Supervisor: prof. dr. J.P. Elhorst
The “spatial trap” of aid to Africa
Mathilde de Wilt (S1998935)
126 June 2015
Abstract
By investigating 37 African countries over 24 years, this research illustrates that aid has made no negative or positive contribution to institutional quality, also when taking into account spatial dependence. Extending the model towards a spatial model is efficient and indicates that the quality of institutions is related to spatial interaction effects. The quality of institutions of an individual country depends upon the institutional quality of neighbouring countries. This spatial dependence also explains the significance of the geographical variables in regressions with aid as dependent variable. However, it should be noted that the results are robust to most model specifications but by including country specific effects the significance of the spatial interaction effects disappears. In addition, a gap in the literature has been identified towards the relationship between mortality rates and institutional quality.
Keywords: Aid • institutions • spatial dependence JEL classification: C21, E02, F35
2
1. Introduction
Following the recent crisis, the volume of aid is at the centre of debate as countries try to decrease expenditures to improve government budgets (BRON?). In addition, the effectiveness of aid is doubtful which also raises questions about the appropriate levels of aid. This research focuses on Africa which is the most aided region in the world, but the development performance has been rather disappointing in almost all African countries (Lancaster, 1999). Although there is quite some heterogeneity in growth rates of African countries. In general, it can be said that the combination of low growth rates and high population growth resulted in falling average per capita income levels since 1970. It is clear that the high levels of aid to Africa did not lead to the intended results.
A possible reason for the ineffectiveness of aid is that high levels of aid may have adverse effects on institutions because among other things, aid supports repressive governments and aid reduces the need for taxes. Especially in Africa, where foreign aid has merely a permanent character, it can negatively influence institutions. Secondly, large volumes of trade can have negative effects on the exchange rate, also known as the ‘Dutch Disease’ (Lancaster, 1999). Others have argued that aid works best in countries with high quality institutions (see e.g. Burnside and Dollar, 2000). The outcome of the research of Burnside and Dollar (2000) resulted in a conditional character of aid programs from the IMF (Easterly, 2003). Moreover, Africa is a continent which has low quality institutions because of among other things, its high level of political instability and colonial past. This low quality of institutions may provide a reason why aid did not result in high growth rates. However, this argument is at odds with the traditional argument, saying that countries need aid in order to develop good institutions (Moss, Petterson and Van de Walle, 2005).
The discussion whether institutions deliver high growth rates or that good institutions are the result of economic growth, has been going on for decades. However, recently a broad consensus has been identified that ‘good’ institutions stimulate economic growth and are not the result of economic growth (e.g. Acemoglu, Johnson; Robinson, 2001). It is clear that ‘good’ institutions are not widely present in African countries and that might be a reason for the low growth rates in Africa. This paper will conduct research about the determinants of the quality of institutions for Africa. It will look at the effects of foreign aid on institutional quality also when taking into account spatial dependence. The hypothesis of this research is that aid has a negative effect on institutions. Aid helps to maintain the status quo, will keep repressive governments in place and will have adverse effects on economic growth via lower institutional quality. This is because aid reduces the need to collect taxes and reduces the accountability of the government.
3
setting. Also, droughts, quality of infrastructure and epidemics, such as the outbreak of Ebola, are not bound to borders. Moreover, refugees driven from their home country turn to neighbouring countries, increasing the need for aid to these countries. If the quality of institutions is influenced by spatial interaction effects, then this result can provide valuable insights for Africa since the low quality of institutions in that continent limits economic growth. By increasing the understanding about the determinants of institutional quality, appropriate policy inference can be based on the increased knowledge. This paper is the first research that includes spatial interaction effects in relation to the quality of institutions in Africa.
Some researchers have identified geographical factors as important drivers behind the ineffectiveness of aid (Sachs and Warner, 1997; Dalgaard, Hansen and Tarp, 2004). According to many researchers, this result is rather unsatisfactory since there is no strong theory connecting the effectiveness of aid and geography (Rajan and Subramanian, 2008; Dalgaard et al. 2004). This research hypothesizes that geographical factors turn out significant because these studies fail to take into account spatial dependence. According to Anselin (2006, p. 901) ‘Spatial dependence is a special case of cross-sectional dependence, in the sense that the structure of the correlation or covariance between observations at different locations is derived from a specific ordering, determined by the relative position (distance, spatial arrangement) of the observations in geographic space’. In other words, spatial dependence measures the influence of relative geographical location of countries. By not considering spatial interaction effects while these are important, research will render biased results. Another contribution of this paper, is the usage of exogenous instruments. Many researchers acknowledge the endogeneity of aid but fail to include exogenous instruments in their regression since it is hard to find exogenous instruments. By examining the spatial impact of institutions, many new potential instruments arise, because exogenous variables from other regions may serve as an instrument for aid, as proposed by Drukker, Egger and Prucha (2013). More on this in section 8.
The third contribution is that a gap in the literature has been identified. There is no literature on the relationship between mortality rates and institutional quality at one specific point in time. This paper argues that high mortality rates will disrupt society and will lower institutional quality. In addition, empirical evidence will be provided to strengthen this claim.
The remainder of this paper is as follows. Section 2 discusses the theory on institutions, aid and spatial dependence. Also, the existing literature will be reviewed. The third section introduces the data. Section 4 deals with the instrumentation strategy, whereas section 5 introduces the model. In section 6 the results of the non-spatial model will be discussed. Thereafter, there is a seventh section about how to model spatial dependence and section 8 deals with the instrumentation strategy for the spatial model. In addition, the results of the model with spatial interaction effects will be discussed and the paper is completed with a conclusion in section 10.
2. Literature review
4
Institutions are broadly defined as the rules of the game. Elements of efficient institutions are protection of property rights and contract enforcement. According to North (1990), economically efficient institutions are institutions that motivate self-interested individuals to act in ways that contribute to collective welfare and economic development. ‘Institutions provide the incentive structure of an economy; as that structure evolves, it shapes the direction of economic change towards growth, stagnation, or decline’ (North, 1990, p. 97).
Countries with high quality institutions, less distortionary policies and more secure property rights will invest more in human and physical capital, and will use these factors more efficiently to achieve a greater level of income (Acemoglu, Johnson and Robinson, 2001). High quality institutions may be good for growth because it stimulates innovation and investment and thereby economic growth. A system of well-defined property rights spurs investments since the investor is assured that he will reap the benefits of its investment. This will also stimulate innovation and structural change.
It is clear that good institutions and economic growth are related but the establishment of causality has long been a hotly debated topic. Are institutions good for economic growth or does economic growth lead to the evolvement of better institutions since the economy is more complex? This empirical question has been convincingly solved by Acemoglu et al. (2001). In their paper the dependent variable is economic growth and mortality rates of soldiers, sailors and bishops are used to instrument current institutions. The reasoning behind this is that more extractive and low quality institutions are created, when there is high (potential) settler mortality in a colony. The authors document a negative relation between (potential) settler mortality and economic performance. This relationship can only run via the institutional channel (see figure 1 for a schematic summary of their theory). Their research gained importance because it is difficult to find a better exogenously given instrument and by this means, the authors solved the problem of endogeneity. In other words, it is (almost) impossible to connect growth performance to (potential) settler mortality other than via the institutional channel. They resolved the discussion and provided evidence that in order to develop an economy, one need sound public institutions. Recent, a broad consensus has emerged among economist that ‘good’ institutions are a prerequisite for economic growth and not a result (Rodrik, 2000; Acemoglu et al. 2001). However, Albouy (2012) is very critical on the choice of instruments of Acemoglu et al. (2001). He claims that the instruments suffer from severe measurement problems, which leads to unreliable results.
Figure 1: Schematic overview of the theory of Acemoglu et al. (2001).
According to many researchers, the low institutional quality in Africa is the reason for its disappointing growth performance. Africa’s problem is a lack of institutional development and appropriate policies, as stated in Lancaster (1999). More than a decade ago, the World Bank argued that “underlying the litany of Africa’s development problems is a crisis of
5
governance”2. Also empirical research point to this problem as several scholars are able to
identify the significance for development of sound public institutions. Rodrik, Subramanian and Trebbi (2004) conduct research on factors that can explain differences in income levels around the world. They try to disentangle the contributions of geography, trade and institutions to income levels. The finding is that institutions are by far the most important contributor to differences in income levels. Easterly and Levine (1997) claim that Africa’s growth tragedy of low quality of institutions, is because of the high ethnic diversity in this region. High ethnic diversity may lead to tensions, political instability, low schooling etc. Africans are still harmed by their colonial past and political instability shows the dissatisfaction of the African population.
Aid and institutions
If high institutional quality is that important for economic growth, the question arises how Africa can achieve better institutions. One way to accomplish this and stimulate institutional change, is via the provision of aid. Aid is supposed to provide temporary financial assistance in order to encourage certain long-term behaviors such as revenue collection, investment in physical and human capital and the establishment of the institutions of a developmental state. In Botswana for instance, this strategy worked and aid assisted the country to support local efforts and the country’s aid was gradually decreasing (Moss, Petterson and Van de Walle, 2005). However, there are also many examples where aid discourages revenue collection, results in distorted decision-making, and undermines the incentive to build state-capacity. In this case, aid can be viewed as harmful for institutional development. This is also known as the aid-institutions paradox since aid is supposed to benefit the quality of institutions but in fact it may harm it (Moss et al., 2005). Theory on aid provides conflicting guidance. On one hand, aid can release governments from revenues constraints and enable them to reform, provide training and pay more to civil servants (Knack and Bräutigam, 2004). On the other hand, provision of aid can have negative effects on institutional quality. Aid may reduce the need to reform because the government receives already money via the aid-channel, so there is no need to reform and improve the government budget. Secondly, receiving aid helps to maintain corrupt and repressive governments (Szirmai, 2004). Moreover, obtaining aid may lead to dependence on aid because aid flows may get a permanent character. The idea behind development aid is that is contributes to building up productive potential and building up the institutional structure of a country in order to promote self-reliance of a country. The central dilemma is the tension between self-reliance and dependence (Szirmai, 2004). Especially in Africa where the aid flows seem to have a permanent character, this is a huge problem.
Bräutigam and Knack (2004) summarize some mechanisms that could explain a negative relationship between institutional quality and foreign aid. High levels of aid can make it more difficult to solve the collective action problems that are deep-rooted in reform efforts. The collective action problem was first considered by Olson (1965) and illustrates that individual and collective interests do not always coincide. This may lead to an undersupply of public goods (Osborne, 2009). Aid may aggravate this collective action problem by increasing a ruler’s resources. If a ruler has more resources, the burden to provide public goods is higher and this may lead to an even more skewed distribution of public goods across income.
2World Bank, Sub-Saharan Africa: From Crisis to Sustainable Growth: A Long- Term Perspective Study (Washington, D.C.:
6
Other problems that may arise via the provision of high levels of aid is moral hazard. Moral hazard can exist both on the recipient-side and the donor side. Moral hazard problems are greater when the aid-levels are higher. This is because higher protection in term of aid goes hand in hand with higher potential distortive problems (Gruber, 2013). Furthermore, higher levels of aid reduces the need to collect taxes, may soften budget constraints which will decrease the need to reform and will also lead to a lower level of accountability for the government. Associated with this, is the theory of the tragedy of the commons. The tragedy of the common theory was brought up by Hardin (1968) and points out that individuals have a tendency to overuse common property. In this case, common property is the government budget, and providing aid will lead to softer budget constraints and higher problems of tragedy of the commons. Which effect, the positive effect of aid or the negative one, dominates is not clear a priori so empirical research is needed to resolve this matter.
Some researchers document a positive relation between aid and development. Kosack (2003) finds that aid can be effective in improving quality of life in a democratic environment. In addition, Tezanos, Quiñones and Guijarro (2013) investigate the relationship between levels of aid and inequality-adjusted GDP per capita growth rate in Latin-America and in the Caribbean. They document a positive effect of aid on growth rates. By distinguishing between developmental and non-developmental aid, Minoiu and Reddy (2010) find that developmental aid provides long run growth. Another article which shows that aid can be effective, is the research of Gupta, Pattillo and Wagh (2009). The authors conduct research on the effect of remittances on financial development. The finding is that remittances, which are a stable, private transfer, have a direct poverty-mitigating effect, and promote financial development. The advantage of this research is that remittances are not endogenous and therefore no instruments are needed. The positive effect of remittances indicate that also aid can potentially benefit society.
7
Conditionality of aid
The conditionality of aid was part of the development-aid debate because of conducted research by Dollar and Burnside (2000), they find that ‘aid has a positive impact on growth in developing countries with good fiscal, monetary, and trade policies but has little effect in the presence of poor policies’ (Dollar and Burnside, 2000, p. 847). This made the World Bank decide to condition the provision of aid on the presence or development of good policies and institutions. However, this goes against the traditional arguments of aid, which sees aid as a mean to provide good policies and institutions and not a condition (Szirmai, 2004). According to a report from the World Bank conditionality of aid only works if reforms are welcomed by the government and considered useful (Devarajan, Dollar and Holmgrem, 2001).
Spatial dependence
Sachs and Warner (1997) also investigate the forces behind the slow growth of Africa and find that, among other things, geographical factors such as tropical climate and lack of access to sea contributed to the slow growth of Africa. Dalgaard, Hansen and Tarp (2004) find in their paper that aid is significantly less effective in tropical areas. They find it very hard to believe that geographical factors can explain the effectiveness of aid and point to further research on this. I hypothesize that the significant results on geographical factors influencing aid effectiveness can be explained by spatial interaction effects among countries. Because spatial dependence is not accounted for in the research, this effect shows somewhere else. In this case the effect of spatial dependence `contaminates´ geographical variables, as it is much related to spatial interaction and location. Spatial dependence models a special kind of geographical influence and is backed by theory.
Brueckner (2003) has illustrated that strategic interaction among governments exists by analyzing a spatial model. Brueckner (2003) shows that the tax rates of neighbouring countries coincide, which suggests that countries look at neighbours before setting their tax rate. If governments indeed interact with each other, then governance could also be related between areas. Seldayo et al. (2010) investigate if spatial dependence in governance quality is present. Their results point out that spatial dependence is relevant and that the quality of governance is clustered. Measuring the quality of governance while not considering these spatial interaction effects render biased results. This suggests that in Africa the quality of governance may also be influenced by its neighbouring countries and spatial interaction effects matter.
8
Spatial dependence may also be important in the distribution of aid in Africa. As an illustration, when a country suffers from a civil war, a neighbouring country may receive aid in order to take care of the refugees from the civil war. The outbreak of Ebola has also shown that such outbreak is not bound to borders and has a regional character. The aid flows received by these countries in order to help these countries also have a regional character. Moreover, areas of drought are also spread beyond country borders and aid to improve this will also benefit other countries. Rajan and Subramanian (2007) provide evidence that aid is associated with weak governance because aid removes the need for taxes. However, Brueckner (2003) has pointed out that governments strategically interact over the tax rates so one should take that results into consideration by including spatial interaction effects, to prevent biased results.
3. Data
Variables
To estimate the relationship between aid-levels and institutional quality, several variables are needed. The dependent variable in the regression is institutional quality. There are two possible candidates used most in the academic literature, to measure institutional quality and it is not clear a priori which one to select, therefore both variables will be used in the model. The International Country Risk Guide (ICRG) dataset provides a dataset which measures the quality of institutions. For this matter, three factors of the ICRG-dataset are selected: rule of law, bureaucratic quality and corruption and an average score of institutional quality is calculated. ICRG-score is widely used among researches because of its reliability and large coverage, especially in Africa (Knack and Bräutigam, 2004).
Another variable which also measures the soundness of public institutions, but then via the quality of governance, is the Kaufmann-index. This index contains worldwide governance indicators, is recent and is used by Seldayo et al. (2010). To calculate institutional quality all 6 dimensions: Voice and Accountability, Political Stability and Absence of Violence/Terrorism, Government Effectiveness, Regulatory Quality, Rule of Law and Control of Corruptions are selected and the average is calculated.
To measure aid, the variable net ODA (Official Development Assistance) received as percentage of GNI is retrieved from the World Bank. According to the World Bank website ‘net official development assistance (ODA) consists of disbursements of loans made on concessional terms (net of repayments of principal) and grants by official agencies of the members of the Development Assistance Committee (DAC), by multilateral institutions, and by non-DAC countries to promote economic development and welfare in countries and territories in the DAC list of ODA recipients. It includes loans with a grant element of at least 25 percent (calculated at a rate of discount of 10 percent)’.
9
reforms, as trade reforms reduce corruption and rent seeking activities. Furthermore, GDP per capita is included as a control variable. A higher income may increase the demand for institutional development (Rodrik, Subramanian and Trebbi, 2004). The control variable ethnic tensions is from the ICRG-dataset and is included since ethnic tensions may lead to social unrest and under these circumstances it is harder to develop sound public institutions. Ethnic tensions is measured as the degree of tension attributable to racial, national, or language divisions. Lower ratings (higher risk) are given to countries where tensions are high because opposing groups are intolerant and unwilling to compromise. This variable is included since Easterly and Levine (1997) underline the importance of ethnic fractionalization in the development of Africa. Another control variable is the mortality rate, which is retrieved from the World Bank. Only Acemoglu et al. (2001) document a relationship between mortality rates and the quality of institutions. This relationship runs via colonization and does not imply a direct relationship between these two variables, it only implies a relationship over time. There is no literature that documents a direct link between mortality rates and quality of institutions in one specific point in time, while this link is present. If a country suffers from high mortality rates, it needs resources to deal with these problems and it will disrupt society. Consequently, this will lower institutional quality. I have identified a gap in the literature since mortality rates influence the institutional quality but there is no literature about it.
Because of possible endogeneity of the aid-variable, several potential instrumental variables are needed. Data on area is received from Rajan and Subramanian (2008), data on population size is retrieved from the World Bank. The colonial-index which measures the length of the past-colonial relationship is self-constructed and measured as current year minus year of independence. For example if a country regained independence in 1962, the variable for the year 1984 will be 22. If a country did not have past-colonial relationship a value of 100 was assigned to that specific country. Data on year of independence is taken from Bertocchi and Cenova (2002). Also data on colonial heritage dummies representing a dummy for each country which had colonial history with a specific European country, is retrieved from Bertocchi and Cenova (2002).
Data on aid and institutions
10
Figure 2: Development of aid supplies to Africa over time measured as Net ODA as % of GNI.
Note: The dots represent African countries which have received aid.
Figure 3 depicts a scatterplot with aid as percentage of GNI on the y-axis and the quality of institutions measured using the ICRG-dataset on the x-axis. It is clear by this scatterplot, that aid and quality of institutions are negatively related. If a certain country has received much aid, it has also lower institutional quality compared to a country which has received less aid. However, this relationship needs not to be a causal one. One may argue that aid is send to the countries which have unsound public institutions, because they need money to develop this. This will imply that countries with lower institutional quality, receive more aid and that may explain the relationship. Conversely, one can also argue that considering the recent developments on conditionality of aid, it is the other way around. Due to important research of Burnside and Dollar (2000), who document that aid has a positive impact on growth in developing countries with sound fiscal, monetary, and trade policies, the World Bank decided on selectivity of aid. This means that a country receives more aid, if reforms have been made to improve institutional quality. If that is the case, a positive relationship is expected. To solve the potential endogeneity of aid, instruments for aid are needed. This is because the amount of aid given to countries is related to institutional quality and therefore aid is not an exogenous variable. Including an endogenous variable as independent variable in the regression leads to biased and inefficient estimates (Verbeek, 2012). Therefore, other variables are needed to instrument aid and to make sure the results are reliable and unbiased.
0 10 20 30 40 50 60 70 80 1980 1985 1990 1995 2000 2005 Aid a s % o f GN I Year
11
Figure 3: Scatterplot of aid as % of GNI and the ICRG-score of quality of institutions.
Note: The dots represent African countries and the data is from four periods.
Periods and panel
Before going into the instrumentation strategy, something needs to be told about the empirical model and the data. In this empirical research a panel model will be used to test the hypothesis. The advantages of a panel model is that it increases the reliability of the estimates, because using a panel model enables to include fixed effects and control for time invariant characteristics. This is compared to a cross-section model beneficial, since cross-section models do not have such advantages (see Rajan and Subramanian, 2008; Verbeek, 2012). To be able to estimate a balanced panel, the dataset needs to be divided into periods. It is chosen to divide the data into different periods since the quality of institutions is a persistent variable, which means that this variable will not change much over time and its effects are more longer term. Therefore, the institutional quality will be measured as an average value over a period instead of looking at the value of institutional quality every year. Most, if not all, researchers (see e.g. Rajan and Subramanian, 2008; Burnside and Dollar, 2000) divide their periods ad hoc in equally sized periods without looking at the data. Clearly this method can be improved upon because to arrive at significant results, as much variation as possible should be captured by the model. In order to ensure no valuable information will be lost while picking the periods, the institutional quality is graphed over time in figures B1 and B2 in appendix B. On basis of these graphs, the periods will be classified. Instead of deciding ad hoc which periods to distinguish, the periods are based on the characteristics of the data.
The sample comprises all African countries that received aid and for which data is available (see Appendix A for a list of included countries). In the ICRG-data (see figure B1 in the Appendix) there is a peak visible around 1995 and 1996. Because it is a peak, it is important that these years should be at the end or at the beginning of a period to make sure this variation is not averaged out in the model. Therefore, for the ICRG-data the following periods are decided upon: 1984-1990 1990-1996 1996-2002 2002-2008. For the Kaufmann-dataset, there is no clear pattern in the graph so it is hard to base the periods on the graph (see figure B2 in the Appendix). The Kaufmann-dataset runs from 1996 to 2012 and in the first couple of years the observations are only once in the two years. Therefore, it is decided to include 7 years in
0 10 20 30 40 50 60 70 80 0 0,5 1 1,5 2 2,5 3 3,5 4 4,5 Aid a s % o f GN I
12
the first period, and 6 years in the next periods to include sufficient variation and information in the first period. This results in the following periods: 1996-2002, 2002-2007, 2007-2012.
4. Instrumentation strategy for non-spatial model
Selecting instruments
Finding appropriate instruments is a daunting, time-consuming and difficult task. Potential instruments have to satisfy certain conditions to be suitable as instrument. A suitable instrument must be relevant. This means that the instrument must correlate with the instrumented variable. Selecting a relevant instrument makes sense since the whole idea is to measure aid by selecting other variables, so these variables should be relevant to aid. Secondly, suitable instruments must be exogenous. Exogeneity of the instrumentation set is needed to arrive at unbiased and efficient estimates. This can be tested by applying the overidentifying restriction test, the so-called Sargan test (Verbeek, 2012). Moreover, the suitable instruments should not have explanatory power in the conventional regression3 with
aid included as variable, they should only serve as an instrument. At last, one can test if aid is really endogenous and aid can be used or if the suitable instruments are needed. This can be done by applying the Hausman-test for endogeneity. To make the process of finding appropriate instruments and the conditions that have to be fulfilled more insightful, figure 4 shows the complete process of selecting appropriate instruments.
Figure 4: Process of finding instruments
The selection of potential instruments is based on previous research on aid. By looking at empirical literature, some potential instruments arise. Dollar and Burnside (2000) use data on arm’s imports, the logarithm of initial income and dummies for specific regions to instrument aid. The reasoning behind it is that arm’s import may reflect strategic reasons to supply aid, just as initial income levels. A second paper instrumenting for aid is the paper of Knack and Bräutigam (2004), they model recipient need by including infant mortality rates and literacy rates as instruments for aid. They also include colonial heritage dummies indicating the colonial history of a country as a factor for providing aid. For instance, Algeria was colonized by France so France may be given aid to Algeria because of a shared past. Djankov, Montalvo and Reynal-Querol (2008) model ´donor strategic interest´ to instrument aid. Instead of looking at the demand for aid, they investigate the supply decision to provide aid. Djankov et al. (2008) look at logarithm of population because if a country has a small population it is easier to gain some control over it and therefore countries are more willing to provide help to these countries. As robustness, the authors include a dummy if a country has had colonial experience. Rajan and Subramanian (2008) also try to model the supply of aid, as opposed to modelling the demand for aid. Rajan and Subramanian (2008) use, among other variables, the
13
logarithm of area and dummies for country of origin to measure the supply of aid. In sum, the above mentioned variables seem to be useful to instrument aid.
However, some variables were not feasible to include in the model because of data limitations and other issues. Data on arm’s import and literacy rates was not sufficiently available for Africa. The dummies for specific regions is not applicable in the research since this paper is focused solely on Africa. Moreover, dummies indicating if a country had colonial experience will not work for Africa either, since almost all countries had colonial experience.
Another issue that I would like to raise is the use of not exogenously given instruments by researchers. An example is the variable mortality rates, which is included as a possible instrument for aid, as used in Knack and Bräutigam (2004). To come up with valid instruments, the instruments need to be relevant to aid but should not have any explanatory power with respect to institutional quality (see box 3, figure 4). Institutional quality is related to mortality rates and should therefore not be included in the instrumentation set. However, there is no literature on the influence of high mortality rates on institutional quality directly and therefore a gap in the academic literature has been identified. Although Acemoglu et al. (2001) use mortality rates to instrument for institutional quality, the relationship between mortality rates and institutional quality in that paper is over time and it is not a direct link because it runs via settler mortality. However, high mortality rates could play an important role in institutional quality. High mortality rates in a country may disrupt society, will costs resources and will lower institutions quality since the government has to provide resources in order to diminish these high mortality rates. Less resources will be left to improve institutional quality. Including mortality rates as an instrument for aid, produces biased estimates. Moreover, the logarithm of initial income, as used as instrument by Dollar and Burnside (2000), may also be endogenous and cannot be used as instrument. The demand for higher institutional quality will rise with income and therefore income is not an exogenous varaible (Rodrik, Subramanian and Trebbi, 2004). Even though initial income is taken and average institutional quality is measured over a period, there is still a relation between these two variables. Initial income may be a predetermined variable but that does not imply exogeneity. Therefore, including initial income as an instrument, will contaminate variables and leads to biased estimates.
To be able to include the colonial experience of African countries, a colonial index is constructed. Colonial relationships may be an important driver behind the decision to send aid so somehow it needs to be included. A dummy for colonies will not work for Africa, as explained earlier. Therefore, a new variable has been constructed which has to do with Africa’s colonial past. From Bertocchi and Cenova (2002) the years for which the African countries became independent were retrieved. A colonial index is constructed which is measured as the current year minus year of independence. This index includes a score, a low score means that this country was not long ago still a colony. It is plausible to assume that the connections between the ruler and the colonial country decrease over time, after independence. A strong relationship between the ruler and the colony, increases the likelihood that the ruler gives aid to the past-colonial country.
14
have to be done to ensure validity of the instruments . Figure 4 shows all the conditions potential instruments have to fulfill to become suitable instruments for aid. In the next section, every test will be described.
It should be noted that by measuring institutional quality via governance indicators in the Kaufmann dataset, the selected variables to instrument aid were not valid (the instrumentation set did not pass the conditions named in figure 4). Moreover, the Kaufmann dataset suffered from many missing observations. Therefore, it is decided to continue with the ICRG-dataset and exclude governance indicators from the Kaufmann dataset as dependent variable. This is because in the case of the ICRG-dataset valid instruments were found, as will be discussed in next section.
Checking validity of the instruments
Every box in figure 4 represents a condition that has to be fulfilled to ensure validity of the instruments. At box 1 the relevance of instruments to aid is tested. This is done by regressing aid on all potential instruments and then apply an F-test indicating all coefficients are zero. It turned out all instruments were relevant to aid since the null-hypothesis can be rejected that all coefficients should be equal to zero at the 1% level of significance (p=0.00, F=10.75). Another piece of evidence which measures the strength of the instrumentation set is the R2
of the first stage. However, according to the literature it is suggested to look at the value of the partial R2 . By looking at the partial R2 the explanatory power from the instrumentation
set is 0.462, which is sufficiently high. The partial R2 measures the correlation between the
instrumented variable and the instrumentation set after partialling out the effect of the strongest instrument. To ensure the fact that the R2 will not be inflated because of a strong
correlation between one instrument and the instrumented variable (aid), Bound, Jaeger and Baker (1995) have promoted using the partial R2. Moreover, as a rule of thumb, the F-test of
the first stage regression should give a value above 10 to ensure strong instruments (Verbeek, 2012). In this case, the F-test gives a value of 10.75 and therefore the instrumentation set is sufficiently strong.4 Strong instruments imply that the selected instruments correlate much
with aid. This is beneficial for the empirical research later on in this paper, since in case of weak instruments, the standard IV-estimator will be biased, standard errors misleading and hypothesis tests are unreliable (Verbeek, 2012).
The second test tests the exogeneity of the instruments (see box 2 in figure 4). By testing the instruments it became clear that by removing the colonial-index variable all instruments were exogenous at the 5% level of significance. This implies that all our instruments are valid.
Thirdly, to come up with valid instruments, the selected instruments should not be relevant to institutions if aid is included. The question is: should the instrumental variables be added as control variables, if aid is already included? This cannot be tested since aid is instrumented with these instrumental variable. Therefore, a different approach is taken and it is tested if the instruments have any explanatory power in the conventional regression, in equation (1a) (the regression with institutional quality as independent variable, see figure 4, box 3). Some colonial heritage dummy variables and the colonial index turned out to be significant in the conventional regression. However, based on the theoretical model, colonial heritage dummies
15
and the colonial-index say nothing about institutional quality and aid is not included in equation (1a), so it is decided to continue with the same set of instruments.
In brief, all three conditions are satisfied by the potential instruments so the instruments are now to be called suitable instruments for aid. The test left to do is box 5, which tests if aid is endogenous and aid should be instrumented or Ordinary Least Squares (OLS) model can be used. According to the Hausman test, which tests the endogeneity of aid, the null-hypothesis cannot be rejected (p=0.99), differences in coefficients are not systematic and this implies that OLS is consistent if we base our analysis solely on test statistics. However, on theoretical grounds and based on previous research, it is clear that aid is endogenous and therefore aid will be instrumented (see box 6a of figure 4).
Furthermore, estimating a panel model requires to make a choice between fixed or random effects. Because the dependent variable shows not much variation within countries, the explanatory power must come from variation between countries. Estimating a country fixed effects model will eliminate all these differences across countries because such a model controls for all time invariant differences between the countries (Kohler and Keuter, 2008). Therefore, the country fixed effect model will often not be estimated by researchers, for instance Brautigam and Knack (2004) do not include fixed effects. However, by excluding fixed effects, one fails to take into account time invariant differences among the observations. If a statistic test points out the inclusion fixed effects is needed and researchers fail to do so, they also do not control for time invariant differences across all independent variables in the regression while important and this may lead to invalid conclusions (Baltagi, 2008). Therefore, both the country fixed effects model and the model without fixed effects will be discussed in this paper, to enhance the reliability of our conclusions. Depending on the test outcome, either the fixed- effects model or the random-effects model will be estimated. The fixed effects model controls for country specific effects in order to led such effects not influence the estimates and ensure reliability.
16
5. Econometric model
The model that will be tested in this research is based on Two-Stages-Least-Squares (2SLS). This is a method to estimate the model in two ‘stages’. The first stage is used to estimate the IV-equation for aid. Secondly, the outcome of the first equation will be used in the second equation as instrumented variable for aid (Verbeek, 2012). Variables have a subscript t, where t runs from 1 to T with steps of seven years and 1 represents 1984 and T represents 2008. Thus, an equation will be estimated with 148 data points (4 periods times 37 countries). For institutional quality, the seven-year average is taken because of the persistent nature of the variable. The effects of improved institutional quality are visible in the longer term and by taking seven-year averages most variation in this variable is included (see section 3). For the other variables, the value of the initial year is taken (e.g. for the first period, the value of 1984 is taken). Subscript i denotes an individual country. To test the hypothesis that aid has a negative impact on institutional quality, the following equations are postulated:
17
6. Results and discussion of non-spatial model
Table 1: Impact of aid in institutional quality in different model specifications. Dependent variable is institutional quality.
(1) OLS (2) IV (3) IV CFE (4) IV TFE (5) IV BOTH Openness -0.003 -0.003 -0.008 -0.003 -0.005 (0.002) (0.002) (0.004)** (0.002) (0.004) Mortality -0.006 -0.006 -0.002 -0.008 -0.011 (0.002)*** (0.002)*** (0.004) (0.002)** (0.005)** Ethnic tension 0.227 0.227 0.133 0.189 0.153 (0.039)*** (0.039)*** (0.079)* (0.040)*** (0.056)***
GDP p. capita -1.22e-06 -1.22e-06 4.81e-07 -7.51e-08 2.92e-07
(1.05e-06) (1.05e-06) (1.06e-06) (8.61e07) (9.44e07)
Aid -0.007 -0.009 0.021 -0.007 0.014 (0.004) (0.005)* (0.047) (0.008) (0.020) Constant 2.284 2.293 2.229 2.371 2.584 (0.000)*** (0.000)*** (0..928)** (0.274)*** (0.509)*** Period 1 0.192 0.335 (0.123) (0.158)** Period 2 0.491 0.467 (0.109)*** (0.159)*** Period 3 0.151 0.187 (0.108) (0.136) Hausman 44.51 NA -16.24 FE/RE (0.00) Observations 138 138 138 138 138 R-squared 0.339 0.337 0.013* 0.389* 0.242*
Notes: Standard errors are in brackets; ***, ** and * denote significance at 1%, 5% and 10% respectively. CFE implies the inclusion of country fixed effects, TFE means time fixed effects and BOTH is time fixed effects and country fixed effects are included.
18
quality of institution. The insignificance of both variables imply that from an empirical point of view, there is no relation between GDP per capita and openness with institutional quality. The aid-variable is not significant but has a negative sign indicating that aid is supplied to countries with low institutional quality.
Because of the endogeneity of aid, the second specification is the instrumental variable approach. Again, the coefficients of ethnic tensions and mortality rates have the expected sign and are strongly significant. In this specification, aid has a negative effect on the quality of institutions and is significant at 10%. This result could point to a negative relationship between aid and institutional quality. Such indicative evidence supports the hypothesis that aid has no advantageous effect on quality of institutions. Again, GDP per capita and openness have the signs but are also not significant.
The third specification controls for country specific effects. In this case, all time-invariant country specific characteristics are controlled for. By doing this, the coefficient for aid loses significance. Moreover, the control variable mortality rates also loses significance. This result is not surprising since many coefficients have to be estimated with a short time-dimension. This decreases precisions, hence significance. In other words, institutional quality is a persistent variable and does not change much over time, including fixed effects will remove all variation across countries and therefore it is hard to come up with significant estimates. The ethnic tension variable is significant and also the coefficient of openness. The significance of the openness variable is counterintuitive since the coefficient is negative indicating that a more open country, will have lower institutional quality. It should be noted that spatial interaction effects are not accounted for yet and if spatial dependence is important all these coefficients are biased. The country fixed effects model is preferred according to the Hausman test, but the R2 did not increase. The reason for this is that Stata does not take into account
the additional explanatory power from the country dummies into the R2 .
The fourth specification includes time fixed effects. For period 1, period 2 and period 3 time dummies are included in the model to control for variations in the business-cycle. The fifth specification controls for both time and country fixed effects. By including time fixed effects the model improves since the period 2 time dummy is significant. Moreover, the R2 is higher.
In both specification the control variables mortality rates and ethnic tensions are significant. By controlling for country fixed effects the aid-variable is positive but not significant. The insignificance implies that from an empirical point of view, no relationship is present between aid levels and institutional quality.
19
7. Modelling spatial dependence
In this section spatial dependence will be modelled. Brueckner (2003) provides theoretical evidence that governments strategically interact and Seldayo et al. (2010) find empirical evidence that the quality of governance is influenced by spatial dependence. This suggests that institutional quality may also be influenced by spatial dependence. To illustrate this may be the case, Figure 5 and 6 below show informal evidence that the quality of institutions is clustered together. Both figures display every country’s institutional quality against the distance to the country with the highest (Morocco) and the lowest (Republic of Congo) institutional quality in 1996. The distances to Morocco and Republic of Congo are measured as the kilometre-converted great circle distances (𝑑𝑖𝑗) on the sphere:
𝑑𝑖𝑗 = arccos[ (sin ∅𝑖sin ∅𝑗) + (cos ∅𝑖cos ∅𝑗cos|𝜕𝛾|) ] (2)
The latitude and longitude for each country is based on the TAGEO-dataset, and is calculated as the average latitude and longitude of all cities with more than 50.000 inhabitants weighted by city size based on its inhabitants. Where ∅𝑖 and ∅𝑗 are the latitude of country i and j
respectively, and |𝜕𝛾| denotes the absolute value of the difference in longitude between i and j. To follow Tobler’s first law of geography, these distances are substituted into a distance-decay function of the form 𝑓(𝑑𝑖𝑗) = (𝑑𝑖𝑗)−1. Equation (2) is based on the Pythagorean
Theorem, while taking into account the fact that the earth is not flat but round. This approach is based on Seldayo et al. (2010).
Figure 5: Quality of institutions and distance to best practice, 1996
20
Figure 6: Quality of institutions and distance to worst practice, 1996
Figures 5 and 6 suggest that good and bad practice of institutional quality is clustered together. A country closer to a best practice country, is more likely to have higher institutional quality. This result also yields conversely for the countries close to the worst-practice country. Therefore, it is useful to allow for spatial dependence of institutional quality. This finding of clustered institutional quality may also explain the core of the problem of Africa: this continent has not succeeded in high-growth rates because low institutional quality is bundled together and it is hard to get out of this ‘trap’ since an individual country is influenced by its neighbouring countries. Both figures document positive spatial dependence in institutional quality across countries. These figures represent informal evidence of spatial dependence, but do not take into account explanatory variables for institutional quality. To formalize our analysis, a regression will be analysed which considers spatial dependence.
Figure 7: The relationship between different spatial models, Halleck and Elhorst (2015)
In the spatial econometrics literature, there are three types of spatial interaction effects to be considered. There could be endogenous interaction effects present, where the dependent variable y in country A depends on the dependent variable in country B and vice versa.
21
Another interaction effect is the exogenous interaction effect where the decision to behave in a particular way depends on independent explanatory variables of other countries. The third type of interaction effect is a spatial error term where omitted variables are correlated. To take into account all three types interaction effects, the general nesting spatial model should be estimated (see figure 7). However, this may lead to overparameterization which means that different type of interaction effects cannot be distinguished (Elhorst, 2014). The solution to this problem, as proposed by LeSage and Pace (2009) is to exclude the spatially autocorrelated error term. This may imply a loss of efficiency, however by ignoring the spatial interaction effects in the dependent and/or independent variables the costs are even higher. Ignoring spatial dependent variables or independent variables while important, would imply omitted variable bias which renders biased estimates (Verbeek, 2012). Especially in this research where the IV-approach is followed, the error term is already contaminated since the explanatory variable is not directly observed but instrumented. Therefore, it is better to ignore the spatial interaction effects among the disturbance term of different units and the spatial Durbin model (SDM) will be estimated. The SDM modelcontains a spatially lagged dependent variable and spatially lagged independent variables. The hypothesis H0 : θ =0 can be used to
test whether the spatial Durbin model can be simplified to the spatial lag model, while the hypothesis H0 : ρ =0 tests if the SDM model can be simplified to the SLX-model (see figure 7).
The spatial Durbin model is specified as:
𝑌𝑡 = 𝜌𝑊𝑌𝑡+ 𝛼1𝑛 + 𝑋𝑡𝛽 + 𝑊𝑋𝑡𝜃 + 𝜀𝑡 (3)
Where the variable WY denotes the endogenous interaction effects, WX the exogenous interaction effects. ρ is called the spatial autoregressive coefficient and θ, just as β, represent a K x 1 matrix with fixed unknown parameters. W is a nonnegative N x N matrix of known constants describing the arrangement of units in space. Its diagonal elements are set to zero by assumption since no country can be viewed as its own neighbour (Elhorst, 2014). In the spatial Durbin model, the IV-approach should be followed since both aid and WY are endogenous. One disadvantage of the IV-approach is the possibility of ending up with a coefficient estimate for ρ outside its parameter space (Elhorst, 2014). This is because ρ is unrestricted and may not fall between the interval (1/rmin, 1)5. More on this in section 9. In the
next section the spatial weights matrix will be discussed in more detail.
To be able to estimate equation (3), a spatial weights matrix, W, should be constructed. The W-matrix is used to describe the spatial arrangements of geographical units in the sample. In this research, the spatial weight matrix describes spatial interaction effects among countries. In the spatial econometrics literature there are several ways to construct a spatial weights matrix. The spatial weights matrix can be based upon shared boundaries, indicating a 1 if a specific country has a shared boundary with another country. Another way to construct a spatial weights matrix is to include q-nearest neighbours in the spatial weights matrix. In that case only the block of q-nearest neighbours can influence a specific country. Others have used inverse-distances matrices, with or without cut-off point to construct the W-matrix (Elhorst, 2014). In this research, the spatial weights matrix is based on an inverse distances matrix without cut off point. The inverse distances matrix is chosen since it reflects at best the relationship across regions based on their distance, compared to p-order binary contiguity matrices which document neighbours and q-nearest matrices because the value in the spatial
5 R
22
weights matrix reflects the degree of spatial dependence based on inverse distances. Moreover, it is chosen for no cut-off point since Africa is a region and all countries are expected to influence each other. In contrary to indicating a 1 if a specific country shares a boundary, this research considers the distances of all countries. Instead of taking the distances between the capitals of all countries, the distance between the countries is measured in a more sophisticated way. The distance between countries is measured as the kilometre-converted great circle distances (𝑑𝑖𝑗) on the sphere, using equation (2) to calculate it.
It is common practice in the spatial econometrics literature, to normalize W such that all elements of each row sum to one. This results in the fact that the weighting operation can be interpreted as averaging values. Another way of normalizing the spatial weights matrix is column normalizing; making all elements of the column sum to unity. In case of row normalization the impact on each unit by all other units is equalized and for column normalization the impact of each unit on all other units is equalized. Row-normalization and column-normalization are not free of criticism since normalization of the elements of the spatial weights matrix will lead to a division of elements by different factors. This may lead to a misspecification problem. In case of inverse-distances matrix this problem is evident, since normalization will lead to a loss of the economic interpretation of the W-matrix (Elhorst, 2014). To solve this issue, Kelejian and Prucha (2010) propose a normalization procedure where each element is divided by its largest characteristic root. In this case, all elements are divided by the same factor, leaving the mutual proportion between the elements unaltered and therefore the W-matrix can be used for economic interpretation. For this reason, the inverse distance spatial weight matrix is normalized by dividing all elements by its largest characteristic root.
Another issue with a spatial weights matrix is that missing observations have a large impact on the data since one missing observation in one country will lead to many missing values in a specific variable. This is why the missing observations were imputed by taking previous years values or by linear interpolation to make sure the W-matrix contains sufficient information.
8. Instrumentation strategy for spatial model
By allowing for spatial dependence and estimating the spatial Durbin model (see equation (3) ) some new variables are created. The WX and WY variables are calculated by multiplying the spatial weights matrix with the dependent and independent variables. The spatial lag of institutional quality (WY-variable) is now included in the model, but this variable is endogenous since it is related to the dependent variable on the left hand side. Treating an endogenous variable as a normal independent variable produces biased estimates and therefore, besides the variable aid, also the spatial lag of institutional quality needs to be instrumented. A welcome feature of including spatial interaction effects in the model, is the availability of new potential instruments. Drukker, Egger and Prucha (2013) point out that WX-variables can be included in the instrumentation set. Moreover, it is also suggested to include W2X-variables into the instrumentation set. This is because, the estimation of the spatial
23
in the SDM model to instrument the WY-variable (Elhorst, 2010). These WX-variables and W2
X-variables have to satisfy the same requirements as other potential instruments (see figure 4). If a WX-variable is relevant to institutional quality, it should be included in the regression as a control variable but if it is not the case, then these WX-variables can be included in the instrumentation set as new potential instruments. Therefore, to arrive at suitable instruments, the tests done in section 4 should be repeated to arrive at suitable instruments
The first step is to ensure the instruments are relevant to aid and the WY-variable, see box 1 in figure 4. The instruments are both relevant to aid and WY-variable (p=0.00 for both) and also the Shea´s partial R2 and Shea´s adjusted partial R2 are quite high (0.85, 0.81 for WY and
0.58, 0.48 for aid) indicating strong instruments.6 In the empirical literature, the adjusted R2
is preferred since the adjusted R2 controls for the fact the more instruments does not
necessarily imply a better fit. The adjusted R2 makes a degrees-of-freedom adjustment to the
R2 (Verbeek, 2012). As can be seen above, both R2 are quite high and a significant
improvement in the R2 of the instruments has been made, as compared to the R2 in the
non-spatial model (0.462) see section 4. This is an indication that the inclusion of WX and W2
X-variables in the instrumentation set was efficient and leads to stronger instruments. Having strong instruments is beneficial because in case of weak instruments the standard IV-estimator is biased, standard errors are misleading and hypothesis tests are unreliable (Verbeek, 2012).
Secondly, the instruments should be exogenous. However, at first the instrumentation set was not exogenously given. A difficult feature of Sargan’s test of overidentifying restrictions is, that it will not identify which instrument is not exogenously given. By systemically removing all instruments one by one, and thereafter removing a combination of two variables it became clear that by excluding the spatial lag of the Italy-dummy, the colonial-index, second order spatial lag of logarithm of population and the second order spatial lag of mortality rates the instrumentation set is exogenous at 5% (p=0.097).7
Thirdly, to come up with valid instruments, the selected instruments should not be relevant to institutions if aid is included. The question is: should the instrumental variables be added as control variables, if aid is already included? This cannot be tested since aid is instrumented with these instrumental variable. Therefore, a different approach is taken and it is tested if the instruments have any explanatory power in the conventional regression, in equation (1a) (the regression with institutional quality as independent variable, see figure 4, box 3). Some colonial heritage dummy variables and the colonial index turned out to be significant in the conventional regression. However, based on the theoretical model, colonial heritage dummies and the colonial-index say nothing about institutional quality and aid is not included in equation (1a), so it is decided to continue with the same set of instruments.
Thirdly, the variables should not be relevant in the usual regression, equation (1a) with aid included. As before, this cannot be tested since aid is instrumented with these instrumental
6 The statistics in this and next sections are arrived by excluding W*Italy, colonial index variable, W2mortality, lw2pop since these variables are not exogenous. Also, the variables which measure openness of neighbouring countries, population size of neighbouring countries, ethnic tensions in neighbouring countries and colonial-index of neighbouring countries are excluded since they are relevant to institutional quality.
24
variables. Therefore, a different approach is taken and it is tested if the instruments have any explanatory power in the conventional regression, in equation (1a) (the regression with institutional quality as independent variable, see figure 4, box 3). It is important to determine which spatially lagged variables are relevant to institutional quality, since these variables should not be considered as potential instruments. By regressing institutional quality on the usual controls, the instrumented aid-variable and one by one adding a WX-variable to the regression, it became clear that the variables which measure openness of neighbouring countries, population size of neighbouring countries, ethnic tensions in neighbouring countries and colonial-index of neighbouring countries are relevant to the quality of institutions and should be included as controls (see box 3 in figure 4).
The instrumentation set that remains covers the: logarithm of population, logarithm of area, colonial heritage dummies and the remaining WX-variables such as colonial heritage dummies of neighbouring countries, GDP per capita of neighbouring countries, mortality rates of neighbouring countries and the remaining second order spatial lagged variables. These variables are used to instrument aid and institutional quality of neighbouring countries (WY-variable).
The last test to do is the Hausman test, which turned out to give a probability of 0.00 so the null-hypothesis that aid and the spatial lag of institutional quality is exogenous, can be rejected and the instrumental variable approach should be followed. In the non-spatial model, it became clear that the inclusion of time fixed effects and/or country fixed effects significantly improved the model. Therefore, also in the spatial model these effects are controlled for.
Now there are two specifications of the spatial model, SDM-model with taking into account country fixed effects (no time dummies) and one including both time fixed effects and country fixed effects. It can be tested whether taking into account spatial dependence is appropriate by testing the hypotheses: ρ=0 and θ=0. There are three cases to be distinguished. If one of the hypothesis is accepted, the spatial Durbin model can be simplified to a spatial lag model or a SLX-model (see figure 7) and allowing for spatial dependence is appropriate. If both hypotheses are rejected, this is evidence that taking into account spatial dependence is correct and the SDM model fits the data best. A third option would be that both hypothesis are accepted, implying that spatial interaction effects are not relevant.
9. Results and discussion by including spatial interaction effects
Table 2: Impact of aid in institutional quality in different model specifications. Dependent variable is institutional quality.
25
(0.040)*** (0.056)*** (0.043)*** (0.054)***
GDP per capita -7.51e-08 2.92e-07 -1.10e-07 9.82e-09
(8.61e07) (9.44e07) (8.88e-07) (9.02e-07)
Aid -0.007 0.014 -0.007 -0.007 (0.008) (0.020) (0.007) (0.014) Constant 2.371 2.584 5.609 3.391 (0.274)*** (0.509)*** (3.736) (75.368) Period 1 0.192 0.335 0.007 0.728 (0.123) (0.158)** (0.372) (3.448) Period 2 0.491 0.467 -0.060 0.831 (0.109)*** (0.159)*** (0.361) (2.197) Period 3 0.151 0.187 -0.112 0.158 (0.108) (0.136) (0.258) (1.121) W*Dep. Var. 0.983 -0.812 (0.475)*** (0.791) W*Population -0.367 0.204 (0.398) (7.722) W*Openness -0.025 -0.044 (0.013)* (0.022) W*Ethnictension -0.216 0.095 (0.236) (0.265) W*colonial 0.010 0.034 (0.020) (0.030) Observations 138 138 138 138 R-squared 0.389* 0.242* 0.402* 0.242* H0 : ρ=0 / Chi2 4.28 1.05 (prob.) (0.039)** (0.305) H0 : θ=0 / Chi2 10.58 5.80 (prob.) (0.032)** (0.215)
Notes: Standard errors are in brackets; ***, ** and * denote significance at 1%, 5% and 10% respectively. CFE implies the inclusion of country fixed effects, TFE means time fixed effects and BOTH is time fixed effects and country fixed effects are included.
In table 2 the estimation results of the spatial extension of the model are recorded. Specification (1) is the non-spatial model following the instrumental variable approach with taking into account time fixed effects and can be compared with specification (3) the spatial Durbin model with time fixed effects. Specification (2) is the non-spatial model with taking into account country specific and time fixed effects. The results of specification (2) can be compared to specification (4), the SDM model with time and country fixed effects.
26
but not significant. This implies that from a statistical point of view no significant relation exist between institutional quality and aid. The coefficient for the spatial lag of institutional quality in specification (3) is quite large and really close to 1. The coefficient is strongly significant and positive indicating that positive spill-overs exists for institutional quality. In other words, the institutional quality of a specific country depends positively on the institutional quality of neighbouring countries. Moreover, most WX-variables are not significantly different from zero but jointly they are. This indicates that the SDM model is the appropriate model, both endogenous interaction effects and exogenous interaction effects exist are present and the model cannot be simplified. The variable which measures the openness of neighbouring countries is weakly significant with a negative coefficient. This indicates that if neighbouring countries have a high level of openness, the institutional quality in a specific country is lower. An explanation could be that because neighbouring countries are more open, they receive more aid compared to a specific country and therefore institutional quality is lower for that specific country because that country receives less aid. However, the hypothesis of this research was that aid negatively influences institutional quality and not positive.
Another, more plausible, explanation has to do with the disadvantage of the SDM model; the possibility of ending up with a coefficient estimate for ρ outside its parameter space (Elhorst, 2014). The interval for which ρ is defined is (1/rmin, 1)8. The coefficient estimate for ρ is close
to one, and documents almost a one to one correspondence between institutional quality of one country and its neighbours. Although the coefficient estimate for ρ falls just in the interval, the large value is unlikely. Because of the large estimated coefficient for ρ, the endogenous interaction effects are overestimated and the exogenous interaction effects could have reversed signs. This could also provide a reason for the fact that almost all WX-variables show counterintuitive results.
For example, it should make sense that the coefficient for the spatial lag of ethnic tensions is positive. If neighbouring countries suffer from ethnic wars, refugees go to the neighbouring country and this may negatively influence the quality of institutions there. Since ethnic tensions is measured as a high score means low ethnic tensions, a positive coefficient is expected and not a negative one. Also for the coefficient of the spatial lag of the colonial-index the sign is unexpected. The spatial lag of the colonial-index might be related to aid received by neighbouring countries. For a higher value of colonial index, less aid is received by neighbouring countries, more aid goes to a specific country so lower institutional quality in that country. A negative sign is expected but a positive coefficient is estimated. Only for the coefficient of the spatial lag of population, the sign is as expected. Higher population in neighbouring countries implies less aid is received in these countries so more aid is supplied to that specific country which harms institutional quality.
The explanation of the reversed signs of the coefficients in front of the WX-variables can be further explained by rewriting the spatial Durbin model in (3) as:
𝑌 = (𝐼 − 𝜌𝑊)−1(𝑋𝛽 + 𝑊𝑋𝜃) + 𝜀 (4)
8 R
