijer International Institute of Ecological Economics, Stockholm. Davidson, James E. H., David F. Hendry, fhnk Srba, and Stephen Yeo. 1978. "Econometric Modelling of the Aggregate Time-Series Relationship between Consumers' Expenditure and Income in the United Kingdom." Economic Journal 88(352):661-92. Dixit, Avinash, Peter Hammond, and Michael Hoel. 1980. "On Hartwick's Rule for Regular Maximin Paths of Capital Accumulation and Resource Depletion." Review of Economic Studies 47(3): 551-56. Easterly, W. 1999. "The Ghost of Financing Gap: Testing The Growth Model Used in the International Financial Institutions." Journal of Development Economics 60(2):423-38. Easterly, W., and R. Levine. 2001. "It's Not Factor Accumulation: Stylized Facts and Growth Models." World Bank Economic Review 15(2):]77-219. Ferreira, Susana, and Jeffrey R. Vincent. 2005. "Genuine Savings: Leading Indicator of Sustainable Development?" Economic Development and Cultural Change 53(3}:737-54. Gelb, Alan, and Associates. 1988. Oil Windfalls: Blessing or Curse? New York: Oxford University Press. Hamilton, James D. 1994. Time Series Analysis. Princeton, N.].: Princeton University Press. Hamilton, Kirk. 2002. "Sustaining per Capita Welfare with Growing Population: Theory and Measurement." Paper presented at the Second World Congress of Environmental and Resource Economists, June 27-30, Monterey, California. Hamilton, Kirk, and Giles Atkinson. 2006. Wealth, Welfare and Sustainability: Advances in Measuring Sustainable Development. Cheltenham, UK: Edward Elgar. Hamilton, Kirk, and Michael Clemens. 1999. "Genuine Savings Rates in Developing Countries." World Bank Economic Review 13(2}:333-56. Hamilton, Kirk, and John M. Hartwick. 2005. "Investing Exhaustible Resource Rents and the Path of Consumption." Canadian Journal of Economics 38(2):615-21. Hartwick, John M. 1977. "lntergenerational Equity and the Investing of Rents from Exhaustible Resources." American Economic Review 67(5}:972-74. Levine, Ross, and David Renelt. 1992. "A Sensitivity Analysis of Cross-Country Growth Regressions." American Economic Review 82(4}:942-63. Perrings, Charles, and Jeffrey R. Vincent eds. 2003. Natural Resource Accounting and Economic Development: Theory and Practice. Cheltenham, UK: Edward Elgar. Pritchett, Lant. 2000. "The Tyranny of Concepts: CUDIE (Cumulated, Depreciated, Investment Effort) Is Not Capital." Journal of Economic Growth 5(4):361-84. Sachs, Jeffrey D., and Andrew M. Warner. 1995. Natural Resource Abundance and Economic Growth. NBER Working Paper 5398. Cambridge, Mass.: National Bureau of Economic Research. 248 THE WORLD BANK ECONOMIC REVIEW Sala-I-Mattin, X. 1997. "I Just Ran Two Million Regressions." American Economic Review 87(2): 178-83. United Nations. 1993. Integrated Environmental and Economic Accounting. Studies in Methods, Series F No. 61. New York: United Nations. Weitzman, Mattin L. 1976. "On the Welfare Significance of National Product in a Dynamic Economy." Quarterly Journal of Economics 90(1):156-62. - - - . 2003. Income, Wealth, and the Maximum Principle. Cambridge, Mass.: Harvard University Press. World Bank. 1997. Expanding the Measure of Wealth: Indicators of Sustainable Development. ESD Studies and Monographs Series 17. World Bank, Washington, D.C. - - . 2006. Where Is the Wealth of Nationsi' Measuring Capital for the 21st Century. Washington, D.C.: World Bank. World Bank. Various years. World Development Indicators. Washington, D.C.: World Bank. Conlparison of Net Benefits of Incentive-Based and Command and Control Environmental Regulations: The Case of Santiago, Chile Raul O'Ryan and Jose Miguel Sanchez The ambient permit system proposed in the literature for cost-effective pollution reduction is difficult to implement and may result in lower net benefits than using another instrument. The article develops a model for comparing the environmental net benefits of three policy instruments for Santiago, Chile, when the policy problem is to meet a given ambient quality standard. Two market-based instruments-the amhient permit system and a simpler emission permit system-are examined along with an emission standard, a command and control instrument usually favored by regulators. Both emission permit system and emission standard are costlier than the amhient permit system, sometimes in large part because they improve ambient emis sion concentrations beyond the required target in much of the city, but the ambient permit system requires a lower degree of control to comply with the standard. The somewhat costlier emission permit system and emission standard provide much higher net benefits than the ambient permit system when the health benefits of their "exces sive" air quality improvements are taken into account. These benefits are different from the fact that an amhient permit system is administratively costlier to implement. JEL code: Q25 Theory suggests that when a regulator wants to obtain a cost-effective (or minimum cost) solution for improving environmental quality in a given airshed or watershed, tradable permits or pollution taxes are the appropriate instru ment. For the simple case of a uniformly distributed pollutant, the solution is a unique emission tax or an emission permit system that allows one-for-one emission trades among sources in different locations. This simplifies implemen tation, requiring only the total allocated emission permits that allow reaching Raul O'Ryan (corresponding author) is an associate professor of economics in the Department of Industrial Engineering at Universidad de Chile, Santiago; his email addressisroryan@dii.uchile.c1. Jose Miguel Sanchez is professor of economics in the Instituto de Economia at Pontificia Universidad Cat6lica de Chile, Santiago; his e-mail address is jsanchez@faceapuc.cL The authors would like to thank Juan Pablo Montero for helpful comments and suggestions and Rodrigo Bravo, Jaques Clerc, and Carlos Holz for excellent research assistance. They also benefited greatly from the comments of three anonymous referees. An earlier draft of this article was presented at the Second World Congress of Environmental and Resource Economists at Monterey, California, in June 2002. The authors gratefully acknowledge financial support from Fondecyt grant 1990617. THE WORLD BANK ECOl'OMlC REVIEW, VOL.22, No.2, pp. 249-269 doi:l0.l093/wber/ihm013 Advance Access Publication August 31, 2007 The Author 2007. 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@oxfordjournals.org 249 250 THE WORLD BANK ECONOMIC REVIEW the required air quality target. A unique price for each emission permit would result independent of the location of the emitting source. However, when the pollutant is not uniformly distributed-as is the case for particulates and many other local pollutants-the optimal system requires that pollution permits be issued not for the amount of emissions at the source, as in an emission permit system, but for the deposition at each receptor point in the airshed, through an ambient permit system. The required overall air quality must be obtained when measured by depositions at each receptor point. As a result, different prices for each unit of concentration reduction emerge at each receptor location. The design and implementation of the instrument become quite complex, requiring multiple interactions among sources that are not based on one-for-one emission trades. To ease implementation, a simple-but not optimal-approximation is to define different trading zones within which sources can trade on a one-for-one basis. Any trading between zones, if allowed at all, must be based on transfer coefficients that consider how pollutants disperse. An example is the Regional Clean Air Management Program in Southern California (RECLAIM), which defines two different zones. Emission permits have been issued for each zone, but trading between these zones is not allowed. 1 Simulation studies for both developed and developing economies of the static efficiency gains from the use of incentive-based instruments, in particular of an ambient permit system, rather than of command and control instruments or an emission permit system, conclude that the cost reductions produced by an ambient permit system are significant in some cases and not very large in others (Atkinson and Lewis 1974; Hahn and Noll 1982; Seskin, Anderson, and Reid 1983; Krupnick 1986; McGartland and Oates 1985; Spofford and Paulsen 1988; Portney 1990; O'Ryan 1996)? An important caveat, however, is that ambient concentrations in many receptor locations are higher under the ambient permit system than under the emission permit system or command and control instruments, while still meeting the pollution reduction target. As a result, the magnitude of the cost reductions from an ambient permit system stems both from the efficiency gains related to equalizing the pollutant reduction marginal costs or cost per unit of pollutant concentration at the receptor location-a true efficiency gain-and from the lower degree of overall required pollution control (Tietenberg 1985).3 1. This assumes that emissions from one zone do not affect the other zone, which is a simplification that allows implementing the system (www.aqmd.gov/reclaim). 2. This ranking of instruments based on cost-effectiveness assumes no uncertainty of benefits and costs, perfectly monitored emissions, complete enforcement, and no asymmetric information. The magnitude of the efficiency gains depends on numerous factors, including dispersion characteristics of the pollutant, relative size and abatement costs of sources, and number of emitting sources (see Tietenberg 1985). 3. This result is true for a unique or dominant receptor location under the ambient permit system. Otherwise, with many receptors the marginal cost of emission reduction for each source is equal to the sum across all receptors of the shadow price of the pollutant concentration at each receptor times the impact of the source's emissions on that receptor. O'Ryan and Sanchez 251 If no value is assigned to the higher overall level of pollution reduction achieved by the emission permit system and the command and control instru ments, these instruments will be considered less desirable from a social perspec tive than the ambient permit system. The problem is that cost-effective approaches implicitly assign a shadow price of zero to improvements that exceed the target. If, "however, reduced concentrations below the level of the standards bring with them improvements in health or the environment, command and control instruments approaches will produce greater benefits tban incentive based approaches" (Oates, Portney, and McGartland 1989, p. 1233). Consequently, comparision of instruments without correcting for these benefits is unfair and may be misleading. Two approaches can be used to overcome this problem. One is to eliminate the lower degree of required control component by requiring that all instruments comply with the same air quality standards in all receptor locations, as is done by O'Ryan (1996). The comparison in this case is still in a cost-effectiveness framework. A second approach is to determine the net benefits for each instrument, allowing for a more complete comparison using a cost-benefit analysis. This article compares the net benefits of an ambient permit system, an emission permit system, and an emission standard, a command and control instrument, in Santiago, Chile, using cost-benefit analysis. Its contribution to the literature is to point out that regulatory schemes that are simpler to implement than the ambient permit system can also yield higher net benefits. 4 Which pollution control system yields the highest net benefits is an empirical question. The authors are not aware of any of the study that answers this question in a developing economy, and there are few studies that address the question in developed economies. In a compari son of a uniform standard and an ambient permit system in Baltimore, Md, Oates, Portney, and McGartland (1989) conclude that the resulting net benefits of the uniform standard are only slightly lower (US$6 million). In developing economies, where few pollution control efforts have been undertaken, abatement costs are usually not very high and the health benefits of improving air quality can be significant. As a result, the net benefits of improving air quality may favor the use of the emission permit system and command and control instruments. The health benefits of improved air quality under these instruments will outweigh their relative cost disadvantage com pared with an ambient permit system. To examine this hypothesis, Santiago'S emission permit system, the Sistema de Compensaciones, is compared with an ambient permit system and an effluent concentration standard (a command and control instrument), The next section presents an overview of the air pollution problem in Santiago. Section II addresses the compliance costs of reaching given air quality targets using market-based instruments and command and control instruments. 4. The additional benefits of reduced transaction costs from a simpler system are not evaluated in this analysis. 252 THE WORLD BANK ECONO>V!IC REVIEW A linear programming model is used to establish the total costs of achieving a desired air quality standard for each instrument. The following sections present the population-based health benefits associated with each instrument, and then compare the net benefits of applying the ambient permit system and the two second-best policies. The last section presents the main policy conclusions and suggests future research lines. L SANTIAGO'S AIR POLLUTION PROBLEM Santiago, Chile, like many large cities in developing economies, suffers from severe air pollution. During winter, concentrations of particulate matter of ten micrometers in diameter (PM10) constantly exceed the established ambient stan dards. An extensive international epidemiologic literature reports illness and pre mature deaths due to exposure to airborne particulate matters. Studies have found that 5.2 million inhabitants were affected in the city because of these high levels of PM10 pollution. s The city's policy-makers have been struggling since the early 1990s to improve air quality, implementing Decontamination Plans in 1990 and 1997 (for details, see ORyan and Larraguibel 2000). For particulate matter emissions from large stationary sources-industrial boilers and processes, and large residential and commercial heaters-a relatively stringent effluent concentration standard was established in 1992. To introduce flexibility, an emission permit system for particulates was introduced in March 1992, under which existing pollution sources can sell or a buy permits, depending on whether their estimated emissions are below or above their grandfathered permits. The system does not consider emission banking. Permits are expressed in kilograms per day and are traded at a one-for-one ratio. All trades require approval by the regulatory agency. Annual compliance inspections reconcile emis sions with the number of permits held by each source. A source that fails to cover its emissions with permits incurs heavy penalties, including the possibility of a temporary shutdown. 6 While an emission permit system was known to be subop timal from a cost-effectiveness perspective, a more complicated ambient permit system was rejected because the required models for implementing it were not available and trades were believed to be unnecessarily complicated. 7 However, there was no explicit evaluation of this decision or of its impacts. 5. Ostro and others (1996) found a strong association between PM10 and daily mortality rates among Santiago residents after controlling for several potential other factors. Ostro and others (1999) found a statistically significant association between PM10 and medical visits for lower respiratory tract illness in children. 6. For an analysis of the emission trading Program see Montero, Sanchez, and Katz (2002) and O'Ryan (2002). 7. Ambient permit systems are difficult to implement because of information and model requirements. In particular, implementing such a system would require knowing the contribution to concentrations at different receptor locations of each of the sources included in the system. Additionally, the acceptability of the instrument by sources is negatively affected since two otherwise similar sources would face different trading rules simply because they are in different locations. O'Ryan and Sanchez 253 FIGURE 1. Baseline Emissions of Particulate Matter (PM10) in Santiago, Chile, in 1998 _500-600 -400-500 0300-400 0200-300 _100-200 _0.0-100 Source: Authors' analysis based on data from CONAMA (2000). To examine the spatial configuration of emissions from fixed-point sources in Santiago, the city can be divided into a 34 x 34 kilometer grids of 289 (2 x 2 kilometers) cells that contain the relevant sources of air pollution in Santiago, as well as most of the exposed population. This area of the city con tains 1,098 fixed-point sources. Total PM10 emissions in the city from these sources reached 2.55 tons a day in 1998 (CONAMA 2000).8 Figure 1 presents average daily PM10 emissions from each cell in the grid, for that year. Point sources are clustered in a few zones. The cell with highest emissions is in the northwestern part of the city and emits 594 kilograms per day, 23 percent of the total PMlO emitted by point sources in the city.9 Of the 289 cells of the grid, only 7 are highly polluting (emit more than three percent of total emis sions) and the 14 most polluting cells account for 65 percent of total emissions. These emissions spread over the rest of the city, affecting air quality in each celL 8. Even though this value seems low, together with emissions by mobile sources (roughly double those by fixed point sources) and the serious thermal inversion problem in Santiago, air quality concentrations exceed the standards discussed previously. 9. This cell includes a power plant with both natural gas- and diesel-powered generators, the largest single emitting source in the city and the only power plant in Santiago. Despite the magnitude of the source, it is included in this analysis since no strategic behavior should be observed. Additionally, there does not seem to be any important incentive for the power plant to hoard permits since it is the only power plant in the city and there is no possibility that another one will be authorized to operate in the city. As a result, the plant has been included in the current tradable permit program. 254 THE WORLD BANK ECONOMIC REVIEW II. COSTS OF IMPROVING AIR QUALITY UNDER ALTERNATIVE REGULATORY INSTRUMENTS The general setting is that there are n sources of pollution spatially distributed in the city. Air quality is measured at K receptor points, and a ton of pollution emitted by the firm i has a different impact on air quality at receptor k than a ton emitted by the firm j. Generally, the regulator wants to reach a vector Q'~ = (q~ . . .q~ ... q~) of maximum permitted ambient pollution concentrations. As is usual in policy formulation, the same standard is imposed on all locations-for all k, q~ = q.~.10 The Policy Instruments Three policies are evaluated for Santiago: two market-based instruments (ambient permit system and emission permit system) and one command and control instrument (an effluent concentration or emission standard). For the spatially differentiated ambient permit system, it is assumed that permits, defined in units of concentration at each receptor, are distributed to achieve the desired unique air quality goal at each receptor. Trades are not undertaken on a one-for-one emissions basis. This is the traditional cost effective benchmark policy. Under the marketable emission permits system, total allowable emissions are established for fixed sources in the airshed. Permits in an amount equal to these emissions are distributed to polluters, who can then buy and sell them on a one-for-one emissions basis. The number of permits each source buys or sells is the result of the cost minimization of compliance costs by each source. Under the uniform effluent concentration standard, all point sources are required to emit at concentrations lower or equal to a unique stack concen tration standard. Total compliance costs are then the sum of the compliance costs for each source needed to at least meet the stack concentration standard. Conceptual Framework for Comparing the Compliance Costs of Each Instrument To compare policy instruments, it is necessary to impose the condition that they reach the desired air quality goal at all receptor locations. However, differ ent policy instruments typically result in different concentrations at each recep tor location. To stay as close to reality as possible, it is usually accepted that the target has been reached when at least one receptor location has a concen tration of q* -the binding receptor-and the others are the same or lower. For this reason, to compare compliance costs, the command and control scheme and emission permit system will be defined so as to achieve the same aIr quality standard at their binding receptors as the ambient permit system. 10. Primary standards that are established to protect health are usually required by law to be the same everywhere in the country. O'Ryan and Sanchez 255 Formally, the cost-effective ambient permit system instrument is used to obtain the least cost solution to achieve a maximum permitted ambient concen tration of q* at K receptor points in the city. This can be expressed as the fol lowing problem (Montgomery 1972): (1 ) s.t.Q* 2: ED; E 2: 0 where D is an n x K dispersion matrix (dik is the impact of a ton of pollution emitted by source i on concentrations at receptor k), E is a 1 x n vector of emissions by n firms in the city, Cj(ei) is the cost to firm i of emitting ei, and Q~' the K-component vector of target concentrations. Under the ambient permit system, there are K types of permits (one for each receptor) that give firms the right to increase ambient concentrations at each receptor. It is well known that as long as permits totaling q* are given out for each receptor and the K sets of permits are traded in competitive markets, the ambient permit system minimizes the cost of achieving Q" (Montgomery 1972). Under an emission permit system, permits equal to E" tons of emissions are distributed to the n firms, and firms trade emission permits one-for-one basis. If the permit market is competitive, the emission permit system is a solution to the following problem: (2) The emission vector that solves problem (2), E', implies a vector Q' E'D. Plotting total costs against the largest element of Q' (qmax) gives the cost of achieving q* = qmax under the emission permit system. If problem (1) has been solved for a given q*, then E* has to be varied until the largest elements of Q' coincide with q*. Under the emission standard, each source's emissions depend on the size of the source (gas flow) and hours of operation per day. The resulting emissions C vector after the standard is applied, E will imply a vector of ambient concen , trations, QC= ECD. The standard that would make the largest elements of QC coincide with q* is the standard to be compared with the ambient permit system and emission permit system. Empirical Estimation of Abatement Costs and Concentrations To estimate the abatement costs under each instrument required to reach the concentration target, it is necessary to know both the abatement cost function, Cj (ej), and the matrix D relating the vector of emissions to concentrations. The cost of abatement for each source depends on the applicable control 256 THE WORLD BANK ECONOMIC REVIEW alternatives. On the basis of the literature (Bretschneider and Kurfurst 1987; Vatavuk 1990; Aranda 1996; Bravo 2000) and expert opinion, two categories of abatement alternatives were identified for the main processes in Santiago: collection devices such as cyclones, multicyclones, bag filters, and wet scrub bers, and for some sources, a change of fuel. Each control option was also assigned an abatement efficiency value. 11 The costs of collection devices were estimated based on estimates of the net discounted cash flow of total capital investments and net annual operating costs incurred each year over the useful life of the equipment. The present value of switching to cleaner fuels was stimated based on estimates of the cost of transformation and the cost differential associated with using a different fuel. Control devices of different sizes were costed. Analytical cost relations were established for each control alternative (see supplemental Appendix 5.1). For each option, the minimum cost required to reach the required standard was used. However, the lack of flexibility may impose a high cost on some point sources, resulting in overall costs that are higher than that under an ambient permit system. To relate concentrations to emissions, the natural systems model is rep resented by the environmental "transfer" coefficient, d ib of the dispersion matrix D. A tool that simulates the dispersion process for Santiago was used to obtain these coefficients, based on a multiple cell model that is solved using mass conservation equations. 12 The wind fields had to be averaged over the day, and meteorological conditions reflecting episode conditions (days in which the air quality standard is exceeded) had to be selected.13 A total of 28 episode days were used, and the corresponding transfer coefficients were averaged. As a result, the transfer coefficients reflect the impact of a unit of emissions on con centration levels in each cell of the grid for adverse meteorological conditions. 14 The Simulation Model Each policy instrument is defined using different policy targets: air quality at each receptor location for the ambient permit system, total emissions for the emission permit system, and a uniform stack concentration target for the emis sion standard. To compare the compliance costs of these policy instruments, 11. These are presented in supplemental Appendix S.1. For the model each source was assigned only the options applicable to it. It is not assumed that existing abatement technologies are dismantled when there is a fuel switch, and the conservative assumption is made that no extra reductions are obtained when control equipment exists. Mixtures of more abatement and fuel switching were not considered, based on expert opinion that suggested that the technical options that were economically feasible are those considered in table S.l of supplemental Appendix S.1. 12. The coefficients are derived in Munoz (1993). 13. The results are presented in supplemental Appendix S.2, "Transfer Coefficients," and discussed in detail in Munoz (1993). , 14. These concentrations do not include secondary particulate matter generated by nitrogen dioxide and sulfur dioxide emissions, as there are no models available for this for Santiago. However, efforts are being initiated to estimate the impact of these emissions in the city. O'Ryan and Sanchez 257 TABLE 1. Level of Application of Emission Permit System and Emission Standard to Reach Target Concentration Ambient permit system concentration Emission permit system Required emission standard target (micrograms per cubic meter) (kilograms per day) (micrograms per cubic meter) 29.1 1,832 90.0 27.8 1,556 37.0 26.5 1,324 13.0 25.2 1,063 8.0 22.9 955 2.3 Source: Authors' analysis based on data from Bravo (2000). the ambient concentration reached in the binding receptor under each instru ment is used as a common target. lS Specifically, different reductions in pol lution concentrations relative to the binding receptor are used as targets for each policy instrument. Table 1 presents the level of application of each instru ment required to reach the same concentration target as defined by the ambient permit system. For the simulation exercise, the problem for each instrument is specified as a linear programming model with binary variables. For the ambient permit system, the model considers the objective function and environmental con straint-a concentration target at each receptor location-presented above. The solution determines which control option was used by each of the sources con sidered to comply at minimum cost. Summing individual compliance costs over all sources results in total compliance costs. To simulate the other two policy instruments, only the environmental con straint has to be modified. The emission permit system is similar to the ambient permit system, but must comply with an overall emission target-total emissions must be lower than a predetermined target. Under the emission stan dard, each emitting source must comply with a target effluent concentration. 16 Once each source has made its cost-minimizing decision, the resulting emis sions in each cell are added to obtain the aggregate emissions on an average episode day. These emissions are then transformed into concentrations at each point of the grid using transfer coefficients, making it possible to compare the average daily concentration reductions in episode days under each instrument and the costs of reaching these reductions. Specifically, the compliance costs under each policy instrument are estimated using the following model that considers an objective function and two con straints: a technological constraint common to all instruments and an environ mental constraint specific to each one. The model considers a total of 1,098 15. This guarantees that in all other receptor locations, air quality is the same or better. 16. The result of global minimization of costs is identical in the case of the emission standard to the individual cost minimization problem and for this reason the same model can be applied. In both cases, the source will choose the unique technology that enables complying at minimum cost. 258 THE WORLD BANK ECONOMIC REVIEW emission sources and a maximum of 10 abatement options for each source. Each of the 289 cells of the Santiago grid is a receptor location. The simulation is carried out for a linear programming model with a binary variable. The model is formulated using the GAMS software and results are obtained with CPLEX solver. The general model is as follows: 1.089 10 Objective function: MinI: I : CTi,tXi,t 1=1 t=1 where CT1,t is the annual cost of applying technology t to source I, and Xi, t is the binary variable that determines whether technology t is applied to source i. 10 Technological I:X;,t = 1 'Vi = 1", ,,1,098 constraint: t=1 Environmental constraint: specific to each instrument. For the ambient permit system, the specific environmental constraints are 1,098 10 289 I : I: I : EiHOi(l'kl,kUBj,d1 - EFFi,t)Xi,t 1=1 t=1 k=1 :=;Qk 'Vk 1, ... ,289 where (l'k',kis the transfer coefficient representing the effect emissions in zone k have on concentrations at location k', HOi is the hours of operation of source i per day, UBi,k is a dummy variable taking a value of one if source i is located in cell k and zero otherwise, EFFj,t is the efficiency in emission reductions of technology t being applied to source i, and Qk is the air quality target for location k (and for all receptor locations). For the emission permit system, the specific environmental constraints are 1,098 10 I : I : Ei HO i(1 EFFi,t)Xi,t:=; E i=1 t=1 where Ej is the total emission of source i (in kilograms per hour), and E is the aggregate emission target. Q'Ryan and Sanchez 259 For the emission standard, the specific environmental constraints are 10 L:0i(1-EFFi,t)Xi,t:::;O Vi 1, ...,1,098 t=l where 0i is the effluent concentration level (in milligrams per cubic meter) of source i obtained as emissions divided by flow, and 0 is the effluent concen tration standard (milligrams per cubic meter). For programming purposes, the targets defined by each instrument are set through Qk, E, and n. Targets implying concentrations at each receptor location ranging from 29.1 to 22.9 micrograms per cubic meter were evaluated. Lower targets are not possible in the worst receptor location without reducing activity at some sources or closing them down, options not considered in this study. 17 Compliance Costs under Alternative Policies The model yields both costs and concentrations per cell of the grid. Before pre senting the results, it is necessary to make a correction to current emissions. Natural gas had only been introduced in 1998 in Santiago, and many sources that could profitably switch to this fuel had not done so yet. To eliminate any distortionary effect on source decisions, it is assumed that all sources that can profit from switching to natural gas do so at the start of the program. Consequently, only the costs and benefits from additional reductions are considered. As expected, the ambient permit system is clearly the most cost-effective instrument. The maximum reduction can be obtained with an annual cost for participating sources of almost US$20 million, less than half the cost for the other policy instruments. The annualized compliance costs and resulting reductions in concentrations for each policy for fixed-point sources in Santiago are presented in table 2. 18 The reduction in compliance costs for the ambient permit system is consider able. The emission permit system is particularly expensive when small reductions are required, for example, for a 29.1 micrograms per cubic meter concentration, the target emission permit system costs are 45 times those of ambient permit system. However, over the range of reduction options for con centration targets lower than 28.7 micrograms per cubic meter, the costs for similar reductions are only 3-20 times higher with an emission permit system than with an ambient permit system. Compliance costs under the emission standard are even more expensive, between 3 and 35 times higher than the ambient permit system for most of the reduction range. The emission standard 17. Even with the best available control technology, concentrations cannot be reduced more for the thermoelectric megasource. 18. Concentrations consider only fixed-point sources. When mobile sources are included, concentrations increase about 50 percent. 260 THE WORLD BANK ECONOMIC REVIEW TABLE 2. Annualized Compliance Costs for Different Concentration Reductions Relative to the Worst Cell (in million U.S. dollars) Concentration target (micrograms Ambient permit Emission permit Emission per cubic meter) system system standard 29.3' 29.1 0.01 0.45 0.07 28.7 0.06 0.96 0.56 28.2 0.1 2 2 27.8 0.2 2 3 27.4 0.2 2 7 26.9 0.3 3 9 26.5 0.5 4 9 26.1 0.7 6 10 25.6 1.1 7 12 25.2 2 12 17 24.8 2 14 21 24.3 3 14 30 23.9 6 24 31 23.4 10 27 34 23.0 13 45 38 22.9 19 51 48 'Current concentration level in the binding receptor. Source: Authors' analysis based on data from Bravo (2000). is also more expensive than the emISSIon permit system for most of the reduction range except for extremely small or large reductions. For very low and very high values of the target, the emission permit system is more costly than the command and control instrument. This is not an unex pected result, because the emission permit system is not cost-effective and so can be more costly than a command and control instrument for some specific reduction goals. This type of result, documented in other studies (Tietenberg 1985), depends on the target, the relative compliance cost functions, and the relative size and number of sources (O'Ryan 2006). Air Quality at Each Receptor Location and Population- Weighted Concentrations A key result is that concentration reductions are different in each receptor location-for the same target-under each policy instrument. This shows that part of the cost reductions from the ambient permit system is not related to effi ciency gains, but is because of the lower degree of required control. Since health effects are related both to pollutant concentrations and to the size of the exposed population in each cell, estimation of pollution exposure under each instrument for each target requires estimation of population-weighted concentrations in each cell and summation of them across all cells. For the four receptors with the highest pollutant concentrations, population-weighted O'Ryan and Sanchez 261 FIGURE 2. Population-Weighted Pollution Concentrations as a Function of the Target Concentration for Selected Receptor Locations, by Pollution Control Instrument Concentration at receptor location 16 15 14 13 -+-- Ambient permit system 12 - - Emission permit system 11 Emission standard 10 9 8 28.7 28.2 27.4 26.1 25.2 23.9 22.9 Concentration target (IJm m-3) Source: Authors' analysis based on data from Chilean National Institute of Statistics. concentrations are higher under the ambient permit system than under the other two instruments, reflecting the lower degree of abatement required under the ambient permit system (Figure 2). As a consequence, the health benefits from applying each instrument will be different. In particular, the emission permit system, which imposes larger improvements in population-weighted air quality, would be expected to result in higher health benefits than the other instruments. The following section esti mates these health benefits. III. THE HEALTH BENEFITS OF IMPROVED AIR QUALITY The damage function approach, frequently used in environmental cost-benefit analysis, is used to estimate the health-related benefits of improved air quality (see, for example, Ostro 1996; Environment Canada 1997; European Commission 1998; USEPA 2000). The methodology involves four steps. First, the change in emissions is determined for each policy instrument. Second, the resulting impact on concentrations is estimated. Third, the effects of the reductions in pollutant concentration on various health outcomes are estimated. The changes in health outcomes are quantified using dose-response functions for a set of health effects for which there are well-established statistical relations in the environmental epidemiologic literature. These dose-response functions are applied to the exposed population to determine the population-weighted health effects. Forth, these health effects are valued in monetary units and summed over the different effects, the individuals exposed, and time. Dose-Response Functions The dose-response functions used were obtained from the environmental epi demiologic literature. For mortality the dose-response function used was 262 THE WORLD BANK ECONOMIC REVIEW TABLE 3. Dose-Response Coefficients for Human Health Effects from Particulate Matter (PMIO) and Unit Costs of Health Effects Concentration Unit cost in response 1998 (in U.S. Source Health effect category parameter dollars)U Ostro and Acute mortality (ICD 460) (percent 0.1 % 700,000 others (1996) increase per one microgram per cubic meter change in annual average PM10 concentration) Burnett and Hospital admissions for respiratory illness 6.73 x 10-4 1,600 others (1995) (lCD 480-86) (individual risk factor per one microgram per cubic meter change in annual average PM10 concentration) Burnett and Hospital admissions for cardiac illness 6.4 x 10- 4 3,500 others (1995) (lCD 410,413,427, and 428) (individual risk factor per one microgram per cubic meter change in annual average PM10 concentration) Emergency room visits for respiratory 80 illness (a parameter that relates total emergency room visits to the total number of hospital admissions in 1995 is used instead of a dose-response function. Emergency room visits were six times the number of hospital admissions) Ostro (1990) Restricted activity days, adult population 0.0168 16 (individual risk factor per one microgram per cubic meter change in annual average PM10 concentration) Dockery and Lower respiratory illness in children 0.0011 170 others (1996) (individual risk factor per one microgram per cubic meter change in annual average PM10 concentration) Abbey and Chronic bronchitis, population over age 25 6.1 x 10- 5 140,000 others (1993) (individual risk factor per one microgram per cubic meter change in annual average PM10 concentration) Krupnick, Acute respiratory symptoms (individual risk 0.1679 9 Harrington, and factor per one microgram per cubic Ostro (1990) meter change in annual average PMI0 concentration) Whittemore and Asthma attacks, among asthmatic 0.059 170 Korn (1980) population (individual risk factor per one microgram per cubic meter change in annual average PM10 concentration) Note: ICD is international classification of diseases. UNumbers have been rounded up to avoid giving a sense of false precision. SouTce: Authors' analyses based on data sources shown in table 3 and Holz and Sanchez (2000) for unit costs. O'Ryan and Sanchez 263 estimated for Santiago (Ostro and others 1996). For the other health effects, the functions were obtained from epidemiologic studies estimated for other populations, although the selection criteria used followed Ostro and others (1996). A large body of literature relates adverse health effects with ambient concen trations of PMlO. The concentration response parameters reported in table 3 are typically obtained as the mean value reported by epidemiologic studies selected as providing the most reliable results. Most of the studies estimated linear and log-linear models, which imply a continuum of effects even at low concentration levels. This is justified by the fact that studies have failed to find thresholds for effects associated with particulate matter. In addition, many recent epidemiologic studies have found an association between particulate matter and health effects throughout the whole range of concentrations, even for levels under the primary air quality standards of the U.S. Environmental Protection Agency. There is also little evidence that the slopes of the dose response functions diminish significantly at lower concentrations (Ostro 1996, p. 4). As a consequence, the functions used in this study assume that the slope of the dose-response function is the same regardless of the concentration level. 19 Finally, since these dose-response functions consider average annual PMlO concentrations, the average daily episode concentrations estimated previously had to be converted to annual values. For this, the factors estimated by Jorquera (2002a, b) were used, which represent average dispersion conditions for each month in Santiago at four different receptor locations. Since his results do not vary much by location, the average results for the four locations were used. Average dispersion conditions in the worst winter month (June) are more than four times as bad as in the best month (January) (table 4). To estimate the average annual reduction in PM10 concentrations, these factors are assumed to represent the average dispersion conditions for each month relative to the episode conditions (which has a factor of 1). Consequently, the average is a weighted average, where the weights are the number of days in the month rela tive to the total annual number of days times the relative dispersion factor. Monetary Valuation of Health Effects For valuing a reduction in mortality from lowering pollution levels, the concept of the value of a statistical life is used, estimated from willingness to pay studies. The value of a statistical life is the average of 13 studies selected 19. See also European Commission (1998, vol. 7, pp. 133-134): "for many of these pollutants, there is clearly a threshold at the individual level, in the sense that most people are not realistically at risk of severe acute health effects at current background levels of air pollution. There is however no good evidence of a threshold at the population level; i.e., it appears that, for a large population even at low background concentrations, some vulnerable people are exposed some of the time to concentrations which do have an adverse effect. This understanding first grew in the context of ambient particles, where the no threshold concept is now well established as a basis for understanding and for policy.» 264 THE WORLD BANK ECONOMIC REVIEW TABLE 4. Relative Dispersion Factors for Each Month Month Relative dispersion factor Number of days January 0.239 31 February 0.279 28 March 0.366 31 April 0.579 30 May 0.805 31 June 1.000 30 July 0.859 31 August 0.646 31 September 0.431 30 October 0.279 31 November 0.251 30 December 0.251 31 Source: Authors' analysis based on data from Jorquera (2002a, b). by the U.S. Environmental Protection Agency that report the lowest values. The values were deflated using the gross national product (GNP) per capita in purchasing power parity terms estimated for 1999 by the World Bank to account for differences in GNP per capita between the United States and Chile. For reductions in illnesses, no willingness to pay studies are available; therefore, the cost of illness estimates from Holz and Sanchez (2000) was used. 20 This approach considers direct treatment costs plus lost income as a measure of pro ductivity loss during illness. This method is simple, but it has several limit ations. It is a lower bound estimate of the true willingness to pay for reductions in illness because it does not consider other costs, such as pain and inconvenience. In addition, it does not consider the fact that people can take defensive actions. The third column of table 3 presents the unit values for each health effect used for the monetary valuation in this analysis. Health Benefits The ambient permit system results in substantially lower health benefits than the emission permit system and the emission standard (figure 3). The differ ences are largely because each policy imposes different reductions in each cell. The annual benefits obtained are on the order of tens of millions of dollars a year, similar to the annualized costs of reducing emissions. 20. The value of a statistical life estimate used in this study is lower than that estimated by Rizzi and Ortuzar (2003) for Chile using a stated choice approach in which individuals are asked to choose among alternatives. Their estimation adjusted to the 1998 U.S. dollars is approximately $800,000. However, in a recent paper by Rizzi (2005), also using stated-choice surveys, estimated a value of a statistical life for Chile of between US$200,000 and US$300,000. O'Ryan and Sanchez . 265 FlGURE 3. Annual Population-Weighted Health Benefits Associated with Ambient Particulate Matter (PMI0) Concentration Targets, by Pollution Control Instrument Health benefits (millions of U.S. dollars per year) 60 ! 50 n- PI . '." -n ~ --+- Ambient permit system ~ 40 I · Emission permit system .. IA" ~ Emission standard i 30 ~ ... .....LX - 20 ~ ..... .A ... 10 o .I " ...... ...... 4 -Y r+- ..... ....... 29.329.1 28.728.227.827.426.926.526.1 25.625.2 24.8 24.3 23.9 23.4 23 ~ I I Micrograms per cubic meter Source: Authors' analysis based on data from Chilean National Institute of Statistics. IV. COMPARING COSTS AND BENEFITS Subtracting the annual costs of each policy instrument from the annual benefits yields the net annual benefits to be expected from each policy instrument (figure 4). The net benefits are significantly higher for the emission permit FIGURE 4. Annual Net Benefits Associated with Ambient Particulate Matter (PMI0) Concentration Targets, by Pollution Control Instrument Net benefits (millions of U.S. dollars per year) 35r--r----~!--~~--r-~~--r--r~r-,-~--r--r-, 30 i -_ -- ...... .... 1--j- -+rIII.-=>.;-+---+-_ . '--- IF' 1 i · \ 25+-~~~...~~-+--.~~~~)/7~+--+-- i\ ..~ I i ~ . ...-r ____ '. , -+-Ambientpermitsystem 20 f-.. J"" , .- __ Emission permit system i II-1IL+,___ ,---+----'t---I---+--- f..---I--+-'!f--+\I-----I '--_ _---' 15 -t---:---(,j'--LI"+---1- I . \ Emission standard 10 V I! -----+--+-.....-:r+---..,-t -I..~ ..... il,' 5 . . - . ~ 1\ o .l ....-.L.-~~ i \l \ 29.3 i 29.1 ! 28.7 28.2127.8 27.4 26.9 26.5 26.1 -5+-~--'-'-+---+--+---+----'--+-- 25.6' 25.2 24.8 24.3 23.9123.4 :1 1 -10~~~--~~~--~~~--~~~--~~~--~~ I I, I Micrograms per cubic meter Source: Authors' analysis based on data from Bravo (2000) and Chilean National Institute of Statistics. 266 THE WORLD BANK ECONOMIC REVIEW system and the emission standard than that for the ambient permit system. The maximum net benefit is obtained at a PM10 concentration of 25.2 micrograms per cubic meter using the emission permit system. These net benefits are approximately US$32 million per year, almost four times higher than the maximum net benefits under the ambient permit system. On average, net benefits are ten times higher under the emission permit system and almost six times higher under the emission standard than under the ambient permit system, with the difference even higher in many cases. For example, for a 28.2 micrograms per cubic meter concentration target, the net benefits of the emission permit system are 22 times higher than those from the ambient permit system, and 12 times higher than those of the emission stan dard. In other cases, the difference is small. For example, for a PM10 concen tration level of 24.3 micrograms per cubic meter, net benefits from the emission permit system are only 3.4 higher and those from the emission stan dard are only 2.7 times higher than those from the ambient permit system. Requirements to achieve concentration levels below 23 micrograms per cubic meter have negative net benefits because of the sharp increases in cost, even when using flexible instruments. The implication is that the regulatory authority must determine the reduction targets carefully to capture most of the net benefits. A difference as small as three micrograms per cubic meter in the required reduction target can result in significantly lower net benefits. For mortality, different values of a statistical life do not change the ranking of instruments and the net benefit-maximizing concentration target. For example, with a lower value of a statistical life of US$300,000 the maximum net benefits are achieved with an emission permit system at a concentration level of 25.6 micrograms per cubic meter. The net benefits are, of course, lower, reaching only US$19 million a year. As conjectured, in a developing country such as Chile, where little effort has previously been undertaken to reduce air pollution, the benefits of better air quality associated with an emission permit system or an emission standard out weigh the relatively small compliance cost reductions obtained with the more cost-effective ambient permit system. Clearly, the decision to apply an emission permit system for Santiago is correct when both costs and benefits are taken into account. V. CONCLUSIONS The choice of instrument to regulate PM10 pollution that yields the highest net benefit is an empirical matter. For Santiago, a simulation model was used to rank policy instruments using given transfer coefficients, emission coefficients, cost estimates, coefficients for health effects, and unit costs of health effects. The analysis assumed away some of the issues currently being discussed in the theoretical literature on instrument choice, such as imperfect emission O'Ryan and Sanchez 267 monitoring, information asymmetries, dynamic incentives for innovation, and incomplete enforcement of regulation. Correcting for the difference in benefits associated with each instrument makes a significant difference in the choice of policy instrument to be used when the air quality goal is fixed and uniform across the airshed, as is usually the case. When only a cost-effectiveness criterion is used, the ambient permit system is clearly the preferred option for Santiago, reducing costs significantly compared with the emission permit system and the emission standard over a relevant range of pollution concentration levels. However, when the benefits associated with the overcontrol achieved using these two instruments are included, the emission permit system has the highest net benefits and the ambient permit system has the lowest net benefits over a wide range of plaus ible reduction possibilities. In this latter case one of the main advantages of an ambient permit system plays against it. Since it is able to impose reductions that closely match the uniform standard in different parts of the city, it does not take advantage of the significant health benefits from reducing concentrations more than required by the standard. The efficiency gains of the ambient permit system are much smaller than the economic losses from the health impacts resulting from the higher pollutant concentrations allowed by this instrument. While in principle an ambient permit system could be designed to exactly emulate the concen trations reached by the other two instruments, and this would then clearly be the best option, in practice regulators set up a uniform air quality standard within an airshed rather than a system of differentiated standards. The emission permit system and the emission standard have higher net benefits than the ambient permit system. An emission permit system is a par ticularly good policy choice for Santiago. Even though there are efficiency losses compared with an ambient permit system, these are more than compen sated for by the health benefits obtained as a result of the reductions in pollu tant concentrations in excess of the required standard. An emission permit system is also much simpler to implement than a trading system that involves spatial complexities in each trade. These results may be applicable to other developing economies where control costs are not extremely high because emissions control is at an early stage. The health benefits from an emission permit system or an emission stan dard may outweigh the lower abatement costs from an ambient permit system. For developed economies, which do not face the same initial conditions (Oates, Portney, and McGartland 1989), the significant reductions in control costs associated with an ambient permit system might outweigh the losses in health benefits compared with other policies. VI. SUPPLEMENTARY MATERIAL Supplementary material is available online at http://wber.oxfordjournals.orgl 268 THE WORLD BANK ECONOMIC REVIEW REFERENCES Abbey, D.E., F. Petersen, P.K. Mills, and W.L. Beeson. 1993. "Long Term Ambient Concentrations of Total Suspended Particulates, Pzone and Sulphur Dioxide and Respiratory Symptoms in a Non-Smoking Population." Archives of Environmental Health 48( 1 ):33-46. Aranda, P. 1996. "Costos y Efectividades de Opciones de Control para la Contaminaci6n del Aire en Santiago." Thesis, Universidad de Chile, Santiago. Atkinson, S., and D. Lewis. 1974. "A Cost-effectiveness Analysis of Alternative Air Quality Control Strategies." Journal of Environmental Economics and Management 1(3):237-50. Bravo, R. 2000. "Proposici6n y Evaluaci6n de Instrumentos de Incentivo Econ6mico para mejorar la CaUdad del Aire en Santiago: Aplicaci6n al Caso de las Fuentes Fijas." Thesis, Universidad de Chile, Santiago. Bretschneider, B., and J. Kurfurst. 1987. Air Pollution Control Technology. Fundamental Aspects of Pollution Control and Environmental Science. Vol. 8. New York: Elsevier. Burnett, R.T., D. Krewski, R. Vincent, T. Dann, and J. Brook. 1995. "Associations Between Ambient Particulate Sulfate and Admissions to Ontario Hospitals for Cardiac and Respiratory Diseases." American Journal of Epidemiology 142(1):15-22. CONAMA (Chilean National Environmental Commission). 2000. Mejoramiento del Inventario de Emisiones de la Region Metropolitana. Final Report. Santiago. Dockery, D.W., J. Cunningham, A.I. Damokosh, L.M. Neas, J.D. Spengler, P. Koutrakis, J.H. Ware, M. Raizenne, and F.E. Speizer. 1996. "Health Effects of Acid Aerosols on North American Children: Respiratory Symptoms." Environmental Health Perspectives 104(5):500-5. Environment Canada. 1997. "Human Health and Environmental Benefits of Reducing Sulphur Dioxide Emissions." In Towards a National Acid Rain Strategy. Submitted to the National Air Issues Coordinating Committee by The Acidifying Emissions Task Group October 1997. (www.ec.gc.ca/ acidrainlacidr/acidctoc_e.htm). European Commission. 1998. "ExternE: Externalities of Energy-A Research Project of the European Commission. Vol. 7: Methodology 1998 Update." European Commission, Directorate-General XII, Science Research and Development, Office for Official Publications of the European Communities, L-2920 Luxembourg. (www.externe.info). Hahn, R., and R. Noll. 1982. "Designing a Market for Tradable Emission Permits." In W. Magat ed., Reform of Environmental Regulation. Cambridge, Mass.: Ballinger. Holz, J.C., and J.M. Sanchez. 2000. "Estimation of Unitary Costs of Mortality and Morbidity and its Application to Assess the Health Benefits of the Santiago Decontamination Plan." Instituto de Economia, Pontificia Universidad Cat6lica de Chile, Santiago. Jorquera, H. 2002a. "Air Quality in Santiago, Chile: A Box Modeleling Approach: I. Carbon Monoxide, Nitrogen Oxides, and Sulfur Dioxide." Atmospheric Environment 36(2):315-30. 2002b. "Air Quality in Santiago, Chile: A Box Modeling Approach: II. PM2.5, Coarse and PMI0 Particulate Matter Fractions." Atmospheric Environment 36(2):331-44. Krupnick, A. 1986. "Costs of Alternative Policies for the Control of Nitrogen Dioxide in Baltimore." Journal of Environmental Economics and Management 13(2):189-97. Krupnick, A.J., W. Harrington, and B. Ostro. 1990. "Ambient Ozone and Acute Health Effects: Evidence from Daily Data." Journal of Environmental Economics and Management 18( 1): 1-18. McGartiand, A., and W. Oates. 1985. "Marketable Permits for the Prevention of Environmental Degradation." Journal of Environmental Economics and Management 12(3):207-28. Montero, J.P., J.M. Sanchez, and R. Katz. 2002. "A Market-based Environmental Policy Experiment in Chile." Journal of Law and Economics 45(1):267-87. Montgomery, David. 1972. "Markets in Licenses and Efficient Pollution Control Programs." Journal of Economic Theory 5(3):395-418. o 'Ryan and Sanchez 269 Muiioz, Ricardo. 1993. "Desarrollo de un Modelo de Celdas Multiples para la Simulacion de la Dispersion de Contaminantes en la Atmosfera: Aplicacion a la Ciudad de Santiago.» Thesis, Universidad de Chile, Santiago. Oates, W., W. Portney, and A. McGartiand. 1989. "The Net Benefits of Incentive-Based Regulation: A Case Study of Environmental Standard Setting." The American Economic Review 79(5):1233-48. O'Ryan, R. 1996. "Cost-effective Policies to Improve Urban Air Quality in Santiago, Chile." Journal of Enl'ironmental Economics and Management 31(3):302-13. - - - . 2002. "Emissions Trading in Santiago: Why Has it Not Worked, but Been Successful?" Paper presented at Concerted Action for Tradable Emissions Permits Workshop, University College, March 24-25, London. - - - . 2006. "Factors that Determine the Cost-effectiveness Ranking of Second-best Instruments for Environmental Regulation." Journal of Regulatory Economics 30(2}:179-98. O'Ryan, R., and L LarraguibeL 2000. "Contaminacion del Aire en Santiago: Estado Actual y Soluciones." Center for Applied Economics Working Paper 75. Universidad de Chile, Santiago. Ostro, B. 1990. "Associations Between Morbidity and Alternative Measures of Particulate Matter." Risk Analysis 10(3):421-27. - - . 1996. "A Methodology for Estimating Air Pollution Health Effects." World Health Organization, Office of Global and Integrated Environmental Health, Geneva. Ostro, B., J.M. Sanchez, c. Aranda, and G. Eskeland. 1996. "Air Pollution and Mortality: Results from a Study of Santiago, Chile." Journal of Exposure Analysis and Environmental Epidemiology 6(1):97-114. Ostro, B., G. Eskeland, J.M. Sanchez, and T. Feyzioglu. 1999. "Air Pollution and Health Effects: A Study of Medical Visits among Children in Santiago, Chile." Environmental Health Perspectives 107(1 ):69-73. Portney, P.R. 1990. Public Policies for Environmental Protection. Washington, D.C.: Resources for the Future. Rizzi, Lt 2005. "The Subjective Value of Road Safety in Chile." Proceedings of Road Safety Conference on Four Continents, Warsaw, October 5 -7, Poland. Rizzi, Ll., and J.D. Ortu.zar. 2003. "Stated Preference in the Valuation of Interurban Road Safety." Accident Analysis and Prevention 35(1):9-22. Seskin, E., R. Anderson, and R. Reid. 1983. "An Empirical Analysis of Economic Strategies for Controlling Air Pollution." Journal of Environmental Economics and Management 10(2}:112-24. Spofford, W., and C. Paulsen. 1988. "Efficiency Properties of Source Control Policies for Air Pollution Control: An Empirical Application to the Lower Delaware Valley." Resources for the Future Discussion Paper QE18-13. Resources for the Future, Washington, D.C. Tietenberg, T.H. 1985. Emissions Trading, an Exercise in Reforming Pollution Policy. Washington, D.C.: Resources for the Future. USEPA (United States Environmental Protection Agency). 2000. Regulatory [mpact AnalysiS: "Heavy Duty Engine and Vehicle Standards and Highway Diesel Fuel Sulfur Control Requirements." Office of Air and Radiation. Washington, D.C. Vatavuk, W. 1990. Estimating Costs of Air Pollution Control. Chelsea, Mich.: Lewis Publishers. Whittemore, A. S., and E. L Korn. 1980. "Asthma and Air Pollution in the Los Angeles Area." American Journal of Public Health 70(7):687-96. Women·s Power, Conditional Cash Transfers, and Schooling in Nicaragua Seth R. Gitter and Bradford L. Barham The Social Safety Net (Red de Protecci6n Social, RPS) program in Nicaragua is one of many conditional cash transfer programs that pay households cash stipends in exchange for school attendance and regular visits to health clinics by the children. A key feature is that payments go to the female head of household. Previous research suggests that exogenous transfers to women are more likely to be spent on their children's health, nutrition, and education and thus to reinforce the goals of these programs. Randomized experimental data from RPS are used to test for heterogeneous program impacts on school enrollment and spending based on a woman's power, as proxied by her years of schooling relative to her husband's years of schooling. The results confirm previous findings that more household resources are devoted to chil dren when women are more powerful. However, when a woman's power greatly exceeds her husband's, additional female power reduces school enrollment. RPS impacts on schooling are much larger than the expected income effects estimated from the control group, although no evidence is found that female power alters the impact of RPS on school enrollment. The conditionality of RPS is probably decisive. While RPS significantly increases food and education expenditures, the impact is attributable primarily to income effects. JEL codes: D13, H31, 120 Many poverty alleviation programs in developing countries stipulate that pay ments or benefits be given to the female head of household (Rawlings and Rubio 2005). The justification for targeting women is based on theoretical models and empirical findings that show that payments received by women are more likely to be spent on improving the welfare of children (for theoretical work, see Kanbur and Haddad 1994; Haddad, Hoddinott, and Alderman 1997; Basu 2006; for empirical research, see Schultz 1990; Thomas 1990; Doss 1996). This article explores the impact of this requirement in Nicaragua's Seth R. Gitter (corresponding author) is an assistant professor of economics at Towson University; his email address is sgitter@towson.edu. Bradford L. Barham is a professor of agricultural and applied economics at the University of Wisconsin-Madison; his email address is barham@aae.wisc.edu. The authors thank IFPRI for providing the data and Michael Carter, Jean-Paul Chavas, Jeremy Foltz, Carolyn Heinrich, John Maluccio, Jaime de Melo, seminar participants at the University of Wisconsin Madison, and three anonymous reviewers for their guidance and comments on earlier drafts. pp. 271-290 THE WORW BA]>,l( ECDNOMIC REVIEW, VOL. 22, No.2, doi:l0.l093/wberllhn006 Advance Access Publication May 22, 2008 © The Author 2008. 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: journa)s.permissions@oxfordjournals.org 271 272 THE WORLD BANK ECONOMIC REVIEW Social Safety Net (Red de Proteccion Social, RPS), a conditional cash transfer program that pays women cash if their children attend school and they make regular visits to health care clinics. Empirical evidence is limited on the effectiveness of targeting conditional cash transfers to women in order to raise school enrollment and affect other consumption outcomes. Three critical components of conditional cash transfers confound efforts to cleanly identify the impacts on school enrollment: income and two nonincome effects, conditionality and intrahousehold effects. The nonincome effects of targeted conditional cash transfers potentially include both the conditionality requirements of program participation (essentially a price effect) and the intrahousehold effects of providing women with the trans fer. In other words, the education outcomes are also shaped by two distinct effects that are both part of the program's treatment. Would a cash transfer without conditions achieve similar school enrollment outcomes (because education is a normal or even superior good for low-income families)? As for identifying the intrahousehold impacts of cash transfers tar geted to women, an ideal experiment would randomly provide some transfers to men and some to women to determine how the impacts differ. Absent such a study design, one could examine household spending patterns of the treat ment group (that are not conditional) to determine whether intrahousehold effects matter and in what ways. One could also look at the effects of intrahou sehold differences in the control group (or the baseline data of the treatment group) to determine whether preexisting differences in education and spending patterns are consistent with power differences between men and women. This article explores how RPS shapes education and spending patterns, with an eye on all three effects: income, conditionality, and intrahousehold impacts. On the intrahousehold side, the intention is to identify whether preexisting gender power structures are at work and to determine whether they are mitigated by the program, either through conditionality or by targeting transfers to women. By providing transfers directly to women, RPS also has the potential to empower women by increasing the resources they control. However, household resources are potentially fungible, raising a concern that other family resources may be real located away from children, offsetting the impact of the transfer. This phenom enon could be captured empirically by demonstrating smaller effects of conditional cash transfers on key outcomes in households in which men have more power. By targeting transfers to women, RPS has the implicit goal of helping ensure that money is spent on women and children, who might otherwise receive smaller shares of household resources in male-dominated households. Thus, it is also possible that the impacts of conditional cash transfer programs could be higher in male-dominated households if the transfers have the effect of changing behavior in the family that did not contribute to salutary outcomes for children. The empirical analysis uses experimental methods that compare treatment and control groups. It adds to previous studies of the impact of conditional cash transfers by estimating heterogeneous program impacts based on Gitter and Barham 273 intrahousehold power differences. The power measure used is based on the ratio of years of school completed by the female and male heads of household. Women's intrahousehold power is assumed to increase as the female to male education ratio rises. This measure is arguably better in terms of exogeneity than male and female wage earnings used in some other studies, because earn ings are endogenous to intrahousehold decision making and correlated with child wages, both of which could affect schooling decisions. The article is organized as follows. Section I places this work within the context of the current literature and identifies its conceptual contributions. Section II presents the empirical approach to analyzing the impact of power and RPS on schooling and household spending. Section III provides back ground information on RPS along with descriptive statistics on variables of interest. Results of the estimations are reported in section IV, with conclusions and suggestions for further study provided in the last section. I. LITERATURE REVIEW This article links three related streams of literature. The first is the intrahouse hold bargaining literature, which suggests that heterogeneous preferences between men and women can lead to different household decisions depending on power relations. The second attempts to measure the impacts of conditional cash transfer programs, with a focus on which aspects of the program (con ditions or cash) are more effective in obtaining the desired results. The third seeks to determine whether there are demonstrable effects of targeting con ditional cash transfers to women. The theoretical and empirical literature on how households make decisions is well developed (Schultz 2002; Basu 2006). Two basic types of household models have been used to study decisions on child schooling and labor and the allocation of consumption expenditures between private and shared goods. Unitary models assume either that there is a benevolent dictator or that house hold members share the same preferences and pool their resources to maximize a single household utility function (Becker 1981). Households with hetero geneous preferences and a set balance of power are guided by a single utility function, even when one member is a nonbenevolent dictator. In these models, targeting transfers to women should have no impact on a household's allo cation of spending (except through household income effects; Attanasio and Lechene 2002). Nonunitary models generally examine decisions made by men and women who have distinct preferences and make decisions somewhere along a spectrum between full cooperation and conflict (McElroy and Horney 1981; Chiappori 1992; Basu 2006). Differences in bargaining power influence whose preferences gain greater expression in the household's choices. These models often assume that women have stronger preferences for child schooling and health outcomes; they therefore predict distinct effects of increases in nonwage income 274 THE WORLD BANK ECONO:\lIC REVIEW depending on who receives the transfer. The motivation for giving conditional cash transfers to women is· the assumption that women's higher propensity to spend on household shared goods will augment program effects. Power relations between fathers and mothers have been shown empirically to affect child schooling outcomes (Binder 1999; Adato and others 2003; Iyigun and Walsh 2007), with relative income increases for women raising child school attendance. Thomas (1990) and Schultz (1990) show that nonwage income received by mothers is more likely than income received by fathers to be spent on children's health or schooling. The child's gender may also affect the resources received. Thomas (1994) shows that Brazilian mothers' nonwage income positively affected their daughter's health but not their sons'. Duflo (2003) shows that the impacts of exogenous income transfers through old-age pensions in South Africa were more likely to increase health outcomes of granddaughters of grandmothers than any other grandparent grandchild relation. Emerson and Souza (2007) find that in Brazil fathers' edu cation has a greater impact than mothers' education on sons' attainment, while mothers' education matters more to daughters' attainment. Attanasio and Lechene (2002) and Adato and others (2003) examine the intrahousehold decision-making effects of conditional cash transfer programs. Both consider Progresa (now known as Oportunidades), a Mexican conditional cash transfer program. Attanasio and Lechene test the impact of Progresa and women's bargaining power as measured by the relative wages (potential and actual) of men and women on the share of household expenditures devoted to different goods (food, alcohol, transportation, services, and clothing).1 The importance of women's power is supported by results that show that an increase in the relative income of women, including from Progresa's targeted cash transfer, has a positive relation to the share of expenditures allocated to children's clothing and food. Using a qualitative approach, Adato and others (2003) find that Progresa decreased the likelihood that husbands reported being the sole decision maker regarding spending on child health care, school attendance, and clothing, suggesting that the targeted cash transfer increased women's bargaining power. One critical methodological and empirical issue in the intrahousehold litera ture is how to measure bargaining power. Adato and others (2003) suggest that each member's bargaining power is based on four factors: control over resources, influence over the bargaining process, interpersonal networks, and basic attitudinal attributes. Most research suggests that those with greater own assets or income (actual or potential) can exert more power, because they can withdraw from the household more easily (Doss 1996). In this sense, con ditional cash transfers could increase the viability of women's exit options and 1. Because wage data are not available in the RPS sample, the results of Attanasio and Lechene cannot be compared with those obtained here. Gitter and Barham 275 strengthen their bargaining power, as long as women receive the transfer even if they leave the household. This article uses the ratio of the number of years of school completed by the female to the number of years of school completed by the male head of house hold as a measure of power? It assumes that as the female to male education ratio increases, women are likely to have more decision-making power. This measure is similar to but less crude than the literacy ratio Basu, Narayan, and Ravallion (2001) use (they use literacy because they assume that a literate member can withhold information from illiterate members to gain an advan tage). One advantage of the education ranking approach over power measures that rely on relative wages or income is that education is exogenous to current income levels, which are themselves endogenous to fundamental household decisions regarding labor allocation. (Wages could not have been used in any case, because the RPS sample data did not collect wage information.) Many studies suggest that women's power is both positively and monotoni ca Ily related to spending on children and school enrollment. This assumption has been questioned by some recent work, however (Felkey 2005; Basu 2006; Lancaster, Maitra, and Ray 2006). Using an intra household theoretical frame work, Basu (2006) shows that if the woman has more power than the man, she will garner a greater share of the income produced by child labor. Based on this result, he posits that as her power continues to increase, she will receive more benefits from child labor, while the benefits of schooling may stay the same. He therefore concludes that if women become sufficiently more powerful than men, additional female power may actually result in a decline in school enrollment. Lancaster, Maitra, and Ray (2006) and Felkey (2005) provide empirical evidence in support of Basu's hypothesis, using samples from India and Bulgaria. In Nicaragua even when women have as much education as their husbands, they still may not have equal power, because of cultural norms. However, at a certain point women with more education than their husbands could have suf ficient power to sustain the nonmonotonic result suggested by Basu. Basu's hypothesis is tested here by examining the nonlinear effects of the female to male education ratio on child school enrollment and household spending outcomes. The article also adds to these previous studies by testing whether cash payments made to mothers are likely to increase their power. Previous work has shown that RPS and Progresa have been effective at increasing school enrollment rates and encouraging spending on food (for RPS, see Maluccio and Flores 2004; for Progresa, see Schultz 2004; Hoddinott and Skoufias 2004). The regression specification used by Hoddinott and Skoufias includes total consumption (including the transfer) as well as program 2. The female-male power ratio is usually given on a 0 to 1 scale. The ratio here ranges from 0 to 9. Though it could be normalized to a 0 to 1 scale, a non-normalized ratio is used for easier interpretation of the coefficients. 276 THE WORLD BANK ECONOMIC REVIEW participation indicator variables. This combination helps provide estimates for the income and nonincome impacts of Progresa on food spending. Hoddinott and Skoufias (2004) find that nonincome effects account for about half of the total impact of Progresa for total food expenditures and a higher percentage for expenditures on fruit, vegetables, and animal products. They place much of the credit for these impacts on lectures women received as part of Progresa that encourage proper nutrition through expenditures on fruit, vegetables, and milk. Attanasio and Lechene (2002) contend that the impacts may also be tied to targeting payments to women. Both could be correct: the health education lectures provided by Progresa could shape preferences, and targeted transfers could enhance women's bargaining power and thus their capacity to reveal those preferences. What is not dear is whether those expen ditures may also have been viewed implicitly by the recipients as part of the conditionality of Progresa. In a simulation of the Bolsa Escola Program, a Brazilian conditional cash transfer, Bourguignon, Ferreira, and Leite (2003) find that both the conditionality of school attendance and income effects increase school enrollment. Other research suggests that preexisting household conditions can shape the impact of a transfer. de Janvry and Sadoulet (2006) argue that conditional cash transfer programs can improve their results by moving from a uniform transfer size to one tied to easily observable household characteristics that alter program impacts. The relative education levels of parents are used here as an easily observable characteristic that may create heterogeneous impacts based on differences in preferences and power between men and women. Attanasio and Lechene (2002) find that payments made to women increase expenditures on food and schooling by increasing women's power (as measured by the ratio of female to male income), but they do not test for nonlinearities in this relation. It is possible that transfers to less powerful women may increase their power enough to participate in decision making and thus augment the targeting effect (as suggested by Adato and others 2003). Another possibility is that less powerful women may not be able to keep the whole transfer or that men may withdraw funds from the household to increase personal expenditures or leisure time. In the case of RPS, there is little evidence for this occurring, as increases in total household consumption were equivalent to the size of the transfer (Maluccio and Flores 2004). de Janvry and Sadoulet (2006) include parental literacy in their estimation of the impact of Progresa on the child schooling decision. The impacts of the literacy of the mother and father are estimated as separate effects, not relative to one another as a measure of power. They find that both father's and mother's literacy increase schooling and decrease the size of the transfer required for the child to attend school. Their regression does not include con trols for income, however, so parental schooling may well be capturing an income effect. Most important, they do not compare across households with Gitter and Barham 277 different female to male education ratios or other relative power measures to test for intrahousehold effects. II. AN EMPIRICAL STRATEGY FOR ESTIMATING THE IMPACTS OF POWER AND RPS Three components of household schooling and resource allocation decisions are examined here. The first is the effect of power structures ex ante of program effects on education outcomes and household spending patterns. The goal of this test is to see whether the power measure provides results that are consistent with the previously cited literature-that is, whether children of more powerful women are more likely to attend school and receive a larger share of resources. The second component is an estimate that identifies income and nonin come effects. The control group is used to estimate income impacts on schooling and household spending. The income effects of a cash transfer in the control group which is the size of that of the RPS are then compared with the total effects of RPS, with the difference being an estimate of nonin come effects.The third component is the effect of women's power on program impacts on school attendance and household expenditure patterns. This component is measured by interacting variables that measure program impact and the power measure to test for heterogeneous program impacts by power. The conventional approach to analyzing the treatment effects of conditional cash transfer programs is to use cross-sectional or panel data to compare out comes in treatment and control groups. When the dependent variable of inter est (school enrollment or consumption share) is not substantially different in the baseline year in control and treatment communities, program impacts can be measured using cross-sectional data in the treatment year. However, if initial conditions (in either the dependent or independent variables) are differ ent in the treatment and control communities, then the full panel data should be used. Difference-in-difference is the standard method used to measure impacts when initial conditions are not the same in control and treatment commu nities. This method measures the difference in the changes of the outcome of interest in treatment and control communities between the first year of treat ment (year 1) and the baseline (year 0). If, for example, the outcome of interest in time period t is denoted as Ct for control communities and It for those in the treatment (intervention) group, the difference-in-difference program impact, denoted 01, is determined by 01 = (II - 10 ) - (C 1 - Co). If through randomization in the baseline the outcome of interest is equally likely in both groups, the difference-in-difference impact is equivalent to 11 - C1 · 278 THE WORLD BANK ECONOMIC REVIEW Maluccio and Flores (2004) present a basic estimation equation for difference-in-difference (equation 1). Program impacts are measured using the difference-in-difference variables; 01 the coefficient on the term Treat, which is the interaction of two binary dummy variables for treatment year (T = 1); and the treatment status of the household (RPS 1 for households in a treatment community).3 where E ict = outcome variable of interest for household (or individual) i in community c at time t, Al = 1 if year is 2001, A2 1 if year is 2002, Treat 1 if treatment year is 2001 or 2002 and household is in RPS interven tion in community c, Uic = all (observed and unobserved) household-level (or individual-level) time-invariant factors, Viet unobserved idiosyncratic house hold (or individual) and time-varying errors, and a's and o's = unknown parameters. The number of years of school completed by the female head of household divided by the number of years of school completed by the male head of house hold-relative female power by schooling years (rFPSY)-is used to measure power. As 49 percent of males have completed zero years of school, 1 is added to both numbers of school years to create a defined ratio for all households: rFPSY = (Number o(,,>!~'!.~s of school completed by female head + 1) (Number of years of school completed by male head + 1) The variable rFPSY is used to measure the impact of female power on school enrollment and household expenditures. The average rFPSY of both control and treatment groups was 1.4; comparison between treatment and control groups does not show statistically significant differences between the two groupS.4 The square of rFPSY is also used in order to test for the possible nonlinearity of the relation between power and these outcomes. The power measure is interacted with the treatment impact measure Treat to estimate the interactive effects of the power measure and RPS. The square of the power measure and the treatment impact measure are interacted to test for a nonlinear relation between power and RPS impacts. Schooling of the male and female heads of household (male_schooling and female_schooling) is added directly into the equation to control for the impact of the individual education levels. Finally, total per capita consumption (peC) is included to estimate and control for income effects, including those from RPS transfers. When pee is included, 3. For ease of interpretation, the two years are combined into a single measure of the impact of the treatment in a treatment year; doing so does not substantially affect the results. 4. A simple t-test of the mean of rFPSY between the treatment and control group yields at-statistic of 0.37. Gitter and Barham 279 the estimated impacts of nonincome effects Treat * RPS in equation (2) for all households is represented by 01' The estimated impacts of power on RPS effects are represented by 02 and 03, respectively. E iet ao + alAI + a2A2 + +a3MaleSchooling + a4FemaieSchooling + asrFPSY + a6rFPSYA2 (2) + ooRPS + 01 Treat + 02 Treat * rFPSY + 03 Treat * rFPSY A2 + f31 In Consumptionet + f32 In Sheet + Uic + Vict where E jct = 1 if child i in community c at time t is enrolled in school, and 0 otherwise; or, for expenditure data, E is expenditures for household i in community c at time t; In Consumption = log (total consumption) for household c in year t (baseline); and In Sizet = log(household size) in year t. For the first two components, this regression specification is similar to that of Hoddinott and Skoufias (2004), who estimate the impact of Progresa on food consumption-with some important distinctions. Their specification includes household characteristics, including the education of the head; the spe cification presented here includes the education of both the household head and his or her spouse separately and as a power measure. The same method used by Hoddinott and Skoufias (2004) is adopted here to separate income effects from nonincome effects by including total consumption in the regression as a control for income (including the transfer) as well as program effect measures. As Hoddinott and Skoufias note, if a conditional cash transfer alters consumption other than directly through transfers, total consumption becomes endogenous and may bias the results. This does not appear to be the case, as Maluccio and Flores (2004) find that the ex post increases in con sumption for the treatment group are not statistically significantly different from the transfer. The final component of the specification is the measurement of heterogeneous impacts of RPS based on household characteristics. The approach used is similar to that of two previous studies that measure the effect of economic shocks on RPS (Maluccio 2005) and Progresa (de Janvry and others 2006). In these studies, the heterogeneity across households is deter mined by exposure to these shocks. A measure of exposure to shocks is then interacted with the program eligibility variable. The approach here is similar, except that heterogeneity comes from the power measure rather than exposure to shock. Both types of models (school enrollment and expenditure levels) are esti mated using ordinary least squares (OLS). Estimating marginal effects is diffi cult using qualitative variable methods because of the interaction terms. Gitter and Barham (2006) find that OLS estimations of the enrollment impacts of 280 THE WORLD BANK ECONOMIC REVIEW RPS are similar to probit predictions. In all of the estimations, errors are clustered at the community level to control for unobserved heterogeneity across communities. Because the household decision on school attendance may be different for boys and girls, separate estimates are performed for boys and girls. III. SUMMARY OF THE RPS PROGRAM AND DESCRIPTIVE STATISTICS The RPS was implemented in 21 randomly selected communities in northwes tern Nicaragua (in Madriz and Matagalpa). Another 21 communities in the region served as the control group. Three survey rounds were conducted in all 42 communities, one in 2000, before program implementation, and two during the program, in 2001 and 2002. This analysis uses a subsample of the 1,300 total households in which there is a head of household who is married. This subs ample includes 1,129 households. s Participation in treatment communities was extremely high, with uptake rates of more than 95 percent of those eligible to participate. 6 Benefits include a C$2,880 ($224) annual food security transfer. 7 Households with children ages 7-13 who had not completed the fourth grade were eligible for a bimonthly transfer for school attendance of C$1,440 per year and an additional C$275 for school supplies. The average household received C$3,885, or about 18 percent of total annual household consumption expenditures. Baseline comparisons between treatment and control groups on outcomes and explanatory variables support the use of experimental methods to test for impact results. The average school enrollment for children of eligible age in the baseline sample was 77 percent, with about a 0.1 percent difference between treatment and control groups. The difference in aggregate total consumption and other consumption measures in treatment and control groups was not sig nificantly different from zero. In more than 40 percent of the households, male and female heads had com pleted the same number of years of school. The other 60 percent of households were divided evenly between those in which the male head had more schooling and those in which the female head had more schooling (table 1), The control and treatment groups had a similar average rFPSY. However, the control group had slightly more (45 percent compared with 40 percent) households in which the male and female households had the same number of years of completed 5. See Maluccio and Flores (2004) for information on the program design. They show that sample attrition rates were similar in both control and treatment communities. 6. Ninety-five percent of households were eligible to participate (Maluccio and Flores 2004). Program participation does not appear to have been affected by adult literacy, household income, or marital status. 7. Exchange rate of C$12.85 = US$l is as of September 2000. Gitter and Barham 281 TABLE 1. Descriptive Statistics of Total Household Consumption and School Enrollment Baseline Difference-in-difference Item rFPSY' Control Treatment T-statisticb 2001 c 2002 d Total household con rFPSY < 1 25,160 24,427 0.51 6,694 .... 2,773*" sumption (cordobas) rFPSY= 1 22,206 21,634 0.64 6,492 .... 5,114"" rFPSY> 1 24,291 24,051 0.81 4,164 .... 4,489"" Total 23,624 23,147 0.70 5,851" 4,284'" School enrollment, ages rFPSY< 1 78 78 -0.04 17.... 14'" 7 -13 (percent) rFPSY= 1 72 74 -0.65 15 .... 10 .... rFPSY> 1 86 80 1.50 20** 12 .... Total 77 77 -0.21 17"" 11'" "Significant at the 10 percent level. .... Significant at the 5 percent level. a(years of schooling completed by female head + l)/(years of schooling completed by male head +1). bComparison of baseline control and treatment. C(Treatment2oo1-ControhoOl) - (Treatment2ooo-ControI2ooo). d(Treatmentzooz-Control20oz) (Treatmentzooo-ControI2ooo). Source: Authors' analysis based on data described in text. schooling, while the treatment group had slightly more (31 percent compared with 26 percent) households in which women had more years of schooling. Consumption in households with more powerful females (rFPSY> 1) is similar to that in households with more powerful males (rFPSY < 1)8; con sumption is lower in households in which rFPSY = 1. This result likely reflects the fact that this group includes a significant number of households in which neither spouse completed a year of school. Previously cited literature suggests that female power is linked to higher school attendance and spending on children. The predicted relation is found in table 1, which shows that households with more powerful women (rFPSY> 1) have average baseline school enrollment rates of 82 percent (86 percent and 80 percent for the control and treatment groups, respectively), while households in which rFPSY < 1 have school enrollment of 78 percent. A t-test on average enrollment between the two groups yields a t-statistic of 1. 78. The relation between power and spending can be seen in some of the other explanatory variables (table 2). The previously cited literature suggests that households in which women have more power spend more on food and edu cation of their children. However, in the RPS sample, there is weak evidence in 8. A t-test comparing the total consumption of households with rFPSY < 1 and rFPSY> 1 yields a t-statistic of 0.6. Relative to households with unequal levels of schooling, households in which rFPSY = 1 have a t-statistic of 3.6. 282 THE WORLD BANK ECONOMIC REVIEW TABLE 2. Descriptive Statistics of PCC of Food, Milk, and Education (cordobas) Baseline Difference-in-difference Item rFPSya Control Treatment T-statistic b 2001' 2002 d Per capita food rFPSY< 1 2,838 3,100 -0.97 673"" 716 expenditures rFPSY= 1 2,755 2,829 -0.31 745"" 321 rFPSY> 1 2,744 2,919 -0.77 611 603 Total 2,801 2,969 1.20 618"" 514 Per capita milk rFPSY < 1 53 64 -0.65 57** 64"* expenditures rFPSY= 1 29 30 -0.09 24 34** rFPSY> 1 75 66 0.52 67"" 31** Total 49 50 0.20 48** 42** Per capita education rFPSY< 1 70 52 -0.93 23 75"" expenditures rFPSY= 1 30 51 -1.90 7 19 ,FPSY> 1 78 94 -0.50 -22 47 Total 56 65 -0.70 2 43 * *Significant at the 5 percent level. '(years of schooling completed by female head +1)/(years of schooling completed by male head +1). bComparison of baseline control and treatment. C(TreatmentzoOl-Controlzooll - (Treatment2ooo-Controhooo). d(Treatmentzoo2-Controlzoo2} - (Treatment2ooo-Controlzoool· Source: Authors' analysis based on data described in text. terms of total food spending, though food expenditures account for such a high proportion of total consumption (70 percent) that the deep poverty of these families may blunt differences in food expenditures evident elsewhere. Expenditure data come from self-reported household surveys on food consumption over a two-week period, which was scaled up for a year (Maluccio and Flores 2004). Children in households with a powerful woman might receive a larger pro portion of the household's food. Unfortunately, data on individual food con sumption are not available, however. One way of determining whether this is the case is to look at milk consumption (including infant formula), which is more likely to benefit children. Milk consumption does appear to be related to women's power: in the baseline data, households with rFPSY> 1 consume more milk than those with rFPSY < 1 (the difference is significant at the 10 percent level using a simple t-test). Maluccio and Flores (2004) use difference-in-difference estimates to measure program outcomes in their analysis of the total impact of RPS. Tables 1 and 2 provide basic difference-in-difference estimations for each of the rFSPSY measures for the outcomes of school enrollment, expenditures, and expenditure shares. In terms of school enrollment, the impacts are larger in households in Gitter and Barham 283 which the woman is powerful. All households saw at least a 15 percentage-point increase in enrollment the first year and a 10 percentage point increase the second year. Given that enrollment was at least 95 percent in all treatment com munities, conditionality appears to be playing the dominant role. The effects were greater, however, in households with more powerful women. One common concern is that men might withdraw money from the house hold for shared goods as women receive income from the transfer and use it for private consumption. If this concern were evident in the data, one would expect male-dominated households to have smaller expenditure impacts from RPS. In fact, in all cases except milk expenditures in 2002, impacts from RPS treatment as measured by difference-in-difference estimates show larger impacts for male dominated households than for female-dominated households. This suggests that RPS transfers to women are having the intended impact of strengthening their potential to influence household consumption and investment choices rather than being captured by men who had pretransfer power advantages. IV. ECONOMETRIC RESULTS This section presents the results of econometric estimations of factors shaping school enrollment and household expenditures. There are three major com ponents of these influences: the effect of female power ex ante of program impacts, income versus nonincome impacts, and variation in program impacts by female power. Two sets of regression results are reported: impacts on school enrollment and impacts on per capita expenditures for food, education, and milk. The regression specification is supported by the finding of an ex ante impact of female power on school enrollment and household expenditures on education. The results also show both income and nonincome effects from RPS, with nonincome effects being more important for schooling and income effects being more important for household spending patterns. The econometric analysis of school enrollment outcomes for children ages 13 includes three sets of regressions: one for all children and one each for boys and girls (table 3). The impact ex ante of gender power differences can be seen through the two rFPSY measures. The coefficients on both rFPSY (posi tive) and rFPSy 2 (negative) are statistically significant for the sample of all chil dren and girls. Children's schooling is positively associated with maternal power, except when the rFPSY ratio is larger than 5 (the case for about 3 percent of the children in the sample), at which point further maternal school ing begins to reduce enrollment. These results are consistent with the nonmo notonic relation between power and schooling found by Basu (2006). However, as the negative effect is observed only at the far tail of the distri bution, it could also indicate that there are monotonic but diminishing returns to power and schooling. For boys the quadratic term is not statistically significant, and the results suggest a positively monotonic relation between female power (rFPSY) and school enrollment. N 00 .j>. ..., x TAB L E 3. Regression on School Enrollment: Impacts of Power and RPS '" E 0 All Children Boys Girls ,.. :>' v Standard Standard '" ;> Variable Coefficient Error Coefficient Error Coefficient Standard Error z ~ RPS 1 if treatment group -0.Q18 0.018 0.026 0.025 -0.068"* 0.024"" '" I'l 0 1 if year 2001 0.048** 0.016"* 0.055*" 0.024** 0.036 0.022 z 0 1 if year 2002 0.062*" 0.016** 0.074** 0.024"* 0.044** 0.022** ~ I'l Male household head school years 0.016** 0.004** 0.022** 0.006** 0.010 0.006 :>' Female household head school years 0.008'* 0.004** 0.008 0.006 0.009 0.006 rFPSY 0.053**" 0.018"· 0.056·'" 0.028 0.044*"" 0.024** '" <: ;;; rFPSY"2 -0.006** 0.002"* -0.004 0.004 -0.007*"" 0.003** E 1 if RPS group and treatment year (Treat) 0.166** 0.030"* 0.153*"" 0.045** 0.184"* 0.041 ** Treat "rFPSY -0.030 0.026 -0.034 0.041 -0.025 0.033 Treat 'rFPSY" 2 0.004 0.005 0.003 0.008 0.005 0.006 Ln(household consumption) 0.054** 0.010*"" 0.047"" 0.015*"' 0.063"* 0.014"" Ln(household size) 0.000 0.016 0.002 0.022 -0.003 0.023 Constant 0.153 0.098 0.172 0.146 0.137 0.132 R-squared 0.09 0.10 0.09 Number of observations 4,593 2,337 2,256 ** Significant at the 5 percent level. "Joint F-test of coefficients on rFPSY = rFPSY"2 = 0 significant at the 5 percent level. Source: Authors' analysis based on data described in text. Gitter and Barham 285 Additional controls were added to the regressions for the number of school years completed by the male and female heads of household. The pooled regression shows that an additional year of school for the male head is twice as important as an additional year for the female head. These differences break mainly along child gender lines. In the sample with just boys, schooling of the male head has a larger impact (0.022) for each year of schooling compared with an extra year for the female head (0.008), although an F-test of male_years = female-JIears is not statistically significantly different from zero. This result suggests that additional school years of the male head and female head of house hold may have equal impacts. The coefficient estimates on both heads of house hold are equal for girls, although neither is statistically significant. The second component of interest-the comparison of income and nonin come effects-is captured by the RPS impact measures (Treat), because income effects are controlled for by using total household consumption (including RPS transfers). The RPS nonincome impacts on school enrollment for both years are measured at 16.6 percent for the total sample, with the impact on girls slightly higher but not statistically significantly so. This estimate is lower than that of Maluccio and Flores (2004), who estimate total impacts (income and nonincome) of 22 percent for 2001 and 18 percent for 2002. This difference suggests that the income effects are on the order of 1,4-5,4 percentage points, or about 25-33 percent of the nonincome effects. Another way to estimate income effects is to use the coefficient estimate on the variable of the natural log of total household consumption, In Consumption. The difference in the average In Consumption between treat ment and control was 0.35 in 2001 and 0.24 in 2002. With a coefficient esti mate of 0.054 on total household consumption, these differences would suggest that transferring the size of RPS would increase schooling by 1-2 per centage points. 9 This impact is slightly less than but consistent in magnitude with the difference between the estimated nonincome effects obtained here and the total effects obtained by Maluccio and Flores (2004). The combined impacts of power and RPS on school enrollment are exam ined through the interaction of the nonincome treatment impact measure (Treat) and the power ratio (rFPSY). This interaction term and its square are not statistically significant, suggesting that the impacts of RPS treatment do not vary depending on the power of the female head of household. Furthermore, when the interaction of Treat and FPSy2 is omitted, the relation between RPS impacts and power (Treat*rFSPY) is negative but not statistically significant. 1o 9. One concern is that with treatment the impact of total consumption on schooling may vary when compared with ex ante consumption patterns. Models that separately estimate the impact of consumption on schooling for only the control group yield coefficients that are not substantially different from the model presented above. These results are available from the authors on request. 10. These results are omitted because of space constraints; they are available from the authors on request. 286 THE WORLD BANK ECONOMIC REVIEW TABLE 4. Regression on per Capita Expenditures by Category: Impacts of Power and RPS Per Capita Spending on Per Capita Spending on Per Capita Spending Food Education on Milk Standard Standard Standard Variable Coefficient Error Coefficient Error Coefficient Error RPS 1 if treatment 56.3 106.0** 1.4 11.9 1.8 8.7 group 1 if year 2001 -370.8"" 99.3** 33.3** 11.2'· 21.1** 8.0·· 1 if year 2002 -423.0*· 96.9** 56.4 10.9' * 3.4 7.8 Male household -20.7 25.8 .... 4.6 2.9 1.1 3.1 head school years Female household 58.2 .... 25.0 .... 5.7"" 2.8 .... 9.6*'" 4.4** head school years Relative female 171.3 113.1 18.2 12.7 11.3 11.7 power by schooling years (rFPSY) rFPSY"2 16.7 14.7 -2.7 1.7 1.6 1.4 1 if RPS group and 445.2** 175.9 0.7 19.8 72.0 n 26.6** treatment year (Treat) Treat *rFPSY -·237.0" 144.7 1.1 16.3 -45.6" 32.3 Treat ·rFPSY"2 25.3 24.5 0.1 2.8 8,4 7.0 Ln(household 1,895.3** 52.3 .... 113.0** 5.9** 24.0'· 7.2 .... consumption) Constant 1,5524.0 521.0 1,078.7 58.6 193.8 70.2 R-squared 0.37 0.16 0.03 Number of 2,550 2,550 2,550 observations **Significant at the 5 percent leveL aJoint F-test of coefficients on rFPSY = rFPSY"2 o or Treat·rFPSY = Treat·rFPSY"2 = 0 significant at 5 percent level. Source: Authors' analysis based on data described in text. The impacts of power, RPS, and income on three types of expenditures (education, food, and milk) are estimated next (table 4). Consistent with the enrollment results, for most households female power as measured by rFPSY has a positive relation with spending on education and a negative quadratic effect. Similar to the enrollment results, the maximum value of power for enrollment occurs at an rFPSY of about 4 (which applies to about 4 percent of the sample). Unlike education, spending on food or milk in particular does not show a statistically significant relation with power. However, all three expendi ture categories show statistically significant impacts of the number of years of schooling of the female household head. The interactive effects of female Gitter and Barham 287 power and RPS on the three types of expenditures do not yield statistically sig nificant coefficients. The impacts of RPS on expenditures can be. seen through two variables: Treat, which represents nonincome effects, and InConsumption, which captures income effects measured through household consumption. Consumption of food, education, and milk increased with an increase in total household consumption, including consumption increases from RPS. Examination of the nonincome impacts of RPS as measured by the variable Treat shows significant positive impacts on spending for milk and food but not for education. The nonincome impacts on milk and food are substantial. The estimated nonin come impact on milk expenditures per capita is $Cn, more than twice the average baseline consumption. The estimated impact of RPS on food con sumption per capita is $C445, nearly a 15 percent increase over baseline consumption. The empirical analysis yields three main results. First, more female power generally leads to higher school enrollment and greater spending on education. However, consistent with the emerging literature, for households with extre mely powerful women, more female power may begin to reduce schooling or at least have no additional marginal impact. Second, nonincome effects of RPS are extremely important for school enrollment, which may not be surprising given the conditionality of the program. Nonincome impacts are evident on both food and milk per capita expenditures. Although the RPS program encourages spending on these items, such spending is not required, suggesting that non income effects other than conditionality had an impact. Two likely possibilities are the targeting of transfers to women and the nutrition education programs. Third, there is no evidence of a decreased impact of RPS on spending or schooling when women are less powerful. l l Overall, these results support the hypothesis that the goals of school enrollment and nutrition can be improved by directing funds to women and requiring school attendance. V. CONCLUSION A large body of literature on intrahousehold bargaining suggests a posltIve relation between women's power and the amount of resources devoted to children. This article uses a power measure based on the ratio of years of schooling of female to male household heads to study the impacts of a conditional cash transfer program in Nicaragua. This measure is generally consistent with the expected positive relation between women's power and child schooling, although, as suggested by Basu's (2006) model, past a certain 11. In a separate model that omits the quadratic interaction term, Treat'FPSY"2, the linear term is negative and statistically significant. 288 THE WORLD BANK ECONOMIC REVIEW point the marginal impact of additional female power on children's enrollment may be zero or negative. In targeting transfers to women, RPS and other conditional cash transfer programs seek to increase women's potential to spend money on children's schooling and other goods, such as food, that can improve children's human capital. The analysis provides evidence of the effectiveness of RPS transfers in improving the allocation of household resources toward women and children. The nonincome effects of the program are responsible for most of the nearly 20 percent increase in school enrollment; the targeting of transfers to women plays a secondary role. Running the enrollment regressions separately for girls and boys reveals that the mother's relative education level always has a positive impact on boys' edu cation outcomes. The results for girls are consistent with the nonlinear relation suggested by Basu (2006): when women's power passes a certain threshold girls' enrollment falls. Basu hypothesizes that parental power may influence the percentage of benefits from child labor garnered by each adult. This percentage may also depend on the child's gender. The nonmonotonic relation for girls but not boys suggests that when girls leave school, the percentage of the benefits received by the female head of household is larger than it is for boys. The expenditure analysis supports the effectiveness of targeting transfers to women: RPS nonincome effects accounted for a more than doubling of milk expenditures and 15 percent of the increase in food expenditures. This effect may be shaped as much by women's education as it is by their power. However, the expenditure analysis shows that the education level of the female head has a positive impact on expenditures, but that the impact of their relative power is weaker. Overall, the empirical results suggest that targeting transfers to women has been effective at increasing key welfare outcomes for all house holds, even those with greater male power. But these estimates are inferences from econometric analyses and not direct measures of treatment effects of targeting transfers to women from a randomized experiment. If one goal of conditional cash transfer programs is to strengthen and broaden the quality of information regarding the efficacy of targeting transfers to women, more detailed questions on how households allocate their resources or possibly experiments that provide targeted and nontargeted transfers should be used in future program designs. REFERENCES Adato, M., B. de la Briere, D. Mindek, and A.R. Quisumbing. 2003. "The Impact of PROGRESA on Women's Status and Intrahousehold Relations." In A.R. Quisumbing, ed., Household Decisions, Gender, and Development: A Synthesis of Recent Research. Washington, D.C.: International Food Policy Research Institute. Attanasio, 0., and V. Lechene 2002. "Test of Income Pooling in Household Decisions." Review of Economic Dynamics 5(4):720-48. Gitter and Barham 289 Basu, K. 2006. "Gender and Say: A Model of Household Behavior with Endogenously Determined Balance of Power." Economic Journal 116(511):558-80. Basu, K., A. Narayan, and M. Ravallion. 2001. "Is Literacy within Households Shared?" Labor Economics 8(6):649-65. Becker, G. 1981. A Treatise on the Family. Cambridge, Mass.: Harvard University Press. Binder, M. 1999. "Schooling Indicators during Mexico's 'Lost Decade.''' Economics of Education Review 18(2):183-99. Bourguignon, F., F. Ferreira, and P. Leite. 2003. "Conditional Cash Transfers, Schooling, and Child Labor: Micro-Simulating Brazil's Bolsa Escola Program." World Bank Economic Review 17(2):229-54. Chiappori, P. 1992. "Collective Labor Supply and Welfare." Journal of Political Economy 100(3): 437-67. de janvry, A., F. Finan, E. Sadoulet, and R. Vakis. 2006. "Can Conditional Cash Transfers Serve as Safety Nets to Keep Children at School and out of the Labor Market?" Journal of Development Economics 79(2):349-73. de janvry, A., and E. Sadoulet. 2006. "Making Conditional Cash Transfers More Efficient: Designing for Maximum Effect of the Conditionality." World Bank Economic Review 20(1):1-29. Doss, C. 1996. "Testing among Models of Intrahousehold Resource Allocation." World Development 24(10): 1597 -609. DuRo, E. 2003. "Grandmothers and Granddaughters: Old-Age Pensions and Intrahousehold Allocation in South Africa." World Bank Economic Review 17(1):1-25. Emerson, P., and A.P. Souza. 2007. "Child Labor, School Attendance, and Intrahousehold Gender Bias in Brazil." World Bank Economic Review 21(2):301-16. Felkey, A. 2005. Husbands, Wives, and the Peculiar Economics of Household Puhlic Goods. Working paper. Ithaca, N.Y.: Cornell University, Department of Economics. Gitter, S.R., and B.L. Barham. 2006. "Conditional Cash Transfers, Credit, Shocks, and Education: An Impact Evaluation of Nicaragua's RPS." University of Wisconsin, Madison, Wisc. Haddad, L., J. Hoddinott, and H. Alderman. eds. 1997. Intrahousehold Resource Allocation in Developing Countries: Methods Models and Policy. Baltimore, Md.: John Hopkins University Press. Hoddinott, J., and E. Skoufias. 2004. "The Impact of PROGRESA on Food Consumption." Economic Development and Cultural Change 53(1):37-61. Iyigun, M., and Randall P. Walsh. 2007. "Endogenous Gender Power, Household Labor Supply and the Demographic Transition." Journal of Development Economics 82(1): 138-55. Kanhur, R., and L. Haddad. 1994. "Are Better Off Households More Equal or Less Equal?" Oxford Economic Papers 46(3):445-58. Lancaster, G., P. Maitra, and R. Ray. 2006. "Endogenous Intra-household Balance of Power and Its Impact on Expenditure Patterns: Evidence from India." Economica 73(291):435-60. Maluccio, J. 2005. "Coping with the 'Coffee Crisis' in Central America: The Role of the Nicaraguan Red de Protecci6n Social." FCND Discussion Paper 188. International Food Policy Research Institute, Washington D.C. Maluccio, J., and F. Flores. 2004. "Impact Evaluation of a Conditional Cash Transfer Program: The Nicaraguan Red de Proteccion Social." FCND Discussion Paper 184. International Food Policy Research Institute, Washington D.C. McElroy, M.B., and M.J. Horney. 1981. "Nash-Bargained Household Decisions: Toward Generalization of the Theory of Demand." International Economic Review 22(2):33-49. Rawlings, L., and G. Rubio. 2005. "Evaluating the Impact of Conditional Cash Transfer Programs." World Bank Research Observer 20(1):29-55. Schultz, T. 1990. "Testing the Neoclassical of Family Labor and Fertility." Journal of Human Resources 25(4):599-634. 290 THE WORLD BANK ECONOMIC REVIEW 2002. "Why Governments Should Investment More in Girls' Education.» World Development 30(2):207-25. 2004. ~School Subsidies for the Poor: Evaluating the Mexican Progresa Poverty Program." Joumal of Development Economics 74(1):199-250. Thomas, D. 1990. "Intra-household Allocation: An Inferential Approach." Joumal of Human Resources 25(4):635-64. 1994. "Like Father, Like Son; Like Mother, Like Daughter: Parental Resources and Child Height." Joumal of Human Resources 29(4):950-88. Does Aid for Education Educate Children? Evidence from Panel Data Axel Dreher, Peter Nunnenkamp, and Rainer Thiele Most of the aid effectiveness literature has focused on the potential growth effects of aggregate aid, with inconclusive results. Considering that donors have repeatedly stressed the multidimensionality of their objectives, a more disaggregated view on aid effectiveness is warranted. The impact of aid on education is analyzed empirically for almost 100 countries over 1970-2004. The effectiveness of sector-specific aid is assessed within the framework of social production functions. The Millennium Development Goals related to education, particularly the goal of achieving universal primary school enrollment, are considered as outcome variables. The analysis suggests that higher per capita aid for education significantly increases primary school enroll ment, while increased domestic government spending on education does not. This result is robust to the method of estimation, the use of instruments to control for the endogeneity of aid, and the set of control variables included in the estimations. JEL codes: F35, 011, H52, 122 There is heated debate over whether foreign aid is effective in promoting economic development. According to Sen (2006) and Tarp (2006), Easterly's (2006) claim that aid has done "so much ill and so little good" obscures the fact that aid can work if done right. Dalgaard, Hansen, and Tarp (2004) find that overall aid has indeed been effective. Even recent surveys of the literature on aid and growth come to sharply opposing conclusions. While Doucouliagos and Paldam (2005) conclude that the aid effectiveness literature has failed to Axel Dreher (corresponding author) is a senior researcher at the KOF Swiss Economic Institute, ETH Zurich, and a research affiliate at the Institute for "::Conomic Research, Center for Economic Studies (CESifo), at the University of Munich; his email addressismail@axel-dreher.de. Peter Nunnenkamp is Head of the Research Area Global Division of Labour at the Kid Institute for the World Economy; his email addressispeter.nunnenkamp@ifw-kiel.de. Rainer Thiele is Head of the Research Area Poverty Reduction, Equity, and Development at the Kid Institute for the World Economy; his email addressisrainer.thiele@ifw-kiel.de. The authors thank Christian Conrad, Martin Gassebner, Katja Michaelowa, participants at a seminar at the Kiel Institute for the World Economy, and three anonymous referees for their useful comments and suggestions, and Michaela Rank for excellent research assistance. A supplemental appendix to this article is available at www:llwber. oxfordjournals.org TIlE WORLD BANK ECONOMIC REVIEW, VOL.22, :-l0. 2, pp. 291-314 doi:10.1093/wber/lhn003 Advance Access Publication April 11, 2008 © The Author 2008. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development I THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals. permissions@oxfordjournals.org 291 292 THE WORLD BANK ECONOMIC REVIEW establish that aid works, McGillivray and others (2005) stress that almost all research published since the late 1990s finds that it does. What both camps tend to ignore is that different types of aid are unlikely to have the same economic effects on recipient countries. In much of the litera ture, it is still common to run panel regressions with aggregate aid flows as the explanatory variable. Work by Clemens, Radelet, and Bhavnani (2004) on short-impact aid has initiated a shift toward using disaggregated aid data. It is open to debate whether a verdict on the effectiveness of aid can be reached at all as long as the analysis is restricted to the aid-growth nexus. Donors have repeatedly stressed that they pursue multiple objectives when granting aid (see, for example, Isenman and Ehrenpreis 2003). Specific-purpose aid intended to support donors' policy statements, including the empowerment of the poor through better education, tends to escape analyses narrowly focused on aid and growth. Against this backdrop, it seems appropriate to pursue a different avenue for assessing the effectiveness of aid. This article focuses on more specific outcome variables than growth. It uses disaggregated aid data to investigate the link between aid granted to the education sector and education outcomes. Education figures prominently among the Millennium Development Goals. Donors have committed themselves to helping countries achieve universal primary education by 2015 and eliminate gender disparities in education. To this end, donors have devoted an increasing share of aid resources to the edu cation sector (Thiele, Nunnenkamp, and Dreher 2007). Yet, it is open to debate whether more resources necessarily translate into better education outcomes or how aid can play a role in achieving universal primary education (Roberts 2003). The effectiveness of aid for education is assessed here within the framework of social production functions. The results show that higher per capita aid significantly increases primary school enroll ment. This outcome is robust to the method of estimation and to the set of control variables included. The article is organized as follows. Section I provides the analytical back ground and discusses the literature on aid and education. Section II addresses data issues. The method of estimation and main results are presented in Section III, and various tests for robustness are performed in Section IV. The article closes with some concluding remarks about the implications of the findings for policy and future research. I. ANALYTICAL BACKGROUND AND RELEVANT LITERATURE A social production function with education-related outcomes as the left-hand-side variable is estimated in which aid for education enters as an explanatory variable. Schultz (1988) proposed a production-demand frame work to model the educational system of countries in the 1980s. The concept of a social production function has also been used in the literature on the link Dreher, Nunnenkamp, and Thiele 293 between government expenditure and social outcomes. Hanushek (1995, p. 2) considers this concept to be "most appealing and useful" to assess the relation between school outcomes and measurable educational inputs. Recent examples following this approach include Bennell (2002), Roberts (2003), and Baldacci, Guin-Siu, and de Mello (2003). While the exact specification of social production functions varies, the common feature is that it includes major demand and supply factors (Roberts 2003). For the production function for education, demand factors typically considered include per capita income (a proxy for household poverty), adult lit eracy (a proxy for the educational status of parents), the relative size of the school population, and the level of urbanization. The "price" of schooling also affects demand, although it is typically not included in empirical cross-country analyses because of lack of data. School fees are supposed to inhibit enroll ment, which is why free universal primary education has been advocated (see, for example, Bruns, Mingat, and Rakotomalala 2003).1 Public spending on education figures prominently among the supply factors considered in social production functions. Other potentially relevant supply factors include the pupil-teacher ratio and the unit costs of education. The regression analysis conducted below extends the production function concept by adding aid for education as an additional supply factor. Various studies find that demand factors explain most of the variance in school attendance (enrollment, completion rates) and educational attainment (youth literacy, test scores) across countries. Surveying the literature, Roberts (2003) concludes that per capita income tends to be the most powerful driving force of school performance; supply-side factors, in particular public expendi ture on education, are statistically insignificant in most instances. 2 Roberts' own cross-country regression analysis corroborates the finding of ineffective public expenditure and finds that adult literacy is the main demand-side factor. Clemens (2004) shows that school enrollment rises only slowly over time and that the impact of education policy is relatively small compared with that of long-term economic changes. 3 Very few studies consider foreign aid for education as a possibly important supply-side factor in the production function; studies that do so (Michaelowa and Weber 2006; Wolf 2007) are inconclusive. The aid literature focuses on whether aid helps achieve economic and social objectives by providing 1. The success of the Mexican anti-poverty program Progresa, in which educational subsidies in the form of conditional transfer payments to poor families increased enrollment, provides evidence for this notion (Behrman, Sengupta, and Todd 2005). 2. See Clemens (2004) and the references given there. Cross-country studies include Filmer and Pritchett (1999), Mingat and Tan (2003), and Baldacci, Guin-Siu, and de Mello (2003). 3. While this rather bleak picture concerning the ability of government spending on education to raise education outcomes appears to represent the majority view in the literature, there are some notable exceptions, including Gupta, Verhoeven, and Tiongson (1999) and Baldacci and others (2004). 294 THE WORLD BANK ECONOMIC REVIEW additional resources for financing pro-poor expenditure and on the extent to which aid is fungible. Although Pettersson (2006) finds a high degree of fungibility, some other studies show the aid-expenditure link to be important. Gomanee and others (2003) and Mosley, Hudson, and Verschoor (2004) find that aid alleviates poverty through its effect on public expenditure. Gomanee and others (2005) reach the opposite conclusion in a more recent version of their article. The effects of aid working through public expenditure could be captured by estimating a system of equations that includes a public expenditure equation with aid as one explanatory variable (Mosley, Hudson, and Verschoor 2004).4 However, such an approach suffers from several problems. It is conceptually demanding, because the specification of the equations should ideally be based on a complete theoretical model, and the determinants of all dependent vari ables would have to be included in the estimations. The more conventional approach of instrumenting potentially endogenous variables in the production function for education has the advantage that instruments need explain only some fraction of the variation in the instrumented variables. Furthermore, the interpretation of coefficients in the public expenditure equation is plagued with problems, particularly regarding the aid coefficient. A coefficient that is not significantly different from zero does not necessarily imply that aid is highly fungible and that aid does not add to overall (foreign and domestic) resources devoted to education. This implication would hold only if (most) aid were accounted for in the public budget of the recipient country.s However, project aid for education-the most important mode of aid delivery, at least until recently-often remains outside the budget (Roberts 2003). If all aid remain outside the budget, full fungibility implies an aid coeffi cient of -1. As it is impossible to determine the proportion of aid outside and inside the budget, estimation of the public expenditure equation offers no meaningful insights. For this reason, estimating a system of equations is not the preferred option. The approach taken here follows the seminal contribution of Borenszstein, de Gregorio, and Lee (1998) on the economic growth effects of foreign direct investment (FDI). They consider both foreign and domestic investment in asses sing whether foreign investment is more productive in raising growth. Analogously, an enrollment equation is estimated here that includes both aid for education and government expenditure on education as explanatory variables. Aid for education may be more effective in raising enrollment rates than government expenditure on education, for several reasons (Roberts 2003). 4. This approach was adopted in an earlier version of this article. 5. The authors thank an anonymous referee for alerting them to the critical importance of this assumption. Dreher, Nunnenkamp, and Thiele 295 First, at least 75 percent of government expenditure typically consists of teacher salaries, which were not covered by donors until recently (Michaelowa and Weber 2006). Donors provide aid for building schools, supplying teaching materials, improving school management, and reforming curricula, in the hope of improving the learning environment, the efficiency of schools, and the quality of education, which may provide stronger incentives to attend school. Second, government expenditure on education is often biased against the poor, the population segment for which enrollment rates tend to be relatively low (Bennell 2002). This contrasts with donor strategies emphasizing poor and disadvantaged groups, in particular girls, for whom school attendance is often lower than for boys. Third, leakage of local funds appears to be high (Reinikka and Svensson 2001) and capture by producers and privileged consumers to be common. External donors may succeed at least partly in mitigating leakage and capture of aid funds by not channeling aid through the public budget of the recipient country, by involving local authorities in aid allocation processes, and by increasingly applying performance-based allocation rules. Measures such as these are recent, however; before the 1990 Jomtien Conference, which empha sized the importance of universal primary education, donor support concen trated on higher levels of education (Thiele, Nunnenkamp, and Dreher 2007). Consequently, it remains an empirical question whether foreign aid has been more effective than domestic public spending on education in promoting primary school enrollment. II. DATA ISSUES The data for assessing the impact of aid for education on education outcomes are far from perfect. The aid data-on commitments of sector-specific aid, including aid for education--come from the Creditor Reporting System (CRS) of the Organisation for Economic Co-operation and Development! Development Assistance Committee (OECDIDAC). These data are imperfect because commitment data tend to overstate actual aid flows (commitments may not be fully disbursed) and because sector-specific commitments go partly unreported (Michaelowa and Weber 2006). These measurement problems, which work in opposite directions, cannot be resolved, because sector-specific disbursement data are not available before 1990. The correlation between com mitments and disbursements of aid for education over the period for which both series are available is fairly high, with correlation coefficients of 0.70 for 1990-94,0.71 for 1995-99, and 0.80 for 2000-04. It can be argued that employing sector-specific aid data understates the con tribution of aid to education objectives in recent years. Several donors now favor general budget support over project aid for specific targets. The extent to which general budget support is ultimately used for educational objectives is not known. 296 THE WORLD BANK ECONOMIC REVIEW A similar argument can be made about multisector aid. However, the evalu ation of the composition of aid in Thiele, Nunnenkamp and Dreher (2007) suggests that this is unlikely to pose serious problems. In contrast to donor announcements, the shares of general budget support and multi sector aid in total aid were actually lower in 2002-04 than in the early 1990s. Nevertheless, the robustness of results is checked by replicating the estimates for a shorter period of observation (excluding recent years, in which donors may have increasingly supported educational objectives through aid that is not picked up in sector-specific aid data). Data limitations with respect to education outcome variables are well known (Roberts 2003; Bennell 2002). Ideally, the outcome variable of the pro duction function should go beyond enrollment rates to include educational attainment and the quality of education. Enrollment rates may provide a mis leading picture. Clemens (2004) draws on detailed country studies to show that rising enrollment rates came at the cost of deteriorating quality of education in some countries, as reflected by much higher pupil-teacher ratios, higher failure and repetition rates, and lower test scores. Furthermore, some countries report unreasonably high net enrollment ratios (more than 100). Qualitative dimensions of education, such as improved literacy and test scores, are not available for a sufficiently large number of countries over a suffi ciently long period of time. However, distortions resulting from the short comings of enrollment rates as an education outcome variable were minimized in several ways. First, completion rates were considered as an alternative indi cator. Second, near universal enrollment rates were checked against reported ratios of boys to girls (enrollment rates of almost 100 percent are inconsistent with gender imbalance): except in Tajikistan, no major discrepancy was detected. Third, additional estimates were run for a reduced sample, eliminat ing countries with exceptionally large increases in enrollment rates. Another data problem concerns the time-series dimension. In 2003, the United Nations Educational, Scientific, and Cultural Organization (UNESCO), the original source of the World Bank data used here, revised its estimates of net primary enrollment for 1998-2001. For some countries, this revision is associated with a major break in the series on primary enrollment. These countries were identified by comparing the old and new data in years for which both series are available (normally 1998-2000) (see Clemens 2004 for a similar approach). Discrepancies were minor (less than 2 percentage points) in 69 of the 119 sample countries for which this comparison was possible. The revision resulted in major discrepancies (more than 10 percentage points) in 15 countries, in 8 of which the old series appears to have overstated enrollment rates. This problem was dealt with in the tests for robustness by replicating the analysis for a shorter period of observation, 1970-97. In this way, the risk of inconsistencies over time can be reduced, even though the old series may suffer from systematic over- or underreporting by some countries. Dreher, Nunnenkamp, and Thiele 297 III. METHOD OF ESTIMATION AND BASE RESULTS Net primary school enrollment is the dependent variable throughout this section. The main explanatory variables of interest are aid to the education sector and government spending on education. (Using aid and spending on primary education is conceptually superior but would leave an insufficient number of observations for estimation.) Aid and government expenditure are measured on a per capita basis. In aid-growth regressions, aid is typically defined relative to the recipient country's gross domestic product (GDP). This provides a reasonable measure of the importance of foreign support relative to the recipient country's overall resources. Aid per capita is more appropriate than the aid to GDP ratio in assessing aid effectiveness with respect to specific Millennium Development Goals. In particular, achieving universal primary education requires accounting for the number of people among whom the resources devoted to education must be shared. For other relevant covariates, the analysis closely follows the literature on education production functions in considering four demand-side variables: adult literacy, per capita GDP,6 share of the population under 15, and share of the urban population in total population. Lagged education outcome is included as an additional explanatory variable in order to account for the poss ible persistence in outcomes. Additional supply-side variables suggested in the literature are added later to test for the sensitivity of results. Pooled time-series cross-section (panel data) regressions are estimated for a maximum of 96 low- and middle-income countries between 1970 and 2004 (with the exception of data on aid, which are available only since 1973).7 As some of the data are not available on an annual basis, all data are five-year averages. (The definitions of and sources for all variables are listed in table A-1; summary statistics are reported in table A-2.) The basic equation takes the following form: schooli,1 = a + .81 schooli,t_1 + .82aidi,t + .83spendingi,t + BX + 11i + e;,t (1) where school;,t represents the logarithm of primary school enrollment in country i in year t; aid;,t is per capita foreign aid to the education sector; and spending;,t is per capita government expenditure on education. X is the vector of control variables, 11i represents country fixed effects, and e;,t represents the disturbance term. The dependent variable, school;,t> is limited by a lower bound of zero and an upper bound of 100. The upper bound may lead to biased results in the sense 6. Following most earlier studies estimating social production functions (for example, Gupta, Verhoeven, and Tiongson 1999; Roberts 2003; and Michaelowa and Weber 2006), per capita GDP is measured in levels. 7. High-income countries were excluded because they receive no aid. The World Bank (2007) defines high-income countries as those with a 2005 GNI per capita of at least $10,726. 298 THE WORLD BANK ECONOMIC REVIEW that aid can have little effect on enrollment in reclplent countries with enrollment rates of close to 100 percent. 8 One way to deal with this problem is through logistic rather than log linear estimation (Fielding, McGillivray, and Torres 2005). A different route is taken here. While the analysis follows the standard approach of the literature, which includes enrollment either in levels or in logs, it tests robustness by excluding recipient countries where enrollment rates exceed a certain threshold and may thus bias results downward. Aid cannot reasonably be expected to be exogenous to school enrollment: donors typically grant more aid to countries that are less developed. Nevertheless, as a first step fixed- and random-effects models that ignore the potential endogeneity are estimated before we present specifications that allow for the endogeneity of aid and government expenditure. 9 The qualitative findings do not depend on the inclusion of random or fixed country effects. However, the random-effects specification is rejected in favor of the fixed-effects model (Hausman test, P = 0.00), so only fixed-effects esti mates are reported (table 1, column 1). An F-test also shows that fixed country effects cannot be omitted (P 0.00). By contrast, fixed-period effects did not turn out to be significant, so the estimates do not include them. The estimates reveal a considerable degree of inertia in primary school enrollment. The lagged dependent variable is highly significant. 1o Clearly, it has some explanatory power, rendering insignificant most of the other covari ates included in the regression. The fixed-effects model indicates that literacy, the share of the population under the age of 15, and the degree of urbanization do not significantly affect enrollment. 11 Most surprisingly perhaps, per capita GDP does not have the expected posi tive impact on school enrollment, even when per capita GDP is alternatively specified in logs (not shown). As noted earlier, cross-country evidence suggests that per capita income is an important determinant of enrollment. 12 By con trast, the explanatory power of per capita GDP is rather weak in the few panel studies that have been conducted. The insignificance of per capita GDP in table 1 is similar to the results obtained by Michaelowa and Weber (2006) as 8. The authors thank an anonymous referee for pointing this out. 9. The estimator is biased and inconsistent in a short panel because of the inclusion of the lagged dependent variable and fixed- or random-country effects. The GMM estimator presented later takes this into account. 10. A second lag of the dependent variable turned out to be completely insignificant. 11. Estimates based on a panel-corrected standard error model (not shown) point to a very strong relation between adult literacy and enrollment across countries, corroborating the findings of Gupta, Verhoeven, and Tiongson (1999) and Roberts (2003). 12. Even among cross-country studies, there are notable exceptions. Gupta, Verhoeven, and Tiongson (1999) do not find a robust effect of per capita GDP on enrollment. Mingat and Tan (1998) show that the cross-country correlation between per capita GDP and education is relatively weak when higher-income countries are excluded from the calculation, as they are in this article. TABLE 1. School Enrollment and per Capita Aid to Education, 1970-2004 Variable (1) (2) (3) (4) Lagged dependent 0.38 (3.82)· .. · 0.65 (6.99)· .. · 0.44 (4.63) .. · .. 0.40 (3.52) ...... Expenditure on education (per capita) 0.000034 (0.37) -0.000198 (1.47) 0.000220 (0.75) 0.001523 (1.04) Aid for education (per capita) 0.0026 (2.89)"** 0.0029 (2.34)" 0.0268 (2.69) .. ·· 0.0243 (2.24) .... Literacy rate 0.0034 (1.35) 0.0019 (1.75)' 0.0046 (1.19) 0.0081 (1.43) GDP per capita -0.000003 (0.61) 0.000008 (1.42) 0.000028 (1.29) 0.000003 (0.08) Population under 15 0.001 (0.34) 0.001 (0.88) 0.019 (2.16)'" 0.020 (2.17) .... Urbanization - 0.00022 (0.09) - 0.00015 (0.30) 0.00420 (0.97) 0.00035 (0.05) Number of countries 94 94 61 61 Number of observations 267 267 156 156 Maximum number of periods 6 4 4 4 Method Fixed effects System GMM 2SLS 2SLS R-squared (overall) 0.93 0.41 0.40 Number of instruments 53 9 9 tl Hansen test (Prob > ,I) 0.60 0.11 0.14 ~ ::s Difference-in-Sargan test (Prob > ,I) 0.17 .~ Arellano-Bond test (Prob > z) 0.24 First-stage F-test 6.32 5.55 ~ ;:t ~ Note: Numbers in parentheses are t-statistics, estimated robustly. The dependent variable is the logarithm of primary school enrollment in country i at year t. ., ;:t .".. ·Significant at the 10 percent level. ~ ., ...Significant at the 5 percent level. ..... Significant at the 1 percent level. .... ;:t ;i Source: Authors' analysis based on data described in table A-1. ~. " N \0 \0 300 THE WORLD BANK ECONOMIC REVIEW well as Baldacci and others (2004).13 The weak explanatory power of per capita GDP in panel studies may at least partly reflect the fact that these studies capture short- to medium-run effects through the time dimension, whereas the effects in cross-section analyses are purely long term. The results show a positive correlation between aid and school enrollment, with coefficients significant at the 1 percent level. As in much of the previous literature, government expenditure on education does not affect enrollment sig nificantly. Because both government expenditure and aid enter the regression, the significant coefficient of aid suggests that it is more productive than dom estic government spending in raising school enrollment. In the next step, the potential endogeneity of aid and government expenditure is taken into account (columns 2-4), starting with the system generalized method of moments (GMM) estimator, as suggested by Arellano and Bover (1995) and Blundell and Bond (1998). The dynamic panel GMM estimator exploits an assumption about the initial conditions to obtain moment conditions that remain informative even for persistent data and is considered most appropri ate in the presence of endogenous regressors. Results are based on the two-step estimator implemented by Roodman (2005) in Stata, including Windmeijer's (2005) finite sample correction. Aid and government expenditure are treated as endogenous and the additional covariates as strictly exogenous. The validity of these assumptions was tested by applying the Hansen test (amounting to a test for the exogeneity of the covariates) and the Arellano-Bond test of second-order autocorrelation, which must be absent from the data for the estimator to be consistent. Both tests turned out to be borderline when the first lags of the aid variable and the dependant variable are included as instruments. This may suggest that aid once lagged is endogenous and thus invalid as an instrument. Therefore, the first lags were excluded from the list of instruments. The test stat istics do then clearly not reject the specification at conventional levels of signifi cance. 14 A difference-in-Sargan test was does also performed on the additional instruments in the system GMM. This test also does not reject the specification. The results remain qualitatively unchanged when the system GMM estima tor is employed (column 2 in table 1)Y Results obtained by rerunning the regression using the Arellano and Bond (1991) GMM difference estimator (not reported) corroborate the main results of the system GMM estimation, although the literacy rate no longer significantly affects school enrollment. 13. Baldacci and others (2004), who do not include aid as a regressor, find per capita GDP to be a significant determinant of enrollment only in some of their estimated equations. 14. Additional tests were performed for third- and fourth-order serial correlation, which does not seem to be present in the data. 15. Excluding the first lag of the endogenous variables reduces the number of instruments to 53. The regression was replicated excluding the second and third lags to ensure that the results do not depend on this still substantial number of instruments. This step reduced the number of instruments to 28. The relevant test statistics do clearly not reject the specification, and the results are not affected. In particular, the difference-in-Sargan test (prob > 1') is 0.42. Dreher, Nunnenkamp, and Thiele 301 Two-stage least squares (2SLS) regressions are estimated instrumenting for aid as an alternative procedure for addressing the endogeneity issue. The International Country Risk Guide, the Fraser index of economic freedom, and the mortality rate of children under age five serve as instruments. These variables-the first two proxies for governance, the third a proxy for need have been shown to be related to aid allocations (McKinlay and Little 1977; Hout 2004; Thiele, Nunnenkamp, and Dreher 2007). They are indeed highly correlated with aid for education in the sample and not significantly correlated with school enrollment once the other relevant regressors are controlled for. 16 Government expenditure on education is taken as an exogenous regressor in column 3; both government expenditure and aid are instrumented in column 4. The International Country Risk Guide index for ethnic tensions is used as an instrument for government expenditure on education. While the pre vious literature does not come up with reliable instruments for government spending (Feldmann forthcoming), ethnic tensions have been shown to affect expenditure. 17 The overidentifying restrictions are not rejected at conventional levels of significance, suggesting that the model is well specified, and there is no sign that the instruments are endogenous. The instruments are jointly significant at the 1 percent level in the first-stage regressions, indicating some power. However, the F-test statistic falls short of the rule of thumb threshold of 10 proposed by Staiger and Stock (1997). Nelson and Startz (1990) show that the distribution of the 2SLS estimator and the t-statistics are only poorly approxi mated by the asymptotic representation with weak instruments. This is likely to imply test statistics with nonnormal distributions, making 2SLS misleading (Staiger and Stock 1997). According to Cruz and Moreira (2005), weak instru ments exacerbate the finite sample bias inherent in the 2SLS estimator. Given that the specifications include the lagged dependent variable, 2SLS estimations may also suffer from dynamic panel bias. These estimations, therefore, have to be interpreted with caution. As all relevant test statistics do not reject the GMM specification, the preferred results are those based on the GMM regression (column 2 of table 1). The effect of aid for education on primary school enrollment remains posi tive and significant at least at the 5 percent level when the two instrumental variable techniques are used. The lagged dependent variable is significant at the 1 percent level; per capita GDP and urbanization are insignificant. As before, so is government expenditure on education. The results on adult lit eracy are mixed. The share of the population under 15 has a completely 16. Initially, a measure of democracy was also included in the list of instruments. However, given that democracy might be important for the effect of aid on enrollment (as argued below), only the former three variables are used. The results are not affected by this choice. 17. Von Hagen (200S) argues that ideological and ethnic divisions result in higher government spending because each segment of society tends to neglect the tax burden falling on other segments, exacerbating the common pool problem. 302 THE WORLD BA:- .,.. 0 -; :r: '" 1Iil ,. 0 t'"' 0 TABLE 2. School Enrollment and Aid: Tests for Robustness, System General Method of Moments, 1970-2004 (6) (7) '" ;;. Z (1) (2) (3) (4) (5) School School (8) (9) (10) (11) (12) (13) (14) (IS) :-: Variable School School School enrollment enrollment enrollment Comp)etion School rates enrollment enrollment levels enrollment School School b School 1970-97 enrollment;\ enrollment enrolJment C School enroHment School School School School enrollmem enrollment enrollment enrollment '" (") 0 Z 0 lagged dependent 0.845 0.530 0.65 0.821 0.603 0.516 0.617 0.574 0.588 0.617 0.631 0.626 0.602 0.594 0.597 :i:: (7.36)'" (5.40)" , 18.46)'" (8.30)'" (5.20)'" 14.46)'" 16.42)'" (7.48)'" (5.98)'" 16.42)'" (5.26)'" (6.33)'" (7.05)'" 15.47)'" (6.50)'" (:; Expenditure on -0.00006 0.00007 -0.00009 0.00023 -0.00102 -0.00008 -0.00534 -0.00004 -0.00008 -0.00001 -0.00002 -0.00005 -0.00010 -0.00009 ,. education (per (0.60) (0.69) 11.21) (1.41) (0.16) (0.94) (0.73) (0.57) (0.94) (0.10) (0.24) (0.74) (1.07) (1.16) '" < disbursements 0.0033 (0.89) '"l 1Ii (per capita) Total aid 0.00004 disbursements (0.32) 1per capita) Aid lor education 0.0008 commitments (2.47)' · (per child) Aid for education -0.002 0.002 0.182 0.003 0.004 0.002 0.003 0.002 0.00.1 0.001 0.003 0.009 commitments (0.46) 12.10)" 11.79)' 12.11)" 11.75)' (1.79)' 12.11)" 13.37)'" (2.54)" (0.29) (2.51)" 10.97) (per capita) Literacy rate 0.001 0.003 0.002 0.001 0.002 0.224 0.002 0.002 0.003 0.002 0.002 0.00.1 0.003 0.003 0.003 (0.52) (2.55)" (2.28)" 10.78) 11.72)" 12.77)'" (2.23)"" (2.30)" (2.27)" 12.23)" (2.65)'" (2.30)" (2.19)" (2.23)" 12.43)" Pupil~teacher ratio 4.13E - 04 (0.23) Unit cOSts of -0.002 enrollment (1.63) Aid·squared 0.00003 (0.61) lnternarional Cftsis 0.16 (1.91)" Democracy 0.004 (0.60) Del1l(X;racy'" aid -0.001 (0.78) Number of countries 94 96 96 92 99 96 96 90 87 96 89 95 96 96 96 Number of 199 269 269 194 306 269 269 221 238 269 176 268 269 269 269 observations Maximum number of 3 6 6 6 6 3 6 6 periods Number 01 21 42 42 24 36 42 42 42 42 42 34 43 62 62 44 instruments Hansen test 0.14 0.58 0.35 0.43 0.25 0.32 0.42 0.31 0.4.5 0.42 0.34 0.43 0.39 0.43 0.30 (Frob> chi') 0.06 0.92 0.13 0.94 0.28 0.19 0.41 0.21 0040 0.21 0.24 0.18 0.73 0.16 0.25 test (Prob> Arellano·Bond test 0.12 0.18 0.06 0.61 0.19 0.29 0.82 0.19 0.20 0.19 0.28 0.18 (Prob > z) Note: Numbers in parentheses are t-statistics, estimated robustly. The dependent variable is in logarithms. * Significant at the 10 percent level. "Significant at the 5 percent level. U * Significant at the 1 percent level. aExcludes countries in the highest enrollment quartile. tl bExcludes Botswana, Indonesia, Malawi, Rwanda, Togo, and Uganda (poor countries for which enrollment rates have risen particularly rapidly, according to Clemens 2004). ~ ~ cExcludes Angola, Kuwait, Liberia, Lesotho, Malawi, and Tanzania (reported enrollment rates increased by more than 20 percent in a single year and by 10 percentage points at least once over the period under consideration). t Source: Authors' analysis based on data described in table A-1. ~ ~ ~ ~ ~ ~ ~. "" w o '" 306 THE WORLD BANK ECONOMIC REVIEW The reason to exclude government expenditure on education is that the aid coefficient may be biased downward when government expenditure is included and (part of) aid runs through the budget. The aid coefficient is supposed to capture the expenditure-augmenting effect of aid in addition to effects on enrollment that are attributable to a higher productivity of aid relative to gov ernment expenditure. The fact that the coefficient of aid is hardly affected when comparing column 5 in table 2 with column 2 in table 1 suggests that the expenditure-increasing effect of aid is not relevant, either because aid is not accounted for in the budget or because it is highly fungible. Using enrollment levels instead of logs (column 6) leaves aid significant at the 10 percent level. One additional dollar of per capita aid increases school enrollment by about 0.2 percentage points. The years after 1997 are excluded because the results may be distorted by the revision of educational data since 1998 and by the recent shift from sector-specific aid, including aid for edu cation, toward general budget support and multisector aid. Yet the key result is not affected: aid increases enrollment at the 5 percent level of significance (column 7). The results reported in column 8 are based on a restricted sample that excludes all countries in the highest quartile in enrollment rates. In this way, a check can be run to determine whether the upper bound of the dependent vari able implies a downward bias for aid effectiveness when including recipient countries with enrollment rates already close to 100 percent. The aid coeffi cient turns out to be somewhat larger for the restricted sample compared with column 2 in table 1. The same pattern applies to the fixed-effects estimates (column 8 in supplemental appendix table S-3 and column 1 in table 1). Standard errors are larger for the restricted sample, which is not surprising given the smaller variance of enrollment rates and the smaller number of observations. Next some potentially influential outliers are excluded, though to little effect. Botswana, Indonesia, Malawi, Rwanda, Togo, and Uganda are treated as outliers (column 9). Clemens (2004) identifies these countries as examples of poor countries for which enrollment rates reported by UNESCO have risen particularly rapidly, at least in some cases at the cost of deteriorating education quality (as reflected in high failure and repetition rates in Rwanda and Togo, steeply rising pupil-teacher ratios in Malawi, and lower test scores in Uganda). Alternatively, Angola, Kuwait, Liberia, Lesotho, Malawi, and Tanzania are excluded (column 10), because reported enrollment rates increased by more than 20 percent in a single year and by 10 percentage points at least once over the period under consideration. 21 21. Both relative and absolute changes are considered when defining the cut-off point, because the two deviate widely at the tails of the distribution. A moderate absolute rise in enrollment from 20 to 25, for example, implies a relative increase of 25 percent. Dreher, Nunnenkamp, and Thiele 307 Alternative specifications include the pupil-teacher ratio, the unit cost of education (government expenditure on education divided by the population under 15, as a percent of per capita GDP), and a dummy variable for a crisis of international scale in the recipient country as additional control variables (columns 11, 12, and 14). The pupil-teacher ratio and the unit cost of pro duction are applied as additional supply factors in the social production func tion framework. 22 The dummy variable for international crisis is included because enrollment might be expected to decline in such years and crises might be correlated with aid. With the exception of the crisis dummy variable, none of the additional vari ables is significant at conventional levels. In all cases, the impact of aid remains significant at the 5 percent level. Aid-squared has often been used in the litera ture on aid and growth (see, for example, Dalgaard, Hansen, and Tarp 2004), because aid may suffer from decreasing returns once its optimal degree is exceeded. Thus aid-squared is included in column 15. Although aid and aid-squared are not individually significant, they are jointly significant at the 5 percent level. Finally, the possibility that the impact of aid on school enrollment might depend on democratic governance in the recipient countries is taken into account. This issue relates to the ongoing discussion of whether donors should target aid to better-governed countries. According to Svensson (1999), demo cratic institutions provide an institutionalized check on governments, encoura ging them to use aid more productively. The impact of aid on education outcomes is thus hypothesized to be greater the greater the degree of democ racy. The test of whether aid is more effective under conditions of good govern ance treats governance as exogenous to aid, which is in line with much of the previous literature. 23 The interaction between aid and democracy is supposed to reveal a differential impact of aid. As a proxy for democratic governance, an index is constructed with data provided by Freedom House (2004). Data on this variable are available for a large number of countries and over most of the years under study. While foreign aid does not increase enrollment individually in a significant way, aid and its interaction with democracy are jointly significant at the 10 percent level (column 15). The level of democracy and its interaction with aid are neither individually nor jointly significant at conventional levels. 22. Ideally, one would also want to control for the "price" of schooling. The lack of data does not permit this to be done. 23. The seminal contribution of Burnside and Dollar (2000) triggered the debate on whether aid is more effective under good policy conditions, treating the policy variables (openness, inflation, budget surplus) as exogenous. In a subsequent article, Burnside and Dollar (2004, p. 4) note that "researchers coming from the left, the right, and the center have all concluded that aid as traditionally practiced has not had systematic, beneficial effects on institutions and policies." Clemens, Radelet, and Bhavnani (2004) use instruments for institutional and policy variables; their use does not affect their results significantly. 308 THE WORLD BANK ECONOMIC REVIEW It therefore appears that, in contrast to the hypothesis derived by Svensson (1999), the impact of aid does not depend on democracy. Aid for education may help achieve universal education even in countries characterized by less-advanced democratic institutions. Whether this result is robust to the speci fication of the model and the measurement of democracy is an interesting ques tion for future research.24 V. CONCLUSIONS The effectiveness of sector-specific aid on education is assessed empirically within the framework of social production functions for almost 100 developing economies over the period 1970-2004, with the education Millennium Development Goals, particularly primary school enrollment, considered as outcome variables. The results suggest that higher per capita aid significantly increases primary school enrollmenr and domestic government spending on education does not. This result is robust to the method of estimation, the inclusion of instruments to control for the endogeneity of aid, and the set of control variables included in the estimations. These findings are in sharp contrast to Easterly's (2006) verdict that foreign aid has done "so little good." At the same time, the analysis underscores the need to disaggregate aid in order to assess its effectiveness. Aid specifically devoted to the education sector modestly but not negligibly contributes to achieving universal primary education in developing economies. The preferred (GMM) specification (column 2 in table 1) implies that an additional dollar of per capita aid to the education sector increases school enrollment by about 0.3 percent. Consequently, school enrollment could improve considerably if donors kept their promise to double current aid efforts. Aid that is effective in improving education should also have favorable long term effects on economic growth, which might not be measurable with conven tional econometric methods. 25 But even if the link between education and growth turned out to be weak, the improved education outcome would be important in its own right, because "schooling has a large number of direct beneficial effects beyond raising economic output, such as lower child mor tality" (Pritchett 2001, p. 388). The positive effects of aid on education outcomes in recipient countries not withstanding, the analysis points to some caveats that donors should keep in mind when giving aid. In contrast to some other studies (Gomanee and others 2003; Mosley, Hudson, and Verschoor 2004), this study finds no evidence that aid works by increasing government spending on education. Estimates are 24. In column 16 of supplemental appendix table S-3, results are also reported after omitting the lagged dependent variable from the fixed-effects specification. The main result is unchanged. Under this specification, the literacy rate positively affects enrollment, with a highly significant coefficient. 25. The longer-term growth effects of aid are difficult if not impossible to capture, as Clemens, Radelet, and Bhavnani (2004) argue. For a different view, see Rajan and Subramanian (2005). Dreher, Nunnenkamp, and Thiele 309 hardly affected when accounting for both the productivity-enhancing and the expenditure-augmenting effects of aid (by excluding government spending from the estimations). It remains open to debate whether this is mainly because donors deliberately decided to grant aid outside the budget or because aid was highly fungible in the past. This question may be resolved in future research if donors increasingly shift from project-related aid to general budget support. For the time being, the complete insignificance of government spending on edu cation in the estimations cautions against expecting too much from the "new form of conditionality" proposed by Mosley, Hudson, and Verschoor (2004). It remains to be seen whether the ability to use aid for education as a means of strengthening the poverty orientation of government spending on education has improved with the advent of Poverty Reduction Strategy Papers. Lacking conclusive evidence that conditionality works donors should be selective when determining the allocation of aid for education. The most obvious criterion for selectivity is the need for aid in education, as reflected in particularly low enrollment and completion rates. The targeting of aid for edu cation according to need leaves much to be desired, as Thiele, Nunnenkamp, and Dreher (2007) show. Another selectivity criterion stressed by many donors-the quality of govern ance in recipient countries-might be less important than widely believed. Investigation of this issue is left to future research. Finally, it would be desirable to assess whether the effectiveness of aid for education could be enhanced by shifting donor resources within the sector toward basic education. Basic education accounted for about one-third of total aid for education by donors in 2002-04 (Thiele, Nunnenkamp, and Dreher 2007). This low level of spending is not only in conflict with Millennium Development Goal2, which would require greater concentration on basic edu cation, but also with findings that social returns to primary education tend to be particularly high in low-income countries (World Bank 1995). Future research may be able to address this issue by disaggregating aid and education data more finely once longer time-series are available for a larger number of countries. ApPENDIX TAB LEA -1. Definitions and Sources Variable Description Source School enrollment Ratio of number of children of official World Bank (2005) school age enrolled in school to number of children of official school age (net enrollment ratio) (Continued) 310 THE WORLD BANK ECONOMIC REVIEW TABLE A-1. Continued Variable Description Source Primary completion rate Number of students successfully World Bank (2005) completing the last year of (or graduating from) primary school in a given year divided by the number of children of official graduation age in the population Expenditure on education Public spending on public education plus World Bank (2003, subsidies to private education at the 2005) primary, secondary, and tertiary levels. Variable measured both per capita and as percent of GDP Aid for education Aid commitments by all donors according OECD (2006) (commitments) to CRS Purpose Code 110. Includes aid for basic education, secondary education, postsecondary education, and unspecified levels of education. CRS guidelines require sector-specific program assistance and budget support in the form of sector-wide approaches to be subsumed under education when meant to benefit this sector. Variable measured both per capita, per child below age 15, and as percent of recipient country's GDP Aid for education Aid disbursements by all donors OECD (2006) (disbursements) according to CRS Purpose Code 110, Form 2. Coverage is same as for commitments. Variable measured per capita Total disbursements Total aid (in all sectors) disbursed by all OECD (2006) donors. Variable measured per capita Literacy rate Percentage of people 15 and older who World Bank (2005) can, with understanding, read and write a short simple statement on their everyday life GDP per capita Per capita GDP in purchasing power World Bank (2005) parity terms (2000 international dollars) Population under 15 Percentage of total population under 15 World Bank (2005) Urbanization Share of total population living in areas World Bank (2005) defined as urban in each country Democracy index [8 (political rights index + civil liberties Freedom House (2004) index)] 12 Pupil-teacher ratio Number of pupils enrolled in primary World Bank (2005) school divided by number of primary school teachers (regardless of their teaching assignment) (Continued) Dreher, Nunnenkamp, and Thiele 311 TABLE A-I. Continued Variable Description Source Unit costs of enrollment Government expenditure on education World Bank (2005) divided by population under 15, as percent of per capita GDP International crisis Dummy variable that takes value of 1 Wilkenfeld and Brecher when a country is involved in an (2006) international crisis Government stability Assesses government'S ability to carry out ICRG (2005) its declared programs and its ability to remain in office. Risk rating assigned is sum of three subcomponents (government unity, legislative strength, popular support), each with a maximum score of four points and a minimum score of 0 points. A score of 4 points indicates very low risk and a score of 0 points very high risk Under-five mortality Probability that newborn baby will die World Bank (2005) before reaching age five if subject to current age-specific mortality rates. Variable expressed as rate per 1,000 Economic Freedom Composite 0-10 index of economic Gwartney and Lawson freedom; higher values reflect greater (2004) freedom Source: Authors' description based on cited data sources. T\BLE A-2. Summary Statistics (Estimation Sample, Table 1, Column 2) Standard Variable Mean Minimum Maximum deviation School enrollment (logarithm) 4.36 3.10 4.68 0.31 Primary completion rate (logarithm) 4.26 2.77 4.72 0,46 Expenditure on education (per capita) 95.78 3.56 697.50 111.94 Expenditure on education (percent of GDP) 4.28 0.83 11.58 1.88 Aid for education, commitments (per capita) 3.67 0.00 64.33 6.51 Aid for education, commitments (percent of 0,43 0.00 3.23 0.57 GDP) Aid for education, commitments per child 10.56 0.00 292.84 22.38 Aid for education, disbursements per capita 0.86 0.00 21.90 1.90 Total aid disbursements per capita 33.50 -2.34 304.07 37.50 Literacy rate 73.79 10.30 99.80 22.75 GDP per capita 4783 494 16050 3497 Population under 15 36.89 15.02 51.09 8.62 Urbanization 48.16 5.14 92.63 21.55 Democracy, index 4.04 1.00 7.00 1.76 Pupil-teacher ratio 31.53 10.28 77.03 13.20 International crisis 0.07 0.00 0.80 0.16 Unit costs of enrollment 13 1.76 38 7 Government stability 7.84 2.28 11.08 1.78 (Continued) 312 THE WORLD BANK ECONOMIC REVIEW TABLE A-2. Continued Standard Variable Mean Minimum Maximum deviation Under-five mortality 75.14 7.50 320.00 65.48 Economic freedom 5.76 2.90 7.50 0.98 Source: Authors' analysis based on data described in table A-I. REFERENCES Arellano, Manuel, and Stephen Bond. 1991. "Some Tests for Specification for Panel Data: Monte Carlo Evidence and an Application to Employment Equations." Review of Economic Studies 58(2): 277-97. Arellano, Manuel, and Olympia Bover. 1995. "Another Look at the Instrumental Variable Estimation of Error-components Models." Journal of Econometrics 68(1):29-51. Baldacci, Emanuele, Maria Teresa Guin-Siu, and Luiz R. de Mello. 2003. "More on the Effectiveness of Public Spending on Health Care and Education: A Covariance Structure Model." Journal of International Development 15(6):709-25. Baldacci, Emanuele, Benedict J. Clements, Sanjeev Gupta, and Qiang Cui. 2004. Aid, Social Spending, Human Capital, and Growth in Developing Countries: Implications for Achieving the MDGs. IMF Working Paper WP/04/217. Washington, D.C: International Monetary Fund. Behrman, Jere, Piyali Sengupta, and Petra Todd. 2005. "Progressing through PROGRESA: An Impact Assessment of a School Subsidy Experiment in Rural Mexico." Economic Development and Cultural Change 54( 1 ):237 - 75. Bennell, Paul. 2002. "Hitting the Target: Doubling Primary School Enrollments in Sub-Saharan Africa by 2015." World Development 30(7):1179-94. Blundell, Richard, and Stephen Bond. 1998. "Initial Conditions and Moment Restrictions in Dynamic Panel Data Models." Journal of Econometrics 87(1):115-43. Borenszstein, Eduardo, R., Jose. de Gregorio, and Jong-Wha Lee. 1998. "How Does Foreign Direct Investment Affect Economic Growth?" Journal of International Economics 45(1):115-35. Bruns, Barbara, Alain Mingat, and Ramahatra Rakotomalala. 2003. Achieving Universal Primary Education by 2015: A Chance for Every Child. Policy Research Working Paper. World Bank, Washington, D.C. Burnside, Craig, and David Dollar. 2000. "Aid, Policies, and Growth." American Economic Review 90(4):847-68. 2004. Aid, Policies, and Growth: Revisiting the Evidence. Policy Research Working Paper 3251. World Bank, Washington, D.C. Clemens, Michael A. 2004. The Long Walk to School: International Education Goals in Historical Perspective. Center for Global Development Working Paper 37. Washington, D.C. Clemens, Michael A., Steven Radelet, and Rikhil Bhavnani. 2004. Counting Chickens When They Hatch: The Short-Term Effects of Aid on Growth. Center for Global Development Working Paper 44, Washington, D.C. Cruz, Luiz M., and Marcelo J. Moreira. 2005. "On the Validity of Econometric Techniques with Weak Instruments: Inference on Returns to Education Using Compulsory School Attendance Laws." Journal of Human Resources 40(2):393-410. Dalgaard, Carl-Johan, Henrik Hansen, and Finn Tarp. 2004. "On the Empirics of Foreign Aid and Growth." Economic Journal 114(June):F191-216. Dreher, Nunnenkamp, and Thiele 313 Doucouliagos, Hristos, and Martin Paldam. 2005. The Aid Effectiveness Literature: The Sad Results of 40 Years of Research. Department of Economics Working Paper 2005.05. Denmark: University of Aarhus. Easterly, William. 2006. The White Man's Burden: Why the West's Efforts to Aid the Rest Have Done So Much III and So Little Good. New York: Penguin Press. Feldmann, Horst, "Government Size and Unemployment in Developing Countries." Forthcoming Applied Economics Letters. Fielding, David, Mark McGillivray, and Sebastian Torres. 2005. "Synergies between Health, Wealth, Education, Fertility and Aid: Implications for Achieving the Millennium Development Goals." World Institute for Development Economics Research, Helsinki. (www.wider.unu.eduJresearchJ 2004-2005/2004- 2005 -1/mdg-paperslfielding-mcgillivray-torres-160805.pdf). Filmer, Deon, and Lant Pritchett. 1999. "The Effect of Household Wealth on Educational Attainment: Evidence from 35 Countries." Population and Development Review 25(1):85-120. Freedom House. 2004. Annual Report. (www.freedomhouse.org). Gomanee, Karuna, Oliver Morrissey, Paul Mosley, and Arjan Verschoor. 2003. "Aid, Pro-Poor Government Expenditure and Welfare." CREDIT Research Paper 03/03. University of Nottingham, United Kingdom: Centre for Research in Economic Development and International Trade. 2005. "Aid, Government Expenditure, and Aggregate Welfare." World Development 33(3):355-70. Gupta, Sanjeev, Marijn Verhoeven, and Erwin Tiongson. 1999. Does Higher Government Spending Buy Better Results in Education and Health Care? IMF Working Paper WP/99/21. Washington, D.C.: International Monetary Fund. Gwartney, James D., and Robert A. Lawson. 2004. Economic Freedom of the World: 2004 Annual Report. Vancouver: Fraser Institute. (www.freetheworld.com). Hanushek, Erik A. 1995. "Interpreting Recent Research on Schooling in Developing Countries." World Bank Research Observer 10(2):227-46. Hout, Wi!. 2004. "Political Regimes and Development Assistance: The Political Economy of Aid Selectivity." Critical Asian Studies 36(4):591-613. ICRG !International Country Risk Guide). 2005. International Country Risk Guide Political Risk Indicator. (www.countryrisk.comlguide/archiveslcaccountry_risk_ratings.html). Isenman, Paul, and Dag Ehrenpreis. 2003. "Results of the OECD DAClDevelopment Centre Experts' Seminar on 'Aid Effectiveness and Selectivity: Integrating Multiple Objectives into Aid Allocations.'" DAC Journal 4(3):7-25. McGillivray, Mark, Simon Feeny, Niels Hermes, and Robert Lensink. 2005. "It Works; It Doesn't; It Can, But That Depends. 50 Years of Controversy over the Macroeconomic Impact of Development Aid." Research Paper 2005/54. Helsinki: United Nations University, World Institute for Development Economics Research. McKinlay, Robert D., and Richard Little. 1977. "A Foreign Policy Model of U.S. Bilateral Aid Allocation." World Politics 30( 1):58-86. Michaelowa, Katharina, and Anke Weber. 2006. "Aid Effectiveness Reconsidered: Panel Data Evidence for the Education Sector." HWWA Discussion Paper 264. Germany: Hamburg Institute of International Economics. Mingat, Alain, and Jee-peng Tan. 1998. The Mechanics of Progress in Education: Evidence from Cross-Country Data. Policy Research Working Paper 2015, World Bank, Washington, D.C. ---.2003. "On the Mechanics of Progress in Primary Education." Economics of Education Review 22(5):455-67. Mosley, Paul, John Hudson, and Arjan Verschoor. 2004. "Aid, Poverty Reduction, and the 'New Conditionality.'" Economic Journal 114(496):F217-43. Nelson, Charles R., and Richard Startz. 1990. "Some Further Results on the Exact Small Sample Properties of the Instrumental Variable Estimator." Econometrica 58(4):967-76. 314 THE WORLD BANK ECONOMIC REVIEW OECD (Organisation for Economic Co-operation and Development). 2006. International Development Statistics (IDS) Online: Databases on Aid and Other Resource Flows. (www.oecd.orgldataoecd/50/ 1715037721.htm). Pettersson, Jan. 2006. KForeign Sectoral Aid Fungibility, Growth and Poverty Reduction." Paper pre sented at the World Institute for Development Economics Research Conference "Aid, Principles, Policies and Performance," June 16-17, Helsinki. Pritchett, Lant. 2001. "Where Has All the Education Gone?" World Bank Economic Review 15(3):367-91. Rajan, Raghuram G., and Arvind Subramanian. 2005. Aid and Growth: What Does the Evidence Really Show? IMF Working Paper WP/051127. Washington, D.C: International Monetary Fund. Reinikka, Ritva, and Jacob Svensson. 2001. "Explaining Leakage of Public Funds." WIDER Discussion Paper 20011147, Helsinki: United Nations University, World Institute for Development Economics Research. Roberts, John. 2003. Poverty Reduction Outcomes in Education and Health: Public Expenditure and Aid. ODI Working Paper 210. London: Overseas Development Institute. Roodman, David. 2005. "Xtabond2: Stata Module to Extend Xtabond Dynamic Panel Data Estimator." Washington, D.C: (',enter for Global Development.(http://econpapers.repec.orglsoft warel bocbocode/s435901.htm). Schultz, T. Paul. 1988. "Expansion of Public School Expenditures and Enrollments: Intercountry Evidence on the Effects of Income, Prices, and Population Growth." Economics of Education Review 7(2):167-83. Sen, Amartya. 2006. "The Man without a Plan." Foreign Affairs (March/April). (www.nyu.edulfas/ institute/drilEasterlylFilelForeign %20Affairs %20- %20The %20Man %20Without%20a %20Plan % 20- %20Amartya %20Sen.pdf). Staiger, Douglas, and James H. Stock. 1997. "Instrumental Variables Regression with Weak Instruments." Econometrica 65(3):557-86. Svensson, Jacob. 1999. "Aid, Growth and Democracy." Economics & Politics 11(3):275-97. Tarp, Finn. 2006. "Scaling Up Aid: Challenges and Opportunities: A Few Comments." Paper presented at the World Institute for Development Economics Research Conference "Aid: Principles, Policies and Performance," June 16-17, Helsinki. Thiele, Rainer, Peter Nunnenkamp, and Axel Dreher. 2007. "Do Donors Aid in Line with the Millennium Development Goals? A Sector Perspective of Aid Allocation." Review of World Economics 143(4):596-630. Von Hagen, Juergen. 2005. "Political Economy of Fiscal Institutions." GESY Discussion Paper 149. Germany: Mannheim University. Wilkenfeld, Jonathan, and Michael Brecher. 2006. "International Crisis Behavior Project." Center for International Development and Crisis Management, College Park, MD, USA. (www.cidcm.umd. edulicb). Windmeijer, Frank. 2005. "A Finite Sample Correction for the Variance of Linear Efficient Two-Step GMM Estimators." Journal of Econometrics 126(1):25-51. Wolf, Susanna. 2007. "Does Aid Improve Public Service Delivery?" Review of World Economics 143(4 ):650-672. World Bank. 1995. Development in Practice: Priorities and Strategies for Education. Washington D.C - - . 2003. World Development Indicators CD-ROM. 2005, and 2007 Washington, D.C World Bank Lending and Financial Sector Development Robert Cull and Laurie Effron A new database of World Bank loans to support financial sector development is used to investigate whether countries that received such loans experienced more rapid growth on standard indicators of financial development than countries that did not. Self-selection is accounted for with treatment-effects regressions. The results indicate that borrowing countries had significantly more rapid growth in M2/GDP than non borrowers and swifter reductions in interest rate spreads and cash holdings (as a share of M2). Borrowers also had higher private credit growth rates than nonborrowers in some treatment-effects regressions but not in standard panel regressions with fixed country effects. On the whole, the results indicate some significant advantages in financial development for borrowers over nonborrowers. JEL codes: F33, G21, 016. The World Bank has been making loans to governments of member countries since 1946. Over time, World Bank lending shifted from supporting post World War II reconstruction to supporting economic growth and poverty alle viation. During its first four decades, it concentrated on financing investments in infrastructure and directly productive activities in agriculture and industry. This approach was driven by the assumption that the scarcity of foreign exchange for capital investments was the main constraint hindering economic growth in developing countries. With the shift in analytic focus on constraints to growth in the 1980s (initially to the economic policies of developing countries and later to their institutional capacities), the World Bank introduced fast-disbursing loans for balance of payments support conditioned on changes in policies and institutions. Robert Cull is a senior economist in the Development Economics Research Group at the World Bank; his email address is rcull@worldbank.org. At the time of her retirement, Laurie Effron was a lead evaluation officer in the Independent Evaluation Group of the World Bank; her email address is leffron@gmail.com. The authors thank Thorsten Beck, Gerard Caprio, Pedro Carneiro, Stijn Claessens, George Clarke, Charles Goodhart, Greg Ingram, James Hanson, Patrick Honohan, Aart Kraay, Daniel Lederman, David McKenzie, Caglar Ozden, Kyle Peters, Claudio Raddatz, Sergio Schmukler, Colin Xu, and three anonymous referees for helpful comments and Maria del Pilar Casal, Sarojini Hirshleifer, and Gulmira Karaguishiyeva for excellent research assistance. A supplemental appendix to this article is available at http://wber.oxfordjournals.orgl. pp. 315-343 THE WORlD BA.r-'K ECONOMIC REVIEW, VOl. 22, No.2, doi:10.1093/wber/lhn004 Advance Access Publication May 15, 2008 The Author 2008. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development I the world bank. All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjoumals.org 315 316 THE WORLD BANK ECONOMIC REVIEW Consistent with its sharper focus on the importance of good macroeconomic policies and institutions for economic growth, the World Bank became an early proponent of supporting appropriate policies and capable institutions in the financial sector, which, it asserted, could also contribute to economic growth (World Bank 1989). These positions were later supported by extensive research establishing a causal link between financial development and economic growth (Levine and Zervos 1998; Rajan and Zingales 1998; Beck, Levine, and Loayza 2000; Levine, Loayza, and Beck 2000; Levine 2005) and by research showing that less government control over financial systems and institutions leads, over time, to deeper, stabler, and more efficient systems (Caprio, Honohan, and Stiglitz 2001; Barth, Caprio, and Levine 2001a,b; La Porta, Lopez-de-Silanes, and Shliefer 2002; Beck, Demirguc-Kunt, and Levine 2003). In the early 1990s, most developing countries maintained policies and institutions that were considered detrimental to financial sector development namely, government-administered financial systems, fixed interest rates, directed credit, and market dominance of publicly owned financial institutions. The World Bank began to target lending support to financial sector reforms addressing these constraints. Between 1992 and 2003, about one quarter of World Bank lending-some $56 billion-included support for financial sector reforms aimed at reducing direct government control over credit allocation, interest rates, and financial institutions and increasing government oversight of domestic financial markets and institutions by strengthening banking super vision and prudential regulations. The basic objective of such support was to establish a strong enabling environment in which well-governed financial insti tutions would mobilize resources, allocate credit, and manage risks efficiently. Most reforms focused on the banking sector and within the banking sector on the restructuring or privatization of state-owned banks. Reforms also sought to strengthen banking legislation, regulation, and supervision. An independent evalu ation found that government ownership of banks in countries with Bank loans that included conditionality on bank privatization decreased substantially (with the per centage of banking assets owned by governments dropping 60 percentage points within a decade) and by more than that in countries that did not borrow for bank privatization (where the decrease was 35 percentage points) (lEG 2006). The results were more ambiguous for differences in banking legislation and regulation; infor mation on changes in banking supervision was not available for many countries. Did these changes increase the mobilization of resources, allocate more credit, and make the financial sector more efficient? This article addresses these questions. Using quantitative indicators to measure changes in depth, efficiency, and credit to the private sector, it examines whether Bank assistance between 1992 and 2003 helped develop financial sectors in client countries and tests whether progress was greater in countries receiving loans for this purpose than in client countries that did not receive such loans. The article contributes to a broader literature on the effects of economic reform programs in developing countries. Much of that literature focuses on the impact Cull and Effron 317 of adjustment lending by the World Bank and the International Monetary Fund (IMF) on broad macroeconomic aggregates, notably real per capita growth. 1 This article examines the effects of World Bank lending on measures of financial sector development, which has been linked to economic growth. Improvements in indicators of financial development were generally signifi cant for borrowers-and more pronounced for borrowers than non borrowers. Treatment-effects regressions are used that explicitly account for nonrandom selection (the possibility that borrowers tended to be countries that were likely to have improved their financial sectors without the loans). Additional robust ness checks test whether the findings are specific to particular regions and whether improvements in financial indicators preceded or followed World Bank loans. (If improvement preceded the loans, it would seem unlikely that the loans had a large causal impact in borrowing countries.) A final set of checks incorporates additional controls for countries' readiness for and experi ence with financial reform. The results of these tests reinforce the main find ings: borrowing countries tended to experience substantial improvement in their financial indicators, significantly more than the typical improvement in non borrowing countries, even after accounting for selection. The remainder of the article is organized as follows. Section I describes the data, including the indicators used to assess outcomes and the variables that summarize World Bank lending in support of financial sector reform from 1992 to 2003. Sections II and III describe the basic regression models and sum marize the base results, and section IV presents results using estimation tech niques that address selection problems. Section V runs additional robustness checks, including regional regressions and models that attempt to control for countries' history of and readiness for reform. The last section briefly summar izes the results of the various methods. 1. DATA The analysis relies on standard indicators, such as M2/GDP and private creditl GDP, that have been shown to be robustly associated with long-run economic growth (Beck, Levine, and Loayza 2000; Levine, Loayza, and Beck 2000; Levine 2005).2 This analysis is restricted to banking indicators, because banks hold the vast majority of financial sector assets in developing countries. 3 1. Easterly (2005) describes this literature as including Barro and Lee (2002), Conway (1994), Corbo and Goldstein (1987), Corbo and Fischer (1995), Devarajan, Dollar, and Holmgren (2001), Dicks-Mireaux, Mecagni, and Schadler (2000), Goldstein and Montiel (1986), Haque and Khan (1998), Hutchinson (2001), Kapur, Lewis, and Webb (1997), Khan (1990), Killick, Gunatilaka, and Marr (1998), Knight and Santaella (1997), Summers and Pritchett (1993), Przeworski and Vreeland (2000), Svensson (2003), and Van de Walle (2001). 2. For descriptions of standard indicators of financial development and their use, see Beck, Demirgul;-Kunt, and Levine (2000). 3. The ratio of private credit to GDP can, however, include lending by nonbank financial institutions. 318 THE WORLD BANK ECONOMIC REVIEW An advantage of these indicators is that they are available for many countries throughout the decade of this analysis. They do create some pro blems, however. For example, as part of the restructuring or privatization of problem banks, the value of nonperforming assets may be reduced or the loans eliminated from bank balance sheets, thereby reducing the private credit ratio. Successful restructuring efforts contribute to a healthier banking sector. Because such efforts reduce private credit, however, they would be viewed as detrimental to financial development in the models presented here. Moreover, the private credit ratio does not provide information about which segments of society receive credit or about the quality of the loans made, because data on nonperforming loans are not available on a standardized basis across countries. The M2/GOP ratio provides information on deposit levels, but that information is not broken down by the income level of the depositors. Increases in the ratio may not mean that all segments of society are availing themselves of formal banking services. Additional indicators-namely, the spread between the lending and deposit interest rates and the ratio of cash held outside of banks to M2 (a measure of the lack of confidence in the formal banking sector)-are incorporated to round out the assessment of financial development. These indicators were chosen largely because of data availability. They, too, have limitations. For example, interest rates were controlled in a number of developing countries at some point during the sample period; spreads are unlikely to be an accurate measure of efficiency in these instances. Measures of capital adequacy, portfolio quality, and profitabil ity are not available in a standard format across countries. 4 These caveats notwithstanding, it seems likely that taken together, M2/GOP, private credit/GOP, cashlM2, and interest rate spreads provide a reasonably com plete picture of both short- and long-term banking development between 1992 and 2003. 5 In the short term, movements in the ratio of cashlM2 can show depositors' reactions to policy changes. The private credit ratio, while subject to short-term perturbations, tends to capture long-term financial development. M2/ GOP and interest rate spreads are arguably somewhere between the two extremes. 4. In the robustness checks presented in the supplemental appendix, an index of financial sector efficiency and freedom developed by the Heritage Foundation is used as the dependent variable in the regressions. The results are qualitatively similar to those for the four quantitative indicators used here. 5. These measures could render an imprecise picture, because the opaqueness of the financial sector makes the effects of reform difficult to observe. As robustness checks, the base models were also rerun using credit ratings-which presumably offer a more nuanced assessment of the success of reforms from three agencies (Moody's, Standard and Poor's, Fitch IBeA) as dependent variables. The unit of observation is the average rating of all sovereign debt issues for each country in each year. Though the coefficient on that variable was positive for borrowers and negative for non borrowers, neither was significant. This is likely because there are so few observations to work with since only a subset of developing countries issued enough sovereign debt on a regular basis. This illustrates the key advantage of the quantitative indicators used in this paper, which is their yearly availability for a wide set of developing countries. Cull and Effron 319 Standard financial outcome variables are combined with data on World Bank loans from a review by the World Bank's Independent Evaluation Group (lEG 2006) (formerly the Operations Evaluation Department). lEG examined 556 lending operations between 1992 and 2003 that involved support for reform of the financial sector,6 including adjustment loans, technical assistance, and lines of credit? The analysis focuses on adjustment loans because they were larger (in dollar terms) and included conditions aimed at spurring broad financial sector development. In contrast, lines of credit often focused on channeling funds through specific financial institutions; technical assistance loans tended to be much smaller than adjustment loans and focused on narrower issues. 8 Sixty-eight countries received at least one adjustment loan with conditions tied to financial reform (table A-I). A control group of 38 countries received no such loans during the period under review. All developing countries for which reasonably complete financial indicator data were available were included in the analysis. Comparisons between the two groups form the basis of most of the statistical analysis that follows. 9 II. METHODOLOGY The base results rely on simple fixed-effects panel regressions of the form where i represents the country and t represents time, measured in years since 1991. The time variable takes on values of 1 to 12. The estimated coefficient /3wb thus represents the average growth rate of the indicator of financial devel opment (Y) for countries that received adjustment loans to support financial reform (table 1 provides descriptions and summary statistics for all the 6. The database of World Bank loans starts in 1992; country and financial sector data start in 1991. The data set was gathered as part of an lEG evaluation of Bank lending to the financial sector. lEG began its analysis in 1992 in an effort to avoid duplication of a previous study (which covered 1984-98) while still covering a sufficiently long period. This type of censoring could conceivably affect the results. 7. Lines of credit are funds passed through an intermediary for demand-driven purposes. The end-user has to repay the loan, usually with interest. 8. Regressions were run with various samples of loans. For models based solely on investment loans, technical assistance loans, or lines of credit, there were no robust significant differences between borrowers and non borrowers. Including smaller subcategories, such as loans for technical assistance, along with the adjustment loans, did not change the qualitative differences between borrowers and non borrowers for the adjustment loan-only sample. Variables based on small subcategories of financial sector loans (such as those devoted to pensions) did not produce stable significant differences between borrowers and non borrowers. 9. The World Bank maintained a policy dialogue throughout the period with some countries in the control group. Nine of the 38 countries in the control group borrowed for financial sector reform before the period of study, although in almost all cases the borrowing consisted of a single loan, often granted as part of a multi sector operation in which financial reforms were not central. These factors could make it harder to find statistically significant differences between the rwo groups in the regressions that follow. 320 THE WORLD BANK ECONOMIC REVIEW TABLE 1. Variable Descriptions and Summary Statistics Variable Description Mean Median High Low Growth rate Equal to the year minus 1991. Estimated 6.68 7 12 coefficient measures annual growth rate for dependent variable in question Policy variables Adjustment loans Cumulative number of adjustment loans 0.70 0 6 0 by country in question at time t. In some models, total includes loans for technical assistance Bank privatization Number of adjustment loans with 0.24 0 4 0 emphasis on bank privatization Regulation and Number of adjustment loans with 0.39 0 4 0 supervIsion emphasis on bank regulation and supervision Other banking Number of adjustment loans with 0.48 0 4 0 emphasis on banking reform other than privatization, regulation, or supervision Auditing and Number of adjustment loans with 0.06 0 2 0 accounting reform emphasis on accounting and auditing reform Capital market Number of adjustment loans with 0.19 0 3 0 development emphasis on capital and securities market development General financial Number of adjustment loans with 0.19 0 3 0 sector reform emphasis on general financial sector development not covered under other variables Rural finance Number of adjustment loans with 0.03 0 2 0 emphasis on rural financial sector development Microfinance Number of adjustment loans with 0.01 0 1 0 emphasis on development of micro finance Nonbank financial Number of adjustment loans with 0.04 0 2 0 sector institutions emphasis on development of nonbank financial institutions Dependent variables Private credit/GDP Claims on private sector (International 25.1 17.7 158.5 0 Financial Statistics [IFS] line 22d) divided by GDP (IFS line 99b) multiplied by 100 M2/GDP Money (IFS line 34) plus quasi-money 33.2 26.9 148.2 0.002 (IFS line 35) divided by GDP (IFS line 99b) multiplied by 100 CashlM2 Currency outside deposit money banks 23.1 19.4 82.5 0 (IFS line 14a) divided by M2 (IFS line 34 + line 35) multiplied by 100 Interest rate spread Lending rate (IFS line 601) minus deposit 11.4 8.6 163.5 -6.9 rate (IFS line 60p), multiplied by 100 (Continued) Cull and Effron 321 TABLE 1. Continued Variable Description Mean Median High Low Concentration Percentage share of total banking sector 62.0 59.5 100.0 14.9 assets held by three largest banks (based on asset information in Bankscope) Macro/institutional controls CPIA score Proxy for institutional development 3.23 3.28 5.35 1.0 Surplus (deficit)! Overall budget balance, including grants, -2.72 -2.11 10.26 -31.63 GDP multiplied by 100 (World Bank) Annual GDP growth Annual GDP growth (World Bank, 3.00 3.94 106.3 -50.2 I percent) various years) Inflation (percent) GDP deflator (World Bank, various years) 78.1 9.41 6041.6 -25.2 Selection equation variables Government checks Variable equals one if there is no chief 2.65 2.50 10.1 1.0 executive. It rises by one under each of the following circumstances: there is a chief executive, the chief executive is competitively elected, and the opposition controls the legislature. In presidential systems, it rises by one for each chamber of the legislature, unless the president's party has a majority in the lower house and a closed-list system is in effect (implying stronger presidential control of the party and therefore of the legislature). It also rises by one for each party coded as allied with the president's party that has an ideological (left-right-center) orientation closer to that of the main opposition party than to that of the president's party. In parliamentary systems this variable rises by one for every party in the government coalition as long as the parties are needed to maintain a majority and for every party in the government coalition that has a position on economic issues (right-left-center) that is closer to that of the largest opposition party than to that of the party of the executive. In parliamentary systems, the prime minister's party is not counted if there is a closed rule in place (in this case the prime minister is presumed to fully control the party). The highest possible score is 18. Average checks 1991-2000 are calculated for each country. (World Bank Database on Political Institutions; see Beck and others 2001.) Debt (as percent of Average external debt 1970-89 (World 56.2 48.1 222.2 4.0 GNI) Bank, various years) (Continued) 322 THE WORLD BANK ECONOMIC REVIEW TABLE 1. Continued Variable Description Mean Median High Low IMF credit Average IMF credit 1970-89 (World 462.0 93.5 9,370.0 0 (millions of Bank, various years) constant dollars) Total debt service Average debt service 1990-99 (World 5.3 4.2 0.3 20.3 (percent of GNI) Bank, various years) Capital formation Gross fixed capital formation 1990-99 22.3 21.0 6.9 64.8 (percent of GDP) (World Bank, various years) Note: Figures are calculated over all observations for which at least one dependent variable was available. variables). A test of whether /3wb = /3nonwb indicates whether adjustment loans had a beneficial impact on financial sector development. To the extent that the growth rates for the control countries were the same as (or greater than) those of countries that received World Bank assistance, the value of that assistance could be questioned. All regressions also include aj, a country-specific fixed effect. Results should be interpreted as changes relative to the country-specific mean for the indicator in question. (More direct methods for addressing potential selection problems are presented later.) First loan measures the number of years since a country received its first loan with financial sector conditions. It is included because improvements in financial indicators were more likely to have materialized in countries that received loans early in the period. lO Including the first financial sector loan variable offers a more precise test of whether improvements in financial indicators occurred after the receipt of loans. For example, if /3wb is positive and significant but the coefficient for the time since first loan variable is insignificant, it would suggest that as a group borrowing countries were more likely to improve their financial indicators regardless of when they received loans from the World Bank. By contrast, if the first loan variable is significant while the simple borrower dummy variable is not, it would suggest that improvement in indicators occurred after the receipt of World Bank loans. 10. For the base regressions, the fairest tests of whether World Bank lending contributed to financial development should include country-specific fixed effects; specification tests confirm that they should be included. Therefore, the dummy variable received Bank adjustment loans is set equal to one throughout the period, regardless of whether the country received its first loan in the first year or the twelfth year. The dummy for no Bank adjustment loans is set equal to one throughout the period for nonborrowers. Had the variables not been coded in this way, all countries that received no loans would have been lost from the observation set because of the country fixed effects; only countries whose borrowing status changed during the period would be used to examine the effects of Bank lending. Such models would have offered comparisons for borrowing countries before and after receiving a loan, but they would not have facilitated comparisons between borrowers and nonborrowers, the focus of this article. Cull and Effron 323 Some specifications include adj, the cumulative number of adjustment loans. Some countries received as many as six adjustment loans with financial sector reform components between 1992 and 2003. Repeated structural adjustment lending from the World Bank or the IMF failed to produce improvement on multiple macroeconomic outcomes, including growth (Easterly 2005).11 Models with adj therefore test whether similar results hold for the financial sector. As with the macroeconomic and insti tutional variables, all policy reform variables are lagged one year in the regresSIOns. Ref is a vector of variables summarizing reform areas covered under adjustment loans (bank privatization; bank regulation and supervision; banking reform not focused on privatization, regulation, or supervision; auditing and accounting reform; capital market development; reform of nonbank financial institutions; general financial sector reform; rural finance; and microfinance).12 Because the data set is a country-level panel of finan cial sector outcomes, the project-level data must be aggregated into country-year reform packages. The cumulative number of loans that had conditions in the policy areas in question are explanatory variables in the regressions that follow.13 X is a vector of macroeconomic and institutional controls, including inflation, real growth, and M2/GDP.14 All of the macroeconomic and insti tutional controls are lagged one year in the panel regressions that follow to mitigate problems arising from the simultaneous determination of the controls and the dependent variables. Inflation should slow financial development if it makes loan contracting over extended periods more difficult. Real growth will accelerate financial development, because it is likely to stimulate demand for financial services. Because World Bank lending to all sectors could spur growth and growth could spur financial development, this is an important control for isolating the effect of financial sector loans on indicators of financial development. 11. Easterly notes that the repeated extension of loans to a country is itself a sign that lending was not effective, "One might expect that it would take more than one loan to accomplish 'adjustment,' but it is hard to see why it would take such a large number" (2005, p. 6). 12. The intention was to specify the policy areas that had the greatest chance of imptoving financial indicators, which are largely bank based. Adjustment loans devoted solely to small- and medium-size enterprise finance or pensions were therefore excluded from the database (very few loans focused only on these areas). 13. Similar qualitative results hold when the number of loans in a given year covering that policy area, simple dummies indicating that a policy area was covered, or dummies indicating that the policy area was covered at some point during the sample period are used. 14. Government budget deficits were included in initial specifications, but they tended to be insignificant. Since inclusion of that variable reduced the sample size by almost half, it was eliminated from the final specifications. Its inclusion does not greatly alter the comparison between borrowers and non borrowers. 324 THE WORLD BANK ECONOMIC REVIEW The base model includes M2/GDP as a general measure of the level of finan cial development. It is not clear a priori whether the level of financial develop ment should have a positive or negative effect on subsequent financial development. On the one hand, a high level of M2/GDP could signal a high level of future development. In that case, lagged M2/GDP can be viewed as a proxy for a country's willingness and ability to pursue financial sector reform. On the other hand, a low level of M2/GDP could signal greater potential for improvement. The Country Policy and Institutional Assessment (CPIA) index is included as a broad measure of institutional development. The World Bank conducts this assessment annually to assess the quality of a country's policy and institutional framework. The index is based on 20 criteria, grouped into four clusters: econ omic management, structural policies, policies for social inclusion and equity, and public sector management and institutions. The CPIA is available for a large sample of countries for the whole period. It incorporates information about the conduciveness of a country's policy framework for reform. Including it reduces concerns that the borrower variable might be picking up a country's general ability to achieve reform. Until 2005, this variable was not available outside the World Bank, and details of its construction were not well known. In the robustness checks, therefore, CPIA is replaced with proxies for insti tutional development that are more readily available and (arguably) less endogenous. One could view the basic model as a one-lag vector autoregression in four variables (CPIA score, inflation, real growth, and M2IGDP). There is no guarantee, however, that this is the correct reduced-form model. A series of models tested for the appropriate included variables and lag lengths by adding lags for each of the explanatory varia bles until the last lag added was not significant. This measure was taken in order to ensure that the underlying model of the indicators of financial development is as complete as possible before adding the treatment variables. Including the additional lags reinforces the conclusions about the relative performance of borrowers and nonborrowers. Indeed, differences between borrowers and nonborrowers are larger in the specifications with multiple lags. To reduce clutter in the specifications and for ease of exposition, the analysis uses the one-lag models as the base specifications (results from the multiple-lag specifications are also discussed below). III. RESULTS In the base specifications, percentage changes in the indicators of financial development are measured by taking their logs (table 2}.15 Two of the simplest 15. Taking logs also helps reduce the influence of outliers in the estimated coefficients. Cull and Effron 325 TABLE 2. Base Results: Fixed-Effects Panel Regression l!:itvtioolli ...HlH.rolS 0.004 -OJ)Q7 -0.004 0.001 0,016 0,033 HMH?) IO,tH?) {0.017} {(!'oM} {OJ).H) (0.0]4) O.OO3·~ O.OO3~· n 0.003'" 0.003... · OJXL'i OJ.104" (0,001) 10,(01) (O.OOt} 10.(0 1) iV,002) (0002) ((1.002) (0.002) -(j,OOOOOI -i),OOOOJ -6.0001 0.0001 -000016 -!}JJOO17 -0.0001 0.0002 IO.OOOi} (0.0001) ((1.0001) :0.(001) 10.00(3) l{i.OOOJl lOjJOO.ll \;),0003) Q.Ot5 .. ~ o.015u~ O.(HS..... O.O}4]..... O.OO'!u" OJ)(W~'" O.OO9"U jI).OOt) :O.OOli (O.OO]} lO.OOll IO.OOlj j{tOO1) i{tool) PO)IIC1"l\nabl~ CUltlu',lUve OJ}l~ -OJ)47~U ~llustmrnt (0.013) 10 .016 ) lOJ}28) k.I1'1S Numbnorloar,/or 0.041* (3)18 b,mk 10.022) (0,046) p',YlloutW:rl Numh"f 1)( ljJ.n~ for 0.041 -OJH4 It)l.Ulafivnl 10.030) (OJJ61) '1 [1Crvmlln Numt>..rofloJl'"foI -0.156 tlll'ltrre(ofilu (O.OH) (0.053) Numht·rollmll'>{OJ: 0.040 Q.3jl)H+ 3uditins/ !0.(41) (OJ}U) .. ~()unnug Humh.:t of loaf>, for 0.083"· 0.115"''' (0,029) (0.061) Numftttof loar., -0.015" -O.{)j9 fw~r... 1 lO.Dl~1 (0,052) ;;rum;:llI] Runtl !:"llane< 0.199" it.2l:;-· lO.on) (0.110) -0.059 ~0.247" 10.(69) (0.142) Nooh.utk Mannal -0.199"" -0.14J (0.1)44) (0.092; 2.71· ... 2.111"·' 1.80~·" 2.110·· .. 2.6J"H 1.37U. 2.39~·· 2,)7'" 2.30·" (0.062) ,D.OM) !0.(67) (0.066) 10.032) (0.115) (0.136) ,0.137) 10.137) 611 611 611 611 '" 610 610 610 610 observanoos NumhN of COUflrrues R'Squared " 0.42 " OA3 " 0.44 " (lA9 " 0.06 " 0.09 " 0.09 . 0.10 . 0.17 I",.thln) 2.1119.7) 2.t2(9,S) 2.11 2.14 1.Hl 2.19 2.20 (-9.12! RKdH·dWorld "(1028 U " -O_OJ4<-U -0.001 -0.017 -OJ}24* -11005 -0,011 0.013... · o.ft7,..... ""k 10.0(3) (0,00$) (OJ}Bj 10.013) ((}.014) iOOO7l lllOO9) {0.021, (0012) 10022) (Continued) 326 THE WORLD BANK ECONOMIC REVIEW TABLE 2. Continued Log(cashlM2) log(tnfert"stspread) 2.11(9.7) 2.12(9.8) 2.13 2.14 2.15(9.9) 2.16(9.10) 2.17(9.11) 2.18 2.19 2.20(9.12) Old not receive -0.017· U 0.001 0.0001 0.001 O.OOO.~ O.OlS 0.024" 0.024" World Bank (0.004) (0.006) (0.006) (0.006) (0.006) (0.010) (0.012) (0.011) (0.011) (0.011) adJustment]oan H o:f3,.,t.=/3"""wb No No No No No reJected?(p =0.05) Years smce fir5( -0.034" -0.042'" -0.045'" -0.116'" -0.113'" -0.111'" :ldju51meotloan (0.016) (0.016) (0.016) (0.026) (0.027) (0.027) No No No Yn Yn ,,,,,.:d,,rejected? ""'-010,= /3n".".'~ rejected? MacroecoROmicalld lflsmutJonal controls CPIAscore, I -0.025 -0.014 -0.030 -0.030 0.014 0.048 0.054 0.046 (O.OJO) (0.030) (0.030) (0.030) (0.050) (0.049) (0.050) (0.0$0) GOP growth, J 0.001 0.0003 0.0001 0.001 -0.005 -0.006 -0.006 -0.007" (0.002) (0.002) (0.002) (0.002) (O.(Hl4) (0.004) (0.004) (0.004) Inilanon,_! -0.0004' -0.0004 -0.0003 -OJJ006 0.001' 0.001" 0.0005 (0.0002) (0.00024) (0.0002) (0.0003) (0.0003) (0.0003) (0.000.1) (0.0004) M2IGDP,_J -O.OOS'" -0.008'" -O.OOS··· -0.007'" O.OOS·" 0.010'" 0.01O*" 0.009'" (0.002) (0.002) (0.002) (0.002) (0.003) (0.003) (0.003) (0.003) Policy vanables CumulatIVe 0.OS6··· 0.105 0 " -0.025 0.022 adlU5lment (0.023) (0.027) (0.038) (0.050) loam Number of loans for 0.037 0.069 bank (0.038) (0.073) privatization Numberofloan~ for 0.041 0.099 regulation! (0.052) (0.111) ~upervi~ion Number of loans for 0.030 -0.110 orherrt"forms (0.047) (0.087) Number of loans for -0.206 u , 0.170 auditmg! (0074) (0.133) accounting Numberoflo~ns for -0.149'" capltalmarkt:( (0.052) (0,)14) Number of loans 0.010 0.053 forgenenl (0045) (0.092) financial Rut;!.i finance -0.2J~u -0.012 (0.090) (0.134) Mlcrofinance 0121 -0.107 (0.101) (0.228) NOllbankfinancial 0.Q25 0.112 (0.OS5) (0.161) J.OJ"· 3.39'" 3.29'" 3.35'" ].lS··· lAO'" 1.50'" (0.017) (0.109) (O.lIS) (0.117) (O.IIS) (0.044) (0.185) (0.196) (0.199) (0.205) Numbcrof 1119 671 67] 671 67] 532 366 366 366 366 observanons Numberofcouurnes 98 87 87 87 87 60 57 57 57 57 R-squared 0.09 0.15 0.16 0.18 0.22 0.01 0.07 0.12 0.13 0.16 (within) ·Significa.nT al [he 10 percell[ level; "slgmficant at th("" 5 percent level; "'sIgnificant at the 1 per~entlevel. Note: All models Include country fi:. All macroeconomic and inmtunonal comrol vanables are lagged one year. Standa.rd errors appear In parentheses. 50urct>: Authors' anal}sls based on data $Ource~ described In the text and in [able 1. regressions (2.1 and 2.11) indicate that borrowers had significantly more rapid growth in M2/GDP and more rapid declines in cashlM2, both signs of better financial development. This pattern holds when macro/institutional controls Cull and Effron 327 are included in the regressions (models 2.2 and 2.12).16 Interest spreads tended to widen for non borrowers and decline for borrowers, though neither coeffi cient is significant in model 2.16. When the macroeconomic and institutional controls are included in model 2.17, the non borrower coefficient is positive and significant, and the borrower coefficient remains negative and insignificant. Borrowers thus tended to outperform nonborrowers in terms of spreads. The pattern of results is different for private credit, which grew faster in nonborrowing than in borrowing countries, though the difference is not signifi cant in either the simplest regression (2.6) or the regression that includes con trols (2.7). In addition, private credit growth was positive and significant for borrowers in both specifications, which might come as a surprise given the number of borrowers that suffered financial crises during the period (see table A-1 for a list of borrowers). If World Bank loans are designed to spur financial development, borrowers should enjoy significantly faster credit growth than nonborrowers. One possible explanation for the fact that they did not is that the non borrower growth rates are "too high." Private credit growth rates are more than twice as large as M21 GDP growth rates for nonborrowers in the simplest specification, and the non borrower coefficient does not achieve significance in the M2/GDP specification when the controls are introduced. Such a pattern may be possible over a short period; over longer periods, it is likely to be destabilizing and unsustainable. Indeed, several articles indicate that rapid growth in indicators of financial depth, particularly those related to credit, can be so destabilizing that they lead to crisis (Demirgii~-Kunt and Detragiache 1998; Kaminsky and Reinhart 1999; Honohan 2004; Loayza and Ranciere 2004).17 By contrast, the growth rates for borrowers seem more reasonable (1 percent for private creditlGDP and 2-4 percent for M2/GDP). Improvements were largest for indicators that are (arguably) better suited to capturing short- and medium-term financial sector development (cashlM2, spreads, and M2IGDP). Therefore, it could be argued that sufficient time had not elapsed to see the full effects of reform on private credit for borrowers. This does not explain why borrowers would perform worse than non borrowers over this period, however. Tests described below indicate that selection could be driving the results in table 2: borrowers came to the World Bank partly because they were less likely than nonborrowers to generate private credit 16. These results and those that follow also hold when additional lags for the macroeconomic and institutional control variables are included. In the base specification for M2fGDP, for example, the borrower coefficient is 0.017, which is significant at the 1 percent level, while the nonborrower coefficient is 0.002, which is insignificant. For the multilag specification, both the borrower coefficient (0.032) and the non borrower coefficient (0.017) are significant at the 1 percent level. In both the base and the full-lag specifications, the hypothesis that the borrower and nonborrower coefficients are equal is rejected at the 1 percent level. The full-lag specifications are available from the authors upon request. 17. Loayza and Ranciere (2004) show that a positive long-run relation between financial depth and growth coexists with a largely negative short-run relation. 328 THE WORLD BANK ECONOMIC REVIEW growth on their own. In the third specification for each indicator in table 2 (models 2.3, 2.8, 2.13, and 2.18), the number of years since first adjustment loan variable is introduced. With the exception of the private credit specifica tions the years since first variable is significantly associated with improved financial development, as expected. This is consistent with the notion that financial reform being a gradual process, countries that received loans earlier in the period experienced greater improvements. Perhaps more important, when the years since first variable enters the regression, the borrower variable is no longer significant. This provides a strong indication that improvements in financial indicators occurred after the receipt of World Bank loans. It, therefore, seems unlikely that borrowing countries were more likely than non borrowers to improve their financial devel opment indicators regardless of whether they received loans. The coefficient for the years since first loan variable (for borrowers) is also statistically distinguish able from that of the non borrower variable for all indicators except private credit. These patterns also hold when the cumulative number of adjustment loans is included (in models 2.4, 2.9, 2.14, and 2.19). The cumulative loan variable is either insignificant or associated with less financial development across specifications. This result is consistent with findings on repeated struc tural adjustment lending (Easterly 2005). In the private credit specifications, years since first loan is always posi tive, though never significant. The borrower variable is also positive and insignificant in all specifications. When the two coefficients are jointly eval uated, the null hypothesis that their sum is equal to zero is rejected at the 5 percent level or better in specifications 2.9 and 2.10. Thus there is some statistical support for the idea that private credit grew in borrowing countries and that the improvements occurred after the receipt of Bank loans. However, the null hypothesis that the difference between the coeffi cients for borrowers and nonborrowers is zero cannot be rejected. Borrowers did not outperform non borrowers in private credit growth in any of the specifications in table 2. The final specification for each indicator (models 2.5, 2.10, 2.15, and 2.20) includes variables that summarize the policy reform areas covered under World Bank loans. Their inclusion does not alter the comparisons between borrowers and nonborrowers, but the interpretation of the results changes slightly: the coefficient on the borrower variable now indicates the impact of participation if there were no conditions attached to loans in any of the policy areas that are controlled for. With the exception of the capital markets development and rural finance variables, the policy variables tend not to be significant across indicators, and the borrower coefficient is similar to that when policy variables are not included in the specification. While one could come up with explanations for the patterns of the policy coefficients in table 2, it is best not to invest too much effort in this direction. The policy variables are the best that have been put together to study the Cull and Effron 329 effects of reform on financial development, but they have some limitations. First, not all loans that covered a policy area did so in the same way. Some loans may have devoted substantial resources to the policy area, while others may not have done so. Second, because the classifications are based in part on the objectives stated in the documents describing the loans, these measures summarize ex ante indications of planned reform rather than actual ex post reforms. Therefore, the policy variables are a set of coarse controls, included to examine whether the primary results on borrowers versus non borrowers hold up. The focus is on the simplest decisions-that is, whether or not to borrow and how many loans to take out-rather than on a painstaking qualification of the nature of the reforms to produce variables that are unlikely to explain vari ation in country-level aggregate financial indicators. Finally, the control variables that are significant tend to be associated with the financial indicators in the ways one would expect (higher inflation and slower growth retard financial development, for example).18 CPIA scores are not significant, perhaps because of collinearity with M2/GDP, which is associ ated with improved financial development for all indicators except interest spreads. 19 The M2IGDP coefficients suggest that the variable could be viewed as a proxy for a country's willingness and ability to undertake financial sector reform. IV. SELECTION EFFECTS The sample of borrowers is unlikely to be random. Selection bias could work in either direction. Countries with the greatest potential for financial develop ment might prefer to pursue reform on their own rather than incur World Bank debt and have to negotiate and adhere to conditions. Alternatively, countries that are ill prepared to achieve financial reform may find themselves ineligible for Bank adjustment loans on mutually acceptable terms. Nonrandom selection of borrowers can be dealt with in at least two ways. One possibility is to use treatment-effects regressions, which consider the effect of an endogenously chosen binary treatment (in this case, the choice to borrow) on another endogenous continuous variable (in this case, indicators of financial development), conditional on two sets of independent variables. The first set of independent variables is used to estimate a selection equation that describes the participation choice. Information from the selection equation is then used in the financial development regression. The key 18. For robustness year dummy variables were also included in the base regressions to control for global factors that might have affected financial development in all countries. These dummy variables were significant only in the interest rate spread regressions; the qualitative comparisons between borrowers and nonborrowers were similar to those for the base regressions. 19. The CPIA variable becomes positive and significant when M2IGDP is dropped from the private credit specifications. 330 THE WORLD BANK ECONOMIC REVIEW difficulty is finding an appropriate set of exogenous variables for use in the selection equation. A second option for facilitating fairer comparisons between borrowers and nonborrowers is propensity-score matching. The intuition underlying this method is that certain country types (for example, the most institutionally sound) are more apt to respond favorably to the treatment than others. To the extent that the control group is more (or less) heavily weighted toward types that are less likely to respond favorably, comparisons with the treatment group can be misleading. The propensity-matching technique therefore matches treat ment and control observations based on relevant observable characteristics: apples are compared with apples and oranges with oranges. However, it can be difficult to judge a good match when treatment and control group observations can be compared on multiple observable dimensions. Propensity-score match ing can reduce that dimensionality by summarizing the impact of observables in a single equation. A standard probability model (logit or probit) is used to estimate the conditional probability of receiving the treatment (in this case adjustment loans) given a set of covariates. Because the equation is used only to reduce the dimensionality of the conditioning, no behavioral assumptions are attached to it. Thus, unlike in the treatment-effects regressions, the exo geneity of the covariates is not a concern. Contemporaneous variables can be used, and higher-order transformations of those variables are typicaL Applications of these techniques usually involve matching a relatively small set of treatment observations to a subset of a relatively large pool of nontreat ment observations. In this case, the set of nontreatment observations is limited, because there are only 38 non borrowers in the sample (see table A_1).2o Treatment-effects regressions are, therefore, relied on. Propensity-matching techniques were also applied to these data (the results are presented and dis cussed in supplemental appendix S-l, available at http://wber.oxfordjournals. org/). In general, propensity matching yields results that favor borrowers a bit more than the base results do. In many Heckman-type selection models, the dependent variable is observa ble only for individuals (or households or countries) that received the treat ment. In this analysis, indicators of financial development are observable for borrowers and non borrowers alike. Treatment-effects models are, therefore, estimated in which (2) Yj = 0' + /3Xj + 5Zi + 8i 20. In principle, it would be possible to increase the number of observations by going back to the panel data set. However, the nearest matching control group observations would almost certainly be from borrowing countries in years when no adjustment loan was in place. As in a fixed-effects regression with a dummy variable for current borrowing status, this would provide information only about those countries that changed their borrowing status during the period. Because the goal here is to compare countries that borrowed with those that did not, applying propensity matching to the panel data set was not appropriate. Cull and Effron 331 where Y is an indicator of financial development; X is the vector of macroeconomic, institutional, and policy control variables; and Z is the endogenous treatment variable indicating whether or not country i borrowed. 21 As is typical in this literature, the decision to borrow is modeled as the outcome of an unobserved latent variable Z*, which is a function of exogenous covariates Wand a random component u: (3) The researcher observes that (4) ZI 1,if Z; > 0 Z· I ootherwise. Because there is an element of self-selection in borrowing from the World Bank and the error term of the model that summarizes this choice (equation 3) could be correlated with the error term in the regression of interest (equation 2), a valid set of instruments is needed. These instruments should be highly cor related with the endogenous regressor (the borrowing dummy variable) but contemporaneously uncorrelated with the error term in equation 2 (that is, truly exogenous). It is difficult to find exogenous variables for use in the selection equation. It is very likely, for example, that proxies ror borrowing needs, as reflected in measures of countries' fiscal health and indebtedness, are themselves endogenous. Appropriate instruments are found by turning to the literature on the politi cal economy of international financial institutions' lending to test whether strong or weak potential reformers are more likely to receive Bank adjustment loans to promote financial sector development. Vreeland (2004) offers the following propositions regarding IMF lending. The head of the executive branch in a developing country is more likely to enter into a lending arrangement with the IMF when the governmental structure dictates that the executive face a large number of veto players. And the IMF is more likely to lend to countries that have fewer veto players. The intuition underlying the first proposition is that reform-minded execu tives in developing countries use IMF support to help overcome opposition to potentially unpopular policies. The idea is that after the executive reaches an agreement with the IMF, failure to achieve reform is more costly, because rejec tion of those policies is also seen as a rejection of the IMF, which all domestic 21. As a robustness check, the total number of loans was also treated as endogenous in specifications that are not presented. The total number of adjustment loans was used to create a dummy variable for "high participation," defined as more than five loans. The high-participation dummy variable was then corrected for selection bias using a treatment-effects regression. These results, which are similar to those for the simple borrower/nonborrower dummy variable, are available upon request from the authors. 332 THE WORLD BANK ECONOMIC REVIEW politicians and interest groups may recognize as costly.22 The likelihood that a head of government uses IMF agreements in this way depends on the checks and balances the executive faces. Leaders facing no veto players (that is, dicta tors) would have no need for IMF support. Leaders facing too many veto players are unlikely to be able to overcome opposition despite IMF support. Because the IMF prefers to finance successful reform projects, it is likely to be unwilling to enter into agreements with executives who face a large number of veto players. The combination of these two effects should result in a nonlinear relation between the number of veto players and the probability of a loan. In some intermediate range, IMF agreements should be most prevalent, because they are more likely to achieve the desired objective of overcoming the opposition of veto players. World Bank adjustment loans could serve a similar purpose. These concepts are operationalized using data on the number of checks and balances stipulated in country constitutions (Beck and others 2001). The number of checks and the squared number of checks are included in the selection equations that follow to test whether Vreeland's hypotheses are valid for this data set. If they are, the coefficient for the checks variable should be positive and the checks-squared variable negative in the selection equation. Thus, the likeli hood of receiving a World Bank loan for financial sector development would first increase as countries move away from dictatorship (as a result of self selection by the country) and then decrease when the number of veto players passed some threshold value (as a result of the Bank's selection criteria). A country's borrowing needs may also affect the likelihood of receiving World Bank adjustment loans. World Bank lending commitments are positively related to an increase in debt service payments and negatively related to the level of international reserves of the borrower (Ratha 2005). As noted, however, contemporaneous measures of countries' fiscal health and indebted ness are likely to be endogenous. Information on fiscal health and indebtedness from 1970 to 1989 is therefore included in the selection equation, which is by definition not contemporaneously correlated with the error term in the finan cial development regressions (which use data from 1992 to 2003). It is also conceivable that developing countries-particularly countries with a relatively large stock of World Bank debt-use the proceeds of new Bank loans to repay old loans (evergreening). Beyond some point, however, debt accumulation becomes problematic, making future agreements less attractive, especially from the Bank's point of view. For countries with little past borrow ing, predictions about future borrowing are difficult to make. If the lack of bor rowing reflects a preference for self-reliance, one would expect little future borrowing. If demand for loans is cyclical, lending would decline during up cycles and increase during down cycles. 22. According to Vreeland (2004, p. 2), "The IMF may restrict access to loans, it may preclude debt rescheduling with creditors who require an IMF arrangement to be in good standing, and decreased investment may result if investors take cues from the IMF." Cull and Effron 333 A variety of variables was used to measure countries' past and current indebtedness and overall fiscal health to test these hypotheses. Squared terms enter the selection equation to capture any nonlinearities between past indebt edness and the likelihood of receiving a World Bank loan. Using historical data to predict whether countries borrowed for financial sector development makes it impossible to estimate a selection effect that varies by year for each country. The likelihood of receiving at least one adjustment loan since 1992 is estimated based on data from 1970 to 89. For this reason, the subscript t does not appear in equation (2). The (largely time-invariant) governmental checks variable is better suited to the cross-sectional regressions than to the panel regressions. The coefficients from a simple probit regression that uses the borrower dummy variable as the dependent variable are as follows: Borrower; = 0.47 + 0.29Checks; 0.07Checks7 - 0.030ebt; (0.69) (0.29) (0.04)* (0.02)* + 0.00030ebtf + 0.0016IMF Credit; (0.0002)* (0.0008)** Number of observations: 79 Pseudo R-squared: 0.15 Standard errors in parentheses "significant at the 10 percent level .... significant at the 5 percent level. The coefficients from the pro bit regression and those from the selection equations in the treatment-effects models that follow provide support for the hypotheses in this section. The checks and checks-squared coefficients imply that Bank loans are most likely for an intermediate level of checks. Various measures of past fiscal health were tried, including the current account balance, tax revenues as a percentage of GOP, and the overall government budget balance. Because there are relatively few observations, only two such variables-total IMF borrowing (in millions of constant dollars) and total external debt (as a share of GOP) from 1970 to 1989-are included in the selection equations. The debt variable is negative and its square positive, imply ing a U-shaped relation with the probability of receiving Bank loans. Thus countries with little past borrowing were more likely to receive loans than those with intermediate levels, possibly indicating that borrowing needs are cyclical. However, heavily indebted countries from 1970 to 89 were the most likely to borrow for financial sector reform from 1992 to 2003, providing additional support for the evergreening hypothesis. 23 The positive coefficient on the IMF borrowing coefficient is also consistent with evergreening. 23. Only a linear and a quadratic term for debt are included in the selection equation, making it impossible to test whether the probability of borrowing eventually declines for extreme levels of indebtedness. The qualitative results for the financial development regressions are similar when the quadratic debt term is excluded from the selection equation. 334 THE WORLD BANK ECONOMIC REVIEW Although the cross-sectional approach is more promIsmg than the panel approach for handling the selection problems faced here, the approach could make it more difficult to find significant results, because standard errors are likely to be larger in regressions with few observations. Skeptics of the panel results above could argue that because the error terms from multiple obser vations from the same country are likely to be correlated, the number of inde pendent observations is the same as the number of countries in the data set. Restricting the observation set to the cross-section of countries can, therefore, be viewed as an additional test of whether borrowing countries outperformed nonborrowers in terms of financial development. For the treatment-effects regressions, growth in Y in year t is calculated as Y/Yt - 1 · The average of annual growth rates over the whole period for each country is used to derive one observation per indicator per country. These country averages are used as dependent variables in the ordinary least squares (OLS) and treatment-effects regressions in table 3. The OLS results in table 3 are similar to those from the panel regressions in table 2, indicating that those results were not solely the product of multiple observations for each country. In particular, M2/GDP grew and cashlM2 declined significantly more rapidly among borrowers. Borrowers' interest spreads declined more rapidly than those of nonborrowers, but the result is not significant in the cross-sectional OLS regression, possibly because there are only 47 observations for that variable. As in the panel regression in table 2, borrowers had slower rates of private credit growth than non borrowers, although the difference is not statistically significant. After correction for self-selection using the treatment-effects model, the results show that borrowers outperformed non borrowers by a wider margin. 24 The change is most pronounced for M2 growth (models 3.2 and 3.3) and private credit growth (models 3.5 and 3.6). At the risk of reading too much into these models, this suggests that the typical World Bank borrower had rela tively poor prospects for financial development. Once this is accounted for econometrically, the positive effects of Bank involvement are easier to detect. Treatment-effects regressions for cashlM2 are more volatile than those for private credit and M2/GDP. In model 3.8, which does not include control variables, the borrower dummy variable is insignificant. Multiple variables are significant in the selection equation, and the likelihood ratio test at the bottom of table 3 indicates that errors from the selection and cashlM2 equations are independent. Thus, the OLS results are valid, and there is no need to perform treatment-effects regression. In model 3.9, which includes institutional and macroeconomic controls, the borrower dummy variable is positive and 24. All treatment effects models in table3 are estimated using maximum likelihood estimation, These models were also estimated using the two-step version of the treatment-effects model. The results were qualitatively similar, except that the borrowers dummy variable was no longer significant in the private credit growth models. On efficiency grounds, the maximum likelihood results are preferred. TABLE 3. Cross-sectional, OLS, and Treatment-Effect Regressions Average change in M2IGDP Average change in private credit/GDP Average change in cashlM2 Average change in interest rate spread Item 3.1OLS 3.2 Treatl 3.3 Tre.t2 3.4 OLS 3.5 Treatl 3.6 Treat2 3.7 OLS 3.8 Treatl 3.9 Treat2 3.IOOLS 3.11 Treatl 3.12 Treat2 Received Bank 0.019' 0.075'" 0.074'" "0.0005 0.144" 0.138'" -0.031'" -0.005 0.059'" -0.128 -0.159 -0.243 adjustment (0.009) (0.014) (0.014) (0.019) (0.020) (0.026) (0.011) (0.036) (0.017) (0.090) (0.244) (0.183) loan CPIA score 0.010 0.012 -0.0001 -0.008 -0.014 -0.010 -0.168' -0.165' (0.010) (0.010) (0.022) (0.018) (0.012) 10.011) (0.094) 10.091) Inflation 0.0001* 0.00005 -0.0001 -0.00016 0.00002 0.00001 0.0001 0.0002 (0.00005) (0.0001) (0.0001) (0.00012) (0.0001) (0.00004) (0.001) (0.0005) Real growth 0.005' 0.004 0.015'" 0.011'" "0.005 "0.006" 0.006 0.008 (0.003) (0.003) (0.006) (0.004) (0.003) (0.003) (0.025) (0.025) M2IGDP -0.0001 -0.0002 -0.00005 0.0001 0.001 0.002 (0.0002) (0.0002) (0.0005) (0.0003) (0.002) (0.002) Constant 0.971 H . 0.987'" 0.939'" 0.988'" 0.953'" 0.950'" 1.06""10 0.984'" 1.00· ao ll< 1.54.. ·· LOS .... · 1.59.. · .. (0.032) (0.010) (0.031) (0.066) (0.018) (0.055) (0.036) (0.021) (0.035) (0.287) (0.141) (0.283) Selection equation External debt -0.034' -0.029' -0.018 -0.005 -0.035 -O.ot8 -0.019 -0.011 1970-89 (0.018) (0.017) (0.011) (0.005) (0.022) (0.017) (0.035) (0.034) (percent of GOP) Extern.l debt 0.0003" 0.0003' 0.0001 0.00004" 0.0003' 0.0001 0.0002 0.0002 squared (0.0001) (0.0001) (0.00008) (0.00002) (0.0002) (0.0002) (0.0003) (0.0003) Number of 0.210 0.164 0.234 0.258 0,478 0.616' 0.260 0.239 governmental (0.228) (0.227) (0.145) (0.170) (0.315) (0.357) (0.538) (0,483) checks Number of -0.050 -0.044 -0.035' -0.034 -0.089" -0.097 -0.056 -0.047 governmental (0.034) (0.032) (0.020) (0.024) (0.045) (0.061) (0.081) (0.069) checks squared IMF borrowing 0.0015** 0.0014" 0.00084' 0.0006 0.0017" 0.0015" (0007) 0.0013 ~ l:> 1970-89 (0.0006) (0.0006) (0.00045) (0.0005) (0.0008) (0.0010) ;:t ;:... (constant $ millions) Constant 0.608 0.532 0.137 -0.233 0.140 -0,412 -0.061 (1.05) -0.089 (0.541) (0.534) (0.294) (0.320) (0.711) (0.546) (0.960) (Continued) w w v, <...> <...> TABLE 3. Continued 0\ Average change in M2/GDP Average change in private creditlGDP Average change in cashlM2 Average change in interest rate spread o-j ::r: Item 3.10LS 3.2 Treat1 3.3 Treat2 3.4 OLS 3.5 Treat1 3.6 Treat2 3.7 OLS 3.8 Treat1 3.9 Treat2 3.100LS 3.11 Treatl 3.12 Treat2 '" :!iJ Number of 73 72 72 73 76 76 70 74 74 47 47 46 o :>:> countries 0.16 0.14 0.18 0.13 r I:) 94.99 98.04 47.90 53.28 82.61 89.96 -36.76 -35.S4 '" ;,. Z 29.24 37.08 50.67 45.38 0.02 32.13 0.43 5.71 0.000 0.000 0.000 0.000 0.8995 0.000 0.5138 0.35 '" tTl n squared o Likelihood ratio 10.85 8.92 17.79 16.70 0.39 6.36 0.01 0.34 Z test of o i!:: independent n eqns (rho 0) :>:> P > Chi squared 0.001 0.003 0.000 0.000 0.5308 0.0117 0.9191 05591 '" <: tTl 'Significant at the 10 percent level. "significant at the 5 percent level; · "significant at the 1 percent level. ~ Note: All policy and control variables (CPIA, inflation, deficitslGDP, M2IGDP) are averaged over 1991-2000. External debt and IMF borrowing are annual averages for 1970-89. Number of governmental checks .s from Beck and others (200 1). errors a ppear in parentheses. Source: Authors' analysis based on data sources described in the text and in table 1. Cull and Effron 337 significant, indicating less confidence in the financial system. Although fewer variables are significant in the selection equation than in model 3.8, the likeli hood ratio test indicates that errors from the selection and cashlM2 equation are not independent; the treatment-effects results are thus preferred over the OLS results. Because the cashlM2 results are highly sensitive to slight pertur bations in either the selection equation or the equation of primary interest, it is difficult to draw a strong conclusion for that variable based on table 3. In contrast, the selection equations for M2/GDP produce many significant coefficients, and the likelihood ratio test indicates that the treatment-effects model is preferred to the OLS model. In the treatment models, the borrower coefficient is nearly identical whether or not controls are included. This rela tively stable pattern of results lends credibility to the conclusion that borrowing countries performed better than non borrowers on that dimension. Significance levels in the selection equation for private credit are somewhat lower than for M2/GDP, but the coefficients are similar. As in the simple probit discussed above, the governmental checks and checks-squared coefficients from the selection equations indicate that the prob ability of receiving an adjustment loan increases from one to three checks but declines thereafter (figure 1). Loans from international financial institutions are therefore most likely for intermediate levels of checks. The debt and debt-squared coefficients indicate that countries with low levels of debt in the 1970s and 1980s were more likely to be borrowers in this data set than those with moderate levels of debt (figure 2), a finding that is consistent with the hypothesis that debt levels may be cyclical. The selection equations therefore provide plausible results in many of the treatment-effects regressions. FIGURE 1. Probability of Receiving World Bank Adjustment Loan for Financial Sector Development as a Function of Number of Political Checks Probability (percent) 60 40 20 O+-------r------,r-----~------_.--~ 3 5 7 9 Number of political checks Source: Authors' calculations based on table 3, model 3.2. Data sources are as described in the text and in table 1. 338 THE WORLD BANK ECONOMIC REVIEW FIGURE 2. Probability of Receiving World Bank Adjustment Loan for Financial Sector Development as a Function of Level of External Debt Probability (percent) 100 80 60 40 20 0+------.-----.------.-----,------, o 50 100 150 200 250 External debt (percent of 1970-89 average GNI) Source: Authors' calculations based on table 3, model 3.2. Data sources are as described in the text and in table 1. These regressions reinforce conclusions about the pOSItIve assocIatIOn between borrowing and financial development. For M21GDP and private credit/GDP borrowers outperform nonborrowers in the treatment-effects models. Unlike some of the base models for private credit that do not control for selection, none of the treatment-effects specifications indicates that bor rowers underperform non borrowers. For interest spreads the hypothesis that the errors from the first- and second-stage regressions are independent cannot be rejected, in which case no correction for selection is required. The base results in table 2 and the OLS results in table 3 are thus valid. The coefficient for borrowers is negative in both sets of regressions, highly significant in the base results, and nearly significant in the OLS results. For cashlM2 the treatment-effects results are unstable; it is thus not possible to draw strong con clusions from them. V. ADDITIONAL ROBUSTNESS CHECKS A series of tests indicates that the main findings are not driven by the regional composition of borrowers and nonborrowers and are robust to the inclusion of variables that measure a country's readiness for and experience with reform and to the substitution of ratings of financial sector development for quantitat ive measures of financial sector development. In the readiness for reform regressions, CPIA scores are replaced with a measure of the degree to which countries adhere to the rule of law developed by the International Country Risk Guide (ICRG). Several findings emerge from this analysis. First, in Latin America and the Caribbean financial sector development was stronger in countries that Cull and Effron 339 borrowed from the World Bank than in countries that did not (supplemental appendix S2). Second, the basic pattern of results holds when countries from Europe and Central Asia and Sub-Saharan Africa are dropped from the analysis (supplemental appendix S3). Countries in Europe and Central Asia might have been driving the base results, because many of them began the period of study with low indicators of financial development that improved largely as a result of macroeconomic stabilization. Countries in Sub-Saharan Africa tended to be in the non borrowing group; the base results might have been picking up differ ences in financial development between them and countries from other regions. Neither of those concerns is supported by the data. Third, the main findings are robust to the inclusion of controls for whether a country was ready for reform, what other reforms it had already taken when it received financial sector adjustment loans, and what other agencies were involved in its reforms (supplemental appendix S4). Fourth, results are similar when an index of banking and financial sector freedom replaces the quantitative indicators as the dependent variable (supplemental appendix S5): countries that borrowed from the \X'orld Bank experienced greater improvement on the index than those that did not. VI. CONCLUSIONS Evidence based on analysis of a new data set on Bank adjustment loans that supported financial sector reform from 1992 to 2003 indicates that borrowing countries performed better than nonborrowers on multiple measures of banking sector development, including M2/GDP, interest spreads, and cash! M2. They performed worse than non borrowers on private creditlGDP in OLS regressions. Improvements in financial indicators occurred after the inception of adjustment lending, even after controlling for the adverse selection effects associated with repeated lending to the same country. The main findings hold both in panel regressions that incorporate fixed-country effects and in cross sectional regressions that use average growth in financial indicators over the full period for each country as dependent variables. The cross-sectional regressions indicate that the panel results are not driven by multiple obser vations from the same country, which can artificially reduce standard errors. A series of models accounts for potential selection effects. Nonlinear selec tion equations capture concepts from the political economy literature on the relations between international financial institutions and developing countries. This approach, therefore, distinguishes countries that prefer not to borrow from these institutions, because they are relatively self-sufficient from those that international financial institutions prefer not to deal with because reform is unlikely to succeed. Addressing nonrandom selection using treatment-effects regressions reveals that the rate of growth of private credit and M2 was signifi cantly larger for borrowers than for non borrowers. For interest rate spreads test statistics indicate that the errors from the selection and financial 340 THE WORLD BANK ECONOMIC REVIEW development regressions are independent, obviating the need to correct for nonrandom selection. For cashlM2 the treatment-effects results are highly sen sitive to small perturbations in the specification, but some models indicate that no correction for selection effects is necessary. Robustness checks indicate that the results are not driven by the regional composition of borrowers and nonborrowers and are robust to the inclusion of proxies for countries' readiness and ability to reform. Taken in their entirety, these results suggest that the World Bank adjustment loans studied here had some positive effects on financial sector outcomes. ApPENDIX A-1. Countries That Did and Did Not Receive World Bank Adjustment Loans for Financial Sector Reform between 1992 and 2003 Countries that received World Bank Countries that did not receive World Bank adjustment loans adjustment loans Albania Angola Algeria Benin Argentina Banngladesh Armenia Belarus Azerbaijan Botswana Bolivia Cambodia Bosnia and Herzogovina Chile Brazil China Bulgaria Congo, Dem. Rep. Burkina Faso Costa Rica Cameroon Cote d'Ivoire Cape Verde Czech Republic Central African Rep. Dominican Republic Chad Egypt, Arab Rep. of Colombia Estonia Croatia Ethiopia Ecuador Gabon El Salvador Gambia, The Georgia India Ghana Iran Guatemala Kenya Guinea Lebanon Guyana Lesotho Honduras Mali Hungary Mauritius Indonesia Nepal Jamaica Nigeria Jordan Panama Kazakhstan Papua New Guinea Korea, Rep. of Paraguay Kyrgyz Rep. Senegal Lao, PDR South Africa (Continued) Cull and Effron 341 ApPENDIX A-I. Continued Countries that received World Bank Countries that did not receive World Bank adjustment loans adjustment loans Latvia Sri Lanka Lithuania Swaziland Macedonia Togo Madagascar Trinidad and Tobago Malawi Venezuela, R. B. de Malaysia Zimbabwe Mauritania Mexico Moldova Mongolia Morocco Mozambique Nicaragua Niger Pakistan Peru Philippines Poland Romania Russia Rwanda Sierra Leone Slovak Republic Slovenia Tajikistan Tanzania Thailand Tunisia Turkey Uganda Ukraine Uruguay Uzbekistan Vietnam Yemen Zambia Note: The 106 countries in this table are those that appear in at least one regression. The maximum number of countries in any regression is 98. Source: Independent Evaluation Group database of World Bank loans for financial sector reform. REFERENCES Barro, Robert, and Jong-Wha Lee. 2002. IMF Lending: Who Is Chosen and What Are the Effects? NBER Working Paper 8951. Cambridge, Mass.: National Bureau of Economic Research. Barth, James, Gerard Caprio, Jr., and Ross Levine. 2001a. "Banking Systems around the Globe: Does Regulation and Ownership Affect Performance and Stability?" InFrederic Mishkined., Prudential 342 THE WORLD BANK ECONO'vIlC REVIEW Supervision, What Works and What Doesn't. National Bureau of Economic Research Conference Report. Chicago: University of Chicago Press. - - - . 2001b. "Bank Regulation and Supervision: What Works Best?" Policy Research Working Paper 2725. World Bank, Washington, D.C. Beck, Thorsten, Asli Demirgu/l-Kunt, and Ross Levine. 2000. "A New Database on the Structure and Development of the Financial Sector." World Bank Economic Review 14(3):597-605. - - - . 2003. "Bank Concentration and Crises." Policy Research Working Paper 3041. World Bank, Washington, D.C. Beck, Thorsten, Ross Levine, and Norman Loayza. 2000. "Finance and the Sources of Growth." Journal of Financial Economics 58(1-2):261-300. Beck, Thorsten, Alberto Groff George Clarke, Philip Keefer, and Patrick Walsh. 2001. "New Tools in Political Economy: The Database of Political Institutions." World Bank Economic Review 15(1):165-76. Caprio, Jr., Gerard, Patrick Honohan, and Joseph Stiglitzeds. 2001. Financial Liberalization: How Far, How Fast? New York: Cambridge University Press. Conway, Patrick. 1994. "IMF Lending Programs: Participation and Impact." Journal of Development Economics 45(2):365-91. Corbo, Vittorio, and Morris Goldstein, eds. 1987. Growth-Oriented Adjustment Programs. Washington, D.C.: International Monetary Fund. Corbo, Vittorio, and Stanley Fischer. 1995. "Structural Adjustment, Stabilization and Policy Reform: Domestic and International Finance." InJere Behrman, and T.N. Srinivasan, eds., Handbook of Development Economics, Vol. III. Amsterdam: Elsevier Science. Demirgli~-Kunt, AsH, and Enrica Detragiache. 1998. "The Determinants of Banking Crisis in Developing and Developed Countries." IMP Staff Papers 45(1 ):81-109. Devarajan, Shantayanan, David Dollar, and Torgny Holmgren, eds. 2001. Aid and Reform in Africa. Washington D.C.: World Bank. Dicks-Mireaux, Louis, Mauro Mecagni, and Susan Schadler. 2000. "Evaluating the Effect of IMF Lending to Low-Income Countries." Journal of Development Economics 61(2):495-526. Easterly, William. 2005. "What Did Structural Adjustment Adjust? The Association of Policies and Growth with Repeated IMF and World Bank Adjustment Loans." Journal of Development Economics 76(1):1-22. Goldstein, Morris, and Peter Montiel. 1986. "Evaluating Fund Stabilization Programs with Multi-Country Data: Some Methodological Pitfalls." IMF Staff Papers 33(2):304-44. Haque, Nadeem UI, and Moshin S. Khan, 1998. "Do lMF-Supported Programs Work? A Survey of the Cross-Country Empirical Evidence." IMF Working Paper WP/981169. International Monetary Fund, Washington, D.C. Honohan, Patrick. 2004. "Financial Development, Growth, and Poverty: How Close Are the Links?" Policy Research Working Paper 3203. World Bank, Washington, D.C. Hutchinson, Michael M. 2001. A Cure Worse than the Disease? Currency Crises and the Output Costs of IMF-Supported Stabilization Programs. NBER Working Paper 8305. Cambridge, Mass.: National Bureau Economic Research. lEG (Independent Evaluation Group). 2006. lEG Review of World Bank Assistance for Financial Sector Reform. Washington, D.C.: World Bank. IMF (International Monetary Fund). Various years. International Financial Statistics. Washington, D.C.: International Monetary Fund. Kaminsky, Graciela L., and Carmen M. Reinhart. 1999. "The Twin Crises: The Causes of Banking and Balance-of-Payments Problems." American Economic Review 89(3):473-500. Kapur, Devesh, John P. Lewis, and Richard Webb. 1997. The World Bank: Its First Half Century. Vol. 1: The History. Washington, D.C.: Brookings Institution. Cull and Effron 343 Khan, Moshin S. 1990. "The Macroeconomic Effects of Fund-Supported Adjustment Programs." IMF Staff Papers 37(2):195-231. KiIlick, Tony, Rumani Gunatilaka, and Ana Marr. 1998. Aid and the Political Economy of Policy Change. London: Routledge. Knight, Malcolm, and Julio A. Santaella. 1997. "Economic Determinants of IMF Financial Arrangements." Journal of Development Economics 54(2):405-36. La Porta, Rafael, Florencio Lopez-de-Silanes, and Andrei Shleifer. 2002. "Government Ownership of Banks." Journal of Finance 57(1):265-301. Levine, Ross. 2005. "Finance and Growth: Theory and Evidence." InPhilippe Aghion, and Steven Durlauf, eds., Handbook of Economic Growth. Amsterdam: Elsevier Science. Levine, Ross, Norman Loayza, and Thorsten Beck. 2000. "Financial Intermediation and Growth: Causality and Causes." Journal of Monetary Economics 46(1}:31-77. Levine, Ross, and Sara Zervos. 1998. "Stock Markets, Banks, and Economic Growth." American Economic Review 88(3}:537-58. Loayza, Norman, and Romain Ranciere. 2004. "Financial Development, Financial Fragility, and Growth." Policy Research Working Paper 3431. World Bank, Washington, D.C. Przeworski, Adam, and James Raymond Vreeland, 2000. "The Effect of IMF Programs on Economic Growth." Journal of Development Economics 62(2):385-421. Rajan, Raghuram G., and Luigi Zingales. 1998 "Financial Dependence and Growth." American Economic Review 88(3):559-86. Ratha, Dilip. 2005. "Demand for World Bank Lending." Economic Systems 29(4):408-21. Summers, Lawrence H., and Lant H. Pritchett, 1993. "The Structural-Adjustment Debate." American Economic Review 83(2):383-89. Svensson, Jakob. 2003. "Why Conditional Aid Doesn't Work and What Can Be Done about It?" Journal of Development Economics 70(2}:381-402. Van de Walle, Nicolas. 2001. African Economies and the Politics of Permanent Crisis, 1979-1999. Cambridge: Cambridge University Press. Vreeland, James. 2004. "Institutional Determinants of IMF Agreements." Department of Political Science, Yale University, New Haven, Conn. World Bank. 1989. World Development Report 1989: Financial Systems and Development. New York: Oxford University Press for the World Bank. - - - . Various years. World Development Indicators. Washington, D.C.: World Bank. HIV Pandemic, Medical Brain Drain, and Economic Development in Sub-Saharan Africa A/ok Bhargava and Frederic Docquier Country-level longitudinal data at three-year intervals over 1990-2004 are used to analyze the factors affecting emigration of physicians from Sub-Saharan countries and the effects of this medical brain drain on life expectancy and number of deaths due to AIDS. Data are compiled on emigrating African physicians from 16 receiving Organisation for Economic Co-operation and Development (OECD) countries. A comprehensive longitudinal database is developed by merging the medical brain drain variables with recent data on HIV prevalence rates, public health expenditures, phys icians' wages, and economic and demographic variables. A triangular system of equations is estimated in a random effects framework using five time observations for medical brain drain rates, life expectancy, and number of deaths due to AIDS, taking into account the interdependence of these variables. Lower wages and higher HIV prevalence rates are strongly associated with the brain drain of physicians from Sub Saharan African to OECD countries. In countries in which the HIV prevalence rate exceeds 3 percent, a doubling of the medical brain drain rate is associated with a 20 percent increase in adult deaths from AIDS; medical brain drain does not appear to affect life expectancy. These findings underscore the need to improve economic con ditions for physicians in order to retain physicians in Sub-Saharan Africa, especially as antiretroviral treatment becomes more widely available. JEL codes: C33, C5, F22, 112, 011, 055. The AIDS pandemic in Sub-Saharan Africa is affecting all dimensions of social and economic life. In many countries, gains in life expectancy achieved over the past several decades have been wiped out. Reductions in life expectancy are Alok Bhargava (corresponding author) is Professor of Economics at the University of Houston; his email address is bhargava@uh.edu. Frederic Docquier is a research associate at the National Fund for Economic Research and professor of economics at the Universite Catholique de Louvain; his email address is frederic.docquier@uclouvain.be. The authors thank C. Ozden, M. Schiff, and A. Winters, of the World Bank, for encouragement and the International Migration and Development Research Program for research support. They also thank P. Ghys and K. Stanecki, of UNAIDS; L.J. Johnson and A. Wittrup, of the International Labour Organization; T. Tan-Torres and N. Van de Maele, of the WHO; and statistical and medical agencies in 16 OECD countries for help in obtaining the data used in the analysis. This article benefited from the helpful comments of three reviewers and the journal editor. The article is dedicated to the memory of Enid M. Fogel, who inspired many economists by her warmth and who devoted herself to improving the lives of young people. TIlE WORLD BANK ECONOMIC REVIEW, VOL. 22, No.2, pp. 345-366 doi:l0.1093/wberllhn005 Advance Access Publication May 15, 2008 © The Author 2008. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development I TIlE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org 345 346 THE WORLD BANI< ECONOMIC REVIEW detrimental at the macroeconomic level, because they reduce economic growth (Bhargava and others 2001). At the micro level, early parental deaths have created the enormous problem of AIDS orphans (USAID, UNICEF, and UNAIDS 2003; Subbarao and Coury 2004). Orphaned children have lower levels of psychological well-being and school attendance, which is critical for learning and for increasing awareness of HIV transmission routes (Bhargava 2005b). Informed policy formulation to deal with these problems requires ana lyses of data at the micro and macro levels. The formulation of policies in the wake of the HIVIAIDS pandemic is complex and would benefit from research on broader biomedical and social science issues. For example, the World Health Organization (WHO) advocates antiretroviral treatment provided by public clinics for patients in HIV/AIDS Stage 4 or with a CD4 cell (type of white blood cell) count below 200 (Gutierrez and others 2004). Studies on the effects of antiretroviral treatment on productivity of undernourished populations can provide insights leading to enhanced treatment strategies. In a similar vein, a recent report by Physicians for Human Rights (2004) emphasizes the need for Sub-Saharan African countries to invest more in training physicians and nurses. The training of additional healthcare staff is hampered by low tertiary enrollment rates in Sub-Saharan African countries; the region could benefit from strategies such as training its physicians in Asia (Bhargava 2005a). Individual-level surveys in six African countries indicate that more than half of all physicians would like to emigrate to developed countries, in search of better working conditions and more comfortable lifestyles (Awases and others 2003). Very large proportions of healthcare staff-38 percent in Ghana, 45 percent in Cameroon, 49 in Senegal, 58 percent in South Africa and Zimbabwe, 62 percent in Uganda-report being "stressed" by caring for HIVI AIDS patients. The risks associated with caring for HIV/AIDS patients and the possibility of children of healthcare staff contracting HIV as they enter adoles cence may exacerbate medical brain drain (Awases and others 2003; Bhargava 2005a). Higher HIV prevalence rates can create a vicious circle, by increasing emigration of physicians and nurses, which can in turn increase deaths from AIDS and the numbers of orphaned children. The trends underlying the international migration of skilled and unskilled labor are complex and reflect several aspects of labor supply and demand conditions in developing and developed countries (Ozden and Schiff 2006). Emigration of skilled workers in general increases remittances and creates business and infor mation networks that can enhance economic performance in the countries of origin. The net effects of emigration of physicians and nurses from developing countries are likely to depend on the domestic demand for healthcare services over a long period. The AIDS pandemic makes it difficult for Sub-Saharan African countries to withstand attrition of already scarce healthcare workers. The results of numerous household-level studies of HIVIAIDS in Sub-Saharan Africa indicate that factors such as multiple sex partners and Bhargava and Docquier 347 sexually transmitted infections can exacerbate HIV transmission (Caral and Holmes 2001). HIV prevalence rates constructed from individual-level data that take account of survival time after contracting HIV are useful for estimat ing life expectancy. The effects of medical brain drain on indicators of well-being-such as life expectancy and the number of adult deaths due to AIDS----cannot be investi gated using data from household surveys; country-level data are needed. Longitudinal data are useful for modeling the relation between HIV prevalence rates and brain drain. Analyses at the country level can provide insights for designing surveys investigating economic and social factors underlying medical brain drain. Variables reflecting medical brain drain are not available for Sub-Saharan African countries. Fortunately, most statistical and medical agencies in receiving countries of the Organisation for Economic Co-operation and Development (OECD) keep longitudinal information on immigration of physicians (infor mation on nurses is not compiled in the same detailed fashion). Databases such as the World Development Indicators (World Bank 2005) contain limited information on HIV prevalence rates. UNAIDS (2006) recently expanded AIDS-related variables on a longitudinal basis for Sub-Saharan African countries. Data on wages from the International Labour Organization (ILO) (2005) and on public health expenditures from the WHO (2006) can be merged with economic and demographic variables from the World Development Indicators to create a comprehensive longitudinal database. This article estimates medical brain drain rates, the number of adult deaths due to AIDS, and life expectancy for 1990-2004 from longitudinal data. Alternative specifications are tested using econometric techniques. The article is structured as follows. Section I briefly describes the data on medical brain drain and other variables. Section II develops the analytical framework for spe cification of the relations, outlining the likely forms of interdependence among medical brain drain rates, the number of deaths due to AIDS, and life expect ancy. Section III describes the empirical models, and section IV the econo metric methods used to estimate and test the models. Section V presents the results from estimating random effects models for medical brain drain rates, the number of adult deaths due to AIDS, and life expectancy for Sub-Saharan African countries at five points between 1990 and 2004. Certain exogeneity hypotheses for the variables are tested to assess the validity of the model assumptions. The last section summarizes the article's main conclusions and identifies some areas for further research. I. THE DATA This article examines the 16 most important OECD countries (Australia, Austria, Belgium, Canada, Denmark, France, Germany, Ireland, Italy, New Zealand, Norway, Portugal, Sweden, Switzerland, the United Kingdom, and 348 THE WORLD BANK ECONOMIC REVIEW the United States) for which longitudinal data on foreign-born physicians are available. These countries account for 93 percent of skilled immigrants in the OECD (Docquier and Marfouk 2006). The medical brain drain can be evaluated in terms of stocks and rates, following Carrington and Detragiache (1999) and Docquier and Marfouk (2006). The rate of medical brain drain m for country i in time period t can be written as: (1) where M jt denotes the stock of physicians from country i working abroad and P denotes the number of physicians working in the home country. Docquier and Bhargava (2007) developed an annual database covering 1991-2004 from data provided by national agencies. For the data extracted from national censuses, two or three data points are usually available; data for the remaining years were interpolated using a log-linear adjustment. Data on the country of qualification of immigrants are available from medical associ ations in Canada, France, New Zealand, Norway, United Kingdom, and the United States; these data cover 73 percent of the sample. When the country of qualification could not be determined, data on country of birth were obtained from national censuses and registers in Australia, Austria, Belgium, Denmark, Ireland, and Sweden; these data cover 18 percent of the sample. For countries for which these data were not available (Italy, Germany, Portugal, and Switzerland), emigrants were defined according to their citizenship; these data cover 9 percent of the sample. 1 The data reveal that medical brain drain rates from 44 Sub-Saharan African countries to 16 OECD countries rose in most countries between 1991 and 2000 (figure 1). Only Angola, Benin, Burkina Faso, Chad, The Gambia, Ghana, Kenya, Malawi, Mauritania, Mozambique, Niger, Senegal, South Africa, and Uganda experienced declines in medical brain drain over this period. Comprehensive longitudinal data on HIV prevalence rates and the number of adult deaths due to AIDS were recently released by UNAIDS for 1990 2004 (UNAIDS 2006). They reveal skyrocketing levels of HIV prevalence rates in many countries between 1991 and 2000 (figure 2). Longitudinal information on government expenditures on health by Sub-Saharan African countries are available for 1996-2004 (WHO 2006). Because of missing observations for 1990-95, average government health expenditures during 1996-2004 are treated as time-invariant variables in the 1. Because the medical brain drain rate is treated as an endogenous variable in the estimation, alternative definitions of emigrants are not critical. Moreover, highlighting differences in the definitions across DEeD countries should promote a more unified approach in the future. Bbargava and Docquier 349 FIGURE 1. Medical Brain Drain Rates for Sub-Saharan Africa, 1991 and 2000 35 T'-~-------~-----'--~---'-------~----'-'--~--'-----~----------'---------.--~----- 111991 112000 30 25 Source: Docquier and Bhargava 2007. FIGURE 2. HIV Prevalence Rates in Sub-Saharan Africa, 1991 and 2000 Source: UNAIDS 2006. 350 THE WORLD BANK ECONOMIC REVIEW econometric modeling. The ILO (2005) provides data on physicians' wages expressed in terms of average physicians' wages in the United States (see also Vujicic and others 2004). Data on gross domestic product (GDP) (in 2000 dollars) are from the World Development Indicators (World Bank 2005). Additional information from UNESCO (2004) and the World Bank (2006) were used to reduce the numbers of missing observations. Estimates of the proportion of the labor force with secondary or tertiary education are from Barro and Lee (2001), Docquier and Marfouk (2006), and Cohen and Soto (2007). The variables in the database were averaged to create five three-year interval time points over 1990-2004. Alternative data sets were created using two- and four-year averages; the three-year average figures are used here because of the structure of the data and the stochastic properties of the variables (Bhargava 2001); interpolations of variables such as life expect ancy can create difficulties for econometric modeling. A steady increase in HIV prevalence rates and the number of adult deaths due to AIDS is evident from the sample means (table 1). The number of phys icians emigrating from Sub-Saharan African countries rose between 1990 and 2004, as did medical brain drain rates. Average life expectancy fell about two years between 1991 and 2003. There was an increase over time in net school enrollment rates in primary and secondary education. Physicians' wages in Sub-Saharan Africa declined slightly relative to physicians' wages in the United States. II. ANALYTICAL FRAMEWORK The relations between HIV prevalence rates, medical brain drain, number of adult deaths due to AIDS, and life expectancy estimated using country-level data are of interest to policymakers. The nature of HIV transmission through sexual intercourse and the lags between contracting HIV and the onset of AIDS have important implications for the specification of macroeconometric models that go beyond the usual difficulties of deducing the effects of disease prevalence rates on aggregate economic indicators. It is even more complex to explain HIV prevalence rates at the country level, where information on the average number of sex partners, the prevalence of sexually transmitted infections, and patterns of migrant labor are unavailable. The effects of HIV prevalence rates on medical brain drain can nevertheless be analyzed, while allowing for the possibility that the HIV prevalence rate is potentially an endogenous variable in the system and may be influenced by medical brain drain. Most rural residents in Sub-Saharan Africa have limited access to basic healthcare, and only a small proportion of people with HIV receive antiretro viral treatment. The lags between HIV infection and the onset of AIDS are thus likely to depend mainly on the natural rate of disease progression. That TABLE 1. Sample Means and Standard Deviations of Selected Variables in Sub-Saharan Africa, 1990-2004 (3-year averages) 1991" 1994 1997 2000 2003 Standard Standard Standard Standard Standard Variable Mean deviation Mean deviation Mean deviation Mean deviation Mean deviation HIV prevalence 2.98 3.72 5.037 5.592 6.594 7.047 7.073 7.907 7.085 8.096 ( percent) Number of AIDS 3,960.82 7,777.64 10,253.64 16,672.51 28,002.43 30,245.11 40,974.4 36,963.61 54,538.9 deaths Number of 0.15 0.24 0.15 0.22 0.16 0.24 0.16 0.24 0.16 0.24 physicians per population Number of 151.35 528.16 169.10 586.36 188.77 650.39 211.26 687.94 269.75 925.89 physicians emigrating Medical brain 0.09 0.11 0.10 0.12 0.10 0.11 0.10 0.11 0.11 0.12 drain b Life expectancy 50.79 8.21 50.13 8.28 49.36 7.75 48.59 7.88 48.21 8.56 (years) GDP per capita 752.84 1,134.56 749.14 1,179.60 818.13 1,292.17 876.92 1,400.12 906.38 1,415.34 (2000$) b:l GDP per capita 1.67 1.55 1.95 2.13 4.18 8.23 2.81 2.86 3.63 3.74 ::l ;:,. growth rate ~ ;:,. ~ ;:,. Population 11,536 16,747 12,511 18,310 13,515 19,799 14,539 21,223 15,581 ;:! (thousands) " t:J Literacy rate 50.76 19.73 53.73 19.72 56.76 19.62 59.70 19.46 62.57 19.21 0 (percent of adult " .tl. ;: population) 1>' .... (Continued) w Vt ...... TABLE 1. Continued Vol v. N 1991 a 1994 1997 2000 2003 -i Standard Standard Standard Standard Standard ::t Variable Mean deviation Mean deviation Mean deviation Mean deviation Mean deviation '" ~ 0 :<> Primary-school 58.11 25.01 60.20 25.41 61.00 23.31 62.57 20.98 67.00 19.71 r :::; enrollment (percent) '" :> z :><: Secondary-school 19.11 15.95 20.64 17.34 22.06 18.22 23.91 18.72 25.58 18.62 enrollment '" n 0 (percent) z 0 Percentage of 13.2 10.2 14.8 11.7 16.0 12.6 17.0 12.9 18.0 13.7 s:: population with n secondary or '" m In(physicians' wages in home 0.036** 0.019 -0.057** 0.030 -0.035** 0.019 0.066* * 0.032 z ~ country/physicians' wages in m ("l United States) o z In (percent school enrollment 0.115** 0.045 0.236** 0.069 0.124** 0.048 0.248*" 0.081 o secondary) :l: ("l In(GDP per capita) -0.035 0.033 -0.102** 0.057 0.037 0.034 -0.085 0.066 ~ In(HIV prevalence) 0.079** 0.017 0.082** 0.022 0.071** 0.021 0.107*" 0.036 m < Lagged dependent variable 0.921 ** 0.028 0.811 ** 0.053 0.914** 0.031 0.769*" 0.Q35 Between/within variance ratio 0.445 0.354 '" ~ Within variance 0.051 2 (maximized log-likelihood 512.75 461.79 570.55 300.45 function) Chi-squared test for random 50.96** effects degrees of Chi-squared test for exogeneity 12.04 .... of HIV prevalence rate (5 of freedom) Dependent variable: Logistic (medical brain drain rate). HSignificant at the 5 percent level. Note: Data on 39 countries, with five time observations at 3-year intervals. "HIV prevalence rate is treated as an endogenous variable. bSpecification uses 4-year averages at three time points. Source: Authors' estimation results. Bhargava and Docquier 359 majorities-68 percent in Cameroon, 77 percent in Zimbabwe, 78 percent in South Africa, 84 percent in Uganda, 85 percent in Ghana-report a desire to earn more as a motivation (Awases and others 2003). Third, net enrollment in secondary education is a positive and significant predictor of medical brain drain, with an estimated short-run elasticity of 0.12. This result is not surprising, as higher enrollments in secondary education entail greater expenditures on education; physicians educated in such environ ments are likely to have better emigration prospects. Fourth, the HIV prevalence rate is a significant predictor of medical brain drain, with the short-run elasticity (0.07) robust across the first three specifica tions. Moreover, the coefficient of the lagged dependent variable is estimated at 0.91 in specification 3, indicating that the long-run impacts of the explanatory variables are about 11 times greater than the short-run coefficients. Thus the long-run elasticity of the medical brain drain rate with respect to HIV preva lence is about 0.8. This means that a doubling of the HIV prevalence rate implies an 80 percent increase in the medical brain drain rate in the long run. This is a large effect, with important policy implications, especially given that the average number of physicians per 1,000 people is only 0.15 in Sub-Saharan Africa (see table 1). Furthermore, higher ratios of physicians to the population wi II be needed as more people with HIV develop AIDS. Fifth, the large estimated coefficients of the lagged dependent variable suggest that emigration patterns in Sub-Saharan African countries are becom ing well established, presumably as a result of stable demand from OECD countries for physicians trained in specific countries. While specification 2 is rejected in favor of specification 1 by the likelihood ratio test, the estimated between/within variance ratio is not significant in specification 2, possibly because of the relatively small number of countries in the sample. Sixth, the results from specification 4 employing four-year averages are similar to those from specification 1, indicating robustness of the results from three-year averages at five time points. In fact, use of four-year averages entails a loss of information, because the data from 2002, 2003, and 2004 cannot be used in specification 4. Finally, a model similar to equation 2 was estimated for HIV prevalence rates with medical brain drain rate as an explanatory variable to investigate possibJe reverse causality. Coefficients of medical brain drain rate were not sig nificant in any of the specifications, thereby supporting the model formulation and exogeneity assumptions. Effect on Numbers of Adult Deaths Due to AIDS and on Life Expectancy The results for the numbers of adult deaths due to AIDS are shown for four specifications (table 3). Specification 2 includes an interaction term between medical brain drain and HIV prevalence rates; this term is included because higher HIV prevalence rates can exacerbate the effects of medical brain drain on adult deaths caused by AIDS. In specification 3, the current HIV prevalence W 0'\ 0 -I :I: '" ~ 0 " r t:l '" :> z ~ TABLE 3. Maximum Likelihood Estimates for Number of Adult Deaths due to AIDS, 1990-2004 '" () 0 Specification 1a Specification za Specification 3 b Specification 4< z 0 ::; () Standard Standard Standard Standard Explanatory variable Coefficient error Coefficient error Coefficient error Coefficient error '" " < m Constant -2.833** 0.422 -3.412** 0.165 -5.011*' 0.203 3.553' · 0.481 ~ InC population) 0.450" 0.017 0.480** 0.006 0.635'" 0.008 In(proportion of population with -0.045 0.027 -0.005 0.026 0.024 0.016 -0.114 0.085 secondary or tertiary education) In(GDP per capita) -0.014 0.025 -0.072** 0.025 -0.008 0.006 -0.093' · 0.047 In(HIV prevalence) 0.542" 0.039 0.778" 0.Q38 0.660" 0.052 In(HIV prevalence)_1 0.836" 0.016 In(medical brain drain rate) 0.095' · 0.019 -0.059·' 0.023 0.080·' 0.020 In(number of physicians in home -0.007 0.173 country/1,OOO people) In(number of physicians abroad) 0.113" · 0.035 [In (number of physicians in home 0.001 0.028 countryl1,OOO people)f [ In(number of physicians abroad))2 0.014' · 0.006 In(medical brain drain) x In(HIV 0.049·' 0.008 prevalence) In(medical brain drain) x In(HIV 0.061" 0.005 prevalence) -1 In(HIV prevalence) x In {physicians -0.032.... 0.011 abroad} Lagged dependent variable 0.510H 0.012 0.498*" 0.001 0.357 .... 0.008 0.535 .... 0.Q15 2 x (maximized log-likelihood 1,243.35 1,258.20 1,282.88 835.39 function) Chi-squared test for exogeneity of 104.03** 116.63** 96.94*" medical brain drain and HIV prevalence rates (10 degrees of freedom) Dependent variable: In (number of adult deaths caused by AIDS). "* Significant at the 5 percent level. Note: Data on 39 countries with five time observations at 3-year intervals are used in the estimation. aMedical brain drain rate (ratio of number of physicians from Sub-Saharan Africa working in 16 OECD countries to the number of physicians working in Sub-Saharan and OECD countries) and HIV prevalence rates are treated as endogenous variables. bHIV prevalence rate is lagged one period. 0) or it does not (Sit 0). A Tobit model is formulated in terms of a latent variable model to determine the relation between FDI and the rate of product innovation: S7t = CtlFCit~l + Ct2FD1jt-l + Ct3F1Nit--l + Ct4Xit-l (1) + CtsFINit - 1*FDI jt - 1 + Ct6RDit-l *FDI;t-l + Dr + D j + D t + eit Sit= 0 if Sit ~ 0 Sit Sit if Sit> 0 where the dependent variable S is defined as the share of innovation output products involving the use of new process innovation or novel technology in total output. 3 This variable, which measures the output of the innovation process, is a more suitable measure than R&D, which is an input into the inno vation process (see Criscuolo, Haskel, and Slaughter 2005). The D variables in equation (1) are full sets of regional (r), industry (j), and time (t) dummy variables. X is a vector of firm-level determinants of innovation. It includes R&D intensity, the ratio of employee training expenditure to the total wage bill, export intensity, subsidies, age, and the firm's market share within its three digit industry. The choice of these firm-level covariates is guided by theoretical considerations as well as evidence from the empirical literature. R&D is an important input into the innovation process and thus is included in the model. Human capital is also an important determinant of innovation. One proxy for human capital is the amount of training provided by a firm, which is included in the empirical analysis. Criscuolo, Haskel, and Slaughter (2005) provide evi dence that firms that are active on export markets are more innovative than others. The model here captures this notion by controlling for firms' export intensities. Because subsidies can help firms engage more in innovation (see Gorg and Strobl 2007), a measure of the level of production-related grants is also included in the model. As Jefferson and others (2006) argue, the age of a firm may also be important in explaining innovation activity (as a proxy for a firm's experience) and hence the possibility for learning effects. Their approach is adopted here by including fum age in the equation. Aghion and others (2005) discuss the role of competition for innovation; Aitken and Harrison (1999) show that multinationals may affect the competitive landscape in the domestic economy, leading to an increase in competition for domestic firms. 3. Definitions of all variables, plus summary statistics, are provided in table 1, discussed in the next section. Girma, Gong, and Giirg 371 To take account of these findings, the model includes a firm's market share as an indicator of its competitive position. FIN is a measure of a firm's access to finance (measured by its ability to obtain loans from domestic banks). Financial constraints are a serious impedi ment to innovation activity (Hall 2002). This effect may be particularly pro nounced in China, where the financial sector is highly regulated and inefficient, and lending is skewed toward inefficient state-owned enterprises (Huang 2003). FC is a measure of foreign capital participation in firm i. It captures the central concern of this article-the impact of FDI on innovative activity in Chinese domestic firms. FC is included to allow for the fact that firms with some share of foreign capital may be more innovative than other firms, for the reasons discussed above. FDI is a vector of industry-region-specific FDI indices. It captures the potential spillover or crowding-out effects of FDI at the industry leveL The effect of FDI is allowed to vary based on a firm's R&D activity and access to finance by includ ing two interaction terms in the empirical estimation of equation (1), namely FDI and R&D intensity and FDI and FIN. The interaction of FDI and R&D intensity captures the notion that firms with higher absorptive capacity are better able to benefit from the technology transferred by incoming FDI.4 The interaction of FDI and FIN allows firms with better access to finance to benefit more from inward FDI; because they are less financially constrained, they may be better able to implement the new technology and less affected by reductions in the avail ability of domestic finance caused by demand for loans by foreign firms. All covariates in the empirical model are lagged by one period to mitigate potential endogeneity concerns. Nevertheless, some firm-level variables in the specification may be endogenous. One is R&D intensity, which is a major input into the product innovation process. The choice of this input is likely to be correlated with factors that determine the firm's decision to innovate. Similar arguments can be made regarding the potential endogeneity of the other firm-level variables. To deal with this possible problem, all lagged firm level variables except age are considered potentially endogenous. The instru mental variables technique for Tobit models developed by Smith and Blundell (1986) is used to estimate this model:' 4. See Girma (2005) for a discussion of the importance of absorptive capacity and an empirical ilIusrration using firm-level data for the United Kingdom. 5. The estimation of Tobit models with endogenous regressors involves two steps. The first is to genuate residual terms from linear regressions of each endogenous variable on the instrumental variables and all other exogenous regressors. The second is to estimate a standard Tobit model by including the residual terms from the first step in the list of covariates. The standard errors are boorstrarped to take account of the fact that residual terms are generated regressors. The residual terms are correction terms for the endogeneity problem; jointly statistically significant coefficients can be takell as evidence in favor of the hypothesis that instrumented variables are indeed endogenous. A one-step variant of this estimator involving stronger distributional assumptions is also available (Newey 1987). However, it fails to attain convergence in the data used here. This type of convergence problem IS frequently encountered when there is more than one endogenous regressor. 372 THE WORLD BANK ECONOMIC REVIEW Twice-lagged values of the potentially endogenous variables are used as instruments. The assumption is that conditional on the regressors, these vari ables are asymptotically uncorrelated with the error term of the model. Ultimately, of course, this is an empirical issue, tested using the Sargan Hansen test for the validity of instrumental variables. Additional instruments are also used. They include the share of state-owned enterprises in a region or industry, the share of loss-making state-owned enter prises in a region or industry, the level of regional financial development (bank loans to the private sector as a share of total loans), and whether the firm is politically affiliated with local, regional, or central governments. These instru ments are designed to account for the endogeneity of sector-level FDI and access to finance. The share of the state sector, for example, is a proxy for state dominance in the region or industry; to the extent that access to finance is different for state-dominated sectors and regions, this is a reasonable instru ment for firm-level access to finance. Similar arguments can be made for the share of loss-making state-owned enterprises and the level of regional financial development. A large number of enterprises in China are affiliated with some level of gov ernment administration. The function of the relevant government body is to offer credit guarantees and political protection to the affiliated firms. This poli tical affiliation variable is strongly related to firms' access to finance, because China's financial system remains dominated by the four large state banks. Different levels of political affiliation are used as instruments to reflect the rea listic assumption that the main effect of political affiliation on innovation comes through its effects on finance. Ultimately, however, the relevance of the instruments is an empirical issue that is tested for in the estimation below. II. DESCRIPTION OF THE DATABASE AND CONSTRUCTION OF VARIABLES The econometric analysis draws on confidential micro data that underlie the Annual Reports of Industrial Enterprise Statistics, compiled by the China National Bureau of Statistics. The reports cover all firms with annual turnover of more than 5 million yuan (about $600,000). The firms in the data set account for an estimated 85 - 90 percent of total output in most industries. The data set includes information on firm ownership structure, industry affiliation, geographic location, establishment year, employment, gross output, product innovation, R&D, value added, net fixed assets, exports, and employee training expenditures. 6 The whole sample (1.3 million observations from about 446,000 firms) is used to construct the variables of interest (such as the share of foreign firms in an industry or region or firms' market share). 6. Nominal values are deflated using industry-specific ex-factory price indices obtained ftom the China Statistical Yearbook 2006 (China National Bureau of Statistics 2007). Girma, Gong, and Gorg 373 The econometric work is confined to domestic-owned enterprises, the focus of this article. The China National Bureau of Statistics assigns a categorical variable to each firm in the database indicating its ownership status. It is also possible to con struct a continuous measure of ownership composition from the database by looking at the fraction of paid-in capital contributed by the state and by private (domestic and foreign) investors. This measure of ownership is used here. Firms are defined as state-owned, collectively owned, or privately owned based on majority ownership of the firm. The information necessary for the econometric estimation is available for 239,085 domestic firms (630,900 total observations). The data set provides information on the extent of foreign capital partici pation at the level of the firm. This makes it possible to calculate the share of foreign ownership in the domestic enterprise and identify the direct effects of FDI on domestic firms' innovative activity. A different method is used to esti mate the indirect (spillover) effect of FDI at the industry level. For each of the 171 three-digit industries and 31 provinces, the proportion of output accounted for by companies with foreign ownership in the industry and region is calcu lated. 7 Alternative measures of industry and region FDI are the proportion of new products accounted for by multinational companies (labeled FDI inno vation), and the share of domestic bank loans extended to foreign multina tionals (FDI loan). The data reveal no substantial relation between firm ownership on the one hand and innovation activity or the level of R&D on the other (table 1). As expected, on average state-owned enterprises receive higher shares of bank loans and larger subsidies from the government. They are less export intensive and receive more modest inflows of foreign capital than privately or collectively owned firms. The pattern of product innovation by state-owned enterprises across indus tries at the two-digit reveals three noteworthy points (table 2). First, the pro portion of innovating firms rose over time in most sectors. In contrast, the share of new product sales in total sales, while generally significant, declined slightly in most sectors. Second, labor-intensive sectors (such as food manufac turing and paper products) have the lowest proportion of innovators. In contrast, export-competing labor-intensive sectors (such as textiles) exhibit a 7. Officially, foreign-owned multinationals are defined as enterprises with at least a 25 percent share of foreign capital. Domestically owned enterprises that have foreign capital participation of less than 25 percent are not considered in this definition. The richness of the data set is exploited by weighing the output of firms with foreign capital by the extent of their foreign participation, measured by the share of foreign capital at the firm level. Under this definition of sectoral FOr, firms classified as domestic but that have some foreign capital also contribute (proportionally) to the aggregate output of the foreign sector. The recent literature on productivity spillovers from For notes that domestic firms may benefit not only from horizontal but also from vertical spillovers through customer-supplier linkages (see Javorcik 2004). Vertical measures (backward and forward spillovers) were calculated by the authors but found not to be consistently statistically significant. They are therefore not included in the analysis that follows. v.> TABLE 1. Variable Definitions and Summary Statistics '-J ~ State-owned Collective enterprise Private enterprise enterprise ..., :r: Standard Standard Standard '" ~ 0 Variable Definition Mean deviation Mean deviation Mean deviation i" t""' 0 Product Share of output involving new process or product innovation 0.041 0.150 0.034 0.151 0.021 0.116 innovation '" ;.. z ~ Restricted sample of firms with nonzero product innovation 0.319 0.295 0.390 0.350 0.369 0.331 R&D R&D expenditure divided by sales 0.002 0.021 0.002 0.013 0.001 0.007 '" ('l 0 Restricted sample of firms with nonzero R&D expenditure 0.013 0.046 0.012 0.337 0.008 0.D18 z 0 Labor training Employee expenditure per employee 0.007 0.027 0.008 0.035 0.008 0.037 ;:: Restricted sample of firms with nonzero labor training outlay 0.015 0.037 0.021 0.053 0.037 0.053 ('l i" Export intensity Share of exports in total sales 0.043 0.164 0.127 0.298 0.110 0.281 Restricted sample of exporters 0.307 0.328 0.580 0.380 0.594 0.374 '" < Market shares Firm's share of sales in total three-digit industry region sales 0.044 0.134 0.022 0.082 0.021 0.073 '" ~ Domestic finance Domestic bank loans normalized by total assets 1.806 2.519 0.825 1.874 0.876 1.886 Subsidy Log of production subsidy from local and central 0.983 2.335 0.642 1.843 0.796 2.052 governments Age Log of years since establishment 3.147 0.914 1.848 0.935 2.569 0.806 Foreign capital Share of foreign multinationals capital in firm's total capital 0.002 0.033 0.004 0.050 0.006 0.060 FDI Share of foreign multinationals' sales in three-digit 0.143 0.181 0.194 0.191 0.175 0.182 industry-region total sales FDI innovation Share of multinationals' innovative output in three-digit 0.097 0.196 0.140 0.212 0.127 0.206 total innovation FDI loan Share of multinationals' domestic bank loans over total 0.099 0.161 0.168 0.198 0.148 0.184 domestic bank loans Number of firms 239,085 34,549 148,694 55,842 Number of 630,900 125,357 316,461 189,082 observations Source: Authors' analysis based on data from China National Bureau of Statistics. Girma, Gong, and Gorg 375 TABLE 2. Sectoral and Temporal Pattern of Product Innovation by State-Owned Enterprises (percent) Share of New product sales as innovators share of total sales Two-digit industry classification 1999 2005 1999 2005 13: Food processing 2.0 10.1 32.3 16.6 14: Food production 4.3 11.6 29.2 23.9 15: Beverage industry 6.0 12.0 27.2 25.1 16: Tobacco processing 12.3 21.1 14.9 15.2 17: Textile Industry 17.3 17.2 30.7 29.6 18: Garments and other fiber products 3.5 6.5 45.0 45.3 19: Leather, furs, down and related products 4.1 8.1 49.4 39.7 20: Timber processing 2.8 6.8 46.2 23.0 21: Furniture manufacturing 4.2 10.0 36.0 21.4 22: Papermaking and paper products 4.0 7.2 37.1 19.0 23; Printing and record medium reproduction 1.8 5.9 37.5 35.0 24: Cultural, educational, and sports goods 9.4 9.2 33.5 38.9 25: Petroleum refining and coking 5.0 6.4 28.9 20.9 26: Raw chemical materials and chemical products 9.2 10.7 31.3 33.2 27: Medical and pharmaceutical products 24 25.2 35.8 37.2 28: Chemical fiber 14.0 10.4 26.7 39.4 29: Rubber products 12 9.8 32.0 30.5 30; Plastic products 9.1 10.2 38.2 33.9 31: Nonmetal mineral products 3.7 10.7 38.1 23.0 32; Smelting and pressing of ferrous metals 5.8 6.9 29.6 24.8 33; Smelting and pressing of nonferrous metals 6.0 9.7 32.9 335 34; Metal products 6.1 7.9 33.4 31.1 35: Ordinary machinery 14.2 13.2 29.5 320 36: Special purpose equipment 17.8 17.2 34.8 37.3 37: Transport equipment 14.1 15.5 35.5 34.7 39: Other electronic equipment 14.8 14.0 36.1 41.8 40: Electric equipment and machinery 26.8 23.2 47.6 53.3 41: Electronic and telecommunications 25.7 25.7 35.3 46.0 42: Instruments and meters 5.7 7.0 39.2 33.0 '\our,;e: Authors' analysis based on data from China National Bureau of Statistics. relatively large number of innovators. Third, the intensity of product inno vation is remarkably similar across labor-intensive, capital-intensive, and technology-intensive sectors. III. DISCUSSION OF THE RESULTS The benchmark Tobit model controls for firm heterogeneity by allowing for firm random effects (table 3, column 1). The model also includes two additional dummy variables for private and collectively owned firms. The estimation shows that R&D intensity exerts a positive and significant influence on the rate of product innovation. This is as expected, given that 376 THE WORLD BANK ECONOMIC REVIEW TABLE 3. Innovation Spillovers from FDI and Access to Finance: Results from Alternative Estimators (2) Tobit (3) Linear generalized (1) Random effects instrumental method of moment Variable Tobit model variables model model R&D 2.312""· (30.4) 4.323"· (19.0) 2.118*** (10.8) Labor training 0.48P"" (11.5) 0.862*"" (7.22) 0.700·"" (3.65) Export intensity 0.252 ..... (39.3) 0.238*** (24.4) 0.217"" (18.5) Market share 0.576 .... · (41.6) 0.600""" (29.7) 0.901 ..... (21.0) Finance 0.0413**" (46.9) 0.0696 ...... (35.6) 0.0684 ...... (23.8) Subsidy 0.0271 ...... (38.8) 0.0313 ...... (24.7) 0.0399 ...... (18.3) Age 0.0505·"" (27.4) 0.0487 ...... (19.2) 0.00197 ...... (6.69) Foreign capital 0.135 ...... (6.85) 0.217 ...... (5.54) 0.168· .... (3.65) J... C) 0; ~ IN ;:cl 380 THE WORLD BANK ECONOMIC REVIEW Several striking differences are apparent across the three ownership types. First, the relation between access to finance and innovation is strongest among private and collectively owned firms, which receive less favorable treatment from China's financial system than state-owned enterprises do. Second, the coefficient on foreign capital is largest for state enterprises, suggesting that injections of foreign capital are associated with the highest positive impact on innovation for this type of firm. This may reflect the fact these state-owned firms are more inefficient than other firms and therefore offer the greatest opportunities for improvement as a result of the influx of foreign capital. 10 Third, and perhaps most striking, the interaction term of FDI and access to finance is positive for private and collectively owned firms but statistically insignificant for state-owned enterprises. Access to domestic finance plays no role in generating spillovers to state-owned enterprises, which are largely ineffi cient but enjoy preferential access to domestic financial resources. Profit-making firms can be distinguished from loss-making firms, most of which are owned by the state (columns 4 and 5 in table 4). The results are in line with expectations: access to finance has no effect on innovation in loss making enterprises, and also does not matter for indirect effects from sector level inward FDIY Sector-level FDI can affect domestic innovation by transferring technology to or creating credit opportunities for domestic firms. The next step of the analysis tries to distinguish these two channels by calculating two different FDI measures. The first measure is aggregate innovation by foreign multinationals, calculated as innovation output by foreign multinationals in a sector or region divided by total innovation output. The second measure is aggregate borrowing by foreign multinationals, calculated as the share of domestic bank loans in total bank loans in the sector or region. The results in columns 1-3 of table 5 show that the effects of the two vari ables are broadly similar to those for private and collectively owned firms. FDI has a positive effect only if the firm is active in R&D and has access to bank loans. State-owned enterprises that invest in their own R&D also benefit more from technology transfer by multinationals than those that do not, but the firms' financial position does not mitigate the effect of FDI technology. The effect of FDI on credit opportunities has no statistically significant relation with state enterprises' ability to innovate. This result suggests that preferential access to domestic financial resources means that finance is not a constraint for state enterprises. In alternative estimations in columns 4 and 5, the data are broken into loss making and profit-making enterprises. Results for loss-makers closely resemble 10. This result is in line with the work by Bartel and Harrison (2005) that shows state-owned enterprises in Indonesia benefit greatly from foreign ownership in the enterprise. 11. As a robustness check, table 4 was reestimated using the endogenous Tobit estimator. The results are very similar and hence not reported here. Girma, Gong, and Gorg 381 those for state-owned enterprises, suggesting that access to finance has little effect on innovation for such firms. 12 IV. CONCLUSIONS The econometric analysis conducted here shows that access to finance is an important issue for firms' innovation activity and their ability to benefit from inward FDI. This is mainly the case for private and collectively owned firms, however. It is far less important for state-owned firms, which receive preferen tial treatment under the current domestic financial system. Firms with foreign capital participation and those with good access to domestic bank loans-that is, firms with less binding financial constraints innovate more than others. Inward FDI at the sectoral level is positively associ ated with domestic innovative activity only if firms engage in their own R&D activities (that is, have some absorptive capacity) or have good access to dom estic finance. This finding points to the possible adverse effect of domestic credit constraints on firms' ability to benefit from inward FDI. Grouping firms by ownership type reveals that access to finance plays a role only among firms that are not state owned. Although state-owned enterprises are largely ineffi cient, they enjoy preferential access to domestic financial resources; access to finance thus provides no bottleneck for them. Sector-level inward FDI has two effects. It transfers technology and may increase domestic credit opportunities. The effect on credit is of very little sig nificance for state-owned enterprises and is independent of their access to finance. In contrast, it is an important channel through which FDI affects the innovation of domestic private and collectively owned enterprises. REFERENCES Aghion, Philippe, Nicholas Bloom, Richard Blundell, Rachel Griffith, and l'eter Howitt. 2005. "Competition and Innovation: An Inverted-U Relationship." Quarterly Journal of Economics 120(2):701-28. Aitken, Brian J., and Ann E. Harrison. 1999. "Do Domestic Firms Benefit from Direct Foreign Investment? Evidence from Venezuela." American Economic Review 89(3):605-18. Amiri, Mary, and Beata S. Javorcik. 2008. "Trade Costs and Location of Foreign Firms in China." J,Jurn.:t1 of Development Economics 85(1-2):129-49. Bartel, Ann P., and Ann E. Harrison. 2005. "Ownership versus Environment: Disentangling the Sources of Public-Sector Inefficiency." Review of Economics and Statistics 87(1):135-47. Berthelemy, Jean-Claude, and Sylvie Demurger. 2000. "Foreign Direct Investment and Economic Growth: Theory and Application to China." Review of Development Economics 4(2):140-55. China National Bureau of Statistics. 2006. Annual Reports of Industrial Enterprise Statistics. Beijing. 12. One other robustness check was conduced, which included squared terms of foreign capital, FDI innovation, and FDI loans in the estimations. The results, which are not reported here, are robust to this slight modification of the estimating equation. 382 THE WORLD BANK ECONOMIC REVIEW - - - . 2007. China Statistical Yearbook 2006. Beijing: China Statistics Press. Criscuolo, Chiara, Jonathan E. Haske!, and Matthew J. Slaughter. 2005. Global Engagement and the Innovation Activities of Firms. NBER Working Paper 11479. Cambridge, Mass.: National Bureau of Economic Research. Girma, Sourafel. 2005. "Absorptive Capacity and Productivity Spillovers from FDI: A Threshold Regression Analysis." Oxford Bulletin of Economics and Statistics 67(3):281-306. Girma, Sourafel, and Yundan Gong. Forthcoming. "FDI, Linkages and the Efficiency of State-Owned Enterprises in China." Journal of Development Studies. Gorg, Holger, and David Greenaway. 2004. "Much Ado about Nothing? Do Domestic Firms Really Benefit from Foreign Direct Investment?" World Bank Research Observer 19(2):171-97. Gorg, Holger, and Eric Strobl. 2007. "The Effect of R&D Subsidies on Private R&D." Economica 74(294):215-34. Hall, Bronwyn H. 2002. "The Financing of Research and Development." Oxford Review of Economic Policy 18(1):35-51. Harrison, Ann E., and Margaret S. McMillan. 2003. "Does Direct Foreign Investment Affect Domestic Credit Constraints?" Journal of International Economics 61 (1):73-1 00. Hu, Albert G. Z., Gary H. Jefferson, and Jinchang Qian. 2005. "R&D and Technology Transfer: Firm Level Evidence from Chinese Industry." Review of Economics and Statistics 87(4 ):780-86. Huang, Yasheng. 2003. Selling China: Foreign Direct Investment during the Reform Era. New York: Cambridge University Press. Javorcik, Beata S. 2004. "Does Foreign Direct Investment Increase the Productivity of Domestic Firms? In Search of Spillovers through Backward Linkages." American Economic Review 94(3):605-27. Jefferson, Gary H., Bai Huamao, Guan Xiaojing, and Yu Xiaoyun. 2006. "R&D Performance in Chinese Industry." Economics of Innovation and New Technology 15(4-5):345-36. Keller, Wolfgang. 2004. "International Technology Diffusion." Journal of Economic Literature 42(3):752-82. Lemoine, Franc;:oise. 2000. "FDI and the Opening Up of China's Economy." CEPII Working Paper 2000-1l. Centres d'Etudes Prospectives et d'Informations Internationales, Paris. Markusen, James R. 2002. Multinational Firms and the Theory of International Trade. Cambridge, Mass.: MIT Press. Newey, Whitney K. 1987. "Efficient Estimation of Limited Dependent Variables Models with Endogenous Explanatory Variables." Journal of Econometrics 36(3):231-50. OECD (Organisation for Economic Co-operation and Development). 2004. Trends and Recent Developments in Foreign Direct Investment. Directorate for Financial, Fiscal and Enterprise Affairs, Paris. Pack, Howard. 2006. "Econometric vs. Case Study Approaches to Technology Transfer." In Bernard M. Hoekman, and Beata S. Javorcik, cds., Global Integration and Technology Transfer. London, United Kingdom: Palgrave Macmillan and World Bank. Redding, Stephen. 1996. "The Low-Skill, Low-Quality Trap: Strategic Complementarities between Human Capital and R&D." Economic Journal106(March):458-70. Smith, Richard J., and Richard W. Blundell. 1986. "An Exogeneity Test for a Simultaneous Equation Tobit Model with an Application to Labor Supply." Econometrica 54(4):679-86. Wei, Yingqi. 2003. "Foreign Direct Investment in China." Working Paper 2003/002. Lancaster University Management School, United Kingdom. ACCELERATED ONLINE PUBLICATION READ PAPERS FROM THIS JO.YFiNAl. ONl.INE. WEEKS IN ADVANCE OF SEEING THEM IN PRINT Th.e advantage f:o:rreaders: Rt;ad the very latest research omine aheaci()f pa.pers appearing i1); a journal issue. The advantage for authors: .See your paper receiving ci~tions faster th,anever t.efore. Visit OxfordJournars website at www.oxfordjournals.org, select this journal from. the list to reach its homepage,and click on Advance Access. Bookmark the page and regularly check for the latest accepted papers. www.oxfordjournals.org STAY ALERT TO THE LATEST CONTENT FROM OXFORD '"'OURNALS.... FOR FREE! WE WILL AUTOMATICALLY EMAIL YOU THE FORTHCOMING TABLES OF CONTENTS FOR NEW ISSUES OF JOURNALS PUBLISHED BY OXFORD JOURNALS Go to this journal's homepage and click on the link for .email table of contents' under ~erting Services'. www.oxfordjournals.org /;~~~ "'I ;::'~i' ~:;;;gf~; OXFORD ,JOURNALS '- ,.r". (~ ~ ~MC - Forthcoming papers in THE WORLD BANK ECONOMIC REVIEW A SYMPOSIUM ON ACCESS TO FINANCE · Systemic Risk, Dollarization, and Interest Rates in Emerging Markets: A Panel-Based Approach Edmar L. Bacha, Mdrcio Holland, and Fernando M. Gonfalves · Qyantitative Approaches to Fiscal Sustainability Analysis: A Case study ofTurkey since the crisis of 2001 Nina Budina and Sweder van Wijnbergen · Mental Health Patterns and Consequences: Results from Survey Data in Five Developing Countries Jishnu Das, Quy- Toan Do, Jed Friedman, and David McKenzie · Psychological Health Before, During, and After an Economic Crisis: Results from Indonesia, 1993-2000 Jed Friedman and Duncan Thomas ISBN-13: 978-0-19-954709-8 9 780199547098