Public Spending Reallocations and Economic Growth Across Different Income Levels

This article examines the effects of public spending reallocations on economic growth. Assembling a disaggregated public spending dataset of 83 countries over the 1970&#8211;2011 period, we show that spending reallocations toward education, from health and social protection, have significant growth&#8208;promoting effects across a wide range of countries' income levels. However, income heterogeneity matters, particularly when reallocations involve infrastructure spending. Specifically, a reallocation from this spending to education also promotes growth, albeit primarily when a country's income level is low. This occurs because the effects of infrastructure spending are particularly weak in low&#8208;income countries, possibly due to the low quality of governance.


I. INTRODUCTION
Can a government promote long-run growth by changing the composition of public spending while keeping the total level fixed? For example, is there room for a rise in public education spending to foster growth, when it is financed by a fall in other spending items, such as defense, health, social protection, and infrastructure spending? Answering these questions is important at least for two reasons. First, even when high levels of public indebtedness prevent a government from raising the level of public spending in several years to come (as is often the case in the aftermath of the financial crisis of 2007-09), a government may still seek to promote growth by reallocating spending components. Second, when facing an expected rise in health and social protection spending amid ongoing population aging, governments may need to decide which other types of spending to cut over the next few decades, while still trying to preserve growth. 1 Various previous studies shed light on this important policy question. For example, Barro (1990) theoretically shows that if a government increases "utility-enhancing" public consumption while reducing "production-enhancing" spending, growth rates fall regardless of the level of total spending. 2 However, as theoretical studies often classify spending in a relatively abstract manner (e.g., utility-enhancing spending), concrete policy implications are often difficult to draw. Empirically, Devarajan, Swaroop, and Zou (1996), for example, find that a spending reallocation toward economic infrastructure is associated with slower growth, particularly in developing countries. 3 However, because this and other related previous empirical works normally do not clarify which specific spending component is used as a financing item (i.e., the component for which spending is decreased, to keep the level of total spending constant), their policy implications may still be limited. In short, the literature has not yet clarified how a government could potentially promote growth through specific spending reallocations. The objective of this article is to fill this gap.
Importantly, however, even when certain specific spending reallocations promote growth in one country, they may not have the same positive effects in others. For instance, Pritchett (2000) argues that infrastructure spending in developing countries can be particularly unproductive, because government officials, often unconstrained by institutions of low quality, may be allowed to pursue their private rents in the public infrastructure management. 4 This implies that for all else equal, a reallocation toward this spending item may not foster growth in these countries. 5 More generally, Gemmell, Misch, and Moreno-Dodson (2013) point out that private production requires different types of public inputs depending on countries' development levels, implying the potentially heterogeneous growth effects of spending reallocations. Overall, therefore, we pay particular attention to the potential role of countries' income levels in the spending reallocations-growth nexus.
To proceed, we first assemble a disaggregated public spending dataset. Based on the International Monetary Fund's (IMF) Government Finance Statistics (GFS) Yearbook, it covers 3. Ghosh and Gregoriou (2008), using data from 15 developing countries, show that a reallocation toward capital spending has a negative effect on growth.
4. Keefer and Knack (2007) find that the level of capital spending increases in the worsening of governance quality, suggesting that this type of spending may be unproductive in countries with poor governance. Given that infrastructure spending is largely capital intensive, their finding relates to Pritchett (2000).
5. Although Devarajan, Swaroop, and Zou (1996) show the growth-retarding effect of infrastructure spending (without specifying its financing item), their reasoning is different. Their argument is that the negative effect is attributed to the excessive use of this type of spending in developing countries, not necessarily highlighting the quality of this spending. functional components including defense, health, social protection, education, and transport and communication, for 83 countries from different income levels over the 1970-2011 period. 6 One key innovation of this dataset is that, to address the potential importance of spending decentralization movements in certain spending items such as health and education, it has a version which utilizes not only central government (CG) level-, but also general government (GG) level-data. 7 We then examine the growth effects of various spending reallocations across different income levels. Notably, we address potential endogeneity concerns in various ways. First, we address fiscal variables' association with business cycles by basing the analyses mainly on an 8-year framework, yielding five observations per country at most. 8 Second, we mitigate reverse causality issues by adopting a lag structure of public spending variables. Third, we tackle other potential endogeneity issues by using the generalized method of moments (GMM) dynamic panel data estimation approach. 9 Our main results are summarized as follows. First, considering the growth effects of spending reallocations without taking account of income heterogeneity across countries, we show that reallocations toward education, specifically from health and social protection spending, exhibit a particularly robust growth-promoting effect. However, we find that income heterogeneity matters in the spending reallocations-growth nexus, particularly when infrastructure spending (proxied by transport and communications) is involved. Specifically, while the abovementioned reallocations toward education tend to promote growth across a wide range of income levels, reallocations away from infrastructure spending, to items including education, also foster growth, albeit primarily when a country's income level is low. What happens is that 6. The list of the 83 countries is given in Appendix S1. 7. To explain briefly, we would ideally like to use only GG-level data, because it is more comprehensive, covering the local/state government levels, as well as the CG level. However, GG-level data are generally much scarcer than CGlevel data and it is unfeasible to conduct analyses based solely on it. Thus, this version of the dataset adopts a hybrid approach, replacing CG-level with GG-level data, particularly when the degree of spending decentralization is known to be high in a country. Details are explained below. 8. We also consider a 5-year framework as a part of the robustness checks.
9. This approach is developed by Arellano and Bond (1991), Blundell andBond (1998), andHoltz-Eakin, Newey, andRosen (1988). while the growth-promoting effects of education spending prevail independently of income levels, the effects of infrastructure spending exhibit a distinctly positive association with them. Evidence suggests that in line with Pritchett (2000), this particularly weak growth-promoting effect of infrastructure spending in low-income countries (LICs) may be driven by the low quality of governance, as often observed in these countries. For instance, if political officeholders are not properly constrained, the effectiveness of infrastructure spending may particularly suffer in these countries, because this type of spending, being capital intensive, tends to be a convenient vehicle for rent seeking. 10 In the literature, several previous works identify the key channels through which different spending items may affect growth. For example, theoretical models by Blankenau and Simpson (2004) and Glomm and Ravikumar (1997) suggest that public education spending may foster growth through promoting human capital accumulation. Agénor (2010) and Aísa and Pueyo (2006) show that health spending can also promote growth by enhancing life expectancy, making agents more patient, and thus encouraging private saving. Barro and Sala-i-Martin (2004) argue that defense spending fosters private investment and thus growth by enhancing entrepreneurs' property rights. Further, studies such as Glomm and Ravikumar (1997) and Turnovsky and Fisher (1995) show that infrastructure spending can promote private firms' productivity and thus growth, albeit with the caveats indicated by Pritchett (2000) for developing countries. Notice, therefore, that with each spending component appearing to have a growth-promoting potential on its own, it is not obvious, a priori, which spending item is more growth promoting. 11 In this regard, this article adds to the literature by empirically studying reallocation effects, that is, by practically comparing the effects of different spending items, across different income levels. 10. While acknowledging that the low quality of governance can also be a problem for education spending, the fact that this type of spending, largely comprising teachers' wage payments, is often predetermined outside officials' discretion means that there may be less room for their rent-seeking behavior.
11. Even social protection spending may promote growth by itself. For instance, Lambrecht, Michel, and Vidal (2005) show that in cases where parents do not leave bequests to their children, social security spending can enhance growth, in contrast to cases where parents leave bequests.
Further, several studies examine the growth effects of public spending for different financing sources, including a rise in revenue or budget deficits, rather than spending reallocations. For instance, Kneller, Bleaney, and Gemmell (1999), using a panel of 22 Organization of Economic Co-operation and Development (OECD) countries over the 1970-1995 period, find that a rise in "productive" expenditure financed by an equal rise in "non-distortionary" taxes promotes growth, whereas a rise in "unproductive" expenditure, coupled with a rise in "distortionary" taxes, reduces growth. Bose, Haque, and Osborn (2007), using a panel of 30 developing countries during the 1970s and 1980s, show that a rise in education spending, offset by nontax revenue, is associated with higher growth, while a rise in transport spending is not. Meanwhile, Gupta et al. (2005), focusing on 39 developing countries during the 1990s, show that a rise in capital spending has a positive effect on growth, when financed through budget deficits, but that a rise in current spending, particularly wages, has the opposite effect. While these studies can be seen as parallel with this article, a spending reallocation could often be a more viable policy option. This is because securing a given financing through revenue may not be straightforward, due to the elusive relationship between tax rates and tax revenue (as reflected in the Laffer curve), while deficit financing may be implausible for many countries in the face of high levels of indebtedness. 12 The remainder of the article is structured as follows. Section II describes the dataset. Section III explains our empirical methodology and Section IV presents regression results. Lastly, Section V presents concluding remarks.

A. Construction of the Dataset
Classification of Public Spending. To conduct analyses on public spending reallocations and growth, we assemble a new disaggregated public spending dataset. Based on historical data reported in the IMF's GFS Yearbook between 1970 and 2011, it classifies public expenditure 12. See Trabandt and Uhlig (2011), for example, for a recent examination of Laffer curves for different taxes. according to its function. 13 One innovation of this dataset is to bridge methodological changes caused by the introduction of the new GFS manual (GFSM2001) from the mid-1990s to the early 2000s. 14 In particular, realizing that GFSM2001 divides total spending into 10 major functional categories, while the old manual, GFSM1986, divides them into 14 categories (see Wickens 2002 for details), we recategorize spending items under GFSM1986 into the items defined by GFSM2001, to make the unified series comparable. 15,16 While following the GFSM2001 categorization, however, we do not attempt to cover all the 10 major components, because data on some of the components are quite scarce. Specifically, our dataset covers the following eight of them: general public services; defense; economic affairs; housing and community amenities; health; recreation, culture, and religion; education; and social protection. 17 Among these, we highlight, largely in accordance with the literature, the following five components: defense; transport and communication (transport, for short), which is a part of economic affairs; health; education; and social protection, while combining the remaining spending into one component (i.e., the remaining three major items and the rest of economic affairs). To note, since data on transport spending, being a subcomponent itself, are sometimes unavailable even when data on economic affairs are available, we do not disaggregate economic affairs unless the reallocation under consideration involves a change in this spending.
13. Disaggregation of public spending can also be based on economic classification, with expenditure allocated by its economic characteristics such as wages, subsidies, social benefits, and capital spending. However, because detailed disaggregation following this classification tends to reduce the sample size drastically, this article focuses on the functional classification instead.
14. The exact years of migration to the GFSM2001 differ depending on the country.
15. Admittedly, the unification process is not exact, but approximate, primarily due to the change in the form in which governments report statistics, from the predominant use of a cash basis under GFSM1986 to the recommended use of an accrual basis under GFSM2001, resulting in the coexistence of both bases under the latter manual. (Under an accrual basis, flows are recorded at the times that transactions occur, independently of the flow of cash; under a cash basis, transactions are recorded when cash effectively flows.) 16. When it happens that spending data are reported based on both accrual and cash bases under GFSM2001, we use the data on the former basis, as it is economically more relevant.
17. The two remaining, relatively small, items are public order and safety, and environmental protection.
Institutional Coverage of the Government. As another notable innovation, we create two versions of the dataset, differing in terms of coverage of the government level. To explain, note first that under GFSM1986, countries report data at most at the consolidated CG level, whereas under GFSM2001, they provide data not only at the CG level but also often at the consolidated GG level, which also covers state and local government levels. 18 With our aim of constructing long data series unifying GFSM1986 and GFSM2001, the first version of the dataset uses data only at the consolidated CG level. However, it is important to realize that spending in certain functional components such as health and education is heavily decentralized in some countries, rendering CG-level data a poor proxy for more comprehensive GG-level data. To address this, we assemble the second version of the dataset. Briefly, the idea is that while using CG-level data as a base, we replace it with GG-level data if the country's public finances are known to be highly decentralized. Specifically, we take the following procedure: (1) we calculate, to the extent the data are available, the ratios of spending at the CG-to GG-levels under GFSM2001 for all the individual components covered in our analyses (as explained above) and take country-level averages; (2) if the average ratio is less than 50%, even in one of these spending items for a country, we abandon the entire CG-level data available for this country and instead use the available GG-level data; (3) while GG-level data under GFSM2001 are used in such cases, when the OECD provides larger GG-level data series, we instead use the data from this alternative source. 19,20 Overall, this alternative dataset, though smaller due to the relative scarcity of GG-level data in general, is less susceptible to the possible effects of decentralization movements. In what follows, the first version of the dataset is denoted CG-based, while the second is CG-GG-based. Table 1 presents the dataset used in the subsequent regression analyses. They follow an 8-year 18. When the data for the different government units are consolidated, statistics are presented as if they constitute a single unit, netting out transfers among them.

B. Description of the Dataset
19. In robustness tests, we use another dataset with the CG/GG ratio threshold value of 75%. 20. Since the OECD follows the same functional classification as GFSM2001, their data are comparable. 9 .7 7 .9 0 .7 3 9 Transport/total spend 5.9 3 .4 0 .8 1 9 .5 6 .2 3 .4 0 .2 1 9 .5 Health/total spend 7.9 5 0 .7 2 3 .5 8 .8 4 .9 0 .9 2 3 .5 Educ/total spend 11.7 6 .1 1 2 9 .6 1 3 .5 5 .7 1 2 9 .6 Soc prot/total spend 25 16 0.8 6 2 .2 2 3 .2 1 5 .7 0 .8 6 2 .2 Rest/total spend 39.9 1 3 .7 1 3 .9 8 2 .4 3 8 .6 1 4 .1 1 3 .9 8 0 .7 Initial GDP p.c. (log) 9.1 1 6 1 0 .9 9 1 .1 6 1 0 .9 Initial level of schooling Notes: Statistics for CG-and CG-GG-based datasets are based on 206 and 180 observations from 82 countries. All the fiscal variables are the initial values in each 8-year period. The initial GDP p.c. is the log of 2005 US$ per capita. Initial level of schooling years are the average years of schooling for population aged between 25 and 64. Investment share and population growth rates are averages over an 8-year period. The figures except for initial GDP and schooling years are in percent. framework, that is, 1972-1979, 1980-1987, … , 2004-2011 (thus yielding five observations per country at most), where all the public spending items take the initial values in each period. 21 The purpose of taking this measure is threefold. First, compared with the framework comprising annual data, this setup is less prone to business cycle effects on spending components (e.g., a possible rise in unemployment benefits in recessionary periods). Second, the lag structure, focusing on the initial spending values, helps mitigate reverse causality issues. For example, it is unlikely that governments' anticipation of higher future growth rates over the next 8 years prompts a change in the spending share of education in the current period. Third, this structure squares with the apparent reality that public spending tends to have delayed effects on growth. Before describing this dataset, one clarification about the lag structure is in order. That is, to deal with the fact that disaggregated spending data is 21. While we examine below a 5-year framework too, we prefer an 8-year framework as a baseline. This is primarily because 5 years may be too short to capture the effects of certain spending components such as education. This choice of period length is in line with, for example, Bleaney, Gemmell, and Kneller (2001), who, using a sample of OECD countries during 1970-1995, show that eight annual lags are required to fully account for the long-run effects of fiscal policy on growth. Nonetheless, it is important to acknowledge that using 8-year averages is not without issues: in particular, it may not be the most suitable way of addressing business cycle effects, since 8 years may be too long for the length of business cycles. unbalanced, we calculate initial spending figures by taking an average over certain years, as long as at least one observation is available over these years. In the reference analyses below, we form initial spending by taking an average over 4 years up to the initial year of each period, that is, for the 1980-1987 period, for example, an average is taken over 1977-1980. 22,23 Table 1 presents the summary statistics, corresponding to the reference analyses below. For brevity, it focuses on the case where transport spending, a subcomponent of economic affairs, is available. Ensuring the availability of all the spending items of interest, the final CG-(CG-GG-) based datasets contain 206 (180) observations from 82 countries. 24 Starting with spending variables in the CG-based dataset, the share of total spending in gross domestic product (GDP) is 27.4% on average. 25 The total is then divided into the respective components of our particular interest. For instance, the mean values of health and 22. For the first period only, we take an average over 3 years, that is, 1970-1972, because the data start in 1970.
24. In case economic affairs spending is not disaggregated, the number of observations is 213 (185), for CG-(CG-GG-) based datasets, from 83 countries.
25. Because the dataset does not cover public order and safety, and environmental protection (due to the severely limited data availability), this total spending figure does not equal the actual total. However, since they generally account for small shares, this figure still approximately equals the actual total. education spending shares in the total are 7.9% and 11.7%, respectively. The average growth rate is 18.7% over the 8 years, corresponding to an annual growth rate of about 2.3%. Further, the table covers the other explanatory variables, including initial GDP, initial level of schooling, investment relative to GDP, and population growth rates, whose rationale is commented on below. 26 Turning to the CG-GG-based dataset, the key difference is found in the different spending shares of certain spending items. Specifically, consistent with the well-known facts that health and education components are often particularly decentralized, these shares in the CG-GG dataset are noticeably larger, that is, 8.8% and 13.5%, respectively.

A. Empirical Specification
Our empirical specification is motivated by neoclassical growth models. The models generally relate the growth of real GDP per capita to two types of variables: state and control/environmental (control, for short) variables (see Barro and Sala-i-Martin 2004). The former variables describe the initial position of the economy, whereas the latter determine the steady state. A key prediction of such models is that when the initial position of the economy is controlled for, an increase in steady state output leads to higher growth rates during the seemingly long adjustment period toward the steady state. 27 Based on this prediction, we examine how public spending reallocations affect the steady state and thus the growth of real GDP, while assessing the role of countries' income levels. 28 26. Detailed data sources of the variables are found in Appendix S2.
27. The steady state growth rate is determined exogenously in these growth models.
28. Endogenous growth models also predict that fiscal policies play a role in the determination of growth rates. For example, they can indicate that the share of "productive" government spending to GDP directly determines the longrun GDP per capita growth rate (e.g., Barro 1990). However, because neoclassical models naturally assign a role to income levels through convergence effects, and this characteristic, in turn, facilitates us to examine the role of income levels in the effects of public spending reallocations through an interaction approach (as seen below), we regard neoclassical growth models as a more suitable theoretical foundation for our analyses.
Formally, we consider the following regression model: Explanatory variables on the right-hand side (RHS) include initial real GDP per capita, y i,t − x , and initial average years of schooling, h i,t − x , as state variables. The former is a convergence variable, while the latter proxies initial human capital. Turning to control variables, s i,j,t − x represents the share of a public spending component, j, in total spending. As explained, spending variables also take initial values. We then have interaction terms between the spending shares and initial GDP per capita. Next, we control for the ratio of total spending to GDP, e i,t − x , to focus on analyses on the growth effects of the composition of spending (as in Devarajan, Swaroop, and Zou 1996). Further, motivated by the Solow growth model, z i,t include additional control variables such as investment rates and population growth rates, as their periodaveraged values. 29,30 ν i captures unobserved country-specific effects, while ξ t is a time dummy, reflecting international shocks during each period.
With Equation (1), we examine the spending reallocations-growth nexus as follows. First, notice that due to exact multicollinearity caused by the fact that ∑ m j=1 s i,j,t−x = 1, we need to proceed by omitting at least one component share from the equation. If, for instance, the share of 29. While we admit that it is more relevant to use labor force growth instead of (total) population growth, using the former variable severely restricts the sample size, particularly for the LICs. Thus, in line with other similar cross-country regressions in the literature which also cover LICs (e.g., Levine, Loayza, and Beck 2000), we use the latter variable.
30. In the robustness checks, we additionally control for demographic variables, specifically the fractions of the population below 15 and above 65 years old. Further, we consider inflation rates, the degree of trade openness, and the ratio of credit extended to the private sector to GDP too.
Importantly, this is the difference of marginal effects of spending component j and m, thus capturing the marginal growth effect of spending reallocations to the former component from the latter, as a function of income levels. In what follows, we examine the effects of different reallocation combinations, while considering the potential role of initial income levels.

B. Estimation Strategy
To estimate the dynamic panel data model of Equation (2), we follow the GMM approach. The reasons for this choice are as follows. First, the framework is flexible enough to accommodate our unbalanced panel. Second, it enables us to handle country fixed effects. 31 Third, it helps us address potential endogeneity of public spending variables through the use of internal instruments (i.e., instruments based on lagged values of those variables), complementing the use of period averages and the lag structure.
While the GMM approach yields consistent estimators, the original "difference" GMM estimators developed by Arellano and Bond (1991) and Holtz-Eakin, Newey, and Rosen (1988) may suffer from finite sample biases. These biases arise particularly when time series are persistent. In fact, Bond, Hoeffler, and Temple (2001) suggest that such biases are likely to be large in the context of empirical growth models, as output tends to be a largely persistent variable. They therefore recommend the use of the alternative "system" GMM estimators 31. To handle fixed effects, we transform variables through "orthogonal deviations" (Arellano and Bover 1995) rather than first differencing, because this transformation maximizes the sample size in our unbalanced panel. developed by Arellano and Bover (1995) and Blundell and Bond (1998), which augment the difference estimator by combining the regression in differences with the regression in levels in a system in which the two equations are separately instrumented. We use this system procedure in what follows. 32,33 Specifically, we treat the state variables of the model, initial output and human capital (y i,t − x and u i,t − x ), and the initial spending variables (e i,t − x and s i,j,t − x ) as predetermined variables, while the other control variables are assumed as endogenous, unless specified otherwise. 34 In what follows, we avoid the problem of instrument proliferation by using only one lag as an internal instrument. 35 To consider the validity of this system GMM approach, we conduct various specification tests. The first is the Arellano-Bond test, whose purpose is to examine the hypothesis that the error term is not serially correlated, which is implicitly assumed in the orthogonality conditions. The second is the Hansen 32. We considered the choice over difference and system GMM estimators explicitly in our context. Specifically, focusing on the coefficient on the lagged dependent variable (as in Bond, Hoeffler, andTemple 2001 andRoodman 2009a), and comparing estimates based on these estimators to the range of estimates set by OLS levels and fixed effects (the former as the upper bound and the latter as the lower bound of good estimates of the true parameter), we found that the point estimate from using difference GMM often lies below the corresponding fixed effects estimate, while the estimate from system GMM largely lies between the range. This suggests that difference GMM is seriously biased downwards, that is, even more biased than fixed effects estimates, while system GMM appears to address this bias observed in difference GMM.
33. Notably, some other works on fiscal policy and growth use the mean-group (MG) and/or pooled mean-group (PMG) estimators developed by Pesaran and Smith (1995) and Pesaran, Shin, and Smith (1999), respectively (e.g., Arnold et al. 2011 andSanz 2011). However, one potential disadvantage of using these alternative approaches is that because the use of annual data is usually required (to have a large number of time series observations), business cycle effects can be more problematic than in our 8-year framework. More fundamentally, however, the fact that our highly disaggregated spending dataset is unbalanced practically prohibits the use of these alternative estimators.
34. That is, with the 8-year framework, we assume that ε i,t , the country-specific time-varying shocks from year t − 7 to t, are uncorrelated with initial GDP, human capital, and fiscal variables in t − 8 and earlier, while they are uncorrelated with the average of the other control variables from t − 15 to t − 8 and earlier.
35. As discussed in Roodman (2009b), the key issue of having "too many" instruments is that it weakens the Hansen test of the instruments' joint validity (mentioned below). Although there is little formal guidance on how many instruments are "too many," a rule of thumb is that the number of instruments should not outnumber individual units in the panel (Roodman 2009a). test, which checks the overall validity of the various instruments of the system. The third is the difference-in-Hansen test, which examines the validity of instruments used in the level part of the system.

A. Spending Reallocation Effects without Income Level Considered
As a benchmark, we first consider the growth effects of spending reallocations without taking account of income levels, that is, by assuming ζ j = 0 for all j in Equation (1). In this case, the regression equation simplifies to (cf. Equation (2)) where a coefficient on spending item j, γ j − γ m , captures the reallocation effect to j from m, independent of income levels. Table 2 presents the estimation results using the CG-based dataset, with the 8-year framework as a reference. For example, in Column (4), which includes the shares of education and the composite of all the items except education and defense (denoted as "Rest excl. financing comp/total") and thus omits defense, the coefficient on education spending captures the growth effect of spending reallocation to education from defense. Specifically, the coefficient of 0.85, being statistically significant at the 5% level, indicates that a 1 percentage point rise in the share of education spending throughout an 8year period, financed by a corresponding fall in defense spending, is associated with a 0.85 percentage points increase in the growth rate over the period, or an approximately 0.11 percentage points annual rise. Columns (5), (6), and (10), which investigate the effects of education spending, when it is financed by health, social protection, and transport spending, respectively, show that reallocations from these components also significantly promote growth. 36 Turning to the 36. In Column (10), the coefficient on transport spending is significantly negative when education spending is omitted. This implies that a spending reallocation toward education from transport fosters growth. remaining combinations, reallocations to defense and social protection from health significantly foster growth (Columns (1) and (3)), while reallocations to defense and social protection, away from transport spending, also exhibit positive growth effects.
Regarding the other variables, the ratio of total spending to GDP mostly has an insignificant growth effect. This happens because the level effect of total spending is offset by a change in other fiscal components to satisfy the government budget constraint. For instance, to the extent that a rise in possibly growthpromoting total spending is financed by a rise in distortionary taxes, the combined effect can be insignificant. Next, the coefficient on the initial GDP per capita, expressed in percent, is significantly negative, being consistent with the conditional convergence hypothesis. 37 The years of schooling, a proxy for initial human capital, has a significantly positive effect in most cases. Lastly, the effect of the ratio of investment to GDP is significantly positive, while the population's growth rate has a negative effect, though insignificant. To note, the diagnostic test results support the use of system GMM estimators. 38 Table 3 summarizes the reallocation effects using the CG-GG-based dataset instead. There, each of the five columns (rows) designates the component whose spending share is increased (decreased) and each cell gives respective reallocation effects. To ease presentation, we only show the sign of the effects, that is, + (−) indicating growth promoting (reducing) effects, with star superscripts in case they are statistically significant. 39 (See Table S1, Supporting Information, for full results.) Notice that the results on education and transport components still stand. That is, the former promotes growth when it is financed by a fall in all the other individual components, while reallocations away from 37. The convergence rates of around 6%-7% over the 8 years imply annual average convergence rates of about 0.8%-0.9%, somewhat smaller relative to the often-reported rates of around 2%. This may be because we consider here a rather heterogeneous group of countries.
38. First, the Arellano-Bond test for AR(2) indicates the absence of serial correlation of the error term. Second, the Hansen test validates the instruments used both in the difference and level parts of the system as a whole. Third, the difference Hansen test, which examines the validity of instruments used in the level part of the system as a whole, also supports the use of system GMM estimators.
39. Notice that focusing on the upper triangular part of the matrix suffices, due to the symmetric nature of the analysis. That is, if increasing item A and decreasing item B promotes growth, decreasing A and increasing B suppresses it.

Notes:
System GMM estimations for dynamic panel data models. Constant and time dummies are not shown for brevity. All explanatory variables were treated as predetermined except for investment rate and population growth rate, which were treated as endogenous. Orthogonal deviation was used to transform variables. Only one lag was used as an internal instrument to reduce the number of instruments. Robust standard errors are in parentheses. All the diagnostic tests present corresponding p values. "Rest excl. financing comp/total" differs across columns (e.g., in Column (1), it is the share of the composite of all the spending items excluding health and defense in the total). ***p <.01, **p <.05, *p <.1.
the latter also show the positive effects, except when reallocated to health. 40 However, the positive effects of reallocations away from health (to defense and social protection), observed with the CG-dataset, are absent, indicating the potential importance of taking into account the effects of decentralization in health spending. 41 Robustness Checks. We now examine the robustness of the above results. First, we change the decentralization threshold from 50% to 75%, meaning that if the ratio of spending at the CG to GG level is less than 75% even in one of the components in a country, we replace the entire CG-level data with GG-level data for the country. This measure mitigates the potential bias from decentralization, albeit it reduces the number of observations further. Second, we form initial spending figures by 3-(and 5-year), instead of 4year, averages. That is, for the 1980-1987 period, for example, an average is taken over 1978-1980 (1976-1980), rather than 1977-1980, as long as at least one observation is available over the years. Third, we control for demographic variables, because while aging societies tend to raise 40. Additionally, we report that education spending tends to be more growth enhancing than the spending items which are not highlighted here. Specifically, we find that spending reallocations to education from general public services and the rest of economic affairs other than transport and communication exhibit significantly positive growth effects, for both the CG-and CG-GG-based datasets. The detailed estimation results are available from the authors upon request. 41. Put simply, differences in results between Table 2 and  Table 3 are only related to the significance of coefficients involving health spending reallocations. That is, while the signs of all possible reallocations are the same across alternative datasets in both tables, only under the CG-based dataset do reallocations to health, from defense and social protection, have a significantly negative effect. Given recent trends in decentralization of health spending, particularly in advanced economies, these differences in results may not be surprising. health and social protection spending, this demographic feature, implying a smaller fraction of the working age population between 15 and 65 years old, can affect growth negatively (e.g., Barro and Sala-i-Martin 2004). Specifically, we add the percentages of the population below 15 and above 65 years old to the reference specification. 42 Lastly, we change the way the entire sample period is divided, examining 5-year frameworks instead (with at most eight observations per country). 43 Results are summarized in Table 4. Overall, while spending reallocations to education from all the other items tend to exhibit growthpromoting effects, the effects appear to be particularly robust when reallocated from health and social protection. Meanwhile, reallocations away from transport spending, toward the other components, may not necessarily exhibit significantly positive effects, except when reallocated to education. Lastly, notice that, although spending reallocations away from health to defense and social protection show some sign of growthpromoting effects, they indeed do not appear robust to the use of GG-based data. 44 42. To avoid the issue of instrument proliferation, we treat these additional variables as exogenous.
43. With the larger time dimension (i.e., eight periods), the issue of instrument proliferation again arises, even when only one lagged value is used as an internal instrument. Therefore, in this case, while still treating investment rate and population growth rate as endogenous (and using one lag for instruments), we use the option collapse of the command xtabond2 for Stata, for these variables.
44. To note, we also examined the robustness by adding the extra controls often considered in growth regressions, namely, inflation rates, the degree of trade openness, and the ratio of credit extended to private sector to GDP. However, since the key results obtained here still stand, the result of this robustness check is not shown for brevity. CG-GG-data with a 75% decentralization threshold Share decreased Defense n/a − − + − * Health n/a + + *** − Socprot n/a +*** −* Education n/a −*** Transport n/a CG (and CG-GG) data with initial spending as 3-year averages Notes: + (−): growth enhancing (reducing). When using CG-GG data with a 75% threshold (and without economic affairs disaggregated), results are based on 83 countries covering 171 observations. For the case with initial spending as 3-year average (initial spending as 5-year average, demographic variables, 5-year periods) using CG-based dataset without economic affairs disaggregated, results are based on 80 (83, 82, 84) countries covering 202 (224, 211, 354) observations. All the underlying estimation results are available from the authors upon request. n/a, not applicable. ***p <.01, **p <.05, *p <.1.

B. Role of Income Levels in the Spending Reallocation-Growth Nexus
Although the above analyses shed some light on the growth effects of public spending reallocations, they may mask the potentially important roles of income heterogeneity across countries. Acknowledging this, we now estimate Equation (2) with the interaction terms present. Figure 1, using the estimation results based on the CGbased dataset, graphically illustrates how the marginal effects of spending reallocations change across the observed range of income levels, for all the different reallocations considered above. (See Table S2 for the estimation results.) For example, the left subfigure in the third row presents the marginal effect (solid line) of a reallocation to education from health, together with the corresponding 90% confidence intervals (dashed lines). Notice that the effects, rather independent of income levels (corresponding to the insignificant coefficient on the interaction term), are significantly positive for a wide range of income levels. Specifically, the effects are significantly positive when the log of real GDP p.c. is between 8.1 and 10.8, roughly corresponding to the middle-to high-income range. (Incidentally, the histogram depicting the distribution of income levels shows that about 80% of the entire observations are located within this region of positive significance.) A similar comment can be made for the reallocation to education from social protection and also from defense, albeit for the former finance source, the positive effects are rather observed over the low-to middle-income range. 45 45. We acknowledge that when countries' income levels are lower, social protection spending tends to be lower and Meanwhile, the subfigures in the fourth and fifth rows describe the reallocation effects toward transport, from all the other components. Notice that the marginal effects do exhibit a distinct positive slope, consistent with the significantly positive coefficients on interaction terms in all the cases. Relatedly, the effects are significantly negative primarily in the low-income region (except for the case when it is financed by a fall in health spending), which indicates that a reallocation away from transport, to items such as education, promotes growth particularly in LICs. Overall, these findings from the interaction analyses highlight the importance of income heterogeneity in the spending reallocations-growth nexus, particularly when transport spending is involved. Lastly, Figure S1 indicates that a similar message can be drawn when using the CG-GG-based dataset, although generally larger standard errors (associated with the smaller number of observations) tend to narrow the income regions where the marginal effects are significant. 46 thus there is smaller room to be reallocated from. However, since our analysis focuses on the marginal effects of reallocations, the results are still valid.
46. Running regressions with interaction terms using the CG-GG-based dataset requires a further restriction of instrument counts, as implied by the Hansen statistic of almost unity for the majority of the regressions. Thus, when running these regressions, we use the option collapse (as mentioned above), for investment rate and population growth rate, while still using only one lag for instruments. The detailed results on CG-GG-based dataset are available from the authors upon request.

Examining
Reallocation Effects Using Subsamples. Next, to check the robustness of the role of income levels, we estimate the growth effects of spending reallocations using subsamples of the entire sample, without interaction terms. Specifically, the following two subsets are considered: one without "poor" countries (57 countries) and another without the "very rich" ones (74 countries). 47 Table 5 summarizes the results on reallocation effects for both subsets. 48 First, the upper subtable shows that when poor countries are excluded, reallocations toward transport spending tend to exhibit positive effects, albeit mostly insignificant, unlike significantly negative effects often observed when considering the entire sample (cf. Table 2  and Table S1). In contrast, the lower subtable 47. We take the following procedure to classify countries into rich and poor. First, for each of the 42 years (1970-2011, the period our fiscal dataset spans), we sort the 183 countries available in the IMF's World Economic Outlook (WEO) according to their GDP per capita level (PPP prices) into two groups: the upper and lower half. Then, counting the number of times each country appears in those groups during the sample period, we classify countries that appear in the upper half most frequently as rich and the remaining countries as poor. Next, to define "very rich" countries, we take a similar approach, sorting the 183 countries into 10 equal groups. Then, we classify countries that appear in the top 10 percentile most frequently as very rich. The reason for choosing this threshold of 10 percentile (say, rather than 25 percentile) is to ensure that sufficient countries/observations are left in the remaining sample.
48. To address instrument proliferation, we use here the option collapse for investment rate and population growth rate. reveals that when using a sample without very rich countries, the negative effects are restored. Indeed, these results appear consistent with the ones from the interaction analyses. Second, reallocations toward education from health and social protection exhibit positive effects in both subsamples, implying that these reallocations tend to have growth-promoting effects over a wide range of countries. Again, this is in line with the results from the interaction analyses.
Further Robustness Checks. We further examined the robustness of the results on the role of income levels, by repeating some of the checks conducted above (cf. Table 4), for both interaction and subsample analyses. 49 Specifically, we examined the use of CG-and CG-GG-based datasets with initial spending as 5-year averages and also the addition of demographic variables. 50 In short, the interaction analyses indicate that the growth-promoting effects of reallocations to education from health and social protection tend to be rather indifferent to income levels, whereas reallocations away from transport to others such as education promote growth primarily in lowincome regions. The subsample analyses yield consistent results.

C. Why Does Income Heterogeneity Matter?
However, why do spending reallocations toward transport, a proxy for infrastructure, have particularly weak growth-promoting effects in LICs, while education does not exhibit such effects in these countries? Our basic conjecture is that, while both spending items have potentially strong growth-promoting effects (e.g., the former through promoting private firms' productivity; the latter through enhancing human capital), the efficiency of infrastructure spending may be particularly susceptible to the low quality of governance often observed in these countries. 51 In our view, there may be at least two bases for this conjecture.
To explain the first, it is useful to acknowledge that infrastructure projects tend to be capital intensive, unlike education spending for which 49. All the test results are available from the authors upon request.
50. Given that the interaction approach requires more predetermined variables, that is, interaction terms, and that the subsample analyses severely restrict the sample size, certain tests, particularly the ones causing the issue of instrument proliferation, are difficult to conduct. 51. We define governance as the general manner in which authority is exercised in a country. its largest share can be rather classified as current spending, such as public wages for teachers. Then, as emphasized by Keefer and Knack (2007) and Tanzi and Davoodi (1997), since capital spending is often a convenient vehicle for government officials' rent seeking (e.g., in the form of commissions from private firms seeking for public capital projects), the efficiency of this type of spending can easily be compromised, particularly when officials are not properly constrained. Meanwhile, in the case of wages, there is less room for rent seeking by officials, because they tend to be predetermined outside officials' discretion. Therefore, the degree of political constraints, a key aspect of governance quality, may influence the effectiveness of infrastructure spending in a particularly critical manner. Second, the quality of bureaucracy in general, another key aspect of governance, can be a particularly important determinant of the efficiency of infrastructure spending. This is because, in general, a proper conduct of infrastructure projects often involves a particularly long-lasting, potentially complicated, management processes, including appraisal, selection, implementation, and evaluation (see Dabla-Norris et al. 2012 for detailed public investment management processes). Therefore, for a project to be carried out efficiently, a high-quality bureaucracy, which is capable of managing these potentially complicated processes properly, is often required. To clarify, while other public spending components should also benefit from such strong bureaucracies, the particular complexity of infrastructure projects may render their successful completion more dependent on their presence.
Is there any evidence for this conjecture? While providing thorough evidence is beyond the scope of this article, we present here suggestive evidence for the possible importance of governance quality, particularly in the effects of infrastructure spending. For this purpose, we consider the following interaction model similar to Equation (1): where the key difference is that this model rather highlights interactions between the shares of spending components, s i,j,t − x and the governance variable, g i,t − x . Following works such as Keefer and Knack (2007), the governance level is proxied by a composite of institutional indicators provided by the International Country Risk Guide (ICRG). 52,53 To proceed, we address again multicollinearity by omitting (at least) one spending share. For simplification, we do not specify the financing factor here. 54 That is, by defining item m as the composite of all the spending items except item j, and omitting this composite, we estimate Then, the marginal growth effect of spending share j, (γ j − γ m ) + (ζ j − ζ m )g i,t − x , captures the effect of spending j, relative to the one of the composite of all the remaining items, m, as a function of the governance level. Figure 2 describes how the effects of education and infrastructure (transport) spending change across the observed range of governance levels, where a larger value corresponds to the better quality of governance. (See Table S3 for the underlying estimation results.) The lower subfigures suggest that for both CG-and CG-GG-based datasets, the effects of infrastructure spending, relative to all the other spending items, are increasing in the governance quality, consistent with the significantly positive coefficients on the interaction terms in the respective regressions. Relatedly, the effects are negatively significant in the low-quality governance region. Meanwhile, the upper subfigures show that the growth-promoting effects of education spending 52. The indicators used are law and order, corruption, bureaucracy quality, and investment profile. The governance proxy is then created as the unweighted addition of these indicators. (We used investment profile, as a substitute for "the risk of expropriation and of repudiation of contracts by government" used by Keefer and Knack 2007, since this variable is no longer available in the recent versions of the dataset.) A similar ICRG index is also used by Shi and Svensson (2006), particularly to capture the level of rents extracted by politicians while in power.
53. Consistent with the spending shares, we use initial governance proxies in each period. To note, since ICRG index often becomes available only since 1984, to ensure that we cover here at least four (8-year) periods since the 1980-1987 period, we extrapolate the oldest available figure to 1980, by assuming that governance level did not change over the first few years.
54. This is to deal with the fact that the use of ICRG makes country coverage and observations smaller. are rather unaffected by governance quality, in line with the insignificant coefficients on the interaction terms. 55 Thus, the implication is that the above finding of weak growth effects of infrastructure spending in LICs is potentially driven by low-quality governance, as often observed in these countries.

V. CONCLUDING REMARKS
Having examined the effects of public spending reallocations on growth, we found that reallocations toward education, from health and social protection, robustly promote growth across a wide range of countries' income levels. However, income heterogeneity matters in the spending reallocations-growth nexus, particularly when infrastructure spending is involved. Specifically, only when income levels are low, reallocations away from this spending, to items such as education, enhance growth. This happens because while the growth-promoting effects of education spending tend to prevail across different income levels, the effects of infrastructure spending exhibit a positive association with them. However, the fundamental reason behind why infrastructure spending has particularly weak growth-promoting effects in LICs may be that inefficiencies caused by their low quality of governance, exemplified by unconstrained officials' rent-seeking behavior, mitigate the positive effects of this spending.
Our findings have various policy implications. First, while governments may attempt to foster growth through public spending reallocations under different circumstances (say, when adopting fiscal austerity measures), at least from a growth perspective, they may wish to increase education spending and instead curb spending in other categories such as health and social protection. Next, particularly in LICs, increasing infrastructure spending through reallocations does not appear to be a growth-promoting policy. Rather, to the extent that low-quality governance is weakening the effects of this spending in these countries, promoting institutional reforms to restrain officials' rent-seeking opportunities in public investment management may be all the more beneficial to them. 55. Further, we report that none of the effects of the other spending items, that is, defense, health, and social protection, exhibits a clear positive association with governance level, as infrastructure spending does. However, one important caveat is in order regarding policy implications. That is, despite its undeniable importance, economic growth is not the only criterion that a government utilizes in determining the composition of public spending. While this article focuses on growth, there are other potential elements, such as employment and income inequality, that are of critical importance. Indeed, although social protection spending may not be growth enhancing, it may help reduce income inequality. Therefore, from a broader policy perspective, it is important to consider the effects of spending reallocation policies on these other key elements of economic well-being as well. A careful examination of these effects thus appears to be a fruitful topic for future research.

FIGURE 2 Effects of Spending Reallocations across Governance Levels
Finally, another possible topic for future research is to carefully examine the role of political and economic freedom in the relation between public spending and growth. It is possible that these freedom concepts, the former defined as the general public's rights which guarantee their own political participation, and the latter as their rights to pursue their own economic decisions without excessive regulations, may play a different role in the growth effects of different spending items. For example, while this article suggests that the low quality of governance is associated with the weak growth effects of infrastructure spending, this may also indicate the important role of political freedom in this association. Further investigation of the possible role of these freedom concepts may shed brighter light on the conditions under which different public spending policies have growth-promoting effects.

SUPPORTING INFORMATION
Additional Supporting Information may be found in the online version of this article: Appendix S1. List of the 83 countries in reference regressions Appendix S2. Data sources Appendix S3. Supplementary results Table S1. Spending reallocations and growth: using CG-GG-based dataset Table S2. Role of income levels in reallocations-growth nexus: using CG-based dataset Table S3. Effects of spending reallocations across governance levels Figure S1. Effects of spending reallocations across income levels: using CG-GG-based dataset