To donate or to invest? : a comprehensive dynamic panel data study on the economic effects of foreign aid and foreign direct investment in developing countries

(1)

.

To Donate or to Invest?

A comprehensive dynamic panel data study on the

economic effects of foreign aid and foreign direct

investment in developing countries.

Abstract – This study investigates the much debated relationship

between foreign aid and gross domestic product per capita in

developing countries, and includes foreign direct investment in the

model. Using a comprehensive dynamic panel data approach and a

completely new dataset, a significant and positive effect of foreign

direct investment on the gross domestic product per capita is found.

Foreign aid is split into four categories, which show to have

significantly different, but not very robust effects. This study

demonstrates that the use of a dynamic model in research on aid

effectiveness leads to improved results, and that different types of aid

and developing countries are heterogeneous, emphasizing the need for

approaches that may overcome this.

Keywords: Foreign Aid, Foreign Direct Investment, Economic

Growth, Aid Effectiveness, Development Economics, Panel Data.

.

Master Thesis written as part of the Master of Science program in

Econometrics at the Amsterdam School of Economics.

Author:

J.J. (Joppe) Arnold (10067639)

Date:

November 24, 2015

First supervisor:

Prof. Dr. J.F. Kiviet

Second supervisor:

Dr. M.J.G. Bun

(2)

(3)

“There's no copy and paste system in foreign aid.”

Atifete Jahjaga, current President of Kosovo.

(4)

Statement of originality

This document is written by Student Johannes Jacob (Joppe) Arnold who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it.

The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

Preface

I would like to thank Professor Jan Kiviet for his engagement, feedback and time, from the first lecture on panel data to the handing in of this thesis. In between we have had many discussions about technical aspects, modelling choices and small details. During this discussions I have learned a lot about econometrics itself, specific techniques, but also how to critically review other work. Next to this, I also had the pleasure to get to know him more personally, with the highlight of visiting him to enjoy the good life Singapore has to offer. He made sure I remained on track and that I delivered a result of which I can be proud.

Although doing empirical research is not always easy and straightforward, I enjoyed the process of gathering data and the editing and writing of this study. Although on the one hand I feel glad to have finished it, on the other hand I will miss applying econometrics on this particular subject. It had lead me to further strengthen my beliefs that econometrics is not just a method to create abstract and complicated investment models, but that econometricians can do (more than) their fair share in helping to analyze and solve some of the real issues this world faces, and help make it a better place throughout.

I would also like to thank Maurice Bun for taking the time out to read this study. Furthermore, I want to thank Yoram Vanmaekelbergh for his support, feedback and help during the writing of this thesis. I would also like to thank my friends and family who have supported me during my time of absence and my housemates for cooking many dinners on intensive days. Finally, I want to thank my parents, who have supported me from my first days to where I am now, at the end of my studies. I believe all of this would not have been possible without the love of God, to who I am grateful for his everlasting support.

(5)

1 Introduction

Economic literature has given attention to the effectiveness of foreign aid to developing countries through numerous studies. Since the sixties the real amount of foreign aid from rich towards poorer countries has vastly increased by more than 50% to over 100 US$ every year. However, research has always questioned whether the economic effect of this wealth redistribution is positive or negative. Investing in economic factors as infrastructure, energy and communication may help to create a better economic climate in a country, but it may also be a message that donors give aid anyway, and thus not motivate countries to create autonomous growth. This makes these countries less competitive, something known as the Dutch disease. Investments in the social sector, for instance in education, health and sanitation, may not have an economic effect in the short term, but through the development of individuals within a population it may have effects on economic growth in the longer term. Various studies have tried to distinguish between the long- and short-term and between economic and social aid, however, the results have been mixed, which is mainly attributed to weak data and poor modelling.

In the redistribution of wealth across countries a severe endogeneity problem is present, according to previous research. Donor countries may favor giving aid towards countries that already perform well to provide a final push towards the second world, possibly creating a severe bias in the estimation of effectiveness. When countries perform better than a certain level, donors may decide to stop helping those countries and start helping poorer countries as the now richer developing countries do not seem to need this aid anymore. These developing countries are now possibly more viable to receive foreign direct investment (FDI), boasting further economic growth and possibly compensating for the reduction in foreign aid towards more developed countries, thus inducing a reverse causality.

This thesis focuses mainly on the relationship between foreign aid (distinguishing different types of aid), FDI and the gross domestic product (GDP) per capita. Using dynamic GMM methods on panel data on developing countries, these relationships are analyzed exploiting a newly assembled dataset. The remainder of this report is organized as follows. In Section 2 an introduction into recent qualitative and quantitative research on foreign aid, FDI and growth is given, including the methods used to analyze these relationships. In Section 3 the research method used in this study will be introduced, clearly describing the models and dataset. Section 4 elaborates on the results that are obtained by estimating the models using the dataset. In Section 5 the conclusions are summarized and Section 6 lists the possible shortcomings and limitations of this study.

(7)

2 Previous research

In this section a theoretical introduction into the effectiveness of aid and FDI is given using a theoretical growth model. Hereafter, recent empirical literature on the relationship between foreign aid, FDI and economic growth is given. Lastly, an attempt is made to replicate and amend two recent empirical studies, namely Rajan and Subramanian (2008) and Minoiu and Reddy (2010), using their original datasets. As the focus in this study is mainly on applying panel data analyses, the panel data findings and methods of earlier research receive most attention in the review of the empirical literature.

2.1 Introduction to foreign aid and foreign direct

investment

Before moving towards theoretical and empirical research on foreign aid, FDI and economic growth, a clear definition of aid and investment is needed. Foreign aid is defined as a transfer of money from a country, official sector organization or a multilateral organization to a developing country, with a granting element of at least 25%. This money must be invested in projects improving development, such as infrastructure, education, food production, loan forgiveness or other humanitarian projects.

A large proportion of foreign aid that is given to developing countries is used for specific projects on a microeconomic scale. Examples include the building of schools, hospitals, water wells, power plants and increasing the fertility of land. Although, according to foreign governments and non-governmental organizations most of these projects are said to be successful, the direct effects of these projects on the GDP of a country are not easily and directly measurable. Theoretical frameworks are often used to provide a possible explanation for how microeconomic changes influence macroeconomic figures as the GDP.

Hansen and Tarp (2000) provide an analytical framework for the macroeconomic effects of foreign aid which is based on the assumption that aid increases domestic investment. This is done by using a model based on Solow (1956), which is a widely applied model in many studies describing economic growth under the assumption that the growth of different countries converges. In this model the growth of income per worker in an economy is determined by a constant and a logarithmic and thus non-linear effect of domestic investment within a country. Under the simplifying assumption that FDI does not exist, the investment variable is modelled using savings and foreign aid as explanatory variables. By putting this investment variable into the growth equation, growth can be modelled using foreign aid and domestic savings. Their proposed model is

𝑔𝑖,𝑡 = β0+ β1log (𝛾1𝑠𝑖,𝑡+ 𝛾2𝑎𝑖,𝑡) − 𝜌log (𝑦i,0), (1)

where gi,t is the growth in income in country i at time t, s is the amount of

savings and a the amount of foreign aid towards country i at time t, while the last term is added in to control for deviations from the steady state of growth which are already present at the beginning of the series. As a Solow growth model

(8)

assumes convergence of income to a steady state, the inclusion of this term allows for the possibility that countries with a lower initial income have higher growth rates, as elaborated on by Mankiw, Romer and Weil (1992). Hansen and Tarp (2000) approximate (1) using a second order Taylor approximation which is

𝑔𝑖,𝑡 = 𝛽̃ + 20 𝛽₁𝛾₁ 𝜏̅ 𝑠𝑖,𝑡+ 2 𝛽₁𝛾₂ 𝜏̅ 𝑎𝑖,𝑡− 1 2 𝛽₁𝛾₁2 𝜏̅2 𝑠𝑖,𝑡 2₋1 2 𝛽₁𝛾₂2 𝜏̅2 𝑎𝑖,𝑡 2₋𝛽1𝛾1𝛾2 𝜏̅2 𝑠𝑖,𝑡𝑎𝑖,𝑡− 𝜌log (𝑦i,0), (2)

where 𝜏̅ is the sample mean, which is used as the point of expansion in the approximation. Since the savings within a country are often not known proxies are used to account for their effects. The proxies are for example the government consumption as in Minoiu and Reddy (2010) or policy variables as used by Rajan and Subramanian (2008).

Although other variants of this growth equation are also used in literature on aid effectiveness by for example Dalgaard, Hansen and Tarp (2004) who use an overlapping generation model to show that foreign aid has an effect on productivity in the long-term, the main theoretical support for the macroeconomic effect of foreign aid on economic growth is based on the assumption that (a proportion of) foreign aid triggers investment. This investment is the base for economic growth under the assumption that the Solow model is a good approximation of economic growth. This model indicates that when analyzing empirical data the chosen specification should allow for possible nonlinear and interaction effects.

It is likely that foreign aid has an effect on macroeconomic indicators not only via investment but also via for example better education which causes a higher level of productivity and competitiveness, improved health circumstances which raise productivity for a worker and improved policies which raise the level of attractiveness for foreign companies to invest in a specific country. Evidence for these effects which are very likely to improve the economic situation is often found, for example by Feyzioglu, Swaroop and Zhu (1998). However, since the growth modelling field for developing countries focuses on the aforementioned investment parameter, most theoretical frameworks do not include the foregoing and other plausible factors. This framework serves more as a reasoning for why foreign aid is expected to have a positive macroeconomic influence, although little support is found that these frameworks exactly describe economic processes.

Foreign direct investment (FDI) is defined as a certain type of money inflow into foreign countries in the private sector. In a FDI transaction, a foreign entity invests money in a domestic business in order to obtain an amount of shares in return, which is set at a minimum of 10%. This 10% threshold is internationally agreed upon as a stake of 10% guarantees a form of control in the domestic business. FDI may take the form of buying a company in another country or by expanding already existing operations in the target country. In the theoretical framework of Hansen and Tarp (2000) the effects of FDI can be included, but this requires many more interaction and quadratic terms, which is beyond the (empirical) scope of this study.

(9)

2.2 Recent empirical studies

A recent influential study on several possible effects of foreign aid on economic growth is that of Rajan and Subramanian (2008). They investigate this relationship using both a cross-sectional and a panel data analysis. Under a single framework of control variables they examine many different possible impacts of aid. They look at the difference between short- and long-term aid, explore data covering different time periods and distinguish multilateral and bilateral aid. In their cross-sectional analysis they use a new type of instrument to control for the endogeneity problem, being the fitted received aid divided by initial GDP. The reasoning behind this is that the decision from a donor to give money to a recipient is often made on the bases of non-economic reasons, as stated by Alesina and Dollar (2000), and thus not correlated with actual economic growth. After estimating a model to find the expected ratio of aid to GDP in a specific period of years using historic factors as colonization and language and a number of variables that are related to the size of the country, they use this instrument to estimate several models. These models range from investigating both short- and long-term effects of aid to models distinguishing different types of aid.

Panel regressions are also run using both the Arellano-Bond and Blundell-Bond estimation techniques. These procedures to deal with panel data are further explained in Subsection 3.2. In these procedures the exogenous and endogenous variables are kept the same throughout different specifications and the number of internal lags for instrumentation is also kept constant. This implicates that in this research only one specification is examined, with just some slight changes in the interaction variable. The choices that are made concerning the number of lags and the treating of variables as endogenous or (weakly) exogenous are poorly motivated and relevant test statistics are missing, making the presented results hard to interpret and leaving room for improvement. Throughout, the estimated effect of aid on economic growth is found to be negative, albeit mostly insignificant. This conclusion holds for different geographic regions, differences in quality of economic policies of countries, different time horizons and different types of aid. This negative effect is known as the micro-macro paradox, first introduced by Mosley (1987), where local projects funded by foreign aid are successful but the effects are not measured on a countrywide scale. Despite the seemingly robust finding, later studies question the validity of the used instrument. An example is the research done by Arndt, Jones and Tarp (2010), which questions the outcomes and tries to improve by using a better instrumentation strategy. Although no panel analysis is carried out, the outcomes of the latter paper indicate a significantly positive effect of aid on economic growth, contrary to the result of Rajan and Subramanian (2008).

Rajan and Subramanian (2011) look deeper into these negative effects and find that they may be caused by the fact that a large foreign inflow of money leads to negative exchange rate effects, weakening both the export and labor intensive sectors. This indicates that countries may not be able to absorb large inflows of aid that are donated upfront. Although not literally a case of the Dutch disease, these donations of aid do lead to decreased levels of competitiveness in

(10)

developing countries, leading to a lower level of economic growth. They argue that it is better to first create a basic economic infrastructure and have skilled workers before donating aid, after which the amount of aid disbursements can slowly be increased.

Minoui and Reddy (2010) put further efforts in distinguishing the effects of social and economic aid and allow for the effect of growth to occur after a longer period of time. In their analysis they also specifically compare the differences in effects between groups of donors. This gives an insight into whether aid from a specific group of donors is more effective than that of other groups. A panel analysis is included to control for country-specific effects. Although more attention is given to the endogeneity and exogeneity of (lagged) variables and to some test statistics that justify their choices, there are still relevant statistics missing. In their specification they include aid lagged for a period of 5, 15 or 25 years and split it into three different groups of donors. It is found that developmental aid has a significant positive effect on growth decades later, albeit smaller for some groups of donors, as is also found using cross-sectional analysis. This leads to the main conclusion that certain types of aid are effective and do promote growth in the long run, as opposed to research stating that all foreign aid has a negative effect on growth. For the variable that is lagged 25 years in the panel analysis it is noted that the sample size shrinks in such a way that the validity and relevance might become questionable, as Bond, Hoeffler and Temple (2001) show.

Harms and Lutz (2006) use panel data on a large number of developing and emerging countries in the last decennium of the 20th_{century to investigate the}

effect of foreign aid on FDI. Using total foreign aid as explanatory variable and several measures of FDI as dependent variable they find that foreign aid has no significant effect on this type of investment. However, when focusing on countries with a large regulatory burden they find that the effect of foreign aid on FDI is significantly positive.

Selaya and Sunesen (2012) investigate the effect of foreign aid on FDI using data on 99 countries over 30 years. Additional to earlier research such as that of Harms and Lutz (2006) they split aid in two sectors, namely complementary inputs, which is aid towards the social and economic sector, and physical capital, which is aid given towards production sectors. This is done since aid invested in the social and economic sector is supposed to benefit countries more via spillover effects, while aid directly invested in the production sector is aid that could typically also be replaced by (foreign direct) investment. The main finding is that foreign aid does have a positive effect on FDI overall. However, there are strong differences between the complementary inputs aid, which has a significantly positive effect, and physical capital aid, which has a significantly negative effect on FDI. This supports the idea that when investigating the effects of aid given towards a developing country not only the amount of aid matters, but that the purpose of this aid is also important.

The relationship between FDI and economic growth is less clear. Even though investments in general should lead to higher economic output in the theoretical

(11)

model of Hansen and Tarp (2000), empirical research has shown that this intuition only holds under certain conditions. Borensztein, De Gregorio and Lee (1998) find that FDI is more productive than domestic investment. Since well-established firms in a particular developing country have more knowhow on that specific country, foreign investors must compensate for that by higher productivity via better management and technological advantages. This seems to especially be the case for FDI in developing countries, which are also the countries that are more likely to receive foreign aid. They also find that the effects of FDI on economic growth depend on the skills of the labor force within the receiving country, while this is not the case with domestic investment. Their results suggest that the main effect of FDI on growth is not via a higher capital inflow but via increased efficiency and productivity.

Carkovic and Levine (2002) find no evidence for a significant effect of FDI on economic growth. Using a panel dataset concerning 68 countries for 7 periods of 5 years, they are the first to jointly control for endogeneity, country-specific effects and dynamics by including lagged growth variables. Although there is a correlation between FDI and economic growth, this correlation is assumed to be caused by the fact that policies that increase FDI also increase growth itself, as no statistical evidence of an influence of FDI on GDP is found.

2.3 Replication of research on the effects of aid on growth

This subsection focuses on the panel data studies reported in Rajan and Subramanian (2008), in this subsection referred to as R&S, and in Minoiu and Reddy (2010), in this subsection referred to as M&R. Employing the very same datasets their results are replicated and thereafter their approaches are critically evaluated, followed by suggestions for possible improvements. Finally the effects of employing a more general dynamic specification then has been used before are examined. From this the conclusion is drawn that satisfactory results cannot be obtained from the datasets that are used by R&S and M&R.

2.3.1 Replicating Rajan and Subramanian (2008)

R&S do not distinguish different types of foreign aid, which results in the aid variable being the sum of aid to all different sectors. The control variables included are initial income per capita, life expectancy, institutional quality, quality of policy, inflation, M2 supply, budget balance, number of revolutions, a possible interaction variable, ethnic fractions, time dummies and three geography variables. All variables are operationalized using their averages over 5 years, while the growth variable is operationalized as the percentage of real economic growth. A complete description of the variables that are used can be found in their Appendix A. All variables, excluding those regarding geography, time and ethnic fraction, are assumed to be endogenous, and thus need instrumentation. For this, all available lags are used without collapsing. The first differences of the assumed exogenous variables are included to serve as standard instruments.

(12)

The specifications are estimated using the Arellano-Bond and Blundell-Bond two-step estimators.

Using the exact code and dataset which were generously made available by R&S the results of Table 1 are obtained, which corresponds to Table 9 in R&S. The results are identical to the results in the original study, except for the p-value of the Hansen test, which is invariably 1.000 according to the calculations in this replication. The estimated model is

𝑔𝑟𝑜𝑤𝑡ℎ𝑖,𝑡 = 𝛽0𝑎𝑖𝑑𝑖,𝑡/𝐺𝐷𝑃𝑖,𝑡+ 𝛿0′𝑥𝑖,𝑡+ 𝜇𝑖+ 𝜋𝑡+ 𝜀𝑖,𝑡, (3)

where the four examined specifications only differ regarding one particular interaction term and where

𝑔𝑟𝑜𝑤𝑡ℎ𝑖,𝑡 = 1 5∑ 𝐺𝐷𝑃/𝑐𝑎𝑝𝑖𝑡𝑎𝑖,𝑗−𝐺𝐷𝑃/𝑐𝑎𝑝𝑖𝑡𝑎𝑖,𝑗−1 𝐺𝐷𝑃/𝑐𝑎𝑝𝑖𝑡𝑎_{𝑖,𝑗−1} 𝑡∗5 𝑗=(𝑡−1)∗5+1 . (4)

Table 1. Replication of main panel results of Rajan & Subramanian (2008) using the Arellano-Bond two-step (AB2) estimator.

specification1 _[1] _[2] _[3] _[4] coefficient aid/GDP -0.151 (0.0770)**2 -0.0145 (0.207) -0.168 (0.140) 0.163 (0.140) coefficient interaction term - -0.00514 (0.00525) -0.0216 (0.0496) 0.376 (0.113)*** interaction term included none aid/GDP * aid/GDP aid/GDP * policy aid/GDP * geography observations 167 167 167 167 instruments 120 135 134 135 p-value AR(1) 0.304 0.181 0.194 0.167 p-value AR(2) 0.198 0.269 0.255 0.199 p-value Hansen 1.000 1.000 1.000 1.000

Note that Table 1 reports that the number of instruments is very large compared to the number of observations in all four specifications. Although no official rule of thumb exists, a 0.75 instruments-observations ratio is too big, as for example Ruud (2000) states. With a more acceptable number of instruments compared to observations the Hansen test will better be able to provide information on the presumed validity of the overidentification restrictions, on which more information can be found in Roodman (2009). Too many instruments may result in a bias in the coefficients according to Bond, Hoeffler and Temple (2001). Therefore the model will be re-estimated using a collapsed set of instruments.

1_{Different variations of estimated models are numbered within a (sub)subsection}

using square brackets, starting with [1] in each (sub)subsection. If an estimated specification of another (sub)subsection is referred to, this is explicitly noted.

2_{In this study all standard errors are robust regarding heteroscedasticity, and in}

the case of two-step GMM, Windmeijer corrected, and shown in parentheses. Additionally, ***, ** and * indicate significance at respectively the 1%, 5% and 10% level.

(13)

Furthermore, first order autocorrelation should be present in first differences by construction, however, the test for AR(1) does not provide statistical evidence of this autocorrelation. This may indicate inconsistency and serious bias in the estimator and its standard errors, implying that the negative value of the effect of aid on growth and its suggested significance in specification [1] are wrongful.

Moreover, although the first specification suggests a significantly negative coefficient of aid3_{, the inclusion of the interaction term between aid and}

geography, as R&S do in specification [4], invalidates this result. If a variable that is added turns out to be significant, this indicates that the original model is incorrectly specified. R&S examine some interaction variables as shown in Table 1, however, other interactions, lagged or squared variables have not been considered. Adding some of these extra variables may invalidate their models and thus their conclusions. As Elbadawi, Ndulu and Ndung’u (1997) find a significant effect for the lagged squared budget deficit on economic growth, the present and lagged squared government budget balance will be added into (3). When an interaction variable is included in the model, that variable will also be exploited for internal instrumentation.

Table 2. Extension of main panel results of Rajan & Subramanian (2008) using the AB2 estimator,collapsing instrumentsandadding variables.

specification [4] [5] [6] [7] aid/GDP 0.163 (0.140) 0.200 (0.118)* 0.0722 (0.141) 0.137 (0.127) aid/GDP * geography 0.376 (0.113)*** 0.598 (0.162)*** 0.276 (0.143)* 0.325 (0.144)** squared budget balance - - 0.0110 (0.00480)** 0.00778 (0.00883) squared lagged budget

balance - - 0.00827 (0.00216)*** 0.00754 (0.00678) squared budget

balance included no no yes yes

observations 167 167 162 162

instruments collapsed no yes no yes

instruments4 ₁₃₅ ₅₈ ₁₄₀ ₃₅

p-value AR(1) 0.167 0.226 0.205 0.124 p-value AR(2) 0.199 0.391 0.100 0.111 p-value Hansen 1.000 0.342 1.000 0.856

Table 2 shows the results of modifying specification [4] of R&S. As specification [4] of R&S shows that the effect of the interaction term between aid and geography is significant this specification is used as the baseline model.

3_{Assuming the assumptions needed to apply Arellano-Bond are satisfied, even}

though the AR(1) test is not satisfactory.

4_{All available internal lags that are assumed to be valid (lags two and higher)}

are used as instruments in specifications [4], [5] and [6]. Specification [7] includes only lags two to four as instruments.

(14)

Firstly the set of instruments is collapsed in [5], secondly the present and lagged squared budget balance are included in [6]. Specification [7] combines the two aforementioned changes and further reduces the number of instruments by using only lags two until four of the endogenous variables as internal instruments, instead of using all available lags.

It can be seen that neither including extra variables nor lowering the number of instruments has resulted in statistical evidence for first-order autocorrelation in the first differences. Reducing the number of instruments does result in more realistic outcomes for the Hansen test however, indicating that the Hansen test is indeed weakened by the large number of instruments in the original study by R&S. In addition to this effect, the reduction of the number of instruments also leads to more positive effects of aid for specification [5].

Furthermore, it can be seen by comparing the results of specifications [4] and [6], that adding the squared and lagged budget balances to the original model leads to significant results for these variables, implying that there is much room for improvement in the model of R&S.

As the results of the test for AR(1) are still unsatisfactory a further modified model is estimated, excluding the contemporaneous initial income per capita, and including the lagged initial income per capita as explanatory variable. This is written as

𝑔𝑟𝑜𝑤𝑡ℎ𝑖,𝑡 = 𝛽0𝑎𝑖𝑑𝑖,𝑡/𝐺𝐷𝑃𝑖,𝑡+ 𝛿0′𝑥𝑖,𝑡+ 𝛿1𝑥𝑖,𝑡−1+ 𝜇𝑖+ 𝜋𝑡+ 𝜀𝑖,𝑡, (5)

where xi,t no longer includes initial income per capita and xi,t-1 includes only initial

income per capita of that period. As the contemporaneous income is part of the economic growth variable by construction5_{the inclusion of this term may lead to}

invalid results. Table 3 shows the results of specification [4] of R&S, specification [7] of Table 2 and the results of the alteration of both these specifications by adding lagged initial income and removing contemporaneous initial income. Table 2 shows that these alterations slightly improve the outcomes of the different test statistics, however, the results are still unsatisfactory. In [8] the number of instruments is too high, while in [9] the test statistics improve only slightly, still indicating a possible misspecification.

A reason for the low number of observations compared to the number of instruments is that for some of the included variables there is no data for several countries, resulting in a drop in the number of cross-sectional observations. While data on aid is available for 1085 cross-sectional observations, the inclusion of all control variables results in only 239 available observations.

In addition to the foregoing alterations to the specification of R&S it can be examined that the aggregation of aid that is used leads to possibly incorrect estimates. Splitting aid between economic and non-economic aid results in significant estimates for both coefficients, of which the results are available upon request from the author of this study. Although the test statistics are still not satisfactory, this may further indicate that the specification of R&S is incorrect.

5_{The growth in the first year of a five year period is the first term of (4),}

(GDP(t-1)*5+1-GDP(t-1)*5)/GDP(t-1)*5 and GDP(t-1)*5 is a transformation of the log of per capita income, namely exp(income(t-1)*5)*population(t-1)*5.

(15)

Table 3. Extension of main panel results of Rajan & Subramanian (2008) using the AB2 estimator and altering initial income.

specification [4] [8] [7] [9] aid/GDP 0.163 (0.140) 0.0664 (0.0957) 0.137 (0.127) 0.0895 (0.163) aid/GDP * geography 0.376 (0.113)*** 0.332 (0.0996)*** 0.325 (0.144)** 0.263 (0.209) squared budget balance - - 0.00778 (0.00883) 0.00556 (0.0144) squared lagged budget

balance - - 0.00754 (0.00678) 0.0130 (0.00870) squared budget

balances included no no yes yes

instruments collapsed no no yes yes

instruments6 ₁₃₅ ₁₃₂ ₃₅ ₃₅

p-value AR(1) 0.167 0.003 0.124 0.077 p-value AR(2) 0.199 0.711 0.111 0.147 p-value Hansen 1.000 1.000 0.856 0.709

2.3.2 Replicating Minoiu and Reddy (2010)

The results of M&R are harder to reproduce as even though their dataset is freely available, the exact specifications of the models that they estimate using Stata are not that obvious. Although the specification is clearly described in their analysis section, the author of this study fails to obtain exact replications. The model that is estimated is

𝑔𝑟𝑜𝑤𝑡ℎ𝑖,𝑡 = 𝛽1′𝑎𝑖𝑑/𝐺𝐷𝑃𝑖,𝑡−1+ 𝛿0′𝑥𝑖,𝑡+ 𝜇𝑖+ 𝜋𝑡+ 𝜀𝑖,𝑡, (6)

where aid is a vector of three by one including bilateral aid from a specific group of donors, bilateral aid from all other countries and multilateral aid donated to the recipient, and growth is defined as in (4).

In the original article the explanatory variables institutional quality and revolutions are assumed contemporaneously uncorrelated with growth and thus all lagged realizations of these variables may be used as internal instruments. The time and geography variables are strictly exogenous and their differences are therefore included as standard instruments. The other controlling variables, namely initial income per capita, inflation, policy and government consumption/GDP and the different forms of aid7_{are assumed endogenous,}

implicating that all lags later than and including t - 2 may be used as internal instruments. All possible instruments are used in the estimation, which is done using two-step Blundell-Bond, thus including level equations. The results of the

6_{All available internal lags that are assumed to be valid (lags 2 and higher) are}

used as instruments in specifications [4], and [8]. Specifications [7] and [9] include only lags 2 to 4 as instruments.

7_{Contemporaneous aid is not present in the functional form (only one-period}

lagged aid which is not endogenous) but aid itself is assumed to be an endogenous variable.

(16)

replication can be found in the Table 4. This table shows the results for replicating the first three columns of Table 6 by M&R.

In this analysis aid is grouped by donor, where G1 consists of main aid donors, and the G2 and G3 groups comprises G1 plus an additional set of donor countries. Multilateral aid is aid that is not directly allocated from one country to another country and is mostly donated by charities and NGOs. Note that in the replication aid has been multiplied by 100, in order to be comparable with the results of R&S.

Table 4. Replication of main panel results of Minoiu & Reddy (2010) using the Blundell-Bond two-step (BB2) estimator.

specification [1] [2] [3] 5-year lagged aid/GDP 0.273 (0.102)*** 0.225 (0.0841)*** 0.241 (0.0960)** 5-year lagged aid/GDP of other countries 0.0203 (0.0499) 0.0337 (0.0564) 0.0299 (0.0562) 5-year lagged multilateral aid/GDP -0.102 (0.0877) -0.105 (0.0962) -0.0911 (0.0900) group of donor countries included in above aid variable8

G1 G2 G3 observations 468 468 468 instruments 235 235 235 p-value AR(1) 0.000 0.000 0.000 p-value AR(2) 0.395 0.419 0.378 p-value Hansen 1.000 1.000 1.000

Table 4 shows that the different AR statistics are satisfactory, where first order autocorrelation is present and no evidence for second order autocorrelation is found. This result seems confusing, as the economic growth and income variables are taken from the dataset of R&S. Even when taking contemporaneous aid instead of lagged aid and using the Arellano-Bond estimator, as in R&S, these statistics remain satisfactory. This difference is likely to be caused by the differences in specification. M&R add and remove some explanatory variables, resulting in more available observations and an altered model.

Similar to R&S, the number of instruments is again very high compared to the number of observations and the Hansen test is again 1.000 throughout all three specifications, indicating the possible weakness of this test. The high number of instruments may lead to a bias in the coefficients and standard errors, which implies that the significance should be taken with a grain of salt. Also, the validity of the assumption that is made in order to include level equations cannot be tested with a weakened Hansen test. Therefore the model is re-estimated using a reduced set of instruments to obtain more realistic outcomes for this test.

8_{The included groups are the same as in the original study of M&R. The exact}

(17)

Furthermore, following the approach of the replication of R&S, a variable on squared government consumption is added in order to find out whether the results may be affected by an omitted variable bias. Table 5 reports the results of the various changes to the specification, where the baseline model is specification [1] of Table 4.

Table 5. Extension of main panel results of Minoiu & Reddy (2010) using the BB2 estimator, collapsing instruments and adding variables.

specification [1] [4] [5] [6] 5-year lagged aid/GDP of G1 countries 0.273 (0.102)*** 0.280 (0.135)** 0.143 (0.0888) 0.216 (0.129)* 5-year lagged aid/GDP of non-G1 countries 0.0203 (0.0499) -0.0412 (0.0498) 0.0444 (0.0495) 0.00105 (0.0433) 5-year lagged multilateral aid/GDP -0.102 (0.0877) -0.152 (0.119) -0.0499 (0.0804) -0.0865 (0.0103) squared government consumption - - 0.00343 (0.00150)** 0.00544 (0.00441) squared government consumption lagged - - -0.00168 (0.00055)*** -0.000582 (0.00136) squared government

consumption included no no yes yes

instruments collapsed no yes no yes

instruments 235 73 262 80

p-value AR(1) 0.000 0.000 0.000 0.000 p-value AR(2) 0.395 0.294 0.500 0.640 p-value Hansen 1.000 0.639 1.000 0.363 p-value Hansen test

BB9 1.000 0.286 1.000 0.531

Table 5 shows that collapsing the instrument sets leads to less significant results, however, the results indicate more or less the same effects of different types of aid on economic growth. The outcomes of the tests for autocorrelation remain satisfactory and the Hansen test demonstrates more realistic values. Furthermore, the test for the validity of the assumption imposed for Blundell-Bond provides no evidence to reject this estimation technique. These findings are robust to further reducing the number of instruments by limiting the number of lags used for internal instrumentation. Specification [5] shows that including the (lagged) squared government consumption results in significant outcomes, indicating that this variable should be included in (6), although this variable becomes insignificant when reducing the number of instruments as in [6]. Including the (lagged) squared government consumption also results in a decline in the estimated effect of aid and its significance. Results which are available

9_{The incremental test for possible invalidity of the assumption imposed in order}

(18)

upon request show that when limiting the number of lags used as instruments the lagged squared government consumption is still significant. The significance of this variable combined with the change in the effect of aid when including government consumption shows that the specification of M&R is open to improvement.

From Table 5, it can be concluded that one-period lagged aid from countries in the G1 group has a significantly positive effect on economic growth, and that this finding is somewhat robust to reducing the set of instruments and adding extra variables. However, M&R only include aid lagged either 5, 15 or 25 years, neglecting more recent aid when estimating the effects. The idea that aid donated 15 years ago has an effect while aid donated in the current year or 5 years ago does not have an effect is intuitively unsound. Neglecting more recent aid as M&R do throughout their specifications may only be done if aid exclusively has an effect after 5, 15 or 25 years Therefore specifications [4] and [6] are re-estimated by altering (6) to also include contemporaneous aid. The results of this extension are shown in Table 6.

Table 6. Extension of main panel results of Minoiu & Reddy (2010) using the BB2 estimator and including contemporaneous aid.

specification [4] [7] [6] [8] current aid/GDP of G1 countries - 0.232 (0.506) - 0.452 (0.494) 5-year lagged aid/GDP of G1 countries 0.280 (0.135)** 0.202 (0.395) 0.216 (0.129)* 0.0151 (0.358) squared government consumption - - 0.00544 (0.00441) 0.00173 (0.00636) squared government consumption lagged - - -0.000582 (0.00136) -0.000215 (0.00137) contemporaneous aid

included no yes no yes

squared government

consumption included no no yes yes

instruments collapsed yes yes yes yes

instruments 73 73 80 80

p-value AR(1) 0.000 0.000 0.000 0.000 p-value AR(2) 0.294 0.347 0.640 0.471 p-value Hansen 0.639 0.571 0.363 0.386 p-value Hansen test

BB 0.286 0.272 0.531 0.713

Table 6 shows that including the contemporaneous value of aid renders all results of the effect of aid on economic growth insignificant. This indicates that even though aid may have an effect on economic growth in the longer term, there is no empirical support for the conclusion that this effect is significant, neither is it found that aid has a significant contemporaneous effect on economic growth.

(19)

Unreported results show that only including contemporaneous aid while leaving lagged aid out does not result in significant effects for contemporaneous aid as well. Table 6 shows that the sample is inadequate, as it exhibits too much multicollinearity to provide empirical evidence on the dynamic effects of aid.

Concluding, scrutiny of the approach of M&R shows that the significantly positive effect of aid on economic growth, as reported in Table 4, which replicated their results, is not very robust to alterations of their research design, although the coefficient signs remain more or less similar. Reducing the number of instruments, adding extra variables and including more recent periods of aid leads to a drop in significance levels and shows that the main conclusion drawn by M&R, namely that aid has a significant positive effect in the long run, is not prone to these small alterations. Furthermore, an exact elaboration on the dataset and description of the different variables is not available10_{, making it}

hard to interpret the results. M&R split aid by different groups of donor countries (G1, G2 and G3) and assume that differences in the estimates are caused by how these groups allocate their aid to different sectors. Although this may be an improvement over the aggregating of aid done by R&S, these differences can be incorporated into a model more directly by grouping donations by sectors instead of donor countries. This latter method can lead to much better insight in how money meant for developmental aid can be put to effective use throughout different donor countries.

2.3.3 Estimation of a dynamic model

As shown in the previous subsubsections the models in both R&S and M&R are affected by an omitted variable bias and their claimed significance is not robust to reducing the number of instruments. Furthermore, the estimated model of R&S is static, while the model of M&R is of the simply distributed lag type. This choice implies that lagged economic growth does not affect current economic growth, and that variables only affect economic growth for a finite spell of time. In this subsubsection the dataset of M&R, which is the more comprehensive of the two aforementioned studies, is exploited to examine the possibility of an autoregressive dynamic relationship by taking lagged economic growth into account. Although both M&R and R&S include some kind of unclear dynamic relationship by including the initial income per capita, the real dynamic effects are not clear from their models.

The model that is being estimated in this subsubsection is 𝑔𝑟𝑜𝑤𝑡ℎ𝑖,𝑡 = 𝛼1𝑔𝑟𝑜𝑤𝑡ℎ𝑖,𝑡−1+ 𝛽0′𝑎𝑖𝑑/𝐺𝐷𝑃𝑖,𝑡+ 𝛽1′𝑎𝑖𝑑/𝐺𝐷𝑃𝑖,𝑡−1+

𝛿0′𝑥𝑖,𝑡+ 𝜇𝑖+ 𝜋𝑡+ 𝜀𝑖,𝑡, (7)

where the aid variable is operationalized as either aid split into economic and non-economic aid, an addition to R&S, or as aid split into specific donor groups as done by M&R. The set of control variables consists of inflation, number of revolutions, (squared) government consumption, initial quality of policy and

10_{The author of this study has contacted the authors but received the}

(20)

interactions between government consumption and total aid and between revolutions and total aid. Furthermore, the lagged realization of growth is included. Except for the exogenous time dummies and the revolutions variable, which is considered to be weakly exogenous, all variables are considered endogenous as in M&R. Variables that are constant over time but differ per country are removed from the model, as these are differenced away when applying Arellano-Bond. These variables may still be used as instruments, but unreported results show that this only leads to marginal changes in the effects while leading to a higher number of instruments.

As in the preceding subsubsection aid and government consumption are expressed in ratios of GDP. Four different cases regarding the model in (7) are examined: distinguishing between aid split into economic and non-economic aid in [1] and split by donor groups in [2] without collapsing the set of instruments, and with collapsing the set of instruments, respectively [3] and [4]. Table 7 presents their Arellano-Bond two-step estimates.

Table 7. Combination and extension of Rajan & Subramanian (2008) and Minoiu & Reddy (2010), estimated using AB2.

specification [1] [2] [3] [4] economic aid/GDP 0.259 (0.334) - 1.16 (0.661)* - non-economic aid/GDP -0.382 (0.182)** - -0.351 (0.271) - G1 aid/GDP - 0.186 (0.268) - 0.473 (0.356) multilateral aid/GDP - 0.0191 (0.152) - 0.190 (0.116) lagged economic / G1 aid/GDP11 0.153 (0.226) -0.140 (0.208) -0.166 (0.323) -0.406 (0.285) lagged non-economic / multilateral aid/GDP12 0.0265 (0.0551) -0.0243 (0.119) 0.0887 (0.0888) 0.0865 (0.108) lagged economic growth -0.310 (0.0872)*** -0.0769 (0.0698) -0.382 (0.0787)*** -0.170 (0.0612)*** observations 274 385 274 385 instruments 180 256 67 79 p-value AR(1) 0.034 0.001 0.017 0.004 p-value AR(2) 0.309 0.100 0.397 0.062 p-value Hansen 1.000 1.000 0.270 0.447

The outcomes of the tests for autocorrelation are satisfactory if aid is split by purpose (economic and non-economic), as in [1] and [3]. When aid is split by donor country, as in [2] and [4], the test for AR(2) indicates second-order autocorrelation at the 10% level, implying a misspecification. As expected, the

11_{In specifications [1] and [3] this is lagged economic aid, in [2] and [4] this is}

lagged aid donated by G1 countries.

12_{In specifications [1] and [3] this is lagged non-economic aid, in [2] and [4] this}

(21)

Hansen test appears to be weakened with the large instrument sets, resulting in a p-value of 1.000. The results for the collapsed instrument sets show more realistic outcomes, a result similar to that throughout this subsection.

It can be seen in Table 7 that the composition of aid has a clear influence on the outcomes. Using the formula (𝛽0+ 𝛽)/(1 − 𝛼1), which is further introduced

in Subsubsection 3.2.4, a long-term positive effect of economic aid and negative effect of non-economic aid is found. Since the amount of non-economic aid donated is vastly more than the amount of economic aid, combining economic and non-economic aid results in a total negative effect of aid on economic growth. When grouped by donor country the effects are mostly positive, except for the multilateral aid in [2], which is very close to 0. Combining aid donated by G1 countries and other types of aid results in a positive total effect of aid in this operationalization of the aid variable, despite the fact that the amount of multilateral aid donated is vastly higher. The author of this study has tried to find out the cause of this difference, however, the exact definitions and raw data for the different replicated studies are not provided. Therefore the author was not able to thoroughly look into these differences and the differences between the definitions of various components of aid.

Furthermore, the lagged economic growth variable differs significantly for the different types of aid, but is highly significant in [1], [3] and [4]. Again, the exact cause is not known, but it is suspected to be related to collinearity between initial income and lagged economic growth. As the authors use aid relative to GDP and define growth as in (4) this leads to severe collinearity between the different included variables when also including lagged growth and initial per capita income. A possible method to deal with this collinearity may be to take absolute instead of relative GDP in order to formulate a model that is clearer. However, this cannot be calculated accurately as the included data on initial income is inaccurate in comparison to available data by the World Bank (2014).

In the foregoing replications some limitations of the dataset have already been mentioned. An important one is that the exact definitions of the variables are not always available, especially for the different operationalizations of the aid variable, leading to confusing and contrary results. Furthermore, the exact sectoral disbursements of aid are not available, although the results indicate that the sector to which aid is disbursed is very important. This is also shown by Clemens, Radelet and Bhavnani (2004) who find that the effects of different types of aid differ by up to 200% compared to taking aggregate aid. Lastly, taking the growth rate as in (4) and discarding the dynamic structure of GDP limits flexibility in the estimation process. Taking the absolute instead of the relative GDP offers more flexibility, as one can always decide to transform the variables as in R&S and M&R. To overcome the aforementioned issues a new dataset is put together in Section 3.

(22)

3 Research method

In this section the research method is elaborated on. In the first subsection the dataset, sources of the used data and the different operationalizations are described. In the second subsection the different methods and specifications that are used to analyze this data are explained in detail.

3.1 Data

Although datasets used in previous studies on the influence of aid and investment on growth are publicly available a new dataset is being used in this study. This is done to extend the number of observations available and to improve the accuracy and flexibility of the data. The dataset that is used consists of several key elements, namely data on foreign aid, FDI, GDP and the population. Next to these main variables several control variables are included to control for political and economic factors. The first two subsubsections introduce the main variables and control variables, after which the method used to transform the data is introduced. This is followed by a subsubsection on gaps in the data and how these are dealt with. Finally a subsubsection is dedicated to different subsets of the data that may be worth looking into.

3.1.1 Main variables

The aid data for this study is constructed from aiddata.org as maintained by Tierney, Nielson, Hawkins, Roberts, Findley, Powers, Hicks et al. (2011). This dataset consists of aggregate disbursement flows and commitment flows per sector. The dataset includes bilateral and multilateral aid but does not include aid given through NGOs. As sectoral data on aid disbursements is only available for years later than 1992 a more widely used approach (see for example Nunnenkamp & Thiele (2006)) is carried out to obtain sectoral data on disbursements, the actual money flows, using data on commitments.

Data on commitments is tracked by donor governments, which are usually more developed countries. This data is available for a longer time period, and believed to be accurate from 1973 onwards as in that year the Creditor Reporting System was established. This dataset keeps track of the amount of money committed towards receiving countries and the projects the money is committed to in detail, allowing insight in the different types of aid disbursed. Data on actual yearly disbursements is taken from the OECD database on aid (2014) and is the aggregated net amount of aid given towards a country per year. The commitments to a specific sector j within a country i at time t are given by comi,j,t and the total disbursements to that country i at time t are given by disi,t.

The actual sectoral disbursements (actdisi,j,t) to sector j in country i at time t

can then be calculated via

𝑎𝑐𝑡𝑑𝑖𝑠𝑖,𝑗,𝑡= 𝑑𝑖𝑠𝑖,𝑡∗

𝑐𝑜𝑚𝑖,𝑗,𝑡

∑ 𝑐𝑜𝑚𝑗 𝑖,𝑗,𝑡, (8)

(23)

actual disbursement intended for a specific sector in a country in a year is thus approximated by calculating the total aid given to that country in a year and multiplying that total aid by the ratio of commitments meant for a specific sector to the total amount of commitments. By doing so a comprehensive dataset of aid flows is calculated which can be used to not only give insights into the total effect of aid but moreover on the effects on economic growth of donating aid to specific sectors. The amounts that are calculated are expressed in constant US$ (2013) to control for inflation and currency exchange effects. As reliable data on foreign aid is the most important part of the dataset this data determines the time period for which other data is gathered, namely 36 years from 1975 to 2010.

Aid is divided into four different categories. Firstly there is economic aid, which is aid meant to improve infrastructure, communication, energy production and the banking and general business sector. The second category is social aid, which is aid meant for education, health and water, social welfare and governmental improvements. Thirdly there is aid meant for improving production within a country, for example agriculture, fishing, forestry, industry and mineral resources production and improving trade. Finally there is general aid, which is general budget support, loan forgiveness, humanitarian aid, environmental and other multisector aid. The categories and their respective DAC5 purpose codes can be found in Appendix A.

Data on the real gross domestic products (GDP) and population sizes of all countries and years is obtained from the Penn World Tables by Feenstra, Inklaar and Timmer (2015). Similar to data on foreign aid this data is also based on a constant US$. Since the raw data is based on a constant US$ (2005) this value is multiplied by the relative change in the US consumer price index between 2005 and 2013, to obtain values based on a constant US$ (2013). This relative change is calculated using data provided by the U.S. Bureau of Labor Statistics (2015). Data on FDI has been retrieved from the database of the World Bank (2014). This variable is defined as net inflows of FDI, which is the sum of all inflows minus disinvestment in a country. As FDI is only available in either current US$ or in percentage of GDP, the percentage of GDP is taken and multiplied by the aforementioned real GDP. All these main variables are in real amounts (of US$, except for population), ensuring as much freedom in estimating as possible.

3.1.2 Control variables

Next to these variables on aid, size of economy, FDI and population several other variables which are known to affect economic growth are also included in the dataset. Firstly data on political and institutional quality is taken from the Polity IV database, maintained by Marshall and Jaggers (2014). This dataset keeps track of the yearly political situation and stability for all countries with more than 500,000 inhabitants. By combining two key variables on political competition and the organization of a country a combined policy index is created, ranging from 1 (no political competition, dictatorial, autocratic) to 30 (high political competition, democratic, open). This variable serves as a proxy for

(24)

institutional quality and openness of countries, following the study of Panizza (2001).

In addition to the political variable a variable on geography is added. The replication of Rajan and Subramanian (2008) showed that the interaction between geography and aid may affect economic growth. The geography variable is defined as the normalized13_{proportion of land with more than 5 frost days per}

month in the winter plus the normalized ratio of land area in the tropics. The definition of this geography variable is similar to that of Rajan and Subramanian (2008) and is based on Bosworth and Collins (2003). Data on frost days is taken from Masters and McMillan (2001) while data on tropical areas is taken from Gallup, Mellinger and Sachs (2001). This geography indicator serves as a proxy on how good the climate in a specific country is for agricultural development for example, which has shown to be of influence on economic development by Gallup, Sachs and Mellinger (1999).

Furthermore, two indicators on health conditions are included, namely life expectancy at birth measured in years and the under-5 mortality rate per 1,000 live births. Data on these variables is taken from the World Bank World Development Indicators (2014), and these variables serve as a proxy for living conditions in a country.

Next, three variables on the composition of the economy are included. The first variable is the amount of land used for agricultural production, in square kilometers. The second variable is related and is the percentage of GDP that consists of agricultural production. This variable is multiplied by total GDP to obtain the total agricultural production in dollars. Diao, Hazell, Resnick and Thurlow (2007) show that agricultural development is an important factor in economic development, justifying the inclusion of agricultural variables. The last variable is the total natural resources rents in percentage of GDP, which is again multiplied by total GDP to obtain the real value. This is the sum of earnings made by a country on the production and extraction of natural resources such as oil. This indicates to what extent a country is dependent on the world market for natural resources. As Sachs and Warner (2001) show, countries which have a large wealth in natural resources have a slower economic growth even after controlling for geographical factors. All three variables concerning the composition of economies are taken from the World Bank (2014).

Finally, four variables on financial development are included in the dataset. The first financial variable is inflation, which is operationalized as the GDP deflator in annual percentages. Inflation has shown to affect economic growth in a negative way by Bruno and Easterly (1998). The second and third financial variables are domestic credit provided by the financial sector and gross savings as a percentage of GDP. Domestic credit and savings are proxies for financial development and depth, following the approach of both earlier replicated studies to include a measure for financial development. Next to that, domestic credit and thus investment is also important in the theoretical framework introduced

13_{This normalization is done by subtracting the average and dividing by the}

(25)

before, as theoretically domestic investment should lead to economic growth and thus should be included in an empirical review. As the replications in this study have shown that government expenditure is an important variable when modelling economic growth, government consumption as a percentage of economic growth is included as fourth financial variable. The data on all four financial control variables are taken from the World Bank (2014). The ratios on domestic credit, savings and government consumption are multiplied by real GDP to obtain the terms in real dollars of these variables.

To complete the raw dataset all non-developing countries are removed. Developing countries are defined as having a gross national income of US$ 11,905 per capita or less as is specified by the World Bank. The list of developing countries that meet this requirement is taken from the International Statistical Institute (2013).

3.1.3 Data transformation

As al data is in real total US$ the data has to be transformed to obtain a usable dataset. Consider for example the case where both a small and large country change from a dictatorial regime to a democracy, changing the policy variable from 5 to 25. While the size of the economy may relatively change 10% for both countries, in absolute terms the effect of the change is much larger for the large economy. A data transformation helps to prevent the necessity of adding interaction terms with population for each monetary variable. Therefore a simple to interpret transformation will be carried out, namely transforming all variables to per capita. This transformation is defined as xtransformed = x/population which

is a common used transformation in previous literature on aid effectiveness. This transformation is also suggested in Roodman (2008) as a method to tackle several estimation problems. In this per capita transformation all cumulative variables are transformed, namely GDP, aid, FDI, domestic credit, national savings, government consumption, the amount of used agricultural land, agricultural production and the profit on the extraction of resources, resulting in per capita values for all those variables. This study will focus exclusively on per capita effects, although other transformations may be worth examining in further research, for example taking variables relative to GDP.

3.1.4 Dealing with gaps

Even though data on total foreign aid, GDP, FDI and population is available without gaps for most countries, the data on commitments is scarce for some smaller countries. Countries for which data on aid or the size of economy is not available for more than two consecutive gap years are completely removed from the dataset. These are mostly small countries which receive aid once every few years and for which the conversion from sectoral commitments to sectoral disbursements cannot be made, or countries which have been in a state of war for a long period between 1975 and 2010. For control variables a linear approximation is used to determine a value to fill in the gaps if sufficient data is

(26)

available before and after these gaps. Linear approximation assumes little variation and a linear trend, and is therefore a quite strict assumption. However, it can be seen in the raw dataset that in most cases there indeed seems to be an approximate linear trend when values are changing but not missing, justifying the use of a linear approximation to keep the loss of observations as small as possible. This type of linear approximation is done in only a few occurrences, namely for less than 0.1% of the total dataset. Removing these countries would have a much larger effect on the size of the dataset and validity of the results and can therefore not be justified, especially since the lack of data has proven to be a major problem in previous studies on foreign aid.

It should be noted that deleting countries based on missing data might introduce a bias in the results. Since the decision to delete a country is based on the lack of data for a specific variable in two consecutive years only countries that lack data are deleted. Since these countries are mostly countries in a state of war, countries which lack good administrative institutions or countries that lack data because of similar reasons this does not purport a random subsample out of the total sample. As the remaining sample then intuitively consists of more developed countries, this leads to the possible existence of a bias. Because of this fact, the findings in this study cannot be generalized to all developing countries, but only apply to the included countries.

Some countries also have a large number of missing data points either in the beginning or at the end of the series. These countries are mostly countries which did not exist until the fall of Soviet Union, which resulted in many new developing countries. As the methods used to analyze the relationship between FDI, aid and GDP are able to deal with this kind of missing data without problems there is no reason to drop these countries from the dataset, however, they will be excluded in certain subsets, as described in the next subsubsection. The total number of included countries and available observations per included control variable can be found in Table 8.

Table 8. The declining number of available countries and observations when including more explanatory variables.

proliferating set of included explanatory variables aid per category + FDI + savings + domestic credit + all other variables # countries 111 108 100 97 76 # annual observations 3511 3090 2514 2419 1995

(27)

3.1.5 Subsets

Earlier studies have shown that the effects of foreign aid and FDI on economies are heterogeneous across countries and regions. Asiedu (2002) provides evidence that the positive effects of FDI are smaller for countries in sub-Saharan Africa than for countries in other regions. Durbarry, Gemmell and Greenaway (1998) and more recently Minoiu and Reddy (2010) also find evidence for different effects determined by geography using dummy variables for specific regions.

To examine these different regional effects the dataset used in this study will be split into several smaller datasets excluding data for one specific continent or region, next to the analyses carried out on the complete dataset. The three different regions to be examined are America, Africa and Asia. The use of these specific different regions is common practice in development economics. If regional effects do exist, the analysis of these subsets should lead to differences in the estimated effects.

Next to splitting by geography this study also examines the subset where relatively rich countries are not included. This is done by setting an arbitrary limit at the 75% percentile of GDP per capita, which is 5,558$ in this dataset, and filtering out observations for countries once they have reached this level. Furthermore, former Soviet countries will be excluded from the dataset, as the reason they are developing may be determined by recent history and they have only recently been founded, possibly leading to different effects.

In addition to the aforementioned subsets this study will look at two operationalizations of the time period. Mainly, the data is examined based on periods of 4 years. As using aggregated time periods is the common practice in related studies this method will also be the preferred method of this study. Longer time periods are used to control for possible outliers and measurement errors in the available data. Aggregation is done using the mean of variables in all available periods of 4 years, resulting in a maximum of 9 observations per country, as the dataset concerns 36 time periods in total. Aggregating the data results in a total of 525 observations on 76 countries when including all control variables. The data will also be examined on a yearly base as all included data is available on a yearly level. The results of the analyses using this yearly data are used as a robustness check of the results obtained using the aforementioned aggregated data.

To donate or to invest? : a comprehensive dynamic panel data study on the economic effects of foreign aid and foreign direct investment in developing countries