Political decentralization and the effectiveness of aid Thushyanthan Baskaran Arne Bigsten Zohal Hessami University of Gothenburg & Gothenburg Centre of Globalization and Development, thushyanthan.baskaran@economics.gu.se University of Gothenburg & Gothenburg Centre of Globalization and Development, arne.bigsten@economics.gu.se University of Konstanz, zohal.hessami@uni-konstanz.de October 11, 2011 Abstract Several international organizations, for example the World Bank, advocate public sector decentralization as a way to improve economic and social outcomes in the developing world. But decentralization is not the only means through which international organizations and donors attempt to foster development. Another important instrument of development policy is aid. In order to evaluate whether decentralization does indeed advance socio-economic progress, we need to understand how it interacts with other development policies. This paper studies whether aid is more effective in decentralized countries by focusing on three important development goals: increasing economic growth, lowering infant mortality, and overcoming poverty. We estimate TSLS models using a cross-section dataset with country averages for 81 countries over the period 1990-2009. The results suggest that aid is not more effective in decentralized countries. Decentralization, therefore, is not a promising strategy to enhance the effectiveness of aid. Keywords: Development aid, Decentralization, Economic and social development JEL codes: : O1, O2, H7 Corresponding author: Department of Economics, University of Gothenburg, Vasagatan 1 (P.O.Box 640), SE 411 24 Gothenburg, Sweden, Tel: +46-(0)31-786 1374, Fax: +46-(0)31-786 1326. Arne Bigsten thanks Sida-SAREC for financial support within the project Global Development and Poverty Reduction: The Role of Institutions and Policies.
1 Introduction Is the vertical organization of the state the extent to which power is centralized at the national capital an important determinant of aid effectiveness? Do more centralized countries make better use of aid? Or should aid receiving countries give as much autonomy as possible to local officials? Experts employed by international organizations and policy makers in the developing world alike have traditionally favored centralized government. While this preference for centralization had many reasons, an important one was the belief that aid resources will be used more efficiently in centralized countries. The standard argument was that the national government has a comprehensive view of the country s investment needs and prospects, and could allocate scarce aid resources and exploit economies of scale more effectively than a multitude of local governments (Smoke, 2001). But after decades of development policies produced unsatisfactory results, both academics and policy makers begun to question their priors about the benefits of centralization. Rather than leading to a more efficient allocation of resources, centralized policy making in the developing world came to be associated with waste, mismanagement, and economic distortions. Because of this change in attitude towards centralization, international organizations and donors are increasingly giving a new policy prescription to developing countries: decentralization. Decentralization has a multitude of dimensions, but on an abstract level, it can be understood as vesting subnational policy makers with sufficient autonomy and authority to purse policies independently from the central government. More specifically, decentralization involves granting local officials some meaningful degree of political independence and fiscal power vis-a-vis the center. Most developing countries claim to have initiated moves toward more decentralized government at the beginning of the 1990s (Dillinger, 1994). While many of these countries begun to decentralize primarily for political reasons in particular to strengthen democracy international organizations such as the World Bank quickly supported the decentralization initiatives in the expectation that they will also lead to improvements in socio-economic conditions (World Bank, 2000). It was thought that through decentralization, the accountability of public officials could be increased and development programs could be better targeted. Indeed, the academic literature on fiscal federalism gave some credence to such expectations. This literature often associates decentralization with better preference matching (Oates, 1972), competitive and therefore more efficient government (Brennan and Buchanan, 1980), and increased public sector innovation (Strumpf, 2002). Nevertheless, there are other academic contributions that identify a number of drawbacks of decentralized policy making. These drawbacks include harmful tax competition and a race to the bottom in tax rates if subnational governments have too much tax autonomy (Zodrow and Mieszkowski, 1986), distortions in the fiscal incentives of subnational policy makers because of intergovernmental transfer schemes (Goodspeed, 2002), or the introduction of additional veto players into the political system (Tsebelis, 1995). While the relevant articles discuss these potential drawbacks of decentralization not explic- 1
itly in the context of aid policies, it is possible that such problematic features might diminish the effectiveness of aid in decentralized countries. Given the huge resources that are devoted to aid, even a small increase in its effectiveness may translate into large improvements in the social and economic conditions of the recipient countries. Against this backdrop, this paper investigates whether the hopes invested in decentralization for development are justified by focusing on the question of aid effectiveness. We explore whether aid is more or less effective in decentralized countries. More specifically, we consider the interactions between development aid and different indicators of decentralization on three indicators of economic and social well-being. The indicators of well-being considered in this paper are (i) economic growth, (ii) infant mortality, and (iii) the poverty rate. The indicators of decentralization that we analyze are whether a country is (i) constitutionally a federation, whether (ii) autonomous subnational regions exist, whether there are (iii) subnational elections, and whether (iv) subnational officials are appointed locally. 1 These are obviously indicators for the political dimension of decentralization. We deliberately omit measures of fiscal decentralization, even though subnational autonomy over revenues and expenditures is an important element of any meaningful decentralization process. But unfortunately the only available data for fiscal decentralization for the countries in our sample are based on the IMF s GFS database, which is notoriously unreliable when it comes to measuring the real decision making power of subnational governments (Ebel and Yilmaz, 2002). All independent and dependent variables are cross-section averages. The data covers 81 countries 2 and (at most) the period 1990-2009. We limit our sample to this period for two reasons. First, most decentralization initiatives were begun in the late eighties and early nineties. Data from earlier periods are therefore of less interest for studying the effects of decentralization. More importantly, the end of the Cold War may have changed the incentives of donors. There is evidence that donors reduced the amount of aid given for geo-political reasons. Our empirical strategy is to specify linear models with interaction effects between the different measures of decentralization and development aid. To account for the endogeneity of development aid, we rely on TSLS regressions. We use the log of population size as our main instrument because instruments based on population size have been shown to perform well by the aid and growth literature (Rajan and Subramanian, 2008; Bazzi and Clemens, 2010). In our regressions, we find little evidence for a beneficial effect of decentralization on aid effectiveness. In fact, the existence of state elections even appears to reduce aid effectiveness when it comes to promoting economic growth. From a policy perspective, these results indicate that decentralization should not be advocated as a means to increase the effectiveness of aid. Since the implications of decentralization for aid effectiveness have been neglected in the econometric literature, it is difficult to set our results in perspective. To our knowledge, the only existing econometric studies that ana- 1 See below for more complete definitions. 2 Even though we have data available for more countries, we decided to drop those that are not commonly regarded as developing. In particular, we drop Easter-European countries and countries such as Korea. 2
lyze this question are Lessman and Markwardt (2010, 2011). However, these authors are mostly interested in the impact of fiscal decentralization on aid effectiveness. Moreover, they only study economic growth and consequently do not explore the decentralization aid nexus with respect to other measures of well-being. 3 The remainder of this paper is structured as follows. The next section elaborates on the theoretical link between decentralization and aid effectiveness. Section 3 describes the empirical model and discusses the data. Section 4 presents the result. Section 5 concludes. 2 Decentralization and development aid: theory and evidence Why should decentralization matter for aid effectiveness? The answer is that most aid resources are given as grants to recipient countries; and government officials in these countries are those who are charged with channeling the resources toward individual projects. In general, aid can be given either with or without conditions. If it is given without any conditions (for example as budget support), national policy makers will have considerable autonomy on what projects to spend resources. If it is given with conditions i. e. if developing countries are required to use the resources for specific projects chosen by the donors the government remains responsible for the implementation of these projects. The nature of the government of recipient countries, therefore, plays an important role for how aid affects socio-economic outcomes. If the government is efficient, competent, and accountable to its citizens, then it stands to reason that aid will translate into significant improvements in living conditions. If, on the other hand, the government is characterized by incompetence and corruption, aid resources may be wasted or even stolen. One important characteristic of the government is its vertical organization. A large literature asserts that whether or not a state is decentralized will have significant consequences for how a country is governed. In essence, these contributions propose a causal effect of decentralization on the quality of government. Depending on the idiosyncratic viewpoints of the authors (i. e. whether they favor centralization or decentralization), their works indicate that by either increasing or decreasing the extent of decentralization, countries can improve the quality of their government. As higher-quality governments should make better use of aid receipts, it is a straightforward conjecture that there might be a link between decentralization and aid effectiveness. Little systematic evidence exists hitherto on whether decentralization indeed increases aid effectiveness. Some conjectures based on the existing theoretical and empirical literature can be formulated, however. Theoretically, decentralization can have beneficial or harmful consequences for governance. As discussed in the introduction, the literature on fiscal federalism has identified a set of beneficial features. For example, it is a straightforward corol- 3 However, Lessman and Markwardt (2010) do study some dimensions of political decentralization, such as whether a country is a federation and the number of tiers of government. They find that those aspects of political decentralization that they analyze in their study are in general insignificant. 3
lary to the famous decentralization theorem (Oates, 1972, 1999), that decentralization will enable countries to take account of different preferences and other forms of heterogeneity when allocating development aid to individual projects. Following the literature on laboratory federalism (Strumpf, 2002), we may also expect that in decentralized countries, subnational governments can experiment with different strategies for the implementation of development programs. Those that prove to be successful can then be adopted by all regions. On balance, this should improve aid effectiveness. According to the literature on yardstick competition, decentralization is beneficial because it allows citizens to compare the performance of their local government with those of neighboring regions, i. e. to use the latter as a yardstick (Salmon, 1987; Besley and Case, 1995). This would increase transparency and force politicians to be efficient. Similarly, the public choice tradition argues that decentralization will tame the leviathan by fostering competition between subnational governments (Brennan and Buchanan, 1980). Another closely related idea is that federalism can function as a market-preserving mechanism, as the recent experience in China seems to show (Qian and Weingast, 1997). These contributions, therefore, suggest that decentralization might force governments to use aid more efficiently. There are, of course, also some potentially negative effects of decentralization for aid effectiveness. First, decentralization may simply increase transaction costs and lead to coordination failures (De Mello, 2000). Consequently, it is possible that neither positive nor negative externalities in aid programs are internalized. Related to this argument, a strand of the literature argues that decentralization will result in inefficient strategic interactions between tiers of government. The literature is mostly concerned with horizontal (Zodrow and Mieszkowski, 1986) and vertical (Keen and Kotsogiannis, 2002) tax competition, but decentralization may also result in inefficient interactions in expenditures (Case et al., 1993). In other words, decentralization may give rise to strategic considerations in how aid resources are spent. For example, it is possible that local officials display a bias toward spending resources on marginally productive infrastructure projects in order to attract capital away from neighboring regions. As a consequence, fewer resources could be allocated to poverty alleviation, even if aid resources would result in much larger improvements in overall living conditions if it would be spent to combat poverty. Inefficient policy choices at the local level have also been at the focus of the literature on soft budget constraints in intergovernmental relations. This literature argues that if subnational governments expect that they will receive a bailout from the center if they decide to spend too much or tax to little, they may have an incentive to do just that (Wildasin, 1997; Goodspeed, 2002). Consequently, another negative implication of decentralization could be that local officials choose to implement projects that are not the priority of the national government, forcing the latter to step in ex-post and use additional resources to carry out its favored projects. Alternatively, local governments may approve projects that require more financing then they have currently available with the explicit intent of petitioning the central government for additional resources once funds are exhausted. Furthermore, decentralization involves a trade-off between the informational advantages of local actors and the possibility that subnational govern- 4
ments are captured by local elites. Bardhan and Mookherjee (2006), for example, show formally that decentralization can lead to distortions in local service delivery such that resources are allocated away from the poor and toward the local elites. Similarly, Prud homme (1995) and Tanzi (1996) argue that decentralization may lead to increased corruption. If these arguments are correct, then allocating aid through local officials will not result in better outcomes; in fact, decentralized aid allocations might have unwelcome distributional consequences. A final concern against decentralization is the quality of the local bureaucracy: since there are better career opportunities for bureaucrats at the center, central agencies will possess more qualified personal (Bardhan, 2002). Consequently, it is possible that the implementation of aid programs could be handled better at the center. The empirical evidence regarding the impact of decentralization on governance is mixed. At the micro-level, Faguet (2004) finds that public investments for local infrastructure such as education, water, or sanitation increased in Bolivia after a move toward decentralization, and did so according to the needs of individual municipalities. Gallaso and Ravallion (2005) find for an antipoverty program in Bangladesh that there were interactions between withinvillage targeting of project funds and both village and project characteristic. For a antipoverty transfer program in Mexico, Coady (2001) finds that while the center is capable to target poor localities, the localities themselves were not particularly successful in targeting poor households. At the macro-level, Fisman and Gatti (2002) find for a cross-section of countries that there is a negative relationship between fiscal decentralization and corruption. Fan et al. (2009) explore the relationship between different dimensions of political decentralization and various measures of corruption. In general, they find that there is a positive relationship between corruption and their measures of decentralization. Enikolopov and Zhuravskaya (2007) find that fiscal decentralization can have positive effects on economic growth and government quality, but that the overall effect depends, inter alia, on the degree of political decentralization. Blanchard and Schleifer (2001) argue, based on the Russian experience, that locally elected officials are prone to be captured by local interests and therefore suggest some degree of centralization would be appropriate for Russia. Treisman (2002) does not identify a relationship between various measures of political decentralization and corruption. On the other hand, Gerring and Thacker (2004) and Kunikova and Rose-Ackerman (2005) find a positive relationship between federalism and corruption. 4 Considering these highly diverse results in both the micro- and macro literature, there appears to be little empirical consensus regarding the socioeconomic effects of decentralization. In particular with respect to the implications of political decentralization, the evidence, which is scarce to begin with, does not lead to unambiguous conclusions. Consequently, it is difficult to predict how decentralization will affect aid effectiveness; a fact that only reinforces the need to study this question empirically. The predicament of ambiguous results plagues the aid and growth literature as well. As this is not the place to review this large literature, we will only focus on some of the most recent contributions. Rajan and Subramanian 4 A good review of the literature is provided by Kyriacou and Roca-Sagales (2010). 5
(2008), for example, find in an influential study that aid has no significant effect on growth. Doucouliagos and Paldam (2009) draw the same conclusion from their comprehensive review of the literature. On the other hand, there are some studies that do find a significantly positive impact of aid. For example, results in Arndt et al. (2010), obtained by the application of a treatment method, suggest that aid has beneficial effects. Clemens et al. (2004) focus on specific categories of aid (such as budget and balance of payments support, investments in infrastructure, agriculture and industry) and find that these have positive growth effects. Similarly, Minoui and Reddy (2009) find that certain types of aid increase growth. Nevertheless, the results in the growthaid literature are fragile (Roodman, 2007). Since aid is given also to achieve a range of goals other than growth, this should maybe not come as a surprise. Besides exploring the average effect of aid, it is also of interest to know whether its effectiveness varies by the recipient s characteristics. One strand of the literature has tied aid effectiveness to political institutions. Burnside and Dollar (2000) find that aid has a positive effect on growth, but only in countries with good policies. Other contributions investigate further conditioning factors for aid effectiveness, such as geographical location (Dalgard et al., 2004) and socio-political stability (Islam, 2005). In view of the contributions which advocate that researchers should consider interactions between recipients characteristics and aid, we hypothesize that the vertical organization of the state affects the effectiveness of aid. And if the effect of aid indeed differs between decentralized and centralized countries, then ignoring this heterogeneity may result in a wrong estimates. For example, aid might be effective in centralized countries while fueling local capture, corruption, and other policy distortions in decentralized ones or vice versa. An empirical model that does not take this heterogeneity into account only estimates an average effect. 3 Empirical model and data To explore the effect of decentralization on aid effectiveness, we estimate the following model y =c+x i α+β 1 Aid/GDP i + β 2 decentralization i + β 3 Aid/GDP decentralization i + ǫ i, (1) where y is either Economic Growth, Log(Infant Mortality), or the Poverty Rate; and decentralization is one of the four following variables: Federation, Autonomous Regions, State Elections, Local Officials). x denotes a vector of control variables listed in Table 1, c is the constant, and ǫ is the error term. All variables are cross-section averages. The available time dimensions vary, of course, between variables and countries. The data cover at most the period 1990-2009. Model (1) is estimated with TSLS. The endogenous variable is the development aid to GDP ratio. While there are several reasons why aid could be endogenous (omitted variables, measurement error, etc.), the most likely one in the current context is reversed causality: donor countries could allocate aid according to the prevailing economic and social conditions in a country. That 6
is, countries with persistently low growth rates or high levels of infant mortality and poverty could receive more aid. On the other hand, donors could allocate more aid to countries that have done well in the past believing that aid resources will be put to good use in these countries. 5 Whatever the true explanation, there is the possibility of an effect of the dependent variables on aid flows, and this possibility must be taken into account in the econometric approach. We, therefore, rely on the instrumental variables technique to identify the effect of aid. In the following, we describe in detail the decentralization variables, control variables, and the instruments that we use to estimate Model (1). All variables and sources are listed in Table 1. Summary statistics can be found in Table 8. 3.1 Decentralization variables As indicated in equation 1, we use four distinct (but to some extent overlapping) measures to capture different dimensions of decentralization. First, we use a dummy, taken from Treisman (2007), which is 1 if a country is considered to be a federation. Second, we use a dummy indicating whether a country has autonomous regions (which is not the same as a state or province: for a country to possess autonomous regions, the constitution must designate at least some regions as autonomous, independent, or special ). Third, we use a dummy that is 1 when when the state or provincial governments or legislatures or both are locally elected. The last two variables are obtained from the Database of Political Institutions provided by Beck et al. (2001). 6 Fourth, we use a a dummy that is 1 when local officials are locally chosen (appointed, elected, or otherwise), and 0 when regional officials are appointed by the central government or when no regional governments exist. 7 This variable is taken from Clark and Regan (2010). We include in the regressions the cross-section averages of the decentralization variables. Note that several countries have changed regimes during the sample period for some dimensions of decentralization. Thus, averaging over the sample period produces for these measures a value that lies between 0 and 1; the decentralization variables capture in these cases the average time period a country has possessed a particular dimension of decentralization. The distribution of the four decentralization measures can be inferred form Figure 1. In general, we find that for most decentralization variables a significant number of countries have switched regimes over-time. The only exception is the variable that measures whether a country is constitutionally designated as a federation: it is constant for all countries throughout the sample period. 5 Collier and Dollar (2001), for example, suggest that donors should target aid to countries with good policies. 6 The original variable for state elections in the DPI is coded on a three-point scale, according to whether neither the state executive and legislature, either one of both, or both are elected locally. We rescaled the original variable to a 0-1 dimension. 7 As for the DPI state election variable, our original source uses a three-point scale: (i) state officials locally appointed, (ii) state officials appointed by central governments, (iii) no state governments. We rescaled the variable to a binary dimension. 7
3.2 Control variables We use in all regressions the same set of control variables that are regularly included in studies on the economic and social implications of aid. 8 First, we include initial 9 GDP per capita. Second, we include as a measure of political rights the initial value of the combined polity score from the Polity IV database (Marshall and Jaggers, 2002). Third, we control for initial economic openness (trade to GDP ratio) using data from the Penn World Tables (Heston et al., 2009). Fourth, we control for the initial inflation rate and for the M2 to GDP ratio (i. e. money and quasi-money) using data from the World Development Indicators (WDI). Fifth, we include the ethnic fractionalization measure provided by Alesina et al. (2003). We furthermore include in all regressions the initial surplus to GDP ratio, initial life expectancy at birth, and the initial amount of arable land per person. The data source for these three variables is the WDI. We also include a measure for revolutionary wars from the Political Instability Task Force. Finally, we include dummies for sub-saharan and East-Asian countries. The rational for including these control variables is that they are plausibly correlated with aid receipts. At the same time, they affect in all likelihood at least some of the dependent variables. That is, initial GDP per capita is used to control for convergence effects in economic growth; and it is also likely that the level of economic development at the beginning of the sample period is related to subsequent infant mortality and poverty rates. Similarly, the level of democracy might affect both economic and social variables. Openness can lead to higher growth and lower poverty rates, for example in countries that adopt export-based growth strategies. The average inflation rate is included to proxy macroeconomic instability. The M2/GDP ratio is a proxy for financial depth. Ethnic fractionalization proxies collective action problems and propensity for conflict. The surplus/gdp ratio is included as a measure of fiscal stability. Life expectancy controls for the physical well-being of the population. The amount of arable land per person is a proxy for geography. Revolutionary wars are a measure for political instability and actual violent conflicts. Finally, the sub-saharan and East-Asian country dummies take account of specific features of countries located in these regions and not already captured by the other control variables. 3.3 Instruments Finding appropriate instruments for aid receipts has been one of the great challenges of the aid and growth literature. A great number of candidates have been proposed. We use in this paper the log of population size as our primary instrument given the results in Rajan and Subramanian (2008). They construct a well-performing instrument that relies on the relative population sizes of donor and recipient countries. The rationale for why this might be a good instrument is that donors will tend to give aid to (relatively) smaller 8 In particular, Rajan and Subramanian (2008) include conceptually similar controls in their cross-section regressions. 9 Whenever possible we include initial values. If initial levels are not available for a country, we include the earliest available value. This holds also for all other variables for which the initial values were missing. 8
countries because donors will have more influence in smaller countries. Bazzi and Clemens (2010), however, show that the Rajan and Subramanian (2008) instrument essentially captures only the population sizes of the recipient countries. In other words, Rajan and Subramanian (2008) identify the effect of aid for all practical purposes only with variation in the recipients population sizes. In our first stage regressions, the log of population is indeed negatively related to aid receipts. Population size should be reasonable exogenous to aid receipts, even though we cannot test formally for instrument validity due to the lack of overidentifying restrictions. On the other hand, population size is certainly not a perfect instrument. There might be, for example, omitted variables that are correlated with both the dependent variables and population size or population size might have a direct effect on the dependent variables (in addition to the indirect one through aid receipts). While there are such concerns, our aim in this paper is not to contribute to the discussion on the proper instrument strategy for aid receipts. Given our research question, it is reasonable to rely on a variant of a variable that has already been successfully used in the literature. Nevertheless, we report further below the results from a set of robustness tests with the Rajan and Subramanian (2008) instrument. A more serious concern is that Model (1) consists, in addition to development aid, of a set of interaction variables that are constructed with development aid. We require for these interaction variables additional instruments. Instead of expanding the instrument set with further (and possibly invalid) instruments, we apply a procedure described in Wooldridge (2002). This procedure works as follows. First, development aid is projected onto the space spanned by the in- and excluded instruments (i. e. the control variables in the second stage regression and population size). Thereafter, the predicted values are interacted with the relevant decentralization variable. This interaction between predicted development aid and the decentralization variable is then used as an instrument for the interaction between the actual aid receipts and the relevant decentralization variable. More specifically, we first estimate the model Aid/GDP i = d+x i γ+δlog(population) i + decentralization i + ν i, (2) where x is the vector of control variables included in Model (1) (i. e. the included instruments), and log(population) is the instrument excluded in the second stage regression. d is the constant and ν i is the error-term in this auxiliary regression. Using the estimate for the coefficients of Model (2), we construct predicted values for Aid/GDP ratio, Pred. Aid/GDP i, that rely exclusively on exogenous variables. These predicted values can then interacted with each of the four decentralization variable. Thus, the instrument for the interaction variable in equation (1) is Interactions instrument i = Pred. Aid/GDP i decentralization i, (3) where decentralization is one of the four measures mentioned further above. In other words, we use as instruments for each interaction between actual development aid and a decentralization variable a separate interaction variable 9
constructed by interacting predicted development aid and the relevant decentralization variable. In the first stage regressions, these instruments are positively and significantly related to the interactions between development aid and the relevant decentralization variable. 10 4 Results This section presents and discusses our regression results. We begin with a set of baseline results in Section 4.1. Thereafter, we explore the robustness of our findings in section 4.2. 4.1 Baseline results We discuss the results form the baseline regressions for economic growth, infant mortality, and the poverty rate in turn. All regressions are conducted after excluding outliers with the Hadi-Method. 11 Considering the summary statistics in Table 8, it is indeed likely that there are outliers with respect to some control variables. For example, there is one country (Peru) that has an initial inflation rate of over 6000%. Out of the 81 countries in the database, 13 are excluded as outliers. 12 Economic growth Table 2 presents the baseline regressions with economic growth as dependent variable. In each of the four columns, a different indicator for decentralization is used. The results in the table are to be interpreted in the following way (Brambor et al., 2006). The estimated coefficient for the Aid/GDP ratio measures the marginal effect of aid when the indicator for decentralization is 0, i. e. the effect of aid on growth in a country that does not possess the relevant dimension of decentralization. Each of the interaction variables tests whether there are significant differences in aid effectiveness between decentralized and centralized countries. If an interaction variable between a measure of decentralization and aid is significant, we can conclude that there are significant differences in aid effectiveness between decentralized and centralized countries. Finally, we report at the bottom of the table, in the row entitles Aid/GDP Decent. = 1, the marginal effect of development aid in countries that posses the relevant dimension of decentralization. The respective p-values are reported in parentheses beneath the marginal effects. 10 All first stage results are available from the authors. 11 We do not include the decentralization variables in the multivariate model that is used to identify outliers. The reason is that the decentralization variables are derived from dummy variables, and as such assume values between 0 and 1, and both extremes are reasonable and expected. We also do not include all control variables because the statistical package we use to implement the Hadi-procedure reports that there might be multicollinearity in some models (causing the abortion of the program) We include, however, all control variables for which outliers seem likely: e. g. inflation or the surplus/gdp ratio. 12 The following 13 countries are excluded in the regressions: Sierra Leone, Botswana, Jordan, Uruguay, Liberia, Niger, Zambia, Dem. Rep. of Congo, Brazil, Nicaragua, Peru, Lebanon, Argentina. Data for poverty was not available for Lebanon and Argentina to begin with; these regressions exclude technically only 11 outliers. 10
The coefficient for the interaction variables are, with one exception, insignificant in Table 2. The exception is the state election variable. Aid seems to lead to lower growth rates in countries that allow for state elections. The marginal effects of aid for Decent=1 even suggests that it has a negative effect on growth in countries that have state elections (coefficient: -0.179, p-value: 0.066). The numerical values imply that in countries with state elections, a one percentage point increase in the aid to GDP ratio reduces growth by 0.18 percentage points. Other dimensions of decentralization, however, do not affect the effectiveness of aid when it comes to promoting growth. Infant mortality Tables 3 reports the regression results for infant mortality. The structure and interpretation of this regression table is identical to the table for economic growth. The coefficient of the interaction effect is always insignificant. In other words, the regressions do not suggest that decentralization enhances the effectiveness of aid in reducing infant mortality. There is furthermore no evidence that aid reduces infant mortality in decentralized countries. The estimate for Aid/GDP Decent. =1 is never significant. Poverty In Table 4, we report the results for the impact of aid and decentralization on poverty. There is no effect of decentralization on aid effectiveness. The interaction term is consistently insignificant. Moreover, the marginal effect of aid in decentralized countries (Aid/GDP Decent.=1) is also consistently insignificant, thereby suggesting that an increase in aid does not lead to a reduction in the poverty rate in decentralized countries. 4.2 Robustness In this section, we explore whether the finding of an insignificant effect of decentralization on aid effectiveness is robust to four robustness checks. The first concern with the baseline regressions is that excluding the outliers might result in the loss of some valuable information. To explore the robustness of our results to this problem, we estimate models with all outliers included. The second concern is that averaging the dummy variables that measure whether a country possesses a certain dimension of decentralization results for some measures and countries in values that lie in between 0 and 1. In these cases, countries have switched regimes. Consequently, the decentralization measures (since they are cross-section averages) capture the average amount of time countries were characterized by a particular dimension of decentralization. As our decentralization variables are, therefore, in several cases not real dummy variables, we explore whether our results change if we transform them such that they are either 0 or 1. Our transformation rule is that the value of each of the decentralization measures is transformed to 1 for countries that where characterized during the majority of the sample period with the 11
relevant dimension of decentralization (i. e. countries for which the crosssection average is larger or equal to 0.5); the value is transformed to 0 when that particular aspect of decentralization was absent in a country during most of the sample period (i. e. for countries where the cross-section average was smaller than 0.5). That is { 1 if decentralization 0.5 dec. binary = (4) 0 else, where decentralization is one of the four decentralization variables considered. Note that these regressions will not produce different estimates for the federalism variable because it does not display over-time variation (we therefore omit these robustness tests for the federalism variable). The third concern is possible invalidity of the log of population size as instrument. One obvious way to address this concern is to use different instruments. Therefore, we replicate our baseline regressions with the instrument used in Rajan and Subramanian (2008) after appending to their dataset our development aid and decentralization data. In particular, we use their control variables and their instrument based on donor country influence. We calculate the instruments for the interactions according to the procedure described above. The fourth concern is multicollinearity. Note, for example, that few of the control variables in Table 2 are significant. While multicollinearity in the control variables does not necessarily result in bias, it might result in insignificant estimates for the variables of interest. More specifically, if aid receipts are strongly correlated with some of the control variables, aid or its interactions with decentralization might have no statistically detectable effect even if they are economically important (see also footnote 11). To check for this possibility 13, we run regressions after dropping all control variables except initial GDP per capita, the aid to GDP ratio, the decentralization variables, and the interactions between the aid to GDP ratio and the decentralization variables. Economic growth Table 5 reports the results of the robustness tests for the growth regressions. For brevity, we only show the results for the aid, decentralization, and interaction variables. In addition, we report the number of observations and the weak identification test statistic. When outliers are included in the regressions, the results are not significantly different from the baseline findings. The interaction between decentralization and aid continues to be insignificant except when state elections are considered. In countries with state elections, aid still appears to be less effective for promoting growth. (Yet, we do not find any more that aid reduces growth in countries with state elections). Using binary decentralization variables does not result in significant differences from the baseline regressions. The interaction term continues to be insignificant for almost all decentralization variables. Moreover, the existence of state election appears to reduce aid effectiveness as in the baseline regres- 13 When initial GDP per capita is dropped as well, the weak identification test statistics are very low: in this sense, initial GDP per capita appears to be an important control variable. 12
sions. The interaction term between state elections and development aid is negative and almost significant. The regressions with the Rajan and Subramanian (2008) dataset also confirm the baseline findings. The interactions are insignificant except in the case of state elections. The only noteworthy difference to the baseline results is that the weak identification test statistics are somewhat lower when the Rajan and Subramanian (2008) instrument is used, in particular for the regressions with the federalism dummy. Finally, dropping all control variables except initial GDP per capita does not lead to significantly different results than in the baseline regressions either. The interaction between the aid and decentralization variables is consistently insignificant. One noteworthy difference is, however, that the interactions are even insignificant in the state election regressions. Infant mortality Table 6 reports the findings from the robustness checks for infant mortality. The results do not change much compared to the baseline regressions. Even if outliers are included, the interaction effect remains insignificant. This is also true when binary decentralization measures and the Rajan and Subramanian (2008) instrument are used. The only exception are the results when other control variables are dropped. In this case, the interaction term is always positive and significant for three decentralization variables. These regressions, therefore, suggest that decentralized countries are less efficient in using aid to reduce infant mortality. Obviously, this result should be viewed with caution given that there might be a serious omitted variables bias. Nevertheless, these results show that if anything, aid is less and not more effective in decentralized countries. Poverty Table 7 presents the robustness checks for poverty. The results are again similar to the baseline estimates. The interaction terms remain insignificant when outliers are included. Using binary decentralization variables leads to essentially the same results as in the baseline regressions. Applying the Rajan and Subramanian (2008) instrument does not change the results either. And omitting control variables results in insignificant interaction effects as well. 5 Conclusion Both the baseline regressions and the robustness checks suggest that there is no systematic link between decentralization and aid effectiveness. Only in some regressions, notably when exploring the effects of state elections, we find significant interactions. But these regressions suggest that this dimension of decentralization reduces the effectiveness of aid in promoting growth. Our results, therefore, indicate that decentralization will not improve aid effectiveness. As our theoretical discussion shows, there are a number of reasons why decentralization could have an irrelevant or even a negative effect on aid effectiveness. Yet, the reduced from regressions presented in this paper 13
cannot shed light on the structural relationships. It is an interesting avenue for future research to explore whether local capture, less able bureaucracies, or inefficient strategic interactions are to blame for the fact that decentralized states are not able to use aid more efficiently than centralized ones. The sobering findings in this paper, however, do not necessarily imply that decentralization is unsuitable for developing countries. Even if decentralization does not increase the effectiveness of aid, there are good reasons beyond aid effectiveness to grant more autonomy to subnational governments. After all, many developing countries begun their decentralization initiatives to foster democracy and not to allocate aid more efficiently. It is possible that enhancing democracy through decentralization and allocating aid efficiently are to some extent inconsistent goals. For example, governments of decentralized states could be forced to allocate aid among regions according to political and not socio-economic considerations if democratic institutions are not sufficiently stable. In other words, countries may have structured their decentralization initiatives with the primary goal of maintaining their fledging democracies. The cost of setting priorities in this way could be less effective use of aid. Nevertheless, it might be a cost worth paying. Acknowledgment We thank Ranghuram Rajan and Arvind Subramanian for sharing their data. References Alesina, A., A. Devleeschauwer, W. Easterly, S. Kurlat, and R. Wacziarg (2003). Fractionalization. Journal of Economic Growth 8(2), 155 194. Arndt, C., F. Jones, and F. Tarp (2010). Aid, growth, and development: have we come full circle? Journal of Globalization and Development 1. Article 5. Bardhan, P. (2002). Decentralization of governance and development. Journal of Economic Perspectives 16, 185 205. Bardhan, P. and D. Mookherjee (2006). Decentralization and accountability in infrastructure delivery in developing countries. Economic Journal 116, 101 127. Bazzi, S. and M. A. Clemens (2010). Blunt instruments: a cautionary note on establishing the causes of economic growth. Center for Global Development Working Paper. Beck, T., G. Clarke, A. Groff, P. Keefer, and P. Walsh (2001). New tools in comparative political economy: the database of political institutions. World Bank Economic Review 15(1), 165 176. Besley, T. and A. Case (1995). Incumbent behavior: vote-seeking, tax setting, and yardstick competition. American Economic Review 85(1), 25 45. Blanchard, O. and A. Schleifer (2001). Federalism with and without political centralization: China versus Russia. IMF Staff Papers 48. 14
Brambor, T., W. R. Clark, and M. Golder (2006). Understanding interaction models: improving empirical analyses. Political Analysis 14(1), 63 82. Brennan, G. and J. Buchanan (1980). The power to tax: analytical foundations of a fiscal constitution. Cambridge: Cambridge University Press. Burnside, C. and D. Dollar (2000). Aid, policies, and growth. American Economic Review 90, 847 868. Case, A. C., H. Rosen, and J. Hines (1993). Budget spillovers and fiscal policy interdependence: evidence from the states. Journal of Public Economics 52(3), 285 307. Clark, D. and P. Regan (2010). Institutions and elections project. Insitutions data in Stata. http://www2.binghamton.edu/political-science/institutionsand-elections-project.html. Clemens, M. A., S. Raddled, and R. Bhavani (2004). Counting chickens when they hatch: the short-term effect of aid on growth. Center for Global Development Working Paper 44. Coady, D. P. (2001). An evaluation of the distributional power of Progresa s cash transfers in Mexico. International Food Policy Research Institute. Collier, P. and D. Dollar (2001). Can the world cut poverty in half? How policy reform and effective aid can meet the international development goals. World Development 29, 1787 1802. Dalgard, C.-J., H. Hansen, and F. Tarp (2004). On the empirics of foreign aid and growth. Economic Journal 114, 191 216. De Mello, L. R. (2000). Fiscal decentralization and intergovernmental fiscal relations: a cross country analysis. World Development 28(2), 365 380. Dillinger, W. R. (1994). Decentralization and its implications for urban service delivery. Urban Management and Municipal Finance Discussion Paper No. 16. Doucouliagos, H. and M. Paldam (2009). The aid effectiveness literature: the sad results of 40 years of research. Journal of Economic Surveys 23, 433 461. Ebel, R. D. and S. Yilmaz (2002). On the measurement and impact of fiscal decentralization. World Bank Policy Research Working Paper No. 2809. Enikolopov, R. and E. Zhuravskaya (2007). Decentralization and political insititutions. Journal of Public Economics 91, 2261 2290. Faguet, J.-P. (2004). Does decentralization increase government responsiveness to local needs? Evidence from Bolivia. Journal of Public Economics 88, 867 893. Fan, C. S., C. Lin, and D. Treisman (2009). Political decentralization and corruption: evidence from around the world. Journal of Public Economics 93, 14 34. 15