Does Foreign Aid
Foster Corruption?
A quantitative comparison of various
estimation methods
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
Name: Luka Bastiaans
University: Universiteit van Amsterdam
Master Specialization: Development Economics Student Number: 5876095
Supervisor: Jan Jan Soon Date: August 2015
!
!
Statement of Originality
!
This document is written by Student Luka Bastiaans who declares to take full
responsibility for the contents of this document.
I declare that the text and the work presented in this document is original and
that no sources other than those mentioned in the text and its references have
been used in creating it.
The Faculty of Economics and Business is responsible solely for the
supervision of completion of the work, not for the contents.
!
Abstract
This study examines the effects of foreign aid on corruption by making a comparison between various estimation methods. Several studies have studied the causal relation between foreign aid and corruption by making use of either OLS, IV, GMM, or the QR method. Results of these studies differ widely. The contribution of this thesis is to examine whether the use of different estimation methods is a significant cause of variation in the conclusions on the matter. A large panel and cross sectional dataset is created with variables also used in other studies. Estimation results vary widely and support the hypothesis that differences in the conclusions are caused by the use of different estimation methods. Providing multilateral aid seems to be less self-oriented than providing bilateral aid, but these results are also dependent on the estimation method used. In a larger perspective this signals that it is very important not to follow conclusions blindly, but to be critical on the assumptions of the studies and to keep in mind that conclusions are often partly arbitrary.
!
!
Table of contents Page
1. Introduction 6
2. Literature Review 8
2.1. Foreign Aid 8
2.2. Possible Causes for the Effects of Aid on Corruption 8 2.3. Effects of Aid on Corruption 10
2.3.1. Ordinary Least Squares 10
2.3.2. Instrumental Variables 11
2.3.3. Generalized Method of Moments 15
2.3.4. Quantile Regression Method 18 2.4. Why Bilateral Aid May Not Be Significant 19 3. Research Design and Methodology 20
3.1. Estimation Methods 20
3.2. Equations 21
3.3. Data 24
3.4. Choice of Variables 25
3.4.1. Controls 25
3.4.2. Instruments for Aid (IV) 27
3.4.3. Instruments for Aid (GMM) 28 3.5. Prediction According to the Literature 28
4. Regression Results 30
4.1. Introduction 30
4.2. Cross Sectional OLS and IV Analysis 30
4.2.1. Introduction 30
4.2.2. Average Cross Sectional OLS 31
4.2.3. Average Cross Sectional IV 33
4.2.4. Difference Cross Sectional OLS 36
4.2.5. Difference Cross Sectional IV 38
4.2.6. Robust Cross Sectional OLS and IV 43
4.2.7. Post Estimation 47
4.2.8. Conclusion Cross Sectional Estimation 47
4.3.1. Introduction 48
4.3.2. OLS Results 50
4.3.3. IV Results 59
4.4. Quantile Regression Analysis 60
4.4.1. Introduction 60 4.4.2. Results 61 4.5. GMM Regression Analysis 66 4.5.1. Introduction 66 4.5.2. Results 67 4.6. Limitations 72 5. Conclusion 73 6. Reference List 76 7. Appendix 80 7.1. Variables 80
7.2. Ordinary Least Squares 84
7.3. Instrumental Variables 85
7.4. Generalized Method of Moments 86 7.5. Quantile Regression Method 89
7.6. Regression Results Tables 90
7.6.1. Cross Sectional OLS Results 90
7.6.2. Cross Sectional IV Results 102
7.6.3. Panel OLS Results 134
7.6.4. Panel IV Results 166
7.6.5. Quantile Regression Results 183
1. Introduction
Foreign aid is an important source of funding for developing countries and as long as there will be developing countries and crises, it will be a necessity. Despite the positive effects of foreign aid, there are also some negative externalities. Among them is a possible rise in corruption in the aid receiving countries due to large aid flows. With large aid flows, individuals may increase their consumption through rent-seeking activities of the revenues of the government. This reduces the amount of public goods provided (Svensson, 2000).
There have been several studies that investigated the effect of foreign aid on corruption. Most studies are based on the influential Alesina and Weder (2002) study. They estimated the effects using OLS, which suffers from endogeneity bias because of reversed causality between aid and corruption. Their results show that an increase in aid increases corruption. Others tried to improve the Alesina and Weder (2002) study by using instrumental variable estimation. However, results differ widely. Whereas, for example, Bräutigam and Knack (2004) have found that higher aid flows cause more corruption or erodes government quality, Tavares (2003) found that more foreign aid decreases corruption. Moreover, Coviello and Islam (2006) concluded that foreign aid has no impact on corruption by using both IV and GMM estimation (Generalized Method of Moments). Also Djankov et al. (2008) conducted IV and GMM estimation. They have found that foreign aid has a negative effect on institutions.
Some have argued that the differences in results can be caused by the differences between multilateral and bilateral aid. Bilateral aid refers to transfers from government to government, whereas multilateral aid is given through multilateral institutions - such as the World Bank - which pool aid from one or more sources and disperse it among recipients. Bilateral donors may have more strategic driven incentives when allocating aid (Alesina and Dollar, 2000). When looking at studies that make a distinction between multilateral and bilateral aid, results are still puzzling. Alesina and Weder (2002) concluded that there are no systematic differences between bilateral and multilateral donors, and that both kinds of aid will increase corruption. Asongu (2012) drew the same conclusions by using both IV and GMM. Okada and Samreth (2012) criticized the use of OLS and IV estimation. They argued that these methods only examine the effect of aid on corruption using mean values. Instead they used the quantile regression (QR) methodology, which measures the effect of aid on corruption at different intervals throughout the corruption distribution. They concluded that multilateral aid reduces corruption and that bilateral aid has no significant effect on corruption.
Thus earlier studies drew very diverse conclusions, as well for total foreign aid as while making a distinction between bilateral and multilateral aid. These diverse conclusions may be explained by using different estimation methods, by using different instruments and controls, or by using differentiated datasets. Overall, there is no clear conclusion on the effect of foreign aid on corruption. In this study one large panel dataset (and a corresponding cross sectional dataset) will be constructed on base of variables used in earlier studies on this subject. Then the various estimation methods used by other authors are utilized to test for the effect of foreign aid (total foreign aid, multilateral aid, and bilateral aid) on corruption.
What motivates the set up of this study is that the conflicting results in the literature linking foreign aid and corruption may be due to different types of estimation methods used. The contribution of this study is trying not to have a tunnel vision towards the use of one single estimation method, but to see the bigger picture of whether the use of different estimation methods is a significant cause of variation in conclusions on the matter. Therefore the hypothesis of this thesis is that variation in the conclusions in the study of the effects of foreign aid on corruption is caused by the use of different estimation methods. By using one dataset variation in the variables used is canceled out as optional cause for variation in the conclusion. The purpose of this thesis is to be able to see the bigger picture on the matter and to be critical towards the use of just one (or several) estimation method(s) and to be critical on generalizing conclusions on the base of a small set of assumptions. With this approach we will be able to see if this study can draw a more homogeneous conclusion - on base of the use of more estimation methods but still on the (small set of) assumptions of this study - towards the subject.
This thesis will be structured as follows. First, the different kinds of aid are discussed and there will be argued why aid might foster corruption. Then an extensive literature review is presented whereby the different studies about the effects of foreign aid on corruption will be discussed. Also the distinction between bilateral and multilateral aid will be made. Following this the research design is presented and then the test results of the different estimation methods used (OLS, IV, QR, and GMM) are presented and discussed. After discussing the test results possible limitations are pointed out and an conclusion is given.
!
!
!
!
2. Literature Review
2.1. Foreign Aid
Developing countries receive approximately 10%-20% of their GNI (Gross National Income) as net ODA (Official Development Assistance) (World Bank Indicators, 2014). This is a large share of their income and therefore it is important to determine whether aid has effects on the quality of institutions, and in particular, on corruption.
Amounts of development aid are often measured by ODA. ODA is defined by aid flows received by developing countries that are present on the DAC (Development Assistance Committee) list of ODA recipients. ODA is: “administered with the promotion of the economic development and welfare of developing countries as its main objective and is concessional in character and conveys a grant element of at least 25 per cent.” (AidFlows, 2014, p.5). Institutions providing ODA are both bilateral institutions and multilateral institutions. Bilateral transactions are directly passed from the donor country to the aid recipients. Multilateral transactions are international agencies, institutions, or organizations of whom the members consist of governments or a fund autonomously managed by an international agency (AidFlows, 2014).
Differences between the provision of bilateral and multilateral aid may be explained by different objectives for providing aid. For bilateral aid, colonial past, trade relationships, and economic performance of the recipient country are important determinants for the aid provision (Alesina and Dollar, 2000; Berthélemy and Tichit, 2004). Moreover, Younas (2008) found that more aid is allocated by bilateral donors to the countries who import capital goods in the donor countries. Thus bilateral donors seem to be self-oriented. However, there is evidence of heterogeneity among bilateral donors (Berthélemy, 2006). Multilateral aid, given through the World Bank, United Nations, and Development Banks, for example, take income per capita and human development of the recipient country into account. However, they also tend to focus on colonial past (Neumayer, 2003).
Milner (2004) argues that direct influencing the aid recipient countries is easier with bilateral than with providing multilateral aid. Multilateral aid is often given to the most needy countries, and it consists of grants, whereby bilateral aid often consists of loans.
!
2.2. Possible Causes for the effects of Aid on Corruption
Almost two-third of foreign aid is allocated to government consumption (Alesina and Dollar, 2000). When governments have lack of revenues to build up proper bureaucracy and legal systems, aid could help financing these institutions (Knack, 2001). Aid not only finances institutions, but can
also improve these by passing on skills and knowledge. Moreover, aid may stimulate reforms in the recipient country (Coviello and Islam, 2006). Donors of foreign aid have tried to improve the quality of governments by providing conditional aid, but this seems to be largely ineffective in improving institutions (Knack, 2001).
Foreign aid may influence corruption levels in the recipient country. Several authors argue that aid can reduce the cost of doing nothing. Moral hazard is often a problem in recipient countries. Because of the high levels of aid countries are less dependent of tax resources, and therefore they may not use these resources in an efficient manner. The government may allow corruption, because they still receive aid independent of the corruption issues. Governments may spent more than they are normally supposed to spent, because they expect deficits to be carried by cheap donor loans. Thus the governments may not have enough incentives to stay on budget and therefore may not raise taxes enough or fight corruption well enough. They may make large deficits in lending programs which places a burden on future governments who have to repay these loans (Bräutigam and Knack, 2004).
Because these governments become less dependent on tax resources, accountability to their people worsens. There is no need of reforming the government to attract more taxes or private capital (Knack, 2001; Bräutigam and Knack, 2004; Coviello and Islam, 2006). By performing rent seeking activities the government provides more private than public goods. This opportunity gets larger when more aid is allocated (Svensson, 2000). Moreover, donors themselves suffer from moral hazard, in the sense that there are often no consequences if an aid projects fails to improve conditions. This can induce riskier behavior with projects (Bräutigam and Knack, 2004).
Implementing a large amount of aid projects can lead to weakened institutions in developing countries, because the government needs to oversight and report the projects, which can keep the government from other priorities in their country. Moreover, because of the large number of projects, governments are not able to manage everything, thus donors set up their own project implementation units. In setting up these units, donors need local people and therefore these people are not available for their own government and private sector. Therefore, less suitable people may come in power. Next to this, because management of the projects is not all done by the recipient governments, government learning is limited. Another disadvantage of aid projects is that it may reduce tax revenues, because the projects are not paying income taxes for their foreign personnel and are not paying import duties for equipment of their personnel, while this may substitute government or private sector work (Bräutigam and Knack, 2004).
2.3. Effects of Aid on Corruption 2.3.1. Ordinary Least Squares
One of the first studies about the effects of foreign aid on corruption is the study of Alesina and Weder (2002). Their study covers three questions. At first they find when using OLS that there is no evidence that more foreign aid is received by countries with the lowest corruption levels. The first question is conducted with data over the time period 1975-1995 and the authors use controls for other determinants of the allocation of foreign aid taken from the well-known Alesina and Dollar (2000) study about aid allocation. As a corruption measure Alesina and Weder (2002) use both the ICRG corruption index (from the International Country Risk Guide) and six corruption indices from other sources.
Using the results from their first question Alesina and Weder (2002) investigate whether foreign aid reduces or fosters corruption. Because the authors did not find an effect of corruption on the allocation of foreign aid in the first question, they assume that they are able to test for the reverse effect of aid on corruption using OLS. Using a yearly panel ranging from 1984-1995 and lags for both the corruption (dependent) variable and aid (independent) variable they investigate whether percentage changes in aid have an effect on the change in corruption within a country. Corruption is measured by the ICRG corruption index, and the authors find that an increase in aid correlates with an increase in corruption.
Regarding their results there is some discussion. Although it is not proven that corruption affects the levels of aid, this regression may be biased due to the possible causality between aid and corruption. Therefore there still could be reverse causality. Bräutigam and Knack (2004) argue that OLS estimates are biased if donors allocate more aid to countries with higher levels of corruption. Then the OLS estimates are too large (upward biased) and therefore the effect of aid on the quality of institutions would be too large. If instrumental variables aid coefficients are larger than OLS coefficient, then donors do not allocate more aid to countries with worse institutions.
Another assumption of OLS is that there should be zero correlation between the error terms of different observations. When making use of panel data you would expect a correlation between the error terms of one country across time. Therefore, if there is an unobserved variable that influences the error term in one period, this variable can influence the error term in other time periods as well, which makes the OLS estimator biased (Stock and Watson, 2007). For these reasons, OLS is not the best estimator. However, the Alesina and Weder (2002) study has been of large influence on the topic of the effects of foreign aid on corruption.
!
Next to these questions, Alesina and Weder (2002) investigate whether donors differ in their willingness to discriminate against corrupt governments. They make a distinction between multilateral and bilateral aid and between different donor countries. To answer their question, the authors use a TOBIT regression since there is the presence of zeros in their analysis, because some donors do not give aid to all recipient countries. As a dependent variable they use the amount of aid per capita given by each donor. Results show there are some cross-country differences, but no significant differences between multilateral and bilateral donors. Overall, their study shows that neither multilateral nor bilateral donors seem to give less aid to more corrupt countries.
!
2.3.2. Instrumental Variables
Because of the endogeneity problems with OLS estimation several authors have studied the relationship between foreign aid and corruption using instrumental variables estimation. In the first stage of this estimation method the explanatory variables are regressed on instruments not correlated with the dependent variable in the second stage. Also exogenous controls could be added. Now instrument exogeneity can be assumed: the instruments are uncorrelated with the error term, therefore IV estimation will solve for the possible endogeneity bias of OLS (Stock and Watson, 2007).
Next to using instrumental variables instead of ordinary least squares, most authors use cross sectional data (data on different sections (countries) for a single time period). An advantage of using cross sectional data is that differences between countries can easily be compared because changes over time are not incorporated into the data (the data cover one time period). A disadvantage is that using data from one time period only reduces the sample size and the data is possible affected by shocks in the studied period. With panel data, more time periods are investigated and therefore shocks have a smaller effect in the analysis. However, with panel data there can be autocorrelation.
Svensson (2000) studies the relationship between foreign aid and rent-seeking during 1980-1995. He develops a game-theoretic rent seeking model and tries to test this model empirically with a panel dataset which is divided into three 5-year periods. By studying multiple periods he wants to increase the sample size and also look at variation over time. The author separates countries suffering more and less from competing social groups and predicts that aid will increase the level of rent-seeking activities in countries that suffer more from competing social groups. A two-stage instrumental variable model is constructed, whereby in the first stage effects of
rent-seeking on aid are tested, and the second stage consists of measuring the effect of existing powerful social groups on the average level of rent-seeking activities in a certain time-period in a certain country. Svensson (2000) uses a measure of ethnic diversity in a country as a proxy for competing social groups. He states that when there are more groups and when they are more equal in size there will be more competition. Rent-seeking is measured by the ICRG corruption index. The model is tested by Two Squares Least Squares (TSLS) with standard errors adjusted for country-specific random effects, because effects could differ in, for example, Sub Saharan Africa and Central America. Results of the first stage show, as in Alesina and Weder (2002), that donor countries do not significantly allocate more aid to countries with less corruption. However, the same problem (reverse causality) as with the Alesina and Weder (2002) study applies here. Second stage estimates show that the more fractionalized a country is, the larger is the effect of aid on corruption. Svensson (2000) also shows that in countries suffering less from competing social groups, aid has a weakly significant negative effect on corruption. Moreover, results show that more democratic countries tend to experience lower corruption, probably because in a democracy there are less conflicts with competing social groups.
Hoffman (2003) tests using instrumental variables whether aid influences the quality of institutions. The data are cross sectional whereby the author averages corruption, aid, and the controls over 1980-2000. As a dependent variable for the second stage different determinants of the ICRG index are used. Just as Svensson (2000), Hoffman (2003) used the log of population and infant mortality as aid instruments. In his second stage, he controls for ethno-linguistic fractionalization (just as Svensson (2000)), but surprisingly, compared to the Svensson (2000) results, the coefficient for aid is not significant. Results show that aid has a negative impact on the quality of institutions, but the aid coefficient is not significant when the dependent variable is the ICRG corruption variable.
Whereas Alesina and Weder (2000) find a positive effect of aid on corruption using OLS, Svensson (2000), using IV estimation, finds only a positive effect when a country suffers competing social groups. When there is the presence of a democracy or less competing social groups, aid has a weakly negative effect on corruption. Knack (2001) investigates - just as Hoffman (2003) - aid dependence and the quality of governance, including corruption, using cross sectional data. Data are from 1982-1995, and the author investigates by using IV whether the change in the end of period corruption measure (ICRG corruption variable) minus the initial value is caused by changes in aid, measured by ODA (Official Development Assistance) as a percentage of GNP averaged by country
over the 1982-1995 period. Knack (2001) argues that using the change in the ICRG index from 1982 to 1995 as the dependent variable is convenient because within country differences are unlikely to matter much. The authors’ OLS estimates are insignificant for the corruption variable, however the IV estimates show that foreign aid deteriorates institutional quality. Because IV aid coefficients are larger than OLS coefficient, donors do not allocate more aid to countries with worse institutions (Bräutigam and Knack, 2004).
Knack (2001) discusses the study of Svensson (2000) by using another ethnicity measure, namely that of Sullivan (1991), which uses the percentage of population belonging to the largest group in a country as an ethnic measure. This ethnic measure interacted with aid is significantly negative, which states that aid causes a better quality of institutions (instead of more corruption) with more ethnic heterogeneity. Thus the conclusion of Svensson (2000) is opposed by Knack (2001).
The study of Bräutigam and Knack (2004) is comparable to the Knack (2001) study, however it examines the quality of governance in Sub Saharan Africa only. Data are cross sectional and from 1982-1997. In their Instrumental Variables estimation they use the ICRG measure and they focus on bureaucratic quality, rule of law, and corruption. As a measure of aid average ODA as a percentage of GNP is used. The dependent variable is just as in Knack (2001) the end of period value minus the initial value for the ICRG index. The results show that aid coefficients are larger in the TSLS tests than in the OLS tests (OLS understates the coefficient of aid) and that aid increases corruption. According to Bräutigam and Knack (2004) this shows that aid donors do not allocate more aid to corrupt countries.
Coviello and Islam (2006) and Djankov et al. (2008) both use instrumental variables to test for the effects of foreign aid on the quality of institutions. Djankov et al. (2008) make use of panel data with 5-year periods. For their dependent variables data from two different periods are studied. Using the DPI database (Database of Political Institutions), they create a variable measured from 1977-1999 that captures the number of decision makers whose agreement on the issue is necessary before the policies can be changed. When this variable is low, democracy levels are low. Next to this variable they use a variable that measures the scale of democracy for 1960-1999 in a range from 0 to 10. Aid is measured as net ODA over GDP, as we also have seen in earlier studies. Terms of trade shock and the size of rents of oil over GDP are used in the second stage regression, because these may also generate institutional changes. The IV results show that foreign aid has a negative effect on the quality of institutions. To check robustness of the results with panel data,
Djankov et al. (2008) also use cross sectional data from 1977-1999. Results remain the same as with 5-year panel data.
Coviello and Islam (2006) make use of cross sectional data and take a more specific corruption variable than Coviello and Islam (2006) as a dependent variable. Among others they also use the ICRG corruption index for 1994-2002. Next to ODA over real GDP they also use Effective Development Assistance over real GDP (AID) as a measure for aid, and the difference between ODA-AID, which shows the amount of technical assistance. Country effects (fixed effects) are omitted by the authors, to be able to account for cross-country differences. Results for OLS and IV estimation show that aid has no significant effect on corruption.
Both Tavares (2003) and Kangoye (2011) oppose the former conclusions by stating that foreign aid decreases corruption. Again, IV estimation is used to estimate the effects. However, aid is instrumented using another approach. Both make use of instruments focused on the donor country, instead of instruments focusing on the aid recipient countries (used by the studies discussed earlier). Kangoye (2011) argues that this approach weakens the chance of the instruments to be correlated with the dependent variable of the second stage. Countries having weak institutions often perform also poorly in other determinants that are included as instruments (such as income, infant mortality, political violence, trade shocks).
For Tavares (2003), the time period of the data is not clear from the paper, however, the data are computed as panel 5-year averages. Again the ICRG variable of corruption is used as a dependent variable. As earlier said, aid is instrumented using another approach: the value of aid outflows of the eleven largest OECD countries is used and multiplied by a dummy variable for a common land border, a dummy for having the same majority religion and one for the same official language. Using this approach Tavares (2003) finds that aid decreases corruption, instead of the studies discussed earlier, who argued that aid increases corruption. Kangoye (2011) draws the same conclusion as Tavares (2003) by using a cross sectional dataset over 1984-2004. The ICRG corruption variable is used as a dependent variable and the main measure of aid is ODA as a percentage of GDP. On top, Kangoye (2011) investigates aid unpredictability and finds that when the amount of aid is uncertain, corruption increases. The author states that aid unpredictability is comparable with revenues from natural resources. Unpredictable aid weakens the “conditionality” effect, because with uncertainty recipient governments are less committed to reform. Also the “liquidity effect” is weakened, because the recipient governments are not being provided with stable resources. Due to these effects corruption increases.
Charron (2011) and Asongu (2012) try to dig deeper into the multilateral/bilateral differences. Charron (2011) argues that from the mid-1990 onwards there is an anti-corruption movement (ACM) going on, and therefore he only expects effects of foreign aid after this time period. As a dependent variable Charron (2011) uses the ICRG corruption variable with the time period ranging from 1986 to 2006 (panel data). For the ACM he uses before- and after 1997 data. Bilateral and multilateral ODA levels are from the World Bank, measured as ODA over GNP. The author argues that there is no perfect solution for the endogeneity problem of ODA and corruption. He argues that a proper method is to use TSLS regression with lagged aid, corruption, and control variables to take into account the possible endogeneity of corruption. As an instrument for aid lagged values of ODA are instrumented through recipient based variables. TSLS results show no significant effects of bilateral and multilateral aid in the pre-ACM period. In the post-ACM period only multilateral aid significantly reduces corruption.
Asongu (2012) uses two panel estimation techniques on data of 52 African countries with data from the African Development Indicators of the World Bank from 1996 to 2010. As a measurement of corruption he uses the ‘control of corruption’ and the ‘corruption perception’ indexes. The aid variable is net ODA over GDP. Instruments include variables focused on the aid recipients, which are used before in the literature. IV results show that foreign aid as a whole as well as both multilateral aid and bilateral aid from DAC countries significantly increase corruption. The coefficient for multilateral aid is larger than for bilateral aid, implying that multilateral aid causes higher corruption levels in Africa.
!
To sum up, IV results differ widely. Studies previously conducted make use of two different kinds of instruments, i.e. donor focused and recipient focused instruments. Most studies use cross sectional data whereby averages over the entire period are used as variables, others use data over 5-year periods in a panel, and some authors make use of 5-yearly panel data. Also there are studies that look at total aid, and studies which make a distinction between multilateral and bilateral aid. Moreover, results that use the same kind of instruments and variables show various results.
!
2.3.3. Generalized Method of Moments
When the error terms are heteroskedastic because of variation between countries in the panel, IV estimation provides no longer the best estimator, which place has now been taken by the
generalized method of moments (GMM) (Stock and Watson, 2007; Heij et al., 2004). There are several versions of these method, whereby the quality of the estimator improved over time. Arellano and Bond (1991) were the first ones to transform the linear GMM system by differencing the regressors and thereby creating the Difference GMM estimator. The Difference GMM estimator takes out possible fixed effects in the data. Blundell and Bond (1998) then augmented this to ensure that the first differences of the instrumental variables - used as instruments for aid and other endogenous variables - are not correlated with the fixed effects present in the GMM regression (see appendix 7.4). This System GMM estimator - named after the system of two equations that are created by the transformation - creates more instruments and is more efficient than the Arellano and Bond (1991) Difference GMM estimator. Windmeijer (2005) made a correction to the standard errors of the Blundell and Bond estimator and proved that otherwise the standard errors are biased downwards (Roodman, 2009). This correction is nowadays widely used.
The original (linear) GMM estimator has more instruments than regressors and therefore the number of moment conditions (each instrument has a moment condition, see appendix 7.4) is larger than the number of regressors. This causes the specification to be over-identified and therefore the system cannot be solved by instrumental variables. This GMM estimator is consistent but often biased because the instruments are correlated with the endogenous parts of the regressors which have been instrumented. The Difference and System GMM estimators can be implemented on panel data with dynamic processes (the dependent variable is, among others, a function of past values), fixed country (or individual) effects (which is not possible with cross section data used for OLS and IV), endogenous regressors, heteroskedasticity and serial correlation in the error term, independent variables which are a function of past values, a small number of time periods and a large number of countries, and also on panel data with instrumental variables based on instruments which are lags of these variables. The general model is explained in the appendix (Roodman, 2009).
!
Coviello and Islam (2006) and Djankov et al. (2008) both use - next to IV - the Generalized Method of Moments (GMM) to test for the effects of foreign aid on the quality of institutions. Their IV results (with instruments focused on the recipients of aid) show for the latter that foreign aid has a negative effect on the quality of institutions and shows for the former no significant effect of aid. However, Djankov et al. (2008) argue that IV estimators may not be efficient when there is presence of heterogeneity. According to Coviello and Islam (2006), this unobserved heterogeneity could be caused by different factors which are caused by within country variation instead of only
cross-country variation. To include the possibility of within cross-country variation next to cross-cross-country differences Coviello and Islam (2006) construct a panel data set. They make use of GMM because the GMM estimators allow to control for potential endogeneity problems of the explanatory variables, country fixed effects, and temporary measurement error.
Coviello and Islam (2006) use both the Difference estimator (Arellano and Bond, 1991) and the System estimator (Blundell and Bond, 1998) for testing and use the Windmeijer (2005) correction. Their results for the most efficient GMM estimator (the System estimator) show that, just as with their IV estimation, aid has no significant effect on corruption. Djankov et al. (2008) make use of the System GMM estimator (Blundell and Bond, 1998) with the Windmeijer (2005) correction. Their results for GMM show the same effect as with IV estimation, namely that foreign aid has a negative effect on institutions.
Charron (2011) also conducted GMM estimation based on panel data. As said earlier, GMM has the advantage of better estimates in the presence of heteroskedasticity. Moreover, compared with IV estimates, as the length of the panel increases GMM produces a larger number of valid instruments. The author uses the Arellano and Bond (1991) Difference estimator and his GMM results differ from the IV results and show that bilateral aid has a positive effect on corruption in the pre-ACM period (anti-corruption movement, see Charron (2011)), and no significant effect in the post-ACM period. Multilateral aid also has a positive effect on corruption in the pre-ACM period, however, in the post-ACM period multilateral aid causes lower levels of corruption. Therefore the overall results indeed show differences between the pre- and post-ACM period, but also significant differences between multilateral and bilateral aid.
IV results of Asongu (2012) - while using data from 1996-2010 for African countries - show that more foreign aid, multilateral aid, and bilateral aid from DAC countries significantly increase corruption. For GMM estimation the author uses the Blundell and Bond (1998) second-step System GMM estimators (see appendix section 7.4 for details). Results for this method show that foreign aid increases corruption for the ‘control of corruption’ index, but not for the ‘corruption perception’ index. The multilateral aid coefficient is - just as with instrumental variables - larger than the coefficient for DAC countries, but it is only significant for the restricted regression without a constant. Bilateral aid is only significant for the ‘control of corruption’ index, however, the magnitude of the estimate is much smaller than for IV estimation. Overall, both GMM and IV show a positive effect of aid on corruption. However, the magnitudes and the sizes of the coefficients differ.
2.3.4. Quantile Regression Method
Okada and Samreth (2012) oppose the methods used in earlier studies. They argue that estimation methods such as OLS, IV, and GMM estimate parameters of interest only at the mean. Therefore they use the quantile regression (QR) methodology to examine the effects of foreign aid at different intervals throughout the corruption distribution (see appendix section 7.5 for details). This method is especially preferable compared to OLS when there is the presence of outliers and a highly non-normal distribution of the dependent variable. Asongu (2012) states that quantile regression does not require a normally distributed error term and argues that the criticism of Okada and Samreth (2012) is valid with respect to OLS, but not when compared to some instrumental and dynamic panel estimation techniques. Therefore, Asongu (2012) uses IV and GMM estimation.
Corruption data of the study from Okada and Samreth (2012) are from the Worldwide Governance Indicators (WGI) from the World Bank. Foreign aid is measured by ODA over GDP. Data are from 1995-2009 and the authors use a panel dataset of five-year averages for each variable to be able to correct for short-term economic fluctuations. Results for foreign aid as a whole show that foreign aid reduces corruption. Its effect is larger in less corrupt countries which are represented by a lower quantile. To account for differences in bilateral and multilateral aid, the authors differentiate between bilateral aid from France, Japan, the UK and the US. Multilateral aid seems to have a larger reduction impact on corruption than foreign aid as a whole. Results for bilateral aid show that except for Japan, bilateral aid has no significant effect on corruption. The differences for Japan are possibly caused by the lack of a close relationship between Japan and former colonies. When Okada and Samreth (2012) simultaneously use all types of aid (multilateral and bilateral from four different countries) as determinants of corruption, multilateral and Japanese aid seem to reduce corruption, whereas aid from the UK and the US seem to increase corruption for some quantiles. Foreign aid from France has no significant effect.
Asongu (2013) extended his earlier Asongu (2012) paper by stating that the effects of foreign aid are possible affected by the initial level of institutional quality. Therefore in Asongu (2013) the quantile regression method is used. Again, only data from African countries are used with panel data from 1996-2010. The corruption variables used are from the World Development Indicators (WDI) and the Corruption Perception Index (CPI) from Transparency International. The results show no significant effect of aid on corruption for all quantiles. Whereas both GMM and IV showed to have a positive effect of aid on corruption (Asongu, 2012), quantile regression (which not makes use of differences between multilateral and bilateral aid) shows no effect of foreign aid
on corruption at all. This is a surprising results, which may tell something about the different methods used.
!
2.4. Why Bilateral Aid May Not Be Significant
As earlier said, there is evidence of heterogeneity among bilateral donors (Berthélemy, 2006). The absence of an effect of bilateral aid on corruption can be caused by differences between bilateral donors. For example, Alesina and Weder (2002) find that countries that have no former colonies allocate more aid to less corrupt governments. Other countries, such as the United States, have other focus points when providing aid and therefore allocate more aid to corrupt governments. Charron (2011) finds no significant effects of bilateral aid on corruption, whereas multilateral aid had a significant negative coefficient. He therefore argues that bilateral donors possible act with self interest in their aid allocation. Okada and Samreth (2012) therefore state that the allocation of bilateral aid is related to strategics and a shared history with the recipient countries. This contrasts with multilateral institutions which normally require recipients of aid to reduce their corruption. Their results show that countries with tight relationships with their former colonies care less about the quality of institutions in the recipient countries.
Among other studies, Alesina and Dollar (2000) find that the larger donor countries are more likely motivated by political and strategic goals. Aid from countries in the North of Europe showed to be more focused on poverty, democracy, and human rights (Gates and Hoeffler, 2004). Because of differences in priorities between bilateral donors and therefore differences in the results of the effects of their aid on corruption, it is possible that those different results cancel each other partly out when considering bilateral aid as a whole. This can possibly be of influence on the results for bilateral aid.
!
!
3. Research Design and Methodology
3.1. Estimation Methods
The studies earlier conducted on this topic use various methods for estimating the effect of foreign aid on corruption. This study will use the same estimation methods, namely OLS, IV, GMM and QR, to investigate whether aid influences corruption. Estimation will be done with the Stata software (StataCorp, 2011). The ways of testing the hypothesis are also based on the studies discussed earlier. By estimating all these different methods on a large dataset, which has not been done before, it can be shown whether these different estimation methods will cause different results or possible only different sizes of the coefficients. If the first is the case, then using different estimation methods largely influences the results, and therefore it will not be possible to draw one single conclusion on the effects of foreign aid on corruption.
In this study the effect of foreign aid on corruption will be estimated using panel datasets for OLS, IV, QR, and GMM. There will be one yearly panel dataset, however, by using a panel with 3-year averages (using 5-3-year averages will not lead to enough time periods for testing) it may be easier to find a clear pattern because that smoothes the data (yearly data has often a lumpy pattern) (Alesina and Weder, 2002). Therefore both a yearly panel and 3-year panel will be used. The advantage of panel data is that data for several countries can be investigated over a larger time period, which causes larger datasets that anticipate well on shocks (shocks do not largely affect outcomes). However, there are also some disadvantages for OLS and IV. For panel data the i.i.d. assumption ! or ! is often not realistic. If unanticipated shocks have an effect on the same country at all points in time, then the error terms ! and ! must be correlated for all ! . Also, if unanticipated shocks have an effect on all countries at one point in time, then the error terms ! and ! must be correlated for all ! . Therefore, also a cross sectional dataset is used for OLS and IV estimation, which is constructed by transforming the panel dataset (Davidson and Mackinnon, 2009). The most efficient GMM methods (the Difference and System GMM methods; for explanation see section 2.3.3. and Appendix section 7.4.) are designed for panel data and therefore for GMM only estimation with panel data is conducted. For QR estimation a yearly and a 3-year panel will be used to be able to examine the effects of foreign aid at different intervals throughout the corruption index. Estimation with cross sectional data for QR will produce only a small number of observations, which causes less accurate results.
E(u X) = 0 E(u Z ) = 0
uit uis t ≠ s
The cross sectional dataset is thus only used for OLS and IV testing. These results may discover differences between panel and cross sectional results, where the use of either panel data or cross sectional data for estimation is also a conscious decision (next to deciding which estimation method to use for testing). The cross sectional dataset for OLS and IV will consist of averages from 1995-2012 when making use of the CPI corruption variable and averages from 1996-2012 when making use of the WGI Control of Corruption variable (see section 3.2). The dependent and independent variables and the controls are then calculated as averages over the data period. Also the change in corruption from the data period will be used as a dependent variable with the independent variables as averages over time (Coviello and Islam, 2006; Knack, 2001; Bräutigam and Knack, 2004).
!
3.2. Equations
The equations that will be estimated in the regression analysis are discussed below. Details about the estimation methods can be found in the appendix (7.2 - 7.5).
For the cross sectional dataset for OLS and IV the following equation will be estimated:
! , (1)
where ! represents the average corruption level for country i and the! represent the average level of foreign aid for country i (either total ODA, multilateral ODA or bilateral ODA depending on the regression). The ! ’s are the controls used, the ! ’s are the initial values in the equation and the ! ’s are the dummy variables in the regression. The error term is represented by ! .
The first stage for TSLS (instrumental variables) is represented by:
! , (2)
where ! are the levels of aid for country i (either total ODA, multilateral ODA or bilateral ODA depending on the regression), ! is the constant, the ! ’s are the instruments, and the ! ’s are the controls. The error term is represented by ! .
!
For the OLS and IV cross sectional dataset with the dependent variable being the change in corruption the following equation (based on Coviello and Islam (2006)) will be estimated:
! , (3)
where ! represents the change in corruption and ! represents the initial value of corruption for country i. The ! represents the average level of foreign aid for country i (either total ODA,
yi=α + β2Ai+β3iXi+β4iη0i+β5iDi+εi yi Ai Xi η0i Di εi Ai=α +γ1iZi+γ2iWi+νi Ai α Zi Wi νi Δyi=α + β1y0i+β2Ai+β3iXi+β4iη0i+β5iDi+εi Δyi y0i Ai
multilateral ODA or bilateral ODA depending on the regression). The ! ’s are the controls used, the ! ’s are the initial values in the equation and the! ’s are the dummy variables in the regression. The error term is represented by ! .
The first stage for TSLS (instrumental variables) is represented by:
! , (4)
where ! are the levels of aid for country i (either total ODA, multilateral ODA or bilateral ODA depending on the regression), ! is the constant, the ! ’s are the instruments, and the ! ’s are the controls. The error term is represented by ! .
!
For the panels for OLS and IV (second stage) fixed effects, random effects, and first-differences estimation will be used.
Each panel regression starts with a simple model:
! , t = 1, 2, …, T, (5)
where ! represents the corruption value for country i in period t, ! is the constant, ! are the levels of aid for country i in year t (either total ODA, multilateral ODA or bilateral ODA depending on the regression), whereas! represent the controls for country i in year t. The ! ’s are the initial values for the countries in the equation and the! ’s are the dummy variables (which are constant over time) in the regression. The error term for country i in year t is represented by ! .
When testing this simple model without augmenting biased estimates will be produced, because there is not controlled for the correlation of the error terms over time for a country.
With the fixed effects method, for each country the average equation over time is estimated, where the fixed effects over time stay constant (equations are based on Wooldridge (2013)):
! , t = 1, 2, …, T, (6)
If equation (2) is subtracted from equation (1) we obtain:
! , t = 1, 2, …, T, (7)
or
! , t = 1, 2, …, T. (8)
Equation (8) is time-demeaned and therefore the (partly unobserved) fixed effects are now omitted. The equation can be tested by pooled OLS. Stata covers the fixed effects method with one simple function (StataCorp, 2011). Xi η0i Di εi Ai=α +γ1iZi+γ2iWi+νi Ai α Zi Wi νi yit =α + β1Ait+β2Xit+β3η0i+β4Di+εit yit α Ait Xit η0i Di εit yi=α + β1Ai+β2Xi+β3η0i+β4Di+εi yit− yi= +β1(Ait− Ai)+ β2(Xit− Xi)+εit−εi !!yit =β1A!!it+β2X!!it+ !!εit
The random effects model starts with the same simple model (equation (5)). Random effects assumes that the unobserved fixed effects are uncorrelated with each independent variable (Wooldridge, 2013). The error term is defined as (equations are based on Wooldridge (2013)):
! , t = 1, 2, …, T, (9)
where the ! are the unobserved fixed effects. Because the unobserved fixed effects are in the error term, the error term is serially correlated across time. Random effects transforms the data into quasi-demeaned data. An advantage of random effects is that independent variables that are fixed over time (initial values and country dummies) can be incorporated in the model (Wooldridge, 2013). Again, Stata has a function for random effects estimation (StataCorp, 2011).
The first differences method uses time dummies for each year and assumes that the idiosyncratic errors (error for each country) are uncorrelated with the independent variables in each year. This means that the independent variables are exogenous when the unobserved fixed effects are omitted (Wooldridge, 2013). Therefore the estimated equation for first differences becomes (equations are based on Wooldridge (2013)):
! , t = 1, 2, …, T, (10)
where ! represents the change in corruption for country i in two successive time periods, ! is the constant for the first time period, the other d’s are the time dummies, the ! ’s represent the changes in levels of aid for country i in two successive time periods (either total ODA, multilateral ODA or bilateral ODA depending on the regression), whereas the ! ’s represent the controls. The error term is represented by ! . The effects that are fixed over time, are differenced away in first-differences estimation.
The base equation for the first stage of TSLS (instrumental variables) is represented by:
! , (11)
where ! are the levels of aid for country i in year t (either total ODA, multilateral ODA or bilateral ODA depending on the regression), ! is the constant, the ! ’s are the instruments, and the ! ’s are the controls. The error terms are represented by ! . This equation is adjusted when used with fixed effects, random effects, or first differences respectively.
!
For the panels for QR the same equation as Okada and Samreth (2012) will be estimated: vit = ai+εit
ai
Δyit =α0+α3Δd3t+α4Δd4t+ ...+αTΔdTt+β1ΔAit+β2ΔXit+ Δεit
Δyit α0 ΔAit ΔXit Δεit Ait=α +γ1iZit+γ2iWit+νit Ait α Zit Wit νit
! , (12)
where ! . ! is the ! th quantile estimator (for details see Appendix 7.5).
Variable ! represents the corruption level for country i in one period and the ! ’s are the explanatory variables used in the regression.
!
For the panels for GMM the Difference GMM estimator (Arellano and Bond, 1991) and the System GMM estimator (Blundell and Bond, 1998) with the Windmeijer (2005) correction will be used (for details see Appendix 7.4). Equations are based on Coviello and Islam (2006):
! , t = 1, 2, …, T, (13)
! , t = 1, 2, …, T, (14)
where ! is the lag of corruption and the other variables in equation (13) have the same meaning as in equation (5). Equation (14) is the regression in differences and therefore the initial values and the dummy variables disappear (these are constant over time). Equation (13) is the regression in levels and equation (14) the regression in differences. The Difference GMM estimator consists of equation (14) with additional assumptions (see Appendix 5.4), whereas the System GMM estimator consists of the system (13) and (14) with additional assumptions.
!
3.3. Data
The data that will be used to estimate the effect of foreign aid on corruption will be selected by making use of the variables utilized in the studies discussed earlier in this thesis. Data will be collected from 1995-2012, and by constructing such large datasets several estimation methods can be tested over a long time period.
As main dependent variable the Control of Corruption variable from the Worldwide Governance Indicators (WGI) (World Bank, 2014) will be used and the Corruption Perception Index (CPI) from Transparency International (2014). The Control of Corruption index is available from 1996-2012 and the CPI from 1995-2013. Unfortunately the widely used International Country Risk Guide (ICRG) corruption variable (data from 1980 onwards) is not available without making large costs.
The Corruption Perceptions Index (from now on called the CPI corruption variable) is based on a 0-100 scale. A 0-score means a highly corrupted country, whereas a score of 100 is a country
min β ∈Rk θ yi− xi ' β + (1−θ ) yi− xi'β i∈{i:yi<xi'β }
∑
i∈{i:yi≥xi'β }∑
" # $ $ % & ' ' 0 < θ < 1 θ θ yi xiyit =αyit−1+β2tAit+β3tXit+β4iη0i+β5iDi+εit Δyit =αΔyit−1+β2tΔAit+β3tΔXit+ Δεit
without corruption (Transparency International, 2014). The Control of Corruption index from WGI (from now on called the WGI corruption variable) is also adjusted to a 0-100 scale so that a higher index value indicates a less corrupt country (World Bank, 2014).
According to Svensson (2005) the three corruption measures mentioned have a strong correlation. The Control of Corruption and the Corruption Perception Index have a correlation of 0.97 and the Control of Corruption or the Corruption Perception Index and the ICRG variable are correlated at 0.75 (2001-2003 data). The largest differences between the indexes are the years and countries covered. All corruption indexes consist of ordinal variables. The used CPI and Control of Corruption indexes are made of several rankings of different institutions.
The amount of net ODA over GNI will be the measure of aid used. Net ODA over GNI is the official ODA measure of the World Bank (2014). Moreover, next to net ODA over GNI, also multilateral and bilateral aid over GNI will be investigated. Control variables will be chosen based upon controls used before and which have been proved to be important to include in the analysis. ODA over GNI and the controls will partly be available through the World Bank Indicators (2014), whereas AidFlows (2014), which is a partnership between the OECD, World Bank, the Asian Development Bank, and the Inter-American Development Bank, has data available on multilateral and bilateral aid.
The number of countries consulted for this study are a frequent number of countries from the various regions of the world, whereas Africa, Asia, and South-America are the most important because of their low-income countries in their region. In total, data from 204 countries will be investigated.
!
3.4. Choice of Variables 3.4.1. Controls
The controls that will be used in the regressions are based on the controls used in earlier studies. Hereby a motivation for the use of the controls will be given.
The initial corruption value is used as a control by four studies that use the change in corruption as the dependent variable. Coviello and Islam (2006) state that the initial corruption value captures a regression to the mean effect and Knack (2001) states that this variable also controls for the limited ability of countries with high corruption values to increase their corruption values.
The growth of (initial) per capita GDP (log of initial per capita GDP) is used as a control in half of the studies investigated. The variable GDP per capita is available via the World Bank Indicators and can be augmented to the log of initial per capita GDP (World Bank, 2014). Kangoye (2011) argues that the effect of this control is not clear. It is possible that a larger income level creates more corruption, but on the other hand the quality of institutions improves with a higher per capita GDP. Tavares (2003) states that institutional quality should improve with a larger GDP per capita if the institutions are efficient and transparent. Therefore, the effect of the control variable is ambiguous.
To control for the ethnic situation in the countries investigated a variable called ethno-linguistic fractionalization is used in several studies. This variable measures the chance that two citizens are part of a different ethnic or linguistic group (Tavares, 2003). The presence of competing groups that are not likely to cooperate because of, for example, large cultural differences can cause more corruption in the sense that rulers may give preference to their own group members (Kangoye, 2011). The data for ethno-linguistic fractionalization in this study are from Alesina et al. (2003). In this dataset, ethnic and linguistic fractionalization are two different variables. Where the data for ethnic fractionalization are from the 1990s onwards (and some older), the data for linguistic fractionalization are from 2001.
Just as many studies choose for a control variable for the growth of (initial) population in a particular country (log initial population) as for ethno-linguistic fractionalization. The log of initial population is a proxy for donors’ interests (Svensson, 2000). The population variable is from the World Bank Indicators (2014) and adjusted to the log of initial population. Countries with a larger population are expected to have more corruption, because larger countries have more difficulties to control their institutions, which creates a larger incentive for corruption (Coviello and Islam, 2006). Svensson (2005) considers human capital also as an important determinant for the quality of institutions (and thus corruption) and proves that this variable is significantly correlated with corruption. To proxy for human capital the log of the years of schooling can be used. Data are from Barro and Lee (2014), who provide 5-year data about years of schooling.
Next to these controls several dummies will be used as controls in the regressions. One of them is a dummy for a country being a democracy or not. A country having a democracy as a political system is likely to have lower corruption levels (Charron, 2011). This dummy can be related to ethno-linguistic fractionalization. When there is the presence of several groups that cooperate, it comparable to the opposite of a democracy, where every citizen has an equal vote and
therefore the chance that corruption is present is less likely than with competing social groups. The democracy dummy is gained from the Polity IV Project (Polity, 2014). The dataset is augmented so that a democracy value of 0-5 is stated as not being a democracy and a value from 6-10 (being a democracy and full democracy according to Polity (2014)) has as dummy value 1 for being a democracy.
Another common dummy is the legal origin dummy. Data are from LaPorta et al. (2008). The legal origin dummies are available for legal origin systems from the United Kingdom, France, Germany, Scandinavia, and countries that have a socialist legal origin. This dummy defines the origin of the countries’ legal system. The dummy has value 1 for a legal system described above and zero otherwise. This control variable accounts for past influences on the quality of institutions (Coviello and Islam, 2006). For example, countries with a legal system of French origin are expected to regulate more, opposed to countries with a British system of origin. More regulation is expected to lead to more corruption (Svensson, 2005). Related to this dummy is the colony dummy, used by Tavares (2003). If a particular country has been a colony after 1825, the dummy has a value of 1. Having been a colony can possible cause more corruption, because the institutions were built to benefit the colonizer, not the original inhabitants of the country (Svensson, 2005). This dummy is composed by looking up for each country whether it has been a colony. Also regional specific dummies are used by several authors to control for unobserved regional specific effects (Kangoye, 2011). In this study dummies for the regions Sub-Saharan Africa, Latin America, East Europe, Middle East, North Africa, South and East Asia will be used. Region dummies are from the OECD Aid Statistics, where the countries per region are specified (OECD, 2014).
!
3.4.2. Instruments for Aid (IV)
In the first stage of instrumental variables, instruments for aid are used. The instruments used in this study will be based on the instruments used in earlier studies. Most studies focused on instruments that focused on the aid recipient countries. This approach will also be used in this thesis.
Just as in the second stage of TSLS, several studies (six) use the log of (initial) per capita GDP and the (initial) log of population in their first stage. Data in this study are from the World Bank Indicators (World Bank, 2014). Burnside and Dollar (2004) found that the initial log of population is highly correlated with aid. They also used the log of initial per capita GDP in their aid regressions. On average, more aid is allocated to poorer countries (Alesina and Dollar, 2000). Knack (2001) argues that the log of initial GDP per capita is an appropriate indicator of recipient
need of aid. Countries with a larger population often receive more aid than small countries (Alesina and Weder, 2002). Both are significant variables in an aid regression.
The infant mortality rate is also used as an instrument for aid by three studies. Just as the log of initial GDP per capita, Knack (2001) sees the infant mortality rate as an indicator of recipient need of aid. With higher infant mortality rates, a country probably needs more aid. Countries with high rates of infant mortality are often the poorest countries with severe hunger and diseases. Younas (2008) found that donors focus on physical miseries and therefore they allocate more aid to countries with high infant mortality rates. Bräutigam and Knack (2004) also consider the illiteracy rate as an indicator of recipient needs. With high illiteracy rates people are often poor(ly educated) and therefore these countries receive more aid. The illiteracy rate variable is adjusted from the literacy rate variable from the World Bank Indicators (World Bank, 2014).
Colony and region dummies are used - just as they are used as controls for corruption - by several studies. Donors are likely to give more aid to former colonies (see section 2.1). The region dummies again control for unobserved regional specific effects (Kangoye, 2011). Alesina and Dollar (2000) consider a democracy dummy as a possible influential variable in aid allocation. Some donors are proved to allocate more aid to countries with a democracy. Therefore, this dummy is also used in the first stage. The sources of the dummies used in this study are mentioned in section 3.4.1.
!
3.4.3. Instruments for Aid (GMM)
As earlier said, GMM uses internal instruments. Internal instruments are instruments which are made of lagged values of the explanatory variables (for the difference estimator). For the regression in levels instruments are based on the lagged differences of the explanatory variables (Coviello and Islam, 2006). Coviello and Islam (2006) use lags (t-2) to (t-4) to avoid the model to be having more instruments than countries (over fitting). Lag (t-1) is not used because there is often too much correlation with time t. The number of lags used in the other studies using GMM are not clear. In this study also lags (t-2) to (t-4) are used.
!
3.5. Prediction According to the Literature
As seen in the literature review, the results of the conducted studies differ widely. There seems to be no specific trend in the results. Some studies find a positive effect of foreign aid on corruption, whereas others find the opposite or an insignificant effect. The effect can also differ for various
quantiles of the distribution. Therefore it is hard to predict test results. It will be interesting to investigate whether using the same variables but various estimation methods will cause large differences in the outcomes of the variables.
!
!
!
4. Regression Results
4.1. Introduction
Below the results of the various estimation results are discussed. First the cross sectional OLS and IV results are evaluated and then the panel OLS and IV results, QR results, and GMM results are discussed. In the introduction of the result analyses the estimation method is further introduced, whereafter the results of the corresponding methods are discussed. Tables with the most relevant information are presented in the text. The complete tables with full information are presented in the appendix section 7.6. All standard errors of the regressions are in parentheses. Stars indicate the following: * = p < 0.05, ** = p < 0.01, *** = p < 0.001. Bold variables and -test results indicate respectively a significant variable or test. Tables in the text have the same table numbers as in the appendix.
!
4.2. Cross Sectional OLS and IV Analysis 4.2.1. Introduction
As earlier said, two variants of cross sectional regression are considered. First, average corruption is regressed on averages of the independent variables (except for the dummy control variables, initial values and values that are measured at one point in time). Data from 1995-2012 are used for the CPI corruption variable, and data from 1996-2012 for the WGI corruption variable. Second, a cross sectional regression will be held over the change in corruption over the data period. Again, average values of the independent variables are used in the regression. Three time periods are considered to be able to see whether results differ when the time periods are becoming smaller and more recent. For the CPI corruption variable relatively few data are available for 1995-2002 and therefore the time period 1995-2012 is excluded for this dependent variable. Therefore, the CPI corruption variable the periods 2002-2012 and 2007-2012 are used. For the WGI corruption variable the time periods 1996-2012, 2002-2012, and 2007-2012 are used for the cross sectional regression.
Cross sectional regression is used for both ordinary least squares and instrumental variables estimation. Robust standard errors are produced to correct for possible heteroskedasticity. With OLS estimation the R-squared for the goodness of fit of the model is used, the t-tests for checking the significance of the independent variables, and an F-test for the overall significance of the independent variables. With IV estimation the t-tests and F-tests are also considered and post estimation test are used. To be able to decide whether the instruments are valid, tests for instrument relevance and instrument exogeneity are used. For instrument relevance a first stage F-statistic is
