80518 Volume 25 • Number 3 • 2011 ISSN 0258-6770 (PRINT) ISSN 1564-698X (ONLINE) THE WORLD BANK ECONOMIC REVIEW Volume 25 • 2011 • Number 3 THE WORLD BANK ECONOMIC REVIEW What Constrains Africa’s Exports? Caroline Freund and Nadia Rocha Does the Internet Reduce Corruption? Evidence from U.S. States and across Countries Thomas Barnebeck Andersen, Jeanet Bentzen, Carl-Johan Dalgaard, and Pablo Selaya Do Labor Statistics Depend on How and to Whom the Questions Are Asked? Results from a Survey Experiment in Tanzania Elena Bardasi, Kathleen Beegle, Andrew Dillon, and Pieter Serneels SYMPOSIUM ON ENTREPRENEURSHIP AND DEVELOPMENT Entrepreneurship and Development: The Role of Information Asymmetries Leora Klapper and Inessa Love Getting Credit to High Return Microentrepreneurs: The Results of an Information Intervention Suresh de Mel, David McKenzie, and Christopher Woodruff The Impact of the Business Environment on Young Firm Financing Larry W. Chavis, Leora F. Klapper, and Inessa Love Does a Picture Paint a Thousand Words? Evidence Pages 361–561 from a Microcredit Marketing Experiment Xavier Giné, Ghazala Mansuri, and Mario Picón Entrepreneurship and the Extensive Margin in Export Growth: A Microeconomic Accounting of Costa Rica’s Export Growth during 1997-2007 Daniel Lederman, Andrés Rodríguez-Clare, and Daniel Yi Xu www.wber.oxfordjournals.org 2 THE WORLD BANK ECONOMIC REVIEW editors Alain de Janvry and Elisabeth Sadoulet, University of California at Berkeley assistant to the editor Marja Kuiper editorial board Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on August 19, 2013 Harold H. Alderman, World Bank (retired) William F. Maloney, World Bank Pranab K. Bardhan, University of California, David J. McKenzie, World Bank Berkeley Jaime de Melo, University of Geneva Scott Barrett, Columbia University, USA Juan-Pablo Nicolini, Universidad Torcuato di Asli Demirgüç-Kunt, World Bank Tella, Argentina Jean-Jacques Dethier, World Bank Nina Pavcnik, Dartmouth College, USA Quy-Toan Do, World Bank Vijayendra Rao, World Bank Frédéric Docquier, Catholic University of Martin Ravallion, World Bank Louvain, Belgium Jaime Saavedra-Chanduvi, World Bank Eliana La Ferrara, Università Bocconi, Italy Claudia Paz Sepúlveda, World Bank Francisco H. G. Ferreira, World Bank Joseph Stiglitz, Columbia University, USA Augustin Kwasi Fosu, United Nations Jonathan Temple, University of Bristol, UK University, WIDER, Finland Romain Wacziarg, University of California, Paul Glewwe, University of Minnesota, Los Angeles, USA USA Dominique Van De Walle, World Bank Ann E. Harrison, World Bank Christopher M. Woodruff, University of Philip E. Keefer, World Bank California, San Diego Justin Yifu Lin, World Bank Yaohui Zhao, CCER, Peking University, Norman V. Loayza, World Bank China The World Bank Economic Review is a professional journal used for the dissemination of research in development economics broadly relevant to the development profession and to the World Bank in pursuing its development mandate. It is directed to an international readership among economists and social scientists in government, business, international agencies, universities, and development research institutions. The Review seeks to provide the most current and best research in the field of quantita- tive development policy analysis, emphasizing policy relevance and operational aspects of economics, rather than primarily theoretical and methodological issues. Consistency with World Bank policy plays no role in the selection of articles. The Review is managed by one or two independent editors selected for their academic excellence in the field of development economics and policy. The editors are assisted by an editorial board composed in equal parts of scholars internal and external to the World Bank. World Bank staff and outside researchers are equally invited to submit their research papers to the Review. For more information, please visit the Web sites of the Economic Review at Oxford University Press at www.wber.oxfordjournals.org and at the World Bank at www.worldbank.org/research/journals. Instructions for authors wishing to submit articles are available online at www.wber.oxfordjournals.org. Please direct all editorial correspondence to the Editor at wber@worldbank.org. THE WORLD BANK ECONOMIC REVIEW Volume 25 † 2011 † Number 3 What Constrains Africa’s Exports? 361 Caroline Freund and Nadia Rocha Does the Internet Reduce Corruption? Evidence from U.S. States and across Countries 387 Thomas Barnebeck Andersen, Jeanet Bentzen, Carl-Johan Dalgaard, and Pablo Selaya Do Labor Statistics Depend on How and to Whom the Questions Are Asked? Results from a Survey Experiment in Tanzania 418 Elena Bardasi, Kathleen Beegle, Andrew Dillon, and Pieter Serneels __________________ SYMPOSIUM ON ENTREPRENEURSHIP AND DEVELOPMENT Entrepreneurship and Development: The Role of Information Asymmetries 448 Leora F. Klapper and Inessa Love Getting Credit to High Return Microentrepreneurs: The Results of an Information Intervention 456 Suresh de Mel, David McKenzie, and Christopher Woodruff The Impact of the Business Environment on Young Firm Financing 486 Larry W. Chavis, Leora F. Klapper, and Inessa Love Does a Picture Paint a Thousand Words? Evidence from a Microcredit Marketing Experiment 508 Xavier Gine ´n ´ , Ghazala Mansuri, and Mario Pico Entrepreneurship and the Extensive Margin in Export Growth: A Microeconomic Accounting of Costa Rica’s Export Growth during 1997-2007 543 Daniel Lederman, Andre ´guez-Clare, and Daniel Yi Xu ´ s Rodrı 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 25, 3 Issues, 2011): Institutions—Print edition and site-wide online access: £168/ $252/E252, Print edition only: £154/$231/E231, Site-wide online access only: £140/$210/E210; Corporate—Print edition and site-wide online access: £251/$376/E376, Print edition only: £230/$345/E345, Site-wide online access only: £209/$314/E314; Personal—Print edition and individual online access: £43/$64/E64. US$ rate applies to US & Canada, EurosE applies to Europe, UK£ applies to UK and Rest of World. There may be other subscription rates available; for a complete listing, please visit www.wber.oxfordjournals.org/subscriptions. Readers with mailing addresses in non-OECD countries and in socialist economies in transition are eligible to receive complimentary sub- scriptions on request by writing to the UK address below. Full prepayment in the correct currency is required for all orders. Orders are regarded as �rm, and payments are not refundable. Subscriptions are accepted and entered on a complete volume basis. Claims cannot be con- sidered more than four months after publication or date of order, whichever is later. All subscriptions in Canada are subject to GST. Subscriptions in the EU may be subject to European VAT. If registered, please supply details to avoid unnecessary charges. For subscriptions that include online versions, a proportion of the subscription price may be subject to UK VAT. Personal rates are applicable only when a subscription is for individual use and are not available if delivery is made to a corporate address. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on August 19, 2013 The current year and two previous years’ issues are available from Oxford University Press. Previous BACK ISSUES: volumes can be obtained from the Periodicals Service Company, 11 Main Street, Germantown, NY 12526, USA. E-mail: psc@periodicals.com. Tel: (518) 537-4700. Fax: (518) 537-5899. CONTACT INFORMATION: Journals Customer Service Department, Oxford University Press, Great Clarendon Street, OxfordOX2 6DP, UK. E-mail: jnls.cust.serv@oup.com. Tel: þ44 (0)1865 353907. Fax: þ 44 (0)1865 353485. In the Americas, please contact: Journals Customer Service Department, Oxford University Press, 2001 Evans Road, Cary, NC 27513, USA. E-mail: jnlorders@oup.com. Tel: (800) 852-7323 (toll-free in USA/Canada) or (919) 677-0977. Fax: (919) 677-1714. In Japan, please contact: Journals Customer Service Department, Oxford University Press, Tokyo, 4- 5-10-8F Shiba, Minato-ku, Tokyo, 108-8386, Japan. E-mail: custserv.jp@oup.com. Tel: þ 81 3 5444 5858. Fax: þ 81 3 3454 2929. POSTAL INFORMATION: The World Bank Economic Review (ISSN 0258-6770) is published three times a year, in February, June, and October, by Oxford University Press for the International Bank for Reconstruction and Development/THE WORLD BANK. Send address changes to The World Bank Economic Review, Journals Customer Service Department, Oxford University Press, 2001 Evans Road, Cary, NC 27513-2009. Periodicals postage paid at Cary, NC and at additional mailing of�ces. Communications regarding original articles and editorial management should be addressed to The Editor, The World Bank Economic Review, The World Bank, 3, Chemin Louis Dunant, CP66 1211 Geneva 20, Switzerland. ENVIRONMENTAL AND ETHICAL POLICIES: Oxford Journals, a division of Oxford University Press, is committed to working with the global community to bring the highest quality research to the widest possible audience. Oxford Journals will protect the environment by implementing environmentally friendly policies and practices wherever possible. Please see http://www.oxfordjournals.org/ethicalpolicies.html for further information on environmental and ethical policies. DIGITAL OBJECT IDENTIFIERS: For information on dois and to resolve them, please visit www.doi.org. PERMISSIONS: For information on how to request permissions to reproduce articles or information from this journal, please visit www.oxfordjournals.org/jnls/permissions. Advertising, inserts, and artwork enquiries should be addressed to Advertising and Special Sales, ADVERTISING: Oxford Journals, Oxford University Press, Great Clarendon Street, Oxford, OX2 6DP, UK. Tel: þ 44 (0)1865 354767; Fax: þ 44(0)1865 353774; E-mail: jnlsadvertising@oup.com. DISCLAIMER: Statements of fact and opinion in the articles in The World Bank Economic Review are those of the respective authors and contributors and not of the International Bank for Reconstruction and Development/THE WORLD BANK or Oxford University Press. Neither Oxford University Press nor the International Bank for Reconstruction and Development/THE WORLD BANK make any representation, express or implied, in respect of the accuracy of the material in this journal and cannot accept any legal responsibility or liability for any errors or omissions that may be made. The reader should make her or his own evaluation as to the appropriateness or otherwise of any experimental technique described. The World Bank Economic Review is printed on acid-free paper that meets the minimum require- PAPER USED: ments of ANSI Standard Z39.48-1984 (Permanence of Paper). INDEXING AND ABSTRACTING: The World Bank Economic Review is indexed and/or abstracted by CAB Abstracts, Current Contents/Social and Behavioral Sciences, Journal of Economic Literature/EconLit, PAIS International, RePEc (Research in Economic Papers), and Social Services Citation Index. COPYRIGHT # 2011 The International Bank for Reconstruction and Development/THE WORLD BANK All rights reserved; no part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise without prior written permission of the publisher or a license permitting restricted copying issued in the UK by the Copyright Licensing Agency Ltd, 90 Tottenham Court Road, London W1P 9HE, or in the USA by the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923. Typeset by Techset Composition Limited, Chennai, India; Printed by Edwards Brothers Incorporated, USA. What Constrains Africa’s Exports? Caroline Freund and Nadia Rocha Africa’s share of global exports has dropped by 50 percent over the last three decades. To stem this decline, aid for trade to the region has increased rapidly in recent years. Assistance can target improvements in three important components of trade facili- tation: transit times, documentation, and ports and customs. Of these, transit delays have the most economically and statistically signi�cant effect on exports. Speci�cally, a one day reduction in inland travel times leads to a 7 percent increase in exports, after controlling for the standard determinants of trade and potential endogeneity. Put another way, a one day reduction in inland travel times translates into a 2 percentage point decrease in all importing-country tariffs. By contrast, longer delays in the other areas have a far smaller impact on trade. Large transit delays are relatively more harmful because they are associated with high (within-country) variation, making delivery targets dif�cult to meet. Finally, the results imply that transit times are pri- Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 marily about institutional features—such as border delays, road quality, fleet class and competition and security—and not geography. JEL codes: F13, F14, O55. Sub-Saharan Africa’s share of world merchandise exports fell from 1.6 percent in 1980 to 0.8 percent today.1 Africa’s weak export performance is also appar- ent in an examination of export levels. Controlling for the standard determi- nants of trade, export volumes in Africa are about 16 percent below what is expected. In sum, over the last 30 years Africa has halved its world market Caroline Freund (cfreund@worldbank.org; corresponding author) is the Chief Economist of the Middle East and North Africa Region of The World Bank. Nadia Rocha (nadia.rocha@wto.org) is an Economic Of�cer in the Economic Research and Statistics Division, of the World Trade Organization. Authors would like to thank Allen Dennis for providing disaggregated data from the Doing Business report and the GPS team at the World Bank for providing detailed GPS data on Sub-Saharan Africa travel distances and times. In addition, authors would like to thank the editor of the journal, three anonymous referees, and seminar participants at the World Bank Trade seminar, the Nottingham Conference on Trade Costs and International Integration in Venice, the Geneva Trade and Development Workshop, and the European Trade Study Group (ETSG) conference for useful comments and suggestions. This article received �nancial support from the governments of Finland, Norway, Sweden and the United Kingdom through the Multidonor Trust Fund for Trade and Development. The views presented in this article are those of the authors and do not reflect the views of World Bank or the World Trade Organization. 1. In this article, Africa refers to Sub-Saharan Africa. Analysis uses a balanced sample of 167 countries (44 in Africa) reporting trade over the period. THE WORLD BANK ECONOMIC REVIEW, VOL. 25, NO. 3, pp. 361 –386 doi:10.1093/wber/lhr016 Advance Access Publication May 30, 2011 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 361 362 THE WORLD BANK ECONOMIC REVIEW share for exports, and today Africa exports less than expected given its income, population, location and other characteristics. Africa’s relatively poor export performance is worrisome for a number of reasons. First, export growth can substitute for lagging domestic demand. This is especially important in small economies, such as in Africa, where foreign markets are likely to be the main engines for growth (Bhagwati 1996, Krueger 1998). Numerous studies examine the effect of openness to trade on income growth, and in general �nd positive effects.2 Second, robust export growth yields both more jobs and better jobs. In particular, exporting �rms create jobs that pay higher wages and offer better working conditions than otherwise similar import-competing �rms.3 Third, strong export growth induces a more ef�cient production structure. This happens through compositional shifts, as the most productive exporting �rms grow most rapidly when exports boom.4 Finally, strong export growth helps prevent against the negative effects of balance of payments crises. This article investigates what constrains Africa’s exports, with a focus on Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 trade facilitation. The results show that long delays in getting export products from the factory gate and onto the ship explain much of Africa’s weak export performance. This is consistent with earlier work that �nds that domestic export delays hinder exports signi�cantly, and to a much greater extent than foreign tariffs (Hummels (2001), Djankov, Freund and Pham (2010), and Portugal and Wilson (2009)). This is especially debilitating for Africa’s exports because of extreme and often unexpected delays. This suggests that improving trade facilitation in Africa would signi�cantly boost exports. But there are different ways to accomplish this, as the domestic delay has three distinct components: documentation, transit time, and port handling and customs clearance. In this article it is explored whether these delays are equally burdensome or whether one of these binds relatively more, using detailed data on average trade times from the World Bank’s Doing Business report. This is important from a policy perspective in order to target aid for trade initiatives to their most productive uses. This is especially relevant for Africa—where aid for trade has increased more rapidly than any other region in recent years— making Africa now the second largest recipient (after Asia) of aid for trade (OECD/WTO 2009). Of the various delays, bureaucratic ones are the longest, taking 19 days on average. There is a lot of variation across countries. For example, it takes 36 days to process export documents in countries such as Angola, Zambia and Niger. In contrast, in Swaziland or in the Seychelles, it takes only 7 and 5 days 2. See Winters (2004) for a summary of the literature on trade and growth. 3. Bernard and Jensen (1995) report detailed statistics for the United States. A number of papers followed their approach and �nd similar results in both developing and developed economies. Schank, Schnabel and Wagner (2007) provide a summary of these papers and offer similar evidence for Germany. Bernard, Jensen, Redding, and Schott (2007) also provide a summary. 4. See Bernard, Jensen, Redding, and Schott (2007) for a summary of the literature. Freund and Rocha 363 respectively to produce all necessary export documents. Bureaucratic delays may be especially burdensome if they change often, making them dif�cult to predict, or if of�cials use them as means to extract rents. In contrast, documen- tation procedures may be less problematic if they are predictable and can be done in advance, or if there is learning by doing. Customs and ports delays are the second longest, taking on average 9 days. They are less variable than documents. Customs and ports could be especially restrictive if there is a hold-up problem. Once the goods arrive, customs and port authorities could extract high rents by delaying goods. In contrast, if customs and ports are reliable (but slow) or if exporters can pay for faster service they may cause fewer problems. Transit costs are on average the shortest, taking 7 days. But, again, there is a lot of variation. For example, it takes 31 days for the goods to be shipped from the city to the port in Chad and only one day within Gabon. Transit costs may be less burdensome if economic activity has developed endogenously, close to ports and borders when transit costs are large. However, they may be more Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 constraining if there is a lot of uncertainty that cannot be avoided. The main contribution of this article is to understand whether different types of export costs affect trade differently. A modi�ed gravity equation that controls for importer-�xed effects and exporter remoteness is used. An impor- tant concern with this approach is that the volume of trade may directly affect trade costs. The marginal value of investment in trade facilitation is higher when trade volumes are large since cost savings are passed on to a larger quan- tity of goods. In addition, many time-saving techniques, such as computerized container scanning, are only available in high-volume ports. Alternatively, increased trade volumes could increase congestion and lessen the ef�ciency of trade infrastructure. Thus, while more ef�cient trade facilitation may stimulate trade, trade is also likely to directly influence trade facilitation. Three distinct strategies are used to deal with the potential effect of export volumes on export times. First, the effect of trade facilitation on trade in new products is examined. These are goods that have not been exported in the past. The intuition is that trade in new products cannot affect the quality of trade facilitation infrastructure or the bureaucracy that is in place for exporting. Second, the effect of requirements in the transit country on exports from land- locked countries is analyzed. This controls for endogeneity because trade facili- tation in transit countries is likely to be exogenous from the perspective of a landlocked country. Finally, it is tested whether lengthy delays have a greater effect on exports of time-sensitive goods. The intuition is that these products make up a small share of total trade so are unlikely to affect trade facilitation. All three different techniques used to analyze the effect of export times of key components on trade values lead to the same conclusion: inland transit delays have a robust negative effect on export values. The estimation results imply that a one day increase of inland transit times reduces export values by about 7 percent. This effect is higher for time-sensitive goods with respect to 364 THE WORLD BANK ECONOMIC REVIEW time-insensitive goods. In contrast, the effects of documents, customs and ports on exports are much smaller. Why would delays in one area affect trade relatively more than in other areas? One potential answer is the variability of each time component. To evaluate this explanation, the effect of within-country export-time uncertainty on export values is examined for each type of delay. While an increase in transit-time uncertainty has a negative and signi�cant effect on trade values, the other time components show no such effect. This suggests that long and unexpected delays in transit make it especially dif�cult for producers to meet export deadlines. These results have important implications for policy. While reducing bureau- cratic delays and improving ports and customs may have positive effects on trade, the binding constraint in most African countries to expanding exports is inland transit. Improving inland transit is unlikely to be easy or cheap, but it is likely to boost exports and have broad positive economic effects. Beyond these direct implications for policy, the results presented in this Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 article also contribute to the broader debate about the influence of geography versus institutions on income. This literature has focused on the effects of climate versus governance on income, and potential interactions between the two.5 Here, the focus is on a single component of income (exports), and the variables of interest reflect geography and institutions to different extents. The dominance of transit time in hindering exports seems to suggest that geography is the main culprit. To test this, data from a GPS system on geo- graphical distance from the port to the economic center and on the estimated time of travel is gathered and included in the regression equation. The differ- ence between travel time in the GPS data and the Doing Business data is that the former are based solely on travel distance and estimated speed of travel by type of road ( paved or unpaved). These data do not incorporate delays due to the type of vehicles, borders, security, traf�c, or other road conditions. The results show that GPS distance negatively affects exports, but GPS travel time does not. Moreover, neither the economic effect nor the statistical signi�cance of the Doing Business inland transit-time variable changes when these vari- ables are included. This suggests that the problem for inland transit lies in the quality and security of the roads, border delays and the ef�ciency of security checkpoints, the age of the truck fleet and competition in trucking. These are factors which are more closely linked with institutions than geography. The article proceeds as follows. The next section discusses the data. Section II presents the estimation strategy. Section III describes the main results and robustness checks. Section IV examines the effect of uncertainty on exports. Section V determines the importance of purely geographical transit costs. Section VI concludes. 5. See, for example, Hall and Jones (1999), Acemoglu, Johnson, and Robinson (2000), and McArthur and Sachs (2001). Freund and Rocha 365 I . D ATA Data on trade times is based on answers to a comprehensive World Bank ques- tionnaire completed by trade facilitators at freight-forwarding companies in 146 countries in 2007 and is collected as part of Doing Business, a World Bank project that investigates the scope and manner of business regulations. Freight-forwarders are the most knowledgeable to provide information on trade costs because most businesses use their services to move their products across borders. They are estimated to handle approximately 85% of global international trade. Their services range from �nding the most appropriate route for a shipment, preparing documentation to meet customs and insurance requirements, arranging payments of fees and duties, and advising on legisla- tive changes and political developments that could affect the movement of freight. Overall, 345 trade facilitators participated in the survey, with at least two per country.6 To document the procedures and export times, the survey respondents are asked about a stylized transaction. The exporter is a local business (100% Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 owned by nationals), has 201 employees, and is located in the country’s most populous city. The exporter does not operate within an export-processing zone or an industrial estate with special export privileges. Each year, more than 10% of its sales go to international markets, i.e., management is familiar with all the trading rules and requirements. The purpose of de�ning the exporter speci�cally is to avoid special cases. Assumptions are also made on the cargo to make it comparable across countries. The traded product travels in a dry-cargo, 20-foot, full container load. It is not hazardous and does not require refrigeration. The product does not require any special phytosanitary or environmental safety standards other than accepted international shipping standards, in which cases export times are likely to be longer. Finally, every country in the sample exports this product category.7 The questionnaire also asks respondents to identify the likely port of export. For some countries, especially in Africa and the Middle East, this may not be the nearest port. For example, Cotonou, Benin’s main port, is seldom used due to perception of corruption and high terminal handling fees. The novel part of the data used for this investigation is on the distinctions by type of trade time. The data provide detailed information on the different kinds of costs an exporter faces when moving his goods from the principal city to the port of exit. More precisely, the survey asks respondents the average and 6. Follow-up conference calls were conducted with all respondents to con�rm the coding of the data. As a further quality check, surveys were completed by port authorities and customs of�cials in a third of the sample (48 countries). 7. These assumptions yield three categories of goods: textile yarn and fabrics (SITC 65), articles of apparel and clothing accessories (SITC 84), and coffee, tea, cocoa, spices and manufactures thereof (SITC 07). 366 THE WORLD BANK ECONOMIC REVIEW F I G U R E 1. Export Procedures by Category Source: Authors’ calculations on Doing Business data. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 the maximum times in calendar days it takes for completing a series of export procedures. Each procedure can be classi�ed into one of four main categories: documentation, inland transportation, customs, and ports. The �rst category records the time it takes for an exporter to complete all documentation activities such as securing a letter of credit, assembling and pro- cessing export and international shipping certi�cates and realizing all pre- shipment inspections and clearance. Inland transportation includes the time it takes for the merchandise to be moved from the principal city to the port of exit, as well as the time spent arranging transport and waiting times for the merchandise’s pick up and loading into a carriage. For landlocked countries, total transport times also include waiting times at the crossing border. The customs category includes the time necessary to realize the technical controls of the merchandise. In addition, for landlocked countries this cat- egory comprises the total time it takes from the submission of request of clearance until the completion of the inspection and clearance procedure in the transit country. Finally, the ports category represents terminal handling times, including storage if a certain storage period is required, the waiting times for loading the containers into the vessel and customs inspection and clearance times. An example illustrates the data. An exporter in Rwanda spends 43 days on average to complete all requirements for shipping its merchandise abroad: 17 days each on delays resulting from documentation and inland transit, while port and custom procedures take respectively 6 and 3 days on average (see Figure 1). Freund and Rocha 367 T A B L E 1 . Times to Export Descriptive Statistics by Geographic Region Region Statistics Documents Customs and ports Inland transit East Asia & Paci�c (23) mean 12 8.5 3.9 sd 9.7 4.9 3.1 min 1 2 1 max 39 19 15 Europe & Central Asia (25) mean 13.8 7.6 9.6 sd 9 6.4 14 min 1 2 1 max 32 34 58 Latin America & Caribbean (30) mean 11.2 7.3 3.9 sd 6.7 3.6 3.1 min 4 2 1 max 30 18 18 Middle East & North Africa (12) mean 10.3 6.1 3.6 sd 3.4 2.7 1.9 min 5 3 2 max 18 13 8 OECD (24) mean 5 3.1 2 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 sd 3.2 1 1 min 2 2 1 max 14 6 4 South Asia (8) mean 16.3 8.6 7.6 sd 11.5 2.6 6.9 min 9 5 1 max 44 12 21 Sub-Saharan Africa (45) mean 18.7 9.4 7.2 sd 9 4.2 7 min 5 2 1 max 41 28 31 Notes: 1. The unit of measure is number of days. 2. Number of countries for each region in parenthesis. Source: Authors’ analysis based on Doing Business data. The summary statistics for each of the components representing the total time8 necessary to ful�l all the requirements for exporting by region and regional arrangement are presented in Table 1. The data show that across regions, docu- mentation procedures times are the longest. Furthermore, while getting a product from the factory to the ship is relatively quick in developed countries, this is not the case for Sub-Saharan Africa, where all time costs categories are on average higher compared to all the other regions. Customs and ports procedures and inland transportation take on average three times more in African countries than in OECD countries. In addition, documentation procedures take four times longer in African countries compared with developed countries. 8. The time delays reported in the survey are probably at the lower end of the time it takes to move the average product from factory to ship. This is because the products are chosen so that they do not require cooling or any technical inspections based on use of hazardous materials. 368 THE WORLD BANK ECONOMIC REVIEW The rest of the data are from standard sources. The trade data are both from the UN Comtrade database and the IMF Direction of Trade database. GDP and Population are from the World Bank’s World Development Indicators. Gravity variables such as country-pair distances, language and colony are taken from the Mayer and Zignago dataset. Country’s Capital abundance information is avail- able for 2005 and comes from GTAP 7 database. Simple average tariffs at 6 digit level are taken from the TRAINS dataset. Aid for trade data is collected for the 2006 and comes from the OECD/DAC Creditor Reporter System on disburse- ments. Table 2 presents correlations between the variables used in the analysis. II. ME T H O D O LO GY In order to examine the effect of trade cost an augmented gravity equation is estimated as a �rst step: LnExportsij ¼ b1 Ln GDPi þ b2 Ln Popi þ b3 LnDistij þ b4 Ln Remotei ð1Þ Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 þ b5 landlocki þ Xij þ mj þ 1ij ; where the i and j subscripts correspond to the exporter and the importer, respectively.9 The dependent variable is the log of bilateral exports from country i to country j. The standard determinants of trade are: mj, importer- �xed effects, which control all importers-speci�c characteristics; GDPi and POPi are respectively the Gross Domestic Product and the total population of the exporting country; Distij is the distance between country i and country j. Xij is a vector of dummy variables associated with the exporter and the impor- ter such as sharing the same of�cial language or border or past colony/coloni- zer relationship. Landlocki is a dummy variable equal to one if the exporter country is landlocked and zero otherwise. Remotei is a measure for the exporter’s remoteness and is calculated following Head (2003), Remotei ¼ P GDP 1 .10 To examine whether African trade is different, this j =Distij j equation is estimated on all available data and a dummy that is one if the exporter is an African country is included in the regression. A negative coef�- cient would imply that Africa undertrades relative to other countries. The main purpose of this article, however, is to investigate how three diverse trade costs—completing documentation, inland transit delays, and customs and ports times—affect Africa’s trade volumes. Longer time delays act as a tax on 9. The times representing terminal port handling and customs and technical control were aggregated due to their very high correlation (See table 2). 10. It is important to control for remoteness in the regressions for two reasons. First, there is evidence that a country’s trade with any given partner is dependent on its average remoteness to the rest of the world (Anderson and Van Wincoop (2003)). Furthermore, remoteness is correlated with factory-to-port time delays hence not including it into the regression would produce biased estimates of the impact of trade facilitation on export volumes. T A B L E 2 . Correlation of Explanatory Variables Travel Total Inland Aid GPS Dist. time Uncert Uncert. export Inland Docs Customs Ports transp. for city to city to Uncert custom & Inland GDP POP time Docs Customs Ports transp. (TC) (TC) (TC) (TC) Remote Trade port port Docs Ports Transit GDP 1 POP 0.73 1 Total Export time 0.04 0.25 1 Documents 0.01 0.19 0.84 1 Customs 0.22 0.22 0.31 0.10 1 Ports 0.28 0.17 0.35 0.13 0.39 1 Inland transport 2 0.11 0.12 0.72 0.35 0.07 2 0.02 1 Docs (TC) 0.03 0.14 0.47 0.66 0.20 0.26 2 0.06 1 Customs (TC) 0.16 0.08 0.26 0.14 0.76 0.36 2 0.01 0.35 1 Ports (TC) 0.22 0.05 0.33 0.16 0.28 0.89 2 0.01 0.27 0.36 1 Inland transp. 0.07 0.17 0.36 0.18 0.18 0.24 0.34 0.26 0.15 0.14 1 (TC) Remote 2 0.15 2 0.08 2 0.24 2 0.11 2 0.23 2 0.12 2 0.23 2 0.10 2 0.10 2 0.13 0.02 1 GPS dist. city 0.03 0.21 0.65 0.48 0.09 2 0.07 0.71 2 0.03 0.00 0.03 2 0.06 2 0.38 1 to port GPS travel time 0.00 0.19 0.68 0.47 0.06 2 0.07 0.80 2 0.06 2 0.04 2 0.02 2 0.02 2 0.34 0.98 1 city to port Aid for Trade 0.37 0.55 0.16 0.07 0.26 0.00 0.17 2 0.16 0.09 2 0.11 0.10 2 0.11 0.27 0.23 1 Uncert. Docs 0.01 2 0.18 2 0.19 0.01 2 0.23 2 0.17 2 0.25 0.18 2 0.27 2 0.11 2 0.03 0.69 2 0.29 2 0.24 2 0.12 1 Uncert. Custom 0.45 0.14 2 0.01 0.14 2 0.53 2 0.13 2 0.03 2 0.11 2 0.55 0.26 2 0.10 0.27 0.16 0.17 2 0.06 0.32 1 & Ports Uncert. Inland 0.05 2 0.22 0.57 0.58 2 0.19 0.10 0.26 0.14 2 0.27 0.13 2 0.27 0.09 0.44 0.44 2 0.13 0.09 0.24 1 Transit Notes: TC stands for transit country. Freund and Rocha Source: Authors’ analysis based on data sources discussed in the text. 369 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 370 THE WORLD BANK ECONOMIC REVIEW exports, especially on high-value goods, since they are effectively depreciating during the delay. In addition, the exporter must spend capital on the exporting process and storage/transport of the goods during the delay. For this analysis, equation (1) is estimated including the Doing Business (DB) variables: LnExportsij ¼ b1 Inland transit i þ b2 Customs & Portsi þ b3 Documentsi þ b4 Ln GDPi þ b5 Ln Popi þ b6 LnDistij ð2Þ þ b7 Ln Remotei þ b8 landlocki þ Xij þ mj þ 1ij : The variables of interest are the export times for transit, customs and ports, and documents. The coef�cient on each represents the effect in percent of trade of a one day increase in that component. The variables are analyzed in levels, so that the coef�cients are comparable—the effect of a one day change. However, for robustness, the regression with the three variables in logs is also estimated. The previous speci�cation could be subject to omitted-variables bias given that the error term might include the effect of country-speci�c policy variables Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 that affect both trade flows and the time-to-export variables. A potential vari- able of concern is aid for trade. Aid affects trade directly through productive capacity assistance or technical assistance for trade policy and also through trade-related infrastructure assistance. To control for this, a lagged variable capturing the total amount of aid-for-trade assistance received by the exporting countries is included. There is a potential reverse causality problem in the regressions because time-to-export variables are likely to be correlated with country exports. An improvement of infrastructure and administrative time costs has positive effects on exports. However, countries that export more may have higher returns to enhance local trade facilitation and invest more in time ef�cient means. In addition, some types of export processing might only be available in high volume ports. Hummels and Skiba (2004), for example, provide evidence that trade volumes affect the adoption time of containerized shipping, which greatly reduces shipping costs. Finally, it might be the case that in countries with higher volumes of trade, export procedures will be affected by congestion effects. In this case, the presence of reverse causality will lead to an underesti- mation of the coef�cient on time costs. To control for the possibility that more trade leads to improved trade facili- tation, the effects on exports of new products are investigated.11 The intuition is that exports of new products cannot have had an impact on the historical develop- ment of infrastructure or the type of bureaucratic procedures in place. In addition, because they are a very small share of total trade, they are unlikely to be associ- ated with congestion effects. The approach of Djankov, Freund, and Pham (2010) is also followed. First, trade times of transit countries are used as instruments for 11. New products are de�ned as those products that were not exported in the years 2002-2004 and that entered into the export market in the time interval 2005-2007. Freund and Rocha 371 trade costs in landlocked countries and second, it is examined whether trade times affect time-sensitive goods relatively more. II I. R ES ULT S As a �rst step, Africa’s exports are examined in comparison with the rest of the world. The augmented gravity equation (1) is estimated on all countries with available data, and a dummy for sub-Saharan Africa is introduced. Recall, con- trols for income, population, distance, common border, language, colonial heritage, landlocked, remoteness and importer-�xed effects are included in this regression. Results are reported in Table 3. The negative and signi�cant value of the Africa dummy (column 1) implies that countries in the region export about 16 percent (e20.18 2 1 ¼ 2 0.16) less than expected.12 Column (2) includes the level of the total export time and the coef�cient on the Africa dummy falls from Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 2 0.18 to 2 0.11 and becomes signi�cant at the 10 percent level only. This implies that trade costs are a signi�cant part of the explanation for Africa’s weak performance. In the next two columns, the effect of trade costs in all countries is compared with the effect of trade costs in Africa. Column (3) reports the basic regression with export times for all countries and column (4) reports the same regression only for sub-Saharan Africa. The time to export is a more important determinant of African trade than overall trade. Finally, the last two columns report the results with the time variable in logs and the nature of the results remains—time is a much bigger deterrent to African trade than general trade. While it would be interesting to perform this type of analysis for each region, this article focuses on Africa because export growth has been relatively weak and trade costs are especially important. In addition, it is likely that the 12. Rodrik (1997) and Coe and Hoffmaister (1999) �nd that Africa’s low trade can be explained with enough control variables. This is not inconsistent with our results, as the Africa dummy loses signi�cance when trade costs are included. In addition, if only income, distance, and partner effects are included, the coef�cient on the Africa dummy more than doubles in magnitude, highlighting that Africa’s weak trade performance is in part explained by the large number of landlocked countries and the general remoteness of the region. Still, there are some important distinctions between the methodology used in this article methodology and theirs. Both Rodrik and Coe and Hoffmaister examine exports and imports together, and many African countries run persistent trade de�cits. Rodrik uses exports plus imports relative to GDP (openness) as the dependent variable, which makes the regression less theoretically based and not a test of export performance. In addition, he controls also for Latin American, East Asian, and OECD countries with dummies—so his measure of Africa’s performance is as compared with the Middle East, Eastern Europe and Central and South Asia, other traditionally weak trade performers. Coe and Hoffmaister examine only trade flows between South and North countries. Still, they �nd very similar results in their basic speci�cation: Africa’s trade is signi�cantly below what is expected in a regression with income, population, and distance. They show that if you control for openness to trade, type of exports, language and a number of �xed effects it may go away. However, even these variables do not change their other main result, that Africa’s trade performance has signi�cantly weakened over recent decades. 372 T A B L E 3 . Is Africa Different? All countries All countries All countries Sub-Saharan Africa All countries Sub-Saharan Africa Dependent Variable: (levels) (levels) (levels) (levels) (logs) (logs) ln (Aggregate exports) (1) (2) (3) (4) (5) (6) GDP 1.248*** 1.136*** 1.149*** 1.080*** 1.104*** 1.060*** [0.014] [0.019] [0.017] [0.053] [0.020] [0.053] Population 2 0.098*** 0.013 0.001 0.004 0.042* 0.047 [0.017] [0.021] [0.020] [0.050] [0.022] [0.051] THE WORLD BANK ECONOMIC REVIEW Distance 2 1.474*** 2 1.474*** 2 1.474*** 2 1.186*** 2 1.474*** 2 1.170*** [0.027] [0.027] [0.027] [0.127] [0.027] [0.128] Sub-Saharan Africa dummy 2 0.180*** 2 0.114* [0.069] [0.069] Total time to export 2 0.024*** 2 0.024*** 2 0.056*** 2 0.626*** 2 2.005*** [0.002] [0.002] [0.006] [0.055] [0.203] Observations 16,085 16,085 16,085 3,494 16,085 3,494 R-squared 0.675 0.678 0.678 0.525 0.678 0.524 Notes: 1. Robust standard errors in brackets. ***p , 0.01, **p , 0.05, *p , 0.1. 2. Other control variables: partner FE, common language, common border, colony, remoteness, landlocked. Source: Authors’ analysis based on data sources discussed in the text. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Freund and Rocha 373 relative importance of the type of trade costs varies by region. This would make a single analysis of the three types of costs in �ve regions quite complex. To focus on the three types of costs in the African sample, the augmented gravity equation from expression (2) is estimated.13 The linear regression results for a sample of 45 Sub-Saharan Africa countries are reported in Table 4. In order to control for the presence of zero trade flows, a nonlinear estimation with censored data is also performed (see column (2) of Table 4).14 The �rst two columns show the results from estimation on all trade. In both cases all three variables are signi�cant and their coef�cients are similar, though it is somewhat higher for inland transit.15 However, this column does not deal with the problem of endogeneity of the right hand side variables. In column (3), results for trade in new goods only are reported. The time variables are less likely to be endogenous to trade in new goods, since this trade was not around in the past when procedures and infrastructure for trade were devel- oped. The results are somewhat different. While the coef�cient on inland transit is little changed from column (1), the other coef�cients fall consider- Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 ably, suggesting that the previous column was also picking up the effect of trade on documentation procedures and customs and ports. In particular the results imply that a one day increase in transit time leads to a nearly 7 percent decline in exports. In the next �ve columns, robustness tests are reported. Columns (4)-(6) report the results of each variable independently and total time. This helps to deal with potential multicollinearity between the variables and also informs us whether each variable is signi�cantly different from total time in its effect on exports. Only inland transit has an independent effect on exports. Moreover the total effect of inland transit, equivalent to 0.066 (0.049 þ 0.017), is nearly four times as large as the effect of the other components of time. This outcome holds after the inclusion of foreign import tariffs in the regression (see column (7)).16 Including foreign tariffs also allows interpreting a day in terms of tariffs. A one day delay is roughly equivalent to a 2 percent point reduction in all importer-country tariffs. Importer tariffs are on average 10 percent and on average transit time is 7 days. This implies that cutting the transit time in half would expand trade by 30 percent; while all importers cutting tariffs in half 13. Equation (2) is estimated excluding trade in oils and minerals from the dependent variable to control for the surge in commodities exports that took place in some African countries during the 2007. 14. A Tobit regression with left-censoring at zero is run. 15. Results from OLS and Tobit estimations are qualitatively similar. However, in the second case the magnitude of the coef�cients more than doubles. This might be explained by the fact that almost 30% of the observations are left-censored. 16. To control for the fact that some African countries bene�t from preferential tariffs a country-pair speci�c dichotomous variable that takes the value of 1 when the partner grants lower preferences to the reporter through the generalized system of preferences (GSP) program and zero otherwise is also introduced. This variable is never signi�cant, but is included in the regressions as a control (coef�cients not reported). 374 T A B L E 4 . The Effect of Export Time Components on Aggregate Exports (OLS Regression) New New New New New New OLS Tobit Products Products Products Products Products Products Dependent variable: (levels) (levels) (levels) (levels) (levels) (levels) (levels) (logs) ln (Aggregate exports) (1) (2) (3) (4) (5) (6) (7) (8) Inland transit time 2 0.073*** 2 0.198*** 2 0.066*** 2 0.049*** 2 0.070*** 2 0.417*** [0.012] [0.025] [0.015] [0.017] [0.024] [0.134] Customs and ports time 2 0.052*** 2 0.129*** 2 0.018** 0.022* [0.007] [0.014] [0.007] [0.013] Documents time 2 0.045*** 2 0.128*** 2 0.016 0.012 [0.015] [0.027] [0.014] [0.016] GDP 1.070*** 1.979*** 0.982*** 1.019*** 0.992*** 0.984*** 1.030*** 0.975*** [0.054] [0.109] [0.063] [0.060] [0.062] [0.061] [0.085] [0.061] THE WORLD BANK ECONOMIC REVIEW Population 2 0.017 0.050 2 0.323*** 2 0.351*** 2 0.328*** 2 0.324*** 2 0.414*** 2 0.281*** [0.055] [0.117] [0.066] [0.065] [0.066] [0.065] [0.090] [0.067] Distance 2 1.210*** 2 3.565*** 2 0.899*** 2 0.889*** 2 0.868*** 2 0.900*** 2 1.029*** 2 0.868*** [0.128] [0.287] [0.161] [0.161] [0.160] [0.161] [0.196] [0.160] Aid for Trade 0.039 0.368*** 0.079* 0.088** 0.074* 0.080* 0.038 0.073* [0.037] [0.074] [0.042] [0.042] [0.041] [0.041] [0.056] [0.041] Total export time 2 0.040*** 2 0.029*** 2 0.017*** 2 0.016* 2 0.502** [0.010] [0.006] [0.006] [0.008] [0.220] Tariffs (simple average) 2 0.048** [0.023] Observations 3,494 5,737 2,054 2,054 2,054 2,054 1,142 2,054 R-squared 0.525 0.425 0.423 0.422 0.425 0.433 0.424 Notes: 1. Robust standard errors in brackets. ***p , 0.01, **p , 0.05, *p , 0.1. 2. Other control variables: partner FE, common language, common border, colony, remoteness, landlocked and GSP (whenever import tariffs are included in the regressions). Source: Authors’ analysis based on data sources discussed in the text. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Freund and Rocha 375 would expand trade by 25 percent.17 Finally in column (8) results using logs of the time variables are reported. Again, it can be seen that only transit time is independently signi�cant (results for other variables are not reported). The second strategy to deal with the potential endogeneity of the export time variables is to use a sample of landlocked countries and use the variables for the transit country(ies) as the instrument. This follows from Djankov, Freund and Pham (2010).18 Results are reported in Table 5. The �rst column reports OLS regression results for this sample. With the exception of docu- ments, the coef�cients are much larger for this sample than for the full sample (column (1) of Table 4). One explanation is that the endogeneity problem is greater here. For example, when landlocked country trade is small, customs and ports authorities (which must be located in neighboring countries) give them the lowest priority. To control for endogeneity, each time variable is instrumented with the corresponding variable faced by exporters in the transit country. The F-statistics of the �rst stage regressions (see Table 1A of the Appendix) indicate that none of the instruments is a weak instrument.19 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Second stage regression results, with and without foreign tariffs, are reported in levels in columns (2) and (3) and in logs in columns (4) and (5). Transit trade from a given country tends to be a small share of total coastal country trade, on average 27 percent. Still, this can be a problem if large land- locked exporters have more influence on times getting through transit countries. As robustness check, the estimated GPS distance in km from the border between the landlocked and the transit country and the port of the latter20 is also used as an alternative instrument for inland transit times. Results in levels and logs are reported in columns (6) and (7). For landlocked countries, the instrumental variables (IV) results show that only inland transit has a robust negative and signi�cant effect on trade.21 Moreover, the magnitude from column (2) is similar to the result using all 17. From column 7, the aggregate coef�cient on transit is 2 0.086 ( 2 0.070 þ 2 0.016) and the coef�cient on tariffs is 2 0.048. Thus, the effect of 3.5 days is 3.5* 2 0.086 ¼ 2 0.30, while the effect of a 5 percentage point cut in tariffs is 5*.048 ¼ 2 0.24. 18. There are two differences with this study. First, Djankov, Freund, and Pham use a difference gravity equation on similar exporters. This strategy is not necessary for this investigation since in the regressions only sub-Saharan Africa countries are considered. Second, while they use the actual times in the transit country as instrument, here the time for the transit county’s trade in the transit country is used. 19. Since this regression is perfectly identi�ed, it is not possible to test whether the excluded instruments are not correlated with the error terms. In table A2 of the Appendix the regressions of table 4 are replicated but this time including separate instruments for customs and ports. In this case the Sargan overidenti�cation test supports the validity of the instruments. 20. This variable is computed taking into account geography and type of road. 21. This result does not only reflect the fact these countries are more isolated. Even though delays in inland transport are higher with respect to coastal countries (15 days versus 4 days on average), delays in documentation (24 days on average) and customs and ports (9 days on average) procedures are even higher for exporters in landlocked countries. 376 T A B L E 5 . The Effect of Export Time Components on Aggregate Exports, Landlocked Sample Regression OLS IV IV IV IV IV IV Dependent Variable: (levels) (levels)a (levels)a (logs)a (logs)a (levels)b (logs)b ln (Aggregate exports) (1) (2) (3) (4) (5) (6) (7) Inland transit time 2 0.125*** 2 0.097*** 2 0.082*** 2 1.788*** 2 1.453** 2 0.126*** 2 2.212*** [0.015] [0.020] [0.031] [0.501] [0.726] [0.033] [0.530] Customs and ports 2 0.252*** 0.113 0.066 0.027 2 0.169 0.208 1.193 [0.053] [0.190] [0.228] [1.706] [2.259] [0.204] [3.287] Documents time 2 0.047*** 0.012 0.019 2 0.412 2 0.083 0.053 0.818 [0.012] [0.047] [0.051] [1.054] [1.281] [0.052] [2.087] GDP 0.389** 0.120 0.700* 0.042 0.670** 0.355 0.433* THE WORLD BANK ECONOMIC REVIEW [0.178] [0.257] [0.370] [0.218] [0.329] [0.311] [0.254] POP 0.555** 2 0.322 2 0.493 2 0.051 2 0.245 2 0.836 2 0.512 [0.249] [0.519] [0.654] [0.424] [0.612] [0.610] [0.800] Distance 2 1.102*** 2 0.940*** 2 1.386*** 2 0.995*** 2 1.388*** 2 1.441*** 2 1.466*** [0.293] [0.284] [0.353] [0.288] [0.347] [0.361] [0.357] Tariffs (simple av.) 2 0.055* 2 0.054* 2 0.070** 2 0.070** [0.030] [0.029] [0.029] [0.030] Observations 991 991 479 991 479 479 479 R-squared 0.569 0.542 0.544 0.563 0.560 0.524 0.541 Notes: 1. Robust standard errors in brackets. ***p , 0.01, **p , 0.05, *p , 0.1. 2. Other control variables: partner FE, common language, common border, colony, remoteness, landlocked aid for trade and GSP (whenever import tariffs are included in the regressions). a. Instruments: documents, customs and ports and inland transit times in transit countries. b. Instruments: documents and customs and ports times and GPS distance from border between the landlocked and the transit country and the port of the latter. Source: Authors’ analysis based on data sources discussed in the text. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Freund and Rocha 377 T A B L E 6 . The Effects of Export Time Components on Time Sensitive Products (OLS regression) Countries exporting at Countries exporting least one product 70% of the products Dependent Variable (levels) (logs) (levels) (logs) ln (Aggregate Exports by industry) (1) (2) (3) (4) Inland transit time*Time sensitivity 2 0.023* 2 0.173* 2 0.038** 2 0.226** [0.013] [0.090] [0.016] [0.101] Customs and ports time*Time sensitivity 2 0.002 0.003 2 0.002 0.039 [0.022] [0.191] [0.024] [0.208] Docs time*Time sensitivity 0.014 0.247 0.015 0.219 [0.009] [0.187] [0.009] [0.196] K abundance*Canned product 0.513** 0.544** 0.689*** 0.715*** [0.222] [0.222] [0.233] [0.231] Observations 626 626 519 519 R-squared 0.519 0.518 0.542 0.541 Notes: Robust standard errors in brackets ***p , 0.01, **p , 0.05, *p , 0.1. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Source: Authors’ analysis based on data sources discussed in the text. countries and new trade (column (3), Table 4): a one day reduction in transit delays leads to about 9 percent more exports.22 Finally, the effects of documentation, inland transit and customs and ports times on the exports of time-sensitive products are analyzed. All time delays should have a greater effect on exports of time-sensitive goods. To examine the extent to which they are hampered, the methodology in Djankov, Freund and Pham (2010) is followed and an estimation of a difference-in-difference gravity equation is performed using trade data of agricultural ( processed and unpro- cessed) products for which time matters the most and the least. This method- ology reduces the endogeneity problem coming from reverse causality because it controls for country- and industry-�xed effects. In addition, the products considered account for only a small fraction of trade in agricultural goods on average (less than 10 percent) so it is unlikely that they have a large impact on establishing trade facilitation processes. The de�nition of time-sensitive agricultural products is based on the infor- mation of their storage life (Gast 1991), which includes a range of products going from a minimum storage life of 2 weeks or less, such as apricots, beans, currants, and mushrooms to 4 weeks or longer, for example apples, cranberries and potatoes and canned products. Goods with a very long storage life such as dry fruits with a maximum storage life of between 6 months and one year and canned products with a storage life ranging from 1 to 5 years, depending on 22. When the log of total time is included as the only trade cost variable in the regression, the estimated coef�cient is about 2 1, the same as the results for developing countries in Djankov, Freund and Pham (2010), although they use a slightly different approach. 378 THE WORLD BANK ECONOMIC REVIEW the good’s acidity are also included. To measure time sensitivity, the inverse of the median storage life of each product is used. To study the joint effect of industry time-sensitivity and country-time delays on exports the following difference-in-difference gravity regression is estimated LnExportsik ¼ ai þ ak þ b1 ðTime Sensitk Þ Â ðInland transit timei Þ þ b2 ðTime Sensitk Þ Â ðCustoms & Ports timei Þ ð3Þ þ b3 ðTime Sensitk Þ Â ðDocs timei Þ þ b4 ðK abundancei Þ Â ðcanned productk Þ þ 1ik , where ai and ak represent country- and industry-�xed effects. The coef�cients b1, b2 and b3 capture the joint effect of time-sensitive products and time delays in inland transit, customs and ports, and documentation on export values. The term (Kabundancei)  (canned productk) is introduced to control for the fact that more capital abundant countries are more likely to have the necessary resources and technologies to process fresh food into canned products. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 With this speci�cation it is tested whether exports of time-sensitive goods are more responsive to time delays in each of the key components of time to export than exports of time-insensitive products. For example, if b1 is negative and signi�cant, it means that a longer transit delay reduces time-sensitive exports more than time-insensitive exports. The key advantage of this approach is that it controls for industry- and country-�xed effects, so signi�cant results cannot be a result of more trade being associated with more ef�cient trade facilitation, and shorter delays. The simple correlation between the share of time-sensitive goods23 in trade and transit times is low (0.02), so it also cannot be due to countries with greater exports of time-sensitive goods having lower times. Results for time-sensitive agricultural products controlling for countries’ capital abundance are presented in Table 6. Results are reported for countries exporting at least one product and countries exporting at least 70 percent of the products, and also with the variables in logs and levels. In all cases, the coef�cient on the interaction term of inland transit times with time sensitivity is negative and signi�cant, and highly signi�cant when intensive exporters are considered. This implies that an increase in inland transit times reduces exports of time-sensitive goods relatively more than time-insensitive goods. In contrast, interactions with documents and customs and ports times are never signi�cant. Transit delays affect the composition of trade, preventing countries from exporting time-sensitive agricultural goods. Time-sensitive goods also tend to have higher value, implying that some of the effects of inland transit delays on aggregate exports results from countries with poor trade-facilitation programs concentrating on low-value, time-insensitive goods. 23. To calculate the share, products with a storage life of 2 weeks or less are classi�ed as time-sensitive. Freund and Rocha 379 In sum, three different ways to examine the effects of various trade delays on trade flows are used, each of which should reduce the endogeneity problem inherent in the analysis. All three point to the same conclusion: delays during inland transit affect trade flows to a much greater extent than delays because of documentation or at the port. These results imply that reducing time spent on inland transit will signi�cantly stimulate trade in Africa. I V. W H Y D O E S I N L A N D T R A N S I T M A T T E R M O R E ? All else equal, a one day delay should affect exports the same way no matter when it occurs. However, one reason it may not is if there is more uncertainty associated with high delays in some procedures than in others. Uncertainty will reduce exports because it makes delivery deadlines harder to meet. In this section it is investigated whether greater uncertainty related to inland transport times makes costs related with documents, customs and ports become a sec- ondary priority relative to travel costs for existing exporters. The effects of time uncertainty in each component of export times are esti- Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 mated for a sub-sample of 22 Sub-Saharan countries for which there is infor- mation on the maximum and the average number of days it takes for an exporter to complete each of the exporting procedures:24 LnExportsij ¼ b1 Inland transit time uncerti þ b2 Customs & Ports time uncerti þ b3 Docs time uncerti þ b4 Ln GDPi þ b5 Ln Popi ð4Þ þ b6 LnDistij þ b7 Ln Remotei þ b8 Landlocki þ Xij þ mj þ 1ij Time uncertainty is de�ned as the difference between the maximum time and the average time it takes to conclude each of the different phases representing the total time to export. Unfortunately, since estimates are only from two or three freight forwarders in each country, it is not possible to use more sophisticated measures of uncertainty like the standard deviation of times in the country. Results from Table 7 show a negative and signi�cant impact of inland transit time uncertainty on trade values, with a one day increase in this variable leading to a reduction of exports of more than 15 percent (column (1)). Or in logs (column (2)), a one percent increase in uncertainty leading to about a one percent reduction in exports. In contrast, uncertainty in the other variables is not robustly signi�cant in reducing exports. The coef�cient on documentation time uncertainty is negative and signi�cant only when considered in levels and its magnitude is much smaller compared to inland transit uncertainty. Ports ˆ te d’Ivoire, Ghana, 24. Benin, Botswana, Burkina Faso, Cameroon, Republic of the Congo, Co Kenya, Madagascar, Malawi, Mali, Mauritania, Mozambique, Namibia, Nigeria, Rwanda, Sierra Leone, South Africa, Tanzania, Uganda and Zambia. 380 T A B L E 7 . The Effect of Time Uncertainty on Aggregate Exports, Full Sample OLS Regression Dependent Variable: Levels Logs Logs Levels Logs Levels Logs ln (Aggregate Exports) (1) (2) (3) (4) (5) (6) (7) Inland transit time uncertainty 2 0.154*** 2 0.978*** 2 0.104*** 2 0.467*** [0.021] [0.131] [0.022] [0.149] Ports and customs time uncertainty 0.021 0.399** 2 0.066 [0.018] [0.179] [0.115] Documentation time uncertainty 2 0.017** 2 0.174 [0.008] [0.115] GDP 1.655*** 1.623*** 1.482*** 1.377*** 1.319*** 1.147*** 1.116*** THE WORLD BANK ECONOMIC REVIEW [0.091] [0.091] [0.087] [0.091] [0.100] [0.071] [0.071] Population 2 0.736*** 2 0.662*** 2 0.410*** 2 0.286** 2 0.252** 2 0.058 2 0.125 [0.137] [0.136] [0.129] [0.121] [0.117] [0.108] [0.107] Distance 2 1.332*** 2 1.305*** 2 1.329*** 2 1.411*** 2 1.439*** 2 1.428*** 2 1.464*** [0.206] [0.204] [0.205] [0.189] [0.188] [0.177] [0.177] Inland transit Time 2 0.079*** 2 0.676*** 2 0.117*** 2 0.972*** [0.017] [0.137] [0.015] [0.104] Observations 1,602 1,602 1,668 1,713 1,713 1,713 1,713 R-squared 0.607 0.608 0.591 0.603 0.606 0.598 0.604 Notes: 1. Robust standard errors in brackets. ***p , 0.01, **p , 0.05, *p , 0.1. 2. Other control variables: partner FE, common language, common border, colony, remoteness, landlocked aid for trade. Data for uncertainty is available for 22 countries. See footnote 24. Source: Authors’ analysis based on data sources discussed in the text. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Freund and Rocha 381 and customs time uncertainty shows a positive and signi�cant coef�cient only when entered in logs. This counterintuitive result might be due to the presence of multicollinearity with other time uncertainty variables. In fact, when the regression is performed including only customs and ports uncertainty, its coef�- cient becomes negative and insigni�cant (column (3)). These results imply that high uncertainty in road transport times jeopardizes delivery targets. In addition, even if documentation requirements take more time than inland transit, they can either be done in advance or there may be learning by doing, such that exporters become more familiar with the procedures and uncertainty is limited. Finally, while exporters may be able to pay in the port to get things out more quickly, nothing can be done on the road. In columns (4) and (5) both inland travel times and inland travel uncertainty are included in the regression. The coef�cients reflecting both variables are signi�- cant. When only inland transit is included (see columns (6) and (7)) in the same sample the coef�cient is larger (as compared with columns (4) and (5)), implying that part of the effect of transit time on exports stems from uncertainty. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 V. G E O G R A P H Y V E R S U S I N S T I T U T I O N S The dominant role of transit suggests that geography is the main component of trade facilitation. But transit is about more than geography. It is also about road quality, fleet class and competition, border delays, road security and other institutional issues. Next it is investigated how much of the transit effect is pure geography and how much is institutional. Speci�cally, to control for dom- estic geography the GPS estimated distance and time based solely on geography and type of road is used. The regressions include the road distance in km from the principal city to the port of export (which is the relevant distance for which transport is calculated in the data). In addition, GPS-estimated total travel time is included. This variable is calculated as the total time it takes to get from the principal city to the port of exit by assuming a speed of 40 km per hour for unpaved roads and 80 km per hour for paved surfaces.25 If transit is primarily a geography effect then the GPS variables should pick up such effect. Indeed, both GPS variables are highly correlated with inland transport (0.71 for distance and 0.82 for time, Table 2). In addition, the difference between the Doing Business time and the GPS time is included, which should reflect institutions, since the GPS records how long it should take if there are no truck problems, road problems, borders, etc. The results using the full sample are reported in Table 8. The �rst column is for reference. The second and third columns include GPS distance and GPS time. Both coef�cients are negative, but only distance is signi�cant. Neither alters the effect of inland transit time on exports. 25. No information on road condition is used in the calculation of GPS travel time. Furthermore, delays at the border (or otherwise) are not included. 382 T A B L E 8 . Geography versus Institutions Dependent Variable: Levels Levels Levels Levels Logs ln (Aggregate Exports) (1) (2) (3) (4) (5) Inland transit time (levels) 2 0.089*** 2 0.087*** 2 0.080*** [0.012] [0.013] [0.014] GDP 1.079*** 1.111*** 1.121*** 1.121*** 1.071*** [0.054] [0.057] [0.056] [0.056] [0.056] POP 2 0.108** 2 0.120** 2 0.209*** 2 0.209*** 2 0.064 [0.055] [0.055] [0.057] [0.057] [0.060] Distance 2 1.166*** 2 1.157*** 2 1.226*** 2 1.226*** 2 1.226*** THE WORLD BANK ECONOMIC REVIEW [0.131] [0.131] [0.134] [0.134] [0.134] GPS distance principal city to port (km) 2 0.073*** [0.022] GPS travel times city to port (days) 2 0.110 2 0.190 2 0.465*** [0.133] [0.126] [0.126] DB times - GPS times 2 0.0800*** 2 0.803*** [0.014] [0.111] Observations 3,494 3,494 3,347 3,347 3,347 R-squared 0.514 0.515 0.518 0.518 0.523 Notes: 1. Robust standard errors in brackets. ***p , 0.01, **p , 0.05, *p , 0.1. 2. Other control variables: partner FE, common language, common border, colony, remoteness, landlocked aid for trade. Source: Authors’ analysis based on data sources discussed in the text. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Freund and Rocha 383 Next, the geographical and institutional components of transit are explored separately. Speci�cally, the GPS transit time (geography) and the difference between Doing Business transit time and GPS transit time (institutions: DB times – GPS times) are introduced. The latter should reflect road delays that are not due to geography and distance. Column (4) shows the results. Only the institutional component is signi�cant. Finally, Column (5) includes both vari- ables in logs. In this case both the geographical and institutional parts of transit times are signi�cant, but the coef�cient on the institutional component is signi�cantly larger (F-Stat ¼ 3.46 of test that they are equal, p-value 0.06). In sum, the results imply that the distance from city to port and whether roads are paved are not the main reason for long delays in transit. There might be other factors such as the quality of the roads and vehicles, accidents, compe- tition in trucking, road blocks or border waiting times which affect the total time for an exporter to get his goods form the factory to the port of exit. This is good news in the sense that these institutional aspects of transit are likely to be more amenable to change than geographical ones. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 VI. CONCLUSION Export performance in Africa has been poor. It has been argued that this is a result of relatively slow income growth in the region. But that explanation is not entirely satisfactory, as export growth surely contributes to income growth. It could just as convincingly be argued that the failure of African exports to surge—as they did in the fast-growing developing regions over this period—is the cause of the lacklustre income growth. Moreover, even controlling for income and population, exports in Africa fall short of expectations. In this article an attempt is made to understand one of the important constraints to Africa’s exports. Detailed data on key components on the time it takes to move containerized products from the factory gate to the ship is used to estimate whether and how diverse trade costs affect export volumes in Sub-Saharan Africa. An augmented gravity equation is estimated by regressing aggregate bilateral exports on differ- ent time delay components such as inland transit, documentation, ports and customs, and other standard gravity variables. To control for the possibility that more trade leads to improved trade facili- tation, the effects that documentation, inland transport, customs and ports times respectively have on the exports of new products is analyzed. Exports in these products are unlikely to have an impact on the historical development of infrastructure. As a robustness check an instrumental variables approach is also used to examine the effect of time trade costs in transit countries on the exports of landlocked countries. Finally, a “difference-in-difference� regression is estimated on a sub-sample of agricultural products to determine whether trade costs affect exports of time-sensitive and time-insensitive goods, ranging 384 THE WORLD BANK ECONOMIC REVIEW from perishable products where time is most critical relative to preserved goods such as tinned food, differently. Our results imply that while inland transit delays have a robust negative impact on export values, higher times in other areas have much smaller effects in reducing Africa’s exports. A one day increase in inland transit time reduces exports by 7 percent on average. Put another way, a one day reduction in inland travel times translates into nearly a 2 percentage point decrease in all importing-country tariffs. In addition, this effect is higher for time-sensitive goods compared to time-insensitive goods. It is shown that long times are associated with high uncertainty in road transport, which jeopardizes expor- ters’ delivery targets. The empirical results have important policy implications. Export tariffs in Sub-Saharan African countries are already at a very low level. Furthermore these countries have preferential access to markets such as the United States and the European Union. Hence, while the bene�ts from a further decrease in tariffs among trading partners might be very small – or even negative in terms Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 of preference erosion if tariff reductions are MFN – reducing transport times will signi�cantly increase their exports. Trade facilitation programs should therefore prioritize those programs directly affecting truck fleets and the infra- structure and security of Sub-Saharan Africa’s road systems. REFERENCES Acemoglu, D., S. Johnson, and J.A. Robinson (2001). “Reversal of Fortune: Geography and Institutions in the Making of the Modern World Income Distribution,� NBER Working Papers 8460. Anderson, J.E., and E. Van Wincoop (2003). “Gravity with Gravitas�. American Economic Review, vol. 93(1), pp. 170–192. Bernard, A., and B. Jensen (1995). “Exporters, jobs, and wages in U.S. manufacturing: 1976–1987� Brookings Papers on Economic Activity. Microeconomics, pp. 67– 119. Bernard, A., B. Jensen, S. Redding, and P. Schott (2007). “Firms in International Trade� Journal of Economic Perspectives, vol. 21(3), pp. 105– 130. Bhagwati, J. (1996). “The Demands to Reduce Domestic Diversity among Trading Nations� in J. BhagwatiR. Hudec (eds) Fair Trade and Harmonization. Prerequisites for free Trade. Cambridge, Massachusetts; London. The MIT press. Coe, D., and A. Hoffmaister (1999). “North-South Trade: Is Africa Unusual?� Journal of African Economies, vol. 8(20), pp. 228 –256. Djankov, S., C. Freund, and C.S. Pham (2010). “Trading on time,� Review of Economics and Statistics, vol. 92(1), pp. 166–173. Gast, K. (1991) “Postharvest Management of Commercial Horticultural Crops� Kansas State University Agricultural Experiment Station and Cooperative ExtensionService. Hall, R., and C. Jones (1999). “Why Do Some Countries Product so Much More Output per Worker than Others?� Quarterly Journal of Economics, vol.114 (1), pp. 83– 116. Head, K. (2003). “Gravity for Beginners� Mimeo University of British Columbia, Vancouver, B.C. Hummels, D. (2001). “Time as a Trade Barrier�. GTAP Working Papers 1152, Center for Global Trade Analysis, Department of Agricultural Economics, Purdue University. Freund and Rocha 385 Hummels, D., and A. Skiba (2004). “A Virtuous Circle: Regional Trade Liberalization and Scale Economies in Transport� in A. Estevandeordal, and D. Rodrik (eds) FTAA and Beyond: Prospects for Integration in the America. Harvard University Press. Krueger, A.O. (1998). “Why Trade Liberalization is Good for Growth� in The Economic Journal, vol.108 (450), pp. 1513–1522. McArthur, J.W., and J.D. Sachs (2001). “Institutions and Geography: Comment on Acemoglu, Johnson and Robinson (2000)� NBER Working Papers N. 8114. OECD and World Trade Organization (2009). Aid for Trade at a Glance 2009: Maintaining the Momentum. Paris: OECD and Geneva: WTO. Portugal, A., and J. Wilson (2009). “Why Trade Facilitation Matters to Africa� World Bank, Policy Research Working Paper, 4719. Rodrik, D. (1997). “Trade Policy and Economic Performance in Sub-Saharan Africa� Mimeo, Harvard University. Schank, T., C. Schnabel, and J. Wagner (2007). “Do exporters really pay higher wages? First evidence from German linked employer–employee data� Journal of International Economics, vol. 72(1), pp. 52– 74. Winters, L.A. (2004). “Trade Liberalization and Economic Performance: An Overview� Economic Journal, vol. 114 (493), pp. F4– F21. World Bank (2007). Doing Business. World Bank, Washington DC, www.doingbusiness.org. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 APPENDIX T A B L E A 1 . Summary Results for First Stage Regressions Partial R2 F statistic p-value Inland transit time (levels) 0.4123 90.98 0.000 Customs and ports time (levels) 0.3529 70.70 0.000 Documents time (levels) 0.3741 77.50 0.000 Inland transit time (logs) 0.3007 55.75 0.000 Customs and ports time (logs) 0.4607 110.78 0.000 Documents time (logs) 0.3895 82.72 0.000 Source: Authors’ analysis based on data sources discussed in the text. T A B L E A 2 . IV Landlocked Sample Regressions using Customs and Ports as Separate Instruments Dependent variable: Levelsa Levelsa Logsa Logsa Levelsb Logsb ln (Aggregate exports) (1) (2) (3) (4) (5) (6) Inland transit time 2 0.096*** 2 0.096*** 2 1.789*** 2 1.640** 2 0.124*** -2.176*** [0.018] [0.028] [0.484] [0.713] [0.030] [0.478] Customs and ports 0.084 0.229 2 0.007 1.374 0.177 0.637 [0.104] [0.148] [1.372] [1.867] [0.152] [1.288] Docs time 0.005 0.055 2 0.434 0.698 0.043 0.444 [0.025] [0.033] [0.845] [1.080] [0.029] [0.529] GDP 0.153 0.468* 0.042 0.566* 0.392 0.439* (Continued ) 386 THE WORLD BANK ECONOMIC REVIEW TABLE A2. Continued Dependent variable: Levelsa Levelsa Logsa Logsa Levelsb Logsb ln (Aggregate exports) (1) (2) (3) (4) (5) (6) [0.177] [0.278] [0.212] [0.323] [0.264] [0.247] POP 2 0.251 2 0.907* 2 0.044 2 0.587 2 0.757 2 0.388 [0.328] [0.489] [0.361] [0.545] [0.495] [0.425] Distance 2 0.934*** 2 1.396*** 2 0.994*** 2 1.420*** 2 1.439*** 2 1.457*** [0.281] [0.362] [0.285] [0.354] [0.358] [0.348] Tariffs (simple av.) 2 0.062** 2 0.063** 2 0.070** 2 0.069** [0.030] [0.029] [0.029] [0.028] Observations 991 479 991 479 479 479 R-squared 0.546 0.520 0.564 0.540 0.531 0.554 p-value of Sargan 0.859 0.368 0.977 0.188 0.820 0.864 statistic Notes: 1. Robust standard errors in brackets. ***p , 0.01, **p , 0.05, *p , 0.1. 2. Other control variables: partner FE, common language, common border, colony, remoteness, land- Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 locked aid for trade and GSP (whenever import tariffs are included in the regressions). a. Instruments: documents, customs, ports and inland transit times in transit countries. b. Instruments: documents customs and ports times and GPS distance from border between the landlocked and the transit country and the port of the latter. Source: Authors’ analysis based on data sources discussed in the text. Does the Internet Reduce Corruption? Evidence from U.S. States and across Countries Thomas Barnebeck Andersen, Jeanet Bentzen, Carl-Johan Dalgaard, and Pablo Selaya We test the hypothesis that the Internet is a useful technology for controlling corruption. In order to do so, we develop a novel identi�cation strategy for Internet diffusion. Power disruptions damage digital equipment, which increases the user cost of IT capital, and thus lowers the speed of Internet diffusion. A natural phenomenon causing power disruptions is lightning activity, which makes lightning a viable instru- ment for Internet diffusion. Using ground-based lightning detection censors as well as global satellite data, we construct lightning density data for the contiguous U.S. states and a large cross section of countries. Empirically, lightning density is a strong instru- ment for Internet diffusion and our IV estimates suggest that the emergence of the Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Internet has served to reduce the extent of corruption across U.S. states and across the world. JEL Classi�cation codes: K4, O1, H0 Corruption is commonly perceived to be a major stumbling block on the road to prosperity. Aside from retarding growth (Mauro 1995), corruption entails �scal leakage, which reduces the ability of poor countries to supply essential Thomas Barnebeck Andersen (corresponding author) is professor at the Department of Business and Economics, University of Southern Denmark, Campusvej 55, DK-5230 Odense M. Jeanet Bentzen is Ph.D. student, Carl-Johan Dalgaard is professor, and Pablo Selaya is assistant research professor at the Department of Economics, University of Copenhagen, Studiestraede 6, DK-1455 Copenhagen K, Denmark. Contact: Thomas Barnebeck Andersen (barnebeck@sam.sdu.dk), Jeanet Bentzen ( jeanet. bentzen@econ.ku.dk), Carl-Johan Dalgaard (carl.johan.dalgaard@econ.ku.dk), and Pablo Selaya ( pablo.selaya@econ.ku.dk). We owe a special thanks to William A. Chisholm for generously sharing his expertise on the subject of lightning protection. We also thank three anonymous referees, Phil Abbott, Shekhar Aiyar, Mark Bills, Christian Bjørnskov, Matteo Cervellati, Areendam Chanda, Oded Galor, Chad Jones, Sam Jones, Phil Keefer, Pete Klenow, David Dreyer Lassen, Ross Levine, Norman Loayza, Ben Olken, Ola Olsson, Elena Paltseva, Nancy Qian, Martin Ravallion, Yona Rubinstein, Paolo Vanin, Dietz Vollrath, David Weil, Nadja Wirz, and seminar participants at the University of Bergen, Brown University, University of Copenhagen, the Jerusalem Summer School in Economic Growth at Hebrew University, ETH Zu ¨ rich, the ZEW Conference on Economics of Information and Communication Technologies, the 2008 Nordic Conference in Development Economics, the Growth Workshop at the 2008 NBER Summer Institute, the Southern Economic Association Annual Meeting 2008, the Latin American Econometric Society Meeting 2009, and the World Bank. Errors are ours. THE WORLD BANK ECONOMIC REVIEW, VOL. 25, NO. 3, pp. 387 –417 doi:10.1093/wber/lhr025 Advance Access Publication May 27, 2011 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 387 388 THE WORLD BANK ECONOMIC REVIEW public services such as schooling and health care (Reinikka and Svensson 2004; World Development Report 2004). Corruption is unquestionably a governance failure one would like to dispose of, yet combating it has not proven easy. In the present paper we hypothesize that the Internet is a useful technology for combating corruption around the world. We test this hypothesis using data for the 48 contiguous U.S. states as well as for a cross section of countries. Our estimates support the proposition that the Internet has worked to reduce corruption since its inception. There are several reasons why the Internet could serve as a corruption- reducing technology. First, the Internet is among a select group of innovations that are considered General Purpose Technologies (e.g., Jovanovic and Rousseau 2005). General Purpose Technologies are fundamental and pervasive innovations that (with time) hold a �rst order impact on economic growth. Insofar as economic growth works to lower corruption, the Internet is likely to act as a corruption suppressor due to its positive impact on growth.1 Second, rapid technological change usually encourages investment in human capital, which may instigate lower levels of corruption.2 Accordingly, human capital Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 accumulation is another potential transmission channel linking Internet diffu- sion and reductions in corruption levels. Third, the World Wide Web is a major source of information. Spreading information about of�cial wrong-doing inevitably increases the risk of detection for politicians and public servants, thus making corrupt behavior less attractive.3 Fourth, the Internet is the chief vehicle for the provision of E-government worldwide (West 2005). By allowing citizens access to government services online, E-government obviates bureau- crats’ role as intermediaries between the government and the public, thus limit- ing the interaction between potentially corrupt of�cials and the public. Moreover, online systems require standardized rules and procedures. This reduces bureaucratic discretion and increases transparency as compared to the 1. See Andersen et al. (2010) for evidence that the Internet has stimulated growth across U.S. states, and e.g. Gundlach and Paldam (2009) for evidence of a causal impact of income on corruption. 2. See Schultz (1975) and Foster and Rosensweig (1996) for evidence of a positive impact of technological change on the return to human capital investments, and Glaeser and Saks (2006) for evidence on the impact of human capital on corruption. 3. A nice illustration of this mechanism at work is found in a 2001 scandal from India, which nearly toppled the government. Reporters from the online news site , www.Tehelka.com . posed as arms dealers and documented negotiations with top politicians and bureaucrats over the size of required side payments to get a contract; in some instances the reporters even got the delivery of the bribe on camera. Consequently, numerous politicians and top of�cials had to resign, chief among them the defence minister George Fernandes. See The Sting That Has India Writhing by Celia W. Dugger, The New York Times (March 16, 2001). Andersen, Bentzen, Dalgaard and Selaya 389 arbitrariness available to civil servants when dealing with the public on a case-by-case basis.4 The present paper examines the reduced form effect of the Internet on corruption levels across the U.S. and across the World. Thus, our analysis addresses the key issue of whether the Internet has had a causal impact on corruption. At the same time, the analysis admittedly does not clarify the exact mechanism(s) through which the Internet affects corruption; it may be either one of the mechanisms mentioned above, or some combination of them. The Internet is a new technology. Indeed, if we identify the Internet with the World Wide Web, it only emerged in 1991.5 To examine the impact of the Internet we therefore estimate the impact of changes in Internet users on changes in corruption levels from the early 1990s to 2006. Using data for the 48 contiguous U.S. states as well as cross-country data we establish a strong partial correlation between the rate of changes in Internet users and the evol- ution of corruption, consistent with the proposed hypothesis. However, since the speed of Internet diffusion is likely endogenous, OLS estimates might be Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 misleading. In an effort to establish causality we develop an identi�cation strategy designed to isolate exogenous variation in the speed of Internet diffusion. The theory underlying our instrument of choice is the following. Computer equip- ment is highly sensitive to power disruptions: power surges and sags lead to equipment failure and damage. Consequently, a higher frequency of power disruptions implies higher costs of IT equipment, either through elevated IT capital depreciation or due to incurrence of additional costs in order to protect equipment from power disruptions. Frequent power disturbances are also likely to reduce the productivity of IT capital (or its marginal bene�t), as power disturbances produce downtime, generate data glitches, etc. A natural phenom- enon which produces power disturbances is lightning activity. In fact, one third of all power disruptions in the U.S. are related to lightning activity. We there- fore hypothesize that higher lightning intensity is a viable candidate instrument for Internet diffusion. Using state level measures of lightning density (ground strikes per square km per year) and global satellite data on lightning activity, 4. The celebrated Bhoomi program (located in the state of Karnataka in India) constitutes a good example of the effectiveness of E-government in limiting the interface between civil servants and the public. Starting in 1998 the program aimed to computerize land records, and at the time of writing more than 20 million landholdings belonging to the state’s 6.7 million landowners have been registered. Before online registration was available citizens had to seek out village accountants to register, a process which involved considerable delays and the need for bribes to be paid; a typical bribe could range from Rs.100 to Rs. 2,000 (US $2 to $40) (Bhatnager 2003). With the online system there is no longer a need for the of�cial middlemen, implying that the online system probably has saved locals hundreds of millions of Rs. in bribes. See also Andersen (2009). 5. TheWorld Wide Web was launched in 1991 by CERN (the European Organisation for Nuclear Research). See Hobbes’ Internet Timeline v8.2 , http://www.zakon.org/robert/internet/timeline/ . . In this paper, we de�ne the Internet/WWW as the network of networks using the TCP/IP/HTTP protocols, which was spawned by the launch of WWW. 390 THE WORLD BANK ECONOMIC REVIEW we establish its influence on Internet diffusion: areas (states and countries) with a higher flash density have experienced a slower speed of Internet adoption. A concern of �rst-order importance is whether, conditional on Internet diffusion, lightning density acts as a stand-in for factors that are correlated with corruption. To �x ideas, suppose areas with high lightning density just happen to be characterized by an abundance of natural resources as well, which influences corruption in its own right (i.e., conditional on Internet penetration). If so, then lightning is not a valid instrument for the Internet. Notice, however, that the proposed instrument is most likely to be (spur- iously) correlated with factors that are exerting a time persistent effect on cor- ruption. To stay with the example from before: lightning prone areas that were rich in natural resources in (say) 1970 most likely remain so, and if the pres- ence of natural resources fueled corruption in the 1970s, it would probably also spur corruption at the beginning of the 21st century. Accordingly, if light- ning density is picking up this sort of influence we would expect to see a fairly Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 time invariant correlation between lightning and (changes in) corruption over the last three or four decades. However, if lightning density is picking up the influence from processes instigated by the emergence of the Internet (be that a growth surge, an acceleration in human capital accumulation, improved diffu- sion of information, or E-government), one would expect to see a time varying correlation between lightning and (changes in) corruption. In fact, instrument validity would require that there is no correlation between lightning density and changes in corruption before 1991 (the founding year of the World Wide Web). These considerations form the basis of a falsi�cation test: If lightning is correlated with (changes in) corruption prior to 1991, it is unlikely to satisfy the exclusion restriction. As documented below, lightning density exhibits a time-varying correlation with corruption; the reduced form relationship between lightning density and corruption does not exist prior to the inception of the World Wide Web. Using U.S. state-level data, as well as cross country data, we are able to establish that the lightning/corruption correlation only exists after 1991. This falsi�cation test makes probable that the lightning instrument is capturing processes insti- gated by the emergence of the Internet, and it therefore supports the use of lightning as an instrument for Internet diffusion. Against this background we employ lightning density as an instrument for Internet diffusion. Our 2SLS estimates corroborate the OLS results: Rising Internet use over the 1990s reduced corruption in the U.S. and across countries. Our results are admittedly stronger, statistically speaking, for the U.S. Still, our cross-section analysis does suggest that the U.S. state-level results generalize to an international setting. The present research is related to the literature which studies the determi- nants of the level of corruption. Notable contributions include Ades and di Tella (1999), Treisman (2000), Brunetti and Weder (2003), Persson, Tabellini, Andersen, Bentzen, Dalgaard and Selaya 391 and Trebbi (2003), Glaeser and Saks (2006), and Licht, Goldscmidt, and Schwartz (2007). Since the Internet is a central source of information, the paper is also related to the political economy literature that studies the impact of information on governance more generally. This literature suggests that a better informed public serves to discipline the political establishment, thus affecting governance (e.g. Besley and Burgess 2002; Stro ¨ mberg, 2004; Reinikka and Svensson 2004; Eisensee and Stro ¨ mberg 2007; Ferraz and Finan 2008). Finally, the paper is related to the literature which studies the determinants of the spread of the personal computer (e.g. Caselli and Coleman 2001) and the Internet (e.g. Chinn and Fairlie 2007) across countries. We add to this literature by documenting a link between lightning density and the speed of Internet diffusion. The paper is structured as follows. In Section I we present our empirical speci�cations of choice. Section II outlines the identi�cation strategy in detail; in particular, we explain how lightning activity impacts digital equipment. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Section III provides an analysis of how the Internet has affected corruption across the U.S. states, whereas Section IV provides cross-country evidence. Section V concludes. I. SPECIFICATION As argued above, we wish to understand the impact of the Internet (or World Wide Web) on the evolution of corruption, from the emergence of the former and onwards. Since the Internet is a General Purpose Technology it seems infeasible to try and isolate any particular mechanism linking the Internet to corruption (e.g., growth spurts, human capital, enhanced dissemination of information, etc.), for which reason we focus on the reduced form. Accordingly, in the analysis to follow we mainly rely on the following parsi- monious speci�cation: DCi ¼ a0 þ a1 DINTERNETi þ a2 Cinitial; i þ 1i ; ð1Þ where DCi is the change in corruption levels between an initial and a �nal year, C final; i À Cinitial; i , and DINTERNETi is the change in Internet penetration, INTERNET final; i À INTERNETinitial; i .6 When DCi . 0 means increasing cor- ruption, the key hypothesis under investigation is whether a1 , 0 or not. There are several reasons why we adopt this lagged dependent variables spe- ci�cation as our preferred speci�cation. First, corruption is a naturally bounded variable. In the U.S. setting we employ corruption convictions as our measure of C, which is bounded from below at zero. Similarly in the 6. In the cross-state sample we also condition on state population size. State population enters as a control because we use total corruption convictions as our measure of state corruption; by including state size we thereby ensure that all scale effects are pruned from the data in a simple way. 392 THE WORLD BANK ECONOMIC REVIEW cross-country analysis we employ a corruption index (the ICRG index), which is con�ned to a particular interval (zero to six) by construction. Obviously, for all country observations near either of the two endpoints, and in states near the zero boundary, one should expect the evolution of corruption to be subject to mean reversion. Unless we thus control for Cinitial; i , our empirical model is likely to be misspeci�ed. To spell it out formally, assume that equation (1) is the true population model, and suppose that Covð1; DINTERNET Þ ¼ 0. If we mistakenly ignore the lagged dependent variable and instead focus on the �rst difference equation, in an attempt to kill off an imagined �xed effect, we would perform OLS on DCi ¼ a0 þ aFD 1 DINTERNETi þ yi , where yi ; a2 Cinitial; i þ 1i . The probability limit of a1 obtained via the �rst differenced estimator is CovðCinitial ; DINTERNET Þ plim aFD 1 ¼ a1 þ a2 : ð2Þ VarðDINTERNET Þ Unless CovðCinitial ; DINTERNET Þ ¼ 0 we have that aFD 1 = a1 . Indeed, Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 a priori aFD 1 % 0 would be a likely outcome. To see this, note that the mean reversion process implies that a2 , 0. Moreover, if places with high initial corruption levels tend to see a slower diffusion of the internet then CovðCinitial ; DINTERNET Þ , 0. Taken together this implies that a2 � CovðCinitial ; DINTERNET Þ . 0, which implies that the expected negative point estimate for a1 is biased towards zero (see equation (2)). Second, another virtue of including the initial level of corruption is that it automatically controls for (a potentially large set of ) variables which may influence the evolution of corruption. To see the latter point more clearly observe that equation (1) is equivalent to a levels regression with a lagged dependent variable: C final; i ¼ a0 þ a1 DINTERNETi þ ða2 þ 1Þ Cinitial; i þ 1i : ð3Þ Accordingly, all time invariant structural characteristics affecting the level of corruption will be picked up by Cinitial; i. This reduces the scope for omitted variable bias in contaminating the estimate of a1 (Wooldridge 2003, p. 300).7 Third, the core part of our analysis involves 2SLS estimation, where lightning density is invoked as an instrument for DINTERNETi . Since lightning density and the level of corruption are highly correlated, one may harbor legitimate concerns about the exclusion restriction if the latter is omitted from the empirical model; if the evolution of corruption is subject to mean reversion (i.e., if a2 , 0 ), trouble arises (see again equation (2)). Naturally, one might also worry that the exclusion restriction is jeopardized by the omission of �xed effects in equation (1). But in the analysis below we 7. See also Angrist and Pischke (2009, Ch. 5.3), for a discussion of the virtues of using the ` -vis the �rst difference model. speci�cation in equation (1) vis-a Andersen, Bentzen, Dalgaard and Selaya 393 are able to gauge the likely severity of this problem by performing the falsi�ca- tion test mentioned above: If the true empirical model involves �xed effects (i.e., a0 in equation (1) should be country speci�c), and if our instrument is correlated with these unobservables, we would expect to see that lightning density is correlated with DCi before as well as after 1991 (the founding year of the World Wide Web). As shown below, however, we are unable to reject the absence of a correlation between our instrument and changes in corruption prior to 1991. This holds true across U.S. states as well as across countries. As a result, we view the lagged dependent variables speci�cation as the relatively safe choice when trying to elicit information about the causal impact of the Internet on corruption. Nevertheless, we will in the interest of completeness report the results from the �rst difference speci�cation as well; that is, we also report the results from assuming a2 ¼ 0 in equation (1). I I . I D E N T I F I CAT I O N S T R AT E GY Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 There is good reason to believe that the adoption of new technologies, such as the Internet, is endogenous to governance. New technologies may create politi- cal as well as economic losers, for which reason incumbent entrepreneurs and politicians may try to block adoption (Mokyr 1990; Parente and Prescott 1999; Acemoglu and Robinson 2001). It seems plausible that places with wide- spread corruption, for example, may have adopted the Internet later, due to the influence of politicians, civil servants, or both. This mechanism rationalizes a positive impact of governance on the number of Internet users. Consequently, OLS estimation is unlikely to identify the impact of the Internet on corruption. To address this concern we employ an IV approach, the logic of which we now describe. Computers are highly sensitive to even ultra brief power disruptions. Such disruptions are likely to cause down-time, though sudden power surges may also damage the equipment and randomly destroy or alter data. As observed in The Economist:8 “For the average computer or network, the only thing worse than the electricity going out completely is power going out for a second. Every year, millions of dollars are lost to seemingly insigni�cant power faults that cause assembly lines to freeze, computers to crash and networks to collapse.� The reason why IT equipment is so sensitive to power disturbances is that computers are constructed to work under a clean electrical current, featuring a particular frequency and amplitude of voltage. The alternating power emanat- ing from the commercial power plant is converted into direct current, after which transistors turn this small voltage on and off at several gigahertz during digital processing (Kressel 2007). However, if the input, in the form of the alternating current, is disturbed or distorted the conversion process is 8. The power industry’s quest for the high nines, The Economist, March 22, 2001. 394 THE WORLD BANK ECONOMIC REVIEW corrupted, which may in turn result in equipment failure and damage. Indeed, voltage disturbances measuring less than one cycle are suf�cient to crash servers, computers, and other microprocessor-based devices; that is, at a 60 Hz frequency (the standard in the U.S.) this means that a power disturbance of a duration less than 1/60 th of a second is enough to crash a computer (Yeager and Stalhkopf 2000; Electricity Power Research Institute 2003). Importantly, this issue is unlikely to diminish over time as the sensitivity to small power distortions increases with the miniaturization of transistors, which is the key to increasing speed in microprocessors (Kressel 2007). Accordingly, in areas with more power disturbances, the user cost of IT capital will be higher due to a higher rate of IT capital depreciation (Hall and Jorgenson 1967). By implication, the desired IT capital stock will be lower, reducing IT investments and the speed of Internet diffusion. Of course, steps may be taken to protect the equipment from power disturbances. A high- quality surge protector provides protection against voltage spikes, for example. High-tech companies install generators to supplement their power needs, Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 thereby insuring themselves against power failure. They also add uninterrupti- ble power sources relying on batteries to power computers until generators kick in. However, these initiatives will in any case increase the costs of acquir- ing digital equipment, and thereby the user cost of IT capital. The crux of the matter is that if one lives in an environment with low power quality, this adds to the costs of a computer.9 To this one may add that in areas with frequent power disruptions and outages, the marginal bene�t of owning a computer is probably lowered as well. Obviously, in countries where �rms and consumers face regular power outages it will be dif�cult to employ IT ef�ciently. But even if power disrup- tions are infrequent and of very short duration, power disruptions lead to glitches and downtime which serves to lower the productivity of IT equipment. Both mechanisms, higher marginal costs and lower marginal return/bene�t, imply that poorer power quality should lead to a slower speed of Internet diffusion. Naturally, power quality is not exogenous; it may well be determined by governance.10 As a result, we employ a variable which generates exogenous variation in power quality, and thus IT costs and bene�ts, as an instrument for Internet diffusion. A natural phenomenon which interferes with digital equip- ment, by producing power failures, is lightning activity (e.g., Shim et al. 2000, Ch. 2; Chisholm 2000). By all accounts, the influence of lightning on power quality is substantial. According to some estimates, lightning is the direct cause 9. Besides, the above mentioned protective devises are not necessarily enough to ensure against damage. According to the National Oceanic and Atmospheric Administration (NOAA), a typical surge protector will not protect equipment from a nearby lightning strike. Generators, in turn, do not react fast enough and can deliver dirty power; batteries are expensive to maintain and may also not react fast enough. See e.g., The power industry’s quest for the high nines, The Economist (March 22, 2001). 10. See Fredriksson et al. (2004) for evidence that corruption affects energy supply. Andersen, Bentzen, Dalgaard and Selaya 395 of one third of all power quality disturbances in the United States (Chisholm and Cummins 2006). Moreover, the probability of lightning-caused power interruptions or equipment damage scales linearly with lightning density (Chisholm 2000; Chisholm and Cummins 2006).11 As a result, in areas with greater lightning density (strikes per square km per year) the (expected) rate of IT capital depreciation will tend to be larger. This implies higher IT investment costs, and possibly lower IT productivity as well. The problems associated with lightning activity in the context of IT equip- ment have not escaped the attention of the popular press. A recent article in The Wall Street Journal highlights the practical relevance of lightning activity, and stresses the dif�culty in shielding IT equipment:12 “Even if electricity lines are shielded, lightning can cause power surges through unprotected phone, cable and Internet lines - or even through a building’s walls. Such surges often show up as glitches. "Little things start not working; we see a lot of that down here," says Andrew Cohen, president of Vertical IT Solutions, a Tampa information-technology consulting �rm. During the summer, Vertical gets as Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 many as 10 calls a week from clients with what look to Mr. Cohen like lightning-related problems. Computer memory cards get corrupted, servers shut down or �rewalls cut out.� Against this background we propose lightning density as an instrument for the speed at which Internet use per capita changed over the period in question. Schematically, we can express the theory underlying our identi�cation strategy in the following way: LIGHTNING DENSITY ! POWER DISTURBANCES ! INTERNET USE; ð4Þ where the second arrow implicitly subsumes the impact of power disturbances on the costs and bene�ts of IT capital. Lightning is certainly external in the sense of Deaton (2010). But this, of course, does not imply that it ful�lls the exclusion restriction required for instrument validity. In particular, it could conceivably correlate with �xed factors (natural resource endowment, say) which themselves exert a persistent effect on the evolution of corruption. In order to examine whether this is likely to be the case or not, we perform a falsi�cation test below: Under the hypoth- esis that lightning influences the evolution of corruption, via the Internet, changes in corruption should only be correlated with lightning after the emer- gence of the World Wide Web. We can in fact reject that lightning and changes in corruption are correlated prior to the invention of the World Wide Web 11. This linear scaling can be expressed precisely. LetNS denote the number of strikes to a conducter per 100kmof power line length,hthe average height (in meters) of the conducter above ground level, andGFD the ground flash density, then NS ¼ 3.8 . GFD . h 0.45 (see Chisholm 2000). 12. There Go the Servers: Lightning’s New Perils. The Wall Street Journal, August 25, 2009. 396 THE WORLD BANK ECONOMIC REVIEW (i.e., prior to 1991) using data for both the U.S. States and our world sample. This suggests quite strongly that lightning is not – spuriously – correlated with country speci�c factors that persistently affect the path of corruption, thus supporting the exclusion restriction. I I I . C RO S S - S TAT E E V I D E N C E The cross-state analysis proceeds as follows: We �rst provide details on the data used for the analysis. Next we provide evidence on the partial correlation between changes in Internet usage and changes in corruption levels. Finally, before proceeding to our 2SLS estimates, we provide an independent check of the validity of our identi�cation strategy. Data In measuring corruption in the U.S. we follow Glaeser and Saks (2006) by Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 employing corruption convictions. The data derives from the Justice Department’s Report to Congress on the Activities and Operations of the Public Integrity Section, a publication which provides statistics on the nation- wide federal effort against public corruption, including the number of federal, state, and local public of�cials convicted of a corruption-related crime by state. As argued by Glaeser and Saks (2006), federal conviction levels capture the extent to which federal prosecutors have charged and convicted public of�cials for misconduct. There are potential problems with using conviction rates to measure corruption: in corrupt places, the judicial system is itself likely to be corrupt, meaning that fewer people will be charged with corrupt practices. This problem, however, is diminished when using federal convictions, the reason being that the federal judicial system is somewhat isolated from local corruption. Consequently, it should treat people similarly across states (Glaeser and Saks, 2006). In concrete terms, we measure the change in corruption between 1991 and 2006 by the log difference in (one plus) the total number of corruption convic- tions in the two years; positive values in the rate of change reflect increasing corruption, and the choice of initial year follows from the fact that the Internet (in the sense of the World Wide Web) was introduced in 1990 (�rst Web page went online in 1991).13 The second key variable is Internet use, which we measure as the percentage of households with Internet access. It is based on data collected in a sup- plement to the October 2003 Current Population Survey (CPS), which includes 13. Note that an increased use of the Internet will both increase the risk of detection for a corrupt of�cial (the detection technology is improved) as well as lower the incentive to commit corrupt acts. Hence, in theory, increased Internet use could increase the number of convictions if the former effect dominates. It might thus seem as if the Internet increases corruption. However, empirically the net effect is negative, implying that the incentive effect dominates, as documented below. Andersen, Bentzen, Dalgaard and Selaya 397 questions about computer and Internet use.14 The CPS is a multi-stage prob- ability sample with coverage in all states. The sample was selected from the 1990 Decennial Census �les and is continually updated to account for new resi- dential construction. To obtain the sample the United States is divided into 2,007 geographic areas, and about 60,000 households are eligible for inter- views. Since U.S. corruption data goes back to 1991, the launch date of the WWW, we de�ne the change in Internet use by state population as DINTERNETi ¼ INTERNETi;2003 À INTERNETi;1991 ¼ INTERNETi;2003 , since INTERNETi; 1991 ¼ 0 for all i (i.e., for all states). In some speci�cations we also control for initial state population, or the growth of state population, which derives from the U.S. Bureau of Economic Analysis (BEA). The natural log of initial state population is denoted log(POP1991) and population growth is D log(POP) ¼ log(POP2006 /POP1991). The purpose of including these controls is to make sure that all scale effects are pruned from the corruption data. Obviously, the total number of corruption convictions is likely greater in more populous states. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Finally, we measure lightning as the number of ground strikes per square kilometer per year. Accurate cloud-to-ground data exist for the 48 contiguous U.S. states, are measured by the U.S. National Lightning Detection Network (NLDN), and are provided by Vaisala for the period 1996-2005. NLDN con- sists of numerous remote, ground-based lightning sensors, which instantly detect (at a very high level of accuracy) the electromagnetic signals given off when lightning strikes Earth’s surface.15 Summary statistics are reported in Appendix Table A.1. Partial Correlations Table 1 documents the partial correlation between changes in Internet use and changes in corruption for the 48 contiguous U.S. states. As a �rst step we report, in Columns 1 and 2, the results from the �rst differ- ence speci�cation where we thus omit initial corruption as a control; Column 1 is the basic speci�cation, whereas population growth is included in Column 2 alongside Internet users. States that have seen larger increases in Internet use have also experienced larger reductions in corruption convictions, consistent with a corruption dampening effect of the Internet. As explained in Section 2 we are concerned about the exclusion of the initial corruption level in the �rst difference speci�cation, as changes in corruption convictions likely are subject to mean reversion. Speci�cally, in states with zero initial convictions one would almost inevitably expect an increase in convic- tions subsequently, regardless of Internet use. In order to check whether this is an actual problem in the U.S. sample, we experiment, in Columns 3 and 4, with omitting states with zero corruption 14. , http://www.census.gov/population/socdemo/computer/2003/tab01B.xls . . 15. The data can be obtained, free of charge, here: , http://www.vaisala.com/thunderstorm . 398 THE WORLD BANK ECONOMIC REVIEW T A B L E 1 . OLS Regressions, U.S. Sample Dependent variable: Change in corruption convictions 1991-2006 (1) (2) (3) (4) (5) (6) DINTERNET 2 0.058** 2 0.057** 2 0.060** 2 0.057* 2 0.058** 2 0.051*** (0.027) (0.028) (0.029) (0.030) (0.024) (0.019) Dlog(POP) 2 0.243 2 0.914 (1.068) (0.972) log(POP1991) 0.849*** (0.145) log(1 þ CC1991) 2 0.386*** 2 0.873*** (0.096) (0.135) Constant 3.667** 3.655** 3.577** 3.543** 4.392*** 2 7.790*** (1.448) (1.469) (1.584) (1.578) (1.312) (2.201) Observations 48 48 40 40 48 48 R-squared 0.09 0.09 0.11 0.13 0.29 0.55 Notes: Robust standard errors in paranthesis. Asterisks ***, ** and * signify, respectively, p-value , 0.01, p-value , 0.05 and p-value , 0.1. All variables are de�ned in the main text. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Source: Authors’ analysis based on data described in the text. convictions in 1991. This should in theory reduce the extent of mean reversion in the sample. Further, in Columns 5 and 6, we employ the lagged dependent variable speci�cation on the full sample: our preferred speci�cation. As seen from Columns 5 and 6, initial corruption is a signi�cant correlate with subsequent changes in corruption; it adds considerably to the explanatory power of the regression model, as evidenced by the increase in R 2 when com- paring Columns 1 and 2 with Columns 5 and 6.16 In Columns 3 and 4 we truncate the sample by excluding states with zero initial corruption. If one includes initial corruption in this setting, it is insigni�cant (not shown). This shows that the signi�cance of initial corruption in Columns 5 and 6 is due to mean reversion in the U.S. sample. Nevertheless, despite the presence of mean reversion in corruption we obtain nearly identical results for DINTERNET in all columns. The reason is simply that, statistically speaking, CovðCinitial ; DINTERNET Þ ¼ 0 in the U.S. sample. As a result, by equation 17 (2), the �rst difference estimator is consistent; i.e., plim aFD1 ¼ a1 : In any event, the bottom line is that the partial correlation is robust to the exact choice of speci�cation. The partial correlation is also robust to a long list of additional controls beyond what is reported in Table 1. Maxwell and Winters (2005) propose four sets of fundamental traits of U.S. states, which should have predictable effects 16. We have also experimented with including the growth rate of population, rather than its level, in the lagged dependent variables speci�cation (Column 5 and 6), but much like in Column 2 and 4 it comes out insigni�cant. 17. By including the lagged corruption level, we can, however, reduce the error variance. Andersen, Bentzen, Dalgaard and Selaya 399 on corruption: (i) number of corruptible government bodies, (ii) the size of the state, (iii) socio-ethnic homogeneity, and (iv) civic-minded and well-informed political cultures. The authors consider seven additional control variables including income growth, general tax revenue and campaign expenditure restrictions. We have experimented with the inclusion of all the proposed cor- ruption determinants (alongside initial corruption), and we �nd that changes in Internet use (1991–2006) remains signi�cant. The results are available upon request. Despite these encouraging results concerns about causality may legitimately be raised. We therefore further scrutinize the Internet/corruption nexus by way of instrumental variables estimation. Instrument Falsi�cation Test 2SLS estimates are obviously no better than the invoked instrument. In the present context we propose the use of lightning density as an instrument for Internet use, based on the theoretical argument we presented in Section II. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 According to the identi�cation strategy lightning is only allowed to influence changes in corruption via Internet use. If, by contrast, it is correlated with unobserved determinants of the evolution of corruption, the exclusion restriction fails, and the 2SLS estimates are no better than the OLS estimates discussed above. Under the null of instrument validity, however, lightning should exert a time varying impact on the path of corruption. The reason is simply that the Internet is a new technology which cannot possibly have affected corruption prior to its inception and widespread use. By extension, determinants of Internet use should not affect corruption prior to the emergence of the World Wide Web. Hence, in an effort to try to falsify our instrument we run the following regression: DCi ¼ a0 þ a1 logðLIGHTNINGi Þ þ a2 Cinitial; i þ 1i ; ð5Þ on two different time periods. First, we examine the link between lightning and changes in corruption during the Internet era: 1991 to 2006. In this period we expect a signi�cantly positive point estimate for lightning, suggesting that high lightning states have experienced slower Internet diffusion and therefore smaller reductions in corruption levels than states with less lightning. Second, we examine the link between lightning and changes in corruption prior to the emergence of the World Wide Web. For instrument validity, the partial corre- lation between lightning and corruption should not be signi�cantly different from zero. In the event of a signi�cant correlation between lightning and changes in corruption we are forced to conclude that lightning is correlated with factors beyond the Internet (and the initial level of corruption), which influences changes in corruption. If so, the instrument is invalid. 400 THE WORLD BANK ECONOMIC REVIEW The falsi�cation test is potentially helpful in another respect: If changes in corruption are not correlated with lightning strikes in the pre-Internet age, the scope for lightning operating on corruption via other electronic technologies is limited. Indeed, it would appear unlikely that the impact date of such technol- ogies coincides with the inception of the WWW. The introduction of the microprocessor, for instance, goes back to the early 1970s. By that time com- puters were widely used by the U.S. government. Therefore, if we think that increased and more ef�cient information storage capacity, say, reduces the scope for corruption, we should observe a correlation between lightning and corruption in the pre-Internet era. To the extent that this relationship is absent, the microprocessor per se is unlikely to influence the evolution of corruption. Table 2 reports the results from this check. In Columns 1-4 we examine the pre-Internet periods 1976–1990, where we have missing observations for cor- ruption in three states, and 1978–1990, where we have a full sample. The take- away is that lightning is not correlated with changes in corruption during these sub-periods, in keeping with the requirement that lightning is uncorrelated Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 with other factors which exert a persistent impact on the evolution of corrup- tion. Figure 1 provides a visual illustration of these �ndings. In Columns 5 and 6 we shift attention to the period following the emergence of the World Wide Web. During this period lightning emerges as a strong cor- relate with changes in corruption levels: In states with more lightning we observe a slower rate of reduction in corruption compared to states with less lightning. This is consistent with the hypothesis that the Internet has worked to lower corruption, and that lightning has worked to slow down Internet diffu- sion. Figure 2 provides a visual illustration of the partial correlation between lightning and changes in corruption 1991–2006. It is important to observe that the insigni�cance of lightning during the pre-Internet era is not simply a matter of imprecise estimates. As is evident from inspection of the results across columns, the size of the estimate itself rises many fold when moving from Columns 1 –4 to Columns 5 –6. It may also be noted that the falsi�cation test is consistent with lightning not operating through other electronic technologies, which would arguably have required that the correlation between lightning and corruption was in existence before 1991. Overall, we �nd that these results constitute compelling evidence in favor of the exclusion restriction in the U.S. context. Hence, we now move to instru- mental variables estimation where lightning is used as an instrument for Internet diffusion. 2SLS Estimates Table 3 reports our 2SLS results; Panel A reports �rst stage results, whereas Panel B reports the second stage. In keeping with the approach taken in Table 1, we experiment with both the �rst difference speci�cation and the lagged dependent variables speci�cation. In the OLS setting (Table 1) we found T A B L E 2 . Falsi�cation test, U.S. Sample Change in corruption convictions Change in corruption convictions Change in corruption convictions 1976– 1990 1978– 1990 1991– 2006 Dependent variable: (1) (2) (3) (4) (5) (6) log(LIGHTNING) 2 0.068 2 0.027 0.019 0.016 0.344** 0.278** (0.183) (0.110) (0.198) (0.113) (0.145) (0.109) log(1 þ CC1976) 2 0.388*** 2 0.954*** (0.143) (0.162) log(POP1976) 0.908*** (0.185) log(1 þ CC1978) 2 0.352** 2 0.930*** (0.170) (0.185) log(POP1978) 0.908*** (0.153) log(1 þ CC1991) 2 0.452*** 2 0.916*** (0.109) (0.142) log(POP1991) 0.833*** (0.147) Constant 1.654*** 2 11.220*** 1.469*** 2 11.304*** 0.787** 2 10.723*** (0.313) (2.733) (0.358) (2.235) (0.347) (2.051) Observations 45 45 48 48 48 48 R-squared 0.15 0.49 0.10 0.44 0.29 0.54 Notes: Robust standard errors in paranthesis. Asterisks ***, ** and * signify, respectively, p-value , 0.01, p-value , 0.05 and p-value , 0.1. All variables are de�ned in the main text. Source: Authors’ analysis based on data described in the text. Andersen, Bentzen, Dalgaard and Selaya 401 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 402 THE WORLD BANK ECONOMIC REVIEW F I G U R E 1 . Reduced form in the pre-Internet era, U.S. sample Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Note: The �gure shows the association between lightning and changes in corruption convictions over the 1978– 1990 period, with the influence of initial population partialled out. Source: Figure is based on data described in the text. F I G U R E 2 . Reduced form in the Internet era, U.S. sample Note: The �gure shows the association between lightning and changes in corruption convictions over the 1991– 2006 period, with the influence of initial population partialled out. Source: Figure is based on data described in the text. virtually no difference in estimation results using either the �rst differenced or the lagged dependent variables speci�cation. This is not true in the 2SLS setting. In the �rst stage results are very similar across speci�cations. Lightning proves to be a strong instrument in explaining changes in Internet use; the �rst T A B L E 3 . 2SLS regressions, U.S. sample Panel A: First stage Dependent variable: Change in Internet users 1991– 2006 (1) (2) (3) (4) (5) (6) log(LIGHTNING) 2 3.583*** 2 3.505*** 2 3.821*** 2 3.764*** 2 3.801*** 2 3.805*** (0.604) (0.606) (0.756) (0.758) (0.651) (0.678) Dlog(POP) 3.443 3.947 (5.099) (5.935) log(POP1991) 0.050 (0.794) log(1 þ CC1991) 0.721 0.693 (0.526) (0.703) Constant 60.54*** 59.83*** 61.41*** 60.70*** 59.55*** 58.86*** (1.080) (1.336) (1.403) (1.547) (1.378) (10.870) F-test (H0: log(lightning) ¼ 0) 35.23 33.49 25.54 24.66 34.11 31.48 Panel B: Second stage Dependent variable: Change in corruption convictions 1991– 2006 DINTERNET 2 0.058 2 0.056 2 0.104*** 2 0.102*** 2 0.090** 2 0.073** (0.043) (0.046) (0.035) (0.037) (0.036) (0.030) Dlog(POP) 2 0.248 2 0.606 (1.117) (1.026) log(POP1991) 0.836*** (0.146) log(1 þ CC1991) 2 0.386*** 2 0.866*** (0.097) (0.137) Constant 3.670 3.626 5.959*** 5.928*** 6.171*** 2 6.416** (2.343) (2.413) (1.924) (1.924) (1.949) (2.708) Observations 48 48 40 40 48 48 R-squared 0.09 0.09 0.05 0.07 0.26 0.54 Andersen, Bentzen, Dalgaard and Selaya Notes: Robust standard errors in paranthesis. Asterisks ***, ** and * signify, respectively, p-value , 0.01, p-value , 0.05 and p-value , 0.1. All vari- ables are de�ned in the main text. Source: Authors’ analysis based on data described in the text. 403 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 404 THE WORLD BANK ECONOMIC REVIEW stage F-statistic exceeds 10 in all columns, documenting that the 2SLS analysis in the U.S. sample is not plagued by weak identi�cation (cf., Staiger and Stock 1997). In the second stage results differ markedly across speci�cations (Panel B). In the �rst differenced speci�cation Internet use is estimated imprecisely, with a point estimate virtually identical in size and sign to what we found in the OLS setting. In contrast, when we exclude either states with zero initial corruption (Columns 3 and 4), or when we control for initial corruption levels directly (Columns 5 and 6), changes in Internet use is estimated with high precision. What should we make of the differences in size and signi�cance of the Internet across speci�cations? The most straightforward explanation runs as follows: Recall from Section 1 that, under the maintained hypothesis that equation equation (1) is the true speci�cation, the �rst difference equation is given by DCi ¼ a0 þ aFD 1 DINTERNETi þ yi , where yi ; a2 Cinitial; i þ 1i . When we employ 2SLS with lightning as an instrument, the exclusion restriction is that CovðLIGHTNING; yÞ ¼ 0; since CovðLIGHTNING; 1Þ ¼ 0 by assumption, Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 CovðLIGHTNING; yÞ ¼ 0 requires that CovðLIGHTNING; Cinitial Þ ¼ 0. This latter condition is violated in the U.S. sample. The failure of the exclusion restriction implies that the 2SLS estimates are biased towards OLS, which explains the similarity of the results reported in Columns 1-2 in Tables 1 and 3. In contrast, if we either eliminate the influence from initial corruption on changes in corruption by way of sample truncation (i.e., omission of states with zero initial corruption), or by including the initial level of corruption directly, the exclusion restriction is much more likely to be satis�ed. The results reported in Columns 3-6 are therefore much more likely to identify the causal influence of Internet diffusion on changes in corruption. In all four columns where the exclusion restriction is plausible we �nd evidence that the Internet has worked to lower corruption in states where its diffusion was rapid, com- pared to states where Internet diffusion occurred at a slower rate. What is the economic signi�cance of the Internet? To answer this question we begin with the levels speci�cation associated with equation (3): logð1 þ C2006 Þ ¼ a0 þ a1 DINTERNET þ ða2 þ 1Þ logð1 þ C1991 Þ: ð6Þ If we linearize, treating C1991 as a constant, the following simple approxi- mation emerges: DC2006 % a1 ð1 þ C2006 Þ INTERNET2003 ; ð7Þ where we have used that DINTERNET ¼ INTERNET2003 . Next, to gauge economic signi�cance, consider moving from the median to the third quartile in the distribution of Internet users in 2003; this is equivalent to an increase of 3.1 Internet users per 100 people. Using the 2SLS results reported in Column 6 of Table 3 (the most conservative estimate) in equation (7) we �nd that DC2006 % ðÀ0:073Þ � ð1 þ 16Þ � ð3:1Þ ¼ À3:85 yearly convictions. This would Andersen, Bentzen, Dalgaard and Selaya 405 correspond to moving from the median to the 33rd percentile in the U.S. state corruption convictions ranking in 2006. Accordingly, our results suggest that the introduction of the Internet has reduced U.S. corruption levels below what would otherwise have been observed absent this technology. I V. C R O S S - C O U N T R Y E V I D E N C E In this Section we examine whether the results obtained above generalize to a cross-country sample. After providing details on the data used in the cross- country setting, we discuss the partial correlation between changes in corrup- tion and changes in Internet users, present an instrument falsi�cation test, and discuss our 2SLS estimates. Data As our main measure of corruption we employ the well-known ICRG index. This index has the useful property that it is available from 1984, which enables Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 us to perform the type of instrument falsi�cation test employed in the context of the U.S. state-level analysis. The coverage for the pre-Internet era is shorter (six years, 1984-1990) than in the U.S. state-level analysis but hopefully long enough for us to clarify whether lightning is correlated with changes in corrup- tion prior to the emergence of the World Wide Web. Changes in the ICRG index are calculated as the absolute change in the index between relevant years (1991-2006 and 1984-1990, respectively). The indicator is bounded between zero and six, with larger values of the index meaning less corruption. As a matter of robustness we have also checked our main results (OLS and 2SLS) using the control of corruption index compiled by Kaufmann et al. (2007). This indicator is often used in cross-national studies (Treisman 2007). But it is unfortunately only available for the period 1996 onwards. Accordingly, the falsi�cation test of our instrument cannot be performed with this indicator, for which reason we have chosen to stick with the ICRG index. But it should be noted that our OLS and 2SLS results for the Internet era are robust to the use of the control of corruption indicator; these results are avail- able upon request. Our key explanatory variable is the number of Internet users per 100 people. Increasingly, the number of Internet users is based on regular surveys. In situations where surveys are not available, an estimate can be derived based on the number of subscribers. Data is compiled by the International Telecommunication Union (ITU) and made available in the World Development Indicators (WDI) 2007. Changes in Internet users are calculated as in the context of our U.S. state-level analysis. When we examine the period 1991-2006, using the ICRG indicator, we use the approximation that INTERNETi; 1991 ¼ 0, as in the analysis of the U.S. states. In the U.S. context data on lightning density derived from ground detectors. Naturally, such data is not available across the world. Instead we employ 406 THE WORLD BANK ECONOMIC REVIEW satellite data on lightning intensity. The raw data (strikes per km 2 per year) is provided by the National Aeronautics and Space Administration (NASA). Speci�cally, we rely on the data from the so-called Optical Transient Detector (OTD), a space based sensor launched on April 3, 1995. For a period of roughly 5 years the satellite orbited Earth once every 100 minutes at an altitude of 740 km. At any given instant it viewed a 1300 km  1300 km region of Earth. Lightning is determined by comparing the luminance of adjoining frames of OTD optical data. When the difference was larger than a speci�ed threshold value, an event was recorded.18 These satellite-based data are archived and cataloged by the The Global Hydrology and Climate Center, where they are also made publicly available.19 We apply the data from a high- resolution (0.5 degree latitude  0.5 degree longitude) grid of total lightning bulk production across the planet, expressed as a flash density, from the com- pleted 5 year OTD mission.20 Figure 3 provides a world map of the average flash density over the 5 years period. We construct average flash densities for each country by �rst mapping the Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 corresponding geographic areas into the lightning data grid and then taking the average of flash densities within each of these areas. The coordinates describing the areas are taken from the GEOnet Names Server (GNS) at the U.S. National Geospatial-Intelligence Agency’s (NGA),21 and the U.S. Board on Geographic Names’ (U.S. BGN) database of foreign geographic names and features.22 We used the GNS database released on October 7, 2008.23 The GNS data covers the entire planet with the exception of the U.S. and the Antarctica. The area for the U.S. was estimated using geographic features for the 48 contiguous U.S. states, contained in the database released on August 15, 2008 by the Geographic Names Information System at the U.S. BGN.24 A potential problem with the OTD data is that it only provides observations on total lightning events; i.e., intra-cloud, cloud-to-cloud, cloud-to-sky, and cloud-to-ground lightning. In other words, OTD data does not separate out cloud-to-ground lightning incidences. The pertinent characteristic of lightning in the evaluation of risk to electronic equipment and electric power systems is the cloud-to-ground flash density. In an effort to examine whether the satellite data is likely to be a good proxy for ground strikes we compared our data on ground strikes for 18. Basically, these optical sensors use high-speed cameras designed to look for changes in the tops of clouds. By analyzing a narrow wavelength band (near-infrared region of the spectrum) they can spot brief lightning flashes even under daytime conditions. 19. , http://thunder.msfc.nasa.gov/data/#OTD_DATA . 20. , ftp://microwave.nsstc.nasa.gov/pub/data/lightning-satellite/lis-otd-climatology/HRFC/ LISOTD_HRFC_V2.2.hdf . 21. , http://earth-info.nga.mil/gns/html/name�les.htm . 22. , http://geonames.usgs.gov/domestic/download_data.htm . 23. , ftp://ftp.nga.mil/pub2/gns_data/geonames_dd_dms_date_20081007.zip . 24. , http://geonames.usgs.gov/docs/stategaz/NationalFile_20080815.zip . Andersen, Bentzen, Dalgaard and Selaya 407 F I G U R E 3 . Global lightning map Note: Average flash density (flashes per year per km2). The �gure is constructed using the OTD Global Lightning Distributions for the period April 12, 1995 to December 31, 1999. Source: Figure is based on data described in the text. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 individual U.S. states with corresponding satellite-based lightning data. Figure 4 provides a scatter plot of total lightning against cloud-to-ground light- ning for the U.S. sample. Note that by de�nition total lightning should be at least as large as cloud-to-ground lightning, which is con�rmed by the �gure; all observations are below the 45 degree line.25 Figure 4 shows that there is agreement between the two measures of light- ning. With a correlation above 0.95, total lightning appears to be a reasonable proxy for cloud-to-ground lightning in the contiguous U.S. states sample. This is also expected to be the case in the cross-country data (Chisholm and Cummins, 2006). Partial Correlations Table 4 reports the results from OLS estimation using cross-country data. As in Table 1 we experiment with various speci�cations: In Column 1 we use the �rst difference speci�cation; in Column 2 we truncate the sample to limit the extent of mean reversion by excluding observations within one standard deviation from either endpoint of the 0-6 “ICRG interval�; in Column 3 we estimate the lagged dependent variables model; and �nally in Column 4 we report the results from estimating our preferred model by way of an outlier robust estimator (Least Absolute Deviations, LAD). Compared to the state-level analysis we �nd more variation in results when we move from the �rst difference speci�cation to the lagged dependent vari- ables speci�cation. In Column 1 changes in Internet penetration is 25. NASA’s flash densities of total lightning are calculated for the 1995-1999 period, while Vaisala’s cloud-to-ground measures are calculated for the 1996-2005 period. In addition, Vaisala uses mile 2 as the area unit; these where converted into km 2 by dividing by the mile 2 numbers by 1.6092. 408 THE WORLD BANK ECONOMIC REVIEW F I G U R E 4 . Total lightning versus cloud-to-ground lightning, U.S. sample Note: The �gure shows a scatter plot of total lightning flash rate density (flashes per km2 per Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 year) based on the Optical Transient Detector high resolution data (horizontal axis) versus cloud-to-ground flash rate density (flashes per km2 per year) based on the U.S. National Lightning Detection Network (vertical axis). The full line is the 45-degree line. Number of observations is 48. Source: Figure is based on data described in the text. insigni�cantly correlated with changes in the ICRG index during the Internet era, whereas a positive correlation emerges when we employ the lagged depen- dent variables approach, thus suggesting a corruption suppressing influence of the Internet. It is worth observing that the initial ICRG level is highly signi�- cant and contributes considerably to the overall �t, as is evident from the increase in R 2 going from Column 1 to 3. The results from Column 2, where we truncate the sample, are somewhere in between. But even when we truncate the sample markedly to suppress mean reversion, the initial corruption level, ICRG in 1991, remains signi�cant if included (not shown but results are avail- able upon request); the partial correlation is weaker, as it should be in theory, but it is not rendered insigni�cant. These results demonstrate that the evolution of corruption is subject to mean reversion in the world sample; perhaps even to a greater extent than in the U.S. sample, since even sample truncation is not suf�cient to fully eliminate the correlation between initial corruption and subsequent changes in corrup- tion. Now using both mean reversion (a2 , 0) and the fact that a high ICRG score in 1991 ( low initial corruption) is associated with faster Internet diffu- sion (i.e., CovðCinitial ; DINTERNET Þ . 0) in equation (2) immediately tells us that the �rst difference model is misspeci�ed in the world sample; that is, since plim aFD1 , a1 the �rst difference model underestimates a1. This is of course exactly what we observe in Table 4: the parameter estimate for DINTERNET rises monotonically when we go from Column 1 to 3. Andersen, Bentzen, Dalgaard and Selaya 409 T A B L E 4 . OLS and LAD regressions, world sample Change in ICRG 1991– 2005 Dependent variable: (1) (2) (3) (4) DINTERNET 0.004 0.0138* 0.0361*** 0.0352*** ( 2 0.003) ( 2 0.007) ( 2 0.006) ( 2 0.006) ICRG1991 2 0.747*** 2 0.772*** ( 2 0.089) ( 2 0.097) Constant 2 0.923*** 2 1.044*** 0.981*** 1.072*** ( 2 0.140) ( 2 0.145) ( 2 0.235) ( 2 0.281) Observations 102 67 102 102 R-squared 0.01 0.04 0.57 Estimator OLS OLS OLS LAD Notes: Robust standard errors in paranthesis. Asterisks ***, ** and * signify, respectively, p-value , 0.01, p-value , 0.05 and p-value , 0.1. In column 2 the sample is truncated; only countries with ICRG scores one standard deviation away from the ICRG index minimum (0) and maximum (6) value are included. Larger values for the ICRG index means less corruption. In column 4 the standard errors are bootstrapped with 1000 repetitions. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Source: Authors’ analysis based on data described in the text. There is a straightforward explanation of why the strong correlation between initial corruption and Internet diffusion, which we observe in the world sample, has no parallel in the U.S. sample. In the world sample, factors influencing the level of corruption, or the level of corruption per se, probably also influence the process of Internet diffusion; perhaps the over-all insti- tutional quality of a nation influences not only corruption and its evolution, but also the speed of technology adoption. This seems plausible, and it can account for the positive correlation between the ICRG index and Internet dif- fusion in the world sample. It appears reasonable to expect that similar forces are not at play in the U.S. sample of relatively more homogenous states, for which reason a similar correlation between state-level corruption and state-level Internet use is absent. Therefore, in the world sample, but not in the U.S. sample, the end result of (erroneously) omitting the level of initial corruption is that the OLS estimate for DINTERNET is biased towards zero. The omitted variables bias is dampened when we truncate the sample (Column 2), as the correlation between the level of corruption and changes in corruption is reduced; the parameter estimate for DINTERNET therefore rises in absolute value. In sum, the lagged dependent variables speci�cation should be con- sidered the appropriate empirical model. In an effort to check the robustness of the correlation we also report, in Table 4, the results from estimating our preferred model by way of an outlier robust estimator. As can be seen, comparing Column 3 and 4, the results are virtually the same whether we use OLS or LAD, implying that outliers are not carrying the correlation. This conclusion is further reinforced by invoking 410 THE WORLD BANK ECONOMIC REVIEW the Hadi (1992) outlier detection procedure; no observations are found to represent (multivariate) outliers. As another check we have examined the robustness of the DINTERNET =DICRG correlation to the inclusion of the corruption determi- nants discussed in Treisman’s (2007) survey. Treisman includes variables that constitute historical and controls; political controls; and �nally a set of rents and competition controls. The correlation between DINTERNET and DICRG remains when these variables are included in the lagged dependent variables model (one-by-one). Finally, the above results carry over if we employ the control of corruption indicator (Kaufmann et al., 2007), instead of the ICRG index. These results are available upon request. Instrument Falsi�cation Test Following the approach taken in the U.S. state-level analysis our 2SLS analysis on the World sample is preceded by an instrument falsi�cation test. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Accordingly, Table 5 reports the results from estimating the reduced form on different time periods: pre-Internet (1984-1990) and post-Internet (1991-2005). As explained above, the pre-Internet period is limited to six years, due to lack of availability of the ICRG index beyond this. We report three sets of results on the two sub-periods: full sample OLS; OLS on a sample pruned for outliers detected by the Hadi (1992) method; full sample and outlier robust estimator (LAD). Turning to the results, we �nd a statistically signi�cant correlation between changes in ICRG and lightning during the Internet era (i.e., 1991 onwards). The partial correlation between lightning and changes in corruption reported in Column 1 can be inspected visually in Figure 5. It is immediately obvious from the �gure that Iceland (ISL) have consider- able influence on the signi�cance of lightning; i.e., Iceland may represent an outlier. Using the Hadi (1992) procedure, we con�rm that Iceland is indeed an outlier; and the only one. When we re-estimate the model, omitting Iceland (Column 2), we �nd essen- tially the same partial correlation as in the full sample. This remains true if we, rather than omitting Iceland from the sample, estimate the model by way of an outlier robust estimator (Column 3). The LAD point estimate is close to the OLS estimates. Hence, as in the U.S. sample, we �nd evidence that lightning prone regions have experienced lower reductions in corruption during the Internet era, consistent with the theory underlying the instrument. In Columns 4-6 we turn to the pre-Internet era. At �rst sight the test seems to falsify the instrument as we detect a signi�cant correlation between lightning and changes in ICRG during the pre-Internet era in Column 4. Taken at face value this means lightning is picking up forces that exert a time persistent influ- ence on the evolution of corruption, which is inconsistent with the theory that lightning captures (solely) Internet diffusion. The OLS based partial correlation T A B L E 5 . Falsi�cation test of instrument, world sample Change in ICRG 1991– 2005 Change in ICRG 1984–90 Dependent variable: (1) (2) (3) (4) (5) (6) log(LIGHTNING) 2 0.230*** 2 0.216*** 2 0.295*** 2 0.0931* 2 0.0938 2 0.0164 ( 2 0.070) ( 2 0.0785) ( 2 0.0777) ( 2 0.0527) ( 2 0.0619) ( 2 0.0588) ICRG, 1991 2 0.531*** 2 0.530*** 2 0.565*** ( 2 0.0825) ( 2 0.0825) ( 2 0.0952) ICRG,1984 2 0.232*** 2 0.232*** 2 0.0267 ( 2 0.0492) ( 2 0.0494) ( 2 0.101) Constant 1.362*** 1.331*** 1.517*** 1.040*** 1.042*** 0.132 ( 2 0.377) ( 2 0.385) ( 2 0.387) ( 2 0.244) ( 2 0.256) ( 2 0.469) Observations 124 123 124 99 98 99 R-squared 0.36 0.36 0.16 0.16 Estimator OLS OLS LAD OLS OLS LAD no outlier no outlier Notes: Robust standard errors in paranthesis. Asterisks ***, ** and * signify, respectively, p-value , 0.01, p-value , 0.05 and p-value , 0.1. Column 2 and 4 omit an outlier (Iceland) detected using the Hadi (1992) procedure. Larger values for the ICRG index means less corruption. In columns 3 and 6 the standard errors are bootstrapped with 1000 repetitions. Source: Authors’ analysis based on data described in the text. Andersen, Bentzen, Dalgaard and Selaya 411 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 412 THE WORLD BANK ECONOMIC REVIEW F I G U R E 5 . Reduced form in the Internet era, world sample Note: The �gure shows the association between lightning and changes in the ICRG index over Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 the 1991– 2005 period, with ICRG in 1991 partialled out. Source: Figure is based on data described in the text. F I G U R E 6 . Reduced form in the pre-Internet era, world sample Note: The �gure shows the association between lightning and changes in the ICRG index over the 1984– 1990 period, with ICRG in 1984 partialled out. Source: Figure is based on data described in the text. between lightning and changes in the ICRG index from Column 4 is illustrated in Figure 6. Visual inspection of the �gure reveals that Iceland once again appears to rep- resent an outlier, possibly even explaining the signi�cance of lightning during the pre-Internet period. Employing the Hadi (1992) procedure one can con�rm Andersen, Bentzen, Dalgaard and Selaya 413 that Iceland indeed is a multivariate outlier; it is the sole observation (as in the post-Internet setting) that the Hadi (1992) test singles out. When we subsequently re-estimate the model, omitting Iceland, the signi�- cance of lightning evaporates (Column 5). The fragility of the lightning /DICRG correlation for the 1984-90 period is further illustrated by the LAD regression on the full sample (Column 6): In this setting, the size of the esti- mate is reduced considerably compared to the OLS result, and is far from being statistically signi�cant at conventional levels. Overall, these falsi�cation tests are reassuring. When the outlier is omitted, or an estimation method that is robust to extreme observations is invoked, there is no statistically signi�cant correlation between lightning and changes in corruption prior to the emergence of the World Wide Web. Moreover, moving beyond statistical signi�cance, it is clear that lightning exhibits a much stronger correlation with changes in corruption during the post-Internet era, compared to the pre-Internet era. As can be seen from Table 5, the OLS estimate in Column 1 exceed that of Column 4 by no less than a factor of 2.5. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 In sum, we �nd the evidence reported in Table 5 suf�ciently supportive of the identi�cation strategy (albeit perhaps less strong than in the U.S. case) so as to allow us to proceed to 2SLS estimation using lightning intensity as an instru- ment for Internet diffusion. 2SLS Estimates Table 6 reports the 2SLS estimates of the impact from the Internet on the evol- ution of corruption around the World. Panel A reports the �rst stage, whereas Panel B displays second stage results. In keeping with the approach taken so far, we report results from three separate exercises: full sample, �rst difference speci�cation (Column 1); truncated sample, �rst difference speci�cation (Column 2); and our preferred lagged dependent variables speci�cation on the full sample (Column 3). Broadly speaking, the results mirror those attained using U.S. state-level data. First, the lightning instrument is statistically strong in the full sample of countries (Columns 1 and 3); the �rst stage features F-statistics well above ten. Second, the �rst difference speci�cation produces insigni�cant estimates for DINTERNET in the second stage, whereas the lagged dependent variables spe- ci�cation leads to a statistically signi�cant point estimate for DINTERNET . Our interpretation of this variation in results across speci�cations is the same as in the case of the U.S. state-level analysis. Changes in corruption display mean reversion, for which reason the initial level of corruption needs to be controlled for in the regression. When it is erroneously omitted, it jeopar- dizes identi�cation since lightning and the level of corruption are correlated. A glance on the lightning map (Figure 3) above is enough to realize that high lightning intensity is more prevalent in the poorer regions of the World, near the equator, where countries also tend to be characterized by relatively poor 414 THE WORLD BANK ECONOMIC REVIEW T A B L E 6 . 2SLS regressions, world sample Panel A: First stage Dependent variable: DINTERNET (1) (2) (3) log(LIGHTNING) 2 10.619*** 2 4.4499** 2 6.9537*** (1.2479) (1.5087) (1.2863) ICRG1991 6.5275*** (1.1915) Constant 38.985*** 20.899*** 10.252* (2.7145) (3.4299) (5.7635) F-test (H0: log(LIGHTNING) ¼ 0) 71.62 7.76 25.08 Panel B: Second stage Dependent variable: Change in ICRG 1991– 2005 DINTERNET 0.00011 0.0216 0.0435*** ( 2 0.006) ( 2 0.0285) ( 2 0.010) ICRG1991 2 0.820*** Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 ( 2 0.114) Constant 2 0.841*** 2 1.137*** 1.079*** ( 2 0.176) ( 2 0.329) ( 2 0.244) Observations 102 67 102 Notes: Robust standard errors in paranthesis. Asterisks ***, ** and * signify, respectively, p-value , 0.01, p-value , 0.05 and p-value , 0.1. In column 2 the sample is truncated; only countries with ICRG scores one standard deviation away from the ICRG index minimum (0) and maximum (6) value are included. Larger values for the ICRG index means less corruption. Source: Authors’ analysis based on data described in the text. governance. Accordingly, identi�cation is only plausible in the case of the lagged dependent variables speci�cation. The results from Column 3 of Table 6 corroborate our OLS �ndings (Table 4): From the early 1990s onward the Internet has worked to suppress corruption. Hence, it would appear that our main results, which pertain to the U.S., carry over to a cross-country setting. In closing, it should be noted that results are similar when we use the Kaufmann et al. (2007) corruption indicator instead of the ICRG data; these results are available upon request. V. C O N C L U D I N G R E M A R K S In the present paper we have examined the hypothesis that technological change may cause improvements in governance. Speci�cally, we have studied the influence from the Internet on the evolution of corruption across U.S. states and across the World. The results are, statistically speaking, perhaps somewhat stronger in the U.S. setting but in both samples we �nd evidence that the Internet has worked to suppress corruption since its emergence. Andersen, Bentzen, Dalgaard and Selaya 415 Our results should be interpreted with some care: they are reduced form esti- mates of the total impact from the Internet on corruption. The analysis there- fore does not speak to the mechanisms linking the Internet to corruption. As observed in the Introduction, the Internet is a General Purpose Technology, for which reason it probably influences the economy in a great many ways; growth, human capital accumulation, the dissemination of information, and E-government are the most obvious intervening factors that are both associated with the Internet and impinge on the level of corruption in a state or country. The task of disentangling these separate pathways of influence from the Internet technology to corruption is left open for future research. The identi�cation strategy developed in the present paper may be useful in other contexts. With a plausibly exogenous instrument for the Internet, researchers may potentially make new progress in the study of the impact from the Internet on other outcomes, such as the return to skills or productivity growth more broadly. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 APPENDIX T A B L E A . 1 . Summary statistics Panel A: Cross-country sample Obs. Mean Median Std. Dev. Min Max ICRG 2005 102 2.60 2.14 1.30 0 6 ICRG 1991 102 3.44 3 1.46 0 6 INTERNET 2005 113 22.56 15.24 21.66 0.21 86.94 Average lightning density 113 9.00 6.59 8.62 0.02 44.38 Panel B: U.S. sample CC 2006 48 20.15 9.50 21.48 0 83 CC 1991 48 13.44 6.00 21.14 0 108 INTERNET 2003 48 54.39 55.00 5.88 39.50 65.50 POP 1991 48 52.22 36.46 56.16 4.59 305.00 Average lightning density 48 10.54 8.58 7.31 0.89 27.34 Source: Authors’ calculations based on data described in the text. REFERENCES Acemoglu, Daron, and James Robinson, 2001. Political Losers as a Barrier to Economic Development. American Economic Review, 90, 126– 130. Ades, Alberto, and Rafael di Tella, 1999. Rents, Competition and Corruption. American Economic Review, 89, 982– 993. Andersen, Thomas B., 2009. E-government as an anti-corruption strategy. Information Economics and Policy, 21, 201–210. 416 THE WORLD BANK ECONOMIC REVIEW Andersen, Thomas B., Jeanet Bentzen, Carl-Johan Dalgaard, and Pablo Selaya, 2010. Lightning, IT Diffusion, and Economic Growth across U.S. States. Review of Economics and Statistics, forthcoming. Angrist, Joshua, and Stefan Pischke, 2009. Mostly Harmless Econometrics, Princeton University Press. Besley, Tim, and Robin Burgess, 2002. The Political Economy of Government Responsiveness: Theory and Evidence from India. Quarterly Journal of Economics, CXVII, 1415–1451. Bhatnagar, Subhash, 2003. E-government and access to information. Global Corruption Report 2003, Transparency International. BSI, 2004. Bundesamt fu ¨ r Sicherheit in der Informationstechnik, IT-Grundschutz Manual. Available online at , http://www.bsi.de/english/gshb/index.htm . . Brunetti, Aymo, and Beatrice Weder, 2003. A free press is bad news for corruption. Journal of Public Economics, 87, 1801– 1824. Caselli, Francesco, and Wilbur J. Coleman III, 2001. Cross-Country Technology Diffusion: The Case of Computers. American Economic Review, 91, 328–335. Chinn, Menzie D., and Robert W. Fairlie, 2007. The determinants of the global digital divide: a cross- country analysis of computer and Internet penetration. Oxford Economic Papers, 59, 16 –44. Chisholm, William A., 2000. Lightning Protection. Chapter 4.10 in: The Electrical Power Engineering Handbook. Edited by L Grigsby., IEEE Press. Chisholm, William A., and Kenneth Cummins, 2006. On the Use of LIS/OTD Flash Density in Electric Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Utility Reliability Analysis. Paper presented at The Lightning Imaging Sensor International Workshop. Available online at , http://lis11.nsstc.uah.edu/Papers/Proceedings.html . . Deaton, Angus, 2010. Instruments, Randomization, and Learning about Development, Journal of Economic Literature, 48, 424– 455. ¨ mberg, 2007. New Droughts, New Floods and U.S. Disaster Relief. Eisensee, Thomas, and David Sto Quarterly Journal of Economics, 122, 693–728. Electricity Power Research Institute, 2003. Electricity Technology Roadmap: 2003 Summary and Syntheses. Available online at , http://www.iea.org/Textbase/work/2004/distribution/presentations/ Gellings_bkgd_%20paper.pdf . . Ferraz, Claudio, and Frederico Finan, 2008. Exposing Corrupt Politicians: The Effect of Brazil’s Publicly Released Audits on Electoral Outcomes. Quarterly Journal of Economics, 123, 703– 745. Glaeser, Edward L., and Raven E. Saks, 2006. Corruption in America. Journal of Public Economics, 90, 1053– 1072. Foster, Andrew D., and Mark R. Rosenzweig, 1996. Technical Change and Human-Capital Returns and Investments: Evidence from the Green Revolution. American Economic Review, 86, p. 931– 53. Fredriksson, Per G., Herman R. J. Vollebergh, and Elbert Dijkgraaf, 2004. Corruption and energy ef�- ciency in OECD countries: theory and evidence. Journal of Environmental Economics and Management, 47, 207– 231. Gundlach, Erich, and Martin Paldam, 2009. The transition of corruption: From poverty to honesty. Economics Letters, 103, p.146–148. Hadi, Ali S, 1992. Identifying multiple outliers in multivariate data. Journal of the Royal Statistical Society 54, 761–771. Hall, Robert E., and Dale Jorgenson, 1967. Tax Policy and Investment Behavior. American Economic Review, 57, 391–414. IEEE, 2005. How to Protect Your House and Its Contents From Lightning. IEEE Guide for Surge Protection of Equipment Connected to AC Power and Communication Circuits. Standards Information Network, IEEE Press. Jovanovic, Boyan, and Peter L. Rousseau, 2005. General Purpose Technologies. In P. Aghion, and S. Durlauf (eds.) Handbook of Economics Growth, Elsevier, North-Holland. Kaufmann, Daniel, Aart Kraay, and Massimo Mastruzzi, 2007. Governance Matters IV: Governance Indicators for 1996–2006. Available online at , http://www.worldbank.org/wbi/governance/ wp-governance.html . . Andersen, Bentzen, Dalgaard and Selaya 417 Kressel, Henry, 2007. Competing for the Future: How Digital Innovations are Changing the World, Cambridge University Press. Licht, Amir N., Chanan Goldschmidt, and Shalom H. Schwartz, 2007. Culture Rules: The foundations of the rule of law and other norms of governance. Journal of Comparative Economics, 35, 659– 88. Mauro, Paolo. 1995. Corruption and Growth. Quarterly Journal of Economics 110, 681–712. Maxwell, Amanda E., and Richard F. Winters, 2005. Political Corruption in America. Working Paper (Dartmouth College). Mokyr, Joel, 1990. The Lever of Riches: Technological Creativity and Economic Progress, Oxford University Press. NASA, 2007. Lightning Detection From Space: A Lightning Primer. Available online at , http://thunder .msfc.nasa.gov/primer . . National Energy Technology Laboratory, 2003. The Value of Electricity When It’s Not Available. Available online at , http://www.netl.doe.gov/moderngrid/docs/The_Value_of_Electricity_When_ It’s_Not_Available.pdf . . Parente, Stephen L., and Edward C. Prescott, 1999. Monopoly Rights: A Barrier to Riches. American Economic Review, 1216– 1233. Persson, Torsten, Guido Tabellini, and Francesco Trebbi, 2003. Electoral rules and corruption. Journal of the European Economic Association, 1, 958– 989. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Reinikka, Ritva, and Jakob Svensson, 2004. The power of information: Evidence from a newspaper campaign to reduce capture. Policy Research Working Paper No. 3239, World Bank. Schultz, Theodore W., 1975. The Value of the Ability to Deal with Disequilibria. Journal of Economic Literature, 13, p. 827– 46. Shim, Jae K., Anique Qureshi, and Joel G. Siegel, 2000. The International Handbook of Computer Security. Glenlake Publishing Company Ltd. Staiger, Douglas, and James H. Stock, 1997. Instrumental variables with Weak Instruments. Econometrica, 65, 557–586. ¨ mberg, David, 2001. Mass Media and Public Policy, European Economic Review, 45, 652– 663. Stro ———, 2004. Radio’s Impact on Public Spending, Quarterly Journal of Economics, 119, 189 –221. Svensson, Jakob., 2005. Eight Questions about Corruption. Journal of Economic Perspectives, 19, 19– 42. Treisman, Daniel, 2000. The causes of corruption: a cross-national study. Journal of Public Economics, 76, 399– 457. ———, 2007. What Have We Learned About the Causes of Corruption From Ten Years of Cross-National Empirical Research? Annual Review of Political Science, 10, 211– 244. Yeager, Kurt, and Karl Stahlkopf, 2000. Power for a Digital Society. Prepared for the Department of Energy CF-170/1-1-DOE. Available online at , http://www.rand.org/scitech/stpi/Evision/Supplement/ yeager.pdf . . WDI, 2007. World Development Indicators 2007, World Bank. West, Darrell M., 2005. Digital Government, Princeton University Press. ———, 2008. Improving Technology Utilization in Electronic Government around the World. Brookings Governance Studies. Wooldridge, Jeffrey., 2003. Introductory Econometrics: A Modern Approach, 3E, South– Western. World Development Report 2004. Making Services Work for the Poor, The World Bank. Do Labor Statistics Depend on How and to Whom the Questions Are Asked? Results from a Survey Experiment in Tanzania Elena Bardasi, Kathleen Beegle, Andrew Dillon, and Pieter Serneels Labor market statistics are critical for assessing and understanding economic develop- ment. However, widespread variation exists in how labor statistics are collected in household surveys. This paper analyzes the effects of alternative survey design on employment statistics by implementing a randomized survey experiment in Tanzania. Two features of the survey design are assessed – the level of detail of the employment questions and the type of respondent. It turns out that both features have relevant and statistically signi�cant effects on employment statistics. Using a short labor module without screening questions induces many individuals to adopt a broad de�nition of employment, incorrectly including domestic duties. But after reclassifying those in Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 domestic work as ‘not working’ in order to obtain the correct ILO classi�cation, the short module turns out to generate lower female employment rates, higher working hours for both men and women who are employed, and lower rates of wage employ- ment than the detailed module. Response by proxy rather than self-report has no effect on female labor statistics but yields substantially lower male employment rates, mostly due to underreporting of agricultural activity. The large impacts of proxy responses on male employment rates are attenuated when proxy informants are spouses and individuals with some schooling. JEL CODES: J21, C83, C93. Elena Bardasi and Kathleen Beegle (corresponding author) are Senior Economists at the World Bank; their email addresses are ebardasi@worldbank.org, and kbeegle@worldbank.org, respectively. Andrew Dillon is a research fellow at the International Food Policy Research Institute; his email address is a.dillon@cgiar.org. Pieter Serneels is associate professor at the University of East Anglia, United Kingdom; his email address is p.serneels@uea.ac.uk. The authors would like to thank Economic Development Initiatives, especially Joachim de Weerdt, the supervisory staff, enumerators, and data entry teams for thorough work in the �eld. They also thank Gero Carletto, Louise Fox, Annette Ja ¨ ckle, David Newhouse, Dominique van de Walle, seminar participants at IFPRI, ASSA annual meetings, the IZA/World Bank Conference, CSAE University of Oxford, the Institute for Social and Economic Research (ISER) at the University of Essex, and the World Bank, as well as the editor and two anonymous referees for very useful comments. This work was supported by the World Bank Research Support Budget and the Gender Action Plan (GAP). All views are those of the authors and do not reflect the views of the World Bank or its member countries. THE WORLD BANK ECONOMIC REVIEW, VOL. 25, NO. 3, pp. 418– 447 doi:10.1093/wber/lhr022 Advance Access Publication June 14, 2011 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 418 Bardasi, Beegle, Dillon, and Serneels 419 I. INTRODUCTION Labor market statistics are critical for assessing how an economy functions, but they may be sensitive to the survey method by which they are collected. This paper provides evidence of the impact of the respondent type (self- reporting vs. proxy informant) and the level of detail of the labor module on labor force participation, hours worked, earnings, sector distribution, and employment status from a randomized survey experiment in a low income country. The �ndings con�rm that labor market statistics are indeed sensitive to the survey method. In particular, male employment is especially sensitive to the selection of the informant, while female employment varies in relation to the inclusion of screening questions at the beginning of the employment module. The experiment is carried out in a low-income country – Tanzania – and contributes to the scarce literature on survey methodology in developing countries. A wealth of evidence exists on the quality and reliability of labor statistics in household surveys, coming largely from the United States (see Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Bound et al., 2001, for a thorough review). However, few studies in low- income countries have attempted to rigorously measure the effect of speci�c features of the survey on the employment statistics it generates. These effects may vary across groups in the population – for example, employment statistics for women and children may be particularly sensitive to the survey method.1 When the wording of the employment questions emphasizes the main activity, this may lead to underestimating the number of economically active women because of the large female presence among unpaid agricultural and family workers (Dixon-Mueller and Anker, 1988). Child and teenage work may be similarly underreported. One way to explore whether alternative survey designs impact the labor stat- istics they produce would be to examine data from two surveys of different design but covering the same country and time period. This is the approach of Guarcello et al. (2009) in examining child labor. However, a review of national surveys in low-income countries yielded few relevant examples of surveys measuring employment and carried out in the same country at suf�ciently close moments in time. This reflects the fact that for most low-income countries, the national surveys are either not annual or, if they are, they are topic speci�c (such as the Labor Force Survey, followed the next year by the Demographic Health Survey, followed by the Household Budget Survey). For Tanzania, the Integrated Labour Force Survey (ILFS) 2000/01 reports labor force partici- pation rates of 90.6 percent for men and 89.5 percent for women (NBS, 2003), while the Household Budget Survey (HBS) reports 91.1 and 82.4 percent, 1. Guarcello et al. (2009) review discrepancies in child labor statistics across surveys in several low-income countries. Dillon et al. (2010) study the effect of survey design on child labor statistics using data from the same experiment as used for this paper. 420 THE WORLD BANK ECONOMIC REVIEW respectively, for the same year (NBS, 2002). The large difference in labor force participation of women between two nationwide surveys that refer to the same year may reflect genuine differences in samples or in the timing of the survey, or may reflect differences caused by the use of distinct survey instruments.2 Another interesting example is offered by the Malawi Integrated Household Survey 2004/05. Although not designed as an experiment, this survey included questions on labor both in the household roster (main activity of members reported by the head) and in a module listing several non-mutually exclusive activities (supposed to be self-reported). The results show that the main activity question, answered by the head of the household, understates the percentage of individuals in farming and paid employment and overstates the fraction of inactive people (both men and women), compared with the activity-speci�c set of questions, answered by each individual.3 While in this case the effects due to the interviewer and the timing of the survey are controlled for, it is not possible to attribute these differences to either the type of question or the type of respondent; only a randomized design would allow for this. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Assessing the internal validity of survey measures, although common in the psychological sciences, remains scarce in economics. In this paper, the focus is on two key survey characteristics: the level of detail of the employment ques- tions and the informant type. There is an experimental setting to estimate the impact of each feature. The results show that there are signi�cant differences in labor force participation, type of economic activity, and hours worked across survey designs. Using a labor module with no screening questions generates lower female labor force participation once domestic work is reclassi�ed as no work to be in line with standard de�nitions, and higher average working hours 2. The question on employment indeed differs between the HBS and the ILFS. The HBS uses one single question to collect information about participation in employment and type of activity (or category of inactivity): “During the last 7 days what was your main activity?� The individual can choose among eleven categories of employment (farming/livestock keeping; �shing; mining; tourism; government employee; parastatal employee; NGO/religious organization employee; private or other employee; self-employed with employee; self-employed without employee; unpaid family helper in non-agriculture business), unemployment, and seven categories of inactivity (no activity; household chores; student; not active: retired; not active: sick; not active: disabled; not active: other). The question is repeated for the second activity. All categories are listed without any explicit distinction between employment and non-employment categories; quite interestingly, four categories are explicitly labeled as ‘not active’, when in fact there are another three categories of inactivity. In the ILFS, one �rst question asks about the usual activities during the last 12 months, to be chosen among a list of 43 economic activities, with options as detailed as “agriculture: cash crop: cotton�, or “construction: farm buildings or fences�. Multiple answers are allowed. Information about current activity is also recorded, with reference to the same list of economic activities used to identify the usual activity. Household duties and other categories of inactivity are explicitly labeled as such and offered as an option in a later question (“What was your main activity when you were not doing economic activity and not available for work during that period?�). By providing a detailed list of economic activities clearly de�ned as ‘work’ the ILFS explicitly de�nes what employment is, while the HBS does not. It is not possible, however, to determine that this is the source of the discrepancy between the employment statistics for women because other survey elements differ as well (the number of categories, the sequence of questions, etc.). 3. Detailed results are available from the authors upon request. Bardasi, Beegle, Dillon, and Serneels 421 for both men and women. Response by proxy rather than self-report yields substantially lower male labor force participation, lower male working hours, and lower employment in agriculture for men. This indicates that the survey design matters to measuring labor outcomes and, moreover, that comparisons across surveys with different design can be compromised by these differences. The structure of the paper is as follows. In the next section, key �ndings from studies from high-income countries and developing countries are dis- cussed. In Section 3, the experimental design is described. Section 4 provides a description of the data, while Section 5 presents the results. Section 6 concludes. II. BACKGROUND AND LITERATUR E Bias in statistics from surveys can arise from several sources. Besides sampling error (related to sample size) and poor representativeness of the sample (due to non-response bias, under-coverage of certain groups of the population, or Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 respondent self-selection), an important source of bias in surveys is measure- ment error. Measurement error – the difference between the value of a charac- teristic reported in the survey and the (“true� and unknown) value sought by the researcher – is related to the data-collection process. Its main sources are the questionnaire (question selection, sequencing, and wording), the type of informant, the data-collection method, and the interviewer. One should be concerned about measurement error because it may bias both survey statistics as well as estimates of relationships between measures of employment and other variables – hence the importance of understanding how different survey methods may impact the accuracy of the data collected. The sources of measurement error have been studied mostly in the context of developed countries (Bound, Brown, Mathiowetz, 2001; Biemer et al., 1991). A recent review of the main issues concerning survey design in develop- ing and transition countries relies almost entirely on research in developed countries when discussing measurement error (Kasprzyk, 2005). As a matter of fact, however, very little research exists for developing countries and it is not clear how relevant the methodological literature from high-income countries is for low-income countries. One main difference is represented by the variables of interest – even when focusing on the restricted area of employment stat- istics. In developed economies, for example, a lot of emphasis has been placed on the correct measurement of unemployment, especially in relation to the use of panel data for the measurement of unemployment duration and the tran- sitions to and from unemployment.4 In low-income countries, however, the concept of unemployment, as de�ned by the International Labour Organization (ILO) and understood in developed societies, seems less relevant; 4. Some examples include Poterba and Summers (1986, 1995), Sinclair and Gastwirth (1998), and Singh and Rao (1995). 422 THE WORLD BANK ECONOMIC REVIEW it is rather the concept of employment (and its quality and intensity) that is at the same time important and elusive for the researcher. This explains why employment is a main variable of interest, alongside earnings and hours of work. The experiment focuses on two sources of measurement error, speci�cally on the effects of (i) using detailed probing questions vs. a single, shorter ques- tion, and (ii) using proxy informants instead of self-reports. In the brief review of the literature on these two sources of measurement error, references are mostly to the studies that are relevant for this analysis – those papers analyzing impacts on employment and possibly adopting an experimental framework. The speci�c wording and style of employment questions are posited to have a large influence on labor statistics. This may be particularly relevant in a setting where a signi�cant proportion of individuals are employed in household-owned enterprises or home production and are not directly remun- erated in the form of a salary or wage. For example, the standard question “Did you work in the last 7 days?� is hypothesized to systematically under- Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 count persons who work in household enterprise activities without direct wage payments (e.g., unpaid family workers), who may have dif�culties in identify- ing themselves as ‘working’. Likewise, employment questions that only cover the past 7 days may produce incorrect statistics on employment participation in settings where employment is highly seasonal or where a signi�cant proportion of workers are casual laborers. A number of studies have focused on the style of different questions (open vs. closed questions, positive vs. negative statements, etc.) and the effects of their placement in the survey questionnaire (see the review in Kalton and Schuman, 1982). Mostly they have con�rmed that question-wording effects are important, although the direction of these effects is often unpredictable. Studies have been carried out in the context of the revision of the employment questions in the U.S. Current Population Survey (CPS) to investigate the concern that irregular, unpaid, and marginal activities may be underreported partly because people do not think of themselves as working. In the Respondent Debrie�ng Study, respondents were asked to classify hypothetical situations (“vignettes�) in terms of “work,� “job,� “business,� and so on. Generally, the majority of respondents were able to classify the situations con- sistently with de�nitions of the CPS. However, for each vignette, large min- orities of respondents gave incorrect answers – for example, 38 percent of the respondents included non-work activities under the “work� classi�cation (Campanelli, Rothgeb, and Martin, 1989).5 An experiment carried out in 1991 to assess the revision of the CPS questionnaire using vignettes and direct screening questions for unreported work con�rmed that questionnaire wording 5. Esposito et al. (1991) discuss methodological tools used to obtain diagnostic information to evaluate the effect of questionnaire revisions on reporting of work activities, including hypothetical vignettes and direct screening questions. Bardasi, Beegle, Dillon, and Serneels 423 and sequence of questions affect the respondent’s interpretation of work and, therefore, the employment statistics (Martin and Polivka, 1995). Moreover, the use of direct screening questions was found particularly useful to detect under- reporting of work done in connection with the household business or farm, as well as underreporting of teenage employment. The 1991 CPS study noted above also pointed to the existence of gender dimensions of these effects. In particular, the revision of the questionnaire, aimed at better capturing unpaid work in a household business or farm, increased the female employment rate. In developing countries, the gender effects may be even more dramatic than in developed countries. Many studies have expressed concerns about the underreporting and undervaluing of women’s work when using the most common methods of employment data collection (Anker, 1983; Dixon-Mueller and Anker, 1988; Charmes, 1998; Mata Greenwood, 2000). In developing countries, women workers tend to have a prominent role in agriculture and informal sector activities and, because of assigned cultural roles, may be considered by others and themselves as inac- Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 tive even when they perform economic activities. In this context, it may be par- ticularly dif�cult to capture women’s work (Mata Greenwood, 2000). In addition to key features of questionnaire design, different surveys adopt different approaches to designating the respondent to the questionnaire. Standard surveys in developing countries, like Household Budget Surveys (HBS), Income and Consumption Expenditure Surveys, and Core Welfare Indicator Questionnaires (CWIQ) typically ask the household head employ- ment questions about all household members. However, proxy informants may not always provide accurate information and this can cause biases in estimation of employment (Hussmanns, Mehran, and Verma, 1990). An alternative approach is to ask each household member above a certain age directly as in the Living Standards Measurement Study surveys (LSMS) (Glewwe and Grosh, 2000) and in Labor Force Surveys (LFS). Requiring all individuals to self-report makes the �eldwork quite burdensome and expensive, creating a trade-off between the accuracy of the information and the cost to obtain it. Most survey experiments6 that study the effects of using proxy informants in lieu of self-respondents on employment statistics are from developed countries. In a study for the U.K., Martin and Butcher (1982), in comparing the answers of husband and wife, found that employment variables had less than a 10 percent disagreement rate, while approximately 20 percent of the proxies did not know the income of their spouse. In a similar U.S. survey, Boehm (1989) found that self and proxy responses resulted in the same labor force classi�- cation 83 percent of the time. However, this study was based on a small 6. Experimental studies are especially useful in assessing the “true effect� of using proxy vs. self-respondents. Non-experimental studies tend to suffer from the problem of self-selection (Hill, 1987; Moore, 1988) – that is, proxies may be individuals who happen to be at home. These proxy informants will typically have different characteristics than those who are absent from the household and those characteristics are generally correlated with the type of information that it is collected. 424 THE WORLD BANK ECONOMIC REVIEW sample of 84 individuals from a group of participant volunteers. In general, the little experimental evidence and the non-experimental studies indicate that self- respondents produce higher household and person non-interview rates, but proxies produce higher item non-response rates, especially for wages and income variables (Biggs, 1992). The use of proxies may amplify recall errors or affect the reporting of hours of work, especially in the case of irregular or mul- tiple activities (Hussmanns, Mehran, and Verma, 1990). Moreover, the use of proxies is also considered to be a potential source of gender bias in a context where women’s participation in economic activity may be underestimated (ILO, 1982). In their study of proxy reports in the United Kingdom, The Of�ce for National Statistics (2003) found that no one proxy informant is best placed to provide reliable proxy information for all questions. Moreover, they reject the notion of an “ideal� proxy informant in terms of personal characteristics given the variation across households. The reasons why there could be discre- pancies between proxy and self-reports are reviewed in Blair et al. (1991). Experiments they conducted to analyze the strategies used by individuals to Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 self-report or proxy report a speci�c event, opinion, or behavior indicate that characteristics of the questionnaire as well as individual characteristics of the self- and proxy-respondent affect the strategies used to respond and the conver- gence of their answers. Unfortunately their experiments do not relate to employment issues. III. THE SURVEY EXPERIMENT The survey experiment conducted and analyzed here seeks to inform the method by which labor statistics are collected in household surveys in low- income countries, and, therefore, the information base for analytical work on employment. Employment is de�ned as time spent in an economic activity, regardless of a wage associated with it or its formal or informal nature. In this study, working includes time spent in any work for pay (as wage or salaried worker), pro�t (as employer, self-employed, or own-account worker), or family gain (as paid or unpaid worker in a family farm or family business). It does not include domestic work such as housekeeping, child rearing, and preparing meals – which are not comprised within the System of National Accounts (SNA) production boundary. Because of the reasons indicated in Section 2, unemployment is not a labor market measure here, as it would require a speci�c conceptual and methodological approach. The survey experiment was designed and implemented to focus on two key dimensions of labor survey design: the level of detail of the questionnaire (speci�cally the use of screening questions to establish employment status) and the type of informant. To investigate the impact of screening questions, a detailed and a short labor module was developed. The short labor module reflects the approach in shorter questionnaires, such as the Core Welfare Indicator Questionnaire (CWIQ). Bardasi, Beegle, Dillon, and Serneels 425 Many countries regularly �eld CWIQ-type surveys (such as Welfare Monitoring Surveys), especially with increasing demands to produce sub- regional household survey statistics. This shorter module is often used to gener- ate statistics with a higher frequency, for example with annual regularity, in lieu of complex multi-topic household surveys. The detailed labor module reflects the approach in longer questionnaires typically used in multipurpose household surveys, such as the LSMS. In this survey experiment, the detailed module differed from the short module in two ways: in the set of screening questions to determine employment status and in asking about second and third jobs. Here the focus is on the effect of including screening questions. The detailed module starts with three questions to determine employment status: speci�cally, whether the person has worked for someone outside the household (as an employee), whether s/he has worked on the household farm, and whether s/he has worked in a non-farm household enterprise. For each of these three questions, the response is “yes� or “no.� These questions were asked with respect to the last 7 days (the refer- Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 ence period for identifying those who are “employed�) and, if the person has not reported to work in the last 7 days, the questions are asked with respect to the last 12 months. In the short module, there was only one question to deter- mine the employment status with respect to the last 7 days: whether s/he did any type of work, with as response also “yes� or “no.� As in the detailed module, the question was asked twice – for the last 7 days and the last 12 months. Annex Table 1 presents the questions to determine whether the indi- vidual is employed. The complete short and detailed employment modules are reported in Bardasi et al. (2010). In the second dimension of the experiment, there is variation as to whether questions were asked directly to the subject or to a proxy informant. Response by proxy rather than individuals themselves reflects the common practice to interview an informed household member (often the household head or spouse), rather than each individual him or herself. In practice proxy infor- mants are often used when individuals are away from the household or other- wise unavailable in the time allotted in an enumeration area to conduct interviews. In the survey experiment, the proxy informant was randomly chosen among household members at least 16 years old.7 This age threshold reflects common practice in �eldwork to choose an adult to be a proxy infor- mant (for children or adults) in the household. The proxy informant is thus either the head of household, spouse of the head, or an older child or relative living in the household. The persons selected to be the proxy informants then 7. The Tanzanian CWIQ 2006 data indicate that the average Tanzanian household has between two and three adults who could serve as a proxy with a minimum age of 16. This informed the design of our survey, and, in fact, our sample households had 2.5 members 16 years and older. 426 THE WORLD BANK ECONOMIC REVIEW reported on themselves and on up to two other randomly selected household members age 10 or older.8 In this paper, the responses, either proxy or self- reported, of those who are age 16 and above (which are de�ned as ‘adults’) are analyzed. In actual implementation of surveys, proxy informants are not ran- domly chosen, but are normally selected by interviewers on the basis of their knowledge and availability. In this sense, the experiment did not exactly mimic the actual conditions that result in proxy responses in household surveys. However, by randomly selecting proxy informants and using the information about the relationship between the proxy informant and the subject, this study can assess whether different types of proxy informants give different types of responses.9 However, because a typical survey does not generally identify the proxy in relation to the person for whom the information is collected, the study cannot determine what the results imply in terms of potential “bias� of a typical survey due to the use of proxies.10 The assumption is that the self-reported information is more accurate than proxy reports. However, it is not tested whether this is true and speci�c Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 reasons for the potential discrepancy cannot be identi�ed. For example, if proxy informants report lower participation in employment, one cannot differ- entiate between explanations such as (1) proxies are not fully knowledgeable of the employment activity of the other household members, either because 8. Random selection of proxies was conducted in the �eld by the enumerator who �rst listed, in each household, all eligible proxies (all household members aged 16 or older) in a Table (let’s call it Table A) on a proxy selection questionnaire page in the same order they appeared in the household roster. Table A was then matched with Table B, generated uniquely for each questionnaire, listing in the �rst column a sequence of numbers from 1 to N, where N was the total number of eligible proxy respondents in that household, and in the second column a randomly generated number in the range (1, N). The proxy respondent was chosen by selecting the individual ordered Mth in Table A, where M was the random number associated with the Nth row in Table B. The selection of the household members to be responded for by the proxy was made using a similar procedure, after excluding the selected proxy from the list of eligible members (aged 10 and older). The random selection of respondents in the self-reported sample was also made using the same procedure, but simpli�ed to only one step. 9. There are two other reasons why the survey experiment was designed to select proxies at random. First, this design attempts to remove the influence of interviewer effects, since better interviewers will select better/more appropriate informants (and, in our view, our interviewers were well above average and had greater supervision than in a typical survey). Second, the structure of the �eld work suggested that if not randomly assigned, the data would be better from proxy informants than a “normal� survey because the teams were in the enumeration area for 17 days to conduct a simultaneous consumption survey experiment allowing for more time to locate the best informant. 10. An alternative research design to assess the effect of proxies would have been to interview two members of the household who report on their own labor activities and proxy report on the other. This design was not implemented because it proved to be too dif�cult to ensure proper implementation for a medium to large sample. After consultation with counterparts in Tanzania, it was concluded that it would be dif�cult to assure that proxy and self-responses would be independent and would remain unaffected by the knowledge that another household member reports on the same information, given the normally social nature of an interview. The speci�c concern was that the design (and open communication about this design within the village) would trigger either a coordinated response by household pairs and/or accommodation of response to the other’s expectations, which would introduce potentially much larger (unobserved) respondent biases. Bardasi, Beegle, Dillon, and Serneels 427 individuals hide their employment participation from other members, or simply because it is dif�cult to “keep track� of what others are doing, especially in large households; (2) proxies tend to have a “low opinion� of other household members and are likely to think that what they do does not qualify as work even when it does; or (3) the opposite, proxy informants are more likely to respond objectively and it is the individual who overstates his or her employment to make it appear that s/he works because “it looks better.� Although proxy informants and self-reporting are both commonly used, the detailed self-report questionnaire is generally considered to be the “best prac- tice� approach of household surveys. The use of multiple questions to deter- mine whether the subject is employed or not is recommended by the ILO, especially when some categories of workers (especially casual workers, unpaid family workers, apprentices, women engaged in non-market production, workers remunerated in-kind, etc.) may not be able to correctly interpret a question about “any type of work� as referring to their situation (Hussmanns, Mehran, and Verma 1990). The focus of this analysis is therefore whether Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 short questionnaires provide the same information as detailed ones, and whether responses by proxy informants deviate from self-responses. For those identi�ed as working in the last 7 days, either through the set of three questions (in the detailed module) or through the single question (in the short module), information on the occupation, sector, employer, hours, and wage payments was collected for the main job. These questions are identical across assignments. Participation in domestic duties, while conceptually not included in the de�nition of employment, is commonly collected in surveys. This is usually done by adding domestic duties as a possible answer to the question about the main sector of activity and this approach was followed in both the short and detailed modules. For all the survey assignments, in addition to the labor module, the question- naire also included six other modules: household roster, assets, dwelling charac- teristics, land, food consumption, and non-food expenditures. In the detailed and short questionnaire, the questions followed the same sequence; identical types of questions follow the same phrasing and recall periods are the same. From an analytical perspective, the objective is to assess the effects of the change in survey assignment ( presence of screening questions and type of respondent). The design of our experiment introduced an imbalance in the composition of the proxy and self-report experimental groups with respect to several demographic characteristics. Proxy informants can exist only in house- holds with at least one individual aged 16 or older and at least another one aged 10 or older; moreover, the random procedure to select proxy informants, individuals to be reported for by proxy, and self-reports is such that similar individuals have different probabilities of being selected in the two samples. We addressed the former problem by retaining for our analysis only those 428 THE WORLD BANK ECONOMIC REVIEW households with at least two persons eligible to be a proxy informant (two persons 16 and older). The second problem was addressed by using survey weights calculated as the inverse of the selection probability. If M is the number of household members aged 10 þ , the probability of being selected to self-report (in the self-report households) is 2/M if M . 1 and 1 if M ¼ 1. In proxy households, the probability of being selected as a proxy informant (and thus also be a self-report) is p ¼ 1/L, where L is the number of household members aged 16 þ . The probability of being selected as an individual responded for by a proxy informant corresponds to the probability of not being selected as a proxy times the probability of being selected out of all remaining individuals eligible to be responded for, that is w ¼ (1-p)  r, where r ¼ min[1,2/(M-1)]. After appropriately de�ning the samples to be compared and weighting each observation for the inverse of the probability of being selected, means can be compared across samples. Because questions on hours, earnings, and sector are identical across assign- ments, variations in statistics across survey assignments are not due to question Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 wording. However, the response to labor force participation determines whether statistics on those other dimensions are collected at all for the individ- ual (in other words, these statistics are conditional on the individual being classi�ed as employed). In the case of self-respondents, the screening questions that differentiate the start of the short and detailed modules entirely explain variations in selection into employment and therefore variations in hours, earn- ings, and sector statistics. In the case of proxy informants, variations in stat- istics for these other outcomes derive from both the quality of reporting by the proxy informant on a speci�c variable (e.g., how well the wife knows how many hours her husband works) and the accuracy of reporting on employment status (if the husband does not report that his wife works, then he will not be asked her hours). Only the latter is a selection issue. I V. D A T A AND CONTEXT The survey experiment was implemented in Tanzania, which has different types of labor market surveys, including CWIQs, LFSs, and multi-purpose household surveys, like the Household Budget Survey (HBS). The survey exper- iment conducted was the Survey of Household Welfare and Labour in Tanzania (SHWALITA). The �eld work was conducted from September 2007 to August 2008 in villages and urban areas from 7 districts across Tanzania: one district in the regions of Dodoma, Pwani, Dar es Salaam, Manyara, and Shinyanga and two districts in the Kagera region. The sampling is a two-stage design in each region. First, villages (or urban clusters) were randomly selected proportional to their population size. Second, 12 households were randomly Bardasi, Beegle, Dillon, and Serneels 429 selected from a household listing in each sample village (urban cluster).11 Three of the selected 12 households were then randomly assigned to each of the four survey designs. The total sample is 1,344 households (with two of these households being replacement households selected from the original listing exercise for two households that refused to participate), with 336 house- holds randomly assigned to each of the four survey assignments. Although the sample of 1,344 is not designed to be nationally representative of Tanzania, the districts were selected to capture variations between urban and rural areas as well as along other socio-economic dimensions. The basic characteristics of the sampled households generally match the nationally representative data from the Household Budget Survey (2006/07) (results not presented here). Household interviews were conducted over a 12-month period but, because of small samples, the survey assignment effects across seasons (such as harvest time with a peak in labor demand and dry seasons with low demand) are not explored. The random assignment of house- holds is validated when examining a set of household characteristics (results Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 not presented here, but available in Bardasi et al., 2010). The individuals are classi�ed on the basis of the survey assignment that they actually received. An individual’s actual survey assignment is the result of the initial assignment of their household among one of the four survey assign- ments, whether the individual is selected to be a proxy informant or a self- report, and whether the self-report or proxy assignment is realized. In the case of the self-report modules, up to two persons age 10 or older are randomly selected to self-report. If a person randomly selected to self-report are unavail- able, an alternative person is selected at random. In the case of proxy assign- ment, one person in the household age 16 or older is selected to self-report (to maximize the number of observations in the sample) and to proxy report on up to two random household members. Because the survey experiment highly emphasized the importance of avoiding proxies, the project was fairly success- ful at completing self-reports when assigned. In about 5 percent of the cases, the team was unable to interview a person selected for self-report and used a proxy informant instead. The results presented in this paper are unchanged if the observations which deviated slightly from the planned design are excluded. In this paper, the focus is on the sample of subjects age 16 and older; issues related to child labor (age 10-15) are examined in another study (Dillon et al., 2010). We further restrict the sample to households with at least two persons 11. The selection of a �xed number of 12 households for each village does not reflect the different size of the villages. This issue should not be a concern because (1) the sample is not meant to be representative of either the whole country or meaningful parts of it, and (2) the focus of the paper is a comparison across ‘similar’ groups of individuals with the purpose of highlighting differences in statistics rather than discussing the levels and meaning of those statistics for the Tanzanian labor market. For this reason, it was decided not to correct the household weights to reflect the unequal household selection probabilities across villages given that this correction would be irrelevant to the analysis. 430 THE WORLD BANK ECONOMIC REVIEW T A B L E 1 . Individual and household characteristics, by survey assignment Individual survey assignment Detailed Detailed Short Short F-test of equality of coef�cients Self-report Proxy Self-report Proxy across groups Individual characteristics Female (%) 54.8 50.3 50.0 50.4 0.309 Age 35.5 36.0 36.2 37.0 0.657 Highest school grade attended 4.9 4.8 4.9 5.0 0.922 Married (%) 63.1 65.7 68.0 61.6 0.161 Household characteristics Head: female (%) 14.6 15.2 14.3 16.2 0.781 Head: age 47.9 48.1 48.8 49.6 0.457 Head: highest school grade attended 4.8 4.8 4.8 4.8 0.993 Head: married (%) 83.9 82.8 83.9 81.8 0.674 Household size 6.6 6.5 6.6 6.9 0.438 Adult equivalence household size 5.4 5.3 5.3 5.5 0.431 Share of members less 6 years 18.3 17.8 18.2 17.8 0.798 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Share of members 6-15 years 25.1 23.9 24.3 23.9 0.443 Number of adults 16 þ years 3.5 3.6 3.5 3.7 0.293 Concrete/tile flooring (non-earth) (%) 25.8 27.7 23.6 24.5 0.669 Main source for lighting is electricity/ 11.2 11.0 11.6 13.5 0.723 generator/solar panels (%) Owns a mobile telephone (%) 33.4 35.3 33.7 37.4 0.384 Bicycle (%) 50.1 49.5 54.4 54.5 0.615 Owns any land (%) 82.3 82.4 84.4 84.3 0.859 Acres of land owned (including 0s) 3.7 3.4 3.7 3.7 0.526 Urban (%) 33.5 33.2 31.9 33.0 0.925 Month of interview (1 ¼ Jan, 5.6 5.4 5.5 5.5 0.673 12 ¼ Dec) N of individuals 734 373 720 366 Notes: See NBS (2002) for details on the adult equivalence scales. The F-test tests the equality of coef�cients across the groups by regressing the group indicators on the household character- istics with clustered household standard errors. Includes person-weights de�ned in the text. eligible to be a proxy informant (two persons 16 and older). Summary statistics for the sample are presented in Table 1. V. R E S U L T S The presentation of the results of the experiment is divided into two parts. In the �rst part, differences across the survey assignments are examined for key employment statistics on the individual’s main activity: labor force partici- pation, weekly hours, daily earnings, the sector of work, and type of work (employment status). The statistics (averages) both between the short module and the detailed module, and between responses given by proxy and self- reported responses are compared. Because a slight unbalance across experimen- tal groups persists even after weighting for unequal selection probabilities Bardasi, Beegle, Dillon, and Serneels 431 ( probably due to the relatively small sample sizes of the groups – see Table 1), we decided to run regressions (with weights) to fully control for discrepancies in the composition of the experimental groups introduced by the survey design: yi ¼ a þ bS Sh þ bP Pi þ lXi þ gDh þ 1h ðEq:1Þ where yi are the different labor statistics (like labor force participation, labor supply, earnings, and occupational choice) for the i th individual, Sh is an indi- cator variable for the short questionnaire treatment of individuals in household h, Pi is an indicator variable for the proxy treatment of individual i in house- hold h, Xi is a vector of individual and household characteristics for the i th individual, D captures district indicators, and 1 is the stochastic error term, which is randomly distributed across households. The marginal treatment effects are estimated using standard models (OLS, probit, and multinomial logit). In the second part, the impact of the characteristics of the proxy informants on the employment statistics are examined, speci�cally whether there are Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 “types of proxy informants� who generate statistics that are closer to self-reports. Differences in Labor Statistics across Survey Assignment Table 2 presents the �ndings, disaggregated by gender, for employment, weekly hours, and daily earnings. In each case, the difference in means across survey assignments is tested using a t-test. Row 1 of Table 2, for instance, reports the employment rate of men from the short module (91.2 percent) and from the detailed module (85.7 percent), and �nds that the difference (5.5 percentage points) is statistically different from zero at the 1% level. When looking at the employment rates based on the informant’s classi�- cation (i.e., derived from the one question “Did you do any type of work in the last 7 days� in the short module, and from the three screening questions specifying three main groups of economic activities in the detailed module), the short module produces higher employment rates than the detailed module, for both men and women (Table 2, top panel). This result is in contrast with what was expected a priori – that a generic and vague question about “work� would miss people in marginal activities and activities with no remuneration. However, after re-classifying domestic duties into “no work� as per the ILO de�nition, shifts in employment rates are observed, especially for women (Table 2, second panel). For men, the decrease in the employment rate is small in the short and even smaller in the detailed questionnaire, so that there is no statistically signi�cant difference in the eventual employment rates produced by the two survey instruments (88 and 85 percent, respectively). For women, however, there is a substantial number of reclassi�cations needed when using the short questionnaire. Because a very large percentage of women gets classi- �ed as “working� but is carrying out domestic duties, the percentage in 432 THE WORLD BANK ECONOMIC REVIEW T A B L E 2 . Labor statistics by survey assignment and sex A. B. Number of Short Detailed Diff Proxy Self-rep Diff observations Participation in employment (informant’s classi�cation)a (%) Men 91.2 85.7 5.5*** 81.1 92.7 2 11.6*** 1062 Women 89.5 81.5 8.0*** 84.4 85.8 2 1.4 1131 Participation in employment (after reclassi�cation of domestic duties)b (%) Men 87.9 85.0 2.9 78.5 91.1 2 12.6*** 1062 Women 75.0 80.1 2 5.1* 76.1 78.5 2 2.4 1131 Weekly hours last week unconditional on employment b (mean) Men 32.8 30.5 2.3* 28.6 33.5 2 4.9*** 1059 Women 24.6 25.9 2 1.3 25.5 25.2 0.3 1128 Weekly hours last week among working (if employment ¼ 1) b (mean) Men 37.4 35.9 1.5 36.4 36.8 2 0.4 924 Women 32.9 32.3 0.6 33.5 32.1 1.4 880 Conditional daily earnings (Tshillings) (if employment ¼ 1 and earnings . 0)b (mean) Men 5,064 3,871 1,193 5,729 3,696 2,033** 168 Women 4,803 4,505 298 5,211 4,255 956 82 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Notes: ‘Diff’ indicates the difference between the averages reported in the two preceding columns. Includes person-weights de�ned in the text. *** indicates statistical signi�cance at 1%, ** at 5%, * at 10%. a For the short questionnaire, this is the percentage of those who answer “Yes� to Question 1 (Annex Table 1, �rst column); for the detailed questionnaire, this is the percent of “Any yes� to Questions 1, 3, and 7 (Annex Table 1, second column). b Participation in employment after re-classifying those who indicated domestic duties as their main work activity (Annex Table 1 Question 4 in the �rst column and Question 9 in the second column, for the short and detailed questionnaire, respectively) into non-employment, according to the ILO de�nition. employment according to the short questionnaire decreases from almost 90 to 75 after correct reclassi�cation. The variation is much smaller in the detailed questionnaire, with only a handful of women re-classi�ed as non-working. As a result, eventual female employment based on the short module becomes about 5 percentage points lower than in the detailed module. Using proxy informants generates male employment rates that are more than 10 percentage points lower than when using self-reports. After re-classifying domestic duties into no-work, the difference between the proxy and self- reported male employment statistics remains large and statistically signi�cant (about 13 percentage points lower for proxy reports). Comparing the �rst and second panels of Table 2, for both proxy informants and self-reports, employ- ment rates are “inflated� when domestic duties are not re-classi�ed. Female employment rates, by contrast, do not differ substantially between proxy infor- mants and self-reports, although both are lower after the reclassi�cation of domestic duties into non-work. When the variation in conditional and unconditional weekly hours are examined across survey experiments, the average unconditional number of Bardasi, Beegle, Dillon, and Serneels 433 weekly hours based on reports by proxy informants is signi�cantly lower for men (about 5 hours less per week on average), but not for women; however, for employed men proxy informants report the same weekly hours as self- reports (about 36.5 per week). This result is driven by the propensity of proxy informants to report a much lower participation in employment for men. If marginal jobs are those that are being underreported with the short or proxy surveys, one would expect to see lower hours among the employed for these two groups with respect to the means generated by the detailed questionnaire and the self-reports. This is what is observed for men, for whom conditional hours are larger in the short (37.4 per week) than in the detailed module (35.9 per week – see fourth panel of Table 2), while for women the difference is smaller (32.9 and 32.3 hours per week, respectively). Results for average hours are similar when employment of subjects working more than 40 hours per week are examined (results not presented). There are differences in daily earnings between survey assignments, but because of the small number of observations (most individuals employed in Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 agriculture or as unpaid family members do not derive earnings from their activity) they are mostly not statistically signi�cant. The detailed module tends to produce lower average earnings for both men and women, and self-reporting also generates lower earnings for both men and women; however only for men the difference between proxy and self-reporting is especially large and statisti- cally signi�cant. In Table 3, the distribution across main activities by assignment is presented. Activities are classi�ed into four categories. Employed individuals are distribu- ted between agriculture and other sectors.12 The category “domestic duties� (included as a possible answer alongside other industrial sectors of activity) is kept as a separate category.13 The fourth category in Table 3, “no work� cor- responds therefore to the informant’s de�nition. Panel A in Table 3 shows that the higher male employment rate in the short module observed in Table 2, although not statistically signi�cant, stems from men being less likely to be in “no work� (about 6 percentage points). By con- trast, the lower female employment rate in the short module results primarily from a large “participation� in domestic duties as women’s main activity. In the case of the detailed module, women are much more likely to be classi�ed as “not working� than in domestic duties; in the detailed module, only 1 12. The non-agricultural sectors are too small to consider in a disaggregated manner. These include: mining/quarrying/manufacturing/processing, gas/water/electricity, construction, transport, buying and selling, personal services, education/health, and public administration. Buying and selling activities are the most frequently reported of these activities (4-7 percent, depending on the sub-group). 13. Although “domestic duties� is listed as a potential “sector of main activity,� the interviewers received clear instructions to include any domestic duties contracted outside the household in the category “personal services� (counted as employment) and classify under “domestic duties� only domestic and household work done for the household where individuals live. Careful debrie�ng con�rmed that these guidelines were strictly followed. 434 THE WORLD BANK ECONOMIC REVIEW T A B L E 3 . Sector of main activity by survey assignment and sex Men Women A. Short or Detailed Short Detailed Diff Short Detailed Diff Main activity^ Agriculture 65.4 59.1 6.3** 64.9 67.2 2 2.3 Other sectors 22.6 25.9 2 3.3 10.1 12.9 2 2.8 Domestic Duties 3.3 0.7 2.6*** 14.5 1.4 13.1*** No work 8.8 14.3 2 5.5*** 10.5 18.5 2 8.0*** N 539 523 547 584 Main activity among workers^ Agriculture 74.3 69.5 4.8 86.6 83.9 2.7 Other sectors 25.7 30.5 2 4.8 13.4 16.1 2 2.7 N 476 451 413 470 B. Proxy or Self-report Proxy Self-rep Diff Proxy Self-rep Diff Main activity^ Agriculture 53.7 67.2 2 13.5*** 63.6 67.5 2 3.9 Other sectors 24.9 23.8 1.1 12.5 11.0 1.5 Domestic Duties 2.6 1.7 0.9 8.2 7.3 0.9 No work 18.9 7.3 11.6*** 15.6 14.2 1.4 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 N 360 702 379 752 Main activity among workers^ Agriculture 68.3 73.8 2 5.5* 83.6 86.0 2 2.4 Other sectors 31.7 26.2 5.5* 16.4 14.0 2.4 N 286 641 292 591 Notes: Includes person-weights de�ned in the text. Other sectors are speci�cally listed on the questionnaire and include mining/quarrying, manufacturing/processing, gas/water/electricity, con- struction, transport, trading, personal services, education/health, public administration, and other. ‘Diff’ indicates the difference between the averages reported in the two preceding columns. *** indicates statistical signi�cance at 1%, ** at 5%, * at 10%. ^ Percentages by group may not sum to 100 due to rounding. percent of women are classi�ed in domestic duties, about 13 percentage points less with respect to the short module. This suggests that the additional ques- tions at the beginning of the detailed employment module succeed in �ltering out the large majority of individuals who would otherwise be classi�ed under domestic duties. These three questions – explicitly mentioning and exemplify- ing farm work, wage work, and work in household enterprises – frame the notion of work to exclude domestic duties in the minds of respondents. The sector composition of employment (for those employed; i.e., the distribution between agriculture and other sectors) is not signi�cantly affected by the short or detailed modules for either men or women. Comparing proxy and self-reports, the distribution across sectors is affected for men but not for women (Table 3 panel B). Proxy informants report lower employment for men as a result of a higher percentage of “no work� (11.6 per- centage points difference between proxy and self-reporting). Interestingly, the decline in male employment when moving from self-report to proxy is almost entirely accounted for by lower participation in agriculture (13.5 percentage Bardasi, Beegle, Dillon, and Serneels 435 T A B L E 4 . Employment status among employed by survey assignment and sex Men Women A. Short or Detailed ^ Short Detailed Diff Short Detailed Diff Paid employee 13.0 20.7 2 7.7*** 5.0 12.2 2 7.2*** Self-employed, with employees 3.4 4.9 2 1.5 1.1 0.9 0.2 Self-employed, no employees 61.0 57.7 3.3 23.8 21.8 2.0 Unpaid family worker 22.5 16.7 5.8** 70.1 65.1 5.0* N 474 447 411 467 B. Proxy or Self-report Proxy Self-rep Diff Proxy Self-rep Diff Paid employee 15.9 17.1 2 1.2 9.0 8.9 0.1 Self-employed, with employees 2.8 4.8 2 2.0 0.9 1.0 2 0.1 Self-employed, no employees 62.7 57.8 4.9 24.4 21.9 2.5 Unpaid family worker 18.6 20.3 2 1.7 65.7 68.2 2 2.5 N 285 636 291 587 Notes: ‘Diff’ indicates the difference between the averages reported in the two preceding columns. Includes person-weights de�ned in the text. *** indicates statistical signi�cance at 1%, ** at 5%, * at 10%. ^ Percentages by group may not sum to 100 due to rounding Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 The sample excludes respondents who indicated domestic duties as their main work activity (see Table 2 note b). points less as reported by proxy), while participation in other sectors is almost unaffected. This suggests that a large proportion of men who classify them- selves as working in agriculture are not recognized as working by other adult household members. For women, signi�cant differences between proxy and self-reports are not observed. The distribution of those in employment by status can also be examined with the data: paid employees, self-employed with employees, self-employed with no employees (often farmers), and unpaid family worker. The results are presented in Table 4. Comparing the short and detailed modules, signi�cant variations in several categories are observed, with the short module producing less paid employment for both men and women (more than 7 percentage points difference), compensated by more self-employment without employees (although not statistically signi�cant) and more unpaid work (statistically sig- ni�cant) for both men and women. The difference in paid employment is large and unexpected – one would have assumed that equating paid employment with work should have been easy even when the concept of work is not pre- cisely de�ned. The impact of asking by proxy is not very large and never stat- istically signi�cant, and results in higher self-employment without employees and lower percentages of unpaid family workers for both men and women. In this case we do not observe substantial variations in the proportion of workers who are paid employees. We estimate equation 1 to measure the marginal effect of each treatment, including the weights described above and controlling for observable character- istics. Controls include individual characteristics of the subjects (age, sex, and 436 THE WORLD BANK ECONOMIC REVIEW T A B L E 5 . Probit and OLS regressions of labor statistics by survey assignment and sex Participation in Conditional daily employment Conditional weekly hours earnings (1) Men (2) Women (3) Men (4) Women (5) Men (6) Women Short 0.015 2 0.044* 0.122** 0.055 0.142 0.081 (0.021) (0.027) (0.052) (0.050) (0.144) (0.311) Proxy 2 0.112*** 2 0.012 2 0.066 0.042 0.183 0.256 (0.024) (0.028) (0.056) (0.050) (0.161) (0.246) N 1,062 1,131 924 880 168 82 Notes: Standard errors in parentheses. Includes person-weights de�ned in the text. Columns 1-2: marginal effects of a probit regression for labor force participation are reported. See Table 2 note (b) regarding excluding domestic duties as employment. Columns 3-4: coef�cients of an OLS regression on log hours are reported. Columns 5-6: coef�cients of an OLS regression on log daily earnings are reported. Other, covariates included but not presented are characteristics of the subject (age, gender, education) his/her household (household composition and assets) and district �xed effects. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 *** indicates statistical signi�cance at 1%, ** at 5%, * at 10%. education), household characteristics, and district dummies. The results reported in Table 5 show the large effect of proxy informants on the male employment rate – 11 percentage points less by proxy than self-reporting, the omitted category. By contrast, in the case of women, it is the short question- naire that is responsible for 4 percentage points less employment with respect to the detailed module (employment in this table correctly excludes domestic duties). These results are similar to the �ndings presented in Table 2 (second panel). The results by urban and rural location are also examined (results not presented here).14 A lower employment rate for men using proxy informants is found in rural areas, but not in urban areas, where employment rates for men do not differ between proxy and self-reports. By contrast, the lower employ- ment rate for women obtained with the short module is driven by the urban areas, while in rural areas no difference between the short and detailed modules was found. Columns 3 and 4 in Table 5 report the estimates of an OLS regression for (log) weekly hours. The short questionnaire is associated with higher average hours for men (about 12 percent more), which may be explained by some underreporting of “short� or marginal jobs in the short questionnaire. The use of proxy informants, by contrast, does not affect average weekly hours.15 To assess whether this is the correct interpretation, probit models are estimated for the probability of working at least a certain number of hours (20, 30, and 40 14. In urban locations in our sample, agriculture is still a major sector, accounting for about one-third of all jobs. 15. By urban and rural location, these results are only found among rural households. Bardasi, Beegle, Dillon, and Serneels 437 weekly hour thresholds), conditional on working. Including the same covari- ates and treatment effects as in Table 5 in the regressions, the coef�cient of the short module is signi�cant for the 20-hours threshold for both men and women; speci�cally, using the short rather than the detailed module produced a 7 percentage point higher probability of jobs longer than 20 hours for those working. However, the coef�cient is smaller and not signi�cant when adopting higher thresholds (with the exception of men for the 30 hour threshold which is signi�cant at 10 percentage points). This result can be interpreted as a higher propensity of the short module to “miss� shorter jobs, while longer jobs (with the number of hours equal to or larger than 30 hours/week, for example) are equally likely to be reported by both the short and detailed modules (and this is true for both men and women). Finally, with respect to daily earnings, there are no signi�cant differences (see Columns 5 and 6 of Table 5). Using a multinomial logit, the effects of the survey assignment on the distri- bution across three employment categories is estimated: agriculture, other sectors, and the omitted category “not in employment�, including also dom- Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 estic duties. The results shown in Table 6 con�rm the large effect of proxy informants on male participation in agriculture (17 percentage points less than self-reports) with respect to “not in employment�. The different survey instru- ments do not generate signi�cant differences for women.16 Following a similar estimation approach for employment status (see Table 7), controlling for proxy assignment, the short module decreases the probability of men and women of being in wage employment compared with unpaid work, and increases the probability of men being self-employed without employees (albeit the coef�cient is not statistically signi�cant). Relying on proxy informants does not impact status in the main job for either men or women. The Impact of the Characteristics of Proxy Informants Because proxy informants were randomly selected, the dataset contains some variability of “proxy types�, which can be used to address the question of whether the proxy’s characteristics affect the proxy’s responses to the employ- ment questions.17 Ideally, these data would be used to identify proxy infor- mants’ characteristics that produce the “best� results, in order to advise survey methodologists on how to choose the proxy informant. In the absence of better 16. As a robustness check, the equations presented in Tables 4 and 5 are re-estimated dropping from the sample the responses of proxy informants who self-reported on themselves before reporting on other household members. The results are virtually unchanged – the only difference is that the coef�cient of the short questionnaire in the conditional hours equation is no longer statistically signi�cant for either men or women. This effect may be partly due to the reduction in statistical power because of the small sample size. These equations were also re-estimated excluding the set of controls, and the results are unchanged as expected given the randomized design of the survey assignment. 17. Analysis of the impacts on hours, earnings, and status in employment is not possible because of sample size considerations. 438 THE WORLD BANK ECONOMIC REVIEW T A B L E 6 . Multinomial logit of main activity by survey assignment and sex Men Women Agriculture Other Sectors Agriculture Other Sectors Short 0.031 2 0.010 2 0.034 2 0.014 (0.037) (0.031) (0.039) (0.011) Proxy 2 0.173*** 0.046 2 0.020 0.001 (0.040) (0.032) (0.043) (0.012) N 1,062 1,131 Notes: Marginal effects are reported. Includes person-weights de�ned in the text. Other cov- ariates included but not presented are characteristics of the subject (age, sex, education), his/her household (household composition and assets) and district �xed effects. The three categories of the multinomial logit model are agriculture, other sectors, and “out of work� (domestic work and no work – omitted category). See note to Table 3 for an explanation of other sectors. *** indicates statistical signi�cance at 1%, ** at 5%, * at 10%. T A B L E 7 . Multinomial logit of employment status in main job in the last 7 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 days by survey assignment and sex Men Women Self-employed Self-employed Wage with Self-employed, Wage with Self-employed, employee employees no employees employee employees no employees Short 2 0.066** 2 0.001 0.043 2 0.062*** 0.000 0.009 (0.031) (0.001) (0.036) (0.023) (0.000) (0.027) Proxy 2 0.037 2 0.001 0.053 0.009 0.000 0.005 (0.027) (0.001) (0.034) (0.022) (0.000) (0.028) N 921 878 Notes: Relative risk ratios are reported. Includes person-weights de�ned in the text. Other cov- ariates included but not presented are characteristics of the subject (age, gender, education), his/ her household (household composition and assets) and district �xed effects. The multinomial logit model is used where the four categories include wage employee, self-employed with employ- ees, self-employed without employees, and unpaid family worker (the omitted category). *** indicates statistical signi�cance at 1%, ** at 5%, * at 10%. criteria, the “best� is de�ned as the proxy report that converges to the self- report, i.e., a concordant answer (this is the de�nition adopted by other researchers, including for example Blair et al., 1991). Unfortunately, both a proxy response and a self-report for the same individual is not available – in which case, one could analyze the characteristics of the proxy informants whose answers approach the self-report for the same person. Another compli- cation arises from the fact that the characteristics of the proxy informants are not unrelated to the characteristics of the individuals on whom they report. For example, a test of whether parents are “better� proxies than other adult house- hold members can only be carried out for young people who reside with at Bardasi, Beegle, Dillon, and Serneels 439 least one parent and at least one other adult other than a parent. In comparing the reports of proxy informants with self-reports, therefore, the sample should be selected in such a way to create the correct comparison group, to make sure that the differences in employment statistics only reflect the “proxy effect� and not genuine differences in employment behavior driven by differences in house- hold composition. The �rst strategy to address this concern is to control for observed individual and household characteristics – household composition variables being especially important. However, unobserved traits related to sample selection may still be a concern. The second strategy, again, controls for observed characteristics, but also conditions the sample to the “relevant� one. This is the sub-sample of those households whose composition allows for all possible selections of the informant considered in each speci�c regression (the econo- metric motivation is one of a non-parametric matching between the proxy and self-report households restricting the matched sample to the common support). For example, when examining the impact of the gender of the proxy informant, Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 only select those households that have at least one adult male and one adult female in addition to the subject (an adult of each gender who could possibly be selected to provide proxy information). Unfortunately, this strategy results in small sample sizes and a loss of precision of the estimates. Instead, the focus is on four speci�c characteristics of proxy informants: gender, education, marital status, and headship. The results are reported in Table 8. Note that the sample changes with each regression. In all regressions, the small and not-signi�cant effect of proxy infor- mants on female employment rates is con�rmed; the substantial and statisti- cally signi�cant effect of proxy informants on male employment rates is also con�rmed. The size of the proxy effect is now generally larger, but these coef�- cients are not directly comparable with those presented in the previous regressions, because they refer to a different subsample of individuals in each column.18 All proxy informants substantially underestimate male employment participation, but proxies who are better educated and (female) spouses do better than others – the bias in both cases is “only� about 12 percentage points (for educated proxies, column 3) and 7 percentage points (for female spouses, column 5). Choosing the head of the household to report on male par- ticipation (column 7) does not produce “better� statistics than when other adults are selected as proxies. However, the extremely large coef�cients esti- mated in this case for a non-head proxy respondent (about 23 percentage points) is very likely driven by the characteristics of the subject. Men who are not the head and reside in a household with at least one other adult who is not the head are likely to be adult children. They may be disproportionately found in marginal, casual, or irregular employment. In both cases, the proxy may 18. This also explains why the impact of the short vs. detailed questionnaire, which is signi�cant for women in the general speci�cation, now differs across household types. 440 T A B L E 8 . Participation in employment by proxy characteristics (1) (2) (3) (4) (5) (6) (7) (8) Subject sample: Men Women Men Women Men Women Men Women Short 0.023 2 0.051 0.025 2 0.099 0.017 2 0.045* 0.032 2 0.049* (0.033) (0.048) (0.026) (0.075) (0.021) (0.027) (0.042) (0.029) Proxy 2 0.166*** 2 0.016 2 0.183*** 2 0.083 2 0.167*** 2 0.029 2 0.229*** 2 0.057 (0.044) (0.058) (0.053) (0.094) (0.030) (0.036) (0.053) (0.049) Proxy is man 0.009 2 0.018 (0.045) (0.078) Proxy has schooling 0.062** 0.050 (0.025) (0.094) Proxy is spouse 0.096*** 0.041 THE WORLD BANK ECONOMIC REVIEW (0.019) (0.045) Proxy is head 0.015 0.063 (0.060) (0.046) N 382 421 211 182 1,062 1,131 316 1,014 Notes: Standard errors in parentheses. Includes person-weights de�ned in the text. Marginal effects of a probit regression for participation in employment are reported. Domestic duties are classi�ed as non-employment (see Table 2 note (b)). Other covariates included but not presented are characteristics of the subject (age, sex, education), his/her household (household composition and assets), and district �xed effects. Columns 1 and 2 are restricted to subjects who reside in a household with at least 1 other adult male and 1 other adult female. Columns 3 and 4 are restricted to subjects who reside in a household with at least 1 other adult with schooling and 1 other adult with no schooling. Columns 5 and 6 are restricted to subjects who reside in a household with at least 1 spouse and one other adult. Columns 7 and 8 are restricted to subjects who are not the head and reside in a household with at least 1 other adult who is not the head. *** indicates statistical signi�cance at 1%, ** at 5%, * at 10%. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Bardasi, Beegle, Dillon, and Serneels 441 have poor knowledge of the male subject’s participation in employment. Still, in the case of women, choosing the head does not offer any obvious advantage – heads and non-heads both do pretty well. Cost Implications of Survey Design Alternative survey designs will have cost implications that have to be weighed against a subjective valuation of “better� data. Using the detailed module added only a few minutes to the average duration of the interview compared with the shorter module, according to �eld work reports from enumerators and supervisors. The cost implication of using a detailed rather than a short module, therefore, is small. The additional cost of printing slightly longer ques- tionnaires and the extra data entry requirement are also only marginal. By contrast, using proxy instead of self-reports involves substantial savings. The use of self-reports increases the length of �eld work because more days are Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 spent in each sample village to locate and interview respondents. Since this survey experiment was carried out in conjunction with a larger consumption expenditure experiment, it required survey teams to spend a full two weeks in each village. This ensured that self-reports were completed but, as a conse- quence, it is dif�cult to report on precisely how many extra days are needed to complete self-reports compared with proxy labor modules. However, based on �eld experience, a rough calculation can be made that for two days spent in a village using proxy informants, the survey team would need at least one more day to track down self-reports. This corresponds to a 50 percent increase in the length of time spent on actual �eld work. All variable costs of �eld staff ( per diems, lodging costs), often the largest category of survey costs, are assumed to increase by 50 percent. Transport costs may also increase if �eld teams use a team vehicle to track down respondents for self-reports. Our results make the valuation of the trade-off between better data and lower survey costs complicated. For women, proxy reporting has no effect on female labor statistics in the sample. Given the higher costs associated with self-reporting, there seems to be no substantial bene�t for insisting on self- reporting for women. This is similar to other �ndings on survey design and child labor (Dillon et al., 2010). Men’s labor statistics, however, particularly employment rates and participation in agriculture, are substantially affected by proxy reporting. Given the heterogenous effects of survey design within this sample and the magnitude of the proxy effect on men’s labor force partici- pation and probability of working in agriculture, serious reflection by survey designers with respect to the objectives and precision necessary in their surveys is paramount. For national surveys on labor activities, self-reporting appears to be the best practice. However, further research on the reliability of self- reporting itself is necessary to con�rm this conjecture, as well as validation of the results in other contexts. 442 THE WORLD BANK ECONOMIC REVIEW VI. CONCLUSIONS Despite the importance of household survey instruments as a source for labor statistics, there is a dearth of evidence on the best practices for collecting these statistics in developing countries. The differences in survey design for national labor statistics over time within a country and across countries have serious implications for both measuring labor market outcomes and carrying out research on labor activities. This paper presents a survey experiment focusing on two key aspects of survey design: the questions to identify participation in employment and the choice of the respondent. The statistics generated from short and detailed questionnaires, and by a proxy informant and self-reporting are compared. The �ndings suggest that both types of survey design decisions have statisti- cally signi�cant effects on labor statistics. These effects are largest on employ- ment participation rates, but also exist for weekly hours of work, main activity, and type of work. The effects vary depending on the labor market outcome and remain signi�cant even after controlling for individual, house- Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 hold, and village characteristics and sample composition. The short question- naire captures a large percentage of women “employed� who are actually engaged in domestic duties; but reclassi�cation of these cases into “no-work� produces a lower female employment rate than the detailed questionnaire. Moreover, it also tends to generate higher average weekly hours conditional on working, especially for men, as well as a lower share of paid employees among the employed. Response by proxy leads to substantially lower male employ- ment rates due to lower reporting of participation in agriculture – while for women there is no effect. The discrepancies between proxy and self-reporting are reduced when the spouse is selected as the proxy and when the proxy is an individual with some education. In this paper, only two dimensions of survey design are the focus; future work is needed to look at other issues, such as the wording of the questions. But even the (limited) results presented here provide some clear advice for survey design. First, the impacts are not consistently associated with one speci�c survey feature but differ for different types of individuals. Using a short rather than a detailed module produces lower employment for women but not for men; using proxies rather than asking the individual directly strongly impacts (negatively) employment for men but not for women. This indicates that the “best� approach – if it exists – may differ depending on the purpose of the survey, which determines the type of variables and the type of sample for which information is collected. Second, the inclusion of a “domestic duties� category as a possible answer to the “main sector of activity� can produce ambiguous results, especially in a short employment module, and in particular for women. Individuals may clas- sify themselves (or be classi�ed) as “employed� when responding to a very direct yes/no question about working, but then de facto rule themselves out of Bardasi, Beegle, Dillon, and Serneels 443 employment (based on the SNA de�nition of economic activity) by indicating that their main activity was being engaged in domestic duties. When using a short module, ambiguity remains as to how these individuals interpret the meaning of “employment� and “domestic duties� and whether they tend to give priority to “domestic duties� as a main activity over other activities in which they may have worked as well. Without information about other jobs, this ambiguity is left unresolved. Third, in a low-income setting, the distinction between agricultural work and “no work� can be subject to interpretation. The proxy informants are much more likely to categorize men out of work rather than working in agri- culture with respect to self-reports. Future work may want to investigate this issue further. Fourth, given the – sometimes unexpected – variations that distinct survey designs can generate, these results underline the importance of staying with the same design if the aim is to make comparison, both over time or across contexts. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 If the detailed self-reported module is best practice – as recommended by ILO guidelines – these results provide an assessment of the implications for data quality of using less expensive approaches, such as response by proxy or using a shorter module. Note that the trade-off between higher quality of data and lower implementation costs is especially driven by the choice of using proxy informants vs. self-reports, and not so much by the small number of additional questions in the detailed questionnaire. These �ndings suggest that there is no substantial bene�t of self-reporting for women (a similar result was found for children, as discussed in Dillon et al., 2010), but there is for men, whose employment rates and distribution across sector of activity are affected by proxy reporting. The results also imply that the detailed questionnaire should be maintained as best practice approach, because using the short ques- tionnaire does not involve relevant cost savings but may affect some of the stat- istics (in this case, female employment rates, male working hours, and the employment status of both men and women). This research also points to a number of fruitful avenues for future work. A common frustration with this type of work is that there is no clear evidence of a gold standard (yet). While self-reporting is generally accepted as best practice, further research on the reliability of self-reporting itself is necessary to con�rm this conjecture. The experiment analyzed the convergence of proxy and self- reports, but did not address the issue of whether any of these reports actually correspond to the “truth�. Moreover, the large percentage of informants that identify “domestic duties� with “employment� or are unable to distinguish between working in agriculture and being out of work suggests that the notion of “employment� should not be taken as given, but needs to be de�ned and clari�ed, especially in low-income countries. This is as much an analytical as a methodological question. It leaves open the question of how questionnaires that have been mostly developed with reference to advanced economies can be 444 THE WORLD BANK ECONOMIC REVIEW further improved to make them fully adequate to capture “employment� in completely different settings. A n n ex T A B L E 1 . Key employment questions in the short and detailed questionnaires Short questionnaire Detailed questionnaire 1. During the past 7 days, has [NAME] worked for someone who is not a member of your household, for example, an enterprise, company, the government or any other individual? YES . . .1 (»3) NO . . . 2 (question repeated for the past 12 months – question 2) 3. During the past 7 days, has [NAME] Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 worked on a farm owned, borrowed or rented by a member of your household, whether in cultivating crops or in other farm maintenance tasks, or has [NAME] cared for livestock belonging to a member of your household? YES . . .1 (»5) NO . . . 2 (question repeated for the past 12 months – question 4) 5. During the past 7 days, has [NAME] worked on his/her own account or in a business enterprise belonging to him/her or someone in your household, for example, as a trader, shop-keeper, barber, dressmaker, carpenter or taxi driver? YES . . .1 (»7) NO . . . 2 (question repeated for the past 12 months – question 6) 1. Did [NAME] do any type of work in the last 7. CHECK THE ANSWERS TO seven days? QUESTIONS 1, 3 AND 5. (WORKED Even if for 1 hour. IN LAST 7 DAYS) YES . . .1 (»3) ANY YES.. 1 NO . . . 2 ALL NO . . . 2 (»37) (question repeated for the past 12 months – question 2) 3. What is [NAME]’s primary occupation in 8. What is [NAME]’s primary occupation in [NAME]’s main job? [NAME]’s main job? (Continued ) Bardasi, Beegle, Dillon, and Serneels 445 ANNEX TABLE 1. Continued Short questionnaire Detailed questionnaire (MAIN OCCUPATION IN THE LAST 7 DAYS) (MAIN OCCUPATION IN THE LAST 7 DAYS) a. OCCUPATION a. OCCUPATION b. OCCUPATION CODE b. OCCUPATION CODE 4. In what sector is this main activity? 9. In what sector is this main activity? AGRICULTURE. . . . . . . . . . . . . . . . 1 AGRICULTURE. . . . . . . . . . . . . .1 MINING/QUARRYING. . . . . . . . . . . . .2 MINING/QUARRYING. . . . . . . . . . 2 MANUFACTURING/ PROCESSING. . . . . . . 3 MANUFACTURING/ PROCESSING. . . . .3 GAS/WATER/ELECTRICITY. . . . . . . . . 4 GAS/WATER/ELECTRICITY. . . . . . . 4 CONSTRUCTION. . . . . . . . . . . . . . . 5 CONSTRUCTION. . . . . . . . . . . . . 5 TRANSPORT. . . . . . . . . . . . . . . . . 6 TRANSPORT. . . . . . . . . . . . . . . 6 BUYING AND SELLING. . . . . . . . . . . . 7 BUYING AND SELLING. . . . . . . . . .7 PERSONAL SERVICES. . . . . . . . . . . . 8 PERSONAL SERVICES. . . . . . . . . . 8 EDUCATION/HEALTH. . . . . . . . . . . . .9 EDUCATION/HEALTH. . . . . . . . . . 9 PUBLIC ADMINISTRATION. . . . . . . . . 10 PUBLIC ADMINISTRATION. . . . . . .10 DOMESTIC DUTIES. . . . . . . . . . . . . 11 DOMESTIC DUTIES. . . . . . .. . . . 11 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 OTHER, SPECIFY. . . . . . . . . . . . . .12 OTHER, SPECIFY. . . . . . . . . . . 12 REFERENCES Abowd, J.M., and A. Zellner 1985. “Estimating Gross Labor-Force Flows.� Journal of Business & Economic Statistics 3(3):254–283. Anker, R. 1983. “Female Labour Force Participation in Developing Countries: A Critique of Current De�nitions and Data Collection Methods.� International Labour Review 122(6):709– 724. Bardasi, E., K. Beegle, A. Dillon, and P. Serneels 2010. “Do Labor Statistics Depend on How and to Whom the Questions Are Asked? Results from a Survey Experiment in Tanzania.� World Bank Policy Research Working Paper 5192. Biemer, P., R. Groves, L. Lyberg, N. Mathiowetz, and S. Sudman 1991. Measurement Error in Surveys. New York: John Wiley & Sons. Biggs, B. 1992. “Self/Proxy informant Rules and Data Quality.� Research Paper, Income Research Paper Series, Ottawa: Statistics Canada. Blair, J., G. Menon, and B. Bickart 1991. “Measurement Effects in Self vs. Proxy Responses to Survey Questions: An Information-Processing Perspective.� in P P. Biemer, R. Groves, L. Lyberg, N. MathiowetzS. Sudman (eds.) Measurement Error in Surveys. New York: John Wiley & Sons. Boehm, L.M. 1989. “Reliability of Proxy Response in the Current Population Survey.� in Proceedings of the Survey Research Methods Section, American Statistical Association. Bound, J., C. Brown, and N. Mathiowetz 2001. “Measurement Error in Survey Data.� In Handbook of Econometrics Vol. 5. ed. J. Heckman, and E. Leamer Amsterdam: North-Holland, Elsevier Science. Campanelli, P., J.M. Rothgeb, and E.A. Martin 1989. The Role of Respondent Comprehension and Interviewer Knowledge in CPS Labor Force Classi�cation. American Statistical Association Proceedings (Survey Research Methods Section). Charmes, J. 1998. Women Working in the Informal Sector in Africa: New Methods and New Data. Paris: Scienti�c Research Institute for Development and Co-operation. Dillon, A., E. Bardasi, K. Beegle, and P. Serneels 2010. “What Explains Variation in Child Labor Statistics? Evidence from a Survey Experiment in Tanzania.� World Bank Policy Research Working Paper 5414. 446 THE WORLD BANK ECONOMIC REVIEW Dixon-Mueller, R., and R. Anker 1988. Assessing Women’s Economic Contributions to Development. Training in Population, Human Resources and Development Planning Paper number 6, Geneva: International Labour Of�ce. Esposito, J.L., P.C. Campanelli, J. Rothgeb, and A.E. Polivka 1991. “Determining which Questions Are Best: Methodologies for Evaluating Survey Questions.� Proceedings of the American Statistical Association (Survey Research Methods Section). P. Glewwe, and M. Grosh (eds). 2000. Designing Household Survey Questionnaires for Developing Countries: Lessons from 15 Years of the Living Standards Development Study. Oxford University Press (for the World Bank). Guarcello, L., I. Kovrova, S. Lyon, M. Manacorda, and F.C. Rosati 2009. “Towards Consistency in Child Labour Measurement: Assessing the Comparability of Estimates Generated by Different Survey Instruments.� Understanding Children’s Work Project Draft Working Paper. Hausman, J.A. 2001. “Mismeasured Variables in Econometric Analysis: Problems from the Right and Problems from the Left.� Journal of Economic Perspectives 15 (4):57–67. Hausman, J.A., J. Abrevaya, and F.M. Scott-Morton 1998. “Misclassi�cation of the Dependent Variable in a Discrete Response Setting.� Journal of Econometrics 87: 239– 269. Hill, D.H. 1987. “Response Errors in Labor Surveys: Comparisons of Self and Proxy Reports in the Survey of Income and Program Participation (SIPP).� In Proceedings of the Bureau of Census, Third Annual Research Conference. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Hussmanns, R., F. Mehran, and V. Verma 1990. Surveys of Economically Active Population, Employment, Unemployment and Underemployment: An ILO Manual on Concepts and Methods. ILO: Geneva. Hyslop, D.R., and G.W. Imbens 2001. “Bias from Classical and Other Forms of Measurement Error.� Journal of Business & Economic Statistics 19(4): 475– 481. ILO. 1982. Resolution Concerning Statistics of the Economically Active Population, Employment, Unemployment and Underemployment. Adopted by the Thirteenth International Conference of Labour Statisticians, Geneva: International Labour Organization. Judge, G., and L. Schechter 2009. “Detecting Problems in Survey Data Using Benford’s Law.� Journal of Human Resources 44(1):1– 24. Kalton, G., and H. Schuman 1982. “The Effect of the Question on Survey Responses: A Review.� Journal of the Royal Statistical Society 145(1):42 –57. Kasprzyk, D. 2005. “Measurement error in household surveys: sources and measurement.� In Household Sample Surveys in Developing and Transition Countries, New York : United Nations, pp. 171–198. Martin, E., and A.E. Polivka 1995. “Diagnostics for Redesigning Survey Questionnaires: Measuring Work in the Current Population Survey.� Public Opinion Quarterly 59:547– 567. Martin, J., and B. Butcher 1982. “The Quality of Proxy Information: Some Results from a Large Scale Study.� The Statistician 31:293– 319. Mata Greenwood, A. 2000. Incorporating Gender Issues in Labour Statistics. Geneva: International Labour Of�ce, Bureau of Statistics. Mathiowetz, N.A., and R.M. Groves 1985. “The Effects of Respondent Rules on Health Survey Reports.� American Journal of Public Health 75(6):639– 644. Moore, J. 1988. “Self/Proxy Response Status and Survey Response Quality: A Review of the Literature.� Journal of Of�cial Statistics 4:155–172. National Bureau of Statistics Tanzania. 2003. Integrated Labour Force Survey, 2000/01 – Analytical Report, see http://www.nbs.go.tz/labourforce/index.htm ———. 2002. Household Budget Report, 2000/01, Chapter 5. Dar es Salaam; http://www.tanzania. go.tz/hbs/Index_FinalReport_HBS.htm Of�ce for National Statistics (ONS). 2003. Labour Force Survey User Guide: Volume 1 Background and Methodology. United Kingdom. Bardasi, Beegle, Dillon, and Serneels 447 Poterba, J.M., and L.H. Summers 1986. “Reporting Errors and Labor Market Dynamics.� Econometrica 54(6):1319 – 1338. ——— 1995. “Unemployment Bene�ts and Labor Market Transitions: A Multinomial Logit Model with Errors in Classi�cation.� Review of Economics and Statistics 77(2): 207–216. Sinclair, M.D., and J.L. Gastwirth 1998. “Estimates of the Errors in Classi�cation in the Labour Force Survey and Their Effect on the Reported Unemployment Rate.� Survey methodology 24(2): 157– 169. Singh, A.C., and J.N.K. Rao 1995. “On the Adjustment of Gross Flow Estimates for Classi�cation Error with Application to Data from the Canadian Labour Force Survey.� Journal of the American Statistical Association 90(430): 478– 488. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Entrepreneurship and Development: The Role of Information Asymmetries Leora F. Klapper and Inessa Love This article reviews the literature on the relationship between entrepreneurship and economic development and introduces four symposium articles. A common thread is that information asymmetries are important determinants of access to �nance in young entrepreneurial �rms. Policy recommendations are proposed that would increase the positive role of entrepreneurship in economic development. JEL codes: G18, G38, L51, M13 The relationship between entrepreneurship and economic development is often Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 studied but seldom agreed on, perhaps because of the complexity of the term entrepreneurship. There are numerous de�nitions of entrepreneurship but little agreement among researchers on the appropriate analytical framework (Venkataraman 1997). This article briefly surveys the research on entrepreneur- ship in a development context and introduces the four articles included in this special symposium issue. Most discussions of entrepreneurship and economic development arrive at some sort of distinction between innovative entrepreneurs and replicative entre- preneurs (Baumol 2010). Innovative entrepreneurs introduce a new technology or bring an existing idea to a new market (also known as gazelles, high impact entrepreneurs, and opportunity entrepreneurs). Schumpeter (1934) was likely referring to innovative entrepreneurs when he extolled entrepreneurs as the engine driving the creative destruction central to capitalism. Others, too, see entrepreneurship as a driving force of economic growth (Hause and Du Rietz 1984; Schramm 2006; Acs, Desai, and Hessels 2008; Baumol 2010). This positive view of entrepreneurship has led policymakers to focus time and resources on programs to aid entrepreneurs. For instance, U.S. Leora Klapper (corresponding author) is lead economist in the development research group at the World Bank; her email address is lklapper@worldbank.org. Inessa Love is senior economist in the development research group at the World Bank; her email address is ilove@worldbank.org. The research for this article was �nanced by the Ewing Marion Kauffman Foundation and the World Bank Group. The authors thank Douglas Randall for outstanding research assistance. This article’s �ndings, interpretations, and conclusions are entirely those of the authors and do not necessarily represent the views of the World Bank, their Executive Directors, or the countries they represent. THE WORLD BANK ECONOMIC REVIEW, VOL. 25, NO. 3, pp. 448– 455 doi:10.1093/wber/lhr044 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 448 Klapper and Love 449 F I G U R E 1. The Relationship between Economic Development and Formal Entrepreneurship, Average 2004–09 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Source: World Bank 2010. policymakers have called for more government support of entrepreneurs to help the country regain a competitive lead in the world economy (Schramm 2006; Baumol and Strom 2007). The European Union’s Lisbon Declaration of March 2000 explicitly identi�es entrepreneurship as the key to the European Union becoming the most competitive world region by 2010 (Naude ´ 2008). Cross-country comparisons are greatly hampered by the shortage of interna- tionally comparable data on entrepreneurship. The 2010 World Bank Entrepreneurship Snapshots, which measure new �rm entry in the formal private sector in 112 economies, make a valuable contribution.1 Using these data, Klapper and Love (2010) �nd a positive and signi�cant relationship between income level and new �rm entry (�gure 1). An important drawback of the data is the exclusion of the informal sector, where the majority of survival entrepreneurs function. However, because high-growth entrepreneurs are more likely to formally register their �rms, the data allow researchers to consider which form of entrepreneurship is more likely to have a positive impact on growth and development. However, if entrepreneurship is de�ned by its most common manifestation, a different conclusion emerges. Worldwide, most entrepreneurs are replicative entrepreneurs (also known as necessity entrepreneurs and survival entrepre- neurs) rather than innovative entrepreneurs. A large number of these entrepre- neurs have been forced into entrepreneurship and are likely to be 1. Data are available at http://econ.worldbank.org/research/entrepreneurship. 450 THE WORLD BANK ECONOMIC REVIEW F I G U R E 2. The Relationship between Economic Development and Self-Employment, 2008 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Source: Gallup World Poll 2011. self-employed, without other employees in their establishment. The relationship between this form of entrepreneurship and economic development has been consistently documented to be negative (Kuznets 1966; Acs, Audretsch, and Evans 1994; Acs and Varga 2005; Schneider, Buehn, and Montenegro 2010). Data on self-reported employment status from the 105 countries included in the 2010 Gallup World Poll (Gallup 2011) con�rm these �ndings (�gure 2). The articles in this symposium issue cover many forms of entrepreneurship. It may therefore make sense to consider the group as a whole, which leads to the literature arguing for a U-shaped relationship between entrepreneurship and economic development (GDP per capita). By this view, poor countries have high levels of necessity entrepreneurs; middle-income countries have few entrepreneurs, as their economies are driven by large manufacturing �rms; and high-income economies have many innovative entrepreneurs, likely concen- trated in the services sector (Blau 1987; Acs, Audretsch, and Evans 1994; Carree and others 2002; Wennekers and others 2005). This aligns with Porter, Sachs, and McArthur’s (2002) classi�cation of the three stages of economic development: factor driven, ef�ciency driven, and innovation driven. The U-shaped model has been criticized, however, for its empirical limitations and lack of explanatory depth (Carree and others 2007; Acs, Desai, and Hessels 2008). In response, Acs (2008) uses data from the Global Entrepreneurship Monitor to develop an opportunity entrepreneur –necessity entrepreneur ratio, which results in a low ranking for countries with high levels of replicative entre- preneurship and restores a positive linear relationship between economic devel- opment and entrepreneurship. (For a more detailed discussion, see Acs 2008). Klapper and Love 451 Still unanswered, however, is whether the relationship between entrepreneur- ship and economic development (whether U-shaped or linear, positive or nega- tive) is a result of distortionary policies or whether it represents the ef�cient distribution of capital and labor at different stages of development. If the relationship is indeed U-shaped, can countries move more quickly from stage two to stage three by implementing reforms to spur new �rm creation, or will the rate of new �rm creation remain stubbornly tied to the natural pace of economic development? Gollin (2008) provides a partial answer using a Japan-based model to demonstrate that it is ef�cient in poor countries for many lower skilled people to remain self-employed. If there is any consensus emerging from the entrepreneurship literature, it is that more research is needed on both a cross-country and a micro level. The articles in this issue each contribute cutting-edge research and methods to the study of entrepreneurship and expand our understanding in several directions. I. INFOR MATIO NAL BA RRIERS TO ENTREPRENEURSHIP Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 It is dif�cult to understand the macroeconomic dimensions of entrepreneurship and the appropriate policy measures to promote high-growth entrepreneurship without understanding the micro dimensions of how �rms are created, grow, and survive. Understanding �rm dynamics, especially for young �rms, is a valuable stream of entrepreneurship research. An important question is how the informational asymmetries of new and young �rms can impede access to new capital. Unlike large �rms, which can leverage their internal funds and reputations to access formal �nancing, small �rms must often resort to costly informal �nancing, which can be risky and constrain �rm growth (Myers and Majluf 1984; Bulan and Yan 2009; Brealey and Myers 2002; Djankov and others 2002; Carpenter and Rondi 2000). In the �rst article in this symposium issue, Chavis, Klapper, and Love (2011) use the World Bank Enterprise Survey data,2 a comprehensive dataset on 70,000 �rms in over 100 countries, to empirically demonstrate systematic differences in the use of �nancing sources for new and older �rms. Across all country income groups, the authors �nd that younger �rms rely less on bank �nancing and more on informal �nancing. For instance, the percentage of �rms using bank �nancing increases as �rms mature and almost doubles by the time �rms reach 13 years, relative to new �rms. This work contributes to the litera- ture linking information asymmetries and access to �nance. For instance, new �rms have more severe �nancing constraints because of lack of existing banking relationships (Carpenter and Rondi 2000; Bulan and Yan 2009), while mature �rms can leverage their internal funds and reputations to obtain bank �nancing (Myers and Majluf 1984; Bulan and Yan 2009; Brealey and Myers 2002; Carpenter and Rondi 2000). 2. Survey data are available at www.enterprisesurveys.org. 452 THE WORLD BANK ECONOMIC REVIEW Informational asymmetries also likely contribute to the disconnect between formal �nancing and small �rms in low-income countries, where formal edu- cation and �nancial access are often restricted to the upper class. De Mel, McKenzie, and Woodruff (2011) make an important contribution to the emer- ging body of research on informational asymmetries impeding �nancial access and �rm growth. Using a randomized experiment in Sri Lanka designed to reduce informational and procedural barriers to access to �nance, the authors �nd that loan take-up increases signi�cantly when entrepreneurs receive infor- mation on loan products and lending requirements are relaxed. Their work contributes to the literature on impact of well-targeted information on behavior (Jensen 2010; Dupas 2006; Duflo and Saez 2003). Taking a marketing angle, Gine, Mansuri, and Picon (2011) also extend the literature on informational asymmetries in access to �nance. Their work touches on the importance of women in entrepreneurship, a topic receiving greater attention over the past decade (Yunus 1999). In a randomized con- trolled trial, the authors �nd a gender bias in marketing for a micro�nance Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 product targeting entrepreneurs. Men were more likely to respond to images of male business owners, while potential women entrepreneurs had no signi�cant response to images of woman business owners. The results suggest that more powerful cues might be necessary to generate responses to role models for women. This work also relates to the literature on the role of information in the behavior of entrepreneurs. The article by Lederman, Rodriguez-Clare, and Xu (2011) provides a novel description of the contribution of new exporting �rms, products, and market destinations to the growth of exports of a successful middle-income country, Costa Rica, and compares the results with those in the nascent international trade literature focusing on the “extensive� margin of trade. The �ndings might appear to be consistent with the pessimistic view that new formal exporting �rms contribute little to annual export growth, even though new exporting �rms account for more than 30 percent of total exporting �rms in a given year. More than 95 percent of annual export growth is due to incumbent exporting �rms. Since the authors’ data from customs transactions cover only formal �rms, this �nding might appear at odds with the view that entrepre- neurship is low in middle-income countries, but it is consistent with the view that entrepreneurship contributes little to overall economic growth (to the extent that export growth is progrowth). However, the real contribution of export entrepreneurs to national export growth might transcend what is visible in accounting exercises: new exporting �rms or new export products might provide information to incumbent �rms about how and where to export new products, even if the new entrepreneurs fail. Thus, the informational contri- bution of new exporting �rms can be an important support to growth and development (see also Lederman, Olarreaga, and Payton 2006; Fajnzylber, Guasch, and Lo ´ pez 2009; Volpe and Carballo 2008). Klapper and Love 453 II. CONCLUSION The articles in this symposium issue highlight the information barriers for new �rms and the importance of information sharing and transparency for �rm entry, growth, and survival. The articles suggest that policymakers need to recognize the importance of relaxing information asymmetries to promote high-growth entrepreneurship that can contribute to job creation and macro- economic growth. In particular, the articles address some of the key information asymmetries of new and young �rms, an endemic barrier to entrepreneurship worldwide. First, entrepreneurs’ limited track record with suppliers and buyers affects new �rms’ ability to obtain external �nance. Second, inadequate information on the loan application process and loan terms may limit the ability of entrepreneurs to obtain loans and grow their businesses. Third, gender biases may lessen the ability of targeted interventions and policies to increase access to �nance. Finally, new exporters can help overcome information disadvantages, enabling incumbent �rms to grow their exports. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Taken together, these articles suggest that policies to improve institutions by promoting information sharing may disproportionately help new �rms. For instance, credit bureaus and registries that allow entrepreneurs to build per- sonal and business credit histories can alleviate information asymmetries with creditors and improve access to start-up �nancing. Similarly, �nancial consu- mer protection laws requiring disclosure of credit terms, fees, and policies to improve business education and �nancial literacy can aid less experienced small borrowers. Finally, export-promotion policies and agencies can assist small exporting �rms in dealing with the challenges of information gathering on prices, market, and �nancing opportunities. REFERENCES Acs, Z. J. 2008. “Foundations of High Impact Entrepreneurship.� Foundations and Trends in Entrepreneurship 4 (6): 535–620. Acs, Z. J., D. B. Audretsch, and D. S. Evans. 1994. “Why Does the Self-employment Rate Vary across Countries and over Time?� CEPR Discussion Paper 871. Center for Economic and Policy Research, London. Acs, Z. J., S. Desai, and J. Hessels. 2008. “Entrepreneurship, Economic Development, and Institutions.� Small Business Economics 31 (3): 219–34. Acs, Z. J., and A. Varga. 2005. “Entrepreneurship, Agglomeration, and Technological Change.� Small Business Economics 24 (3): 323–34. Baumol, W. J. 2010. The Microtheory of Innovative Entrepreneurship. Princeton, NJ: Princeton University Press. Baumol, W. J., and R. J. Strom. 2007. “Entrepreneurship and Economic Growth.� Strategic Entrepreneurship Journal 1 (3– 4): 233– 37. Blau, D. 1987. “A Time-series Analysis of Self-employment in the United States.� Journal of Political Economy 95 (3): 445 –46. 454 THE WORLD BANK ECONOMIC REVIEW Brealey, R. A., and S. C. Myers. 2002. Financing and Risk Management. New York: McGraw-Hill. Bulan, L., and Z. Yan. 2009. “The Pecking Order of Financing in the Firm’s Life Cycle.� Banking and Finance Letters 1 (3): 129– 40. Carpenter, R. E., and L. Rondi. 2000. “Italian Corporate Governance, Investment, and Finance.� CERIS-CNR Working Paper 14/2000. Institute for Economic Research on Firms and Growth, Moncalieri, Italy. Carree, M., A. van Stel, R. Thurik, and S. Wennekers. 2002. “Economic Development and Business Ownership: An Analysis Using Data of 23 OECD Countries in the Period 1976– 1996.� Small Business Economics 19 (3): 271– 90. ———. 2007. “The Relationship between Economic Development and Business Ownership Revisited.� Entrepreneurship and Regional Development 19 (3): 281– 91. Chavis, L. W., L. Klapper, and I. Love. 2011. “The Impact of the Business Environment on Young Firm Financing.� World Bank Economic Review 25 (3). Djankov, S., I. Lieberman, J. Mukherjee, and T. Nenova. 2002. “Going Informal: Bene�ts and Costs.� World Bank, Washington, D.C. Duflo, E., and E. Saez. 2003. “The Role of Information and Social Interactions in Retirement Plan Decisions: Evidence from a Randomized Experiment.� Quarterly Journal of Economics 118: 815 –42. Dupas, P. 2006. “Relative Risks and the Market for Sex: Teenagers, Sugar Daddies and HIV in Kenya.� Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Dartmouth University, Hanover, NH. ´ pez, eds. 2009. Does the Investment Climate Matter? Fajnzylber, P., J. L. Guasch, and J. H. Lo Microeconomic Foundations of Growth in Latin America. Washington, D.C.: World Bank. Gallup, Inc. 2011. Gallup World Poll Database. https://worldview.gallup.com/signin/login. aspx?ReturnUrl=%2f. Gine, X., G. Mansuri, and M. Picon. 2011. “Does a Picture Paint a Thousand Words? Evidence from a Microcredit Marketing Experiment.� World Bank Economic Review 25 (3). Gollin, D. 2008. “Nobody’s Business but My Own: Self-employment and Small Enterprise in Economic Development.� Journal of Monetary Economics 55 (2): 219– 33. Hause, J. C., and G. Du Rietz. 1984. “Entry, Industry Growth, and the Microdynamics of Industry Supply.� Journal of Political Economy 92 (4): 733 –57. Jensen, R. 2010. “The (Perceived) Returns to Education and the Demand for Schooling.� Quarterly Journal of Economics 125 (2): 515–48. Klapper, L., and I. Love. 2010. “The Impact of the Financial Crisis on New Firm Registration.� Policy Research Working Paper 5444. World Bank, Washington, D.C. Kuznets, S. 1966. Modern Economic Growth: Rate, Structure, and Spread. New Haven, CT: Yale University Press. Lederman, D., M. Olarreaga, and L. Payton. 2006. “Export Promotion Agencies: What Works and What Doesn’t?� Policy Research Working Paper 4044. World Bank, Washington, D.C. Lederman, D., A. Rodriguez-Clare, and D. Y. Xu. 2011. “Entrepreneurship and the Extensive Margin in Export Growth: A Microeconomic Accounting of Costa Rica’s Export Growth during 1997– 2007.� World Bank Economic Review 25 (3). de Mel, S., D. McKenzie, and C. Woodruff. 2011. “Getting Credit to High Return Microentrepreneurs: The Results of an Information Intervention.� World Bank Economic Review 25 (3). Myers, S., and N. Majluf. 1984. “Corporate Financing and Investment Decisions When Firms Have Information That Investors Do Not Have.� Journal of Financial Economics 13 (2): 187–221. ´ , W. 2008. Entrepreneurship in Economic Development. UNU-WIDER Research Paper 2008/20. Naude Basingstoke, U.K.: Palgrave Macmillan. Porter, M., J. Sachs, and J. McArthur. 2002. “Executive Summary: Competitiveness and Stages of Economic Development.� In M. Porter, J. Sachs, P. K. Cornelius, J. W. McArthur, and K. Schwab, eds., The Global Competitiveness Report 2001–2002. New York: Oxford University Press. Klapper and Love 455 Schneider, F., A. Buehn, and C. E. Montenegro. 2010. “Shadow Economies All Over the World: New Estimates for 162 Countries from 1999 to 2007.� Policy Research Working Paper 5356. World Bank, Washington, D.C. Schramm, C. J. 2006. The Entrepreneurial Imperative. New York: Harper Collins. Schumpeter, J. 1934. Capitalism, Socialism, and Democracy. New York: Harper & Row. Venkataraman, S. 1997. “The Distinctive Domain of Entrepreneurship Research: An Editor’s Perspective.� In J. Katz and R. Brockhaus, eds., Advances in entrepreneurship, �rm emergence, and growth. Greenwich, CT: JAI Press. Volpe, C. M., and J. Carballo. 2008. “Is Export Promotion Effective in Developing Countries? Firm-level Evidence on the Intensive and the Extensive Margins of Exports.� Journal of International Economics 76 (1): 89– 106. Wennekers, S., A. van Stel, A. R. Thurik, and P. Reynolds. 2005. “Nascent Entrepreneurship and the Level of Economic Development.� Small Business Economics 24 (3): 293–309. World Bank. 2010. World Bank Group Entrepreneurship Snapshots database. World Bank, Washington, D.C. http://econ.worldbank.org/research/entrepreneurship. Yunus, M. 1999. Banker to the Poor. London: Aurum Press Ltd. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Getting Credit to High Return Microentrepreneurs: The Results of an Information Intervention Suresh de Mel, David McKenzie, and Christopher Woodruff Small-scale entrepreneurs typically cite access to �nance as the most important constraint to growth. Recent randomized experiments have shown the return to capital to be very high for the average microenterprise in Sri Lanka. An intervention was designed to improve access to credit among these high-return microenterprises without subsidizing interest rates or requiring group lending. The intervention con- sisted of information sessions providing details of the micro�nance loan product offered by a regional development bank and a reduction from two to one in the number of personal guarantors required for these loans. Ten percent of the microen- terprises invited to the information meetings received a new loan, doubling the pro- portion of �rms receiving loans over this period. However, the loans do not appear to Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 be going to particularly high-return �rms but rather to �rms with more household assets. Many more �rms would like loans but are constrained by an inability to �nd personal guarantors and by other bureaucratic procedures. The results suggest that information alone is unlikely to be enough for most �rms and point to the need for credit bureaus that cover micro�nance loans and for continuing innovation in loan products that can reach the urban microenterprise sector. JEL classi�cation: G21, D24, D83, O12 Access to �nance is the most common constraint to growth cited by entrepre- neurs in a broad range of countries. Surveys of small and medium-size �rms in 54 countries indicate that �nancing constraints are particularly severe for small Suresh de Mel (sdemel@pdn.ac.lk) is senior lecturer, Department of Economics and Statistics, University of Peradeniya, Sri Lanka. David McKenzie (dmckenzie@worldbank.org; corresponding author) is senior economist, Development Research Group, at the World Bank. Christopher Woodruff (c.woodruff@warwick.ac.u) is professor of economics, University of Warwick. The authors thank the editors, three anonymous referees, Shawn Cole, and conference participants for helpful comments; Susantha Kumara, Susil Nawarathna, and Jayantha Wickramasiri for outstanding research assistance; the Ruhuna Development Bank for its collaboration in designing the loan product and implementing this intervention; and Sri Lankan enterprise owners for their continued involvement in the Sri Lanka Microenterprise Survey. AC Nielsen Lanka administered the surveys on which the data are based. Financial support from National Science Foundation grant SES-0523167 and the World Bank is gratefully acknowledged. THE WORLD BANK ECONOMIC REVIEW, VOL. 25, NO. 3, pp. 456– 485 doi:10.1093/wber/lhr023 Advance Access Publication May 18, 2011 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 456 de Mel, McKenzie, and Woodruff 457 �rms. Moreover, �rms identifying �nance as a constraint are found to grow more slowly (Beck, Demirgu ¸ -Kunt, and Maksimovic 2005), suggesting that ¨c bridging the �nance gap is important for stimulating entrepreneurship. In most countries, microenterprises are even less likely than small and medium-size �rms to have access to �nance from formal channels, resulting in very large returns to capital for these �rms. Recent randomized experiments in Mexico and Sri Lanka providing grants of $100–$200 to microenterprises found mean real monthly returns to capital of 20 percent or more in Mexico and more than 5 percent in Sri Lanka (McKenzie and Woodruff 2008; de Mel, McKenzie, and Woodruff 2008). These high returns—well in excess of microlending interest rates in these countries—motivate the research and policy question of this article: How can the formal �nancial system get credit to these high-return microentrepreneurs? The traditional government response to the credit needs of microenter- prises has been subsidized interest rate programs. Such programs have been criticized as too costly, politically directed, and damaging to incentives for Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 the �nancial sector to �nd innovative ways to expand access (World Bank 2008). For �rms with high returns, loan access is more important than interest rates, and access to micro�nance has recently succeeded in expand- ing access to credit to some segments of the poor. But the group lending approach that has worked well with rural women has required adaptation to meet the demands of urban business owners; in urban areas, individual loans offered by micro�nance providers are often larger and often require collateral (Armenda ´ riz and Morduch 2007). Sri Lanka has the fourth highest micro�nance penetration in the world, with 4.3 percent of the popu- lation as loan clients (Honohan 2004). Yet only 5.2 percent of the Sri Lankan microenterprises in this study sample received a micro�nance loan in the year before the intervention. Why aren’t more microenterprises receiving credit? The low frequency of observed credit may reflect credit supply issues. Borrowers may not be able to meet the criteria set by formal �nancial institutions. The lending criteria include collateral requirements, legal title, records and forms, and other lender requirements that potential borrowers might not be able to supply. In a sample of much larger �rms than those studied here, Banerjee and Duflo (2008) show that supply constraints are an important determinant of credit to small �rms. Alternatively, the low levels of observed credit may reflect a lack of demand from microenterprises. Entrepreneurs may not want credit at available interest rates. On the surface, this seems at odds with the high average returns to capital in microenterprises. But there is considerable heterogeneity in these returns, and even when the expected returns exceed market interest rates, some owners may object to loans on religious grounds or because of risk aversion. Alternatively, entrepreneurs may want credit but not know where or how to get it. This possibility, which concerns the role of information and �nancial lit- eracy, has received far less attention in the literature. Information and �nancial 458 THE WORLD BANK ECONOMIC REVIEW literacy problems are likely to be particularly prevalent in a developing country setting, where schooling levels are low, use of outside �nancial advisors or accountants is rare, and people have little experience with the �nancial system (Miller 2008).1 This article reports on a �nancial literacy intervention carried out by the authors in Sri Lanka that permits assessing the role of these three explanations in accounting for the low level of credit among microenterprises. The interven- tion aimed to increase access to credit at market interest rates by providing information about the availability of loans and procedures for obtaining loans from a local development bank. Providing information has recently been found to have large effects in other domains of behavior, such as schooling decisions in the Dominican Republic (Jensen 2010), risky sexual behavior among girls in Kenya (Dupas 2011), and employee decisions about retirement plans at a large U.S. university (Duflo and Saez 2003). Holding information sessions is less costly and controversial than providing interest subsidies, so, if effective, pro- viding information could become a useful tool for increasing �nancial sector Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 participation by the poor. Microentrepreneurs in two districts (similar to states) received a letter informing them of an existing loan product offered by the Ruhuna Development Bank (RDB) and inviting them to an information meeting with bank staff to hear about the loans available and how to �ll out an application. Discussions with branch managers also led RDB to relax requirements on the number of guarantors required for loans from two to one and make other minor modi�cations of requirements. Thus, the intervention both provided new information and slightly relaxed the loan requirements. There was considerable interest among microenterprises in obtaining a loan; 62 percent of the enterprises contacted attended the information meet- ings. Difference-in-differences and �xed effects methods show that the inter- vention doubled the number of �rms receiving a loan in a three month period, a sizable effect. However, this still represents new loans for only 10 percent of the �rms invited to the meetings and is likely an upper bound on the impact of information. For �rms attending the meeting that did not obtain loans, the main reasons were inability to �nd a guarantor for the loan and the fact that applicants had to travel to other banks and micro�nance institutions to obtain endorsements showing they had no loan commitments to these organizations, because there were no credit bureaus to provide this information. Data from the randomized capital injections described in de Mel, McKenzie, and Woodruff (2008) show that the �rms receiving loans did not have the highest predicted returns to capital but 1. In a study in Zambia, only half the adult population knew how to use the most basic �nancial products such as a savings account (DFID 2008). In India and South Africa, even clients of �nancial institutions were found to have little knowledge of compound interest, products available, or why a bank charges fees (Cohen and others 2006). de Mel, McKenzie, and Woodruff 459 rather had higher household assets and capital stock than other �rms. Even at the margin, the �nancial system is thus directing loans to �rms with higher assets rather than to �rms that stand to gain the highest returns from the loans. The article is structured as follows. Section I describes the �nancial system in Sri Lanka and the intervention. Section II discusses the data, and section III describes the results of the intervention. Section IV discusses the policy implications of the results. I. SETTING AND THE INTERVENTION Sri Lanka has a population of approximately 20 million, with GDP per capita of $1,070 in 2006. The �nancial sector consists of a wide range of institutions including commercial banks (both state owned and private), six state-owned regional development banks, the privately owned Sanasa Development Bank, Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 �nance and leasing companies, cooperatives, nongovernmental organizations, and Samurdhi banks, which are part of the state-run Samurdhi welfare program (Duflos and others 2006; GTZ 2007). Outreach is fairly extensive, with a recent study �nding that 78 percent of urban households had accessed �nancial institutions for savings and 40 percent for loans (GTZ 2007). Micro�nance institutions, which include regional development banks, Sanasa, Samurdhi banks, nongovernmental organizations, and community-based organizations, are an important part of the �nancial system, with more than 15 million deposit accounts and 2 million outstanding microloans in 2004 (Duflos and others 2006). However, the GTZ report also found large unmet demand for credit, with collateral requirements, excessive documentation, rigid terms and conditions, and long processing periods all viewed as key barriers, particularly among poorer households. Sri Lanka’s credit bureau reports only on loans over 100,000 rupees (Rs.) (approximately $1,000), and is accessible only to its shareholders, which are commercial banks, licensed specialized banks, and the regional development banks. There is no organized credit information sharing among micro�nance providers. This lack of information, coupled with the large number of competing micro�nance providers, leads to concerns about clients taking out multiple loans (GTZ 2007). Direct evidence of credit constraints among microenterprises can be seen in the results of a randomized experiment conducted by the authors in which grants of $100–$200 were given to a randomly chosen subset of microenter- prises surveyed in the Galle, Kalutara, and Matara districts of Sri Lanka (de Mel, McKenzie, and Woodruff 2008). The study found real monthly returns to capital averaging 5 percent, with the highest returns for �rm owners with greater ability and lower wealth. The intervention examined in this article was designed in light of these high returns. 460 THE WORLD BANK ECONOMIC REVIEW The Ruhuna Development Bank The Ruhuna Development Bank (RDB) was the �rst state-owned regional development bank to be incorporated under the Regional Development Bank Act of 1997. It began operations in July 1998 through the amalgamation of three regional rural development banks in the districts of Galle, Hambantota, and Matara along the southern coast of Sri Lanka. RDB is keen to position itself as the “premier development bank of the masses of Ruhuna� (RDB 2007). At the end of 2007, it had 440 employees, 408,000 savers, and 171,000 active borrowers with an average loan balance of $254.2 RDB, with its regional focus on the southern districts of the country and its emphasis on promoting self-employment activities for development, was a natural partner for implementation of the micro�nance intervention. The Loan Product Discussions with RDB management enabled the design of a loan scheme that Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 was comparable to other loan products offered by RDB to similar micro-level enterprises. The loan product offered to the research project’s enterprises had the following features and conditions (summarized in table A1 in the appendix): † Loan amount: Enterprises could apply for any loan in the Rs. 5,000– 25,000 range (approximately $50–$250). Recall the average loan balance among all RDB customers is $254, so these loans were at the lower end of what the bank offered. The bank had flexibility in making the �nal decision on the loan amount, subject to usual investigations and procedures. † Interest rate: All enterprises were charged RDB’s regular annual lending rate of 16 percent. This rate is competitive in the local market—compar- able to the interest rates offered by other development banks and cheaper than interest rates of around 22 percent offered by local micro�nance organizations. † Loan repayment: All loans had a two-year repayment period. Normal RDB practice allows for a grace period of three months when necessary. The bank offers either a reducing balance (the interest payment, and therefore the installment, drops as principal is paid off)3 or an equal installment loan. Because of the nature and size of the enterprises and the education levels of the entrepreneurs, an equal monthly installment 2. http://www.mixmarket.org/m�/ruhuna/data [accessed December 17, 2008]. 3. Under a reducing balance system, the �rms would pay more in the �rst months, with the repayment amount falling with the balance. For a Rs. 20,000 loan, �rms would pay Rs. 1,100 in the �rst month, declining slowly to Rs. 844 in the last month. The equal installment method averages these payments out and is calculated by the RDB so that the same amount of interest is paid over two years (Rs. 3,333 on a Rs. 20,000 loan) as would be paid under reducing balance. de Mel, McKenzie, and Woodruff 461 method was preferred. The monthly installment was Rs. 486 for a Rs. 10,000 loan and Rs. 972 for an Rs. 20,000 loan. † Collateral: RDB usually requires collateral; however, for small loans the branch has the discretion to relax this requirement and did so for this research project loan scheme. To stress the importance of loan repay- ment, loan of�cers had the discretion to take an inventory of household assets. † Guarantors: RDB usually requires two cosigners as loan guarantors. The guarantor should have a steady income from wage/business employment or access to liquid assets. For the research project loan scheme, the bank relaxed requirements to allow guarantors to be family members or mutual loan applicants (“interperson� guarantee). Individual branches also had the discretion to reduce the number of guarantors from two to one. † Endorsements by other �nancial institutions: RDB usually requires all loan applicants to provide information on bank accounts and loan com- Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 mitments. In the absence of a micro�nance credit bureau, applicants must obtain a written endorsement from each bank and micro�nance intermediary in the area. For the purposes of the research project loan scheme, this process was limited to the �ve leading micro�nance interme- diaries in the area (Sanasa, Samurdhi, SEEDS, Rural Bank, and Ceylinco Grameen) and veri�cation against the centralized RDB database. † Residence veri�cation by Grama Niladhari of�cer: RDB usually requires that all applicants obtain veri�cation of their residential address by the Grama Niladhari (the smallest administrative unit) of�cer. For the research project loan scheme, this requirement was imposed as usual. † Oath of attestation: RDB usually requires applicants to furnish an oath of attestation, signed in the presence of a justice of the peace, certifying the �xed and movable assets owned by the applicant. For the purposes of the research project, the branches were given discretion to waive this requirement. † Salary certi�cation by employer: RDB usually requires employee loan applicants to provide salary certi�cation from their employers. Since the research project is targeted to self-employed enterprise owners, this con- dition was not imposed. † Business registration: Required for larger loans, this is usually not imposed as a precondition for approval of small loans. † Age: If the enterprise owner is more than 55 years old, RDB usually requires a joint application with a younger immediate family member. This condition was imposed as usual. † RDB account holder condition: All loan applicants are required to be RDB savings account holders. Existing account holders were asked to 462 THE WORLD BANK ECONOMIC REVIEW bring their savings account books. Those without accounts were required to open new savings accounts with an initial deposit of Rs. 250 (approximately $2.50). The usual three-month waiting period between opening a savings account and loan application was not imposed. † Loan application-related administrative costs: RDB usually charges an administrative fee on all loan applicants that is deducted from the loan amount credited to the account. These costs (Rs. 250 per loan applicant) were imposed as usual but were paid by the research project rather than loan applicants. In sum, the loan offered was similar to existing RDB loan products, but with fewer constraints in terms of guarantors and other minor adjustments where discretion was often already given. The main reason RDB agreed to relax the constraint was that the number and size of loans were small compared with the overall portfolio. From the project viewpoint, this mild relaxation of conditions was viewed as potentially helping to ensure there would be some loan appli- cants to learn from since it was unclear ex ante how many �rms would meet Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 the requirements. Information Intervention RDB regional managers divided the 383 enterprises in the Galle and Matara districts of the Sri Lanka Microenterprise Survey (described later) among eight bank branches—four in each district—based on location. In late August 2006, the enterprises were sent a letter informing them about the RDB loan scheme. The letter stated that loan amounts would range from RS. 5,000 to Rs. 25,000, with the actual amount at RDB’s discretion; the interest rate would be 16 percent; the loan would be repayable in equal monthly installments over 24 months (with examples given for Rs 10,000 and Rs. 20,000 loans); and appli- cants would not have to pay the administrative fees of the loan application (as the project would pay them) The letter emphasized that this would not be a grant but a loan subject to the prevailing rules and regulations of the bank and that this was a transaction between the loan applicant and the bank, with the research project playing only a facilitating role.4 Anyone interested in obtaining more information about the loan offer and the application process was invited to attend a meeting at an identi�ed venue (such as a community hall) in the area. Loan applicants were also requested to bring their existing savings accounts books or the initial minimum Rs. 250 deposit and identity card to open an RDB account. Although the costs of going to a bank branch to learn the conditions and details of a loan product might be modest, �rms are unlikely to make inquiries 4. Given previous randomized interventions with equipment and cash grants with these enterprises in 2005 (see de Mel, McKenzie, and Woodruff 2008), special care was taken to identify the current intervention as being a loan and to differentiate it from the previous project’s grants. This fact was reemphasized at the loan awareness meetings held in September 2006. de Mel, McKenzie, and Woodruff 463 if they have not heard of the bank or that it offers products to �rms like theirs and is willing to spend time talking with them. Thus, the information interven- tion was considered important in providing credible information about speci�c loan products and the willingness of RDB to lend to enterprise like those in the sample. These awareness meetings were held in mid-September 2006 at eight locations and attended by both RDB and research project staff. At the meet- ings, bank staff explained the loan features and the conditions. RDB staff opened 118 new accounts; 72 attendees were existing account holders. Loan application forms were distributed and explained. RDB and project staff answered questions about the loan and the application form. Loan application forms could not be submitted at the meetings because of the need for various third-party endorsements. The median elapsed time between receiving the application form and sub- mitting it was seven days. Median times reported by �rms to ful�ll the different requirements were approximately one hour to �ll out the application form, two Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 hours to obtain the residence veri�cation, seven hours to obtain a guarantor and �ll out the associated forms, three hours to obtain the endorsements of other �nancial institutions, and one hour to submit the application. The median applicant spent 26 hours completing these steps. The completed applications, including the required endorsements, were turned in to the assigned RDB branch. During September–October 2006, 41 enterprises completed the application process. After veri�cation of the infor- mation on the application by the RDB loan of�cer, which included a �eld visit, 38 of the loans were approved and 3 were rejected.5 Funds were dispersed during November –December 2006.6 Why Design the Intervention This Way? By building on the knowledge gained from the authors’ prior experiment with these �rms, this intervention offered two advantages. First, there was good knowledge of the returns to capital and characteristics of the �rms, including several observations on the �rms before the intervention. Second, using an existing sample of �rms enabled learning about the constraints to �nance with relatively low marginal costs, making the intervention budgetarily feasible. However, this decision also raised three key issues. The �rst issue is whether participation in the prior experiment—in which randomly selected �rms had received grants—would change their behavior in the loan experiment. This raises the issue of external validity. Firms that had received grants might have felt obligated to attend the meetings, or 5. RDB loan of�cers usually conduct a �eld visit before approving a loan. In this case, it was mainly to verify the existence and status of the enterprise. If deemed necessary, loan of�cers also had the discretion to visit the residence and take an inventory of household assets. 6. Except for two loan applicants that handed in their application material late and had loans approved in early January 2007. 464 THE WORLD BANK ECONOMIC REVIEW those that had not received grants might show up in expectation of receiv- ing bene�ts this time. This would imply higher meeting attendance than would be the case in another sample of microenterprises. Or, if the grants had alleviated credit constraints, that might have reduced the demand for loans, resulting in lower attendance among �rms that received the earlier grants. The article examines this concern by testing whether prior receipt of a grant is a signi�cant determinant of loan-seeking behavior and �nds a small and insigni�cant effect. This suggests that these concerns may not be a �rst-order issue. The second issue is that this is not a randomized experiment—the inter- vention was carried out in the two districts where RDB operates and not in Kalutara where it does not. Given the sample size and the small number of RDB branches in Galle and Matara, it did not make sense to randomly introduce the intervention to, say, only four of the eight branches in these districts. Randomization at the individual �rm level was rejected because of the possibility of information spillovers and reduced sample size. As a Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 result, the study relies on nonexperimental estimation. However, the setting offers several characteristics that make the nonexperimental estimation reasonably convincing. The comparison district is adjacent to those treated, with similar characteristics. The six rounds of panel data available on both intervention and control �rms before the intervention make the difference-in-difference approach more credible. A matching exercise was also conducted to ensure that comparisons are made only among �rms with similar characteristics. The �nal issue is that the intervention is a combination of information provision and modest changes in the loan approval process. These changes were made to improve the chances of at least some �rms getting loans; but the drawback is that any impacts cannot be ascribed to information pro- vision alone. Rather, the results can be interpreted as an upper bound on the effect of information provision. Firms were asked detailed questions to help unpack the dimensions along which the intervention mattered. Nevertheless, this limitation must be acknowledged. The current study is a �rst step in learning about the constraints to getting access to credit, which future experiments can build on. I I . D ATA The main data source is the Sri Lanka Microenterprise Survey (SLMS), a panel survey of 617 microenterprises in three southern and southwestern districts of Sri Lanka: Galle, Kalutara, and Matara. The baseline survey was carried out in April 2005, with eight additional waves conducted at quarterly intervals through April 2007. Two additional waves were conducted in October 2007 and April 2008. Finally, preliminary data from a December 2010 revisit provide additional evidence. The survey was designed in part to study the de Mel, McKenzie, and Woodruff 465 recovery from the December 26, 2004, Indian Ocean tsunami, and the sample was selected to draw equally from areas along the coast, where �rms suffered direct damage from the tsunami; areas slightly inland, where �rms did not suffer direct damage; and inland areas, where neither assets nor demand were affected. A door-to-door screening survey in 25 Grama Niladhari divisions in these three districts was used to select full-time self-employed entrepreneurs running manufacturing, services, and retail trade �rms with invested capital of Rs. 100,000 (about $1,000) or less, excluding investments in land and build- ings. See de Mel, McKenzie, and Woodruff (2008, 2010) for more details on this survey. This study focuses on the 574 �rms remaining in the sample after six rounds of the panel survey (July 2006 was the last survey round before the intervention occurred). The 383 qualifying �rms in Galle and Matara dis- tricts received invitations to the RDB meetings, and the 191 �rms in the Kalutara district did not because they were outside RDB’s coverage area. In Galle and Matara, �rm owners were approximately evenly split between Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 male and female owners (table 1). The average owner was 43 years old, married, with nine years of schooling. Mean monthly pro�ts in June 2006 were Rs. 5,814 (about $58), and mean capital stock aside from land and buildings was Rs. 41,428 ($414). A quarter of �rms had been in business for three years or less at the time of the baseline survey. The �rms were largely informal: only 20 percent were registered with the Pradeshiya Saba or District Secretariat, only 27 percent kept business records, and 70 percent operated out of the owner’s home. GPS coordinates of the location of the bank branches and �rms were used to calculate the straight-line distance between the �rm and the bank branch where meetings were held and loan applications were delivered. Distance from the branch was 3.6 kilometers for the mean �rm and 2.2 kilometers for the median �rm, with the top quartile by distance 5.2–32 kilometers away. Straight-line distances understate the travel distance to the banks but should be a reasonable approximation for most �rms, assuming a reasonably dense urban road network (Gibson and McKenzie 2007). The SLMS contains rich data on owner characteristics. For demand for credit, potential attributes of interest include measures of general and entrepre- neurial ability (education, digitspan recall, and self-assessed entrepreneurial self-ef�cacy); risk aversion (measured as the coef�cient of relative risk aversion obtained from playing lottery games with the �rm owners for real money and as self-assessed risk-seeking in general and �nancial domains on a 10-point scale, as used in the German Socioeconomic Panel (a well-known large-scale survey); a household asset index (the �rst principal component of a set of indicators of ownership of selected durable goods); number of wage workers in the household (who can provide a source of �nancing for the �rm); and the religion of the owner (7 percent of owners in the Galle and Matara sample are Muslim). 466 THE WORLD BANK ECONOMIC REVIEW T A B L E 1 . Difference in Means Galle and Matara �rms Kalutara Invited to Attended Applied for Application Characteristics �rms meetinga meetingb loanc refusedd Owner characteristics Female owner 0.43 0.52* 0.54 0.55 0.33 Age of owner 40.9 42.8* 42.7 42.3 46.0 Married 0.80 0.82 0.86** 0.80 0.33** Years of education 9.63 9.23 8.97** 9.05 10.00 Muslim 0.16 0.07*** 0.06 0.00* 0.00 Digitspan recall 5.94 5.89 5.83 5.83 6.67 Risk aversion (lottery-based) 0.41 0.08** 0.03 2 0.02 2 0.16 Overall Risk-seeking behavior 6.93 6.24*** 6.12 6.28 5.67 Financial risk-seeking 6.10 5.40*** 5.38 5.33 5.33 behavior Entrepreneurial self-ef�cacy 33.8 29.9*** 29.9 30.1 29.7 Household asset index 0.10 0.06 2 0.08* 0.41** 0.70 Number of wage workers in 0.69 0.66 0.65 0.65 1.67** Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 household Financially literate 0.49 0.70*** 0.71 0.74 0.67 Firm characteristics Capital stock (excluding land 45259 41428 38748 50952* 45750 and buildings) in June 2006 Real pro�ts in June 2006 6789 5814* 6000 6370 3465 Real revenues in June 2006 23345 17137*** 17638 18977 21545 Owner hours in June 2006 50.0 47.8 48.6 52.4 54.7 Registered �rm 0.18 0.20 0.17 0.30** 0.33 Manufacturing/services 0.36 0.57*** 0.59 0.55 0.33 Three years or less in age 0.40 0.24*** 0.22 0.28 0.67 Business operated out of the 0.46 0.70*** 0.70 0.65 0.00** home Business records kept 0.24 0.27 0.23** 0.28 0.33 Received grant 0.59 0.59 0.58 0.60 0.00** Had previously had a loan in 0.24 0.25 0.25 0.23 0.00 March 2005 Received a formal loan in the 0.24 0.21 0.20 0.15 0.00 year to June 2006. Directly affected by the 0.34 0.34 0.33 0.30 0.33 tsunami Distance to the bank branch n.a. 3.59 3.54 3.99 3.34 (kilometers) Sample size 191 383 237 40 3 *Difference in means signi�cantly different from 0 at the 10 percent level; **difference signi�- cant at the 5 percent level; ***difference signi�cant at the 1 percent level. a. Difference in means between Kalutara and Galle/Matara �rms b. Difference in means compared with nonattendees among Galle/Matara �rms. c. Difference in means compared with nonapplicants among attendees. d. Difference in means compared with approved loans among loan applicants. Source: Authors’ analysis based on data described in the text. de Mel, McKenzie, and Woodruff 467 Supplementing the data from the SLMS are RDB administrative data on loan applications and loan decisions and a questionnaire administered during the RDB meetings to 128 �rm owners.7 The baseline survey asked whether the business had ever received a loan from a private bank, government bank, micro�nance organization, Samurdhi, Sanasa, or Integrated Rural Development Program (IRDP) or Rural Economic Advancement Program (REAP).8 A loan from any of these sources is con- sidered a formal loan. The survey also asked whether the business had ever received an informal loan—money from a moneylender or family and friends. Each round of the survey asks whether the �rm has received a loan in the past three months from any of these sources. In Galle and Matara, 25 percent of �rms had had some form of formal loan in the baseline survey, and 21 percent had received a formal loan in the year before the intervention. None of these lenders emerged as a preferred lender. Formal loans were spread across a variety of sources: 3.9 percent had received a loan from a private bank, 5.0 percent from a government or development bank, 5.2 percent from a Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 micro�nance organization, 6.0 percent from Samurdhi, 3.7 percent from Sanasa, and 1.6 percent from IRDP/REAP. Informal loans favored family or friends (5.0 percent of �rms) over moneylenders (2.6 percent). The study survey included two questions on �nancial literacy used by Lusardi and Mitchell (2006) in the United States to assess whether individuals understand basic concepts of compound interest and inflation: † Suppose you had Rs. 100 in a savings account and the interest rate was 2 percent a year. After �ve years, how much do you think you would have in the account if you left the money to grow: more than Rs. 102, exactly Rs. 102, or less than Rs. 102? † Imagine that the interest rate on your savings account was 1 percent a year and inflation was 2 percent a year. After 1 year, would you be able to buy more than, exactly the same as, or less than today with the money in this account? In Galle and Matara, 91 percent of �rm owners got the �rst question correct, and 74 percent the second. The 70 percent who got both questions correct are 7. In particular, administrative records were examined on who applied for a loan and who was granted a loan. Survey self-reports closely matched these administrative records, and the results of the estimations are qualitatively similar when survey self-reports of applying for or receiving a loan are used in place of the administrative records. 8. Samurdhi is the main government poverty alleviation program, begun in 1995. Although it is mainly a direct welfare grant, it has a credit component that includes group savings and intragroup emergency credit, credit schemes implemented by the two state banks (Bank of Ceylon and People’s Bank), and micro�nance loans from Samurdhi banking societies. The Sanasa Development Bank, set up as a licensed specialized bank in 1997, functions as the apex of the thrift and credit cooperative movement and provides micro�nance loans to its membership. The Integrated Rural Development Program (IRDP) and the Rural Economic Advancement Program (REAP) are lending programs at the regional level. 468 THE WORLD BANK ECONOMIC REVIEW referred to as �nancially literate. This compares favorably with the sample of older Americans in Lusardi and Mitchell (2006), where only 56 percent got both questions right. While general basic �nancial literacy is reasonably high, product-speci�c literacy is more limited. In the short survey given at the RDB meetings, attendees were asked about their knowledge of RDB before receiving the letter: 36 percent of �rms said they did not think that RDB offered loans to small businesses, 62 percent said they had no idea what interest rate RDB charged on a loan of Rs. 10,000, and 10 percent knew that the annual rate was 16 percent. Agreement seems to be emerging among �nancial literacy trainers that product-speci�c �nancial literacy training is more effective than general �nancial literacy training. For example, telling someone how to apply for a loan is less effective than walking them step by step through applying for a loan from a speci�c bank using the bank’s forms.9 Thus, despite the reasonably high levels of basic �nancial literacy, the lack of RDB-speci�c �nancial literacy suggests considerable scope for the interven- Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 tion to provide new information. I II. RE S U LTS This section reports on the results for meeting attendance and loan application and outcome. Meeting Attendance Meeting attendance was high, with 237 of the 383 invited �rm owners (62 percent) attending.10 Based on pre-meeting data, a t-test shows signi�cant difference in means between meeting attendees and nonattendees (see table 1 column 3). Meeting attendees are more likely to be married, to have slightly less education and household durable assets, and to be slightly less likely to keep business records than nonattendees. The �rst three columns of table 2 present the results of a probit estimation of the correlates of meeting attendance. Since the loan meeting invitation mentioned only the availability of the loan, the interest rate, and repayment period but not the relaxation of guarantor requirements, it seems reasonable to view this reduced form as revealing the correlates of demand for loans. 9. While related, this information provision differs from the marketing efforts of banks. The point of marketing is to sell services to new customers, and marketing outreach by some �nancial intermediaries includes �nancial literacy training. This intervention did not try to sell the �rms on all the things they could use a loan for or encourage them to get a loan. Rather, it provided information on how to apply for a loan at this particular �nancial institution if one wanted such a loan. 10. In survey round 7 immediately following the meetings, 231 �rm owners said they attended the meeting. The study research assistant recorded 208 owners attending. Firms are classi�ed as having attended if either the research assistant recorded them as attending or the �rm owners say they attended. The results are robust to using either measure separately. T A B L E 2 . Determinants of Attending Meeting and Applying for Loan: Marginal Effects from Probit Estimation Attending meeting Applying for loan Characteristics (1) (2) (3) (4) (5) (6) Owner characteristics Female owner 0.00202 0.0407 2 0.0243 2 0.0199 (0.071) (0.054) (0.066) (0.053) Age of owner 2 0.00334 2 0.00381 0.000269 2 0.00117 (0.0029) (0.0025) (0.0029) (0.0026) Married owner 0.203** 0.138* 2 0.129 2 0.0309 (0.085) (0.071) (0.11) (0.074) Years of education 2 0.0118 2 0.0189** 2 0.0105 2 0.00512 (0.012) (0.0096) (0.0099) (0.0087) Muslim owner 2 0.197 2 0.176* (0.12) (0.11) Digitspan recall 0.00745 2 0.0103 0.0322 0.0108 (0.025) (0.021) (0.022) (0.019) Risk aversion (lottery based) 0.00329 0.0164 (0.021) (0.020) Overall risk-seeking behavior 2 0.0230 2 0.00443 (0.016) (0.014) Financial risk-seeking behavior 0.00709 2 0.00348 0.00806 2 0.00102 (0.016) (0.012) (0.013) (0.010) Entrepreneurial self-ef�cacy 0.000737 0.000904 (0.0059) (0.0065) Household asset index 2 0.0146 2 0.0266 0.0351* 0.0396** (0.020) (0.016) (0.020) (0.016) Number of wage workers in household 2 0.0123 2 0.0148 (0.037) (0.038) Financially literate 0.0691 0.0646 0.0839 0.00876 (0.070) (0.058) (0.057) (0.058) de Mel, McKenzie, and Woodruff Firm characteristics Log capital stock (excluding land and buildings) 2 0.0219 2 0.0505** 2 0.0137 2 0.000988 (0.031) (0.025) (0.030) (0.022) 469 (Continued ) Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 TABLE 2. Continued Attending meeting Applying for loan 470 Characteristics (1) (2) (3) (4) (5) (6) Log real pro�ts 2 0.00772 0.0333 0.000532 2 0.000359 (0.034) (0.029) (0.032) (0.024) Log real sales 0.0364 2 0.000940 (0.023) (0.023) Owner hours worked per week 0.00132 0.00119 (0.0015) (0.0016) Registered �rm 2 0.0244 2 0.0153 0.105 0.215** (0.086) (0.074) (0.10) (0.099) Manufacturing/services 0.0240 0.0626 0.107 0.0683 (0.082) (0.067) (0.071) (0.055) Age three or younger 2 0.0205 2 0.0241 0.166 0.121 (0.081) (0.069) (0.11) (0.077) Business operated out of home 0.00696 2 0.0463 0.0112 0.00449 THE WORLD BANK ECONOMIC REVIEW (0.079) (0.068) (0.074) (0.058) Business records kept 2 0.114 2 0.116* 2 0.0182 0.00377 (0.075) (0.064) (0.074) (0.062) Received grant 2 0.0361 2 0.000615 2 0.00583 0.0110 (0.065) (0.056) (0.066) (0.051) Had previously had a loan by March 2005 0.0456 2 0.0116 (0.071) (0.064) Received a formal loan in the year to June 2006 2 0.103 2 0.0272 2 0.145*** 2 0.0949** (0.078) (0.067) (0.047) (0.048) Directly affected by the tsunami 2 0.0751 2 0.0416 2 0.0349 2 0.0164 (0.068) (0.060) (0.059) (0.053) Log distance to the bank branch 2 0.00480 0.00630 2 0.00998 0.0130 (0.032) (0.026) (0.031) (0.025) Number of observations 282 369 340 163 215 209 * p , 0.1, ** p , 0.05, *** p , 0.01. Note: Numbers in parentheses are robust standard errors. Source: Authors’ analysis based on data described in the text. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 de Mel, McKenzie, and Woodruff 471 Firm owners attending the meetings and those not attending are similar in many respects, and the differences shown in table 1 continue to hold after controlling for other variables. The probability of attending is lower for owners with more education, who are unmarried, who are Muslim, and with more household assets or whose �rms have higher pre-intervention capital stock. Uncorrelated with meeting attendance are the owner’s risk aversion and entrepreneurial ability, pre-intervention pro�tability of the �rm, industry, and level of exposure to the 2004 tsunami. Also uncorre- lated with meeting attendance are receipt of a grant during the prior ran- domized experiment on returns to capital (de Mel, McKenzie, and Woodruff 2008) and distance to the RDB bank branch holding the meeting. Firm owners who did not attend the RDB meetings were asked why (table 3). Personal reasons topped the list, such as illness and family emer- gency (42 percent of nonattendees). That so many owners would be affected by illness or family emergency seems implausible; more likely, this became a Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 catch-all category for such reasons as did not get around to it and did not want to make the effort. The next two most common responses were no need for a loan and do not like taking loans (14 percent each). One in ten nonattendees reported not knowing about the meeting. Only 22 percent of nonattendees reported one of several responses indicating that the loan requirements or conditions were responsible for their lack of interest in a loan. Loan Application and Loan Outcomes Only 41 �rm owners submitted a loan application, and 38 of these were granted a loan. Only one �rm owner submitting a loan application had not attended the meeting. The majority of applications were submitted in the last few days of September 2006 and in October 2006, with most loans being approved and disbursed in November and December 2006. The last approval and disbursement occurred on January 8, 2007. Most (30 out of 38) loans were for Rs. 25,000 ($250); two �rms received Rs. 10,000, �ve received Rs. 20,000, and one received Rs. 75,000. The importance of product-speci�c information is evidenced by the fact that before the inter- vention 19 percent of those receiving loans did not know the location of the RDB branches, 38 percent did not think that RDB offered loans to small businesses like theirs, and only 15 percent had any idea what interest rate RDB charges. The fourth column of table 1 provides summary statistics for owners who attended the meetings and who applied for a loan, indicating where t-tests show a difference in means between them and owners who attended the meetings but did not apply for a loan. Columns 4–6 of table 2 show the marginal effects from probit estimations of the determinants of loan application among those attending the meeting. These reflect the joint 472 THE WORLD BANK ECONOMIC REVIEW T A B L E 3 . Reasons Given for Not Attending the Loan Information Meeting Percent stating Reason this reason Personal reasons (e.g., illness, family emergency) 41.9 I dislike/do not believe in taking loans 14.2 I do not currently have any use for a loan 14.2 I did not know about the meeting 10.8 Interest rate on loan is too high 6.1 I am still paying back a loan from another institution 5.4 I don’t have con�dence I could make regular loan repayments 5.4 Loan amount is insuf�cient 4.7 I will not be able to ful�ll other RDB criteria other than guarantees/collateral 4.7 Dislike closing the business to attend 2.7 Time and effort to �ll out application forms, etc. 2.7 Inability to �nd guarantors and collateral for the loan 2.7 Bank branch is too far/too inconvenient 1.4 I do not have much faith in development banks such as RDB 1.4 Sample size 149 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Source: Authors’ analysis based on responses from October 2006 Sri Lanka Microenterprise Survey (round 7). outcome of demand for loans and supply restrictions imposed by lenders, and so are simply a description of who got credit.11 Although those attend- ing the meetings were somewhat poorer and more likely to own a �rm that was not formally registered than those not attending, the subsample of �rm owners who applied for a loan have more household assets and their �rms have higher capital stock and are more likely to be registered than meeting attendees who did not apply. Conditional on other covariates, owners who had received a formal loan in the past year were 10 –15 percentage points less likely to apply for a loan. During the meetings, owners with existing loans were dissuaded from making new loan applications. The only other signi�cant variable is the dummy variable for Muslim, as no one who reported their religion as Muslim applied for a loan. Since loan appli- cations were rejected for only three �rm owners, the characteristics of owners approved for a loan and those who applied are very similar; the last column of table 1 shows how the three rejected applicants differ from those whose applications were approved. One plausible reason why more �rms do not apply for credit is risk rationing (Boucher, Carter, and Guirkinger 2008). Even if average returns to capital are high, risk-averse individuals may not choose to borrow because of the risk of 11. One approach would be to try to jointly model supply and demand for loans as a system of equations. Given the small number of �rms that ultimately received loans and the lack of convincing identifying restrictions to separate supply and demand at this second stage, this approach is not pursued here. de Mel, McKenzie, and Woodruff 473 T A B L E 4 . Reasons Given by Meeting Attendees for Not Applying for a Loan Reason Percent stating this reason Inability to �nd guarantors and collateral for loan 24.7 I will not be able to ful�ll other RDB criteria other than guarantees 21.3 Interest rate on loan is too high 21.0 I dislike/do not believe in taking loans 18.7 I do not currently have any use for a loan 13.1 Time and effort in �lling out loan applications, collecting signatures 12.0 I am still paying back a loan from another bank/institution 10.9 I don’t have con�dence I could make regular loan repayments 9.7 Loan amount is insuf�cient 7.9 I do not have much faith in development banks such as RDB 4.5 Bank branch is too far/too inconvenient 3.4 I have already taken a loan from RDB and am still repaying 1.5 Sample size 267 Source: Authors’ analysis based on responses from October 2006 Sri Lanka Microenterprise Survey (round 7). Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 losing collateral or putting their guarantor under pressure if they cannot repay the loan. While the wealth measures discussed above provide some support for this supposition, other measures suggest that this is not the prime reason. Table 2 shows that the measures of risk aversion are not signi�cantly associated with either meeting attendance or loan application. Tables 3 and 4 show that only 5.4 percent of owners reported a lack of con�dence that they could make regular loan repayments as a reason for not attending the meeting, and only 9.7 percent of those attending who did not apply for a loan cited uncertainty about making payments as a reason. Finally, in December 2010 �rm owners were asked about the consequences of missing a loan payment. Only 3.7 percent believed that the bank would seize collateral, and 17 percent thought the bank would ask the guarantor for payment, while 74 percent thought that they would have to pay additional interest but would get extra time to repay, and 30 percent said that they would be given extra time to repay without paying additional interest. Thus, most owners believed that there was some scope for flexibility in repayment if they could not pay a particular loan installment. DID THE FIRMS THAT DID NOT APPLY SIMPLY NOT NEED CREDIT? Given that 21 percent of �rms had some formal loan at baseline and that recent studies have found modest take-up rates for credit, it is reasonable to ask whether lack of demand explains the low loan application rate.12 This study provides some 12. See, for example, Banerjee and others (2010) [ please add to reference list], where 27 percent of urban households take a loan after micro�nance is introduced in newly urbanized areas of Hyderabad. One difference worth noting in the Banerjee and others study is that the lender was offering a group-lending micro�nance product directed to women only. Overall demand for credit without these gender and group-lending restrictions might have been higher. 474 THE WORLD BANK ECONOMIC REVIEW evidence from direct questioning and economic estimation that suggests that �rms might not have applied for loans even though they were, by standard de�nitions, credit-constrained. The October 2006 survey, which occurred in the month following the meetings, asked �rm owners whether they had applied for a loan, had decided not to apply, or had not applied but planned to do so in the near future. Fifty-seven owners said that they planned to apply in the near future, but only eight did (they are included among the 41 owners who applied for a loan). The most important reasons given for not applying were not meeting the loan criteria: 25 percent reported an inability to �nd guarantors or collateral, and 21 percent reported not being able to ful�ll other criteria (see table 4). Next in importance was lack of demand at the going interest rates: 21 percent said the interest rate was too high, 19 percent said they did not like loans, and 13 percent said that they did not need a loan at the time. Only 3.4 percent said that the bank branch was too far away or too inconvenient to get to. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Those who said that they intended to apply in the near future were asked the reasons for the delay. Half the owners reported dif�culty �nding guaran- tors, 19 percent could not decide immediately whether to apply, and 16 percent identi�ed the process of getting endorsements from other banks. When asked directly during a revisit to the �rms in December 2010, 62 percent of �rms in the sample that did not have formal loans said they would apply for a Rs. 20,000 Rs loan at the same interest rate and terms as the loan in the inter- vention if they knew a bank would approve their application. Thus, at least half the �rms that did not apply indicated a demand for credit when asked directly. From the authors’ previous randomized experiment, there is also econometric evidence of returns far in excess of interest rates (de Mel, McKenzie, and Woodruff 2008). These high returns to capital generated by grants are more likely to result from credit constraints than from risk aversion. Returns to the grants were higher for male enterprise owners, for owners with more education and higher digitspan recalls, and for owners with lower house- hold assets and fewer wage workers in the household. The higher returns suggest that these owners are more credit constrained. Marginally signi�cant evidence is found that �rm owners with fewer house- hold assets and lower capital stocks are more likely to attend the meetings, and clearer evidence that the less educated are more likely to attend. Moreover, among those attending, there is clear evidence that �rms with higher capital stock and more household assets are more likely to apply. That is, while owners whose �rms have the highest predicted returns are more likely to have attended the meetings, they are not more likely to have obtained a loan through the intervention. The main determinant of receiving a loan was higher household assets. One might question whether the inability to �nd a guarantor reflects a lack of suitable candidates or whether it reflects the fact that those in a position to de Mel, McKenzie, and Woodruff 475 guarantee a loan know the business and think it unlikely to succeed. Ninety percent of those obtaining a loan have a government employee as their guaran- tor. Firm owners were asked whether they had parents or parents-in-law, sib- lings, uncles, aunts, or cousins who were government wage workers. Among those giving lack of a guarantor as a reason for not applying, only 2 percent of had a parent or parent-in-law who was a government wage worker, 24 percent had a sibling, and 43 percent had a more distant relative. Thus, the majority of those who said they could not �nd a guarantor did not have a close family member in a position to act as a guarantor. If potential guarantors were basing their decision on the expected returns to the loan, more educated, higher ability, less wealthy �rm owners would be expected to have guarantors and to receive a loan. Table 5 compares the characteristics of those approved for a loan and those who said that the inability to �nd a guarantor was the reason for not applying for a loan or for delaying (and then not applying). Those who said that they could not �nd guarantors have lower levels of household durable assets and their �rms have Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 lower capital stock (although not signi�cantly so) and are less likely to be for- mally registered. Thus, wealth seems the main factor determining whether a guarantor is available to those wishing to apply for a loan or who would apply if they had a guarantor. Gender, the ability of the owner, and factors that de Mel, McKenzie, and Woodruff (2008, 2009) associate with higher returns do not appear linked to the availability of a guarantor. Guarantors thus appear to be considering the same factors as the bank, ensuring that the potential loan applicant has suf�cient collateral rather than deciding based on expected returns to capital. WHAT DID FIRMS RECEIVING CREDIT CONSIDER THE MOST IMPORTANT PART OF THE INTERVENTION? The intervention consisted of providing information and relax- ing two loan conditions: reducing the number of guarantors from two to one and paying the Rs. 250 loan application fee. Firms that received loans were asked how important each of these components of the intervention was in enabling them to get the loan. Firms considered information provision to be the most important component of the intervention (table 6). Three-quarters of the borrowers said that learning that RDB would lend to small �rms like theirs was very important, and 51 percent said that it was the most important com- ponent of the intervention. Next in importance was learning that business registration was not required (60 percent of �rms viewed this as very important and 24 percent as the most important component). Relaxing the guarantor requirement was important for some �rms (40 percent considered it very important, but only 8 percent said it was the most important component). Asked whether they could have provided two guarantors, 48 percent of �rms answered yes. Paying the administrative fee was viewed as very important by 11 percent of the �rms, and 86 percent said that they would have paid it had the project not done so. 476 THE WORLD BANK ECONOMIC REVIEW T A B L E 5 . Differences in Characteristics between Those Receiving Loans and Those Who Could not Find Guarantors Characteristics Approved No guarantor p-value Owner characteristics Female owner 0.55 0.49 0.514 Age of owner 42.4 41.5 0.645 Married 0.84 0.87 0.718 Years of education 8.95 8.52 0.492 Muslim 0.00 0.11 0.032 Digitspan recall 5.74 5.73 0.978 Risk aversion (lottery-based) 2 0.01 2 0.09 0.784 Overall risk-seeking behavior 6.32 6.13 0.647 Financial risk-seeking behavior 5.34 5.39 0.912 Entrepreneurial self-ef�cacy 29.9 29.4 0.545 Household asset index 0.41 2 0.47 0.004 Number of wage workers in household 0.55 0.67 0.539 Financially literate 0.76 0.61 0.119 Firm characteristics Capital stock (excluding land and buildings) in June 2006 50308 36219 0.195 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Real pro�ts in June 2006 6556 5312 0.328 Real revenues in June 2006 18290 17407 0.844 Owner hours in June 2006 52.4 48.3 0.341 Registered �rm 0.32 0.11 0.005 Manufacturing/services 0.58 0.61 0.737 Three years or less in age 0.24 0.14 0.209 Business operated out of the home 0.71 0.78 0.421 Business records kept 0.26 0.30 0.677 Received grant 0.63 0.54 0.367 Had previously had a loan in March 2005 0.24 0.26 0.825 Received a formal loan in the year to June 2006. 0.18 0.19 0.951 Directly affected by the tsunami 0.32 0.32 0.944 Distance to the bank branch (kilometers) 4.03 3.32 0.231 Sample size 38 90 Note: Numbers in bold are values that are signi�cant at conventional levels. Source: Authors’ analysis based on responses from Sri Lanka Microenterprise Survey (various rounds). DID THE INTERVENTION INCREASE ACCESS TO CREDIT? Ten percent of targeted �rm owners (38 of 383) received a loan from the RDB. It is possible that all 38 �rms would have received loans in the absence of the intervention, either from the RDB or from another �nancial institution. The panel data are examined to determine whether the intervention increased access to credit to microenter- prises in Galle and Matara. Figure 1 summarizes the identi�cation strategy. It plots the share of enter- prises receiving a new loan in the past three months by survey wave in Galle and Matara, where the intervention took place, and in neighboring Kalutara, where it did not. On average 7.2 percent of �rms in Galle and Matara and 8.7 percent of �rms in Kalutara obtained a new loan in any given three month de Mel, McKenzie, and Woodruff 477 T A B L E 6 . What Do Firms Receiving Loans See As the Most Important Part of the Intervention? Percent Percent ranking ranking as very as most Intervention component important important Providing information that RDB would lend to small �rms like theirs 75.7 51.4 Providing information that RDB would lend without collateral 62.2 13.5 Providing information that business registration was not necessary 59.5 24.3 Showing them how to �ll out the application forms 56.8 0.0 Reducing the number of guarantors from two to one 40.5 8.1 Providing information on terms and requirements of loan 24.3 2.7 Paying the Rs. 250 administrative cost charged by the bank 10.8 0.0 Source: Authors’ analysis based on responses from Sri Lanka Microenterprise Survey (2007 survey rounds). Figure 1. Proportion of Firms Getting a Formal Loan in the Last Three Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Months, by Geographic Region Source: Authors’ analysis based on data described in the text. period over the 15 months preceding the start of the intervention (survey rounds 2 through 6). The share of �rms receiving new loans is likely overstated because of double-counting: 36 percent of those receiving a loan in the last three months report also having received a new loan in the previous survey round, and 39 percent who received a loan in survey round t but not in t-1 report receiving a new loan in survey round t þ 1. Since most loans are for periods of one to two years, this likely represents double counting of loans. 478 THE WORLD BANK ECONOMIC REVIEW A new variable, unique formal loan, for �rms reporting a formal loan in the last three months but not in the previous survey round �nds that 5.9 percent of Galle and Matara �rms and 6.7 percent of Kalutara �rms obtained a unique formal loan in the 12 months preceding the intervention (survey rounds 3 through 6). Loan applications were processed during the seventh survey wave; a large spike in new loans appears in the eighth survey wave in January 2007, when most of the loans had been approved. There is some spillover into the ninth survey wave in April 2007—due both to double-counting and to a few loans being approved in January 2007. No spike is apparent in Kalutara. Difference-in-differences and �xed effects estimations are used to identify the impact of the intervention on new loans issued. The difference-in- differences estimation uses the 574 �rms in all surveyed areas to estimate the following equation: FormalLoani;t ¼ a þ bTreati;t þ l0Xi þ dt þ 1i;t ð1Þ Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 where FormalLoani,t is an indicator of whether �rm i received a new loan in the three months up to time t, Treati,t is a dummy variable that takes a value of 1 in survey round 8 for the Galle and Matara �rms and 0 otherwise, Xi are controls for characteristics of the �rm and owner, and the dt are survey round �xed effects. A linear model is used to ease interpretation of the coef�cients— the signs and signi�cance remain the same when panel logit models are used. Standard errors are clustered at the �rm level. Column 1 of table 7 presents the ordinary least squares difference-in- difference estimates without �rm or owner characteristics included as controls. Although �gure 1 shows similar trends for new loans for �rms in Kalutara and in Galle/Matara in the period preceding the intervention, table 1 revealed some differences in baseline characteristics between �rms in the two areas. Column 2 of table 7 adds controls for these characteristics. Column 3 uses propensity score matching to match Kalutara and Galle/Matara �rms on characteristics prior to the intervention, and the analysis is restricted to �rms in the domain of overlapping support with estimated propensity score above 0.10 and below 0.90. Column 4 shows that including both controls and propensity matching slightly reduces the estimate of the treatment effect, lowering it from 8.0 per- centage points to 5.2–7.7 percentage points. Column 5 presents the panel �xed effects estimate, which is a 7.0 percentage point increase in formal loans. Columns 6 and 7 use the unique loans measure to show robustness to potential double-counting of loans. The sample uses waves 3 through 10 of the SLMS, since it is not possible to distinguish which of the loans reported in the last three months in the wave 2 survey were also reported in the last three months in the baseline survey. The results are robust to using this measure of new loan uptake. T A B L E 7 . Did the Intervention Increase the Proportion of Firms with Formal Loans and Crowd Out Informal Loans? Formal loans Unique formal loansa Informal loansb Ordinary least squares Ordinary least Ordinary least Fixed effects squares Fixed effects squares Fixed effects (1) (2) (3) (4) (5) (6) (7) (8) (9) Treatment dummy variablec 0.080*** 0.063** 0.077** 0.052 0.0702*** 0.054** 0.0701*** 0.015 0.0164 (0.027) (0.029) (0.032) (0.032) (0.025) (0.026) (0.024) (0.015) (0.014) Owner and �rm controlsd No Yes No Yes No Yes No Yes No Matched sample No No Yes Yes No Yes No Yes No Firm*Period observations 5,089 4,830 3,536 3,536 5,089 3,143 4,515 3,536 5,089 Number of �rms 574 537 393 393 574 393 574 393 574 ** p , 0.05, *** p , 0.01. Note: Dependent variable is the proportion of �rms with a loan in last three months. Numbers in parentheses are standard errors clustered at the �rm level, All interventions also include survey wave dummy variables. a. Loans for which there is no new formal loan in the previous survey round. b. Loans from moneylenders and family. c. Takes a value of 1 in round 8 (January 2007) and 0 in other survey rounds. d. Include gender, age, marital status, education, Muslim dummy variable, digitspan recall, risk aversion, overall risk seeking behavior, �nancial risk seeking behavior, entrepreneurial self-ef�cacy, household asset index, number of wage workers in household, and dummy variables for �nancial literacy, formality, manufacturing and services, age of �rm less than three years, home business, keeps business records, ever received grant, and previously received credit in baseline survey. These variables were also used for propensity score matching, and the matched sample is trimmed to include only �rms with propensity scores above 0.10 and below 0.90. Source: Authors’ analysis based on responses from Sri Lanka Microenterprise Survey (various rounds). de Mel, McKenzie, and Woodruff 479 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 480 THE WORLD BANK ECONOMIC REVIEW T A B L E 8 . What Did Borrowers Report Using Loans For? Average Share Share Proportion amount of excluding spending on spent total Rs. 75,000 Reported loan use this use (rupees) spent outlier Household durables 0.08 608 2.5 2.8 Food for home consumption 0.16 788 3.3 3.6 School supplies or fees 0.08 297 1.2 1.3 Religious festivals and ceremonies 0.03 270 1.1 1.2 Repairs to the house 0.08 1064 4.4 4.8 Repayment of loans 0.14 851 3.5 3.9 Savings 0.30 3935 16.3 14.8 Inventories and raw material for business 0.59 9195 38.1 41.7 Equipment for business 0.16 2504 10.4 11.3 Inventories and equipment for another business 0.11 2432 10.1 4.9 Other 0.11 2162 9.0 9.8 Total 24106 Source: Authors’ analysis based on responses from January 2007 Sri Lanka Microenterprise Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Survey (round 8). Both the difference-in-difference matching and �xed effects estimates there- fore show that the intervention increased access to credit, leading to a 5.4–7.0 percentage point increase in the proportion of �rms receiving loans in the past three months. Since only 5.9–7.2 percent of Galle and Matara �rms were receiving loans in the period prior to the intervention, this represents a doubling of the proportion receiving loans, a sizable treatment effect. Finally, columns 8 and 9 of table 7 examine whether the increase in formal loans substituted for or crowded out informal loans from moneylenders, friends, and family. Use of informal loans is limited, and the results show that the treatment had no signi�cant effect on use of informal credit. Thus, the increase in credit from the intervention did not come from a substitution away from informal credit. WHAT WAS THE IMPACT OF THIS CREDIT ON THE FIRMS RECEIVING IT? The January 2007 survey found that the most common use for the loan was buying inven- tories and raw materials: 59 percent of �rms had done this by January 2007, with the amount spent averaging 38 percent of the loan amount—42 percent when the outlier �rm that received a Rs. 75,000 loan is removed from the analysis (table 8). The next most common items were savings (reflecting that in January 2007 many �rms had just received the loan and had not yet spent it) and business equipment. Together, business assets and savings accounted for 67 percent of the loan amount. Household uses accounted for 14 percent; capital stock for other household businesses, 5 percent; and repayment of prior loans, 4 percent. The remaining 10 percent was spent on miscellaneous items. The categories and shares are similar to those observed in a previous project in de Mel, McKenzie, and Woodruff 481 which �rms received grants of Rs. 10,000–20,000 (de Mel, McKenzie, and Woodruff 2008). Of course, money is fungible, so simply asking people how they spent the loan may not reflect its true marginal effect. While an ideal comparison group against which to measure marginal impacts is lacking, a before-after compari- son and a comparison with the group of �rms not receiving a loan provide sug- gestive evidence on the marginal returns to capital. Fixed effects panel instrumental variables are used to estimate the return to capital for the �rms receiving loans (only �rms with approved loans are included): profitsi;t ¼ ai þ bKi;t þ 1i;t ð2Þ Capital stock, Ki,t, is instrumented, with a dummy variable that takes a value of 0 in the period before the loan was approved and 1 afterwards. Loan approval does signi�cantly predict capital stock; the �rst stage F-statistic is 13.53. This estimates the return to the change in capital coming from the loan. Although equation (2) is estimated on a self-selected group of borrowers, the Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 timing of loan approval was dictated by the loan intervention rather than the �rm’s decision to seek out a loan, providing at least some level of exogeneity. Moreover, so long as the return to capital is stable over the survey period, equation (2) will still correctly identify the return to capital for the �rms receiv- ing the loans. De Mel, McKenzie, and Woodruff (2008) �nd evidence consist- ent with constant returns to capital among the full sample of enterprises. Alternatively, the Galle and Matara �rms that did not receive a loan can be used as controls for time effects in the following estimate, using instrumental variables: profitsi;t ¼ ai þ bKi;t þ dt þ 1i;t ð3Þ Column 1 of table 9 then reports the implied real monthly return to capital under equation (2), which is 5.7 percent, and column 2 reports the implied return under equation (3), which is 6.6 percent. Returns to capital are similar under the two estimation strategies, suggesting that it is reasonable to assume that returns to capital are fairly constant for these �rms over the period studied. The estimated return is very similar to the 5.8 percent average return estimated in de Mel, McKenzie, and Woodruff (2008) for the full sample of �rms not directly affected by the tsunami and lower than the 9.9 percent average return for tsunami-affected �rms reported in de Mel, McKenzie, and Woodruff (2008).13 Since 32 percent of �rms receiving loans were directly affected by the 13. De Mel, McKenzie, and Woodruff (2008) discuss adjustments to account for changes in the labor hours worked by the owner in response to the intervention, which lower the returns to capital to closer to 5.0 percent. Given the small number of �rms receiving loans and the sensitivity of some of these adjustments to outliers, no adjustment is made here for hours worked, and the results are compared to the speci�cations in de Mel, McKenzie, and Woodruff (2008) that also do not adjust for hours worked. 482 THE WORLD BANK ECONOMIC REVIEW T A B L E 9 . Impact of Loan on Return to Capital (instrumental variable, �xed effects estimation) Before– after Other Galle/Matara �rms as controls (1) (2) Capital Stock 0.0577* 0.0663 (0.031) (0.085) Firm*Period observations 345 3434 Number of Firms 37 376 * p , 0.1. Note: Dependent variable is real pro�ts (rupees). Numbers in parentheses are robust standard errors clustered at the �rm level. Receipt of loan is used as an instrument for capital stock. Source: Authors’ analysis based on responses from Sri Lanka Microenterprise Survey (various rounds). tsunami, the weighted average of the 5.8 and 9.9 returns is 7.1 percent. For comparison with returns to the grant, the 1.3 percent monthly interest rate is Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 subtracted from pro�ts when calculating the return to the loan, yielding a 5.8 percent real return. The fact that the point estimate of the return to capital for �rms receiving loans is the same as that for the full sample using grants suggests that loans were not made to �rms with particularly high returns to capital. DID THE FIRMS REPAY THEIR LOANS? The intervention succeeded in increasing the number of �rms with loans. These loans were associated with higher capital stocks and higher pro�ts for the �rms receiving them. A key ques- tion for banks considering extending credit to such marginal �rms is whether the loans will be repaid. Data from wave 10 of the SLMS in October 2007 provide evidence on repayment during the �rst year of the loan. Monthly installments ranged from Rs. 500 ($5) to Rs. 2,000 ($20), with a median of Rs. 1,325. One of the 38 �rms was approved but ulti- mately decided not to take the loan. All 37 other �rms had begun repay- ments, and no �rms had defaulted. Five of the 37 �rms say they had delayed a loan payment at least once: 2 had delayed one payment, and 3 had delayed two payments. However, at the time of �nal repayment, admin- istrative data indicate that repayment rates exceeded the RDB average. Preliminary data from a resurvey of �rms in December 2010 show that 35 of 37 �rms had fully repaid their loans. Only 5 of the 37 had subsequently taken another RDB loan, and only 12 currently had a loan from any formal source. Those who took the �rst RDB loan are still more likely to have a formal loan than the other �rms in Galle and Matara or than the average �rm in Kalutara. When asked the reason for not taking another loan, approximately two-thirds of the RDB loan sample gave reasons related to lack of demand (for example, no need for a loan, it is too dif�cult to repay a loan), while one-third gave reasons related to loan supply (for example, bank rejected my application, dif�- cult to �nd guarantor, my age is higher than the threshold the bank uses now de Mel, McKenzie, and Woodruff 483 so they rejected the loan). Supply reasons still appear to affect the ability to obtain loans. I V. D I S C U S S I O N AND CONCLUSIONS The average microenterprise in Sri Lanka has returns to capital well above market interest rates but is not receiving credit. The intervention provided information on loans offered by one �nancial institution, reduced the require- ment for loan guarantors from two to one, and made minor additional changes. One in 10 microenterprises participating in the intervention received a new loan, doubling the proportion of �rms receiving a loan in the last three months. Information therefore appears to have had some impact. One-third of the microenterprises receiving the new loans had not previously thought that the bank would lend to �rms like theirs, and before the interventions few knew the interest rate and other loan terms. Since the general �nancial literacy of microenterprise owners was relatively high, it can be concluded that providing Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 product-speci�c �nancial information can improve access to credit for microen- terprises. Since repayment rates exceeded the RDB average, neither reducing the number of guarantors nor expanding the pool of borrowers appears to have had an adverse effect on the quality of the loans. This is consistent with Karlan and Zinman’s (2010) �nding that expanding the borrowing pool is pro�table for lenders. However, because the intervention also modestly relaxed lending conditions, this is likely an upper bound on the impact of information. Information alone is not enough for most microenterprises to get credit. Among those not receiv- ing loans, demand and supply constraints both contributed to the lack of credit. Some 36 percent of nonborrowers said that they did not apply for reasons attributable to a lack of demand at prevailing interest rates. However, 44 percent said they did not apply for reasons related to bank requirements for obtaining a loan. The most important of these was the need to �nd a guaran- tor. A second administrative barrier, necessitated by the absence of a credit bureau for micro�nance, was the requirement to travel to other �nancial institutions and obtain letters of endorsement showing that the �rm did not have any outstanding loans. The results suggest several avenues for policymakers and �nancial insti- tutions seeking to expand access to credit to small-scale entrepreneurs. The �rst is development of a credit registry that includes information from micro�nance organizations and development banks. This would remove the barrier of having to seek endorsements from other lenders. Second, banks and other �nancial institutions can expand their customer base through product-speci�c �nancial education. Third, more innovative ways of deliver- ing collateral-free individual loans to microenterprises need to be developed. The establishment of a credit registry that includes micro�nance would also aid in this regard. 484 THE WORLD BANK ECONOMIC REVIEW APPENDIX T A B L E A 1 . Summary of Loan Conditions and Modi�cations Usual Ruhuna Development Condition Bank (RDB) condition Modi�cation Loan amount (rupees) 5,000– 25,000 None Interest rate (annual, percent) 16 None Repayment period 2 years, equal installment or Equal installment offered declining balance Collateral Often waived on small loans Waived Guarantors Two guarantors required Allowed family members and mutual loan applicants to act as guarantors; branch discretion on number Endorsements from other Required None �nancial institutions Residence veri�cation by Grama Required None Niladhari Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Oath attested by justice of the Required Branch discretion peace Salary certi�cation by employer Not required for self-employed None Business registration Not required for self-employed None Age If over 55, joint application None with younger family member RDB account holder Must be existing account 3 month waiting period holder, 3 month waiting waived period Loan application fee (rupees) 250 None Source: Authors’ detailing of loan requirements. REFERENCES ´ riz, Beatriz, and Jonathan Morduch. 2007. The Economics of Micro�nance. Cambridge, MA: Armenda MIT Press. Banerjee, Abhijit, and Esther lo. 2008. “Do Firms Want to Borrow More? Testing Credit Constraints Using a Directed Lending Program.� MIT Department of Economics Working Paper 02-25. MIT, Cambridge, MA. Banerjee, Abhijit, Esther Duflo, Rachel Glennerster, and Cynthia Kinnan. 2010. “The Miracle of Micro�nance? Evidence from a Randomized Evaluation.� BREAD Working Paper 278. Bureau for Research and Economic Analysis of Development. http://ipl.econ.duke.edu/bread/papers.htm. ¨c Beck, Thorsten, Asli Demirgu ¸ -Kunt, and Vojislav Maksimovic. 2005. “Financial and Legal Constraints to Growth: Does Firm Size Matter?� Journal of Finance 60 (1): 137–77. Boucher, Stephen, Michael Carter, and Catherine Guirkinger. 2008. “Risk-rationing and Wealth Effects in Credit Markets: Theory and Implications for Agricultural Development.� Journal of Agricultural Economics 90 (2): 409–23. Cohen, Monique, Elizabeth McGuinness, Jennifer Sebstad, and Kathleen Stack. 2006. “Market Research for Financial Education.� Micro�nance Opportunities Working Paper 2. Micro�nance Opportunities, Washington, DC. de Mel, McKenzie, and Woodruff 485 de Mel, Suresh, David McKenzie, and Christopher Woodruff. 2008. “Returns to Capital in Microenterprises: Results from a Field Experiment.� Quarterly Journal of Economics 123 (4): 1329–72. ———. 2009. “Are Women More Credit Constrained? Experimental Evidence on Gender and Microenterprise Returns.� American Economic Journal: Applied Economics 1 (3): 1– 32. ———. 2010. “Enterprise Recovery Following Natural Disasters.� Policy Research Working Paper 5269. World Bank, Washington, DC. DFID (Department for International Development). 2008. “UK Backs Lessons in Banking to Help Africa’s Poor.� January 25. www.d�d.gov.uk/news/�les/alexander-lessons-in-banking.asp. Duflo, Esther, and Emmanuel Saez. 2003. “The Role of Information and Social Interactions in Retirement Plan Decisions: Evidence from a Randomized Experiment.� Quarterly Journal of Economics 118 (3): 815–42. Duflos, Eric, Joanna Ledgerwood, Brigit Helms, and Manuel Moyart. 2006. Country-level Effectiveness and Accountability Review: Sri Lanka. Washington, DC: Consultative Group to Assist the Poor. Dupas, Pascaline. 2011. “Do Teenagers Respond to HIV Risk Information? Evidence from a Field Experiment in Kenya.� American Economic Journal: Applied Economics 3 (1): 1– 36. Gibson, John, and David McKenzie. 2007. “Using the Global Positioning System (GPS) in Household Surveys for Better Economics and Better Policy.� World Bank Research Observer 22 (2): 217– 241. GTZ (German Agency for Technical Cooperation). 2007. “Outreach of Financial Services in Sri Lanka Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 2007.� GTZ ProMiS. Colombo, Sri Lanka. Honohan, Patrick. 2004. Financial Sector Policy and the Poor: Policy Issues and Selected Findings. Working Paper 43. Washington, DC: World Bank. Jensen, Robert. 2010. “The (Perceived) Returns to Education and the Demand for Schooling.� Quarterly Journal of Economics 125 (2): 515– 48. Karlan, Dean, and Jonathan Zinman. 2010. “Expanding Credit Access: Using Randomized Supply Decisions to Estimate the Impacts.� Review of Financial Studies 23 (1): 433– 64. Lusardi, Annamaria, and Olivia Mitchell. 2006. “Financial Literacy and Planning: Implications for Retirement Wellbeing.� Dartmouth University, Hanover, NH. Miller, Margaret. 2008. “The Irresistible Case for Financial Literacy.� World Bank, Washington, DC. McKenzie, David, and Christopher Woodruff. 2008. “Experimental Evidence on Returns to Capital and Access to Finance in Mexico.� World Bank Economic Review 22 (3): 457–82. RDB (Ruhuna Development Bank). 2007. Annual Report. Galle, Sri Lanka: RDB. World Bank. 2008. Finance for All? Policies and Pitfalls in Expanding Access. Washington, DC: World Bank. The Impact of the Business Environment on Young Firm Financing Larry W. Chavis, Leora F. Klapper, and Inessa Love A unique dataset of over 70,000 �rms, most of which are small, in over 100 countries, is utilized to systematically study the use of different �nancing sources for new and young �rms. Consistent age-related patterns emerge. Across all countries younger �rms rely less on bank �nancing and more on informal �nancing. There is a clear substi- tution effect: as �rms mature, more �rms switch out of informal �nance toward bank �nance, while the total proportion of �rms using external �nance remains relatively unchanged. Importantly, these relationships hold for �rms of different sizes, �rms in different sectors, and �rms located in countries with different income levels and on different continents. Thus, these patterns of young �rm �nancing show clear universal tendencies. Given that even small �rms increasingly use formal bank �nancing over Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 time, these results suggest that information asymmetry plays an important role in decreasing a young �rm’s ability to obtain bank �nance. JEL codes: G2, G21, O16 Access to external �nance and the ability to undertake pro�table investment opportunities is an important ingredient for success of any new business and ultimately for economic development and growth (see Levine 2005). However, liquidity constraints hinder potential entrepreneurs from starting a business (see, for example, Evans and Jovanovic 1989) and reduce growth rates, especially in small businesses (Beck, Demirguc-Kunt, and Maksimovic 2004). Relaxing these constraints can promote new �rm entry and success. For example, a cross-country study of 35 European countries �nds that that entry is higher in more �nancially dependent industries in countries that have greater �nancial development (Klapper, Laeven, and Rajan 2006). Larry W. Chavis (larry_chavis@unc.edu), assistant professor, Kenan-Flagler Business School, University of North Carolina at Chapel Hill, Leora F. Klapper (corresponding author, lklapper@ worldbank.org), lead economist, Development Research Group, The World Bank, Inessa Love (ilove@ worldbank.org), senior economist, Development Research Group, The World Bank. The authors thank the participants of the Kauffman Foundation Conference on Entrepreneurial Finance, the Cleveland Fed-Kauffman Foundation Conference on Entrepreneurial Finance, and The Second Annual Searle Center Research Symposium on the Economics and Law of the Entrepreneur and three anonymous referees for helpful comments. This paper was prepared with outstanding assistance from Douglas Randall. The opinions expressed do not necessarily represent the views of the World Bank, its Executive Directors or the countries they represent. THE WORLD BANK ECONOMIC REVIEW, VOL. 25, NO. 3, pp. 486– 507 doi:10.1093/wber/lhr045 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 486 Chavis, Klapper and Love 487 Financing options of mature �rms could be explained by the pecking order theory of �nancing (Myers and Majluf 1984). These �rms generally have more internal funds (retained earnings) due to higher pro�tability and lower growth opportunities and, therefore, might prefer to use internal funds �rst (Bulan and Yan 2009; Brealey and Myers, 2003). Furthermore, a good reputation, such as a long credit history, mitigates the adverse selection problem between bor- rowers and lenders. Mature �rms might therefore be able to obtain loans on better �nancial terms compared to their younger �rm counterparts (Bulan and Yan 2009; Carpenter and Rondi 2000) and generally use debt before equity for their �nancing needs (Bulan and Yan 2009). There is evidence from around the world that new �rms without proven track records experience more severe �nancing constraints. For instance, studies con- ducted in China, Italy and the U.S. �nd that information asymmetry signi�cantly limits the debt capacity of young �rms (Shirai 2009; Bulan and Yan 2009; and Carpenter and Rondi 2000, respectively). In addition, higher �nancing con- straints reduce the likelihood of starting a business in Thailand, especially in Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 poorer regions (Paulson and Townsend 2004). In comparison, having an existing bank relationship increases the chances of starting a business with hired employ- ees in Bosnia and improves the odds of survival for the new entrepreneur. Furthermore, according to studies of German and Canadian �rms, a higher equity ratio in new �rms has a particularly positive effect on investment in R&D, while such an effect has not been found in older �rms (Mu ¨ ller and Zimmermann 2006; Baldwin and others 2002). Without access to formal �nancing, start-up �rms might resort to informal sources. For example, family and friends provide affordable and accessible funding to Indian SME’s in start-up and growth phases (Allen and others 2006). Yet �nancing from friends and family might be “unreliable, untimely� and bearing “signi�cant non-�nancial costs� (Djankov and others 2002, p. 72). For instance, a study across 29 countries �nds that �rms choose infor- mal �nancing over more formal routes when government of�cials are corrupt as a way to avoid paying bribes (Mehnaz and Wimpey 2007); thus �rms might be willing to bear the costs of informal �nancing if there is the added bene�t of evading corruption. A study of Chinese �rms �nds that while more �rms use informal �nancing than bank �nancing, only bank �nancing is associated with higher growth rates (Ayyagari and others 2010). To the best of our knowledge, there is no systematic cross-country study of the usage of �nancing by new and young �rms. Our study attempts to �ll this gap by examining a vast �rm-level database constructed from 170 World Bank Enterprise Surveys (WBES). This database includes about 70,000 �rms, most of which are small and medium sized (SMEs), in 104 developing and developed countries, including many low-income countries.1 This database is used to study what types of �nancing are important for new �rms, relative to older 1. The complete questionnaire and database is available at http://www.enterprisesurveys.org/. 488 THE WORLD BANK ECONOMIC REVIEW �rms, and explore rich cross-sectional variation in �rm and country types. Our research addresses two questions: (1) What is the relationship between �rm age and the use of external �nancing? and (2) Does this relationship hold for differ- ent types of �rms and different groups of countries? We start by investigating whether �rms use any external �nance or rely solely on internal funds and how age affects the use of external �nance. Next, we examine the relationship between �rm age and usage of different �nancing sources, including local and foreign bank �nancing, leasing, trade credit, credit cards, family and friends, and informal lenders. We focus on the use of formal versus informal �nance over the lifecycle of �rms, controlling for �rm charac- teristics, such as sector and ownership. We �nd that around the world younger �rms use less formal (bank) �nance and more informal �nance, relative to older �rms. Bank �nance and informal �nance serve as substitutes: as �rms age, bank �nance replaces informal �nance, while the overall use of external �nance remains relatively unchanged. Interestingly, these relationships between �rm age and type of Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 �nancing hold for �rms of different sizes, �rms in different industries, �rms located in countries in different income groups and in different continents. Thus, the pattern of young �rm �nancing appears universal. Our results also highlight that the likely reason for lower levels of bank �nance in young �rms is asymmetric information. Speci�cally, this follows from our �nding that even smaller �rms increase their reliance on bank �nance (and decrease reliance on informal �nance) as they get older and build a longer credit history. While indirectly, our results suggest that improving the avail- ability and quality of credit information is important for alleviating �nan- cing constraints of young �rms. We consider �rm age as a useful proxy for entrepreneurial �rms. For example, Schumpeter wrote that a person is an entrepreneur “only when he actually carries out new combinations, and looses that character as soon as he has built up his business and settled into running it� (Schumpeter 1942). Thus, new and young �rms are more likely to retain the “entrepreneurial spirit� alluded to by Schumpeter. Other useful proxies for an entrepreneur might be �rms where an individual or family is the largest shareholder; whether the owner is also manager of the �rm; and if the �rm is registered as a sole-proprietor (relative to a limited-liability partnership or corpor- ation.) In this paper we focus mainly on young �rms as a proxy for ‘entre- preneurial’ �rms, while also exploring other samples of �rm ownership and legal types. The rest of the paper is organized as follows: Section 2 presents our data. Section 3 presents our main results on the patterns of �nancing and �rm age. Section 4 presents sample breakdowns by size, sector, income group and conti- nent. Section 5 presents our results over time. Section 6 presents multivariate regression analysis. Section 7 concludes. Chavis, Klapper and Love 489 I . D ATA The WBES dataset includes �rms across multiple sectors (manufacturing, services, agriculture, and construction). The database includes both quantitat- ive and qualitative information on �rm characteristics, including sources of �nance, barriers to growth, access to infrastructure services, legal dif�culties, and corruption. The dataset also includes some measures of �rm performance, such as multiple years of historical data on employment and sales. The database includes over 70,000 randomly sampled �rm-level obser- vations collected in 170 cross-sectional surveys in 104 countries, i.e., many countries include multiple years of data.2 The database is globally represented, which we summarize by region: Sub-Saharan Africa (AFR), East Asia (EA), South Asia (SA), Eastern Europe and Central Asia (ECA), Latin America and Caribbean (LAC), and the Middle East and Northern Africa (MENA). The database also includes a few industrialized countries (IND). Figure 1 shows the distribution of countries and observations by region. A notable difference between the two panels is that surveys in Africa include a relatively small Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 number of �rm observations and fewer countries have multiple survey years, while surveys in East Asia include a relatively large number of �rms and include multiple years. A complete list of countries and �rm observations is shown in Appendix Table 1. The distribution of �rms across income groups highlights the uniqueness of our dataset (Figure 2). Unlike similar studies of entrepreneurial �nance, which focus on �rms in the U.S. or other developed countries, �rms in our database are distributed across income groups, with a focus on developing countries— which are most likely to face barriers in the business environment. Our data- base includes 38 low-income and 37 lower-middle income countries, which account for over 73 percent of observations. The surveys were conducted over the span of eight years, 1999–2006 (Figure 2). Figure 3 shows the distribution of �rms in our sample across sectors, owner- ship, and output markets. We expect these �rm characteristics to affect the use of external �nancing, relative to other young �rms. A caveat is that some surveys focused exclusively on manufacturing companies, so in part by design, the majority of �rms in our sample are manufacturing �rms (60 percent), fol- lowed by services (30 percent) and construction (6 percent). Since manufactur- ing companies are likely to be more capital intensive, sources of entrepreneurial �nance should be particularly illustrative of country-level bar- riers to access to credit. Next, we �nd 4 percent of �rms with state ownership ( particularly in lower income countries) and 5 percent with foreign ownership. Both types of �rms might receive preferential access to �nancing. Finally, about 2. We are unable to control for whether an individual �rm is included in multiple survey years, although the likelihood of a �rm being included more than once is insigni�cant. 490 THE WORLD BANK ECONOMIC REVIEW F I G U R E 1. Distributions of Surveyed Countries and Firm Observations by Region Source: Authors’ analysis based on World Bank Enterprise Surveys. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 F I G U R E 2. Distribution of Total Firms, by Country-Level Income and Year Source: Authors’ analysis based on World Bank Enterprise Surveys. 23 percent of �rms in our sample are identi�ed as exporters, which might have greater access to overseas customer and bank �nancing. Importantly for our analysis, the WBES data is a random sampling of �rms. An important caveat, however, is that many country surveys do not include new �rms: 79 percent of surveys have a minimum �rm age of one; 8 percent a minimum �rm age two; 4 percent a minimum �rm age three, and 9 percent a minimum �rm age four.3 Therefore, summary tables include an increasing number of country and �rm observations along the age dimension. Figure 3 shows the distribution of total �rms by age. Over 8 percent of the total sample is three years old or younger, while 58 percent of all �rms are 12 years or 3. In robustness regressions, reported in section 5, we discuss our results using only surveys that include one year old �rms, i.e., eliminating the sampling bias in some countries. Chavis, Klapper and Love 491 F I G U R E 3. Distribution of Firms by Sector, Ownership, Export Status and Age Source: Authors’ analysis based on World Bank Enterprise Surveys. younger. The largest number of observations is for age four, which includes observations from all country/year surveys, regardless of survey-speci�c minimum �rm age. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 II. FINANCIAL PRODUCTS AND FIRM AGE We begin by examining whether age is related to the use of a bank line of credit (L/C) or overdraft facility. In our data, 44 percent of �rms have a L/C, compared to 32 percent of �rms that use bank �nancing to fund working capital or new investment; the correlation between a L/C and bank �nancing is about 0.50, and signi�cant at 1 percent.4 The usage of L/Cs increases with �rm age, from about 20 percent of new �rms to over 40 percent of �rms age six and older (Figure 4). This supports the hypothesis that access to bank and other sources of formal �nancing is related to �rm age. Next, we observe the complete distribution of sources for working capital and new investment �nancing, by age. The percentage of �rms that use select types of �nancing for either working capital or new investment—disaggregated by �nan- cing source and aggregated by �nancing categories—are reported by �rm age (Table 1). Throughout the paper we focus on �ve distinct categories of external �nancing sources: informal sources and family and friends (“Informal Finance�), foreign and domestic bank �nancing (“Bank Finance�), “Leasing,� “Trade Credit,� and “New Equity,� which includes equity, grants, and other sources.5 4. Bank �nance and L/C are analyzed separately because the data is drawn from two separate survey questions. Respondents are �rst asked which sources of �nancing they utilize (banks, leasing, trade credit, etc.) to fund working capital and new investment. They are then asked if their �rm has a line of credit or overdraft facility. 5. We have also experimented with aggregating leasing and trade credit into one category that could be thought of as “operational �nance,� following Allen and others (2005) and Beck and others (2008). Our results were not materially affected and are available on request. We choose to report leasing and trade credit separately because leasing is asset-backed �nance, while trade credit is largely “relationship based.� 492 THE WORLD BANK ECONOMIC REVIEW F I G U R E 4. Access to Letter of Credit by Firm Age Source: Authors’ analysis based on World Bank Enterprise Surveys. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 The columns do not sum to 100 percent since most �rms use more than one source of �nancing. The percentage of �rms using retained earnings is fairly consistent across our age categories. Similarly, there is no clear relationship between �rm age and the use of new equity or trade credit. However, like the availability of a L/C, the use of both bank �nancing and leasing are more likely as �rms mature. The relationship between age and informal �nancing runs in the oppo- site direction. Not only does the likelihood of using informal �nancing decrease over time, there is also a sharp decrease in this category—in both �nancing from friends and family and informal sources—during the �rst few years after a �rm begins operations. In addition, we construct an indicator variable called “no external �nan- cing,� which is equal to one if the �rm does not use any external sources of �nancing. This measure has somewhat of an inverted U-shape—it is low for very young �rms, higher for �rms aged 3–8 years, and then drops for older �rms. We speculate that start-up �rms need to rely more on external �nance— although in emerging markets, this �nancing is mostly from informal sources, which is signi�cantly higher in the �rst two years of �rm life than in sub- sequent years. As �rms mature, they are more likely to use formal sources of external �nance (and therefore less likely to be classi�ed into the “no external �nance� category). Table 2 shows the percentage of total Working Capital (Panel A) and New Investment (Panel B) �nancing provided by each type of �nancing (the columns do not sum to 100 percent because credit cards are excluded). The primary source of working capital �nancing for all �rm ages is retained earnings. In other words, on average, �rms in all size buckets rely primarily on their own funds for over half of their �nancing needs, which is in line with the pecking Chavis, Klapper and Love 493 T A B L E 1 . Financing Patterns (Working Capital or New Investment) by Firm Age (percent) 1– 2 3–4 5–6 7–8 9 – 10 11 – 12 13 þ Total Panel A: By Financing Type Retained Earnings 85 85 83 85 85 86 83 84 Local Banks 17 20 25 28 27 31 37 30 Foreign Banks 2 2 2 5 5 2 4 3 Leasing 3 5 7 6 7 8 7 7 Trade Credit 30 22 27 21 29 26 29 27 Credit Cards 1 2 2 2 2 3 2 2 Family & Friends 22 15 14 13 14 10 9 12 Informal Sources 10 7 4 4 3 3 3 4 Grants 3 2 3 3 3 3 4 3 Equity 10 10 10 10 10 10 10 10 Other 4 6 8 7 7 5 9 7 Panel B: By Financing Category Bank Financing 18 21 26 29 31 32 39 32 Leasing 3 5 7 6 7 8 7 7 Trade Credit 30 22 27 21 29 26 29 27 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Informal Finance 31 20 16 15 16 12 10 14 New Equity 15 17 19 18 18 17 21 19 No External Finance 38 45 43 44 36 39 35 39 Note: Panel A shows the percentage of �rms that use the �nancing source for either Working Capital or New Investment (i.e., the reported percentage of total �nancing is greater than zero). Panel B shows the percentage of �rms that use the �nancing source, by aggregated categories: “Bank Finance� includes local and foreign banks, “Leasing,� “Trade Credit,� “Informal Finance,� which includes family and friends and informal sources, and “New Equity� includes equity, grants, and other sources. Source: Authors’ analysis based on World Bank Enterprise Surveys. order theory of capital structure. However, the reliance on different types of external �nancing shows a monotonic relationship with age: for instance, older �rms use a larger percentage of bank �nance and leasing and rely less on infor- mal sources. As �rms mature, they might shift their dependence from informal sources of �nance to more formal sources.6 Trade credit is also an important source of working capital �nancing for all �rms, and becomes slightly more important as �rms get older—and establish longer relationships with their sup- pliers. Young �rms are more likely to receive infusions of new equity capital, relative to older �rms; however, these are likely to be the owner’s own funds. Summary statistics of loan characteristics are shown in Table 3. There is not a notable difference between the use of collateral, the percentage of loan size collateralized, interest rates, or maturity across �rm age. However, new �rms are less likely to have audited accounts. Table 4 shows summary statistics of the variables used in our econometric speci�cations. Panel A shows summary statistics of all �rms. (Complete 6. In addition, �rms with access to formal �nancing may be more likely to survive (to an older age). 494 THE WORLD BANK ECONOMIC REVIEW TA B L E 2 . Distribution of Firm Financing by Firm Age (percent) 1– 2 3– 4 5–6 7–8 9 – 10 11– 12 13 þ Total Panel A: Working Capital Retained Earnings 61 68 64 67 65 66 60 63 Bank Financing 9 8 10 11 11 11 16 13 Leasing 0 1 1 1 1 1 1 1 Trade Credit 12 7 11 8 9 9 11 10 Informal Finance 8 7 6 5 5 4 3 5 New Equity 9 9 9 9 9 8 9 9 Panel B: New Investment Retained Earnings 66 66 65 66 62 64 59 62 Bank Financing 9 10 12 13 16 17 20 16 Leasing 1 2 3 2 3 3 3 3 Trade Credit 3 3 5 3 4 3 3 3 Informal Finance 10 9 5 4 4 3 3 5 New Equity 10 10 10 10 10 8 11 10 Note: This table shows the percentage of total Working Capital (Panel A) or New Investment (Panel B) �nancing provided by each of these sources. “Bank Finance� includes local and foreign Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 banks; “Leasing,� “Trade Credit,� “Informal Finance,� which includes family and friends and informal sources, and “New Equity� includes equity, grants, and other sources. Columns do not sum to 100 percent as credit cards are excluded. Source: Authors’ analysis based on World Bank Enterprise Surveys. T A B L E 3 . Summary Statistics of Loan Characteristics by Firm Age 11– 1 –2 3–4 5–6 7– 8 9 – 10 12 13 þ Total Loan requires collateral (percent) 76 77 78 73 81 80 75 76 Value of collateral, relative to loan 129 132 193 130 131 138 138 141 value (percent) Interest rate (percent) 13 14 15 13 14 14 13 13 Loan duration (months) 34.7 33.2 32.2 35.2 31.5 32.3 36.2 34.5 Audited �nancial statements 42 37 43 52 54 50 61 53 (percent) Source: Authors’ analysis based on World Bank Enterprise Surveys. variable de�nitions are shown in Appendix 2). The average �rm age in our sample is 15 years, with a maximum age capped at 80 years. 27 percent of �rms are identi�ed as “Micro,� with less than 10 employees; 39 percent are identi�ed as “Small,� with less than 50 employees; and the remaining �rms as “Medium/Large.� In our sample 23 percent of �rms are exporters, 48 percent are corporations, and 52 percent have audited statements. Finally, 4 percent of �rms are identi�ed as state-owned and 5 percent as foreign-owned. Panel B shows mean tests between characteristics of �rms that use external �nance versus �rms that use only retained earnings. Older and larger �rms are more likely to use external �nance; limited liability companies (LLCs) are sig- ni�cantly more likely to use external �nance than other types (e.g., sole T A B L E 4 . Summary Statistics and Mean Tests by Category Panel B: Comparing Firms With and Without Panel C: Comparing Firms by Panel A: All Firms Financing Age Mean Obs. No External Financing Uses External Financing Firm age 5 Firm age . 5 Firm Age (years) (percentage of �rms) 15.9 68,419 14.8 16.6*** 3.6 19.4*** Micro 27 62,900 36 22*** 44 23*** Small 39 62,900 38 39** 35 40*** Medium/Large 34 62,900 26 39*** 21 37*** Limited Liability Company 48 66,701 38 54*** 40 50*** Sole Proprietorship 27 66,701 33 22*** 38 23*** Owner Manager 32 68,419 35 31*** 32 32 Exporter 23 67,480 18 26*** 16 25*** Audit 52 67,453 45 56*** 39 55*** Foreign Owned 5 68,419 5 6* 4 6*** State Owned 4 68,419 4 3*** 1 4*** Note: Complete variable descriptions are shown in Appendix 2. Panels A and B shows summary statistics for all �rms that report �nancing sources. Panel C shows summary statistics for �rms that are (i) less than or equal to �ve years old and (ii) greater than �ve years old. The second column in Panels B and C shows t-statistics for mean difference. This compares �rms with and without �nancing in Panel B and �rms more or less than �ve years old in Panel C. Asterisks *, **, and *** indicate signi�cance at 10 percent, 5 percent, and 1 percent, respectively. Source: Authors’ analysis based on World Bank Enterprise Surveys. Chavis, Klapper and Love 495 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 496 THE WORLD BANK ECONOMIC REVIEW proprietorships and partnerships); while �rms run by owner-managers are less likely to use external �nance ( possibly because they rely more on the owner’s own funds). Exporting and foreign owned �rms are more likely to use external �nance, as are those with audited �nancial statements. Panel C shows summary statistics and mean tests disaggregated by �rms (i) less than or equal to �ve years old, and (ii) �rms older than �ve years. These preliminary statistics show large and signi�cant differences between younger and older �rms. For instance, young �rms are almost twice as likely to be micro (de�ned as less than 10 employees) and signi�cantly less likely to be exporters or to have audited �nancial statements. III. BREAKDOWN BY GNI, REGION, FIRM SIZE, AND SECTOR In this section, we consider how the patterns described above vary across �rms operating in countries with different levels of economic development (measured by income level) and across regions, and the relationship with �rm size and Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 sector. We begin with income groups. Figure 5 shows the patterns of bank �nance, informal �nance, and leasing. Across income groups, the patterns appear similar for these three key measures. Bank �nancing (Figure 5) increases with age in all groups—almost doubling by the time a �rm reaches 13 years of age, relative to new �rms—and is highest (for all ages) in high income countries. The use of informal �nance (Figure 5) declines with age across all income groups, and it is highest in lower-income countries across all age groups. The percentage of �rms using leasing �nance is signi�cantly larger in countries with middle-high and high income (Figure 5). Leasing �nance is a more sophisticated type of �nance and is rarely available in lower income countries. In high income countries there is a clear positive relationship between leasing and age. The use of trade credit (not shown) does not seem to be systematically related to income. Our measure of “no external �nancing� does not show signi�cant changes over the life of the �rm (not shown). The relatively stable proportion of �rms with no external �nance is likely be explained by substitution between bank �nance and informal �nance—as �rms mature they may switch from one source to another, while their reliance on external �nance does not change dramatically over the life cycle. (Later we show a more formal empirical test for whether this relationship is statistically signi�cant in a multivariate framework). We also consider whether there are regional variations in bank �nance and informal �nance using World Bank Country classi�cations for the sample of developing countries (Figure 6). Similar patterns emerge in all regions—an increasing use of bank �nance with age and a declining use of informal �nance with age. Again, there are no strong patterns for our “no external �nance� measure (not reported). Next, we present the relationship between age and sources of �nancing for �rms of varying sizes. For every age level, smaller �rms use less bank �nance Chavis, Klapper and Love 497 F I G U R E 5. Financing Patterns by Income Groups Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Source: Authors’ analysis based on World Bank Enterprise Surveys. and more informal �nance than larger �rms (Figure 7). In fact, the lines for micro, small, and medium/large �rms are almost parallel (and upward slop- ping). These patterns are consistent with stylized facts in the literature on the difference between �rm sizes and their usage of bank �nance (Beck, Demirguc-Kunt, and Maksimovic 2008). While banking �nancing increases, informal �nance gradually declines for all �rm sizes (Figure 7). The fact that even small �rms use more bank �nance as they age points to informational asymmetries as an important reason for lack of �nance for younger �rms, for example, �rms can build positive credit histories over time. There is no strong pattern between the category “no exter- nal �nance� and �rm age, however, smaller �rms use less external �nance for all age groups (not shown). 498 THE WORLD BANK ECONOMIC REVIEW F I G U R E 6. Financing Sources by Region Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Source: Authors’ analysis based on World Bank Enterprise Surveys. Finally, in regard to sectoral differences, bank �nance is increasing and informal �nance is declining in all four sectors. Firms in the manufacturing sector are most likely to use bank �nance and �rms in the service sector are least likely to do so (Figure 8). This is possibly because of collateral require- ments—service �rms are less likely to have any signi�cant �xed assets to use as collateral. Service �rms also use the most informal �nance, especially in their early years (Figure 8). There are no signi�cant differences across sectors in the use of external �nance (not shown). To summarize, our �ndings suggests that bank �nance increases with �rm age and informal �nance declines with �rm age for all groups of �rms and countries in our sample. I V. F I N A N C I N G PAT T E R N S OV E R T I M E To study the relationship over time, we examine a sample of countries with surveys in multiple years. For a subset of 22 countries in Europe, we have three Chavis, Klapper and Love 499 F I G U R E 7. Financing Sources by Size Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Source: Authors’ analysis based on World Bank Enterprise Surveys. repeated surveys: in 1999, 2002 and 2005.7 We show the average of each main source of �nancing and the “no external �nance� indicator over these three repeated cross-sections. Among this group of countries, bank �nance was the lowest in 2002, and the highest in 2005 (Table 5). Leasing and trade credit do not show changes over time, while the use of informal �nance decreased over this time frame from about 20 percent of all �rms using informal �nance in 1999 to about 13 percent in 2005. These patterns reflect the increasing use of formal �nancing by these �rms, and, in parallel, the declining reliance on infor- mal �nance. The proportion of �rms without any external �nance is higher in 2002 and 2005 relative to 1999. The previously outlined trend of the use of banking increasing with �rm age while the use of informal �nancing decreases continues to hold in this sample (Figure 9). An important caveat to this analysis is that this is not a panel of identical �rms, but rather repeated cross-sections. Therefore, the sample composition may affect these results. 7. This is the sample of Business Environment and Enterprise Performance Surveys (BEEPS), which is a joint initiative of the European Bank for Reconstruction and Development (EBRD) and the World Bank Group. 500 THE WORLD BANK ECONOMIC REVIEW F I G U R E 8. Financing Sources by Sector Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Source: Authors’ analysis based on World Bank Enterprise Surveys. T A B L E 5 . Financing Patterns Over Time (percent) Survey Year Banking Leasing Trade Credit Informal No External Financing 1999 24.9 12.7 16.5 19.9 40.4 2002 19.2 11.5 20.0 18.0 45.4 2005 32.2 11.6 14.8 13.3 46.6 Note: Table shows the percentage of �rms that use each �nancing source for either Working Capital or New Investment by aggregated categories: “Bank Finance� includes local and foreign banks; “Leasing�; “Trade Credit�; “Informal Finance,� which includes family and friends and informal sources; and “New Equity� includes equity, grants, and other sources. This is based on the 22 European countries for which surveys were carried out in three separate years. Source: Authors’ analysis based on World Bank Enterprise Surveys. In addition, we investigate whether there was “mean-reversion� over time. Countries that fell below average in the usage of bank �nance in the earlier part of the sample experienced faster growth in bank �nance (Figure 10). This Chavis, Klapper and Love 501 F I G U R E 9. Financing Patterns Over Time Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Source: Authors’ analysis based on World Bank Enterprise Surveys. relationship is even more signi�cant with regards to informal �nance (Figure 10). Thus, countries with the highest usage of informal �nance in 1999 experienced the largest declines in the use of informal �nance (i.e., a negative change). Similar patterns occur for “no external �nance� measures (Figure 10). These graphs show a convergence in �nancing indicators over time and less variation in these measures in the later part of the sample than in the earlier years. V. M U L T I VA R I A T E R E G R E S SI O N A N A LY S I S In this section we investigate the relationship between our four main categories of external �nance and age using a simple multivariate regression framework. 502 THE WORLD BANK ECONOMIC REVIEW F I G U R E 10. Mean Reversion in Financing Patterns Over Time Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Source: Authors’ analysis based on World Bank Enterprise Surveys. While we cannot claim causality, this analysis serves to verify the patterns observed earlier, while controlling for different sample compositions across countries. Table 6 reports regressions for each of the sources of �nance de�ned above: bank �nance, leasing, trade credit, and informal �nance. In this table, the dependent variables are equal to one if the �rm uses a type of �nancing for either working capital or new investment, and zero otherwise. In addition, we include a dummy which is equal to one if a �rm uses a line of credit or over- draft facility. We estimate the model by probit with standard errors clustered by country and year (because several countries have more than one survey). In our main regressions (columns 1–5), we exclude �rms that use zero external �nancing, since we are unable to disentangle whether these �rms rely on T A B L E 6 . Is There a Relationship between Sources of Finance and Firm Age? (1) (2) (3) (4) (5) (6) Line of Credit Bank Finance Leasing Trade Credit Informal Finance No External Financing Ln Firm Age 0.039 0.041 2 0.001 0.017 2 0.054 2 0.012 [0.00]*** [0.01]*** [0.74] [0.07]* [0.00]*** [0.16] Micro 2 0.258 2 0.230 2 0.051 2 0.081 0.163 0.179 [0.00]*** [0.00]*** [0.00]*** [0.00]*** [0.00]*** [0.00]*** Small 2 0.173 2 0.150 2 0.025 2 0.003 0.059 0.067 [0.00]*** [0.00]*** [0.00]*** [0.90] [0.01]*** [0.00]*** Limited Liability 0.050 0.045 0.010 0.013 2 0.036 2 0.083 [0.02]** [0.01]*** [0.15] [0.57] [0.01]** [0.00]*** Exporter 0.045 0.030 0.005 0.014 2 0.018 2 0.009 [0.00]*** [0.08]* [0.23] [0.64] [0.17] [0.59] Audit 0.088 0.050 0.009 0.000 2 0.027 2 0.050 [0.00]*** [0.02]** [0.08]* [0.98] [0.04]** [0.00]*** Foreign Owned 2 0.008 0.043 2 0.011 0.069 2 0.142 0.014 [0.82] [0.54] [0.06]* [0.49] [0.00]*** [0.76] State Owned 2 0.016 2 0.167 2 0.036 2 0.012 2 0.118 0.063 [0.81] [0.00]*** [0.00]*** [0.66] [0.00]*** [0.05]* Observations 37,434 37,083 27,485 37,049 37,061 58,771 Censored Obs. (percent) 43.80 51.10 13.40 43.90 21.90 38.90 Pseudo R2 0.27 0.19 0.21 0.2 0.14 0.12 Note: Table 6 reports probit estimates with country �xed effects. The dependent variable in the �rst column is a dummy equal to one if the �rm reports using a line of credit or overdraft facility; the second column is a dummy equal to one if the �rm uses local or foreign bank �nancing (“Bank Finance�), the third column is a dummy equal to one if the �rm uses leasing; the fourth column is a dummy equal to one if the �rm uses trade credit, and the �fth column is a dummy equal to one if the �rm uses informal �nancing. We exclude �rms that do not use any source of external �nance (i.e., retained earnings equals 100 percent). All variables are de�ned in Appendix 2. All regressions include sector �xed effects (manufacturing, services, and construc- tion), country-level �xed-effects, and survey year �xed effects. P-values based on standard errors clustered by country and survey year are reported. Chavis, Klapper and Love Asterisks *, **, and *** indicate signi�cance at 10 percent, 5 percent, and 1 percent respectively. Source: Authors’ analysis based on World Bank Enterprise Surveys. 503 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 504 THE WORLD BANK ECONOMIC REVIEW internal �nancing by choice or because they have been rejected by external creditors. However, in column 6 we present regressions for our “no external �nancing� indicator to test whether there is a signi�cant relationship with age and other controls. The regressions control for a number of �rm characteristics, such as dummies indicating micro and small sized �rms (medium/large �rms are the omitted category), exporters, �rms with audited statements, legal status (limited liability vs. other types), and state and foreign ownership. We also include dummies for sector �xed effects (manufacturing, services, and construc- tion), country-level �xed effects, and survey year. The key variable of interest is the log of �rm age. The results are similar to the univariate results discussed before: bank �nance is gradually increasing with age, while informal �nance is gradually decreasing with age. The usage of L/Cs follows a similar increasing trend over the life of the �rm. We do not �nd a signi�cant pattern for leasing �nance, though trade credit use is slightly increasing with age (and the likely length of Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 supplier relationships). We also observe a negative, but not statistically signi�- cant, relationship between �rm age and the use of external �nance, which is similar to the graphical analysis presented earlier. The multivariate regressions show that the descriptive results discussed earlier are not driven by different composition of �rms across countries, or different country-level characteristics (which are captured by �rm-level control variables and country dummies). However, these regressions simply show partial corre- lation and do not attempt to establish causality of �rm age and access to external �nance or types of �nance, which is a limitation of our dataset. VI. CONCLUSIONS In this paper, we systematically study the use of different �nancing sources for new and young �rms. We use a unique dataset from over 170 surveys, which contain about 70,000 �rms, most of which are small and medium sized (SMEs) in 104 developing and developed countries, including many low-income countries. We use this dataset to examine corporate �nancing decisions, focus- ing on the relationship between �rm age and sources of external �nancing. Across all countries, younger �rms have less reliance on bank �nancing and more reliance on informal �nancing. The relationship with leasing and trade credit is less associated with �rm age. This relationship holds for �rms of different sizes, �rms in different industries, and �rms located in countries in different continents and different income groups. Over time, countries tend to converge; that is, countries with low bank �nancing increase reliance on bank �nancing faster than countries with initial higher levels of bank �nance. Chavis, Klapper and Love 505 APPENDIX T A B L E A 1 . List of Countries and Number of Observations Albania 537 Germany 1,196 Nepal 223 Algeria 557 Greece 546 Nicaragua 452 Angola 540 Guatemala 455 Niger 125 Argentina 1,063 Guinea 327 Nigeria 232 Armenia 647 Guinea-Bissau 296 Oman 337 Azerbaijan 657 Guyana 163 Pakistan 965 Bangladesh 1,001 Honduras 450 Panama 604 Belarus 707 Hungary 1,007 Paraguay 613 Benin 197 India 2,722 Peru 1,208 Bhutan 98 Indonesia 713 Philippines 716 Bolivia 1,284 Ireland 501 Poland 1,829 Bosnia 509 Kazakhstan 982 Portugal 505 Botswana 444 Kenya 284 Romania 980 Brazil 1,642 Korea, Rep. of 598 Russia 1,659 Bulgaria 1,228 Kosovo 329 Rwanda 340 Burkina Faso 51 Kyrgyz Republic 609 Saudi Arabia 681 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Burundi 407 Lao PDR 246 Senegal 262 Cambodia 503 Latvia 547 Serbia 550 Cameroon 119 Lebanon 354 Slovakia 528 Cape Verde 47 Lesotho 75 Slovenia 536 Chile 948 Lithuania 756 South Africa 603 China 3,948 Macedonia, FYR 506 Spain 606 Colombia 1,000 Madagascar 293 Sri Lanka 452 Costa Rica 343 Malawi 160 Swaziland 429 Croatia 550 Malaysia 902 Syrian Arab Rep. 560 Czech Republic 760 Mali 155 Tajikistan 483 Dominican Republic 250 Mauritania 361 Tanzania 760 Ecuador 453 Mauritius 212 Thailand 1,385 Egypt 1,973 Mexico 1,480 Turkey 2,544 El Salvador 465 Moldova 766 Uganda 3,099 Eritrea 79 Mongolia 195 Ukraine 5,004 Estonia 521 Montenegro 100 Uzbekistan 660 Ethiopia 427 Morocco 1,709 Vietnam 1,650 Gambia, The 301 Mozambique 194 Zambia 207 Georgia 374 Namibia 429 Source: Authors’ analysis based on World Bank Enterprise Surveys. 506 THE WORLD BANK ECONOMIC REVIEW T A B L E A 2 . Variable De�nitions and Mean Statistics Variable Name De�nition Mean Measures of Access to Finance Bank Finance Dummy (0/1) ¼ 1 if the �rm uses local or foreign bank �nance 0.32 for working capital or new investment, and ¼ 0 otherwise. Leasing Dummy (0/1) ¼ 1 if the �rm uses leasing for working capital or 0.7 new investment and ¼ 0 otherwise. Trade Credit Dummy (0/1) ¼ 1 if the �rm uses trade credit for working capital 0.27 or new investment and ¼ 0 otherwise. Informal Finance Dummy (0/1) ¼ 1 if the �rm uses informal �nance or family and 0.14 friends for working capital or new investment and ¼ 0 otherwise. Equity Finance Dummy (0/1) ¼ 1 if the �rm uses new equity, grants, or ‘other’ 0.19 �nancing for working capital or new investment and ¼ 0 otherwise. Retained Earnings Dummy (0/1) ¼ 1 if the �rm uses retained earnings for 100 0.20 percent of working capital and new investment �nancing and ¼ 0 otherwise. No External Dummy (0/1) ¼ 1 if the �rm does not use any external �nancing 0.39 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Financing for working capital or new investment and ¼ 0 otherwise. Survey responses suggest owners in many countries confused equity and retained earnings, thus we do not consider new equity to be external �nancing in de�ning this variable. General Firm Characteristics Firm Age Continuous variable equal to �rm age in years. 16.08 Micro Dummy (0/1) ¼ 1 if the �rm has less than 10 employees and ¼ 0 0.26 otherwise. Small Dummy (0/1) ¼ 1 if the �rm has 10-49 employees and ¼ 0 0.39 otherwise. Med/Large Dummy (0/1) ¼ 1 if the �rm has 50 or more employees and ¼ 0 0.35 otherwise. (Excluded category). Exporter Dummy (0/1) ¼ 1 if the �rm exports more than 10 percent of its 0.23 goods and ¼ 0 otherwise. Limited Liability Dummy (0/1) ¼ 1 if the �rm is a publicly listed company or a 0.48 privately held, limited company and ¼ 0 otherwise. (cooperatives, partnerships and sole-proprietors are the excluded categories). Audit Dummy (0/1) ¼ 1 if the �rm has audited �nancial statements and 0.53 ¼ 0 otherwise. Foreign Owned Dummy (0/1) ¼ 1 if the largest shareholder is a foreign company 0.5 and ¼ 0 otherwise. State Owned Dummy (0/1) ¼ 1 if the �rm has state ownership and ¼ 0 0.4 otherwise. Country Characteristics Low Income Group Dummy (0/1) ¼ 1 for countries with GNI per capita less than 0.37 $766, and zero otherwise (WB-WDI). Lower– Middle Dummy (0/1) ¼ 1 for countries with GNI per capita between 0.36 Income Group $766 and $3,035, and zero otherwise (WB-WDI). (Continued ) Chavis, Klapper and Love 507 TABLE A2. Continued Variable Name De�nition Mean Upper – Middle Dummy (0/1) ¼ 1 for countries with GNI per capita between 0.19 Income Group $3,036 and $9,385, and zero otherwise (WB-WDI). High Income Group Dummy (0/1) ¼ 1 for countries with GNI per capita in excess of 0.8 $9,385, and zero otherwise (WB-WDI). REFERENCES Allen, Franklin, Rajesh Chakrabarti, Sankar De, Jun Qian, and Meijun Qian. 2006. “Financing Firms in India.� World Bank Policy Research Working Paper No. 3975. Washington, D.C. Ayyagari, Meghana, Asli Demirgu ¨ c-Kunt, and Vojislav Maksimovic. 2010. “Formal versus Informal Finance: Evidence from China.� Review of Financial Studies 23 (8): 3048– 3097. ´ rie Gaudreault. 2002. “Financing innovation in new small Baldwin, John R., Guy Gellatly, and Vale �rms: new evidence from Canada.� Statistics Canada analytical studies series. Working Paper No. 190. Beck, Thorsten, Asli Demirguc-Kunt, and Vojislav Maksimovic. 2005. “Financial and Legal Constraints Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 to Firm Growth: Does Firm Size Matter?� Journal of Finance 60 (1): 137– 177. Beck, Thorsten, Asli Demirguc-Kunt, and Vojislav Maksimovic. 2008. “Financing Patterns around the World: Are Small Firms Different?� Journal of Financial Economics 89 (3): 467– 487. Brealey, Richard A., and Stewart C. Myers. 2003. Financing and Risk Management. New York: McGraw-Hill. Brown, Gregory, Larry Chavis, and Leora Klapper. 2008. “A New Lease on Life: Institutions, External Financing, and Business Growth.� World Bank Policy Research Working Paper. Washington, D.C. Bulan, Laarni, and Zhipeng Yan. 2009. “The Pecking Order of Financing in the Firm’s Life Cycle.� Banking and Financing Letters 1 (3): 129–140. Carpenter, Robert E., and Laura Rondi. 2000. “Italian Corporate Governance, Investment, and Finance.� Emperica 27 (4): 365– 388. Djankov, Simeon, Ira Lieberman, Joyita Mukherjee, and Tatiana Nenova. 2002. “Going Informal: Bene�ts and Costs.� In Boyon Belev, ed., The Informal Economy in the EU Accession Countries. So�a: Center for the Study of Democracy, 63– 80. Evans, David S, and Boyan Jovanovic. 1989. “An Estimated Model of Entrepreneurial Choice under Liquidity Constraints.� Journal of Political Economy 97 (4): 808–27. Klapper, Leora, Luc Laeven, and Raghu Rajan. 2006. “Barriers to Entrepreneurship.� Journal of Financial Economics 82(3): 613– 641. Levine, R. 2005. “Finance and Growth: Theory and Evidence.� In Philippe Aghion and Steven Durlauf, eds., Handbook of Economic Growth. Amsterdam: Elsevier Science. Mehnaz, Safavia, and Joshua Wimpey. 2007. “When Do Enterprises Prefer Informal Credit?� World Bank Policy Research Working Paper No. 4435. Washington, D.C. ¨ ller, Elisabeth, and Volker Zimmermann. 2009. “The Importance of Equity Finance for R&D.� Mu Small Business Economics 33 (3): 303–318. Myers, Stewart, and Nicholal S. Majluf. 1984. “Corporate Financing and Investment Decisions when Firms have Information that Investors Do Not Have.� Journal of Financial Economics 33 (3): 187–221. Paulson, Anna L., and Robert Townsend. 2004. “Entrepreneurship and �nancial constraints in Thailand.� Journal of Corporate Finance 10 (2): 229–262. Shirai, Sayuri. 2009. “Banks’ Lending Behavior and Firms’ Corporate Financing Patterns in People’s Republic of China.� Keio University, Faculty of Policy Management Working Paper. Tokyo. Schumpeter, Josef. 1942. Capitalism, Socialism and Democracy. New York: Hampers and Brothers. Does a Picture Paint a Thousand Words? Evidence from a Microcredit Marketing Experiment ´ *, Ghazala Mansuri, and Mario Pico Xavier Gine ´n Female entrepreneurship is low in many developing economies partly due to con- straints on women’s time and mobility, often reinforced by social norms. We analyze a marketing experiment designed to encourage female uptake of a new microcredit product. A brochure with two different covers was randomly distributed among male and female borrowing groups. One cover featured 5 businesses run by men while the other had identical businesses run by women. We �nd that both men and women respond to psychological cues. Men who are not themselves business owners, have lower measured ability and whose wives are less educated respond more negatively to the female brochure, as do women business owners with low autonomy within the household. Women with relatively high levels of autonomy shown the male brochure have a similar negative response, while there is no effect on female business owners with autonomy shown the female brochure. Overall, these results suggest that women’s response to psychological cues, such as positive role models, may be mediated by their autonomy and that more disadvantaged women may require more intensive interventions. JEL codes: G21, D24, D83, O12 Women in developing countries face numerous barriers to participation in economic life. In addition to constraints on time and limited access to capital, women’s exposure to markets and business networks is often also limited by * Gine ´ (Corresponding author): Development Research Group, The World Bank (e-mail:xgine@ worldbank.org). Mansuri: Development Research Group, The World Bank (e-mail: gmansuri@ worldbank.org). Pico ´ n: Development Research Group, The World Bank (e-mail: mpicon@worldbank. org). We are indebted to three anonymous referees for extremely insightful comments which have substantially improved the paper. We also thank Jonathan Zinman and Tahir Waqar for valuable discussions. We are especially grateful to the following for their help and support in organizing the experiment: Shahnaz Kapadia, at ECI Islamabad, for her help with designing the brochure; Irfan Ahmad at RCONs, Lahore, for managing all �eld operations and data collection; Dr Rashid Bajwa, Agha Javad, Tahir Waqar and the �eld staff at NRSP for implementing the intervention; Qazi Azmat Isa, Kevin Crockford and Imtiaz Alvi at the World Bank Of�ce in Islamabad, and Kamran Akbar at the Pakistan Poverty Alleviation Program (PPAF) in Islamabad for their support and encouragement. This project was jointly funded by the World Bank (Development Research Group, South Asia Region, Poverty Reduction and Equity Network, Gender Group), the PPAF and the Kaufmann Foundation. Santhosh Srinivasan provided outstanding research assistance. The views expressed herein are those of the authors and should not be attributed to the World Bank, its executive directors, or the countries they represent. THE WORLD BANK ECONOMIC REVIEW, VOL. 25, NO. 3, pp. 508– 542 doi:10.1093/wber/lhr026 Advance Access Publication July 10, 2011 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 508 ´ , Mansuri, and Pico Gine ´n 509 low mobility and lack of education. Weak decision-making power within the household often reinforces these disadvantages, further limiting women’s ability to secure time or resources for their own productive activities. One manifestation of this is the comparatively low prevalence of female-owned businesses. Even as entrepreneurs, however, women operate businesses that are far smaller in scale and pro�tability than male businesses. The recognition that women may therefore be most in need of credit for small businesses has played an important role in the woman-centered microcre- dit movement of the last two decades (Yunus, 1999). In Pakistan, however, micro�nance as well as other support for micro-entrepreneurship has focused primarily on men while women remain at the margins of economic life. This paper reports on an experiment designed to encourage female uptake of a new microcredit product that allowed eligible borrowers the opportunity to borrow up to four times the typical loan size.1 In cooperation with the micro�nance institution, we designed two different brochures that provided information about the characteristics of the loan and described the application process. The brochures were identical, except for the cover page: one featured 5 different businesses with men operating them, while the other had the exact same �ve businesses, but with women entrepreneurs, instead. Groups of bor- rowers targeted to be offered the new product were randomized to receive either the brochure with the female or the male pictures. We �nd that this form of marketing affects both male and female clients but quite differently. In the full sample of clients, the brochure has little impact on loan demand among either men or women. When we focus on business owners, however, we �nd that exposure to the female brochure substantially decreases demand for the larger loan among women clients but has no impact on male clients. Importantly, this is not an artifact of business scale. While women typically operate much smaller businesses, we show that women who operate businesses which are comparable in scale to male businesses also react negatively to the female brochure. However, once we allow the response to the brochure picture to vary by individual characteristics, we �nd that this negative effect is concentrated among women business owners with low decision making power. For this group of women, we conjecture that it is the male decision maker’s reaction to the female brochure that matters. We �nd that men also react negatively to the female brochure, but only among non-entrepreneurs, individuals with poor digit span recall (correlated with edu- cation and entrepreneurship) or with wives who are relatively poorly educated. This could be interpreted as evidence of af�nity (Evans, 1963; Mobius and Rosenblat, 2006), but it is also consistent with men’s low regard for females as business owners (see Beaman et al. 2009 who study perceptions about female 1. While supply constraints may also be an important determinant of lack of credit to small �rms (Banerjee and Duflo, 2008; de Mel et al. 2010), our sample consists of individuals that are already micro�nance clients. 510 THE WORLD BANK ECONOMIC REVIEW politicians). Interestingly, female business owners with high decision making autonomy shown the male brochure also react negatively by roughly the same magnitude, while there is no effect on female business owners with autonomy shown the female brochure. These �ndings contribute to the relatively limited evidence on the impact of marketing on behavior change and in this sense, Bertrand et al. (2010) is the paper closest to ours. They analyze a direct mail �eld experiment in urban South Africa that randomized advertising content, loan price, and loan offer deadlines simultaneously. Their subjects, like ours, were existing micro�nance clients. Bertrand et al. �nd that advertising content matters, especially among men, but it is mostly treatments that appeal to intuition (such as a picture) as opposed to reason (like a comparison of interest rates across lenders) that influ- ence behavior. The reason why reflexive cues are less relevant is perhaps because individuals that received the loan offers had rationally decided not to borrow, and so were already familiar with the terms of the loan made salient by reflexive treatments (Kahneman, 2003). In Bertrand et al (2010) men respond positively to the image of women credit of�cers in a context where this creates no threat to relative male auth- ority within the household or within the community. In our experiment, however, the picture of a female entrepreneur on a brochure advertising a larger loan challenges local norms of relative power. In this sense, the paper is relevant to the growing literature on social norms which uses traditional mar- keting techniques to alter attitudes and behavior by changing individual per- ceptions (see, for example, La Ferrara et al., 2008; Jensen and Oster, 2007 and Paluck, 2009). The remainder of the paper is structured as follows. Section I describes the context in Pakistan and the marketing experiment. Section II discusses the data, Section III describes the empirical strategy and Section IV reports the results of the experiment. Section V discusses the policy implementations of the results and concludes. I. SETTING AND INTERVENTION Pakistan has a population of over 162 million, with over 60% living in rural areas. Although the agricultural sector continues to be important for overall growth, 45 percent of the rural poor rely on non-farm activities as important sources of income.2 Pakistan’s �nancial system has grown signi�cantly in the past few years due in part to the successful implementation of various �nancial sector reforms, including the granting of banking licenses to a number of new private banks in the early 1990s, the modernization of the governance and regulatory 2. Source: World Development Indicators database, April 2009 and World Development Report 2008. ´ , Mansuri, and Pico Gine ´n 511 framework of the banking sector in the late 1990s, and the privatization of major public sector banks since the early to mid-2000s. Despite these recent achievements, access to �nancial services is still quite limited, especially in rural areas. According to the Access to Finance dataset (Nenova et al. 2009), only 14 percent of households interviewed reported using a �nancial product or service (including savings, credit, insurance, pay- ments, and remittance services) from a formal �nancial institution.3 When informal �nancial access is taken into account, however, this �gure rises to just over 50 percent. Overall, rural �rms and households account for about 7 percent of total credit disbursement (about Rs 130.7 billion) and the bulk of this is for agricul- tural �nance (Rs 108.7 billion), including both farm and nonfarm credit (see Akhtar, 2008). While microcredit volumes are skewed towards rural areas, microcredit currently accounts for only 17 percent of total rural credit and serves some 1.7 million clients. Comparative rates of micro�nance penetration in the South Asian region are 35 percent in Bangladesh, 25 percent in India, and 29 percent in Sri Lanka. Among micro�nance providers, Khushali Bank is active in 86 districts; National Rural Support Program (NRSP) comes second with a presence in 51 districts while Kashf Foundation has some presence in 24 districts.4 These three micro�nance entities account for approximately 70 percent of the sector’s active clients (MicroWatch, 2008). Unlike most other countries, the micro�nance sector has focused primarily on men rather than women on the grounds that there is less demand for credit from women give their low mobility levels and cultural norms around women as economic actors. Consistent with this, Pakistan continues to under-perform on a range of social indicators relative to other countries at similar levels of per capita income and rural development. According to the 1998 Human Development Report, for example, Pakistan ranked 138 out of 174 on the Human Development Index, 131 out of 163 on the Gender Development Index (GDI), and 100 out of 102 on the Gender Empowerment Measure (GEM). National Rural Support Program Established in 1991, NRSP is the largest of the Rural Support Programs in the country in terms of outreach, staff and development activities. It is modeled after the Aga Khan Rural Support Program, established in the early 1980s as a not-for-pro�t rural development organization. During the early 1990s, NRSP remained small, but the establishment of the Pakistan Poverty Alleviation Fund in 2000, a second-tier funding and capacity-building apex, provided critical 3. In comparison, 32 percent of the population has access to the formal �nancial system in Bangladesh, and this �gure amounts to 48 percent in India and 59 percent in Sri Lanka (World Bank, 2008). 4. Both NRSP and Kashf obtain a large fraction of their loan funds from the Pakistan Poverty Alleviation Fund (PPAF) which supported this work. 512 THE WORLD BANK ECONOMIC REVIEW funding that fueled NRSP and other partner NGOs’ growth. As part of its growth strategy, NRSP applied for a micro�nance bank license in 2008 and became a micro�nance bank in early 2010, falling under the supervision of the State Bank of Pakistan. Micro�nance banks now account for almost half of the outreach of the micro�nance sector. NRSP makes loans largely to members of a community organization. Its staff supports the creation of community organizations (CO) by a process of social mobilization which includes the creation of a community co-�nanced and co-managed infrastructure project and skill and group management train- ing. Members of a CO typically live close to each other and meet regularly. Most also contribute towards individual and group savings. Up to date, NRSP has organized more than a million poor households into a network of more than 100,000 COs across the country. Roughly one-half to two-thirds of CO members are also active borrowers and group meetings serve as the venue for the receipt and repayment of loans for most members. NRSP has three main credit products: a single installment loan for agricul- tural inputs (fertilizer, seeds, etc) with maturity of 6 to 12 months; enterprise loans and loans for livestock that have 12 monthly installments each. The maximum amount that can be borrowed depends on the number of loans suc- cessfully repaid (loan cycle). A new borrower starts with a small loan limit of Rs 10,000 (USD 117)5 which can increase in intervals of up to Rs 5,000 per loan cycle. As a point of comparison, a cow costs around Rs 60,000. All loans have joint liability at the CO level although new loans are issued even if some CO members are overdue.6 Besides credit, NRSP offers training in various vocational skills and provides up to 80 percent �nancing for infrastructure projects in the village. The Experiment and Marketing Intervention The study was conducted in �ve branches in the districts of Bahawalpur, Hyderabad, and Attock, spanning different agro-climatic regions of Pakistan.7 Figure 1 shows the location of the study districts. NRSP staff conducted a complete listing of the occupation of CO members in the study branches to identify those who were engaged in a non-farm activity. After the listing, a baseline survey was conducted in November 2006 in a sample of 747 COs, selected so that their membership was between 5 and 26 members. The original sampling framework included all CO members that according to the listing exercise had a non-farm business and �ve other members selected at random from each CO. In practice, enumerators ended up 5. Currency converter accessed online on July 2nd, 2010. 6. Borrowers are required to �nd two guarantors, who can be members of the same CO. NRSP staff uses guarantors as a means of exerting peer pressure, rather than enforcing repayment from them. 7. These branches are as follows: Matiari and Tando Muhammad Khan in Hyderabad, Attock in Attock and Bahawalpur (rural and urban) in Bahawalpur. ´ , Mansuri, and Pico Gine ´n 513 F I G U R E 1. Pakistan Study Districts interviewing everyone that attended a special CO meeting that was called to conduct the baseline survey. The resulting sample consisted of a total of 4,162 members interviewed, and 2,284 members (54.9%) that were in good standing. The timeline of the experiment is presented in Figure 2. Using data from the listing exercise, COs were randomly allocated into two groups, one of which was assigned to receive business training. Training ses- sions were held, from February to May 2007. Each session lasted for 6 to 8 days (see Gine´ and Mansuri (2011a) for more details about the business train- ing intervention). After completing the business training sessions, members from all study COs were invited to an orientation meeting that introduced the possibility of bor- rowing a larger loan amount. Most orientation sessions took place in regularly scheduled CO meetings and lasted for about an hour and a half. Attendance at these sessions was high, with more than 90 percent of members attending. Message consistency during the orientation was maintained by providing 514 THE WORLD BANK ECONOMIC REVIEW F I G U R E 2. Timeline training to all NRSP credit of�cers and other staff who were in charge of deli- vering the orientations.8 During the orientation meeting, members who were in good standing i.e., those who had successfully repaid at least one loan on time received one of two versions of a marketing brochure. Orientations occurred successfully in 596 COs. In the remaining 151 COs orientation meetings could not be held because the CO had either disbanded or was newly formed so that none of its members was eligible for the lottery.9 The brochure was identical in all respects except one. In one version, the entrepreneurs manning the business were male while in the other they were female. To ensure that only the gender of the busi- nessperson differed between both versions of the brochure, the exact same business was photographed twice, the �rst time with a man as owner and the second with a woman. Figure 3 shows the front and back of the brochure along with the picture of the businesses �rst with men (Male brochure) and then with women (Female Brochure). The businesses in the brochure were chosen to be representative of the type of businesses typically run by NRSP microentrepreneurs. The brochure thus contained two agribusinesses, two retail businesses and one tailoring business. According to our baseline data, 49.41 percent of male businesses were agribusiness, 26.87 percent were in retail and 8.53 percent were involved in handicrafts and tailoring, thus accounting for almost 85 percent of all male businesses. Among female businesses, 19.88 percent were agribusinesses, 17.60 percent were retail businesses and 56.90 were in handicrafts and tailoring, accounting for almost 95 percent of all female businesses. All members of a CO were given one of the two brochures, which were randomly allocated across COs. The goal of the brochure was to explain how to apply for a larger loan via a lottery. Appendix A provides the translated text of the brochure. According to this, all eligible members could make a loan request of up to Rs. 100,000. The request was subject to all the usual technical and social reviews conducted by 8. There were 12 teams of two NRSP staff each in Attock, 29 in Bahawalpur and 7 in Hyderabad. 9. First time borrowers were not eligible to participate in the lottery. NRSP felt it did not have suf�cient credit history for this group to allow them access to the much larger loans available to lottery winners. ´ , Mansuri, and Pico Gine ´n 515 F I G U R E 3. Brochures NRSP credit of�cers, who could also determine the loan amount they were willing to approve for each borrower. Approved loans which were larger than the usual limit of Rs. 30,000, were to be forwarded to headquarters, where the result of the lottery were maintained.10 Lottery winners could borrow the approved amount, while those who lost the lottery could borrow up to their regular loan size. Although the brochure encouraged members to borrow for productive purposes, in practice there were no restrictions on the use of the loan. In addition, qualifying members who already had an outstanding loan with NRSP were allowed to apply for the larger loan, subject to the condition that part of the new loan would be used to pay off the outstanding debt. Eligible CO members had seven months, from November 2007 to June 2008, to apply for the larger loan. Of the 2,284 eligible CO members, 713 (31.2 percent) applied. NRSP approved 532 loans (74.6 percent). Most appli- cants had their loan amounts reduced. Credit of�cers reported that this was due to concerns that borrowers would not be able to make the required monthly installments. Of the customers approved, 254 were assigned to win the lottery (47.7 percent) and 211 ended up borrowing (83 percent). Among the 278 loan applicants that lost the lottery, only 161 borrowed (58 percent). Among the reasons cited for changing their mind are time elapsed from request ´ and Mansuri 10. The lottery was designed so that the chance of winning was 50 percent. See Gine 2011 for more details. 516 THE WORLD BANK ECONOMIC REVIEW F I G U R E 4. Distribution of Loans by Date of Disbursement (Source: Administrative records, NRSP) to approval (average time was 2 months), and for losers the fact that the new loan size was not too different than the loan they currently had. Figure 4 shows the distribution of loans by disbursement date and gender. The vertical axis measures number of loans applied for in each disbursement month. It is clear that women did the bulk of the borrowing in the �rst three to four months after the lottery started. Males, on the other hand, waited for the months of May and June to ask for loans, coinciding with the agricultural season. A follow-up survey was conducted in December 2008. This was six months after the loan lottery concluded and about 13 months after the loan orientation meetings. In the follow up, some 45 percent of eligible CO members recalled attending the loan lottery orientation meeting. Among those who recalled attendance, about 70 percent recalled receiving a bro- chure and of these about half were able to correctly recall the picture they were shown. Of the 211 lottery winners who took the larger loan, 125 reported loan use in the follow up. Table 1 reports the share of loans used for different purposes by gender, along with the p-value of a test of differences by gender. On average, about 48 percent of the loan was used for working capital, with men signi�cantly more likely to use loans for this purpose (50 percent as compared to 36 percent for women). In contrast only 6 percent of loan proceeds were used for purchasing business equipment, but here women were 3 times more likely to report this use (12 percent compared to about 4 percent for men). For other uses, men and women look roughly similar, though women use loans ´ , Mansuri, and Pico Gine ´n 517 T A B L E 1 . Reported Larger Loan Use All Male Female P-val of t-test (2)-(3) (1) (2) (3) (4) Household Durables 0.038 0.027 0.080 0.161 Food consumption 0.041 0.041 0.041 0.999 School Supplies 0.003 0.003 0.005 0.584 Festivals and Ceremonies 0.010 0.007 0.022 0.250 Household Repairs 0.070 0.081 0.027 0.293 Previous Loan Repayment 0.059 0.057 0.068 0.787 Savings 0.041 0.036 0.063 0.438 Inventories/ raw materials for 0.475 0.505 0.355 0.110 main business Equipment for main business 0.055 0.038 0.124 0.064 Inventories and Equipment for 0.061 0.053 0.096 0.361 other household businesses Other uses 0.146 0.153 0.118 0.606 Number of Obs. 125 100 25 Note: Authors’ analysis based on data from the follow-up survey conducted in December 2008. Columns report the average fraction of loans used for each purpose. more frequently for consumer durables (8 percent of loan proceeds on average versus about 3 percent for men), while men use loan proceeds more frequently for housing improvements (8 percent versus about 3 percent for women). I I . D ATA Baseline data collected in November 2006, prior to the business training and loan lottery orientations, included questions on the CO member, the member’s household and the business. Besides the usual set of variables, such as age, edu- cation, marital status etc, individual characteristics include measures of entre- preneurship, digit span recall, risk preferences and decision making autonomy across a range of household and business outcomes. Household characteristics include information on the income generating activities of all household members, total household assets including livestock and past and current bor- rowing and saving of household members. Business characteristics, including age, location and type of business activity, as well as the scale of the business as measured by its assets, hired workers and monthly sales. Summary statistics from the baseline survey are presented in Table 2, and variable de�nitions are provided in Appendix B. CO members are about 38 years old, have 4.1 years of education, own some 3.5 acres of land and have average household expenditures of about Rs 5,200 per month (roughly 61 USD). This places them signi�cantly above the bottom half of the popu- lation of the villages in which they reside (see Mansuri, 2011). Women are T A B L E 2 . Summary Statistics 518 Matched Business Business P-value of All members P-value of Owners P-value of Owners t-test t-test (4)-(5) t-test (7)-(8) (10)-(11) N. Obs All Male Female Male Female Male Female (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) Take-up Shown Female brochure (Yes ¼ 1) 3,451 0.52 0.49 0.56 0.00 0.47 0.57 0.02 0.45 0.58 0.02 Eligible for loan lotttery (Yes ¼ 1) 3,451 0.55 0.62 0.62 0.00 0.69 0.67 0.00 0.71 0.68 0.01 Offered Business Training (Yes ¼ 1) 3,451 0.54 0.52 0.53 0.00 0.55 0.53 0.00 0.54 0.56 0.04 Applied for loan (1 ¼ Yes) 2,149 0.31 0.42 0.18 0.00 0.45 0.17 0.00 0.44 0.22 0.00 Approved, conditional on applying (1 ¼ Yes) 664 0.76 0.76 0.76 0.01 0.78 0.75 0.05 0.81 0.74 0.35 Borrowed, conditional on being approved 503 0.69 0.75 0.52 0.62 0.75 0.55 0.85 0.78 0.63 0.87 (1 ¼ Yes) Amount borrowed (‘000 Rs) 445 33.00 33.83 30.00 0.06 34.59 29.55 0.02 36.04 32.50 0.43 THE WORLD BANK ECONOMIC REVIEW Baseline Characteristics Individual Characteristics Female (Yes ¼ 1) 3,451 0.46 2 2 2 2 2 2 2 2 2 Age 3,451 37.88 38.18 37.51 0.01 37.5 36.97 0.29 38.49 39.02 0.98 Years of Education (0-16) 3,451 4.09 5.31 2.63 0.00 5.53 2.73 0.00 5.29 2.65 0.00 Married (Yes ¼ 1) 3,451 0.83 0.82 0.84 0.12 0.82 0.86 0.13 0.84 0.87 0.09 Digit Span Recall (0-8) 3,451 3.31 3.84 2.68 0.00 4.03 2.74 0.00 3.87 2.48 0.00 Member of a Mixed Group (Yes ¼ 1) 3,451 0.06 0.04 0.09 0.00 0.05 0.09 0.00 0.03 0.13 0.00 Index of Female Mobility 1,571 0.06 2 0.06 2 2 0.12 2 2 0.15 2 Index of No Purdah 1,571 0.15 2 0.15 2 2 0.13 2 2 0.37 2 Business Owner (Yes ¼ 1) 3,451 0.60 0.61 0.58 0.01 2 2 2 2 2 2 Risk Tolerance (0 ¼ Risk Averse; 10 ¼ Risk 3,451 3.61 3.81 3.37 0.00 3.84 3.39 0.00 4.08 3.59 0.00 Lover) Months as Member 3,451 26.55 24.57 24.56 0.00 26.92 22.11 0.00 28.38 25.35 0.01 Household Characteristics Years of Education, Spouse (0-16) 3,451 3.31 2.12 4.75 0.00 2.24 4.66 0.00 2.06 4.53 0.00 (Continued ) TABLE 2. Continued Matched Business Business P-value of All members P-value of Owners P-value of Owners t-test t-test (4)-(5) t-test (7)-(8) (10)-(11) N. Obs All Male Female Male Female Male Female (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) Total Household Income (log) 3,451 11.53 11.62 11.44 0.00 11.71 11.45 0.00 11.53 11.61 0.99 Expenditures (log) 3,451 8.28 8.29 8.27 0.00 8.32 8.29 0.00 8.19 8.39 0.40 Number of Children under 9 3,451 2.65 2.89 2.37 0.00 2.89 2.43 0.00 2.77 2.31 0.01 Land (area) 3,451 4.48 5.94 2.74 0.00 5.84 2.57 0.01 4.99 3.48 0.01 Credit constraints (Yes ¼ 1) 3,451 0.14 0.12 0.16 0.00 0.12 0.12 0.00 0.12 0.13 0.01 Family ever in business (Yes ¼ 1) 3,451 0.61 0.61 0.61 0.31 0.75 0.7 0.06 0.74 0.69 0.14 Decision Making (0-8) 3,451 2.65 3.34 1.84 0.00 3.35 1.59 0.00 3.20 1.79 0.00 Sources of Credit % borrowing Formal Sector 2006-08 2,931 0.05 0.07 0.03 0.01 0.08 0.03 0.02 0.09 0.01 0.01 Amount borrowed 2006 (‘000s) 2,931 5.94 9.67 1.37 0.08 11.57 0.85 0.19 13.35 0.15 0.31 % borrowing Micro�nance Institutions / 2,931 0.82 0.79 0.86 0.00 0.86 0.93 0.00 0.87 0.91 0.00 Micro�nance Banks 2006-08 Amount borrowed 2006 (‘000s) 2,931 31.17 34.23 27.42 0.00 37.67 28.79 0.00 39.62 31.54 0.13 % borrowing Friends and Family (other than CO 2,931 0.22 0.21 0.23 0.11 0.2 0.22 0.30 0.19 0.22 0.47 members) 2006-08 Amount borrowed 2006 (‘000s) 2,931 5.57 6.19 4.81 0.00 6.19 4.78 0.01 6.04 4.61 0.13 % borrowing Informal Lenders 2006-08 2,931 0.48 0.50 0.47 0.16 0.47 0.42 0.27 0.49 0.45 0.20 Gine Debt to informal lenders 2006 (‘000s) 2,931 15.07 23.21 5.06 0.00 26.45 4.32 0.00 18.38 4.73 0.02 Businesses Agribusiness, Dairy, Livestock (Yes ¼ 1) 1,962 0.40 0.54 0.22 0.00 0.54 0.22 0.00 0.56 0.56 2 Retail and Food Services (shopkeeping) (Yes ¼ 1) 1,962 0.24 0.28 0.19 0.00 0.28 0.19 0.00 0.29 0.29 2 Handicraft, Tailoring, Vocational Trade 1,962 0.29 0.09 0.55 0.00 0.08 0.55 0.00 0.05 0.05 2 (Yes ¼ 1) Other (Yes ¼ 1) 1,962 0.07 0.09 0.04 0.00 0.09 0.03 0.00 0.09 0.09 2 ´ , Mansuri, and Pico Sales (‘000 Rs) 2,065 14.51 22.3 4.74 0.01 22.3 4.74 0.00 7.06 7.01 0.48 ´n Note: All variables are from the baseline survey (November 2006), except for Sources of Credit information, which is from the followup and recalls credit for the 2006-2008 period. See Appendix B for de�nition of variables. Columns (2)-(4), (6)-(7) and (9)-(10) report means. 519 520 THE WORLD BANK ECONOMIC REVIEW only slightly less likely than men to report owning a business (61 percent of sample men and 58 percent of sample women owned a business at baseline).11 However, there are signi�cant differences between male and female CO members in almost all other dimensions, and these differences are sustained when we focus only on those who owned a business at baseline. Women tend to have less education, perform signi�cantly below men on digit span recall and are less risk tolerant (on a 0 to 10 scale). Women members are also more likely to come from households that have less land wealth, as compared to the households of male CO members. This indicates some selection of women CO members by wealth and is consistent with more stringent female seclusion prac- tices among landed rural households (see Jacoby and Mansuri, 2011). Importantly, women members also report far less decision making autonomy than men do on a range of household, individual and business outcomes. Among business owners, women report having complete autonomy over roughly 1.5 decisions out of a total of 8, whereas men claim to have complete autonomy over more than 3 decisions.12 Women are also more likely than men to belong to a mixed gender CO in our sample (about 63 percent of members in mixed COs are women). While this is a small sample overall, given that only 7 percent of COs have both men and women, there are some interesting differences between women in mixed and single gender COs. In particular, the former seem to have less ‘voice’ in the COs they belong to. While the odds of holding of�ce in the single gender COs are about the same for men and women, at 20 and 18 percent respect- ively, 48 percent of men and only 7 percent of women in mixed COs report holding any of�ce. Women in mixed COs are also signi�cantly more likely to observe purdah, even though mixed COs are more than twice as likely to be composed of members of the same zaat (caste).13 They are also younger and have more young children under age 9. They are also less risk tolerant, signi�- cantly less optimistic and have less trust in the formal institutions. However, they do about as well as other women on digit span recall and the index of entrepreneurial literacy. They are also about as well educated as women in female only COs and come from households that are, on average, wealthier than those of other women in the sample (Appendix Table A1). Bothe men and women in mixed COs are also far less likely to be eligible for the larger loan, and this is even more the case for men, perhaps because they have been CO members for a much shorter time, on average. Entrepreneurs are also different from other CO members. They are more likely to be older, more educated and to report a family history of business ownership. In addition, business owners have better digit span recall, have a 11. This is not surprising given that the men and women in our sample are all micro�nance clients. 12. Gine´ and Mansuri (2011a) also �nd that women members are often not the effective managers of businesses they claim to own. 13. Seclusion practices are much less stringent within zaat/biradari (caste) groups. See Jacoby and Mansuri (2011). ´ , Mansuri, and Pico Gine ´n 521 T A B L E 3 . Percentage Borrowing and Reasons for not Borrowing by Credit Source Credit Source Commercial Informal Relatives and Bank MFI Lenders Friends Percent borrowing from [source] in 5.07 83.43 33.77 21.12 2006 The main [reason] for not borrowing from [source]? Do not like/need to borrow 52.55 81.39 70.81 74.69 Inadequate collateral 18.50 1.98 11.94 6.77 Lender’s procedures are too 14.49 9.31 6.70 5.82 cumbersome Lender’s loan terms are 5.48 0.59 3.35 0.91 unfavorable Lender is too far away 3.18 0.40 0.78 1.35 Need to pay bribes 2.02 0.79 2.18 0.08 Past default with lender 1.96 0.79 0.89 1.15 Members not willing to lend to me 1.25 4.55 2.51 6.53 Bad credit history 0.56 0.20 0.84 2.69 Notes: Data come from the baseline survey (November 2006) better outlook on life and not surprisingly, also score higher on a business knowledge test (see Gine ´ and Mansuri, 2011a for details). They are also more likely to be risk tolerant. Among women, those that have a business are less likely to report mobility constraints, but are somewhat more likely to observe purdah, perhaps because they are relatively wealthier. Table 2 also suggests that the scale at which male and female businesses operate is quite different (see also de Mel et al. 2009). Men’s businesses yield monthly sales that are more than 4 times as large as monthly sales for women. Women also tend to operate more businesses from home. While there is con- siderable overlap in the type of businesses owned by men and women, men are far more likely to be in the agribusiness sector, with much greater contact with wider markets, while women are concentrated in small scale manufacturing, handicrafts and tailoring in particular. Table 3 shows the main sources of credit for sample households at baseline. Only 5 percent of members have any loans from formal �nancial sources, mainly commercial banks. In contrast, 34 percent report borrowing from infor- mal lenders, mainly shopkeepers, and 21 percent report borrowing from rela- tives.14 Not surprisingly, most borrow from an MFI, including NRSP. Average loan sizes vary greatly by source. While the average loan from a commercial bank is around 100,250 Rs (1,172 USD), the average MFI loan is only 12,000 14. Informal lenders also include traders and wholesalers and, to a smaller extent, professional moneylenders and landlords. 522 THE WORLD BANK ECONOMIC REVIEW Rs (USD 140) and loans from informal lenders are typically in about the same range. There is little variation in the relative share of lenders by gender, however, though female CO members tend to borrow less overall. Table 3 also reports reasons for not borrowing for those who reported not using a credit source. Interestingly, most respondents without loans report an unwillingness to borrow – either because they have no need for loans or because they dislike borrowing – as the main reason for not taking a loan and this varies little across lenders. Lack of collateral and cumbersome loan appli- cation procedures come in next, and are particularly important when dealing with a formal lender. Table 4 presents the means of baseline variables for the sample. Columns 2, 3 and 4 report the means for CO members who were exposed to the male and female brochure, respectively and the p-value of the test that the difference in means is signi�cant. We also report the means and p-value of the same test for the sample of males (columns 5 to 8) and females (columns 9-12) separately. Overall, we �nd balance between the two groups. The difference in means for members receiving the male and female brochure is signi�cant at conventional levels for only two of the 16 baseline variables we consider. Study participants who received the male brochure borrow more from informal sources and are less likely to be members of a mixed group (both signi�cant at the 10% level). For the sample of males, the difference in means in the two groups is again sig- ni�cant for only two of the 16 variables considered. Male participants who received the male brochure are somewhat less likely to belong to a mixed CO (signi�cant at the 10% level)15 and are somewhat less likely to borrow from friends and relatives (signi�cant at the 5% level). For the sample of females the difference is only signi�cant for own education. Women who received the male brochure appear to have higher formal education by about one third of a year. Table 3 also reports the p-value of an F-test that all baseline characteristics are jointly insigni�cant. We cannot reject this hypothesis in any of the three samples (lowest p-value is 0.58). III. EMPIRICAL STRATEGY Because the type of brochure is assigned randomly at the CO level, its impact can be estimated via the following regression equation: Yij ¼ a þ bFB j þ gXij þ eij ð1Þ where Yij is the outcome of interest for individual i in CO j (for e.g., and indi- cator variable that takes the value 1 if the respondent had applied for a larger loan, and 0 otherwise), FBj is an indicator variable that take the value 1 if the 15. Since brochure-type was assigned at the CO level, the lower representation of men in mixed COs likely explains this. T A B L E 4 . Veri�cation of Randomization All members Male Female P-val of P-val of P-val of Means t-test Means t-test Means t-test (1)-(2) (4)-(5) (7)-(8) N. Male Female N. Male Female N. Male Female Obs. Brochure Brochure Obs. Brochure Brochure Obs. Brochure Brochure (1) (1) (2) (3) (5) (4) (5) (6) (9) (7) (8) (9) Characteristics at Baseline Age 3,451 37.95 37.81 0.74 1,880 37.98 38.40 0.89 1,571 37.91 37.20 0.85 Number of children 3,451 1.73 1.83 0.77 1,880 1.92 1.99 0.53 1,571 1.45 1.40 0.17 under 9 Business 3,451 0.60 0.59 0.57 1,880 0.63 0.59 0.75 1,571 0.57 0.60 0.79 Owner(Yes ¼ 1) Digit Span Recall 3,451 3.36 3.28 0.85 1,880 3.82 3.88 0.83 1,571 2.72 2.65 0.82 Risk Tolerance 3,451 3.71 3.52 0.25 1,880 3.81 3.81 0.53 1,571 3.57 3.22 0.49 (0 ¼ Risk Averse; 10 ¼ Risk Lover) Land 3,451 4.38 4.58 0.82 1,880 5.83 6.05 0.77 1,571 2.36 3.05 0.55 Years of Education of 3,451 3.69 3.71 0.69 1,880 2.72 2.73 0.86 1,571 5.05 4.75 0.55 Gine Spouse Member of a Mixed 3,451 0.04 0.08 0.10 1,880 0.02 0.06 0.06 1,571 0.07 0.10 0.89 Group (Yes ¼ 1) Own Education 3,451 4.26 3.95 0.14 1,880 5.27 5.36 0.78 1,571 2.84 2.47 0.07 Decision Making 3,451 2.74 2.57 0.35 1,880 3.41 3.26 0.77 1,571 1.82 1.85 0.61 N. Obs. 1,662 1,789 967 913 695 876 Pct. Borrowing ´ , Mansuri, and Pico from . . . at the time of ´n Baseline* (Continued ) 523 524 TABLE 4. Continued All members Male Female P-val of P-val of P-val of Means t-test Means t-test Means t-test (1)-(2) (4)-(5) (7)-(8) N. Male Female N. Male Female N. Male Female Obs. Brochure Brochure Obs. Brochure Brochure Obs. Brochure Brochure (1) (1) (2) (3) (5) (4) (5) (6) (9) (7) (8) (9) Commercial Bank 2,931 2.38 1.7 0.64 1,616 3.69 2.71 0.60 1,315 0.51 0.69 0.74 Micro�nance 2,931 71.34 67.62 0.30 1,616 69.44 63.74 0.98 1,315 74.06 71.74 0.46 THE WORLD BANK ECONOMIC REVIEW Institution Friends and Relatives 2,931 7.71 6.58 0.48 1,616 7.85 5.16 0.05 1,315 7.5 8.09 0.14 Informal lenders 2,931 0.7 0.2 0.07 1,616 0.83 0.26 0.19 1,315 0.51 0.13 0.23 N. Obs. 1,427 1,504 841 775 586 729 Pct. Offered Business 3,451 52.47 51.82 0.83 1,880 51.29 52.03 0.93 1,571 54.1 51.6 0.48 Training Member is eligible for 3,451 63.72 69.00 0.54 1,880 64.63 59.69 0.82 1,571 62.45 62.21 0.66 loan lottery (Yes 5 1) P-val of F-test that all 0.65 0.58 0.65 baseline characteristics are jointly insigni�cant Notes: * denotes variable measured at follow-up, conducted in December 2008. Other variables are from baseline conducted in november 2006. Pct. Offered Business Training and Member eligibility come from administrative data from NRSP. ´ , Mansuri, and Pico Gine ´n 525 respondent was shown the female brochure, Xij is a vector of individual charac- teristics collected at baseline and e ij is a mean-zero error term. Standard errors are clustered at the CO level throughout since the CO level treatment assign- ment creates spatial correlation among members of the same CO (Moulton, 1986). The vector Xij includes the following baseline characteristics reported in Table 2: eligibility, being offered business training, a dummy if decision- making power is above the median in the same gender sample, own education, digit span recall, spouse education, landholdings, membership in a mixed group, age, number of children and risk tolerance. It also includes �eld unit (branch) dummies, our strati�cation variable. This set of variables includes characteristics for which there is imbalance as well as variables that are likely to affect the decision to borrow. 16 The coef�cient b captures the impact of being shown a brochure with pictures of female entrepreneurs on the cover and is the coef�cient of interest. We also examine interactions between the type of brochure shown and base- line characteristics which could proxy for attitudes towards women Yij ¼ a þ rðFBà j Zij Þ þ bFB j þ gXij þ eij ð2Þ Zij is a subset of individual characteristics included in the vector Xij that could represent an individual’s attitudes towards women. The coef�cient r on the interaction term FBj * Zij reveals the extent to which the impact of the female picture (FB) on loan uptake varies according to the members attitudes towards women. I V. E M P I R I C A L R E S U L T S This section presents evidence on the impact of the female brochure on two main outcomes: the decision to borrow and the loan amount requested. Table 5 presents regression results from the estimation of equation (1) for the decision to borrow. Columns (1)-(3) present results for the full sample, combined and disaggregated by gender. On average, only 14.6 percent of members, roughly 31 percent of the eligible, applied for a larger loan. This is somewhat higher for men, at 20 percent, and correspondingly lower for women, among whom only 8 percent of the eligible applied. Given that the sample consists of seasoned micro�nance clients, this number appears low. When asked anecdotally, many borrowers report either that the monthly installment amount for the larger loan was too high or that the maturity period was too short. Both of these indicate that the clients of NRSP are not, by and large, constrained by the current loan size, but that some could bene�t from larger loans. Among men, older members and business owners were more 16. The regression results to come are not substantially affected by the exclusion of this set of control variables. T A B L E 5 . Uptake of Larger Loan 526 OLS All members All Business Owners Matched Business Owners All Males Females All Males Females All Males Females Sample: (1) (2) (3) (4) (5) (6) (7) (8) (9) Female Brochure 2 0.026 2 0.033 2 0.015 2 0.029 2 0.017 2 0.051** 2 0.029 2 0.007 2 0.043 (0.021) (0.034) (0.019) (0.027) (0.044) (0.023) (0.035) (0.058) (0.039) Offered Business 0.048** 0.046 0.046** 0.058** 0.067 0.037 0.071** 0.056 0.095** Training (0.021) (0.035) (0.019) (0.027) (0.044) (0.024) (0.035) (0.055) (0.047) Female 2 0.016 2 0.034 2 0.029 (0.021) (0.031) (0.046) Age 0.006*** 0.008** 0.000 0.006 0.009 2 0.003 0.009 0.011 2 0.003 THE WORLD BANK ECONOMIC REVIEW (0.002) (0.004) (0.004) (0.004) (0.006) (0.006) (0.006) (0.008) (0.011) Age ^ 2 2 0.000** 2 0.000** 0.000 0.000 0.000 0.000 2 0.000* 0.000 0.000 (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) High 0.001 2 0.003 2 0.022 0.002 0.009 2 0.016 2 0.002 2 0.037 2 0.039 Decision-making power (1 ¼ yes) (0.002) (0.020) (0.017) (0.003) (0.025) (0.022) (0.005) (0.051) (0.036) Own Education 0.002 0.005 0.000 0.003 0.004 0.001 0.004 0.004 0.007 (0.002) (0.003) (0.003) (0.003) (0.004) (0.004) (0.005) (0.007) (0.006) Digit Span Recall 0.003 0.002 0.001 2 0.001 2 0.002 2 0.002 0.002 2 0.001 2 0.001 (0.003) (0.005) (0.004) (0.005) (0.008) (0.005) (0.008) (0.013) (0.011) Risk Tolerance 0.001 2 0.001 0.002 0.001 2 0.001 0.001 0.004 2 0.002 0.006 (0.002) (0.003) (0.003) (0.003) (0.004) (0.003) (0.005) (0.008) (0.006) Yrs. Of Education 0.000 2 0.005 0.002 2 0.001 2 0.003 0.000 2 0.004 2 0.011* 2 0.001 of Spouse (0.002) (0.003) (0.002) (0.002) (0.004) (0.003) (0.003) (0.006) (0.004) (Continued ) TABLE 5. Continued OLS All members All Business Owners Matched Business Owners All Males Females All Males Females All Males Females Sample: (1) (2) (3) (4) (5) (6) (7) (8) (9) Number of 0.003 0.003 0 0.007 0.007 0.006 0.01 0.017 0.001 children under 9 (0.003) (0.004) (0.004) (0.004) (0.006) (0.007) (0.008) (0.011) (0.011) Business Owner 0.028** 0.040* 0.01 (Yes ¼ 1) (0.014) (0.022) (0.018) Land 0.000 0.001 0.000 0.000 0.001 0.000 -0.001 0.000 -0.002 (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.001) (0.001) (0.002) Member of a 2 0.092*** 2 0.086* 2 0.105** 2 0.150*** 2 0.121* 2 0.196*** 2 0.166** 0.021 2 0.272*** Mixed Group (Yes ¼ 1) (0.032) (0.047) (0.048) (0.043) (0.066) (0.070) (0.073) (0.236) (0.100) Eligibility 0.188*** 0.246*** 0.130*** 0.203*** 0.299*** 0.111*** 0.219*** 0.304*** 0.148*** (0.015) (0.026) (0.016) (0.019) (0.033) (0.018) (0.024) (0.040) (0.032) Gine Mean Dependent 0.15 0.20 0.08 0.15 0.20 0.08 0.18 0.25 0.11 Variable N. Observations 3,451 1,880 1,571 2,065 1,149 916 726 363 363 R-squared 0.17 0.18 0.11 0.18 0.17 0.13 0.17 0.17 0.16 Note: Authors’ analysis based on survey and administrative records data. The dependent variable takes value 1 if the member applied for a larger loan. ´ , Mansuri, and Pico ´n All regressions include branch �xed effects and are estimated using OLS methods. Standard errors are clustered at the borrower group level.The following symbols *, * * and *** denote signi�cance at 10, 5 and 1 percent level, respectively. See Appendix B for variable de�nition. 527 528 THE WORLD BANK ECONOMIC REVIEW likely to apply, while among women, being offered business training increases loan demand. Membership in a mixed CO dampens loan demand for both men and women even after controlling for eligibility, perhaps because group dynamics provide disincentives to borrow, since only half of the members of mixed COs are eligible to borrow (see Appendix Table A1). Exposure to the female brochure does not seem to have any effect, however. In columns (4)-(6) we focus on all CO members who had a business at base- line. We �nd, perhaps surprisingly, that the female brochure impacts negatively the uptake by women entrepreneurs. The effect is not small. From a base of 8.3 percent, the point estimate indicates a reduction of 39 percent the prob- ability of applying for the larger loan. In contrast, businessmen appear indiffer- ent to the picture in the brochure and, as before, being offered business training increases loan demand while membership in a mixed CO depresses it, but there are no differences by gender in these or any other characteristics. Taken face value, the result among women appears counterintuitive: if any- thing, one might expect that a woman who owns a business would be attracted to a loan product that makes salient her identity as a business woman. But since women operate businesses that are much smaller than those of men, it is possible that the negative impact is an artifact of the business scale. Speci�cally, women may �nd the scale of the business pictured in the brochure too large and they may be discouraged from applying.17 To explore this hypothesis further, we restrict attention to female business owners who operate at a scale comparable with men.18 Speci�cally, we construct a sample of matched male and female businesses by sector.19 Given differences in business scale by gender, this matched sample necessarily consists of businesses in the upper (lower) tail of the distribution of female (male) businesses. In particular, female businesses in the matched sample have on average Rs 4,016 higher sales compared to the sample of all female businesses, while male businesses in the matched sample have Rs 8,129 lower sales, on average. Sample size is also considerably reduced. Only 363 female businesses can be matched with a corresponding male business. Columns (7)–(9) report these results. The coef�- cient for female brochure remains negative but we lose precision. The coef�cient falls by about 16 percent while the standard error increases by 17. We have some anecdotal evidence that businesswomen felt that some of the businesses featured in the brochure were run on a larger scale than the typical female business and that this fact discouraged them from borrowing. As mentioned, many women operated each of the three business types featured in the brochure, which together account for 95 percent of all businesses run by women. In addition, the brochure clearly stated that the loan could be used for any purpose and for any business, but nonetheless, the picture may have triggered a more deliberative response (Kahneman, 2003). 18. We are grateful to an anonymous referee for suggesting this analysis. 19. The matched sample is generated as follows. For each female business, we compute the absolute difference in sales with each male business in the same sector and then pick the business with the smallest difference. A male business is matched only once with a female business. In the �nal sample we keep only those female businesses where the absolute difference in sales is less than Rs 1000. ´ , Mansuri, and Pico Gine ´n 529 about 70 percent. In sum, we cannot conclusively rule out a negative effect of exposure to the female brochure even among women business owners who run larger businesses. Note, however, that in this matched group, being offered business training increases the probability of a woman applying for a larger loan by 86 percent from a base of 11 percent and appears to be driving the overall increase in loan applications due to the offer of business training. On the other hand, decision making power does not appear to encourage loan applications from women in any of these samples. Table 6 reports the regression results from speci�cation (2), for males (columns 1-3) and females (columns 4-6) on all three samples. All regressions include the baseline controls used in Table 5. The �rst striking fact is that a subset of men does respond to the psychological content of the brochure. Speci�cally, business owners with low scores on the digit span recall question, a proxy for ability, and those whose wives are poorly educated ( p-value of FB x Years Education of Spouse is 0.11), respond negatively to the female bro- chure. In this group, loan demand falls by about 13 percentage points (more than a 50% decline over the base demand). While our experimental design does not allow us to distinguish between lack of af�nity towards women and an outright distaste for female-run businesses as possible explanations for the negative response, it does suggest one channel through which exposure to the female brochure among female business owners could depress loan demand. Speci�cally, women who have low decision making power may turn to their husbands for per- mission to borrow and may face a negative response from them when shown the female brochure. We do �nd some evidence in support of this. Females in mixed groups, which as mentioned before appear to have less ‘voice’, as well as business women with low decision making power in their household react negatively to the female brochure. Interestingly, female business owners with high decision making autonomy shown the male brochure also react negatively by roughly the same magnitude, while there is no effect on female business owners with autonomy shown the female brochure ( p-value is 0.28).20 This suggests that women with decision-making autonomy react negatively to the brochure of opposite sex by about as much as men without business, with low digit span and poorly educated wives. Given the disadvantaged position of women in rural Pakistan, we conjecture that men may have less respect for female businesses whereas women may feel more alienated when shown the male brochure. 20. More formally we test whether the coef�cient on the female brochure plus the coef�cient on decision-making power above median plus the coef�cient on the interaction between the female brochure and decision-making power above median is different from zero, that is b þ g þ r ¼ 0 following the notation of Equation (2) . 530 T A B L E 6 . Heterogeneous Effects of Larger Loan Uptake OLS Male Female All All Business Matched Business All All Business Matched Business Members Owners Owners Members Owners Owners (1) (2) (3) (4) (5) (6) Female Brochure (FB) 2 0.129** 2 0.131* 2 0.157 0.001 2 0.062 2 0.053 (0.051) (0.072) (0.114) (0.034) (0.038) (0.052) Business Training 0.064 0.101* 0.072 0.029 0.018 0.044 (0.045) (0.054) (0.069) (0.025) (0.032) (0.050) FB x Business Training 2 0.045 2 0.065 0.007 0.039 0.035 0.082 (0.067) (0.082) (0.110) (0.036) (0.045) (0.074) High Decision 2 making power 2 0.016 2 0.008 2 0.038 2 0.021 2 0.051* 2 0.095* THE WORLD BANK ECONOMIC REVIEW (1 ¼ yes) (0.025) (0.031) (0.068) (0.024) (0.029) (0.054) FBx High Decision making power 0.007 0.007 0.006 0.024 0.079** 0.137** (0.005) (0.007) (0.014) (0.030) (0.039) (0.068) Digit Span Recall 2 0.005 2 0.014 2 0.016 0.010* 0.009 0.017 (0.006) (0.010) (0.017) (0.005) (0.007) (0.013) FB x Digit Span Recall 0.016** 0.023* 0.023 2 0.01 2 0.015 2 0.023 (0.008) (0.012) (0.021) (0.006) (0.009) (0.016) Yrs. Education of Spouse 2 0.039 2 0.065* 2 0.111 2 0.011 2 0.012 0.031 (0.028) (0.038) (0.069) (0.025) (0.033) (0.065) FB x Yrs. Education of Spouse 0.038 0.09 0.124 0.044 0.043 2 0.016 (0.040) (0.055) (0.101) (0.032) (0.039) (0.072) Member of a Mixed Group 2 0.100* 2 0.113 0.058 2 0.047 2 0.125*** 2 0.152*** (Yes ¼ 1) (0.053) (0.069) (0.242) (0.034) (0.039) (0.051) FB x Mixed Group 0.053 0.009 2 0.087 2 0.135** 2 0.168** 2 0.242*** (Continued ) TABLE 6. Continued OLS Male Female All All Business Matched Business All All Business Matched Business Members Owners Owners Members Owners Owners (1) (2) (3) (4) (5) (6) (0.073) (0.091) (0.215) (0.055) (0.067) (0.079) Business owner (Yes ¼ 1) 0.031 0.046** (0.030) (0.022) FB x Business owner 0.031 2 0.059* (0.042) (0.031) Mean Dependent Variable 0.20 0.24 0.25 0.08 0.08 0.11 N. Obs. 1880 1149 363 1,571 916 363 R 2 squared 0.19 0.19 0.19 0.12 0.13 0.19 Gine Note: The dependent variable takes value 1 if the member applied for a larger loan. All regressions are estimated using OLS methods and control for eligibility, own education, landholdings, membership in a mixed group, age, number of children, risk tolerance, as well as branch �xed effects. Standard errors are clustered at the borrower group level.The following symbols *, * * and ** * denote signi�cance at the 10, 5, and 1 percent level, respectively. See Appendix B for de�nition of variables. 531´ , Mansuri, and Pico ´n 532 THE WORLD BANK ECONOMIC REVIEW In the matched sample (column 6), the results for female business owners with high decision making autonomy become stronger (the point estimate increases in absolute value). Table 7A reports the impact of the female brochure and baseline character- istics on the loan amount requested (columns 1-4) and approved (columns 5-8) among loan applicants. Table 7B reports the same regressions among loan applicants with a business at baseline. In both tables we �nd a positive and sig- ni�cant effect of the picture on the loan amount requested among males but not females. The result in both tables is driven primarily by selection because men without a business and low digit span recall tend to borrow less and decide not to borrow when shown a female brochure. V. C O N C L U S I O N S We designed a marketing experiment to test whether exposure to positive role models could encourage women’s uptake of a new credit product in a context where women face barriers to participation in economic life. Brochures adver- tising the new product, a much larger loan, were varied such that the cover page featured either men or women as entrepreneurs running �ve otherwise identical businesses. The brochures were randomly assigned to existing clients in good standing. The results suggest that both men and women respond to psychological cues embedded in this type of social norms marketing. However, men’s response is conditioned by relative economic status and ability while women’s response is conditioned by relative status within the household. In particular, men who are not themselves business owners, have lower measured ability and whose wives are less educated respond more negatively to the female brochure, as do women business owners who have low autonomy within the household. Women with relatively high levels of autonomy shown the male brochure have a similar negative response. In contrast, the reaction of high autonomy female business owners shown the female brochure is no different from that of low autonomy women shown the male brochure. We conjecture that loan demand for low autonomy women is mediated through men who respond positively to the male brochure. This suggests that social norms marketing can often be more salient for the more disadvantaged (Paluck and Ball, 2010) but, as our results suggest, this could generate perverse responses in some contexts. Finally, our results also suggest that exposing women to positive role models or information that challenges prevailing norms may meet different levels of success depending on the level of autonomy enjoyed by women. In particular, women with low levels of autonomy may require more intensive interventions, consistent with other work which has used information campaigns to change stereotypes. Gine ´ and Mansuri (2011b), for example, �nd that the response to a voter awareness campaign directed at rural women in Pakistan was most T A B L E 7 A . Loan Size, All Members OLS Log of Amount Requested Log of Amount Approved All Members All Members Males Females Males Females (1) (2) (3) (4) (5) (6) (7) (8) Female Brochure 0.088 0.085* 0.161*** 2 0.035 0.0036 2 0.006 0.008 2 0.098 (0.054) (0.049) (0.052) (0.104) (0.029) (0.030) (0.030) (0.069) Offered Business Training 2 0.041 2 0.032 2 0.092 0.004 2 0.01 0.007 (0.042) (0.046) (0.104) (0.029) (0.031) (0.068) Female 0.078 2 0.054 (0.079) (0.055) Age 0.010 0.015* 2 0.007 0.010** 0.013** 0.002 (0.007) (0.009) (0.018) (0.005) (0.005) (0.017) Age^2 0.000 2 0.000* 0.000 2 0.000* 2 0.000** 0.000 (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) High Decision 2 making power (Yes ¼ 1) 0.015*** 0.087*** 0.024 0.008* 0.021 0.1 (0.005) (0.028) (0.071) (0.005) (0.026) (0.064) Own Education 0.018*** 0.016*** 0.016 0.016*** 0.014*** 0.016^ (0.005) (0.004) (0.010) (0.003) (0.003) (0.010) Digit Span Recall 2 0.012 2 0.018** 0.024 2 0.013** 2 0.012* 2 0.001 Gine (0.009) (0.008) (0.024) (0.006) (0.006) (0.016) Risk Tolerance 0.003 0.001 0.005 2 0.002 2 0.004 2 0.001 (0.004) (0.004) (0.013) (0.004) (0.005) (0.008) Yrs. Of Education of Spouse 0.002 0.003 0.002 0.004 2 0.004 0.018** (0.004) (0.005) (0.010) (0.005) (0.005) (0.008) Number of children under 9 0.004 0.004 0.007 0.004 0.003 2 0.017 (0.009) (0.009) (0.029) (0.006) (0.006) (0.022) ´ , Mansuri, and Pico ´n Business Owner (Yes ¼ 1) 0.023 2 0.006 0.064 0.050 0.058 0.027 (Continued ) 533 534 TABLE 7A. Continued OLS Log of Amount Requested Log of Amount Approved All Members All Members Males Females Males Females (1) (2) (3) (4) (5) (6) (7) (8) (0.032) (0.036) (0.069) (0.035) (0.042) (0.061) THE WORLD BANK ECONOMIC REVIEW Land 0.006*** 0.005*** 0.007 0.001 0.000 0.005** (0.001) (0.001) (0.005) (0.001) (0.001) (0.002) Member of a Mixed Group (Yes ¼ 1) 0.096 0.377** 2 0.233 0.025 2 0.06 0.076 (0.153) (0.170) (0.167) (0.176) (0.176) (0.197) Mean Dependent Variable 10.93 10.93 10.92 10.97 10.67 10.67 10.72 10.52 N. Observations 664 664 491 173 503 503 372 131 R 2 squared 0.21 0.27 0.32 0.33 0.27 0.34 0.26 0.44 Note: The dependent variable is log of loan amount requested (columns 1-4) and log of loan amount approved (columns 5-8). The sample includes loan applicants. All regressions include branch �xed effects and are estimated using OLS methods. Standard errors are clustered at the borrower group level.The following symbols *, * * and ** * denote signi�cance at the 10, 5, and 1 percent level, respectively. See Appendix B for de�nition of variables. T A B L E 7 B . Loan Size, Business Owners OLS Log of Amount Requested Log of Amount Approved Business Owners Business Owners Males Females Males Females (1) (2) (3) (4) (5) (6) (7) (8) Female Brochure 0.093 0.08 0.135** -0.061 0.027 0.011 0.002 -0.068 (0.057) (0.050) (0.052) (0.129) (0.032) (0.033) (0.030) (0.081) Offered Business Training -0.032 -0.058 -0.021 0.054** 0.026 0.032 (0.046) (0.047) (0.145) (0.027) (0.026) (0.098) Female 0.089 -0.094 Gine (0.082) (0.064) Age 0.013 0.017 -0.024 0.007 0.015** -0.006 (0.009) (0.011) (0.021) (0.005) (0.006) (0.023) Age^2 -0.000* -0.000* 0.000 0.000 -0.000** 0.000 (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) High Decision-making power (Yes ¼ 1) 0.025 0.095** -0.022 0.029 -0.007 0.094 (0.038) (0.038) (0.091) (0.030) (0.024) (0.070) ´ , Mansuri, and Pico Own Education 0.022*** 0.020*** 0.027 0.012*** 0.013*** 0.012 ´n (0.005) (0.005) (0.017) (0.004) (0.004) (0.010) (Continued ) 535 TABLE 7B. Continued 536 OLS Log of Amount Requested Log of Amount Approved Business Owners Business Owners Males Females Males Females (1) (2) (3) (4) (5) (6) (7) (8) Digit Span Recall -0.014 -0.019* 0.003 -0.009 -0.009 0.001 (0.011) (0.010) (0.034) (0.006) (0.007) (0.020) Risk Tolerance 0.002 0.001 0.014 -0.001 0.000 -0.003 (0.005) (0.005) (0.017) (0.005) (0.006) (0.011) Yrs. Of Education of Spouse 0.001 0.007 -0.007 0.002 -0.005 0.013 (0.005) (0.005) (0.013) (0.006) (0.006) (0.010) Number of children under 9 0.004 0.008 -0.032 0.001 -0.004 -0.011 THE WORLD BANK ECONOMIC REVIEW (0.011) (0.011) (0.037) (0.007) (0.007) (0.019) Land 0.005*** 0.005*** 0.005 0.002 0.000 0.004 (0.001) (0.002) (0.005) (0.001) (0.001) (0.003) Member of a Mixed Group (Yes ¼ 1) 0.138 0.314 -0.293 -0.007 -0.256*** 0.030 (0.167) (0.213) (0.293) (0.239) (0.086) (0.331) Mean Dependent Variable 10.92 10.92 10.92 10.92 10.69 10.69 10.69 10.69 N. Observations 459 459 356 103 355 355 278 77 R-squared 0.23 0.30 0.34 0.42 0.28 0.35 0.29 0.58 Note: The dependent variable is log of loan amount requested (columns 1-4) and log of loan amount approved (columns 5-8). The sample includes loan applicants. All regressions include branch �xed effects and are estimated using OLS methods. Standard errors are clustered at the borrower group level. The following symbols *, * * and ** * denote signi�cance at the 10, 5, and 1 percent level, respectively. See Appendix B for de�nition of variables. ´ , Mansuri, and Pico Gine ´n 537 effective among women who had voted in the past and hence had overcome some of the barriers to participation in public life. APPENDIX A: BROCHURE SCRIPT NRSP (in collaboration with the World Bank and the Pakistan Poverty Alleviation Fund) has initiated an Enterprise Development Program. As part of this program, the possibility of obtaining a larger loan is now available. T A B L E A 1 . Baseline Member Characteristics of Mixed vs single gender COs Male Female P-val of P-val of Means t-test Means t-test (1)-(2) (4)-(5) Mixed Single Mixed Single Gender gender Gender gender CO/CG CO/CG CO/CG CO/CG (1) (2) (3) (4) (5) (6) Individual Characteristics Age 36.26 38.27 0.16 34.06 37.85 0.00 Own Education 7.22 5.22 0.06 2.54 2.65 0.61 Business Owner 0.63 0.61 0.66 0.59 0.58 0.84 (Yes ¼ 1) Digit Span Recall 3.87 3.85 0.94 2.39 2.71 0.12 Risk Tolerance 4.09 3.80 0.38 2.80 3.43 0.02 (0 ¼ Risk Averse; 10 ¼ Risk Lover) Married (Yes ¼ 1) 0.77 0.82 0.01 0.83 0.85 0.48 Index of Optimism -0.23 -0.02 0.21 -0.87 -0.29 0.00 Business Literacy 0.97 0.66 0.07 0.14 0.23 0.67 Index Female Mobility 2 2 2 0.08 0.00 0.36 Index No Purdah 2 2 2 -0.30 0.19 0.00 Trust in Formal -0.01 -0.03 0.31 -0.28 -0.14 0.00 System Months as Member 16.48 28.74 0.00 20.28 24.98 0.45 Holds of�ce in Group 0.48 0.20 0.00 0.06 0.18 0.01 (Yes ¼ 1) Eligibility 0.49 0.63 0.00 0.53 0.63 0.00 Household Characteristics Household size 7.88 7.93 0.70 7.45 7.00 0.00 Years of Education of 3.87 2.67 0.02 4.93 4.88 0.91 Spouse Number of children 1.78 1.97 0.40 1.69 1.40 0.05 under 9 Land 4.38 6.01 0.39 4.48 2.58 0.30 (Continued ) 538 THE WORLD BANK ECONOMIC REVIEW TABLE A1. Continued Male Female P-val of P-val of Means t-test Means t-test (1)-(2) (4)-(5) Mixed Single Mixed Single Gender gender Gender gender CO/CG CO/CG CO/CG CO/CG (1) (2) (3) (4) (5) (6) Fraction of Members 0.41 0.43 0.12 0.45 0.21 0.00 of same Zaat (caste) Ever in Business 0.60 0.62 0.11 0.54 0.61 0.76 (Yes ¼ 1) Log of Value of 5.77 9.36 0.00 6.99 4.78 0.00 Livestock Distance to meeting 6.24 7.50 0.07 5.91 9.17 0.00 place Credit Constraints 0.09 0.12 0.28 0.19 0.16 0.31 (Yes ¼ 1) Expenditures (in 1,000 5.42 4.98 0.91 4.86 4.58 0.36 Rupees) Decision Making (0-8) 3.10 3.35 0.71 1.95 1.83 0.39 N. Obs 82 1798 137 1434 Which CO members qualify for this larger loan? All CO members that have a good borrowing record with NRSP (that is, have successfully repaid at least one loan) will be eligible to put in a request for a larger loan to fund their business activity. Who will obtain the larger loan? Among the applicants who qualify for the larger loan, a lottery will be implemented to determine who gets the larger loan. Winners of the lottery will receive the larger loan approved by NRSP. Losers of the lottery will be offered a normal size loan according to their credit history with NRSP. Each qualifying CO member will have a 50% chance of winning the lottery. To ensure trans- parency and fairness, the loan lottery will be done in the NRSP head of�ce in Islamabad. Procedure to apply for a larger loan The following steps are involved in accessing larger loans through this program (1) COs will pass a resolution identifying a larger loan need for their eligible and interested members, as such demands come in. (2) Each time there is a demand for a larger loan, a social and technical appraisal will be done to assess the applicant’s credit history and repay- ment capacity and the loan size that the candidate can safely repay. ´ , Mansuri, and Pico Gine ´n 539 (3) If a larger than normal loan is approved, the loan application will be forwarded to headquarters where the results of the lottery will be checked and disbursement made accordingly. (4) Borrowers who win the lottery will actually get the larger loan. Borrowers who lose in the lottery will be offered the normal loan amount set by NRSP. Amount of Loan Maximum up to Rs. 100,000/- (One hundred thousand only) according to the need Purpose of Loan Loan will be taken and subsequently utilized for productive purposes only Duration of Loan Up to maximum of One year Loan Repayment Loan repayment will be made according to the prevalent NRSP procedures, whereby the borrower will be given a repayment schedule and he/she will have to repay in installments, in accordance with the schedule, in the nearest NRSP village branch and take a repayment receipt. APPENDIX B: VARIABLE DEFINITIONS Data used in this paper come from two surveys: a baseline conducted in November 2006, a follow-up survey in December 2008. We also used adminis- trative data about loan take-up, obtained from NRSP’s internal records. TREATMENTS AND TAKE UP (FROM ADMINISTRATIVE RECORDS). † Female picture brochure takes the value of 1 if the member was shown a brochure with female business owners on the cover, 0 if the brochure showed the same businesses with men. † Eligibility, takes the value of 1 if member is eligible for the loan lottery, 0 otherwise. † Business Training, a dummy taking the value of 1 if the individual was offered business training, 0 otherwise. † Applied for larger loan, 1 if the member actually requested a loan, 0 otherwise. † Approved, conditional on applying for a loan, takes a value of 1 if the member was approved by NRSP, 0 otherwise. 540 THE WORLD BANK ECONOMIC REVIEW † Borrowed, takes a value of 1 if the individual actually borrowed money from NRSP; while conditional on approval, not all those that applied and were approved took actually a loan. † Amount borrowed, is measured in thousands of Rupees. BASELINE CHARACTERISTICS INDIVIDUAL. † Female equals 1 for women and 0 for men. † Age is respondent’s age in years. † Years of education is years of completed schooling, and is top-coded at 16. † Married, a dummy taking the value of 1 if member is married, 0 if single, divorced or widowed. † Digital span recall reports the number of digits correctly recalled after being shown an eight digit number for 30 seconds. † Member of a mixed group, dummy takes the value of 1 if the member belongs to a borrowing group with mixed gender, 0 if the group is of the same gender. † Index of female mobility and No purdah index are principal components of several variables with negative values indicating less mobility (or observing more types of purdah). † Business owner equals 1 if the member had a business at baseline, 0 otherwise. † Aversion to risk general is measured on a 0-10 scale where 0 indicates the most risk averse and 10 the most risk-tolerant/lover. † Months as member, number of months as member of NRSP group. † Holds Of�ce in Group, takes value 1 if member has or has had in the past a leadership position in group. † Fraction of Members of same Zaat (caste), is a percentage of members in the group that share the same cast of the member. HOUSEHOLD. † Education of spouse is years of completed schooling of the respondent’s partner, if any. Top coded at 16. † Total HH income and Expenditures, transformed to logs for analysis. † Number of children under 9 † Land is the total owned land inside and outside the village. ´ , Mansuri, and Pico Gine ´n 541 † Credit constraints, dummy taking a value of 1 if the member faced any type of credit constraint, formal or informal. † Ever in business, captures business experience within the household. Equals 1 when this is the case, 0 otherwise. † Decision Making, is the number of household decisions out of a total of eight that the member usually takes on his or her own. The decisions are: children’s schooling, consumption expenditures, major investments in business or land, the respondent’s participation in community or pol- itical activities, the respondent’s spouse participation in community or political activities, whether or not the respondent should work for an income, whether or not the spouse should work for an income and how much the household saves. In the analysis, a dummy is used that takes value 1 if the variable is above the median for each gender subsample. BUSINESS CHARACTERISTICS. † Type of business, dummy variables for businesses shown on brochure † Business Literacy, scores of component 1 of a PCA for a set of questions about knowledge about how to run a business, and of competition. † Sales in ‘000 rupees, sales of business in an average month at the time of baseline. ANALYSIS-RELATED VARIABLES. † Field Unit, refers to the NRSP branch and is the strati�cation variable. There are six �eld units in the three study districts. REFERENCES Akhtar, Shamshad, 2008. “Financial Sector– Ten Year Vision and Strategy�. Presentation at the inaugu- ration of the Development Finance Conference on “Expanding Frontiers of Financial Access in Pakistan� organized by State Bank on the eve of 60th Anniversary of the State Bank of Pakistan, Karachi Banerjee, Abhijit, and Esther Duflo, 2008. “Do Firms Want to Borrow More? Testing Credit Constraints Using a Directed Lending Program�, MIT Department of Economics Working Paper No. 02-25. Beaman, L. Raghabendra Chattopadhyay, Esther Duflo, Rohini Pande, and Petia Topalova, 2009. “Powerful Women: Does Exposure Reduce Bias?�, Quarterly Journal of Economics 124(4): 1497–1540. Bertrand, et al., 2010. “What is advertising content worth? Evidence from a consumer marketing �eld experiment�, Quarterly Journal of Economics, 125, 1:pp. 263– 306. De Mel, Suresh, David McKenzie, and Christopher Woodruff, 2009. “Are Women More Credit Constrained? Experimental Evidence on Gender and Microentreprise Returns�, American Economic Journal: Applied Economics, 1(3):1–32. ———, 2010. “Getting Credit to High Return Microenterprises: The Results of an Information Intervention� mimeo, World Bank. Evans, Franklin B., 1963. “Selling as a Dyadic Relationship�, American Behavioral Scientist 6: 76– 79 542 THE WORLD BANK ECONOMIC REVIEW ´ , Xavier, and Ghazala Mansuri, 2011a. “Money or Ideas? A Field Experiment on Constraints to Gine Entrepreneurship in Rural Pakistan�, mimeo. ———, 2011b. “Together We Will: Evidence from a Field Experiment on Female Voter Turnout in Pakistan�, mimeo. Jacoby, Hanan, and Ghazala Mansuri, 2011. “Crossing Boundaries: Gender, Caste and Schooling in Rural Pakistan�, mimeo. Jensen, Robert, and Emily Oster, 2007. “The Power of TV: Cable Television and Women’s Status in India�,NBER Working Paper No. 13305. Kahneman, Daniel, 2003. “Maps of Bounded Rationality: Psychology for Behavioral Economics�, American Economic Review, 93(5): 1449– 1475. La Ferrara, Eliana, Alberto Chong, and Suzanne Duryea, 2008. “Soap Operas and Fertility: Evidence from Brazil�, BREAD Working Paper No. 172. Mansuri, Ghazala, 2011. “Bottom up or Top Down? Participation and the Provision of Local Public Goods�, mimeo. Mobius, Markus M., and Tanya S. Rosenblat, 2006. “Why Beauty Matters�, American Economic Review 96(1): 222 –235. Moulton, Brent, 1986. “Random Group Effects and the Precision of Regression Estimates�, Journal of Econometrics, 32, 3:p. 385– 397. Nenova, T., C. Nenang, and A. Ahmad, 2009. “Bringing Finance to Pakistan’s Poor: A Study on Access to Finance for the Underserved and Small Enterprises�, World Bank report. Washington, DC: World Bank. Paluck, Elizabeth L., 2009. “Reducing intergroup prejudice and conflict using the media: A �eld exper- iment in Rwanda,� Journal of Personality and Social Psychology 96(3): 574– 587. Sen, A., 1999. Development and Freedom, New York: Anchor Books. UNDP, 1998. Human Development Report 1998: Consumption for Human Development. New York: Oxford University Press. Yunus, M., 1999. Banker to the Poor, London: Aurum Press Ltd. World Bank, 2008. “Finance for All? Policies and Pitfalls in Expanding Access�, World Bank: Washington, DC. Entrepreneurship and the Extensive Margin in Export Growth: A Microeconomic Accounting of Costa Rica’s Export Growth during 1997-2007 ´ s Rodrı Daniel Lederman, Andre ´guez-Clare, and Daniel Yi Xu Successful exporting countries are often seen as successful economies. This paper studies the role of new exporting entrepreneurs – de�ned as �rms that became expor- ters – in determining export growth in a fast growing and export oriented middle- income country i.e., Costa Rica during 1997-2007. It provides a detailed description of the contribution of export entrepreneurs in the short and long run, and comparing the observed patterns with an emerging literature on the role of the “extensive� margin in international trade. On a year-by-year basis, the rate of �rm turnover into and out of exporting is high, but exit rates decline rapidly with age (i.e., the number of years the �rm has been exporting). On average, about 30 percent of �rms in each Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 year tend to exit export activities, and a similar percentage of �rms enter. The exiting and entering �rms tend to be signi�cantly smaller than incumbent �rms in terms of export value (e.g., entrants export about 30 percent less on average than incumbent �rms). These �ndings are consistent with existing evidence for other middle income Latin American countries. However, in the long run new product-�rm combinations (i.e., product-�rm combinations not present in 1997) account for almost 60 percent of the value of exports in 2007. Surviving new exporters actively adopted new products (for the �rm, but not necessarily new for the country) and abandoned weaker existing products they start with, and their export growth rates were very high during a period (1999-2005) when those of incumbent exporting �rms were actually negative. JEL codes: F14 Daniel Lederman (corresponding author) is a lead economist at the International Trade Department of World Bank. Andre ´guez-Clare is a professor of economics at the Pennsylvania State University ´ s Rodrı and a faculty research associate of NBER. Daniel Yi Xu is an assistant professor of economics at New York University and a faculty research fellow of NBER. The authors gratefully acknowledge �nancial support from the World Bank’s Latin American and Caribbean Regional Studies Program under the proyect on the “Quality of Trade� led by William F. Maloney and from the Costa Rica Competitiveness Project led by Jose ´ Luis Guasch, Thomas Haven, and Jose ´ Guilherme Reis. The following colleagues provided useful insights and guidance during preliminary discussions: Jose ´ Luis Guasch, Por�rio Guevara, Bill Maloney, Martha Denise Pierolo, Roberto Alvarez, and Augusto de la Torre. Javier Cravino and Oana Luca provided stellar research assistance at various stages. Ronald Arce (PROCOMER), Francisco Gamboa (Research Director, PROCOMER), and Ricardo Matarrita (INCAE, formerly with PROCOMER) spent time providing and explaining the PROCOMER data during recent years. Last but not least, we gratefully acknowledge invaluable comments provided by two anonymous referees and Betty Sadoulet. THE WORLD BANK ECONOMIC REVIEW, VOL. 25, NO. 3, pp. 543 –561 doi:10.1093/wber/lhr031 Advance Access Publication July 28, 2011 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 543 544 THE WORLD BANK ECONOMIC REVIEW Successful exporting countries are often seen as successful economies. Most govern- ments use export promotion policies and have established export promotion agencies, regardless of the level of development or institutional capacities (Lederman, Olarreaga and Payton 2010). The World Bank consistently argues that promoting trade and exports in particular is a recipe for promoting �rm and national productivity (e.g., Fajnzylber, Guasch, and Lo ´ pez 2009). East Asian economies used export targets as part of their development strategies in the 1970s and 1980s (Noland and Pack 2003, Pack 1997). Furthermore, export activities are also seen by policymakers as a means to improve the productivity or other outcomes of small and medium enterprises, and thus export promotion policies are often designed to serve these �rms rather than large or multinational corporations (Volpe and Carballo 2008). In this paper we study the role of new export entrepreneurs in determining export growth. We focus on the case of Costa Rica, a successful middle-income economy, during the period 1997-2007. This article provides a detailed description of the contribution of export entrepreneurs in the short and long run, and compar- ing the observed patterns with an emerging literature on the role of the “extensive� Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 margin in international trade (e.g., Besedes and Prusa 2006, 2010; Eaton et al. 2007; Alvarez and Fuentes 2009; Brenton et al. 2010; Freund and Pierola 2010). The empirical analysis relies on customs data compiled by the Costa Rican export promotion agency, PROCOMER, which is part of the Ministry of Foreign Trade. The data includes all �rms with positive exports during 1997-2007 and provides information on export values by year classi�ed by product categories as well as quantities measured in kilograms and destination markets. Although the information is limited, because it does not provide information about domestic sales or other variables that would be needed to calculate standard indicators of �rm productivity, it does allow us to decompose export growth and explore the role of �rm turnover in and out of exporting activities, as well as the main drivers of export growth across �rms of different export sizes (in terms of value). The main �ndings are as follows. On a year-by-year basis, the rate of �rm turnover into and out of exporting activities is quite high. On average, about 30 percent of incumbent exporting �rms in each year tend to exit export activities while a similar percentage of �rms are new exporters. The exiting and entering �rms tend to be signi�cantly smaller than incumbent �rms in terms of export value (e.g., entrants export about 30 percent less on average than incumbent �rms). Over 40 percent of �rms exit exporting after one year, and the exit rate thereafter hovers around 20 percent.1 To put these numbers in an international comparative perspective, Freund and Pierola (2010) report that 34 percent of Peruvian �rms that export agricultural and agribusiness products exit after one year. Brenton et al. (2010) report that for middle-income economies only about 51 percent of product-destinations survive past the �rst year, with the rate of sur- vival stabilizing just below 20 percent in subsequent years. 1. These numbers were calculated by the authors, based on data provided by PROCOMER and cleaned by the authors – see Section II below. ´guez-Clare, and Xu Lederman, Rodrı 545 In the long run, the main driver of export growth in Costa Rica was the introduction of new products by surviving �rms. New product-�rm combi- nations (i.e., product-�rm combinations not present in 1997) account for almost 60 percent of the value of exports in 2007 Surviving new exporters actively adopt new products (for the �rm, but not necessarily new for the country) and abandon weaker existing products they start with. The rest of this paper is organized as follows. Section I briefly compares Costa Rica’s export and growth performance during 1997-2007 with other countries and regions. Section II discusses the PROCOMER data by comparing the series with other data on Costa Rican exports. This section also describes how the �rm data was cleaned and discusses key features of the resulting data set, including the lack of signi�cant changes in the composition of exports and exporting �rms across broad industrial categories. Section III presents the microeconomic account- ing frameworks used to assess the contribution of incumbent and new �rms, pro- ducts and export market destinations, to total export growth in the short run and in the long run. Section IV concludes with a brief summary of the results. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 I . CO S TA R I CA ’ S E X PO R T A N D G ROW T H P E R FO R M A N C E I N CO M PA R AT I V E P E R S P E C T I V E , 1 9 9 7- 2 00 7 Table 1 shows average growth rates of exports (measured in current U.S. dollars) and Gross Domestic Product per capita (adjusted for Purchasing Power Parity with international prices of 2005) for the period 1997 – 2007. Costa Rica’s average export growth during this period was 9 percent per year. This is a bit higher than the average growth rate for Latin America and the Caribbean region, and clearly superior to the Central America and Caribbean region, but lowers than the average for all the other regions. Within Latin America, Costa Rica’s export growth performance is dominated by Peru and Chile, which achieved export growth rates of 16.4% and 15.7%, respectively. The picture is slightly better when looking at GDP per capita growth rates. Costa Rica’s economic growth rate of 3.3% is higher than Peru and Chile, and higher than the average for all regions except Europe and Central Asia and South Asia. This suggests that Costa Rica’s overall economic performance was relatively more impressive than its export growth rate I I . T H E P RO COM E R D ATA Before conducting a detailed analysis of the microeconomics of export growth, we need to ascertain the quality of the data. Figure 1 compares the value of total Costa Rican merchandise exports in the PROCOMER data with the World Bank’s data on total merchandise exports, which come from of�cial government sources. The two series are not identical, which is worrisome. However, the over time correlation is very strong, and thus the PROCOMER data does capture the direction if not the exact magnitude of merchandise export growth 546 THE WORLD BANK ECONOMIC REVIEW T A B L E 1 . Annual Merchandise Export and GDP per Capita Growth Rates, 1997-2007 Exports GDP per Capital (PPP) Country/Groups Mean Median Mean Median Costa Rica (1) 0.090 0.129 0.033 0.042 Export Over Achievers Peru (2) 0.164 0.160 0.028 0.033 Chile (2) 0.157 0.147 0.026 0.030 China (3) 0.215 0.215 0.088 0.088 Cambodia (3) 0.200 0.173 0.073 0.074 Azerbaijan (4) 0.317 0.361 0.140 0.100 Albania (4) 0.253 0.266 0.065 0.055 Libya (5) 0.210 0.230 0.016 0.026 Lebanon (5) 0.198 0.168 0.013 0.012 Buthan (6) 0.199 0.127 0.061 0.045 India (6) 0.158 0.193 0.055 0.062 Chad (7) 0.520 0.058 0.046 0.039 Sierra Leone (7) 0.447 0.480 0.043 0.039 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Slovak Republic (8) 0.207 0.154 0.048 0.041 Czech Republic (8) 0.190 0.156 0.035 0.037 Equatorial Guinea (9) 0.374 0.422 0.190 0.143 Trinidad and Tobago (9) 0.215 0.236 0.076 0.074 Kazakhstan (10) 0.240 0.243 0.080 0.089 Lybia (10) 0.210 0.230 0.016 0.026 Sudan (11) 0.340 0.291 0.045 0.040 Azerbaijan (11) 0.317 0.361 0.140 0.100 Chad (12) 0.520 0.058 0.046 0.039 Sierra Leone (12) 0.447 0.480 0.043 0.039 Other Comparator Countries Singapore 0.100 0.129 0.034 0.040 Ireland 0.087 0.078 0.047 0.037 Hong Kong 0.067 0.090 0.032 0.040 Regional and Income Groups Latin America and Caribbean 0.084 0.083 0.020 0.019 Central America and Caribbean (1) 0.055 0.051 0.022 0.023 Latin America (2) 0.114 0.123 0.017 0.014 East Asia and Paci�c (3) 0.097 0.088 0.026 0.020 Europe and Central Asia (4) 0.168 0.171 0.062 0.060 Middle East and North Africa (5) 0.163 0.162 0.023 0.026 South Asia (6) 0.126 0.104 0.042 0.042 Sub-Saharan Africa (7) 0.121 0.094 0.017 0.016 High income OECD members (8) 0.097 0.087 0.026 0.023 High income non OECD economies (9) 0.112 0.095 0.035 0.025 Upper-middle income economies (10) 0.105 0.093 0.034 0.030 Lower-middle income economies (11) 0.135 0.114 0.035 0.026 Low income economies (12) 0.119 0.094 0.017 0.018 Note: This table presents mean and median annual growth rates of merchandise exports (current US$) and GDP per capita PPP (constant 2005 international $) for each of the described groups. The regional and income country classi�cations come from the World Bank (as of July 2008). Each country’s group is indicated inside parentheses. All data are from the World Bank’s World Development Indicators. ´guez-Clare, and Xu Lederman, Rodrı 547 F I G U R E 1. Costa Rica: Merchandise Exports versus PROCOMER Data, 1997-2007 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Source: Data are from PROCOMER and World Bank, and both series are expressed in current U.S. dollars. observed in the macroeconomic data. The divergence between the two series is largest in 1997 and 1998, which converge to a constant gap in 1999. Consequently, the PROCOMER data tend to exaggerate the total export growth rates in 1998 and 1999, but the subsequent growth rates are compar- able, as shown in Table 2. Nonetheless, the underlying microeconomic dynamics observed in the PROCOMER data are informative for understanding the sources of total export growth during this time period. Unfortunately there are other issues with the PROCOMER data, which would affect the microeconomic analyses. Table 3 lists the steps taken to clean the data and the main features of the resulting data set. The cleaning focused on issues that are relevant for understanding the microeco- nomic sources of export growth, including the role of incumbent, new and exiting �rms, products and destinations, as well as unit values (the ratio of export value over quantities exported measured in kilograms). Hence we removed duplicate observations of �rms-product-destinations, deleted observations where the descrip- tion of the product was empty, entries where the reported quantity was zero, obser- vations that reported Costa Rica as the destination market, etc. After this phase of the cleaning, the total number of observations of �rms-products-year-destinations declined by about 6,000 observations and total export value declined by 0.26 percent. The next step was to remove any remaining duplicate observations that were due to the recording of the same �rm, but with spelling mistakes or other 548 THE WORLD BANK ECONOMIC REVIEW T A B L E 2 . Costa Rica: Merchandise Export Growth Rate Procomer Export Growth (%) Export Growth WDI 1998 56.8 29.1 1999 26.7 19.3 2000 2 12.1 2 10.8 2001 2 13.6 2 14.4 2002 5.0 4.8 2003 15.6 15.9 2004 2.5 3.3 2005 11.4 11.5 2006 17.8 16.7 2007 13.9 14.2 Source: Data are from PROCOMER and World Bank, and both series are expressed in current U.S. dollars. T A B L E 3 . Cleaning the PROCOMER Data Share of Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 # of Total real export initial value Step observations value (1997 U.S.$) (%) Raw data 296,238 59,219,688,885 100 (-) Duplicate entries 10 174,453 0.00 (-) Entries with desc ¼ "" 56 3,584,951 0.01 (-) Entries with quantity 0 4,739 3,638,070 0.01 (-) country: Costa Rica 217 2,144,049 0.00 (-) country: Alta Mar 4 44,239 0.00 (-) country: Generico 189 4,821,894 0.01 (-) country: Zonas Francas de Exportacion 678 148,249,409 0.25 Subtotal 290,345 59,057,031,820 99.73 Cleaning �rm-product-year-country 289,549 59,057,031,820 99.73 observations (consolidation of observations with similar �rm names) (-) Abbott entries (product-destination-years) 1,157 1,816,626,930 3.07 (-) Intel entries (product-destination-years) 3,277 13,689,799,703 23.12 Cleaned data 285,115 43,550,605,188 73.54 Source: Authors’ calculations based on data from PROCOMER. errors in the reported name of the �rm. This step reduced the total number of �rm-product-destination-year observations by about 800. Finally, given that we are interested in learning about Costa Rican �rm dynamics in export activities, we removed the observations corresponding to two big multinational corporations, INTEL and Abbott Labs, whose export experience is better suited for �rm case studies. INTEL alone accounted for over 23 percent of the value of exports of Costa Rica during 1997-2007, while Abbot contributed over 3 percent of the total. Table 4 describes the main features of the resulting data set. Regarding export growth, the resulting data shows a weaker performance than the aggre- gate export data presented in Figure 1. The year 1998 remains an outlier with a real growth rate of 29 percent, even after removing the exports from INTEL ´guez-Clare, and Xu Lederman, Rodrı 549 T A B L E 4 . Summary of the Cleaned PROCOMER Data: Real Export Growth, Firms, Products, and Destinations Total exports Year (1997 USD) Growth (%) Firms HS6 products HS10 products Destinations 1997 3,217,752,622 - 2,200 2,454 3,283 121 1998 4,157,732,964 29 2,328 2,629 3,513 129 1999 3,911,376,028 26 2,432 2,599 3,505 116 2000 3,695,897,353 26 2,392 2,594 3,557 121 2001 3,573,232,631 23 2,493 2,626 3,641 127 2002 3,827,746,595 7 2,531 2,592 3,707 129 2003 3,860,064,635 1 2,670 2,736 3,869 133 2004 4,055,665,596 5 2,760 2,774 3,915 134 2005 4,083,061,065 1 2,863 2,800 3,933 138 2006 4,507,923,206 10 2,937 2,833 4,087 136 2007 4,660,152,494 3 2,973 2,878 4,293 151 1997-2007 (total) 8,865 4,568 7,941 189 1997-2007 (continuing) 554 1,262 1,232 88 Continuing/Unique (%) 6.2 27.6 15.5 46.6 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Source: Authors’ calculations based on data from PROCOMER and World Bank. Export growth is expressed in constant U.S. dollars of 1997, using the U.S. Producer Price Index as deflator. and Abbot.2 This year was also an outlier in the aggregate export data. In the remaining years, the real export growth rates were rather mediocre when com- pared to the aggregate data, with negative real growth rates during the years of the East Asian �nancial crisis of 1997-1999, the beginning of the U.S. slow- down in 2000-2001, and the incorporation of China as a full member of the World Trade Organization in 2001. Thereafter the real export growth rates were positive but modest, except in 2006 when it reached 10 percent. The most striking feature of the data, however, is the low number of �rms that reported exports in every year during 1997-2007, which were 554. This is a mere 6.2 percent of the total number of �rms that appear in the sample during the eleven years. The number of continuing products, measured either at the 6- or 10-digit product categories, is also small relative to the total number of products exported at any time during this period (27.6 and 15.5 percent, respectively). The difference between these ratios also indicates that the level of aggregation of the product categories affects the accounting of the contribution of the new products. In contrast, the number of export desti- nations was relatively constant over time, and over 46 percent of destinations were serviced every year. Overall, the cleaned data suggests that the rate of turnover of exporting �rms and products is quite high, with very few continu- ing �rms and products, the latter being sensitive to the product nomenclature. Table 5 provides a standard analysis of the composition of trade. It shows the export shares of broad industrial sectors. If the aforementioned high turnover 2. It is noteworthy that both companies began exporting from Costa Rica in 1998. T A B L E 5 . Broad Inter-Sectoral Changes are Absent 550 A. Sectoral Shares in Total Exports (%) Ind 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 1 10.3 8.6 4.4 4.2 4.4 4.1 4.0 3.4 3.6 3.0 3.0 2 33.1 29.9 31.0 27.0 26.9 25.5 27.9 27.0 26.1 26.7 27.9 3 8.5 8.2 8.7 9.0 9.4 10.2 10.1 11.2 11.0 11.6 13.5 4 0.8 0.7 0.8 1.5 1.4 1.6 0.9 0.3 1.0 0.9 1.0 5 5.1 5.0 5.8 6.6 7.7 7.9 8.5 8.5 8.2 7.8 8.1 6 3.7 3.2 4.0 4.5 4.9 5.2 5.5 6.0 6.6 6.5 6.0 7 1.1 1.4 1.5 1.2 1.2 1.0 1.2 1.8 1.7 1.4 1.1 8 3.0 2.7 3.0 2.9 3.0 2.9 2.6 3.3 3.3 3.4 3.7 9 6.0 16.7 18.0 19.1 18.0 17.4 13.9 11.7 10.4 8.1 6.8 10 0.5 0.5 0.4 0.2 0.1 0.0 0.0 0.0 0.0 0.0 0.0 11 2.0 2.5 2.6 2.2 2.5 2.7 2.3 2.1 2.4 2.9 2.5 12 2.8 2.6 2.8 3.0 3.0 3.0 3.1 3.7 4.2 4.9 5.7 13 8.0 10.8 9.6 12.4 11.5 12.6 13.1 14.2 13.6 14.6 11.4 THE WORLD BANK ECONOMIC REVIEW 14 0.2 0.3 0.4 0.8 0.9 0.5 0.7 0.4 0.5 0.5 0.7 15 1.6 3.7 4.0 4.1 5.1 5.4 6.0 6.5 7.3 7.6 8.8 16 13.2 3.1 2.9 1.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 B. Sectoral Share in Total Firms (%) 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 1 6.9 6.9 6.5 5.6 6.3 6.0 6.1 5.5 5.0 4.9 4.4 2 28.5 27.1 27.4 29.8 30.7 29.1 28.2 27.6 26.1 28.1 28.8 3 9.2 9.8 10.1 10.6 10.4 9.8 9.8 10.5 10.5 10.7 10.3 4 3.6 4.7 4.1 4.0 3.7 3.7 3.6 3.8 3.7 3.8 3.6 5 14.1 15.0 15.2 16.0 16.3 16.8 16.6 16.7 16.6 16.7 16.4 6 15.2 18.1 18.5 19.1 18.7 18.4 19.4 20.2 21.2 21.6 21.6 7 3.0 3.4 3.2 3.4 3.6 3.2 3.9 4.1 3.9 3.6 3.5 8 18.8 19.7 20.1 19.9 20.8 20.4 21.1 21.2 22.2 20.4 20.8 9 11.1 11.4 12.3 12.3 11.4 9.9 10.0 10.6 10.3 9.1 9.5 10 2.9 3.0 2.7 2.9 2.5 3.0 3.5 3.4 3.1 3.2 3.5 11 6.2 6.8 6.5 5.8 5.9 6.3 7.0 6.7 7.4 7.3 6.6 12 15.1 16.4 16.4 15.7 16.6 16.7 16.9 17.8 18.1 19.1 18.2 13 22.4 25.3 26.1 26.2 25.8 26.6 26.9 26.7 28.2 28.4 27.8 14 3.6 5.2 4.2 4.6 4.9 5.1 4.2 5.0 4.8 5.3 5.4 15 14.5 17.1 17.6 18.3 18.6 20.0 18.4 19.0 20.8 20.0 18.8 16 9.5 5.8 2.5 2.3 0.6 0.2 0.1 0.1 0.0 0.0 0.1 Source: Authors’ calculations based on data from PROCOMER. Ind ¼ industry. See industry classi�cation in the Appendix. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 ´guez-Clare, and Xu Lederman, Rodrı 551 rates of �rms and products were associated with structural change across indus- tries, we should also observe dramatic changes in the composition of trade. The data in this regard is a bit noisy and needs to be analyzed with caution. The last industry, which is broadly labeled “services�, captures over 13 percent of total exports at the beginning of the sample but falls to zero by 2001. Also, the indus- try labeled “Miscellaneous� experienced a dramatic increase in its share, but it is dif�cult to interpret these fluctuations in export shares as structural change precisely because these industries are loosely identi�ed. Perhaps more interest- ingly, between 2001 (when “services� were more appropriately recorded as having a zero share in the merchandise export accounting) and 2007 the most dramatic decline happened in textiles and apparel, an industry that we know faced tough competition from Chinese exports to third markets, including the United States, as a result of its WTO accession and the dismantling of the Multi-Fiber Arrangement that had historically maintained high levels of protec- tion for costly producers in high-income countries (see Ozden and Sharma 2006 and Ozden 2006).3 In brief, the changes of the inter-industry structure of Costa Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Rican exports between 1997 and 2001 seems to be a recording illusion, due to the role of “miscellaneous� and “service� merchandise exports, and after the latter category disappeared in 2001, the only notable structural change is related to the textile and apparel sectors. More importantly, these structural changes are very small when compared to the evidence concerning the rather high rates of �rm and product turnover rates during 1997-2007. Thus, most of the action in terms of �rm and product dynamics was within sectors, rather than across broad industrial categories. This observation alone makes the fast growing literature on �rm heterogeneity and within industry dynamics (e.g., Melitz 2003) of par- ticular interest to the case of Costa Rica. In this light, the following sections assess the contribution of �rm, product and export-destination dynamics to overall export growth. I I I . AC CO U N T I N G FOR MICROECONOMIC SOURCES OF EXPORT GROWTH What are the main micro sources that drive export dynamics? We borrow the insights from the literature of industry dynamics to view �rms’ exporting behav- ior as a process of “creative destruction.� A consistent message from the indus- try dynamics literature, both empirical and theoretical, is that new �rms are born small and suffer a high hazard rate of exit. Yet, in the medium to long run, the new �rms that manage to survive grow rapidly and take over the incumbents. This in turn forces the inef�cient incumbent �rms to quit the market. 3. Ozden and Sharma (2006) found that preferential margins on apparel exports to the U.S. from bene�ciaries of the Caribbean Basin Initiative declined signi�cantly during 2000-2002 when compared to 1992-2000. Ozden (2006) found that Costa Rican apparel exports to the U.S. declined after 2000, but that average unit values (prices) of apparel exports to the United States rose. 552 THE WORLD BANK ECONOMIC REVIEW While a lot of previous studies have looked at the above-described process using data of �rm domestic sales, entry, and exit, few have investigated this with data on �rm export market dynamics. The related work in this area includes Colombia (Eaton et al., 2007), Chile (Alvarez and Fuentes 2009), and Peru (Freund and Pierola 2010). We believe that looking at trade transactions data brings some empirical advantage compared with previous studies based on industry census data. First, customs data provides richer information on how new �rms expand and prosper over time. They can expand by selling products in more markets, or adopting existing product lines, or creating brand new products. While each of these activities will be reflected in the growth of export sales, they tend to have different implications for competition, resource allocation, and welfare. Second, the fact that exporting �rms potentially serve multiple markets at different points in time provides rich variation in controlling for �rm’s initial heterogeneity and avoids selection bias. To assess the contribution of microeconomic dynamics related to �rm, product and export destinations, we conduct two sets of export-decomposition exercises. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 The �rst concerns the contribution of microeconomic dynamics to short-term growth, namely on annual growth rates during 1997-2007. The second explores the contribution of new export entrepreneurs, products and export destinations in the longer run, which is de�ned as a �ve-year period of export growth (1999-2004). Short-run decomposition framework The export-growth decomposition framework used by Eaton et al. (2007) to study export growth in Colombia is given by equation (1): 0 1 1 CN CN X t À X t À1 ðX t À1 þ Xt Þ CN CN Xt À XtÀ1 C B ¼ @2 à A 1 1 1 CN ðXtÀ1 þ Xt Þ ðXtÀ1 þ Xt Þ ðXtÀ1 þ XCN t Þ 2 2 2 0 1 EN BNENtÀ1 ÃXtÀ1 Xt À NENtÀ1 XtÀ1 C þ@ þ A ð1Þ 1 1 ðXtÀ1 þ Xt Þ ðXtÀ1 þ Xt Þ 2 2 0 1 EX B NEXtÀ1 ÃXtÀ1 XtÀ1 À NEXtÀ1 XtÀ1 C þ @À À A 1 1 ðXtÀ1 þ Xt Þ ðXtÀ1 þ Xt Þ 2 2 where Xt equals total exports in period t; X  tÀ1 is the average exports (across �rms) in period t-1; CN, EN and EX are indexes for variables corresponding to continuing, entering and exiting �rms, respectively (continuing �rms are those that exported both in t and t-1, entering �rms are those that exported in t but not in t-1, and exiting �rms are those that exported in t-1 but not in t); and NENt and NEXt is the number of entering and exiting �rms in t, ´guez-Clare, and Xu Lederman, Rodrı 553 respectively. The denominator in the export growth ratio is the average of exports in t and t-1, which Eaton et al. use for convenience so that the growth rates do not depend on the one year. In any case, the results, discussed further below are not signi�cantly affected by this choice of denominator. In a nutshell, the decomposition exercise separates the contributions to annual export growth of incumbent, entering and exiting �rms. The contri- bution of incumbent �rms is simply the product of the share of exports of incumbent �rms times their export growth. This contribution appears in the �rst term inside brackets in equation (1). The contribution of entering �rms has two components, both appearing inside the brackets of the second term on the right-hand side of (1). The �rst is simply the number of entering �rms as a share of average number of total �rms in t-1 and t. In (1), this is written as the number of entrants times the average exports per �rm in t-1. The second component concerns the deviation of the average exports of new �rms from the average exports of incumbent �rms, the latter being equal to the number of new �rms’ times the average exports Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 per �rm in the previous year. The contribution of exiting �rms is analogous to the decomposition of the contribution of entrants. The results from these decompositions of annual export growth rates are presented in the following section and Table 6a. As discussed below, the annual growth rates are dominated by the contribution of incumbent �rms. Thus, we explore the contribution of new and exiting products and export destinations by incumbent �rms in Tables 5b and 5c. Short-run results The �rst column of Table 6a shows the annual growth rate of real exports observed in the cleaned PROCOMER data. The second column shows the share in total exports in the previous year due to incumbent �rms, and the third column contains the export growth of these incumbent �rms. The evidence clearly shows that incumbent �rms dominate export growth from year to year, as over 95 percent of total exports in the previous year were due to �rms that remained as exporters in the following year. Consequently the export growth rate of the incumbent �rms closely tracks the annual growth of total exports. Interestingly, the rate of turnover of exporting �rms is large. The number of entrants account for more than 27 percent of the number of �rms in every year – see column 4. Similarly, the rate of exit is higher than 25 percent every year – see column 6. Furthermore, the average exports of entrants and departing �rms tended to be low relative to the average export value of incumbent �rms. This is reflected in the value gap of entrants and departing �rms, which was around 30 percent during the period (i.e., entrants were 30% smaller than incumbents, and similarly for �rms that stopped exporting). Thus, incumbent �rms’ export growth dominates year by year export growth in spite of the rather high turnover rate of exporting �rms, because both new and exiting �rms export very low values. 554 THE WORLD BANK ECONOMIC REVIEW T A B L E 6 a . The Contribution of Firm Turnover in the Short Run Export Share cont. Growth cont. Entry Entry value Exit Exit value growth (%) �rms (%) �rms (%) (%) gap (%) (%) gap (%) 1998 25 97 26 30 2 27 2 25 23 1999 26 95 27 35 2 30 2 31 26 2000 26 96 24 30 2 28 2 32 27 2001 23 97 25 34 2 31 2 30 28 2002 7 98 7 30 2 28 2 29 27 2003 1 98 1 34 2 32 2 29 26 2004 5 98 7 32 2 30 2 28 25 2005 1 99 1 33 2 31 2 29 28 2006 10 98 11 29 2 27 2 27 24 2007 3 99 3 27 2 26 2 25 24 Source: Authors’ calculations based on data from PROCOMER. Cont. ¼ continuing or incum- bent �rms. The results on �rm dynamics for Costa Rica seem high. However, the Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 evidence from Colombia and Chile also suggest that �rm turnover in export activities tends to be high. For the case of Colombia, Eaton et al. (2007) report entry rates that average over 45 percent and the average exit rates surpass 43 percent per year during 1996-2005. Alvarez and Fuentes (2009) report compar- able rates for Chilean manufacturing exports during 1991-2001 when annual entry rates averaged over 29 percent while exit rates ranged between 8 and 28 percent. Freund and Pierola (2010) also found high entry and exit rates of export entrepreneurs in the agricultural and agribusiness sectors in Peru during 1994-2007.4 Thus the results for Costa Rica in terms of �rm dynamics seems consistent with data from other case studies. Next, we investigate the growth and decline path of incumbent �rms. The customs data provides us with two important dimensions of incumbent �rms’ export dynamics: destinations and products. Regarding the role of new export destinations in shaping the growth of exports of incumbent �rms, Table 6b shows the results from the decomposition of the annual export growth of incumbent �rms into incumbent destinations, new destinations, and exiting destinations. Not surprisingly, incumbent destinations account for most of the observed export growth of incumbent �rms, but we do observe non-trivial entry and exit of new destination markets. Table 6c presents the results concerning the contribution of the new and exiting products exported by incumbent �rms. The entry and departure rates of products exported by incum- bent �rms are very high, even higher than the �rm and export-destination 4. Freund and Pierola (2010) report entries and exits in the annual data ranging from under 100 at the beginning of the period to close to 200 by the end. They also report that the total number of �rms in the sectors they investigate peaked at 593 in 2007. Both Freund and Pierola (2010) and Besedes and Prusa (2010) propose theoretical models that rely on ex-post realizations of �xed costs of exporting to explain these high rates of entry and exit. ´guez-Clare, and Xu Lederman, Rodrı 555 turnover rates reported in tables 5a and 5b. Also, the value gaps are larger for new and exiting products than for entering and exiting �rms or destinations. In sum, in the short-run, the growth rate of exports by incumbent �rms is the main factor behind the aggregate export growth rate, but this occurs with vigorous �rm dynamics. These dynamics are characterized by high �rm entry and exit from export activities, experimentation by incumbent �rms with new markets and especially new products. The introduction of new products and the shedding of existing products by incumbent �rms tend to be the largest source of renewal for Costa Rica’s exports. Firm dynamics and export growth in the long run The previously discussed results concern annual export growth rates. Given that incumbent �rms dominate aggregate export growth in the short run, but with notably high churning at the �rm and �rm-product level, it is worth asking whether new �rms, products, or destinations made signi�cant contri- butions to Costa Rica’s total export growth between 1997 and 2007. In the Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 long run we can expect much higher contributions to total exports from �rm dynamics. For instance, in the case of Colombia, existing evidence indicates that the contribution of new �rms to total export growth (of 7.6 percent) between 1996 and 2005 rose from 10 percent on an annual basis to 26 percent for the whole period. The case of Chilean manufacturing exports is more dra- matic: the contribution of entering �rms to total export growth during 1991-2001 (which averaged 11.5 percent per year) rises from 23 percent on an annual basis to over 83 percent over the whole period. In the case of Costa Rica, new product-�rm combinations (i.e., product-�rm combinations not present in 1997) account for almost 60 percent of the value of exports in 2007.5 However, what are the �rm dynamics behind this high long-run contri- bution of export entrepreneurship? Is it that new exporters grow faster than incumbent �rms when they survive for a few years? If so, is this export growth by new export entrepreneurs associated with changes in products? To better understand the microeconomic dynamics underpinning the long run result we use the period 1999-2005 and examine the contribution of new exporters relative to incumbents. We start with year 1999 to identify the “new� exporters. We de�ne any �rm that never appears in the customs records before 1999 (i.e. on 1997 and 1998) and enter the record on or after 1999 as a “new� �rm. This helps to alleviate the concern of the initial status problem since our sample starts in 1997. With a similar rule, we de�ne a �rm as an “exiting� exporter if it appears in custom records on or before 2005 and never appears in the customs records after 2005 (i.e. on 2006 and 2007). Incumbent exporters are de�ned as those that enter before 1999 and survive until 2005. 5. This �nding is also consistent with an analysis of the contribution of new products and new destinations utilizing SITC 4-digit product level data provided by an anonymous referee: In 2008, ninety three percent of total exports were due to product-destination relationship that existed in 1999. 556 THE WORLD BANK ECONOMIC REVIEW T A B L E 6 b . The Contribution of Destination Turnover among Incumbent Firms in the Short Run Growth cont. Share cont. Growth cont. Entry Entry value Exit Exit value �rms (%) �rm_dest (%) �rm-dest (%) (%) gap (%) (%) gap (%) 1998 26 96 25 28 2 23 2 23 20 1999 27 95 25 27 2 24 2 30 24 2000 24 97 24 28 2 26 2 26 23 2001 25 97 24 26 2 23 2 27 24 2002 7 96 8 25 2 22 2 26 22 2003 1 98 1 27 2 25 2 24 22 2004 7 98 7 26 2 24 2 24 22 2005 1 97 21 28 2 25 2 25 23 2006 11 98 10 26 2 23 2 24 22 2007 3 97 4 28 2 25 2 24 20 Source: Authors’ calculations based on data from PROCOMER. T A B L E 6 c . The Contribution of Product Turnover among Incumbent Firms in Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 the Short Run Entry Growth cont. Share cont. Growth cont. Entry value gap Exit Exit value �rms (%) �rm_prod (%) �rm_prod (%) (%) (%) (%) gap (%) 1998 26 94 30 50 2 45 2 41 34 1999 27 94 21 52 2 49 2 51 42 2000 24 97 24 54 2 51 2 49 46 2001 25 96 27 53 2 48 2 49 46 2002 7 96 6 46 2 41 2 45 41 2003 1 96 5 53 2 50 2 46 40 2004 7 96 6 48 2 44 2 46 43 2005 1 98 0 50 2 48 2 46 45 2006 11 98 10 48 2 46 2 44 42 2007 3 92 2 51 2 43 2 48 41 Source: Authors’ calculations based on data from PROCOMER. The bene�t of these de�nitions is that the group of incumbent exporters is �xed over the whole period. This allows us to evaluate the contribution of “new� exporters to long-run export growth despite their extremely high year-by-year turnover rate. Table 7 summarizes the share of Costa Rica’s total exports by entrants versus incumbents. It illustrates that the new exporters as a group experienced an increase in exports by almost a factor of ten during the sample period. In contrast, the sales of incumbent exporters declined by 25% over the same period. This is consistent with earlier empirical �ndings that older �rms grow slower and their growth rate eventually becomes negative (see, for instance, Dunne, Roberts, and Samuelson 1989 on manufacturing plants in the United States). Overall, it is fair to say that new exporters are the main driving force ´guez-Clare, and Xu Lederman, Rodrı 557 of export growth in this sample period. If we think of “exporting entrepreneur- ship� as entry of new �rms into exporting, we can say that over the medium term export entrepreneurship is the main driver of export growth in Costa Rica. What are these new exporters doing? How are they growing? This is what we explore next. To further understand the path of new exporter dynamics, we can also look at each cohort of these entrants. We de�ne the year 1999 cohort as the new exporters that appear for the �rst time in custom records in year 1999. In Table 8 we trace out the annual export sales of surviving 1999 cohort �rms and exiting 1999 cohort �rms. As shown in Table 8, although the majority of the 1999 cohort exited before the end of our sample (only 152 out of 732 remain), the exiting �rms as a group account for very little in terms of total export sales of the whole cohort. As a result, total exports of this cohort were dominated by surviving �rms. The results of Table 8 are consistent with the view that although new exporters enter small and have a high failure rate, the surviving ones tend to catch up with incumbents Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 quickly. A natural question here is whether the new product margin is an impor- tant channel for growth among surviving new exporters. To answer this question, we separate new exporters’ products into two categories. We de�ne “initial products� as the 6-digit products that the new exporters sell in the �rst year of their export market participation. We further de�ne “added pro- ducts� as those 6-digit products that are added in their later years of export- ing. In Table 9a, we report the sales coming from “initial products� and “added products� for the surviving exporters within the 1999 and 2000 cohorts. Both cohorts exhibit a similar pattern: the new product margin and the initial product margin are equally important in contributing to new exporters’ sales growth in the long run. The “added products� category explains close to 40% of new exporters’ growth after 5 years, while the within “initial product� category growth is also very strong over the similar time span and explains 60%. In contrast, in Table 9b, we report the value of “continuing products� and “added products� of incumbent �rms which survive until year 2005. For this group of �rms, the “continuing products� are 6-digit products they’ve already exported before 1999. Although this group of �rms as a whole declines over time, their “added products� still grows quite substantially from 1999-2005. However, compared with the new exporters of 1999 and 2000, the growth rate of “added products� from incumbent �rms is lower. Table 9c documents how new exporters drop their initial products and how the remaining initial products grow. We again focus on the 1999 cohort. The evidence suggests that the creative destruction process on the product margin within exporters also has a strong selection effect. New exporters keep only the strongest products that they start with and drop the weaker products along the way. Table 9d reports how incumbent exporters drop their continuing products 558 THE WORLD BANK ECONOMIC REVIEW T A B L E 7 . The Contribution of New vs. Incumbent Exporters in the Long Run New Incumbent Sales Share (%) Number Sales Share (%) Number 1999 5 732 95 823 2000 13 942 87 817 2001 18 1,221 82 821 2002 25 1,391 75 818 2003 32 1,616 68 834 2004 36 1,791 64 846 2005 39 1,885 61 922 T A B L E 8 . Total Sales of 1999 New Exporters: Survivors vs. Exits (millions of US $) Survivors from 1999-2005 Exit at Year x Survive Year x But Exit Later Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 1999 157 15.0 36.1 2000 332 9.46 50.4 2001 270 6.85 45.0 2002 337 1.63 22.7 2003 422 12.1 4.96 2004 581 3.05 2005 641 and how the surviving products grow. Similar to new exporters, incumbent exporters also drop their weaker products over the years. However, their sur- viving products also decline gradually, while the surviving initial products of new exporters grow strongly. This explains a large fraction of the difference in export growth between new and incumbent exporters Overall, Tables 6–8 provide a coherent picture of Costa Rica’s export growth from 1999–2005. We �nd that the surviving new exporters are the major contributing force to aggregate export growth. Meanwhile, they actively adopt more products and abandon weaker existing products they start with. So in this sense, the survival of new exporters itself is not random: it is partially determined by a �rm’s active experimentation with their export product lines. IV. CONCLUSION Costa Rica’s export growth was not stellar when compared to other countries, and even less so without the contributions of two large multinational corpor- ations. Inter-sectoral adjustments across broad industries were negligible, both in terms of export-value shares and in terms of the number of exporting �rms ´guez-Clare, and Xu Lederman, Rodrı 559 T A B L E 9 a . Total Sales of 1999/2000 New Exporters: Initial Products vs. Added Products (millions of US $) Cohort 1999 Cohort 2000 Initial Products Added Products Initial Products Added Products 1999 157 2000 324 7.87 79 2001 241 29.0 188 12.8 2002 298 39.7 200 41.2 2003 358 63.3 229 76.3 2004 442 140.0 245 82.1 2005 406 235.0 209 102.0 T A B L E 9 b . Total Sales of Incumbent Exporters: Continuing Products vs. Added Products (millions of US $) Incumbent Exporters - 2005 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Continuing Products Added Products 1999 2510 58.9 2000 2480 147 2001 2260 222 2002 2250 363 2003 2160 497 2004 2260 677 2005 2160 695 as a share of the total number of exporting �rms. Hence most of the action seems to be in within-industry dynamics. In the short run, by far the major contribution to export growth came from incumbents exporting more of the same products to the same markets, but there are high rates of turnover in �rms, products and destinations that become very important in the long run. Almost 60 percent of export growth was due to incumbent �rms exporting products that were not originally exported in 1997. Most of this number is due to surviving new exporters who actively add new products and drop weaker ones. Overall, we �nd that the country’s export performance was primarily limited by the inability of �rms to survive the test of exporting. In contrast, the usual suspected obstacles to export growth, such as the inability of small �rms to enter exporting activities or to grow their exports, do not seem to be the binding constraints. In fact, new exporting �rms experienced the fastest growth in their export values, so that over the long run they contribute almost as much to overall export growth as incumbents. 560 THE WORLD BANK ECONOMIC REVIEW T A B L E 9 c . Total Sales of 1999 Cohort Surviving Exporters: Initial Products Dropped (millions of US $) Initial Products Initial Products Total Surviving 99-05 Dropped Before 05 1999 132 25.5 157 2000 284 39.6 324 2001 217 23.6 241 2002 290 7.39 298 2003 357 1.04 358 2004 441 .563 442 2005 406 406 T A B L E 9 d . Total Sales of Incumbent Exporters: Continuing Products Dropped (millions of US $) Continuing Products Continuing Products Total Surviving 99-05 Dropped Before 05 Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 1999 2180 327 2510 2000 2230 250 2480 2001 2060 205 2260 2002 2070 180 2250 2003 2100 61.5 2160 2004 2230 27.5 2260 2005 2160 2160 APPENDIX T A B L E A 1 . Broad Industrial HS Classi�cation Used in Table 5. 1 Animal and Animal Products 2 Vegetable Products 3 Food stuffs 4 Mineral Products 5 Chemicals and Allied Industries 6 Plastics / Rubbers 7 Raw Hides, Skins, Leather, and Furs 8 Wood and Wood Products 9 Textiles / Apparel 10 Footwear / Headgear 11 Stone / Glass 12 Metals 13 Machinery / Electrical 14 Transportation 15 Miscellaneous 16 Service ´guez-Clare, and Xu Lederman, Rodrı 561 REFERENCES Alvarez, Roberto, and Rodrigo Fuentes. 2009. “The Quality of Trade: Unit Value Dynamics.� Mimeo. Central Bank of Chile and Department of Economics, Catholic University of Chile, Santiago, Chile. Besedes, T., and T.J. Prusa 2010. “The Role of the Extensive and Intensive Margins and Export Growth.� Journal of Development Economics, forthcoming, Besedes, Tibor, and Thomas Prusa. 2006. “Ins, Outs and the Duration of Trade.� Canadian Journal of Economics 104(1): 635– 54. Brenton, Paul, Christian Saborowski, and Erik Von Uexkull. 2010. “What Explains the Low Survival Rate of Developing Country Export Flows?� The World Bank Economic Review, forthcoming Dunne, Timothy, Mark J. Roberts, and Larry Samuelson. 1989. “The Growth and Failure of U.S. Manufacturing Plants.� The Quarterly Journal of Economics 104(4): 671–98. Eaton, Jonathan, Marcela Eslava, Maurice Kugler, and James Tybout. 2007. “Export Dynamics in Colombia: Firm-Level Evidence.� NBER Working Paper 13351, National Bureau of Economic Research, Inc., Cambridge, MA. ´ L. Guasch, and J. Humberto Lo Pablo Fajnzylber, Jose ´ pez, editors. 2009. Does the Investment Climate Matter? Microeconomic Foundations of Growth in Latin America. World Bank: Washington, DC. Freund, Caroline, and Martha Denisse Pierola. 2008. “Export Entrepreneurs: Evidence from Peru.� World Bank Policy Research Working Paper 5407, Washington, DC. Downloaded from wber.oxfordjournals.org by guest on October 18, 2011 Lederman, Daniel, Marcelo Olarreaga, and Lucy Payton. 2010. “Export Promotion Agencies: Do They Work?� Journal of Development Economics 91: 257–65. Melitz, Marc. 2003. “The Impact of Trade on Intra-Industry Reallocations and Aggregate Industry Productivity.� Econometrica 71(6): 1695–1725. Noland, Marcus, and Howard Pack. 2003. Industrial Policy in an Era of Globalization: Lessons from Asia. Institute for International Economics: Washington, DC. Ozden, Caglar. 2006. “Caribbean Basin Initiative Bene�ciary Countries and the Apparel Sector: Same Preferences, Different Responses.� Mimeo, Development Research Group, The World Bank, Washington, DC. Ozden, Caglar, and Gunjan Sharma. 2006. “Price Effects of Preferential Market Access: Caribbean Basin Initiative and the Apparel Sector.� World Bank Economic Review 20(2): 241–260. Pack, Howard. 1997. “The Role of Exports in Asian Economic Development and Lessons from Latin America.� In Pathways to Growth, edited by N. Birdsall, and F. Jaspersen. Baltimore: Johns Hopkins University Press. Volpe Martinicus, Christian, and Jero ´ nimo Carballo. 2008. “Beyond Average Effects: The Distributional Impacts of Export Promotion Programs in Developing Countries.� Mimeo. Inter-American Development Bank, Washington, DC. Forthcoming papers in THE WORLD BANK ECONOMIC REVIEW • When Should We Worry about Inflation? Raphael Espinoza, Hyginus Leon, and Ananthakrishnan Prasad • Is Economic Volatility Detrimental to Global Sustainability? Yongfu Huang • How to Deal with Covert Child Labor and Give Children an Effective Education, in a Poor Developing Country Alessandro Cigno • Empirical Evidence on Satisfaction with Privatization in Latin America Céline Bonnet, Pierre Dubois, David Martimort, and Stéphane Straub • The Discriminatory Nature of Specific Tariffs Sohini Chowdhury • Skills, Exports, and the Wages of Seven Million Latin American Workers Irene Brambilla, Rafael Dix-Carneiro, Daniel Lederman, and Guido Porto • Resource Windfalls and Emerging Market Sovereign Bond Spreads: The Role of Political Institutions Rabah Arezki and Markus Brückner THE WORLD BANK 1818 H Street, NW Washington, DC 20433, USA World Wide Web: http://www.worldbank.org/ E-mail: wber@worldbank.org