Targeted Aid and Capture in Development Projects more

Targeted Aid and Capture in Development Projects Matthew S. Winters University of Illinois at Urbana-Champaign mwinters@illinois.edu July 2011 Previous versions of this paper have been presented at the 2009 American Political Science Association Annual Meeting, the 2009 International Political Economy Society Annual Meeting, the 2010 Midwest Political Science Association Annual Meeting, the 2010 PoliticalEconomy of International Organizations Conference, Brown University and the University of Iowa. For comments and useful discussions, thanks to Kate Baldwin, Sarah Bermeo, Jake Bowers, Jose Cheibub, Xinyuan Dai, Simone Dietrich, David Epstein, Macartan Humphreys, Kate Ivanova, Phil Keefer, Dan Nielson, Laura Paler, Pablo Pinto, Dane Rowlands, Thania Sanchez, Alex Scacco, Gisela Sin, Tracy Sulkin, Udaya Wagl´, Rebecca Weitz-Shapiro, e Cara Wong, Jong-Sung You and members of the IGERT Program on International Development and Globalization at Columbia University. Thanks to Amanda Cronkhite for valuable research assistance. Thanks to the Niehaus Center for Globalization and Governance at Princeton University and the Mellon Fellows program at the Institute for Social and Economic Policy and Research at Columbia University for support. 1 Abstract If development projects are to be effective, a minimum requirement is that the funding reaches its intended destination. Yet the history of international development is replete with examples of this not happening. I propose that there will be fewer problems with corruption or other diversions of funding — which I jointly label capture — in more precisely targeted projects. More well-defined targeting results in end-user constituencies that have a clearer sense of the lines of responsibility for a project, more information about project outputs and lower collective action costs. I use an original cross-country, cross-project dataset on the incidence of capture in World Bank investment projects to explore the theory. The data show a negative relationship between targeting and capture, and I demonstrate that this relationship is robust to the inclusion of a set of potentially confounding country- and project-level covariates. In addition, at the country level, I find that there is a higher likelihood of capture in projects in those countries perceived as more corrupt according to commonly used survey-based measures from Transparency International and the Worldwide Governance Indicators, cross-validating those measures and my own. 2 “The only poverty that we have alleviated has been that of those in power who have plundered [World] Bank-funded projects along with their national treasuries and anything else they could get their hands on”.1 In January 1996, the World Bank approved a $115 million credit to Kenya as the Urban Transport Infrastructure Project, intended to improve roads in 26 different cities across the country (World Bank Project ID P001319). The project was to fund routine maintenance on 2,400 km of roads and rehabilitation of an additional 400 km. By the end of the project, however, little of this work had been accomplished, and auditors discovered “multiple indicators of fraud and corruption” in the project.2 As a result of bid-rigging in the project, three World Bank staff members were fired and 11 companies temporarily barred from bidding on Bank projects. The Bank’s review of the project concluded that “most of the outputs that were anticipated [have not] been accomplished and the project objectives [have not] been fully met.”3 Also in January 1996, the World Bank approved a $50 million credit to Kenya to fund improvements to 500 km of the road running from Nairobi to Mombasa (World Bank Project ID P035691.) This project concluded at approximately the same time as the Urban Transport Infrastructure Project, yet “the reconstruction and strengthening of the Nairobi-Mombasa road [was] rated highly satisfactory,” and the Bank congratulated itself by saying, “The El Nino rains destruction of this road would have resulted into [sic] a major economic crisis in the East African region had this project not come in time.”4 The problems with corruption that resulted in project failure in the first case appear not to have been present in the second case Why do we see this variation in project outcomes for two development projects funded by the same international donor over the same period of time in the same country? I argue that the more-focused design of the second project facilitated the existence of superior Steve Berkman. The World Bank and the Gods of Lending. Sterling, VA: Kumarian Press, 2008, p. 22. World Bank Report 34061, p. 8. 3 World Bank Report 34061, p. 19. 4 World Bank Report 31525, p. 4. 2 1 3 accountability relationships such that corruption was less likely to be a problem in the project. The key claim of this paper — which I will examine using original data from almost 600 World Bank-funded investment projects – is that more precisely targeted development projects will be less likely to suffer from problems with corruption or other forms of capture. This argument opens up a new dimension in the study of the relationship between accountability and corruption. In important literature from the last decade, political scientists and economists have examined the ways in which corruption may correlate with different political institutions5 or be affected by broad bureaucratic reforms.6 There has been less attention, however, to the relationship between the design of specific projects and the levels of corruption they experience. Building on the notion that “clarity of responsibility” creates incentives for politicians to avoid abusing public office and to combat bureaucratic corruption,7 I argue that the specificity of projects creates incentives for governments to be more faithful and better implementers. My argument ultimately implies that, when considering the strength of accountability relationships, we can and should look beyond overarching government institutions to also look at the design of particular government projects and programs. I test the argument in the context of World Bank-funded projects; however, it is worth noting at the outset that there is nothing limiting the scope of the theory to foreign aid See, for example, Roger B. Myerson. “Effectiveness of Electoral Systems for Reducing Government Corruption: A Game-Theoretic Analysis.” Games and Economic Behavior 5.1 (1993), pp. 118–132; John Gerring and Strom C. Thacker. “Political Institutions and Corruption: The Role of Unitarism and Parliamentarism.” British Journal of Political Science 34.2 (2004), pp. 295–330; Jana Kunicova and Susan Rose-Ackerman. “Electoral Rules and Constitutional Structures as Constraints on Corruption.” British Journal of Political Science 35 (2005), pp. 573–606; Eric C.C. Chang and Miriam A. Golden. “Electoral Systems, District Magnitude and Corruption.” British Journal of Political Science 37 (2007), pp. 115–137; Margit Tavits. “Clarity of Responsibility and Corruption.” American Journal of Political Science 51.1 (2007), pp. 218–229. 6 See, for example, Caroline Van Rijckeghem and Beatrice Weder. “Bureaucratic Corruption and the Rate of Temptation: Do Wages in the Civil Service Affect Corruption and By How Much?” Journal of Development Economics 65.2 (2001), pp. 307–331; Rafael Di Tella and Ernesto Schargrodsky. “The Role of Wages and Audigint During a Crackdown on Corruption in the City of Buenos Aires.” Journal of Law and Economics 46 (2003). 7 G. Bingham Powell and Guy D. Whitten. “A Cross-National Analysis of Economic Voting: Taking Account of the Political Context.” American Journal of Political Science 37.2 (1993), pp. 391–414; G. Bingham Powell. Elections as Instruments of Democracy. New Haven: Yale University Press, 2000; Tavits, “Clarity of Responsibility and Corruption.” 5 4 projects alone; among distributive projects that are of purely domestic origin, specificity should similarly facilitate superior government implementation (e.g. funding construction of a single bridge versus funding a nationwide project of unspecified small-scale infrastructure improvements). In the next section, I describe the logic of why project specificity will reduce capture. Then I present the new dataset that I will use to examine this hypothesis. After defining the key explanatory variable, I show the empirical results. Why Targeting Makes Capture Less Likely As is well established in comparative politics, accountability refers to citizens’ ability to identify those responsible for policy decisions and to punish those individuals or aggregate groups (e.g. political parties) if their performance is found lacking.8 Given the different levels of accountability in different systems, a large body of work has proposed that different institutional rules will lead to different types of policy outcomes.9 Other work exploring accountability relationships links different institutional rules to the overall quality of governance in countries.10 In her article “Clarity of Responsibility and Corruption,” Margit Tavits applies the pioneering work of G. Bingham Powell and Guy Whitten11 to the area of corruption. She argues that “political corruption depends on the effectiveness of the democratic process, i.e., the ability of voters to monitor their representatives, to detect those responsible for unsatisfactory outcomes, and to hold them accountable by voting them out of office” (2007: 218). She Adam Przeworski, Susan C. Stokes, and Bernard Manin. Democracy, Accountability, and Representation. New York: Cambridge University Press, 1999; Torsten Persson and Guido Tabellini. The Economic Effects of Constitutions. Cambridge, MA: MIT Press, 2003. 9 See the review in Persson and Tabellini, The Economic Effects of Constitutions, Ch. 2. 10 Torsten Persson, Gerard Roland, and Guido Tabellini. “Separation of Powers and Political Accountability.” Quarterly Journal of Economics 112 (1997), pp. 1163–1202; Gerring and Thacker, “Political Institutions and Corruption: The Role of Unitarism and Parliamentarism”; Kunicova and Rose-Ackerman, “Electoral Rules and Constitutional Structures as Constraints on Corruption”; Chang and Golden, “Electoral Systems, District Magnitude and Corruption.” 11 Powell and Whitten, “A Cross-National Analysis of Economic Voting: Taking Account of the Political Context”; Powell, Elections as Instruments of Democracy. 8 5 then argues that certain constellations of political institutions create more clarity of responsibility, which facilitates this monitoring, detection and use of accountability mechanisms. My argument here is similar, except that instead of focusing on political institutions, I focus on the design of development projects and argue that certain types of projects better facilitate monitoring, detection and the use of accountability mechanisms. This should be true in both democracies and non-democracies, where the difference is in the type of accountability mechanisms available to citizens. In particular, the argument here is that development projects designed to benefit a more specific constituency will be less likely than broader development projects to suffer from capture.12 This specific constituency might be geographic (e.g. targeting an infrastructure program at a single city rather than multiple cities around a country) or social (e.g. targeting a cash transfer program at a small subset of the population rather than making it a nationwide program). Three mechanisms explain why precision targeting yields more monitoring, a greater threat of detection and an increased likelihood of the use of accountability mechanisms, which in turn reduce corruption. First, more specific projects have clearer lines of accountability within the government, making it easier for a constituency to identify those responsible for poor performance and lodge complaints about problems. Second, it is easier for a constituency to monitor outputs received in a more delimited project. Third, it is easier I use the term “capture” to refer to the diversion of funds. Corruption is generally defined as “the use of public office for private gain” (Joseph Nye. “Corruption and Political Development: A Cost-Benefit Analysis.” American Political Science Review 61.3 [1967], pp. 412–427; World Bank. Helping Countries Combat Corruption: The Role of the World Bank. Washington, D.C.: The World Bank Group, 1997), which is a narrower concept than the behavior under study here, where funds may be diverted for private gain or by the government for political gain. For a similar use of the term, see Ritva Reinikka and Jakob Svensson. “Local Capture: Evidence from a Central Government Transfer Program in Uganda.” Quarterly Journal of Economics 119.2 (2004), pp. 679–705. It is used elsewhere in the political science and economics literatures to refer to special interest groups dominating policy making (e.g. Jean-Jacques Laffont and Jean Tirole. “The Politics of Government Decision Making: A Theory of Regulatory Capture.” Quarterly Journal of Economics 106 [1991], pp. 1089–1127; Pranab Bardhan and Dilip Mookherjee. “Capture and Governance at Local and National Levels.” American Economic Review 90.2 [2000], pp. 135–139; Jakob Svensson. “When is Foreign Aid Policy Credible? Aid Dependence and Conditionality.” Journal of Development Economics 61 [2000], pp. 61–84) or to the government establishing a monopoly on rents in a particular arena (e.g. Robert F. Rollison. “Rent Seeking: A Survey.” Kyklos 35.4 [1982], pp. 565–602; Era Dabla-Norris and Elisabeth Paul. “What Transparency Can Do When Incentives Fail: An Analysis of Rent Capture.” International Monetary Fund Working Paper WP/06/146. 2006). 12 6 for smaller groups to overcome collective action problems and organize for the purposes of monitoring, detection and accountability. First, more delimited projects generally have clearer lines of accountability. If a project is supposed to go to a specific area, then the set of political and bureaucratic representatives responsible for ensuring the project’s delivery may be clearer.13 Where the lines of responsibility are clearer, capture or corruption is a less likely outcome because individual politicians or individual bureaucrats can more easily be held accountable and therefore are more likely to experience consequences in the event of improper or incomplete project implementation.14 These consequences might be electoral (i.e. citizens not voting for politicians associated with a poorly-implemented project), contentious (i.e. citizens protesting) or even personal (i.e. citizens using social sanctions against a local politician or bureaucrat because of malfeasance). The idea that local-level accountability is more meaningful or effective in development projects has been demonstrated in studies which find that within-village targeting of anti-poverty programs is superior to between-village allocations of program resources.15 Second, smaller projects are simply easier to monitor. If goods and services are supposed to arrive in a single city or a single province, the local population has more opportunity to tell whether or not they actually have arrived as compared to knowing the fate of deliverables in a nationwide project. When goods and services are to be spread across multiple cities or multiple provinces, it is easier for a government or implementing agency to claim that they are being delivered elsewhere — that the project is being implemented even though some This will not be true in every case. A project targeted at a single city might have a large number of components with an array of implementing agencies. 14 Powell and Whitten, “A Cross-National Analysis of Economic Voting: Taking Account of the Political Context”; Paul Seabright. “Accountability and Decentralisation in Government: An Incomplete Contracts Model.” European Economic Review 40 (1996), pp. 61–89; David Samuels. “Presidentialism and Accountability for the Economy in Comparative Perspective.” American Political Science Review 98.3 (2004), pp. 425– 436; Tavits, “Clarity of Responsibility and Corruption”; Mariano Tommasi and Federico Weinschelbaum. “Centralization vs. Decentralization: A Principal-Agent Analysis.” Journal of Public Economic Theory 9.2 (2007), pp. 369–389. 15 Emanuela Galasso and Martin Ravallion. “Decentralized Targeting of an Antipoverty Program.” Journal of Public Economics 89.4 (2005), pp. 705–727; Pranab Bardhan and Dilip Mookherjee. “Pro-Poor Targeting and Accountability of Local Governments in West Bengal.” Journal of Development Economics 79.2 (2006), pp. 303–327. 13 7 subset of citizens is not seeing the results. Such a claim is more likely to be debunked when made about a finely-targeted project. Referring to the two cases at the beginning of the paper, the Kenyan government could easily claim that the funds for the first project were being used elsewhere if one particular locale complained about the lack of outputs, whereas the intended destination of funding in the Nairobi-Mombasa road project was much clearer and therefore more easily monitored.16 Third, with regard to group organization, one of the essential insights of collective action theory is that smaller groups are easier to organize.17 Participating in an action in pursuit of a collective goal is costly; therefore, there is a baseline incentive for individuals to refrain from participating in the action even if they would benefit from achieving the collective goal. As group size shrinks, however, it becomes easier to convince people to participate in pursuit of the common goal, either because the benefits begin to outweigh the costs of participation or else through the provision of selective incentives or the threat of sanctions. Development projects are rarely targeted at so small a constituency that I expect there to be extensive organization in advance of problems in the project.18 However, in the event that problems do arise in a development project, insofar as that project is targeted at a more specific constituency, a response to those problems from the affected citizens is more likely. In part, this is for organizational reasons: it is easier to hold a meeting of affected people in a single city; it is easier to saturate local news media with a local issue; it is easier to get in contact with local political representatives; it is easier to hold a local protest. As the size of the group that needs to be organized becomes smaller, organization becomes easier. In addition, individual citizens are more likely to identify with their local community. Even in large-scale projects, it may be possible to provide local information. Reinikka and Svensson demonstrate how a newspaper campaign in Uganda that provided local information about expected resource transfers from a nationwide program drastically reduced the amount of capture in the program.(Ritva Reinikka and Jakob Svensson. “Fighting Corruption to Improve Schooling: Evidence from a Newspaper Campaign in Uganda.” Journal of the European Economic Association 3.2-3 [2005], pp. 259–267) 17 Mancur Olson. The Logic of Collective Action: Public Goods and the Theory of Groups. 2nd Edition. Cambridge, MA: Harvard University Press, 1971. 18 There may be organization to get the project. I believe that this is less true in the empirical case with which I deal here — foreign-funded aid projects — but in the case of domestic distribution projects, it is quite likely. 16 8 This means that they will be more outraged if they know that something has been taken from them locally or if money from which they or those near them were going to benefit instead has been redirected elsewhere or lost to corruption.19 Corruption in a broader project is less likely to evoke the same emotional sentiment among those who might take action. Importantly, this organization among the end-users does not necessarily have to come to fruition. (Insofar as I look empirically for a lower incidence of capture among well-targeted projects, I, in fact, assume that the threat of organization decreases project problems ex ante.) The fact that the government anticipates such organization in the event of problems will either inhibit the government from redirecting resources or else will make the government a more effective monitor of project implementation.20 In precisely-targeted projects, the threat of organization constrains the government — the threat of a reaction induces ex ante accountability, leading to less capture.21 In more specifically targeted projects, the lines of responsibility for the project will be clearer; the costs of gathering information about project implementation will be lower; and the costs of collective action among the end-user constituency will be lower. All three characteristics make it easier to monitor project implementation, detect problems in the project and invoke accountability relationships when necessary.22 Therefore, I expect governments Peter A. Furia. “Global Citizenship, Anyone? Cosmopolitanism, Privilege and Public Opinion.” Global Society 19.4 (2005), pp. 331–359; Cara J. Wong. Boundaries of Obligation in American Politics: Geographic, National and Racial Communities. New York: Cambridge University Press, 2010. 20 The idea that the government is both constrained from diverting resources and also from shirking on monitoring bureaucrats overlaps with the idea that “politicians are responsible for not only their own corrupt activities but also for the failure to combat bureaucratic corruption.” (Tavits, “Clarity of Responsibility and Corruption,” p. 220) (Gerring and Thacker, “Political Institutions and Corruption: The Role of Unitarism and Parliamentarism,” See also). 21 This is the same logic that holds across studies of the correlation between political institutions and corruption. In those studies, the logic is that politicians fear the risk of being held accountable (and losing office) to a greater extent in some systems as compared to others, which then leads to less ex ante corruption in those systems. Tavits writes, “[I]n a cross-national context, given the credible threat of punishment in high-clarity countries, one would expect that the overall levels of corruption are kept low” (2007: 221). Reinikka and Svensson make a similar point at a more local level with regard to the newspaper campaign that reduced capture of education funding in Uganda. They write, “Importantly, since an actual protest and the threat of voice may have discouraged the local political elite from diverting resources intended for the schools, in equilibrium, there is no reason to believe that the incidence of voice and local diversion of funds should be correlated.” (Ritva Reinikka and Jakob Svensson. “The Power of Information in Public Services: Evidence from Education in Uganda.” Journal of Public Economics 95 [2011], pp. 956–966, p. 959) 22 It is worth noting that even if the precision of targeting is the same across particular projects, it may be 19 9 both to be less likely to divert money from targeted projects and also to take more care in implementing targeted projects. In either case, we should observe less capture among more precisely targeted projects. Looking for Capture and Corruption in World BankFunded Development Projects In the literature so far, there is not comparable cross-country data on capture in development projects. A number of authors have provided descriptions of instances of corruption in individual projects.23 In a few cases, the study of corruption has been systematized within a given project. For instance, Reinikka and Svensson study capitation grants to village schools in Uganda using an expenditure tracking survey and find that villages received, on average, only 13 percent of allocated funds before changes were made to give local end-users more information.24 Olken studies a World Bank-funded community-driven development project in Indonesia, measuring the actual labor and materials used in road construction against what was reported on invoices and labor records.25 Sometimes, estimates of the overall the case that there is variation in the organizational capacity of the groups at which the projects are targeted (Carol Graham. “From Safety Nets to Social Sector Reform: Lessons from the Developing Countries for the Transition Economies.” Social Development in Latin America. Ed. by Joseph Tulchin and Alison Garland. Boulder: Lynn Reinner, 2000, pp. 71–86; Judith Tendler. “Safety Nets and Service Delivery: What are Social Funds Really Telling Us?” Social Development in Latin America. Ed. by Joseph Tulchin and Alison Garland. Boulder: Lynn Reinner, 2000, pp. 87–115; Rebecca Weitz-Shapiro. “Partisanship and Protest: The Politics of Workfare Distribution in Argentina.” Latin American Research Review 41.3 [2006], pp. 122–147). For instance, a largely middle-income city is likely to be more equipped to organize for monitoring, detection and accountability than is a largely poor city, even though the level of targeting (i.e. one city) is the same. For the current cross-national project, I acknowledge but set aside this source of variation and instead look only at variation in the level of targeting. In other empirical contexts, it will be more possible to assess the relationship between targeting, group strength and capture. 23 See, for example, Robert Klitgaard. Tropical Gangsters: One Man’s Experience with Development and Decadence in Deepest Africa. New York: Basic Books, 1991; Graham Hancock. The Lords of Poverty: The Power, Prestige, and Corruption of the International Aid Business. New York: Atlantic Monthly Press, 1994; Berkman, The World Bank and the Gods of Lending. 24 Reinikka and Svensson, “Local Capture: Evidence from a Central Government Transfer Program in Uganda.” 25 Benjamin A. Olken. “Monitoring Corruption: Evidence from a Field Experiment in Indonesia.” Journal of Political Economy 115.2 (2007), pp. 200–249. 10 amount of development funds being lost in a particular country have been produced.26 Until now, however, no one has systematically compiled a dataset of corruption across multiple development projects or across multiple countries. In the dataset that I use here, I derive the coding of the dependent variable from an official World Bank report that is available across numerous development projects from all of the countries to which the World Bank lends. The document describes the outputs and sometimes outcomes from the project and describes the functioning of various processes during project implementation. In this section, I first describe at a general level the types of behaviors that fall under the definition of capture. Then I describe the Implementation Completion Report and how I make use of the document to assign an outcome coding to each project for whether or not there is evidence of capture. Corruption is commonly defined as the misuse of public office or public power for private gain.27 Common forms of corruption include bribe-taking either to accomplish or expedite official duties, collusion with goods- or labor-suppliers for kickback payments, the manipulation of wage payments that allow an official to pocket the difference between reported and paid wages and the inflation of labor or goods expenditures that likewise allow an official to pocket the difference between reported and actual amounts.28 In any of these cases, corruption is harmful because it means that the anticipated quantity (or quality) of goods and services does not reach the impoverished end-users. Many of these types of corruption occur within an implementing bureaucracy. In addition, it is possible that a government will unfairly discriminate in favor of certain beneficiaries when distributing development funds.29 While not corruption per se, this is a Jeffrey A. Winters. “Criminal Debt.” Reinventing the World Bank. Ed. by Jonathan R. Pincus and Jeffrey A. Winters. Ithaca, NY: Cornell University Press, 2002, pp. 101–130; Berkman, The World Bank and the Gods of Lending. 27 Susan Rose-Ackerman. Corruption and Government: Causes, Consequences and Reform. New York: Cambridge University Press, 1999; Jakob Svensson. “Eight Questions about Corruption.” Journal of Economic Perspectives 19.3 (2005), pp. 19–42. 28 Benjamin A. Olken. “Corruption and the Costs of Redistribution: Micro-Evidence from Indonesia.” Journal of Public Economics 90.4 (2006), pp. 853–870. 29 This is a problem in particular with external funding. Bueno de Mesquita and Smith propose that certain types of governments will make use of foreign aid to satisfy the demands of particular constituencies.(Bruce 26 11 form of capture that diverts development finance from its intended destination. For instance, in a national transportation infrastructure project, if the government chooses, despite need in other areas, to allocate funding to districts that are a source of electoral support or home to co-ethnics with the ruling party, this favoritism implies that development aid is being captured in a way that prevents the resources from reaching their intended recipients. Another form of capture occurs when the government reallocates foreign aid funds for an alternative purpose from that for which they were intended. Although some scholars are concerned with indirect fungibility — where the presence of international assistance frees up other resources within a national budget to be used for alternative purposes30 — I only consider direct fungibility to be capture. Direct fungibility implies a reallocation of the actual transfer to some new use.31 In all three of these scenarios — bureaucratic corruption, biased selection of beneficiaries and the direct diversion of funding — development funds are not reaching their intended destination. This is a necessary condition to say that capture has occurred. In addition, there is a sufficient condition: the failure of the money to reach its intended destination must have been the result of a purposeful act. Significant amounts of development financing are simply wasted because of inefficiencies and the lack of bureaucratic capacity in developing Bueno de Mesquita and Alastair Smith. “A Political Economy of Aid.” International Organization 63.2 [2009], pp. 309–340) Morrison describes how foreign aid generally helps governments remain in power.(Kevin Morrison. “Oil, Nontax Revenue, and the Redistributional Foundations of Regime Stability.” International Organization 63.1 [2009], pp. 107–138) 30 Howard Pack and Janet Rothenberg Pack. “Foreign Aid and the Question of Fungibility.” Review of Economics and Statistics 75.2 (1993), pp. 258–265; Tarhan Feyzioglu, Vinaya Swaroop, and Min Zhu. “A Panel Data Analsysis of the Fungibility of Foreign Aid.” World Bank Economic Review 12.1 (1998), pp. 29–58; Vinaya Swaroop and Shantayanan Devarajan. “The Implications of Foreign Aid Fungibility for Development Assistance.” World Bank Policy Research Working Paper No. 2022. Washington, D.C., 1998; Vinaya Swaroop, Shantayanan Devarajan, and Andrew Sunil Rajkumar. “What Does Aid to Africa Finance?” World Bank Policy Research Working Paper No. 2092. Washington, D.C., 1999; Dominique P. Van de Walle and Dorothyjean Cratty. “Do Donors Get What They Paid For? Micro Evidence on the Fungibility of Development Project Aid.” World Bank Policy Research Working Paper 3542. Washington, D.C., 2005; Dominique P. Van de Walle and Ren Mu. “Fungibility and the Flypaper Effect of Project Aid: Micro-evidence for Vietnam.” Jorunal of Development Economics 84.2 (2007), pp. 667–685. 31 So even if the presence of an internationally-financed irrigation project implies that the government has redirected other (domestic) funds away from the agricultural sector, farmers still may get everything to which they are explicitly entitled under the irrigation project, in which case none of the project money per se has been captured. 12 countries, and although this is an unfortunate loss that involves intended beneficiaries not receiving certain goods and services, it is not a purposeful undertaking in the way that capture is. Capture involves the conditions established by the international donor being willfully skirted. Therefore, in assessing whether or not capture has occurred in World Bank projects, I pay attention both to the diversion of funds and to the intentionality of the act. The Implementation Completion Report The Implementation Completion Report (ICR, more recently known as the Implementation Completion and Results Report) is the World Bank’s main mechanism for reviewing project operations and effectiveness. The Bank describes the reports in the following fashion: When a project is completed and closed at the end of the loan disbursement period ... the World Bank and the borrower government document the results achieved; the problems encountered; the lessons learned; and the knowledge gained from carrying out the project. A World Bank operations team compiles this information and data in an Implementation Completion and Results Report, using input from the implementing government agency, co-financiers, and other partners/stakeholders. The report, prepared by Bank operational staff, is submitted to the Bank’s Board of Executive Directors for information purposes.32 It is important to note that these reports are prepared by operational staff — often staff involved in the project. Therefore, there may be incentives to downplay problems within projects. The reports, however, are reviewed by the World Bank’s Independent Evaluation Group (IEG, known as the Operations Evaluation Department until 2006), which is external to the Bank’s operational staff.33 World Bank, ”How We Measure Results,” http://go.worldbank.org/WERUQ6XI10. All ICRs undergo desk review by the IEG. Currently, one in four projects also undergoes a field review by the IEG, which results in a new evaluation document called the Project Performance Assessment Report (PPAR, formerly known as the Project Performance Audit Report). I do not make use of PPARs in this dataset, although they may provide a useful resource in the future for checking the validity of the capture coding. 33 32 13 ICR-like instruments have been in use since 1973, and the Bank produced general guidelines for these reports in 1977.34 In 1994, the instruments were renamed and standardized in response to NGO criticisms of Bank operations.35 New guidelines for ICR preparation were released five years later in 1999. Beginning in August 2001 — as part of a revised Bank Disclosure Policy — ICRs became immediately available to the public upon their completion.36 Only select ICRs from before that date are available, although their number has increased substantially with the Bank’s recent change to its Access to Information Policy that took effect in July 2010.37 ICRs include performance ratings for the Bank, the national government and the projectimplementing agency; a description of the quality of the project at entry; a description of project outputs and outcomes; an analysis of the economic and financial rates of return; an assessment of the project’s institutional impact; a discussion of the factors that influenced the project’s outcome; and a list of the lessons learned from the project. Coding ICRs for Evidence of Capture In general, ICRs do not directly describe problems with capture: there is not, for instance, a quantitative measure of the percentage of funds that reached their intended destination. In only a small minority of cases is there direct mention of corruption. For instance, one ICR describes “credible evidence that all consulting contracts may have been rigged” (World Bank Report 34061, 8), which implies a loss of development funds because of inefficient Patrick G. Grasso, Sulaiman S. Wasty, and Rachel V. Weaving, eds. World Bank Operations Evaluation Department: The First 30 Years. Washington, D.C.: The World Bank, 2003; Mervyn L. Weiner. “Institutionalizing the Evaluation Function at the World Bank, 197584.” World Bank Operations Evaluation Department: The First 30 Years. Ed. by Patrick G. Grasso, Sulaiman S. Wasty, and Rachel V. Weaving. The World Bank, 2003, pp. 17–30; Christopher Willoughby. “First Experiments in Operations Evaluation: Roots, Hopes, and Gaps.” World Bank Operations Evaluation Department: The First 30 Years. Ed. by Patrick G. Grasso, Sulaiman S. Wasty, and Rachel V. Weaving. The World Bank, 2003, pp. 3–16. 35 H. Eberhard K¨pp. “Promoting Professional and Personal Trust in OED.” World Bank Operations o Evaluation Department: The First 30 Years. Ed. by Patrick G. Grasso, Sulaiman S. Wasty, and Rachel V. Weaving. The World Bank, 2003, pp. 55–60, pp. 57-58. 36 World Bank, “World Bank Revises Disclosure Policy,” http://go.worldbank.org/L9KE2D2OM0. 37 World Bank, “World Bank Broadens Public Access to Information,” http://go.worldbank.org/L3HF51WOX0. 34 14 selection of contractors. Projects where the ICR describes actual corruption immediately are coded as suffering from capture in the dataset. In most cases, however, it is necessary to look for observable implications of capture — descriptions in the ICR of financial management, audit or procurement problems that imply corruption or descriptions of political interference that imply the diversion of funds. ICRs, for instance, report “allegations of financial mismanagement,”38 “problems with procurement and financial management,”39 a failure to meet “fiduciary and financial management standards,”40 or the delivery of services being “at risk of political interference.”41 I code as instances of capture those projects where there are descriptions of financial management or procurement “problems”; financial management, procurement or auditing processes that were “irregular,” “unsatisfactory” or “not meeting standards”; “late,” “missing” or “incomplete” audits; “non-transparent” government involvement or “political interference”; resources that were “mismanaged” or “improperly utilized.”42 In each case, the description implies development funds not reaching the end-user constituency for which they were intended. However, in all cases where the ICR made clear that the problem was due to a lack of bureaucratic capacity — rather than purposeful deceit — I do not code the project as subject to capture.43 Similarly in cases where fraud or corruption was discovered and then remedied by the government, I do not code the project as subject to capture.44 I also do not code as capture one of the most frequent complaints in ICRs: the lack of counterpart World Bank Report 34384, p. 11. World Bank Report 34513, p. 10. 40 World Bank Report 34745, p. 5. 41 World Bank Report 33541, p. 17. 42 Full coding scheme available on the author’s website. 43 For instance, in the Fourth Social Investment Fund Project in Honduras, the ICR describes “an unsatisfactory rating for financial management” that was “corrected within a few months” after the Bank sent technical assistance missions to improve the implementing agency’s reporting procedures. World Bank Report 25481: p. 23. 44 An example of this is the Health Sector Recovery Program in Mozambique in which there was “fraud of US$300,000 in the Special Account,” but then “[t]he government fired two accountants and repaid the funds.” World Bank Report 26963: p. 16. 39 38 15 funding. Across perhaps a majority of World Bank projects, the government fails to live up to its commitment to provide cofinancing. Although this is certainly detrimental to project success, it is not a diversion of the international funds coming from the World Bank and therefore is not capture per se. I code for capture only in the cases where I find evidence that a World Bank dollar expected to reach a certain destination did not reach that destination. In summary, I code projects as subject to capture when there are • direct descriptions of corruption • negative descriptions of financial management, • negative descriptions of procurement practices, • negative descriptions of audit practices, or • descriptions of direct political interference in allocation decisions unless there also is a clear explanation for the problems in terms of the lack of bureaucratic capacity or a clear and immediate response to the problems by the government. Ultimately, this coding scheme will miss cases of capture — corruption is difficult to detect. As the World Bank’s Nigeria Country Team noted at one point, “Even with much experience handling procurement matters, in some cases it is almost impossible to detect misrepresentation/fraud (estimated at 30-40%)”.45 And where the World Bank is making excuses about a country’s lack of administrative capability, they may actually be masking instances of fraud and corruption. And it is possible that some Bank staff have more or less of an incentive to announce certain problems: for instance, it might be wise to report corruption in a failing project in order to shift blame onto the recipient government. Acknowledging these potential limitations, this dataset nonetheless is the first to include a single measure of capture across a broad set of development projects that span nearly all developing and middle-income countries in the world. 45 Berkman, The World Bank and the Gods of Lending, p. 78. 16 Description of the Data I gathered the set of all Implementation Completion Reports available as of early 2006. Although the ICR has been in use since 1994, it did not become standard Bank practice to disclose them to the public until 2001. Therefore, at the time of data collection, only a limited number of ICRs (38) were available from before 2001. The majority of the data come from projects completed between January 2002 and June 2005. Of 941 total projects — including both investment projects and also programmatic budget support lending — completed in the period 2002-2004, ICRs exist for 748 of them. I check to see if certain project or country characteristics predict whether or not there will be a completed ICR. There was less likely to be an ICR for smaller projects, projects where the World Bank share was smaller and projects in more populous countries. Projects where the main theme was human development were more likely to have an ICR. Importantly, the quality of a country’s governance was not a significant predictor of whether or not there would be an ICR produced for a given project. Table 1 summarizes the number of ICRs in the dataset by the year of project completion, dividing them into investment project loans and “non-project” loans, by which I primarily mean programmatic lending (i.e. budgetary support or structural adjustment lending) or technical assistance lending (i.e. lending in support of improving capacity within national or sub-national government ministries). (I also include in the non-project category the small number of loans that the World Bank made directly to sub-national governments (7), to non-government entities (23) or to multiple countries (10).) In the analysis that follows, I look only at the 598 investment projects in the dataset. Columns (4) and (5) of table 1 show the breakdown by year of the dependent variable: the number of projects suffering from capture. For the period 2002-2005, the proportion of projects each year evincing problems with capture ranges from 15 to 32 percent. Overall, approximately one in five World Bank investment projects in the dataset show evidence of capture. 17 Table 2 shows the reasons (and number of reasons) for the positive capture codings. For half of the cases (62), the ICR only provides one reason for a positive capture coding (although the ICR may describe multiple instances of this single reason — for instance, multiple cases of misprocurement or qualified audits). For over one-third of the cases (48), two reasons for a positive capture coding are found. And for the remaining 12 cases, more than two reasons for a positive capture coding were present. In the ICRs, the World Bank provides an overall ranking of the project outcome as well as rankings of the Bank and borrower’s performance in the project. Table 3 shows that projects coded as experiencing capture can be rated as satisfactory by the Bank — and in fact are more likely to be rated satisfactory than unsatisfactory — and that projects can be rated unsatisfactory by the Bank for reasons other than those covered by the capture coding. Unsatisfactory ratings in projects where I do not find evidence of capture often are related to a lack of counterpart financing or inept — but not criminal — administration of the project such that the project fails to meet its development objectives. Even when we look at the borrower’s performance rating, as in table 4, the ratio of unsatisfactory ratings for projects experiencing capture is relatively unchanged. This suggests that I am observing a phenomenon that the World Bank has not measured with its own ratings system (and that therefore is not analyzed in papers that use the World Bank’s outcome ratings as their dependent variable46 ). As prima facie evidence for the validity of the capture measure, I correlate the capture coding with perceptions of country-level corruption. In table 5, I present the results from linear probability models in which the country’s level of perceived corruption — as measured in column (1) by the Worldwide Governance Indicators’ Control of Corruption index and Daniel Kaufmann and Yan Wang. “Macroeconomic Policies and Project Performance in the Social Sectors: A Model of Human Capital Production and Evidence from LDCs.” World Development 23.5 (1995), pp. 751–765; Jonathan Isham, Daniel Kaufmann, and Lant H. Pritchett. “Civil Liberties, Democracy, and the Performance of Government Projects.” The World Bank Economic Review 11.2 (1997), pp. 219–242; David Dollar and Jakob Svensson. “What Explains the Success or Failure of Structural Adjustment Programmes?” The Economic Journal 110 (2000), pp. 894–917; David Dollar and Victoria Levin. “Sowing and Reaping: Institutional Quality and Project Outcomes in Developing Countries.” World Bank Policy Research Working Paper 3524. Washington, D.C., 2005. 46 18 in column (2) by Transparency International’s Corruption Perceptions Index — predicts whether or not an individual project will suffer from capture. I cluster the standard errors on country-years to account for the fact that there are multiple projects within countries and that the corruption perceptions measures change on an annual basis. For both country-level measures of corruption — according to which higher values imply less corruption — there is a lower likelihood of World Bank projects experiencing capture as the value of the index goes up. The relationship is highly statistically significant.47 In other words, there is a greater incidence of capture in World Bank projects in those countries otherwise judged to be corrupt. This result, therefore, provides cross-validation of both the new measure of capture introduced here and these traditional perceptions-based measures of corruption. In countries that are subjectively perceived as being corrupt, we are more likely to observe a positive coding for capture within a given World Bank project. Insofar as the capture coding is an objective measure of corruption within a country, this correlation suggests the validity of these commonly-used perceptions-based measures of country-level corruption. Targeting and the Incidence of Capture The key explanatory variable for this study is the level of targeting. As described above, projects that are targeted to a more specific set of end-users are expected to suffer from a lower likelihood of capture because of the increased clarity of responsibility within the project and the increased ability and desire of potential beneficiaries to mobilize. The World Bank does not have an explicit set of categories that it uses to describe how its projects are targeted. Therefore, for the World Bank investment projects in the dataset, I classify the targeting of the project using nine categories that correspond to the concentration or diffusion of project outputs. The results are identical in terms of substantive association and significance levels in a logistic regression model using clustered standard errors. 47 19 There are five straightforward geographic codings, where a project is targeted at (1) a single city, (2) multiple cities, (3) a single region or (4) multiple regions or else is (5) national. The rule for choosing between multiple cities and a region was that a project became regional if it was to reach more than half of the cities within a region. A similar rule was used for choosing between multiple regions and the national coding. Two other codings are not strictly geographic but rather involve projects intended for specific types of places within a country. Projects for (6) the rural sector fund agricultural or other rural projects across the whole country (as compared to within a single region), while projects for (7) the urban sector fund urban infrastructure across the whole country (as compared to within a specified set of cities). Some projects are targeted at (8) specific social groups, such as ex-combatants or students enrolled in higher education. And finally some projects have (9) business or industry as the intended end-users; the deliverables in these projects are private goods or private benefits for a specific (type of) business or industry. Figure 1 shows the incidence of capture for each type of targeting. Looking first at the geographic categories, projects targeted at a single city or a single region are less likely to suffer from capture than the average project. (The average rate of capture is 0.20.) Projects targeted at multiple cities or multiple regions or targeted at the entire rural or urban sector, on the other hand, have higher-than-average rates of capture. In addition, projects targeted at social groups or business and industry have very low rates of capture. These two types of projects share characteristics with those targeted at more limited geographic areas: they target delimited constituencies that likely have an increased capacity to organize as a group and an increased level of information about whether goods and services are being delivered. (Whether or not projects targeted at business/industry and social groups have clearer lines of accountability within the government is not clear.) Table 6 splits the data to compare the incidence of capture among projects targeted at single cities, single regions, business/industry or social groups to its incidence among nationwide projects or projects targeted at multiple cities, multiple regions, the rural sector 20 or the urban sector. The crosstab shows that there is a significant difference in the proportion of capture in the more precisely-targeted versus the less precisely-targeted projects. As my theory predicts, there is less capture in the set of more-precisely targeted projects. Based on this relationship, I create an indicator variable called concentrated targeting for those projects that fall into the more precise targeting categories (i.e. single city, single region, business/industry or social group). I use this indicator in a logistic regression model to show that this negative relationship between the precision of targeting and capture is robust to the inclusion of a number of possible confounding covariates.48 Possible Confounders Predicting Both Targeting and Capture To avoid spurious correlation, the regression equation includes other variables that might be correlated with both the level of targeting and the likelihood of observing capture within a project. In terms of project characteristics, with larger projects, which may be less targeted, there is more opportunity for corruption and other forms of diversion. Therefore, I include a variable that measures total project size, as reported in the World Bank’s Project Database. Since the project size variable includes money from the government and other donors, I also examine a specification that controls only for the size of the World Bank’s contribution to the project. In both cases, I take the logarithm of the value. It is possible that World Bank borrowers will be more or less careful depending on whether the money is market-rate borrowing (in which case it is more costly) or concessional lending. If governments treat concessional lending in a less stringent fashion, we might see higher rates of capture in these projects. In addition, there may also be differences in targeting strategies across the World Bank’s concessional and non-concessional branches. Therefore I include Note that this variable coding, although in accordance with the theory, is inductive; it was not a priori clear whether social group and business/industry targeting should have been included in the concentrated targeting category. For the purposes of hypothesis testing and calculating accurate significance values, the inductive coding may be problematic, since information in the data already has been used to define the explanatory variable. Therefore, the results from the regressions shown in the next section remain suggestive rather than definitive. In the future, they can be confirmed using newly coded data (i.e. an out-of-sample test). 48 21 an indicator for whether or not the loan consists solely of market-rate International Bank for Reconstruction and Development (IBRD) funding or whether it is a mix of IBRD and concessional International Development Association (IDA) funding (so-called ‘blend’ loans). (The omitted category is pure IDA funding.) I expect capture to be lower in IBRD-only or blend projects because the money in these projects is more costly for the recipient country. I also control for several country characteristics that might influence the likelihood of capture and the World Bank’s targeting strategy. To provide a measure of the level of corruption in a country, I use the Worldwide Governance Indicators control of corruption measure.49 In alternative specifications, I also use Transparency International’s Corruption Perceptions Index and an aggregate of the six Worldwide Governance Indicators. In all models, I include a variable measuring the quality of democracy (the Freedom House index), since democracy conceivably could have a different effect from governance on targeting and capture. A country’s level of development — typically measured using GDP per capita — may also alter the World Bank’s targeting strategy if, for instance, state capacity is higher in more developed countries such that it is easier to use nationwide projects because the capacity exists to implement them across the entire country; the cross-country evidence also shows that there are fewer problems with corruption in more developed countries.50 The measure of GDP per capita comes from the Worldwide Development Indicators. Since the projects in the dataset span a number of years, whereas most country-level data is recorded by year, I average country-level variables for the life of a project (from its year of approval through its closing year).51 When country-level data is not available for some year of project operation, I take the average over the remaining years. For data where Note that this variable, which is the outcome variable in many political science studies of corruption, is a predictor variable in the current study. This obviates the need to control for many of the common explanations of national-level corruption in the literature. These variables would need to be included only insofar as they confounded the effect of targeting separately from their effect on the overall level of corruption within the country. 50 Daniel Treisman. “The Causes of Corruption: A Cross-National Study.” Journal of Public Economics 76.3 (2000), pp. 399–457. 51 Alternative operationalizations could take the relevant country-level data from the middle year of the project or the average over some smaller period that is more likely to include only actual implementation years and not years during which the project is still in a planning or set-up phase. 49 22 coverage does not begin until a project already is in progress — for example, 1996 is the first year of coverage in the World Governance Indicators dataset, but many of the projects in the dataset begin before 1996 — I make use of available observations, which may mean that the value assigned to a particular project disproportionately measures national conditions from the latter part of the project.52 Thinking about the internal dynamics of the World Bank, it is possible that different departments of the Bank have different standards for reporting problems within the Implementation Completion Reports or different baseline patterns of project targeting. In order to account for this, I include regional fixed effects based on the six World Bank lending regions (Africa, East Asia and the Pacific, Europe and Central Asia, Latin America and the Caribbean, the Middle East and North Africa, and South Asia). The World Bank also classifies projects according to their major sector and major theme, which implies the involvement of different substantive departments within the Bank. It is possible that projects in some sectors or covering some themes are more or less likely to be susceptible to capture and are also more or less likely to be targeted. I group projects according to the first “major sector” reported in the World Bank’s database, which may account for anywhere between 20 and 100 percent of the project’s expected output. Excluding the categories of information/communication and finance — where the total number of projects is small — projects in the industry and trade sector experience the lowest rate of capture (15 percent), whereas those in the water, sanitation and flood prevention category experience the highest (28 percent). In terms of thematic content, projects that aim for social protection or risk management have the lowest proportion of ICRs demonstrating evidence of capture (12 percent), whereas those aiming for social or gender inclusion have the highest (43 percent) (ignoring economic management, given the small number of projects that fall under that theme). I include a set of indicator variables in some regression specifications for project sectors and themes. 52 The Worldwide Governance Indicators dataset also lacks values for the years 1997, 1999 and 2001. For those years, I have used a linear interpolation of the values from the preceding and subsequent year. 23 Table 7 presents the results of these regressions. In the first column of the table, targeting remains a significant negative predictor of capture even after controlling for the general level of corruption in the country, the level of development, the level of democracy, the total project size and the branch of the World Bank from which the project originated. The second and third columns show that the relationship continues to hold using a different measure of the national corruption level or a measure of the overall national governance quality. In the fourth column, I show that if the size of the project is measured by looking only at the World Bank’s contribution, rather than the overall project size, there is little change in the magnitude or significance level of the targeting variable. In the fifth and sixth columns, I add sector and theme fixed effects respectively, and the overall results remain unchanged. In the seventh column, I add country fixed effects and also fixed effects for the year of the ICR; because of collinearity, the country-level variables are estimated with substantially more uncertainty, but the coefficient and significance level of the targeting variable remains the same. Therefore, controlling for a range of possible confounding covariates, we see a consistent, substantively and statistically significant relationship between targeting and capture. Based on the coefficient estimate in column (1) of table 7, the probability of a targeted project experiencing capture is 12.0 percentage points less as compared to a non-targeted project, holding all other variables at their means and setting the IBRD and blend indicators equal to zero.53 Targeting appears to result in a substantial reduction in the likelihood of project capture. I also checked the robustness of the results to alternative definitions of the outcome variable. In table 8, I present results from a series of models in which I change the capture coding to 0 for cases that have been positively coded based on only one of the possible indicators and then rerun the baseline model. If there is any systematic bias in the coding of cases based on particular indicators of capture, then targeting may cease to be a significant 53 The 95 percent confidence interval around the estimated change in probability is [0.06, 0.18], calculated using the delta method. 24 predictor of capture. Across four different definitions of the outcome variable in columns (1) through (4), there is little change in the magnitude or significance of the coefficient on the targeting indicator. In column (5), I show the results of an ordered logistic regression where the outcome variable is coded 0 if there is no evidence of capture, 1 if the ICR contains one indicator of capture and 2 if the ICR contains two or more indicators of capture. The coefficient estimates for this regression are almost identical with those from the regular logistic regression. A change from a non-targeted to a targeted project — holding all other variables at their means and the IBRD and blend indicators at 0 — leads to an equal decrease in the probability of being in either of the two corruption outcome categories (six percentage points), suggesting that there is little gained by treating the outcome variable as ordinal rather than dichotomous. Is There Selection Bias in the Analysis? Previous work has shown that overall levels of World Bank funding and the proportion of subnational aid within World Bank project lending are responsive to the governance characteristics of recipient countries. Insofar as the World Bank has a higher probability of using targeted projects in countries where there are concerns about corruption, this might bias the coefficient estimate on the targeting variable if the factors affecting selection are not included as covariates. Under this scenario, since targeted projects would be more likely to be found in countries where there are greater concerns about corruption, we would be less likely to see the negative relationship between targeting and capture that has been found above. That is, systematic use of targeted projects in high corruption countries should bias the coefficient estimate on targeting downward. However, when I divide the current data between projects in countries with a control of corruption score below the median and projects in countries with a control of corruption score above the median, there actually is a higher proportion of targeted projects within the less corrupt countries. Out of 313 projects in the low corruption countries, 100 are targeted (32 25 percent); out of 285 projects in the high corruption countries, 65 are targeted (23 percent).54 So in fact, in the dataset, there are more targeted projects absolutely and proportionally in low corruption countries, which could artificially induce a negative correlation between targeting and capture. However, as was seen in all of the results above, the correlation is robust even controlling for a country-level corruption measure. In addition, I use an interaction model to see if there is a significantly different effect of targeting on the likelihood of capture in high corruption countries as compared to low corruption countries, which would suggest a bias in the estimation of the average effect reported above. Introducing an interaction term between targeting and control of corruption to the model (1) specification in table 7 results in a very similar coefficient estimate for the effect of targeting (β = −0.91, p < 0.05) and a statistically insignificant coefficient on the interaction term (β = −0.10, p < 0.89). The estimated change in the probability of capture for a country with a control of corruption score at the 10th percentile is a 14 percentage point decrease; for a country with a control of corruption score at the 90th percentile, it is a 9 percentage point decrease. As expected given the insignificant coefficient on the interaction term, the 95-percent confidence intervals for the two estimated changes overlap. Therefore, within the current data, we do not see evidence of the type of selection we would expect (in which more poorly governed countries are more likely to see targeted projects). Controlling for the country’s level of corruption helps us to believe that the estimate on the targeting variable is not being affected by this propensity for the Bank to target projects in less corrupt countries. In addition, an interaction model shows that targeting has a statistically indistinguishable effect on the probability of capture in high and low corruption countries. The high corruption countries receive both a smaller number of projects and also a significantly lower amount of World Bank funds during the time period under study: $24 billion as compared to $47 billion. These numbers are based only on the projects in the dataset and are therefore not comprehensive. 54 26 Conclusion In this paper, I have used an original dataset of capture in World Bank projects to examine the impact of targeting on development project outcomes. The data is the first of its kind — it allows us to compare across projects and across countries to see what factors might make capture more or less likely. Building on the notion of clarity of responsibility, I have argued that capture is less likely when a project is more precisely targeted at a particular constituency. In the data, I find evidence of this: the incidence of capture among projects targeted at single cities, single regions, business and industry or particular social groups is lower than the incidence among nationwide projects or more diffusely-targeted projects. I show that this correlation is robust to the inclusion of potential confounding variables at the country- and project-level. This result is an important addition to the literatures on corruption and accountability. It reveals the way in which project design can create superior accountability relationships that lead to less corruption. Whereas past work has focused on how the structure of government correlates with corruption, this article has looked at the way that the government implements projects (or agrees to implement projects) correlates with corruption. Insofar as political institutions are relatively difficult to change and likely to exist in an equilibrium determined by factors besides the risk of corruption, the institutional literature in some ways does not offer much hope for improving the clarity of responsibility and reducing corruption; I argue that these things can be achieved by thinking about project design. In terms of foreign aid effectiveness, donors could improve aid effectiveness by using more targeted projects in more corrupt countries. But the implications of the theory are not limited to foreign-funded development projects; within countries, locally-financed development projects might also be improved through targeting, such that the government receives better feedback from citizens who are more equipped to monitor, detect and hold accountable because of the clearer and more delimited project design. 27 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 Total Implementation Reports Evidence of Capture (1) (2) (3) (4) (5) (6) Total Non-Project Project No Capture Capture Percent 6 2 4 4 0 0 4 0 4 3 1 25 5 2 3 3 0 0 3 1 2 2 0 0 9 0 9 8 1 11 11 2 9 7 2 22 212 74 138 106 32 23 270 86 184 157 27 15 262 93 169 134 35 21 117 41 76 52 24 32 899 301 598 476 122 20 Table 1: Number of ICRs and Incidence of Capture by Year of Project Completion. Non-project category includes programmatic loans, technical assistance loans and loans to non-government borrowers. Capture codings assigned only to project loans. Capture Across Levels of Targeting 200 Capture No Capture 150 Number of Projects 0.21 0.23 100 0.09 50 0.12 0.14 0.11 0.3 0.31 0.31 0 Bu s Ind ine us ss So try cia lG ro up On eR eg ion On eC ity Na tio na l Mu Re ltip gio le ns Ru ra lS ec tor Mu lt Ci iple tie Ur s ba nS ec tor Figure 1: Incidence of Capture across Targeting Categories. Targeting types are ordered from lowest incidence of capture to highest. or 28 Information Procurement Problems Political Interference Direct Mention of Corruption Financial Mismanagement Audit Problems Total with One Type of Indicator Financial and Procurement Problems Corruption and Financial Problems Financial and Audit Problems Procurement and Audit Problems Corruption and Procurement Problems Financial and Political Problems Other Combinations of Two Indicators Total with Two Types of Indicators Financial, Procurement and Audit Problems Other Combinations of Three Indicators Total with Three Types of Indicators Combinations of Four Types of Indicators Total Cases of Capture Number of Cases 21 14 12 12 3 62 10 6 6 6 5 5 10 48 3 6 9 3 122 Table 2: Indicators That Lead to a Positive Capture Coding. Description of the number and type of indicators that led to the set of positive capture codings. World Bank Project Rating Evidence of Capture Satisfactory Unsatisfactory Total Capture 75 47 122 No Capture 423 53 476 Total 498 100 598 Table 3: Capture Coding and Project Outcome Rating. χ2 statistic for H0 : no relationship between satisfaction ratings and capture — 52.3 (p < 0.00). 29 World Bank Borrower Rating Evidence of Capture Satisfactory Unsatisfactory Total Capture 72 50 122 No Capture 419 57 475 Total 490 107 597 Table 4: Capture Coding and Borrower Performance Rating. χ2 statistic for H0 : no relationship between satisfaction ratings and capture — 58.4 (p < 0.00). WGI TI Control of -0.13∗∗∗ -0.09∗∗∗ (0.04) (0.08) Corruption N 597 505 Table 5: Linear Probability Models Predicting Capture. Robust standard errors clustered on country-year in parentheses. *** - p < 0.01 Capture No Capture Single Cities, Single Regions, 19 146 Businesses or Social Groups (12 percent) (88 percent) Nationwide 103 330 or Other Targeting (24 percent) (76 percent) Total 122 476 (20 percent) (80 percent) Total 165 433 598 Table 6: Capture in Single-City, Single-Region, Business/Industry or Social Group Projects versus Other Types of Projects. χ2 statistic for H0 : no relationship between targeting and capture — 11.08 (p < 0.01). 30 Concentrated Targeting Control of Corruption (WGI) Corruption Perceptions (TI) Average Governance Quality (WGI) Log(GDP per capita PPP) Freedom House (1) -0.95*** (0.27) -1.06*** (0.39) (2) -1.07*** (0.27) (3) -0.95*** (0.27) (4) -0.86*** (0.27) -1.00** (0.40) (5) -0.97*** (0.28) -1.07*** (0.40) (6) -0.95*** (0.28) -1.06** (0.42) (7) -0.94*** (0.31) 1.50 (2.91) -0.65*** (0.19) -1.54*** (0.40) -0.25 (0.26) 0.07 (0.06) 0.08 (0.08) -0.20 (0.30) 0.07 (0.07) 0.07 (0.10) -0.14 (0.24) 0.17** (0.07) 0.12 (0.08) 0.11 (0.10) 0.20 (0.39) -0.56 (0.36) N N N N 581 111 0.07 -276.75 0.80 0.00 0.24 (0.42) -0.48 (0.38) N N N N 494 90 0.09 -233.50 0.79 -0.01 0.14 (0.39) -0.52 (0.38) N N N N 581 111 0.08 -274.53 0.79 -0.01 0.16 (0.39) -0.60* (0.37) N N N N 553 109 0.07 -266.71 0.79 0.00 0.19 (0.38) -0.68* (0.38) Y N N N 579 111 0.08 -272.63 0.80 0.00 0.19 (0.41) -0.67 (0.42) N Y N N 578 111 0.08 -273.12 0.81 0.05 -0.06 (0.55) -0.03 (0.62) N N Y Y 436 60 0.17 -209.64 0.77 -0.11 -0.30 (0.28) 0.07 (0.07) -0.22 (0.27) 0.06 (0.06) 0.12 (0.09) -0.24 (0.28) 0.08 (0.07) 0.09 (0.09) 3.41 (3.92) 0.06 (0.31) 0.17 (0.13) Log(Total Project Size, $US) Log(World Bank Contribution, $US) IBRD Project Blend Project Sector Effects Theme Effects Country Effects ICR Year Effects N J Pseudo-R2 Log-Likelihood Pct Correctly Predicted Prop Reduction in Error Table 7: Logistic Regressions Predicting Capture. Each observation is a project. Robust standard errors clustered on country in parentheses. All models include region effects (with regions defined by the World Bank) and a constant. * - p < 0.10; ** - p < 0.05; *** p < 0.01. 31 Outcome Variable Concentrated Targeting Control of Corruption Log(GDP per Capita PPP) Freedom House (1) (2) No No Financial Procurement -0.89*** -0.92*** (0.28) (0.35) -1.11** (0.44) -0.37 (0.26) 0.06 (0.06) 0.10 (0.09) 0.20 (0.40) -0.32 (0.47) -0.88** (0.38) -0.19 (0.24) 0.04 (0.07) 0.06 (0.09) 0.02 (0.40) -1.80*** (0.43) (3) (4) (5) No No Ordered Audit Political Scale -0.88*** -1.22*** -0.96*** (0.27) (0.31) (0.26) -1.12*** -1.24*** -1.04*** (0.41) (0.40) (0.38) -0.28 (0.26) 0.06 (0.06) 0.09 (0.09) 0.24 (0.40) -0.46 (0.35) -0.10 (0.27) 0.02 (0.07) 0.10 (0.09) 0.13 (0.45) -0.55 (0.36) -0.28 (0.25) 0.06 (0.06) 0.10 (0.09) 0.18 (0.39) -0.61 (0.39) 1.11 (2.57) 2.01 (2.57) 581 111 0.06 -359.53 Log (Total Project Size, $US) IBRD Project Blend Project Cutpoint 1 Cutpoint 2 Y=1 N J Pseudo-R2 Log-Likelihood Pct Correctly Predicted Prop Reduction in Error 107 581 111 0.06 -259.54 0.82 -0.00 100 581 111 0.08 -245.26 0.83 0.01 116 581 111 0.07 -270.40 0.80 -0.00 106 581 111 0.09 -248.90 0.82 -0.00 Table 8: Logistic Regressions Predicting Different Definitions of Capture. Model in column (5) is an ordered logistic regression. Each observation is a project. Robust standard errors clustered on country in parentheses. All models include region effects (with regions defined by the World Bank) and a constant. * - p < 0.10; ** - p < 0.05; *** p < 0.01. 32