TIIF WORLD BANK ECONOMIC REVIEW EDITOR Jaime de Melo, University of7Geneva ASSISTANT T O T H E EDITOR Marja Kuiper EDITORIAL BOARD Chong-En Bai, Tsinghua University, China Jan Willem Gunning, Free University, Jean-Marie Baland, University ofNamur, The Netherlands Belgium Hanan Jacoby, WorldBank Kaushik Basu, Cornell Univers~ty,USA Graciela Kaminsky, George Washington Alok Bhargava, Houston University, USA University, USA Fran~oisBourguignon, WorldBank Peter Lanjouw, World Bank Kenneth Chomitz, WorldBank Thierry Magnac, Universitidi. ToulouseI, Maureen Cropper, University ofMaryland, France USA Jonathan Morduch, New York University, Jishnu Das, World Bank USA Klaus Deininger, WorldBank Juan-Pablo Nicolini, Universidad Torcuato Asli Demirgu5-Kunt, WorldBnnk di Tella,Argentina Stefian Dercon, University of Oxford, UK Boris Pleskovic, WorldBank Ishac Diwan, WorldBank Martin Rama, WorldBank Augustin Kwasi Fosu, WorldInstitutefor Ritva Reinikka, WorldBank Development Economics Research Elisabeth Sadoulet, University of Calfornia, (WDER), Helsinki, Finland Berkeley, USA Alan Harold Gelb, WorldBank Joseph Stiglitz, Columbia University, USA Paul Gertler, WorldBank Jonathan Temple, University ofBristol, UK Indermit Gill, WorldBank L. Alan Winters, WorldBank 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 ..,.,I.,,;, ,,,Lee;,,:,,. ,,,l:,-., relo.l.l,re Anern+;Anql qr.-\or+r ,,-r\nr\m;,-c r.>tL-r thon THE WORLD BANK ECONOMIC REVIEW Volume 21 -2007-Number 2 Symposium in the Memory of Riccardo Faini Migration, Remittances, and the Brain Drain Migration, Remittances, and the Brain Drain: a Symposium in Memory of Riccardo Faini-an Introduction 173 Jaime de Melo Remittances and the Brain Drain: Do More Skilled Migrants Remit More? 177 Riccardo Faini Brain Drain in Developing Countries 193 Fre'de'ric Docquier, Olivier Lohest, and Abdeslam Marfouk Are Remittances Insurance? Evidence from Rainfall Shocks in the Philippines 219 Dean Yang and HwaJung Choi A New Data Base Measuring International Skilled Migration: A New Database Controlling for Age of Entry 249 Michel Reine, Fre'de'ric Docquier, and Hillel Rapoport The Anarchy of Numbers: Aid, Development, and Cross-Country Empirics David Roodman Incremental Reform and Distortions in China's Product and Factor Markets 279 Xiaobo Zhang and Kong-Yam Tan Child Labor, School Attendance, and Intrahousehold Gender Bias in Brazil 301 Patrick M. Emerson and Andre' Portela Souza Tracking Poverty Over Time in the Absence of Comparable Consumption Data 31 7 David Stifel and Luc Christiaensen 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 21, 3 Issues, 2007): Academic libraries-Print edition and site-wide online access: US$158/f99/149, Print edition only: US$150/f94/141, Site-wide online access only: US$150/f94/141; Corporate-Print edition and site-wide online access: US$237/f148/222, Print edition only: US$226/f141/212, Site-wide online access only: US$226/f141/212; Personal-Print edition and individual online access: US$56/f35/53. Please note: f Sterling rates apply in the UK, in Europe, and US$ elsewhere. There may be other subscription rates available; for a complete listing, please visit www.wber.oxfordjournals.org/subscriptions Readers with mailing addresses in non-OECD countries and in socialist economies in transition are eligible to receive complimentary subscriptions on request by writing to the UK address below. Full prepayment in the correct currency is required for all orders. Orders are regarded as firm, and payments are not refundable. Subscriptions are accepted and entered on a complete volume basis. Claims cannot be 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. C0N.rAC.r INFORMATION: Journals Customer Service Department, Oxford University Press, Great Clarendon Street, OxfordOX2 6DP, UK. E-mail: jnls.cust.serv@oxfordjournals.org. Tel: +44 (0)1865 353907. Fax: +44 (0)1865 353485. In the Americas, please contact: Journals Customer Service Department, Oxford University Press, 2001 Evans Road, Cary, NC 27513, USA. E-mail: jnlorders@oxfordjournals.org. Tel: (800) 852-7323 (toll-free in USAICanada) or (919) 677-0977. Fax: (919) 677-1714. In Japan, please contact: Journals Customer Service Department, Oxford University Press, 1-1-17-SF, Mukogaoka, Bunkyo-ku, Tokyo, 113-0023, Japan. E-mail: oupjnl@po.iijnet.or.jp. Tel: (03)3813 1461. Fax: (03)3818 1522. POSTAL INFORMATION: The World Bank Economic Review (ISSN 0258-6770) is published by Oxford University Press for the International Bank for Reconstruction and DevelopmentlTHE WORLD BANK. Send address changes to The World Bank Economic Reviezo, Journals Customer Service Department, Oxford University Press, 2001 Evans Road, Cary, NC 27513-2009. Communications regarding original articles and editorial management should be addressed to The Editor, The World Bank Economic Review, The World Bank, 3, Chemin Louis Dunant, CP66 1211 Geneva 20, Switzerland. DIGITAL OBJECT IDENTIFIERS: For information on dois and to resolve them, please visit www.doi.org. PERMISSIONS: For information on how to request permissions to reproduce articles or information from this journal, please visit www.oxfordjournals.org/jnls/permissions. ADVERTISING: Inquiries about advertising should be sent to Helen Pearson, Oxford Journals Advertising, PO Box 347, Abingdon OX14 lGJ, UK. E-mail: helen@oxfordads.com. Tel: +44 (0)1235201904. Fax: t 4 4 (0)8704296864. 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 De~el~pment/THE WORLD BANK or Oxford University Press. Neither Oxford University Press nor the International Bank for Reconstruction and Development/TH~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 239.48-1984 (Permanence of Paper). INDEXING AND ABsmAsrlNG: The World Bank Economic Review is indexed andlor abstracted by CAB Abstracts, Current ContentslSocial and Behavioral Sciences, Journal of Economic LiteraturelEconLit, PAIS International, RePEc (Research in Economic Papers), and Social Services Citation Index. COPYRIGHT T; The International Bank for Reconstruction and Development/TH~WORLD BANK 2007 All rights reserved; no part of this publication may be reproduced, stored 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 W1P 9HE, or in the USA by the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923. Migration, Remittances, and the Brain Drain: a Symposium in Memory of Riccardo Faini-an Introduction Jaime de Melo JEL codes: F02, F22 It seems inevitable that international migration will be one of the major chal- lenges of the twenty-first century. Lower transaction and communication costs have already greatly eased the formation of migrant networks and reduced migration costs, for long a deterrent to migration from developing countries to developed countries. Demographics also keep migratory policies near the top of policy agendas in developed countries, where dependency ratios continue to rise. Migratory pressures will also increase as falling dependency ratios in developing countries contribute to the swelling of their labor forces. Designing immigration policies in industrial countries will be a challenge, especially in the European Union, where negative attitudes toward immigrants from poor, nonEU countries persist.' Even if survey results suggest a slightly more favorable attitude toward immigrants from countries that have recently joined the European Union, facing up to renewed migratory pressures is likely to remain a challenge. To see why, suppose that migratory decisions depend on wage differentials between destination and source regions and on amenities in each region, but that citizens also have a home-country bias in their locational preference and that amenities enter the utility function as a normal good. Then, unless citizens in source countries face a severe liquidity constraint that prevents them from Jaime de Melo is a professor of economics at the University of Geneva; his email address is demelo@ecopo.unige.ch. The author thanks Ckline CarrPre, Florence Miguet, Tobias Muller, and Andri. Sapir for their comments. 1. As reported by the European Social Survey in 2002 and 2004, about half those surveyed answered "a few" or "none" to the question of how many immigrants they wanted to see come and live in their country (Jowell and Central Co-ordinating Team 2005). The survey also indicated that immigrants coming from rich EU countries are in general much more welcome than immigrants from poor, non-EU countries. rm WORLD BANK ECONO~Y<: RWIEW, VOI.. 21, NO. 2,pp. 173-1 76 doi:lO. 1093/wber/lhm009 Advance Access Publication 18 June 2007 ((? The Author 2007. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE wo~1.1)BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org covering migration costs, with amenities a normal good in the utility function an all-around wage increase that preserves the wage differential between locations will reduce the propensity to migrate. This is so because increasing income for non-liquidity-constrained citizens will increase their demand for local amenities, thereby reducing the propensity to emigrate. In their study of the migration experience of Southern Europeans, Faini and Venturini (1994) found supportive evidence of an inverted U-shaped relation between migration rates and income. This pattern may repeat itself soon following the recent EU enlargement, leaving the European Union and other destination countries that have integrated with "close" neighbors grappling with how to craft immigra- tion policies for "more distant" countries. For source countries inward remittance flows in 2004, estimated at $150 billion (Ratha 2005), largely surpass net flows of official development assistance, at $79.5 billion (World Bank 2006). Remittances from migrants to developed countries should then help to close the developing-country- developed-country divide. Yet even if it is generally the case that remittances help reduce poverty, it is not clear how beneficial remittances are in the long run for physical and human capital formation. Money is fungible, and it is hard to disentangle causality in any positive correlation between migration and remittances on the one hand and investment on the other. Neither is it clear that migration is beneficial to growth because of the brain drain effects associated with the migration of high-skill workers. Recent and ongoing research at the World Bank and elsewhere (some of it presented in Ozden and Schiff 2005) on the effects of migration on source countries and on development in general is starting to fill this gap, aided by new data from household surveys with specific modules on migration and by the construction of detailed databases on the stock of skilled migrants. This symposium issue of the World Bank Economic Review is dedicated to the memory of Riccardo Faini. Riccardo had deeply felt interests in policies that he believed would help make the world a better place for all. He wrote extensively on migration policies, adducing evidence that restrictive trade policies may have increased migration pressures in the past and that supportive aid policies, while benefiting developing countries, may have reduced migration pressures less than expected, especially in relatively poor countries (Faini and Venturini 1993). His enthusiasm led him to play a catalytic role in the policy debate on migration policies. He convened several conferences on the topic, including one that resulted in a volume covering many of the issues currently under debate (Faini, de Melo, and Zimmermann 1999). His continuing interest in the subject is evident from Riccardo's comments on several articles in this issue. With colleagues, Riccardo contributed to a better understanding of both the determinants of the decision to migrate and the implications of increased labor mobility on source and destination countries. He was among the first to formally discuss the link between risk aversion and migration in the presence of financial market imperfections. He also produced evidence of a significant correlation between the location choices of migrants and the necessity of "spatially diversify- ing" income at the household level (Daveri and Faini 1999). In the contribution to this symposium (Faini, 2007), Riccardo questions the conventional wisdom that the emigration of high-skill workers leads to rela- tively higher flows of remittances. Taking inspiration from data showing that skilled migrants stay longer in European countries and have a higher propensity to reunite with close family members in the destination country, he posits an altruistic model in which migrants care more about close relatives than about others to show that a higher skill content of migration will not necessarily bring about an increase in remittances, because of composition effects as reuni- fication with close relatives takes place. As is often the case when testing a micro model with aggregate data, the results are not as sharp as one would wish, but the correlation between the share of skilled migrants and remittance flows is negative and the econometric results suggest that the reunification effect leading to lower remittances might be stronger than the wage effect working in the opposite direction. In "Are Remittances Insurance?" Yang and Choi (2007) give new evidence of the crucial role of remittances in poor countries. High-income volatility in rural areas combined with the near-absence of insurance markets requires households to find ways to cope with aggregate and household-specific risk. One way is to rely on remittances from overseas. Using panel data for the Philippines, Yang and Choi resolve the problem that income and remittances are jointly determined by exploiting the exogeneity of rainfall shocks that affect household incomes. They estimate that 60 percent of the shortfall in income resulting from aggregate income shocks is compensated for by remit- tances from overseas. Their econometric strategy includes accurate checks con- firming the validity of weather events as instruments for changes in domestic income. Rainfall shocks are shown to be uncorrelated with changes in domestic labor supply or with heterogeneous exchange rates shocks experienced by households during the 1997 Asian financial crisis. These results provide fresh justification for policies facilitating international migration and easing remittance flows. Further research is needed to understand whether remittances are also an effective strategy for coping with idiosyncratic income shocks. On a less optimistic note, the contributions by Docquier, Lohest, and Marfouk (2007) and by Beine, Docquier, and Rapoport (2007) provide evi- dence that small, poor, and often isolated countries are subject to important brain-drain effects, adding to their already high vulnerability. Even after modi- fying the common definition of brain drain (migrants with a post-secondary education) to include only skilled emigrants over the age of 22 when they left their country (so as to not attribute skill formation to the source country that is acquired in the destination country), Beine, Docquier, and Rapoport find staggeringly high emigration rates of skilled labor-reaching 80 percent in the small island economies of the Caribbean and around 50 percent for several countries in sub-Saharan Africa. Moreover, in their detailed analysis of bilateral flows of skilled migrants to Organisation for Economic Co-operation and Development countries, Docquier, Lohest, and Marfouk find that an increase in the education level of natives generates a less than proportional increase in the education level of emi- grants. Their estimates indicate that the lower the proportion of educated people in source countries, the higher the brain drain; brain drain rates also increase with the degree of political instability and of fractionalization in source countries. These findings suggest that in sub-Saharan Africa, where schooling rates are relatively low and the prevalence of political instability and of religious and ethnic cleavages is high, policies aimed at increasing education and improving the sociopolitical environment are likely to reduce brain drain. Combining more accurate estimates of brain drain with estimates of sub- sidies to higher education in source countries should help in devising more appropriate policies for higher education in source countries. More reliable estimates should also help in devising burden-sharing policies between desti- nation and source countries. May this symposium contribute to the design of the "win-win" migration policies that Riccardo hoped his work would help to build! Beine, Michael, Frdric Docquier, and Hillel Rapoport. 2007. "Measuring International Skilled Migration: A New Database Controlling for Age of Entry." World Bank Economic Review 21(2); doi:10.1093/wber/lhm007. Daveri, Franscesco, and Riccardo Faini. 1997. "Where Do Migrants Go?" Oxford Economic Papers 51(4):595-622. Docquier, Frdric, Olivier Lohest, and A. Marfouk. 2007. "Brain Drain in Developing Countries." World Bank Econonzic Review 2l(2);doi:lO. 1093/wberllhm008. Faini, Riccardo. 2007. "Remittances and the Brain Drain: Do More Skilled Migrants Remit More?" World Bank Economic Review 21(2);doi:lO.l093/wber/lhm006. Faini, Riccardo, and Alessandra Venturini. (1993). "Trade, Aid, and Migrations: Some Basic Policy Issues." European Economic Review 37(2-3):435-42. -. 1994. "Migration and Growth: The Experience of Southern Europe." CEPR Discussion Paper 964. London: Centre for Economic Policy Research. Faini, Riccardo, Jaime de Melo, and Klaus Zimmermann, eds. 1999. Migration: The Cotztroversiesand the Evidence. Cambridge, U.K.: Cambridge University Press. Jowell, R., and the Central Co-ordinating Team. 2005. Europeatt Social Survey 2004/2005: Technical Report. London: City University, Centre for Comparative Social Surveys. Ozden, Caglar, and Maurice Schiff, eds. 2005. lnterrzational Migration, Remittances, and the Brain Drain. New York: World Bank and Palgrave. Ratha, Dilip. 2005. "Workers' Remittances: An Important and Stable Source of External Development Finance." In Samuel Mairnbo, and Dilip Ratha, eds., Remittances: Development Inzpact and Future Prospecrs. Washington, D.C.: World Bank. World Bank. 2006. World Development indicators 2006. Washington, D.C.: World Bank. Yang, Dean, and H. Choi. 2007. "Are Remittances Insurance? Evidence from Rainfall Shocks in the Philippines." World Bank Economic Review 21(2);doi:lO.l093/wber/lhm003. Remittances and the Brain Drain: Do More Skilled Migrants Remit More? Riccardo Faini In most destination countries, immigration policies are tilted more and more in favor of skilled individuals. Whether this shift hurts economic prospects in sending countries, as argued by the traditional brain drain literature, is somewhat controver- sial. The most recent literature focuses on the link between skilled outmigration and educational achievements in the home country. This article emphasizes a different channel. It considers the argument that skilled migrants raise economic welfare at home by sending a relatively larger flow of remittances. While skilled migrants typi- cally earn more, and so might be expected to remit more, they are also likely to spend more time abroad and to reunite with their close family in the host country. These second two factors should be associated with a smaller propensity to remit. Thus, the sign of the impact of the brain drain on total remittances is an empirical question. A simple model has been developed showing that skilled migrants may indeed have a lower propensity to remit from a given flow of earnings. An empirical equation of remittances is estimated as a measure of the brain drain in developing countries using the Docquier and Marfouk (2004) data set. Evidence is found that the brain drain is associated with a smaller propensity to remit. JEI, codes: F02, F22 Immigration policies in receiving countries are increasingly tilted in favor of skilled migrants (Beine, Docquier, and Rapoport 2003; OECD 2003). This trend is raising considerable concern among policymakers in developing countries, wary of having to bear the cost of educating and then losing their most entrepreneurial and talented workers. Anecdotal evidence is startling. According to Stalker (1994), Jamaica had to train five doctors to retain Editor's note: This paper was submitted in September 2006. The author received comments in November. Riccardo Faini passed away suddenly on January 20, 2007. This version was revised by the editor and reflects helpful comments from three referees. The editor thanks David McKenzie for suggestionsand Celine Carr6re and Mario Piacentini for help. Riccardo Faini was a professor of political economy at Universita di Roma Tor Vergata and a research fellow at the Institute for the Study of Labor (Bonn) and the Center for Economic Policy Research (London).He wished to thank seminar participants at the Royal Economic Society, the Global Trade Analysis Project Conference in Lubeck, the Poverty Redirection and Economic Management Conference at the World Bank, and anonymous reviewers for helpful comments. He also wanted to thank Domenico De Palo for outstanding research assistance. rttt:WORLD BAKK ECONOMIC REVIEW, 21, NO.2, pp. 177-191 VUL. doi:lO. 1093/wber/lhm006 Advance Access Publication 24 May 2007 0TheAuthor 2007.Published by OxfordUniversityPresson behalf of theInternational Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org just one, and Grenada, 22. Central American and Caribbean countries are particularly affected by the brain drain, presumably because of their proximity to the United States. African countries have also suffered from a massive emigration of their skilled workers, with 30 percent of the best educated people working abroad, chiefly in the European Union. Extensive confirm- ation of these trends is given in the article by Docquier, Lohest, and Marfouk in (2007). The brain drain is not an unmitigated curse, however, at least not in theory. The possibility for educated migrants to move abroad should raise the returns to education and, in the end, may even lead to an increase in the number of educated workers who stay at home (Bhagwati and Hamada 1974; Bhagwati 1976; Mountford 1997; Stark, Helmenstein, and Prskawetz 1997, 1998). Moreover, skilled migrants will typically earn more and may therefore remit more (World Bank 2003, 2006), relieving the foreign exchange constraint at home and thus fostering gowth.' This article takes a fresh look at whether the brain drain is associated with a larger flow of remittances. It finds that the conventional wisdom, as often happens, is wrong. There is no evidence that skilled workers remit more. This is not so simply because they come from relatively wealthy families. It also reflects the fact that skilled migrants are more likely to spend a longer time abroad, weakening their ties with the home country. As a consequence, the growing bias in receiving countries toward skilled migration may reduce the flow of remittances to sending countries, although this negative effect is not statistically significant. The impact of the emigration of highly skilled people on the economic pro- spects of origin countries is highly controversial. The early literature (Bhagwati 1976) tended to conclude that the brain drain was detrimental to sending countries. Even when skilled workers were unemployed at home, their social marginal productivity was not necessarily nil, as they could move internally and become productively employed. More crucially, the costs of education are 1. Even the impact of remittances on home country growth has been questioned, however. Much of the descriptive literature on remittances argues that they are used "unproductively," citing micro evidence that remittance income is often wasted on conspicuous expenditures (see, however, Adams 1991, 1998, for an opposite view). More elaborate models conclude that higher remittances may exacerbate moral hazard problems on the recipient side and discourage work effort. Growth may fall as a result. Building on this argument, Chami, Fullenkamp, and Jahiah (2003) argue that remittances are not an effective source of capital for development. Chami, Fullenkamp, and Jahjah (2003) and Faini (2002)give some (conflicting)evidence on the link between growth and remittances. See also Rajan and Subramanian (2005a, 2005b) for an insightful discussion of the different growth impacts of aid and remittances. typically borne by home country taxpayers. The more recent literature is more nuanced. Mountford (1997),Stark, Helmenstein, and Prskawetz (1997, 1998), and Beine, Docquier, and Rapoport (2001, 2003) argue that the possibility of migrating abroad raises the returns to education and may boost investment in education. It could then well be that, ex post, after some educated workers have moved abroad, the home country share of educated workers rises rather than falls. A different line of argument emphasizes the role of remittances. According to World Bank (2003), the negative effects of the brain drain are somewhat offset by inward remittances from migrants. There is indeed some (limited)evi- dence that remittances tend to increase with the level of skills (Johnson and Whitelaw 1974; Rempel and Lobfell 1978). Presumably, skilled migrants earn more and therefore, other things being equal, are likely to remit more. However, there are many unresolved issues with this strand of literature. First, the evidence is not unequivocal, with Rodriguez and Horton (1994) for instance showing for the Philippines that the education level of migrants has no impact on the amount of remittances. Second, skilled workers may come from wealthier and more educated families and therefore have less of an incentive to remit. Finally, they may spend a longer time abroad,' either because they are more eager to reunite with their families in the host country or because they face fewer constraints in doing so. Indeed, a typical finding in the literature is that the flow of remittances tends to decline with the length of migrants' stay (Lucas and Stark 1985). There is substantive evidence that skilled migrants stay longer in the host country. Borjas and Bratsberg (1996) show that, under fairly general con- ditions, return migration will amplify the initial selection bias. Thus, if migrants are initially relatively skilled, then the least skilled will be more likely to return to their home country. Intuitively, if the initial selection bias is positive with the more skilled also more prone to migrate, then the least skilled will be the marginal migrants and will be more likely to reconsider their initial decision. Solimano (2002) reports that, at least in science and engineering, a large fraction of PhD graduates from developing countries remain in the United States after graduating. National Science Foundation data show that, four years after graduation, 88 percent of China's and 79 percent of India's graduates in science and engineering are still working in the United States. More comprehensive evidence comes from Massey and Lindstrom (1994) for Mexican migrants, Reagan and Olsen (2000) for the United States, Bauer and Gang (1998) for Egypt, and Steiner and Velling (1994) for Germany. For the Philippines, Rodriguez and Horton (1994) show 2. More direct evidence on the positive relationship between education and durarion of stay comes from Reagan and Olsen (2000)for the United States. Similarly, the intended durarion of stay is found ro rise with education in Germany (Steiner and Velling 1994). Knerr (1994) also finds that for Pakistan skilled workers tend to stay longer abroad than unskilled workers. 180 T H E WOR1.D B A N K E C O N O M I C R E V I E W that returnees are somewhat less educated than those still abroad. Finally, Borjas (1989) shows that the least successful foreign scientists are more likely to return from the United States. Further evidence on these issues is provided here for the European case using data for 1994-2001 from the European Community Household Panel. The focus is on foreign-born individuals with a view to assessing whether skilled migrants are more likely to reunite with their close family members and less likely to return home. For Europe as a whole and over the entire period low- skill individuals are less likely to be reunited with their spouse and live in the same dwelling (figure 1). The European Community Household Panel is a closed panel and therefore cannot easily be used to study return migration. However, some useful insights come from an analysis of the pattern of attrition. To control for the possibility that factors other than migrants' choice to return home may affect attrition, panel attrition is assumed to be due either to the unwillingness of respondents to be interviewed again or to their choice to move to a different location, poss- ibly to their home country. Accordingly, included among the regressors are the number of visits for the interview (under the plausible assumption that individ- uals are more likely to drop out of the panel when interviews are very time consuming) and a set of time-dependent dummy variables for the immigrant's region of origin to capture the effect that changing conditions in the home FIGURE1. Family Reunification Among Foreign-Born Individuals (Percentage of Households in which Spouses are Reunited) Source: Author's analysis based on data from the European Community Household Panel. country might have on the decision to return. A simple equation is estimated in which the probability of an immigrant remaining in (not dropping out of) the panel is a function of individual and household characteristics. Finally, separate equations are estimated for natives, immigrants from other EU countries, and non-EU immigrants. The key finding (table 1, column 3) is that more educated immigrants from non- EU countries are less likely to drop out of the panel, even after controlling for age, gender, employment status, and length of stay in the host country. This finding may be taken to indicate that educated (non-EU) immigrants are less likely to return home, consistent with the notion that skilled migrants tend to stay longer in the host country. However, it may also be interpreted as showing that educated immigrants are less reluctant to be interviewed again. Which explanation holds is difficult to determine. What can be said, however, TABLE The Pattern of Attrition in the European Community Household 1. Panel Sample (dependent variable: probability that respondent does not drop out of the panel) EU Non-EU Variable Natives immigrants immigrants - Household size Age Highest education Intermediate education Gender Employment Spouse Visits" ~ i n u t e s ~ Immigrant Immigrant EU Immigrant non-EU LengthC< 5 years Length' 6-15 years Lengthc 16-25 years Constant Yes Yes Yes Country dummy variable Yes Yes Yes Time'origin Yes Yes Yes Time dummy variable Yes Yes Yes Number of observations Number of observations censored "p < 0.05; ;'"p < 0.01; """p < 0.001. "Number of visits to complete the interview at time t-1. h~engthof the interview at t-1. 'Immigrant's length of stay in the host country. Source: Author's analysis based on data from the European Community Household Panel, 1994-2001. 182 T H E W O R L D R A N K E C O N O M I C R E V I E W is that this effect is quantitatively much stronger for immigrants than for natives. More crucially, the overall effect of the home country time dummy variable is strongly significant, suggesting that conditions at home influence the attrition outcome. This evidence is consistent with the notion that skilled migrants stay longer in the host country and are more likely to reunite with close family members. The implications of these findings for remittances are illustrated with the help of a simple model. Assume that the household is composed of two groups, one very close to the migrant and the other less close. Only close family members are assumed to reunite with the migrant in the host country. The size of each group is normal- ized to one. Thus, in what follows, fR (with 0 0. Second, for a given C, both the level and the marginal value of the migrant's utility are relatively larger for the close family members: (2) v"(c) > vD(C)and vC'(C)> vD1(C). There are four budget constraints. One is for the migrant: where w denotes the migrant's wage, Riremittances to group i, and 8 the cost for the migrant of bringing relatives to the host country. The migrant's con- sumption is then equal to the migrant's wage minus the sum of remittances and reunification costs. The other three budget constraints are for each of Faini 183 the three other household groups: where household members have two sources of revenue, their own income (Y,) and remittances (Ri). The first-order condition with respect to remittances to, say, nonmigrating close household members (RH),can be written as: After deriving analogous conditions for RR and RD,this becomes: Because of equation (2),this implies: Then, unless YD is significantly smaller than YH (a somewhat implausible case, given that YH will typically fall following the migration of one of the migrant's close relatives), remittances to close nonmigrating household members (RH)will be larger than those to distant family members (RD). Given equation (6),the first-order condition with respect to fR reads simply as: which states that the marginal rate of substitution between fR and CMmust be equal to its cost. This condition plays a key role in the results, as shown subsequently. Now consider the impact of a higher skill content of migration. Migrants with higher skills will typically earn a higher wage abroad. From equations (1) and (3)it can easily be seen that an increase in w will, for a given flow of remittances and an unchanged value of fR, lead to a rise in the migrant's own consumption (CM)and, as a result, to a fall in the marginal utility of CM. Thus, Uf/Ucwill be greater than 6. For a new equilibrium to obtain, the degree of family reunification will also have to rise, so as to bring Uf/Uc back in line with 6. As claimed, (high wage) skilled migrants will therefore be more likely to reunite with their family. What about remittances? Given equation (6),in the new equilibrium both vC'and vDfwill need to fall and, as a result, for a given set of YiYs,remittances to all groups (H, R, and D) will in~rease.~However, the fact that both RH and RD will increase does not necessarily imply that the total flow of "actual" remit- tances will in~rease.~Indeed, there will be two effects. First, as just noted, the amount of per capita remittances will rise for both nonmigrating close and distant family members (the "wage" effect). Second, a larger share of close family members will reunite with the migrant. The composition of nonmigrating members will therefore shift toward the low-remittance group-the distant family members (the "reunification" effect5).If the reunification effect is stron- ger than the wage effect, per capita remittances to nonmigrating household members will decline. The model does not provide an unambiguous answer in this respect. Only empirical analysis can resolve such ambiguity. Finally, the impact of higher income at home can also be examined. If Yi rises, the marginal utility of remittances to group i will fall. Remittances will therefore decline. Empirical analysis of the brain drain has been hampered by the lack of compre- hensive and comparable data. With the pioneering work of Carrington and Detragiache (1998)and the contributions of Dumont and Lemaitre (2004) and Docquier and Marfouk (2004), however, this gap is being filled. Carrington and Detragiache (1998) rely on the 1990 U.S. Census to estimate the number of skilled migrants to the United States from a large set of sending countries. For non-U.S. destinations they have information only on the total number of migrants. They address this shortcoming by assuming that migrants to non-U.S. OECD destinations have the same skill composition as migrants to the United States. Obviously, their origin country data on the brain drain are a valid approximation only for countries that send most of their migrants to the United States. Docquier and Marfouk (2004)overcome this limitation by using national sources of data for destinations other than the United States to esti- mate the skill composition of migranw6 Finally, both Carrington and 3. More formally, R, and fR will increase following a rise in w if they are both normal goods:U,b - UbbUi/Ub > 0 4. The focus next is on "actual" remittances, or the amount of transfers to nonmigrating household members (RH and RD).Indeed, RR (the remittances to close family members who have already reunited with the migrant) is simply an intrahousehold transfer and does not generate any foreign exchange inflow for the home country. 5. Reunification may well be dictated by immigration laws in the host country. But reunification procedures are typically quite generous in most destination countries. Moreover, migrants have considerable leeway in selecting which close family members to bring to the host country. Migrants may also select the host country based on its provisions for family reunification. Overall, therefore, reunification is likely to be largely exogenous. 6. Docquier and Marfouk (2004)also extend the Carrington and Detragiache dataset to 2000. Detragiache (1998)and Docquier and Marfouk (2004) relate the total number of skilled migrants to the Barro and Lee (2001) data set on educational achievements to derive a measure of migration rates for skilled workers, here defined as migrants having completed tertiary education. The following relies on the Docquier and Marfouk (2004) data set to inves- tigate the relationship between remittances and skilled migration. Remittance data come from the International Monetary Fund (IMF, various years) and include workers remittances, compensation of employees, and capital transfers. Population and income data come from the World Bank (2005).Docquier and Marfouk provide data for 1990 and 2000 on the level and composition of migration. Remittance data are averaged over 1990-91 and 2000-01. The estimating equation is inspired by the model described in section I I . ~Two groups of migrants are distinguished, skilled (S) and unskilled (U),with remit- tances of skilled workers denoted by Rs and of unskilled workers by RU.Total remittances are therefore identically equal to: where mi denotes the number of migrants in group i. To implement the model, detailed information is needed on the degree of family reunification and the number of close and distant relatives. While such data are unavailable, an esti- mating equation can nonetheless be derived. The model suggests that for both skilled and unskilled migrants remittances are a function of the migrant's wage and the income of the migrant's family members, leading to the following behavioral equation: where wi denotes the wage level of migrant i, y, denotes the per capita income of household members, and ai > 0 for i = S,U is the propensity to remit. With altruism, remittances should be a declining function of yi (the income of those left behind). Thus, it is expected that Pi > 0 for i = S,U, where Pi > 0 measures the degree of altruism. In equation (lo),under the assumption that the reunifi- cation effect applies mainly to skilled migrants, as < au. Common sense and human capital theory permit the assumption that ws > wu. If the reunification effect is stronger than the wage effect, it could well be that asws < auwu. In the absence of information about the functional form of equation (lo), a linear relationship is postulated in the estimation and then a log-linear functional form. 7. The results differ from those in Faini (2002) for two main reasons: a different and broader data set is used and the estimating equation is more closely related to theory. 186 T H E W O R L D R A N K EC.ONOMIC R E V I E W An additional hurdle arises because there is no separate information on the wage level of skilled and unskilled migrants (wi) or on the income of their household members (y;).Nonetheless, equation (10)can be substituted for I = S,U in equation (9) to estimate an aggregate equation. After some simple algebra, that yields: where M (= ms + mu) is total migration, P (=ps +pu) is total population, and Y (= pSyS puyu) is total income in the home country. Equation (11)can be + estimated by regressing per capita remittances (RIP) on the total stock of migrants abroad relative to population (MIP), the ratio of skilled migrants to the home country population (mslP),and income per capita times the migration rate of unskilled workers (mulpu YIP) and skilled workers (mslpSYIP).The major shortcoming of this approach is the need to assume that the unobservable distributional share parameters, TU -- (1 psyslY), 7s = (psyslY) are invariant - across countries. As a first step, figure 2 reports the scatter plot of per capita remittances on the share of skilled migrants in the total stock of migrants, showing the linear fit between the two variables with the associated 95 percent confidence FIGURE2. Per Capita Remittances and the Share of Skilled Migrants Skilled migrants in total stock of migrants (mslM) Source: Author's analysis based on data from Docquier and Marfouk (2004). interval. The slope is significantly negative at a 10 percent significance level. This preliminary inspection suggests that a larger share of skilled migrants is associated with a lower remittance flow. Estimation of equation (11)leads to: where D, = 1 if t = 2000; the expected signs are y2 = ~ U W U >0, 7 3 = ~ S W S- aUwU50,y4 = BurU< 0, and y5 = -PSrS< 0; and &it is a stochastic error - term. Equation (12) is first estimated separately for each period. Data are then pooled after testing that the relevant coefficients are constant across time and that pooling is appropriate. In the pooled equation a dummy variable D, (which takes the value of one in the year 2000) is added to control for possible time effects. The coefficient on the stock of migrants is expected to be higher in 2000, since wages are likely to have increased between 1990 and 2000. In an experiment with an alternative specification the time dummy variable is inter- acted with the migration stock variable ( yl > 0). The pooled estimates are pre- sented in column 1 of table 2.8 The pooling restriction is not rejected by the data at the 5 percent significance level, with F(4,178)= 0.94. Generally speaking, the coefficients have the expected sign. Most are signifi- cantly different from zero at standard statistical levels. Total migration has a positive sign, as expected. All else being equal, therefore, an increase in the stock of the migrant population relative to the home country population (MIP) will be associated with a larger flow of remittances. The coefficient of mslP is negative, but not significantly different from zero. This coefficient is crucial for the analysis. As shown, (aSwS- aUwU)could be of either sign depending on whether the wage effect is stronger or weaker than the reunification effect. The negative nonsignificant sign indicates that the reunification effect is stronger than the wage effect - as < au.Finally, the estimates suggest that y4 > 0 and that y5 < 0. The degree of altruism, as measured by the coefficient pi, is posi- tive for skilled workers but negative for unskilled workers. When the fact that 64 of 188 observations are censored at zero (using the Tobit maximum likeli- hood estimation procedure) is taken into account, the results are basically 8. Only the specification with the time dummy variable interacted with the migration stock variable is reported. TABLE2. Remittances and the Skill Composition of Migration (1) (2) (3) (4) (5) (RIP) (RIP) (RIP) In(RIP) In(RIP) Estimator Pooled Tobit Pooled Pooled Instrumental ordinary least ordinary ordinary variable squares least squares least squares with R > 0 3.6 (0.4) 3.8 (2.3) - 1.7 (0.2) 0.13 (2.1) -0.014 (1.5) Number of observations R is total remittances, M is migration stock, P is home country's population, ms is skilled migrants, mu is unskilled migrants, p~ is home country's skilled population, pUis home country's unskilled population, Y is GDP, and t is time effect. Note: Numbers in parentheses are t-statistics corrected for heteroskeda~ticit~. Source: Author's analysis based on data from Docquier and Marfouk (2004). unchanged (column 2) with respect to those in column 1, and they remain so even if the sample is restricted to observations with R > 0 (column 3).9 Next, some curvature is introduced in the relationship between remittances and home income. To check the reobustness of the linear functional form of equation (lo), several alternative functional specifications are tested. Remittances are again found to be positively affected by a larger stock of migrants, and the population share of skilled migrants carries a negative coeffi- cient, although it is statistically insignificant. The degree of altruism of both skilled and unskilled workers is now found to be positive. Note that if it is assumed that pU= &-that altruism between skilled and unskilled migrants is equal-the distributional parameter can easily be computed: 7s = (psyslY)= 0.24, a somewhat large but not unreasonable value. Finally, to control for the possibility that the total migration rate is endogen- ous, the equation is re-estimated with an instrumental variable procedure using the log of distance between the home and the host country as an instrument.'' 9. In response to suggestions by a referee, a dummy variable for small island countries (mslP) was introduced both additively and multiplicatively to capture the possibility that remittances are measured less accurately for these countries, which have the largest brain drain, thereby mechanically leading to a negative relation between (mslP) and remittances. Estimated coefficients remained unchanged. The results are also unchanged when the regression is estimated only for 2000 and when income per capita in the source country is introduced by itself as an additional variable. 10. A weighted average of distance to the European Union and the United States is used, with weights reflecting the relative importance of these two destinations. While distance is typically a major determinant of migration, it should not affect a financial flow such as remittances, making it an adequate instrument. The R~ of the first-stage regression is 0.96, and the F statistics is 685.8. Distance is a significant determinant of migration, suggesting that it contains considerable additional and, one hopes, exogenous information. In the second- stage regression (reported in column 5 of table 2), all previous results carry through and, if anything, are even stronger in both coefficient size and statisti- cal significance. However, it could be that financial remittances are inversely related to distance, in which case distance would not be a valid instrument. IV. C O N C L U S I O N S It is often argued that the negative impact of the brain drain might be mitigated by its favorable effect on remittances. This article has shown that this is not generally true. This is both because skilled migrants are more likely to come from wealthier families and because their propensity to remit is relatively lower, presumably reflecting the fact that they are keener (and more able) to bring their closest relatives to the host country. The findings of this article need to be confirmed by further research, especially at the household level. Globally, they show that the migration of skilled workers is unlikely to boost the flow of remittances to the source country. However, this finding assumes that the migration of skilled workers occurs as a one-for-one substitute for unskilled migration. While this appears to be the hope of some policymakers, in practice the total level of migration may increase as skilled migration increases. The extent to which skilled migration may substitute for or complement unskilled migration remains an open question for future research. Adams, R. 1991. "The Economic Uses and Impact of International Remittances in Rural Egypt." Economic Development and Cultural Change 39(4):695-722. . 1998. "Remittances, Investment and Rural Asset Accumulation in Pakistan." Economic Development and Cultural Change 47(1):155-73. Barro, R.J., and J.W. Lee. 2001. "International Data on Educational Attainment: Updates and Implications." Oxford Economic Papers 53(3):541-63. Bauer, T., and I. Gang. 1998. "Temporary Migrants from Egypt: How Long Do They Stay Abroad?" Discussion Paper 3. Bonn, Germany: Institute for the Study of Labor, Bonn University. Beine, M., F. Docquier, and H. Rapoport. 2001. "Brain Drain and Economic Growth: Theory and Evidence." Journal of Development Economics 64(1):275-89. . 2003. "Brain Drain and LDCs' Growth: Winners and Losers." Discussion Paper 819. Bonn, Germany: Institute for the Study of Labor, Bonn University. Bhagwati, J. 1976. "The International Brain Drain and Taxation: A Survey of the Issues." In J. Bhagwati ed., The Brain Drain and Taxation 11: Theory and Empirical Analysis. Amsterdam: North-Holland. 190 THE W O R L D B A N K E C O N O M I C R E V I E W Bhagwati, J., and K. Hamada. 1974. "The Brain, International Integration of Markets for Professionals, and Unemployment: A Theoretical Analysis." journal of Developtnent Economics 1(1):19-24. Borjas, G. 1989. "Immigrant and Emigrant Earnings: A Longitudinal Study." Economic Inquiry 27(1):21-37. Borjas, G., and B. Bratsberg. 1996. "Who Leaves? The Outmigration of the Foreign-Born." Review of Economrcs and Statistics 78(1):165-76. Carrington, W., and E. Detragiache. 1998. "How Big Is the Brain Drain?" IMF Working Paper 981102. Washington, D.C.: International Monetary Fund. Chami, R., C. Fullenkamp, and S. Jahjah. 2003. "Are Immigrant Remittance Flows a Source of Capital for Development?" IMF Working I'aper 031189. Washington, D.C.: International Monetary Fund. Docquier, F., and A. Marfouk. 2004. "Measuring the International Mobility of Skilled Workers (1990- 2000)." Policy Research Working Paper 3381. World Bank, Washington, D.C. Docquier, F., 0 . Lohest, and A. Marfouk. 2007. "Brain drain in developing countries." World Bank Economic Review 21(2);doi:10.1093/wber/lhm008. Dumont, J.C., and G. Lematre. 2004. "Counting Immigrants and Expatriates in OECD Countries: A New Perspective." Social, Employment and Migration Working Papers 25. Paris: OECD Directorate for Employment, Labour, and Social Affairs. Faini, R. 2002. "Developmenr, Trade, and Migration." Revue d'~conomieet du Ddveloppement, Proceedings trom the ABCDE Europe Conference 1-2: 85-116. IMF (International Monetary Fund). Various years. Balance of Payments Statistics Yearbook. Washington, D.C. Johnson, G., and W. Whitelaw. 1974. "Urban-Rural Income Transfers in Kenya: An Estimated Remittances Function." Economic Development and Cultural Change 22(3):473-79. Knerr, B. 1994. "Labour Migration from South Asia: Patterns and Economic Implications." In D. O'Connor and L. Farsakh, eds., Development Strategy, Employment and Migration. Country Experiences. Paris: OECD Development Centre. Lucas, R.E.B., and 0 . Stark. 1985. "Motivations to Remit: Evidence from Botswana." journal of Political Economy 93(5):901-18. Massey, D., and D. Lindstrom. 1994. "Selective Emigration, Cohort Quality, and Models of Immigrant Assimilation." Sucial Science Research 23(4):315-49. Mountford, A. 1997. "Can a Brain Drain Be Good for Growth in the Source Economy?" journal of Deuelopment Economics 53(2):287-303. OECD (Organisation for Economic Co-operation and Development). 2003. Economic Outlook. Paris: OECD. Rajan, R., and A. Subramanian. 2005a. Aid and Growth: What Does The Cross-Country Evidence Really Show? NBER Working Paper 11513. Cambridge, Mass.: National Bureau of Economic Research. .2005b. What Undermines Aid's Impact on Growth?NBER Working Paper 11657. Cambridge, Mass.: National Bureau of Economic Research. Reagan, P., and R. Olsen. 2000. "You Can Go Home Again: Evidence from Longitudinal Data." Demography 37(3):339-50. Rempel, H., and R. Lobdell. 1978. "The Role of Urban-to-Rural Remittances in Rural Development." journal of Development Studies 14(3):324-41. Rodriguez, E., and S. Horton. 1994. "International Return Migration and Remittances in the Philippines." In D. O'Connor and L. Farsakh, eds., Development Strategy, Employment and Migration. Country Experiences. Paris: OECD Developmenr Centre. Solimano, A. 2002. "Globalizing Talent and Human Capital: Implications for Developing Countries." Paper presented at the Annual World Bank Conference on Developmenr Economics for Europe, June 24-26, Oslo, Norway. Stalker, P. 1994. The Work of Strangers: A Survey of lnternational Labotir Migration. Geneva: International Labour Office. Stark, O., C. Helmenstein, and A. Prskawetz. 1997. "A Brain Gain with a Brain Drain." Economic Letters 55(2):227-34. . 1998. "Human Capital Formation, Human Capital Depletion, and Migration: A Blessing or a 'Curse?"' Economic Letters 60(3):363-67. Steiner, V., and J. Velling. 1994. "Re-migration Behaviour and Expected Duration of Stay of Guest Workers in Germany." In G. Steinmann and R. Ulrich eds., The Economic Consequences of I?nmigration to Germany. I-leidelberg: Physlca-Verlag. World Bank. 2003. "Chapter 7." Global Development Finance 2003: Striving for Stability in Development Finance. Washington, D.C. . 2005. World Det~elopmentIndicators 2005. Washington, D.C. -. 2006. Global Economic Prospects 2006: Economic implications of Remittances and Mtgration. Washington, D.C. Brain Drain in Developing Countries Fre'de'ric Docquier, Olivier Lohest, and Abdeslam Marfouk An original data set on international migration by educational attainment for 1990 and 2000 is used to analyze the determinants of brain drain from developing countries. The analysis starts with a simple decomposition of the brain drain in two multiplica- tive components, the degree of openness of sending countries (measured hy the average emigration rate) and the schooling gap (measured by the education level of emigrants compared with natives). Regr6ssion models are used to identify the determinants of these components and explain cross-country differences in the migration of skilled workers. Unsurprisingly, the brain drain is strong in small countries that are close to major Organisation for Economic Co-operation and Development (OECD) regions, that share colonial links with OECD countries, and that send most of their migrants to countries with quality-selective immigration programs. Interestingly, the brain drain increases with political instability and the degree of fractionalization at origin and decreases with natives' human capital. JEL classification codes: F22, 015,J24 The international migration of skilled workers (the so-called brain drain) has attracted considerable attention. Industrial countries such as Canada, Germany, and the United Kingdom worry about the emigration of their talented workers, but it is the detrimental consequences of the brain drain for developing countries that are usually stressed in the literature. By depriving developing countries of human capital, one of their scarcest resources, brain drain is usually seen as a drag on economic development. Yet recent theoretical studies emphasize several compensatory effects, showing that a limited but positive skilled emigration rate can be beneficial for sending countries (Commander, Kangasniemi, and Winters 2004; Docquier and Rapoport 2007; Beine, Docquier, and Rapoport 2001, forthcoming; Schiff 2005 provides a critical appraisal of this literature). However, without reliable comparative data Frtdtric Docquier (corresponding author) is a research associate at the National Fund for Economic Research; professor of economics at the Universite Catholique de Louvain, Belgium; and research fellow at the Institute for the Study of Labor, Bonn, Germany; his email address is docquier@ires.ucl.ac.be. Olivier Lohest is a research is a researcher at the Institut Wallon de I'Evaluation, de la Prospective et de la Statistique, Regional Government of Wallonia, Belgium; his email address is olohest@hotmail.com. Abdeslam Marfouk is a researcher at the Institute for the Study of International Migration, Georgetown University; his email address is a.marfouk@skynet.be. THE WORLD BANK ECONOMIC REVIEW, VOL.. 21, NO. 2, pp. 193-2 18 doi:10.1093/wber/lhm008 Advance Access Publication 1 3 June 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 on international migration by educational attainment, the debate on the causes and consequences of the brain drain has remained essentially theoretical. With the rapid evolution of international migration and the policy issues at stake, the international community must be prepared to address the major chal- lenges raised by the brain drain. Assessing the economic impact of emigration by skilled workers requires better knowledge of the educational structure of international migration and its determinants. This article seeks to characterize the distribution of the brain drain from developing countries in 1990 and 2000 and its main determinants using the new harmonized comprehensive data set on migration stocks and rates by edu- cational attainment recently built by Docquier and Marfouk (2006). Generalizing the pioneering work of Carrington and Detragiache (19981, their method consists of collecting census and registry data on the structure of immi- gration in all Organisation for Economic Co-operation and Development (OECD)countries. In a first step, aggregating these data allows for evaluating the stock of emigrants from all developing countries to OECD countries by level of schooling. In a second step, comparing the number of migrants to that of natives (defined here as residents and emigrants) in the sending country in the same education group gives a relative measure of the emigration rate by educational attainment for 1990 and 2000. Section I presents the data set on the brain drain, as measured by the emigra- tion rate of post-secondary-educated workers, and describes the average brain drain from developing countries by income group and country size. Between 1990 and 2000, the stock of skilled immigrants in OECD countries increased by 64 percent. The rise was stronger for immigrants from developing countries (up 93 percent), especially from Africa (up 113 percent) and Latin America and the Caribbean (up 97 percent). Although the number of skilled workers from developing countries increased, emigration rates decreased slightly. What at first looks like a ~aradoxcan be explained by the general rise in educational attainment in many developing countries between 1990 and 2000. The new brain drain measures are then compared with those in previous studies, showing how they resolve many important sources of bias. Section II decomposes the brain drain into two multiplicative components: the degree of openness, measured by the average emigration rate of working- age natives, and the schooling gap, measured by the relative education attainment of emigrants compared with natives. On average, there is a negative correlation between openness and schooling gap, implying that a high brain drain usually accompanies either strong permeability or a high schooling gap, but not both. This justifies decomposing the brain drain into these two com- ponents and investigating their individual determinants. A preliminary descriptive analysis reveals interesting regularities in the data. On the one hand, openness is strongly affected by country size: small countries exhibit higher average emigration rates than large countries do. On the other hand, the schooling gap is closely related to the average level of schooling among Docquier, Lohest, and Marfouk 195 natives: poor countries exhibit higher schooling gaps. Bilateral schooling gaps vary across destination countries, so destination choices affect the intensity of the brain drain. Other things being equal, the brain drain is stronger in small and poor countries sending most of their emigrants to countries with quality- based immigration policies. Section I11 uses ordinary least squares and instrumental variable regression models to analyze the determinants of openness and the schooling gap. The degree of openness increases as country size declines, as natives' human capital and political instability increase, as colonial links strengthen, and as geographic distance to the major OECD countries declines. The schooling gap depends on natives' human capital, the type of destination country (with or without selective-immigration programs), on distances and religious fractionalization at origin. A rise in human capital stimulates openness and reduces the schooling gap. The second effect dominates: other things being equal, the brain drain is stronger in poor countries where the average level of schooling is low. All these findings improve the understanding of the sources of the brain drain. The analysis builds on the new international migration data set developed by Docquier and Marfouk (2006).The data set was used to compute absolute and relative emigration data by educational attainment for developing countries for 1990 and 2000. First, absolute emigration stocks by educational attainment are computed for every country. Next, these numbers are expressed as percentages of the total labor force born in the sending country (including migrants) with the same education level. Stocks of Skilled Emigrants Emigration statistics provided by origin countries, when available at all, do not give a realistic picture of emigration (see Wickramasekera 2002). Data on emi- gration can be captured only by aggregating harmonized immigration data collected in many receiving countries. Detailed information about the origin and skill of immigrants can usually be obtained from national censuses and registries. The Docquier-Marfouk data set is based on data collected in all OECD countries. It counts as migrants all working-age (25 and older) foreign- born individuals living in an OECD country. The total number of working-age emigrants from country i of skill s in year t is denoted by M:,,. Three levels of schooling are distinguished. Low-skill workers, with a primary education; medium-skill workers, with a secondary education; and high-skill workers, with a post-secondary education. The brain drain is defined as the migration of high-skill workers. The Docquier-Marfouk data set devotes special attention to data homoge- neity and comparability. To this end, several methodological choices were made (see Docquier and Marfouk 2006 for details). Considering the working-age population (ages 25 and older) maximizes comparability between immigration data and data on educational attain- ment in source countries and excludes the large number of students who emigrate temporarily to complete their education.' Restricting the set of receiving countries to the OECD area focuses atten- tion on emigration from developing countries to industrial countries and between industrial countries. While there is a brain drain outside the OECD area as well, based on (less detailed) census data collected from various non-OECD countries, it is estimated that 90 percent of high-skill international emigrants are living in OECD countries. Holding receiving countries constant between 1990 and 2000 allows comparisons over time. Consequently, Czechoslovakia, Hungary, the Republic of Korea, Mexico, and Poland are considered receiving countries in 1990 although they were not then members of the OECD. The number of adult immigrants in the OECD increased from 41.8 million in 1990 to 59.0 million in 2000, and the number of skilled immi- grants increased from 12.5 million to 20.4 million. Defining migration primarily on the basis of the concept of the foreign- born population rather than citizenship better captures the decision to emigrate and is time invariant. Information about the origin country of migrants is available in the large majority of OECD countries, repre- senting 52.1 million immigrants in 2000 (88.3 percent of the total). Information on citizenship is used for the remaining countries (Italy, Germany, Greece, Japan, and the Republic of Korea). While the defi- nition of foreign born is not fully comparable across countries, efforts were made to homogenize the concepts. Using direct data on educational attainment for 24 countries for 2000 and data from Labor Force Surveys, which provide less detailed infor- mation about immigrants' origins, for three countries (Belgium, Greece, and Portugal), means that the educational structure can be obtained or estimated for 27 countries representing 57.9 million immigrants (98.1 percent of the t ~ t a l )For . ~ migrants whose educational attainment is not described, the educational structure is extrapolated from the Scandinavian countries for Iceland and from the rest of the OECD for Japan and Korea. Skilled Emigration Rates Relative emigration measures are obtained by comparing the emigration stocks to the total number of people born in the source country (residents plus emi- grants, which together equal natives) and belonging to the same educational category. Calculating the brain drain as a proportion of the total educated 1. Carrington and Detragiache (1998)also considered individuals ages 25 and older. 2. Figures for 1990 are detailed in Docquier and Marfouk (2006). Docquier, Lohest, and Marfouk 197 labor force provides a better measure of the pressure imposed on the local labor market. Thus, for example, the emigration of 150,000 skilled Egyptians (4.5 percent of their educated labor force) exerts less pressure on the Egyptian labor market than the emigration of 2,500 skilled Seychellians (56 percent of their educated labor force) exerts on the Seychelles labor market. The term emigration rate is thus used to refer to relative stock data and not to immigra- tion flows. Denoting by N:,, the number of residents in country i, of skill s (with s = h for skilled workers) in year t, the skilled emigration rate mt, is defined as: Evaluating N:,,requires data on the size and the skill structure of the working-age population in the countries of origin. Population data by age are provided by the United Nations Population Division (http://esa.un.org/unpp). Population data are split across educational groups using international human capital indicators. Several sources based on education attainment and enrollment variables can be found in the literature. These data sets suffer from important shortcomings. Those published in the 1990s reveal a number of sus- picious features and inconsistencies. And all of them are subject to serious comparability problems because of the variety of educational systems around the world. Three major competing data sets are available: Barro and Lee (2001), Cohen and Soto (2007), and de la Fuente and Domenech (2002).The first two sets depict the educational structure in both developed and developing countries. De la Fuente and Domenech focuses only on 21 OECD countries. Statistical comparisons of these data sets reveal that the highest signal to noise ratio is obtained in de la Fuente and Domenech. For developing countries Cohen and Soto's set outperforms Barro and Lee's in growth regressions. However, Cohen and Soto's data underestimate official statistics in many devel- oping countries. Generally speaking, Cohen and Soto predict extremely low levels of human capital in ~ f r i c a(the ~ share of post-secondary educated is lower than 1 percent in a large number of African countries) and in a few other non-OECD c~untries.~The Barro and Lee estimates seem closer to the African census data obtained for a dozen countries. As the brain drain is par- ticularly important in African countries, the Barro and Lee indicators are used when available. 3. For this reason, Cohen and Soto (2007)exclude African countries from their growth regressions. 4. According to the 1996 South African census, the share of educated individuals amounts ro 7.2 percent. Cohen and Soto report 3 percent (Barro and Lee report 6.9 percent). The Kenyan 1999 census gives 2 percent while Cohen and Soto report 0.9 percent (1.2 for Barro and Lee). In Cyprus the 2001 census gives 22 percent while Cohen and Soto give 4.6 percent ( 17.1 percent in Barro and Lee). Consequently, the Docquier-Marfouk data set relies on de la Fuente and Domenech's indicators for OECD countries, Barro and Lee's measures for most non-OECD countries, and adjusted Cohen and Soto's estimates for countries not in Barro and Lee. For countries for which no data are available, the skill structure of the neighboring country with the closest enrollment rates or GDP per capita is applied. This method gives good approximations of the brain drain rates, broadly consistent with anecdotal evidence. The Brain Drain in Developing Countries Following the 2000 World Bank income classification our analysis distinguishes 54 low-income countries, 58 lower-middle-income countries, and 40 upper-middle-income countries. Among these, three groups are of particular interest: small island developing countries, landlocked developing countries, and the least developed countries as defined by the United Nations. Table 1 gives an overview of absolute and relative emigration rates by country group in 1990 and 2000. In 2000 developing countries accounted for 64.5 percent of total immigrants and 61.6 percent of skilled immigrants in the OECD, 15 percentage points higher than in 1990. About three-quarters of these immigrants live in one of the three most important host countries with selective-immigration policies (Australia, Canada, and the United States). One-fifth of them live in 1 of the 15 member countries of the European Union (EU15). These percentages vary across origin groups: small island countries send many migrants to selective-immigration countries; least developed and landlocked countries send more migrants to the EU15. These destination choices are linked to geographic distances and histori- cal ties. Most small island countries are located in the Caribbean and the Pacific and thus send many migrants to the United States or Australia and New Zealand. Many landlocked countries are located in Africa and have strong colonial links with European countries. In every group the proportion of skilled workers among migrants (on average 33 percent for developing countries) is much higher than the pro- portion of skilled workers among residents (on average 6 percent). Hence, skilled emigration rates (on average 7.3 percent) are much higher than average emigration rates (on average 1.5 percent). These average levels hide a strong heterogeneity across states. The brain drain is extremely small (below 1 percent) in countries such as Bhutan, Oman, and Tajikistan, while it exceeds 85 percent in Grenada and Jamaica. Between 1990 and 2000 the average emigration rate rose from 1.1 to 1.5 percent. Although the proportion of skilled migrants increased, the skilled emi- gration rate decreased from 7.7 to 7.3 percent as the general level of schooling increased in developing countries. The highest brain drain rates are observed in small island developing countries and in the least developed countries, and the lowest rates in large and landlocked developing countries. Setting aside small island economies, the TABLE. Descriptive Statistics by Country Group, 1990-2000 1 Emigration Emigration structure Skilled emigrants by destination Labor force structure (region of origin) rates Total Skilled emigrants emigrants In selective- Total labor Skilled labor (ages 25 (ages 25 Share of immigration In EU15 In rest of force (ages force (ages Share of and older, and older, skilled countries countries OECD 25 and older, 25 and older, skilled Total Skilled Group of origin thousands) thousands) (%) (Yo) (Yo) (%) thousands) thousands) (%) (%) (%) 2000 World" High-income countries Developing countries Low-income countries Lower-medium-income countries Upper-medium-income countries Least developed countries Landlocked developing countries Small developing islands Large developing countries (>40 million) i (Continued) TABLE Continued 1. Emigration Emigration structure Skilled emigrants by destination Labor force structure (region of origin) rates Total Skilled emigrants emigrants In selective- Total labor Skilled labor (ages 25 (ages25 Share of immigration In EU15 In rest of force (ages force (ages Share of and older, and older, skilled countries countries OECD 25 and older, 25 and older, skilled Total Skilled Group of origin thousands) thousands) (%) (%) thousands) thousands) (%) (%) (%) 1990 World" High-income countries Developing countries Low-income countries Lower-medium-income countries Upper-medium-income countries Least developed countries Landlocked developing countries Small developing islands Large developing countries (>40 million) "Sum of emigrants from high-income countries, developing countries, and dependent territories and emigrants who did nor report their country of birth. Source: Docquier and Marfouk 2006. Docquier, Lohest, and Marfouk 201 highest average brain drain rates are observed in Sub-Saharan Africa (13 percent), Latin America and the Caribbean (11percent), and the Middle East and North Africa (10 percent). Comparison with Previous Studies The Docquier-Marfouk data set generalizes the work of Carrington and Detragiache (1998, 1999), which was the first serious effort to compile a har- monized international data set on migration rates by education level. Carrington and Detragiache used 1990 U.S. Census data and general OECD statistics on international migration to construct estimates of emigration rates at three education levels for 61 developing countries.' Although their study clearly initiated new debates on skilled migration, their estimates suffer from important shortcomings: The numbers of immigrants by country of origin are taken from U.S. Census data and from OECD statistics for the remaining countries. Although census data give an accurate picture of U.S. immigration, OECD statistics report the number of immigrants for the major origin countries only (top-10 or top-5 sending countries). This led to underesti- mates of immigration for a large number of sending countries, whose data were aggregated and considered as residual in the entry "other countries." This underreporting bias is reinforced by the fact that 1990 immigration data were missing for three OECD countries (Greece, Iceland, and Turkey) and that three countries (Mexico, Poland, and Slovakia) became OECD members after 1990. Although data based on country of birth are available from many national censuses, the OECD classifies European immigrants by citizen- ship. This is another source of underreporting bias as the number of foreign-born people is usually much higher than the number of foreign citizens (twice as large in the Netherlands and Sweden, for example). OECD statistics give no information on immigrants' age, making it impossible to isolate those ages 25 and older. This introduces an overre- porting bias when the aim is to consider skilled workers. Carrington and Detragiache applied the education structure of U.S. immigrants to immigrants in other OECD countries. For example, Surinamese migrants to the Netherlands are assumed to be distributed across educational categories in the same way as Surinamese migrants to the United States. Since U.S. immigration policy differs from that of many other countries, this assumption is highly tentative, especially for countries with a low migration rate to the United States. 5. Adams (2003) used [he same methodology to compute brain drain rates from 24 countries in 2000. 202 T H E W O R L D BANK E C O N O M I C REVIEW The Docquier-Marfouk (2006) study, which collected census, registry, and survey data from all OECD countries, enables the size of these biases for devel- oping countries to be evaluated. A comparison shows that the brain drain is highly overestimated in countries such as Algeria, Morocco, Siio Tomi. and Principe, Suriname, Tunisia, and Turkey. In transposing the educational struc- ture observed in the United States, Carrington and Detragiache (1998, 1999) and Adams (2003) obtain emigration rates of post-secondary-educated workers for North Africa and Turkey of 35 to 45 percent. The Docquier-Marfouk data set gives much lower skilled emigration rates for these countries of 5 to 20 percent. The brain drain is underestimated in many Sub-Saharan African countries, such as The Gambia, Kenya, Mauritius, and Seychelles, and in small countries sending a small number of emigrants to OECD countries, such as Mauritius. The over- and under-estimation biases range from 51.5 percent for Siio Tomi. and Principe to - 51.2 percent for Mauritius. Figure 1 shows skilled migration rates evaluated under these three measure- ment methods: Docquier and Marfouk (2006), based on national census and administrative data; Carrington and Detragiache (1998) and Adams (2003), based on OECD statistics and U.S. educational attainment data; and an inter- mediate method based on census and administrative data on the number of FIGURE Skilled Emigration Rates under Three Measurement Methods; all 1. Developing Countries, 2000 b -Centus data (Docqu~erand Marfouk 2006) 1 OECD stat~st~cssharlng ot U S cducat~onddta + (Carr~ngtonand Detrag~achc1998) A Cenrus + shar~ngof U S educat~ondata (~ntermed~atemethod) Note: Country codes follow the International Organization for Standardization classification (see www.iso.org/). Countries are ranked in descending order according to the Docquier and Marfouk (2006) method. Source: Authors' analysis based on data from Docquier and Marfouk (2006). Docquier, Lohest, and Marfouk 203 migrants and U.S. educational attainment data on education. In comparison to Docquier-Marfouk, Carrington and Detragiache and Adams underestimate the brain drain for a large majority of countries, while the third method overes- timates the brain drain. The highest skilled emigration rates are observed in small and poor countries (see table 1).Although many factors help to explain the intensity of the brain drain, country size and development levels are key determinants. A simple mul- tiplicative decomposition of the skilled emigration rate can help to explain the distribution of the brain drain across countries. Denoting by Mf, the number of working-age emigrants from country i of skill s (s= h for high-skill workers and s = 1 for low-skill workers) in year t and by N:,, the corresponding number of residents, the skilled emigration rate mtt can be decomposed as following: The first multiplicative component is the ratio of emigrants to natives-the average or total emigration rate of all types of individuals. It reflects the degree of openness of the sending country. The second multiplicative component is the ratio of the proportion of skilled emigrants by the same proportion among natives. This ratio reflects the schooling gap between emigrants and natives. This ratio is always higher than one, indicating that emigrants are more edu- cated than natives in all developing countries. Consider a hypothetical world in which emigration is proportional to popu- lation and the skill structure of emigration is identical to that of the native population. The schooling gap would then be equal to one and all countries would exhibit the same degree of openness. From the decomposition (brain drain = openness index x schooling gap), the brain drain would be homo- geneous across countries. Obviously, observations depart from that hypothetical situation: average emigration rates and schooling gap are strongly heterogeneous. As the next section shows, these two components are closely related to the characteristics of sending countries as well as to proximity variables and characteristics of the main destination countries. First, however, consider four stylized facts related to the process of emigration by skilled workers. Stylized fact 1: Average emigration rates and schooling gaps are negatively correlated. Figure 2 plots the log of the emigration rate and the log of the schooling gap in 2000. Both variables are expressed as differences from the sample mean. Average emigration rates and schooling gaps are negatively FIGURE Average Emigration Rate and Schooling Gap 2. Log of the averageemigration rate aspercent (deviation from the mean) Source: Authors' analysis based on data from Docquier and Marfouk (2006). correlated. The majority of observations fall in the top left panel (low emigra- tion rates and high schooling gaps) and bottom right panel (high emigration rates and low schooling gaps). A small number of observations fall in the top- right panel, but they are quite close to one of the axes. This means that no developing country has both strong openness and a high schooling gap. If a country suffers from a large brain drain it is either because it is very open or because the positive self-selection of migrants is strong. This justifies the decomposition and the analysis of the specific determinants of these two components. Stylized fact 2: Average emigration rates decrease with country size. There is an obvious link between population size in country of origin and number of migrants abroad. In absolute numbers the main emigration countries are the largest ones (China, India, Mexico, Philippines, and Turkey) while the smallest number of emigrants come from small countries (Maldives, Nauru, Palau, Tuvalu, and Vanuatu). However, an increase in population generates a less than proportional increase in emigration. As is well documented in the litera- ture, the average or total emigration rate decreases with population size in the country of origin. Thus the degree of openness is decreasing in the population size at origin. In 2000, the average emigration rate to the OECD ranged from 0.1 percent (for Bhutan, Chad, Lesotho, Niger, Oman, Swaziland, and Turkmenistan) to 53.7 percent (Grenada).The correlation between the log of native population size and average emigration rate is -53 percent (figure 3). In 2000, seven countries had average emigration rates above 40 percent (Dominica, Grenada, Guyana, Saint Kitts and Nevis, Samoa, Suriname, and Tonga): their average Docquier, Lohest, and Marfouk 205 FIGURE3. Average Emigration Rate and Country Size 0.0 1 ... 4 -8.0 Y 2 0 4.0 6.0 8.0 10.0 12.0 -9.0 - Native population in thousand (in logs) Source: Authors' analysis based on data from Docquier and Marfouk (2006) size was 0.237 million and none had a population above 1million. Among the eight countries with a population above 100 million (China, India, Indonesia, Brazil, Russia, Pakistan, Bangladesh, and Nigeria), the emigration rate was 1 percent or lower. Small countries have the highest emigration rates (table 2). Small island developing economies (average population of 1.3 million) exhibit an average emigration rate of 13.8 percent, compared with 1 percent for large developing countries (population of more than 40 million). Obviously, country size is not the unique determinant of openness, as revealed by the strong dispersion of the scatter plot in figure 3. However, differences in country size are important and explain a substantial fraction of the disparities across income groups. Average country sizes are 38 million for low-income countries, 40 million for lower-middle-income countries, and 15 million for upper-middle-income countries. Unsurprisingly, upper-middle-income countries exhibit the highest openness index. Stylized fact 3: Schooling gaps decrease with natives' rising human capital. An interesting major regularity concerns the educational structure of emigra- tion. It is natural that the proportion of educated among emigrants increases with the general level of education of the native population. The most educated diasporas originate from countries where the proportion of educated natives ranges from 10 to 20 percent (such as Jordan, Libya, Mongolia, Oman, Panama, the Philippines, South Africa, and Venezuela). Less educated dia- sporas come mainly from very poor countries (such as Angola, Guinea-Bissau, Mali, Mozambique, and Tuvalu). Six countries had a schooling gap greater than 30 (Lesotho, Malawi, Mozambique, Niger, Rwanda, Uganda): their TABLE Decomposition of Skilled Emigration Rates, 1990-2000 2. Decomposition Openness by destination (%) Schooling gap by destination -- A B C Brain To selective- To selective- drain (%) Openness Schooling immigration To EU15 To rest of immigration To EU15 To rest of Group of origin A = B x C (yo) gap countries countries OECD countries countries OECD 2000 Worlda High-income countries Developing countries Low-income countries Lower medium-income countries Upper-medium-income countries Least developed countries Landlocked developing countries Small island developing economies (Continued) TABLE Continued 2. Decomposition Openness by destination (Oh) Schooling gap by destination Brain To selective- To selective- drain (%) Openness Schooling immigration To EU15 To rest of immigration T o EU15 To rest of Group of origin A = B x C ( % I gap countries countries OECD countries countries OECD Large developing 5.6 1.0 5.8 0.7 0.2 0.1 6.8 3.6 4.5 countries (>40 million) 1990 World" 5.2 1.6 3.3 0.9 0.5 0.1 4.6 1.7 2.4 High-income countries 3.9 3.0 1.3 1.6 1.0 0.4 1.8 0.7 0.9 Developing countries 7.7 1.1 7.1 0.6 0.4 0.1 9.7 3.7 5.0 Low-income countries 5.6 0.5 11.1 0.3 0.2 0.0 16.4 5.6 6.4 Lower medium-income 7.6 0.9 8.3 0.5 0.3 0.1 11.8 3.6 5.3 countries Upper-mediumincome 10.9 4.1 2.6 2.7 1.3 0.2 3.2 1.7 2.4 countries U Least developed 10.8 0.7 14.5 0.3 0.4 0.0 22.3 8.8 11.3 -nX countries E. Landlocked developing 8.5 0.6 14.1 0.3 0.3 0.0 19.3 9.5 12.8 countries 6 3- Small island developing 45.0 11.8 3.8 9.6 2.6 0.1 4.4 2.6 5.4 vi -k economies n Large developing 5.4 0.6 8.3 0.4 0.2 0.0 10.6 3.9 7.0 CL countries (>40 $ million) -& "Sum of emigrants from high-income countries, developing countries, and dependent territories and emigrants who did not report their country of birth. Source: Authors' analysis based on data from Docquier and Marfouk (2006). skilled average was 0.6 percent. Among the 10 countries where the schooling gap is below 1.5, the skilled average was 16 percent (much higher than the average of 6 percent for all developing countries). An increase in the education level of the native population generates a less than proportional increase in the education level of emigrants. Thus, the school- ing gap decreases with a rising human capital level in the country of origh6 In 2000, the schooling gap ranged from 1 in Turkey and Mexico to 92 in Niger. The correlation between the log of the schooling gap and the log of the proportion of educated among natives is - 90 percent (figure 4). The average schooling gap obviously decreases with national income (see table 2). Low-income countries have an index of 10.4, least developed countries an index of 13, and upper-middle-income countries an index of 1.7 (slightly above the average for high-income countries). This regularity explains why, other things being equal, poor countries tend to suffer more from brain drain. Stylized fact 4: Schooling gaps depend on destination choice. The choice of destination affects the size of the brain drain (see table 2). Remember that about three-quarters of skilled emigrants from developing countries live in selective-immigration countries (Australia, Canada, and the United States; see table 1).Thus, average emigration rates to selective-immigration countries are unsur~risingl~stronger than those to the EU15 and the rest of the OECD, where immigration policies focus mainly on family reunion and asylum seeking. FIGURE Schooling Gap and Natives' Human Capital 4. I 1 I -1.5 1 . 5 5 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 Percentage of educated among natives (in logs) Source: Authors' analysis based on data from Docquier and Marfouk (2006). 6. This relationship goes beyond a pure tautological composition effect (i.e. when 100 percent of natives are skilled, the skilled emigration rate equals the average emigration rate and the schooling gap is equal to one). Docquier, Lohest, and Marfouk 209 "Bilateral" schooling gaps also vary across destinations. On average, the schooling gap observed in selective-immigration countries was about twice as large as the gap observed in EUl5 and other OECD countries in 2000. Thus, countries that send many migrants to North America and Australia are likely to exhibit stronger schooling gaps than the others. Although many economic and institutional factors may explain these differ- ences (skill premia, welfare programs, etc.), increasingly quality-selective irnmi- gration policies are likely to play an important role. Since 1984, Australian immigration policy has officially privileged skilled workers, with candidates being selected according to their prospective contribution to the Australian economy. Canadian immigration policy follows similar lines, resulting in an increased share of highly educated people among the selected immigrants. For example, in 1997, 50,000 professional specialists and entrepreneurs immigrated to Canada along with 75,000 additional family members, representing 58 percent of the annual immigration flow. In the United States, since the Immigration Act of 1990 and the American Competitiveness and Work Force Improvement Act of 1998, the emphasis has been on the selection of highly skilled workers through a system of quotas favoring candidates with academic degrees and specific professional skills. The annual number of visas issued for highly skilled professionals (H-1B visas) increased from 110,200 in 1992 to 355,600 in 2000, with the entire increase due to immigration from developing countries. About half these workers now come from India. As argued in Antecol, Cobb-Clark, and Trejo (2003),except for immigrants from Central American countries, the U.S. selection rate is higher than the Canadian or Australian ones. In 1990, the differential between selective-immigration countries and the EU15 was even stronger. The evolution of the differential is partly due to the fact that a growing number of EUl5 countries (including France, Germany, Ireland, and the United Kingdom) have recently introduced programs to attract a qualified labor force through the creation of labor-shortage occupation lists (see Lowell 2002). German Chancellor Schroder announced plans in February 2000 to recruit additional specialists in information technology and by August 2001 German information, communication, and technology firms had the opportunity to hire up to 20,000 non-EU specialists for up to five years. In 2002, the French Ministry of Labor established a system to induce highly skilled workers from outside the EU to live and work in France, and the French government is replacing passive immigration policy with a selective-immigration policy. 111. E M P I R I C A LN A L Y S IOSF THE D E T E R M I N A NOTF ST H E A B R A I ND R A I N This section examines the determinants of average emigration rates and school- ing gaps using empirical regressions. In a two-equation system, the dependent variables are the logistic transformation of the average emigration rate and the log of the schooling gap. The dependent variable is In[ml(l-m)], where 210 T H E W O R L D H A N K E C O N O M I C R E V I E W 0 < m < 1 is the emigration rate. This increasing monotonic transformation expands the range of the variable from (0,l) to (-inf, +inf). Potential Explanatory Variables The economics literature on international migration distinguishes many poten- tial determinants of labor mobility. The regressions here use five sets of expla- natory variables that are common in the empirical literature and that capture traditional proximity and push-pull factors. Because current emigration stocks depend on past as well as present decisions about migration, the average level observed over a long period is used for each explanatory variable when the data are available. The first set, country size at origin, includes the log of the native population (residents plus emigrants), and a dummy variable for small island developing economies. Population is the average of the annual number of people residing in the home country during 1985-2000 and the total number of working-age emi- grants living in an OECD country in 1990 and 2000. Data on population size are from World Bank (2005) and data on emigrants are from the Docquier- Marfouk data set. Although emigrants are likely to exhibit different mortality and fertility patterns than natives, using the native population rather than resi- dent population minimizes the risk of endogeneity. An obvious reverse causality occurs between migration and the resident population. Residents include the immigrant population since immigrants cannot be split by age group and edu- cation level in non-OECD countries. The small island developing economies dummy variable is based on the recent United Nations classification.' A second set of variables accounts for the level of development of the sending country using the log of the proportion of post-secondary-educated natives. Again, using natives rather than residents reduces the risk of endogene- ity. However, the recent literature on brain drain and human capital formation suggests that natives' human capital may depend on emigration prospects (Mountford 1997; Stark, Helmenstein, and Prskawetz 1997; Beine, Docquier, and Rapoport 2001, forthcoming). The risk of reverse causality is important and requires using instrumentation techniques. Also considered are the log of gross national income (GNI)per capita in purchasing power parity, a dummy variable for the least developed countries, and a dummy variable for oil export- ing countries. The native proportion of those with a post-secondary education comes from the Docquier-Marfouk data set. Data on GNI per capita are from World Bank (2005) and are averaged for 1985-2000. The dummy variable for least developed countries is based on the recent United Nations definition. The third set captures the sociopolitical environment at origin. These are created from a mixture of two data sets on governance and fractionalization. These data sets provide many insights on the potential push factors for emigra- tion. Data on governance are given in Kaufmann, Kraay, and Mastruzzi (2003) 7. See http://www.un.org/special-rep/ohrlls/ohrlls/default.htm. Docquier, Lohest, and Marfouk 211 for 1996, 1998, 2000, 2002, and 2004. From the six available indicators in this data set, two are used: political stability and absence of violence and gov- ernment effectivene~s.~ The first indicator measures perceptions of the likeli- hood that the government in power will be destabilized or overthrown by unconstitutional or violent means, including domestic violence and terrorism. The second indicator measures the quality of public service provision, the quality of the bureaucracy, the independence of the civil service from political pressures, and the credibility of the government's commitment to policies. Both are normally distributed between -2.5 (bad governance) and 2.5 (good gover- n a n ~ e )All the available scores are averaged for each country. Indicators of . ~ religious fractionalization from Alesina et al. (2003) are also used. This vari- able gives the probability that two randomly selected individuals from a given country share the same religion. The indicator ranges from about 1 percent to 83 percent. In developing countries, religious diversity often gives rise to con- flicts (Hindus and Muslims in India; Catholics, Orthodox, and Muslims in the former Yugoslavia) or discrimination. Although some studies consider govern- ance as an endogenous variable, political and governance indices are treated as exogenous here. The fourth set of variables accounts for geographic and cultural proximity between developing and OECD countries. Since Greenwood (1969), many studies have stressed distance as a proxy for the monetary and psychic costs of migration. Three variables are distinguished for this purpose: distance from selective-immigration countries (Australia, Canada, and the United States), dis- tance from the EU15 members, and a dummy variable for landlocked develop- ing countries, which suffer from a lack of territorial access to the sea, remoteness, and isolation from world markets. Colonial links, by implying better information about the destination country and thus lower migration costs, also affect the cultural distance between former colonies and destin- ation countries. A dummy variable is used if the sending country is a former colony o f an OECD country or if it shares the same language as a selective-immigration country. The data come from Clair et al. (2004).Finally, to control for the choice of destination, a dummy variable is included if the main destination is a selective-immigration country or if the main destination is an EU1S member state. Econometric Issues The empirical model consists of two equations, one for the average emigration rate and one for the schooling gap. Although dependent variables are available for both 1990 and 2000, most or the explanatory variables are time-invariant (either by nature or because levels observed over a long period are averaged). 8. They are strongly correlated with the four remaining variables and with Transparency International's corruption perception index (www.icgg.org/corruption.cpi~2003.html). 9. Under certain circumstances a country's rating might exceed these thresholds. 212 T H E W O R L D B A N K E C O N O M I C REVIEW Because it would be impossible to understand the effect of time-invariant vari- ables-the variables of primary interest-using a panel regression model with country fixed effects, cross-section empirical models were estimated on 2000 data.'' In a first stage, the general model is estimated with all the potential determi- nants in both equations. The standard ordinary least squares (OLS)regressions are used with White-corrections for heteroskedasticity (model OLS-1). Eliminating nonsignificant variables gives the first set of OLS-robust estimators (model OLS-2). To account for the potential endogeneity of the educated pro- portion of the native population, the parsimonious model is then estimated using a two-stage least square procedure with instrumentation of the educated proportion of the native population (model IV-1). The excluded instruments are the lagged proportion of the educated among natives, and the amount of public education expenditures." To allow comparisons between these models, the same sample size of 108 cross-country observations is used. Finally, a new parsimonious model is estimated using the instrumental variable technique when the sample size is maximized. This model (IV-2) is based on 125 obser- vations for the first equation and 123 for the second. Empirical Findings The first two parsimonious models provide very similar and robust results (table 3). The sign and significance levels of all coefficients are stable, with R~ of about 70 percent and 90 percent, respectively. The exogeneity test12 in the IV-1 model reveals that the educated proportion of the native population cannot be considered exogenous in the first equation. This is consistent with the new brain drain literature, which posits the positive impact of migration prospects on human capital formation in developing countries. There is no endogeneity problem in the second equation. The Sargan test and Hansen J-test of overidentification confirm that both excluded instruments are relevant and valid. Consequently, the IV models seem appropriate for the first equation of openness. The OLS models provide good results for the second equation. The parsimonious model IV-2 uses the largest number of observations. Adding 20 percent of additional observations gives similar predictions for the majority 10. The model was also estimated using random-effect panel techniques and seemingly unrelated regressions. Results are similar and available on request from the authors. The Hausman test rejects the random-effect hypothesis compared with the fixed-effect model. Hence, the random-effect model is clearly a second-best option. Pooling 1990 and 2000 data or working with 1990 data also gives similar results. 11. Public expenditures in primary education (in U.S. dollars) is used. Other tests based on expenditures in secondary and tertiary education give similar results. 12. A Durbin-Wu-Hausman test is used for the first equation. Since the regressions indicate the presence of heteroskedasticity in the second equation of the schooling gap, a C-test was used to obtain a valid endogeneity statistic in a heteroskedastic-robust context (see Baum and Schaffer 2003). I d d 22d-m om I Docquier, - S I 'Ci Lohest, 222 a-v. mom I and I Marfouk 213 TABLE Continued 3. OLS-1 General model OLS-2 Parsimonious model 1V-1 Parsimonious model Larger sample model Variable Openness" Schooling gaph Openness" Schooling gapb Openness" Schooling gaph Openness" Schooling gapb Main destination = EUl5 Same language as a selective- immigration country Constant Observations Adjusted R-squared Overidentification test' Instrument relevance: p-value of F statistic Exogeneity testJ Note: Numbers in parentheses are standard errors. Due to heteroskedasticity, the IV method for the schooling gap equation is a general method of moments estimator. Heteroskedastic-robust standard errors for OLS. "Logistic transformation of the average emigration rate. b~choolinggap in logs. 'p-value of statistic: Sargan test for the openness and Hansen j test for the schooling gap. d~xogeneitytest of natives of proportion skill. p-value of 2:Durbin-Wu-Hausman test for the openness and C-test for the schooling gap. List of instruments: lagged level +public expenditures in primary education (in logs). Source: Authors' analysis based on data from Docquier and Marfouk (2006). Docquier, Lohest, and Marfouk 215 of variables, but it affects the significance of several variables. Eliminating explanatory variables in the parsimonious models retrieves observations from many countries particularly affected by poverty and political instability. Model IV-2 is thus referred for the first equation. Model OLS-2 provides interesting insights for the second equation. All the regressions reveal small values for the variance inflation factor, indicating no real collinearity problem in the regressions.l 3 The empirical analysis confirms that country size is a key determinant of openness (see stylized fact 2), but has no effect on the schooling gap. The average emigration rate decreases with population size and is significantly larger in small island developing countries. This confirms stylized fact 2. The level of development has a very strong effect on openness rates and schooling gaps. Although some collinearity is observed between natives' level of schooling, GNI per capita, the oil exporting dummy variable, and the least developed country dummy variable, the variance inflation factor is below the tolerated value. The proportion of post-secondary-educated natives is the most robust and best predictor of the degree of openness. In developing countries, the higher natives' level of schooling, the higher is the average rate of emigra- tion. This effect can be explained by the fact that educated people can afford to pay emigration costs (self-selection) and are more likely to be accepted in host countries with selective-immigration policies. Natives' level of schooling has a negative impact on the schooling gap. This is compatible with stylized fact 3. The effect on the schooling gap is quantitatively more important than the effect on openness. A simulation exercise reveals that the marginal impact of natives' human capital on the brain drain is always positive, whatever the country size. The lower the natives' level of schooling, the greater is brain drain. That explains why poor regions such as Sub-Saharan Africa and South Asia suffer from the brain drain. After controlling for human capital, GNI per capita has a moderately negative impact on the schooling gap under some spe- cifications. Model IV-2 also reveals that oil exporting countries exhibit lower emigration rates. The least developed country dummy variable is never significant. The sociopolitical environment has a significant impact on openness. In all regressions the religious fractionalization variable has a positive and significant impact on the schooling gap. As fractionalization often induces conflict in developing countries, this suggests that skilled migrants are more sensitive to ethnic and religious tensions. From model IV-2, average emigration rates are also higher in politically unstable countries. Government effectiveness as well as many other variables introduced in alternative specifications did not prove to be significant. Fractionalization and political instability are particularly strong in Sub-Saharan African countries. 13. The strongest co linearit!. concerns the main destination dummies (EUlS and select~ve-immigrationcountries). Proximity significantly affects openness and the schooling gap. The geo- graphic distance between origin countries and major destination regions reduces the emigration rate and augments the schooling gap (also comforming to stylized fact 1,that emigration rates and schooling gaps are negatively corre- lated). Skilled migrants are less sensitive to distance. Lack of territorial access to the sea and remoteness and isolation from world markets strongly reduce the degree of openness of landlocked developing countries. Proximity has a strong impact on the brain drain from Central America, Caribbean and Pacific island countries, and, to a lesser extent, Northern Africa. Unsurprisingly, being a former colony has a positive effect on openness. It has no significant impact on the schooling gap. The effect of colonial links is obtained only in the large samples, but it is highly significant. Countries that send most of their migrants to selective-immigration countries experience stronger schooling gaps. When the main destination is the E UI5, the effect is positive but less strong and the effect is not significant when the sample size is maximized. The literature on migrants' economic assimilation reveals that migrants get a high return on their language skills. Although Chiswick and Miller (1995) among others found a strong correlation between language skills and the earnings of educated migrants, the effect of linguistic proximity with selective-immigration countries on the brain drain is seldom significant. IV. C O N C L U S I O N The article presents new estimates of the brain drain experienced by developing countries based on a new data set that draws on census and register data col- lected in all OECD countries. The analysis starts with a simple multiplicative decomposition of the brain drain into two components: degree of openness of sending countries, as measured by average or total emigration rate, and school- ing gap, as measured by the relative education level of emigrants compared with natives. The approach based on such a decomposition is justified by the facts that no country has both strong openness and a high schooling gap and that these two variables vary with specific determinants. The degree of openness is found to increase with country smallness, natives' human capital, political instability, colonial links, and geographic proximity to major OECD countries. The schooling gap depends on natives' human capital, the type of destination countries (with or without selective-immigration pro- grams), distances, and religious fractionalization in the country of origin. Geographic proximity and natives' human capital have ambiguous effects on the brain drain (they increase openness and reduce the schooling gap). On the whole, the brain drain is stronger in countries that are not too distant from OECD countries and where the average level of schooling of natives is low. Taken together these results increase the understanding of the causes of brain drain. Small islands of the Pacific and the Caribbean clearly suffer from Docquier, Lohest, and Marfouk 217 their smallness and proximity to OECD countries. Proximity is also a key determinant of Central American brain drain. Sub-Saharan African countries combine various disadvantages such as a low level of development, high politi- cal instability, and religious and ethnic fractionalization. The brain drain results from multiple possible causes, many of which cannot be affected by public interventions (such as proximity, historical links, country size, and frac- tionalization). Focusing on areas that can be influenced by public policy, such as promoting education and improving the political climate at origin, could help to reduce the brain drain. This article benefited from helpful comments by Michel Beine, Skbastien Laurent, Caglar Ozden, Hillel Rapoport, and Maurice Schiff. The authors thank Jaime de Melo, Riccardo Faini, and two anonymous referees for helpful comments and suggestions on an earlier draft. Frkdkric Docquier is grateful for financial support from the Belgian French-speaking community's program Action de recherches concertkes (ARC 03108-302) and acknowledges financial support from the Belgian federal government (PA1grant P6107). Adams, R. 2003. "International Migration, Remittances and the Brain Drain: A Study of 24 Labor-Exporting Countries." Policy Research Working Paper 2972. World Bank, Washington, D.C. Alesina, A,, A. Devleeschauwer, W. Easterly, S. Kurlatand, and R. Wacziarg. 2003. "Fractionalization." Journal of Economic Growth 8(2):155-94. Antecol, H., D. A. Cobb-Clark, and S. J. Trejo. 2003. "Immigration Policy and the Skills of Immigrants to Australia, Canada, and the United States." Iournal of Human Resources 38(1):192-218. Barro, R. J., and J. W. Lee. 2001. "International Data on Educational Attainment: Updates and Implications." Oxford Economic Papers 53(3):541-563. Baum, C. H., and M. Schaffer. 2003. "Instrumental Variables and GMM: Estimation and Testing." The StataJournal3(1):1-31. Beine, M., F. Docquier, and H. Rapoport. 2001. "Brain Drain and Economic Growth: Theory and Evidence." Journal of Dez~elopmentEconomics 64(1):275-89. -. (forthcoming) "Brain Drain and Human Capital Formation in LDCs: Winners and Losers." Ecot~omicJournal. Carrington. W. J., and E. Detragiache. 1998. "How Big Is the Brain Drain?" IMF Working Paper 981 102. Washington, D.C.: International Monetary Fund. . 1999. "How Extensive Is the Brain Drain?" Finance and Development 36(2):46-49. Chiswick, B. R., and P. W. Miller. 1995. "The Endogeneity between Language and Earning: An International Analysis." Journal of Labor Economics 13(2):246-88. Clair, G., G. Gaullier, T. Mayer, and S. Zignago. 2004. "Notes on CEPll's Distances Measures." CEPII Explanatory Note." Paris: Centre d'Etudes Prospectives d'Informations Internationales. Cohen, D., and M. Soto. 2007. "Growth and Human Capital: Good Data, Good Results." Journal of Econotnic Growth 12(1):51-76. Commander, S., M. Kangasniemi, and L. A. Winters. 2004. "The Brain Drain: A Review o f Theory and Facts." Brussels Economic Review 47(1):29-44. de la Fuente, A., and R. Domenech. 2002. "Human Capital in Growth Regressions: How Much Difference Does Data Quality Make? An Update and Further Results." CEPR Discussion Paper 3587. London: Centre for Economic and Policy Research. Docquier, F.. and A. Marfouk. 2006. "International Migration by Education Attainment in 1990- 2000." In C. Ozden and M. Schiffeds., International Migration, Renrittances, and the Brain Drain. Washington,D.C.: The World Bank. Docquier, F., and H. Rapoport. 2007. "Skilled Migration: The Perspective o f Developing Countries." In J. Baghwati and G. Hanson eds., Skilled migration: prospects, problems and policies. New York: Russell Sage Foundation. Greenwood, M. J. 1969. " A n Analysis o f the Determinants o f Geographic Labor Mobility in the United States." Revrew of Economics and Statistics 5 l ( 2 )189-94. : Kaufmann, D., A. Kraay, and M. Mastruzzi. 2003. "Governance Matters 111: Governance Indicators for 1996-2002." Policy Research Working Paper 3106. World Bank, Washington,D.C. Lowell, L. B. 2002. Some Developmental Effects of the Internatronal Migration of Highly Skilled Persons. International Migration Papers 46. Geneva: International Labor Office. Mountford, A. 1997. "Can a Brain Drain Be Good for Growth in the Source Economy?" ]ournal of Development Eco?~omics53(2):287-303. Schiff,M. 2005. "Brain Gain: Claims about Its Size and Impact on Welfare and Growth Are Greatly Exaggerated." In C. Ozden and M. Schiffeds., International Migration, Remittances and the Brain Drain. Washington, D.C.: The World Bank. Stark. O., C. Helmenstein, and A. Prskawetz. 1997. " A Brain Gain with a Brain Drain." Economics Letters 55(2):227-34. Wickramasekera, P. 2002. "Asian Labour Migration: Issues and Challenges in the Era o f Globalization." International Migration Papers 57. Geneva: International Labor Office. World Bank. 2005. World Development Indicators 2005. Washington, D.C.: World Bank. Are Remittances Insurance? Evidence from Rainfall Shocks in the Philippines Dean Yang and HwaJung Choi D o remittances sent by overseas migrants serve as insurance for recipient households? In a study of how remittances from overseas respond to income shocks experienced by Philippine households, changes in income are found to lead to changes in remittances in the opposite direction, consistent with a n insurance motivation. Roughly 60 percent of declines in household income are replaced by remittance inflows from overseas. Because household income and remittances are jointly determined, rainfall shocks are used as instrumental variables for income changes. The hypothesis cannot be rejected that consumption in households with migrant members is unchanged in response to income shocks, whereas consumption responds strongly to income shocks in households without migrants. JEL codes: D81, F22, F32, 012, 015 Several facts motivate this study. First, life in developing countries is prone to many kinds of risk, such as crop and income loss due to natural disasters (weather, insect infestations, fire) and civil conflict. Second, international migration and remittance flows are substantial and growing. Between 1965 and 2000, individuals living outside their country of birth grew from 2.2 to 2.9 percent of the world population, totaling 175 million people in 2000.' The remittances that these migrants send to their countries of origin are an impor- tant but poorly understood type of international financial flow. In 2002, remit- tance receipts of developing countries totaled $79 bi~lion.~This amount Dean Yang (corresponding author) is an assistant professor at the Gerald R. Ford School of Public Policy and in the Department of Economics at the University of Michigan; his email address is deanyang@umich.edu. HwaJung Choi is a PhD candidate in economics at the University of Michigan; her email address is hwajungc@umich.edu. The authors are grateful for comments and suggestions from Becky Blank, John Bound, Jaime De Melo, Claudio Gonzalez-Vega, Caglar Ozden, Maurice Schiff, three anonymous referees, and seminar participants at Ohio State University's Rural Finance Program. The World Bank's International Migration and Development Research Program provided important research support. A supplemental appendix to this article is available at http://wber.oxfordjournals.orej. 1 . Estimates of the number of individuals living outside their country of birth are from United Nations (2002),whereas data on world population are from United States Bureau of the Census (2002). 2. The remittance figure is the sum of the "workers' remittances," "compensation of employees," and "migrants' transfers" items in the International Monetary Fund's International Financial Statistics database for all countries not listed as high income in the World Bank's income groupings. TEE WOKLT) BANK FCONOLIIC REVIEW, VOL. 21, ho. 2, pp. 219-248 do~:10.1093/wber/lhm003 Advance Access Publ~catlon9 Mav 2007 ( The Author 2007. Publ~shedbv Oxford Unlversrty Press on behalf of the Intern~t~onalBank for Reconstruction and Development /THE WORLD BANK.All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org 220 T H E W O R L D RANK E C O N O M I C REVIEW exceeded the total official development aid ($51 billion) and equaled roughly 40 percent of foreign direct investment (FDI)inflows ($189 billion) received by developing countries that year.3Understanding the functions of remittances for recipient households is necessary for weighing the benefits to origin countries of developed country policies liberalizing inward migration [as proposed by Rodrik (2002)and Bhagwati (2003),for example].4 What connection, if any, is there between the pervasiveness of risk in deve- loping countries and international remittance flows? Do remittances from over- seas migrants serve as insurance for relatives back home? To shed light on this question, this article examines how income shocks experienced by households in the Philippines affect their receipt of remittances from overseas. To break the simultaneity between income and remittances, rainfall shocks are used to instrument for changes in household income. In households with members who are overseas migrants, changes in income from domestic sources lead to changes in remittances in the opposite direction of the income change: remit- tances fall when income rises and remittances rise when income falls. In such households, the amount of insurance is large: roughly 60 percent of exogenous declines in income are replaced by remittance inflows from overseas. In con- trast, in households without overseas migrants, changes in income from dom- estic sources have no effect on remittance receipts. As a result, the hypothesis cannot be rejected that consumption in households with migrant members is unchanged in response to income shocks, whereas consumption responds strongly to income shocks in households without migrants. Numerous studies have examined the mechanisms through which house- holds in developing countries cope with risk. Among others, Townsend (1994), Udry (1994), Ligon, Thomas and Worall (2002), and Fafchamps and Lund (2003) have documented risk-pooling arrangements among households intended to smooth consumption in response to shocks. Households may also autonomously build up savings or other assets in good times and draw down these assets in hard times (Paxson 1992; Rosenzweig and Wolpin 1993; Udry 1994), increase their labor supply when shocks occur (Kochar 1999), or take steps (such as crop and plot diversification) to reduce the variation in their incomes (Morduch 1993). This article examines a mechanism for coping with shocks ex post on which previous micro-level studies have not focused: remittances from family members overseas. At the international level, it is commonly posited that remit- tance flows from overseas buffer economic shocks in migrants' home countries (e.g., Ratha 2003), but there have been relatively few empirical tests of this 3. Aid and FDI figures are from World Bank (2004).Although the figures for official development aid and FDI are likely to be accurate, by most accounts national statistics on remittance receipts are considerably underreported (see, e.g., Ratha 2003). so the remittance figure may be taken as a lower bound. 4. Borjas (1999)argues that the investigation of benefits accruing to migrants' source countries is an important and virtually unexplored area In research on migration. Yang and Choi 221 claim with micro-level household data." Related research on the role of domes- tic migration in pooling risk within extended families includes Lucas and Stark (1985), Rosenzweig and Stark (1989),and Paulson (2000). A key distinguishing facet of this article is its emphasis on credible identifi- cation of the effect of income shocks on international remittances. Studies of the impact of household income on remittance receipts use cross-sectional data, and so are subject to ~ o t e n t i a l lsevere biases in directions that are not ~ obvious a priori. Reverse causation is a major concern: productive investments funded by migrant remittances can raise household income, leading to positive correlations between household income and remittances. Alternately, remit- tances may reduce households' need to find alternative income sources, leading to a negative relationship between remittances and domestic-source income. Even if reverse causation from remittances to income in migrants' source households were not a problem, it would be difficult to separate the cross- sectional relationship between income and remittances from the influence of unobserved third factors affecting both income and remittances (e.g., the entre- preneurial spirit of household members). Two aspects of the empirical strategy are the key in resolving these identifi- cation problems. First, the focus is on income changes due to shocks that are credibly exogenous-changes in local rainfall-so that bias due to reverse cau- sation is not a ~ o n c e r nBut . ~ the estimated impact of economic shocks in cross- sectional data is still likely to be biased, because the likelihood of experiencing a shock may be correlated with time-invariant household characteristics (in other words, omitted variables are still a concern). For example, if shocks occur more frequently in poorer areas, and more remittances generally flow to poor areas, estimates of the impact of income on remittances will be biased in a negative direction. So the second crucial aspect of this article is its use of panel data, so that estimates of the impact of income shocks can be purged of the influence of unobserved time-invariant household characteristics that are jointly related to remittances and to the likelihood of experiencing a shock. Estimation of the impact of shocks focuses on how shocks are related to changes in remittances rather than on the level of remittances. Section I considers the theoretical role of international remittance flows in sharing risk across family members in different countries. Section I1 describes the data used and provides empirical results. Section I11 discusses some of the policy implications of the findings and recommendations for future research. Further details on the household data sets are provided in the supplemental appendix, available at http://wber.oxfordjournals.org/. 5. On the international macroeconomic level, Yang (2006a) documents that international financial flows (including remittances) at the country level respond positively to economic losses due to hurricanes. For studies with micro-data, see Brown (1997)and Gubert (2002). 6. Other research using rainfall shocks as instruments includes Paxson (1992), Munshi (2003), Miguel (200i),and Maccini and Yang (2006). 222 T H E W O R L D B A N K E C O N O M I C K F V I t W When a household experiences a negative income shock, how should we expect remittance receipts from overseas to change? A basic theoretical result is that if there is a Pareto-efficient allocation of risk across individual entities (in this case, individual household members) in a risk-sharing arrangement, individual consumption should not be affected by idiosyncratic income shocks. Consider households consisting of two members, indexed by i E {1,2). Let one household member be located in the origin household in the Philippines and the other household member be located overseas. Assume that both house- hold members work and are able to send funds back and forth to each other. Individuals have an uncertain income in each period t, y:,, depending on the state of nature st E S. Household member i consumes c:,, and experiences within-period utility of ~,(c:,)at time t. Let utility be separable over time, and let instantaneous utility be twice differentiable with U: > 0 and U: < 0. For the allocation of risk across household members to be Pareto-efficient, the ratio of marginal utilities between members in any state of nature must be equal to a constant: u;(cl,) - w2 , -- for all st and t, u;(c:) - w1 where wl and w2 are the Pareto weights of members 1 and 2. Household members' marginal utilities are proportional to each other, and so consumption levels between members move in tandem. Let utility be given by the following constant absolute risk aversion func- tion: Then, following Mace (1991), Cochrane (1991), Altonji, Hayashi, and Kotlikoff (1992), and Townsend (1994), a relationship between individual household member i's consumption and average consumption across the house- hold members, E,, is obtained by: Efficient risk sharing implies that an individual's consumption level depends here only on mean consumption in the household, c,, and an effect determined by the individual's Pareto weight relative to the other's. Because this latter term is constant over time, changes in consumption for each individual will depend Yang and Choi 223 only on the change in mean household consumption. Said another way, indi- viduals face only household-level risk. How might this within-household (but cross-country) risk sharing be carried out in practice? It is simple to imagine that an individual sends remittances to the other household member when that member experiences a negative shock.7 Adapting Fafchamps and Lund (2003), let consumption of individual i in state st be the sum of income y:t and net inflows of remittances ri: So then equation (3)can be rewritten as: This equation can be transformed into an empirically testable specification as follows. First, separate income y:, into: where jj' is the permanent component of income and z:, is the transitory com- ponent of income. Only the transitory component depends on the state of the world. The function of Pareto weights and the permanent income component y' can be captured by an individual fixed effect, y,. The mean household consumption level, c,,, can be represented by a time effect, 4,. Also, allow a random com- ponent, E , ~ , a mean-zero error term. Then equation (4)becomes: The empirical test of this article is based on equation (6), where the outcome variable is remittances received from overseas. The focus is on a par- ticular type of transitory shock, z:,, changes in income from domestic (Philippine)sources, using rainfall shocks as the instrumental variable. There are two key questions of interest. First, is the coefficient on remit- tances with respect to domestic income z:, less than zero? If yes, then this is the evidence that at least some insurance is taking place. Second, can the null hypothesis of full insurance be rejected, that is, that the coefficient on zit is equal to negative one? 7. micro-economic studies among households of the insurance role of gifts and remittances include Lucas and Stark (1985),Ravallion and Dearden (1988),Rosenzweig and Stark (1989). Platteau (1991), and Cox, Eser, and Jimenez (1998). This section first describes the data and sample constructions and provides descriptive statistics on the sample households. It then discusses the regression specification and some empirical issues and presents empirical results. Finally, tests are conducted of potential violations of the instrumental variable exclu- sion restriction and of an important omitted variable concern. Data and Sample Construction The empirical analysis uses data from linked household surveys conducted by the Philippine National Statistics Office covering a nationally representative household sample: the Labor Force Survey, the Survey on Overseas Filipinos, the Family Income and Expenditure Survey, and the Annual Poverty Indicators Survey. The Labor Force Survey is administered quarterly to inhabitants of a rota- ting panel of dwellings in January, April, July, and October. The other three surveys are administered with lower frequency as riders to the Labor Force Survey. Usually, one-fourth of dwellings are rotated out of the sample in each quarter, but the rotation was postponed for five quarters starting in July 1997, so that three-quarters of dwellings included in the July 1997 round were still in the sample in October 1998 (one-fourth of the dwellings had just been rotated out of the sample). The analysis here takes advantage of this postponement of the rotation schedule to examine changes in households over the 15-month period from July 1997 to October 1998. Survey enumerators note whether the household currently living in the dwelling is the same as the household surveyed in the previous round; only dwellings inhabited continuously by the same household from July 1997 to October 1998 are included in the sample for analysis. Because the impact of domestic income shocks on remittance receipts is likely to vary according to whether households had migrant members, households that reported having one or more members overseas in June 1997 and households that did not are analyzed separately. The comparison between migrant and non-migrant house- holds should be taken as merely suggestive because households with and without migrants differ in many ways. Thus, while differences in results between these two groups of households could be due to whether they have migrants, it could also be due to other factors such as differential access to other risk-coping mechanisms (savings, credit, informal and formal insurance). Rainfall data used in constructing instrumental variables for household domestic income were obtained from the Philippine Atmospheric, Geophysical, and Astronomical Services Administration. Daily rainfall data are available for 47 weather stations, often as far back as 1951. Rainfall variables are con- structed by station separately for the two distinct weather seasons in the Philippines: the dry season from December through May and the wet season from June through November. Monthly rainfall is calculated by summing daily Yatzg and Choi 225 rainfall totals, with missing daily values replaced by the average daily totals in the given station-month for stations that had 20 or more daily rainfall records. For station-months with less than 20 daily rainfall records, monthly rainfall for the station is taken to be the monthly rainfall recorded in the nearest station with 20 or more daily rainfall records. Seasonal total rainfall for each station in each year is obtained by summing monthly rainfall for the months in each wet and dry season. Rainfall shock variables for a given season are constructed as rainfall in that season (in thousands of millimeters) minus rainfall in the same season the year before. Households are assigned the rainfall data for the weather station geo- graphically closest to their local area (specifically, the major city or town in their survey domain), using great circle distances calculated using latitude and longitude coordinates. Because some stations are never the closest station to a particular survey domain, a total of 38 stations are represented in the empirical analysis. The supplemental appendix provides other details on the household surveys and the construction of the sample for analysis (available at http://wber.oxford- journals.org/). Characteristics of Sample Households Table 1 presents summary statistics for the 27,881 households used in the empirical analysis, separately for migrant and non-migrant households. Migrant households are those with overseas workers in June 1997. The 1,655 migrant households represent 5.9 percent of the sample households. The table also presents rainfall data to provide a sense of the instrumental variables used. The rainfall data are deviations (in thousands of millimeters) from the historical mean of each station for the dry and wet seasons. The dry season immediately before the first observation for each household (year 1, with income from January to June 1997) runs from December 1995 to May 1996, and the wet season is from June to November 1996. Correspondingly, the dry season for the second observation for each household (year 2, with income from April to September 1998) is December 1996 to May 1997, and the wet season is from June to November 1997. For migrant households, the dry season in year 1 was on average wetter than normal, with a mean deviation across households of 0.28. In year 2, dry season rainfall was more typical, with a mean deviation of 0.03. Therefore, the mean household experienced a decline in dry season rainfall between years 1 and 2: the mean change in the dry season deviation across households is -0.25. The wet season for year 1 was only slightly wetter than normal, with a mean devi- ation across households of 0.07. In year 2, wet season rainfall was much dryer than normal, with a mean of -0.48. The mean household thus experienced a substantial decline in wet season rainfall between years 1 and 2 of -0.55. These declines in rainfall between the 2 years have been attributed to the 1997 El Niiio weather phenomenon. The mean of the rainfall variables for the TABLE Characteristics of Sample Households, 1997 1. Migrant households Non-migrant households -- - Standard 10th 90th Standard 10th 90th Mean deviation percentile Median percentile Mean deviation percentile Median percentile Rainfall variables (thousands of millimeters) Dry season year 1 Dry season year 2 Change between dry season years 1and 2 Wet season year 1 Wet season year 2 Change between wet season years 1 and 2 Outcome variable Change in household domestic income as share of initial household income Change in household remittance receipts as share of initial household income Change in household expenditure as share of initial household expenditures Change in indicator for overseas worker in household Household financial statistics (January-June 1997) Total expenditure Total income Total income per capita in household Domestic income 58,067 80,815 7,971 38,310 120,317 54,170 75,912 13,076 34,800 109,760 Remittance receipts 36,122 46,752 0 26,000 87,000 1,888 13,182 0 0 0 Remittance recei~tsas share of 0.39 0.31 0.00 0.37 0.85 0.02 0.10 0.00 0.00 0.00 total income Household size, including 6.2 2.4 3.0 6.0 9.0 5.2 2.3 3.0 5.0 8.0 overseas members, July 1997 Located in urban area 0.68 0.58 Household head characteristics July 1997 Age 49.9 13.9 32.0 50.0 68.0 46.7 14.1 30.0 45.0 67.0 Highest education level (indicators) Less than elementary 0.17 Elementary 0.20 Some high school 0.10 High school 0.22 Some college 0.16 College or more 0.14 Occupation (indicators) Agriculture 0.23 0.38 Professional job 0.08 0.06 Clerical job 0.13 0.11 Service job 0.05 0.07 Production job 0.14 0.26 Other 0.38 0.12 Does not work 0.00 0.00 Marital status is single (indicator) 0.03 0.03 3 Number of households 1,665 26,126 2 Note: Rainfall variables are deviations from historical average of each station in corresponding season. For year 1, dry season is December 1995-May % 1996 and wet season is June 1996-November 1996. For year 2, dry season is December 1996-May 1997 and wet season is June 199-November 1997. Rainfall data are collected from 38 stations. Total income includes both domestic-source income and remittances from overseas. Remittance receipts are from overseas only. Source: Authors' analysis based on Philippine National Statist~csOffice surveys described in text. N non-migrant households are generally quite similar to those for migrant households. The changes between years 1 and 2 for the wet and dry seasons are used as the instrumental variables for the change in domestic income in the empirical analysis that follows. The geographic distribution of the rainfall shocks is depicted graphically in figures 1 and 2. FIGURE1. Dry Season Rainfall Shocks, Philippines Note: Rainfall shocks are change in rainfall (in thousands of millimeters) between the December 1996-May 1996 season and the December 1996-May 1997 season. Each number in the figure is centered at coordinates of a rainfall station. Source: Authors' calculations based on data described in the text. Yang and Choi 229 FIGURE. Wet Season Rainfall Shocks, Philippines 2 -0.63 Note: Rainfall shocks are changes in rainfall (in thousands of millimeters) between the June 1996-November 1996 and the June 1997-November 1997 season. Each number in the figure is centered at coordinates of a rainfall station. Source: Authors' calculations based on data described in the text. Total expenditure and total income in the first period (January-June 1997) are higher in migrant households than in non-migrant households. Average total expenditure is 73,576 pesos ($2,830)for migrant households and 47,437 pesos ($1,825) for non-migrant household^.^ Average total income is 94,189 pesos ($3,623) for migrant households and 56,059 pesos ($2,156) for 8. Peso figures were converted to US dollars at the January-June 1997 rate of 26 pesos per dollar. non-migrant households. Remittances have a mean of 36,122 pesos ($1,389) in migrant households and 1,888 pesos ($73) in non-migrant households. Remittances amounted to 39 percent of household income for migrant house- holds and 2 percent for non-migrant households.' Average migrant household size is 6.2 members (including overseas members), whereas non-migrant household size is 5.2 members. Overall, 68 percent of migrant households are located in survey-defined urban areas, com- pared with 58 percent of non-migrant househo~ds.'~Heads are more educated in migrant households than in non-migrant households: around in 30 percent of heads in migrant households have at least a college degree, compared with only 20 percent in non-migrant households. Fewer household heads worked in agriculture in 1997 in migrant households (23 percent) than in non-migrant households (38 percent). Heads in migrant households are also slightly older (mean age of 49.9) than heads of non-migrant households (46.7). Identification Strategy Overseas remittance responses to exogenous changes in household income from domestic sources are examined to determine whether households use remittances as insurance. The following identification strategy is used. The remittance amount received by each household at time t is determined by household characteristics that are constant over time (such as completed education of household adults), time-variant household characteristics (such as household size), time effects common to all households (such as changes in remittance regulations or the nationwide economic situation), and time-varying household income from domestic sources. In addition, there may be time effects that vary systematically according to household characteristics, as when a nationwide economic shock has differential effects on better-educated and less-educated households. For household h at time t, the remittance equation is as follows: where Rht is the household remittance receipts from overseas, Yhtthe house- hold income from domestic sources, &,, a vector of time-variant household characteristics, Wh,a vector of time-invariant characteristics, y, the time effect for period t, Tt a dummy variable for each time period, the T, Whterm allows the time effect to vary systematically with household time-invariant character- istics, and ~h~ a mean-zero error term. 9. Remittance receipts of non-migrant households are not zero because households can receive remittances from non-household memhers (such as distant relatives or friends). 10. Although these may seem to be h~ghurban percentages, the Philippine National Statistics Office appears co use a hroad defin~t~onof an ~lrbanarea, and many areas classified as "urban" are likely to be closely linked to adjacent agricultural areas. Yang and Choi 23 1 The coefficient of interest is P, the coefficient on domestic income Yht. If remittances help insure households from losses of domestic income, this coeffi- cient should be negative. Its magnitude represents the replacement rate of dom- estic income by remittances from overseas. Although a rich variety of information is available on household character- istics that might be included in the vector hr,serious problems remain with obtaining an unbiased estimate of P. First, there is reverse causation: domestic income can itself be a function of remittances, when remittances help fund household entrepreneurial investments. Alternately, households receiving insur- ance through remittances could exert less effort and thus earn lower incomes (Azam and Gubert 2005), leading to a negatively biased estimate of the effect of income on remittances. Such endogeneity concerns motivate this article's empirical strategy-the use of panel data and the use of rainfall shocks as instruments for income. In this context, the focus can be on the impact of exogenous changes in income on changes in remittances. There should there- fore be no concern that the changes in income are endogenous with respect to remittance receipts due to moral hazard or other reasons. Another concern is omitted-variable bias: unobservable household character- istics (say, the entrepreneurial spirit of household members) are likely to jointly determine domestic income and remittances. The identification strategy focuses on reducing bias generated from simultaneity and omitted variables. With two observations for each household, first differences can be used to control for the influence of unobservable household characteristics. Rewriting equation (7)separately for each of the 2 years (1997and 1998)yields: To eliminate the influence of unobservable household time-invariant charac- teristics, Wh, first differences can be taken by subtracting equation (8) from equation (9),and rearranging to obtain: It still remains to deal with time-variant heterogeneity, A&98, and with reverse causation. To do so, the change in household domestic income, AYh98, is instrumented by the change in rainfall over the study period. The change in rainfall should be a valid instrument, as it is likely to have an important effect 232 T H E W O R L D B A N K E C O N O M I C REVIEW on household income in a country such as the Philippines, where most house- holds owe their livelihoods either directly or indirectly to agriculture. In addition, it is also plausible that rainfall affects remittances primarily through the change in household income (the instrumental variable exclusion restric- tion).'' The sample is not limited to households in rural areas, since (as dis- cussed earlier) the definition of an urban area used in the surveys is quite broad, and many households classified as urban (43 percent) do report non-zero agricultural income. Nor would it be desirable to limit the sample to households with agricultural income. Negative agricultural income shocks should reduce demand on the part of agricultural households for non- agricultural goods and services, so that negative rainfall shocks should also affect income in non-agricultural households. The first stage regression is: where ARAIN-DRYh98 and ARA1N-WETh98 are the changes in rainfall in the dry and wet seasons relevant for the change in income between 1997 and 1998, and Wh98 is a mean-zero error term. The inclusion of Wh in the regression allows for heterogeneity in the time trend from 1997 to 1998 across house- holds depending on time-invariant characteristics. The predicted change in income from equation (ll),A Y , ? ~can ~ , be substi- tuted for AYh98in equation (lo),and various terms rewritten to obtain: where 5, a constant term, substitutes for the change in year effects, v for the change in the vector of coefficients (x9*-x ~ ~and, the new error term ) 77h98 for the remaining terms from equation (lo),&hs8 -&h97 + g A &98.(NOWthat the change in household income is instrumented by rainfall, it is plausible to assume that shocks to other household outcomes, A &98, are orthogonal to A ? and so can safely be included in the error term.) ~ ~ ~ Equation (12)is the estimating equation used in the regression analysis. The variables included in the vector of controls, Wh, are a set of household charac- teristics in the first period (January-June 1997): an indicator for urban location; five indicators for the household head's highest level of education completed (elementary, some high school, high school, some college, and college or more; less than elementary omitted); six indicators for head's occu- pation (professional, clerical, service, production, other, not working; agricul- tural omitted); and log per capita household income. 11. Robustness checks, conducted later, examine and reject the existence of important alternative channels (other than household income) for rainfall's effects on remittances. Yang and Choi 233 Regression Results This section describes the impact of rainfall shocks on changes in household domestic income. It then presents the impact of changes in household domestic income (instrumented by rainfall shocks) on changes in household remittance receipts from overseas. It also looks at the impact of instrumented domestic income on total household expenditures and at the number of migrant members in the household. Impact of Rainfall on Domestic Income (First-Stage Estimates) Regression results from the first stage-predicting changes in domestic income using rainfall shocks, as in equation (11)-are presented for two specifications in table 2. The dependent variable in both regressions is the change in house- hold income from domestic sources between the January-June 1997 and April-September 1998 reporting periods, divided by initial (January-June 1997)total household income. (For example, a change amounting to 10 percent of initial income is expressed as 0.1.'~)The mean of the dependent variable is 0.03 for both migrant and non-migrant households, indicating that both types of households experienced increases in domestic-source income on average between the two time periods. Spatial correlation in the outcome variables is likely to be a problem in this analysis, biasing ordinary least squares (OLS) standard error estimates down- ward (Moulton 1986). In particular, the concern is correlation among error terms of households associated with the same weather station, because the rain- fall instrumental variables vary only at this level. So standard errors allow for an arbitrary variance-covariance structure within the coverage areas of 38 weather stations (standard errors are clustered by weather station coverage area). The first column in table 2 presents the coefficient estimates where the rain- fall shock variables are changes in rainfall in the dry and wet seasons. The coefficient on the dry season rainfall shock is positive and statistically signifi- cant at the 5% level. The coefficient on the wet season shock is negative but is not statistically significant. A decline of 500 mm of rainfall in the preceding dry season leads to a 3.3 percentage point decline in initial household domestic income. The F-statistic for the test of joint significance of the rainfall variables is 3.150, with a P-value of 0.054. 12. Dividing by pre-crisis household income achieves something similar to taking the log of an outcome: normalizing to take account of the fact that households in the sample have a wide range of income levels and allowing coefficient estimates to be interpreted as fractions of initial household income. We choose to normalize outcome variables in this way (rather than taking the log) because some second-stage outcome variables (in particular, remittances) often take on zero values. Results are robust to express the dependent variable as the level of income (in pesos) rather than as shares of initial income. T A B L E Impact of Rainfall Shock on Domestic Income, 1997-98: OLS 2. Estimates, First Stage of Instrumental Variable Regression Regression -- Variable (1) (2) Dry season rainfall shock (thousands of millimeters) Square of dry season rainfall shock (thousands of millimeters) Wet season rainfall shock (thousands of millimeters) Square of wet season rainfall shock (thousands of millimeters) Household head characteristics Highest education level (indicators) Elementarv Some high school High school Some college College or more Occupation (indicators) Professional job Clerical job Service job Production job Other job Does not work Household characteristics Log income per capita in household -0.355"'" (0.031) -0.365"'" (0.030) Located in urban area 0.126""" (0.018) 0.115'''" (0.017) F-statistic: joint significance of 3.150 6.780 rainfall shock variables P-value 0.054 0.000 Number of observations 27,781 27,781 R~ 0.03 0.03 "Significant at 10 percent level; "Significant at 5 percent level; ""Significant at 1 percent level. Note: Dependent variable: change in household domestic income as share of initial household income. Each column reports the results of a first-differenced regression. Numbers in parentheses are standard errors, clustered by rainfall station. Domestic income (January-June 1997lApril- September 1998) is household total income excluding remittances from overseas expressed as a fraction of initial (January-June 1997) total household income. Rainfall shocks are changes in rainfall between first and second period. Omitted occupation indicator is agricultural job. Omitted education indicator is less than elementary. See table 1 for other variable definitions. Source: Authors' analysis based on Philippine National Statistics Office surveys described in text. The effect of rainfall on household income may be non-linear, so column 2 of the table presents three results of a regression that also includes the square of the wet and dry season rainfall shocks. The dry season shock and its square are both now positive and statistically significantly. The wet season shock and Yang and Choi 235 its square are also both positive, but only the squared term is statistically sig- nificant. The four rainfall shock variables jointly appear to be quite strong as instrumental variables: the F-statistic for the test of the joint significance of the rainfall variables is 6.78, with a P-value of 0.000. In addition, the coefficients on the main effect and squared terms imply fairly substantial effects on income: a decline in dry season rainfall of 500 mm (roughly the standard devi- ation of the dry season change variable) leads to a 11.5 percentage point decline in household income. Because of the relative strength of the instruments as specified in column 2, this regression is used to create the first-stage prediction of the change in house- hold income, in the instrumental variables analyses. Instrumental Variables Estimates The instrumental variables estimates are based on regression equation 12, using the regression results in column 2 of table 2 to generate the predicted income change. Instrumental variable standard errors are calculated using a bootstrap procedure that takes into account the variation induced by the gener- ated regressor as well as geographic clustering of observations by rainfall station. OLS standard errors simply account for clustering of observations at the level of the rainfall station. REMITTANCERECEIPTS. Table 3 presents OLS and instrumental variable regression results where the outcome variable is the change in household remittance receipts from overseas between the January-June 1997 and April- September 1998 reporting periods, expressed as a share of initial (January- June 1997) household income. On average, both migrant and non-migrant households saw increases in remittances: the mean of the dependent variable was 0.07 for migrant households and 0.01 for non-migrant households. For migrant households, the OLS estimate of the impact of the change in household domestic income on the change in remittances is negative and stat- istically significant at the 10 percent level, but is small in magnitude (-0.080). In contrast, the corresponding instrumental variable estimate is negative, large in magnitude, and statistically significant at the 1 percent level: 62.9 percent of income declines are replaced by new inflows of remittances to the household. That said, the standard error on the -0.629 instrumental variable point esti- mate is large enough that the null hypothesis of full insurance (that the coeffi- cient is equal to negative one) cannot be rejected. The OLS and instrumental variable estimates of the impact of changes in household domestic income on changes in remittances are dramatically differ- ent, highlighting the importance of the instrumental variable approach. A number of factors are likely to help explain this difference. First, measure- ment error in domestic household income will attenuate the OLS coefficient (particularly as this is a regression in first differences). Second, reverse causa- tion may be at work. For example, increases in remittances may reflect 0 910 oogss .*ga &?boo """2"99999 00003 ~m-mo 00 cam sssssg I 00000 000000 --t\v~r """E"" 999999 30000fn ov~*rn- -An--- I I 12 I XXS 00 w-Sj ""w 99 00 C7 I Yang and Choi 237 increased investment in household entrepreneurial enterprises, leading to increased domestic income. This would lead the OLS coefficient to be biased in a positive direction. Finally, there may be omitted variables positively corre- lated with both the change in remittances and the change in income. For example, a need to accumulate resources for a large household purchase (such as a vehicle) or some other lump-sum payment (tuition, medical expenses) might lead to both increased remittances, increased domestic labor supply, and increased domestic income. Omitted variable stories such as these would also cause positive bias in the OLS coefficient compared with the instrumental variable. The contrast with the results for the non-migrant households (the last two columns of table 3) is striking. The OLS coefficient is essentially zero, whereas the instrumental variable coefficient is positive (0.056). Although the instru- mental variable coefficient is small, it is statistically significantly at the 5 percent level: exogenous increases in household income raise remittance receipts from overseas. Further analyses (described later) provide an expla- nation for this result. How valuable is the insurance provided by remittances for Philippine house- holds? One way to gage the welfare gain is to ask what fraction of income households would be willing to give up to reduce rainfall-driven income shocks, both positive and negative, by the amount indicated in column 2 of table 3 (62.9 percent). The actual distribution of rainfall shocks observed in the data is used to calculate predicted 1998 income due solely to this rainfall variation for all households in the data set (only dry season shocks are used because wet season shocks are not statistically significant in table 2). The distri- bution of predicted 1998 income shocks (relative to 1997 income) observed across households is then used to represent the underlying risk to be insured. The calculation assumes constant relative risk aversion utility, U(c)= cl- Y/ (1- y).13A household with a reasonable risk aversion parameter ( y = 1.5) should be willing to give up 0.24 percent of income to achieve income with this degree of smoothness. Although this number may seem small, it is large relative to the fraction of income such a household would be willing to give up to achieve complete smoothness, which is 0.28 percent under the same assumptions. l4 HOUSEHOLDEXPENDITURES.If remittances serve as insurance for migrant house- holds, changes in household expenditures should be relatively unresponsive to changes in household domestic income, because remittances respond so strongly (and in the opposite direction) to changes in household domestic 13. The theoretical section assumed constant absolute rlsk aversion utility for tractability of the emp~ricalderivation, but constant relative risk aversion is typically thought to better characterize indivrdual behavior under uncertainty. 14. In an analogous calculation, Lucas (1987)finds relatively small welfare gains from elimination of aggregate consumption risk in the United States. income. It is also of interest to explore whether expenditures are smoother in migrant households than in non-migrant households in the face of domestic income shocks. Table 4 presents the results from OLS and instrumental variable regressions - where the outcome variable is the change in household expenditures between the January-June 1997 and April-September 1998 reporting periods, expressed as a share of initial (January-June 1997) household expenditures. The mean of the dependent variable is 0.00 for migrant households and -0.04 for non-migrant households (on average, expenditures are roughly stable or slightly declining between the periods). The regression specifications are other- wise the same as those reported in table 3. The OLS results indicate that household domestic income is highly positively correlated with total expenditures for both migrant and non-migrant house- holds. For example, for migrant households, a 10 percentage point increase in domestic household income is associated with a 5.0 percentage point increase in total expenditure; the magnitude of the OLS coefficient is similar for non- migrant households. In the instrumental variable specification, however, the income coefficient for migrant households declines by roughly half (from 0.499 to 0.248) and also declines somewhat for non-migrant households (from 0.623 to 0.508). That the coefficient on the change inhomestic income in the migrant regression declines substantially in the instrumental variable specification (and is not stat- istically significant) is consistent with remittances playing an important role in helping these households maintain their expenditure levels when they experi- ence income shocks. That said, standard errors in the instrumental variable regressions are quite large (the equality of the OLS and instrumental variable coefficients cannot be rejected), so these results should be taken only as suggestive. The relative decline in the coefficient on the change in domestic income is larger for migrant households than for non-migrant households, although again standard errors are too large to allow strong conclusions. This is most appropriately taken as merely suggestive evidence that migrant households are better able to smooth expenditures in the face of exogenous income shocks. The comparison between migrant and non-migrant households should be taken as suggestive because households with and without migrants differ across many observed and unobserved characteristics. Differences in results between these two groups of households could be due to the presence or absence of migrants, but it could also be due to other differences such as variations in access to other risk-coping mechanisms (savings,credit, other types of informal and formal insurance). EFFECT OF SHOCK ON OVERSEASMIGRATION FROM THE HOUSEHOLD.DO exogenous income shocks driven by rainfall also affect whether a household has a member working overseas? Part of the insurance provided by migrants could T A B L4. Impact of Domestic Income Shock on Total Expenditure, 1997-98: OLS and Instrumental Variable Estimates E Migrant households Non-migrant households -- -- - Ordinary least Instrumental Ordinary least Instrumental Variable squares variable squares variable Change in domestic household income (as share of 0.499" :':'(0.071) 0.248 (0.171) 0,623:b:";i (0.087) 0.508*"' (0.156) initial household income) Household head characteristics Highest education level (indicators) Elementary Some high school High school Some college College or more Occupation (indicators) Professional job Clerical job Service job Production job Other job Does not work Household characteristics Log income per capita in household -0.113""'" (0.026) -0.1 10' (0.064) 0.069" (0.034) 0.019 (0.053) Located in urban area 0.033 (0.034) 0.044 (0.037) -0.012 (0.015) 0.006 (0.027) Number of observations 1,655 1,655 26,126 26,126 2 'Significant at 10 percent level; ""Significant at 5 percent level; """Significant at 1 percent level. Note: Dependent variable: change in household expenditure as share of initial household expenditures. Each column of table is a separate first- differenced regression. Instrumental variables for change in domestic household income are rainfall shocks in dry and wet seasons (see table 2 for first- 8 stage regression). Numbers in parentheses are standard errors. OLS standard errors are clustered by rainfall station; instrumental variable standard errors 2, are bootstrapped. See table 2 for other notes and table 1 for variable definitions. Source: Authors' analysis based on Philippine National Statistics Office surveys described in text. t., take the form of delayed return and extended periods of high overseas earnings if their origin households experience negative income shocks. This section shows whether income shocks affect whether a household has a member working overseas. The outcome variable is the change in an indicator for a household having an overseas worker between the July 1997 and October 1998 surveys. For migrant households, this indicator was equal to 1in the first period and could equal 0 or 1in the second period. The mean of the outcome variable is -0.38 for migrant households, meaning that in 38 percent of house- holds with a migrant member in July 1997, all migrant members had returned by October 1998. For non-migrant households, this indicator was equal to 0 in the first period. The mean of the outcome variable for non-migrant households is 0.02, meaning that 2 percent of initially non-migrant households had become migrant households by the second period. Table 5 presents the results from OLS and instrumental variables regressions. Specifications are the same as in tables 3 and 4. For migrant house- holds, both the OLS and instrumental variable coefficients on the change in domestic income are negative, but neither is statistically significant at conven- tional levels. There is no indication for migrant households that remittance responses to income shocks are in part explained by migrants' changing their return decisions. For non-migrant households, the OLS coefficient is close to zero and is not statistically significant. The instrumental variable coefficient is positive and statistically significant. The instrumental variable coefficient (0.075) indicates that a 10 percent increase in domestic income leads to a 0.75 percentage point increase of in the household's likelihood of having an overseas migrant. This is a large effect, given that the mean of the outcome variable among all initially non-migrant households is 2.0 percentage points. This positive causal impact of income on overseas migration among initially non-migrant households helps explain the positive impact of income on remit- tances in these households (table 3, last column). This may reflect the fact that international migration requires fixed up-front costs (such as fees to recruit- ment agencies), so that households facing credit and savings constraints become more willing or able to pay the fixed costs when current income increases. Robustness Checks This section discusses the evidence against alternative channels to income for rainfall's effects and against an important potential confounding factor- exchange rate changes in migrants' overseas locations. POTENTIALVIOLATIONS EXCLUSIONRESTRICTION. important concern when OF An instrumenting for changes in household income using rainfall variation is that rainfall shocks affect all households in a local area. Because of this, at least part of the effects found may be due to changes in locality-level economic TABLE. Impact of Domestic Income Shock on Indicator for Overseas Worker in Household, 1997-98: OLS and 5 Instrumental Variable Estimates Migrant households Non-migrant households -- Ordinary least Instrumental Ordinary least Instrumental Variable squares variable squares variable Change in domestic household income as share of -0.031 (0.021) - 0.068 (0.178) -0.001 (0.001) 0.075" * (0.020) initial household income Household head characteristics Highest education level (i~zdicators) Elementary 0.048 (0.047) 0.051 (0.054) 0.002 (0.002) -0.000 (0.002) Some high school 0.079 (0.057) 0.082 (0.054) 0.006""" (0.003) 0.001 (0.004) High school 0.037 (0.049) 0.043 (0.051) 0.006*** (0.003) -0.002 (0.004) Some college 0.066 (0.043) 0.076 (0.054) 0.013"" (0.005) -0.001 (0.006) College or more 0.017 (0.066) 0.038 (0.100) 0.004 (0.007) -0.024""" (0.011) Occupation (indicators) Professional job -0.112" (0.066) -0.102 (0.065) - 0.006 (0.007) -0.020*" (0.006) Clerical job -0.170""" (0.061) 0.163"; (0.062) -0.004 (0.004) - 0.015"" (0.005) Service job -0.045 (0.070) -0.040 (0.075) - 0.008** (0.003) - 0.017" " (0.004) Production job -0.108'' (0.059) -0.104' (0.054) -0.003 (0.003) -0.006**' (0.003) Other job -0.037 (0.050) -0.025 (0.062) 0.020"" (0.004) 0.001 (0.006) Does not work -0.082 (0.225) -0.070 (0.253) 0.236" (0.139) 0.224 (0.144) Household characteristics Log income per capita in household 0.010 (0.023) - 0.008 (0.072) 0.014"" (0.001) 0.040** (0.007) Located in urban area -0.013 (0.034) -0.006 (0.044) - 0.001 (0.002) -0.011 "* (0.003) 3 Number of observations 1,655 1,655 26,126 26,126 4 Q "Significant at 10 percent level; ""Significant at 5 percent level; """Significant at 1 percent level. CL Note: Dependent variable: change in indicator for overseas worker in household. Each column is a separate first-differenced regression. Instrumental 3 variables for change in domestic household income are rainfall shocks in dry and wet seasons (see table 2 for first-stage regression). Numbers in parenth- 0. eses are standard errors. OLS standard errors are clustered by rainfall station; instrumental variable standard errors are bootstrapped. See table 2 for other notes and table 1 for variable definitions. I\) Source: Authors' analysis based on Philippine National Statistics Office surveys described in text. $ conditions (such as wage rates), rather than merely to changes in household income.'"~his would be a violation of the instrumental variable exclusion restriction, the assumption that the rainfall instruments affect household remit- tances only through their effect on household income. This section tests for potential violations of the exclusion restriction. One way in which rainfall might affect remittances is through changes in the relative returns to various types of work, which could induce households to change their labor supply. This could be problematic if changes in household labor supply lead to changes in remittances independent of their effects on household income. For example, if adults in the household spend more time working, households may hire maids or nannies to provide child care, and remittances may rise to help pay for such help. Or households may invite older relatives to live with them and look after children, and remittances may rise to help support the larger number of household members. If such responses are empiri- cally important, the instrumental variable regression estimates of the impact of the change in domestic income on the change in remittances will be biased in directions that cannot be predicted in advance. To test whether such concerns have any basis, it is useful to test the stability of the instrumental variable regression coefficients to the inclusion of control variables for the change in various alternative channels. In particular, control variables are included for the change in total household hours worked and for the change in household size.I6 Any substantial change in the instrumental variable estimates when these control variables are included would cast doubt on the assumption that the effects of rainfall variability work primarily through changes in domestic income. Table 6 presents the results of this exercise. The coefficient estimates for regressions where the outcome variable is the change in remittances are very similar to those in table 3. For example, the coefficient in the instrumental vari- able specification for migrant households is -0.569 in table 6 (and is statisti- cally significant at the 1 percent level), compared with -0.629 in table 3. There appears to be little reason for concern that rainfall affects remittances through changes in household labor supply or household size independently of rainfall's effects on income. The results for the change in household expendi- ture (row 1)and for the change in the indicator for having an overseas migrant (row 3) are not substantially different from the previous results (tables 4 and 5).The same is true for non-migrant households. AN O~~ITTEDVARIABLECONCERN:CHANGESEXCHANGERATES.Another general IN identification concern arises because 1997-98 was a time of substantial 15. Rosenzweig and Wolpin (2000) raise concerns from using weather events as instrumental variables. 16. Hours worked in the past week are reported for all household members above the age of 10. The change is from July 1997 to October 1998. The change in household size is over the same time period and includes overseas members. Yang and Choi 243 T A B L E6. Impact of Domestic Income Shock on All Outcomes, 1997-98: Fixed Effect OLS and Instrumental Variable Estimates, Controlling for Change in Household Size and Labor Supply Migrant households Non-migrant households -ppp- Ordinary least Instrumental Ordinary least Instrumental Outcome squares variable squares variable Total 0 . 0 8 4 ' (0.043) -0.562" (0.204) -0.003 (0.002) 0.063$" (0.022) remittance Total 0.480"* (0.083) 0.305 (0.160) 0.621'" (0.090) 0.521"' (0.160) expenditure Overseas -0.030 (0.021) -0.006 (0.172) -0.001 (0.001) 0.079*"" (0.018) worker indicator Number of 1,655 1,655 26,126 26,126 observations "Significant at 10 percent level; ""Significant at 5 percent level; ""*Significant at 1 percent level. Note: Each cell presents coefficient estimate on change in domestic household income in a sep- arate regression. Instrumental variables for change in domestic household income are rainfall shocks in dry and wet seasons (see table 2 for first-stage regression). Each regression includes control variables for the change in number of household members and the change in hours worked by household members between 1997 and 1998, as well as other control variables included in tables 3-5 (coefficients not shown). Numbers in parentheses are standard errors. OLS standard errors are clustered by rainfall station; instrumental variable standard errors are boot- strapped. See table 2 for other notes and table 1 for variable definitions. Source: Authors' analysis based on Philippine National Statistics Office surveys described in text. economic fluctuation in the Philippines (and in other Asian countries) due to the Asian financial crisis. The Philippine economy experienced a decline in economic growth after the onset of the crisis in mid-1997. Annual real GDP contracted by 0.8 percent in 1998, following growth of 5.2 percent in 1997 and 5.8 percent in 1996 (World Bank 2004). The urban unemploy- ment rate (unemployed as a share of total labor force) rose from 9.5 percent in 1999 to 10.8 percent in 1998, whereas the rural unemployment rate went from 5.2 percent to 6.9 percent (Philippine Yearbook 2001, table 15.1). Of course, any effects of the domestic economic downturn common to all households are not an issue, because the regressions here use first-differenced variables, so that common economic shocks are captured in the constant term. In addition, the control variables for households' 1997 characteristics included in all regressions (education, occupation, income, and urban indicator) will help account for any differential effects of the 1997-98 crisis that differ across households by socioeconomic status. However, there is another important dimension of heterogeneity that is par- ticularly relevant for migrant households: fluctuations in the exchange rates faced by migrant members. The devaluation of the Thai baht in June 1997 set off a wave of speculative attacks on national currencies, primarily in East and Southeast Asia. Overseas Filipinos work in dozens of foreign countries, includ- ing many countries most affected by exchange rate shocks due to the 1997 Asian financial crisis, such as the Republic of Korea and Malaysia and, to a lesser extent, Taiwan, China, Singapore, and ~ a ~ a n . " An omitted variable concern arises if the 1997-98 exchange rate shocks experienced by households in particular areas happen to be correlated with the rainfall shocks in the same areas over the same period. If, for example, areas with greater declines in dry season rainfall (and thus greater declines in income) also had exchange rate shocks that allowed migrants to send more remittances, then the negative relationship between income and remittances would be overstated. To test whether such concerns are empirically important, the main regressions are repeated for migrant households with the change in the exchange rate (Philippine pesos per unit of foreign currency) experienced by the households' migrants included as a control variable (table 7). The change in the exchange rate is the average of the 12 months leading to October 1998 minus the average of the 12 months leading to June 1997, divided by the second number.'' None of the coefficients is substantially different from the corresponding coefficients in tables 3-5. The exchange rate shocks experienced by household migrants appear to be orthogonal to the rainfall shocks experi- enced by their origin households. There is no evidence that omitted variables bias due to correlation between exchange rate and rainfall shocks is a cause for concern. The incomes of households in developing countries are often highly exposed to environmental risk factors such as weather. At the same time, government- sponsored social insurance is generally poor or non-existent. How do house- holds in poor countries shield themselves from environmental risk? This article documents empirically that some households are able to insure themselves without direct government involvement by sending members to work overseas. Their remittances serve as insurance in times of negative income shocks. In households with overseas migrants, exogenous changes in income lead to changes in remittances of the opposite sign, consistent with an insurance motivation for remittances. In such households, the results show a replacement rate of household domestic income by remittances of roughly 60 percent. The null hypothesis of full insurance cannot be rejected. In contrast, changes in 17. Yang (2006b),examines the impact of these heterogeneous exchange rate shocks on return migration and on investment behavior in migrants' origin households. 18. For further discussion of the exchange rate shock measure, see Yang (forthcoming). Yang and Choi 245 TABLE Impact of Domestic Income Shock on All Outcomes, 1997-98: 7. Fixed Effect OLS and Instrumental Variable Estimates, Controlling for Exchange Rate Shock, Migrant Households Only Migrant households Ordinary least Instrumental Outcome squares variable -- - Total remittance Total expenditure Overseas -0.032 (0.021) -0.107 (0.176) worker indicator Number of observations *Significant at 10 percent level; ""Significant at 5 percent level; ""*Significant at 1 percent level. Note: Each cell presents coefficient estimate on change in domestic household income in a sep- arate regression. Instrumental variables for change in domestic household income are rainfall shocks in dry and wet seasons (see table 2 for first-stage regression). Each regression includes control variable for the exchange rate shock experienced by migrant members between 1997 and 1998, as well as other control variables included in tables 3-5 (coefficients not shown). Numbers in parentheses are standard errors. OLS standard errors are clustered by rainfall station; instru- mental variable standard errors are bootstrapped. See table 2 for other notes and table 1 for vari- able definitions. Source: Authors' analysis based on Philippine National Statistics Office surveys described in text. household income have no effect on remittance receipts in households without overseas migrants. A key question is whether remittance responses to income shocks depend on the performance or availability of alternative methods of coping with risk, such as asset sales, credit markets, and reciprocal transfer networks. In particular, the availability of other risk-coping mechanisms may depend on whether shocks are aggregate (shared by other households) or idiosyncratic (on average uncorrelated with other households). By focusing on income shocks driven by local weather changes, this article assesses the role of remittances as insurance in the face of aggregate shocks to local areas. One reason for the finding of such large responses of remittances to rainfall-driven income shocks could be that such shared shocks make it more difficult to access credit or interhousehold assistance networks that nor- mally help households cope with risk. For example, when a large fraction of households in a local area experiences a negative shock, the demand for credit may rise, pushing up local interest rates. Some substantial fraction of house- holds needing loans may thus be priced out of the credit market. In addition, there may be difficulties in smoothing consumption through asset sales when there are aggregate shocks, because other households simultaneously seek to sell their assets, driving down prices.'9 If local risk-coping mechanisms break down under aggregate shocks, remittance inflows from migrant household members may be used more heavily as a smoothing device. Whether remittances exhibit such large responses to income shocks when the shocks are idiosyncratic, or specific to given households, is therefore an important avenue for future research. An idiosyncratic shock to a given house- hold, if truly uncorrelated on average with shocks experienced by other house- holds, should have negligible effects on the quality of local risk-coping mechanisms, and so households should be better able to use such mechanisms than if the shock were aggregate. Remittances might not respond nearly as much to idiosyncratic shocks precisely because households should still have access to alternative local risk-pooling arrangements. These results provide additional justification for government policies facili- tating international migration and remittance flows. For migration-origin countries, greater opportunities for international migration and improvements in the ease of sending remittances should expand the extent to which remit- tances can serve as social insurance. Policies to ease international migration include provision of information and social services for migrants and their families left behind and oversight of recruitment agencies for overseas jobs. Policies to facilitate remittances include strengthening financial infrastructure and payment systems to lower the cost and broaden the reach of formal remit- tance channels. Migration destination countries also have numerous policy options, such as expanding immigration quotas and loosening restrictions on formal remittance flows to developing countries. To the extent that immigra- tion policies in the rich world remain relatively restrictive, however, in most countries, remittances will be an important source of insurance for only a small minority of households. Thus, the results reported here do not support the wholesale dismantling of existing systems of social protection in migration origin countries, which will remain important for the majority of households. Altonji, J.G., F. Hayashi, and L.J. Kotlikoff. 1992. "Is the Extended Family Altruistically Linked? Direct Tests Using Micro Data." Anzerican Econoniic Revielo 82(5):1177-98. Azam, Jean-Paul, and Flore Gubert. 2005. "Those in Kayes: The Impact of Remittances on their Recipients in Africa." Revue ~ c o n o m i ~ 56(6):1331-58. u e Bhagwati, Jagdish. 2003. "Borders Beyond Control." Foreign Affairs 82(1):98-104. Borjas, George. 1999. "The Economic Analysis of Immigration." In Orley Ashenfelter, and David Card eds., Handbook of Labor Economics. Vol. 3A. Amsterdam: North-Holland. Brown, Richard P.C. 1997 "Estimating Remittance Functions for Pacific Island Migrants." World Developnient 25(4):613-26. 19. This point has been made by Rosenzweig and Wolpin (1993), Fafchamps, Udry, and Czukas (1998),and Lim and Townsend (1998). Yang and Choi 247 Cochrane, John. 1991. "A Simple Test of Consumption Insurance." journal of Political Economy 99 (5):957-76. Cox, Donald, Zekeriya Eser, and Emmanuel Jimenez. 1998. "Motives for Private Transfers over the Life-Cycle: An Analytical Framework and Evidence for Peru." Journal of Development Economics 55(1):57-80. Fafchamps, Marcel, and Susan Lund. 2003. "Risk-sharing Networks in Rural Philippines." Journal of Development Economics 71(2):261-87. Fafchamps, Marcel, Christopher Udry, and Katherine Czukas. 1998. "Drought and Saving in West Africa: Are Livestock a Buffer Stock?" Journal of Development Economics 55(2):273-305. Gubert, Flore. 2002. "Do Migrants Insure Those Left Behind? Evidence from the Kayes Area (Western Mali)." Oxford Development Studies 30(3):267-87. Kochar, Anjini. 1999. "Smoothing Consumption by Smoothing Income: Hours-of-Work Responses to Idiosyncratic Agricultural Shocks in Rural India." Review of Economics and Statistics 81(1):50-61. Ligon, Ethan, Jonathan P. Thomas, and Tim Worall. 2002. "lnformal Insurance Arrangements with Limited Commitment: Theory and Evidence from Village Economies." Review of Economic Studies 69(1):209-44. Lim, Y., and Robert M. Townsend. 1998. "General Equilibrium Models of Financial Systems: Theory and Measurement in Village Economies." Review of Economic Dynamics 1(1):58-118. Lucas, Robert E. 1987. Models of Business Cycles. New York: Basil Blackwell. Lucas, Robert E.B., and Oded Stark. 1985. "Motivations to Remit: Evidence from Botswana." Journal of Political Economy 93(5):901-18. Maccini, Sharon, and Dean Yang. 2006. "Under the Weather: Health, Schooling, and Socioeconomic Consequences of Early-Life Rainfall." University of Michigan. Mace, B.J. 1991. "Full Insurance in the Presence of Aggregate Uncertainty." Journal of Political Economy 99(5):928-56. Miguel, Edward. 2005. "Poverty and Witch Killing." Review of Economic Studies 72(4):1153-72. Morduch, Jonathan. 1993. "Risk, Production, and Saving: Theory and Evidence from Indian Villages." Harvard University. Moulton, Brent R. 1986. "Random Group Effects and the Precision of Regression Estimates." Journal of Econometrics 32(3):385-97. Munshi, Kaivan. 2003. "Networks in the Modern Economy: Mexican Migrants in the U.S. Labor Market." Quarterly Journal of Economics 118(2):549-99. Paulson, Anna. 2000. "Insurance Motives for Migration: Evidence from Thailand." Northwestern University, Kellogg Graduate School of Management. Paxson, Christina. 1992. "Using Weather Variability to Estimate the Response of Savings to Transitory Income in Thailand." American Economic Review 82(1):15-33. Platteau, Jean-Philippe. 1991. "Traditional Systems of Social Security and Hunger Insurance: Past Achievements and Modern Challenges." In E. Ahmad, J. Dreze, J. Hills, and A. Sen eds., Social Security in Developing Countries. Oxford: Clarendon Press. Ratha, Dilip. 2003. "Workers' Remittances: An Important and Stable Source of External Development Finance." In Global Development Finance 2003: Striving for Stability in Development Finance. Washington, D.C.: International Monetary Fund. Ravallion, Martin, and L. Dearden. 1988. "Social Security in a 'Moral Economy': An Empirical Analysis for Java." Review of Economics and Statistics 70(1):36-44. Rodrik, Dani. 2002. Feasible Globalizations. NBER Working Paper 9129. Cambridge, Mass.: National Bureau of Economic Research. Rosenzweig, Mark, and Oded Stark. 1989. "Consumption Smoothing, Migration, and Marriage: Evidence from Rural India." Journal of Political Economy 97(4):905-26. Rosenzweig, Mark, and Kenneth Wolpin. 1993. "Credit Market Constraints, Consumption Smoothing, and the Accumulation of Durable Production Assets in Low-Income Countries: Investments in Bullocks in India." journal of Political Economy 101(2):223-44. Rosenzweig, Mark, and Kenneth Wolpin. 2000. "Natural 'Natural Experiments' in Economics." journal of Economic Literature 38(4):827-74. Townsend, Robert. 1994. "Risk and Insurance in Village India." Econometrics 62(3):539-591. Udry, Christopher. 1994. "Risk and Insurance in a Rural Credit Market: An Empirical Investigation in Northern Nigeria." Review of Economic Studies 61(3):495-526. United Nations. 2002. "International Migration Report." Department of Economic and Social Affairs, Population Division, New York. United States Bureau of the Census. 2002. "International Data Base." Washington, D.C. World Bank. 2004. World Development Indicators 2004. CD-ROM. Washington, D.C. Yang, Dean. 2006a. Coping with Disaster: The Impact of Hurricanes on international Financial Flows, 1970-2002. NBER Working Paper 12794. Cambridge, Mass.: National Bureau of Economic Research. -. 2006b. "Why Do Migrants Return to Poor Countries? Evidence from Philippine Migrants' Responses to Exchange Rate Shocks." Review of Economics and Statistics 88(4):715-35. Forthcoming. "International Migration, Remittances, and Household Investment: Evidence from Philippine Migrants' Exchange Rate Shocks." Economic journal. Measuring International Skilled Migration: A New Database Controlling for Age of Entry Michel Beine, Fre'de'ric Docquier, and Hillel Rapoport Recent data on international migration of skilled workers define skilled migrants by education level without distinguishing whether they acquired their education in the home or the host country. This article uses immigrants' age of entry as a proxy for where they acquired their education. Data on age of entry are available from a subset of receiving countries that together represent 77 percent of total skilled immigration to countries of the Organisation for Economic Co-operation and Development. Using these data and a simple gravity model to estimate the age-of-entry structure of the remaining 23 percent, alternative brain drain measures are proposed that exclude immigrants who arrived before ages 12, 18, and 22. JEL Codes: F22,J24 Recent data sets on international skilled migration (Carrington and Detragiache 1998; Adams 2003; Docquier and Marfouk 2004, 2006; Dumont and Lemaitre 2004) define skilled immigrants as foreign-born workers with university or post-secondary training. This definition, based on the country of birth, does not account for whether education was acquired in the home or the host country and may therefore appear either too inclusive or too exclusive depending on the intended use of the data. For example, the definition would seem too narrow for measuring the extent of a country's skilled diaspora, but may be too inclusive for estimating the fiscal cost of the brain drain for the Michel Beine is professor of economics at the University of Luxemburg and at Universiti. Libre de Bruxelles and a research fellow at the Center for Economic Studies, Institute for Economic Research at the University of Munich; his email address is mbeine@ulb.ac.be. Frkderic Docquier is a research associate at the Belgian National Fund for Economic Research, professor of economics at the IJniversite Catholique de Louvain (Belgium),and a research fellow at the Institute for the Study of Labor (Bonn) and the Center for Research and Analysis of Migration at University College London; his email address is docquier@ires.ucl.ac.be. Hillel Rapoport (corresponding author) is senior lecturer in economics at Bar-Ilan University, a member of EQUIPPE, Universites de Lille (EA CNRS 4018), and a research fellow at the Center for Research and Analysis of Migration at University College London; his email address is hillel@mail.biu.ac.il. This article is part of the World Bank Migration and Development Program, which provided financial support. The authors are grateful to Riccardo Faini, Mark Rosenzweig, and two anonymous reterees for comments. A supplemental appendix to this article is available at http://wber. oxfordjournals.org. THI..wo~1.11 BANK ECONOMIC REVIEW,VOL.21, NO. 2, pp. 249-254 doi:10.1093/wber/lhm007 Advance Access Publication 1June 2007 (., The Author 2007. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development 1THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org 250 T H E W O R L D BANK E C O N O M I C R E V I E W source country, which should consider as skilled emigrants only people who received post-secondary training in their home country. This article uses immigrants' age of entry as a proxy for where education was acquired. Data on age of entry are available from a subset of receiving countries that together represent 77 percent of total immigration of skilled workers to Organisation for Economic Co-operation and Development (OECD) countries. These data are used in a simple gravity model to estimate the age-of-entry structure of skilled immigration for the remaining 23 percent. These estimates can be used to establish alternative measures of the brain drain for both 1990 and 2000 by defining skilled immigrants as those who left their home country after age 12, 18, or 22. These corrected skilled emigration rates are by construction lower than those computed without age-of-entry restric- tions by Docquier and Marfouk (2006). Census and registry data were collected in a sample of OECD host countries for which information on immigrants' age of entry was available: the 1991 and 2001 Australian Censuses, the 1991 Belgian Census, the 1991 and 2001 Canadian Censuses, the 2000 Danish register, the 1999 French Census, the 2001 Greek Census, the 1991 and 2001 New Zealand Censuses, and the 1990 and 2000 U.S. Censuses. In 2000, the sampled countries hosted 77 percent of total skilled immigrants to the OECD. The sample is representative of the OECD in that it includes countries with different demographic sizes, regional locations, development levels, and immigration policy and traditions. In addition to the data on age of entry, the statistical sources provide harmonized bilateral information on migrants' age, education level, and country of birth for a total of 192 origin countries. An age-of-entry structure was constructed for skilled immigrants aged 25 or older at the time of the census to show what proportion of this population arrived after ages 12, 18, and 22. After zeros and a few suspicious observations were eliminated, 1,580 observations remained for each age threshold (1990 and 2000 included).' These observations are used to forecast the age-of-entry structure among the skilled foreign-born for the rest of the OECD area for which such information was not available. Obviously, an approach based on census data is not perfect. As Rosenzweig (2005, p. 9) explains, "information on entry year.. . is based on answers to an ambiguous question-in the US Census the question is 'When did you first come to stay?' Immigrants might answer this question by providing the date when they received a permanent immigrant status instead of the date when 1. Table S.1 of the supplemental appendix glves descriptive statistics on the observed proportions of skilled immigrants arrived after age Ju = 12, 18, and 22). It may be seen that immigrants arriving after ages 12, 18, and 22 represent on average 85.7 percent, 78.2 percent, and 69.1 percent of the total. Beine, Docquier, and Rapoport 25 1 they first came to the US, at which time they might not have intended to or been able to stay."2 Only surveys based on comprehensive migration histories can provide precise information about the location in which schooling was acquired. Still, the census is the only harmonized data source available. Survey data are not available for many countries, and when they are (for example, in the EU Labor Force Survey and in the European Community Household Panel), they do not provide representative cross-sectional pictures of immigrants' characteristics. Their coverage can be very small for countries with few emigrants. And with few exceptions (such as the New Immigrant Survey in the United States)they are not explicitly designed to capture immigrants' characteristics. Hence, extrapo- lating the entry age structure from national surveys can be misleading. To estimate the age structure of immigration for receiving countries for which information on age of entry is missing a simple gravity model is used. It aims to identify the determinants of the proportion of migrants from country i to country f with a tertiary education who arrived after age J = 12, 18, and 22. These bilateral proportions are denoted by a$. Since the proportions of skilled migrants who arrived after a given age lie between 0 and 1, it is appropriate to use a transformation so that the dependent variable is defined on (-m, +m). Therefore, H$= ln[cr,li/(l- a,:)] is used as the dependent variable. More pre- cisely, the following equation is estimated: where X$ (k = 1,. .., nif) is a collection of nif variables capturing proximity between origin and host countries, 23 (k = 1,.. ., ni) are origin country charac- teristics, and ~ f(k = 1, ...,nf) are host country characteristics. These variables k can affect the age-of-entry structure through self-selection mechanisms as well as through outselection mechanisms due to differences in host country immi- gration policies. In addition, a time fixed effect for 2000 is included to account for possible common trends in immigration policies. Included as the proximity variables in x$-are indicators of economic dis- parity between the home and the host countries, indicators of geographic and linguistic distances, and dummy variables for whether the pair of countries 2. In most countries (for example, Australia, France, and New Zealand), immigrants are simply asked about year of arrival or number of years of residence. In Canada, the way the question is asked creates an upward bias: "When did you become a landed immigrant for the first time," a landed immigrant being a person with the right to permanent residence. As one referee noted, another potential source of bias is the possibility that a person was born in country A, educated in country B, and lives in country C. share a colonial link. Included as origin country characteristics in Z: are democracy indicators and measures of public expenditures on primary, second- ary, and tertiary education. And included as host country characteristics in W; are indicators of social expenditures, education e ~ ~ e n d i t u r eand~ degree of s , openness to immigration. The variables used and their data sources are pre- sented in the supplemental appendix, which also discusses several econometric issues and reports all the estimate^.^ All coefficients are highly significant for the parsimonious models, robust across specifications, and affect the structure by age of entry in a very intuitive way. The proportion of skilled migrants that arrived after age J increases with economic disparity (as measured by the ratio of host to origin GDP per capita) and geographic distance and decreases with colonial and linguistic links. Education expenditures at destination favor family migration while social expen- ditures have the opposite effect. The higher the host country immigration rate, the higher the proportion of skilled migrants who arrived as children. Regarding origin country characteristics, the democracy index has no significant effect, and public education expenditures are not significant. Bringing together the census data on age of entry, which represent 77 percent of skilled immigrants to the OECD, and the estimated structure computed using the results of the parsimo- nious model for the remaining 23 percent5 provides alternative measures of the brain drain from which skilled immigrants who arrived before a given age are excluded. These are described in the next section. The Docquier and Marfouk (2006) data set indicates the total number of skilled emigrants from a given origin country i to a given host country f and the number of skilled residents in the home country. Denoting by Mif the number of skilled emigrants from country i to country f and by N, the number of skilled residents in the home country, the skilled emigration rate is then defined as the ratio of skilled emigrants to the total number of skilled natives (residents +emigrants). The method here is to multiply Mif by the estimated proportions of skilled migrants who left their home country after age J (J= 12, 18, 22) to obtain skilled emigration rates controlled for age of entry1 departure. The adjusted skilled emigration rates are then given by: 3. See the supplemental appendix for a discussion of possible reverse causality between our dependent variables and some of the explanatc~r~variables used in the regressions. 4. See tables S.2, S.3, and S.4. 5. See column 4 in tables S.2-S.4. Beine, Docquier, and Rapoport 253 where a;lf is the observed or the predicted proportion of skilled migrants who left after age J. The Docquier and Marfouk measures correspond to the special case where J = 0 or df= 1. As dfdecreases with J, the corrected rates for J = 12, 18, 22 are by construction lower than mrf .6 For the 192 sending countries in the sample, the mi? lm$ ratios range from 74.8 percent to 98.6 percent, the m y /m$ ratios from 59.4 percent to 97.9 percent, and the m y lm; ratios from 48.5 percent to 95.0 percent. The corre- lation between corrected and uncorrected measures is extremely high. Simple regression results of 4on m0 give R' values of 0.9775 for J = 22, 0.9895 for J = 18, and 0.9966 for J = 12. Table S.5 of the supplemental appendix focuses on the countries most affected by the brain drain. As table S.5 shows, control- ling for age of entry does not significantly alter the rankings by degree of brain drain intensity. IV. C O N C L U D I N E RG M A R K S Recent data sets on international migration of skilled workers define skilled migrants by education level independently of where the education was acquired. This leads to evaluations of the magnitude of the brain drain that may appear too broad or too narrow, depending on the objective for which the data are used. This article uses immigrants' age of entry as a proxy for where education was acquired. It combines observations and estimations and pro- poses alternative measures of the brain drain that exclude those who left their home country before age 12, 18, or 22. The corrected rates are obviously lower than those calculated without age-of-entry restrictions. However, the correlation between corrected and uncorrected rates is very high, and the country rankings by brain drain intensities are only mildly affected by the cor- rection. This implies that the results from recent empirical studies on the growth effects of the brain drain (for example, Beine, Docquier, and Rapoport 2001, forthcoming) are likely to be robust to the choice of corrected or uncor- rected skilled emigration rates. Adams, R. 2003. "International Migration, Remittances and the Brain Drain: A Study of 24 Labor-Exporting Countries." Policy Research Working Paper 2972. World Bank, Washington, D.C. Beine, M., F. Docquier, and H. Rapoport. 2001. "Brain Drain and Economic Growth: Theory and Evidence." Journal of De~~elopmentEconomics 64(1):275-89. . forthcoming. "Brain Drain and Human Capital Formation in LDCs: Winners and Losers." Economic /ournal. 6. The complete data set can be found at: http://siteresources.worldbank.org/lNTRES/Resources/ DataSet~BDwith~age~of~entry~DocquierRapoport.xls. Carrington, W.J., and E. Detragiache. 1998. "How Big Is the Brain Drain?" Working Paper 98-102. Washington, D.C.: International Monetary Fund. Docquier, F., and A. Marfouk. 2004. "Measuring the International Mobility of Skilled Workers (1990- 2000)." Policy Research Working Paper 3381. World Bank, Washington, D.C. . 2006. "lnternational Migration by Education Attainment in 1990-2000." In C. Ozden, and M. Schiff eds., International Migration, Remittances, and the Brain Drain. New York: Palgrave Macmillan. Dumont, J.C., and G. Lematre. 2004. "Counting Immigrants and Expatriates in OECD Countries: A New Perspective." Social, Employment, and Migration Working Papers 25. Paris: OECD Directorate for Employment, Labour, and Social Affairs. Rosenzweig, M.R. 2005. "Consequences of Migration for Developing Countries." Paper prepared for the United Nations Expert Group Meeting on International Migration and Development, Population Division, July 6-8, New York. The Anarchy of Numbers: Aid, Development, and Cross-Country Empirics David Roodman The recent literature contains many stories of how foreign aid affects economic growth. Aid raises growth in countries with good policies, or with difficult economic environ- ments, or outside the tropics, or on average but with diminishing returns. The diversity of the results suggests that many are fragile. Seven important aid-growth papers are tested for robustness, using 14 minimally arbitrary tests deriving mainly from differences among the studies themselves. This approach investigates the importance of potentially arbitrary specification choices while minimizing the arbitrariness in testing choices. All of the results appear fragile, especially to sample expansion. JEL Codes: F35, 023, 0 4 0 In early 1981, economist Edward Leamer gave a lecture at the University of Toronto, in which he bemoaned the state of econometrics. Econometrics sought the status of a science, with regressions as its analog for the reproduci- ble experiments of chemistry or physics. Yet an essential part of econometric "experimentation" was too often arbitrary, opaque, and unrepeatable. "The econometric art as it is practiced at the computer terminal involves fitting many, perhaps thousands, of statistical models.. . . This search for a model is often well intentioned, but there can be no doubt that such a specification search invalidates the traditional theories of inference" (Leamer 1983, p. 36). The way out of the quagmire, he argued, was for econometricians to explore larger regions of "specification space," systematically analyzing the relationship between assumptions and conclusions. One econometric debate with hallmarks of the syndrome Leamer describes is that on the effectiveness of foreign aid in developing countries. Since Griffin and Enos (1970), econometricians have parried over the question of how aid affects economic growth. Prominent in contemporary work, Burnside and David Roodman is a research fellow at the Center for Global Development; his e-mail address is droodman@cgdev.org.The author thanks Michael Clemens and William Easterly for advice; Craig Burnside, Lisa Chauvet, Jan Dehn, and Henrik Hansen for data and assistance; and Patrick Guillaumont, Stephen Knack, anonymous reviewers, and participants in an August 2003 seminar at the Center for Global Development for valuable comments. A supplemental appendix to this article is availableat http:Nwber.oxfordjournals.org/. THE WORLD BANK ECONOMIC REVIE\V, 21, VOL. NO. 2, pp. 255-277 doi:10.1093/wberllhm004 Advance Access Publication 13 May 2007 'i-,The Author 2007. Published by Oxford University Press on hehalf of the International Bank for Reconstruction and Development / THF WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org Dollar (2000, p. 847) conclude that "aid has a positive effect on growth in a good policy environment." Their evidence: the statistical significance in cross- country panel growth regressions of an interaction term of total aid received and an indicator of the quality of recipient economic policies (aid x policy). But Burnside and Dollar are just one voice among many. Collier and Dehn (2001), Collier and Dollar (2002, 2004), and Collier and Hoeffler (2004)cor- roborate their finding, whereas others challenge it. From the ongoing debate emerge several stories of the relationship between aid and growth, each turning on a particular quadratic or interaction term involving aid. The stories are not incompatible, but most of the papers support only one. Hansen and Tarp (2001) find that entering the square of aid drives out aid x policy and makes the simple aid term significant too: aid works on average, but with diminishing returns. Guillaumont and Chauvet (2001) also fail to find significance for aid x policy and instead offer evidence that aid works best in countries with difficult economic environments, characterized by volatile and declining terms of trade, low population, and natural disasters. In the same vein Collier and Dehn (2001) find that increasing aid cushions countries against negative export price shocks. Collier and Hoeffler (2004) offer a triple-interaction term: aid works particularly well in countries that are recovering from civil war and that have good policies. Last, Dalgaard, Hansen, and Tarp (2004)say that aid raises growth outside the tropics but not in them. These papers differ not only in their conclusions but in their specifications as well. Within the group there are two choices of period length in the panel data sets, three definitions of policy, three of aid, and four choices of control variable sets. Although probably none of the choices is made on a whim, these differences appear to be examples of what Leamer called "whimsy." From Learner's point of view the studies together represent a small sampling of speci- fication space. And few include much robustness testing. Without further analy- sis, it is hard to know whether the results reveal solid underlying regularities in the data or are fragile artifacts of particular specification choices. This article examines the possibility of fragility systematically. Since by the laws of chance any regression can be broken with enough experimentation, it is essential for credibility that the testing suite itself be minimally arbitrary. The tests derive from two sources: the choices present in the original specifica- tions, and the passage of time, which allows expansion of data sets (as in Easterly, Levine, and Roodman 2004). In all, regressions from seven of the most prominent studies are subjected to this systematic test suite. Section I reviews the approaches and conclusions of the studies that are tested for robustness, section I1describes the tests, and section 111reports the results. The hope has often arisen that a turn to the numbers would shed light on the questions of whether and when foreign aid works. In the view of Hansen and Roodman 257 Tarp (2000) the aid effectiveness literature has gone through three generations. The first generation essentially spans 1970-72 and mainly investigates the aid- savings link. Influenced by the Harrod-Domar model, in which savings is the binding constraint on growth, aid-induced saving is assumed to lead directly to investment and then to growth through a fixed incremental capital-output ratio. The second generation runs from the early 1970s to the early 1990s and directly investigates whether aid affects investment and growth. Hansen and Tarp argue that the ~re~onderanceof the evidence from these first two generations shows that aid increases total savings, but less than one for one, and that aid increases investment and growth. They suggest that studies with more pessimistic results, such as Modey, Hudson, and Horrell (1987),gained disproportionate attention precisely because they are contrarian. The third generation, which began with Boone (1994) and continues to this day, is the focus of this article. It brought several innovations. The data sets cover more countries and years. Reflecting the influence of the new growth theory, regressors are typically included to represent the economic and insti- tutional environment (sometimes together called the "policy environment"). The potential endogeneity of aid is addressed through instrumenting. And the marginal aid-growth slope is allowed to vary, through the incorporation of such regressors as aid2 and aid x policy. The data sets are almost always panels. Burnside and Dollar (2000) test whether an interaction term of aid and an index of recipient country economic policies are significantly associated with growth. Their panel is drawn from developing countries outside the former Eastern bloc, covering the six four-year periods in 1970-93. They incorporate some controls found to be significant in the general growth literature: initial income (log real GDP per capita) to capture convergence; ethno-linguistic frac- tionalization (Easterly and Levine 1997), assassinations per capita (Banks 2002), and the product of the two; the Knack-Keefer (1995) institutional quality variable, the International Country Risk Guide Economic rating (ICRGE); the ratio of M2 to GDP, to indicate financial depth, lagged one period to avoid endogeneity (King and Levine 1993);and dummy variables for Sub-Saharan Africa and fast-growing East Asia. Burnside and Dollar use a measure of aid called effective development assist- ance (EDA; Chang, Fernandez-Arias, and Serven 1998). EDA differs in two major respects from the usual net official development assistance (ODA) measure tabulated by the Organisation for Economic Co-operation and Development-Development Assistance Committee (OECD-DAC 2002). First, EDA excludes technical assistance, on the grounds that it funds not so much recipient governments as consultants. Second, it differs in its treatment of loans. Net ODA counts disbursements of concessional (low-interest) loans only, but at full face value.' As a capital flow concept, it nets out principal but 1. The DAC considers a loan concessional if it has a grant element of at least 25 percent of the loan value, using a 10 percent discount rate. not interest payments on old loans. In contrast, EDA includes development loans, regardless of how concessional (for example, it includes loans on near- commercial terms by the World Bank to middle-income countries such as Brazil), but counts only their grant element-that is, their net present value. Concerned about limited statistical power, Burnside and Dollar combine their economic policy indicators into a single variable. They first run a growth regression without aid terms, but with all controls and three indicators of economic policy-log(1 +inflation), budget balance as a percentage of GDP, and the Sachs-Warner (1995) openness variable. All three policy variables are significantly different from zero at the 0.05 level, so Burnside and Dollar form a linear combination of the three using their coefficients as weights.2 When Burnside and Dollar run their base specification including aid and aid x policy, the term of central interest, aid x policy does not enter signifi- cantly. However, it becomes significant after either of two possible changes. Five outlier observations can be excluded (giving Burnside and Dollar's pre- ferred specification). Or a quadratic interaction term can be added-aid2 x policy, in which case both aid x policy and aid2 x policy appear significantly different from zero, the first with a positive sign and the second with a nega- tive. Burnside and Dollar famously conclude that aid raises growth in a good policy environment, but with diminishing returns. Burnside and Dollar's work has triggered responses, some critical, some sup- portive. Hansen and Tarp (2001)make one prominent attack. They modify the Burnside and Dollar two-stage least-squares (2SLS)regressions in several ways, most importantly by adding aid2. Aid x policy is not significant in their results, but aid and aid2 are, the first with a positive sign and the second with a nega- tive. The implication is that aid is effective on average, but with diminishing returns-regardless of recipients' policies as far as the evidence goes. Hansen and Tarp then criticize both the Burnside and Dollar regressions and their own for failing to handle several standard concerns. There may be country-level fixed effects that correlate with both policies and growth. Failing to purge or control for all such effects could give spurious explanatory power to policy and aid x policy. Also, variables other than aid and its interaction terms, such as fiscal balance, could be endogenous and need instrumenting too. They deploy the Arellano-Bond (1991) generalized method of moments (GMM) estimator, which is designed to handle these ~roblemsin short panels. Hansen and Tarp also add the change in aid (Aaid) and b aid^ as regressors.3 Their 2SLS results on aid and aid2 hold. And Aaid and aid^ are significant too, again, the first with a positive sign and the second with a negative. 2. They also add a constant term to the index, but this has no effect on the regression results of interest here. 3. This is equivalent to adding lagged ard and lagged aid2 since the regressions also control for 'rid and aid2. Roodman 259 Guillaumont and Chauvet (2001)tell a third story. They hypothesize that the economic vulnerability of a country influences aid effectiveness. They call econ- omic vulnerability the "environment," not to be confused with Burnside and Dollar's "policy environment." In this story, aid flows stabilize countries that are particularly buffeted by terms of trade difficulties, other sorts of external shocks, or natural disasters. Guillaumont and Chauvet build an environment index out of four variables: volatility of agricultural value added (to proxy for natural dis- asters), volatility of export earnings, long-term terms of trade trend, and log of population (small countries being more vulnerable to external forces). Their specification is distinctive in using 12-year periods, and in its controls, which include population growth, mean years of secondary school education among adults (Barro and Lee 2000);the Barro-Lee (1994)measure of political instability, based on assassinations and revolutions; ethno-linguistic fractionalization; and lagged M2lGDP. In their OLS and 2SLS regressions, aid x environment appears with the predicted negative sign, indicating that aid works better in countries with worse environments. The term also drives out aid x policy. Collier and Dollar (2002)corroborate Burnside and Dollar with a quite differ- ent data set and specification. Unlike Burnside and Dollar, they perform OLS estimations only. They include former Eastern bloc countries, the Bahamas, and Singapore. They use net ODA rather than EDA. They study 1974-97 instead of 1970-93. They drop all Burnside and Dollar controls except log initial GDP per capita, ICRGE, and period dummy variables. But they add region dummy variable^.^ And they define policy as the overall score from the World Bank's Country Policy and Institutional Assessment (CPIA), a composite rating of countries on some 20 aspects of policies and institution^.^ They add aid2 but then drop the linear aid term from their preferred specification as insignificant. After all the changes aid x policy is again significant, as is aid2, with a negative sign. Starting from the Collier and Dollar core regression, Collier and Hoeffler (2004) analyze how recent emergence from civil war influences aid effective- ness. Sticking to the four-year panel, they create three dummy variables to indi- cate how recently civil war ended. Peace onset is one in the period when a country goes from civil war to peace. Post-conflict 1 is one in the following period and post-conflict 2 is one in the period after that-assuming that civil war does not recur. Aid x policy x post-conflict 1 is significant in Collier and Hoeffler's preferred (OLS) specification: aid works particularly well in a good policy environment a few years after civil conflict. Also corroborating Burnside and Dollar, Collier and Dehn (2001) hew closely to the Burnside and Dollar specification and data set and tell a story 4. The regions are Europe and Central Asia, Middle East and North Africa, Southern Asia, East Asia and I'acific, Sub-Saharan Africa, and Latin America and the Caribbean, as defined by the World Bank. 5. Collier and Hoeffler (2004) make a small correction to the Collier and Dollar data set, excluding five observations where a missing value had been treated as zero. The Collier and Hoeffler version of the Collier and Dollar regression is tested here. that incorporates elements from Guillaumont and Chauvet (2001). They find that adding variables incorporating information on export shocks renders Burnside and Dollar's preferred specification-the one with aid x policy but not aid2 x policy-more robust to the inclusion of Burnside and Dollar's five outliers. First, they add two variables indicating the magnitude of any positive or negative commodity export price shocks. They report that aid x policy is then significant at 0.01 for a regression on the full data set. The negative-shock variable is significant too, with the expected minus sign.' Then Collier and Dehn add four aid-shock interaction terms: lagged aid x positive shock, lagged aid x negative shock, Aaid x positive shock, and Aaid x negative shock. The first and last prove positive and significant in OLS, and the last, Aaid x negative shock, proves particularly robust in their testing. The study buttresses Burnside and Dollar while suggesting that well-timed aid increases ameliorate negative export shocks. This matches the Guillaumont and Chauvet result in spirit. But where Guillaumont and Chauvet interact the amount of aid with the standard deviation of an index of export voltime and other variables, Collier and Dehn's significant term involves the change in aid and the change in export prices. Dalgaard, Hansen, and Tarp (2004) tell a novel aid-growth story. They focus on the share of a country's area that is in the tropics as a determinant of both growth and the influence of aid on growth. This variable surfaces as a growth determinant in Bloom and Sachs (1998),Gallup and Sachs (1999),and Sachs (2001, 2003). The causal links may include institutions and economic policies (Acemoglu, Johnson, and Robinson 2001; Easterly and Levine 2003). Dalgaard, Hansen, and Tarp thus see tropical area as an exogenous "deep determinant" of growth. In the regressions, aid and aid x tropical area fraction are significant, the first with a positive sign and the second with a negative sign and similar magnitude. For countries situated completely in the tropics, the derivative of growth with respect to aid (the sum of the two coefficients) is indistinguishable from zero. Thus, on average, aid seems to work outside the tropics but not inside them. The authors report that their new interaction term drives out both aid x policy and aid2. There are other third-generation studies (Hadjimichael and others 1995; Durbarry, Gemmell, and Greenaway 1998; Svensson 1999; Lensink and White 2001; Chauvet and Guillaumont 2002; Burnside and Dollar 2004). This article focuses on those already highlighted as being among the most influential and, with one exception, having been published. The exception is Collier and Dehn (2001),which is a pillar of the published Collier and Dollar (2004). The testing here applies to what appear to be the authors' preferred regressions (table I). Country by country the tested regressions generate a 6. However, the reproduction using their data gives a t statistic of only 0.42 to aid x policy despite having the same R~ and sample size, so their result may be an error. But the same negative sign does appear in the reproduction on the negative-shock variable. TABI.E 1. Regressions Tested Definition of Former east Study Years1 Outliers Key significant Regression Estimator bloc? Controls period period Aid Policy out? term(s) Burnside and OLS No LGDP, ETHNF, 1970-93 4 EDNreal GDP BB, INFL, SACW Yes Aid x policy Dollar (2000) ASSAS, 5/0LS ETHNF x ASSAS, ICRGE, M2, SSA, EASIA, period dummy variables Collier and OLS NO LGDP, ETHNF, 1974-93 4 EDNreal GDP BB, INFL, SACW No Aid x policy, Dehn (2001) ASSAS, Aaid x negative 3.4 ETHNF x ASSAS, shock ICRGE, M2, SSA, EASIA, period dummy variables Collier and OLS Yes LGDP, ICRGE, 1974-97 4 ODNreal CDP CPIA No Aid x policy, aid2 Dollar (2002) policy, period and 1.2" region dummy variables Collier and OLS Yes LGDP, ICRGE, 1974-97 4 ODNreal GDP CPIA No Aid x policy x Hoeffler policy, period and post-conflict 1 (2004) 3.4 region dummy variables Hansen and Difference No LGDP, ASSAS, 1978-93 4 ODNexchange INFL, SACW No Aid, aid2, Aaid, ;tr Tarp (2001) GMM ETHNF x ASSAS, rate GDP aaid2 8. 3.2 ICRGE, M2, period $j dummy variables (Continued) & L TABLE Continued 1. Definition of Former east Study Years1 Outliers Key significant Regression Estimator bloc? Controls period period Aid Policy out? term(s) Dalgaard, System Yes LGDP, policy, period 1970-97 4 EDNreal G D P ~BB, INFL, SACW No Aid, aid x tropical Hansen, and GMM dummy variables area fraction Tarp (2004) 3.5 Guillaumont 2SLS No LGDP, ENV, SYR, 1970-93 12 ODNexchange BB, INFL, SACW No Aid, and Chauvet POPG, M2, rate GDP aid x environment (2001)5.2 PINSTAB, ETHNF, period dummy variables LGDP, log initial real GDP per capita; ETHNF, ethnolinguistic fractionalization, 1960; ASSAS, assassinations per capita; ICRGE, composite of International Country Risk Guide Economic governance indicators; M2, MZIGDP, lagged; SSA, Sub-Saharan Africa dummy variable; EASIA, fast-growing East Asia dummy variable; ENV, Guillaumont and Chauvet "environment" variable; SYR, mean years of secondary schooling among adults; PINSTAB, average of ASSAS and revolutions per year; BB, budget balance1GDP; INFL, log(1 +inflation);SACW, Sachs-Warner openness; EDA, effective development assistance; ODA, net official development assistance. "As revised in Collier and Hoeffler (2004). b ~extrapolated to 1970-74 and 1996-97 in Easterly, Levine, and Roodman (2004). s Source: Author's analysis based on sources shown in the table. Roodman 263 TABLE Marginal Impact of Aid According to Preferred Regression, Various 2. Studies, for 20 Largest Aid Recipients of 1998 Burnside and Collier Collier Collier Hansen Dalgaard, Dollar and and and and Hansen, Guillaumont (2000) Dehn Dollar Hoeffler Tarp and Tarp and Chauvet Country 5-OLS (2001) (2002) (2004) (2001) (2004) (2001) Bangladesh Bolivia China CBte d'Ivoire Egypt Ethiopia Haiti India Indonesia Kenya Mozambique Nicaragua Philippines Poland Russia Tanzania Thailand Uganda Vietnam Zambia Note: Numbers in parentheses are robust standard errors. All figures are based on reproduc- tlons of the original regressions. All pertain to 1994-97, except the Guillaumont and Chauvet regression, which pertains to 1982-93. Aid is taken as a share of exchange rate GDP in the Hansen and Tarp and Guillaumont and Chauvet regressions and of purchasing power parity GDP in the rest. Blank cells are caused by missing observations of underlying indicators. Source: Author's analysis based on sources shown in the table. diversity of conclusions about the slope of growth with respect to aid at the margin (table 2 illustrates these conclusions for the 20 largest aid recipients in 1998). As an example of the calculations here, the Burnside and Dollar (2000) structural equation is where Y is GDP per capita, A is aid, P is policy, x is a vector of controls, including initial GDPIcapita, and E is the error term. So the implied slope of growth with respect to aid is d(AY)ldA= a +PP, which depends on the recipi- ent's policy level. Applying such formulas to 1998 data, the Burnside and Dollar regression generally predicts benefits from increasing aid while, at the other extreme, the Dalgaard, Hansen, and Tarp (2004)and Guillaumont and Chauvet (2001)regressions express pessimism. The question is what to make of such conclusions. There is some robustness testing in the recent literature on aid-growth connec- tions, albeit focusing on Burnside and Dollar (2000). Lu and Ram (2001) introduce fixed effects into the Burnside and Dollar regressions. Ram (2004) splits the aid variable into the components coming from bilateral and multi- lateral donors and also tests alternative definitions of policy. Dalgaard and Hansen (2001) modify the choice of excluded outliers. Easterly, Levine, and Roodman (2004) extend the Burnside and Dollar data set to additional countries and an additional period, 1994-97. All these tests eliminate the key Burnside and Dollar result. The present study expands Easterly, Levine, and Roodman (2004)along two dimensions. It tests more studies. And it applies more tests. In addition to fragility, the other bugaboo of econometrics is misspecifica- tion. Important questions can be raised about the validity of the regressions tested here. Some exhibit serial correlation in the errors.' The excludability and relevance of instruments are legitimate concerns. Regressors treated as exogen- ous may not be. And term pairs such as aid and aid2 may be multicollinear. But for the sake of concision, this article focuses on the problem of fragility. The Tests The tests applied to these third-generation aid-growth regressions constitute a more systematic sampling of "specification space" than has hitherto been done. 7. An earlier version of this article attempted to address autocorrelation by further modifying the tested specifications, at the expense of complexity in presentation and, arguably, "whimsy." In particular, most of the test regressions included the log of population as a control since Sargan-type tests suggested it was an improperly excluded instrument. This explains the difference in results between this and the earlier version. To limit complexity and minimize arbitrariness, each test involves changing just one aspect of the estimations at a time (the tests are summarized in table 3). The first four groups of tests-relating to the controls, the definition of aid and policy, and period length-transfer one specification's choices to the others.' Last are tests that modify the sample by dropping outliers or expand- ing to new countries and periods. There are six groups of tests: (1)Changing the control set. In Learner's worries about whimsy, the spe- cification choice that concerns him is that of regressors. The studies examined use four different control sets, which give rise to four robustness tests (detailed in table 3). Each substitutes an alternative control set for the original one and examines the effect on the signifi- cance of key terms. Like the authors of the original regressions, and in the spirit of avoiding arbitrariness, the robustness tests here use all the complete observations available for developing countries (including the countries of Eastern Europe). Because different variables are available for differ- ent subsets of countries, changing the regressor set changes the regression sample. One could perform variants of the tests that are restricted to the intersections of the old and new samples in an attempt to distinguish the effects of changing samples and changing variables. This course is not taken here because it would add to the complexity, would still cause sample changes, and would not answer the hypothetical question, "What would the results have been if the original authors had used alternative controls?" The authors almost certainly would have used all available observations. (2) Redefining aid. All the studies take total aid received as a share of recipi- ent GDP. But there are differences in defining both the numerator and denominator of the ratio. Burnside and Dollar (2000),Collier and Dehn (2001),and Dalgaard, Hansen, and Tarp (2004)use EDA in the numer- ator, whereas the rest use net ODA. On the choice of denominator, there is also a split. Hansen and Tarp (2001) and Guillaumont and Chauvet (2001) use GDP converted to dollars at market exchange rates, whereas the others use real GDP from the Penn World Tables (Summers and Heston 1991). A country's relative price level strongly correlates with income per capita, with the poorest countries having price levels 20-25 percent that of the United States. Thus, using purchasing power parities instead of exchange rates will result in relatively larger GDPs and relatively smaller ratios of aid to GDP for the poorest countries. 8. Papers that instrument the variables of interest also differ in their choice of instruments. But since different variables (aid x policy in one regression, say, and aid2 in another) ought to be instrumented differently, the various instrument sets are less interchangeable and of less use for the approach in this article. TABLE3 . Robustness Tests Test Description Changing controls Burnside and Dollar (2000) Control for LGDP, ETHNF, ASSAS, ETHNF x ASSAS, controls ICRGE, M2, SSA, EASIA, period effects, as in Burnside and Dollar (2000),Collier and Dehn (2001), and Hansen and Tarp (2001) Collier and Dollar (2002) Control for LGDP, ICRGE, period and region effects, as in controls Collier and Dollar (2002) and Collier and Hoeffler (2004) Guillaumont and Chauvet Control for LGDP, ENV, SYR, POPG, M2, PINSTAB, (2001) controls ETHNF, period effects, as in Guillaumont and Chauvet (2001) Dalgaard, Hansen, and Tarp Control for LGDP, ICRGE, SSA, EASIA, period effects, as (2004)controls in Dalgaard, Hansen, and Tarp (2004) Changing aid definition EDNreal GDP Effective development assistancelreal GDP, as in Burnside and Dollar (2000), Collier and Dehn (2001), and Hansen and Tarp (2001) ODNreal GDP ODNreal GDP, as in Collier and Dollar (2002) and Collier and Hoeffler (2004) ODNexchange rate GDP ODNexchange rate GDP, as in Hansen and Tarp (2000) and Guillaumont and Chauvet (2001) Changing policy definition INFL, BB, SACW Inflation, budget balance, and Sachs-Warner openness, as in Burnside and Dollar (2000)and Collier and Dehn (2001) INFL, SACW Inflation and Sachs-Warner, as in Dalgaard, Hansen, and Tarp (2004) CPIA Country Policy and Institutional Assessment, as in Collier and Dollar (2002) and Collier and Hoeffler (2004) Changing period length 12-year Aggregate over 12-year periods, as in Guillaumont and Chauvet (2001) Changing sample and data set No outliers Remove Hadi outliers in the partial scatter of the dependent variable and the independent variable of greatest interest Expanded sample New data set. Carried to 2001, except shocks data end in 1997 and Guillaumont and Chauvet (2001)environment variable not updated Expanded sample, no outliers Combine above two changes LGDP, log initial real GDP per capita; ETHNF, ethnolinguistic fractionalization, 1960; ASSAS, assassinations per capita; ICRGE, composite of International Country Risk Guide Economic governance indicators; M2, MZIGDP, lagged; SSA, Sub-Saharan Africa dummy vari- able; EASIA, fast-growing Easr Asia dummy variable; ENV, Guillaumont and Chauvet "environ- ment" variable; SYR, mean years of secondary schooling among adults; PINSTAB, average of ASSAS and revolutions per year; BB, budget balance1GDP; INFL, log(1 +inflation); SACW, Sachs-Warner openness; EDA, effective development assistance; ODA, ner official development assisrance. Sozdrce: Author's analysis based on sources described in the text. Roodman 267 TABLE4. Simple Correlations of Aid and Good Policy Measures, Four-Year Periods. on Available Observations EDA/real GDP ODA/real GDP ODNexchange rate GDP EDNreal GDP 1.00 ODA/real GDP 0.97 ODNexchange rate GDP 0.78 Inflation, budget Inflation, CPIA balance, Sachs-Warner Sachs-Warner Inflation, budget balance, 1.00 Sachs-Warner Inflation, Sachs-Warner 0.98 CPIA 0.53 EDA, effective development assistance; ODA, net official development assistance; CPIA, World Bank's Country Policy and Institutional Assessment. Source: Author's analysis based on sources described in the text. This could have a significant effect on coefficient estimates for aid and its interactions. With two options each for measuring aid and GDP, there are four possible combinations for the aid to GDP ratio. The literature includes all but EDAIexchange rate GDP, and these are the bases for three tests.9 In fact, EDMreal GDP and ODMreal GDP are highly correlated (Dalgaard and Hansen 2001), so switching from one to the other may not stress results much (table 4). (3) Redefining good policy. Three sets of "good policy" variables appear among the tested regressions. One is Burnside and Dollar's combination of budget balance, inflation, and Sachs-Warner openness. A second is inflation and Sachs-Warner only (Hansen and Tarp 2001). And a third is CPIA alone (Collier and Dollar 2002; Collier and Hoeffler 2004). These generate three robustness tests. With Burnside and Dollar's coefficients used to form policy indexes (6.85 for budget balance, -1.40 for inflation, and 2.16 for Sachs-Warner),the first two policy definitions are highly correlated, at 0.98, but the third varies more distinctly (see table 4). But in actual application of the tests, the Burnside-Dollar-style index-forming regression is rerun each time; it includes all regressors except aid and its interaction terms, and the coefficients on the policy 9. The published EDA data (Chang. Fernandez-Arias, and Serven 1998) cover only 1975-95. EDA as used here is extrapolated to the rest of 1970-2001 through a regression of EDA on net ODA. variables are used to make the index, regardless of statistical significance.lo (4) Changing periodization. All but Guillaumont and Chauvet use four-year periods. The lack of higher frequency observations of the Guillaumont and Chauvet environment variable prevents adapting their 12-year regressions to a 4-year-period panel. But the other regressions can be tested on 12-year panels. Notably, key cross-section studies in the growth literature use periods of 10-25 years despite the small samples that result (Barro 1991; Mankiw, Romer, and Weil 1992; Sachs and Warner 1995). (5) Removing outliers. The tested Burnside and Dollar specification excludes five observations that are outliers in aid x policy and highly influential on the coefficient on that term. This raises a general question about the importance of outliers. To investigate, one robustness test reruns the reproductions of the original regressions after excluding outliers. Another does the same for the expanded-sample versions (see sub- sequently). Following Easterly, Levine, and Roodman (2004), outliers are chosen by applying the Hadi (1992) procedure for identifying mul- tiple outliers to the partial scatter of growth and a regressor of interest, using 0.05 as the cut-off significance level." In 2SLS estimations, regres- sors are first projected onto instruments.12 Since the two-dimensional partial scatter plot is not well defined for GMM regressions, in those cases, analogous 2SLS regressions are run to identify outliers. Outliers are not synonymous with influential observations. But even outliers that do not greatly influence coefficients of interest can substan- tially affect reported standard errors. In addition, outliers are the obser- vations most likely to signal measurement problems or structural breaks beyond which the core model does not hold-both of which seem better reasons for exclusion than high influence. That said, outliers do not necessarily signal measurement problems or structural breaks. This is especially possible when the variable of interest is highly non-normal, such as the Collier and Dehn (2001)export price shock variable. In such cases, outliers may contain valuable information about the development process under rare circumstances. 10. The constant term in the policy index is computed in the same manner as in Burnside and Dollar. It is the predicted growth rate in the model when the policy variables and the period dummy variables are zero, and all other variables take their sample-average values. 11. Applying the Hadi procedure directly to a full, many-dimensioned data set typically identified 20 percent or more of observations as outliers. 12. This test is even run on the Burnside and Dollar SIOLS regressions, from which one set of outliers is already excluded. Regardless of the genesis of these regressions' results, it is interesting whether they are driven by a few observations in the remaining sample. Roodman 269 ( 6 ) Expanding the sample. Easterly, Levine, and Roodman (2004) develop a data set that extends that of Burnside and Dollar from 1970-93 to 1970-97 and adds six countries. For the current study that data set has been extended to 2001 and improved in other respects. (See supplemen- tal appendix S.l, available at http://wber.oxfordjournals.org/.) This allows a net expansion in both years and countries for all but the Guillaumont and Chauvet regression, whose 12-year periods and unusual environment variable hinder expansion. Issues in Interpreting Results If Leamer's (1983) extreme-bounds analysis is applied to the results of this testing, then a coefficient will be deemed robustly different from zero only if it is significantly different from zero in every test. However, as Sala-i-Martin (1997) argues, this definition of robustness indeed seems extreme. For example, one could test robustness by averaging together all observations for each global region, generating samples of some six observations. Almost no regression would pass this test. One could argue that this test would be "unfair," in that it would be too weak to generate meaningful results. But there is no sharp div- ision between fair and unfair tests. Indeed, in this test suite the 12-year-period test destroys every regression it can be applied to. It is not obvious whether the test is too strong or the regressions too weak. Thus, robustness should be a con- tinuous rather than a dichotomous concept. Sala-i-Martin offers his own procedure for assessing robustness. In essence, he estimates the cumulative distribution function for a coefficient of interest by running a large number of variants of the regression it comes from. The robust- ness of a coefficient is then the fraction of the density that is on one or the other side of zero. The validity of this concept is based on the assumption, however informal, that the set of regressions actually run is a representative of all possible variants of the original regression. For the collection of tests assembled here, however, that assumption is not valid. For example, one important subset of tests, those expanding the sample, cannot be applied to the Guillaumont and Chauvet (2001) regression. It does not seem plausible that the test results are representative both with and without this important subset of tests. The sampling of specification space that is made here is minimally arbitrary, but cannot be assumed to be representative of all possible tests. Thus, while Leamer's definition of robustness may be too harsh for this context, Sala-i-Martin's has its own limitations. This will be true even if one performs every possible combination of tests in the suite rather than just one at a time. In the end it seems that human judgment applied to the full set of results must substitute for mechanical definitions of robustness. This in turn means there is some value in keeping the tests few enough for the human mind to embrace. TABLE. Coefficients on Key Terms under Specification-Modifying Tests (OriginalData Sets) 5 Dalgaard, Hurnside Hansen. O D N INFL, and Cnllier and Gu~lldurnonr and EDN redl O D N real exchange BB, INFL, Specificat~on Key term O r ~ g ~ n a l Dollar Dollar and Chauvet Chauver CIIP GIIP rate GDP SACW SACW CPlA 12-year Burnside Aid x pol~cy and Dollar Observations (2000) Collier and Ard x polrcy Dehn (2001) Aaid x negatrve shock Observat~ons Coll~erand Aid x polrcy Dollar (2002) Observat~ons Collier and Post-conflrct 1 Hoeffler x Ald x polrcy (2004) Observat~ons Hansen and Ald Tarp (2001) A I ~ ~ Observations TABLE Continued 5. Changing Changing controls Changing aid Changing policy periods -- Dalgaard, Burnside Hansen, O D N INFL, and Coll~erand Guillaumont and EDAI real O D N teal exchange BB. INFL, Specification Key term Original Dollar Dollar and Chauvet Chauvet GDP GDP rate GDP SACW SACW CPlA 12-year Dalgaard, Hansen, and Tarp (2004) Ard x troptcal -0.98"' -1.49"' -1.79"' - 1.66""" Area % (-3.16) (-8.58) (-8.26) (-6.72) Observations 371 354 371 315 Guillaumont Aid x Environment -0.15' -0.12' -0.07 and (-1.79) (-1.68) (-1.35) Chauvet Observat~ons 68 71 73 (2001) *Significant at the 10 percent level; "Significant at the 5 percent level; ')'"Significant at the 1percent level. EDA, effective development assistance; ODA, net official development assistance; INFL, log(1 + inflation); BB, budget balance1GDP; SACW, Sachs-Warner openness; CPIA, World Bank's Country Policy and Institutional Assessment. Note: Numbers in parentheses are t statistics, and all are heteroskedasticity-robust; those for general method of moments (GMM)regressions are also autocorrelation-robust. Except for the original Hansen and Tarp regression, all GMM standard errors incorporate the Windmeijer (2005) finite-sample z? correction. Source: Author's analysis based on sources described in the text. i a TABLE Coefficients on Key Terms under Data Set-Modifying Tests 6. Original data set New data set Specification Key term Full sample No outliers Full sample No outliers Burnside and Dollar Aid x policy 0.19*"* (2.61) -0.05 (-0.45) -0.04 ( -0.23) -0.27 ( - 1.33) (2000) Observations 270 263 436 426 Collier and Dehn (2001) Aid x policy 0.10~(1.70) 0.11 (1.11) 0.06 (1.05) 0.10 (0.75) AAid x negative shock 0.04""" (3.17) - 0.06 ( -1.33) 0.03"*+(3.01) -0.10 (-1.15) Observations 234 224 391 372 Collier and Dollar (2002) Aid x policy 0.14" (2.15) 0.07 (1.06) -0.01 (-0.19) -0.04 (-0.81) Observations 344 341 520 508 Collier and Hoeffler Post-coflict x aid x policy 0.18"*" (3.92) 1.18""(2.12) 0.08' (1.91) -0.06 (-0.19) (2004) Observations 344 333 520 494 Hansen and Tarp (2001) Aid 0.90"' (4.22) 0.96""(2.19) 0.08 (0.41) 0.03 (0.16) ~ i d l -0.02'"" (-3.83) -0.02" ( - 1.95) -0.001 (-0.57) - 0.001 ( -0.42) AAid -0.70""" (-4.91) 0.70"' (-2.73) -0.13 (-0.92) -0.07 ( -0.53) AAid2 0.01"':' (3.64) O.Ol* (1.86) 0.002 (1.03) 0.001 (0.71) Observations 213 212 517 514 Dalgaard, Hansen, and Aid 0.69"' (5.09) 0.34 (0.20) 0.94" "" (4.60) -0.11 (-0.09) Tarp (2004) Aid x tropical area % -0.98""" (-3.16) -0.41 (-0.25) 0.95"- (-3.01) -0.02 (-0.02) Observations 371 362 463 451 Guillaumont and Chauvet Aid x environment -O.l5* (-1.79) -0.11" (-1.96) (2001) Observations 68 67 'Significant at the 10 percent level; ""Significant at the 5 percent level; *""Significant at the 1 percent level. Note: Numbers in parentheses are t statistics; all are heteroskedasticity-robust; those for general method of moments (GMM) regressions also autocorrelation-robust. Except for original Hansen and Tarp regression, all GMM standard errors incorporate the Windmeijer (2005) finite-sample correction. Source: Author's analysis based on sources described in the text. Roodman 273 The first step in the testing is to use the authors' data sets to reproduce their original results (see column 1 of tables 5 and 6). All the reproductions exhibit the same pattern of results as the originals and all but one have the same sample size.'"he Burnside and Dollar (2000), Collier and Dehn (2001), and Hansen and Tarp (2001)reproductions are perfect, and the rest are close. Since the purpose is to test robustness, the inexact matches are not a concern. If the results from the tested regressions are robust, they should withstand what- ever minor changes in data or specification cause the discrepancies in the reproductions. Tables 5 and 6 report results on key terms in all tests.14 Blank cells indicate inapplicable tests. The test involving the definition of aid as the ratio of EDA to real GDP, for example, is not applicable to regressions that originally use it. Using 12-year ~eriodsdoes not work for the Collier and Hoeffler (2004) regression, because the definition of their post-conflict 2 variable assumes four- year periods. Lack of higher frequency data for Guillaumont and Chauvet's (2001) environment variable prevents short-period tests. A total of 71 robust- ness checks are run.'" Results for tests inspired by differences among the original regressions are given in table 5. The Collier and Hoeffler result on post-conflict 1 x aid x policy (or the collinear post-conflict 1 x aid) and the Dalgaard, Hansen, and Tarp results for aid and aid x tropical area fraction do best. All of these center on sharply bimodal variables: the Collier and Hoeffler post-conflict 1 dummy variable is one for only 13 of the 344 observations in their original sample, and there are negative shocks in 38 of the 234 Collier and Dehn observations. In the Dalgaard, Hansen, and Tarp (2004)sample, 233 of the 371 observations are 100 percent tropical and 68 are zero percent, leaving 70 in between. Evidently, regularities involving such variables are more resilient to specifica- tion changes. Results from sample-modifying tests are shown in table 6. The first two result columns are based on regressions on the original authors' data sets-first 13. The Dalgaard, Hansen, and Tarp (2004)regression was executed with the DPD tor Ox ~ackage (Doornik, Arellano, and Bond 2002). [t turns out that a quirk in this software-incomplete observations that create gaps in the time series must always be included in the data file rather than deleted-led to a slight mishandling of the data. The xtabond2 module for Stata (Roodman 20061, used here, handles gaps as users intend. This explains the difference in samples. 14. Full results are available on request. 15. Initial testing revealed multicollinearity in the Collier and Hoeffler regression. In their prcliminary regression 3.1 (not regression 3.4, which is tested here), they include the variables post-conflict I , post-conflict 1 x policy, and post-conflict I x aid', along with the favored post-conflict 1 x aid x policy. In the reproduction of 3.1, post-conflict 1 x aid x policy and post-conflict 1 x aid have a partial correlation of 0.985, making the two statistically indistinguishable. Thus the Collier and Hocffler results ought to be interpreted as pertaining to either post-conflict 1 x aid x policy or post-conflict 1 x aid. Occam's razor argues for the second. for their full sample and second for the sample excluding outliers. The next pair of columns is analogous, for the expanded data set. The corresponding partial scatter plots in supplemental appendix S.2 (available at http://wber. oxfordjournals.org/) illustrate the sample-modifying results and are reminders of the importance of checking for outliers. Except for Guillaumont and Chauvet, all the original OLS and 2SLS results depend on outliers for some or all of their significance. The dependence is particularly heavy for the regressions involving aid x policy. On the other hand, the lack of significance of most of the coefficients under the sample-expansion test is not driven by out- liers. It is worth noting that the Collier and Dehn result on Aaid x negative shock, another interaction term involving a variable with a highly non-normal distribution, is arguably stronger than it looks. The coefficient is reversed by the exclusion of outliers from the original sample. But it is arguably fallacious to draw conclusions about the role of shocks having excluded many of the most dramatic examples. The overall pattern is clear. The 12-year test is the toughest-probably because of the small samples-failing all regressions. The new-data test is not far behind, an important point given that the modification it involves-a mode- rate sample expansion-is much less radical. When the tables are read by rows (test subjects) instead of columns (tests), the Dalgaard, Hansen, and Tarp result on the aid-tropics link is the only one to come through the specification- modifying tests strongly. But it too falls down on the sample-modifying tests after outliers are removed. Four of the nine original-sample outliers are for Jordan, covering 1974-89, a period in which that non-tropical country experi- enced high growth and received considerable aid from its neighbors. This con- firms the conclusion of Rajan and Subramanian (2005) that the aid x tropics result is fragile too. IV. C O N C L U S I O N The results reported here suggest that the fragility found in Easterly, Levine, and Roodman (2004) for Burnside and Dollar (2000) is the norm in the cross- country aid-growth literature. Indeed, in a counterpoint to the focus of Leamer (1983), Levine and Renelt (1992), and Sala-i-Martin (1997) on the choice of controls as a source of fragility, it turns out that modifying the sample gene- rally affects results the most. For example, in the Collier and Dollar (2002) regression, half of the specification-modifying tests leave the t statistic at 1.49 or higher and two more lower it to near 1.00 (see table 5). But adding more years sends it to -0.19-and, after dropping outliers, to -0.81 (see table 6). Does this mean that the various stories of aid effectiveness should be sum- marily dismissed? Are recipient policies, exogenous economic factors, and post- conflict status irrelevant to aid effectiveness? Are there no diminishing returns to aid? Is helping the neediest countries a hopeless task? No. There can be no doubt that some aid finances investment and that domestic policies, governance, external conditions, and other factors these authors study influence the productivity of investment. Why then do such stories of aid effectiveness not shine through more clearly? Aid is probably not a fundamentally decisive factor for development, not as important, say, as domestic savings, inequality, or governance. Moreover, foreign assistance is not homogeneous. It consists of everything from food aid for famine-struck countries to technical advice on building judi- ciaries to loans for paving roads. And much aid is poorly used-or, like venture capital, is like good bets gone bad. Thus the statistical noise tends to drown out the signal. Perhaps researchers will yet unearth more robust answers to the fundamental questions of aid policy. Or perhaps they have hit the limits of cross-country empirics. Either way, robust, valid generalizations have not and will not come easily. Despite decades of trying, cross-country growth empirics have yet to teach us much about whether and when aid works. Acemoglu, D., S. Johnson, and J. A. Robinson. 2001. "The Colonial Origins of Comparative Development: An Empirical Investigation." American Economic Review 91(5):1369-1401. Arellano, M., and S. Bond. 1991. "Some Tests of Specification for Panel Data: Monte Carlo Evidence and an Application to Employment Equations." The Review of Economic Studies 58(2):277-97. Banks, A. 2002. "Cross-National Time-Series Data Archive." Databanks International, New York, N.Y. Barro, R.J. 1991. "Economic Growth in a Cross-Section of Countries." American Economic Review 106(2):407-43. (http://www.nber.org/pub/barro.lee). Barro, R.J., and J. Lee. 1994. Data Set for a Panel of 138 Countries. Cambridge, Mass.: National Bureau of Economic Research. . 2000. International Data on Educational Attainment: Updates and Implications. Working Paper 7911. Cambridge, Mass.: National Bureau of Economic Research. Bloom, D., and J.D. Sachs. 1998. "Geography, Demography, and Economic Growth in Africa." Brookings Papers on Economic Activity 28(2):207-73. Boone, P. 1994. "Aid and Growth." London: London School of Economics. Burnside, C., and D. 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 2834. World Bank, Washington, D.C. Chang. C.C., E. Fernandez-Ar~as,and L. Serven. 1998. "Measuring Aid Flows: A New Approach." Working Paper 387. Washington, D.C.: Inter-American Development Bank. Chauvet, L., and P. Guillaumont. 2002. "Aid and Growth Revisited: Policy, Economic Vulnerability and Political Instability." Paper presented at the Annual Bank Conference on Development Economics: Towards Pro-poor Policies, June 24-26, Oslo. Collier, P., and J. Dehn. 2001. "Aid, Shocks, and Growth." Policy Research Work~ngPaper 2688. World Bank, Washington, D.C. Coll~er,P., and D. Dollar. 2002. "Aid Allocation and Poverty Reduction." European Economic Review 45(1):1-26. . 2004. "Development Effectiveness: What Have We Learnt?" Econonzic lournal 114(496):F244-71. Collier, P., and A. Hoeffler. 2004. "Aid, Policy and Growth in Post-Conflict Societies." European Economic Review 48(5):1125-45. Dalgaard, C., and H. Hansen. 2001. "On Aid, Growth and Good Policies." lournal of Development Studies 37(6):17-41. Dalgaard, C., H. Hansen, and F. Tarp. 2004. "On the Empirics of Foreign Aid and Growth." Economic lournal 114(496):FI91-216. Doornik, J.A., M. Arellano, and S. Bond. 2002. "Panel Data Estimation Using DPD for Ox." (www. doornik.com/download/dpd.pdf). Durbarry, R., N. Gemmell, and D. Greenaway. 1998. "New Evidence on the Impact of Foreign Aid on Economic Growth." CREDIT Research Paper 98r8. Nottingham, United Kingdom: University of Nottingham, School of Economics, Centre for Research in Economic Development and International Trade, University Park. Easterly, W., and R. Levine. 1997. "Africa's Growth Tragedy: Policies and Ethnic Divisions." Quarterly journal of Economics 112(4):1203-50. -. 2003. "Tropics, Germs, and Crops: How Endowments Influence Economic Development." lournal of Monetary Economics 50(1):3-39. Easterly, W., R. Levine, and D. Roodman. 2004. "New Data, New Doubts: A Commenr on Burnside and Dollar's 'Aid, Policies, and Growth (2000)'." American Economic Review 94(3):781-84. Gallup, J.L., and J.D. Sachs. 1999. "Geography and Economic Development." In B. Pleskovic, and J.E. Stiglitz, eds., Annual World Bank Conference on Development Economics, 1998 Proceedings. Washington, D.C.: World Bank. Griffin, K.B., and J.L. Enos. 1970. "Foreign Assistance: Objectives and Consequences." Econowzic Development and Cultural Change 18(3):313-27. Guillaumont, P., and L. Chauvet. 2001. "Aid and Performance: A Reassessment." lournal of Development Studies 37(6):66-92. Hadi, A.S. 1992. ''Identifying Multiple Outliers in Multivariate Data." lournu1 of the Royal Statistical Society Series B(54):761-77. I-Iadjimichael, M.T., D. Ghura, M. Muhleisen, R. Nord, and E.M. Ucer. 1995. "Sub-Saharan Africa: Growth, Savings, and Investment, 1986-93." Occasional Paper 118. Washington, D.C.: International Monetary Fund. Hansen, H., and F. Tarp. 2000. "Aid Effectiveness Disputed." lournal of International Development 12(3):375-98. . 2001. "Aid and Growth Regressions." journal of Development Economics 64(2):547-70. King, R.G., and R. Levine. 1993. "Finance and Growth: Schumpeter Might be Right." American Economic Review 108(3):717-37. Knack, S., and P. Keefer. 1995. "lnstitutions and Economic Performance: Cross-Country Tests Using Alternative Institutional Measures." Economics and Politics 7(3):207-27. Learner, E.E. 1983. "Let's Take the Con out of Econometrics." American Economic Review 73(1):31-43. Lensink, R., and H. White. 2001. "Are There Negative Returns to Aid?" ]ournu1 of Development Studies 37(6):42-65. Levine, R., and D. Renelt. 1992. "A Sensitivity Analysis of Cross-Country Growth Regressions." American Economic Review 82(4):942-63. Lu, S., and R. Ram. 2001. "Foreign Aid, Government Policies, and Economic Growth: Further Evidence frorn Cross-Country Panel Data for 1970-93." Economia Internazlonale 54(1):15-29. ~Mankiw,N.G., D. Romer, and D.N. Weil. 1992. "A Contribution to the Empirics of Economic Growth." American Economic Review 107(2):407-37. Mosley, P., J. Hudson, and S. Horrell. 1987. "Aid, the Public Sector and the Market in Less Developed Countries." Economic]ournal 97(387):616-41. OECD-DAC (Organisation for Economic Co-operation and Development-Development Assistance Committee). 2002. Development Assistance Committee Online, Paris. (www.oecd.org/dac). Rajan, R., and A. Subramanian. 2005. "Aid and Growth: What Does the Cross-Country Evidence Really Show?" Working Paper 051127. Washington, D.C.: International Monetary Fund. Ram, R. 2004. "Recipient Country's 'Policies' and the Effect of Foreign Aid on Economic Growth in Developing Countries: Additional Evidence." Journal of International Developnient 16(2):201-11. Roodman, D. 2006. "How to Do xtabond2: An Introduction to 'Difference' and 'System' GMM in Stata." Working Paper 103. Washington, D.C.: Center for Global Development. Sachs, J.D. 2001. Tropical Underdevelopmerit. Working Paper 8119. Cambridge, Mass.: National Bureau of Economic Research. . 2003. Institutions Don't Rule: Direct Effects of Geography on per Capita Income. Working Paper 9490. Cambridge, Mass.: National Bureau of Economic Research. Sachs, J.D., and A. Warner. 1995. "Economic Reform and the Process of Global Integration." Brookings Papers on Econoniic Activity 25(1):1-118. Sala-i-Martin, X.X. 1997. "I Just Ran Two Million Regressions." American Ecotiomic Review 87(2):178-83. Summers, R., and A. Heston. 1991. "The Penn World Table (Mark 5): An Expanded Set of International Comparisons, 1950-88." QuarterlyJournal of Economics 106(2):327-68. Svensson,J. 1999. "Aid, Growth, and Democracy." Economics and Politics 11(3):275-97. Windmeijer, F. 2005. "A Finite Sample Correction for the Variance of Linear Two-Step GMM Estimators." Journal of Econometrics 126(11):25-51. Incremental Reform and Distortions in China's Product and Factor Markets Xiaobo Zhang and Kong-Yam Tan The purpose of economic reform is to reduce distortions and enhance efficiency. However, when reforms are partial and incremental, individuals and local govern- ments are often able to capture the rent inherent in the gradual transition process. Young (2000)warned that such rent-seeking behavior might lead to increasing market fragmentation. Empirical studies have shown the opposite in the product market. This article argues that as the rent from China's product market has been squeezed out due to deepening reforms, rent-seeking behavior may have shifted to the capital market. Further reforms are needed in the capital market to squeeze out these rent-seeking opportunities, just as those from the product and labor markets were squeezed out earlier. JEL Code: D33, D61, D63, 011, 053, P23. Over the past 25 years, China's transformation from a centrally planned to an increasingly market-driven economy has led to substantial efficiency gains and rapid economic growth (Maddison 1998; Fan, Zhang and Robinson 2003). However, as Young (2000) argues, the reforms may not have been sufficiently complete to improve domestic market integration. This could happen, for example, if increased interregional competition as a result of fiscal decentralization led local governments to impose trade protection measures against each other. Young's work has stimulated a series of studies to investigate trends in market integration. A recent survey by the China State Council Development Research Center (2003) indicates that China's domestic product markets have become more rather than less integrated. Measures of regional protection have also declined significantly over the past decade. Wei and Fan (2004) show that output prices have become more integrated. Huang, Rozelle, and Chang (2004) use evidence from the rice market to argue that China's commodity markets are becoming increasingly integrated as a result of the reforms. Based on a Kong-Yam Tan (corresponding author), formerly a senior economist at the World Bank Beijing Office, is a professor at Nanyang Technological University, Singapore; his email address is kytan@ntu.edu.sg. Xiaobo Zhang is a senior research fellow at the International Food Policy Research Institute and a visiting professor at Zhejiang University, China; his email address is x.zhang@cgiar.org. A supplemental appendix to this article is available at http://wber.oxfordjournals.org/. THF WORLD BANK ECONOMIC REVIEW,VOI..21, NO. 2, pp. 279-299 doi:lo. 1093/wber/lhm002 Advance Access Publication 23 March 2007 ';3 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 panel data set of 32 industries at the two-digit level of aggregation in 29 pro- vinces, Bai and others (2004) find, after an initial decline, an increase in regional specialization of industrial production, suggesting diminishing impedi- ments to regional trade flows. These findings appear to contradict Young's pre- dictions about worsening market fragmentation. Besides the final goods market, it is also possible that distortions occur in factor markets. De Brauw and others (2002) show that there has been a huge transfer of rural labor from the low-productivity farming sector to high- productivity nonfarm sectors over the past two decades, suggesting a shift toward a more integrated rural labor market. Using the population census data to examine labor flows across provinces, Poncet (2003)concludes that the inter- provincial border barriers to labor migration have declined from the 1980s to the 1990s. Zhang and others (2005)find that returns to education in nonpublic enterprises caught up with those in state-owned enterprises, indicating increas- ing labor mobility across sectors. Yet numerous studies suggest that there is still significant segmentation in the labor market (Meng2000; Knight and Li 2005). China has instituted several financial market reforms, such as the establish- ment of a stock market and regionalization of major banks. Yi (2003) argues that these reforms have made China's financial market more efficient. However, several empirical studies reach the opposite conclusion. Fan, Zhang, and Robinson (2003)find that the provincial marginal rates of return to capital in agriculture, urban industry, urban services, and rural enterprises have diverged since 1985. Boyreau-Debray and Wei (2003) use two methods to test the degree of capital market fragmentation based on provincial data for 1978- 2000. The first approach is to examine the correlation of local savings and investment. In an integrated capital market the correlation should be low. The second approach, drawing from the risk-sharing literature, is to check the degree of consumption smoothing across time and space, which is an important indicator of capital mobility and asset market completeness. Both approaches show that the capital market has become more fragmented.' 1. Recent rent-seeking activities in the banking and real estate sectors include those of Yang Xiuzhu, vice chief of the construction department of Zhejiang Province, who extracted bribes from property developers and disappeared (CaijingJuly 23, 2003); Shanghai real estate tycoon, Zhou Zhengyi, who was implicated in an array of illegal loans coupled with default on statutory compensations for relocatees whose homes were improperly demolished for redevelopment projects (Shartghai Daily, September 6, 2003); Chen Kai, a local government official of Fuzhou, Fujian Province, who borrowed an estimated $50 million from six state hanks and provided kickbacks of around 5 percent of the loans to the lending officers (Washirzgton Post, December 17,2003); former chairman of China Everbright Group, Zhu Xiaohua, who was sentenced to 15 years in jail in November 2002 for taking bribes worth 4 n~illionyuan (Caijing, December 25,2002); and Zhu Yaoming, a stock speculator who was arrested in July 2003 for loan fraud involving 2 billion yuan, which he borrowed from securities firms and banks to speculate on stocks in the Shanghai and Shenzhen stock exchanges (Caiiirzg, December 25, 2003). Numerous Communist Party officials have also been ousted for accepting bribes involving property and real estate projects. They include the former mayor of Shanghai, Chen Liangyu; the former general secretary of Guizhou Province, Liu Fangren; former general secretary of Hebei Province, Cheng Weigao; former Minister of Land and Resources, Tian Fengshan; a former vice mayor of Shenzhen City; and a former mayor and a vice mayor of Shenyang City. Zhang and Tan 281 In summary, the empirical literature on trends in market fragmentation and its extent is inconclusive. Most studies focus on either product or factor markets and over a short period only. The objective of this article is to docu- ment the evolution of both ~roductand factor markets using a more integrated framework over a longer period covering the entire course of economic tran- sition and reforms. To assess the degree of factor market fragmentation, the economy is divided into four sectors: urban industry, urban services, agricul- ture, and rural enterprises (all nonfarm activities such as rural industry, con- struction, transportation, and commerce). The analysis is based on estimating production functions for each sector, using provincial time series data for 1978-2001. One side contribution of the analysis is the computation of a capital stock series by sector, using fixed investment data from the National Bureau of Statistics that are not yet fully available publicly. The estimated par- ameters from the regression equations are used to quantify the regional vari- ation in the marginal products of capital and labor by sector. The results confirm that labor markets are becoming more integrated, but also show that capital markets have become more fragmented. As the reforms in the product markets have deepened, distortions seem to have shifted to the capital market. In this sense, Young's (2000) argument is still valid: in a partially reformed economy distortions may beget more distortions. However, the distortions may not necessarily stay in the same sector. The article first reviews the history of market development in China in the second half of the twentieth century. It then presents data on changes in labor and capital productivity across sectors and regions in the Chinese economy over recent decades and explores trends in product market integration. Regional variations in the marginal products of capital and labor are quantified and serve as good indicators of factor market integration. The efficiency gains for economic growth are simulated with the current barriers to factor flows across regions and sectors removed. A supplemental appendix, available at http://wber.oxfordjournals.org provides additional details about the data. This section briefly summarizes market development in China in the twentieth century. Product Market Market fragmentation has a long history in China. In the early 1950s, China adopted a "self-sufficient" agricultural and industrial policy at both national and provincial levels (Lin, Cai, and Li 1996). Provinces were encouraged to develop their own industries and ensure enough grain pro- duction. However, the underlying economic structure was often inconsistent with a region's comparative advantage. Therefore, local governments had to impose various protections on local products. The planning system led to serious shortages in final goods, forcing the government to impose rationing on consumers as well. Since the economic reforms of the late 1970s, China has decentralized its fiscal system to provide more incentives for local governments to develop their economies (Zhang 2006). Under the fierce competition that resulted from fiscal decentralization, interest groups in provinces and cities were eager to protect their local interests. Regional trade wars broke out in the 1980s and early 1990s (Young 2000). In responding to the crises of regional trade blockades, the National People's Congress passed the "Law on Unjust Competition" in 1993, and in 2001 the State Council issued order 303 "Stipulation of the State Council to Forbid Regional Blockade in Market Activities." Labor Market In the 1950s, the government established the hukou system of household regis- tration, confining people to the village or city of their birth, to ensure enough agricultural labor to produce sufficient grain to support the industrial and urban sector. Rural and urban labor markets became totally segmented (Yang and Zhou 1999). Since the 1980s, China has gradually reduced institutional barriers to migration (for more detail on China's labor market development, see Fleshier and Yang 2003 and World Bank 2005). In 1983, farmers were permitted to engage in transport and marketing of their products beyond local markets. In 1988, the central government permitted farmers to work in cities under the condition that they had to provide their own staples. Since the early 1990s, various measures have been introduced to further relax the hukou system and encourage greater rural to urban labor mobility. Some cities have adopted a selective migration policy, issuing permanent residency to migrants who paid a fee, invested in local business, or bought expensive houses in the city. In addition, urban reforms of housing, employment policies, and the social secur- ity system; the lifting of rationing; and expansion of urban nonstate sectors have made it easier for migrant workers to live in cities. Despite progress in reducing institutional barriers to labor mobility, some obstacles still impede population movement across regions (Fleisher and Yang 2003). For instance, most rural migrants in cities are unable to obtain legal residence permits and are treated as second-class citizens. They have to pay much higher fees for healthcare and schools than legal residents. Discriminatory treatment of rural migrant workers in employment and social services is commonplace, particularly in the formal sector. Capital Market In the central planning era, banks were the dominant source of business financing (World Bank 2005). They provided loans primarily to formal state Zhang and Tan 283 enterprises within their locality. The central government exerted direct control over banks. Administrative rather than market forces determined capital movements. The major role of banks was to provide equity financing and to support national development strategies. Since the late 1970s, China has conducted a series of banking sector reforms. In 1983, the four state-owned commercial banks (Bank of China, Agricultural Bank of China, Industrial and Commercial Bank of China, and Construction Bank of China) were reorganized to become more market oriented. In addition to direct vertical control within the bank, local govern- ments were ganted more horizontal controls over bank branches. As the economy developed rapidly, so did demand for credit. Local governments tightened their control over local bank branches by blocking saving deposits from moving elsewhere. Many local governments forced banks in their jurisdiction to extend credit to them, creating serious inflation in the early 1990s. Since 1994 the central government has reasserted its control over the banks, ended local government control of bank branches, and set up regional banks to encourage capital mobility across provinces. However, loopholes remain in the system. In particular, local governments can use land to acquire loans to finance infrastructure (World Bank 2005). Once land is acquired from farmers for public purposes, local governments and developers can use this "state-owned" land as collateral for credit from the local branches of state banks. Land banking is a major driver of the rapid growth in infrastructure investment in China (Zhang2006). Even after the establishment of the Shanghai and Shenzhen Stock Exchanges in December 1990, banks have retained a dominate role in financial markets. In 2000, the banking system accounted for about two-thirds of financial trans- actions, while the bond and stock market accounted for only 5 percent of financial flows (World Bank 2005). There have been many abnormal phenom- ena in the development of the stock market (Lin 2004). Most listed companies are state-owned enterprises and in general perform worse than nonpublic enter- prises (Chen 2003). Many listed companies performed well initially, but their performance deteriorated after the first year. The turnover rate has been much higher than in other countries. The scale of stock market activity is too small to contribute significantly to capital mobility across regions and sectors, some- thing it should be able to do as it grows. Despite the financial sector reforms, rural small businesses still find it harder to obtain credit than do urban-based, state-owned enterprises. The recent arrest and release of millionaire entrepreneur Sun Dawu highlights the problem. Because of the difficulties in raising funds from state-owned banks and credit cooperatives, Mr. Sun solicited deposits from his employees and local rural residents, which violated the state law (Economist 2004). Anecdotal evidence aside, more research is needed to quantify whether the capital market has become more integrated or more fragmented. Driven largely by institutional reforms, the Chinese economy has experienced a dramatic transformation over recent decade^.^ The share of agricultural GDP in total GDP declined from more than half in 1952 to less than 20 percent in 2001, while the share of the rural nonfarm sector increased from almost zero to more than a quarter. Coupled with these structural changes was a massive shift of labor from the lower productivity agricultural sector to the higher pro- ductivity nonfarm sector. Growth in labor and capital productivity by region and sector highlights the dramatic changes in factor markets and economic structure over the period 1978-2001 (tables 1 and 2). Labor and capital productivities are calculated as the ratios of GDP to labor and capital; they are therefore measures of average not marginal productivity. There are large regional variations in labor productivity, and they have widened over time. The northeast region had the highest labor productivity in 1978, but by 2001 it had fallen well behind the eastern region. The regional gap between the west and the rest of China has worsened over time. Compared with labor productivity, the regional disparities in capital productivity are much smaller, and they have narrowed over time. Labor productivity grew fastest in the rural nonfarm sector and slowest in the agricultural sector (see table 2). Labor productivity began at a relatively low level in agriculture, and the gap with other sectors is now much wider. The transfer of rural labor from farm to nonfarm activities will undoubtedly have enhanced overall economic growth and labor productivity. The rural nonfarm sector also experienced the most rapid growth in capital productivity and by 2001 had achieved the highest level of all sectors. These disparities highlight capital market imperfections and the hunger for credit and capital that remains within rural areas for nonfarm activities. Broadening access to credit and investing more in the rural nonfarm sector would enhance economic efficiency and growth. A comparison of the labor productivity of the industrial and service sectors relative to agriculture for China and several other Asian countries helps to put China's economic transformation in a broader international perspective (table 3). The differences are stark. The labor productivity ratio of industry to agriculture is much higher in China than in other Asian countries. Moreover, while the ratios for other countries have generally remained stable or fallen, the ratio for China has risen substantially over the past 20 years. The same is true for the labor productivity ratio between services and agriculture. These extre- mely high ratios for China as well their increasing trends are symptomatic of 2. Lin (1992) provides a good reference for rural reforms; Groves and others (1994) cover the reforms of state-owned enterprises; Lau, Qian, and Roland (2000) explain the rationale behind the successful price reforms. Zhang and Tan 285 TABLELabor and Capital Productivity by Region 1. Productivity China East Central Western Northeast Labor productivity 1978 1984 1990 1995 2001 Growth rate (%) Capital productivity 1978 1984 1990 1995 2001 Growth rate (%) Note: The unit of labor productivity is 1978 yuan; the unit of capital productivity is 1978 yuan per 100 yuan capital stock. East includes the municipalities of Be~jing,Shanghai, and Tianjin, and the provinces of Fuiian, Guangdong, Hainan, Hebei, Jiangsu, Shangdong, and Zhejiang. Central includes Anhui, Henan, Hubei, Hunan, Jiangxi, and Shanxi Provinces. West includes the autonomous regions of Nei Mongol, Ningxia, Tibet, and Xinjiang, and the provinces of Gansu, Guangxi, Guizhou, Ningxia, Qinghai, Shanxi, Sichuan, and Yunnan. Northeast includes Heilongjiang, Jilin, and Liaoning Provinces. Source: Calculated by the authors based on the data for 28 provinces, which are slightly differ- ent from those based on national data. For details on the data see supplemental appendix S.1, available at http://wber.oxfordjournals.org. major distortions in China's factor markets. There appears to be considerable potential for further economic growth simply by reallocating labor and capital among sectors. This section updates Young's (2000) analysis of the trends in product market integration. As in Young, the analysis uses the following sum of the squared deviations of the sectoral output shares of China's provinces from the group average to the degree of product market integration: Unweighted measure : X(S,,3,)' - (1) i j Weighed measure : N*wi(Si,- s,)' (2) i j where Sij denotes the share of sector j in province i's output; Sj is the group average S,, across provinces; wi denotes the province's share of total GDP of N 286 T H E W O R L D B A N K E C O N O M I C REVIEW TABLELabor and Capital Productivity by Sector 2. Productivity China Agriculture Urban industry Urban services Rural nonfarm Labor productivity 1978 868 346 3,245 1,949 623 1984 1,260 509 3,783 2,883 856 1990 1,841 585 5,713 4,615 1,510 1995 3,356 761 8,597 6,275 4,917 2001 5,949 987 23,074 9,573 8,193 Growth rate (%) 8.7 4.7 8.9 7.2 11.9 Capital productivity 1978 36 52 46 19 22 1984 42 74 45 26 30 1990 41 78 38 30 59 1995 53 74 45 33 121 2001 52 57 51 25 192 Growth rate (%) 1.6 0.4 0.5 1.1 9.8 Note: The unit of labor productivity is 1978 yuan; the unit of capital productivity is 1978 yuan per 100 yuan capital stock. Source: Calculated by the authors based on provincial data. For details on the data see sup- plemental appendix S.l, available at http://wber.oxfordjournals.org. provinces and Sj = CwiSji. In the absence of trade, a region would return to an autarky type of Robinson economy, with a production structure diversified to cope with daily needs for food, clothes, shelter, and so on. Therefore, without trade, the likelihood of having a specialized production structure is much smaller than with trade integration. It is expected that the more barriers there are to interregional trade, the more similar the composition of output across provinces and the smaller the value of the measures. Graphing the unweighted and weighted measures of the composition of output shares for 1978-2001 shows similar results-the composition of output con- verges up to the early 1990s and diverges thereafter (figure 1).Product market development follows a U-shaped curve. An initial decline is followed by an upward trend that leads to a higher overall degree of regional specialization in 2001 than in 1978. The convergence between 1978 and the early 1990s replicates Young's (2000) finding that China's product market became more fragmented. However, the upward trend of the measures since the early 1990s indicates that product markets have become more integrated. The evolving pattern of regional integration reported here for a four-sector disaggregation of GDP also echoes the findings of Bai and others (2004) based on a 36-industry breakdown. The turn- ingpoint coincides with the time when the central government took serious measures to remove interregional trade barriers. The initial market reforms may - have brought about more distortions in the short run, but with deepening reform, the barriers in the product markets were broken down over time. Figure 2 presents the standard deviation of the logarithmic provincial GDPs per capita of farming, urban industry, urban service, and rural nonfarm Zhang and Tan 287 TABLETrends in the Labor Productivity of the Industry and Service Sectors 3. as a Ratio of Agricultural Labor Productivity, China and Other Selected Asian Countries, Various Years Countrylyear Industrylagriculture Serviceslagriculture China 1978 7.0 4.9 1988 4.6 3.8 1995 5.4 3.2 2001 7.5 4.0 Philippines 1989 1995 2002 Korea, Rep. 1987 1995 2002 Indonesia 1993 1998 2002 Malaysia 1987 1995 2001 Taiwan, Cbina 1981 1988 1995 2002 Unzted States 1987 1.5 1.6 1995 1.8 1.7 2001 1.4 1.3 Source: World Bank, various years, World Development Indicators. activities. The variations in output per capita of urban industry and urban ser- vices are steady up to 1990 and then increase rapidly. The standard deviation of output per capita of farming increases by 81 percent from 1978 to 1994 and levels off thereafter, while the spatial distribution of rural nonfarm activity becomes increasingly uneven over the whole sample period. However, as Young (2000, p. 1111) notes: "The imposition of trade barriers has clear FIGURE Convergence in the Composition of Output 1. - Unweightedmeasure +Weighted measure Note: The measures are the weighted and unweighted sum of squared deviations of the sectoral output shares of China's provinces from the national average. Source: Authors' analysis based on data described in supplemental appendix S.l, available at http://wber.oxfordjournals.org implications for the interregional variation in output shares; it has no predic- tion regarding the variation in absolute output levels." Nonetheless, the vari- ations of output per capita in the four sectors offer useful information on the evolution of spatial distribution of economic activities. IV. V A R I A T I O N S I N M A R G I N APRODUCTSF C A P I T AALN D L A B O R L O Following the analysis above of recent trends in product market integration, this section turns to an analysis of possible fragmentation in factor markets. Resource allocation is most efficient when the marginal product of each input is equalized across sectors and regions. Thus intersectoral and interregional FIGURE 2. Standard Deviation of In GDP Per Capita , +- Urbanservices++Rural nonfarm 1 Source: Authors' analysis based on data described in supplemental appendix S.1, available at http://w ber.oxfordjournals.org. Zhang and Tan 289 variations in the marginal product of each factor can show the degree of factor market distortions and hence opportunities for achieving greater economic effi- ciency through improved factor all~cation.~ Assume that real value added (GDP) by sector follows a well-behaved, neo- classical production function: where Xi,, is input j for sector i in year t. A thornier question is what functional form of the production function to use. Considering both econometric esti- mation and theoretical consistency, the following Cobb-Douglas functional form can be specified:4 whereAit = aio +aitt +ajttt 2 D, is a set of year dummy variables, and cit is the corresponding coefficient. The parameters in equation (4) corresponding to labor and capital are their elasticities. The estimated function for agriculture includes arable land as a separate input in addition to capital and labor. Because arable land area does not change much and is location specific, provincial dummy variables cannot be used to control for potential heteroscedasticity. As a compromise, dummy variables for the eastern, central, and western regions are added to the pro- duction functions. To capture technological change over time, the time trend and its square are included in one specification. In a second specification, the fixed effects of year dummy variables are added. To estimate production functions for each of the four sectors, data are used for 28 provinces for 24 years (1978-2001), providing a panel of 672 obser- vations. Tibet is excluded mainly because of lack of data. For data consistency, Chongqing and Hainan Provinces are included in Guangdong and Sichuan Provinces, although they were separated in 1987 and 1997. A detailed 3. Desai and Martin (1983)estimated the efficiency loss due to resource misallocation in industry in the former Soviet Union using a similar method. Syrquin (1988)conducted a similar exercise. 4. It is well known that rhe Cobh-Douglas form has caveats. It assumes constant returns to scale and strong separability among inputs. To test the robustness of the results on the first caveat, Zhang and Tan (2004)present an alternarive specification using a varying coefficient model, and the basic findings are the same. Several flexible functional forms have been put forward to address the separability problem. However, their limitations have been increasingly recognized in the empirical literature (Chambers 1988). For example, the multicolIinearity problem inherent among the interactive terms and the fewer restrictions on the underlying production technology often lead to results that do not make much economic sense. description of the data used is provided in the supplemental appendix (avail- able at http://wber.oxfordjournals.org/). The results of the estimated production functions for the four sectors under two different specifications are presented in table 4.5 Because agricultural output is measured as value added, intermediate inputs such as fertilizer are excluded from output measures by definition. Including fertilizer and other intermediate inputs is more appropriate in estimating a production function for gross output. The results under the two different specifications are similar. The adjusted R ~ S are high for all the regressions, indicating a good fit. The year dummy variables in the first specifications are jointly significant in all four regressions. Most coefficients for the time trend variables in the second specifi- cation are statistically significant. The regression results for agriculture indicate that land still plays an impor- tant role in Chinese agricultural production. Among the regressions for all the sectors labor elasticity is larger than capital elasticity, indicating that China's comparative advantage lies in labor-intensive production. Differences in estimated elasticities for the same input across sectors reflect differences in production technology, but on their own do not provide any indi- cation of how efficiently resources are allocated. To obtain such insights, it is necessary to calculate the marginal productivities of each factor. The marginal product of each factor is equal to the product of the estimated elasticity and the corresponding partial factor productivity: Figure 3 presents the marginal product of labor and capital by sector. The mar- ginal product of labor is much higher in urban areas than in the farming and rural nonfarm sector, indicating huge potential gains from rural to urban labor migrations. In 1990, the marginal product of labor in urban industry was about 19 times that of agriculture and the marginal product of labor of urban services was about 13 times that of agriculture. The results are comparable to the findings in Yang and Zhou (1999) that the ratios of the marginal product of labor in the state sector to the agricultural sector was about 15 and 16 between 1988 and 1992. The ratio of the marginal product of labor in the rural nonfarm sector to the farming sector in 1990 was 3.6 in 1990, similar to the 3.7 in 1992 reported by Wang (1997). In 1993, the Company Law was passed to encourage privatization of town and village enterprises. As a result, their share in gross industrial output value jumped from 20 to 25 percent while 5. The calculations of variations in marginal products of factors are rather robust to various specifications in large part because marginal products are determined mainly by factor productivity across sectors rather than by the estimated elasticities. For simplicity, the inequality measures based on several alternative specifications are not reported here but are available on request. TABLEEstimated Production Functions by Sector, China 4. Specification I Specification 11 Agriculture Urban industry Urban service Rural nonfarm Agriculture Urban industry Urban service Rural nonfarm Labor 0.430' (0.026) 0.708" (0.036) 0.601' (0.026) 0.428" (0.026) Capital 0.111' (0.018) 0.263" (0.029) 0.364* (0.031) 0.114' (0.018) Land 0.386' (0.031) 0.386" (0.031) Eastern region 0.081" (0.039) 0.376" (0.039) 0.373' (0.051) -0.325' (0.056) 0.079' (0.039) Central region -0.203' (0.033) -0.152' (0.040) 0.107' (0.051) -0.391' (0.055) -0.203" (0.032) Western region -0.521" (0.035) 0.044 (0.047) 0.018 (0.048) -0.818' (0.057) -0.522' (0.035) Year dummy variable Yes" Yes" Yes" Yes' T 0.071:'(0.005) T~IIOO 0.112' (0.020) Adjusted R' "Significant at the 10 percent level. Note: Figures in parenthesis are standard errors. Source: Authors' analysis based on data described in supplemental appendix S.l, available ac http:Mwber.oxfordjoumals.org. FIGURE Marginal Products of Labor and Capital 3. , 25,000 Marginal product of labor Marginal product of capital I 80 1 +Agricukure -m- Urbanindustry +Agricukure -B- Urbanindustry L-A- Urbanservices +Rural - nonfarm L. --tUrbanservices+e- Rural nonfarm - -_ --A Source: Authors' analysis based on data described in supplemental appendix S.l, available at http://wber.oxfordjournaIs.org. that of state enterprises dropped from 43 to 34 per cent from 1993 to 1995 (China National Bureau of Statistics, China Statistical Yearbook, p. 401, 1996). The large difference in marginal product of labor suggests potential gains in aggregate output from labor mobility across sectors. The graph of the marginal product of capital by sector shows that the nonfarm sector has grown much faster than other sectors and by 2001 has the highest value among the four sectors (see figure 3). The marginal product of capital is lowest in the farming and urban service sectors. Overall, the differences in marginal product of factors across sectors are quite large. A generalized entropy (GE) inequality measure was used to quan- tify the degree of variation in the marginal products of inputs across the 4 sectors and 28 provinces.6 Because each province has four sectors, there are 6. Other measures are also used, and the results are similar. Following Shorrocks (1980), the GE measure in the marginal product of capital (k)can be written as: where MUkdenotes the marginal product of factor k for sector j in province i, /I. is the arithmetic sample mean, and w, is the share of GDP of sector j for province i in total GDP. GF.(O)is the mean logarithmic deviation, GE(1) is the Theil index, and GE(2) equals halt the square of the coefficient of variation. In principle, the GE measures are sensitive to various parts of the distribution depending on the selected value of c. The simplest form of this equation was used in which c = 0. When c = 0, it is the mean logarithm deviation and more sensitive to the bottom part of the distribution. The results are similar for c = 1 and c = 2. The reason for using GE is its appealing property of decomposing overall inequality into between- and within-group subcomponents. Zhang and Tan 293 2,688 observations in all. Figure 4 graphs the variations in the marginal pro- ducts of labor and capital. The marginal product of labor has shown some convergence over the reform period, except in the last five years of the analysis (which may be the result of changes in the way the labor surveys were conducted during those years; see supplemental appendix). Variation in the marginal product of capital, in con- trast, was steady between 1978 and the early 1990s before rising substantially. The divergence in the marginal product of capital during the 1990s indicates greater fragmentation of capital markets. This finding is consistent with that of Boyrau-Debray and Wei (2003). These results suggest that as competition intensified in product and labor markets, distortions may have shifted to banking, real estate, and infrastructure projects. In this sense, the findings support Young's (2000)argument that partial reforms may lead to more distor- tions in the rest of the economy. The GE family of inequality measures can be decomposed into the sum of within- and between-group components for any given partitioning of the popu- lation into mutually exclusive and exhaustive groups. Figure 4 graphs the between- and within-goup (region and sector) components of the variation in the marginal products of capital and labor. The ratio of the between-group component to overall inequality is called the polarization index (Kanbur and Zhang 1999; Zhang and Kanbur 2001). Intersectoral variations in the marginal products of labor and capital contribute far more to overall inequality than interregional variation. In particular, the sectoral polarization index on the marginal product of capital has increased. This provides further evidence that FIGURE Variations in Marginal Product of Labor and Capital 4. Marginal product of labor by sector Marginal product of capital by sector 80 ., 80 1 I Marginal product of labor by region Marginal productof capital by region 80 7 80 ; Note: The blank bars show the within-sector or -region variation, while the solid bars show the between-component variation. Source: Authors' analysis based on data described in supplemental appendix S.l, available at http://wber.oxfordjournals.org. 294 T H E W O R L D B A N K E C O N O M I C R E V I E W as the reform process has deepened in the product market, the capital market has become more distorted. These results indicate that there is room to improve China's overall econo- mic efficiency simply by reallocating factors among sectors and regions. Reversing the entrenched urban-biased investment policies and undertaking in-depth reforms within the financial sector would not only have the greatest impact on economic efficiency but would also promote greater equity as most poor people live and work in rural areas. How large are the potential gains from improving factor market performance? To answer this question, estimated production functions from the first specifi- cation in table 4 are used to calculate the potential increases in national GDP resulting from simulated factor reallocation^.^ Supplemental appendix S.2 reports the underlying models and baseline information. As a first step, the models are calibrated to obtain the constant terms in the production functions of the four sectors based on the estimated elasticities on labor, capital, and GDP information for 2001. Doing that means that the production functions will predict the actual results for 2001. Next, the calibrated models are used in the four sectors to conduct policy simulations. Considering the low level of labor productivity in the agricultural sector, the first experiment is to move additional labor out of that sector. With 2001 as a baseline, three scenarios are evaluated: moving 1, 5, and 10 percent of the agri- cultural labor force out of agriculture and distributing it equally among the other three sectors (table 5). Reallocating even 1 percent of the agricultural labor force could increase national GDP by 0.9 percent. Reallocating 5 percent or 10 percent increases national GDP by 4.4 percent or 8.8 percent. The results are supported by an independent early study by Yang and Zhou (1999),who find gains in aggregate output of 0.7 percent, 3.1 percent, and 5.8 percent based on the same three hypothetical percentage transfers of labor using 1992 as a baseline. The second experiment simulates a change in the current urban-biased poli- cies by shifting capital from cities to rural areas while keeping total capital con- stant. Reallocating 1 percent, 5 percent, and 10 percent of urban capital to rural areas leads to gains in national GDP of 0.5 percent, 2.1 percent, and 3.9 percent. 7. Policy simulations point out only the potential gains from reform. However, questions remain on the mapping from simulations to actual reforms. In addition, there are no standard errors. Therefore the precision cannot be assessed. It is likely that the simulations results depend on the underlying functional forms as well as the accuracy of the data. We are reassured in that simulations based on a varying coefficient model have led to similar findings (Zhang and Tan 2004). In 5, we also check the robustness of the results by undertaking similar simulations with a baseline of higher labor productivity in the agriculture sector. Zhang and Tan 295 TABLE5. Impact of Alternative Policy Simulations on China's GDP Experiment Results Move x% of the agricultural labor force out of 1% 5% 10% farming Change in GDP (%) 0.89 (0.89) 4.42 (4.22) Reallocate x% investment from cities to rural areas 1% 5% Change in GDP (%) 0.46 (0.41) 2.13 (1.90) Add x billion yuan of investment in rural areas 10 50 Change in GDP over 2001 (%) 0.03 (0.03) 0.15 (0.14) Change in GDP over 2001 (billion yuan) 3.66 18.26 Add x billion yuan of investment in urban areas 10 50 Change in GDP over 2001 (%) 0.01 (0.01) 0.04 (0.04) Change in GDP over 2001 (billion yuan) 0.92 (1.03) 4.58 (5.16) The ratio of returns to investment in rural areas to 3.99 (3.60) 3.98 (3.59) urban areas Note: The figures in the parentheses are the simulation results based on adjusted national GDP data. Source: Authors' analysis based on data described in supplemental appendix S.l, available at http://wber.oxfordjournals.org. The third experiment assumes that the government allocates all the addi- tional investment in rural areas and distributes it equally between the agricul- tural and rural nonfarm sectors. The investment is converted into capital stock using a discount rate of 4 percent and a national fixed asset price index.8 An additional 10 billion yuan of investment in rural areas yields a 0.03 percent increase in national GDP, equivalent to 2.9 billion 2001 yuan. Considering that the farm and rural nonfarm sectors are labor intensive, this scenario would likely also boost the incomes of many of the poorest people in China. When investment increases to 50 billion yuan, national GDP rises by 0.15 percent (14.3 billion yuan) and when it increases to 100 billion yuan GDP rises by 0.29 percent (28.4 billion yuan). Because the capital does not vanish immediately, the long-term impact is much higher. Assuming a 4 percent dis- count rate, the annual internal rate of returns to the investment in rural areas is more than 20 percent. The next experiment considers a counterfactual scenario in which all the additional investment is distributed evenly in the two urban sectors. Under the three scenarios of investment of 10, 50, and 100 billion yuan, national GDP increases by 0.92, 4.58, and 9.16 billion yuan, respectively. As shown in the last row of the table, the rate of returns to rural investment is almost four times of that to urban investment. 8. For the period 1991-2001, the national fixed asser price index is available from the China Statistical yearbook. However, it was not published prior to 1991. Therefore, the narional GDP deflator is used a proxy for the period, 1978-91. For the whole period the calculated capital price index is 3.53, compared with the published GDP deflator of 3.33. The National Statistical Bureau adjusted national GDP figures based on the first economic census in 2004. To check the robustness of the results, the con- stant terms in the four production functions were recalibrated as shown in sup- plemental appendix S.2 using the adjusted 2001 GDP data by sector, and the same set of simulations was undertaken. The basic results are similar to those based on original GDP figures (see table 5). The policy simulation highlights the potential economic gains from reallocat- ing factors from low- to high-productivity sectors. Removing barriers to labor movement, reversing the urban bias in government investment policies, and deepening reforms would significantly enhance overall economic growth. In addition, these policy changes could bring about favorable distributional effects by reducing regional and sectoral inequalities. Since large inequalities are a potential source of social conflict and instability, the far-reaching social impact of these policies could be equally important. VI. C O N C L U S I OANNSD P O L I C YI M P L I C A T I O N S The aim of China's reforms is to reduce economic distortions and improve effi- ciency. This article has examined the changing patterns of distortions during the reform process, how past policies have contributed to these distortions, and the estimated cost to the economy through lower output and greater regional and sectoral disparity. The empirical findings indicate that product markets in China have become more integrated after a short period of increasing fragmen- tation in the early reform period. Labor markets also have become increasingly integrated due to a large shift in the labor force from the agricultural sector to nonfarm sectors and relaxed constraints on migration. However, intersectoral differences in the marginal product of capital have grown during the reform period. Local governments, which have been collecting rents in a partially reformed system, are the interim winners from reform. In the short run, distortions might beget more distortions, as Young (2000) has shown. However, in response to the increasing fragmentation in product markets, the government has undertaken measures to remove local protections. Consequently, there are fewer and fewer rents to be collected in the product and labor markets over time, and the distortions have been increasingly squeezed into the financial and land markets (including infrastructure and real estate). For local governments, these are the two last bastions for rent collection, as well as breeding grounds for corruption. Looking only at the product market suggests that the market might have become distorted in the short run. However, as the government responded to the problems with deepening reforms, the market became inte- grated. When all the sectors are considered, however, the results seem to support Young's argument that as some distortions in a partially reformed economy are removed, new distortions may be added. The key is whether the government can continue to add new reforms to squeeze out the distortions in Zhang and Tan 297 the capital market as those in the product and labor markets were squeezed out before them. The continuing large differences in labor and capital productivity across sectors suggests that China still has great potential for further efficiency gains through continued structural change. To realize this potential, however, restric- tions on factor movement, in particular, intersectoral capital movement, need to be removed. Efficient capital markets that can funnel new investment to sectors with higher returns still need to be developed. The higher capital returns in the rural nonfarm sector suggest that more aggressive government policies should be sought to increase investment there or at least not hinder its movement. Such policies will not only improve overall economic performance, but will also narrow the development and inequality gaps between the rural and urban sectors. Similarly, the government should encourage labor move- ment from agriculture to rural enterprises, urban industry, and service sectors since labor productivity in these sectors continues to be much higher than in the agriculture sector. While empirical estimates and policy simulations can provide rough order of magnitude estimates of structural problems, policy recommendations on gradual elimination of these distortions need to take into account complex issues of political feasibility, sequencing, implementation problems, downside risks of policy measures, nature of vested interests and how to overcome them, the need to minimize negative side effects, and the effects on equity, regional disparity, and rural-urban inequality. More research is needed to understand the political economy dimensions that have at times seriously constrained the pace of reform. Nonetheless, simulations of alternative policy proposals and their estimated effects could act as useful inputs to policymaking. VII. S U P P L E M E N T AMRAYT E R I A L Supplementary material is available at: http://www.wber.oxfordjournals.org/ Bai, Chong-En, Yingjuan Duan, Zhigang Tao, and T. Sarah Tong. 2004. "Local Protectionism and Regional Specialization: Evidence from China's Industries." Journal of International Economics 63(2):397-417. Boyreau-Debray, Genevieve, and Shang-Jin Wei. 2003. "How Fragmented is the Capital Market in China?" Paper presented at the Workshop on National Market Integration organized by the World Bank Beijing Office, 6 September, Beijing. Development Research Center of the State Council, Beijing. Chambers, Robert G. 1988. Applied Production Analysis: A Dual Approach. New York: Press Syndicate of the University of Cambridge. Chen, Zhiwu. 2003. "Capital Markets and Legal Development: The China Case." China Economic Review 14(4):451-72. 298 T H E WORLD B A N K ECONOMIC. REVIEW China, National Bureau of Statistics. 1990. Historical Statistical Materials for Proz~~nces,Autonomous Regions and Municipalities 1949-1989. Beijing: China Statistical Publishing House. . 1996. China Fixed Asset Investment Statistical Materials, 1950-95. Beijing: China Statistical Publishing House. . 1997a. Calculation Methods of China Annual GDP. Beijing: China Statistical Publishing House. . 1997b. The Gross Domestic Product of China, 19.52-95. Dalin, China: Dongbei University of Finance and Economics Press. .Various years. China Rural Statistical Yearbooks. Beijing: China Statistical Publishing House. .Various years. China Statistical Yearbooks. Beijing: China Statistical Publishing House. .Various years. China Township and Village Enterprise Statistical Yearbook. Beijing: China Statistical Publishing House. China State Council, Development Research Center. 2003. "Research on Measures, Objectives and Degrees of Local Protection in Chinese Market: An Analysis Based on Sample Survey." Paper pre- sented at the Workshop on National Market Integration organized by the World Bank Beijing Office, 6 September, Beijing. Chow, Gregory C. 1993. "Capital Formation and Economic Growth in China." Quarterly journal of Economics 108(3):809-42. de Brauw, Alan. Jikun Huang. Scott Rozelle, Linxiu Zhang, and Yigang Zhang. 2002. "The Evolution of China's Rural Labor Markets During the Reforms." journal of Comparative Economics 30(2):329-53. Desai, Padam, and Ricardo Martin. 1983. "Efficiency Loss from Resource Misallocation in Soviet Industry." Quarterly journal of Economics 98(3):441-56. The Economist. 2004. "Wealth in Asia: China Speaking Out." January 3: 26-27. Fan, Shenggen, Xiaobo Zhang, and Sherman Robinson. 2003. "Structure Change and Economic Growth in China." Review of Development Economics 7(3):360-77. Fleisher, Belton, and Dennis Tao Yang. 2003. "China's Labor Market." Paper presented at the Stanford Center for International Developmerit conference on "China's Market Reforms," Stanford University, 19-20 September, Palo Alto, Calif. Groves, Theodore, Yongmiao Hong, John McMillan, and Barry Naughton. 1994. "Autonomy and Incentives in Chinese State Enterprises." Quarterly jourrzal of Economics 109(1):183-209. Huang, Jikun, Scott Rozelle, and Ming Chang. 2004. "Tracking Distortions in Agriculture: China and Its Accession to the World Trade Organization." World Bank Economic Review 18(1):59-84. Kanbur, Ravi, and Xiaobo Zhang. 1999. "Which Regional Inequality? The Evolution of Rural-Urban and Inland-Coastal Inequality in China (1983-1995)." journal of Comparative Economics 27(4):686-701. Knight, John, and Shi Li. 2005. "Wages, Firm Profitability and Labor Market Segmentation in Urban China." China Economic Revieui 16(3):205-28. Kohli, Ulrich. 1978. "A Gross National Product Function and the Derived Demand for Imports and Supply of Exports." Canadian journal of Economics 11 (2):167-82. Lau, Lawrence, Yingyi Qian. and Grard Roland. 2000. "Reform without Losers: An Interpretation of China's Dual-Track Approach to Transition." journal of Political Economy 108 (1):120-43. Li, Kui-Wai. 2003. "Ch~na's Capital and Productivity Measurement Using Financial Resources." Economic Growth Center Discussion Paper 851. Yale University, New Haven, Conn. Lin, Justin Yifu. 1992. "Rural Reforms and Agricultural Growth in China." Americari Economic Review 82(1):34-51. Lin, Justin Yifu. 2004. "Viability and the Development of China's Capital Markets." China & World Economy 12(6):3-10. Lin, Justin Yifu, Fang Cai, and Zhou Li. 1996. The China Miracle: Development Strategy and Economic Reform. Hong Kong: The Chinese University Press. Zhang and Tan 299 Maddison, Angus. 1998. Chinese Economic Performance in the Long Run. Paris: Organisation for Economic Co-operarion and Development. Meng, Xin. 2000. Labor Market Reform it1 Cl7ina. New York: Cambridge University Press. Ministry of Agriculture. 1980-94, Various Issues. China's Agricultural Yearbook. Beijing: China's Agricultural Press. Poncet, Sandra. 2003. "Domestic Market Fragmentation and Economic Growth in China." ERSA Conference Paper ersa03pl17. European Regional Science Association. Shorrocks, Anthony. 1980. "The Class of Additively Decomposable Inequality Measures." Econometrics 48(3):613-2.5. Syrquin, Moshe. 1988. "Parterns of Structural Change." In Hollis B. Chenery, and T.N. Srinivasan eds., Handbook of Development Economics. Val. I. Amsterdam: North-Holland. Wang, Xiaolu. 1997. "What Contributed to China's Rapid Rural Industrial Growth during the Reform Period?" Ph.D, diss., Department of Economics. The Australia National University, Canberra. Wei, Xiangdong, and C. Simon Fan. 2004. "Converge to the Law of One Price in China." Paper pre- senled at the Allied Social Sciences Annual Meeting, 3-5 ,Januaty,San Diego, Calif. World Bank. 2005. China: lntegration of National Product and Factor Markets-Economic Benefits and Policy Recommendations. Washington, D.C. . Various years. World Development Indicators. Washington, D.C. Yang, Dennis Tao, and Hao Zhou. 1999. "Rural-Urban Disparity and Sectoral Labor Allocation in China." lournal of Development Studtes 35(3):105-33. Yi, Gang. 2003. "History of Financial Market." Paper presented at the Workshop on Nariorral Market lntegratiorr organized hy the World Bank Beijirrg Ofice, 6 Septenrber, Beijing. Development Research Center of the State Council, Beijing. Young, Alwyn. 2000. "The Razor's Edge: Distortions and lncremenral Reform in the People's Republic of China." QuarterlyJournal of Economics 115(4):1091-35. Zhang, Junsen, Yaohui Zhao, Albert Park, and Xiaoq~ngSong. 2005. "Economic Returns to Schooling in Urban China, 1988 to 2001." journal of Comparative Economics 33(4):730-52. Zhang, Xiaobo. 2006. "Fiscal Decentralization and Political Centralization in China: Implications for Regional Inequalicy." Journal of Comparative Economics 34(4):713-26. Lhang, Xiaobo, and Ravi Kanbur. 2001. ''What Difference Do Polarisation Measures Make? An Application to China." Journal of Development Sttrdies 37(3):85-98. Zhang, Xiaobo, and Kong-Yam Tan. 2004. "Blunt to Sharpened Razor: Incremental Reform and Distortions in the Product and Capital Markets in China." Discussion Paper 13. International Food Policy Research Institute, Development Strategy and Governance Division, Washington, D.C. Child Labor, School Attendance, and Intrahousehold Gender Bias in Brazil Patrick M. Emerson and Andre' Portela Souza An extensive survey data set of Brazilian households is used to test whether intrahousehold gender bias affects the decisions of mothers and fathers to send their sons and daughters to work and to school. An intrahousehold allocation model is - examined in which fathers and mothers may affect the education investment and the child labor participation of their sons and daughters differently because of differences in parental preferences or differences in how additional schooling affects sons' and daughters' acquisition of human capital. Brazilian household survey data for 1998 are used to estimate the impact of each parent's education on the labor market participation and school attendance of their sons and daughters. For labor market participation, the father's education has a greater negative impact than the mother's education on the labor status of sons. The father's education also has a greater impact on sons' labor status than on daughters'. For schooling decisions, the mother's edu- cation has a greater positive impact than the father's education on daughters' school attendance, but fathers have a greater positive impact on sons' school attendance than on daughters'. JEL codes: 520, 0 1 2 , 054 This article uses an extensive survey data set of Brazilian households to test whether intrahousehold gender bias influences the decisions of mothers and fathers to send their sons and daughters to work and to school. The results suggest that fathers generally have a greater impact on decisions about sons and mothers generally have a greater impact on decisions about daughters. These results support models of intrahousehold bargaining in the child labor and schooling decisions of a family, even though most theoretical work on child labor has assumed a unitary family model. Patrick Emerson (corresponding author) is an assistant professor at Oregon State University; his c-mail address is patrick.emerson@oregonstate.edu. Andri. Portela Souza is a professor at the SZo Paulo School of Economics at the Funda~iaGetulio Vargas; his e-mail address is andre.portela.souza@fgv.br. For valuable comments and advice the authors thank John Abowd, Kaushik Basu, Francine Blau, Jaime de Melo, Lawrence Khan, and three anonymous referees. This article benefited from presentations at the 2002 Latin American and Caribbean Economics Association hleetings and the 2001 Western Economics Association International Meetings. A supplemental appendix to this article is available at http://wber.oxfordjournals.org/. THt WOK1 D B4NK ECONOMIC. R~VIEW, 21, VOl. NO. 2, pp. 301-316 do1:10.1093/wber/lhm001 Advance Access Publication 8 March 2007 ( )The Author 2007. 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 The recent theoretical literature on child labor generally assumes (follow- ing Becker 1982) that parents have common preferences and are altruistic toward their children (Baland and Robinson 2000; Bell and Gersbach 2000; Dessy 2000; Emerson and Souza 2003; Basu and Van 1998). Additionally, the empirical literature on child labor has explored mainly the relation between the economic conditions and incentives of the family unit and child labor outcomes (Emerson and Souza 2003; Ray 2000; Grootaert and Patrinos 1999; Jensen and Nielsen 1997). Although a unitary model of intrahousehold allocations is a valid starting point for focusing on the poverty dimension of child labor, this emphasis does not account for other potentially important factors. Recently, a few studies have examined intrahousehold allocations explicitly. For example, Basu (2006) and Ridao-Cano (2000) extend intra- household behavior to child labor decisions. Both suggest that fathers and mothers have different impacts on the labor supply of their children and that this is potentially related to their relative bargaining power. Neither, however, explores gender bias within the intrahousehold allocation decisions.' There is also an extensive literature on gender differences in human capital investments and outcomes that presents some evidence on intrahousehold gender bias. Sen (1990), for example, reports that males significantly out- number females in Asia and North Africa, opposite the pattern found in North America and Europe. Other studies have shown that sons are favored in the intrahousehold allocation of nutrients and have better anthropometric out- comes (Behrman 1988; Sen 1984). Perhaps even more compelling are recent studies that find that the gender bias in child inputs or outcomes is related to the gender of the parent who con- trols the distribution of child resources. In a study of families in Brazil, Ghana, and the United States, Thomas (1994)finds that children's health achievement (as measured by height for age) is linked to the educational attainment and nonlabor income of the parent of the same sex as the child. In other words, sons are healthier (taller) the more education and nonlabor income the father has, and daughters do better the more education and nonlabor income the mother has. This finding suggests that there may be differences in the prefer- ences of the parents. Studies of Brazil also suggest that there may be important gender-based differences in allocations within a household. For example, Thomas (1990) shows that unearned income controlled by mothers has stron- ger impacts on family's health than income under fathers' control (see also Tiefenthaler 1999). None of these studies, however, examines intrahousehold gender bias in the child labor context. In a recent study of child labor, Emerson and Souza (2003) find strong evi- dence of intergenerational persistence in child labor among families in Brazil. 1. Basu's theoretical contribution goes further, to include the possibility that the choices taken by the individuals can affect their bargaining power. Emerson and Souza 303 Specifically, people who start work at a younger age end up with lower earn- ings as adults, and children are more likely to work the younger their parents were when they entered the labor force and the lower the educational attain- ment of their parents and grandparents. These findings are consistent with unitary models of child labor and poverty persistence (Emerson and Souza 2003; Bell and Gersbach 2000; Dessy 2000; Basu 1999; Glomm 1997) in which parents' labor as children reduces their ability to gain human capital through schooling, making them unable to command high enough wages as adults to afford to keep their children out of the labor force. These results also hold when the analysis is performed for sons and daugh- ters separately. That is, there is a persistence of child labor from parents to sons as well as from parents to daughters. This raises the question, is this effect different for sons and daughters based on the individual level of human capital of their mothers and fathers? This study uses Brazilian household survey data to estimate the impact of fathers' and mothers' education on the labor market status and school atten- dance of their sons and daughters separately. It finds compelling evidence that the father's education has a greater impact on the labor status of sons than does the mother's education and a greater impact on the labor status of sons than of daughters. It finds equally compelling evidence that the mother's edu- cation has a greater positive impact on the school attendance of daughters than does the father's education and that the father's education has a greater impact on the school attendance of sons than of daughters. Section I discusses models of household allocations and how altruistic fathers and mothers may have different impacts on their sons and daughters due to differences in their preferences or differences in how additional edu- cation affects their children's acquisition of human capital. Section I1 describes the 1998 Brazilian National Household Survey data used in this study and the variables used in the regression estimations. Section I11 discusses the empirical results, and section IV offers some policy implications. The concept of intrahousehold allocation decisions has a long established place in the literature. Two classes of models typically used in the intrahousehold allocation literature that allow for differences in parental preferences are the family bargaining model (Lundberg and Pollak 1993; McElroy 1990; McElroy and Horney 1981) and the collective model (Browning and Chiappori 1998; Chiappori 1992, 1988). Bargaining models assume that household allocation outcomes reflect a bargaining process in which household members seek to allocate the resources they control to the goods that they individually prefer. The resulting equilibria are sensitive to the threat point definition and equili- brium concept assumed. The collective model leaves unspecified the underlying nature of the allocation process within the household but assumes that the resource allocations are Pareto effi~ient.~ As male and female household heads may have different preferences in general, and different specific preferences for the outcomes for their children depending on the gender of the child, the allocation of resources within a household can be seen as the result of some kind of resolution of the preference differences of the male and female heads. This resolution may depend on the relative bargaining power of each individual involved in the allocation decision, and this power may depend on several factors, most commonly, income and education. In general, in the literature, households are modeled as consisting of two heads (mother and father) and some number of children, who can be sons or daughters. Generally, both fathers and mothers are considered altruistic in that they value the consumption of each member of the household and the human capital achievement of their children. In a low-income country setting where child labor is common, the children of the household can go to school, go to work, or spend time in both activities. The amount of schooling children receive determines the wages they are able to command as adults, and children who work are not able to acquire as much education as those who do not. Therefore, the amount of labor income the father and mother bring into the household depends on how much schooling they received as children. Parents who were child laborers command lower wages and are more likely to be impoverished and so are more likely to demand that their children work to sup- plement the family income, a repetitive pattern that is termed the intergenera- tional persistence of child labor. This intrahousehold bargaining framework generates a number of interesting empirical implications. First, the higher is the parents' human capital attain- ment, the higher is current parental income and the less needed is the child's contribution to current household income. If parents care about the human capital attainment of their children, higher levels of parental education should lead to higher education for their ~ h i l d r e nSecond, the human capital of the . ~ parents may have differential effects on the human capital production functions of the children, and these differences may depend in part on the gender of the parent and of the child. Third, parental preferences for their children's human capital attainment may vary by gender of the child, and this may lead to differ- ential investments based on gender. The focus of this article is on the differ- ential effects of fathers and mothers on the child labor participation and education of their sons and daughters. To fix ideas, consider a four-person household that consists of a father, mother, son, and daughter. Now consider an increase, ceteris ~aribus,in the father's human capital. The effect on the son's optimal education function is 2. For a summary of intrahousehold allocation models, see Behrman (1997) and Strauss and Thomas (1995).Basu (2006)demonstrates that power within the household can be endogenous. 3. Which can in turn lead to the children's being less likely to work (see Emerson and Souza 2003). Emerson and Souza 305 potentially twofold. First, more human capital for the father means more income for the family, and thus the family needs the son to provide less income, and this reduces the amount of child labor. The reduction in child labor means more schooling and thus more human capital for the son, which the family values. Second, the higher human capital of the father may also increase the return to the son's education since having a more educated parent in the household may enhance the learning environment. The family will therefore have an increased incentive to invest more in the son's education. However, and of key interest here, this effect of the increase in the father's human capital need not be the same for all children (see Horowitz and Wang 2004). This should reduce child labor equally across all children if additional schooling increased children's human capital equally. But if additional school- ing affects children's acquisition of human capital differently, the reduction in child labor would vary across children. Thus, the increase in schooling will differ across the two children because of their idiosyncratic human capital technologies. The increase in the father's human capital may also affect the marginal returns to education of the son and daughter differently. Finally, the father may favor one child over the other, and his additional human capital may give him stronger bargaining power to impose his preferences. These differences can lead to different investments in education for the son and daughter of the family. Thus, the idiosyncratic nature of the human capital technology and parental preferences that vary across children can lead to different impacts by each parent on the same child as well as different parental impacts across children. For example, for the same child, it is possible that an increase in the father's human capital will have a different impact than an equal increase in the mother's human capital. In addition, it could be that the effect of an increase in mother's human capital could be different for the son and daughter. Finally, note that these different impacts can also be driven by different parental prefer- ences for the human capital of their children and the relative bargaining power of the father and the mother. The impact of gender on intrahousehold allocations in a child labor environ- ment is explored as an empirical issue in section 111, following the brief data description below. The data used in this study come from the 1998 Brazilian National Household Survey (PNAD) conducted by the Brazilian Institute of Geography and Statistics. The PNAD is an annual labor force survey that covers all urban areas and most rural areas in Brazil. The 1998 PNAD included a special module on the labor market activities of all children seven years old or older. The sample consists of all sons and daughters 7-16 years old who live in a two-parent family.4 Younger children are excluded because compulsory schooling begins at age seven in Brazil, and children over 16 because most definitions consider workers older than 16 as adult laborers, not child laborers. Because the impact of both parents' human capital on the labor status and schooling of children is of primary concern, a sample of observations is used that has complete information on each parent's characteristics, including years of schooling. Thus families with single heads are excluded from the analysis." Each observation consists of information on the child's characteristics, his or her parents' characteristics, and his or her family characteristics. Finally, all observations for which the age difference between the head of the family or spouse and the oldest child is 14 years or less are excluded. The final sample consists of 26,930 sons and 25,435 daughters. A child is considered working if he or she worked any number of hours during the survey week. A second child labor indicator variable was con- structed that equals one if the child worked 20 hours or more per week. This alternative definition of child labor was used to check the robustness of the main results; all the qualitative results remained the same.6 For each child, information was also obtained on school attendance, gender, race, area of resi- dence, number of children in family, and age and years of schooling of each parent. In addition, a variable was constructed for the income of the family net of the observed child's earnings from his or her main labor activity. (Basic statistics for all the variables used in this analysis are presented in appendix table A-1.) Child labor is widespread in the sample of households in Brazil. Almost 18 percent of all sons work some hours in the labor market, as do almost 9 percent of daughters (table 1). School attendance is also quite high, with almost 93 percent of sons and more than 94 percent of daughters attending 4. The data enable families to be distinguished from households and thus enable siblings to be identified. Additionally. PNAD classifies children as sons and daughters if they are the son or daughter of the head of the household or the spouse. This classification means that children may be classified as siblings who are related legally but not biologically to each other and the parents, but who are living as one family. Finally, the father is called the head of the household if the head is identified as male and the mother is called the spouse (if listed as the opposite sex), and the mother is called the head if the head is identified as female and the father is called the spouse (if listed as the opposite sex). 5. This selection criterion may impose some selection bias if, for example, children in single-headed families are more likely to work. But since the study seeks to capture separate impacts of the father's and mother's schooling, nonlabor income, and child labor status on sons and daughters, the sample with two-parent households is used. 6. The results are not presented here but are available on request. Emerson and Souza 307 TABLE Child Labor and School Attendance: Families with at Least One 1. Child Ages 7-16 School Attendance Sons Daughters Child Labor N o Yes Total No Yes Total N o Number Row (Yo) Column ( % I Yes Number Row (Yo) Column (%) Total Number Row ( "0 ) Column (%) Source: Authors' analysis based on data from the 1998 Brazilian National Household Survey. school at least part-time. Among children who work, more than 81 percent of sons and 84 percent of daughters also attend school. A series of bivariate probit models were estimated to test the impact of in- trahousehold gender difference~on child labor and educational outcomes. Because child labor and schooling decisions are likely related, as evidenced by the high proportion of children in the sample who both work and attend school, a bivariate probit model is useful for combining information from the correlation among the errors of the child labor regression and the child school- ing regression. The bivariate probit model does impose some restrictions on the error terms. To ensure robustness, separate probit models were also estimated, as was a multinomial logit model in which the choices are the four combinations of working and attending school. In both cases, the results are essentially qualitat- ively the same. (These results are available in the supplemental appendix at http://wber.oxfordjournals.org.7) The first bivariate probit model estimated is a regression of the child labor indicator variable and the school attendance indicator variable (for daughters 7. Logit models were estimated as well, with qualitatively the same results; these are available on request. 308 THE W O R L D B A N K E C O N O M I C R E V I E W and sons separately) on the father's and mother's years of schooling, control- ling for child's age and other individual characteristics.' The other character- istics are whether the individual is nonwhite, to control for race effects; whether the individual lives in a rural area or metropolitan area, to control for differential labor and education markets; and the age of the parents as an additional control for income and family size, to control for the diluting effects of many children on household resources. As mentioned, parental years of schooling are included as a proxy for the human capital of parents, and the objective is to test for differential impacts of parental human capital on the school attendance and child labor of sons and daughters. The coefficient estimates for child labor suggest that the higher the parent's schooling, the less likely the child is to work (table 2). However, fathers' and mothers' schooling have different impacts on sons' work status. A father's education has a stronger negative impact than a mother's on a son's likelihood to work. The chi-square tests for the parents' schooling coefficients confirm this result, rejecting the hypothesis that a father's and a mother's education have the same impact on a son's probability of working. The coefficient esti- mates for the bivariate probit on school attendance are positive and significant for both the father's and the mother's education levels. More important, the chi-square test reveals that the mother's education has a greater positive impact on a daughter's probability of attending school than does the father's edu- cation. A father's and mother's education do not have different impacts on a son's probability of attending school, however. The results are also presented for the marginal effects (evaluated at the means of the independent variables) and a test of the difference between the marginal effects of each parent's schooling on a son's and a daughter's prob- ability of working (table 3). Because these are separate estimations, the covari- ances are unknown, but considering that there is a strong and positive correlation (0.72) between the two education measures, they are almost surely positive. In these estimations, the covariances are assumed to be zero, providing a stronger test than the standard t-test of the hypothesis that the two marginal effects are not significantly different from zero. A father's education has a stronger and significantly negative impact on a son's probability of working compared with the impact on a daughter, regard- less of school attendance. Also, a father's education has a positive and signifi- cantly greater impact on a son's school attendance than on a daughter's school attendance if they do not work. If they do work, however, a father's education has a significantly larger but negative impact on a son's school attendance than on a daughter's school attendance. 8. Separate bivariate probit regressions are estimated for sons and daughters because the decisionmaking processes for them are likely to be quite different. For example, decisions concerning girls rnay depend on whether girls are needed to help with household chores, whereas that may not be the case tor boys. Emerson and Souza 309 TABLE Bivariate Probit Results on Child Labor and School Attendance 2. Sons Daughters Standard Standard Independent Variable Coefficient Error Coefficient Error Child labor Father's years of schooling - 0.051 0.004 Mother's years of schooling - 0.024 0.004 Age 0.298 0.005 Nonwhite - 0.049 0.023 Rural 0.788 0.026 Metropolitan area - 0.472 0.028 Age of father 0.001 0.002 Age of mother - 0.005 0.002 Number of children 0.043 0.006 Intercept -4.432 0.084 Father's years of schooling = Mother's years of schooling School attendance Father's years of schooling 0.059 0.005 Mother's years of schooling 0.061 0.005 Age -0.105 0.006 Nonwhite - 0.024 0.027 Rural -0.095 0.030 Metropolitan area 0.013 0.031 Age of father --0.001 0.002 Age of mother 0.003 0.002 Number of children - 0.056 0.007 Intercept 2.431 0.098 Father's years of schooling = Mother's years of schooling Rho Wald chi-square Number of observations "Chi-square hP>chi-square Note: White-Huber heteroskedastic consistent errors used in regressions. Source: Authors' analysis based on data from the 1998 Brazilian National Household Survey. In general, the marginal impact of parents' education on child labor is much higher for children who attend school (perhaps the more marginal cases) than for children who do not. For school attendance, the marginal impact of parental edu- cation is higher for children who do not work than for children who do. In fact, parents' years of schooling have a negative marginal impact on children's school attendance if a child works. This could be an artifact of the self-selection mecha- nism: educated parents may send their children to work only if they are desperate, whereas less educated parents may be more willing to make a marginal tradeoff TABLE Marginal Effects of Father's and Mother's Education 3. Sons Daughters Difference dyl Standard Standard dx(S)- Standard dyldx Error dyldx Error dyldx(D) Errora On child Labor I f child does not attend school Father's -0.0012 0.0001 -0.0003 0.0001 years of schooling Mother's -0.0009 years of schooling I f child does attend school Father's -0.0065 0.0006 -0.0019 0.0004 years of schooling Mother's -0.0026 years of schooling On school attendance I f child does not work Father's 0.0120 0.0007 0.0047 0.0006 years of schooling Mother's 0.0082 years of schooling I f child does work Father's - 0.0043 0.0004 - 0.0025 0.0004 years of schooling Mother's -0.0047 years of schooling "Assuming zero covariance. Source: Authors' analysis based on data from the 1998 Brazilian National Household Survey. between some school time and some additional income. Surprisingly, the marginal impacts of parents' education on school attendance if a child does work is nega- tive. This is unexpected and may be due to a selection process in more educated households in which part-time work to supplement family income is uncommon, and only those who have given up on school go to work. Thus, the initial results suggest that a father's years of schooling have a greater impact on both a son's labor status and his school attendance than on a daughter's. The mother has a very small, yet significantly greater mitigating impact on a daughter's labor status if she does not attend school. Emerson and Souza 31 1 To better understand the results, consider the following example. Imagine a son and a daughter in a household where both parents have four years of education (equivalent to the completion of first primary in Brazil). Now con- sider the impact on these children of having parents with eight years of edu- cation instead (equivalent to the completion of second primary in Brazil). The impact of adding four years of schooling for both the father and the mother can be estimated by using the estimated marginal effects from table 3 (which are evaluated at the means of the independent variables). If both the son and the daughter were attending school, the son would be 2.6 percent less likely to work because of the increase in the father's education, but only 1 percent less likely because of the increase in the mother's edu- cation.' The daughter would be 0.8 percent less likely to work because of the increase in the father's education, and 0.9 percent less likely because of the increase in the mother's education. If both the son and daughter were not working, the son would be 4.8 percent more likely to attend school because of the increase in the father's education, but 3.2 percent more likely because of the increase in the mother's education. The daughter would be 1.9 percent more likely to attend school because of the increase in the father's education, but 3.2 percent more likely because of the increase in the mother's education. To check the robustness of these results three additional regressions were performed: one with an additional variable of family income minus the observed child's income, one with the sample size reduced to households with only one son and one daughter, and one with the sample size reduced only to households with mothers at least 40 years old and with exactly three children. The chi-square tests of the equivalence of the estimated coeffi- cients on fathers' and mothers' years of education for all three robustness checks are presented in table 4. The difference in the estimated marginal effects of all three additional regression estimations are presented in table 5. The full results of all three estimations are presented in the supplemental appendix. The first robustness check repeats the estimates presented in tables 2 and 3, but with the additional variable of the family's income net of the observed child's income. Including this variable can be problematic as there is the strong possibility that the variable is endogenously determined with the school and work decisions of the children. Additionally, parental education is generally a very good proxy for family income and wealth, and so correlation between the two regressors is likely to be high. Nonetheless, if the results remain unchanged even with the inclusion of this variable, it enhances the robustness of the results. In this set of regressions, the same pattern emerges as in tables 2 and 3: the father's education has a greater mitigating effect on 9. Based on the estimated 0.65 percent per year decline in a son's probability of working and the estimated 0.26 percent decline in a daughter's probability of working. TABLE Chi-square Tests from Robustness Checks 4. Sons Daughters Chi- P>Chi- Chi- PBChi- Square Square Square Square - - Child labor3 Father's years of schooling = 13.920 0.000 0.033 0.567 Mother's years of schooling School attendance" Father's years of schooling = 0.030 0.855 16.200 0.000 Mother's years of schooling Child laborh Father's ycars of schooling = 4.480 0.034 0.010 0.933 Mother's years of schooling School attendanceb Father's years of schooling = 3.560 0.059 17.430 0.000 Mother's years of schooling Child laborc Father's years of schooling = 5.060 0.025 0.170 0.680 Mother's years of schooling School attendanceC Father's years of schooling = 0.330 0.567 0.640 0.423 Mother's years of schooling 3Bivariate probiton on child Labor and school attendance with family income minus child's earnings. b~ivariateprobiton child Labor and school attendance with at least one son and one daughter ages 7- 16. 'Bivariate probiton child labor and school attendance with three-child families and mothers over age 40. Note: White-Huber heteroskedastic consistent errors uscd in regressions. Source: Authors' analysis based on data from the 1998 Brazilian National Household Survey. the son's child labor status than does the mother's education, and the mother's education has a greater impact on the daughter's school attendance than does the father's education. The father's education has a greater mitigat- ing impact on the son's child labor and school attendance than on the daugh- ter's. And the mother's education has a slightly greater mitigating impact on the daughter's child labor than does the father's education if the daughter does not attend school. The second robustness check is to ensure that the results are not being biased by the comparison of some households with children of the same gender with households with mixed-gender children. Estimations of the models presented in tables 2 and 3 were performed again on a sample of children from households with at least one son and one daughter between the ages of 7 and 16. While this restriction more than halves the sample size, none of the quali- tative results change except that the father's education no longer has a differen- tial impact on school attendance if the child works. Emerson and Souza 3 13 TABLE. Marginal Effects from Robustness Checks 5 Only Families with one Only Families with With Family Income Son and One Daughter Mothers Ages 40+ and Minus Child's Income 7-16 Three Children Standard Standard Standard Differencea ~ r r o r " Differencea ~ r r o r ~ Differencea ~ r r o r ~ I f child does not attend schoolc Father's 0.0001 -0.0010 0.0002 years of schooling Mother's years of schooling I f child does attend school' Father's 0.0007 -0.0046 0.0013 vears of schooling Mother's years of schooling I f child does not workd Father's 0.0010 0.0074 0.0016 vears of schooling Mother's vears of schooling I f child does u!orkd Father's 0.0006 -0.0017 0.0010 years of schooling Mother's years of schooling "Difference. b~ssumingzero covariance. 'Marginal effects of father's and mother's education on child labor. dblarginal effects of father's and mother's education on school attendance. Source: Authors' analysis based on data from the 1998 Brazilian National Household Survey. The final robustness check seeks to eliminate a potential source of bias that may arise from the potential endogeneity of the family size variable. To deal with this problem, the original sample was restricted to families with a mother over age 40 (thusthose who have most likely completed their fertility decisions) and with three children (those who chose to have three children plus some random error). Doing so severely reduces the sample to about 10 percent of its original size, restricting the ability to measure the point estimates with 314 THE W O R L D B A N K E C O N O M I C R E V I E W precision. The qualitative results remain for the child labor regressions but disappear for the school attendance regressions. Together, these results suggest that, as found in previous studies of child health (Thomas 1994), for example, there is intrahousehold gender bias in the allocation of resources in the child labor context, with mothers favoring daugh- ters and fathers favoring sons. In intrahousehold models this can arise because of idiosyncratic technologies that convert education into human capital, or differences in parental preferences related to a child's gender. IV. C O N C L U S I O N This investigation of the impact of parental education and the impact of intra- household gender differences on child labor and school attendance in Brazil finds that higher parental education increases the probability that a child will attend school and decreases the likelihood of a child entering the labor market. These impacts differ across fathers and mothers and across sons and daughters. As education and income are highly correlated, the results suggest that parents may direct resources selectively, with fathers investing more heavily in sons and mothers in daughters. The results have potentially important policy implications. If, as this study finds, gender matters for the intrahousehold allocation of resources, then the gender of the adult recipient of transfers and the gender of the intended target may both matter. For example, if the policy goal is to reduce child labor, and child labor is predominantly a male activity in an economy, then the findings suggest that transfers allocated to the father may be more effective in reducing overall child labor than transfers to the mother. If the goal is to increase female school attendance rates, it may be more effective to allocate transfers to the mother than to the father. From a program effectiveness standpoint, it may be important to determine which parent receives the money, and how much they receive, based on the composition of the family and on the goals of the program. A P P E N D I X TABLE 1 Unweighted Basic Statistics A- Number of Standard Variable Observations Mean Error Minimum Maximum Sons Child's age 26,930 11.517 2.851 7 16 Nonwhite child 26.927 0.519 0.500 0 1 School 26,930 0.927 0.259 0 1 Work 26,930 0.176 0.381 0 1 (Continued) Emerson and Souza 315 TABLE Continued A-1 Number of Standard Variable Observations Mean Error Minimum Maximum Sons Rural Metropolitan area Number of children Age of father Age of mother Father's years of schooling Mother's years of schooling Family income minus child's earnings Daughters Child's age Nonwhite child School Work Rural Metropolitan area Number of children Age of father Age of mother Father's years of schooling Mother's years of schooling Family income minus child's earnings Source: Authors' analysis based on data from the 1998 Brazilian National Household Survey. Baland, Jean-Marie, and James A. Robinson. 2000. "Is Child Labor Inefficient?" journal of Political Economy 108(4):663-79. Basu, Kaushik. 1999. "Child Labor: Cause, Consequence, and Cure." Journal of Economic Literature 37(3):1083-119. Basu, Kaushik. 2006. "Gender and Say: A Model of Household Behavior with Endogenously- Determined Balance of Power." The Economic journal 116(511):558-80. Basu, Kaushik, and Hoang Van Pham. 1998. "The Economics of Child Labor." American Economic Review 88(3):412-27. Becker, Gary. 1982. Treatise on the Family. Cambridge, Mass.: Harvard University Press. Behrman, Jere. 1988. "Intra-E~ouseholdAllocation of Nutrients in Rural India: Are Boys Favored? Do Parents Exhibit Inequality Aversion?" Oxford Economic Papers 40(1):32-54. Bell, Clive, and Hans Gersbach. 1997. "Intra-Household Distribution and the Family." In M.K. Rosenzweig, and and 0.Stark eds., Handbook of Population and Family Economics, vol. 1A. Amsterdam: North-Holland. Bell, Clive. and Hans Gersbach. 2000. "Child Labor and the Education of a Society." IZA Discussion Paper 338. Institute for the Study of Labor, Bonn. Browning, Martin, and Chiappori Pierre-Andr. 1998. "Efficient Intra-Household Allocations: A General Characterization and Empirical Tests." Econometrica 66(6):1241-78. Chiappori, Pierre-Andr. 1988. "Rational Household Labor Supply." Econometrica 56(1):63-89. Chiappori, Pierre-Andr. 1992. "Collective Labor Supply and Welfare." ]ournal of Political Economy 100(3):437-67. Dessy, Sylvain. 2000. "A Defense of Compulsory Measures against Child Labor." ]ournu1 of Development Economics 62(1):261-75. Emerson, Patrick, and Andr Portela Souza. 2003. "Is There a Child Labor Trap? Inter-Generational Pcrsistence of Child Labor in Brazil." Economic Development and Citltural Change 51(2):375-98. Glomm, Gerhard. 1997. "Parental Choice of Human Capital Investment." lournal of Development Econontics 53(1):99- 114. Grootaert, Christiaan, and Harry Anthony Patrinos eds. 1999. Policy Analysis of Child Labor: A Comparative Study. New York: St. Martin's Press. Horowitz, Andrew, and Jian Wang. 2004. "Favorite Son? Specialized Child Laborers and Students in Poor LDC Households." journal of Dez~elopmentEconomics 73(2):631-42. Jensen, Peter, and Helena Skyt Nielsen. 1997. "Child Labour or School Attendance? Evidence from Zambia." ]ournal of Population Econontics 10(4):407-24. Lundberg, Shelly, and Robert A. Pollak. 1993. "Separate Spheres Bargaining and the Marriage Market." Journal of Political Econonly 101(6):988-1010. McElroy, Marjorie B. 1990. "The Empirical Content of Nash-Bargained Household Behavior." Journal of Huntan Resources 25(4):559-83. McElroy, Marjorie B., and Mary Jean Horney. 1991. "Nash-Bargained Household Decisions: Toward a Generalization of the Theory of Demand." International Econoniic Review 22(2):333-49. Ray, Ranjan. 2000. "Analysis of Child Labour in Peru and Pakistan: A Comparative Study." lournal of Population Econontics 13(1):3- 19. Ridao-Cano, Cristobal. 2000. "Child Labor and Schooling in a Low Income Rural Economy." University of Colorado, Boulder, Colo. Sen, Amartya. 1984. "Family and Food: Sex Bias in Poverty." In A. Sen ed., Resources, Value and Developntent. London: Blackwell. Sen, Amartya. 1990. "More than 100 Million Women Are Missing." New York Review of Books 37 (20):61-66. Strauss, John, and Duncan Thomas. 1995. "Human Resources: Empirical Modeling of Household and Family Decisions." In J.R. Behrman, and T.N. Srinivasan eds., Handbook of Development Economics, vol. 3. Amsterdam: North Holland. Tiefenthaler, Jill. 1999. "The Sectoral Labor Supply of Married Couples in Brazil: Testing the Unitary Model of Household Behavior." journal of Population Economics 12(4):591-606. Thomas, Duncan. 1990. "Intra-Household Resource Allocation: An Inferential Approach." lournal of Human Resources 25(4):635-64. Thomas, Duncan. 1994. "Like Father, Likc Son; Like Mother, Like Daughter: Parental Resources and Child Height." ]ournu1 of Hunian Resources 29(4):950-89. Tracking Poverty Over Time in the Absence of Comparable Consumption Data David Stifel and Luc Christiaensen Following the endorsement by the international community of the Millennium Development Goals, there has been an increasing demand for practical methods for steadily tracking poverty. An economically intuitive and inexpensive methodo- logy is explored for doing so in the absence of regular, comparable data o n household consumption. The minimum data requirements for this methodology are the availability of a household budget survey and a series of surveys with a com- parable set of asset data also contained in the budget survey. This method is illustrated using a series of Demographic and Health Surveys for Kenya. JEL codes: C81, I32 The worldwide endorsement of the Millennium Development Goals and the shift to results-based lending in supporting developing countries have intensi- fied the importance of being able to reliably gauge the evolution of poverty. The common approach to measuring poverty is anchored in utility theory and is empirically based on household consumption or income measures, which are usually derived from nationally representative household budget surveys (Ravallion 1996a; Deaton 2003). Obtaining reliable measures of household consumption presents a series of challenges in practice.' These challenges increase when comparing poverty over time. David Stifel is an assistant professor at Lafayette College; his email address is stifeld@lafayette.edu. Luc Christiaensen (corresponding author) is a senior economist in the East .4sia Rural Development and Environment Unit at the World Bank; his email address is Ichristiaensen@worldbank.org. The authors thank the participants at the Kenya Poverty Analysis and Data Initiative workshop in Nairobi and three ar~unymousreferees for their comments, and Christine Cornelius. Peter Lanjouw, Johan Mistiaen, and David Newhouse for providing feedback and facilitating access to data and software. The authors are also indebted to ORC Macro for supplying the Demographic and Health Survey data, particularly Bridgette James for her assistance and prompt responses to queries. 1. Among the challenges are determining the optimal recall period, valuing home consumption, and deciding how to treat consumption of housing, education, and health services, as well as appropriately accounting for the consumption of public gonds. Deaton and Zaidi (2002)providc cxcellent guidelines on how to meet such challenges. THE WORLI) BANK ECONOMIC REVIEW,VOL.21, NO. 2, pp. 317-341 doi:l0.1093/wberilhm010 Advance Access Publication 19 June 2007 $<;The Author 2007. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WOHLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org First, nationally representative household budget surveys are often una- vailable at regular intervals. Second, when available, they are frequently not comparable in design, and appropriate price deflators are usually difficult to come by. That changes in questionnaire design may systematically affect the resulting household consumption and welfare measures has been well documented (Scott and Amenuvegbe 1990; Appleton 1996; Pradhan 2001) and is most vividly illustrated by the "great Indian poverty debate".2 One way to circumvent the absence of regular household budget surveys is to link the annual series of national accounts to existing consumption surveys (Hoogeveen and Demombynes 2004). While the method is straight- forward, the predicted evolution of poverty holds only under a series of stringent assumptions such as distribution-neutral growth, a correct attribu- tion of sectoral GDP growth to households (World Bank 2005), and a close correspondence between growth observed in the national accounts and income or consumption growth measured in household surveys (Ravallion 2003; Deaton and Kozel 2005). Using household data instead, Ravallion (1996b) explores the use of subjective indicators of poverty and Sahn and Stifel (2000) propose tracking poverty by tracking household assets, which they combine into a single index based on statistical association. Building on insights from the Indian poverty debate (Deaton and Drkze 2002; Kijima and Lanjouw 2003; Sen and Himanshu 2004) and developments in the literature on the dynamics of poverty (Simler, Harrower, and Massingarela 2004; Azzarri and others 2006), this article explores an "econ- omic" asset index approach that is anchored in consumption and uses advanced prediction techniques akin to those applied in the poverty mapping literature (Elbers, Lanjouw, and Lanjouw 2003). By linking the different assets directly to consumption, the article provides a theoretical welfarist foundation for aggregating assets, an important enrichment over the "statistical" asset index approach of Sahn and Stifel (2000). Doing so also facilitates estimation of different poverty and inequality measures. Further, in contrast to Kijima and Lanjouw (2003), Simler, Harrower, and Massingarela (2004), and Azzarri and others (2006), this analysis excludes assets with returns that are more prone to change over time (education, labor, and land) and includes location-specific factors (such as rainfall, prices, and malaria incidence) that vary annually and are likely to affect returns to assets. The inclusion of more time-variant variables is an important innovation-it mitigates changes in returns to other assets and improves the capture of (transitory) changes in welfare and poverty. Subjective indicators of welfare could also be included, though they have been found to add little to 2. For a review of efforts to correct poverty estimates for India following a change in the recall periods for expenditure items in the 55th round of the National Sample Survey, see Deaton and Kozel (2005). Stifel and Christiaensen 319 the method's performance and have displayed limited correspondence with changes in poverty based on consumption measures in other settings (Ravallion 1996b; Azzarri and others 2006). They were also not available in the surveys used here. The economic asset index approach is applied to a series of standardized Demographic and Health Surveys for Kenya. The high degree of standardiz- ation in survey and questionnaire design across rounds means that comparabil- ity issues are minor. These surveys are also freely available for many sub-Saharan African countries. Yet the proposed approach can be applied to any carefully selected set of assets or poverty predictors that are regularly col- lected, including those under the Core Welfare Indicator Questionnaire surveys developed by the World Bank. The empirical application uses the asset information from the 1993, 1998, and 2003 Kenyan Demographic and Health Surveys and the con- sumption measure from the 1997 Welfare Monitoring Survey (WMS). Estimates derived from these indicate a continuous decline in poverty between 1993 and 2003 in rural Kenya and stagnation in poverty in urban Kenya. Trends were diverging in Nairobi (poverty declining) and other urban areas (poverty increasing), though these trends were not individually statistically significant. The direction of these findings is broadly consistent with those from the statistical asset index, the national accounts, other rural surveys as well as the initial poverty estimates from the national 2005106 Kenya Integrated Household Budget Survey that became available as this article was going to press. They are also in line with the observed evolution of key nonmonetary indicators such as school enrollment and child malnutrition during this period. Let W(c,) denote the value at time t of a population welfare measure (for example, poverty or inequality) that depends on individual consumption levels c at time t. Given comparable observations on c, at different time intervals, the evolution of W can be tracked. Without such observations, but with compar- able observations for the individual, household, and location assets x, that underpin c, = c,(x,), the evolution of c, (and thus W,) can be tracked, provided that one has an empirical understanding of the mapping of x, into c,.~ Tracking W, by tracking x, requires essentially three steps: developing an accu- rate empirical model of c, as a function of x,; estimating c,+k as a function of xt+k, where k is a positive or negative integer; and generating an estimate of expected Wt+kfrom the estimated ct+k. 3. The components of x, are not restricted to contemporaneous assets. For example, current consumption is likely to depend on lagged rainfall given the lag in agricultural production. 'palsal aq 1squnj ue3 qJ!qM jO . ~ J U E ~ ~ !1~3!1S!IeIS aql 'dl~aaodu! a8ueq3 aq] jo a]em!]sa ]uals!suos e sp[a!d s!qJ U ~ ! S . ~ I J J A O ~ jo ~ ~ U E ! J E A a q ~put ueaur a q ~q ~ o qjo saleu!lsa 1ua1s!suo3 su!e~qos ~ s qpssodo~dq x o ~ d d eaqJ .!io!~d e J E J ~IOU s! X I J ~ A Ou!~ a8ueq3 pa~~ur!~saa q f~o s~!qjo lua~xapue uo!13a~!p a q .~d ~ ~ a ~u!oa%ueqs d pue p a [ paleur!~saaql q ~ o qu! se!q e pue 'Y+'J ul jo uo!lnq!ns!p aql jo asue!len aql jo aleur!lsalapun ue u! %u!l[nsal sf':, UI 13!pa~d 01 'dq+;x ~ I U Oasn ( 9 0 0 ~S)J ~ ~ I O p u ~!JJEZZV ' I ~ E J I U O ~u1 . S ~ ~ a u ualqer~duro3e u! ~ u r aasse awes a q ~uo uopeur~oju!pa13allo3 I E ~ Ilaad ~ a q ~ o u!cd a ~ ~ leuo!leu ~ a q ~ o 01edah~nsla8pnq u n s u p1oqasnorl E n ~ o qam!] 1aAo ( l a ~ aa1~8a~B8eue IE) d ~ ~ a s13!pa~d q x o ~ d d exapu! lasse 3!urouosa u v l ~ o d .snsuaJ (asop 10) lead-am~se 01 L a ~ ~ 1a2pnq ploqasnoq e u r o ~Bu!ddeur dq d l ~ a ~ ojod uo!lnq!l]s!p n s j pa1e8a188es1p i(11e31qde~8oa8E s]>!pald %u!dd~urd l ~ a ~ oIdE ~ Iu! q3eoldd~xapu! 1ass.e >!mouosa aql uroq s~ajj!p Bu!dd~m L I J ~ A O.d~!l~nbau!pue dl~aaodjo suo!~gpa~dea~e-][ems~ o sanb!uq>a~ ~ j s!Jlamouosa padola~apo q '(~007)~ n o ! u epue ' ~ n o ! u e'siaql3 s ~ o l l odla81el qseo~ddeaqL ~ ~ ~ j '+, 'Y+ld '4+lx)m]3 's! ley] W+lnjo SasnlEaj ]euo!lnq!ns!p ayi ySno~ylpaurnssE s! ?+I9 jo amleu 2!~seysols aqL .4+19 1en13e ayl lou ' u ~ o u qs! ytJ9 2pseqs -01s aql X[uo ley1 13ej aql uroq s ~ o ~~rrauoduro2AonAa 9yv~9u~so!p?ayL ~ o j .Jona Su!ldur~s01 ~sa!qnsosle aJe 4 + l jo saleurpsa ' ( ~ 0 0M~~)O ~ ! U E ~ pue ' ~ n o ! u e'ssaqly u! passnss!p ssoua leuo!~e~nduro~pue 'lapour 's!~essuliso!p! ~ ayl 01 uoptppe u1 .pays!nSu!s!p ase lona jo sasmos snod .4+1/k1 Su!leur -!pa U! paaloau! ssona ayl az!ur!u!ur 01 luel~odur!s! I! ' Y + ~ I an11 ayl uo suoyaa -sasqo jo asuasqe ay1 u! ?+Jmjo saleur!~sa1ua1s!suo2 pue as~sasdSu~nssndu1 *xrpuaddeaq, u! pwap UI pau!l~noS! asnpasold ayL -A sMesp luasajj!p 103 aaoqe paqrJssap ssaso~daql Su!~e~nur!sliq pau!e]qo s! y + l ~jo anlea palsadxa ayl jo a~eu~!~sau v .4+13 Su!sn pale1ns1es s! 4+% jo ajeurpsa ue 'dl]eu!d -xapu! lasse s!mouo2a ue jo sa~eur!~sa se uaas aq o s p ues jo saleur!lsa 'Y+ln pue JdySnosq~uo!~durnsuo2u! Su! -soysue s!ayl pue q+lx 30 sanlea ayl uo lipaeay spuadap 4+930 anIeA ayl asu!~ pue y+ln jo suo!lnq!lls!p palaur!lsa ayl uroq Merp auo salouap A alaqM S :spla!X 'Y+Jx ~ l e lasse palepdn ay) y ] ! ~Suole by+Jdp u ~ p :,i2n30 uo!~eu!quro~a y l .]uelsuos aq 01 pamnsse osle s! ssasosd S u ! l e & u a ~ e ~ayl jo alnleu sysepaysolalay e ~ aql Su!u!unalap d!ysuo!lela~ ayl 'ytJn aleur!lsa 01 y+lx y l ! ~palepdn s! Jn 30 uo!~nq!s~s!payl ysnoqlle ' ~ a q l ~ n.awes aq1 a n Id pue ?+ldjo suopnqpls!p d aql 's! ~ey~-arupIaao ~ u e ~ s u ou!euras Id jo suopnq!Jls!p ayl s uopdurnssa ayl sasodur! X8olopoqlaru ay, '0s 8u!op UI '(1) uor~enba8upeurpsa u! pau!elqo Id pue In jo suo!lnq!sls!p paleur!lsa ay] uroq u ~ e s pY+ld pue yiZn jo saleur!lsa Sqsn palelns(e2 ase Win + q+$dy+:x =ytJ9 ul jo sa]eur!lsa ' l x a ~ Stifel and Christiaertsert 321 u ~ + ~is) ]calculated rather than W(ctfk).The model error component arises from the fact that the parameters Pt+k are estimated as well as those describing the distribution of ut+k; that is, E[W(x,+b, b,, tit+k)]is calculated rather than E[W(x,,k, Pttk, As the expectation is often analytically intractable, it is approximated through simulation, thereby generating a computational error. Finally, the sampling error follows from imputing from a survey and not a census; that is, the xt+k terms are obtained from a survey. The size of the idiosyncratic error component depends critically on the size of the target population, with the size of the error declining the larger the target population to which the welfare measure is imputed and increasing the smaller the population. This feature is critical in constructing small-area welfare estimates as it determines "how low one can go" (Alderman and others 2002). The interest in this article is in tracking welfare and poverty measures over time for major groups or areas for which representative data have been collected. Since these populations (rural, urban, province) are usually large, the idiosyncratic error component tends to be small. This idiosyncratic error component further depends on the explanatory power of the x variables in the model.6 Careful selection of the different sub- groups for which the consumption model (equation 1) is estimated and inclusion of key deterministic and stochastic location-specific variables (for example, rainfall variability and rainfall levels) are important to reduce idiosyn- cratic error in the welfare estimate. The magnitude of the model error component is in general determined by the precision of the coefficient estimates, the sensitivity of the welfare indicator to errors in the estimated consumption measures, and the extent to which the levels of the x variables in the target population deviate from the population of origin (Elbers, Lanjouw, and Lanjouw 2002). A second general source of model error derives from the assumption that the estimated distributions of bt and the parameters used to estimate iit+k are ~ t a t i o n a r While it is difficult to theoretically examine the magnitude of ~ . ~ the error introduced this way, theory and past empirical research can provide guidance to mitigate errors due to nonstationarity in specifying the consump- tion model. First, not all parameters are equally prone to change over time. For instance, while returns to labor and especially to education may change with changes in market conditions (Juhn, Murphy, and Pierce 1993; Ferreira and Paes de Barros 1999; Alwang, Mills, and Taruvinga 2002), estimated associations between consumer durables and other less frequently purchased 6. While consumption is clearly measured with error in practice, error-free consumption measures are assumed. Scc Chesher and Schluter (2002)for rules to approximate the effect of measurement error in estimating welfare measures. 7. The assumption of stationary empirical distributions is implicit in practice in many poverty-mapping exercises as household budget survey data and census data are typically collected in adjacent years rather than the same year. items are arguably more stable. The Engel relationship between these items and full consumption is likely to hold (Deaton and Kozel 2005), especially when consumption is predicted only for limited periods into in the future or the past. Second, risks of nonstationarity can be mitigated through the explicit inclusion of the sources of nonstationarity, such as prices and rainfall. Controlling for rainfall patterns is especially important in agriculture-based economies. Further partial corrections can be introduced by updating the x, variables with x,+k in the estimated means and variance-covariance matrices for tit+k. In sum, while potential model error from nonstationarity cannot theoretically be eliminated, careful selection of assets can go a long way in sub- stantially mitigating the risks of such errors. An empirical range of their magni- tude could be established if there were a series of consumption surveys with comparable consumption and asset data.8 A third source of model error arises from differences in the asset variables across the surveys arising from (small) differences in definition or ranking of questions in the questionnaire. To mitigate this potential, selection of the common asset variables is based on careful empirical comparison of the distri- butional characteristics of the x variables. Elbers, Lanjouw, and Lanjouw (2003) find that the computational error depends on the computational method. It is small when a sufficient number of simulations are used. The sampling error depends on the sampling design, the sample size, and the population variance of the consumption measure. The application of this methodology to Kenya focuses on tracking poverty between 1993 and 2003. Three major sources of data are used: the 1997 Welfare Monitoring Survey and district-level malaria data constructed from the 1992-97 WMS;~the 1993, 1998, and 2003 Demographic and Health Surveys; and district-level data on infrastructure from the 1999 census, and district-level rainfall data obtained from the Famine Early Warning no stem." 8. Sen and Himanshu (2004) show that it is sometimes possible to test the stability of the coefficients even in the absence of additional consumption data, though their test was not applicable here. 9. Only the third of a series of WMS surveys between 1992 and 1997 is used for the poverty analysis because differences in the timing of the survey and the questionnaire design rendered the reported poverty numbers noncomparable (World Bank 2003). A new national expenditure survey was fielded in 2005, but the data were unavailable at the completion of this study. 10. Rainfall data were available for 21 of the 36 districts in the analysis. These data were further used to impute rainfall patterns to the remaining 15 districts based on their geographic proximity. Stifel and Christiaensen 323 The 1997 WMS is a national survey containing information on house- hold consumption; household demographics; and individual, household, and community assets. The survey was conducted between February and May 1997 and covered 10,874 households." The consumption measure derived from the data is a geographically deflated measure of aggregated household expenditures including consumption of own production, as revised by the World Bank (2003). The World Bank (2003) estimated the rural poverty headcount at 52.8 percent and urban poverty at 43.1 percent. The 1997 WMS together with the secondary data from the census, the malaria data from 1992-97 WMS, and the Famine Early Warning System are used to estimate the distributional parameters in equation (1). Three Demographic and Health Surveys of about 8,000 households each were carried out in Kenya at five-year intervals between 1993 and 2003.12 As these surveys are not designed for economic analysis, there are generally no data on income or expenditures. However, several of the asset variables col- lected under the 1997 WMS are also tracked in the Demographic and Health Surveys. Furthermore, Demographic and Health Surveys are known for their comparability over time (and across countries). Survey instruments remain largely unchanged, and consistent sampling designs are maintained. Although the samples were intended to be nationally representative, the seven districts not covered in the 1993 and 1998 samples are excluded from the analysis.l3 To construct the economic asset index, a subset of assets ( x ~ +is~selected ) from the larger set of assets that are commonly available in the two surveys. This is the "zero stage" in the poverty mapping literature. Three criteria are used to select this subset, all motivated by an effort to optimize the trade-off between maximizing explanatory power (minimizing idiosyncratic error) and minimizing model error. First, to reduce potential error due to parameter instability, the set of assets is restricted to those for which parameters are likely to remain stable over time. Assets such as labor and education, which are more prone to parameter instability following economic or polity change, are therefore ex~luded.'~Only consumer durables and housing characteristics as well as more time-variant rainfall and individual health variables are included. The first two sets of 11. For reasons of logistics, insecurity, and inaccessibility, Isiolo, Mandera, Samburu, and Turkana districts were not covered. As such, the sample is not entirely representative at the national level, though consistency is maintained in the comparisons over time and across data sets by excluding these districts from all of the data and analysis. 12. The 1989 survey contained very limited household information. 13. These are Garissa, Mandera, and Wajir in North Eastern Province; Samburu and Turkana in Rift Valley Province; and Isiolo and Marsabit in Eastern Province. 14. Deaton and Kozel (2005) consider the inclusion of household size, education of household members, and household land holdings to be the main weakness of Kijima and Lanjouw's (2003) poverty predictions for India, which use a similar technique. variables capture the more nearly permanent part of household consumption, while the second two capture transitory aspects.15 Second, to ensure that the structure of the model estimated in the WMS data is appropriate for the Demographic and Health Survey data, only assets that are similar in the WMS and the adjacent Demographic and Health Surveys are retained. In particular, only variables for which there is 95 percent confidence that their means in the 1997 WMS and the 1998 Demographic and Health Survey do not differ are kept. Assets such as televisions and bicycles, for which one-year changes are possible, are kept if the difference in the means from 1997 to 1998 is no greater than the difference in the means from 1993 to 1998. Third, step-wise regression models using the resulting common subpool of assets are applied to identify the set of variables that are statistically significant (at the 5 percent level) while maximizing the explanatory power (as captured by the R-squared statistic).16 Balancing the trade-off between maximizing explanatory power and minimizing model error also motivated the choice of different regression specifications for rural, other urban, and Nairobi house- holds because the differences in their livelihood systems suggest a different relationship between a household's consumption and its asset base. The assets that are retained based on these three criteria include housing quality and sani- tation variables, consumer durables, child nutritional status, cluster and district averages of the household-level variables, district measures of malaria incidence averaged across the 1992, 1994, and 1997 WMSs, and district rainfall measures (table 1). The table also presents selected household demographic and education variables, which are used for sensitivity analysis in section IV. Since the food expenditure share is on average above 60 percent (1997 WMS), children's nutritional status is likely associated with household con- sumption and is a good proxy for transitory changes in household consump- tion.17 The average height-for-age z-score for children under five in the household is used instead of weight for age, because height is a longer term measure of nutrition that corresponds more closely to the dependent variable, 15. Consumer durables and housing characteristics may display downward rigidity and as a result may not appropriately capture potential declines in welfare resulting from moderate declines in household incomes. Yet households with durables and better housing conditions usually also have better access to credit, enabling them to better smooth their consumption levels in the face of such income shocks. Moreover, the flow of services derived from these goods persists, even when incomes temporarily decline, and should he reflected in the consumption predictions. Nonetheless, it is important to build sufficient flexibility into the model to help capture both the more permanent and the more transitory parts of household consumption. This is achieved through the addition of more time-variant explanatory variables in the model, such as rainfall shocks and health and nutrition variables, that genuinely fluctuate from year to year and likely affect income and welfare. 16. For all variables that were retained after applying the first two criteria, cluster and district averages were also computed and included in the stepwise regression models to better capture location characteristics. The cluster averages are estimated from the survey data, while the district averages are calculated from the 1999 household census. 17. The authors thank an anonymous referee for suggesting the inclusion of nutritional status as an asset. TABLE Levels (proportions) and Changes in Individual, Household, and Location Assets Held by Rural, Other Urban, 1. and Nairobi Households in Kenya During 1993-2003 Rural Other urban Nairob~ --- Changes Changes Changes W M S WMS - WMS -- 1993- 1998- 1993- 1993- 1998- 1993- 1993- 1998- 1993- Variable Housing Characteristics Dummy Vanahle: House Floor 0.79 0.02"' -0.02"' 0.05'"' of Low Quality (Mud, Dung, Sand) ( I = yes) Dummy Variable: House Roof 0.36 -0.06"* - 0.08'"' -O.l.F'*' of Low Quality (Thatch) ( I =yes) Dummy Variable: D r ~ n k ~ n g Warer, P~pedor Public Tap ( I = ycs) Dummy Variable: Flush Toilet 0.01 -0.004" 0.001 0.003 (I =yes) Household Durables 2 Dummy Variable: Owns a Radio 0.59 0.078-* 0.14'*' 0.24""' 0.77 0.04"' 0.02 0.10"" 3 ( I =yes) a 4 Dummy Variable: Owns a 0.04 0.02""" 0.07'"" I I ' 0.31 0.00 0.02 0.09**' Television (1= yes) 9 2. Dummy Variable: Owns a 0.01 0.00 0.01"" 0.01"' 0.13 -0.03' 0 . 0 4 " ' -0.06"' 0.18 0.00 0.07'"' 0.118*' g. Refr~geraror( I =yes) m Dummy Variable: Owns a Bike 0.29 0.02" 0.08'"" 0,11*:>* s II, (1 =yes) (Continued) /=: b TABLE Continued 1. W h, rn Rural Other urban Nairobi - -I .I. Changes Changes Changes m WMS WMS WMS r U 1993- 1998- 1993- 1993- 1998- 1993- 1993- 1998- 1993- P r 1997 1997 2003 2003 1997 1997 2003 2003 1997 1997 2003 2003 U w > Cluster and District Characteristics Zr: Cluster Average Household with 0.19 0.03"' 0.01 0.08"' m n Low Quality Floors 0 Cluster Average Household with 0.79 -0.04"' -0.46"' -0.53"' 0 7. Access to Piped Water 2 n Cluster Average Household 0.01 0.001' 0.005"' 0.005'" 0.11 -0.02" -0.04"' -0.05"' P Owns Refrigerator m 5 District Average Household with 0.06 0.002"' 0.001 -0,001 m Access to Electricity Rainfall and Health Rain, Early Rainy Season -0.21 0.22"' 0.02" 0.45"' -0.12 0.39"' -0.07"' 0.49"' Deviation from Long-run Average Malaria Prevalence in District 0.109 -0.001 -0.003"' -0,002"' 0.147 0.003 -0.001 0.003 (Averagein 1990s) Average Household -1.28 0.14"' 0.07"' 0.16"' Height-for-Age z-score Household Demographics Dependency Ratio 0.52 -0.04*" -0.02"' -0.04"' 0.41 -0.01 0.00 0.00 0.34 0.01 -0.02 -0.03" Household Size 6.42 -0.36"' -0.06 -0.76"' Education Share of Household Members 0.11 0.03"' 0.00 0.02"' with Secondary Education Share of Household Members 0.02 0.02"' 0.01"' 0.02"' with Post-secondary Education (Continued) TABLE Continued 1. Rural Other urban Nairobi Changes Changes Changes WMS WMS WMS - Dummy Variable: Household 0.33 0.01 0.09"' o . ~ o * * ' Head with Primary Education Dummy Variable: 0.20 0.05'" -0.03''' 0.03"' Household Head with Secondary Education Dummy Variable: Household Head with Post-secondary Education Cluster Average Share with Post-secondary Education Cluster Average Head with Primary Education Cluster Average Head with Secondary Education Cluster Average Head with 0.09 0.04"' 0.07"' 0.10"' 0.14 0.10"' 0.09"' 0.19"' 2 Post-Secondary $ $a Education L 3 n 'Significant at the 10 percent level; "'Significant at the 5 percent level; '""Significant at the 1percent level. WMS is Welfare Monitoring Survey. Note: Italicized variables are those to which negative weights are assigned. A decrease in the means of these assets is associated with a predicted $ increase in mean household expenditure. For variables that are not italicized, an increase in the mean is associated with a predicted increase in mean 2 household expenditure. Source: Authors' analysis based on data described in the text. W FIGURE Deviation of Annual Rainfall and Early Onset of Rainfall from 1. Long-Run Average in Kenya, 1992-2003 30 1 Note: Early onset of rain is defined as the percentage deviation in the amount of rainfall during the first month of the long rains from the long-run average (1992-2003) amount for that month. Selection of the first month differs by district and is based on the district-specific rainfall patterns in the Famine Early Warning System data. Source: Authors' analysis based on Famine Early Warning System data. annual consumption per adult equivalent, whereas weight can vary seasonally within a given year. After the stepwise regression is applied, the anthropo- metric variable was only retained in the rural model. From (figure I), 1992 emerges as a very dry year with rainfall starting late and the overall amount well below the long-run average; 1996 and 1997 were above average both in timing (early or on time) and overall level. Similarly, rainfall in 2002 was slightly above normal, though it was slightly later in coming.18 Given the large fluctuations in both level and timing of rainfall and their independent importance for welfare-rural households in Kenya are often unable to protect consumption from drought shocks (Christiaensen and Subbarao 2005)-it is important to account for actual rainfall patterns in tracking poverty over time. The timing of the onset of the long rains rather than the level of rainfall in that year is included as an asset in the predicting models. While they are correlated,19 timing yielded a better fit. 18. Given the lag In agricultural production, the relevant rainfall patterns are mostly those in the year preceding the survey year. 19. The correlation coefficient for these two measures of rainfall is 0.53. Stifel and Christiaensen 329 The coefficients (which can be viewed as asset weights) from the three differ- ent first-stage models (rural, other urban, Nairobi) are presented in table 2.20 - Since these are the results of stepwise regressions, each parameter estimate is significant at the 95 percent confidence level or higher. Although the signs are generally as expected, plausibility of the parameter estimates is not critical since consistency of the predicted dependent variable does not rely on consist- ent estimation of the parameters. For the rural sample of 8,807 households in the WMS, 13 variables are kept in the model. Twenty-one percent of the variation in log per adult equivalent expenditure is explained by the model. The stepwise regressions led to the retention of eight variables resulting in an R-squared of 0.25 for the other urban models and to three variables and an R-squared of 0.35 for the Nairobi models. In the Nairobi models only six explanatory variables passed the com- parability test between the WMS 1997 and the Demographic and Health Survey 1998, from which only three were kept in the stepwise regression. The simulated poverty rates for all four data setslyears are shown in table 3. The 2997 poverty rates constitute the baseline as these were estimated using the sole household budget survey (WMS). Due to a persistent 1-2 percentage point underestimation of the simulated poverty prevalence in the baseline data com- pared with the actual poverty prevalence directly observed in the baseline data using the official poverty line, the poverty line is allowed to be determined endogenously to replicate these 1997 poverty levels. This adjusted poverty line is then applied to all four data sets to maintain comparability. Adjustments were minor and did not affect the magnitudes of the predicted changes in poverty; only the levels were affected. As such, the simulated poverty headcount ratios in 1997, the base year, are consistent with those reported in World Bank (2003).~'They are 52.8 for rural areas, 43.2 for other urban areas, and 40.0 for Nairobi. Taken together, about half of the Kenyan population was estimated to be poor in 1997. The economic asset index suggests that poverty prevalence in Kenya fell from 55.8 percent in 1993 to 50.8 percent in 1997, and continued to fall to 45.0 percent in 2003. Although the fall is not statistically significant after 1997, it is for 1993-1997 and for 1993-2003. With most poor people residing in rural 20. Hausman tests, described in Deaton (1997), were used to determine whether sampling weights should be used in the final regression models. For all three models the weighted versions of the explanatory variables added to equation (I), were jointly significant at the 99 percent level of confidence. Consequently, sampling weights were used in all of the prediction models. 21. The urban and rural poverty lines constructed by the World Bank (2003) were based on a nonparametric approach to adding basic nonfood requirements to the food poverty lines. The food poverty lines were themselves based on the monetary value of a food basket that allowed minimum nutrient requirements (2,250 calories) to be met. Finally, they were adjusted for regional price differences. TABLE2. Estimated Coefficients or Asset Weights from First-Stage Regressions (DependentVariable = Log Consumption Per Adult Equivalent, 1997) Other Rural Urban Nairobi Housing Characteristics Dummy Variable: House Floor of Low Quality (Mud, Dung, Sand) (1= yes) Dummy Variable: House Roof of Low Quality (Thatch) (1= yes) Dummy Variable: Drinking Water, Piped or Public Tap (1= yes) Dummy Variable: Flush Toilet (1 = yes) Household Durables Dummy Variable: Owns a Radio (1= yes) Dummy Variable: Owns a Television (1 = yes) Dummy Variable: Owns a Refrigerator (1= yes) Dummy Variable: Owns a Bike (1= yes) Cluster and District Characteristics Cluster Average of Households with Low Quality Floors Cluster Average of Households With Access to Piped Water Cluster Average of Households Owning a Refrigerator District Average of Households With Access to Electricity Rainfall and Health Rain, Early Onset (Deviation From Long-Run Mean), District Level Rain, Early Onset, Squared Malaria Prevalence in District (Average in 1990s),District Level Average Household Height-for-Age z-Score Among Under Five Year Olds Constant Adjusted R~ Number of Clusters Number of Variables Number of Observations Source: Authors' analysis based on data described in the text. areas (80 percent in 2003), the evolution of rural poverty between 1993 and 2003 is very similar to that observed at the national level. Nairobi also experi- enced a decline in poverty as the prevalence dropped from 40.7 percent in 1993 to 35.1 percent in 2003. In contrast, poverty prevalence in other urban areas rose from 39.0 percent to 46.0. Yet, the simulated poverty changes in both Nairobi and other urban areas were not statistically significant. The (simulated)evolution of the more distribution-sensitive poverty measures (the poverty gap and poverty severity indices) across the rural, other urban, and Nairobi populations are broadly consistent with the picture emerging from the headcount figures. To provide insights into the factors behind the emerging pattern of the evolu- tion of poverty, the average evolution of assets across the different survey years is presented in table 1. The evolution in the asset base is broadly consistent with the observed evolution of poverty across the different groups. Caution should be Stifel and Christiaensen 331 TABLE Asset Poverty in Kenya, 1993-2003 3. Levels Test Statistics for Changes - -- 1997 1993 (WMS, 1998 2003 1993- 1997- 1998- 1993- (DHS) base) (DHS) (DHS) 1997 1998 2003 2003 Economic Asset Index Headcount Ratio (Po) National Rural Other Urban Nairobi Poverty Gap (PI) National Rural Other Urban Nairobi Poverty Severity 0'2) National Rural Other Urban Nairobi Statistical Asset Index Headcount Ratio (Po) National Rural Other Urban Nairobi 'Significant at the 10 percent level; '"Significant at the 5 percent level; ""the Significant at the 1 percent level. WMS is Welfare Monitoring Survey and DHS is Demographic and Health Survey. Note: Numbers in parentheses are standard errors. Source: Authors' analysis based on data described in the text. taken in interpreting the results for Nairobi, however, as household ownership of refrigerators appears to be driving the poverty predictions. This highlights the need to ensure comparability in designing base and target data sets. The substantial reduction in rural poverty between 1993 and 1997 com- pared with the reduction between 199711998 and 2003 is partly related to the underlying rainfall pattern (very bad in 1992, exceptionally good in 1997, and modest in 2002; see figure 1).Nonetheless, the improvements in rural welfare between 1993 and 2003 appear genuine and shared by the poorer segments of the population. To test this notion, poverty was predicted for 2003 using the 2003 Demographic and Health Survey data and the 1992193 Famine Early Warning System rainfall data. The resulting 50.1 percent headcount ratio suggests that the better rainfall accounted for only 22.8 (2.119.2) percent of the overall fall in rural poverty between 1993 and 2003. IV. ARET H E RESULTESM P I R I C A LR Y B U S T ? L O To gauge the reliability of the poverty trends emerging from the economic asset index, these trends are briefly compared with those based on other indicators, and the plausibility of the assumptions underpinning the economic asset-based poverty numbers in the Kenyan context is examined. First, the trends are com- pared with the picture emerging from the trends in the statistical asset index developed by Sahn and Stifel (2000),who apply factor analysis to a set of assets common to the three Demographic Health Since the weights applied to these assets are derived in a purely statistical manner, this index is considered a statistical asset index. The results of this sensitivity analysis appear at the bottom of table 3.2" The trends are similar to those derived from the economic asset index for rural and other urban areas despite different weighting schemes and a different asset bundle. While the direction of change for Nairobi is the same, the declines in poverty are markedly larger with the statistical asset index. The findings here are also broadly consistent with the evolution of per capita consumption observed in the national accounts. Per capita private con- sumption observed in the national accounts grew 2.8 percent a year between 1993 and 1998, consistent with the simulated reduction in national poverty incidence from 55.8 percent to 48.1 percent. While growth in per capita private consumption essentially stagnated thereafter,24 poverty continued to decline according to the simulations in this study, though the change was less 22. These assets include household characteristics (source of drinking water, toilet facilities, and house construction material) and household durables (ownership of radio, television, refrigerator, and bicycle). The 1997 W'MS is excluded to avoid dropping assets for lack of commonality, as was necessary for the economic asset index. 23. The poverty lines arc dcfincd in order to replicate the 1997 poverty rates in the 1998 Demographic and Health Survey. 24. Average annual growth in per capita private consumption between 1998 and 2003 was estimated at -0.1 percent in the national accounts. Stifel and Christiaensen 333 pronounced and no longer statistically significant. However, further decompo- sition of the national accounts between 1998 and 2003 shows that the stagna- tion in overall per capita GDP growth was driven by contraction of the industrial sector and stagnation of the services sector, while per capita agricul- tural GDP continued to grow, albeit at a slightly slower pace.25 This is consis- tent with the simulated decline in rural poverty and the increase in poverty in other urban areas (though not with the decline in poverty in Nairobi). Other surveys also suggest a continuing decline in rural poverty between 1998 and 2003. Nyoro, Muyanga, and Komo (2005)find, for example, that $1 a day poverty dropped between 1997 and 2004 among a panel of 1,500 predo- minantly maize-growing smallholders. Preliminary estimates from the national 2005 Kenya Integrated Household Budget Survey also suggest a decline in rural poverty, with the decline similar in magnitude to that predicted in the approach followed here. The new poverty estimates also point to an increase in poverty in other urban areas and a decrease in Nairobi in 2 0 0 . 5 . ~ ~ Finally, the poverty trends reflected by the economic asset-based poverty indices are broadly similar to the evolution of key nonmonetary indicators of well-being in Kenya, such as primary and secondary enrollment rates and stunt- ing prevalence. Comparison of these indicators across the population in rural areas, other urban localities, and Nairobi between 1993 and 2003 based on the Demographic and Health Surveys (table 4) shows substantial improvements in primary and secondary enrollment rates and stunting prevalence in rural areas, even stronger improvements in these indicators in Nairobi, and a mixed picture in other urban areas, with primary enrollment rates increasing, second- ary enrollment rates falling marginally, and stunting prevalence increasing by 6 percentage points.27 The plausibility of the assumptions underlying the economic asset index may also affect the empirical performance. To reduce the likelihood of model error following from nonstationarity of the estimated parameters, the predic- tion model included rainfall and nutritional status. Further, although there were no dramatic shifts in the economic and political regime during the periods considered here, such assets as household labor supply and educational attainment were excluded as a precaution.28 25. During 1993-98 per capita agricultural GDP grew at 0.9 percent, industrial GDP at -0.56 percent, and service GDP at 0.96 percent and during 1998-2003 they grew at 0.75 percent, -0.99 percent, and 0.14 percent, respectively. 26. These poverty estimates became available as this article was going to press. 27. The evolution of infant mortality, which increased by 9.4 children per 1,000 born in rural areas between 1993 and 2003, appears at odds with the estimated evolution in household welfare. However, in-depth mulrivariate analysis of the determinants of enrollment rates and health outcomes in Kenya using the 1993, 1998, and 2003 Demographic and Health Survey data (Stifel and Christiaensen 2006) shows that while household consutnption is positively associated with (primary) enrollment rates and nutritional status, there is no correlation between consumption and infant mortality. 28. Land and livestock, which are absent from the Demographic and Health Surveys, also fall in this category. 334 T H E W O R L D B A N K E C O N O M I C R E V I E W TABLE4. Nonmonetary Indicators of Well-Being in Kenya, 1993-2003 Deterioration (-) or Level Changes Improvement (+) 1993- 1998- 1993- 1993- 1998- 1993- 1993 1998 2003 1998 2003 2003 1998 2003 2003 National Enrollment Rates Primary (Ages 6-13) 75.6 85.5 90.1 9.8 4.6 14.5 + + + Secondary (Ages 76.8 75.1 77.4 + + - 1.7 2.3 0.6 - 14-17) Stunting Prevalence 33.3 33.0 30.9 -0.2 -2.1 -2.4 + + + Infant Mortality 73.8 78.6 82.4 4.8 3.8 8.6 - - - Rural Enrollment Rates Primary (Ages 6-13) 75.3 85.4 89.8 10.0 4.5 14.5 + + + Secondary (Ages 78.4 77.6 79.8 -0.7 2.1 1.4 - + + 14-17) Stunting Prevalence 34.8 34.7 32.5 -0.1 -2.3 -2.3 + + + Infant Mortality 75.8 81.1 85.2 5.2 4.2 9.4 - - - Other Urban Enrollment Rates Primary (Ages 6-13) 80.8 85.6 91.3 4.9 5.7 10.6 + + + Secondary (Ages 65.8 61.9 65.6 -3.9 3.7 -0.2 - + - 14- 17) Stunting Prevalence 20.7 24.1 26.7 3.4 2.6 6.0 - - - Infant Mortality 62.0 62.0 62.0 0.0 0.0 0.0 Ns Ns Ns Nairobi Enrollment Rates Primary (Ages 6-13) 74.1 87.3 92.9 13.2 5.7 18.9 + + + Secondary (Ages 54.8 56.1 62.9 1.3 6.7 8.1 + + + 14-17) Stunting Prevalence 22.5 25.7 18.5 3.2 -7.2 -4.0 + - + Infant Mortality 55.0 55.0 55.0 0.0 0.0 0.0 Ns Ns Ns Ns = Changes are not statistically significant. Source: Authors' analysis based on data described in the text. Table 5 presents economic asset index-based poverty estimates when these assets are included and compares them with the original predictions. The labor supply variables include household size and the dependency ratio, while the education variables include information about educational attainments at the household and cluster levels as well as for the household head (see table I ) .While ~ ~ the precision of the estimates (as reflected in the lower stan- dard errors) improved slightly (except for the Nairobi poverty estimates), general trends remained unchanged. The larger simulated decline in poverty for Nairobi counsels caution in the use of education variables to track changes in poverty, especially in more sophisticated urbanized settings. Not only is the wage gradient usually much steeper in such settings, it is also likely to be more sensitive to the performance 29. The first-stageparameter estimates are available on request from the authors. Stifel and Christiaensen 335 of the (formal)economy. It is likely that the returns to higher education declined in Nairobi in the face of the rapid expansion of the supply of highly educated professionals30 and the stagnation and contraction of the urban economy.3' That said, the massive investment by households in their education may well have enabled some to escape poverty in Nairobi, despite a decline in the rate of return. Finally, part of the simulated poverty reduction results from the substantial increase in ownership of durables (radios, televisions, and refrigerators). Further inspection indicates that their relative prices (in terms of the overall consumer price index) declined substantially, possibly because of technological innovation, trade liberalization, or exchange rate misalignment (in particular, real exchange rate overvaluation). If these durables miere perfect substitutes for other goods in the consumption basket of the poor, the economic asset index would substantially overestimate poverty reduction since the increase in demand for these durables would have been offset by a decrease in demand for other goods. It is unlikely, however, that the poor substitute electronics and household appliances for food at substantial rates. Consequently, the observed increase in the demand for these goods must result largely from an increase in people's incomes. Moreover, the estimated association between some consumer durables (for example, radios and televisions) and consumption reflects not only wealth levels, but also the flow of services derived from the possession of these goods (such as improved access to information, which may in turn improve the returns to other assets). There is no reason to believe that the utility derived from these goods would change drastically over time. Thus, any downward shift in the distribution of the estimated coefficients on durables between 1997 and 2003 is probably very small, and the simulated poverty reduction is likely only slightly overestimated, if at all. In light of these findings and in the absence of further empirical evidence by way of an empirical counterfactual, a strategy of limiting, on theoretical and empirical grounds, the choice of assets to those whose returns are unlikely to change over time (consumer durables, housing characteristics, rainfall, and health) and of excluding those whose returns are more prone to variation over time (land, labor, education) is appropriate. Although there are no a priori reasons to suspect that the stationarity assumption is substantially violated, the extent to which stationarity holds for each asset and the way violation of stationarity might affect predictions of welfare ultimately remain an empirical matter that can be tested only through another consumption survey. 30. The proportion of households whose head has some post-secondary education increased by 19 percentage points in Nairobi between 1993 and 2003. 31. Average annual per capita growth in the services sector, which accounts for about 55 pcrccnt of the Kenyan economy, stagnated at about 0.14 percent during 1998-2003, while average annual per capita GDP growth in the industrial sector, which accounts for about 20 percent of the Kenyan economy, was estimated at -0.99 percent. TABLE Economic Asset Index Poverty for Different Models 5. Headcount Ratio (Po) Test Statistics for Changes Change in Po 1997 1993 (WMS, 1998 2003 1993- 1997- 1998- 1993- (DHS) base) (DHS) (DHS) 1997 1998 2003 2003 1993-2003 Rural Base 57.2 52.8 50.6 48.0 -1.54 -0.81 -0.87 -2.87""* - 9.2 (2.2) (1.9) (1.9) (2.4) Base+ Demographics 58.4 52.8 49.9 45.9 -2.47"" - 1.32 - 1.44 -4.46"'" -12.5 (1.8) (1.4) (1.7) (2.1) Base $ Demographics +Education 58.7 52.8 50.2 47.9 -2.63""' -1.25 -0.86 -3.95"'" -10.8 (1.8) (1.4) (1.6) (2.1) Other urban Base 39.0 43.2 41.3 46.0 0.70 -0.34 0.73 1.03 +7 (4.5) (4.0) (4.0) (5.0) Base + Demographics 37.5 43.2 41.0 46.6 1.06 -0.41 0.95 1.53 +9.1 (4.0) (3.5) (4.0) (4.3) Base t Demographics +Education 39.5 43.2 41.8 47.2 0.70 -0.31 1.05 1.33 +7.7 (4.1) (3.3) (3.1) (4.0) Nairobi Base 40.7 40.0 38.5 35.1 -0.05 -0.11 -0.28 -0.44 -5.6 (9.5) (10.0) (8.8) (8.7) Base +Demographics 39.7 40.0 38.4 36.2 0.02 -0.11 -0.18 -0.26 -3.5 (10.0) (10.9) (8.9) (9.0) Base + Demographics +Education 44.4 40.0 38.1 30.1 -0.28 -0.13 -0.62 -0.96 -13.3 (11.9) (10.6) (9.0) (9.1) *Significantat the 10 percent level; +"Significant at the 5 percent level, and *+"S~gnificantat the 1 percent level. Note: Numbers in parentheses are standard errors. Source: Authors' analysis based on data described in the text. Stifel and Christiaensen 337 This article contributes to the growing literature on inexpensive and economi- cally intuitive methods for tracking poverty in the absence of comparable con- sumption data. The minimum data requirements are a household budget survey and a series of other surveys with a set of comparable data on assets. An application to Kenya using a series of Demographic and Health Surveys and secondary data provides poverty predictions that are broadly consistent with other indicators for Kenya during the period 1993-2003. Rural poverty declined, while urban poverty stagnated, with diverging trends in other urban areas (an increase) and Nairobi (a decrease), though the urban trends were not statistically significant. The economic asset index approach for tracking poverty proves promising, especially given the high costs involved in collecting comprehensive consump- tion data and the readily available PovMap software to conduct the predic- t i o n ~ It~can be easily extended to track mean consumption levels and . ~ inequality, an important additional advantage over the statistical asset index method. Furthermore, its empirical precision can be strengthened substantially through careful preselection of the tracked assets based on the theoretical and empirical plausibility of their "returns" being constant over time and on their predictive power using econometric analysis of existing household budget surveys. Inclusion of key time-variant variables such as rainfall, health status, and prices is critical both to control for shocks that may affect these returns and to better capture transitory changes in welfare and poverty since consumer durables and housing characteristics may display some downward rigidity. Nonetheless, despite the great care taken in asset selection to avoid violating the stationarity assumption, regular recalibration of the model is advisable, and predicting too far into the future or the past should be avoided. Going forward, comparing economic asset-based poverty measures with those derived from household budget surveys using actual consumption data emerges as an important research agenda for applied economists to shed further light on the empirical validity of the stationarity assumption. Individual consumption is approximated by household log per adult equivalent expenditures and estimated at the geographic levels for which the data are representative and for which comparisons over time are meaningful. Equation (1) is estimated for each region with the vector of disturbances u, distributed F(O,Z). To minimize model error in the predictions, efficient estimates of the P, para- meters are sought by exploring a heteroskedastic specification of the individual 32. T h e software can be found at http://iresearch.worldbank.org/PovMap/index.htm. disturbance terms and estimating equation (1)using generalized least squares (GLS)and an estimate of Z. In doing so it is assumed that u, (ucht= qct e,ht) is made up of a cluster- + or location-specific term (q,,) and a household-specific term (ech,),which are independent and uncorrelated with any of the observable characteristics, x,. This structure allows for both spatial autocorrelation (a "location effect" for households in the same cluster) and heteroskedasticity at the household As such Z is an N x N block-diagonal matrix. To estimate Z, equation (1) is initially estimated by ordinary least square yielding hChtin the form of the residuals from this regression. The location com- ponent is then estimated as the within-cluster mean of the overall residuals, where N, is the number of households in cluster c. The idiosyncratic household component estimate (&) is the overall residual less the location component, To allow for household-specific heteroskedasticity and to estimate &,,h,, &,is modeled using the zch, variables derived from x,~,,, their squares, and interactions that best explain its variation. The conditional variance is esti- mated using a logistic function, as outlined in Mistiaen and others (2002).The - estimated variance of icht (i&,) can then be obtained in a straightforward manner. o,,, the estimated variance of q,,, and its sample variance +(&$) are 2 estimated following Elbers, Lanjouw, and Lanjouw (2002). Armed with ~ 7 and~BzCht,and : thus an estimator for 2; ($), final efficient estimates of the betas in the original first-stage model (equation 1) can be obtained using GLS and the household budget survey data. This GLS esti- mation produces ptGLsand the variance-covariance matrix of this estimator, var(ptGLs),which concludes stage 1. To obtain estimates of the expected welfare indicator in stage 2, a vector of beta coefficients ( E )is first drawn from a multivariate normal distribution with a mean ptCLSand variance-covariance C(ptGLs)and applied to the target data xt+k to predict household log expenditures (xLhttkE).This highlights the importance of acquiring efficient estimates of the beta coefficients. Second, for each simulation the distribution of the location disturbance is allowed to vary. As such, the simulated location disturbance (@2t) is drawn from a distribution with zero mean and simulation-specific variance 33. Following Elbers, Lanjouw, and Lanjouw (2003), heteroskcdasticity is limited to the household-specific term since the number of clusters in the consumption survey is usually too small to allow for heteroskedasticity in the cluster component. Stifel and Christiaensen 339 (a:,)',itself drawn from a gamma distribution defined so as to have a mean of %, and a variance qi?; ,. Third, the simulated idiosyncratic component (ECht+k) is determined by first drawing an alpha coefficient (&:) from a normal distribution with mean &, and variance Q (&,). This is then applied to the data to determine the household - variance, 2 ~ , , , h , + k . ~ ~ Finally, is drawn from a distribution with mean zero and variance, &.&h,+k. Fourth, these three components are combined to simulate the value of household per adult equivalent expenditures, 2Lt+k = ~ X P ( X+~qzt,++ ~ E ~ Ezht+k). Using the full distribution of simulated household expenditures (tzht+k) in the target data, welfare measures (in this case the Pa poverty measures) are calculated for each simulation. This procedure is carried out for 100 simulations and yields a distribution of welfare measures. The means of the poverty measures are reported as the point estimates, and the standard deviations as the standard errors of these measures (see table 3). Various distributional forms for the location (7,)and idiosyn- cratic components of the disturbance term were used. These include normal, t (with varying degrees of freedom), and nonparametric distributions. As the results were robust to these different distributions, only the poverty estimates from simulations with normal distributions are reported. 34. Note that this variance is a function of the target data. 340 T H E W O R L D BANK E C O N O M I C REVIEW Alderman, Harold, Miriam Babita, Gabriel Demombynes, Nthabiseng Makhatha, and Berk zler. 2002. "How Low Can You Go? Combining Census and Survey Data for Poverty Mapping in South Africa." Journal of Afrzcan ~conomres1 l(2):169-200. Alwang, Jeffrey, Bradford Mills, and Nelson Taruvinga. 2002. "Why Has Poverty Increased in Zimbabwe?" Poverty Dynamics in Africa Monograph. World Bank, Washington, D.C. Appleton, Simon. 1996. "Problems of Measuring Changes in Poverty over Time: The Case of Uganda 1989-1992." institute of Development Studies Bulletin 27(1):43-55. Azzarri Carlo, Gero Carletto, Beniamin Davis, and Alberto Zezza. 2006. "Monitoring Poverty without Consumption Data: An Application using the Albania Panel Survey." Eastern Errropean Economics 44(1):59-82. Chesher, Andrew, and Christian Schluter. 2002. "Welfare Mcasurement and Measurement Error." Review of Economic Studies 69(21:357-78. Christiaensen, Luc, and Kalanidhi Subbarao. 2005. "Towards an Understanding of Household Vulnerability in Rural Kenya." Journal of African Economies 14(4):520-58. Deaton, Angus. 1997. The Analysis of Household Surveys: A Microeconometric Approach to Development Policy. Washington, D.C.: Johns Hopkins University Press and World Bank. -. 2003. "Household Surveys, Consumption and the Measurement of Poverty." Economic Systems Research 15(2):135-59. Deaton, Angus, and Jean Drzc. 2002. "Poverty and Inequality in India: A Reexamination." Economic and Political Weekly 37(7):3729-48. Deaton, Angus, and Valerie Kozel. 2005. "Data and Dogma: The Great Indian Poverty Debate." World Bank Research Observer 20(2):177-99. Deaton, Angus, and Salmnn Zaidl. 2002. "Guidelines for Constructing Consumption Aggregates for Welfare Analysis." Living Standards Measurement Study Working Paper 135. World Bank, Washington, D.C. Elbers, Chris, Jean 0. Lanjouw, and Peter Lanjouw. 2002. "Micro-level Estimation of Welfare." World Bank Policy Research Working Paper 2911.World Bank, Washington, D.C. .2003. "Micro-level Estimation of Poverty and Inequality." Econornetrica 71(1):355-64. Ferreira, Francisco, and Ricardo Peas de Barros. 1999. "The Slippery Slope: Explaining the Increase in Extreme Poverty in Urban Brazil, 1976-1996." The Brazilian Review of Econometrics 19(2):211- 96. Hoogeveen, Johannes, and Gabriel Demombynes. 2004. "Growth, Inequality, and Simulated Poverty Paths for Tanzania, 1992-2002." World Bank Policy Research Working Paper 3432. World Bank, Washington, D.C. Juhn, Chinhui, Kevin Murphy, and Brooks Pierce. 1993. "Wage Inequality and the Rise in Returns to Skill." Journal of Political Economy 101(3):410-42. Kijima, Yoko, and Peter Lanjouw. 2003. "Poverty in India during the 1990s: A Regional Perspective." Policy Research Working Paper 3141. World Bank, Washington, D.C. Mistiaen. Johan, Berk zler, Tiaray Razafimanantena, and Jean Razafindravonona. 2002. "Putting Welfare on the Map in Madagascar." Africa Region Working Paper 34. World Bank, Washington, D.C. Nyoro, James K., M. Muyanga, and I. Komo. 2005. Revisiting the Role of Agricultural Performance in Reducing Poverty in Kenya. Kenya: Tegemeo Institute of Agricultural Policy and Development, Egerton University. Pradhan, Menno. 2001. '"Welfare Analysis with a Proxy Consumption Measure: Evidence from a Repeated Experiment in Indonesia." Cornell Food and Nutrition Program Working Paper 126. Ithaca, N.Y.: Cornell University. Stifel and Christiaensen 341 Ravallion, Martin. 1996a. "Issues in Measuring and Modeling Poverty." Economic Journal 106(438):1328-43. . 1996b. "How Well Can Method Substitute for Data? Five Experiments in Poverty Analysis." World Bank Research Observer 11(2):199-221. . 2003. "Measuring Aggregate Wclfarc in Developing Countries: How Well Do National Accounts and Surveys Agree?" Review of Economics and Statistics 85(3):645-52. Sahn, David, and David Stifel. 2000. "Poverty Comparisons Over Time and Across Countries in Africa." World Development 28(12):2123-55. Scott, Christopher, and Ben Amenuvegbe. 1990. "Effect of Recall Duration on Reporting of Household Expenditures: An Experimental Study in Ghana." Social Dimensions of Adjustment in sub-Saharan Africa Working Paper 6. World Bank, Washington, D.C. Sen Abhiiit, and Himanshu. 2004. "Poverty and Inequality in India-I." Economic and Political Weekly 39(38):4247-63. Simler, Kenneth, Sarah Harrower, and Claudio Massingarela. 2004. Estimating Poverty Indices from Simple Indicator Surveys. Washington, D.C.: International Food Policy Research Institute. Stifel, David, and Luc Christiaensen. 2006. A Systematic Review of the Evolution and Prospects of Human Development in Kenya. World Bank, Washington, D.C. World Bank. 2003. "Kenya: A Policy Agenda to Restore Growth." World Bank Country Economic Memorandum 25840-KE. World Bank, Poverty Reduction and Economic Management, Africa Region, Washington, D.C. . 2005. "Well-Being and Poverty in Ethiopia: The Role of Agriculture and Agency." Report 29468. World Bank, Poverty Reduction and Economic Management, Africa Region, Washington, D.C. Forthcomingpapersin THE WORLD BANK ECONOMIC REVIEW Earnings, Schooling, and Economic Reform: Econometric Evidence from Hungary (1986-2004) Nauro Campos and Dean Jollge Comprehensive Wealth and Future Consumption: Accounting for Population Growth Susana Ft~reira,Kirh Hamilton, andJefrey R. Vincent The Impact of Decentralized Data Entry on the Quality of Household Survey Data in Developing Countries: Evidence from a Randomized Experiment in Vietnam Pal// Glewwe and Hai-Anh Hoang Dang Comparing the Net Benefits of Incentive Based and Command and Control Environmental Regulations:The Case of Santiago, Chide Rag/ O'Ryan andJost Miguel Sa'nchez