The effect of new scientific knowledge about effective aid allocation on actual allocation policies

(1)

The effect of new scientific knowledge about

effective aid allocation on actual allocation

policies

Master thesis Economics

Supervised by Prof. E.J.S. Plug

Jessica Wagenvoord

University of Amsterdam

September 30, 2015

Abstract

Burnside and Dollar (2000) have been very influential with their publication in which they combine results in the aid effectiveness literature with economic growth literature into one hypothesis. They showed that aid can only lead to economic growth when the recipient country has sound economic policies. This study investigates whether their publication has been influential on real aid allocation policies, by examining whether aid allocation policies have become more conditional on policies in recipient countries after the publication by Burnside and Dollar (2000). This is done by a linear regression model with donor and year fixed effects and aid determinants as control variables. The results show that the publication of Burnside and Dollar (2000) did not change aid allocation policies with respect to the size of aid commitments. However, the publication did cause policies in recipient countries to play a more important role when donors decide to refrain from providing aid to a country.

(2)

1

1 Introduction

Donors spend billions on aid in order to reduce poverty and to increase welfare worldwide. Whether this money is spend effectively, is very relevant for policy makers. There has been discussion on the effectiveness of aid among both politicians and economists. Politicians of both sides have doubted the effectiveness of the provision of foreign aid to developing countries (Burnside & Dollar, 1997). Left-wing politicians on one hand argue that aid agencies urged unwilling governments to take up certain policies, which in the end did not fulfil the promises made. Right-wing politicians on the other hand argue that aid supports large and inefficient governments, which creates a bad economic environment.

Also economists have addressed the effectiveness of aid in the literature in the past. By now, the literature regarding aid effectiveness is substantial, more than 100 studies have been devoted to it since 1960. Some of them found that foreign aid is effective (Hansen & Tarp, 2001; Lensink & White, 1999; Hadjimichael et al., 1995 among others), while others found it is not effective (Boone, 1996; Rajan & Subramanian, 2008 among others). Within this literature on aid effectiveness, there is one study which has been extraordinarily influential; the study by Burnside and Dollar which was published in 2000.

The influence of Burnside and Dollar (2000) is illustrated by the extraordinarily high number of citations, which is eight thousand for the final publication besides the three thousand citations of the working paper. Furthermore, their results and implications have been incorporated in multiple publications by the World Bank (1998, 2002a,b) and in a paper by the U.K. Department for International Development (2000) which was presented to the U.K. Parliament. Moreover, the results received some attention in one of the speeches by President George W. Bush’s (2002), and in the announcement by the White House on creating the Millennium Challenge Corporation (White House, 2002). Lastly, their results were described in recognized newspapers, such as in the Economist (March 16, 2002), the Washington Post (February 9, 2002), and in a column in the Financial Times (March 11, 2002).

Before the publication by Burnside and Dollar (2000) the literature was inconclusive about the effect of aid on economic growth. In another string of the literature, researchers found that economic growth itself is conditional on economic policies. Burnside and Dollar (2000) combined these two hypotheses and examined whether they are related. They found that aid only has a positive effect on economic growth when the recipient country has sound economic policies. Their study has been innovative with the inclusion of new variables,1_{new institutional}

(4)

3 and political variables and an interaction term of economic policies and foreign aid.2

So, Burnside and Dollar (2000) found that aid is allocated effectively when it is provided to countries with sound policies. As part of their study, they explored whether aid has been allocated as such in the past. They found that multilateral aid has been allocated conditionally on policies between 1970 and 1993, but bilateral aid has not. Since in this period two third of total aid was bilateral, they basically argued that two third of total aid provided to lower income countries between 1970 and 1993 was allocated ineffectively.

This study contributes to the literature by investigating whether the publication by Burnside and Dollar (2000), has not only been influential in the literature but also on actual aid allocation policies. There is reason to believe policymakers took up the policy advice following from the results because of the large amount of attention it received and more specifically because the Monterrey Consensus (United Nations, 2002) and the research policy report Assessing Aid (World Bank, 1998) have supported their findings. In addition, they published their research as a World Bank policy research working paper, and World Bank’s influence on the common definition of good governance and aid allocation policies is reasonably large.

In the literature on aid allocation, no study explored the translation of knowledge following from scientific studies into actual aid allocation policies before. This study aims to fill this gap in the literature. As Burnside and Dollar (2000) have been the first to find evidence for the conditionality of aid effectiveness on policies and due to the fact that it has been so influential within the aid literature, their study is a suitable candidate to find influence of scientific results on actual policies.

The research question of this study is: Is the increased awareness about effectiveness of

aid, when allocated conditionally on policy characteristics in recipient countries, reflected in the data?

The research question is answered by examining the relation between aid allocation and policies in recipient countries, specifically by investigating whether this relation has changed after an influential publication in 2000. This is done with the use of a linear regression model with multiple fixed effects. The variable of interest, an interaction term consisting of a time dummy and a policy indicator, is regressed on both bilateral and multilateral aid flows. This interaction term shows the selectivity of aid allocation on policies after the publication, compared to the years before the publication.

The results show that, the increase in awareness of effective aid allocation if policies in

2_{The use of an interaction term was a new way to capture the non-linearity in the aid-growth relationship. They}

(5)

4 the recipient country are taken into consideration did not affect the size of aid flows. The awareness did, however, affect donors’ decision whether to provide aid to a certain country or not. Policies in recipient countries became more important for donors when they decide not to provide a country with aid.

In order to structurally answer the research question, chapter two provides a literature review and considers methodologies of previous studies on which the methodology of this study is based. The data are described in chapter three and chapter four provides the methodology of this study. Ultimately, chapter five describes the results and this study is concluded in chapter six.

2 Literature review

This chapter will provide an overview of the literature on aid allocation and it describes the methodologies and results of relevant studies. The last part of this chapter will be devoted to the translation from scientific research outcomes into actual policies.

2.1 Previous studies and their results

Several macro studies in the nineties of the twentieth century explored the determinants of economic growth in developing countries. They found that growth is dependent on certain policies and economic environments, such as the level of inflation rates, size of the budget surplus and trade openness (Easterly & Rebelo, 1993; Fischer, 1993; Sachs & Warner, 1995). At the same time other studies were exploring whether aid leads to economic growth. Some of them found evidence to contradict this relation (Boone, 1994; White, 1992 among others), but others found evidence to confirm this (Hanushek & Wößmann, 2007; Bloom, Canning & Sevilla, 2004; Lensink & White, 2001). Therefore, the literature was inconclusive about the effect of aid on economic growth.

So, one the one hand, results show that economic growth is conditional on policies and on the other hand results show that aid might lead to economic growth. From combining these two relations follows the hypothesis that the effect of aid on economic growth is conditional on economic policies.3_{This is the hypothesis which is researched by Burnside and Dollar (2000).} They found evidence confirming that aid leads to economic growth when the recipient country has sound economic policies. When exploring whether aid has been effectively in the past, they

3_{The rationale put forward by the authors is that policies which affect growth directly, make poor countries}

(6)

5 found that this was only the case for multilateral aid (provided by governments but channelled through institutions like United Nations’ agencies, World Bank and NGO’s) and not for bilateral aid (provided by a single country to a single country). Martinez and Winters (2015) found the same result for the period 2004-2010.

The World Bank and the United Nations are important multilateral donors. The World Bank already promoted the idea of conditionality of aid on policies in the study ‘Sub-Saharan Africa: From Crisis to Sustainable Growth’ which was published in 1989. This is the earliest evidence of conditionality of aid allocation on policies in the literature. In 1998, the World Bank published a policy report, again discussing this relation. This report also presents preliminary results of Burnside and Dollar (2000). The United Nations supported the idea of selectivity of aid allocation on policies by including it in the Monterrey Consensus (United Nations, 2002).4

Collier and Dollar (2002), Claessens, Cassimon and Van Campenhout (2009), Martinez and Winters (2015) and Dollar and Levin (2006) found the same result as Burnside and Dollar (2000) but for bilateral aid allocations in 1996, 1999-2004, 2004-2010 and 2000-2003 respectively. These results are contradicting with those of Easterly (2007) who found that bilateral aid has not been allocated conditionally on policies, between 1960 and 2003.

To sum up, the majority of the literature provides evidence indicating that bilateral aid flows responded to the quality of policies and institutions in the recipient country from the mid-nineties of the twentieth century. For multilateral aid this selectivity already started earlier.

2.2 Previous studies and their methodologies

2.2.1 Overview

There have been multiple studies in the past which explained aid allocation as a function of characteristics of the donor or recipient (Bermeo, 2010; Claessens et al., 2009; Collier & Dollar, 2002; Dollar & Levin, 2006; Easterly, 2007; Martinez & Winters, 2015 among others). The relation between aid and these characteristic was examined by linear cross-country regressions with multiple explanatory variables. Where this study is focused on policy are other studies focused on the relation between aid and other characteristics, such as geopolitical importance,

4_{The Monterrey consensus was the outcome of the United Nations International Conference on financing for}

development, which took place in spring 2002. More than 50 heads of state, representatives of the World Bank, the International Monetary Fund (IMF) and the World Trade Organisation (WTO) attended the conference and a new partnership for global development was conceived. The aim of the conference was reflect the development goals set in the previous decade, with a special focus on the aim of halving the number of people living in absolute poverty by 2015.

(7)

6 commercial links, developmental needs and potential of the recipient country. Obviously the explanatory variable of interest, such as policy, should be captured in the regression. Studies which tried to explain aid allocation by policies have used measures such as inflation, trade openness, democracy and rule of law.

In addition to the explanatory variable of interest, those related studies have included more or less the same control variable. Firstly, all relevant studies control for the need for aid of the recipient country, and therefore include measures for income and population size as explanatory variables. Income is the best proxy for the development of a country, and studies control for population size because there tends to be a bias in aid allocation towards small countries (Claessens et al., 2009). In previous studies, these variables are in general found to be large and significant. Secondly, besides controls for need for aid other studies have included measures of dyadic connections, characteristics specific for a certain pair of donor and recipient country. These measures control for commercial ties or alliances between the donor and recipient country as those might bias aid flows (Bermeo, 2010; Claessens et al., 2009; Martinez & Winters, 2015; Dollar & Levin, 2006). Thirdly, previous studies have also included measures of the rate of external debt to GDP, and the rate of investments to GDP (Martinez & Winters, 2015; Claessens et al., 2009).

A drawback of this method is that the results may be sensitive to the choice of the regressors. The next paragraph will discuss comprehensively the methodology of two studies specifically, as these are most relevant for the methodology of this study.

2.2.1 Especially relevant

Two studies are especially relevant for this research although both for different reasons. And since the methodology of this research is also based on the methodologies of these two studies, special attention is devoted to the two.

Claessens et al. (2009) is specifically relevant because it analyses the change in aid selectivity over time, from 1970 to 2004, which is also the aim of this study. In addition, Martinez and Winters (2015) is the latest publication in the aid allocation literature and is especially relevant because they analyse multilateral aid separately from bilateral aid.

Martinez and Winters (2015) and Claessens et al. (2009) as well as all studies described in previous paragraph use a linear regression model with fixed effects and control variables. Martinez and Winters (2015) use a single fixed effects model where Claessens et al. (2009) use a multiple fixed effects model. Claessens et al. (2009) include fixed effects for both donors and recipients to account for any time-invariant historical, geographical, political and cultural

(8)

7 influence that will lead to deviations from average aid flows. Time fixed effects are included to control for changes over time which are unrelated to policy selectivity and aid determinants. The control variables in both studies are similar to each other and also similar to those in the studies described in the previous paragraph.

Again, Claessens et al. (2009) is especially relevant for this study because of its analysis of aid allocation selectivity over time. They analyse this by interacting time dummies with aid determinants. They split up the period they apply their analysis on in several smaller time frames. When determining the structural breaks of these time frames, they take events into account that might have been relevant for aid allocation policies. Neither Claessens et al. (2009) nor other studies using structural breaks treated the publication of Burnside and Dollar (2000) as a relevant event for their study.

Martinez and Winters (2015) made a distinction in donor types because they believe the decision making is different for these types of aid. This is consistent with the view held by Burnside and Dollar (2000) who also separated bilateral from multilateral aid, and found significantly different results for both

2.3 The translation of research outcomes to policies

This study aims to find evidence of a successful translation of the results of Burnside and Dollar (2000) into aid allocation policies. Arguments in the first paragraph will state why general issues would lead to unsuccessful translation of research into policies. In contrast, the second paragraph will argue why successful translation is likely in the case of Burnside and Dollar (2000).

2.3.1 General issues

There is no universal framework on successful communication of conclusions evolving from scientific research. Many researchers and policymakers face problems when exchanging research outcomes, policy messages and implementing evidence-based interventions (Faso, 2014). This can be caused by several issues.

Firstly, there might be a lack of communication between researchers and policymakers. Secondly, there may be difference in timing between policymakers and researchers. Executing a study and analysing data and results usually takes several years, while policymakers may be restricted to a short period of being in office. Thirdly, the relation between researchers and policymakers might suffer from mutual mistrust. Fourthly, the study

(9)

8 might not convince the policymaker due to bad quality or inconclusive and conflicting

findings (Mitton et al., 2007).

2.3.2 The translation from ‘Aid, policies and growth’ into policy

This paragraph will state why there is reason to assume that Burnside and Dollar did overcome (part of) the issues presented previously.

Firstly, their study presents a simple recommendation for aid allocation policymakers, which is clearly stated in the conclusion.

Secondly, the working paper of ‘Aid, policies and growth’ was presented as a World Bank policy research working paper. It was meant to encourage exchange of ideas around this subject while work was still in progress.

Thirdly, both Burnside and Dollar were working at the World Bank when they published their study. As a consequence, most of the influence was channelled through the World Bank. The influence of the World Bank on a common definition of good governance and aid allocation policies is reasonably large. Since the seventies of last century, the World Bank has been working on the ‘right definition’ of good governance. In this process were three indicators developed, of which one is the CPIA index, used by Burnside and Dollar (2000) and another is the WGI index, which will be used in this study. Diarra and Plane (2014) showed that the World Bank influences bilateral aid allocation. To illustrate, Northern European countries have a large part of their aid funds to be managed by the World Bank and the other part is significantly influenced by one of the World Bank policy indicators. Diarra and Plane (2014) show that the World Banks influence on multilateral aid is not only limited to its own funds but that several other multilateral donors allocate their funds based on one of the three governance indicators of the World Bank.

Fourthly, next to the World Bank, there was another channel through which the implications of the publication could reach policymakers, namely the Monterrey Consensus. More than 50 heads of states were present at the United Nations conference which resulted in this consensus. This basically means that the implications were directly presented to and supported by many donor and recipient countries already, possibly influencing bilateral aid allocation policies. The fact that this conference was organised by the United Nations makes it likely that the recommendations from the Monterrey Consensus, including the findings of Burnside and Dollar (2000), also have been taken into account in aid allocation policies of the United Nations. Many of the multilateral donors are departments of the United Nations, such as UNHCR and UNICEF.

(10)

9

3 Data

This chapter describes the different variables, the data and its sources which are used in this study. In addition, table A in the appendix provides a clear overview of the definitions and the data sources of the different variables, table 1 at the end of this chapter provides the descriptive statistics of the variables.

3.1 Dependent variable

This study analyses the decision making process regarding aid allocation in donor countries. The dependent variable should therefore reflect the amount of aid provided by the donor. Aid commitments reflect how much a donor is willing to provide and is focused on the donor side of aid allocation. It reflects the decision-making process for aid allocation within donor governments and aid agencies, and is therefore the best candidate for the dependent variable of this research. The alternative would be the total amount of aid received, but it is more likely that this measure reflects allocation policies with noise. Data on multilateral and bilateral aid, is retrieved from AidData, an agency gathering all relevant data on aid (AidData, 2015). The aid research release 2.1 dataset covers over 1 million aid activities funded by more than 80 donors from 1940 to 2013 and has multiple sources (Tierney et al., 2011).

Aid flows are recorded as aid commitments on project level. More specifically, the aid observations are the amount of aid which the donor agreed upon to provide for the duration of a project. The actual payment may be spread over the following years, but the moment at which the aid is committed is recorded in the dataset (Tierney et al., 2011).

Aid commitments recorded in the research release 2.1 by AidData are presented at 2009 prices and exchange rates (Tierney et al., 2011). The aid flows to developed countries are left out, as this study is focused on aid to developing or low income countries. In 1989, the World Bank divided countries into categories based on its GDP per capita level at that time. Since then, the thresholds have been corrected on a yearly basis for inflation. In 2009 were countries with a level of GDP per capita of USD 12.195 or higher categorized as developed countries (World Bank, 2015b). Consequently, all recipient countries with a GDP per capita level in 2009 categorized as developed countries are excluded from the study. The data are a three-dimensional panel consisting of aid commitments made by 87 donors, both multilateral and bilateral, to 117 countries. The dependent variable is used in logarithmic terms, in order to address skewness in the distribution (Neumayer, 2003 among others).

(11)

10

3.2 Independent variables

This study examines the change in the dependency of aid allocation on the quality of policies in the recipient country. Therefore, the main variable of interest of this study is the level of policies and governance in recipient countries. From the literature review in chapter 2 follows, however, that aid flows are also dependent on other characteristics of the donor and the recipient country. More specifically, aid flows are assumed to be related to the following characteristics:

 Policies in the recipient country: Proxied by the score in Worldwide Governance Indicators

 Donor’s strategic interest: Proxied by the size of bilateral trade, alliances and a shared colonial past.

 Recipient’s need for aid: Proxied by income, debt rate and investment rate

 Population size

The motivation for the inclusion of these characteristics and a description of the datasets will be given below.

3.2.1 Policies in the recipient country

The variable of interest is proxied by the average score in the Worldwide Governance Indicators (WGI). These indicators, recorded by The World Governance Indicator Project and initiated by the World Bank, measure the level of governance and policies in a country by using the perception of governance. It reports aggregate and individual governance indicators for 215 economies over the period 1996-2013, for six dimensions of governance (World Bank, 2015a). The six different indicators are: Control of corruption, government effectiveness, political stability, voice and accountability, regulatory quality and rule of law. These are based on hundreds of underlying variables which reflect perceptions of governance of thousands of survey respondents from the public, private and NGO sectors worldwide. The indicators are recorded in units of a standard normal distribution, with mean zero, standard deviation of one and run approximately from -2,5 to 2,5. Higher values correspond to a better perception of governance (World Bank, 2015a).

Because of the unobservable character of governance, one does not have many possibilities other than relying on the subjective perception of governance. Also agents base their actions on these perceptions (Kaufman, Kraay & Mastruzzi, 2011). For example, the level of corruption is impossible to measure accurately. If a donor finds it likely that part of the aid commitment is lost because it perceives the government of the recipient country as corrupt,

(12)

11 their perception of corruption may negatively influence aid allocation to that country.

Due to this unobservable nature of the actual level of governance in a country, any indicator for governance is imperfect. There is no reason to assume that the WGI are an exception to this. As a consequence, the observations include margins of error for each governance estimate (World Bank, 2015a). Kaufman et al. (2011) argue that users of the WGI should be careful when interpreting the data, because of these margins of error. They, however, found that even after taking these errors into account it is possible to make meaningful comparisons across countries and over time.

Moreover, Hout (2007) states that the WGI are explicitly used by donors, both in aid allocation formulas as well as in the discussion whether a country should receive aid at all. To sum up, the perception of governance is a good candidate to proxy the level of governance, there is reason to believe that the WGI are sufficiently accurate to be suitable for comparisons across countries and over time and the WGI scores are actually used in aid policy making. These three arguments make the WGI suitable indicators for policies in order to answer the research question of this study.

In order to get an overall view of the perception of governance in the recipient country, and to conveniently limit the number of variables of interest to one, an average of the six indicators is determined and yearly recorded on country level. The policy indicator in this study will therefore be the average WGI score of a country.

3.2.2 Donor’s strategic interest

The literature showed that bilateral donors are likely to be responsive to their own interests (McKinlay & Little, 1979; Martinez &Winters, 2015; Claessens et al., 2009). When analysing bilateral aid, one should therefore control for these factors. In the literature studies

consistently find that donor interests play a substantial role in determining the size of aid flows, and how aid is allocated (Alesina & Dollar, 2000; Fleck & Kilby, 2010 among others). Based on previous literature, bilateral trade, colonial past and alliances are the variables which will control for donor’s strategic interests in this study.

Bilateral trade is the first strategic interest variable. The size of bilateral trade, the sum of total imports and exports, between the donor and recipient country is used as a proxy for commercial ties. The data on bilateral trade are gathered from the Direction of Trade Statistics (DOTS) dataset, reported by the International Monetary Fund until 2014 (International

Monetary Fund, 2015). DOTS provided 15 datasets with yearly import and export data between all countries, and these are merged into one dataset.

(13)

12 The second variable which is used as a control for donor’s strategic interest is the colonial past of the recipient country. This is a binary variable which indicates whether the recipient country is a former colony of the donor country is included, to control for historical ties between the donor and the recipient. The data for this variable are gathered from the Quality of Governance dataset (Teorell et al., 2013).

The last variable, alliance, will control for formal military ties between the donor and recipient country. It indicates whether the donor and recipient country are allies at the time of the aid commitments. In the Correlates of War Database (Gibler, 2013) are all formal

alliances recorded which existed from 1816 till 2013, including mutual defence pacts, non-aggression treaties, and ententes.

It is important to make a distinction between multilateral and bilateral donors with respect to this control variable. For multilateral aid, strategic interest variables are irrelevant aid determinants since it is typically less tied to political self-interest or donor strategic interest. Therefore, these strategic interest variables are only included in the bilateral aid analysis and are left out in the multilateral aid analysis. But, one can argue that policies and institutions are different when a country has been colonized by one of the Western powers, compared to the situation in which it has not been colonized. Therefore, when multilateral aid is considered a binary variable is included which indicated whether the recipient country has been colonized by a Western Power.

3.2.3 Recipient’s need for aid

Previous studies have shown that donors respond to the need for aid in the recipient countries. The need for aid has been proxied by different variables in previous literature, those used in this specific study are discussed in next paragraphs. In this study, measure for income, debt rate and investment rate will act as controls for the recipient’s need for aid.

Income is the first variable to control for the need for aid. Previous research confirms that the poorer a country is, the more likely it is to receive aid from donors. More importantly, the literature has also shown that the poorer a country is, the more likely it is to be badly governed (Acemoglu, Johnson & Robinson, 2000; Svensson, 2005). This makes the variable income a possible confounder for the analysis of this study, it should therefore be included as a control variable in the regression. A common and useful proxy for income is the level of GDP per capita. The data on GDP per capita come from the World Development Indicators dataset as recorded in the Quality of Governance dataset (Teorell et al., 2013). This dataset ranges until 2014 and is recorded by the Quality of Government Institute, University of

(14)

13 Gothenburg. Income enters the regression in log terms because of skewness in the distribution and because it is likely that changes in aid flows are proportional to changes income

(Martinez & Winters, 2015).

The second and third variables to control for need for aid are the debt rate and investment rate. Another way of observing the need for aid is by looking at the investment rate, the ratio of the size of investments within a country to its GDP, and the debt rate, the ratio of the size of external debt of a country to its GDP. If a country has a high debt rate, the need for aid is likely to be high. But a high debt rate might also affect the amount and type of aid it gets and the level of policies in a country. Data for both variables also come from the World Development Indicators as recorded in the Quality of Governance dataset (Teorell et al., 2013). The debt rate is recorded as the present value of external debt as a ratio to GDP, while the investment rate is recorded as the ratio of internal investments to GDP.

3.2.4 Population size

Countries with large populations, like India, require more aid than countries with small populations, like Malawi. However, studies have found that countries with small populations receive relatively more aid than countries with large population, possibly because economies of small countries tend to be more open than economies of large countries (Alesina and Dollar, 2000). Due to the possible existence of a factor which is related to both policies in the recipient country and aid, population size must be included as control variable in the analysis. The variation in population size is extreme, as the smallest population consists of less than 10.000 people while China’s population goes up to 1,3 billion. Since almost no countries have a population between 300 million and 1 billion is the variable used in log terms in the

regressions, in order to address skewness.

The time range of the World Governance Indicator project and the time range of the AidData research release 2.1 restricts this study to the years between 1996 and 2010. All six different datasets are merged, and the data are collapsed to the donor-recipient dyad level for all years separately, which makes the dataset ready for use. These modification leads to more than 54.000 aid observations from donor A to recipient B in year X. When also the zero observations are included, this leads to almost 338.000 observations.

(15)

14 Table 1: Variables and Descriptive Statistics

Variable

Number of

observations Mean

Standard

Deviation Minimum Maximum

Dependent variable

All aid commitments (in US dollars)

337.792 7,85 mln 182,9 mln 0 51,6 bln

Non-zero aid commitments (in US dollars) 51680 51,3 mln 465 mln 4,97 51,6 bln Independent variables Policy 197.379 -0,115 0,892 -2,494 1,986 Control of Corruption 198.926 -0,108 0,965 -2,057 2,586 Government Effectiveness 198.926 -0,108 0,966 -2,454 2,408 Political Stability 198.653 -0,104 0,993 -3,321 1,665 Rule of Law 201.565 -0,121 0,970 -2,670 2,002 Regulatory Quality 199.017 -0,110 0,969 -2,675 2,247

Voice and Accountability 201.565 -0,0987 0,993 -2,279 1,826

Income 234.507 10.501 12.327 107,5 74.164 Population 248.794 34,16 mln 129,0 mln 9.264 1,338 bln Investment rate 244.699 61,73 60,10 8,215 1,669 Bilateral trade (in US dollars) 522632 250 bln 1567 bln 0 53.160 bln Alliance 315.315 0.0229 0.150 0 1 Debt rate 159.887 68,94 8,35 2,013 1.381 Colonized 337.792 0,748 0,434 0 1 Former Colony 337.792 0,00546 0,0737 0 1

4 Methodology

This chapter describes which empirical model will be used to answer the research question. It provides a detailed description of the model while motivating choices made with respect to the structure of the regressions and using the variables described in previous chapter. In addition, the paragraphs describing the regressions within the model are clarified by concluding each paragraph with the regression equations.

4.1 Empirical model

This study is aimed to find out whether aid allocation has become more conditional on policies in recipient countries after the publication of Burnside and Dollar (2000). The methodology is mainly based on the different studies described in the literature review, and specifically on Martinez and Winters (2015) and Claessens et al. (2009).

One is able to research the selectivity of donors on policies by regressing policy on aid. To answer the research question of this study, however, one should be able to compare

(16)

15 selectivity of donors between two periods, before and after the publication. The first period, 1996-2000, represents aid allocation policies before the publication of the study by Burnside and Dollar (2000). The second period, 2001-2010, represents aid allocation policies after the publication of the study by Burnside and Dollar. Generating a dummy variable which takes up the value one when the donor committed to provide aid after the publication, helps to find the answer on the research question. This dummy variable is interacted with the policy indicator. The interaction term is the variable of interest, since it indicates the relation between policy in the recipient country, aid flows and the publication year of Burnside and Dollar (2000). Its coefficient will be positive when aid has been allocated more conditionally on policy since the publication of Burnside and Dollar (2000) and vice versa for a negative coefficient. Studies which researched aid allocation selectivity before have used fixed effects models for it. Fixed effects are used to prevent the regression to suffer from omitted variable bias. This bias appears in the estimates of parameters in an incorrectly specified regression since the independent variable is left out while it is correlated with both the dependent variable and one or more independent variables. Also here, a cross country linear regression model with fixed effects is a good candidate, because it tries to explain aid flows between donor x, recipient

y at time t, using a matrix of explanatory variables (Baltagi, 2001).

The inclusion of donor fixed effects accounts for the variation in total aid provided by different donors. It controls for unobservable time invariant characteristics of donor countries and it also means that the coefficient estimates are based on within-donor comparisons.

Including year fixed effects controls for general changes over time which are unrelated to policy selectivity and possible other aid determinants which are included in the regression. Recipient country fixed effects would control for unobservable time invariant characteristics of recipient countries, but these are not included in this study. While fixed effects solve omitted variable bias problems, it can reduce part of the signal in the data by taking up too much of the variation in the variables of interest. This is the case with the inclusion of the recipient country fixed effects, it takes out too much variation in the aid and policy variables, which makes it hard to estimate the relationship between the two. Table B in the appendix shows the results for three single fixed effects models, which are identical except for the entity which is used for the fixed effects. It shows that the coefficients for the policy variable and the interaction term do not differ much when one considers the year and donor fixed effects, but they attenuate reasonably when the recipient fixed effects are used.

Claessens et al. (2009) include lagged values of explanatory variables to prevent the regression to suffer from simultaneity bias, e.g. aid flows might increase GDP per capita or

(17)

16 investment rates. This is not necessary in this study because the coefficients of the aid determinants represent an average value of five or ten years for each dyad.

To sum up, the empirical model for this study will consist of linear multiple fixed effects regressions with control variables in which the variable of interest is an interaction term between the policy indicator and a time dummy indicating the period after the publication by Burnside and Dollar (2000). This is clearly illustrated by the regression models at the end of this chapter.

4.2 Distinction between multilateral and bilateral

The analysis in this study is structured in such a way that bilateral aid and multilateral aid are separated. Multilateral aid is provided by governments but channelled through institutions like United Nations’ agencies, World Bank and NGO’s. Bilateral aid is provided by a single country to a single country. It is important to make this distinction because it is likely that aid allocation policies and decision making processes are not the same for both types of aid, which would lead to biased results if one is not separating it. Earlier research has for example shown that multilateral aid donors are less likely to coordinate on any particular strategic interest than bilateral donors (Girod, 2008). Also, there might be a difference in the (speed of) uptake of scientific outcomes, and both donor types might prefer different types of aid which might lead to different outcomes. The difference in multilateral and bilateral aid has been put forward in several earlier studies (Girod, 2008; Martens, Mummert, Murell & Seabright, 2002 among others). Martinez and Winters (2015) and Burnside and Dollar (2000) also made this distinction in donor types and did a similar analysis.

4.3 Zero observations

An important issue in the aid literature which also has to be inclined here is the fact that donors tend to allocate aid to specific target countries. Because of this, the majority of aid commitment flows in datasets are equal to zero. These observations represent decisions by donors not to provide aid to certain recipient countries. This specific characteristic of aid data might introduce selection bias into the regression. This could happen on the donor side simply because very few is known about the recipient country and therefore aid is not provided. This could happen on the recipient side when the government of the receiving country refuses to engage with the donor because of political reasons, and therefore aid observation is zero (Claessens et al., 2009). From table 1 follows that 85% of the aid commitment observations in this specific dataset are

(18)

17 equal to zero. The literature provides several methods to overcome this issue.5_{Berthélemy and} Tichit (2004) and Bertélemy (2006), however, have shown that the differences in estimates between the different methods are small, when one is using a three dimensional panel dataset for an aid allocation analysis. Others (Linders & De Groot, 2006) explicitly tested this and concluded that omitting the zero observations from the sample is not favourable econometrically, but it leads to acceptable results. From this follows that omitting the zero aid flow observations in a linear regression model will lead to acceptable results and therefore this study will refrain from using other complex econometrical methods to overcome this issue of selection bias.

However, the zero aid flow observations do contain information as well about the decision making process on aid allocation by donors, since not providing aid is also a decision. Therefore, these observations should not be ignored by excluding them completely from this study. As a consequence, the regression analysis of this study will consist of two parts, one addressing only the nonzero observations and the other part addressing both the zero and nonzero observations.

4.4 How much aid to provide?

When one considers a dataset consisting of only the nonzero aid flow observations for the dependent variable, one can analyse how the size of an aid flow is dependent on the level of policies in recipient countries. The decision whether to provide aid to a certain country at all has already been taken, and it focuses only on the question how much aid to provide to the recipient country. The dependent variable is expressed in logarithmic terms, in order to address skewness in the distribution. This, together with all considerations in chapter three and this chapter leads to the regression to estimate the change in the dependency of the size of aid commitments on policies in recipient countries after the publication by Burnside and Dollar (2000):

(1) 𝐿𝑜𝑔(𝐵𝑖𝑙𝑎𝑡𝑒𝑟𝑎𝑙 𝐴𝑖𝑑_𝑑𝑟𝑡) = 𝛼 + 𝐷𝑜𝑛𝑜𝑟_𝑑+ 𝑌𝑒𝑎𝑟𝑡+ 𝛽1∗ 𝑃𝑜𝑙𝑖𝑐𝑦𝑟𝑡∗ > 2000 𝑑𝑢𝑚𝑚𝑦 + 𝛽2∗ 𝐿𝑜𝑔(𝐼𝑛𝑐𝑜𝑚𝑒𝑟𝑡) + 𝛽3∗ 𝐿𝑜𝑔(𝑃𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛𝑟𝑡) + 𝛽4∗ 𝐷𝑒𝑏𝑡 𝑟𝑎𝑡𝑒𝑟𝑡+ 𝛽5∗

𝐼𝑛𝑣𝑒𝑠𝑡𝑚𝑒𝑛𝑡 𝑟𝑎𝑡𝑒𝑟𝑡+ 𝛽6∗ 𝐵𝑖𝑙𝑎𝑡𝑒𝑟𝑎𝑙 𝑇𝑟𝑎𝑑𝑒𝑟𝑑𝑡+ 𝛽7∗ 𝐹𝑜𝑟𝑚𝑒𝑟 𝐶𝑜𝑙𝑜𝑛𝑦𝑟+ 𝛽8∗ 𝐴𝑙𝑙𝑖𝑎𝑛𝑐𝑒𝑟𝑑𝑡

5_{One could do a tobit analysis or restrict the analysis to simple ordinary least squares regression. Other}

possibilities are fixed effects models using only nonzero observations with a Heckman correction for selection bias or a two stage model with a Heckman correction while using all observations or a random effects model (Claessens et al., 2009).

(19)

18 (2) 𝐿𝑜𝑔(𝑀𝑢𝑙𝑡𝑖𝑙𝑎𝑡𝑒𝑟𝑎𝑙 𝐴𝑖𝑑_𝑑𝑟𝑡) = 𝛼 + 𝐷𝑜𝑛𝑜𝑟_𝑑+ 𝑌𝑒𝑎𝑟𝑡+ 𝛽1∗ 𝑃𝑜𝑙𝑖𝑐𝑦𝑟𝑡∗ > 2000 𝑑𝑢𝑚𝑚𝑦 + 𝛽2∗ 𝐿𝑜𝑔(𝐼𝑛𝑐𝑜𝑚𝑒𝑟𝑡) + 𝛽3∗ 𝐿𝑜𝑔(𝑃𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛𝑟𝑡) + 𝛽4∗ 𝐷𝑒𝑏𝑡 𝑟𝑎𝑡𝑒𝑟𝑡+ 𝛽5∗ 𝐼𝑛𝑣𝑒𝑠𝑡𝑚𝑒𝑛𝑡 𝑟𝑎𝑡𝑒𝑟𝑡+ 𝛽6∗ 𝐶𝑜𝑙𝑜𝑛𝑖𝑧𝑒𝑑𝑟

Where bilateral aiddrt are bilateral aid commitments from donor country d to recipient country

r at year t. Where multilateral aiddrt are multilateral aid commitments from donor d to

recipient country r at year t and α is the constant. Policyrt is the policy indicator for recipient country r at in year t and >2000 dummy is a binary variable indicating whether the

observations are done after the publication of Burnside and Dollar(2000). Donord is a donor

fixed effect regressor which captures unobserved time- and recipient invariant fixed effects for donor d. Yeart is a year fixed effect regressor which captures unobserved donor- and

recipient fixed effects for year t. Incomer represents the level of GDP per capita of recipient

country r at time t. Populationrt represents the number of people living in recipient country r

at time t. Debt ratert represents the ratio of external debt to GDP of recipient country r at time

t. Investment ratert represents the ratio of internal investments to GDP in recipient country r at

time t. Bilateral traderdt represents the sum of exports and imports between donor country d,

recipient country r at time t. Former Colonyrd indicates whether recipient country r is a

former colony of donor country d. Alliancerdt indicates whether recipient country r had some

kind of alliance with donor country d at time t. Colonizedr represents whether recipient

country r has been colonized by any of the Western Countries in the past.

4.5 Providing aid: Yes or No?

This study also examines whether the publication by Burnside and Dollar (2000) has affected the decision made by donors to allocate aid to a certain country or not. This is done by regressing the same independent variables as in regression 1 and 2 on a different dependent variable and including the zero observations. The zero aid flow observations need to be included in the analysis since these reflect the decision not to provide aid, whereas the nonzero observations reflect the decisions to provide aid. In addition, the dependent variable is no longer expressed in logarithmic terms, otherwise the zero observations could not be defined. Multilateral aid and bilateral aid are again considered as two different types of aid, and therefore two different dependent variables. The following regressions show how this study examines whether the publication by Burnside and Dollar (2000) has affected the decisions made by donors to allocate aid to a certain country or not.

(20)

19 (3) 𝐵𝑖𝑙𝑎𝑡𝑒𝑟𝑎𝑙 𝑎𝑖𝑑_𝑟𝑡𝑑 = 𝛼 + 𝐷𝑜𝑛𝑜𝑟𝑑+ 𝑌𝑒𝑎𝑟𝑡+ 𝛽1∗ 𝑃𝑜𝑙𝑖𝑐𝑦𝑟𝑡∗ > 2000 𝑑𝑢𝑚𝑚𝑦 + 𝛽2∗ 𝐿𝑜𝑔(𝐼𝑛𝑐𝑜𝑚𝑒𝑟𝑡) + 𝛽3∗ 𝐿𝑜𝑔(𝑃𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛𝑟𝑡) + 𝛽4∗ 𝐷𝑒𝑏𝑡 𝑟𝑎𝑡𝑒𝑟𝑡+ 𝛽5∗ 𝐼𝑛𝑣𝑒𝑠𝑡𝑚𝑒𝑛𝑡 𝑟𝑎𝑡𝑒𝑟𝑡+ 𝛽6∗ 𝐵𝑖𝑙𝑎𝑡𝑒𝑟𝑎𝑙 𝑇𝑟𝑎𝑑𝑒𝑟𝑑𝑡+ 𝛽7∗ 𝐹𝑜𝑟𝑚𝑒𝑟 𝐶𝑜𝑙𝑜𝑛𝑦𝑟+ 𝛽8∗ 𝐴𝑙𝑙𝑖𝑎𝑛𝑐𝑒𝑟𝑑𝑡 (4) 𝑀𝑢𝑙𝑡𝑖𝑙𝑎𝑡𝑒𝑟𝑎𝑙 𝑎𝑖𝑑𝑟𝑡𝑑 = 𝛼 + 𝐷𝑜𝑛𝑜𝑟𝑑+ 𝑌𝑒𝑎𝑟𝑡+ 𝛽1∗ 𝑃𝑜𝑙𝑖𝑐𝑦𝑟𝑡∗ > 2000 𝑑𝑢𝑚𝑚𝑦 + 𝛽2∗ 𝐿𝑜𝑔(𝐼𝑛𝑐𝑜𝑚𝑒𝑟𝑡) + 𝛽3∗ 𝐿𝑜𝑔(𝑃𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛𝑟𝑡) + 𝛽4∗ 𝐷𝑒𝑏𝑡 𝑟𝑎𝑡𝑒𝑟𝑡+ 𝛽5∗ 𝐼𝑛𝑣𝑒𝑠𝑡𝑚𝑒𝑛𝑡 𝑟𝑎𝑡𝑒𝑟𝑡+ 𝛽6∗ 𝐶𝑜𝑙𝑜𝑛𝑖𝑧𝑒𝑑𝑟

Where all variables have the same definition as in regression 1 and 2.

4.6 Disaggregating the policy indicator

The last part of the empirical analysis of this study explores the change in importance of the six Worldwide Governance Indicators for aid allocation since the publication of Burnside and Dollar (2000). This is done by disaggregating the combined governance index into the six original indicators. Multilateral donors might value one of the indicators differently compared to bilateral donors, e.g. perhaps is the perception the level of corruption in a country very important for bilateral donors but prioritize multilateral donors the perception of the rule of law. Due to this possible difference, also here the multilateral donors are analysed separately of bilateral donors. The analysis will examine whether the publication by Burnside and Dollar (2000) changed the importance of the separate indicators on the size of aid commitments. The regression models are shown in model 5 and 6, basically reproducing regressions 1 and 2, with the six different policy indicators instead of the average score of WGI.

(5) 𝐿𝑜𝑔(𝐵𝑖𝑙𝑎𝑡𝑒𝑟𝑎𝑙 𝐴𝑖𝑑𝑑𝑟𝑡) = 𝛼 + 𝐷𝑜𝑛𝑜𝑟𝑑+ 𝑌𝑒𝑎𝑟𝑡+ 𝛽1∗ 𝐶𝑜𝑛𝑡𝑟𝑜𝑙 𝑜𝑓 𝐶𝑜𝑟𝑟𝑢𝑝𝑡𝑖𝑜𝑛𝑟𝑡∗ > 2000 𝑑𝑢𝑚𝑚𝑦 + 𝛽2∗ 𝐺𝑜𝑣𝑒𝑟𝑛𝑚𝑒𝑛𝑡 𝐸𝑓𝑓𝑒𝑐𝑡𝑖𝑣𝑒𝑛𝑒𝑠𝑠𝑟𝑡∗ > 2000 𝑑𝑢𝑚𝑚𝑦 + 𝛽3∗ 𝑃𝑜𝑙𝑖𝑡𝑖𝑐𝑎𝑙 𝑆𝑡𝑎𝑏𝑖𝑙𝑖𝑡𝑦𝑟𝑡∗ > 2000 𝑑𝑢𝑚𝑚𝑦 + 𝛽4∗ 𝑅𝑒𝑔𝑢𝑙𝑎𝑡𝑜𝑟𝑦 𝑄𝑢𝑎𝑙𝑖𝑡𝑦𝑟𝑡∗ > 2000 𝑑𝑢𝑚𝑚𝑦 + 𝛽5∗ 𝑅𝑢𝑙𝑒 𝑜𝑓 𝐿𝑎𝑤𝑟𝑡∗ > 2000 𝑑𝑢𝑚𝑚𝑦 + 𝛽1∗ 𝑉𝑜𝑖𝑐𝑒 𝑎𝑛𝑑 𝐴𝑐𝑐𝑜𝑢𝑛𝑡𝑎𝑏𝑖𝑙𝑖𝑡𝑦𝑟𝑡∗ > 2000 𝑑𝑢𝑚𝑚𝑦 + 𝛽6∗ 𝐿𝑜𝑔(𝐼𝑛𝑐𝑜𝑚𝑒𝑟𝑡) + 𝛽7∗ 𝐿𝑜𝑔(𝑃𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛𝑟𝑡) + 𝛽8∗ 𝐷𝑒𝑏𝑡 𝑟𝑎𝑡𝑒𝑟𝑡+ 𝛽9∗ 𝐼𝑛𝑣𝑒𝑠𝑡𝑚𝑒𝑛𝑡 𝑟𝑎𝑡𝑒𝑟𝑡+ 𝛽10∗ 𝐵𝑖𝑙𝑎𝑡𝑒𝑟𝑎𝑙 𝑇𝑟𝑎𝑑𝑒𝑟𝑑𝑡+ 𝛽11∗ 𝐹𝑜𝑟𝑚𝑒𝑟 𝐶𝑜𝑙𝑜𝑛𝑦𝑟+ 𝛽2∗ 𝐴𝑙𝑙𝑖𝑎𝑛𝑐𝑒𝑟𝑑𝑡  (6) 𝐿𝑜𝑔(𝑀𝑢𝑙𝑡𝑖𝑙𝑎𝑡𝑒𝑟𝑎𝑙 𝐴𝑖𝑑𝑑𝑟𝑡) = 𝛼 + 𝐷𝑜𝑛𝑜𝑟𝑑+ 𝑌𝑒𝑎𝑟𝑡+ 𝛽1∗ 𝐶𝑜𝑛𝑡𝑟𝑜𝑙 𝑜𝑓 𝐶𝑜𝑟𝑟𝑢𝑝𝑡𝑖𝑜𝑛𝑟𝑡∗ > 2000 𝑑𝑢𝑚𝑚𝑦 + 𝛽2∗ 𝐺𝑜𝑣𝑒𝑟𝑛𝑚𝑒𝑛𝑡 𝐸𝑓𝑓𝑒𝑐𝑡𝑖𝑣𝑒𝑛𝑒𝑠𝑠𝑟𝑡∗ > 2000 𝑑𝑢𝑚𝑚𝑦 + 𝛽3∗ 𝑃𝑜𝑙𝑖𝑡𝑖𝑐𝑎𝑙 𝑆𝑡𝑎𝑏𝑖𝑙𝑖𝑡𝑦𝑟𝑡∗ > 2000 𝑑𝑢𝑚𝑚𝑦 + 𝛽4∗ 𝑅𝑒𝑔𝑢𝑙𝑎𝑡𝑜𝑟𝑦 𝑄𝑢𝑎𝑙𝑖𝑡𝑦𝑟𝑡∗ > 2000 𝑑𝑢𝑚𝑚𝑦 + 𝛽5∗ 𝑅𝑢𝑙𝑒 𝑜𝑓 𝐿𝑎𝑤𝑟𝑡∗ > 2000 𝑑𝑢𝑚𝑚𝑦 + 𝛽1∗ 𝑉𝑜𝑖𝑐𝑒 𝑎𝑛𝑑 𝐴𝑐𝑐𝑜𝑢𝑛𝑡𝑎𝑏𝑖𝑙𝑖𝑡𝑦𝑟𝑡∗ > 2000 𝑑𝑢𝑚𝑚𝑦 + 𝛽6∗ 𝐿𝑜𝑔(𝐼𝑛𝑐𝑜𝑚𝑒𝑟𝑡) + 𝛽7∗ 𝐿𝑜𝑔(𝑃𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛𝑟𝑡) + 𝛽4∗ 𝐷𝑒𝑏𝑡 𝑟𝑎𝑡𝑒𝑟𝑡+ 𝛽5∗ 𝐼𝑛𝑣𝑒𝑠𝑡𝑚𝑒𝑛𝑡 𝑟𝑎𝑡𝑒𝑟𝑡+ 𝛽6∗ 𝐶𝑜𝑙𝑜𝑛𝑖𝑧𝑒𝑑𝑟

(21)

20 Where the variables have the same definition as the corresponding variables in models 1,2,3 and 4. The variables Control of Corruption, Government Effectiveness, Political Stability,

Regulatory Quality, Rule of Law and Voice and Accountability represent specific indicators for

perceived governance in recipient country r in year t. These are Worldwide Governance Indicators of which the exact definitions are presented in the appendix table A.

4.7 Endogeneity issues

A variable is endogeneous when it is correlated with the error term of the regression equation. This implies that the regression coefficient is biased (Stock & Watson, 2011). In short, there are three causes of endogeneity, all will be discussed below.

4.7.1 Measurement error

According to one of the producers of the Worldwide Governance Indicators, the observations of the policy indicators are measured with a margin of error. When the variable of interest in a regression is measured with an error, this is called a classical measurement error. The classical measurement error model assumes that the margin of error is unrelated to the regressor. In short, this causes the OLS estimate to suffer from attenuation bias, typically a downward bias. When one does not assume that the margin of error is independent of the regressor, this necessarily implies that the OLS estimate is biased. Also, in the case in which the policy indicator would not be the only independent variable measured with error it is possible that some of the attenuation bias will be picked up by other variables.

There are two reasons why the possible margins of error in the policy indicators are not of concern to this research. Firstly, research showed that meaningful comparisons across time and countries are still possible (Kaufman at al., 2011). Secondly, in case attenuation bias would be of concern, the policy estimators of both before and after 2000 would suffer from downward bias. As this research focuses on comparing the coefficients before and after 2000, biased estimates would not affect the result of this study. Since the methodology of gathering data of Worldwide Governance Indicators did not change over the years, there is no reason to assume that the margins of error are significantly different before and after 2000 (World Bank, 2015a).

4.7.2 Omitted variables

There may exist confounding factors which potentially cause a spurious relation between the dependent variable of interest (policy) and the dependent variable (aid). Specifically, such a

(22)

21 confounding factor would be a variable which is correlated with one or multiple regressors and the dependent variable. When this factor is omitted from the regression, the estimates may suffer from omitted variable bias. The ensuing parameters will be biased and

inconsistent, its variance will be biased, the interaction term will be biased (Stock & Watson, 2011). In this study, a non-causal relation between policy and aid may as a consequence of omitted variables be interpreted as causal. Therefore should variables which correlate with both aid and policy be included in the regression as control variables to reduce the risk for a spurious relation. On the other hand, too many control variables make interpretation of the estimates harder.6 In order to prevent the regressions in this study to suffer from this bias are the choice for the inclusion of the control variables dependent on theoretical relevance of the controls and results from previous studies.

4.7.3 Reverse causality

The last potential source of endogeneity occurs when the regressor and the dependent variable are simultaneously determined. Not only is the dependent variable dependent on the regressor but also vice versa. For example, suddenly there is a positive shock in aid flows (an increase in the error term) and this causes policies in recipient countries to be relatively less important for allocation. This would mean that the policy indicator is positively related with the positive shock in aid, and with the error term. In this way, endogeneity causes the policy coefficient to be upward biased.

In this study, one could be worried about the introduction of simultaneous bias with the inclusion of the investment rate and debt rate variables. One might argue that these variables are post-treatment variables relative to governance. If poor governance causes the investment rate to be low or the debt rate to be high, then the estimates of these variables may be based on the quality of governance in the country (Martinez &Winters, 2015). However, table C and D in the appendix show that the estimates of the policy indicators change only slightly when the investment rate and debt rate are included in the specification.

6_{Table C and D of the appendix show that the coefficients of the variables of interest do not attenuate when the}

control variables are added to the model one by one. This shows that the results are not sensitive to the choice of the regressors.

(23)

22

5 Regression results

This chapter presents the results of the analysis, the core of this study. The first part analyses whether the importance of policy for the size of aid commitments have changed since 2000. The second part analyses whether donors’ decision to provide aid to a country or not has been affected by the publication of Burnside and Dollar (2000). The last part of the analysis

considers the change in importance of the different policy indicators for the size of aid commitments since 2000.

5.1 How much aid to provide?

Table 2a presents the results of running regressions 1 and 2 on the dataset which is restricted to the nonzero observations. The table presents two sets of two regressions, one set for

bilateral aid and one for multilateral aid. This table analyses only those observations for which donors have already decided to allocate aid to a country, and therefore addresses the size of the aid commitment. The first column of each set presents whether the size of aid

commitments was conditional on the quality of policies in the recipient country between 1996 and 2010. The second column of each set shows whether donors have changed their behaviour since the publication of Burnside and Dollar (2000) by making the size of aid flows more dependent on policies in recipient countries.

The estimate in the first column for bilateral aid shows that policies in recipient countries have been very important for the size of bilateral aid commitments between 1996 and 2010. More specifically, a one unit increase in a countries’ average score of the WGI caused bilateral aid flows to that country to be 56% larger. The third column indicates that this was also but slightly less the case for multilateral aid allocation, an one unit increase in the WGI score increased multilateral aid commitments to be 28% larger. These estimates indicate that policies have been remarkably important for aid allocation, which is in line with previous studies (Martinez &Winters, 2015; Alesina & Dollar, 2000).

The negative coefficient in the second column for bilateral aid indicates that policies in recipient countries became less important for the size of bilateral aid commitments after the publication by Burnside and Dollar (2000). The small coefficient means that the decline in importance was minimal and as importance was already large, one can still state that policies remained to be important for bilateral aid allocation after 2000. Also for multilateral aid allocation policies became the quality of policies in recipient countries less important for the

(24)

23 size of the aid commitment. The drop in importance is, however, larger than in the bilateral donor analysis.

5.2 Provide aid: Yes or No?

Before the donor decides how much aid to provide to a country, it decided to provide aid to this country at all. Table 2b presents the results of running regressions 3 and 4 showing two sets of two regressions, again, one set for bilateral aid and one for multilateral aid. The dependent variables in all columns are aid commitments in absolute terms since the zero observations would otherwise not been defined. The first column of each set presents the same regression as the interaction regression in table 2a to function as a baseline regression for the next column. The second column of each set shows whether donors have changed their behaviour since the publication of Burnside and Dollar (2000) by making their decision to allocate aid to a certain country more dependent on the quality of policies in that country.

The estimates in the first and third columns confirm the finding of table 2a, for both types of aid became policies in the recipient country less important when donors decide on the size of aid commitments. When a large amount of zero observations are included in the sample, the coefficient for the interaction term becomes positive. It is, therefore, likely that the zero observations cause the coefficient to become positive. One can conclude that policies in recipient countries became more important, after 2000, when donors decide not to provide aid to a certain country. This conclusion should only carefully be drawn because of two reasons. Firstly, because it is drawn from indirect reasoning since covering the zero observations in a regression with logarithmic dependent variable, such as regression 1 and 2, is not possible. Secondly, because the use of zero observations in an OLS regression might cause estimates to suffer from selection bias as described in chapter 4.

To sum up, policies in recipient countries are very important for donors when they decide whether and how much aid to provide to a country. After the publication by Burnside and Dollar (2000), policies in recipient countries became more important for donors when they determine to provide aid to a country or not. However, after the publication were policies in recipient countries less important when donors decide how much aid to provide to the recipient. This last result is surprising because this contradicts the recommendations following from the publication in 2000.

(25)

24 Table 2a: How much aid to provide?

Bilateral Aid Multilateral aid Variable Base Regression Interaction Regression Base Regression Interaction Regression Policy, 1996-2010 0,524*** 0,559*** 0,279*** 0,520*** (0,0911) (0,0948) (0,0941) (0,139) Policy, after 2000 -0,0499 -0,288* (0,0107) 8,851*** (0,156) Constant 8,833*** (1,399) 10,74*** 10,84*** (1,402) (1,307) (1,315) Observations7 _20.093 _20.093 _10.598 _10.598 R-squared 0,204 0,204 0,236 0,237

Donor FE Yes Yes Yes Yes

Year FE Yes Yes Yes Yes

Controls Yes Yes Yes Yes

*** significant at p<0,01 ** significant at p<0,05 * significant at p<0,1

Note: Numbers in parentheses are robust standard errors. Column 3 presents the results from running

regression (1) whereas column 5 presents the results from running regression (2). The other two columns present the results of the baseline regression, without the interaction term. The control variables which are omitted to save space are Log(Population), Log(Income), Debt rate, Investment rate and Log(Bilateral trade), Military Alliance and Former Colony for bilateral aid and Colonized for

multilateral aid.

Table 2b: Provide aid: Yes or No?

Bilateral Aid Multilateral aid Variable Non-zero Incl. zero Non-zero Incl. zero Policy, 1996-2010 15.64** 8.789** 85.16** 11.69** (6.613) (3.764) (33.35) (4.959) Policy, after 2000 -2.752** 0.531 -29.44 2.074 (1.350) (0.388) -64.17 (23.73) (3.437) Constant -183.0* -1,421 -263.3* (97.54) (39.19) (856.2) (135.8) Observations 21,515 53,175 10,850 59,625 R-squared 0.066 0.037 0.022 0.005

Donor FE Yes Yes Yes Yes

Year FE Yes Yes Yes Yes

Controls Yes Yes Yes Yes

Note: Numbers in parentheses are robust standard errors. Column 2 and 4 cover a dataset with only

non-zero aid flow observations. Column 3 and 5 cover a dataset with both zero as well as non-zero aid flow observations. The control variables which are omitted to save space are Log(Population), Log(Income), Debt rate, Investment rate and Log(Bilateral trade), Military Alliance and Former

Colony for bilateral aid and Colonized for multilateral aid.

7_{The drop in number of observations in the regression has multiple causes. Firstly, and most importantly, there}

are a large number of missing values in the datasets of the policy variable, bilateral trade and the debt rate. The missing values in the bilateral trade variable can easily be explained by the fact that countries are not trading with all countries. The missing values in the other variables can not be explained, they come from missing values in the original datasets. Secondly, the regressions make a distinction between multilateral and bilateral aid, and therefore cover together the complete set of aid commitment observations

(26)

25

5.3 Disaggregating the policy indicator

In this study is the level of perceived policies in the recipient countries indicated by the average score of the six different Worldwide Governance Indicators (WGI). In the last part of the analysis is the WGI average disaggregated. Table 3 presents if the score of a country on the separate indicators became more important for the size of aid commitments after the publication by Burnside and Dollar (2000). Table 3 shows both positive and negative estimates for the different indicators, for which some are significant while others are not. It shows inconsistencies across the two donor types, both confirming and contradicting the results of table 2a and 2b. These findings are difficult to explain. One explanation may be allocation policies with different priorities among donors, with some donors focusing on one indicator and leaving the others out, while others are doing the opposite. This might cause the inconsistency in the signs across the indicators and donor types. One could also argue that donors are using the WGI in their allocation policies, as Hout (2007) states, but instead of using the specific indicators they use an aggregate measure, such as the average, or even a source indicator from the database. This is just speculation, further research is necessary to explain these findings.

Note: Numbers in parentheses are robust standard errors. The control variables which are omitted to

save space are Log(Population), Log(Income), Debt rate, Investment rate and Log(bilateral trade), Military Alliance, Former Colony for bilateral aid and Colonized for multilateral aid.

Table 3: The change in importance of the disaggregated policy indicators

Variable Bilateral aid Multilateral aid

Control of Corruption -0,145 0,471** (0,163) (0,216) Government Effectiveness -0,214 -0,480** (0,181) (0,216) Political Stability 0,0728 0,0412 (0,0728) (0,0999) Regulatory Quality 0,107 -0,0842 (0,188) (0,196) Rule of Law 0,192 -0,0696 (0,184) (0,144)

Voice and Accountability -0,116 -0,136

(0,0991) (0,111)

Constant 9,688*** 10,90***

(1,323) (1,310)

Observations 20.093 10.598

R-squared 0,212 0,243

Donor FE Yes Yes

Year FE Yes Yes

The effect of new scientific knowledge about effective aid allocation on actual allocation policies