MORE AID, LESS DEMOCRACY? A Panel Data Analysis. Athens University of Economics and Business

Transcription

1 MORE AID, LESS DEMOCRACY? A Panel Data Analysis Sarantis Kalyvis and Irene Vlachaki Athens Universy of Economics and Business March 2007 Abstract: Foreign aid flows have increased considerably during the last decades targeting, apart from the economic development of recipients, several goals related to their democratization process. In this paper we investigate whether aggregate aid flows affect the transion of recipient countries to democracy. To this end, we introduce a two-stage binary response model and we use annual data on Net Official Development Assistance covering the period for 64 aid-recipients. We find robust evidence that aid flows are negatively associated wh the likelihood of democratization in the recipient country and that this adverse effect of aid is stronger in countries wh bad economic and social condions. We also find some evidence that the negative effect of aid on the polical regime is moderated when aid flows are accompanied by economic liberalization. Keywords: panel data, binary model, log, prob, endogeney, random effects, democratization, foreign aid JEL classification: D70, F35, C25. Acknowledgements: Financial support through project Pythagoras II co-financed by the European Social Fund and the Ministry of Education in Greece is gratefully acknowledged. Vlachaki would also like to acknowledge financial support by the Propontis Foundation. We have benefed from comments by participants at various conferences and seminars. The usual disclaimer applies. Corresponding author: Department of International and European Economic Studies, Athens Universy of Economics and Business, Patission Str. 76, Athens 10434, Greece. Tel: (+30210) Fax: (+30210) skalyvis@aueb.gr

2 1. Introduction The goal of this paper is to investigate empirically the impact of foreign aid on the democratic transion of recipient countries. This relationship has attracted considerable attention for two main reasons. First, is evident that democratic regimes have gained wide acceptance during the last decades: the percentage of world population living under elected governments wh universal suffrage rose from 31% to nearly 60% between 1980 and 2000 wh more than 80 countries taking significant steps towards democracy (Human Development Report, 2002). This dramatic change in the world polical scene has renewed academic interest in the driving forces of democracy. 1 Second, foreign aid, which is oriented to poorer and typically more autocratic countries, has increased considerably during the last decades and includes several non-developmental goals, like polical freedom, civil liberties and the respect for the rule of law, as established policy objectives proclaiming that they are essential in promoting the democratization process in recipients. Foreign aid can act on the polical regime of the recipient country in various ways. By promoting democratic instutions, good governance and the rule of law, aid flows help preventing the occurrence or reoccurrence of conflict in dysfunctional states. This effect takes mainly place through the strengthening of channels that have been shown to encourage democracy, such as income and education levels (see, among others, Lipset, 1959, Almond and Powell, 1965, Barro, 1996). Another major channel arises through condionaly, which has been extended from the economic realm of good policies to the polical arena and enhances the polical effectiveness of foreign aid. Most donors have placed the promotion of democratic government among the major goals of their aid programs by binding aid transfers wh improved performance in the field of civil liberties, the conduct of elections, and the respect for the rule of law, which are the cornerstone of democracy. However, although the concept of condionaly seems to provide a strong incentive for increasing democratic governance and instutional reform in the recipient countries, the lerature has accepted that wh few exceptions condionaly has not succeeded in systematically affecting policies (see Collier, 1997, Crawford, 1997, Dollar and Prchett, 1998). 1 Starting wh Bollen and Jackman (1985a, 1985b), a series of studies, including Muller (1995), Barro (1999), Feng and Zak (1999), Przeworski et al. (2000), and more recently Acemoglu et al. (2005a, 2005b), Epstein et al. (2006), and Papaioannou and Siourounis (2006), have explored empirically the determinants of democracy. 1

3 In fact, a number of papers have supported the view that foreign aid can be detrimental to polical rights and civil liberties. Bräutigam and Knack (2004) have recently summarized some mechanisms that could explain a negative relationship between foreign aid and democracy, wh the most important being that a large amount of aid can reduce the incentives for democratic accountabily. Excess aid flows can hinder the solution of collective action problems inherent in reform efforts, create moral hazards for both recipients and donors, perpetuate both a soft budget constraint and a tragedy of the commons wh regards to the future budget, and weaken the development of local pressures for accountabily and reform as aid revenues do not depend on the taxes raised from cizens and business. At the same time corrupt government officials will try to perpetuate their rent-seeking activies by reducing the likelihood of losing power. Friedman (1958) has suggested that foreign aid provided to governments increases the relative size of public sector activies and acts as a substute for tax revenues finding; this effect is supported empirically by Remmer (2004). Therefore, aid is likely to impede democratization by hampering governmental accountabily and undermining the cizens control over governing parties through tax paying. Likewise, when aid is given in an authorarian and imposing way that does not perm any parliamentary review or control, fundamental democratic constutions are likely to be violated (Bräutigam, 1992, 2000). Finally, aid transfers may spur competion among interest groups by increasing the size of available resources and inducing corruption, rent seeking and other harmful activies, which in turn lead to less representative polical instutions. 2 The empirical studies that have so far attempted to assess empirically the effect of aid flows on the democratization of recipients have not produced clear-cut results. Crawford (1997) analyzes 29 instances of polically motivated aid sanctions over the period, but finds that these measures did not induce democratization. Goldsmh (2001) finds a posive, statistically significant relationship between official development assistance from Western countries and the level of democracy in forty-eight sub-saharan African countries between 1975 and Al-Momani (2003) uses pooled cross-sectional time-series data for 174 countries from 1976 through 1994 and employs 2 There is a large body of evidence on the rent-seeking activies generated by the reception of foreign aid; see, among others, Grossman (1992), Svensson (2000), Alesina and Weder (2002). 2

4 various measures of democracy. The author does not find any evidence of a direct aid-democracy link and claims that foreign aid has a negative impact on the economic development variable, which makes s overall impact on the level of democracy negative. This might be because recipient countries do not always use aid in a way that would help the economy and only use for polical reasons, which may include repression. Dunning (2004) demonstrates that the small posive effect of foreign aid on democracy is limed to the post Cold War period, a finding that highlights the importance of the geopolical context in condioning the causal impact of development assistance. Knack (2004) uses cross-section data covering the period and finds no evidence that aid promotes democracy; this result is robust to the use of alternative estimation techniques, different proxies for democracy and alternative measures of aid intensy. 3 Recently, Djankov et al. (2006) have claimed that the impact of aid on democratic instutions is negative and even outweighs the corresponding adverse effect of natural resources. In our view, there are three problems wh the empirical lerature on the democracy-aid nexus. First, the existing empirical studies have principally relied on averaged cross-section data, mainly in an attempt to circumvent the low whin-country variabily in democracy levels, whereas the empirical analysis is usually performed whin the context of linear probabily models. However, is well known that cross-sectional analysis whin the context of linear models can be subject to several drawbacks, such as limed robustness in the presence of non-linearies, and parameter heterogeney. Although there are a number of econometric techniques to address these caveats, the suation is likely to be particularly acute regarding the determinants of democratic transion, which is often documented to occur globally in infrequent but massive waves that indicate the presence of strong nonlinearies in polical developments. 4 Second, the allocation of aid is likely to be subject to 3 Notice that since the beginning of the 1990s aid donors began to focus on good governance in the form of increased efficiency of state instutions and changes in the instutional and legal framework regulating the market and the private sector. In a parallel manner, donors placed emphasis on other indirect aspects of (non- )democratisation, such as bureaucracy, corruption and other harmful social activies; for instance, is often claimed that aid flows result in briberies of public officials due to lack of sound public procurement and the associated discretion in awarding contracts. Hence, if democracy is considered a prerequise of good governance the indirect impact of aid on democracy could be assessed through s effect on governance qualy. Bräutigam (2000), Svensson (2000), Knack (2001) and Bräutigam and Knack (2004) have examined the link between aid and the qualy of governance, and find that aid flows seem to increase corruption and hamper bureaucratic qualy, particularly when ethnic fractionalization is high in the recipient country. 4 A classical example is the well-documented surge of democratization involving Sub-Saharan Africa, Latin 3

5 simultaney bias, as may be affected by the donors interests regarding the polical regime of the recipient country. The most clear evidence on the endogeney of aid as a determinant of democracy comes from Alesina and Dollar (2000) who state that..countries that have democratized have received a surge in foreign aid, immediately afterwards[ ] The typical democratizing country gets a 50% increase in aid. Third, the existing studies treat aid recipients as a homogeneous group of countries whose economic and social characteristics do not play a role in shaping the impact of aid on democratization. However, a bulk of anecdotal evidence, surveyed forcefully by Easterly (2006), shows that the democratic impact of aid differs substantially between recipients and depends crically on the domestic condions. To confront these cricisms, we adopt here a Two-Stage Instrumental Variables discrete response framework that is suable for analyzing non-linear patterns in the data at hand, in which we control for the heterogeney of aid recipients by allowing for random effects. We then use data covering the period for 64 aid-recipients to investigate the impact of foreign assistance on their democratization process. Our main finding is that foreign aid flows affect negatively the probabily of democratic transion in the recipient country. Our results are robust to various definions of democracy and the aid measure utilized, to alternative empirical specifications, and to a battery of sensivy tests. We also establish that the negative marginal impact of aid flows on the democratization process of recipients is not uniform, but depends on the general economic and social environment in the recipient country; the more unfavorable this environment is for democracy, the more disastrous are aid flows to democracy. Finally, we take our investigation further by asking whether economic liberalization is important for the impact of aid on polical liberalization and we provide some evidence that the adverse effect of aid on democratization is moderated when aid flows are accompanied by economic liberalization in the recipient country. Our paper thus belongs to the class of studies that have focused on identifying the impacts of aid in recipient countries. In particular, the paper contributes to the lerature in the following ways. First, we are able to identify a clear negative effect of aggregate aid flows on democracy. This result is America and the Caribbean, as well as Eastern Europe, since the late 1980s (often termed as the third wave of democratization after Huntington, 1991). 4

6 important because promoting democracy along wh economic development has become a central objective of foreign policy regarding aid transfers. Second, after taking into account the significant heterogeney of developing countries we are able to establish that the negative effect of aid on the democratic instutions is not uniform, but rather depends on the social and economic stance of recipients. This differentiation is important, because can explain the frequently mentioned stylized fact that aid impedes democracy in some cases, but seems to be less harmful elsewhere. Finally, we explore the interactions of polical liberalization and aid flows wh economic adjustment in the recipient country and we show that the impacts of financial transfers on the recipient s polical progress interact wh s economic transformation. The rest of the paper is organized as follows. Section 2 outlines the econometric model in the context of the present empirical analysis and Section 3 describes the data at hand. Sections 4 and 5 report and discuss the empirical results, section 5 examines the interaction of aid flows wh economic liberalization, and, finally, section 6 concludes the paper. 2. The econometric model In this section we develop the econometric framework used to assess the impact of aid flows on democratization. The main empirical features that need to be addressed are the following. First, democratic transions occur in massive and infrequent waves and thus cannot be captured by linear probabily models. Second, foreign aid flows may not be exogenously determined but are likely to be affected by the level of democracy in the recipient country. To address these issues, we adopt a binary response setup that can capture the non-linear pattern in democratic transions and we follow Rivers and Vuong (1988) and Petrin and Train (2003) to incorporate the hypothesis that aid flows are determined endogenously. Our setup also incorporates random effects to account for potential countryspecific unobserved features. We assume that the polical regime is described by a binary variable, Y, which takes the values 0 or 1 if country i is autocratic or democratic respectively at time t. We then assume that these values are determined by an unobservable latent variable, denoted by * Y, that depends on various countryspecific and time-specific characteristics, including foreign assistance, through the following 5

7 relationship: * Y = α i + β AID + γ k X + u (1) k where AID denotes the measure of aid utilized as received by country i at time t, X includes a set of k 2 observable characteristics of country i and u ~ N(0, σ ) is an i.i.d. disturbance term. Assuming that the time-invariant term, α i, can be spl into a constant part, a, and a random, country-specific part, µ i, 2 wh µ i ~ (0, σ ), so that α i = α + µ i, then (1) can be wrten as: N µ u * Y = α + β AID + γ k X + ε (2) k 2 2 where ε = µ i + u wh ε ~ (0, σ + σ ) and the random country-specific part, µ i, is uncorrelated wh N µ the explanatory variables, i.e. Corr(µ i, X ) = Corr(µ i, AID ) = 0 for all t. u Now, if foreign assistance that country i receives at time t, AID, is affected by unobservable or omted factors captured by the idiosyncratic effects, ε, then AID will be correlated wh the error term. To account for the potential impact of aid endogeney, we use here an Instrumental Variables methodology by following the control function approach suggested by Petrin and Train (2003). This approach decomposes the endogenous regressor, namely AID, as a function of a set of exogenous instruments, g(w) where w is the instrument set, and omted characteristics, ζ, as follows: AID = g(w ) + ζ (3) In our case Corr(AID, ε ) 0 Corr(ζ, ε ) 0 since Corr(w, ε ) = 0, thus implying that the disturbance terms in the equations of aid and regime are correlated. In fact, the direction of the correlation is not obvious. The error terms will be negatively correlated if donors responded to negative democratization shocks by providing more assistance. On the oppose case, countries making progress towards democratization may receive favorable treatment from donors, thus triggering a posive correlation between ζ and ε. Given that Corr(ζ, ε ) 0, ε can be decomposed into a mean condional on ζ given by f(ζ ) called the control function since controls for the part of the original error term, ε, that is correlated wh foreign aid, AID, and a deviation from the mean, ξ, which is 6

8 orthogonal to AID. Following Petrin and Train (2003) and Villas-Boas and Winer (1999), we will assume that the control function is linear in the residuals of the form f(ζ ) = λζ. Thus, we can then rewre the regime function as: * Y = α + β AID + γ k X + λζ + ξ (4) k 2 where ξ ~ N(0, σ ) and i.i.d. The condional probabily of a country i being democratic can then be wrten as: ξ Pr(Y = 1 AID, X, ζ ) = G α + βaid + γκ X + λζ (5) k where G is a function taking values between zero and one. The control function approach adopted here requires a two-step estimation process. In the first step, foreign aid is regressed on the set of instruments via OLS. The estimated residual from the first stage, ˆ ζ, is then incorporated directly as a regressor in the second step, in which Y is regressed on AID, logistic function: ζ and X. Assuming that G is the ˆ ˆ α+ ˆ βaid ˆ ˆ + γκ λζˆ X + k ˆ ˆ ˆ ˆ ˆ e G α + β AID + γκ X + λζ = (6) ˆ α + ˆ β AID ˆ ˆ + γk X + λζˆ k k 1+ e we can estimate a random-effects log model (also called mixed log) by Maximum Likelihood (ML). The two-stage ML log procedure yields consistent and superior estimates compared to other estimation techniques. 5 We can then test for the endogeney of AID by applying a Wald test on ˆ ζ. Alternatively, we will use a similar approach due to Rivers and Vuong (1988) that estimates a prob model wh an endogenous continuous explanatory variable. Estimation of (4) then involves regressing the foreign aid on a set of exogenous controls via OLS and then running a random-effects prob on the saved residuals and the rest explanatory variables. Rivers and Vuong (1988) have shown that the two- 5 See, for instance, Maddala (1983, p ) and Maddala and Lee (1976). Notice that the random effects log requires substantially fewer parameters to be estimated than a typical fixed-effects log and is recommended when the number of cross sections exceeds the number of time uns, as in our case. 7

9 step estimator is consistent and that the usual t statistic on the residuals is a valid test of the null hypothesis that the independent variable of interest, AID, is exogenous. 3. Data In this section, we briefly describe the annual dataset covering the determinants of democratic transion in 64 aid recipient countries (listed in the Data Appendix) for the period We describe the main variables for democratic transion and democracy aid in subsection 4.1 and we then list the control variables in subsection 4.2. A more detailed description along wh the relevant statistics is given in the Data Appendix and in Table A Data on democratic transion and foreign aid To proxy for the dependent variable (polical status) we follow Przeworski et al. (2000), who in turn follow Schumpeter (1950) by defining democracy as a regime in which the executive and the legislature are both filled by contested elections. Conversely, dictatorships are regimes in which eher the executive or the legislature are not filled by contested elections. Contestation implies that multiple parties compete, incumbents have some probabily of losing the elections, and all parties comply wh the results of the elections. Przeworski et al. (2000) have developed a dichotomous measure of regime (updated in Cheibub and Gandhi, 2004), denoted by REG, according to which democracy is a polical system in which key government offices are filled through contested elections. The definion has two parts: key government office, defined as the executive and the legislature, and contested, implying that more than one party has some probabily of winning office through election. The reversed REG dummy variable we employ is coded 1 for democracies and 0 for dictatorships. Transion years are coded as the regime that emerges in that year. To control for foreign assistance we use the standard measure of aid, as provided by the OECD. This measure corresponds to Net Official Development Assistance, which is the net disbursement amount, i.e., disbursements minus amortisation, of those flows classified as Official Development Assistance, a conventional term introduced by the OECD. Official Development Assistance includes Grants or Loans to countries and terrories on developing countries which are: (a) undertaken by the 8

10 official sector; (b) wh promotion of economic development and welfare as the main objective; (c) at concessional financial terms (if a loan, having a Grant Element of at least 25 per cent). In addion to financial flows, Technical Co-operation is included in aid. Grants, Loans and creds for milary purposes are excluded. Transfer payments to private individuals (e.g. pensions, reparations or insurance payouts) are in general not counted. We scale Net Official Development Assistance wh GDP (denoted by NODA), which is the usual weighting mechanism recommended to obtain a proxy for this form of transfers. We also use alternatively Net Official Development Assistance in per capa terms (in constant 2002 prices) as an explanatory variable, denoted by NODA_PC, to examine whether our results differ when alternative weighting schemes of aid are taken into account. Existing studies have concluded that poor aid performance in terms of growth is not affected to the choice of scaling. However, the polical impact of aid may depend on the population size of the recipient country rather than on s economy size for two reasons. First, a heavily populated developing country requires, ceteris paribus, more aid than a less populated one (McGillivray, 1989). Second, expressing assistance in per capa terms might be of particular importance here since macroeconomic data for developing countries rarely reflect the actual size of their economies due to illegal and other underground or unreported activies wh discrepancies reaching sometimes 70% of GDP (see Schneider and Enste, 2000). Hence, foreign aid in per capa terms might be more appropriate given the nature of our dependent variable and also allows us to investigate whether aid effectiveness is affected by the scaling of total aid employed Data on control variables We use a number of control variables to account for economic, polical, social, and structural country-specific determinants of the democratization process in aid recipient countries. The choice of these variables is mainly dictated by the existing theoretical and empirical lerature, and is adjusted according to data availabily for the period under consideration. So, we control for economic condions, demographics, religion, polical, historical, and social condions, as well as for various country-specific characteristics. The lerature on the determinants of democracy usually includes income as a determinant of 9

11 democracy level. Lipset s (1959) general hypothesis is that the more well-to-do a nation, the greater the chances that will sustain democracy. 6 In empirical applications, modernization enters in several forms but the majory of the studies employ the level of education (see, among others, Barro, 1999, Knack, 2004, Acemoglu et al., 2005a). In the present study we use the percentage of lerate population to total population aged Several other country-specific economic characteristics have at times been considered as democracy determinants. According to the lerature, democracy is expected to emerge out of a strategic face-off between the rich minory who is inimical to democracy due to fear of redistribution and the poor majories who try to extract democratic concessions from the rich; thus countries wh higher income inequaly tend to be less democratic (Acemoglu and Robinson, 2006). We use the Theil index to control for the effects of income inequaly on democratization. 7 Although most of the tradional work on the determinants of regime change has focused on the domestic attributes, is also recognized that external factors related to the degree of openness in a country, like international trade, are likely to influence the prospects of democracy. The idea that globalization promotes the diffusion of democratic ideas goes back to Schumpeter (1950), Lipset (1959) and Hayek (1960), who argued that free trade and capal flows enhance the efficient allocation of resources, raise income and lead to economic development that in turn fosters demands for democracy through the spread of innovative ideas and the adoption of more liberal polical systems. 8 We use here a measure of trade openness, namely the sum of exports and imports to GDP, to capture the extent of this influence. According to an argument broadly termed as curse of natural resources, rich countries tend to adopt less democratic ways of governance partly because abundance of natural resources enables the state to buy off society wh low taxation and high welfare spending and thereby allay popular demand 6 Notice that recent empirical work by Knack (2004) and Acemoglu et al. (2005b) shows that this association evaporates once one controls for factors that simultaneously affect income and democracy; we return to this point in subsection 4.2 below. Similarly, Przeworski and Limongi (1997) have argued that there is no systematic relationship between economic factors and the appearance of democratic regimes; however, there is a significant relationship between economic factors and the likelihood of a country remaining a democracy. 7 Some studies use the Gini index as a measure of income inequaly. We opt here for the Theil index due to data availabily. 8 See also Huntington (1991), Whehead (1996), Gledsch (2002), Lopez-Cordova and Meissner (2005). The empirical evidence on this relationship is mixed. Li and Reuveny (2003), Rigobon and Rodrik (2004), and Giavazzi and Tabellini (2005) find eher no impact of trade openness on democracy or even assess an adverse effect. In contrast, Rudra (2005) and Papaioannou and Siourounis (2006) argue that the effect of trade openness on democratization is posive. 10

12 for polical accountabily. It can also distort modernization by spurring the expansion of national income whout inducing the socioeconomic changes that usually accompany an increase in wealth and that are likely to favor democracy. 9 To control for this effect we add a dummy variable for oilexporting countries. Milary expendures (expressed as percentage of GDP) enter the analysis to control for the milary country-specific character of recipient countries. We also use the ethnolinguistic fractionalization index as a measure of population heterogeney, in order to proxy for the number of competing groups and hence for the degree of conflict whin society. We expect social division to affect democratization process negatively since democracy is less likely to prevail in countries that are socially divided and lack cultural and linguistic coherence. 10 We also let religious beliefs enter as a potential determinant of democratization by adding two dummies variables for Islamic and Catholic countries. 11 We also follow some recent studies that regard geographic posion as a factor that contributes to the shaping of polical instutions. For instance, has been argued that in temperate zones the climate is healthier and agriculture is more productive, thus enabling a faster development process that facilates the improvement of instutions (Landes, 1998). La Porta et al. (1999) have established empirically that the latude of a country has a strong posive effect on government performance, especially when one controls for economic performance. We follow this approach and we also introduce the absolute value of latude (normalized between 0 and 1) in the estimated specification. Another important issue is that regimes consolidate over time and become self-sustaining (Barro, 1999, Acemoglu et al., 2005a, 2005b). In the present setup, we examine a simplified version of Barro s (1999) democracy convergence hypothesis according to which the polical regime of a 9 See Karl (1997), Ross (2001), Collier and Hoeffler (2005) and Jensen and Wantchekon (2005). 10 See Lijphart (1977) and Horowz (1993). Notice that the indices do not measure the intensy of conflict between groups but rather, for a given number of ethnic groups in society, the probabily that two randomly selected individuals from the country in question will not belong to the same ethnic group wh a higher value reflecting a greater degree of fractionalization. Due to low availabily of annual data we use each country s average figure over throughout based on the assumption that instutional factors change slowly over time. We then introduce a dummy variable taking the value uny whenever the ethnolinguistic fractionalization index for a country exceeds In 2005 only 10 out of 47 Islamic countries (21%) had electoral democracies, whereas 28 countries (60%) were classified as non-free according to the Freedom-House index of democracy, a combined average of polical rights and civil liberties ratings (Freedom in the World, 2005). Our sample also contains a limed number of countries where Jewish, Protestants, or Hindu are the majory of the population. 11

13 country converges gradually over time toward a (moving) target; Barro (1999) tests this hypothesis whin a linear dynamic model wh two five-year lags of the dependent variable. Knack (2004) also includes an index of inial regime to capture regression-to-the-mean effects due to the limed opportuny of democratic countries to increase their ratings. In this vein, we add the inial values of REG as an exogenous regressor. This specification enables regime ratings to be condional on their starting values and also helps dealing wh serial correlation often met in the dynamic modelling of polical regimes. Finally, we add a dummy variable capturing the Post-Cold war period when democracy experienced a sharp increase worldwide as a result of externally-influenced transparent, participatory, and accountable polical and economic systems, the abandonment of dictators from the West and the acceptance of free trade, human rights, and the rule of law as norms. 12 Since there is a number of control variables aiming at capturing the economic, social, polical and country-specific characteristics of the countries at hand, we anticipate that some of these variables may be highly correlated, which would in turn affect the empirical specifications. Therefore, we report the correlation coefficients of foreign aid wh the instruments involved in stage 1 (Table A2) and the correlation coefficients of the control variables that explain polical regime in the main stage 2 (Table A3). As expected, NODA is negatively correlated wh inial income and inial population. Foreign aid also correlates negatively wh latude and oil exporting activy. In countries where Islam is the predominant religion there is evidence of expanded expendures in the milary sector. The milary index is also posively associated wh trade openness, income inequaly and oil exporting activy, whereas ethnolinguistic fractionalization correlates negatively wh income levels and posively wh income inequaly. 4. Empirical implementation and results Following the approach described in section 2 we model the impact of aid on democratization whin the following two-stage empirical setup: 12 According to the lerature, in the years immediately after the fall of the Berlin Wall in 1989 and the collapse of the Soviet Union in 1991 democracies increased from about 40 percent of all states to 60 percent; see Bratton and van de Walle (1997), Kirschke (2000), and Solt (2001). 12

14 Stage 1: uˆ ˆ ˆ ˆ AID - = a1 + bk + cr i X Z k r (7) Stage 2: ˆ PY ( = 1) = F aˆ ˆ ˆ ˆ ˆ 2 + β X + γaid+ δu k k (8) where X is the vector of the k exogenous regressors of the polical regime function, Z is a vector of r time-invariant, country-specific instruments of AID. Stage 1 given by equation (7) is a reduced-form specification used to explain aid receipts by variables that are not included in the regime regression but are closely related to aid giving. (In subsection 4.1 we provide a more extensive review on the determinants of aid that are likely to serve as potential instruments in our setup). Stage 2 given by equation (8) is then a nonlinear binary-response model wh a dichotomous dependent variable that indicates the probabily of democratization and corresponds to a log or prob specification depending on the distribution of the error term. Assuming that the random effects are uncorrelated wh the explanatory variables, estimation of (8) via ML is unbiased and consistent. i 4.1. Instrumenting for aid flows We explore the potential endogeney of aid by regressing NODA on various pre-aid factors that have been pointed out by the lerature as major aid allocation creria: Boone (1994, 1996), Burnside and Dollar (2000), Easterly et al. (2004), and Knack (2004) have shown that there are several instruments for aid that can be used to address endogeney problems. We follow these studies and include proxies of recipients needs and donors strategic interests. We use inial income (measured by the log of real per capa income in the beginning of the period or the first available observation) and inial population (in logs) to capture recipients needs. We expect a negative correlation between aid and both income and population levels. Regarding the donors interests there is ample evidence that donors direct their aid to low-income countries; however, donors are also influenced by the population size of the recipients ( country-size bias ). Despe the popular belief that aid is primarily There are several reasons why the size of the recipient country may be an important determinant of aid flows. First, both international instutions and bilateral donors hesate to transfer large nominal amounts, as they will come under much greater public scrutiny than relatively smaller amounts. Second, small countries may have relatively higher influence in some international organizations wh the most obvious example being the voting process at the Uned Nations. Finally, small countries may be more willing to sell their influence, as they may 13

15 motivated to assist the poor, substantial evidence points towards important polical and geopolical factors, like strategic alliances of donor countries as major driving forces behind aid programs. 14 To capture these strategic interests we use polical dummy variables aimed at capturing the importance of the country to a particular donor. The dummies include Sub-Saharan Africa (to which most European aid is directed), the Franc zone countries, Egypt (a tradional ally of the U.S.), and Central American countries (also in the U.S. sphere of influence). These dummy variables also aim at capturing regional effects and we expect that they will be posively related wh aid flows. We rely on two tests to evaluate the validy of our instruments. First, we follow Staiger and Stock (1997) and we use the heteroscedasticy and auto-correlation robust F-test of the first stage to test the exclusion of the instruments. The null of the test is that the instrument set is weak and we consider our instruments valid if this F-statistic exceeds the value of 10. Second, we use the Stock and Yogo (2005) test for weak instruments based on Cragg and Donald (1993) to investigate whether our instruments are good predictors of aid flows. 4.2 Results The first-stage estimations for the system of equations (7) and (8) are reported in columns (1) and (4) of Tables 1 to 4. In these specifications foreign aid is regressed on various instruments (pre-aid factors and regional dummies) and the exogenous variables of the main second-stage democracy regression via Ordinary Least Squares (OLS) corrected for heteroskedasticy. 15 We find that foreign assistance is systematically directed to catholic populations and low-development countries. Open economies are potential candidates for aid, whereas oil exporters and milary countries receive proportionally less aid. We also find evidence in favour of the country-size bias effect wh highly populated countries receiving lower amount of assistance. Egypt and Sub-Saharan countries enjoy more aid ceteris paribus. However, we find no evidence that countries in Central America receive gain more from joining a coalion than by acting independently. 14 For some studies on the determinants of aid see Maizels and Nissanke (1984), Frey and Schneider (1986) and Trumbull and Wall (1994). The majory of the studies on the determinants on aid conclude that donors strategic interests are a more important allocation crerion than commercial interests. 15 Notice that inial income, inial population levels and regional dummies can be considered as valid instruments since they can jointly explain aid inflows rather well, whereas their correlation wh the future polical regime is found to be low. 14

16 proportionally more assistance. When Egypt is excluded in the outliers-free estimation we find strong evidence that countries belonging to the Franc Zone have a comparative advantage regarding foreign assistance allocation. Also, donors seem to favor recipients that are located far from the equator and those where income inequaly is high. Finally, there appears a posive structural break in the amount of aid giving during the Post-Cold-War period whereas donors seem to favour less democratic countries. We then test the validy of our approach by the F-test of the first stage; s value is above the usual thresholds implying that both the Staiger and Stock (1997) and Stock and Yogo (2005) creria are satisfied. Finally, we include the predicted reduced-form residuals obtained from the firststage into the second stage and the Wald coefficient test on the Predicted NODA variable always leads to the rejection of the null hypothesis that NODA is exogenous. We then estimate the polical effect of foreign aid based on two alternative specifications. In the first specification we predict the probabily of observing democracy by use of a wide set of economic, social, instutional and religious controls. In a second, augmented specification a proxy for the inial polical condions of recipient countries is added in the estimated equations to capture potential mean-reverting dynamics in democracy. We use the inial values of the dependent variable or the first available observation as a proxy for inial polical condions. Turning first to the estimates for the control variables, our findings confirm the view expressed by Knack (2004) and Acemoglu et al. (2005) who showed that the association between polical change and economic condions falls out once one controls for factors that simultaneously affect income and democracy. Trade openness is not also found to act as a determinant of democracy. Oil exporting activy is found to be a deterrent of democracy wh the estimated coefficient being negative and statistically significant in the majory of specifications. We confirm La Porta et al. (1999) who give a democratic advantage to more distant countries; the coefficient of latude scale is posive and statistically significant in most specifications, meaning that countries that are far from the equator are prone to more democratic governments. We also find strong evidence that religion is a crucial determinant of polical circumstances. Given our results, catholic countries seem to enjoy more democracy whereas Islamic religion does not exhib a strong negative effect on democratization. Other time-invariant factors like milary structures and ethnolinguistic fractionalization are not found 15

17 to exert a statistically significant effect on democracy. Finally, we find that inial polical condions have a posive and statistically significant coefficient showing that there exists considerable degree of persistence (mean reversion) in democracy; nonetheless the inclusion of this variable in the estimated regression leaves the rest of the estimated coefficients virtually unaltered. Regarding the main variable of interest, in Table 1 we find that foreign aid as % of GDP exerts a negative and statistically significant effect on the probabily of democratization. In particular, in column (2) we report estimates of a log model that incorporates both endogeney and heterogeney hypotheses, whereas in column (3) we estimate a prob model. Even when polical history enters our model to account for regression-to-the-mean effects, a statistically significant, negative coefficient of aid emerges in both the log and prob models. Given the dispersion of our observations, the above finding might be due to outliers that work as disturbance factors. In fact, we observe that the allocation of aid exhibs large variation across recipients; for instance, average foreign aid amounts for 5.5% of total income in the developing world, but in some countries exceeds 50% of total GDP. Countries like Guinea Bissau, Mozambique and Sao Tom and Principe are highly dependent on foreign aid, whereas in most recipients aid has increased steadily only recently. To investigate the robustness of our results to the presence of outliers, in Table 2 we rerun the regressions of Table 1 after excluding influential observations detected by the Hadi (1994) procedure. 16 The broad picture confirms the finding of a negative impact of aid on democratization. Finally, we investigate whether our main finding that aid is bad for democracy is affected by the scaling mechanism of aid. In Table 3 we scale aid wh total population (NODA_PC) and we replicate the log and prob estimations for the total and the outliers-free sample sets whin the augmented specification that controls for the inial regime. 17 The estimated coefficients on NODA_PC are found to be negative and statistically significant, thus confirming the negative effect of aid on the democratization process of recipients. 16 To identify outliers in our data sample we use the standard cutoff probabily of As the estimation by use of the original variables yielded large standard errors for most of the variables at hand due to their large dispersion, we report the results for the normalized variables. 16

18 4.3 Robustness tests and goodness-of-f measures Our first robustness test of the evidence reported earlier on follows the reasoning adopted by Papaioannou and Siourounis (2006) who examine only countries that are classified inially as nondemocratic. This setup aims at addressing issues like sample selection and reverse causaly, since is often argued that only countries that have already started their democratization process are eligible for assistance that aims, for instance, at helping improve elections and thus are likely to complete the transion to democracy as a result of this type of flows. In our sample, 51 countries out of 64 (80%) in the whole sample entered inially as members of the non-democratic group. Here we use an outlierfree sample wh 44 out of 64 countries (69%). 18 These figures indicate that most of the countries in our sample were classified as non-democratic in the beginning of the period (see Table A1 in the Appendix for a detailed classification). Specifications (1) to (3) of Table 4 are Hadi outliers-free estimations for the sub-sample of inially non-democratic countries. In both log and prob models the estimated coefficients for NODA retain their negative sign and are found to be statistical significant. We also investigate whether our results are the outcome of spurious correlation between aid giving and polical developments due to the impact of neglected time trends; for instance, a spurious negative relationship could arise if the frequency of democracy is an increasing function of time, whereas aid amounts decline over time. To remove the time-dependent variation of variables we add a linear time trend in both stages. 19 As one can readily see in specifications (4)-(6) of Table 4 the linear time trend exerts a posive and statistically significant effect on both NODA and democracy, indicating that the negative sign of NODA variable cannot be spuriously attributed to time trending. As expected, the coefficient of NODA remains negative and statistically significant, whereas the rest of the results are similar to those reported in section 4.2. Given the existence of alternative codings of polical regimes, another concern is whether the results depend on the measure of democracy utilized. To test if our results are robust to the definion 18 Countries excluded as outliers are Bahrain, Egypt, Jordan, Kuwa, Maurania, Oman, and Saudi Arabia. Our results do not change qualatively when we use the full sample of 51 inially non-democratic countries. 19 Since the temporal variation of variables is captured by the linear time trend we om the Post-Cold-War dummy to avoid multicollineary problems. The negative effect of NODA remains intact even when both temporal variables enter the regression simultaneously. 17

19 of democracy we use Poly2 index as a measure of democracy. 20 Specifically, we construct a dummy variable coded one if the Poly2 score is strictly posive and zero otherwise. 21 We then estimate the corresponding models in specifications (7)-(9) of Table 4 after controlling for inial polical regime and outliers. Although the choice of Poly2 variable result in a moderate decrease in the sample size and the model f, the results obtained are in general similar to those reported in Section 4.2. A final robustness test involves the relationship between foreign aid and democratization in the Post-Cold-War period. It has often been pointed out that the causal impact of aid on democracy during the Cold War and post Cold War eras may differ because the importance of democratization as an explic policy target of aid programmes has risen considerably in the 1990s for several reasons. First, the waves of democratization that started in Latin America and soon prevailed in other parts of the world prompted developed democratic countries to increase financial assistance in an effort to promote stabily and secury. Moreover, the end of the Cold War rendered the support of fledgling democracies an acceptable international objective, whereas reduced the influence of geopolical creria on donors allocation of aid (as authorarian regimes of the developing world lost the support received during the Cold War era). 22 To detect any structural breaks in the effectiveness of AID after the 1990s we separate the sample into two subgroups corresponding to periods and and we use a Wald test on the null hypothesis that there is no statistically significant difference between the coefficients of aid in the two subgroups. We find that although the magnude of the (negative) effect is considerably smaller in 20 Poly2 index ranges from -10 to +10 wh higher values denoting more freedom and equals the difference between the Poly democracy and Poly autocracy index. The 10-point democracy index is based on evaluations of how executives are selected, and of whether or not there are effective instutional checks on their power. The 10-point autocracy index represents some combination of less competive processes for selecting chief executives, and fewer constraints on their authory. Addional detail on these indicators is contained in Gurr and Jaggers (1998). 21 Another widely-used measure of democracy is provided by the Freedom House comprised by a combination of polical rights and civil liberties ratings. We do not use this index as an addional robustness test as Freedom House data begin in 1972 and s correlation wh Poly2 is very high. 22 Knack (2004) also investigates the potential effect of aid on democracy in a subsample that corresponds to the period. The author argues that aid was conceivably more effective in this Post-Cold-War period, in which donors likely gained leverage relative to recipient nations, which could no longer play the superpowers off against one another. In particular, aid was less likely to be used for propping up pro-u.s. dictators, which may have forestalled democratization during the Cold War. However, his empirical findings over the Post- Cold-War period support the main conclusion of his study, namely that aid is unrelated to democratization. 18

20 the Post-Cold-War sample, the calculated Wald statistic (2.18) is substantially lower than the crical value (3.84), which leads to non-rejection of the null hypothesis. 23 We close this subsection by reporting some goodness-of-f measures for our regressions (see the lower part of Tables 1 to 4). The first goodness-of-f measure is the Pseudo-R 2 statistic that ranges from 0.63 to 0.82 and indicates a substantial explanatory power for the models at hand. We also test for the joint significance of all variables by employing a Wald test and we find that we can always reject the null hypothesis that all of the slope coefficients are not significantly different from zero. Next, we test whether the errors terms involve a country-specific component. We use the fact the idiosyncratic errors ε will be serially correlated across t wh Corr(ε, ε is )= σ µ / ( σµ σu ) +, for t s, whereas for t = s this ratio denotes the proportion of the total variance attributed to the panel-level 2 variance component, σ, also known as the Rho measure. We test for the presence of random effects µ by a Likelihood Ratio test, where the null hypothesis corresponds to Rho = 0, implying that the panel level variance component is not important and the panel model is not a significant improvement over the pooled model. The estimates of Rho indicate that the random-effects estimator is superior to the pooled estimator at standard significance levels for all specifications. Finally, we report the percent correctly predicted that equals the percentage of times the predicted Y matches the actual Y. In order to evaluate the overall abily of the model to predict both zero and uny values we calculate a weighted index of the percent correctly predicted. 24 As can be readily seen, the log and prob regressions in Tables 1-4 predict 66-82% of the polical outcome correctly indicating a rather strong predictive power of the empirical models Interpreting the empirical results The discrete response random-effects model of our analysis implies that the probabily of observing democracy in country i at time t is a function of measured explanatory variables and unmeasured factors captured in the country-specific, time-invariant random effect. However, in 23 The detailed results are available upon request. 24 This amount is a weighted average of the percent correctly predicted of zero Y and the percent correctly predicted of uny Y, where the weights are the fractions of zeros and ones in the sample, respectively. 19