This paper's findings, interpretations and conclusions are entirely those of the authors and do not necessarily represent the views of the World Bank, its Executive Directors or the countries it represents.

Abstract

This paper addresses two challenges that the fiscal decentralisation and economic growth nexus faces – namely, endogeneity problems and inaccurate measurement of fiscal decentralisation. We introduce novel instrumental variables based on common legal system origin, common federal system, geographical position and relative country size. The positive relationship between fiscal decentralisation and economic growth that we find remains valid when using these instrumental variables. Using fiscal decentralisation measures that better reflect the autonomy of subnational governments changes this relationship. This finding, however, is the result of the accompanying changes in the sample rather than the use of these alternative measures themselves.

Evidence for economic growth benefits of fiscal decentralisation is on a stronger footing when novel instrumental variables are used.

Benefits of fiscal decentralisation are unaffected when using metrics that better reflect local governments’ policymaking authority.

I. Introduction

Over the past decades, many countries became increasingly open to the idea that the provision of government services should be subject to the principle of subsidiarity. Under this principle, the provision of these services should be performed by the most decentralised government level that is capable of delivering them effectively. The hope is that the proximity of these subnational governments to citizens gives them an information advantage over national governments regarding local preferences, resulting in an improvement of allocative efficiency in the public sector.1 Both developing and developed countries have embraced this principle of subsidiarity and devolved parts of their fiscal policymaking authority to subnational levels of government.

This process of fiscal decentralisation has led to a wide array of studies that look at its possible consequences, and its relationship with economic growth is often analysed.2 Despite the interest in this relationship, there are two relevant issues that have received little attention in the literature so far. To begin, there is the possibility of reverse causality, where fiscal decentralisation affects economic growth but at the same time the reverse holds as well. Together with the potential omitted variable bias, where fiscal decentralisation measures capture some other effects that are not controlled for in the estimation procedure, these sources of endogeneity are the first issue. The second issue is the dissatisfaction with the accuracy with which conventional government expenditure- and revenue-based measures of fiscal decentralisation reflect the true decision-making authority of subnational governments. When these issues are ignored, it is impossible to say what the relationship between fiscal decentralisation and economic growth truly is. This paper addresses both of these issues.

To address the possible sources of endogeneity, we propose instrumental variables for fiscal decentralisation that are based on country characteristics such as the legal system origin, the federal system, country size and geographical position. Our conjecture is that countries that are similar in these characteristics experience not only a similar degree of fiscal decentralisation for a given point in time but also a similar process of fiscal decentralisation. For example, fiscal decentralisation is expected to be alike in countries that are very similar in size, while for countries that are vastly different in size it is not expected to be related. Practically, our conjecture means that rather than using these country characteristics as instrumental variables themselves – a more conventional approach taken in the literature by, for example, Fisman and Gatti (2002) – we use them as a basis for constructing weighted averages of the fiscal decentralisation measures and use these as our instrumental variables. Besides differing from the aforementioned approach, ours also differs from studies that use lagged observations of fiscal decentralisation measures as instrumental variables – for example, Iimi (2005). To our knowledge, we are the first to use this alternative approach and to show whether and how it differs from these others.

To make the fiscal decentralisation measures better reflect the true decision-making authority of subnational governments, we propose measures that incorporate information on the tax autonomy of subnational governments that is collected by the Organisation for Economic Cooperation and Development (OECD). These alternative measures differ from the more traditional measures used in the literature that are solely based on the government expenditure and revenue data from the Government Finance Statistics of the International Monetary Fund (IMF). Martinez-Vazquez and McNab (2003) argue that these traditional measures may not capture all dimensions of fiscal decentralisation, such as the degree of discretion that subnational governments have over expenditures and taxes. For example, if half of a country's government expenditures take place at the subnational government level, then looking at this share over-represents the true degree of fiscal decentralisation when a large part of these expenditures are actually mandated by the national government. While Ebel and Yilmaz (2002) and Thornton (2007) adjust their fiscal decentralisation measures in a similar way, it remains unclear how the relationship between fiscal decentralisation and economic growth changes when using these alternative measures.3 We contribute to the literature in this aspect by clearly identifying whether, how and why results change.

Using a sample from up to 56 countries over the period 1990–2007, we find that fiscal decentralisation is positively related with economic growth. Quantitatively, a one standard deviation change in fiscal decentralisation is associated with, on average, half a percentage point change in economic growth, or one-sixth of a standard deviation.

This relationship remains valid when using our proposed instrumental variables to deal with the endogeneity problems. In addition, we find that the hypothesis of fiscal decentralisation being exogenous cannot be rejected. Together, these results speak in favour, but are not definite proof, of the causality running from fiscal decentralisation to economic growth. More importantly, these results cannot be established by employing the other instrumental variables used in the literature. The instrumental variables that are based on the country characteristics themselves are too weak to be valid.4 Using the instrumental variables that are based on lagged observations of the fiscal decentralisation measure themselves leads to a reduction in the sample size to 37 countries, and the above relationship is much weaker for the reduced sample. Apart from this selection bias, it is not possible to test for the correlation of the instrumental variables with the error terms, something which can be done with our instrumental variables. This means we cannot rule out that the lagged observations of fiscal decentralisation capture some other effect that is not controlled for in the estimation procedure. Hence, our instrumental variables outperform those that are traditionally used in the literature.

When correcting the fiscal decentralisation measure for the tax autonomy data of the OECD, we no longer find the previously-described positive relationship between fiscal decentralisation and economic growth. This finding, however, is not caused by the use of these alternative fiscal decentralisation measures, but is explained by the accompanying changes in the sample. For example, when using the sample of 27 countries for which the alternative measures are available, we find no evidence to support fiscal decentralisation being related with economic growth, regardless of the fiscal decentralisation measure used. For our samples and alternative measures, using traditional government expenditure- and revenue-based fiscal decentralisation measures leads to the same qualitative results as using measures that better reflect the true degree of fiscal autonomy of subnational governments.

The remainder of this paper discusses these findings in more detail and is organised as follows. Section 'Literature' presents a brief overview of the literature. Section 'Empirical methodology and data' discusses our econometric specification, the data, our approach to dealing with endogeneity problems and various measures of fiscal decentralisation. Section 'Estimation results' presents the results and Section 'Conclusions' concludes.

II. Literature

This section provides a short overview of the literature on the relationship between fiscal decentralisation and economic growth to see whether it provides guidance on the choice of the framework in which to address the issues laid out in the introduction. Hereby, we focus on studies that look at the relationship in a cross-country context. There is no consensus on the relationship between fiscal decentralisation and economic growth, despite the attention it has received in the literature. There are almost equal numbers of studies that find evidence for a positive, a negative or no relationship. It is possible that the divergence in these outcomes can be explained in part by differences in the number or type of countries, the time period of the analysis, the estimation method or the empirical specifications used (see Table 1).5 However, the many characteristics that set them apart make it hard to single out a preferred framework. Selecting a framework is further complicated by the fact that little attention is paid to quantifying the relationships found. Many studies focus solely on the sign and significance of the fiscal decentralisation coefficient, a possible non-linear effect or how the relationship depends on the interaction with other explanatory variables. This focus, although not necessarily unimportant, prevents a clear comparison of the quantitative findings corresponding to the frameworks used.

Table 1. Overview of the fiscal decentralisation and economic growth literature

‘Panel’ refers to analyses that allow for country fixed effects and ‘OLS’ to analyses that do not, even though they may have a panel structure.

^{b}

Developing countries are indicated with ‘D’ and industrialised countries with ‘I’.

^{c}

Correction for autonomy is based on the use of data from OECD (1999), Blöchliger and King (2006) and Blöchliger and Rabesona (2009) or from Stegarescu (2005).

Note: The table focuses on the direct effect and abstracts from any interaction effects.

The overview in Table 1 illustrates the limited attention to endogeneity problems. There are only a few studies that use instrumental variables to address the problem of reverse causality. Both Iimi (2005) and Martinez-Vazquez and McNab (2006) use an internal instrumental variable – the lagged observations of the fiscal decentralisation measures themselves – where the former focuses on the variation in fiscal decentralisation between countries and the latter on the variation within countries. Enikolopov and Zhuravskaya (2007), however, use an external instrumental variable – the geographical area of a country – and focus on the variation in fiscal decentralisation between countries. A thorough analysis of the endogeneity problems, where all of the different approaches are discussed, is missing though.

There are studies that correct the fiscal decentralisation measures for the autonomy of subnational governments. Baskaran and Feld (2013), for example, find no relationship between fiscal decentralisation and economic growth when using the conventional government expenditure- and revenue-based measures of fiscal decentralisation. However, they find evidence for a negative relationship once they use the autonomy-corrected measure. Thornton (2007) finds no relationship when using the autonomy-corrected measure, but he does not use the conventional fiscal decentralisation measure. These different findings and approaches show that it is still unclear how correcting conventional measures for the autonomy of subnational governments affects the relationship between fiscal decentralisation and economic growth.

Overall, the empirical literature provides little guidance on selecting our framework. The same holds for the theoretical literature. There are no formal, mathematical models that capture the subsidiarity principle and from which it is possible to directly derive an empirical specification as is done by, for example, Mankiw, Romer and Weil (1992) and Hauk and Wacziarg (2009).6 Hence, the literature on fiscal decentralisation and economic growth does not point to a specific framework that can be used to analyse either the different approaches that address endogeneity problems or the effect of using fiscal decentralisation measures that take into account the policymaking authority of subnational governments.

III. Empirical methodology and data

This section discusses the econometric specification, the data, our approach to addressing possible endogeneity problems and how we correct fiscal decentralisation measures for the autonomy of subnational governments. We use data from up to 56 countries over the period 1990–2007, where variables are three-year averages unless otherwise indicated. Descriptive statistics, the corresponding data sources and sample composition are given in Tables A1–A3 in the online appendix.

1. Econometric specification

Given the lack of guidance in the literature on selecting a framework as discussed in Section 'Literature', we use a ‘Barro-style’, non-formally derived, growth regression:

git=ϕfit+ψ′xit+γ′zit+εit,(1)

where countries are denoted with i = 1,…,N and time with t = 1,…,T. The dependent variable g_{it} denotes the growth rate of real GDP per capita of country i at time t. Fiscal decentralisation is denoted by f_{it} and ϕ is the corresponding parameter. The matrix x_{it} contains explanatory variables that are often used in growth regressions and the matrix z_{it} contains additional explanatory variables, where ψ and γ are the corresponding vectors of parameters.7 The parameter ε_{it} is an independent and identically distributed (i.i.d.) error term.

The matrix x_{it} consists of a constant, the logarithm of initial income per capita, population growth, investment and schooling. Investment is measured by the logarithm of the share of gross fixed capital formation in output, and schooling is defined as the logarithm of the product of gross secondary school enrolment and the population share of secondary school age. These variables are often found to be related with economic growth.8

The matrix z_{it} contains explanatory variables such as the federal system, government size, trade openness, dummies for regions – East Asia & Pacific, Eastern Europe & Central Asia, Latin America & Caribbean, Africa and OECD member countries – and dummies for time, which represent three-year periods starting in 1990. The federal system variable has the value 1 if a country is defined as federal and 0 otherwise. Government size is measured by the logarithm of the share of general government final consumption expenditures in output. Trade openness is the sum of exports and imports of goods and services expressed as a share of output.

We define our main measure of fiscal decentralisation as the share of subnational government expenditures in general government expenditures based on data from IMF (2010), the Government Finance Statistics. Subnational expenditures are defined as expenditures at both the state and local government levels, where the state level refers to the largest geopolitical entity within a country and the local level describes the smaller governmental units below the state level.9 General government expenditures encompass public expenditures at the central, state and local government levels together.

2. Endogeneity problems

When estimating equation (1) by ordinary least squares (OLS), there is the possibility of reverse causality, where fiscal decentralisation affects economic growth but at the same time the reverse also holds. Another possibility is that measures of fiscal decentralisation capture some other effects that are not controlled for in the estimation procedure. When these endogeneity problems are ignored, it is impossible to say what the relationship between fiscal decentralisation and economic growth truly is. However, they can be addressed by using instrumental variables that are correlated with the fiscal decentralisation measure f_{it} but not with the error term ε_{it}.

We propose instrumental variables for fiscal decentralisation that are based on country characteristics such as the legal system origin, country size, the federal system and geographical position. The conjecture is that countries that are similar in these characteristics experience not only a similar degree of fiscal decentralisation for a given point in time but also a similar process of fiscal decentralisation. More specifically, the instrumental variable for a particular country is a weighted average of the fiscal decentralisation measures of all other countries in the sample, where the weights are determined by the similarity of these countries to that particular country. This similarity argument builds on the tax competition literature that uses neighbourliness arguments to construct instrumental variables in the same way.10

Thus, we can define the instrumental variable for fiscal decentralisation of country i at time t as follows:

qit=∑j=1Nωijfjt,ωij≡dij/∑j=1Ndij for i≠j0 for i=j,(2)

where f_{jt} denotes the fiscal decentralisation measure of country j at time t, ω_{ij} denotes a weight and d_{ij} measures the similarity of countries i and j. The weights are normalised so that they lie in the closed interval [0,1] and sum to unity. Since the instrumental variables are based on a weighted average of other countries in the sample, missing observations in a single time period may lead to an inconsistency in the number of countries used in their construction. Therefore, we balance the sample by interpolating the fiscal decentralisation variable for the missing values so that the same number of countries is used in every period when constructing the instrumental variables.11

We use some of the same country characteristics as Fisman and Gatti (2002) and Enikolopov and Zhuravskaya (2007), although we use them as a basis for weights when constructing the instrumental variables rather than as instrumental variables themselves. Our instrumental variables are likely to be more strongly related to fiscal decentralisation than the country characteristics themselves since our instrumental variables are based on more information.

Using the same rationale as Fisman and Gatti (2002), we base one instrumental variable on whether countries have a common legal system origin; it takes the value 1 if countries have the same legal system origin and 0 otherwise. Based on La Porta et al. (1999), we define the following main categories of legal origin: British, French, German, Scandinavian and socialist. The choice for the legal system origin can be motivated by Fisman and Gatti (2002), who find that ‘the proportion of public expenditures accounted for by state/local governments is much lower in French origin countries than in British origin countries’ (p. 337). This observation, they argue, is in line with the affinity of a civil legal code for government centralisation.

Using the same rationale as Enikolopov and Zhuravskaya (2007), we base another instrumental variable on the relative size of countries i and j, which is given by

sij=si+sjsi+sj+ξsi−sj,(3)

where s_{i} and s_{j} denote the geographical area in square kilometres of country i and country j respectively and ξ ∈ [0,∞) is the weight of the absolute size difference. Panizza (1999, p. 104) argues that the distance of the national government from its citizens increases with the size of a country, making it less likely that the national government's policies are in line with the citizens’ preferences and strengthening the case for increased decentralisation.

The remaining instrumental variables are first based on having the same federal system, which corresponds to the definition of whether a country is federal or not12 and is motivated by Lijphart (1984, p. 176), who argues that ‘federalism and decentralization tend to go together’. Second, we consider an instrumental variable based on geographical position. More specifically, we use the inverse of the squared distance between the main cities of countries i and j as a measure of similarity. This captures the idea that governments set their rate of fiscal decentralisation close to that of their neighbouring countries, either by mimicking policies or by sharing similar geographical features.

Table 2a displays the pairwise correlation coefficients of our fiscal decentralisation measure and our instrumental variables. The instrumental variables based on countries with the same federal system and based on the relative size of countries display the strongest correlation with fiscal decentralisation, where the correlation coefficients of the latter become larger when increasing the weight of the absolute size difference, ξ. The instrumental variables based on common legal system origin and distance squared are less strongly correlated.

Table 2a. Correlation coefficients of instrumental variables

Variable

(2)

(3)

(4)

(5)

(5)

(5)

(5)

(5)

ξ=5

ξ=10

ξ=25

ξ=50

ξ=100

Note: The table shows pairwise correlation coefficients of fiscal decentralisation and potential instrumental variables. ^{***}, ^{**} and ^{*} indicate that the coefficient is statistically different from zero at the 0.01, 0.05 and 0.10 levels respectively. Correlation coefficients are calculated using the sample of 56 countries over the period 1990–2007.

Table 2b. Correlation coefficients of fiscal decentralisation measures

(2)

(3)

(3)

(3)

(2)

(3)

(2)

(3)

I

II

III

II

III

Note: The table shows pairwise correlation coefficients of several measures of fiscal decentralisation. ^{***}, ^{**} and ^{*} indicate that the coefficient is statistically different from zero at the 0.01, 0.05 and 0.10 levels respectively. Tax autonomy measures of 1995, 2001 and 2005 are used for the periods 1993–97, 1998–2002 and 2003–07 respectively. ‘I’ takes all categories of tax autonomy into account; ‘II’ takes all categories related to discretion on rates, bases and reliefs into account; and ‘III’ takes only the category with full discretion on rates, bases and reliefs into account. See Table A4 in the online appendix for an overview of all categories.

Countries with a similar federal system, geographical area, legal system origin and geographical position thus appear to have a similar process of fiscal decentralisation. Figure 1 shows the positive relationships between our constructed instrumental variables and fiscal decentralisation. It illustrates that the unexplained variation of fiscal decentralisation differs across the instrumental variables. While individually their variation only explains up to 30 per cent of the variation in fiscal decentralisation, we can use a combination of them. This allows us to test for the correlation of the instrumental variables with the error terms. We think there is no mechanism or omitted variable captured in the error term that affects economic growth of a country and at the same time is related to a weighted average of the fiscal decentralisation measures of all other countries in the sample. Hence, our constructed instrumental variables are likely to be related with fiscal decentralisation but not with the error term, hereby satisfying the conditions of valid instrumental variables.

When addressing the endogeneity problems, we do not control for country fixed effects in our estimation procedure. While we acknowledge that including them deals perfectly with a possible omitted variable bias, we prefer not to do this since it may come at a cost in the presence of measurement error. Hauk and Wacziarg (2009) show that controlling for fixed effects exacerbates the bias arising from measurement error when the time persistence in the covariates is larger than that of the errors in measurement.13 When looking at growth regressions, the authors find that, based on their simulations, cross-sectional estimates are preferred to panel estimates that take into account country fixed effects. They also look at other estimators that, in theory, deal with the omitted variable bias, the measurement error bias and biases arising from reverse causality.14

3. Alternative fiscal decentralisation measures

As in most studies in the fiscal decentralisation and economic growth literature, our main measure of fiscal decentralisation is based on the government expenditure and revenue data from IMF's Government Finance Statistics. Martinez-Vazquez and McNab (2003) criticise this type of measure for not accurately reflecting the true decision-making authority of subnational governments. Thornton (2007) follows this point up and constructs a measure that takes the policymaking authority of subnational governments into account. More specifically, he uses the share of subnational government tax revenue in total government tax revenue and multiplies this measure by an indicator of tax autonomy. This indicator of the subnational government's tax autonomy is based on data from OECD (1999), which distinguishes several categories of tax autonomy, ranging from full discretion on tax rates and reliefs (i.e. credits and allowances) to no discretion on rates and reliefs at all.

Using data from OECD (1999), Blöchliger and King (2006) and Blöchliger and Rabesona (2009), we follow Thornton (2007) and construct our own autonomy-based measures of fiscal decentralisation. More specifically, we use the share of subnational government tax revenue in total government tax revenue and only consider the tax revenue at the subnational government level over which the corresponding government has autonomy. Since there is no clear definition of or consensus on the policymaking authority of subnational governments, we examine three different cases in which we give the tax autonomy categories of the OECD different weights, where in each case we increase the number of restrictions on the autonomy of subnational governments. Table A4 in the online appendix provides an overview of the categories and weights.15

Table 2b presents pairwise correlation coefficients of the different fiscal decentralisation measures. The conventional measures of fiscal decentralisation based on government expenditure and revenue data from the IMF's Government Finance Statistics are highly correlated. Moreover, this correlation remains strong after we correct the tax-revenue-based measure using the tax autonomy data from the OECD. The strength of the correlation declines somewhat with the number of restrictions we impose on the autonomy that subnational governments have. The weaker correlation is caused by countries that become fully centralised when increased restrictions on the autonomy of subnational governments are imposed. Excluding these fully-centralised countries reduces the sample size but increases the correlation between the conventional and autonomy-corrected fiscal decentralisation measures. Our main measure of fiscal decentralisation thus seems a strong approximation for fiscal decentralisation measures that better reflect the true degree of fiscal autonomy of subnational governments.

IV. Estimation results

This section discusses the estimation results. In the analyses, we address possible endogeneity problems and look at various measures of fiscal decentralisation that differ in the degree to which they reflect the autonomy of subnational governments.

1. Endogeneity problems

Table 3 presents our main estimation results and focuses on the comparison of our instrumental variables with those used in the literature. The first column in panel a regresses the growth rate of real GDP per capita on fiscal decentralisation, initial real GDP per capita, population growth, investment, schooling and a constant. We estimate the equation by ordinary least squares (OLS), where robust standard errors are clustered at the country level since we could not reject the error terms being serially uncorrelated. Initial real GDP per capita and population growth both have a negative and significant coefficient, and investment and schooling both have a positive and significant coefficient. These results are in line with most of the growth literature.16 We find no evidence for a relationship between fiscal decentralisation and economic growth in this specification. However, the coefficient of fiscal decentralisation is positive and significant in the second column, where we add federal system, government size and trade openness. Federal system has a negative and significant coefficient, and we find no evidence that the size of the government is associated with growth in real GDP per capita, but trade openness is positively associated with it.

Table 3. Fiscal decentralisation and economic growth

Panel a. Preferred instrumental variables

Dependent variable: growth rate of real GDP per capita

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

OLS

2SLS

Note: ^{***}, ^{**} and ^{*} indicate that the coefficient is statistically different from zero at the 0.01, 0.05 and 0.10 levels respectively. Robust standard errors are clustered at the country level. F statistics for the serial correlation test, as described by Drukker (2003), for columns 1–4 are 11.02, 8.794, 8.794 and 8.235 respectively, which means a rejection of the null hypothesis of no serial correlation in all cases.

We add regional dummies and time dummies in the third and fourth columns respectively, where OECD countries and the period 2005–07 are the bases. The estimation results remain the same in sign and significance except for the coefficients of population growth, schooling and federal system, which are no longer significant. Countries belonging to East Asia & Pacific, Latin America & Caribbean, and Africa grow at a lower rate than OECD countries. Compared with the period 2005–07, growth rates are lower over the periods 1990–92 and 1999–2004. Based on the estimation results of columns 2–4, the relationship between fiscal decentralisation and economic growth is as follows: a one standard deviation increase in fiscal decentralisation is associated with an increase in the growth of real GDP per capita of, on average, half a percentage point, or one-sixth of a standard deviation.17

We estimate these same specifications by two-stage least squares (2SLS) in columns 5–8, where we use our instrumental variables based on common legal system origin and relative country size, with ξ = 50. Following the criterion of Staiger and Stock (1997), these instrumental variables are valid since the joint significance of the instrumental variables in the first-stage regression is large enough; as a rule of thumb, the corresponding F statistic should be larger than 10. The estimation results are similar in sign and significance for almost all variables, and the point estimates for fiscal decentralisation are somewhat higher. In the specification where we include both region and time dummies, the coefficient of fiscal decentralisation is no longer significant at conventional levels.18

The regression-based test of Wooldridge (1995) is used to test whether fiscal decentralisation is endogenous. When the regression-based F statistic is significant, we can reject the null hypothesis of fiscal decentralisation being exogenous and must treat it as an endogenous variable. The robust score test of overidentifying restrictions of Wooldridge (1995) is used to test whether the instrumental variables are uncorrelated with the error term. When the χ^{2} statistic is significant, the instrumental variables used may not be valid.19 In all of the specifications in columns 5–8, the p values are too large for these statistics to be significant. Hence, our instrumental variables are valid according to the criteria in the literature, and we still find a positive relationship between fiscal decentralisation and economic growth. Moreover, we find that the hypothesis of fiscal decentralisation being exogenous cannot be rejected. Together, these results speak in favour, but are not definite proof, of the causality running from fiscal decentralisation to economic growth.

We now look at how these results change when using the instrumental variables that are used in the literature on fiscal decentralisation and economic growth. We start with the use of lagged observations of our fiscal decentralisation measure as an instrumental variable. This approach leads to the loss of one or more cross-sections of data, and the loss of observations is quite large when the sample is unbalanced as it is in our case. Panel b of Table 3 repeats the analyses of panel a with the reduced sample, where the number of observations is reduced from 201 to 126 and the number of countries from 56 to 37. The estimation results are similar for most variables, although we no longer find evidence for a relationship between fiscal decentralisation and economic growth when estimating the equations with OLS. It is very likely that it is not the reduction in observations, but rather the selection of countries, that leads to this finding since we find similar results when repeating the analyses of columns 1–4 using just the 37 countries of the reduced sample.

Columns 13–16 give the estimation results when we estimate the equations by 2SLS using the lagged observations of fiscal decentralisation as the instrumental variable. While we find evidence for a positive relationship between fiscal decentralisation and economic growth, the coefficient of fiscal decentralisation is only significant at the 10 per cent level. As with our instrumental variables, point estimates of the fiscal decentralisation coefficients are somewhat higher, and we cannot reject the null of fiscal decentralisation being exogenous. However, we are unable to test whether the instrumental variables are uncorrelated with the error term since we only have one instrumental variable. This means that, with this approach, we cannot rule out that the lagged observations of fiscal decentralisation capture some other effect that is not controlled for in the estimation procedure.20

The first four columns of panel c show the estimation results when we use legal system origin and country size themselves as instrumental variables for fiscal decentralisation. The values of the F statistic that represents the joint significance of the instrumental variables in the first-stage regression never satisfy the criterion of Staiger and Stock (1997). The instrumental variables that are based on the country characteristics themselves are thus too weak to be valid. This means that the use of the country characteristics as weights in the construction of instrumental variables is preferred over the use of these characteristics as instrumental variables themselves. The last four columns of panel c present the estimation results corresponding to our instrumental variables based on common federal system and distance squared. With the exception of column 24, where the instrumental variables outperform those based on common legal system origin and relative country size, these instrumental variables are also too weak to be valid instruments.

2. Alternative fiscal decentralisation measures

Table 4 shows whether and how the relationship between fiscal decentralisation and economic growth changes when alternative measures of fiscal decentralisation are used. We regress the growth rate of real GDP per capita on initial real GDP per capita, population growth, investment, schooling, federal system, government size, trade openness, a constant and a measure of fiscal decentralisation. In the first column, this measure is defined as the share of subnational government tax revenue in total government tax revenue. Tax revenue at the subnational government level is corrected for tax autonomy in columns 2–4. We consider all tax autonomy categories of the OECD in the second column, where lower weights are given to categories with low degrees of autonomy. In the third column, we restrict subnational tax revenue to those taxes for which subnational governments have full discretion on rates, bases and reliefs and those for which they have full discretion on rates only. We only consider tax revenue for which subnational governments have full discretion on rates, bases and reliefs in the fourth column. Our standard measure of fiscal decentralisation based on expenditure data is used in column 5 as a counterfactual. In all cases, we find no evidence for a relationship between fiscal decentralisation and economic growth.

Table 4. Fiscal decentralisation and economic growth: including autonomy measures

Dependent variable: growth rate of real GDP per capita

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Case:

I

II

III

II

III

Note: ^{***}, ^{**} and ^{*} indicate that the coefficient is statistically different from zero at the 0.01, 0.05 and 0.10 levels respectively. Robust standard errors are clustered at the country level.

When we consider only those countries with positive ratios of fiscal decentralisation after controlling for tax autonomy (columns 6–9), our sample reduces in size. This reduction in sample size changes the results in column 8 compared with column 4, and we now find evidence for a positive relationship between fiscal decentralisation and economic growth. However, we find the same relationship when using our main measure of fiscal decentralisation based on expenditure data and the smaller sample in column 9.21 Overall, these results imply that the discrepancy in these findings is not caused by the use of these alternative fiscal decentralisation measures, but is explained by the accompanying changes in the sample. Although the results suggest that our main measure of fiscal decentralisation is a good approximation for measures that better reflect the autonomy of subnational governments, we cannot say with certainty that this also holds when looking at larger samples or measures related to discretionary spending of subnational governments rather than their autonomy over government revenue.

V. Conclusions

In this paper, we look at the relationship between fiscal decentralisation and economic growth. Using a sample of up to 56 countries over the period 1990–2007, we find evidence for a positive relationship. This relationship remains valid after controlling for possible endogeneity problems with the use of instrumental variables based on common legal system origin and relative country size. We think that this finding and the accompanying result of not being able to reject fiscal decentralisation being exogenous speak in favour, but are not definite proof, of the causality running from fiscal decentralisation to economic growth. More importantly, these results cannot be established by using other instrumental variables used in the literature. Given that our instrumental variables outperform these others, our approach may also be used in studies that look into other possible consequences of fiscal decentralisation. The relationship we find changes when we use alternative measures of fiscal decentralisation that better reflect the autonomy of subnational governments. This result, however, follows from the accompanying changes in the sample rather than from the use of these alternative measures themselves. In our case, where we look at tax-revenue-autonomy-based measures for selected countries, the use of conventional fiscal decentralisation measures leads to the same qualitative results as the use of measures that better reflect the true degree of fiscal autonomy of subnational governments. Here, the former thus seem a good approximation of the latter.

Our choice of instrumental variables is based on the notion that countries with similar characteristics, such as the origin of their legal system or geographical size, have a similar process of fiscal decentralisation. This idea implies that our approach may not work well for samples with a relatively small number of countries that differ in these characteristics. One possible solution to this problem, and direction for future research, is the construction of a broader data set of fiscal decentralisation, covering a larger number of countries over a longer time period, that can be used to derive a general set of instrumental variables. Another direction for further research is to look at alternative country characteristics that may be related to the process of fiscal decentralisation, such as the number of government tiers, number of local governments or measures of autonomy. Finally, the development of a theoretical framework that provides guidance on selecting parameter values would be a useful step in determining, through the use of simulations, which class of estimators is preferred when analysing the relationship between fiscal decentralisation and economic growth.

See Martinez-Vazquez and McNab (2003) and references therein for an excellent review of the channels and issues concerning the relationship between fiscal decentralisation and economic growth.

3

Other related papers are, for example, Bodman (2011) and Baskaran and Feld (2013), who use alternative measures of fiscal decentralisation based on Stegarescu (2005) to capture the dimension of policymaking authority of subnational governments.

4

We follow the criterion of Staiger and Stock (1997) that considers instrumental variables to be valid when the joint significance of the instrumental variables in the first-stage regression is large enough; as a rule of thumb, the corresponding F statistic should be larger than 10.

5

A comprehensive review of these studies is beyond the scope of this paper. Rodríguez-Pose and Ezcurra (2011) and Buser (2011) provide discussions and critical assessments of the literature.

6

While the framework of Davoodi and Zou (1998) is often used as a theoretical justification in the empirical literature on fiscal decentralisation and economic growth, it is not based on the subsidiarity principle and it cannot be used to directly derive an empirical specification. In their model, output is affected by government expenditures at both the national and subnational levels. Under the assumption that government expenditures at both levels are subject to diminishing returns to scale, it is optimal to devolve part of total government expenditures to the subnational level.

7

Given our focus on the endogeneity problems and the measurement of fiscal decentralisation in general, we do not address the identification of the appropriate set of variables to include in our growth specification. The latter is beyond the scope of the paper but could be tackled by using model averaging approaches; see Durlauf, Johnson and Temple (2005) for an overview. Hence, our findings complement the existing findings in the literature on fiscal decentralisation and economic growth, and our choice of explanatory variables should be interpreted as such.

See IMF (2001). Some countries (for example, Spain and the US) have more than one level of government between the central level and the local level. In such cases, IMF (2001) groups the intermediate levels of government together with the level they are most closely associated with.

The number of interpolated observations is 65. The interpolated fiscal decentralisation data show similar associations with size and legal origin to those of the available fiscal decentralisation data. An exception is the association with socialist legal origin, which is stronger for the interpolated data. We only use the extrapolated fiscal decentralisation data when constructing the instrumental variables and not when estimating equation (1) directly.

Unfortunately, the trade-off between the omitted variable bias and measurement error bias cannot be resolved on theoretical grounds, but must be evaluated with the use of simulations. In the fiscal decentralisation and economic growth literature, this sort of analysis is not possible given the lack of a theoretical framework (see Section 'Literature') that can provide guidance on choosing values for the parameters necessary for these simulations. The development of such a theoretical framework is beyond the scope of this paper and is left for future work. Appendix B, available online, discusses how results change when controlling for country fixed effects in the estimation procedure.

14

For example, Arellano and Bond (1991), Arellano and Bover (1995) and Blundell and Bond (1998).

15

In the first case, we take all categories of tax autonomy into account, using lower weights for categories with low degrees of autonomy. In the second case, we take all categories related to discretion on rates, bases and reliefs into account. In the last case, we take only the category with full discretion on bases and reliefs into account.

Appendix B, available online, discusses a wide array of robustness checks. The positive relationship between fiscal decentralisation and economic growth remains valid after changing time periods, the frequency – yearly, six-year averages and nine-year averages – of the data and the composition of government expenditures and when correcting for outliers.

18

In column 4, the p value corresponding to the coefficient of fiscal decentralisation is 0.077, whereas in column 8 it is 0.128, so it is not significant at conventional levels. In Appendix B (available online), we let the value of ξ vary. Higher values of ξ reflect larger restrictions in the degree of similarity, which translate into a better performance as instrumental variables but leads to a decline in the size and statistical significance of the fiscal decentralisation coefficient.

19

We use these tests rather than the standard tests since we cluster standard errors at the country level.

20

In Appendix B (available online), we show that when country fixed effects are included in the estimation procedure, neither our instrumental variables nor the lagged observations of fiscal decentralisation are valid instrumental variables.

21

In Appendix B (available online), we repeat the analyses of Table 4 with yearly data; the qualitative results of these estimations are the same. This means that the results are not caused by a small number of observations.