The relationship between trade openness and government size: Does disaggregating government expenditure matter?

Journal of Macroeconomics 34 (2012) 239–252

Contents lists available at SciVerse ScienceDirect

Journal of Macroeconomics journal homepage: www.elsevier.com/locate/jmacro

The relationship between trade openness and government size: Does disaggregating government expenditure matter? Michael Benarroch 1, Manish Pandey ⇑ Department of Economics, The University of Winnipeg, Winnipeg, Manitoba, Canada R3B 2E9

a r t i c l e

i n f o

Article history: Received 14 February 2011 Accepted 9 November 2011 Available online 3 December 2011 JEL classification: F1 H1 Keywords: Government size Openness Causality

a b s t r a c t This paper is the first to examine the causal relationship between trade openness and government size using both aggregate and disaggregated government expenditure data, including data on social security. Our results indicate that examining the relationship separately for functional categories of government expenditures and based on differences in incomes across countries provide important details on the relationship between the two variables not found elsewhere in the literature. Our causality tests provide little or no support for a causal relationship between openness and aggregate or disaggregated government expenditure. Similar results are obtained when our sample is split into low income versus high income countries. The only evidence of a robust, statistically significant, positive causal relationship is found between openness and education expenditures in low income countries. In no case is there a positive causal relationship between social security and openness. This leads us to conclude that there is no evidence to support the relationship suggested by Rodrik (1998). Ó 2011 Elsevier Inc. All rights reserved.

1. Introduction In recent years there has been considerable interest concerning the effects of greater openness on government size. The idea that openness may be positively related to government size was initially proposed by Cameron (1978) and later developed as the ‘compensation hypothesis’ by Ruggie (1982).2 It was however Rodrik (1998) who first conducted a detailed empirical study of the issue and then combined the empirical analysis with a simple general equilibrium model that provides a plausible explanation for the direct relationship between openness and government size. According to Rodrik, the most likely reason for this association is that countries exposed to a greater amount of ‘‘external risk’’ demand larger governments as a form of social insurance. Based on this explanation, Rodrik (1998) makes a number of hypotheses including the notion that causality runs from ‘‘exposure to external risk to government spending’’ (p. 998). In the current paper, we conduct an empirical analysis using both aggregate government expenditure data and eight categories of disaggregated government expenditure data, including social security, to examine whether there is evidence of a causal relationship between trade openness and government size.3 Further, we extend the analysis to consider whether the relationship between openness and government size differs across low income versus high income countries. While a number ⇑ Corresponding author. Tel.: +1 204 786 9289; fax: +1 204 772 4183. E-mail addresses: [email protected] (M. Benarroch), [email protected] (M. Pandey). Tel.: +1 204 786 9268; fax: +1 204 774 8057. 2 A competing view is the ‘efficiency hypothesis’ which proposes that increases in openness lead governments to reduce spending on welfare programs through pressures to reduce taxes (Garrett, 2001). Gemmell et al. (2008) provide a comprehensive review of the evidence for and against this hypothesis. In the current paper we focus on the direct effect of openness on government spending and investigate the ‘compensation hypothesis’. 3 It is important to note that causality tests ‘‘only indicate that changes in one variable precede changes in another variable of interest (with a positive or negative sign) rather than establishing causation in the traditional sense of the word’’ (p. 136, Casu and Girardone, 2009). 1

0164-0704/$ - see front matter Ó 2011 Elsevier Inc. All rights reserved. doi:10.1016/j.jmacro.2011.11.002

240

M. Benarroch, M. Pandey / Journal of Macroeconomics 34 (2012) 239–252

of the hypotheses made by Rodrik (1998) have been examined throughout the literature, little work has been done examining the causal relationship between openness and government expenditure or distinguishing whether this relationship differs between low and high income countries. The current paper builds on an initial study by Benarroch and Pandey (2008) that employs panel data and aggregate government consumption to test for the causal relationship between openness and government size. They find no support for the positive causal relationship between these two variables. In using government consumption data however, Benarroch and Pandey (2008) cannot consider the impact of openness on transfer payments since such expenditures are not included in government consumption. Further, as shown in the literature, the use of aggregate data can mask the underlying association between openness and specific components of government size (Shelton, 2007; Dreher et al., 2008). Though a causal relationship between aggregate government expenditure and trade openness may not exist, a formal test of Rodrik’s hypothesis should address the issue of whether greater trade openness has a causal relationship with specific components of government expenditure. For instance, if as argued by Rodrik (1998), trade openness leads to greater volatility, then greater openness should have a positive causal relationship with social security spending thereby protecting and insuring workers against such shocks while not necessarily increasing overall spending. Our paper adds to a growing literature that uses both aggregate and disaggregated government spending data to examine whether the results in Rodrik (1998) are robust and whether there is validity in how he rationalizes his findings. Much of this literature finds only weak evidence in support of Rodrik’s findings that ‘‘there is a positive and robust partial correlation between openness, as measured by the share of trade in GDP, and the scope of government, as measured by the share of government in GDP’’ (Rodrik, 1998, p. 998). For example, Alesina and Wacziarg (1998) conclude that country size is the key and that small countries tend to be more open. While Alesina and Wacziarg (1998) cast doubt on Rodrik (1998), they do not completely rule out his findings since their regressions show that if one considers government transfers there is ‘‘some evidence of a direct relationship between openness and the size of government transfers’’ (p. 306). Ram (2009) examines the sensitivity of Alesina and Wacziarg’s results with regards to the relationship between country size and both openness and government size, and concludes that after controlling for country-specific fixed effects and time effects the results do not hold. Likewise, studies that primarily use panel data for high income countries have found little evidence to support Rodrik (1998). Specifically, Islam (2004) and Molana et al. (2004) employ OECD data and conclude that size of government has not changed to moderate against greater external risk. Cavallo (2007) finds that openness leads to less volatility, whereas Liberati (2007) employs mostly European data and rejects the Rodrik’s hypothesis in favor of a hypothesis that capital openness is negatively related to government expenditures across 20 OECD countries. Further, Garen and Trask (2005), using non-budgetary measures of government size, find that the ‘‘scope of government is much larger in less open economies’’ (p. 534). Their results are however, explained by differences in per capita GDP across countries rather than the reasons given by Rodrik (1998). A few recent studies analyze the relationship between disaggregated government expenditure and openness. In addition to Alesina and Wacziarg (1998) who employ disaggregated data, Gemmell et al. (2008) use a dynamic model of 25 OECD countries and find that increases in foreign direct investment shifts government expenditure towards social spending. Conversely, Shelton (2007) concludes that openness is not associated with an increase in any of the categories of government expenditure that insure for risk in a large dataset that includes low income countries, while Dreher et al. (2008) conclude that none of the expenditure categories they consider are affected by globalization. None of the papers discussed above however, consider whether there is a causal relationship between openness and disaggregated government expenditures. The current paper thus contributes to the literature on the relationship between openness and government size in a number of important ways. (1) To our knowledge we are the first to conduct a causality test using aggregate and disaggregated data on government expenditure.4 Our goal is to examine whether increases in trade openness cause government size to expand using a dynamic panel data estimation model. This estimation model allows for an examination of the long-run relationships between openness and government size and Granger causality tests not found elsewhere in the literature. Relative to the fixed effects approach used in most other studies, the estimation of dynamic models control for endogeneity issues when examining whether greater openness in the previous period causes government size to increase in the current period.5 (2) In addition to the above mentioned tests, we also consider whether there is evidence of a causal relationship between openness and various components of government expenditure separately for low income and high income countries. We believe this is important given that high income countries have well established transfer payment programs relative to low income countries implying that it should be easier for this group of countries to provide greater welfare benefits to offset the negative impact associated with increases in openness (Rodrik, 1998; Shelton, 2007).6 Finally, (3) we are the first to test whether the causal relationship between government size and openness is robust to a sub-sample restricted only to democratic countries and to an alternate definition of openness that accounts for financial openness.7,8 4 Note that we also estimate regressions similar to Alesina and Wacziarg (1998), Rodrik (1998), Shelton (2007) and others to provide baseline results that are comparable to the rest of the literature. 5 A number of other advantages of employing the dynamic approach are discussed later in the paper. See also Rodman (2006) for further details. 6 We use the World Bank definition to classify countries as low and high income. For further details see Footnote 10. 7 Recent studies suggest that democratic countries, due to pressures from interest groups and/or electorates, are more inclined to use government spending in response to increased volatility from greater trade openness (Avelino et al., 2005; Adserá and Boix, 2002; Rudra, 2002). We thus test to insure that our results are robust to a division of countries based on the degree of democratic rule within the country. 8 We use the data from the Polity IV Project (url: http://www.systemicpeace.org/inscr/inscr.htm) and define democratic countries as those with a polity2 score greater than zero. Further details are provided in Section 2.

M. Benarroch, M. Pandey / Journal of Macroeconomics 34 (2012) 239–252

241

The results from our baseline fixed effects model using aggregate data indicate that there is no evidence of a statistically significant positive association between trade openness and government expenditure after controlling for country-specific fixed effects. Further, our results show that, as in Alesina and Wacziarg (1998) and Shelton (2007), an analysis using aggregate data masks important details concerning the relationship between openness and government expenditures that are found when the empirical analysis is conducted using disaggregated data. For example, when the data are disaggregated on the basis of differences in income across countries our results provide some evidence that greater trade openness is associated with larger government expenditures in low income countries but the relationship does not hold in aggregate for high income countries. Further, using disaggregated government expenditure data, a positive and statistically significant association with openness is only found to exist for expenditures on defense and health in low income countries. In none of the regressions however is a statistically significant relationship found between social security spending and openness as suggested by Rodrik (1998). When we extend the literature and test whether greater trade openness causes government size to increase using a dynamic panel data estimation model, we conclude that, as in our fixed effects model, there is no positive causal relationship for aggregate government expenditure. Separate tests for high versus low income countries show that these results hold for both groups of countries. The results using disaggregated data confirm the findings using aggregate data. Specifically, of the eight categories of government expenditure considered, no evidence of a causal relationship with openness is found for any of the categories. When we conduct further causality tests for the relationship between openness and the eight categories of government expenditure for low income and high income countries separately, we find evidence in support of a positive, and statistically significant, causal relationship between openness and education expenditures for low income countries, with this relationship holding in the long-run. Our findings consequently suggest that greater openness in low income countries has led to greater demand for government funded education expenditures but not for expenditures on any of the other seven categories. These results are robust to alternative divisions across countries (democratic versus non-democratic) and an alternative definition of openness based a measure of financial openness. Overall, our findings provide evidence against the existence of the mechanism suggested by the compensation hypothesis that increases in openness lead to higher social security payments and other such transfers to counteract some of the negative effects of increased trade. However, we find evidence that an increase in openness has caused an increase in education expenditures in low income countries. The remainder of the paper is organized as follows. Section 2 describes the data used for our analysis and provides some summary statistics. Section 3 uses fixed-effects models to examine the relationship between trade openness and aggregate government expenditure as well as its main components. In Section 4 we investigate whether there is a causal relationship between trade openness and the main components of government expenditure. The potential difference in the relationship between government size and openness between high and low income countries is examined in Sections 3 and 4. Section 5 examines the robustness of the results and Section 6 provides a brief conclusion.

2. Data Most studies that examine the relationship between trade openness and government size use data on government consumption expenditures from the Penn World Tables.9 This data however lacks information on transfer payments, such as social security payments, that by definition are not included in government consumption expenditures. Given that transfer payments could be used by governments in response to greater volatility caused by increased trade openness, a better measure of government size for examining the relationship between openness and government size is total government expenditures including transfer payments. The Government Financial Statistics (GFS) dataset provided by the International Monetary Fund (IMF) and complied by William Easterly provides data on total government expenditure as a percentage of GDP for 119 countries for the period ranging from 1972 to 2000 that is disaggregated by various functional categories.10 For the current analysis eight government expenditure categories are considered: public administration, defense, education, health, social security, housing, recreation and economic services. These categories, on average, account for approximately 95% of total government expenditures. Data on trade openness, defined as total exports and imports relative to GDP, real GDP per capita and population are obtained from Penn World Tables (PWT 6.3).11 Following Rodrik (1998) the dependency ratio and fraction of urban population are used as controls in our analysis. The data for these variables is obtained from the World Development Indicators (WDI). Given that high income countries, relative to low income countries, have more established and comprehensive social security programs, there is evidence to suggest that it is easier for the former group to increase welfare benefits in response to the pressures of globalization (Rudra, 2002). To investigate whether there are differences between low and high income countries, we define low income countries using the World Bank definition based on real Gross National Income (GNI) per 9

See, for example, Alesina and Wacziarg (1998), Rodrik (1998) and Benarroch and Pandey (2008). As with aggregate government expenditure, expenditure on each of the functional categories is available as a percentage of GDP. See Easterly (2001) for further details. The data has been posted on the World Bank website as part of the Global Development Network Growth Database and was downloaded using the following url: http://econ.worldbank.org/WBSITE/EXTERNAL/EXTDEC/EXTRESEARCH/0,,contentMDK:20701055~pagePK:64214825~piPK:64214943~theSitePK:469382,00.html. 11 See Heston et al. (2009) and http://pwt.econ.upenn.edu/php_site/pwt_index.php for details on PWT 6.3. 10

242

M. Benarroch, M. Pandey / Journal of Macroeconomics 34 (2012) 239–252 Table 1 High versus low income countries.a Low income countries Belarus Benin Bhutan Bolivia Bulgaria Burkina Faso Burundi Cameroon Central African Rep. Chad Colombia Congo, Rep. of Cote d‘Ivoire Djibouti Dominican Rep. Egypt El Salvador Ethiopia Gambia Guatemala Guinea-Bissau

Guyana Haiti Honduras India Indonesia Iran Jamaica Kazakhstan Lesotho Liberia Madagascar Maldives Mali Mauritania Moldova Mongolia Morocco Namibia Nepal Nicaragua Niger

Pakistan Paraguay Peru Romania Russia Rwanda Senegal Sri Lanka Suriname Syria Tajikistan Tanzania Thailand Togo Tonga Tunisia Vanuatu Yemen Zambia Zimbabwe

High income countries Argentina Australia Austria Bahamas Bahrain Barbados Belgium Belize Brazil Canada Chile Costa Rica Croatia Cyprus Czech Rep. Denmark Dominica Estonia Finland France Gabon

Greece Hungary Iceland Ireland Israel Italy Japan Korea, Rep. of Kuwait Latvia Luxembourg Malaysia Malta Mauritius Mexico Netherlands Norway Panama Poland Portugal Seychelles

Singapore Slovak Rep. Slovenia South Africa Spain St. Kitts & Nevis St. Lucia Sweden Switzerland Trinidad & Tobago Turkey United Kingdom United States Uruguay Venezuela

a Due to lack of sufficient data, countries in Italics were excluded from the estimation of dynamic regressions examining causality.

capita for the year 2000 and obtain the data from the WDI.12 Table 1 provides a list of high and low income countries in our sample. To account for the long-run relationship between government size and openness, we follow Shelton (2007) and smooth short-run fluctuations by taking 5-year averages of all variables.13 This provides an unbalanced panel dataset with a maximum of six observations for each of the 119 countries.14 Table 2 reports the average values for government consumption and government expenditure as a percentage of GDP as well as trade openness. The number of countries for which data are available varies over the periods between 78 and 94. As one would expect, government expenditure as a percentage of GDP is higher than government consumption on average by about 11%.15 Further, trade openness has steadily grown between 1970 and 2000, with

12 The World Bank definition based on real GNI per capita (GNIpc) in US dollars for the year 2000: low income countries – GNI pc 6 755, lower middle income countries 756 < GNIpc < 2995, upper middle income countries 2996 < GNIpc < 9265, high income countries >9265. We classify a country as low income if it belonged to the first to groups (low income or lower middle income), that is if GNIpc for the country in 2000 was less than $2596. We used a number of cutoffs for per capita income and the findings were qualitatively similar to those reported in the paper. The results for the estimation are available from the authors upon request. 13 As argued by Shelton (2007), the choice of 5-year averages is a common compromise in the growth literature. 14 The periods over which averages are computed are 1970–1975, 1976–1980, 1981–1985, 1986–1990, 1991–1995 and 1996–2000. 15 The pair wise correlation between consumption and expenditure measures of government size is about 0.24. This suggests that using the two variables to measure government size could lead to different conclusions for the analysis of the impact of trade openness on government size. Further, it is worth noting that the decline in the average ratio of government consumption to GDP between 1991–1995 and 1996–2000 is due to differences in the number of countries for which data are available for the two periods.

243

M. Benarroch, M. Pandey / Journal of Macroeconomics 34 (2012) 239–252 Table 2 Openness, government expenditure and government consumption over time. Period

G/Y (consumption)

G/Y (expenditure)

Openness

Observations

1970–1975 1976–1980 1981–1985 1986–1990 1991–1995 1996–2000

17.36 18.83 19.17 19.77 19.74 17.85

24.98 29.13 32.47 30.68 31.01 30.50

67.58 75.85 76.40 72.78 78.38 86.45

78 93 93 91 94 80

Growth rate Corr (X, Open)

0.001 0.03

0.028*** 0.28

0.040*** 1.00

The rate of growth is the estimate for the time trend controlling for country specific fixed-effects. Significance at 5%. Significance at 10%. *** Significance at 1%. 



the post 1990 period experiencing the largest increases. Government size, whether measured using expenditure or consumption, has grown at a slower rate, with the growth in expenditures being higher than consumption. The pair wise correlation between trade openness and government expenditure (0.28) is also higher than that for trade openness and government consumption (0.03) implying that the relationship between government size and openness could be strengthened when government expenditure is used to measure government size. The average values of expenditures relative to GDP over time for each of the eight disaggregated categories of government expenditures are reported in Table 3. The rate of growth for each of the categories as a percentage of GDP suggests that while expenditures on public administration, defense and economic affairs have declined overtime; health, education, social security and housing expenditures have increased. Further, the last row of Table 3, which provides the pair-wise correlation between each of the categories and trade openness, indicates that there are significant differences in the correlation across the categories of government expenditure. While the correlation is positive for all categories, education and housing expenditures have the highest correlation with openness. Such differences across categories of government expenditure suggest that using aggregate government expenditure may mask the underlying association between openness and specific components of government expenditure. Further, given that some of these components of government expenditure could be used to respond to the negative fallout from greater openness, it is important to examine the relationship between openness and specific components of government expenditure for a comprehensive test of the ‘compensation hypothesis’. 3. Fixed effects estimation In this section we use the panel data described in the previous Section to estimate fixed effects models using both government consumption and expenditures as a percentage of GDP to proxy for government size. This approach provides baseline results which are comparable to estimates reported in previous studies, as well as a comparison of the results between the two measures of government size. The regression we estimate is similar to that used by Rodrik (1998) and is given by:

ln g it ¼ a0 þ a1 lnopenit1 þ bX it þ cpd dum þ gi þ eit ;

ð1Þ

where for country i at period t all variables are in natural logarithm terms (ln). The dependent variable g is government size measured by government consumption or expenditure as a percent of GDP; open is trade openness lagged by one period to Table 3 Government expenditure on functional categories to GDP ratio over time. Period 1970–1975 1976–1980 1981–1985 1986–1990 1991–1995 1996–2000 Growth rate Corr (X, Open)

Public admin 3.55 4.09 4.16 3.32 2.82 2.81 0.068*** 0.24

Defense 3.17 3.30 3.11 2.88 2.98 2.34 0.036*** 0.01

Education

Health

Social security

Housing

Recreation

3.33 3.54 3.70 3.52 3.46 3.32

1.87 2.03 2.22 2.20 2.46 2.62

4.89 4.93 5.85 5.93 7.68 7.81

0.67 0.88 0.89 0.99 0.98 0.88

0.38 0.46 0.40 0.40 0.40 0.43

0.021*** 0.32

0.049*** 0.15

0.039** 0.02

0.083* 0.35

0.001 0.11

The rate of growth is the estimate for the time trend controlling for country specific fixed-effects. Significance at 10%. Significance at 5%. *** Significance at 1%. *

**

Economic affairs 5.47 6.69 7.12 6.17 4.76 4.23 0.082*** 0.14

244

M. Benarroch, M. Pandey / Journal of Macroeconomics 34 (2012) 239–252

Table 4 Openness, government expenditure and government consumption: fixed effects estimates. Expenditure

ln opent1

Consumption

1

2

3

4

1

2

3

4

0.104 (0.08)

0.089 (0.08)

0.047 (0.09)

0.369 (0.20)* 0.105 (0.11) 0.254 (0.13)*** 0.038 (0.19)

0.018 (0.09) 0.108 (0.12) 0.392 (0.20)* 0.110 (0.12) 0.232 (0.13)*** 0.025 (0.19)

0.055 (0.09)

0.260 (0.17) 0.102 (0.10)

0.004 (0.09) 0.162 (0.12) 0.307 (0.17)* 0.110 (0.11)

0.196 (0.15) 0.268 (0.08)*

0.203 (0.17) 0.265 (0.08)* 0.004 (0.14) 0.098 (0.19)

0.132 (0.10) 0.282 (0.11)*** 0.278 (0.15) 0.281 (0.08)*

0.168 (0.10) 0.328 (0.12)*** 0.273 (0.17) 0.280 (0.09)* 0.068 (0.14) 0.135 (0.19)

519 119 0.213

511 116 0.231

519 119 0.22

511 116 0.234

519 119 0.131

511 116 0.125

519 119 0.152

511 116 0.152

0.10

0.21

0.14

0.12

ln opent1  LIC ln pop ln gdppc ln urban ln depend Observations Countries R2-within

Ho: lagln open + lagln open  LIC = 0 p-Value

Notes: Robust standard errors based on the Huber–White sandwich estimate of the variance reported below estimates. All models include a constant, fixed effects and time (period) effects. The last row provides the test of significance for the estimated coefficient for openness for low income countries. ⁄⁄ Significance at 5%. * Significance at 10%. *** Significance at 1%.

address the contemporaneous endogenity; X represents the control variables: real GDP per capita (gdppc); population (pop); rate of urbanization (urban); dependency ratio (dep). In addition, we include country-specific fixed effects (g), time effects through period dummies (pd_dum), and an iid error term (e).16 The estimation results for both the consumption and expenditure measures of government size are presented in Table 4. Given that data for the dependency ratio and the rate of urbanization are only available for a smaller number of countries, regression (1) is estimated with and without these variables. The estimated coefficient for population is negative for all specifications and statistically significant for most regressions providing support for the findings of Alesina and Wacziarg (1998) that country size is an important determinant of government size. Similar to Rodrik (1998) among others, the estimated coefficient for real GDP per capita is negative, though statistically significant for only the consumption measure of government size. With regards to the urbanization rate and dependency ratio, the signs of the estimated coefficients differ between the expenditure and consumption measures. While for expenditure both the coefficients are positive (statistically significant only for urban), they are negative and not significant for consumption. Such differences may be explained by the fact that the expenditure measure includes transfer payments not included in the consumption measure.17 The estimated coefficient for lagged openness is positively associated with government size for both the consumption and the expenditure measures, but unlike the findings of Rodrik (1998) who only considers the consumption measure, the association is not statistically significant (Columns 1 and 2 of Table 4 for expenditure and consumption measures).18 To investigate whether there are differences in the relationship between openness and government size between low income and high income countries, an interaction between the dummy for low income countries and the lagged value of openness (ln openit1  LIC) is introduced into regression (1). For the two measures of government size, Columns 3 and 4 in Table 4 present the estimation results with the interaction term. Contrary to expectations, the estimates suggest that there is some evidence of a positive association between openness and government size for low income countries rather than for high income countries. In fact for the consumption measure the association is negative, though not statistically significant, for high income countries but positive and significant for low income countries.19 This differs somewhat from Shelton (2007) who finds that government expenditures in both OECD and non-OECD countries are positively associated with openness.

16 In addition, following the argument in Shelton (2008) regarding differences in the effect of the two age groups, younger than 15 and older than 65 years, on government expenditures, in place of the dependency ratio we introduced the logarithm of the fraction of population younger than 15 and older than 65 years of age as separate independent variables in regression (1). We also experimented with a proxy for federalism using the ratio of transfers to sub-national governments to total government expenditures. The estimated coefficients for these variables were not statistically significant and their introduction did not change the main findings reported in this section. The results, not reported here, are available from the authors upon request. 17 These differences between the consumption and the expenditure measures of government size suggest that transfer payments are positively associated with the dependency rate and the urbanization rate. 18 Benarroch and Pandey (2008) employ data from the Penn World Table version 6.1. As in the current paper they also find a negative, but not statistically significant, relationship between openness and government size. 19 Further, the results for the test reported in the last row of Table 4 suggest that the estimated coefficient for lagged openness for low income countries is positive and significant at the 10%, 12% and 14% level of significance for three of the four specifications that include the interaction term.

245

M. Benarroch, M. Pandey / Journal of Macroeconomics 34 (2012) 239–252 Table 5 Openness and functional classifications of government expenditures – fixed effects estimates. Public admin ln opent1

0.20 (0.19)

0.14 (0.18)

0.37 (0.48) 0.07 (0.20) 0.09 (0.36) 0.02 (0.33) 448 112 0.15

448 112 0.16

ln opent1  LIC ln pop ln gdppc ln urban ln depend Observations Countries R2-within

Defense 0.06 (0.28) 0.39 (0.33) 0.46 (0.49) 0.08 (0.20) 0.01 (0.37) 0.05 (0.33)

Ho: lagln open + lagln open  LIC = 0 p-Value

ln opent1

0.11 (0.19)

ln pop ln gdppc ln urban ln depend Observations Countries R2-within

1.04 (0.45)** 0.25 (0.24) 0.32 (0.62) 0.55 (0.42) 425 109 0.09

425 109 0.10

Ho: lagln open + lagln open  LIC = 0 p-Value

0.09 (0.39) 0.41 (0.18)** 0.12 (0.46) 0.86 (0.41)** 410 106 0.17

410 106 0.19

0.48

0.09 (0.25)

Health 0.41 (0.20)**

0.12 (0.37) 0.07 (0.19) 0.27 (0.20) 0.21 (0.28)

0.20 (0.16) 0.39 (0.19)** 0.03 (0.39) 0.09 (0.19) 0.19 (0.19) 0.24 (0.29)

0.28 (0.56) 0.27 (0.23) 0.02 (0.31) 0.35 (0.39)

0.38 (0.30) 0.04 (0.36) 0.27 (0.59) 0.27 (0.23) 0.03 (0.30) 0.36 (0.39)

446 112 0.07

446 112 0.09

446 112 0.09

446 112 0.09

0.09 Housing

0.02 (0.17) 0.20 (0.28) 1.08 (0.46)** 0.26 (0.25) 0.28 (0.64) 0.56 (0.43)

ln opent1  LIC

0.07 (0.12)

0.13

Social security

Education 0.29 (0.25) 0.62 (0.28)** 0.26 (0.39) 0.43 (0.19)** 0.23 (0.45) 0.94 (0.42)***

0.16 Recreation 0.08 (0.20)

1.03 (0.61)*** 0.56 (0.20)* 1.10 (0.58)*** 0.25 (0.53)

0.54 (0.43) 0.91 (0.46)** 0.80 (0.65) 0.53 (0.20)* 0.94 (0.59) 0.19 (0.54)

433 110 0.13

433 110 0.14 0.15

0.08 Economic affairs 0.15 (0.16)

0.96 (0.62) 0.04 (0.35) 0.02 (0.43) 0.11 (0.52)

0.12 (0.28) 0.06 (0.39) 0.95 (0.66) 0.04 (0.35) 0.01 (0.43) 0.11 (0.53)

0.04 (0.38) 0.26 (0.19) 0.17 (0.30) 0.03 (0.34)

0.05 (0.22) 0.14 (0.27) 0.01 (0.38) 0.26 (0.19) 0.14 (0.32) 0.02 (0.34)

410 108 0.07

410 108 0.07

444 111 0.29

444 111 0.29

0.83

0.83

Notes: Robust standard errors based on the Huber–White sandwich estimate of the variance reported below estimates. All models include a constant, fixed effects and time (period) effects. The last row provides the test of significance for the estimated coefficient for openness for low income countries. * Significance at 10%. ** Significance at 5%. *** Significance at 1%.

As has been argued by Alesina and Wacziarg (1998), Shelton (2007), Dreher et al. (2008) among others, any analysis based on aggregate government expenditure may mask the underlying association between openness and specific components of government size. Specifically, some forms of government expenditures, such as transfer payments, could be used to respond to greater volatility as a result of increased openness. In fact, using disaggregated data, Shelton (2007) concludes that expenditures do not increase across categories ‘‘that constitute social insurance’’ (Shelton, 2007, p. 2245). To determine whether this is the case in the current paper, the relationship between specific components of government expenditure and openness are estimated using regression (1), where we replace ln g with the logarithm of expenditure for each of the specific components as a fraction of GDP. Table 5 presents the results of the estimation for the eight categories of government expenditure considered. When we do not allow for differences in the slope between low income and high income countries, the estimated coefficient for openness is positive and statistically significant only for expenditure on health. Introducing the interaction term between lag openness and the dummy for low income countries for each of the categories, we find a positive and statistically significant association between openness and expenditures on defense and health for low income countries.20 For high income countries, no statistically significant positive association between openness and government expenditure is found for any of the categories of expenditure. Hence, we find that there are significant differences in the relationship between openness and various categories of government expenditure as well as between low income and high income countries. In particular, we do not find evidence

20

The statistical significance of the association is determined using the test result reported in the last row of Table 5.

246

M. Benarroch, M. Pandey / Journal of Macroeconomics 34 (2012) 239–252

that greater openness leads to an increase in social security expenditure, the mechanism proposed by the compensation hypothesis. These results are consistent with those found in Shelton (2007) and others employing disaggregated data. The above findings suggest some evidence of a positive association between openness and government size, in particular for low income countries. These results, however, do not allow us to evaluate whether changes in openness in the previous period lead to changes in government size in the current period. In other words, as Benarroch and Pandey (2008) argue, the positive association between openness and government size does not imply that greater trade openness causes government size to increase, which was the hypothesized relationship in Rodrik (1998) and would be required for the confirmation of the compensation hypothesis. 4. Does greater openness ‘cause’ government size to increase? This section investigates whether there is evidence of a causal relationship between openness and government size or the various categories of government expenditure. A finding confirming a positive causal relationship would suggest that, as proposed by the compensation hypothesis, greater openness in the previous period causes government size to increase in the current period. To determine whether openness ‘Granger’ causes government size to increase, appropriate instruments at the country level and over time are required. Such instruments, however, are not readily available. Dynamic panel regression models resolve the issue of instruments by using suitable lagged levels and lagged first differences of the regressors as instruments. To examine whether openness causes changes in government size we follow Casu and Girardone (2009), Hartwig (2010) and Michauda and van Soest (2008) and estimate the following dynamic regression equation:

lnðgov tÞit ¼ a0 þ

m X p¼1

cp lnðgov tÞitp þ

m X

bp lnðopenÞitp þ cpd dum þ gi þ eit

ð2Þ

p¼1

Eq. (2) is an AR(p) process with country specific fixed effects and period dummies (pd_dum). The test of whether ln(open) ‘Granger’ causes ln(govt) is based on a joint test: b1 = b2 =  = bp = 0.21 Evidence for a long-run relationship between the two variables is determined by testing whether b1 + b2 +  + bp = 0. When evidence exists in favor of a causal relationship, this test provides a means of determining whether the two variables are related in the short-run or the long-run.22 There are several advantages in using this approach for the estimation of the causal relationship between openness and government expenditure: (1) the estimator is designed for small-T large-N panels, which is consistent with our data; (2) it allows the dependent variable to be dynamic; (3) it addresses the issue of endogeneity by allowing independent variables that are not strictly exogenous; (4) allows for country-specific fixed-effects; and (5) allows for heteroskedacity and autocorrelation of variables within panel (Roodman, 2009a). To test for Granger-causality, it is necessary that the two time series be stationary. We examine the time series properties of openness and government size using panel data unit root tests (augmented Dickey–Fuller test and the Phillips–Perron test) and find that both series are stationary.23 Further, the results for any causality test are sensitive to the choice of lag length for the variables. Following Hartwig (2010), regression (2) is estimated using OLS and the choice of the optimal lag length is based on the Schwarz Information Criterion (SIC).24 The optimal lag length is found to be two periods or 10 years. For estimating Eq. (2) the system GMM estimator developed by Arellano and Bover (1995), and Blundell and Bond (1998) is used.25 This procedure employs the variables in levels but uses first differenced lagged values as instruments.26 The estimator allows us to address the endogeneity of explanatory variables in a dynamic formulation and explicitly controls for potential biases arising from country-specific effects. To obtain robust two-step standard errors for all our regressions, we perform the small sample correction proposed by Windmeijer (2005). The system GMM estimator is consistent only if lagged values of variables are valid instruments. As suggested by Chang et al. (2009), the issue of the validity of instruments can be addressed by performing three specification tests (two over-identification tests and a serial correlation test). The two over-identification tests, the Hansen test and the Incremental Hansen (difference-in-Hansen test), examine the validity of the full set of instruments and the additional instruments that are introduced in the levels equations. For both these tests, not rejecting the null hypothesis of no over-identification provides support for the model. The third test, AR2 test, examines whether the first-differenced error term is second-order serially 21 It is important to note that Granger causality tests ‘‘only indicate that changes in one variable precede changes in another variable of interest . . . rather than establishing causation. . .’’ (Casu and Girardone, 2009). 22 Gemmell et al. (2008) use a similar dynamic model to examine the relationship between openness and government size for 25 OECD countries. While they focus on the short-run relationship, our approach examines the long-run causal relationship between openness and government size. 23 The values of the augmented Dickey–Fuller test statistic, which has a chi-squared distribution, are 305.10 for openness, 255.48 and 547.46 for the expenditure and consumption measures of government size. 24 The value for the test statistics for the Schwarz (Bayesian) Information Criterion are 48.67 for one lag, 33.00 for two and 40.14 for three. 25 See Roodman (2009a) for technical details for the system GMM estimator and Cavallo and Cavallo (2010) for a description of the conditions that are required for a system GMM estimator. 26 Roodman (2009a) argues in favor of using the system GMM approach as it is more efficient than the difference GMM estimator developed by Holtz-Eakin et al. (1988). The system GMM estimator addresses the problem of weak instruments, which arises for difference GMM estimators when variables are very persistent. To examine the robustness of our results we estimated regression (2) using the difference GMM approach and obtained similar results.

247

M. Benarroch, M. Pandey / Journal of Macroeconomics 34 (2012) 239–252 Table 6 Government expenditure and consumption: causality results. Government expenditure

Government consumption

All

All

Collapsed instruments All

LIC

HIC

0.917 (0.10)*** 0.135 (0.09) 0.117 (0.11) 0.046 (0.10)

0.764 (0.13)*** 0.285 (0.09)*** 0.277 (0.41) 0.061 (0.11)

0.775 (0.17)*** 0.234 (0.06)*** 0.285 (0.16)* 0.065 (0.13)

0.614 (0.19)*** 0.286 (0.19) 0.360 (0.27) 0.081 (0.11)

283 93 30

283 93 14

115 45 14

Specification tests (p-values) Hansen 0.37 Incremental Hansen 0.52 AR2 0.96

0.57 0.92 0.58

Causality tests (p-values) Test b1 = b2 = 0 Test b1 + b2 = 0

0.79 0.52

ln gt1 ln gt2 ln opent1 ln opent2 Observations Countries Instruments

0.45 0.27

Collapsed instruments All

LIC

HIC

0.838 (0.12)*** 0.292 (0.17)* 0.050 (0.16) 0.026 (0.08)

0.853 (0.18)*** 0.254 (0.25) 0.054 (0.24) 0.067 (0.16)

0.602 (0.31)* 0.213 (0.22) 0.448 (0.25)* 0.022 (0.15)

0.971 (0.26)*** 0.189 (0.27) 0.535 (0.33) 0.183 (0.09)*

168 48 14

283 93 30

283 93 14

115 45 14

168 48 14

0.71 0.31 0.88

0.42 0.18 0.41

0.47 0.32 0.35

0.07 0.06 0.34

0.65 0.44 0.28

0.51 0.38 0.08

0.14 0.06

0.43 0.25

0.94 0.82

0.92 0.94

0.18 0.07

0.07 0.29

Notes: Standard errors in parenthesis below estimates. Time dummies and a constant are included in all regressions, estimated coefficients not reported. ‘Collapsed’ indicates that the collapse option is used to reduce the number of instruments. Estimation method: two-step system GMM with Windmeijer (2005) small sample robust standard error correction. ⁄⁄ Significance at 5%. * Significance at 10%. *** Significance at 1%.

correlated. Second order serial correlation implies that the error term is serially correlated and follows a moving average process. This results in the rejection of lagged values as appropriate instruments for the system GMM estimation. As with the over-identification tests, not rejecting the null of the absence of second-order serial correlation lends support to the model. A criticism of the system GMM estimation procedure has been that the implementation of the estimator can lead to instrument proliferation. As argued by Roodman (2009b), a large number of instruments over-fit endogenous variables and at the same time weaken the Hansen test for over-identification. While there are no formal tests that can determine the number of lags that should be used as instruments, a rule of thumb has been to limit the number of instruments below the number of panel observations, in our case the number of countries. In addition, Roodman (2009b) suggests two ways to reduce the number of instruments. First, combine instruments through addition into smaller sets, and second, limit the number of lagged values. As is conventional for endogenous variables (Roodman, 2009a), our estimation uses lags of only two periods or greater as instruments. To avoid instrument proliferation the number of instruments is also collapsed for most specifications.27 Collapsing instruments reduces the number of instruments but minimizes the loss of information. Table 6 presents the results from estimating model (2) using government consumption and expenditure measures of government size. For the specifications considered, the two over-identification tests (Hansen and Incremental Hansen) and the second-order serial correlation test (AR2) fail to reject the null hypothesis for all specifications except for the specification with the consumption measure.28 The results in Table 6 include estimation when all lags greater than two are used as instruments as well as when the instruments are collapsed. Given that null b1 = b2 = 0 is not rejected when we consider all countries in our sample; we find no evidence of a causal relationship between openness and government size, whether measured by government expenditure or government consumption. The same result holds when we collapse the instruments.29 Given the differences in the relationship between openness and government size between low and high income countries reported above, separate causality tests for the low and high income countries for both the consumption and expenditure

27 Though for the full sample the number of instruments is less than the number of countries, when the sub-samples of low and high income countries are examined, the number of instruments could exceed the number of countries. 28 When we consider all countries in the sample and collapse the instruments, the two over-identification tests indicate the model is over-identified for the consumption measure of government size. This indicates that when the instruments are collapsed, persistence in the variables leads to two lags or higher not being valid instruments requiring three lags or greater. Doing so, leads to the instruments being valid and yields similar results. Further, since we emphasis the use of the expenditure measure, we do not place much weight on the results obtained with the consumption measure. 29 Collapsing instruments reduces the number of instruments from 30 to 14.

248

M. Benarroch, M. Pandey / Journal of Macroeconomics 34 (2012) 239–252

Table 7 Openness and expenditure on government functional categories: all countries. Public

Defense

Education

Health

Social security

Housing

Recreation

Economic services

0.478 (0.80) 0.166 (0.33)

0.148 (0.87) 0.065 (0.16)

0.726 (0.49) 0.113 (0.15)

0.515 (0.40) 0.336 (0.17)*

0.056 (0.67) 0.125 (0.36)

0.342 (0.56) 0.203 (0.25)

0.343 (0.46) 0.084 (0.30)

0.654 (0.55) 0.196 (0.23)

230 80 14

205 73 14

229 80 14

229 80 14

212 74 14

221 77 14

201 75 14

228 79 14

Specification tests (p-values) Hansen 0.28 Incremental Hansen 0.22 AR2 0.49

0.67 0.81 0.68

0.71 0.93 0.37

0.78 0.40 0.45

0.15 0.67 0.84

0.82 0.97 0.71

0.57 0.56 0.38

0.74 0.82 0.13

Causality tests (p-values) 0.53 Test b1 = b2 = 0 Test b1 + b2 = 0 0.35

0.92 0.92

0.11 0.06

0.14 0.53

0.90 0.75

0.28 0.23

0.75 0.62

0.49 0.29

ln opent1 ln opent2 Observations Countries Instruments

Notes: Standard errors in parenthesis below estimates. First and second lags of dependent variable included in all regressions, the estimates are not reported. Time dummies and a constant are included in all regressions, estimated coefficients not reported. Estimation method: two-step system GMM with Windmeijer (2005) small sample robust standard error correction with collapsed instruments.  Significant at 5%. ⁄⁄⁄ Significant at 1%. * Significant at 10%.

measures are undertaken.30 For the expenditure measure, the causal relationship between openness and government size is rejected for both low and high income countries.31 For the consumption measure, the causality test results for low income countries are very similar to that for the expenditure measure. When we examine the relationship for high income countries however, the null hypothesis of a non-casual relationship is rejected at the 7% level. The negative value for the sum of the estimated coefficient for lagged values of openness indicates evidence in favor of a negative causal relationship such that an increase in openness in the previous period results in a decrease in government size in the current period. This provides evidence against the ‘compensation hypothesis’ for high income countries.32 This negative causal relationship for high income countries however, holds only in the short-run. While we find no support in favor of a positive causal relationship between openness and government size, governments may respond to greater openness by increasing expenditure on certain functional categories such as social security. We examine the relationship between openness and our eight categories of government expenditure. To do so, regression model (2) is estimated using the logarithm of the ratio of expenditure to GDP for each of the eight categories as the dependent variable, with the collapse option to reduce the number of instruments. Table 7 presents the results of the causality test.33 For all the specifications estimated, there is support in their favor from both the over-identification tests as well as the AR2 test. Based on the causality test reported in the last panel of Table 7, no evidence of a causal relationship is found for any of the categories at the conventional levels of statistical significance.34 In particular, our findings provide evidence against the existence of the mechanism suggested by the compensation hypothesis that increases in openness lead to higher social security payments and other such transfers to counteract some of the negative effects of increased trade. Next we examine the relationship between openness and the eight categories of government expenditure separately for low income and high income countries. Table 8 presents the results for the causality tests undertaken for the two groups of countries. Based on the causality tests, there is only evidence in support of a positive, and statistically significant, causal relationship between openness and education expenditures for low income countries, with this relationship holding in the longrun. Our findings suggest that greater openness in low income countries has led to greater demand for government funded education expenditure but not greater demand for expenditures from any of the other seven categories. 30 When all countries are considered, government expenditure and government consumption are, on average, equally persistent over time. There are however, differences between low income and high income countries. While government expenditures are more persistent than government consumption for low income countries, the opposite holds for high income countries. Further, government consumption is more volatile than government expenditure for low income countries, which results in the standard errors for the estimates of lagged values of ln g being higher for the former measure of government size. 31 The test for a long-run relationship suggests evidence of a long-run relationship between openness and government size at the 6% level of significance for low-income countries. 32 For the government consumption specification for high income countries the estimated coefficients for the lagged values of the logarithm of openness are jointly statistically significant at the 7% level of significance and the sum of the estimated coefficients is negative. This suggests possible evidence in favor of the efficiency hypothesis that an increase in openness results in a reduction in government size. Given that we focus on the compensation hypothesis in the current paper, we leave the further exploration of the efficiency hypothesis for future research. 33 When disaggregated data are used for the estimation of dynamic models, the number of countries for which data are available varies between 73 and 80. 34 It should however be noted that for education expenditures in low income countries there is evidence of a long-run causal relationship with openness at the 6% level of significance.

249

M. Benarroch, M. Pandey / Journal of Macroeconomics 34 (2012) 239–252 Table 8 Openness and expenditure on government functional categories: low income versus high income countries. Public

Defense

Education

Health

Social security

Housing

Recreation

Economic services

0.089 (1.03) 0.163 (0.85)

0.046 (0.39) 0.061 (0.28)

0.659 (0.43) 0.416 (0.32)

0.429 (0.31) 0.064 (0.20)

0.083 (0.69) 0.254 (0.34)

0.001 (0.72) 0.665 (0.48)

0.321 (0.56) 0.218 (0.67)

0.818 (0.71) 0.505 (0.34)

96 39 14

81 33 14

96 39 14

96 39 14

78 33 14

91 36 14

76 34 14

95 38 14

Specification tests (p-values) Hansen 0.36 Incremental Hansen 0.76 AR2 0.77

0.69 0.83 0.52

0.83 0.80 0.45

0.78 0.37 0.45

0.26 0.49 0.12

0.71 0.69 0.91

0.63 0.80 0.50

0.73 0.79 0.72

Causality tests (p-values) Test b1 = b2 = 0 0.88 Test b1 + b2 = 0 0.64

0.98 0.96

0.03 0.01

0.35 0.18

0.72 0.77

0.12 0.17

0.67 0.40

0.33 0.49

0.342 (0.36) 0.0003 (0.23)

0.311 (0.53) 0.162 (0.35)

0.252 (0.44) 0.021 (0.23)

0.388 (0.39) 0.204 (0.24)

0.438 (0.40) 0.372 (0.22)

0.041 (0.79) 0.035 (0.48)

0.022 (0.66) 0.057 (0.30)

0.394 (0.25) 0.011 (0.27)

134 41 14

124 40 14

133 41 14

133 41 14

134 41 14

130 41 14

125 41 14

133 41 14

Specification tests (p-values) Hansen 0.48 Incremental Hansen 0.14 AR2 0.56

0.27 0.19 0.43

0.40 0.71 0.36

0.52 0.23 0.31

0.80 0.91 0.21

0.55 0.32 0.15

0.87 0.59 0.84

0.54 0.65 0.21

Causality tests (p-values) Test b1 = b2 = 0 0.23 Test b1 + b2 = 0 0.11

0.84 0.66

0.65 0.40

0.61 0.40

0.19 0.78

0.98 0.87

0.96 0.88

0.23 0.15

LIC ln opent1 ln opent2 Observations Countries Instruments

HIC ln opent1 ln opent2 Observations Countries Instruments

Notes: Standard errors in parenthesis below estimates. Significant at 10%, Significant at 5% and ⁄⁄⁄Significant at 1%. The causality tests for each category of government expenditure are based on estimation of regression equation (2). First and second lags of dependent variable included in all regressions, the estimates are not reported. Time dummies and a constant are included in all regressions, estimated coefficients not reported. Estimation method: two-step system GMM with Windmeijer (2005) small sample robust standard error correction with collapsed instruments.

5. Robustness of causality test results This section examines the robustness of the result that no causal relationship between openness and government size exists by (1) limiting the analysis only to democratic countries and (2) using an alternative measure of openness, i.e. financial openness. A number of recent studies (see Avelino et al., 2005; Adserá and Boix, 2002; Rudra, 2002) have argued that democratic countries are more inclined to use government spending in response to increased openness due to pressures from interest groups and/or the electorate.35 Without such demand from the electorate, governments are said to be less likely to increase spending to compensate for the negative fallout of greater openness. These studies thus, argue that the compensation hypothesis is more likely to hold for democratic countries. To test for this possibility, we construct an indicator for democracy using data from the Polity IV project. The data provides a score (polity2) on a scale of +10 (strongly democratic) to 10 (strongly autocratic) for 93 countries in our dataset. The score is constructed as a difference between democracy and autocracy scores. The average of this score for the 1995–2000 period is used to classify countries as democratic if they had an average of greater than 0. In other words, a country is defined as democratic if its democracy score is greater than its autocracy score.36 The results of estimating regression (2) with the expenditure measure of government size as the dependent variable, restricting the sample only to democratic countries, is presented in Table 9. When democratic countries, 72 out of 93 in

35 Mulligan et al. (2004) finds no evidence of differences between democratic and non-democratic countries with regards to public spending policies. The relationship between openness and government size may however differ between democratic and non-democratic countries. 36 The reported results are qualitatively similar when a cutoff of democracy of 5 rather than 0 is used. In addition, we classified countries as democratic and non-democratic based on the average score for the 1970–2000 period and obtain similar results.

250

M. Benarroch, M. Pandey / Journal of Macroeconomics 34 (2012) 239–252 Table 9 Openness and government size: democratic countries. All

LIC

HIC

0.565 (0.76) 0.083 (0.14)

0.384 (0.23) 0.012 (0.14)

0.367 (0.23) 0.015 (0.12)

Observations Countries Instruments

225 72 14

73 29 14

152 43 14

Specification tests (p-values) Hansen Incremental Hansen AR2

0.96 0.71 0.78

0.79 0.51 0.44

0.36 0.09 0.92

Causality tests (p-values) Test b1 = b2 = 0 Test b1 + b2 = 0

0.76 0.48

0.16 0.06

0.07 0.03

Dependent variable: ln(govt_expdr) ln opent1 ln opent2

Notes: Standard errors in parenthesis below estimates. Significant at 10%, Significant at 5% and Significant at 1%. First and second lags of dependent variable included in all regressions, the estimates are not reported. Time dummies and a constant are included in all regressions, estimated coefficients not reported. Estimation method: two-step system GMM with Windmeijer (2005) small sample robust standard error correction with collapsed instruments.

our dataset, are considered no evidence of a causal relationship is found. Conversely, when the sample is divided into low income and high income groups, there is evidence of a negative and statistically significant causal relationship for democratic high income countries. In sum, the results for the democratic sample provide further evidence against a positive causal relationship between openness and government size. While most studies examining the relationship between openness and government size have used the ratio of total trade to GDP as the measure for openness, another measure that has been used in recent studies is financial integration (Garret and Mitchell, 2001; Kimakova, 2009). We follow Kose et al. (2008) and construct a measure of financial integration as the ratio of gross stocks of external liabilities to GDP.37 The data are from Lane and Milesi-Ferretti’s (2007) External Wealth of Nations Database. Table 10 replicates the estimation results presented in Table 6 for the expenditure measure of government size replacing trade to GDP ratio with financial integration as the measure of openness. As with trade openness, we find no evidence of openness ‘Granger’ causing government size even for financial integration. This result holds for both low income as well as high income countries.38

6. Conclusion This paper empirically investigates the existence of positive causal relationship between openness and government size proposed by the compensation hypothesis. Though the hypothesis has been explored in a number of studies (Alesina and Wacziarg, 1998; Dreher et al., 2008; Rodrik, 1998; Shelton, 2007; among others), none of the studies have examined the causal relationship between openness and government size. Moreover, while previous studies examining the compensation hypothesis employ fixed effects models, we estimate dynamic panel data estimation models to examine whether changes in openness precede changes in government size. These models allow us to control for country-specific effects and address the issue of endogeneity of variables present in much of the literature. We use data for 119 countries over the period ranging from 1972 to 2000 and test for the causal relationship between openness and government size and also between openness and eight different components of government expenditure. Further, we divide countries into low and high income to investigate whether income differences across countries matter for the relationship between the two variables. 37 Kose et al. (2008) argue that this measure of inflows ‘‘is most closely related to the notion of openness to foreign capital that could be associated with technological and other spillovers’’ (p. 7). 38 In addition, we investigated the relationship between volatility of financial openness and government size. The standard deviation of the financial openness variable was used as measure of volatility of financial openness. Similar to the results for financial openness, we found no evidence of a causal relationship between volatility of financial openness and government size for all countries or for sub-groups of low income and high income countries. Further, we also investigated the relationship between financial openness and the eight categories of government expenditure and found no evidence of a causal relationship between financial openness and any of the eight categories of government expenditure that we consider. The results are available from the authors upon request.

251

M. Benarroch, M. Pandey / Journal of Macroeconomics 34 (2012) 239–252 Table 10 Financial openness and government size. All

LIC

HIC

0.066 (0.06) 0.003 (0.03)

0.063 (0.12) 0.061 (0.07)

0.058 (0.08) 0.006 (0.06)

Observations Countries Instruments

269 88 14

109 42 14

160 46 14

Specification tests (p-values) Hansen Incremental Hansen AR2

0.56 0.19 0.30

0.30 0.07 0.71

0.56 0.49 0.46

Causality tests (p-values) Test b1 = b2 = 0 Test b1 + b2 = 0

0.44 0.20

0.70 0.98

0.73 0.46

Dependent variable: ln(govt_expdr) ln(fin_open)t1 ln(fin_open)t2

Notes: Standard errors in parenthesis below estimates. Significant at 10%, Significant at 5% and Significant at 1%. First and second lags of dependent variable included in all regressions, the estimates are not reported. Time dummies and a constant are included in all regressions, estimated coefficients not reported. Estimation method: two-step system GMM with Windmeijer (2005) small sample robust standard error correction with collapsed instruments.

Our results indicate that examining the relationship separately for functional categories of government expenditures and based on differences in incomes across countries provide important details on the relationship between the two variables not found elsewhere in the literature. When using data for all countries, we find no evidence of a causal relationship between openness and aggregate government size. Similar results are obtained when our sample is split into low versus high income countries. Investigating this relationship further by disaggregating government expenditure into eight major components, we find evidence in support of a positive, and statistically significant, causal relationship only between openness and education expenditures for low income countries. In other words, increases in openness for low income countries cause education expenditures to increase. For high income countries, contrary to the relationship proposed by the compensation hypothesis, we find some evidence of a negative causal relationship indicating that greater openness may have caused a fall in government size. Given the lack of a positive causal relationship between openness and government size as well as between openness and social security expenditures, our study finds no evidence in favor of the compensation hypothesis. Our findings, however, do suggest that openness may affect government expenditure on education for low income countries. One reason for this may be that with greater openness and international competition, governments in low income countries are diverting expenditures towards education in order to develop the human capital required to remain competitive in the future.

References Adserá, A., Boix, C., 2002. Trade, democracy, and the size of the public sector: the political underpinnings of openness. International Organization 56, 229– 262. Alesina, A., Wacziarg, R., 1998. Openness, country size and government. Journal of Public Economics 69, 305–321. Avelino, G., Brown, D.S., Hunter, W., 2005. The effects of capital mobility, trade openness, and democracy on social spending in Latin America, 1980–1999. American Journal of Political Science 49, 625–641. Arellano, M., Bover, O., 1995. Another look at the instrumental variable estimation of error-components models. Journal of Econometrics 68, 29–51. Benarroch, M., Pandey, M., 2008. Trade openness and government size. Economics Letters 101, 157–159. Blundell, R., Bond, S., 1998. Initial conditions and moment restrictions in dynamic panel data models. Journal of Econometrics 87, 115–143. Casu, B., Girardone, C., 2009. Testing the relationship between competition and efficiency in banking: a panel data analysis. Economics Letters 105, 134–137. Cameron, D.R., 1978. The expansion of the public economy: a comparative analysis. American Political Science Review 72, 237–269. Cavallo, E., 2007. Openness to Trade and Output Volatility: A Reassessment. Inter-American Development Bank Research Department Working Paper #604. Cavallo, A.F., Cavallo, E.A., 2010. Are crises good for long-term growth? The role of political institutions. Journal of Macroeconomics 32, 838–857. Chang, R., Kaltani, L., Loayza, N.V., 2009. Openness can be good for growth: the role of policy complementarities. Journal of Development Economics 90, 33– 49. Dreher, A., Sturm, J.E., Ursprung, H.W., 2008. The impact of globalization on the composition of government expenditures: evidence from panel data. Public Choice 134, 263–292. Easterly, W., 2001. The lost decades: developing countries’ stagnation in spite of policy reform 1980–1998. Journal of Economic Growth 6, 1381–4338. Garen, J., Trask, K., 2005. Do more open economies have bigger governments? Another look. Journal of Development Economics 77, 533–551. Garrett, G., 2001. Globalization and government spending around the world. Studies in Comparative International Development 35, 3–29. Garret, G., Mitchell, D., 2001. Globalization, government spending and taxation in the OECD. European Journal of Political Research 39, 145–177. Gemmell, N., Kneller, R., Sanz, I., 2008. Foreign investment, international trade, and the size and structure of public expenditures. European Journal of Political Economy 24, 151–171. Hartwig, J., 2010. Is health capital formation good for long-term economic growth? – Panel Granger-causality evidence for OECD countries. Journal of Macroeconomics 32, 314–325.

252

M. Benarroch, M. Pandey / Journal of Macroeconomics 34 (2012) 239–252

Heston, A., Summers, R., Aten, B., 2009. Penn World Tables Version 6.3, Center for International Comparisons of Production, Income and Prices at the University of Pennsylvania. Holtz-Eakin, D., Newey, W., Rosen, H.S., 1988. Estimating vector autoregressions with panel data. Econometrica 56, 1371–1395. Islam, M.Q., 2004. The long run relationship between openness and government size: evidence from bounds test. Applied Economics 36, 995–1000. Kimakova, A., 2009. Government size and openness revisited: the case of financial globalization. Kyklos 62, 394–406. Kose, M.A., Prasad, E., Terrones, M., 2008. Does Openness to International Financial Flows Contribute to Productivity Growth? NBER Working Paper No. 14558. Lane, P., Milesi-Ferretti, G., 2007. The external wealth of nations mark II: revised and extended estimates of foreign assets and liabilities, 1970–2004. Journal of International Economics 73, 223–250. Liberati, P., 2007. Trade openness, financial openness and government size. Journal of Public Policy 27, 215–247. Michauda, P., van Soest, A., 2008. Health and wealth of elderly couples: causality tests using dynamic panel data models. Journal of Health Economics 27, 1312–1325. Mulligan, C., Gil, R., Sala-i-Martin, X., 2004. Do democracies have different public policies than nondemocracies. Journal of Economic Perspectives 18, 51–74. Molana, H., Montagna, C., Violato, M., 2004. On the Causal Relationship between Trade-openness and Government Size: Evidence from 23 OECD Countries. University of Dundee Discussion Paper No. 164. Ram, R., 2009. Openness, country size, and government size: additional evidence from a large cross-country panel. Journal of Public Economics 93, 213–218. Rodrik, D., 1998. Why do more open economies have bigger governments? Journal of Political Economy 106, 997–1032. Roodman, D., 2009a. How to do xtabond2: an introduction to difference and system GMM in Stata. Stata Journal 9, 86–136. Roodman, D., 2009b. Practitioners’ corner: a note on the theme of too many instruments. Oxford Bulletin of Economics and Statistics 71, 135–158. Rudra, N., 2002. Globalization and the decline of the welfare state in less-developed countries. International Organization 56, 411–445. Ruggie, J.G., 1982. International regimes, transactions, and change: embedded liberalism in the postwar economic order. International Organization 36, 379–415. Shelton, C.A., 2007. The size and composition of government expenditure. Journal of Public Economics 91, 2230–2260. Shelton, C.A., 2008. The aging population and the size of the welfare state: Is there a puzzle? Journal of Public Economics 92, 647–651. Windmeijer, F., 2005. A finite sample correction for the variance of linear efficient two-step GMM estimators. Journal of Econometrics 126, 25–51.