Universiteit van Amsterdam
Commercial Interests in Aid Allocation
The Role of Recipient Country’s Natural Resources
Master Thesis
Frank van Hoenselaar
Studentnumber: 6082882
Supervisor: prof. dr. Joop Hartog E-mailadres: frank.vh@live.nl Date: July, 2014
Abstract
Literature on aid allocation has always found a strong role for donor country commercial interest as a determinant of aid provision. However, these commercial interests have never been investigated more thoroughly than bilateral trade relationships between two countries. This paper has provided an extension on the role of commercial interests of donor countries
by including a proxy for natural resource abundance of the recipient country. Using a Heckman two-step model and the largest three-dimensional panel ever employed in the field,
this paper finds that indeed more aid is directed towards resource rich economies. The coefficients found on the traditional regressors in aid allocation models are all in line with those found in previous research. The latter strengthens once more the observation that aid is
Acknowledgements
First, I want to thank my supervisor prof. dr. Hartog for his time, input and the critical review of my work
Second, I want to thank my econometrics teacher dr. van Ophem for helping me to find solutions for the several econometric difficulties I encountered.
Table of Contents
1. Introduction 1
2. Literature Review 3
2.1. Aid Allocation Motives 3
2.1.1. Recipient Need 4
2.1.2. Recipient Merit 5
2.1.3. Donor Self-Interest 6
2.1.4. Donor Self-Interest: The Role of Natural Resources 6
2.2. Econometric Methods 7
2.2.1. Theoretical Approaches 8
2.2.2. Econometric Approaches 10
3. Empirical Model and Data 13
3.1. Data 13
3.2. Methodology 16
3.2.1. Tobit type I 17
3.2.2. Cragg Two-part 18
3.2.3. Heckman Two-Step 18
3.2.4. Three-dimensional panel data 19
3.2.5. Discussion on the preferred specification 20
4. Results 23
4.1. Average Donor Behavior 23
4.2. Average donor behavior: Heterogeneity over time 27
4.3. Aid Allocation Behavior: Heterogeneity among donors 29
4.4. Robustness Checks 31
5. Discussion and Conclusion 32
5.1. Discussion 33
5.2. Conclusion 34
6. References 36
1
1. Introduction
Does aid promote macroeconomic growth? This question has been a point of debate within the field of development economics for the past 30 years. Many studies suggest that there is no or at most a very weak link between aid and macroeconomic growth (Boone, 1994; Burnside and Dollar, 2000; Easterly et al., 2003) These findings gave rise to a new and again much debated question: Is aid allocated efficiently? (i.e. towards the countries that need it most and can absorb it best?)
Whereas mission statements of donor country aid agencies almost always state that the allocation of aid is solely based on developmental motives, a large body of literature suggests that donor country self-interest is an important motivation as well. If indeed, foreign aid is partly (or entirely) distributed according to commercial, strategic or political interest of the donor country, it seems unlikely that aid is allocated in an efficient manner. However, this does not imply that humanitarian motives are always conflicting with donor self-interest, it rather implies that the less prominent is recipient need as a determinant of aid allocation, the more likely it is that the effectiveness of aid is reduced. (Harrigan and Wang, 2011).
Alesina and Dollar (2000) used advanced econometric tools to extensively explore the strategic and political considerations in the allocation of aid. They find that strategic and political considerations are so important, that according to their estimations: “an inefficient, economically closed, mismanaged non-democratic former colony politically friendly to its former colonizer, receives more foreign aid than another country with similar levels of poverty, a superior stance, but without a past as a colony.” Following Alesina and Dollar (2000), many authors have attempted to improve the modelling of aid allocation motives and especially donor self-interest. Berthélemy (2006) tested whether donor countries were more likely to allocate aid to countries with outstanding debt contracts, Hoeffler and Outram (2011) tested an empirical link between United Nations voting allies and aid allocation and Mcillivray (2003) included arms transfers between recipient and donor as an explanatory variable.
Among others, Berthelemy and Tichit (2004) find a strong link between bilateral trade between donor and recipient country and receiving developmental aid. This link suggests that donor countries reward trade relationships with more aid assistance. These commercial interests of the donor country have not yet been explored more extensively. This paper hypothesizes that commercial interests in aid allocation are even stronger than previous research suggests. If aid is indeed used as a strategic tool to secure trade, one might expect
2
that certain kinds of commercial activities are preferred to others. Traditionally the OECD countries have preferred trade with developing countries that were able to offer them interesting primary resources such as oil, coal, gas etc. One might therefore argue that aid, as a strategic tool, might be more directed towards economies that have relatively a lot of these natural resources. This paper tests the hypothesis whether aid is more directed towards countries that are resource rich.
The contribution of this paper is three-fold. First, the paper attempts to find an empirical link between aid allocation behavior and natural resource abundance. Second, the paper uses the most extensive three dimensional panel dataset ever employed in the field. The dataset covers 30 years (1980-2010), 23 donors and 152 recipient countries and has close to 70,000 complete observations. Third, the paper settles an ongoing dispute in the aid allocation literature on the preferred econometric method by extensively evaluating the most frequently used approaches.
The paper is structured as follows: Section 2 provides a broad literature review that discusses different aid allocation motivations, theoretical approaches and econometric methods that have dominated the literature. Section 3 describes in more detail how the empirical specification in this paper comes about and why there is only one reasonable econometric specification. The results are presented in section 4. Section 4 discusses average donor behavior, donor behavior for all donors separately and average donor behavior over time. Section 5 provides both a conclusion and a discussion about the implications and limitations of this paper.
3
2. Literature Review
The literature on aid allocation is plentiful. Authors have reached little consensus however, both on the most appropriate econometric approach and the explanatory variables used in their empirical framework. This section will first outline the main motivations for countries to offer bilateral aid that have been described in the literature. Section 2.2 will subsequently describe the discussion that is currently going on about the most appropriate theoretical approach and econometric method to model the aid allocation process.
2.1 Aid Allocation Motives
Official Development Aid (ODA) will in general fulfill three requirements: ODA is undertaken by the official sector (governments), ODAs’ main purpose is the promotion of economic development and ODA has a grant element of at least 25% (OECD, 2014). ODA is provided both bilateral and multilateral. Bilateral aid refers to transfers from government to government, whereas multilateral aid is often delivered through multilateral institutions (World Bank, IMF etc.). Motivations for donor countries to provide either bilateral or multilateral aid are in general very different. Maizels and Nissanke (1984) find that multilateral aid is more responsive to the needs of developing countries, whereas bilateral aid allocation is more driven by the interest of the donor countries. Moreover, Burnside and Dollar (2000) find that multilateral aid is distributed more efficiently i.e. to countries that implement “good policies”. Most recent aid allocation studies have been focusing on bilateral aid because this type of aid more strongly reflects the choices made by individual donor countries.
An early study by McKinley & Little (1979) compared how well different aid allocation motives fitted the actual U.S. aid allocation patterns. Their model regressed U.S. aid provision on different recipient country observables that represented potential motivations. Motivations were categorized by developmental motives, economic interests, political Interests and security interests. The authors find that developmental motives (GDP per capita, capital formation and economic growth) do not fit the data very well and they therefore conclude that humanitarian motives (recipient need) do not play a significant role in U.S. aid allocation behavior. However, they do find support for the model that describes economic interests (investments, trade). Trade appears to be a significant determinant of aid and the authors therefore conclude that aid follows trade flows, a finding that this paper will build upon. Investments do not appear significant in their estimation, but likely exhibit an
4
effect through trade flows, since foreign investments might intensify trade flows. The security (percentage communist votes) and political interests (military resources and population) also have strong explanatory power. It appeared that the U.S. was strongly attracted to give aid to communist countries, countries with large populations and countries with a strong military force. The finding by McKinley & Little (1979) that aid allocation is mainly determined by donor interest rather than humanitarian motives has still not been falsified.
The findings by McKinley & Little (1979) gave rise to the emergence of a large body of aid allocation literature. Authors thereafter maintained the distinction between recipient need variables and donor interest variables but introduced many new proxies to improve the fit of the model. The finding of Burnside and Dollar (2000) that aid is only effective (i.e. yields economic growth) in countries that have good policies, resulted in the introduction of a third category of aid allocation motives: recipient merit (Hoeffler and Outram, 2011). Variables that reflect recipient merit attempt to find whether aid is allocated to countries that can best absorb it i.e. countries with good policies and institutions. The following subsections will discuss the variables that have been used and were found to be significant in each category.
2.1.1. Recipient Need
In Berthélemy (2006) the recipient need variables measured the level of altruistic behavior by donor countries. In other words, to what extent do donor countries provide aid to the countries that need it most? GDP per capita is the only proxy in the recipient need category that always appears with a significant negative coefficient in the aid allocation equation. In other words, aid is mainly provided to countries that are poor. However, Alesina & Dollar (2002) find that the effect of GDP per capita is not linear and refer to this as the middle-income bias. The middle middle-income bias is the observation that the poorest countries receive very little aid until a certain threshold has been reached. The common solution to this observation is the use of the Log of GDP per capita (among others Hoeffler & Outram) or the introduction of a squared term (Alesina & Dollar, 2002).
Other development indicators such as health indicators have been used to model recipient need. McGillivray (2003) uses child mortality, but finds no evidence that aid favors countries that have high child mortality. Thiele et al. (2007) investigated how well aid flows follow the Millennium Development Goals (UN, 2013) by regressing aid allocation on several human development indicators. They find no evidence that donor countries take development indicators other than income into account. Therefore GDP per capita is often
5
considered sufficient to model the recipient need, mainly because GDP per capita is often strongly correlated with other development indicators such as health and education.
2.1.2. Recipient Merit
Burnside and Dollar (2000) concluded that aid only generates growth conditional on a good macroeconomic policy environment. Such an environment was characterized by low inflation, a budget surplus and an open economy. However, they also find that actual aid flows are not directed towards countries that have good policies. Nonetheless, Burnside and Dollar’s (2000) research has been extended in all papers on aid allocation thereafter. Early authors had already been including proxies for human rights and democracy and found that these had very different effects for different countries. For instance, democracy in the receiving country plays a large role in the aid allocation policy of the U.S., UK and the Netherlands, whereas France, Belgium and many others do not allocate more aid towards democracies (Alesina & Dollar, 2002).
By looking at aggregate donor behavior Berthélemy and Tichit (2003) find that only lagged economic growth has a significant positive effect on receiving aid. They state that previous economic growth signals ‘good’ economic policy and is therefore rewarded with more aid. However, as Clist (2011) points out, there might be other reasons for donors to allocate aid to countries that have high growth rates than only a signal for good policy (profits from increased trade, signal that aid efforts are fruitful etc.). Clist (2011) therefore advocates the approach by Neumayer (2003) using the Freedom Index and the Political Terror Scale (PTS) to model recipient country merit. The Freedom Index is an aggregate score from 7 (worst) to 1 (best) that evaluates an individual’s civil liberties and human rights according to seven subcategories: the ability to legitimately vote in elections, the ability to freely participate in the political process, the availability of accountable representatives, the ability to exercise freedoms of expression and belief, the ability to freely assemble and associate, the accessibility of the system of the rule of law and the ability to enjoy social and economic freedoms (Freedomhouse, 2014). The Political Terror Scale measures levels of violence exercised by the state and is a score between 5 (worst) and 1 (best) (Gibney et al., 2014). Clist (2011) finds that the Freedom Index exhibits a significant negative effect for all countries meaning aid favors countries with good civil liberties and human rights conditions. The PTS does not play a significant role for every country, but its sign is never positive, meaning that some donors favor recipient countries with low levels of state violence.
6
2.1.3. Donor self-interest
Donor self-interest has often been described in terms of political, strategic and commercial motives of the donor country. Political motives have mainly been described in the literature as colonial ties. Donor countries are much more likely to allocate aid to countries that have been former colonies (among others Berthélemy & Tichit, 2003). Furthermore because Egypt and Israel have received such large amounts of developmental aid from the U.S., many authors have included separate Egypt and Israel dummies to their regressions that always yield significant positive coefficients (among others McGillivray, 2003).
Bilateral aid might also serve as a strategic tool for a donor country to exhibit influence or strengthen their relation with a certain recipient. Hoeffler & Outram (2011) test whether the UN voting behavior of the recipient plays an important role for the amounts of aid it receives. They compare the voting behavior of the receiving country with the voting behavior of the donor country and find that when voting behavior of the donor and receiving country is closely aligned, the amounts of aid allocated are higher. The main problem with their approach is the high correlation between the voting behavior of the largest donors, which diminishes the explanatory power of this proxy. A more convincing proxy for strategic considerations is the amount of arms transfers between donor and recipient used by McGillivray (2003). He finds a significant positive effect of arms transfers on the amount of aid allocated, implying that military relations influence aid flows.
McKinley & Little (1979) were the first to discover that aid follows trade flows and hence aid might serve as an instrument to secure commercial interest of the donor country. Thereafter authors have included trade into their regressions without exception. Trade is often measured by the imports + exports between the donor and recipient as a percentage of the donor GDP (Berthélemy & Tichit, 2003). Alternatively, authors have used only exports as a percentage of donors' GDP, which yielded very similar results (Berthélemy, 2006). Although commercial interests have been proven to play an important role in the aid allocation process, there has been no further examination in what ways commercial interests determine the direction of aid flows.
2.1.4. Donor self-interest revisited: the role of natural resources
The dominant role of donor-self interest in the aid allocation process is reflected in the research by the relatively large array of proxies that have been used to model these interests. However, the important role of trade between donor and recipient country has never been explored more extensively. When the provision of ODA is used as an instrument to
7
strengthen trade relations, one can imagine that donor countries seek to provide aid to countries that possess interesting trade products. Western countries have always been particularly interested in obtaining the natural resources that many developing countries own. The contribution of this paper is therefore to investigate whether aid flows are directed towards developing countries that are relatively abundant in natural resources. I will use two different measures for natural resources both provided by the World Bank (2014). The World Bank’s measures for natural resources include oil, natural gas, coal, minerals and forests. The first measure is natural resource rents and is the share of GDP that is earned by natural resources for every year. Natural resource rents reflect what is currently earned with natural resources rather than the potential profits that donor countries might be interested in. Therefore I alternatively use a measure of resource wealth as applied by van der Ploeg and Poelhekke (2000) in their famous paper on the natural resource curse. The World Bank has made very precise estimates of the net present value of the natural resources for every country in 2005 (World Bank, 2014). My empirical research will aim to find a link between these two measures of the availability of natural resources and the flows of ODA.
This section has shown the relative importance of the donor self-interest over the actual needs of the recipient country. This is mainly reflected by the fact that the needs of the recipient countries that are taken into account can be summarized by only the negative relation between GDP per capita and Aid. On the other hand, the self-interest of donor countries is still an interesting field to explore and can be modelled using a variety of proxies. The current paper will explore such a new proxy as discussed in the previous paragraph.
2.2 Econometric Methods
Harrigan and Wang (2011) distinguish between three broad categories of aid allocation literature: explanatory, descriptive and prescriptive studies. Explanatory studies aim at explaining the observed aid allocation, descriptive studies compare actual aid allocation against certain normative criteria and prescriptive studies aim at calculating what aid allocation should optimally be. The explanatory studies are by far the most common approach and this paper will take this approach as well. This subsection will present the previous literature in a new manner. I explicitly distinguish between different methodological approaches and different theoretical approaches, because previous research has been confusing on this part.
8
2.2.1 Theoretical Approaches
By the early nineties the explanatory aid allocation literature was already so plentiful that McGillivray and White (1993) refer to six types of studies: recipient need/donor interest models, hybrid models, bias-models, developmental models, administrative/incremental and Limited Dependent Variable models (LDV). In this literature review I will disassociate from this rather arbitrary division, since this approach fails to distinguish between differences in theory and differences in econometric practice. In fact it is more appropriate to distinguish between authors that have used Ordinary Least Squares (OLS) and authors that have used a LDV approach. Within both of these two methodological categories authors have estimated recipient need/donor interest models, bias models, developmental models and administrative/incremental models all built on different theoretical assumptions about the aid allocation process. The following paragraphs will discuss these four different theoretical approaches. The differences between OLS and LDV will be discussed in subsection 2.2.2.
Recipient need and donor interest models (RN-DI models) are among the first methods that were employed to analyze the allocation of aid (among others Levitt, 1968). These studies theorized that the aid allocation process is determined by both the need of the recipient and the self-interest of the donor country. Early authors estimated two different models, one solely based on recipient need (poverty, life expectancy etc.) and one based on the interest of the donor (former colony, same language, military power). Subsequently the authors decided what model best fitted the data. The general result was that donor interest models performed better than the recipient need models. However, there is a major problem with this separation between recipient need and donor interest because it assumes that aid allocation can only be based on one type of motives. In reality it is more likely that there is a mix of motives that determine the distribution of aid, hence such a specification is likely to suffer from omitted variable bias (McGillivray, 2003). As a response to the potential omitted variable bias in the dichotomous RN-DI approach, the use of hybrid models became the new standard. In a hybrid model both the recipient need and donor interest variables are included in one regression. This approach has been first used by McKinley & Little (1979) and has dominated the aid allocation literature since.
Few studies took the biases approach. These studies have been focusing on two biases that have been found in aid allocation: the population bias and the middle-income bias. Note that these biases should be interpreted as a nonlinearity in the aid allocation process rather than a bias introduced by a wrong econometric approach. The population bias refers to the observation that small countries tend to receive more aid in per capita terms. There are two
9
potential explanations for this observation. First, small countries tend to be more specialized and hence their trade shares are relatively high. Therefore the commercial interests of the donor country might play a role via trade promotion especially in small countries. Second, small countries have higher marginal political benefits to a donor country (Dowling & Hiemenz, 1985). Small countries can exhibit relatively substantial political power that can potentially help or hurt a donor, for instance via the use of their UN votes (Harrigan & Wang, 2011). The middle-income bias is the observation that the poorest countries receive less aid until a certain threshold value has been reached. In other words the negative relation between aid and GDP per capita is not entirely linear and aid favors countries that have low, but not the lowest GDP per capita (Alesina & Dollar, 2002). Contemporary authors on aid allocation include quadratic terms or use log-transformations to deal with the nonlinearities caused by the two biases discussed above.
Harrigan and Wang (2011) furthermore distinguish bureaucratic (incremental) models. These models are based on the assumption that aid allocation is mostly determined by previous year's allocation and that changes in aid allocation are only made at the margin. It assumes that aid is allocated every year from a merely fixed pool of money and that due to bureaucracy donor countries do not like to change the allocation amounts too much from year to year. Feeny and McGillivray (2008) were among the first who took this approach. They used timeseries to analyze the determinants of receiving aid for the ten largest ODA recipients. Their model was based on a system of ten equations, estimated for ten recipients separately: : : (1) :
Where RN and DI refer to vectors including recipient need and donor interest variables respectively and where B is the lagged dependent variable (Aid). The main problem with this approach is that all donors are aggregated, hence the dependent variable refers to the total aid received by country j. Donor interest is referred to as the total export to Development
10
Assistance Committee (DAC) donors and the total DAC investments made in country j. Hence their model assumes that all donors exhibit the same behavior. Their conclusion that aid seems to be responsive over time to the needs of the recipient country is therefore likely to be driven by this assumption (Feeny & McGillivray, 2008). Although their specification might be wrong, the underlying theory is very useful for my analysis. It is likely that aid allocation is indeed a bureaucratic process and takes time to implement, therefore it is much more realistically to assume that a donor decides on the allocation of aid according to last year’s indicators.
Developmental models attempt to measure to what extent aid flows follow the development needs recipient countries. In fact it is the estimation of a recipient need model that only uses development indicators (life expectancy, child mortality etc.) as explanatory variables. Thiele et al. (2007) is a well-known example of such an approach. Their paper investigated to what extent aid flows serve the accomplishment of the Millennium Development Goals (UN, 2013). They conclude however, that donor countries do not take development indicators other than GDP per capita into account.
The four theoretical approaches that I distinguish are all built on different assumptions on how the aid allocation process works. All these approaches are still applied and each approach has its own advocates. This paper will take the approach of the hybrid form of the RN-DI model because this approach allows me to test what factors (recipient need or donor interest?) play a dominant role in the aid allocation process.
2.2.2 Econometric approaches
Advocates of every theoretical approach have either used Ordinary Least Squares (OLS) or Limited Dependent Variable (LDV) methods to estimate their model. The following section will outline the differences between OLS and LDV. Early authors that estimated aid allocation models always used a multiple OLS regression. In the example below we assume the author used panel data although this became only a practice from the 1990’s.
(2)
Where i refers to the donor country, j to the recipient country and t to a particular year. X is a vector containing variables that reflect both recipient need and donor interest variables, the other regressors represent fixed effects. An important assumption of the OLS specification is
11
that all observations are independent and identically distributed (i.i.d.). When we look at a sample of potential recipient countries, we observe both countries that receive aid and countries that do not receive aid. McGillivray (2003) describes how authors that used OLS have mainly solved this problem. In fact one can either use only the countries that do receive aid or use the full sample of potential recipients. Figure 1 gives an example of a simplified aid allocation equation. On the y-axis one finds the amount of aid and on the x-axis one finds per capita income, hence the observed negative relation.
Figure 1: OLS regression of Aid on GDP per capita (McGillivray, 2003).
Line A represents a sample with all potential recipients. Line B represents a sample with only actual recipients.
Using a full sample of potential recipients (Line A) will in this case result in a regression line that is inevitably too flat. This is because the countries that receive no aid, are likely to be the countries that have a relatively high GDP per capita. In other words there is non-random selection into the pool of countries that receive aid i.e. these countries have different characteristics and hence the i.i.d. assumption is violated. But what if one uses a sample of only countries that do actually receive aid? This case is represented by line B (the white dots represent the excluded observations). Figure 1 implies that there is a certain level of GDP per capita that cannot be exceeded in order to receive aid. In other words, the donor countries do select recipient countries based on a certain income threshold. One can formalize this as follows (McGillivray, 2003):
(3)
Where Z is the threshold applied by the donor countries to decide upon which country receives aid. Therefore, if one truncates the sample by only using the actual recipients one will violate the OLS assumption of an expected mean zero error term. The recognition of this
12
sample selection problem in the 1990’s completely changed the econometric methodology in the aid allocation literature.
Tarp et al. (1998) state that the allocation of aid is a process that consists of answering two different questions: First, who receives any aid and second, how much aid does a country receive once selected? The first question is about answering a yes or no question which makes the dependent variable censored. The second question is about answering how much aid is allocated to a subset of countries from the original sample, which makes the sample truncated. Guo and Fraser (2010) state that truncation and censoring are the characteristics of Limited Dependent Variables (LDV). Limited Dependent Variables are closely related to what Greene (2007) refers to as nonrandom sample selection which makes them not suitable for OLS. One way to overcome nonrandom sample selection is to explicitly model the selection process (Guo & Fraser, 2010):
(4)
(5)
Where the vector z is a vector of observables that determine the selection of a potential recipient country into the pool of actual recipients. Berthélemy and Tichit (2004) outline three sample selection models that have been applied in aid allocation modeling: Two-part models, Heckman two-step method and Tobit models. In the contemporary aid allocation literature, there does not seem to have emerged a consensus on the best specification.
A Tobit Model: This model assumes that the decision on who receives aid and the decision on how much aid one receives is a simultaneous process. It therefore imposes the condition that the process generating ones and zeros (eligible or ineligible for aid) is the same process that determines the amounts of aid (Harrigan and Wang, 2011). This implies that the beta’s in the selection and level process must be equal, otherwise the results become inconsistent. Think of a situation where the dependent variable is aid per capita. We know that probability for a country of being selected is positively related with the size of its population. But on the other hand we know that small countries relatively get more aid, which implies a negative relation between population and aid per capita in the level process. Tobit yields inconsistent estimates with such a specification.
13
A two-part model: In the first part the selection equation is estimated using a probit regression. Subsequently, using OLS the level of aid commitments is estimated only using the countries that received a positive amount of aid. In fact this method only recognizes the sample selection problem by estimating two separate equations but fails to integrate these two equations into one process.
A Heckman’s two-step method: This approach attempts to correct for the potential selection bias by including the Inverse Mill’s ratio (IMR) from the selection model into the regression of the second step (Berthélemy, 2006). The IMR is the probability density divided by the cumulative density (Greene, 2007). It therefore accomplishes to reflect the fact that the two aid allocation choices (who and how much?) are interdependent. A crucial assumption to this method is the so called exclusion restriction. Such a restriction states that at least one explanatory variable that appears significant in only one of the two equations, needs to be excluded in the equation where it insignificant (Puhani, 2000). This exclusion makes the identification process much stronger than in a case where the nonlinearities induced by the inverse mills ratio are the only identification mechanism. Section 3 will discuss the three LDV models in more depth and list the most important their assumption. In subsection 3.2.5 all these assumptions will be tested using a “toy-model” that in fact is a simplified specification of my real model. From this model it follows that only the two-step Heckman method can reasonably fulfill its assumptions.
3. Empirical Model and Data
In the previous chapter I elaborately discussed the different theories, methods and explanatory variables that have been used in the aid allocation literature. This chapter will discuss the approach of the current paper based on the insights obtained from the previous chapter. The first section will discuss the data that will serve my analysis whereas section 3.2 will discuss the choices I made regarding my estimation method.
3.1 Data: Variables and Sources
Bilateral aid flows are constructed using a large panel dataset that the OECD Development Assistance Committee (DAC) has available for the past 40 years (OECD, 2014) . My analysis will be based on bilateral aid flows from 23 donors to 152 recipient countries over the 1980 – 2010 period. The panel is special because it has in fact three dimensions (donors x recipient x
14
years) which potentially brings the number of observations to over 100.000 (actual number will be lower because of missing data in other sources). This three-dimensional panel data approach will get special attention in section 3.2. The OECD distinguishes between bilateral ODA commitments and disbursements. ODA disbursements are the actual agreed payments made to the recipient countries, whereas the ODA commitments reflect the proposals made by the donor countries. This paper will use ODA commitments because these better reflect the decisions made by the donor country. Disbursements depend in part on the recipients willingness to accept the aid and its administrative capacity to deal with the money (Berthélemy, 2006).
Authors have chosen different specifications for the dependent variable. Berthélemy & Tichit (2004) argue that aid per capita is the most appropriate dependent variable because it makes it possible to compare small and large recipients. Moreover they argue that specifying aid in absolute terms would lead to heteroscedasticity issues because the residuals will be much larger for larger countries. However, Clist (2011) advocates the use of the log percentage of donor i’s aid commitment going to recipient j,
. Such a
specification more likely reflects the decision made by the donor, rather than a donor country deciding how much aid per capita it will distribute to each recipient. Alternatively Clist (2011) argues that yields similar results and also has the advantage of a normally distributed dependent variable (the absolute values are somewhat skewed). This paper will select the first approach of Clist (2011) because it reflects how a donor country decides upon what share of its pool of aid is directed to each recipient.
Following Tarp et al. (1998) the total aid commitments by other bilateral donors will be added as an explanatory variable. However, this variable is scaled by the population of the recipient country. Berthélemy (2006) argues that including such a variable allows you to test whether assistance of other donor countries is perceived as a complement or a substitute to a donor’s own aid commitment. He argues that ‘egoistic’ donors i.e. donors that pursue self-interest are likely to perceive aid commitments by others as negative. Additionally Berthélemy (2006) argues that adding the total amount of aid granted by the donor in a specific year will correct for differences in aid budgets between donors and differences between years. Hence, these two measures will enter the aid allocation equation in this current paper.
Past colonial ties have proven to be a very important reason for donors to provide aid to a recipient (among others Alesina & Dollar, 2000). Previous authors have often introduced
15
dummies that marked past colonial ties between recipient and donor. Moreover, authors have introduced dummies for the special aid relationship between the U.S and its recipients Israel and Egypt. Dummies have also been introduced to capture the strong aid assistance relationship between Japan and many Asian countries. However, the regression methodology that is employed in the current paper is based on a group-wise fixed effects regression. These fixed effects therefore already capture special ties between donor and recipient country pairs, which makes dummies superfluous.
The occurrence of conflicts such as war have an important influence on flows of aid. The International Peace Research Institute of Oslo (PRIO) has an extensive dataset that describes all internal, interstate or international conflicts of the past 100 years for all countries. Following Berthélemy (2006) all thesese different types of conflicts will be aggregated into a single conflict dummy that indicates a conflict in a given recipient country in a given year. A conflict is only considered a conflict if the number of deadly casualties is larger than 12 (PRIO, 2014).
The proxies for “good policy” will be chosen according to Neumayer’s (2003) proposal of using the Political Terror Scale (PTS) and the Freedom House Index (FHI) instead of past economic growth. Both measures are freely available on the websites of the corresponding independent institutes. The FHI consists of a score for civil liberties and a score for political rights In this paper I will aggregate these scores into one single scale measure from 2 (best) to 14. Clist (2011) introduced both measures into his aid allocation equation without acknowledging that we are dealing with a scale variable. I will therefore explicitly take distance from his approach and treat both the PTS and FHI as qualitative variables and hence introduce dummies that mark above median scores for both measures. Additionally, because multilateral aid has been proven to be allocated to countries with better policies, I introduce the multilateral aid per capita for the recipient country. This does not imply that multilateral aid causes good policy, but only captures the fact that they might be closely associated (Berthélemy, 2006).
Penn World Tables (PWT) provide data on population and GDP per capita of both the donor as well as the recipient countries for the 1980 – 2010 period (Heston et al, 2012). The population is used to correct for the fact that large recipients will receive more aid in absolute terms. To correct for the nonlinearities introduced by the population bias as discussed in section 2.2 I will use the log of population. The GDP per capita reflects the recipient need i.e. do the poorest and neediest countries receive more aid? Since the nonlinearities introduced by a potential middle income bias as observed by Alesina and Dollar (2000) I will also use the
16
log of GDP per capita in my regression. Additionally, the World Bank provides data on the infant mortality rate of all countries over the past 30 years. Although it has never been proven significant in previous papers, I will once more test whether this health development indicator exhibits any influence on the aid allocation process.
Bilateral trade between donor and recipient will be included to model the donor’s commercial interests. However, trade might be endogenous to aid i.e. aid might induce trade and trade might induce aid, but due to the fact that my model is based on lagged terms of all explanatory variables, this will not cause any problems. Trade will be defined as following:
. The term is the ratio of total trade to the GDP of the donor country,
hereby indicating the relative importance of the bilateral trade to the donor (Berthélemy & Tichit, 2004). Data on bilateral trade is available on the Correlates of War website (COW, 2014).
The main contribution of my paper is to extend the notion of commercial interests of the donor by looking whether aid flows are directed to countries that have potentially valuable natural resources. As discussed in section 2.1.4. I will employ two measures of the availability of natural resources. Natural resource rents reflect the current exploitation of natural resources as a percentage of GDP. Alternatively natural capital is a measure that more accurately reflects the natural resource wealth of a recipient country now and in the future. Since natural capital is given in absolute terms I will use the log to make it an appropriate measure. Both measures are available in the World Bank database (World Bank, 2011).
Appendix 1 and 2 provide summary statistics and histograms of all described variables. The histograms show that most of the variables are approximately normally distributed. The variables that are clearly not, are the scale variables. The theory provided by Feeny & McGillivray (2008) that aid allocation is a bureaucratic process that takes some time to implement, will also be applied in this paper. My paper will therefore build on the assumption that the aid allocation decision for year t is based on the situation in t-1. This implies that I will lag all my recipient specific variables with one period. All amounts used are in dollars and at current prices.
3.2 Methodology
In order to properly model the aid allocation process, one has to distinguish between two different processes: the selection process and the process that determines the level of aid. Contemporary authors have used three different approaches that incorporate these two
17
different processes: Tobit, Two Part Models and Heckman selection models. Each of these approaches are built on important assumptions and this paper will test which approach is most feasible based on these assumptions. Following sections will first present the three models more mathematically and evaluate the most important assumptions. Section 3.2.5 will test these assumptions using a simplified aid allocation model and make choice on the best specification for my case.
3.2.1. Tobit type I model
As discussed earlier the aid variable can be defined as a normal censored variable, which formalizes to:
(6)
{
(7)
James Tobin (1958) proposed different types of models to deal with latent variables. For a normal censored variable he proposed the so called type I approach. If we translate this approach to the aid allocation story, the cross sectional aid allocation model would look like this:
(8)
With such a model one can estimate two different expected values of the latent variable y. Because y is partly censored, the distribution of y has to be inferred using the distribution of X. Greene (2007) formalizes these inferred expected values of y using the Inverse Mills Ratio (probability density /cumulative density) of X. Where refers to the probability density and refers to the cumulative density:
|
=
] (10)
For my analysis I am only interested in the countries that did receive a nonzero amount of aid, hence the first (9) specification is the most appropriate. The Tobit model has an important implication regarding the effects of X on both the selection of countries and the decision on how much a country receives. The following derivations can clearly illustrate this:
18 | [ ( ) ] (11) | ( ) (12) From equation (11) and (12) it follows that the selection model and intensity equation are both positive multiples of and are hence always affected by changes in X with the same sign. This implies that the Tobit type I model is only feasible in a situation where coefficients in the selection process and the level process are expected to have the same sign.
3.2.2 Cragg Two Part Model
As a response to the strong assumptions that have to be made on the sign of in the Tobit type I model, Cragg (1974) developed a two part model that explicitly allowed for differences in the selection and the level process. Such a model combines the concepts of censoring and truncation. It regresses a dummy indicating nonzero aid on a set of regressors using probit to overcome censoring. For the level equation it uses a truncated normal regression that only looks at nonzero amounts of aid (Burke, 2009). Such a model can be written as:
{
The crucial assumption of such an approach is however that the . But in the setting of the aid allocation process it seems rather unlikely that the selection process is completely independent from the process that determines the level of aid.
3.2.3 Heckman sample selection model (Tobit type-II)
James Heckman (1976) developed a statistical method that combines the best elements of the Tobit type I model and the two-part model.
{
(17) | |
19
=
= (18)
The Heckman method also involves a two-step procedure, but assumes that the selection and the level process are interdependent as can be seen from equation (17). Heckman uses a transformation of the probabilities obtained in the selection equation to correct for nonrandom selection into the level equation (Greene, 2007). The transformation of these probabilities is equal to the Inverse Mills Ratio ( . Note that the IMR used in the Tobit equation is a transformation of the distribution of the vector .
The specification of Heckman allows for coefficients with different signs in both the selection and level equation and moreover allows for nonzero correlation between the error terms of the selection and level equation. However, equation (15) and (16) explicitly distinguish between a vector z and a vector x, implying that the selection process is not determined by the same regressors as the level process. Vector z and x are allowed to contain the same variables, though at least one variable needs to be omitted (a variable that appears significant in the selection equation and insignificant in the level equation or vice versa). This assumption is referred to as the exclusion restriction and according to Puhani (2000) the results become inconsistent when it is not satisfied. To make the identification mechanism even more robust, some econometricians argue that the vector z should not be a subset of vector x (or vice versa). The latter implies that we additionally should exclude a variable from vector x that is also present in vector z. Berthélemy (2006) and Clist (2011) did not add this extra restriction to their specification.
3.2.4 Three-dimensional panel data
Panel data often consists of data in two dimensions an entity (i) and time (t) dimension. However, my paper studies the aid allocation behavior of 23 donor countries to 150 potential recipients over 30 years and hence has an extra (j) dimension. Such a design complicates the estimation of what I would like to call: “average donor behavior”. Average donor behavior refers to an estimation of the model using the full dataset including all donor countries. Software packages such as Stata 12 are not equipped with commands to deal with panel data in more than two dimensions. One way to deal with this problem is to create unique donor-recipient pairs. Fixed or random effects are in such a case based on these donor-donor-recipient pairs and not on donor or recipient specific effects. Another advantage of such an approach is
20
that I do not have to create dummies that represent colonial or special ties between donor and recipient.
Subsequently my analysis will focus on the donor countries separately to determine what donor countries are driven by self-interest and which are driven by rather altruistic motives. The latter approach allows me to drop the three-dimensional panel approach and to work in two panel dimensions. Fixed or random effects are in this case based on the receiving country.
Table 1 Example of Three-Dimensional Panel Data
Donor Recipient Year AID GDP p.c. Population Etc.
Australia Afghanistan 1980 Australia Afghanistan : : : : : Australia Afghanistan 2010 Australia : Australia Zimbabwe 1980 Australia Zimbabwe : : : : : Australia Zimbabwe 2010 : : : : : : : US Zimbabwe 2010
3.2.5 Discussion on the preferred specification
This section will discuss the preferred specification of the aid allocation equation by evaluating the assumptions of the three LDV models presented earlier. With a two-part Cragg estimation I am able to test all the assumptions of the three models. The model I use is a simplified model of my final specification and is very similar to the model used by Berthélemy (2006). Table 2 shows the results of this regression. We can draw some important conclusions just from this table. First we see that in the selection equation GDP per capita appears with a positive sign, whereas the sign is negative in the level equation. This implies that a Tobit regression will yield inconsistent results, remember that from equation (11) and (12) it follows that beta always enters both processes with the same sign. Another important conclusion that we can draw from this table is that in case of a Heckman approach it is likely
21
that the results will satisfy the exclusion restriction. Note that the freedom indicator appears only significant in the selection process.
Table 2: Two-Part Model
Selection Level σ p-value σ p-value Ln(GDPpc) 0.485863*** 0.01669 0 -0.61474*** 0.01483 0 High PTS -0.00089 0.01871 0.962 -0.11016*** 0.01644 0 Low Freedom -0.17774*** 0.02269 0 -0.02278 0.01929 0.238 Ln(pop) 0.689786*** 0.02256 0 0.04916*** 0.01612 0.002 Ln(trade) -0.05086*** 0.00614 0 0.14836*** 0.00655 0 Ln(aidothers) 0.258124*** 0.00874 0 0.36805*** 0.00792 0 α -10.495*** 0.25018 0 -2.78897*** 0.21082 0 R-squared .2713 Observations 81228 49388 *** Significant at 1%
In order for Heckman to be the preferred specification over the Two-Stage model presented above, we need to find evidence of a correlation between the error terms of the selection and level equation. Because we do not observe the error terms, we will predict the residuals and test whether their correlation is significant. Table 3 shows that although the correlation between the residuals is surprisingly low, it is significantly different from zero at a 1% level. We can hereby conclude that the two step Heckman model is our preferred specification. Table 3: Correlation between Residuals
Residual (Selection) Residual (Level)
Residual (Selection) 1
Residual (Level) 0.0472*** 1
*** Significant at 1%
The introduction of either random or fixed effects in the level equation is another important point of discussion. Table 4 shows the output of a Hausman test that tests a fixed effects
22
specification against a random effects specification. The p-value indicates that the fixed effects specification is preferred at 1% significance. Note that the selection model can only be estimated using random effects, because probit estimation cannot be consistently estimated with fixed effects.
Furthermore in a fixed effects regression we might want to control for year fixed effects. To test whether such effects exist, we include year dummies and test their joint significance. This test concludes that year fixed effects are significant at a 1% significance level and will hence be added to the specification.
Table 4: Hausman test Fixed vs Random Effects
Fixed Random β β ∆ σ(∆) lntrade 0,1036 0,1484 -0,0447 0,0038 lnpop -0,2939 0,0492 -0,343 0,0465 lngdppc -0,5182 -0,6147 0,0965 0,0172 highpts -0,1087 -0,1102 0,0015 0,0034 highfreedom 0,0161 0,0228 -0,0067 0,0064 lnaidothers 0,384 0,368 0,0159 0,0028
Test: Ho: difference in coefficients not systematic Chi-squared (6) = 172.6
Prob>Chi-squared = 0.000
Additionally we have to decide whether to include heteroskedasticity robust standard errors in our specification. By performing a modified Wald-test for group-wise heteroskedasticity I have to reject the null hypothesis of homoscedasticity at a 1% significance (see Table 5). Stata 12 has a built in robust option to calculate heteroskedasticity robust standard errors which I will employ in all regressions presented in the results section.
Table 5: Modified Wald-Test for Groupwise heteroskedasticity
for all i
Chi-squared (2941) 2.0e+34
Prop>Chi-squared 0.000**
23
The final specification of the aid allocation model for this paper is therefore a two-step Heckman model with random effects in the selection equation, group fixed effects and time fixed effects in the level equation and heteroskedasticity robust standard errors.
4. Results
This section will discuss the estimations of the aid allocation model we specified in the previous section. Stata 12 is not equipped with a tool to estimate a two-step Heckman approach using panel data. We therefore have to estimate the selection and level equation separately and calculate the Inverse Mills Ratio from the linear prediction of the probit regression. The first part of this section will study the behavior of the average donor country, in other words we estimate the model using our full panel of 23 donor countries, 150 potential recipients over the period 1980-2010. The second part will study how these average effects have changed over time. The third section studies the DAC donor countries separately by performing separate regressions for each donor.
4.1 Average Donor Behavior
{ (19)
Equations (19) and (20) show the exact equations I have estimated. The selection equation 19 is estimated using a random effects probit regression. The fixed effects regression is estimated with the observations for which . Where is the fraction of donor i’s budget allocated to recipient j at time t, is a vector of explanatory variables, represents the donor-recipient pair fixed effects, are the time fixed effects and is the Inverse Mills Ratio calculated with the linear prediction of the probit regression. To fulfill the exclusion restriction, the total aid provided by the donor is ommited from the level equation (20). When the aid budget of a donor is increasing the probability for a recipient country to be selected increases, but the fraction from the budget it receives will not be affected. Additionally the conflict dummy is excluded from the selection equation, because it always appeared insignificant regressions where it was included.
24
Table 6 shows the results of a regression that estimates the average behavior of the 23 DAC donor countries over the past 30 years. The specification uses resource rents as a proxy for natural resource abundance. The dependent variable in the selection equation is a binary variable equal to one if donor (i) allocates an amount of aid bigger than zero to recipient (j) and that is zero otherwise. Our main coefficients of interest are the ones presented in column 2. The level regression model is mainly a log-log specification, implying that we are dealing with elasticities which make the coefficients convenient to interpret.
The dataset consists of 68,812 complete observations and covers the 1980-2010 period. However, because the aid variable is often censored (0) we have to estimate our level equation using only positive amounts of aid, which leaves us with 46,963 uncensored observations in the equation of interest. To my knowledge this is the largest panel dataset ever employed in the field. Table 6 shows that the exclusion restriction has been optimally satisfied by the exclusion of the log of total aid by the donor in the level equation and the exclusion of the conflict dummy in the selection equation. Furthermore we observe a strong correlation (ρ) between the selection and the level process. These two findings strengthen our choice regarding the two-step Heckman procedure.
Table 6: Two Step Heckman using Resource Rents (1980-2010)
(1) (2)
VARIABLES Selection Level
Β(σ) Β(σ)
Ln(trade), lagged 0.0504*** 0.0683***
(0.00743) (0.0132)
Ln (GDPpc), lagged -0.366*** -0.301***
(0.0275) (0.0677)
Infant Mortality Rate, lagged -0.00595*** -0.00901***
(0.000777) (0.00164)
Resource Rents, lagged 0.000530 0.00922***
(0.00105) (0.00181) PTS>Median, lagged -0.0863*** -0.0718*** (0.0216) (0.0225) FHI>Median, lagged 0.0112 -0.000367 (0.0273) (0.0321) Conflict, lagged -0.0394 (0.0318) Ln(multilateralaidpc) 0.0695*** 0.137*** (0.00900) (0.0111) Ln(aidotherspc) 0.196*** 0.357*** (0.0109) (0.0177)
25
Ln(Population) 0.478*** 0.564***
(0.0186) (0.215)
Inverse Mills Ratio 0.930***
(0.0889) Ln(totalaiddonor) 0.611*** (0.0134) Constant -2.422*** -6.042*** (0.349) (2.102) Censored Observations 68,812 Uncensored Observations 46,963 R² 0.081 Number of pairs 3,173 2,825 Group FE Yes Year FE Yes ρ 0.5900
Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
The results in table 6 mainly confirm our expectations about the included regressors. The positive sign of the lagged logartithm of trade confirms the earlier finding by Berthélemy & Tichit (2004) The elasticity implies that every percent increase in trade (imports+exports) between a donor and recipient country results in a 0.0683% increase in the fraction of aid allocated to the receiving country. The following numerical example illustrates how to interpret the elasticities.
For example if a country initially receives 10% of the aid allocation budget of a given donor and increases trade by 20% (as a percentage of donor GDP), it is expected to gain 20*0.0683=1.366% on the budget and now receives 10.1366% of the budget (ceteris paribus). For a donor distributing 10 billion each year, this would imply that this recipient would gain 13.66 million each year in aid.
Furthermore the negative coefficient on the log GDP per capita of the receiving country implies that on average more aid goes to countries that have low GDP per capita. Hence, on average donor countries seem to be responsive to this proxy for recipient country “need”. However, if we look at the infant mortality rate we observe a surprising result. On average donor countries seem to allocate less aid to countries with high infant mortality rates. Previous work never found any link between the infant mortality rate and aid allocation behavior, but the link I find is rather surprising. Although the correlation between GDP per capita and the infant mortality is quiet high (-0.6947), there is no reason to believe that the
26
specification suffers from multicollinearity. The finding confirms the general result that donor countries are not very responsive to the development needs of recipient countries and moreover implies they avoid countries with poor health conditions.
The Political Terror Scale (PTS), Freedom House Index (FHI) and Conflict dummy have been included to check whether donor countries are responsive to recipient country merit. The results show that they do, but they seem only responsive to the level of Political Terror. The negative coefficient implies that countries that have a PTS score above the median, receive significantly less aid. An above median recipient country’s aid is reduced by a factor 0.9307 next year, so if a country initially received 10% of a certain donor’s
budget, it now only gets 9.307%. The FHI dummy does not appear to be significant, which means that donor countries are on average not responsive to the Civil Liberties and Political Rights situation in the recipient country. One might argue that this result might be due to multicolinearity beween the PTS and the FHI. However, a simple correlation check shows that the correlation between these dummies is only 0,2040. The insignificant conflict dummy moreover shows that the existence of a conflict (war) does not result in less or more aid to the receiving country the following year.
The control variables multilateral aid per capita, total bilateral aid per capita by other donors and the population all have the expected sign. Donor countries respond to the fact that multilateral agencies provide aid to a given recipient, by allocating more bilateral aid assistance. The same is true for more bilateral assistance by other donor countries. The significant positive population coefficient implies that larger countries receive a larger fraction of the aid budget of donor (i), which is rather straightforward. The elasticities are relatively high, mainly because the regressors are the log of absolute amounts rather than the log of a ratio. If for instance bilateral aid by other donors goes up by 1%, a given donor will increase its aid by 0.357%.
The main variable of interest for this research is the coefficient on the presence of natural resources. This research hypothesized that donor countries strategically provide aid to countries that have precious natural resources (oil, coal, gas etc.). This hypothesis is tested using two different proxies for natural resources: natural resource rents and natural resource wealth. Table 6 shows that the hypothesis is confirmed for natural resource rents. Donor countries significantly allocate more aid to countries that have relatively more production of natural resources. This finding is significant at the 1% level and implies that for every percentage of GDP that is produced with natural resources, aid goes up by a factor 1.0093
27
( . In other words a country that produces 10% of its GDP with natural resources, receives in general 9.3% more aid than a country that produces no natural resources . For example, if it initially received 10% of donor (i)’s budget, it now receives 10.93%. Alternatively, I have used a proxy for natural resource wealth that possibly better reflects the potential future benefits allocating aid. The main shortcoming is however, that measures for natural resource wealth for every country are only provided every 5 years by the World Bank. This substantially lowers the number of observations in my panel to only 4711 observations. Appendix 4 provides the regression results of this regression. The main conclusion that can be drawn from this regression is that there is no significant elasticity between natural resource wealth and the fraction of the aid budget. Although in column (1) we see that natural resource wealth exhibits a negative influence on the probability of being selected as recipient.
The overall fit of the model seems low (R-squared =0.081), but this is due to the calculation method that Stata 12 employs for fixed effects. Correcting the residual for fixed effects results in an overall R-squared of 0.729.
4.2 Average donor behavior: Heterogeneity over time
Aid allocation behavior has been studied for decades and many results have received a lot of attention from policymakers (for instance Burnside & Dollar (2000) finding that aid effectiveness depends on a good policy environment). Additionally the political environment has been changing and the focus on effectiveness has been increasing with the emergence of more rightwing governments. One might therefore argue that aid allocation behavior has possibly changed over time. This section looks at three separate 10-year periods and discusses the observed changes over time.
Table 7: Two-Step Heckman Estimations (Level Equation)
(1) (2) (3) VARIABLES 1981-1990 1991-2000 2001-2010 β(σ) β(σ) β(σ) Ln(trade), lagged 0.0672*** 0.113*** 0.00866 (0.0247) (0.0195) (0.0061) Ln (GDPpc), lagged -0.370*** -0.371*** -0.402*** (0.139) (0.109) (0.115)
Infant Mortality Rate, lagged -0.00183 -0.00408 -0.00416
(0.00422) (0.00367) (0.00369)
Resource Rents, lagged 0.00323 -0.00184 0.00540**
28 PTS>Median, lagged -0.0291 -0.0701** -0.00643 (0.0390) (0.0348) (0.0292) FHI>Median, lagged -0.0710 -0.0325 -0.120** (0.0547) (0.0433) (0.0510) Conflict, lagged -0.0913* -0.0883** -0.0962** (0.0547) (0.0418) (0.0404) Ln(multilateralaidpc) 0.106*** 0.0903*** 0.0973*** (0.0225) (0.0146) (0.0137) Ln(aidotherspc) 0.251*** 0.213*** 0.341*** (0.0311) (0.0223) (0.0234) Ln(pop) 1.291** -0.421 1.578*** (0.563) (0.488) (0.478) invmills 0.968*** 1.117*** 1.149*** (0.125) (0.0918) (0.109) Constant -13.06** 2.585 -15.70*** (5.175) (4.604) (4.794) Uncensored Observations 16,396 24991 27425 Censored Observations 10,693 16190 20080 R² 0.032 0.037 0.065 Number of pairs 1,534 2,411 2,627
Group FE Yes Yes Yes
Year FE Yes Yes Yes
ρ 0.4006 0.3861 0.2672
Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
The major insight that can be drawn from table 7 is that the effects are not stable over time and that aid allocation behavior has been changing over time. For instance, we observe that the trade variable only exhibits influence in the first two decades of my panel. The coefficient doubles in the 1990’s decade but then drops to almost zero after the millennium. This indicates that self-interest might play less of role after the millennium. However, at the same time we observe that the presence of natural resources starts playing a significant role only after the millennium. Actually we observe a shift in the self-interest strategy from donor countries from aid allocation to trading partners towards counties with relatively a lot of natural resources.
On the recipient need side we observe that the elasticity between GDP per capita and aid has been increasing over the past three decades. Hence, donor countries have directed more aid to the countries that need it most in terms of income. On the other hand, from the millennium onwards we observe that countries that are needy in terms of health (Infant
