3.2. Variables
- (1)
Explanatory Variable: Fiscal Decentralization
Our main explanatory variable is the share of subnational expenditures among the total government expenditures as a measure of fiscal decentralization. This is one of the most widely used measures of fiscal decentralization in the empirical literature [
3,
20,
36,
38]. This GFS-based measure (GFS, which is short for the government finance statistics produced by the International Monetary Fund, considers the share of subnational expenditures (or revenues/taxes) among total government expenditures as a measure of fiscal decentralization. Given the wide use of this measure in empirical studies, many researchers simply call it the “GFS-based measure” for short.) is not perfect and has been criticized because it does not capture the real autonomy of subnational governments well and tends to overstate the actual level of decentralization [
39]. Hence, considerable effort has been dedicated to constructing indices of fiscal decentralization, such as those proposed by Hooghe et al. (2010) [
40]. However, there is no evidence that index measures are more effective than other measures, or that they yield more consistent results [
38]. There is also no empirical evidence that the GFS-based measure is less effective than other measures as a proxy for the relative level of a country’s fiscal decentralization, or that it is subject to systematic measurement errors across countries [
3,
36]. Ultimately, we chose the GFS-based measure of expenditure decentralization and accepted its imperfection because it allows us to utilize large samples that include multiple developing countries, thereby helping us to obtain empirical results with general significance.
- (2)
Endogeneity Problems and Instrumental Variables
Existing studies on the impact of fiscal decentralization on economic development suffer from an endogeneity problem [
41]; whereas fiscal decentralization may directly and indirectly affect economic growth, many governments have embarked on decentralization reforms in the pursuit of accelerated economic growth. In order to address this two-way causal relationship between fiscal decentralization and economic growth, existing studies have used a variety of techniques, including system-GMM and IV approaches [
3,
38].
In the case of fiscal decentralization and the HDI, we can similarly expect the presence of a two-way causal relationship. On the one hand, fiscal decentralization affects economic growth, public health, education, social welfare, and other determinants of national development [
42]. On the other hand, governments may seek to enhance national development by adjusting their decentralization policies in areas such as health, education, social welfare, etc., leading to an endogeneity problem [
43].
In order to solve potential endogeneity problems, we adopted the Geographic Fragmentation Index (GFI) developed by Canavire-Bacarreza et al. (2016) [
44] as an instrumental variable. First, there is a strong correlation between the GFI and fiscal decentralization. Canavire-Bacarreza et al. (2016) [
44] pointed out that the GFI is an important factor affecting fiscal decentralization and empirically tested this argument. As an instrumental variable, the advantage of the GFI over other geographical factors (e.g., total country area) is that it is a time-varying variable and can be used to estimate the time effect of panel data. Secondly, geographical factors are strictly exogenous, that is, they may affect economic activities and national development, whereas economic activities and national development do not affect geographical factors. Thirdly, the GFI has proven to be a suitable instrumental variable for empirical research on fiscal decentralization [
36].
The GFI reflects the weighted probability that two individuals selected at random in a country do not live in similar altitude zones, with the weight matrix calculated as the average distance between altitudes. Thus, the index is simply calculated as follows:
where
is the share of the population by elevation and
measures the distance between altitude
i and altitude
j. The values of this measure range from zero, which corresponds to a case in which all of the population is settled in the same altitude zone, to one, which corresponds to the implausible case in which each person lives at a different altitude. In general, geographic fragmentation increases with the number of altitude zones.
The GFI data were acquired from NASA’s Earth Observing System Data and Information System (EOSDIS) hosted by the Center for International Earth Science Information Network (CIESIN) at Columbia University. The data were available for the years 1990, 1995, 2000, and 2010. Given the low level of variation in the GFI over time, to address the missing values for 2001–2005, we assumed them to be the same as they were for 2006–2010. In addition, because the GFI data were available for only the five periods mentioned above, we also applied a lag period of expenditure decentralization as an alternative instrumental variable.
- (3)
Control Variables
First, we controlled for the variables commonly used for canonical specification in empirical research analyzing economic development issues, namely human capital, population, and openness [
45]. Specifically, we adopted the secondary school enrollment rate as a measure of human capital [
20,
38,
39], population was measured by the natural logarithm of the actual population, and openness was measured by the proportion of total import and export trade relative to the total GDP [
3,
36,
38,
46].
Secondly, although it is not regularly considered in the empirical study of the effects of fiscal decentralization, we controlled for government size, as measured by the revenue-to-GDP ratio. Fiscal decentralization is generally constrained by the government’s financial capacity [
47], and if fiscal decentralization leads to a smaller public sector (because of increased competition among levels of administration), and there is a negative relationship between the public sector size and growth, then there will be a positive bias in the estimation [
48].
Thirdly, we controlled for the dependence on natural resources, as measured by the share of total natural resource rent relative to the total GDP, since the presence of natural resources is likely to affect national development [
49]. Revenues from natural resources such as fossil fuels and minerals can make up a significant proportion of the GDP; however, their exploitation in the present will not only reduce future development and living standards but is also likely to cause environmental pollution and ecosystem damage [
50].
Fourthly, institutional, environmental, and ethnic fractionalization issues are considered to play important roles in economic growth and social harmony; therefore, we chose democracy, corruption, civil liberty, and ethnolinguistic fractionalization as control variables, with reference to [
3,
36,
46]. Democracy was measured by the democracy index according to the Polity IV dataset. Corruption was indicated by the Transparency International Corruption Perceptions Indices of Transparency International. Civil liberty was determined by Freedom House. Ethnic fractionalization was measured by the Ethnolinguistic Fractionalization (ELF) Indices built by Roeder (2001) [
37].
Finally, in this paper, we introduced time and region dummy variables to avoid missing the variable problems caused by time or regional differences [
38]. The region dummy variables were created according to the World Bank’s division method, namely Europe and Central Asia, Latin America and the Caribbean, East Asia and the Pacific, Sub-Saharan Africa, North America, South Asia, and the Middle East and North Africa.
3.3. Data and Imputation
Given the existing data information constraints, we compiled a panel of 50 countries (see
Table 2 for the complete list), including 20 developed countries and 30 developing countries, mainly distributed in six regions, namely Europe and Central Asia, Latin America and the Caribbean, Sub-Saharan Africa, North America, East Asia and the Pacific, and the Middle East and North Africa. The main consideration determining the selection of countries for inclusion in the sample was confirmation that there were data available for the selected countries on expenditure decentralization and other control variables.
- (1)
Sample for Basic Regression
The basic regression sample is a comprehensive (time-series and cross-country) dataset comprising 50 countries for the period of 1991–2020. With reference to Canavire-Bacarreza et al. (2020) [
36], we averaged the values for five-year periods to smooth the data over the macroeconomic cycle, which allowed us to explore the long-run effects. This resulted in a cross-country dataset covering six periods, namely 1991–1995, 1996–2000, 2001–2005, 2006–2010, 2011–2015, and 2016–2020. A summary of statistics for this sample is reported in
Table 3.
- (2)
Sample for Robustness Check
The sample used for the robustness check is a comprehensive panel (time-series cross-country) dataset comprising 50 countries for the period of 2010–2021. A summary of statistics for this sample can be found in
Table 3.
- (3)
Data Imputation
Because some data were missing for some variables, we adopted complementary methods depending on the actual situation. First, the mean value interpolation method was adopted. For example, if data for 2010 and 2012 were available, but data for 2011 were missing, we used the average value of 2010 and 2012 to replace the value for 2011. We also adopted the nearest-neighbor interpolation method, which was used to deal with missing data for variables that were very stable over time, such as the natural logarithm of the total population. The clustering mean interpolation method was also adopted, whereby a missing value was supplemented by the mean value of the region or organization to which the country belonged. If partial data for a variable were missing for a given period (5 years), for example, if there were only 3 years of data, we took the average value of those 3 years as the data for the whole period. Specifically, for the period of 2006–2010, fiscal decentralization data for Turkey were only available for 2008, 2009, and 2010; therefore, we used the average value of these 3 years as the average value for the whole period. Since this method was applied to long periods of time and across all countries, it did not cause any major problems. In addition, for some variables with serious missing data issues, we did not make any supplements or modifications. For example, the expenditure decentralization data for the United Arab Emirates before 2011 were all missing; therefore, we did not make any interpolation.