ARE AID FLOWS EXCESSIVE OR INSUFFICIENT? ESTIMATING THE GROWTH IMPACT OF AID IN THRESHOLD REGRESSIONS Sarantis Kalyvitisa, Thanasis Stengosb, and Irene Vlachakic September 2011 Abstract: Existing empirical studies and policy reports provide ambiguous results on the growth effect of foreign aid flows in the recipient countries. The present paper examines whether there exists an aid threshold that determines the growth impact of foreign aid. We use a threshold regression methodology to estimate growth specifications and the associated aid thresholds in a sample of 42 aid recipients covering the period 1970-2000. Our findings indicate that there is a threshold level of aid, above which the growth impact of aid becomes positive. Keywords: growth, aid, threshold regression, endogeneity. JEL classification: F35, O4, C2. Acknowledgements: We have benefited from comments and suggestions by two anonymous referees. We thank Zeb Aurangzeb for excellent research assistance. a Corresponding author: Department of International and European Economic Studies, Athens University of Economics and Business, Patission Str. 76, Athens 10434, Greece. Tel: (+30210) – 8203151. Fax: (+30210) – 8203137. e-mail: [email protected] b Department of Economics, University of Guelph, Guelph, Ontario N1G 2W1, Canada. e-mail: [email protected] c Department of International and European Economic Studies, Athens University of Economics and Business, Patission Str. 76, Athens 10434, Greece. e-mail: [email protected] 1. Introduction A persistent issue in development economics involves the impact of foreign aid in terms of boosting economic growth in the recipients. The empirical literature on this issue, surveyed by Kanbur (2006), McGillivray et al. (2006) and Doucougliagos and Paldam (2008), has provided ambiguous evidence. Earlier empirical studies, conducted until the 1990s, concluded that aid flows did not have a positive effect on growth. Subsequent studies have been more favourable to a positive impact of aid on growth under the assumption of diminishing returns (Hadjimichael et al., 1995; Hansen and Tarp, 2000, 2001; Lensink and White, 2001; Dalgaard and Hansen, 2001; Clemens et al., 2004; Alvi et al., 2008) or the conditionality of sound policies in the recipient country (Burnside and Dollar, 2000). Yet, recently, Rajan and Subramanian (2008) have conducted an extensive study on the growth impact of aid and have found no evidence of a significant effect over the last decades.1 In parallel to the empirical literature, and viewing the growth-aid nexus from a policy perspective, several reports have highlighted the role of insufficient aid flows in explaining the poor growth results of recipients. Perhaps the most prominent manifestation of this claim involves the attainment of the Millennium Development Goals (MDGs), which require substantial additional funding in terms of foreign aid to the developing world.2 For instance, Zedillo et al. (2001) have estimated that roughly $50 billion a year in additional aid would be required to achieve the MDGs in all developing countries. Similarly, Devarajan et al. (2002) have provided a figure in the range of $40-70 billion, which roughly represents a doubling of official aid over 2000 levels. The Commission for Africa (2005) has called for an additional $25 billion per year in aid to African countries by 2010, with a further $25 billion a year to be implemented by 2015. The present paper attempts to re-examine the growth impact of aid by addressing the following 1 See, however, Arndt et al. (2010) for a critical review of the Rajan and Subramanian (2008) approach. In July 2005, the G-8 agreed to double foreign aid to Africa, from $25 billion a year to $50 billion in an attempt to finance the “big push”, required for African countries to get out of the poverty trap though a large aid-financed increase in investment. Similarly, the European Commission (2005) issued an “EU Strategy for Africa” in which increased aid was required to achieve a significant boost in growth. See also the United Nations (2006) report for a detailed review on related estimates for Africa. 2 1 question. Are aid flows to developing countries excessive or insufficient? In other words, we aim at answering whether there exists an aid threshold, above which the growth impact of aid changes critically. To this end, we use the threshold regression model developed by Hansen (2000) and Caner and Hansen (2004) for estimating variants of standard growth specifications in a panel of 42 aid recipients covering the period 1970-2000. Our central finding is that there is a threshold level of aid, above which the growth impact of aid becomes unambiguously positive. In particular, we find that low levels of aid (measured as a percentage of the recipients’ GDP) exert a negative or insignificant effect on growth. However, the growth impact of aid becomes positive for recipients where aid flows exceed a critical threshold, amounting roughly to 3.4% of their GDP. Our results obviously coincide with recent calls for a major scaling up of aid aiming at helping poor countries achieve the Millennium Development Goals. In particular, Sachs et al. (2004) and Sachs (2005) have put forward an idea that goes back to Rostow (1960), according to which poor countries are stuck in low-saving poverty traps and that a major intervention (‘big push’) is required to eliminate poverty. One simplifying idea behind these calls is that investment is inadequate in developing countries due to low savings (triggered, for instance, by the needs for subsistence consumption) and poor productivity. As a result, these countries converge to a low-growth equilibrium, a situation that is aggravated under credit market imperfections. Alternatively, potential non-convexities in the production process, such as increasing returns on infrastructural capital or threshold effects in human capital, suggest that a large aidinduced rise in domestic investment would have a strong long-run growth impact. In this vein, aid recipients could benefit from a massive inflow of aid oriented towards savings and capital accumulation. It is noteworthy that actual experience and associated empirical evidence have not provided overwhelming support for these mechanisms until now. Although Azariadis and Stachurski (2005) have noted that generally poverty-trap models seem to be lacking testable quantitative implications, some studies have attempted to investigate their predictions in the context of aid flows. Easterly (2006) has claimed that the stylized facts are not consistent with a low-income poverty trap due to insufficient aid, as growth is lower in aid-intensive countries than in similar developing countries that get little aid. In 2 addition, aid to Africa has risen over time (measured as a percent of income), but Africa’s growth rate has fallen at the same time. Kraay and Raddatz (2007) have recently tested whether the savings and increasing returns patterns predicted by poverty-trap models are supported by the data. The authors show that there is no supporting evidence either in the behavior of savings and per capita income, or technological nonconvexities, in favor of a poverty trap. The authors also fail to provide evidence for the existence of a high-growth high-equilibrium that countries might be able to attain with appropriately large aid inflows. Against such a background, the results presented in the current paper seem to offer, for the first time, some compelling evidence that large-scale aid flows can have a significant growth impact over the long run. This empirical regularity is supported by the recent findings of Herzer and Morrissey (2009) who have established that there is substantial heterogeneity in the output effects of aid among 59 recipients: although the estimated long-run effect of aid is negative, almost one-third of the countries examined have enjoyed a positive growth effect. Indeed, the United Nations (2006) report has mentioned several cases of aid success, where aid flows have resulted in boosting domestic investment and growth over the last decades. For instance, the East Asian miracle economies, notably the Republic of Korea and Taiwan, received enormous amounts of aid during the initial and early stages of their development.3 In Africa, both Botswana and Mauritius received remarkably large amounts of aid at key strategic moments in their development as earlier did Tunisia.4 These examples indicate that, despite the often-cited failure of aid in boosting development in recipients, large amounts of well-targeted aid can, in conjunction with other factors, produce remarkable success stories in terms of growth. The present paper belongs to the newer generation of empirical studies that have attempted to investigate heterogeneous policy effects on growth. For instance, Kourtellos et al. (2007) have shown that countries that belong to a growth regime characterized by levels of ethnolinguistic fractionalization above 3 The nearly $6 billion in US economic aid to South Korea between 1946 and 1978 was only marginally lower than the US total aid to Africa in the same period ($6.9 billion). A similar pattern was found in Taiwan, where although its big push began on the back of a greater degree of domestic resource mobilization, aid still accounted for nearly 40% of gross domestic capital formation in the 1950s and was over $4 billion between 1949 and 1967 with per capita aid being higher than that of Korea. 4 Botswana enjoyed initially a very high aid to GDP ratio, which dropped sharply following a sustained period of rapid growth, whereas a similar pattern was found in Mauritius. 3 a threshold value experience a negative partial relationship between aid and growth, while those belonging to the regime with fractionalization below the threshold do not experience any growth effects from aid. Yet, the authors also find that the typology of these regimes may be alternatively well-characterized by institutions or macroeconomic policies, and not just ethnolinguistic fractionalization. Instead, our approach identifies the threshold level of aid that triggers differences in its growth impact. Investigating threshold effects of aid on growth in conjunction with other variables, such as ethnolinguistic fractionalization, or key aggregates, such as inflation or the domestic fiscal stance, seems therefore to offer a promising route for future research. The rest of the paper is structured as follows. Section 2 briefly outlines the empirical methodology and describes the specification utilized and the dataset at hand. Section 3 presents the empirical results and section 4 concludes the paper. 2. Empirical methodology and data The threshold regression model treats the sample split value (threshold parameter) as unknown by internally sorting the data on the basis of some threshold determinants into groups of observations, each of which obeys the same model. The threshold regression approach is parsimonious, but also allows for increased flexibility in functional form and it is not as susceptible to the curse of dimensionality problems as nonparametric methods. Chan (1993) showed that the asymptotic distribution of the threshold estimate is a function of a compound Poisson process. This distribution is too complicated for inference as it depends on nuisance parameters. Using a concentrated least squares (TR-CLS) approach, Hansen (2000) developed a more useful asymptotic distribution theory for estimating both the threshold parameter and the regression slope coefficients in a cross-section of observations, as opposed to simple parametric approaches that set the threshold exogenously.5 In particular, assume that { yi , xi , q, ui }in=1 is strictly stationary, ergodic and ρ-mixing, and that 5 Masanjala and Papageorgiou (2004) point out that the exogeneity assumption in determining the threshold effect is, in fact, constrained to the estimation of dynamic linear panel data models, which is not the case in the context of the present analysis. 4 Eu i |F i−1 = 0 , where yi is the dependent variable (growth), xi is a p × 1 vector of covariates (including aid), and qi is a threshold variable (aid). Consider then the following threshold regression with threshold for aid: yi = xi′ β1 + u1i , qi ≤ γ (1) yi = xi′ β 2 + u2i , qi > γ (2) Equations (1) and (2) describe the relationship between the variables of interest in each of the two regimes with γ being the sample split (aid threshold). Note that qi is observed but the sample split is unknown. Τhe variance covariance matrix of the errors (u1i , u2 i )′ has the following properties: Eu 1i , u 2i = 0 , Eu 21i = σ 21 > 0 , Eu 22i = σ 22 > 0 . In general, if the model involves exogenous slope variables then estimation is based on Concentrated Least Squares (Hansen, 2000). In turn, the heteroskedasticityconsistent Lagrange Multiplier (LM) test introduced by Hansen (1996) is employed to verify whether there is indeed evidence of a sample split; the null hypothesis of the test is that there is no threshold effect and the corresponding p-values are computed by a bootstrap analog. Using a similar set of assumptions, Caner and Hansen (2004) study the case of endogeneity in the slope variables and propose a concentrated two stage least squares estimator (IVTR-C2SLS) for the threshold parameter and a GMM estimator for the slope parameters. To estimate equations (1) and (2) we use data from 42 aid-recipient countries over the period 19702000. Although the sample size is relatively small compared to related studies due to data limitations, it is representative of the population of aid recipients.6 Since our emphasis is on the long-run growth impact of aggregate aid without the inclusion of country fixed effects that traditionally help capture the impact of worldwide business cycles, we estimate long-run horizon cross-country regressions using alternatively 6 The countries included are: Burkina Faso, Bolivia, Brazil, Botswana, Chile, Côte d'Ivoire, Cameroon, Congo Rep., Colombia, Costa Rica, Dominican Republic, Ecuador, Egypt Arab Rep., Ethiopia, Ghana, Gambia, Guinea-Bissau, Guatemala, Honduras, Indonesia, Jamaica, Kenya, South Africa, Sri Lanka, Madagascar, Mexico, Mali, Malawi, Malaysia, Niger, Nicaragua, Peru, Philippines, Paraguay, Singapore, Sierra Leone, El Salvador, Thailand, Trinidad and Tobago, Uganda, Venezuela, and Yemen. 5 whole-period and ten-year averages, rather than four-year averages as is common in a strand of the relevant literature. Hence, growth volatility, which is far higher in poorer countries (Pritchett, 2000), and cyclical factors are unlikely to affect our estimates. In turn, we follow the empirical specification adopted by Dalgaard et al. (2004, Table 3), which allows for a parsimonious representation of the long-run growth equation with aid as one of the determinants. However, we also experiment with additional potential control variables to assess the robustness of our empirical results. Data come from the World Bank database unless otherwise specified. The dependent variable of the estimated regressions is the average growth rate of real per capita GDP. To capture convergence effects the logarithm of initial GDP per capita in constant 1985 dollars (source: Heston et al., 2006) is included as a control variable. Dalgaard et al. (2004) have established that the growth impact of aid is far smaller in the tropical region. In line with these authors, the importance of (non-political) structural characteristics on aid effectiveness is assessed using the fraction of a country’s area that is located in the tropics (source: Gallup and Sachs, 1999). A measure of institutional quality that captures security of property rights and efficiency of the government bureaucracy also enters growth regressions; data are drawn from Knack and Keefer (1995).7 Turning to macroeconomic policy variables and in line with Burnside and Dollar (2000) and Dalgaard et al. (2004), we use the budget surplus as a percentage of GDP to capture fiscal policy in addition to inflation and the revised trade openness dummy variable introduced by Sachs and Warner (1995) and updated by Easterly et al. (2004) and Wacziarg and Welch (2008).8 Regarding data on aid flows, we employ in benchmark regressions ordinary data on Effective Development Assistance (EDA) measured as a percentage of real GDP (constant 1985 dollars) drawn from Roodman (2007). Using EDA allows comparison of our empirical results with studies that have employed EDA as a measure of aid; see, among others, Burnside and Dollar (2000), Dalgaard et al. 7 The dummies for East Asia and Sub-Saharan Africa, which are routinely included in empirical growth specifications, cannot be identified in a threshold regression framework. 8 Sachs and Warner (1995) define closed economies as those having average tariffs on machinery and materials above 40%, or a black market premium above 20%, or pervasive government control of key tradables. 6 (2004), Kourtellos et al. (2007) and Economides et al. (2008).9 A prevalent criticism of aid-growth regressions involves the likely endogeneity of aid as it is often argued that donors might reward countries that have used aid well in the past or, conversely, help countries that have experienced natural disasters, thus inducing a spurious correlation between aid and growth. To avoid endogeneity-induced problems and account solely for the exogenous component of aid, we follow the instrumentation strategy introduced by Rajan and Subramanian (2008). This approach picks instruments directly at the level of the donor, rather than the recipient country, and hence precludes any direct association of excluded instruments with growth rates. In turn, the choice of aid instruments relies on two assumptions: the first assumption is that the greater the extent of historic relationships between a donor and a recipient the more likely it is that a donor will want to give this country aid. This idea introduces colonial links and common language in the aid supply regression. The second assumption is that donors are more likely to want to give aid the more they expect to have influence over the recipient. Thus, the relative size of the donor and the recipient, and also the interaction terms between the relative country size and the colonial links are included in the set of instruments in order to construct the exogenous part of aid flows.10 To assess the relationship between actual and fitted aid for the period 1970-2000 we regress actual aid on fitted aid and the full set of growth covariates. We find that the relationship between actual and fitted aid is remarkably strong, with a t-statistic that exceeds 6. Figure 1, which depicts the residuals from regressions of actual and fitted aid on the growth covariates, indicates that the correlation coefficient is 9 Notice that EDA corresponds to nearly 1.9% of the recipients’ GDP on average, a figure that is close to our study’s average. An alternative measure of aid flows is Net Official Development Assistance (Net ODA). However, as illustrated by Chang et al. (1998), Net ODA focuses on the net flow of grants and concessional loans (entailing a grant element of at least 25%), therefore leading to systematic overestimates of the concessionality of official loans especially after mid-1980s. EDA, on the other hand, captures the overall grant element of all official financial flows, therefore allowing meaningful comparisons between recipients and donors. Net ODA in our sample amounts, on average, to around 6% of the recipient’s GDP and often exceeds 7%, whereas in the population of recipients it ranges from -0.5% to 12% of GDP with only 3% of the recipients (Comoros, Cape Verde, Djibouti, and Liberia) exceeding the maximum value of our study (7%) over the period 1970-2000. 10 In fact, the authors model the supply of aid based on the bilateral (donor-recipient) relationship and then aggregate up the predicted values of aid received from each donor over all donors in order to get a precise measure of aid flows as a percentage of each recipient’s GDP. See Rajan and Subramanian (2008) for more details on the instrumentation strategy. 7 high and statistically significant, reaching almost 0.65. Thus, we can safely infer that, as in Rajan and Subramanian (2008), fitted aid contains a great amount of information about actual aid, which cannot be attributed to growth factors that affect both variables simultaneously. Another important issue in the present context involves the potential endogeneity of institutional quality since more developed countries can simply afford better institutions or because both growth and institutions might be affected by the same factors. In this vein, Acemoglu et al. (2001) have used European settler mortality as a source of exogenous variation in institutions.11 Following this rationale, we adopt the Acemoglu et al. (2001) approach to address the likely endogeneity of institutions and we use their dataset on settler mortality.12 These data correspond to estimates of mortality rates (expressed in logarithms) faced by European soldiers, bishops and sailors in the colonies in the 17th, 18th and 19th centuries. However, the settler mortality instrument is available for a subset of countries that were colonized, which reduces our sample to 36 countries. Table 1 summarizes the descriptive statistics of the variables at hand. Actual aid flows for the 42 aid recipients analyzed here have on average been 1.5% of their GDP ranging between nearly zero and almost 7%. In contrast, the estimated exogenous component of aid amounts to almost 5% of the recipients’ GDP and varies widely between -4.4% and 30%. The annual growth rate of the recipients’ real per capita GDP hardly reaches 1.5%, but in some extreme cases it may exceed 7%. Institutional quality is at medium levels and average inflation amounts to 20%. Average budget surplus is close to zero and almost half of the recipient counties have closed economies, according to the definition of Sachs and Werner (1995). Ninety per cent of the recipients’ land is located in the tropics. Descriptive statistics for settler mortality 11 During colonization in the previous three centuries, Europeans pursued different policies depending on the mortality rate faced by the settlers. Specifically, Europeans were more likely to set up extractive institutions when faced with high mortality, and it is possible that the differences in institutions have persisted to create differences in institutional qualities across countries in the late twentieth century. Notice that this instrumentation approach and its empirical implementation have been criticized by Albouy (2011). Acemoglu et al (2011) provide an extensive reply to Albouy's comments. 12 Estimates of mortality rates correspond to potential settler mortality, measured in terms of deaths per annum per 1,000 “mean strength” (raw mortality numbers are adjusted to what they would be if a force of 1,000 living people were kept in place for a whole year, e.g., it is possible for this number to exceed 1,000 in episodes of extreme mortality as those who die are replaced by new arrivals). For more details on the construction of the mortality rate index, see Acemoglu et al. (2001). 8 implies that in ex-colonies nearly one-quarter of European settlers would die per annum due to unfavourable disease environments.13 3. Empirical findings In this section we present the empirical results of the methodology developed in the previous section. For comparison reasons, we also report results obtained from OLS and 2SLS regressions. As a benchmark we use ordinary aid data (first four columns of Table 2) to assess the growth impact of aid flows. Although this dataset is not purged from endogeneity, we nevertheless report findings to allow comparison with the relevant literature. Column (1) presents estimates obtained from a pooled OLS regression that confirm some standard results of the literature. The control variables appear with expected coefficients, although the majority of them are statistically insignificant. In particular, institutional quality, budget surplus and trade openness are positively signed as expected, but only the latter variable is statistically significant at 1% level. Initial income appears with a negative and statistically significant coefficient, thus confirming the standard convergence hypothesis. In accordance with Dalgaard et al. (2004), inflation and the share of a country’s land located in the tropics enter with negatively signed coefficients, although both are statistically insignificant. Interestingly, for our purposes, the coefficient on aid is negative and statistically insignificant, thus confirming the broad picture from the empirical literature on the aid-growth nexus. To detect any non-linearities in the aid-growth relationship we follow the standard empirical strategy and we augment the linear growth regression of column (1) by adding an aid-squared term. The inclusion of this additional variable improves OLS estimation while leaving intact the effects of the control variables (see column 2 of Table 2). Aid now exerts a statistically significant negative effect on growth, but this adverse effect is progressively weakened at higher aid levels, as indicated by the statistically significant positive coefficient of the squared term. Thus, when using ordinary aid data one finds significant evidence that the marginal growth effect of aid is not uniform across recipients, but rather 13 The extreme value of this variable corresponds to Mali, for which estimated mortality rates exceed 1,000, i.e. all living settlers and also new-born settlers are expected to die in one-year period. Another case with estimated mortality rates above 1,000 is Gambia for which the index reaches 1470. 9 depends on the amount of aid received. Given the aforementioned evidence in favour of aid non-linearities, in columns (3) and (4) of Table 2 we estimate threshold regressions using the same dataset of ordinary aid data. The control variables have the expected signs and are now mostly significant. In particular, in major aid recipients (column 4) the fraction of land in tropics exerts a largely negative effect on growth whereas the budget surplus affects growth positively. For both subgroups of countries trade openness retains its positive significance and the standard convergence hypothesis is validated empirically. Regarding aid, the heteroskedasticity-consistent Lagrange-multiplier (LM) test (Hansen, 1996) reported in the lower part of Table 2 indicates that, when ordinary aid data are employed, there is no threshold of aid above which aid is beneficial for growth. The right panel of Table 2 highlights the main point of the paper. In particular, in columns (5)-(10) we report estimates using endogeneity-free aid data. First, column (5) reports the results obtained from the standard OLS regression. Again, the control variables appear with the expected signs and trade openness, budget surplus and initial income exert a statistically significant growth effect. The OLS coefficient on aid is statistically insignificant, although now it is positively signed. As in the case of ordinary aid data, we let a quadratic term of endogeneity-free aid enter an alternative growth regression in order to detect nonconstant marginal effects of aid (column 6). The results remain virtually the same, but the coefficient of aid becomes now negative (although insignificant), whereas the statistically significant positive coefficient of the squared term indicates that the reverse growth effect of aid is moderated at higher aid levels. In light of this evidence, we move on to threshold regression estimation. The picture here changes starkly: there is an aid threshold below which the growth impact of aid is negative and statistically insignificant, but above which it becomes positive and statistically significant. This result provides prima facie evidence that the adverse growth impact of aid typically reported in the literature is driven by low aid flows, as in the present analysis it is evidently reversed in higher aid levels. In columns (9) and (10) of Table 2 we perform a similar exercise using ten-year averages to test the robustness of the aforementioned results in the medium-run horizon. Our unbalanced sample of 42 countries now consists of 114 observations where 10 endogeneity-free aid data are obtained following the same instrumentation methodology.14 The main picture survives and is now in fact more striking. The estimated aid coefficient above the threshold remains positive and statistically significant, whereas the coefficient below the threshold is found negative and statistically significant. In accordance with these findings, the values of the LM test indicate that one can safely reject the null hypothesis of no threshold at 1% significance level. Thus, we obtain significant evidence in favor of the existence of an aid threshold effect when both cross-sectional and panel data are employed and when endogeneity of aid is controlled for.15 Regarding the rest of the controls, we find that, as in column (1), the fraction of land in the tropics turns out a growth deterrent in major aid recipients, whereas budget surplus exerts an adverse effect for countries below the aid threshold and a positive effect for those above the threshold. We next address the potential endogeneity of institutions discussed in the previous section by using settler mortality data as instruments. Due to data availability our sample is now reduced to 36 observations.16 Column (1) in Table 3 reports results obtained via Two-Stage Least Squares estimation. Evidently, the instrumentation of institutions does not affect the growth impact of aid. The coefficient of aid turns out negative and insignificant, validating the findings of the literature. However, when we account for the presence of a threshold in the growth effect of aid the picture is again different (columns 2 and 3). There is a threshold for aid above which the growth impact of aid is positive and statistically significant, whereas it is insignificant below the threshold.17 Thus, we confirm that the existence of an aid threshold is not affected by the endogeneity of institutions. We also address the endogeneity of institutions using 10-year averages, which reduces our sample to 101 observations. Estimation results are reported in 14 We do not report estimation results of the corresponding OLS regressions, since they are similar to those reported in columns (5) and (6). 15 In order to test and correct for potential threshold endogeneity and omitted-variable bias we use Heckman’s correction method. In the lower part of Table 2 we provide estimates of the Inverse Mills ratio (lambda coefficient), as well as the corresponding probability levels for each set of threshold regressions, where the null hypothesis is that the aid threshold is exogenous. As can be readily seen, the lambda coefficients are always statistically insignificant, thereby validating our finding that the growth impact of aid turns unambiguously positive at above-threshold levels. 16 Botswana, Guinea-Bissau, Malawi, Philippines, Thailand, and Yemen are excluded due to missing observations. 17 The LM test for no threshold is not applicable in the case of endogenous regressors. We also note that the budget deficit is not included in specifications (2)-(3) because it is highly correlated with institutional quality in high aid recipients, which causes numerical problems in the cross-section regression with endogenous institutions. 11 columns (4) and (5) of Table 3. Although the threshold regressions indicate that the coefficient of aid above the threshold is now insignificant, its sign changes again from significantly negative to positive above the threshold. As a final step, to eliminate the possibility of omitted-variable bias and to test the robustness of our results to the inclusion of additional explanatory variables, we follow the recent study by Alvi et al. (2008) and we augment the model by first adding ethnic fractionalization to the benchmark regression. Ethnic fractionalization denotes the probability that two individuals will not belong to the same ethnic group and data correspond to 1960 values as provided by Easterly and Levine (1997). Columns (1)-(2) and (5)-(6) of Table 4 report the results for both exogenous and endogenous institutions. Ethnic fractionalization has a negative effect on growth in high aid-recipient countries. Our main finding persists across the estimated specifications. The coefficient on aid is found to be negative below the threshold and significantly positive above the threshold. Following Alvi et al. (2008), we also replicate estimation using money supply as a share of GDP as a control variable (source: World Bank). Again the regressions corroborate our evidence on the positive growth impact of aid for high aid recipients.18 We close the presentation of our empirical findings by noting that their main implication is that there is a threshold of aid flows, above which their growth impact becomes significantly positive. We stress, however, that when endogeneity-free aid data are employed, the reported threshold parameters do not correspond to actual aid disbursements. Still, one can draw valid inference about the minimum amount of EDA flows needed to make aid work by simply classifying countries into below- and above- threshold groups. Thus, after controlling for endogeneity of aid and institutions (columns 5-6 of Table 4), we find that Effective Development Assistance should exceed 3.4% of a recipient’s GDP, in order to boost domestic growth. Countries exceeding this threshold level (Burkina Faso, Gambia, Guinea-Bissau, Mali, 18 We also experimented with the following additional variables. We introduced assassinations and an interaction term with ethnic fractionalization, but both variables turned out insignificant, whereas the coefficient of aid was not substantially affected. Also, following Clemens et al. (2004) we augmented the model by adding the logarithm of life expectancy at birth in 1970. However, the estimation of this specification generated numerical problems because, although initial log life expectancy is not very strongly correlated with initial log GDP (the correlation coefficient is around 0.7), the correlation between these variables reaches 0.95 for high aid levels. 12 Malawi, Niger, Nicaragua) have benefited from aid during the time period under investigation by enjoying higher growth rates. By contrast, in all remaining recipients aid has had a negligible growth effect. 4. Concluding remarks One major issue in international development is the failure of aid to boost growth in recipient countries. The negative or, at best, insignificant growth effect of aid supported by the majority of studies lies in the central assumption that the relationship between aid and growth is uniform across countries. Using a datadriven threshold regression approach, this paper aimed at investigating whether the growth impact of aid changes beyond a critical threshold. We showed that in standard OLS and Two-Stage Least Squares regressions aid is found to be ineffective in enhancing growth in recipient countries. However, when a threshold regression approach is used we found that high aid flows affect growth positively. We close the paper with a word of caution. Given that most recipients are classified below the estimated threshold, our evidence provides some indication why aid flows have so far been insufficient in terms of exerting a significantly positive effect on the growth rate of recipients. Hence, the evidence seems to favor the view that a substantial increase of aid flows is required for making aid work. Nevertheless, the present approach cannot identify the generating mechanisms and channels through which this growth effect takes place, but only aims at highlighting a robust empirical fact that warrants further exploration. Recently, Ouattara and Strobl (2008) have shown that the negative growth effect of aid comes mainly from financial program aid, whereas project aid affects growth positively but with diminishing returns. Minoiu and Reddy (2010) find that developmental aid has a positive and large effect on growth, while non-developmental aid is mostly growth-neutral. In this spirit, investigating the role of various aid forms on growth through a more refined analysis warrants further investigation. 13 References Acemoglu D., S. Johnson and J. A. Robinson (2001), “The colonial origins of comparative development: an empirical investigation”, American Economic Review 91(5): 1369-1401. Acemoglu D., S. Johnson and J. A. Robinson (2011), “Hither thou shalt come, but no further - Reply to The colonial origins of comparative development: an empirical investigation”, American Economic Review forthcoming. Albouy D. (2011), “The colonial origins of comparative development: an empirical investigation: comment”, American Economic Review forthcoming. Alvi E., D. Mukherjee and E.K. Shukralla (2008), “Aid, policies, and growth in developing countries: a new look at the empirics”, Southern Economic Journal, 74(3): 693-706. Arndt C., S. Jones and F. Tarp (2010), “Aid, growth and development: have we come full circle?”, Journal of Globalization and Development, 1(2): Article 5. Azariadis C. and J. Stachurski (2005), “Poverty traps”, in P. Aghion and S. Durlauf (eds.), Handbook of Economic Growth, North Holland: Amsterdam. Burnside C. and Dollar (2000), “Aid, policies and growth”, American Economic Review 90(4): 847-868. Caner M. and B. Hansen (2004), “Instrumental variable estimation of a threshold model”, Econometric Theory 20(5): 813-843. Chan K.S. (1993), “Consistency and limiting distribution of the least squares estimator of a threshold autoregressive model”, The Annals of Statistics 21(1): 520-533. Chang C., E. Fernandez-Arias and L. Serven (1998), “Measuring aid flows: a new approach”, World Bank Policy Working Paper No 2050. Clemens M.A., R. Bhavnani and S. Radelet (2004) “Counting chickens when they hatch: The short-term effect of aid on growth”, Center for Global Development Working Paper 44: Washington. Commission for Africa (2005), Our Common Interest: Report Of The Commission For Africa, http://www.commissionforafrica.org. Dalgaard C. and H. Hansen (2001), “On aid, growth and good policies”, Journal of Development Studies 14 37(6): 17–41. Dalgaard C., H. Hansen and F. Tarp (2004), “On the empirics of foreign aid and growth”, Economic Journal 114(496): F191–216. Devarajan S., M. Miller and Eric V. Swanson (2002), “Goals for development: history, prospects and costs”, World Bank Policy Research Working Paper No. 2819. Doucougliagos H. and M. Paldam (2008), “Aid effectiveness on growth: a meta study”, European Journal of Political Economy 24(1), 1–24. Easterly W. and R. Levine (1997), “Africa’s growth tragedy: policies and ethnic divisions”, Quarterly Journal of Economics 112(4): 1203–50. Easterly W. (2006), “Reliving the 1950s: the big push, poverty traps, and takeoffs in economic development”, Journal of Economic Growth, 11: 289–318 Easterly W., R. Levine and D. Roodman (2004), “Aid, policies and growth: a comment”, American Economic Review 94(3): 774-780. Economides G., S. Kalyvitis and A. Philippopoulos (2008), “Does foreign aid distort incentives and hurt growth? Theory and evidence from 75 aid-recipient countries”, Public Choice 134(3), 463-488. European Commission (2005), EU Strategy for Africa: Towards a Euro-African pact to accelerate Africa’s development, European Parliament and the European Economic and Social Committee: Brussels. Gallup J.L. and J.D. Sachs (1999), “Geography and economic development”, in: B. Pleskovic and Stiglitz J.E. (eds.), Annual World Bank Conference on Development Economics, World Bank: Washington DC. Hadjimichael M.T., D. Ghura, M. Mühleisen, R. Nord, and E. M. Uçer (1995), “Sub-Saharan Africa: growth, savings, and investment, 1986-93”, Occasional Paper 118, International Monetary Fund. Hansen B.E. (1996), “Inference when a nuisance parameter is not identified under the null hypothesis”, Econometrica 64(2), 413-430. Hansen B.E. (2000), “Sample splitting and threshold estimation”, Econometrica 68(3): 575–604. 15 Hansen H. and F. Tarp (2000), “Aid effectiveness disputed”, Journal of International Development 12(3): 375–98. Hansen H. and F. Tarp (2001), “Aid and growth regressions”, Journal of Development Economics 64(2): 547–70. Herzer D. and O. Morrissey (2009), “The long-run effect of aid on domestic output”, CREDIT Research Paper 09/01, University of Nottingham. Heston A, R. Summers and B. Aten (2006), Penn World Tables Version 6.2, Center for International Comparisons of Production, Income and Prices, University of Pennsylvania. Kanbur R. (2006), “The economics of international aid”, in S. Kolm and J.M. Ythier (eds.), Handbook of the Economics of Giving, Altruism and Reciprocity, Elsevier: Amsterdam. Knack S. and P. Keefer (1995), “Institutions and economic performance: cross-country tests using alternative institutional measures”, Economics and Politics 7(3): 207-227. Kraay A. and C. Raddatz (2007), “Poverty traps, aid, and growth”, Journal of Development Economics, 82: 315–347. Kourtellos A, C.M. Tan, X. Zhang (2007), “Is the relationship between aid and economic growth nonlinear?”, Journal of Macroeconomics 29(3): 515–540. Lensink R. and H. White (2001), “Are there negative returns to aid?”, Journal of Development Studies 37(6): 42–65. Masanjala W.H. and C. Papageorgiou (2004), “The Solow model with CES technology: nonlinearities and parameter heterogeneity”, Journal of Applied Econometrics 19(2): 171-201. McGillivray M., S. Feeny, N. Hermes and R. Lensink (2006), “Controversies over the impact of development aid: it works; it doesn't; it can, but that depends...”, Journal of International Development 18(7): 1031-1050. Minoiu C. and S.G. Reddy (2010), “Development aid and economic growth: A positive long-run relation”, Quarterly Review of Economics and Finance, 50(1): 27-39. Ouattara B. and E. Strobl (2008), “Aid, policy and growth: does aid modality matter?”, Review of World 16 Economics, 144(2): 347-365. Pritchett L. (2000), “Understanding patterns of economic growth: searching for Hills among Plateaus, Mountains and Plains”, World Bank Economic Review 14(2): 221-250. Rajan R.G. and A. Subramanian (2008), “Aid and growth: what does the cross-country evidence really show?”, Review of Economics and Statistics 90(4): 643-665. Roodman D. (2007), “The anarchy of numbers: aid, development, and cross-country empirics”, World Bank Economic Review 21(2): 255-277. Rostow W. (1960), The Stages of Economic Growth: A Noncommunist Manifesto, Cambridge University Press: Cambridge. Sachs J. (2005), Investing in Development: A Practical Plan to Achieve the Millenium Development Goals, UN Millenium Project: New York. Sachs J. and A. Warner (1995), “Economic reform and the process of global integration”, Brookings Papers on Economic Activity 1: 1-118. Sachs J., J.W. McArthur, G. Schmidt-Traub, M. Kruk, C. Bahadur, M. Faye, G. McCord (2004), “Ending Africa's poverty trap”, Brookings Papers on Economic Activity 1:117-240. United Nations (2006), Doubling Aid: Making the ‘Big Push’ Work, New York. Wacziarg R. and K.H. Welch (2008), “Trade liberalization and growth: new evidence”, The World Bank Economic Review, 22(2): 187-231. Zedillo E., A.Y Al-Hamad, D. Bryer D, M. Chinery-Hesse, J. Delors, R. Grynspan, A.Y. Livshit, A.M. Osman, R. Rubin, M. Singh and M. Son (2001), Recommendations of the Highlevel Panel on Financing for Development, UN General Assembly, New York: United Nations. 17 Table 1. Descriptive statistics (42 aid-recipient countries, 1970-2000) Mean Std. dev. Minimum Maximum Real per capita GDP growth 1.39 1.91 -1.93 7.13 Aid (EDA) 1.42 1.50 -0.01 6.73 Aid (endogeneity-free data) 4.77 5.10 -4.39 29.99 Institutional quality 4.62 1.53 2.50 8.94 Fraction of land in tropics 0.90 0.25 0.04 1.00 Budget surplus -0.04 0.04 -0.22 0.05 Initial GDP per capita (log) 7.26 0.74 5.69 8.96 Inflation 0.20 0.20 0.03 0.91 Sachs-Warner openness 0.45 0.26 0.13 1.00 Settler mortality rates* 284.42 533.38 15.50 2940.00 Note: * Descriptive statistics for settler mortality rates correspond to a sub-sample of 36 aidrecipients for which data are available. 18 Table 2. Growth OLS and aid threshold regressions ordinary aid endogeneity-free aid OLS TR OLS ≤ 1.37 TR >1.37 TR >3.32 ≤3.32 CI = [1.29, 1.87] CI = [2.81, 6.48] ≤ 6.18 > 6.18 CI = [5.32, 8.10] (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Aid -0.33 (0.32) -1.20** (0.50) -1.21** (0.53) 0.15 (0.29) 0.05 (0.05) -0.05 (0.07) -0.06 (0.06) 0.25*** (0.06) -0.26*** (0.09) 0.08** (0.04) Institutional quality 0.20 (0.14) 0.18 (0.14) 0.35*** (0.11) -0.02 (0.32) 0.11 (0.16) 0.15 (0.16) -0.18* (0.11) 0.02 (0.21) 0.10 (0.16) 1.06*** (0.29) Fraction of land in tropics -1.12 (1.19) -1.13 (1.35) 1.06* (0.60) -7.26*** (1.58) -1.51 (1.22) -1.37 (1.28) -0.71 (0.44) -1.50 (2.06) -0.83 (0.85) -19.51*** (5.53) Budget surplus 7.86 (10.50) 13.01 (10.68) 3.86 (6.21) 18.30* (10.16) 17.26** (8.64) -0.08 (0.36) 0.99** (0.42) Initial GDP per capita (log) -0.93** (0.40) -1.31*** -1.28*** -2.54*** (0.35) (0.41) (0.91) Dependent variable: per capita growth Inflation -0.45 (1.26) Sachs-Warner openness 3.68*** 3.19*** (1.03) (0.93) -0.56* (0.30) -0.59* (0.31) -0.15 (0.10) -0.79** (0.37) -2.61 (7.19) -11.24 (9.40) -0.62 (0.87) 3.78 (2.50) -0.41 (1.53) -0.16 (1.60) -0.55 (0.41) 2.12 (1.61) -2.73*** (0.63) -3.84*** (0.85) 2.44*** (0.86) 4.28*** (1.33) 4.08*** (0.96) 3.61*** (0.94) 5.75*** (0.49) 3.16*** (0.82) 0.71 (0.48) -0.84 (0.73) 0.16** (0.07) Aid Squared R-squared -0.19 (1.14) 22.22** -29.38*** 41.17*** (10.65) (4.50) (10.71) 0.01* (0.00) 0.54 0.58 - - 0.53 0.55 - - 7.69 (0.00) 12.72 (0.00) - - 5.81 (0.00) 5.94 (0.00) - - - - 0.25 - - 0.00 0.01 42 42 42 42 Inverse Mill’s ratio (Prob) - - 42 2.32 (0.39) - - 42 0.30 (0.87) 42 -1.79 (0.42) No. of observations 42 42 F-statistic on joint-significance (Prob) LM test for no threshold: Bootstrap P-value No of countries 24 18 42 16 26 78 36 Notes: All regressions include a constant. Robust standard errors are in parentheses. *** denotes significance at 1%, ** at 5%, and * at 10%. In regressions (1)-(4) aid data correspond to Effective Development Assistance (EDA), while in regressions (5)-(10) endogeneity-free aid data are employed as in Rajan and Subramanian (2008). Variables in columns (1)-(8) are time-averages for 1970-2000 and variables in columns (9)-(10) correspond to 10-year averages. TR and CI denote Hansen (2000) Threshold Regression and the 95% Confidence Interval, respectively. For the heteroscedasticity-consistent Lagrange-multiplier (LM) test for no threshold the null hypothesis is that there is no threshold effect (Hansen, 1996). 19 Table 3. Growth 2SLS and aid threshold regressions: endogenous institutions 2SLS TR ≤ 5.41 TR > 5.41 CI = [5.28, 6.49]] Dependent variable: Real per capita GDP growth ≤ 4.87 > 4.87 CI = [0.14, 6.00] (1) (2) (3) (4) (5) Endogeneity-free Aid -0.17 (0.15) -0.15 (0.11) 0.38** (0.16) -0.29** (0.13) 0.74 (0.69) Institutional quality 1.53 (1.06) 1.00 (2.75) 1.55*** (0.30) 0.69 (0.79) 3.84 (3.96) Fraction of land in tropics 1.14 (2.16) -0.02 (3.50) -59.41*** (15.88) 0.03 (0.98) -28.68 (24.86) Budget surplus -7.50 (20.57) - - -12.60 (8.06) -48.63 (63.15) Initial GDP per capita (log) -1.06 (0.69) -1.53 (2.55) 0.36 (0.31) -0.55 (0.84) -0.17 (1.15) Inflation 0.12 (1.45) 1.73 (8.13) -4.71*** (1.00) -2.95*** (0.65) -5.68 (3.74) Sachs-Warner openness 1.82 (2.15) 3.03 (2.37) -2.33 (1.89) 0.43 (0.72) -0.96 (2.26) F-statistic on jointsignificance (Prob) 1.68 (0.16) - - No of countries 36 36 36 No. of observations 36 15 21 58 43 Notes: For the Two-Stage Least Squares estimation heteroscendasticity-consistent robust standard errors are reported. In columns (2)-(5) the Caner and Hansen (2004) regressions are used and a heteroskedasticity corrected asymptotic 95% confidence interval for the threshold estimate is computed using a quadratic polynomial as in Hansen (2000). Institutional quality is instrumented using log settler mortality as in Acemoglu et al. (2001). See also Table 2. 20 Table 4. Aid threshold regressions: robustness tests exogenous institutions ≤ 5.76 > 5.76 CI = [2.82, 6.49] Dependent variable: Real per capita GDP growth ≤ 3.32 (2) (3) -0.18*** (0.07) 0.43* (0.25) -0.08 (0.06) Institutional quality -0.03 (0.13) 0.16 (0.42) Fraction of land in tropics -1.20* (0.73) -10.76*** (3.18) Budget surplus -2.41 (10.04) Initial GDP per capita (log) -0.51 (0.35) -0.48* (0.25) -0.12 (0.10) Inflation -1.92 (1.34) -1.13 (2.04) Sachs-Warner openness 3.25*** (0.85) 1.56 (1.86) Ethnic fractionalization 0.25 (0.74) -3.70*** (0.45) (4) No of countries 28 > 6.37 CI = [5.28, 6.49] ≤ 6.37 > 6.37 CI = [5.28, 6.49] (6) (7) (8) 0.27*** (0.06) -0.16 (0.24) 0.58*** (0.15) -0.22 (0.92) 0.44 (0.29) -0.21* (0.11) -0.05 (0.25) 1.61 (6.13) 0.67* (0.36) 2.02 (20.26) 1.18 (0.79) -0.48 (0.42) -2.10 (2.37) 0.92 (8.29) -70.47*** (9.38) 0.89 (19.21) -69.11*** (14.29) - - - - -1.00** (0.45) -2.39 (5.84) 0.59 (0.41) -2.14 (13.60) 0.12 (1.20) 0.06 (0.60) 2.59 (2.03) 4.64 (22.44) -6.11*** (0.96) 4.56 (55.49) -4.19*** (0.46) 5.68*** (0.49) 2.41* (1.33) 3.38 (2.48) 1.93 (2.39) 2.56 (10.53) -4.38 (8.62) -2.52 (7.31) -3.42** (1.53) -0.02 (0.33) 0.08 (0.07) 0.02 (0.01) LM test for no threshold: Bootstrap P-value ≤ 6.37 (5) 35.67*** -26.91*** 42.36*** (12.76) (2.67) (10.50) M2 No. of observations > 3.32 CI = [2.82, 6.49] (1) Endogeneity-free Aid endogenous institutions 0.03 (0.04) 0.09 0.01 - - 41 42 36 36 13 16 Notes: See Tables 2 and 3. 21 26 25 11 25 11 Figure 1. Conditional relationship between Actual and Fitted Aid, 1970-2000 Residuals of Actual Aid (% of GDP) 2 NER NIC BWA MLI TTO GMB 1 VEN SLV YEM HND BFA BOL MDG GHA EGY CIV GTM PER JAM UGA CMR ZAF COL PHL ECU THA SLE MYS GNB SGP CHL MWI 0 MEX CRI -1 IDN PRY COG DOM KEN ETH LKA BRA -2 -5 0 5 10 Residuals of Fitted Aid (% of GDP) 15 Note: The figure plots the first-stage relationship between actual and fitted aid, conditional on all the covariates that enter the second-stage growth regression. 22
© Copyright 2026 Paperzz