





















Study with the several resources on Docsity
Earn points by helping other students or get them with a premium plan
Prepare for your exams
Study with the several resources on Docsity
Earn points to download
Earn points by helping other students or get them with a premium plan
Community
Ask the community for help and clear up your study doubts
Discover the best universities in your country according to Docsity users
Free resources
Download our free guides on studying techniques, anxiety management strategies, and thesis advice from Docsity tutors
Aid remains a key tool for enhancing the development prospects of poor countries. KEYWORDS: foreign aid, growth, aid effectiveness, causal effects. Author Notes ...
Typology: Study notes
1 / 29
This page cannot be seen from the preview
Don't miss anything!
Volume 1, Issue 2 2010 Article 5
Recommended Citation: Arndt, Channing; Jones, Sam; and Tarp, Finn (2010) "Aid, Growth, and Development: Have We Come Full Circle?," Journal of Globalization and Development : Vol. 1: Iss. 2, Article 5.
DOI: 10.2202/1948-1837.
Channing Arndt, Sam Jones, and Finn Tarp
Abstract
The micro-macro paradox has been revived. Despite broadly positive evaluations at the micro- and meso-levels, recent literature doubts the ability of foreign aid to foster economic growth and development. This paper assesses the aid-growth literature and, taking inspiration from the program evaluation literature, we re-examine key hypotheses. In our findings, aid has a positive and statistically significant causal effect on growth over the long run, with confidence intervals conforming to levels suggested by growth theory. Aid remains a key tool for enhancing the development prospects of poor countries.
KEYWORDS: foreign aid, growth, aid effectiveness, causal effects
Author Notes: We have benefited greatly from editorial advice and a careful anonymous referee report. We also thank Tony Addison, Ernest Aryeetey, Pranab Bardhan, Bruce Bolnick, Imed Drine, Jan Willem Gunning, Gerry Helleiner, Paul Isenman, Homi Kharas, Dirk Krueger, David Roodman, Erik Thorbecke, Alan Winters, and Adrian Wood for encouragement and most valuable comments. The same goes for the participants at conferences and seminars held by the African Economic Research Consortium (AERC), the Bergen Resource Center for International Development at CMI Norway, the Brookings Institution, central ministries in Mozambique, Tanzania and Vietnam, Cornell University, the European Union Development Network (EUDN) (organized by CERDI, Clermont-Ferrand, France), the Helsinki Center for Economic Research (HECER), NORAD Norway, OECD Paris, the United Nations HQ (organized by UNU-ONY), the University of Copenhagen, the University of Ghana, the UNU Conference of Directors (CONDIR), and UNU-WIDER. Thanks are also due to Tseday Jemaneh Mekasha for excellent research assistance and to Raghuram G. Rajan and Arvind Subramanian for sharing their original data and STATA files. The usual caveats apply.
Scholarship on the relationship between aid, growth, and development is voluminous. This section provides a brief survey of how the aid-growth debate has evolved.
2.1. Earlier generations
Studies of the aid-growth relationship from the 1970s until recently have been classified into three generations, each influenced by dominant theoretical paradigms as well as available empirical tools (Hansen and Tarp, 2000). The first two generations were inspired by simple models of the growth process, i.e. the Harrod-Domar model and the two-gap Chenery-Strout extension. The underlying idea behind the Harrod-Domar model is of a stable linear relationship between growth and investment in physical capital. Assuming all aid is invested, it is straightforward to calculate how much aid is required to achieve a target growth rate. The impact of aid is positive and helps plug either a savings or a foreign exchange gap. Empirical studies in this tradition consequently focused on the extent to which aid increased savings and investment in recipient countries (Papanek, 1972, 1973). Overall, first generation studies show that aid tends to increase total savings, but not by as much as the aid flow. Quite reasonably, this suggests a non-negligible proportion of aid is consumed rather than invested. Retaining the focus on capital accumulation, a second generation of literature explored the impact of aid on growth via investment. Using data for a cross section of countries, a large number of studies of this kind were produced during the 1980s and early 1990s. These studies consistently pointed to a positive link between aid and investment. While a majority of the aid-growth studies of this generation also suggested a positive impact, the result that captured attention was Paul Mosley’s micro-macro paradox. An influential line of critique of the Harrod-Domar and two-gap approach was the argument that growth is less related to physical capital investment than often assumed (Easterly, 1999). If the productive impact of aid depends more on incentives and relative prices, as well as the policy environment more generally, then it becomes important to consider these broader effects. The second generation of studies also introduced the problem that poorly performing countries may receive more aid precisely because of their poor growth performance. Empirical analyses that do not account for the endogeneity of aid will not reveal aid’s causal impact. Most second generation studies, however, did not deal with this issue. From the early 1990s a third generation of more sophisticated econometric studies came to dominate the discourse about aid. This was motivated by the availability of panel data, allowing analysts to look at changes both across and
2
Journal of Globalization and Development, Vol. 1 [2010], Iss. 2, Art. 5
DOI: 10.2202/1948-1837.
within countries over time. Insights from new theories of economic growth also influenced the research agenda. Mindful of the weaknesses of previous studies, the aid-growth relationship came to be perceived as (possibly) non-linear and the endogeneity of aid was taken more seriously. Among the numerous studies of this generation, the contribution by Burnside and Dollar (2000) came to exert a significant influence on policy. These authors made an argument for conditional aid effectiveness, specifically: “aid has a positive impact on growth in developing countries with good fiscal, monetary and trade policies ... [but] ... in the presence of poor policies, aid has no positive effect on growth” (2000: 847). However, these results were subject to criticism. Hansen and Tarp (2001) found that a story of diminishing returns to aid best captured the non-linear relationship between aid and growth. In a later contribution, Easterly et al. (2004) added that the Burnside-Dollar aid-policy result is fragile when the dataset is expanded (by years and countries). Dalgaard et al. (2004) found that aid has been less effective in tropical areas over the last 30 years and called for further research to help disentangle the channels through which aid matters for productivity. In an empirical review of these contributions, Roodman (2007) argued that the results of this generation are extremely sensitive to methodological choices, concluding that while some aid is likely to increase investment and growth, aid “is probably not a fundamentally decisive factor for development” (2007: 275).
2.2. Recent studies
More recently, a fourth generation of literature has emerged. A distinctive aspect of this generation is the view that aid’s aggregate impact on economic growth is non-existent. A leading paper that appears to establish this result is Rajan and Subramanian (2008). They find no systematic effect of aid on growth regardless of the estimation approach, the time period. and the type of aid. Explanations for non-positive aggregate effects of aid often refer to political economy dynamics. For example, Djankov et al. (2008) argue that aid has effects that are analogous to a natural resource curse. Similarly, Rajan and Subramanian (2007) find that the rate of growth of value added by the manufacturing sector in developing countries has been undermined by a detrimental effect of aid inflows on governance quality. Fourth generation scholars have also become increasingly skeptical about our ability to make valid causal inferences with respect to complex aggregate phenomena, such as the determinants of economic growth. In particular, previous methods used to deal with endogeneity have been subject to criticism. There is increasing awareness that dynamic panel (system) GMM methods – frequently employed in the third generation – are not a panacea. The concern that weak instruments typically bias coefficient estimates towards their unadjusted counterparts (e.g., OLS or panel fixed effects estimates) applies as much to panel GMM as to cross-section estimators. Bun and Windmeijer (2010) show that the
3
Arndt et al.: Aid, Growth, and Development: Have We Come Full Circle?
on earnings. Both problems are likely to be characterized by endogenous selection, heterogeneous treatment responses, and mis-measurement of treatment input (both in terms of quality and quantity). Considerable effort has been expended in the analysis of large, high-quality, schooling datasets by some of the most skilled econometricians in the profession. Even so, debate has persisted, at least until recently, with respect to the net bias of ordinary least squares (OLS) estimates of returns to education (Card, 2001). If the profession has experienced serious difficulties estimating the causal effect of schooling on earnings in developed countries, then it should not be surprising that estimating the impact of aid on growth in developing countries is contentious. However, it is difficult to deny that the aid-growth issue is both compelling and relevant. In developed countries, policy-makers and the wider public continue to ask whether aid is a cost effective use of taxpayer money on aggregate. Today, the attention of both the aid community and decision-makers is on “Dead Aid” (Moyo, 2009), which argues for a complete cessation of aid flows to Africa. We note that the financial crisis of 2008/09 has highlighted the importance of public spending to stabilize and stimulate economic activity. While foreign aid has multiple objectives, economic growth is central among them. If the economics profession as a whole were to abandon the question of aid’s impact on growth, it would leave the issue open to speculative and potentially unhelpful contributions.
2.4. Formulating an appropriate prior
A key aim of empirical analysis is to falsify or discriminate between competing hypotheses. Consequently, it is necessary to make explicit the prior upon which empirical testing is focused. With respect to the effect of foreign aid on economic growth, relatively few studies address the issue of an appropriate prior. A recent exception is Rajan and Subramanian (2008) who consider aid in a standard neoclassical growth model. Assuming that aid only augments physical capital investment and has no effect on productivity, they derive that a one percentage point increase in the ratio of aid to GDP should be expected to raise the growth rate of per capita GDP by around 0.16 percentage points on average. In practice, at least some aid is directed towards consumption or non-growth enhancing activities. As a result, Rajan and Subramanian place the expected growth return at around 0.1 percentage point for each percentage point of aid in GDP. Thus, the implied increase in the growth rate accruing from aid inflows at 10% of GDP should be about 1%, which is considerably less than the predictions based on Harrod-Domar models. In sum, growth theory points towards more modest expectations with respect to the potency of aid (see also Dalgaard and Erickson, 2009).
5
Arndt et al.: Aid, Growth, and Development: Have We Come Full Circle?
A related issue is the appropriate time-frame over which any growth effects accruing from aid can be expected to materialize. Various factors may exert a cumulative but not immediate impact on the rate of income growth. For example, changes in education move only slowly at the aggregate level and have a positive influence on economic growth with a substantial lag. This follows from simple demographics whereby improvements in schooling indicators can take many years to translate into noticeable increases in average education levels among working-age adults. Changes in human capital due to improved health indicators may take even longer to translate into more rapid economic growth. Ashraf et al. (2008) and Acemoglu and Johnson (2007) find that the initial economic impact of gains in life expectancy from the health interventions introduced from the 1940s may be a reduction in per capita incomes due to the increase in population and dependency ratios. The former authors find that it can take 30 years or more for per capita incomes to return to pre-intervention levels. They also find that significant increases in life expectancy at birth only begin to have a modest positive effect on incomes after about a 35 year lag.^2 Overall, a series of considerations indicate that the aid-growth relationship is only likely to emerge over a long time-horizon. Many aid investments, such as in education, health, and institution-building are long term in nature; and growth theory indicates that the contribution of these investments to growth is likely to be relatively modest. When these observations are combined with the volatility of growth in most developing countries and the high degree of measurement error inherent in nearly all the variables of interest, relatively long time-frames would appear to be necessary to reliably detect the aid-growth relationship.
Following Temple’s (2010) recommendation to build explicitly on existing empirical work, our starting point for developing an appropriate empirical strategy is Rajan and Subramanian (2008) (henceforth RS08). RS08 provides a thoughtful and highly influential contribution that is widely understood to have established that aid has no impact on growth. However, a number of concerns question this fundamental conclusion. This section addresses these concerns and suggests suitable improvements. Specifically, Section 3.1 provides a brief summary of RS08’s approach and main results. Section 3.2 presents a detailed analysis of the validity of their instrumentation strategy. This motivates the development of an amended instrument in Section 3.3, while Sections 3.4 and 3. propose modifications to the specification and regression estimators. Once these
(^2) Ashraf et al. (2008) focus on demographic trends as a result of disease eradication. Productivity
effects, demand effects, and complementary policies may speed the realization of growth benefits from health gains.
6
Journal of Globalization and Development, Vol. 1 [2010], Iss. 2, Art. 5
DOI: 10.2202/1948-1837.
Table 1: Alternative models for aid and growth, 1970-
(I) (II) (III) (IV) (V) (VI) (VII) (VIII) OLS 2SLS IV-LIML IV-LIML IV-LIML IV-LIML IV-LIML IV-LIML Aid / GDP -0.08* 0.10 0.10 0.07 0.08 0.12 0.21* 0.19* (0.04) (0.07) (0.08) (0.06) (0.06) (0.10) (0.11) (0.10) Initial per cap. GDP -1.67*** -1.41*** -1.40*** -1.44*** -1.44*** -1.44*** -1.34*** -1.36*** (0.32) (0.43) (0.39) (0.37) (0.36) (0.30) (0.33) (0.33) Initial level of policy 2.28*** 2.14*** 2.13*** 2.16*** 2.15*** 2.28*** 2.29*** 2.28*** (0.47) (0.62) (0.56) (0.52) (0.53) (0.46) (0.52) (0.51) Initial life expectancy 0.02 0.08* 0.08** 0.07** 0.07** 0.04 0.05 0. (0.03) (0.04) (0.04) (0.03) (0.03) (0.04) (0.04) (0.04) Geography 0.39** 0.61** 0.62** 0.58*** 0.58*** 0.25 0.29 0. (0.18) (0.26) (0.24) (0.22) (0.22) (0.22) (0.24) (0.23) Excluded instruments 1 1 7 3 1 7 3 1 Regional dummies SSA, EA SSA, EA SSA, EA SSA, EA SSA, EA SSA, A, LA SSA, A, LA SSA, A, LA N 78 78 78 78 78 78 78 78 R-squared 0.70 0.59 0.58 0.62 0.61 0.69 0.65 0. Weak identification stat. - 31.6 5.50 10.40 33.41 4.32 5.06 14. Stock-Wright LM S stat. - - 11.85 3.05 2.12 18.17 5.40 5. (probability) 0.158 0.384 0.145 0.020 0.145 0. Hansen J stat. - - 7.72 0.29 - 11.88 0.31 - (probability) 0.358 0.865 0.104 0. significance level: * 10%; ** 5%; *** 1%
Notes: only selected variables reported; columns (I) and (II) replicate results in Rajan and Subramanian (2008; RS08); column (III) employs a full set of aggregate instruments (see Table 2) in place of a single generated instrument; column (IV) restricts the aggregate instrument set to the mean population ratio, the colony dummy and their interaction; column (V) uses only log initial (recipient) population as the instrument; columns (VI) to (VIII) replicate columns (III) to (V) but use a preferred set of conditioning variables (not shown, see Section 3.4); regional dummies are included as indicated, SSA = sub- Saharan Africa, EA = East Asia, A = Asia, LA = Latin America & Caribbean; weak identification statistic is the first stage partial-F statistic in col. (II), and the Kleibergen-Paap Wald F statistic elsewhere; initial policy refers to the Sachs-Warner trade policy index; geography refers to the average of the number of frost days and tropical land area; standard errors, given in parentheses, are robust to arbitrary heteroskedasticity; dependent variable is mean real growth rate; Aid/GDP is treated as endogenous in all models except column (I); in columns (I) to (V) Aid /GDP is taken from RS08, in columns (VI) to (VIII) it is re-estimated from OECD-DAC (2008) data treating possible missing values as zeroes (see Section 3.3). Source: authors’ estimates.
3.2. Instrument validity
The supply-side approach to instrumentation developed by RS08 represents the state of the art in the aid-growth literature. Nevertheless, it has been subject to criticism. Clemens and Bazzi (2009) note that different authors have used the same variables as exogenous instruments for a wide range of endogenous variables. This raises the possibility that these exogenous instruments are correlated with other omitted variables, thereby invalidating the exclusion restriction on which valid causal inference depends. They direct specific attention to the reliance of the RS08 (fitted) instrument on the natural logarithm of the aid recipient’s population size. They find that log population has a “statistically
8
Journal of Globalization and Development, Vol. 1 [2010], Iss. 2, Art. 5
DOI: 10.2202/1948-1837.
significant partial relationship with several variables that are plausible growth determinants” (2009: 11) and that are omitted from RS08’s specification. While the existence of a partial correlation between omitted explanatory variables and the chosen instrument indicates that the coefficients in a regression specification may be biased, the extent of bias is, ultimately, an empirical matter. This is recognized by Clemens and Bazzi (2009), leading them to advocate application of a range of empirical tests for instrument validity. On the face of it, straightforward validity checks of the RS08 (generated) instrument based on Sargan or Hansen tests are not possible because their IV model is just identified – i.e., the number of excluded instruments equals the number of endogenous variables. Nonetheless, recalling that the zero stage of the RS08 approach generates a single instrument as a linear combination of variables, it is possible to use modified versions of these same variables as excluded instruments directly in the aggregate aid-growth regressions. This provides for a large number of potential instruments, and therefore permits Hansen tests to be run either on the full set or on specific sub-sets of instruments. Following this logic, we collapse the bilateral aid dataset along the donor dimension, thereby transforming the explanatory variables used in the bilateral zero stage regressions for use at a more aggregate level. For continuous zero stage regressors, such as the donor-recipient population ratio, the corresponding “aggregate” instrument is the mean of the population ratio for each recipient across all donors. For dummy regressors, such as the specific colonizer, it is more appropriate to take the maximum value of the dummy for a given recipient (again, across all donors). Ignoring relatively minor variables, such as currently being a colony and the population-colony interaction terms employed in RS08, this yields a set of eight possible instruments as per the rows of Table 2. Column I of Table 2 verifies whether these aggregate instruments are adequate proxies for the fitted instruments generated from the zero stage regressions. As expected, the explanatory power is high. Moreover, underlining the contention of Clemens and Bazzi (2009), a driving force behind the fitted aid instruments appears to be the population ratio term. Thus, a fundamental issue for the RS08 instrumentation strategy is the validity of the exclusion restriction as it applies to the population-based instruments. Nevertheless, the results from column I of Table 2 indicate that other variables make some (albeit smaller) contribution to the overall fitted instrument. Thus, to further test instrument validity, we re-estimate the RS08 model employing the full set of eight aggregate instruments. The results, reported in column III of Table 1, closely replicate column II; and the Hansen J test reports a probability of 0.358, which fails to reject the validity of the exclusion restriction assumption.
9
Arndt et al.: Aid, Growth, and Development: Have We Come Full Circle?
ratio variables do not give cause for concern, providing comfort as to their suitability as exogenous instruments in these models.
3.3. Improved instrumentation strategy
The results of Section 3.2 indicate that the RS08 instrumentation approach is broadly convincing but is weakened by inclusion of suspect variables in the zero stage. From a theoretical point of view, the validity of using colonizer dummies and their interactions as instruments is questionable. The institutional transplants and broader colonizing strategies pursued by imperial powers were not alike, and they may have a persistent effect on income levels to the present day. This notion is at the heart of the debate concerning the effect of different legal origins (La Porta et al. , 2008), historical events (e.g., Nunn, 2008), and other institutional forms on contemporary economic outcomes. Put simply, the colonial relations variables are not orthogonal to growth and therefore should not be included in the zero stage regression explaining aid. As a first step towards improving the RS08 instrument, we re-run their aid-growth model using a smaller and “less suspect” sub-set of the aggregate instruments used in Section 3.2. These are the population ratio, a dummy for ever having been a colony and their interaction. Results are given in column IV of Table 1, showing that the Hansen J test is now passed with a high level of confidence. Nevertheless, compared against column II, the results also suggest a trade-off between efficiency and transparency in instrument selection. While the use of multiple aggregate instruments is more transparent, it does not exploit the full information about bilateral aid flows contained in the zero stage. This may be one reason why the weak identification statistics are considerably lower in column IV versus column II of Table 1. In fact, as shown in column V of Table 1, even if only one aggregate instrument is employed, namely the log of the recipient’s initial population, the strength of the instrument returns to similar values to those in column II and all coefficients are essentially unchanged.^3 Consequently, using a single instrument is likely to be more efficient but there are also potential information gains from employing a zero stage, especially in small aggregate samples such as those used in (static) cross-country regressions. Thus, to strengthen the instrumentation approach, we return to the zero stage regressions. Aside from removal of suspect terms, additional concerns motivate further modifications. First, there are errors in the calculation of average Aid/GDP in all stages of RS08’s regressions. The OECD-DAC aid dataset used for bilateral aid flows includes numerous missing values. While in some cases these genuinely refer to absent data, in most cases they represent unreported null
(^3) This result further underlines the reliance of the RS08 instrumentation strategy on population
size (also see Clemens and Bazzi, 2009).
11
Arndt et al.: Aid, Growth, and Development: Have We Come Full Circle?
values.^4 RS08 incorrectly treat these as missing. This is material because it distorts estimates for average bilateral aid flows over time. Consequently, it is necessary to re-estimate the bilateral aid variables and calculate period averages for Aid/GDP and aid per capita, setting missing entries to zero. This affects the dependent variable employed in the zero stage regression as well as the endogenous aid variable used in the IV estimations. Second, in the RS08 strategy, recipient GDP occurs in the denominator of the dependent variable in the zero stage regressions. Following Kronmal (1993), inappropriate use of ratio variables may lead to substantial misinterpretation (or bias) in least squares regressions. This may arise if the denominator of the dependent variable is correlated with the RHS variables independently of the numerator of the dependent variable. In the present case, this could arise if donor decision rules do not target the Aid/GDP ratio, and/or if there is a direct association between recipient GDP levels and population size or past colonial experiences. Third, it is apparent that individual donor countries exhibit distinct attitudes to giving foreign aid (Alesina and Dollar, 2000), which reflect cultural and historical factors. These time-invariant influences can be understood as fixed effects and may be included as RHS variables in the zero stage regression. Notably, and unlike the RS08 explanatory variables, these fixed effects may explain a part of the variation in aid allocations that is unrelated to purely strategic or political motives. As such, their inclusion may strengthen the overall validity and interpretation of the generated instrument. To address these concerns, we modify the RS08 specification of the zero stage regression. In place of Aid/GDP, we use aid per capita (Aid/POP) as the dependent variable which accords closely with the explicit aid allocation rules used by donors, such as the World Bank (see Annex 1 of IDA15, 2008).^5 We drop the colonizer-specific variables (and interactions) and only include a dummy for whether a country was ever a colony (COLONY). Adding donor-specific fixed effects (DONOR), our zero stage regression emerges as follows:
where the subscripts d and r represent donors and recipients respectively; CURCOL indicates whether the recipient is a current colony of the donor.
(^4) Confirmed in correspondence with the OECD DAC Secretariat. (^5) Note that in all subsequent regression stages the endogenous variable of interest remains
Aid/GDP.
12
Journal of Globalization and Development, Vol. 1 [2010], Iss. 2, Art. 5
DOI: 10.2202/1948-1837.
as per equation (1). Again, there is a small loss of explanatory power, but the population ratio and its interaction with the colony dummy remain highly significant. Also, the donor fixed effects (coefficients not shown in the table) vary in sign and many are significant. Overall, the RHS variables continue to explain a reasonable share of observed aid allocations. The existence of zero-value aid inflows points to a final possible weakness. In principle, the decision by a donor to provide aid involves at least two distinct choices (Tarp et al., 1999): (i) which recipients should receive aid; and (ii) how much to supply – i.e., the distribution of bilateral aid flows reflects an unobserved selection process. In the absence of an explicit model, one way to address potential bias from unobserved selection effects is to use Heckman’s correction (Heckman, 1979). Column V of Table 3 employs a Heckman selection model (estimated by full information maximum likelihood) to the specification in column IV, where the existence of zero or non-zero aid flows is used as the binary selection variable. Despite these changes, the direction of the results and their interpretation are largely unchanged. However, we reject the hypothesis that there is no selection bias. We therefore retain the Heckman estimator employed in Column V as our preferred zero stage regression.
3.4. Improved specification
Before presenting the results of the aid-growth IV regressions using the improved instrument, it is appropriate to discuss additional areas where the RS08 approach can be strengthened. The first of these is the choice of covariates. Given the relatively small sample available in the aggregate regressions (78 countries), inclusion of redundant variables may lead to a loss of efficiency and/or contribute to undesirable multicollinearity. In the case of RS08, we note that the three macroeconomic initial conditions (inflation, money supply, and budget balance) as well as ethnic fractionalization are insignificant in RS08’s cross-section outcome regressions for all periods. In addition, and as Wooldridge (2005) clarifies, inclusion of contemporaneous outcome variables – i.e., variables which may also be affected by the level of treatment – can invalidate the unconfoundedness assumption required for valid causal inference (Angrist and Pischke, 2008). This is pertinent as RS08’s chosen specification includes two variables that capture average outcomes during the period of analysis – institutional quality and the number of forced changes in the top government elite, labelled “revolutions”. Inclusion of these variables is puzzling in light of the literature which examines the effects of aid on growth through institutional performance. Controlling for such outcomes blocks potential channels through which aid may affect growth and thereby restricts the estimated coefficient on aid to a partial as opposed to a general effect. Such variables may also introduce unwanted reverse causality.
14
Journal of Globalization and Development, Vol. 1 [2010], Iss. 2, Art. 5
DOI: 10.2202/1948-1837.
It is also helpful to consider the appropriate role of regional fixed effects. In RS08’s specification, only East Asia and sub-Saharan Africa are included as regional dummy variables. This appears to be an ex post choice in the sense that prior to the 1980s there was no particular reason to identify these as “special”. Including regional dummy variables helps absorb intra-regional correlations and captures omitted spatial fixed effects such as those arising from geography, shared historical experiences, and trade relationships. A priori , a more plausible approach is to include a fuller set of regional dummies. Finally, it is appropriate to include additional variables that reflect initial socioeconomic conditions such as education and health indicators, as well as additional geographic characteristics such as trading distances. These variables are frequently seen as important determinants of growth and may also proxy for initial conditions; as such, they may explain some of the variation in the expected growth returns to aid. Consequently, we propose a revised covariates specification (denoted AJT). This involves dropping contemporaneous outcome covariates and redundant variables, adding an alternative set of regional dummies and including additional controls. These are selected following Sala-i-Martin et al. (2004) who undertake comprehensive Bayesian averaging of long-run growth estimates. We include variables identified by these authors that are among those with the highest posterior probability of inclusion and refer to initial conditions. To this, we add civil liberties in 1972 and distance to major ports. The first of these captures additional dimensions of initial institutional quality, including the ability of citizens to bring the government to account, which is often deemed relevant for aid effectiveness. Air distance is associated with export transaction costs, and ease of access to developed markets and has recently been identified by Moral-Benito (2009) as a robust correlate of growth.
3.5. Alternative estimators
Another area that can be strengthened concerns the choice of IV estimator. In light of the expected complexity of the growth process as well as the different properties of alternative estimators, it is valid to investigate whether or not empirical results hold across different estimators. While RS08 employ a 2SLS estimator, this is not the only option. Other possible estimators, which offer moderate differences, include LIML (limited information maximum likelihood), Fuller’s modified LIML (with alpha = 1), and a continuously updated GMM estimator (GMM-CU). In the program evaluation literature, the “doubly robust” estimators of Robins and Rotznitzky (1995) are attractive. Imbens and Wooldridge describe these estimators as “best practice” (2009: 25). Various doubly robust estimators have been proposed (see Imbens, 2004); however, none of these can be applied straightforwardly to the current aid-growth problem. They assume a binary
15
Arndt et al.: Aid, Growth, and Development: Have We Come Full Circle?
Table 4: Modified IV regressions, 1970- (I) (II) (III) (IV) (V) (VI) IV-LIML IV-IPWLS IV-LIML IV-IPWLS GMM-CU Fuller Aid / GDP 0.22* 0.21* 0.25** 0.13*** 0.25** 0.24** (0.12) (0.13) (0.12) (0.05) (0.12) (0.12) Initial per cap. GDP -1.34*** -1.92*** -1.03*** -1.33*** -1.03*** -1.05*** (0.40) (0.39) (0.38) (0.27) (0.38) (0.37) Initial level of policy 2.14*** 2.58*** 2.12*** 2.44*** 2.12*** 2.12*** (0.60) (0.62) (0.54) (0.46) (0.54) (0.53) Initial life expectancy 0.09** 0.05 0.04 0.03 0.04 0. (0.04) (0.03) (0.04) (0.04) (0.04) (0.04) Geography 0.63** 0.48** 0.29 0.25 0.29 0. (0.25) (0.24) (0.26) (0.21) (0.26) (0.25) Coastal pop. density in 1965 0.00** 0.00*** 0.00** 0.00** (0.00) (0.00) (0.00) (0.00) Primary schooling in 1960 2.58** 2.26** 2.58** 2.56** (1.15) (0.88) (1.15) (1.13) Malaria risk in 1966 -1.50* -1.06* -1.50* -1.49* (0.85) (0.58) (0.85) (0.83) Invest. goods price, 1960-64 -0.01 -0.01 -0.01 -0. (0.00) (0.00) (0.00) (0.00) Civil liberties in 1972 -1.28* -0.98* -1.28* -1.24* (0.70) (0.50) (0.70) (0.68) Air distance (log) 0.09 -0.03 0.09 0. (0.38) (0.33) (0.38) (0.38) Specification RS08 RS08 AJT AJT AJT AJT Scale of excluded instrument Continuous Binary Continuous Binary Continuous Continuous Regional dummies SSA, EA SSA, EA SSA, A, LA SSA, A, LA SSA, A, LA SSA, A, LA N 78 78 78 78 78 78 R-squared 0.57 0.70 0.59 0.77 0.59 0. Kleibergen-Paap Wald F stat. 29.48 24.42 17.28 39.78 17.28 17. Stock-Wright LM S stat. 4.33 3.53 5.77 6.49 5.77 5. (probability) 0.037 0.060 0.016 0.011 0.016 0. significance level: * 10%; ** 5%; *** 1% Notes: the endogenous variable is Aid/GDP, re-estimated from OECD-DAC (2008) data treating possible missing values as zeroes; in columns (I) and (II) the specification follows Rajan and Subramanian (2008) (only selected covariates shown); all remaining columns use a modified specification, removing contemporaneous and redundant covariates and adding additional initial conditions; chosen estimator is in the column title; initial policy refers to the Sachs-Warner trade policy index; geography refers to the average of the number of frost days and tropical land area; intercept not shown; standard errors, given in parentheses, are robust to arbitrary heteroskedasticity; dependent variable is the average real growth rate. Source: authors’ estimates.
With respect to the implementation of the IV-IPWLS estimator (columns III and IV), a binary instrument is required. This is derived by taking the fitted instrument from RS08’s zero stage regression, sorting countries in ascending
17
Arndt et al.: Aid, Growth, and Development: Have We Come Full Circle?
order (from lowest to highest predicted aid shares), and then selecting the first 30 for the “control” and the rest for the “treatment” group. The motivation for this choice is to identify a sub-sample of countries with the smallest possible average value for predicted aid inflows while still maintaining statistical viability. Thus, in practice, the control group approximately corresponds to all countries falling below the 40 th^ percentile. Besides permitting application of doubly robust techniques, dichotomization of the instrument represents a useful robustness check. If results arising from the binary instrument were not comparable with its continuous counterpart, this might indicate that the latter findings were driven by peculiarities in the distribution of the instrument. It also relaxes the assumption of a constant linear relationship between aid and growth, instead placing emphasis on the average difference between treatment and control groups regardless of the shape of growth’s response to aid. Consequently, possible non-linear effects due to diminishing returns to aid are addressed by this dichotomization. Finally, as the instrument is derived from a zero stage regression, dichotomization provides a check against measurement error or misspecification in the zero stage. Turning to results, the range of test statistics reported in Table 4 indicates that the new instrument continues to perform strongly across different specifications and estimators. Under-identification tests (not shown), which can be interpreted as testing the null hypothesis of a zero correlation between the instruments and the endogenous regressors, are all rejected. The weak identification test (the Kleibergen-Paap Wald F statistic, which uses a finite- sample adjustment of the standard F-statistic to assess the strength of the partial correlation between the excluded instruments and the endogenous variables in first-stage regressions) not only exceeds critical values in all cases but is comparable to the levels achieved using RS08’s original approach (Table 1, column II). Perhaps more importantly, the Stock-Wright S statistic, which is based on the reduced form regression and is robust to the presence of weak instruments (see Baum et al. , 2007), finds a significant (partial) correlation between the instrument and dependent variable in all cases. Moving across the columns of Table 4, we note that the treatment effect – i.e., the coefficient on the endogenous aid variable – is consistently positive, significant, and in a domain that is consistent with the RS08 prior (Section 2.4). The main effect of using the new and strengthened instrument (column I) is that the treatment effect estimate edges upwards (from 0.l0 to 0.22). The doubly- robust estimator leaves this result almost unchanged, but enhances the overall explanatory power of the model. According to the Kleibergen-Paap Wald F statistic, switching to the modified specification (column III onwards) slightly reduces the strength of the instrument in the IV-LIML first stage. However, by placing greater emphasis on the most informative observations, the strength of the instrument is considerably improved for the IV-IPWLS estimator. Finally, the
18
Journal of Globalization and Development, Vol. 1 [2010], Iss. 2, Art. 5
DOI: 10.2202/1948-1837.