Though many studies have referred to an “anti-corruption movement” beginning in the 1990s by major international organizations, none has empirically tested its effectiveness on corruption. The data show that from 1997 onward, the impact of multilateral aid is strongly and robustly associated with lower corruption levels, while bilateral aid is shown to be an insignificant determinant. An increase in any official development assistance (ODA) pre-1997 is associated with higher levels of corruption or has no impact at all. Using panel data from 1986 to 2006, this study reveals a more nuanced relationship between ODA and corruption than previous studies and demonstrates that when disaggregating the time periods, there are sensitive temporal effects of ODA's effect on corruption overlooked by earlier studies, and provides initial evidence of the effectiveness of the international organization anti-corruption movement in the developing world.
In the wake of the “anti-corruption movement” (ACM) of the mid-1990s, has foreign aid assistance (official development assistance, or ODA) had an impact on the corruption levels of recipient countries? Moreover, are there reasons to believe that different types of ODA,1 whether multilateral or bilateral, impact the quality of governance in any systematic way? Though the good governance ACM has been widely discussed, these questions have had surprisingly little empirical attention from literature pertaining to the impact on foreign aid on good governance and corruption. These are of course important questions for scholars and policy makers alike, as the volume of ODA redistributed from the developed to the developing world has increased substantially over the past decade. This complex relationship has drawn the attention of many scholars in recent years and the effects of ODA on “good governance” indicators, whether democracy, bureaucratic quality, or corruption, remains strongly debated in the literature. This study makes a significant contribution in assessing the vastly complex effects of the ACM by theoretically distinguishing and empirically testing the impact of ODA on corruption by disentangling its effects over time and by donor type.
One of the first distinctions made by many studies when looking at the effect ODA has on democracy, economic growth, and so on, is whether the aid is multilateral or bilateral. Bilateral ODA is argued by many to be tied with the political agenda of the donor country and less focused on “good governance” reform in the recipient country for its own sake. Though of course not apolitical, multilateral ODA is seen as relatively more impartial, and the program to fight corruption and improve governance in the developing world has been at the forefront of the agenda of each major Breton Woods organization since the mid-1990s, and thus might be associated with more effective results in curbing corruption.
The second distinction is the time period in which one examines the effect of ODA on corruption. What I argue is that previous studies have overlooked this salient distinction—that in the post-ACM era of the major international organizations (IOs) it is theoretically plausible that we should expect different results than in previous time periods (Armon 2007). It is thus the central contribution of this paper that a more nuanced relationship between ODA and corruption in the developing world is uncovered. Though many scholars have previously found little to no impact on democratization or corruption with higher levels of ODA (Knack 2001, 2004; Ear 2007) my argument follows from Dunning (2004) in that to better understand the more complex relationship,2 disaggregating the sample diachronically is of vital importance. Otherwise, a significant relationship might be overlooked.
I argue that two reasons best explain this relationship. The first comes from the rational side—it is in recipient states' self-interest to consent to the new demands of their multilateral donors for the sake of their international reputation and in order to maintain future ODA. On the donor side, in seeking to maintain their relevance as major international actors of development and governance, the major IO donors have a strong incentive to monitor recipient states to uphold “their end of the bargain.” The second explanation for this more nuanced relationship can be explained by the more normative, constructivist approach. An “anti-corruption” norm was instigated by leading IOs in the mid-1990s and after proliferating to all other major IOs during this time period, was accepted by major actors in the international system. This in turn brought substantial attention to the fight to curb corruption, and recipient states followed suit. I maintain that these two approaches help explain the primary hypothesis and empirical findings in this analysis in a complementary manner.
I test this relationship on 82 ODA recipient countries from 1986 to 2006.3 Using the two-stage generalized method of moments (GMM) I model a number of potential problems of endogeneity between corruption and ODA. The empirical results show that the ACM adopted by all major IOs proved to be an effective strategy in combating corruption in developing states, while the effects of multilateral ODA before this time period had mixed effects. Bilateral ODA is either a negative or insignificant determinant of corruption levels in recipient countries in both time periods. Upon multiple robust checks with alternative specifications and data on corruption, the results hold strongly.
I. THE IMPACT OF AID ON CORRUPTION
Corruption, though difficult to characterize in the abstract (Tanzi 1998) and admittedly even more difficult to detect in the “real world,” is defined as “the abuse of public office for private gain” and has an effect that is “corrosive to the development of a state” (Kaufmann 1997, p. 114). As Alesina and Dollar (2000) point out, almost two-thirds of all foreign aid collected goes to government consumption. These funds are therefore allocated by international sources and end up in the hands of government bureaucrats to be distributed in some form to the general public. Thus, some argue that foreign aid allocations are funds that are “ripe territory for corruption” (Tavares 2003, p. 100). Consequently, the question is: what effect does foreign aid have on corruption and governance in recipient countries? Not surprisingly, numerous previous empirical studies have investigated the relationship between foreign aid, or ODA, and some type of democratic-performance outcome (Goldsmith 2001; Stone 2004; Dunning 2004; Ear 2007; Knack 2001, 2004; Knack and Rahman 2007). Some scholars have argued that there is a positive relationship between ODA-dependence and corruption and have reported empirical evidence to support this claim.
The ways in which ODA could potentially exacerbate corruption and harm recipient governance has been well documented by Knack (2001, 2004). The argument essentially goes as follows: when aid dependence increases (whether measured by ODA/GDP or ODA as a proportion of government consumption), it is expected that recipient states will become less accountable for their own actions, that incentives for domestic corruption will be increased by increasing conflict over aid funds, and that aid will essentially compensate for poor economic policies and weak government institutions by offering a “crutch” (Knack 2001, 2004). Some scholars have found empirical support for the notion that, the more ODA a state receives relative to its GDP, the worse their democratic and bureaucratic performance and corruption levels become (Knack 2001, 2004; Knack and Rahman 2007). For example, Knack and Rahman (2007) estimate the effects of several determinants of bureaucratic quality using the International Country Risk Guide (ICRG) data (PRS Group 2009) and find that the quality of bureaucratic services is negatively impacted as the proportion of a state's ODA rises relative to GNP (Knack and Rahman 2007, pp. 189–92).
On the other side of the debate, there are numerous scholars who argue that in fact ODA has a positive impact on governance and indeed contributes to reducing corruption (Goldsmith 2001; Tavares 2003; Dunning 2004; Ear 2007). The argument in favor of more foreign aid in assisting with democratic development and corruption reduction is that IOs and bilateral donors can bring in certain expertise to developing states that they would otherwise not have. Accountability could in fact be enhanced due to international oversight, along with numerous conditionality measures which stipulate that states must reform their governing practices in order to make them more efficient and less corrupt. Developing states concerned about their reputation will seek to make enough reforms to receive future ODA. Furthermore, the expertise of some IO employees or foreign diplomats could provide the necessary “know-how” for developing states to make critical reforms in order to improve governance. Knack (2001) also provides the argument that increases in ODA can make-up for the shortfalls of resources in some countries that might be used for the salaries of bureaucrats and thus provide less of an incentive for them to practice “petty corruption.”
Some evidence has been reported by scholars to support the notion that ODA improves governance. In studying African states, for example, Goldsmith (2001) finds that increases in ODA as a proportion of GDP are associated with higher levels of democratic performance and economic freedom. Dunning (2004) replicates these results, but when disaggregating the time period into the Cold War and post–Cold War periods, he finds that ODA only improves democratization in the latter. Moreover, Tavares (2003) finds that even when controlling for such factors as economic development, oil resources and political rights, ODA has a strong and statistically significant relationship with curbing corruption. However, there appears to be no clear theoretical or empirical consensus on the effects of ODA on outcome variables such as quality of governance in general or corruption specifically. This analysis builds on these previous contributions while adding a more nuanced explanation of this complex relationship.
II. THE ANTI-CORRUPTION MOVEMENT AND ITS IMPACT ON RECIPIENT STATES
Official development assistance has become an increasingly more relevant source of income in developing regions over time, as shown in Figure 1. Of particular interest to this study is the subsequent impact of the consensus among major IOs to shift significant attention to the agenda of “good governance” in the mid to late 1990s.4 Beginning with the OECD in 1994, discussions on bribery came to the forefront by 1996, when a binding convention on “Combating Bribery of Foreign Public Officials in International Business Transactions” was signed by all 36 OECD member states (Sandholtz and Gray 2003). The World Bank (WB) followed suit with a clear message about fighting corruption and in 1997 began working with the nongovernmental organization (NGO) Transparency International on combating such practices, along with establishing its own anti-corruption institution, the World Bank Institute (WBI). Transparency International and WBI together take on a number of corruption-related problems. Since this time, the WB has tied anti-corruption practices to its list of conditionalities (Pieth 1997; Sandholtz and Gray 2003). The IMF, though not a development institution, in addition to the other Bretton Woods financial institutions began its campaign against corruption in 1996. In 1997, the organization finalized the first round of discussions on policies against corruption and declared its new agenda to combat it (IMF 1997). Moreover the United Nations created its own division called the Management Development and Governance Division (MDGD) in 1995, which by 1997 was elected by the member states to pursue an agenda of government accountability and transparency. Lastly, the World Trade Organization (WTO) compelled each member state to join the Working Group on Transparency in Government Procurement in 1996, which dealt with accountability and corruption issues. In addition to the major global IOs, a number of regional IOs have followed suit in the anti-corruption theme as well, such as the European Union and the Organization of American States signing comprehensive anti-corruption initiatives.
This analysis explores whether the new anti-corruption measures have had any significant impact on corruption levels in recipient countries. The ACM could have had an impact for two reasons, stemming from either the rationalist/utility-based or constructivist/normative perspective. First, on the normative side, the “anti-corruption norm” ensured that multilateral donors were serious about ODA being used for measures that fought corruption and improved governance in recipient states. As Barnett and Finnemore (2001) explain, “having established rules and norms, IOs are eager to spread the benefits of their expertise and often act as conveyor belts for the transmission of norms and models of ‘good’ behavior” and that “developing states continue to be a popular target for norm diffusion by IOs” (pp. 416–17). The inspiration for the anti-corruption norm is ideational or normative in the sense that IOs sought to combat a coercive element found in many developing states in order to improve governance and economic performance worldwide. Much like norms related to women's suffrage, non-proliferation, and human rights, they are driven by a select group of “idealist” states, then accepted by international actors (in this case IOs) and then “cascaded” throughout the system (DuBois 1994; Katzenstein 1996; Finnemore and Sikkink 1998). Whether the norms of “good governance” in this case are accepted domestically is of course a matter of self interest on the part of the developing country itself.
Second, as Finnemore and Sikkink (1998) note, “in addition, international norms must always work their influence through the filter of domestic structures and domestic norms, which can produce important variations in compliance and interpretation of these norms” (p. 893). This “two-level game” (Putnam 1988) between international ideas and domestic change leads into the rational side, where both the donors and recipients have incentives to push for good governance reforms when channels of multilateral ODA are established. In the case of recipient states acting rationally in their own interest, I posit that states are concerned with two simultaneous issues—reputation and future aid. States receiving multilateral ODA that are strongly tied to anti-corruption and good governance stipulations must in return make at least minimal reforms or else face the consequences of obtaining a poor reputation and face foreign aid cuts from skeptical donors in the future. It is therefore in their interest to comply so as to obtain such aid in the future. The reputation concerns also impact the multilateral donors—that the conditional policies tied to ODA actually have an impact after the aid is received by the recipient country. The less reform is observed in recipient countries, the less relevant donor IOs become and future stipulations of reform are likely to be taken less seriously by developing states. Consequently, the donors have incentives to monitor recipients so as to ensure some necessary changes are made with respect to improving governance.
III. THE POST–ANTI-CORRUPTION MOVEMENT MULTILATERAL AID HYPOTHESIS
The central contribution of this study is the idea that both the time and source of ODA matter in the relationship between foreign aid and corruption. Briefly, the drawbacks to bilateral aid are clear—the donor country can set whatever agenda it wishes with the ODA, and oftentimes such aid is “tied aid,” which many scholars have shown can exacerbate wasteful government consumption (see Ehrenfeld 2004) and distort trade (Pratt 1994). Recipient states are perceived to be much more willing to accept counsel from IOs than from other governments directly, which underscores multilateral aid's relative effectiveness as compared to bilateral ODA (Ehrenfeld 2004). However, because of previous mixed empirical results on the relationship between ODA and corruption, I argue that neither multilateral nor bilateral aid will play any significant role in levels of corruption for recipient states before the major shift in focus to the ACM (Bukovansky 2002) in the mid-1990s as outlined above. Due to bilateral ODA being largely strategic and the lack of serious attention before the mid-1990s to combating corruption by multilateral actors, there is no reason to believe that either type of ODA would be an effective determinant in reducing corruption in developing states. Subsequently, the shift by major multilateral donors to “good governance” and fighting corruption created a movement for a new international norm to improve governance conditions in recipient countries.5 What I seek to test empirically in this study is whether or not this shift in focus has had any substantive effect—whether the source (i.e., the major multilateral IOs) that has advanced the “anti-corruption norm” has impacted corruption levels in any significant way, which will be operationalized by the multilateral aid.
Based on the literature on bilateral aid and the findings of Alesina and Weder (2002), I leave open the possibility that bilateral sources may or may not be associated with higher levels of corruption perceptions in developing states. However, the contrast between the two sources is intended only to show possible differences in the effects of multilateral and bilateral ODA after the ACM begins (if any exist), not to have any specific theories about bilateral aid in and of itself. Finally, the effect of multilateral foreign aid prior to 1997 is not expected to play any significant role in determining variation in corruption levels because the “good governance” norms had yet to be universally accepted by all major donors before this time.
IV. DATA, SPECIFICATION, AND MEASUREMENT
The dependent variable in this study is corruption as operationalized by a leading indicator of this concept taken from the Political Risk Services Group's International Country Risk Guide (PRS Group 2009). The PRS Group, a think tank specialized in economic and political risk assessment internationally, has published monthly data for business and investors on over 140 countries since 1980. The PRS Group measure is primarily concerned with accounting for “excessive patronage, nepotism, job reservations, ‘favor-for-favors,’ secret party funding, and suspiciously close ties between politics and business.”6 A primary advantage for this study is that the time period of this indicator ranges from 1984 to 2006 and covers up to 139 countries, while other indicators such as those of the World Bank or Transparency International either have far too little or none of the pre-1997 data needed to test the hypothesis in this study. The data in the analysis has a finite range from 0 to 10, with higher scores indicating lower levels of perceived corruption. The PRS Group data has been used in numerous recent publications on determinants of corruption (Ades and Di Tella 1999; Persson, Tabellini, and Trebbi 2003; Knack and Rahman 2007; Charron and Lapuente 2010). Figure 2 breaks the data down by region across the time period in the study.7 Clearly there is a substantial amount of variance and movement among the regions, with some showing stability in the aggregate over the time period and others, such as the post-Soviet Eastern European bloc, demonstrating substantial declines. Of the 138 countries in the ICRG data for 2005, for example, Finland ranks best with a score of 10, followed by Sweden, New Zealand, and Denmark with 9.16. Developing countries show a significant range of variance: countries like Nigeria (1.67), Pakistan (2.5), and Gabon (1.60) rank on the low end, while Chile (6.67), Botswana (5.0), Jordan (5.0), and Singapore (7.5) rank among the least corrupt outside of the OECD.
The empirical tests are to be conducted as parsimoniously as possible. The primary independent variables, bilateral ODA (Bilateral) and multilateral ODA (Multilateral), are annual data taken from the World Bank. Clearly there are finer distinctions one can make with ODA, yet this simple disaggregation is made to maximize observations and maintain a level of parsimony to the analysis. Following the format of Goldsmith (2001), I take each figure as an annual proportion to a state's GNP, which is thus essentially measuring the level of dependence on ODA of each recipient country. Other significant determinants of corruption are also controlled for in the full model. The level of institutionalized democracy (Democracy) is a measure from Freedom House and measures the level of political rights on a scale of 1 to 7. I invert the scale so that higher numbers indicate higher levels of democracy. Higher levels of democracy are anticipated to be associated with lower levels of corruption, as found by previous empirical studies (Sandholtz and Gray 2003; Treisman 2000). Secondly, I control for economic development with the log of GDP per capita annually from the World Bank (logGDP). Higher levels of economic development are consistently shown in the literature as having a negative impact on corruption levels (La Porta et al. 1999; Treisman 2000). Additionally, I control for a state's level of ethno-linguistic heterogeneity by including Alesina et al.'s (2003) level of ethno-linguistic fractionalization (Ethnic Frac). Several studies have demonstrated a positive relationship between ethnic and linguistic diversity in a country and corruption levels (see Charron 2009). Finally, I control for a country's legal system: common law (UK Common), in particular, has been shown compared to others to be a significant factor in a developing country's transitional phase to well-functioning government institutions (La Porta et al. 1999).8
V. RESEARCH DESIGN
Due to data constraints in the dependent variable, I run models based on a limited time frame between 1986 and 2006. Based on the potential problems of endogeneity and the mixed empirical evidence suggesting that corruption could in fact impact the levels of ODA a country receives—with some claiming that it has a negative relationship with aid (McGillivray, Leavy, and White 2002) and others finding null or mixed results (Alesina and Weder 2002; Alesina and Dollar 2000)—ordinary least squares (OLS) may no longer be the best linear unbiased estimator (BLUE). In the presence of endogeneity from reverse-causality one of the key OLS assumptions (e.g., E(u|x) = 0) is therefore violated.9 I elect to model this problem explicitly.
I employ two estimation methods in the analysis to remedy this problem. One is a simple two-stage least squares (2SLS) model in which lagged values of the ODA types are instrumented variables with several “predetermined” factors, such as colonial heritage and geography, along with past values of corruption. The 2SLS attempts to model the endogenous relationship between corruption and ODA explicitly. Next, I run a series of regressions using GMM estimation on panel data, introduced by Hansen (1982), in which he demonstrates that GMM estimators are consistent and asymptotically normally distributed. The estimation method takes into account problems associated with endogeneity and may produce more efficient and reliable estimates than 2SLS in the presence of heteroskedasticity (Baum, Schaffer, and Stillman 2003). In addition, GMM has advantages over the standard IV estimates because as the length of the panel increases, so does the number of valid instruments (Roodman 2007).
When the number of instruments (K) equals the number of parameters (L) and the equation is exactly identified, and thus K=L, then the method of moments estimator corresponds with the 2SLS estimator. Assessing the validity of the instruments is done by using a post-estimation Sargan–Hansen test. The null hypothesis of the Sargan–Hansen test is that the overidentifying restrictions are valid, in other words that the instrumental variables are uncorrelated with the error term. Although we are to expect first-order autocorrelation (AR1) in the model because the change in the estimation of the error term for observation i at time t (Δνit) is mathematically related with Δνi,t−1 due to the fact that they share the term νi,t−1, Arellano and Bond (1991) argue that the first difference estimates will be consistent if Δνi,t−2 is not correlated with Δvit. Therefore, the validity of instruments also requires that there is no second-order serial correlation in the residuals. When running models using 2SLS, ODA is the instrumented variable. Along with the Sargan–Hansen test, I test the relevance of the first stage with an F-test.
On the issue of outliers, I make two adjustments to the sample. First, there are three extreme outliers with respect to ODA that, if included, violate one of the assumptions of the model and significantly alter the results—Sudan, Niger, and Ghana—as they receive levels of multilateral aid in multiple years from 1995 to 2006 that are well above three standard deviations over the mean sample level.10 Although these two-step estimators are not alone on this point, numerous scholars in the literature on GMM and 2SLS estimation have pointed to the potentially hazardous and misleading effects of extreme outlying cases when estimating models with panel data (see Huber 1981; Baum, Schaffer, and Stillman 2003; Lucas, van Dijk, and Kloek 2007). In particular, in datasets where a few cases demonstrate significantly divergent behavior from the majority of cases in the sample and the number of cross-sectional units is substantially greater than the number of time periods, as in the case of this study, the outliers can substantially impact the estimates in misleading ways (Lucas, van Dijk, and Kloek 2007, p. 2). If the cases are few, then dropping them or controlling for them in the model is an appropriate solution; both approaches are performed in the empirical section.11 Secondly, based on Figure 2, the East Asian countries show a trend that is divergent from the rest of the sample in that there is a sizable and rapid drop-off in corruption scores from 1999 on, while at the same time this group received less ODA due to rapid economic growth. I control for this by reporting models with and without this region.
Two separate time periods are analyzed: before and after 1997.12 The models are regressed on all available data before and after 1997 to test the impact of ODA (both multilateral and bilateral) on corruption.13 I check the robustness of the results by employing the GMM estimation and 2SLS estimations with both ICRG data and an alternative source of corruption data, Transparency International's Corruption Perceptions Index (CPI). Further, I check the sensitivity of the time frame by setting the start year of the ACM at 1998 and 2000 to see if the results are time sensitive. Finally, the sample of states in the empirical analyses is all developing (recipient) states for which corruption data is available.14
The models in both Tables 1 and 2 report the two-stage estimates to capture the effect of the relationship between ODA and corruption levels. Both the 2SLS and GMM estimations are extensions of linear regression and their interpretation is similar to that of OLS.15 The estimates in Table 1 are intended to elucidate the effects of the two types of ODA before and after the ACM.16 Model 1 examines the overall effect of total ODA over the entire time period, demonstrating essentially “what we would miss” if the data was not disaggregated by time and aid source. As several studies have reported, total levels of ODA are actually associated with greater levels of corruption (see Knack and Rahman 2007), yet the coefficient fails to reach an appropriate level of significance in model 1, and is in fact positive.
Table 1. Disaggregating the Effects of Official Development Assistance (ODA) on Corruption: Two-Stage Least Squares (2SLS) Estimation
No Southeast Asia
Note: This table uses 2SLS estimation. International Country Risk Guide (ICRG) corruption scores are used as the primary dependent variable. The first-stage F-test in the linear IV model tests for instrument relevance. The null hypothesis is that all of the instruments are uncorrelated with the endogenous regressors. The Sargan–Hansen J-test performs an overidentification test with a χ2 distribution (p-values reported). The instruments used in the first stage for each type of ODA are colonial origin, regional dummies and a two-to-five-year (averaged)-lagged corruption variable. The year of the anti-corruption movement (ACM) in the models is 1997. Alternative data from Transparency International (CPI) are used only for the time period after 1997 as they have almost no coverage beforehand. They range from 0 to 10, respectively, with higher numbers indicating lower corruption levels. Thus, the five-year-lagged dependent variable (ICRG) used at first as a regressor of the ODA instrumented variable in the first four models is kept so as to address (albeit imperfectly) the endogeneity problem in models 5–8. Ethnic fractionalization and UK colonial heritage were insignificant in models 1–9, so models were run without them in the second stage of the equation. The results were not significantly altered with their inclusion compared with those reported. Robust standard errors correcting for heteroskedasticity are in parentheses.
***, **, and *
represent statistical significance at the 1%, 5%, and 10% level, respectively.
Table 2. Disaggregating the Effects of Official Development Assistance (ODA) on Corruption: Generalized Method of Moments (GMM) Estimation
No Southeast Asia
Note: International Country Risk Guide (ICRG) corruption scores are used as the primary dependent variable. This table uses generalized method of moments (GMM) estimation, where single-stage estimation is reported as recommended by Arellano and Bond (1991) (“xtabond” command in STATA). The instruments are based on lagged values of the explanatory variables. The Sargan–Hansen J-test, which represents a difference-in-Sargan test of exogeneity of the instruments, performs an overidentification test with a χ2 distribution (p-value reported). The tests for first- and second-order autocorrelation are asymptotically distributed as standard normal variables (Arellano and Bond 1991). The Arellano–Bond AR(1) and AR (2) tests test for first- and second-order autocorrelation (z-scores reported). A significant value represents the presence of autocorrelation. The lack of AR(2) indicates that the model has been correctly specified. The year of the anti-corruption movement (ACM) in the models is 1997. Alternative data from Transparency International (CPI) are used only for the time period after 1997 as they have almost no coverage beforehand. CPI ranges from 0 to 10, with higher numbers indicating lower corruption levels. Robust standard errors, correcting for heteroskedasticity are in parentheses.
***, **, and *
represent statistical significance at the 1%, 5%, and 10% level, respectively.
Models 2 and 4 show the impact of multilateral and bilateral aid on corruption before 1997. The control variables in the model, political rights (democracy) and GDP per capita (logged), are significant determinants of corruption levels. However, neither bilateral nor multilateral aid impacts corruption in recipient states significantly during this time period, with multilateral aid actually having a negative coefficient. Yet when moving to models 3 and 5, which elucidate the post-ACM effects of ODA, the estimates show a different impact on the dependent variable for multilateral aid entirely. A one-unit increase in multilateral aid is quite substantial, estimated with a reduction of corruption by almost an entire standard deviation of the ICRG variable (1.62), and, unlike bilateral aid, statistically significant at the 95% level of confidence. However, a one-unit increase (an additional 1% of GDP coming from ODA) would be a sizable increase and is not all that likely according to the data. A more realistic interpretation would be to use one-tenth of a one-unit increase, which would result in an increase of approximately 0.16 in the ICRG corruption score.
The diagnostic tests are included for each model. First, as recommended by Wright (2003) the first-stage F-test provides a “sufficient (but not necessary) test for underidentification” (p. 329). All models have a significant F-statistic, meaning that we would not suspect a “weak instrument” problem in the models. Second, results of a Sargan–Hansen J-test of overidentifying restrictions are reported, which are also asymptotically distributed as χ2. Judging by the Sargan–Hansen statistic (the p-values are reported), none of the tests indicate a decisive rejection of the model's overidentifying restrictions
Models 6–10 test alternative models of corruption. In models 6 and 7, I employ Transparency International's CPI, which is a composite index made up of multiple surveys conducted by numerous sources and is available annually for a substantial number of countries from 1996 onward.17 Since the time range for these variables is limited, only the effects of the post-ACM time period are analyzed with the CPI. Here we see robust support for the results in models 3 and 5 and the results for multilateral aid. An increase in multilateral aid during this time period is strongly associated with lower levels of corruption using the CPI, demonstrating wider support for the hypothesis.
In model 8, I remove a group of potentially problematic outliers (Southeast Asian countries) in the model to find out if their inclusion is driving the results in a substantial way. Based on the aggregate figures from Figure 2, we observe that this group declines significantly in corruption scores after around 1999. This trend occurred while states in East Asia were receiving significantly less ODA during this time period due to their economic development.18 These two simultaneous trends pose a potential problem in that they could be driving the results in the post-ACM period analysis. Looking at the result in model 8, the significance of multilateral ODA is slightly diminished (as indicated by the higher standard error), yet the coefficient is still strongly positive and significant, indicating that despite the exclusion of Southeast Asian countries, higher levels of multilateral ODA are associated with lower corruption levels, ceteris paribus. In the final two models in Table 1, the ACM is moved up to 1998 and 2000, respectively, to test the sensitivity of the year 1997. Establishing a later year for the ACM does not substantially alter the results, although with the reduced number of observations in model 10, the significance for multilateral aid drops to 90%, while the significance of the GDP per capita coefficient, for example, drops below 90%.
As noted previously, Baum, Schaffer, and Stillman (2003) find that GMM may produce more efficient and reliable estimates than 2SLS in the presence of heteroskedasticity and fixed-country effects in the data. I therefore check the robustness of the results in Table 1 with the Arellano and Bond (1991) GMM estimator in Table 2. The variables are now transformed to their first difference, which will take care of any fixed effects in the data which may have altered estimates in Table 1. Lagged variables by two years are used as instruments in all models
In model 1, a similar result is observed compared with model 1 in Table 1, with ODA having an insignificant impact on corruption during the whole time period. In models 2 and 3, both estimates for bilateral ODA are negative, yet are again insignificant, as they were in Table 1. The estimates for multilateral ODA are in the same directions as they were in models 4 and 5 in Table 1, yet when considering their first-difference effects, multilateral ODA is actually associated with higher degrees of corruption before the ACM and lower levels of corruption after the ACM. Models 6 and 7 corroborate this finding using Transparency International's data. Here we find that the effect of multilateral aid on corruption after the ACM remains strongly significant, at the 99% level of confidence, demonstrating strong and robust support for the 2SLS estimates. Thus both estimations in Tables 1 and 2 elucidate that the relationship between ODA and corruption is nuanced—that it depends on the type of ODA and timeframe on which the analysis focuses.
Further robust checks include model 8, which excludes East Asian states from the analysis as in Table 1. The coefficient for multilateral ODA weakens by roughly 25% from model 5, yet the relationship remains positive and significant at the 90% level of confidence. Moreover, whether the ACM is set later in 1998 or in 2000 plays no substantial role in altering the results, as shown in models 9 and 10. Moving to the tests for autocorrelation, we find that AR(1) is present as expected, yet we do not find AR(2) in any model, indicating that the models are correctly specified. The GMM estimation can be considered consistent if the lagged values of the explanatory variables are valid instruments. A look at the difference in Sargan statistic in each of the 10 models is reassuring, confirming that the instruments used in the model are indeed valid.
Briefly moving to the control variables, GDP per capita is again a strong indicator of corruption, with positive changes indicating a reduction in corruption. Interestingly, the coefficient for democracy is negative in most models, which would suggest positive changes toward democratization would mean higher perceived corruption, however, in none of the 10 models is the relationship significant at even the 90% level of confidence.
The results thus far have corroborated the hypothesis that multilateral aid in the post-ACM time period is associated with less corruption, suggesting that the ACM has in fact played a significant role in combating corruption in developing countries. In sum, irrespective of the estimation method, corruption data employed, the start date of the ACM used, or sample, the effects of the ACM (as measured by multilateral aid) are consistently associated with lower levels of corruption. The results however are slightly sensitive to dropping the Southeast Asian countries and changes in the start year of the ACM regime. Thus while the results mainly reveal strong support of the hypothesis, one must interpret them with caution, in particular due to the imperfect operationalization of the ACM regime.
This study provides a more nuanced explanation and empirical examination of the complicated effects of foreign aid on domestic corruption levels. While the data on corruption and the use of multilateral ODA as a proxy for the ACM are of course imperfect indicators of the respective concepts they are intended to represent, it would seem that, given this caveat, there is relatively strong empirical support for the “anti-corruption” hypothesis—that, based on the strong correlation between multilateral ODA and the dependent variable, the ACM has in fact been relatively effective in curbing the level of corruption in recipient countries during the regime. This can be distinguished from bilateral aid, which was shown to be largely an insignificant (or slightly negative) determinant of corruption levels in recipient states. The results demonstrated that multilateral aid in the pre-ACM time period proved ineffective in combating corruption, yet it was revealed to be quite successful in the sample which examines only the time period after 1997 (the year by which all major IOs had signed agreements on fighting corruption). The effects of multilateral ODA dependence are strong and generally robust across both indicators of corruption, using multiple estimation methods and specifications, different start years, including and excluding East Asian states and when controlling for economic development, level of democracy, colonial heritage, and ethnic fractionalization. However, since it was not possible to measure the ACM directly, there is always the possibility that the results are somewhat spurious, and thus should be interpreted with a degree of caution. Further, while the alternative indicator, the CPI, corroborated the findings from the ICRG data, it is not available for the pre-ACM time period. However, a “pre-test,” so to speak, could not be conducted, and we are therefore not privy to whether or not the ACM altered any relation between the two types of ODA and corruption as measured by Transparency International.
In addition, there are methodological difficulties in studies such as these, namely the issue of endogeneity. Scholars have reported that the relationship works both ways (i.e., that ODA impacts corruption and corruption impacts ODA) and thus the researcher must pay a good deal of attention to the modeling of this issue so as to not to produce biased estimates. While there is no perfect solution to this problem, this study used 2SLS regression with a lagged independent variable, modeling corruption as a function of past values of itself and ODA (as well as the control variables) while simultaneously modeling ODA as a function of past values of corruption and two-year-lagged GDP and democracy values along with regional and colonial heritage controls. In addition to 2SLS, the GMM estimation in Table 2 corroborated the findings reported by the 2SLS estimation.
Some practical implications follow from these results. One, that the relatively new worldwide attention to overall “good governance” and more specifically to fighting corruption has been rather effective from a multilateral standpoint. The data show that after the “anti-corruption” norm was accepted in the mid-1990s, multilateral agencies, which can be considered to have less of their own political agenda with respect to the aid that they allocate compared with direct bilateral aid donors, were considerably more effective in producing better governance relative to bilateral ODA investments. The perception remains that bilateral ODA donors tie their own self-interest to the aid that they allocate to recipient countries. The comparison of the results in the two time frames is thus interesting in the sense that neither ODA strategy, multilateral or bilateral, was associated with significantly lower levels of corruption in the early to mid-1990s. Yet when attention was focused on the ACM, the agencies were apparently able, with processes of loan allocation that are mostly more transparent than those of bilateral transactions, to achieve a significant task—bringing down corruption levels in the developing world. However, this study demonstrates only a strong correlation regarding this relationship. In further scholarship, more theoretical development is needed to elucidate the underlying causal mechanism more thoroughly.
Secondly, and further, it is clear that if states are serious about fighting corruption—and there are both economic and moral reasons for the international community to be serious (Bukovansky 2002)—policy should be shifted in order to allocate more resources to IOs for ODA redistribution. Accomplishing this is of course easier said than done—states collect revenues from their citizens, who expect their leaders to spend their money in their interest. However, though more investigation and study of the nuanced relationship between the two variables shown in this analysis, the data and results in this study demonstrate that one channel of foreign aid is more effective than another in accomplishing positive results for improving governance and combating corruption.
ODA is defined as “flows of official financing administered with the promotion of the economic development and welfare of developing countries as the main objective, and which are concessional in character with a grant element of at least 25 percent (using a fixed 10 percent rate of discount). By convention, ODA flows comprise contributions of donor government agencies, at all levels, to developing countries (‘bilateral ODA’) and to multilateral institutions. ODA receipts comprise disbursements by bilateral donors and multilateral institutions” (Organisation for Economic Co-operation and Development, “OECD Glossary of Statistics Terms,”http://stats.oecd.org/glossary/detail.asp?ID=6043, accessed December 2008).
Dunning (2004) demonstrates that there are different effects of ODA on the level of African countries' democratization scores when separating the Cold War era from the post–Cold War era. This conditional effect was overlooked in the Goldsmith (2001) study which demonstrated a less nuanced relationship.
While this dichotomy might seem somewhat crude, it allows me to maximize the number of developing states in the sample, as opposed to using measures such as Department for International Development (DFID) aid. The trade-off with using multilateral aid in the aggregate is of course that the level of abstraction becomes higher; yet using more specific data on aid specifically targeting corruption is too limited and significantly reduces the number of cases.
Summary statistics for all variables in the model can be found in Appendix Table 2.
The post-estimation Wu–Hausman test in the 2SLS and GMM statistic in the two-stage models I run indicate that there is indeed consistent endogeneity in the models, and thus the specification appears to be correct.
For example, in 2003, 6.75% of Ghana's GDP was from multilateral development aid (the sample mean for that year was 0.20) and in 2004, Sudan (along with experiencing extreme civil conflict) received 7.25 of its GDP in multilateral aid. Similarly, Niger reached multilateral aid levels of 3.5% of GDP during the years from 1996 to 1998, and the levels have remained over 1.5% since.
For the sake of space, only models without the extreme outlying cases are reported. Please contact the author for results with their inclusion.
In additional robust checks, I change the ACM year to 1998 and 2000, and remove some of the instruments, such as the colonial heritage, with no significant changes in the results.
I elect to regress bilateral and multilateral ODA on corruption in separate models due to relatively high levels of multicollinearity, which results in more efficient coefficient estimates.
Iraq is omitted as an outlier case due to the huge influx of ODA starting in 2003. Additionally, Afghanistan has no data on corruption.
Running the multilateral and bilateral ODA data in the same model introduces a high level of multicollinearity as indicated by post estimation t-tests. For the sake of efficiency and clarity, I elect to run them separately in this analysis.
The World Bank data was available biannually from 1996 to 2002, and has been available annually from 2002 to the current year. More detail on the World Bank and Transparency International data can be found in the Quality of Government's data codebook: http://www.qog.pol.gu.se/.
I would like to thank an anonymous reviewer for this suggestion.
Table APPENDIX TABLE1. List of States
Removed from the final sample due to extreme outlier figures.