The Revenues-Expenditures Nexus: a Panel Data Analysis of Spain s Regions (Running title: Tax-Expenditure, Expenditure-tax or Fiscal Synchronization)

Transcription

1 The Revenues-Expenditures Nexus: a Panel Data Analysis of Spain s Regions (Running title: Tax-Expenditure, Expenditure-tax or Fiscal Synchronization) ABSTRACT KEY WORDS JEL CODES Manuel Jaen GARCIA Department of Applied Economics, University of Almeria, Ctra. Sacramento s/n Almeria, Spain mjaen@ual.es In this paper an analysis is made of nexus revenue-expenditure for Spain s Autonomous Regions. The particular characteristics of finance for the public sector in Spain lead to the conclusion that a unidirectional causality relationship exists between public revenue to public expenditure. revenues-spending, governments borrow, tax-spending E01, E21, E62 I. Introduction The revenues-spending nexus has been an object of great empirical research over the last twenty years. These analyses have been made on the national, regional, and local levels using different tests of unit roots, co integration, and causality both with time series and panel data. Determining which hypothesis best characterizes an economy is more than an intellectual exercise because it can potentially contribute to the discovery of a solution to the problem of growing budget deficits. In the second section the four hypotheses are discussed. In the third section a review of published empirical literature is provided. Section four contains an overview of the current situation in Spain and is followed by some empirical tests in the fifth section. In the sixth section the empirical test for the case of Spain is made. Conclusions of this paper are provided in the seventh section. Hypothesis of the Tax-spending Debate Literature on tax-spending has considered four hypotheses for the intertemporal relation between public spending and revenues. First, the tax-spending hypothesis formulated by Friedman (1978) argues that raising taxes, increasing the cost of available resources for the public sector in order to decrease the budget deficit, only results in a rise in public expenditure. Indeed, if the revenues have a positive effect on the expenditures, then the reductions in the revenues will lead to reductions in the expenditures. 24

2 An alternative point of view for this hypothesis is Wagner (1970) and Buchanan and Wagner (BW) (1977). They suggest that increases in revenues will result in a decrease in expenditures via fiscal illusion. In this line of thinking, public spending financed by means other than direct tax results in the perception by the people that the cost of public spending is lower than it would be with direct tax. Although the taxpayer can pay less with direct tax, fiscal illusion emerges as the people are paying through indirect tax, via high interest rates and a substitution effect is produced and, moreover, inflation causes displacements in the tax track. The point of view of BW places emphasis on the fact that public expenditure is found by three sources: direct tax, debt emission, and indirect tax through inflation. From a political perspective, the optimal level of revenue sources in order to finance a specific expenditure is directly related to its costs. Therefore, in the line of thinking of BW, a rise in taxes (direct taxation) can lead to a reduction in public expenditure if there is a reduction in direct taxation methods, or rather, a reduction in debt emission as well as in displacements in the tax track. Secondly, the expenditure-tax hypothesis suggests that spending decisions are initially determined by adjustments in tax policy and revenue to meet expenditure demands. Given the Ricardian equivalence proposition that what governments borrow today results in an increase in future tax liabilities, which are fully capitalized by the taxpayers, Barro (1979) questions the fiscal illusion hypothesis considered in BW. Under the Ricardian equivalence model, increases in public spending lead to increases in taxes. Among others, Roberts (1978) and Peacock and Wiseman (1979) propose that temporary increases in public spending due to crisis situations translate to permanently higher taxes (this is the displacement effect hypothesis). Thus, under the expenditure-tax hypothesis, reductions in government spending will lead to reductions in deficits. Thirdly, the fiscal synchronization hypothesis argues that decisions on revenues and expenditures are taken simultaneously or jointly. Musgrave (1966) and Meltzer and Richard (1981) propose that voters compare the marginal benefits and costs of government programmes when they decide on the appropriate level of public expenditure and revenue. Fourthly, the institutional separation hypothesis recognizes that taxation decisions are independent of the allocation of public spending (Wildavsky, 1988; Baghestani and McNown, 1994). Review of Empirical Literature The literature on this subject has produced a wide variety of results due in part to the different methodological approaches, the degree of temporal aggregation, and the reporting periods themselves. Although the specific methodology varies across studies, the empirical literature primarily focuses on the concept of causality in the spirit of Granger. In this context, Friedman's version of the tax-spending hypothesis is supported if unidirectional causality is clear from revenue to expenditure and revenues have a positive impact on spending. The BW version is met if there is unidirectional causality from revenue to expenditure and revenues have a negative impact on spending. The spending-tax version is supported if the causality runs from expenditures to revenues. Fiscal synchronization is confirmed if there is bidirectional causality between revenues and expenditures. Finally, the lack of causality supports the institutional separation hypothesis (Paybe, 2003). 25

3 These studies can be divided roughly between those that perform the analysis at a national level and those that consider the state or local level. Among the first group, most have been done for the USA and major European countries, and a smaller number for the industrialized countries of the Pacific and other less developed countries. There are several key differences between the analysis at national and subnational levels. Firstly, state and local governments are unable to carry out inflationary policies as a means of raising taxes such as national governments. Secondly, many state and local governments govern under legislative and constitutional requirements that attempt to restrict budget deficits. Thirdly, high levels of government subsidies influence the budgets of state and local governments. The presence of subsidies as a source of funding can affect expenditures (flypaper effect) and revenues (substitution effect). For example, changes in spending can be attributed to changes in revenues when in fact they are actually attributable to changes in the subsidies (Payne, 2003). As for the subnational level, there is a limited number of studies with mixed results, which is also the case for the national level. There are many research studies about the revenue-expenditure nexus for specific countries: Anderson et al. (1986), Von Furtensberg et al. (1986), Miller and Russek (1990), Baghestani and McNown (1994) and Westerlund et al. (2009) for the USA; Hasan and Lincoln (1997) for the UK; Payne (1997) for Canada; Darrat (1998) for Turkey; Li (2001) and Chang and Ho (2002) for China; and Chang and Ho (2002) for Taiwan. In addition, Ram (1988a, 1988b), Baffes and Shah (1994), Chang et al. (2002), Reddick and Hassan (2003) and Kollias and Palelogou (2003) for multi-country studies. To subnational level it can be cited, between other, the papers of Joulfaian and Mookerjee (1990) for multicountry, Dahber and Johanson (1998) for Swedish municipalities, Moisio (2000) for Finland, Ho and Huang (2009) for China and many papers for USA, such as Marlow and Manage (1987, 1988), Poterba (1994), Alesina and Bayoumi (1996), Sorensen, Wu and Yosha (2001), Westerlund et al. (2009) and Zapf and Paybe (2009). Situation in Spain An important process of decentralization of the public sector took place in Spain in a relatively short period of time. Currently, the Spanish territorial division is similar to that of most federal countries. Therefore, there are three levels of government: Central. Regional (Autonomous Regions): There are seventeen Autonomous Regions (ARs) created by the Spanish Constitution in 1978 in recognition of their right to autonomy, gradually established between 1979 and Local (provinces and municipalities): The regions each have a number of provinces, but six of them only have one. In Spain, there are fifty provinces, each of which includes several municipalities. The total number of municipalities is approximately eight thousand. The distribution of responsibilities between different levels of government is regulated by the Constitution, the Statutes of Autonomy of the Autonomous Regions and Local Government Law. The central government has exclusive responsibility for defence, foreign 26

4 affairs, economic stabilization, social security, and public order 1. Other responsibilities have been transferred to regional governments, although at different rates and levels according to the Autonomous Region 2. Also, the Local Government Act sets a minimum level of mandatory services for municipalities according to their population size and provinces, although there is some ambiguity and overlap in the allocation of certain responsibilities between regional and local levels. However, the degree of decentralization of public expenditure has not been matched by such progress in terms of revenue. The Constitution establishes two funding systems for the ARs: the Foral regime, which is applied to the Basque Country and Navarre and is regulated by economic agreements; and the common system, for the fifteen remaining ARs, which is reflected in the Financing Act of the Autonomous Regions (Ley Orgánica de Financiación de las Comunidades Autónomas, LOFCA). The Foral regime is the result of recognition of historical rights, maintaining a peculiar tax regime for those territories, which involves a complete fiscal autonomy. It is based on the transfer of revenue and management of most state taxes to the provincial administration of these regions as well as some regulatory powers in relation to those state taxes. On the other hand, the region makes a contribution to the State through an annual global quota as a contribution to all state expenses not directly taken by the Autonomous Region. The common system has gradually taken shape over time, consistent with the growth of responsibilities and services provided by regional governments. The two most important types of resources are: Own taxes and those which are either completely or partially ceded by the State 3. Transfers from the General State Budget. These consist basically of two kinds: a) a general subsidy, as a share of central government resources to finance common services and education, and b) conditional cash transfers (e.g. for health and social services) and the territorial compensation fund. The system has been tending toward greater correspondence between tax revenues generated in each territory and regional financing. However, revenues of the ARs in the common system dramatically depend on central government subsidies, the entity that primarily controls the management of public resources. Moreover, the system does not guarantee full equality in revenues among the regions in order to address their autonomous responsibilities. Empirical Methodology In this paper an analysis is made of revenue-expenditure nexus using data for the Spanish economy from 1987 to 2003 for the 15 Spanish regions with a common fiscal regime. Data 1 In the case of public order, the State shares responsibility with the regional governments of Catalonia and the Basque Country within their territories. 2 Responsibility in education was gradually transferred to the ARs between 1995 and 1999, while all the ARs finally took control of health care in However, the central government has the right to establish basic legislation in these areas. Finally, in some other areas, the responsibility is shared between central and regional government. 3 Currently these are the wealth tax, tax on capital transfers and documented legal acts, inheritance and gift tax and the gambling tax. Also, 33% for income tax, 35% for VAT, and from 40% to 100% for excise tax, depending on the case. 27

5 utilized include those on revenues, expenditures, and gross domestic product (GDP) at constant prices (deflated in 1995) using different deflators 4. The methodology uses the following steps: Unit root panel tests of three series. Different tests are then utilized: Levin, Li and Chu; Breitung, Im, Pesaran and Chin; and Fisher-type test proposed by Maddala and Wu; and the Hadri test. Cointegration tests of heterogeneous panels are also used: Pedroni, Kao and Johansen-Fisher and Granger causality test using the methodology from Hansen and Rand (2006). An empirical analysis has to be carried out carefully in order to verify the nature of the series, because if they are not stationary, problems could arise in the estimation of the regression equation coefficients. Valid estimations require that data be stationary (integrated zero-order) or, if they are not stationary (integrated first-order), that they be cointegrated. More specifically, the first step will be to verify whether the variables are stationary or if they have one or more unit roots. In case they are integrated, an analysis will be made of the possible existence of cointegration between them. Unit root tests for panel data rather than for individual time series have the advantage of being able to increase test power as they simultaneously exploit cross-section and time series information. This study uses tests by Levin, Li and Chu (LLC), Breitung, Im, Pesaran and Chin (IPS), as well as Fisher s test by Maddala and Wu, whose null hypothesis is the existence of a unit root. The first two are processes of common unit roots, while the last three are processes of individual unit roots. Finally, another perspective, in the form of Hadri s test, maintains the null hypothesis that the time series for each cross-section unit is stationary against the alternative of a common unit root. Once the existence of a panel unit root has been established, the question arises whether a long-run equilibrium relationship exists between the variables. Since all of the variables are I(1), an analysis is then done for panel cointegration using different tests. This produces the result in Table 2. The seven variants of cointegration tests derived by Pedroni are then used, bearing in mind that the first four have common AR coefficients and the last three have individual AR coefficients just like the test by Kao. Both tests expand upon the cointegration test technique by Engle and Granger based on an analysis of the residuals of a spurious regression carried out using variables I(1). If the variables are cointegrated, then the residuals will be I(0); while if they are not cointegrated, the residuals will be I(1). Pedroni s seven tests take the null hypothesis of no cointegration with different alternative hypotheses. This allows the endogeneity of the regressors as well as considerable heterogeneity among dynamics, fixed effects, and cointegrating vectors for individual i in panel data, which tests with Kao s approximation in which homogeneity is imposed on cointegrating vectors. Pedroni also demonstrates that his tests are distributed under the standard normal distribution. In the case of the first test, the right tail of the standard normal distribution is needed to reject the null of no cointegration (large positive values imply rejection), while in the other six, large negative values imply that the null of no cointegration is rejected. 4 The used data proceed from Spanish Statistical Institute (INE) for GDP and deflators, BADESPE (database from Finance Ministry) for revenues and expenditures and they are updated to 01/02/

6 Finally, the Johansen-Fisher test is carried out in which Maddala and Wu use Fisher s combined test results in order to propose an alternative approximation for the panel data cointegration test by combining individual cross-section tests to obtain a test statistic for the complete panel. These two preliminary steps are important to ensure that the correct econometric procedure is being used. The estimation of a cointegration relationship using ordinary least squares is generally biased due to problems of variable endogeneity, which is why the corresponding t statistics do not follow a regular Student t distribution; therefore, it is impossible to make any inference about its significance. On the other hand, regressing the first difference of the variables when a long-run equilibrium relationship does actually exist between them leads to the well-known problem of specification for omitted variables. In fact, what disappears in this kind of regression is the error correction term. Analysing the causality we use the methodology from Hansen and Rand (2006). We consider bivariate vector autoregressive (VAR) models for the LnEXP and LnREV 5. Let x it =[LnEXP it, LnREV it ], where subscript i indexes regions (i=1,.n) while t indexes time (t=1, T).The VAR model for x it is specified as where A are matrices of parameters that are allowed to vary across countries, μ and δ are country specific intercept and trend parameters, λ is a time specific component and ε is an idiosyncratic error component assumed to be iid (0,Ω), with country specific positive definite covariance matrices. In this model, Granger non-causality from LnREV to LnEXP is formulated as the hypothesis H 0 : a 12(ji) =0 j=1,...p where a si are the elements in the A matrices. If the hypothesis is rejected, we say that LnREV Granger causes LnEXP. The reverse hypothesis of Granger non-causality from LnEXP to LnREV is given as H 0 : a 21(ji) =0 j=1, p The former VAR model can be remodelled as the error-correction form. Let In this formulation the dependent variables are the change in the variables. The error correction form is a convenient formulation for many other purposes. First of all, the hypotheses of Granger non-causality are unchanged by the linearity of the transformation. In the error correction form the hypotheses are 5 In the empirical analysis we consider three variables: Expenditure (EXP), Revenue (REV), and Gross Domestic Product (GDP). 29

7 Some authors separate the Granger-causality hypothesis into two sub-hypotheses of short and long-run causality. Short-run causality relates to hypotheses about zeros outside the diagonal in the Г matrices while long-run causality is about off-diagonal zeros in Π. In this paper, in keeping with Hansen and Rand (2006), the classical notion of Granger-causality is utilized and we denote the hypotheses about off-diagonal zeros as neutrality-hypotheses. The neutrality hypotheses are interesting because they can be used to relate cross-country studies using long averages over time and time-series and panel studies with annual observations. The relationship is given by the moving average representation of the model, which for large T, can be approximated by In this model the log-run impact of LnREV on LnEXP is estimated by c 12(i). The relationship between C and Π is, assuming the latter is invertible, given by, supposing that the Π matrix is (2x2). From this relationship the notion of neutrality is clear: if π 12(i) =0, then c 12(i) =0 and there is no long run impact from LnREV to LnEXP. This shows that cross-country studies using long averages and time series studies using annual observations may well differ in their conclusion about causality. The first kind of studies is testing neutrality while the second is testing causality. It should be clear that the only direct relationship is that Granger non-causality implies neutrality. Finally, if cointegration between LnEXP and LnREV is found, there will be implications on the computation and interpretation of the long run impact matrices Π i and C i, as both matrices have reduced rank. When it has reduced rank, in this model rank 1, it is convenient to write the matrix as a product of two matrices Π i =α i β i, both (2x1) matrices. In the cointegrated model the test for neutrality can still be based on significance of the parameters in the autorregresive representation because a zero-row in Π i corresponds to a zero-column in C i. If, say, LnREV is neutral for the long-run level of LnEXP, then α 2i =0 and it follows that π 21i =π 22i =0 and c 11i =c 21i =0. However, the interpretation of neutrality is somewhat different in cointegrated systems compared to stationary systems. In particular, even if neutrality of LnREV with respect to LnEXP is accepted, it cannot be concluded that LnREV has no impact on the long run level of LnEXP as they are both non-stationary. But, it can be concluded that the level of LnREV carries no information about the level of LnEXP. Empirical Results for the Spanish Autonomous Regions The series used herein are public expenditure, public revenue and gross domestic product, all of which are in constant euros from The deflators for each series have been used and the data have been taken from the period of Although previous data are available, 1987 was chosen as it is the first year of a new funding system after completing a socalled transition period. Moreover, data for the Basque Country and Navarre have been excluded because they have a Foral regime, different from the common system. 30

8 The causal relationship between public revenue and expenditure has been analysed using the GDP as a control variable. Therefore, two models are used: in the first, only the variables of interest are considered; and in the second, the GDP is included. The incorporation of real GDP in the model is done because both government spending and taxes or other revenues are sensitive to changes in real economic activity. For example, during recessions, government spending automatically rises (due to increased unemployment benefits and public subsidies) and taxes automatically fall (due to lower personal and corporate income). These automatic responses of fiscal variables to business conditions are commonly called the elements of automatic stabilization of the budget. It is therefore important that real GDP be included in the model to prevent contamination of the data with non-political effects (business cycle). Thus, in line with Anderson et al. (1986), von Furstenberg et al. (1986), Baghestani and McNown (1994), Ross and Payne (1998) and Kollias and Paleologou (2006), GDP is included as a variable control in the model. This approach makes it possible to distinguish between direct causal link between revenue and expenditure and indirect causal effects via GDP. Most papers that employ a bivariate approach suffer from the ever-present problem of the missing third variable, resulting in potential results and inaccurate conclusions. Table 1 displays the results of stationary tests for all variables in both levels and first differences. HERE TABLE 1 The tests do not reject the H 0 hypothesis for the existence of a unit root with the level of variables but they do reject the hypothesis with the first differences of the variables. Hadri s test rejects the null hypothesis every time but in this test there is a over-reject of the null hypothesis. Therefore, all variables are I(1) but never I(2). HERE TABLE 2 When the three variables are considered, all tests reject the null hypothesis of no cointegration between the variables, except in two cases. The Johansen-Fisher test provides for the existence of up to two cointegration relations. When the two variables are considered, it is discovered that all tests unequivocally point to the existence of cointegration and the Johansen-Fisher test demonstrates the existence of a cointegration vector. In accordance with the theoretical analysis the estimation is carried out by means of PMGE (Pesaran, Shin and Smith, 1999), which restricts long-run coefficients if they are identical in an error correction approach, but allows short-run coefficients and error variances to differ across the groups. In this way, the following mode is implemented: LnEXP it = 0i + 1i t+ ik LnEXP it-k + ik LnREV it-k + i LnEXP it-1 - i LnREV it-1 + it, where - i / i = i, as the long-run elasticity of public expenditure in relation to the GDP in each region. The estimation can be carried out using three procedures, the first of which is the group mean (MG) that separately estimates the previous equation for the 15 autonomous regions and averages out the estimated coefficients. In this way, the long-run elasticity will be 1/15 i = MG. This estimator is consistent, however, it does not exploit the cross-section dimension of the sample so as to take advantage of the given opportunity in the fact that some parameters may 31

9 be the same among regions. The approximation of the PMG estimator is based on the homogeneity of this coefficient; therefore, for each region i =. However, the speed of adjustment to imbalance over the long-term freely varies across autonomous regions. In this way, the previous equation can be rewritten as LnEXP it = 0i + 1i t+ ik LnEXP it-k + ik LnREV itk+ i (LnEXP it-1 - i LnREV it-1 )+ it. In the causality analysis, the coefficient ik must be non-zero and the coefficient i negative and the coefficient i different from zero. Finally, there is the traditional fixed effects estimator (dynamics) (FE), which allows the constant to differ between regions while the slope estimator is the same for all regions. However, this kind of pool does not permit differentiation between short-run and long-run dynamics. Furthermore, it often leads to inconsistent estimators of parameters (unless the slopes are indeed equal) and this inconsistency does not disappear when the sample size grows, both in its cross-section and time dimensions. Lastly, the PMG estimator allows dynamic specification over the short-term to vary among regions, which is not possible with the FE estimator. The following results are obtained: HERE TABLE 3 AND 4 When the variables LnEXP and LnREV are utilized unidirectional causality is obtained from LnREV to LnEXP both when the estimation is performed using PMG as using MG. There is no causality in the direction from LnEXP to LnREV, which indicates that in Spain public expenditure of ARs with the common system is determined by the revenue they receive. This result is expected due to the fact that approximately 70% of revenues in these regions are collected through central government transfers, while only 30% are AR own revenues (mainly transferred taxes). Similarly, this also happens in the long term when the neutrality hypothesis is rejected in the direction from revenue to expenditure while it is accepted in the direction from expenditure to revenue. When the GDP variable is introduced, results are equivalent to the above. If the estimation is considered using PMG, causality and non-neutrality from revenue to expenditure are once again found while we accept the null hypothesis of no existence of causality and neutrality in the direction from expenditure to revenue. Conclusions This study has utilized an analysis of unit roots, cointegration and Granger causality to study the relationship between public revenues and expenditures in the Spanish Autonomous Regions. The results demonstrate unequivocal unidirectional causality from revenue to expenditure. One implication of this result is that the size of the public sector at a regional level is not determined by the demand for spending but perhaps by the resources offered such as taxes and especially subsidies. Since in Spain much of the financing of the common regime ARs is made through central government transfers, a result like this was expected. 32

10 References Alesina, A. and Bayoumi, T. (1996) The cost and benefit of fiscal rules: evidence from U.S. states, Working Paper No. 5614, NBER, Cambridge, MA. Anderson, W., Wallace, M. S., and Warner, J. T. (1986) Government spending and taxation: what causes what?, Southern Economic Journal, 52, Baffes, J. and Shah, A. (1994) Causality and comovement between taxes and expenditures: historical evidence from Argentina, Brazil and Mexico, Journal of Development Economics, 44, Baghestani, H. and McNown, R. (1994) Do revenue or expenditures respond to budgetary disequilibria?, Southern Economic Journal, 61, Barro, R. J. (1979) On the determination of the public debt, Journal of Political Economy, 81, Buchanan, J. M. and Wagner, R.W. (1977) Democracy in deficit, Academic Press, New York. Chang, T. and Ho, Y-H. (2002) A note on testing tax-spend, spend and tax or fiscal synchronization: the case of China, Journal of Economic Development, 27, Chang, T., Liu, W. R. and Caudill, S. B. (2002) Tax and spend, spend and tax or fiscal synchronization: new evidence for ten countries, Applied Economics, 34, Dahberg, M. and Johanson, E. (1998) The revenues-expenditures nexus: panel data evidence from Swedish municipalities, Applied Economics, 30, Darrat, F. (1998) Tax and spend or spend and tax? An inquiry into the Turkish budgetary process, Southern Economic Journal, 64, Friedman, M. (1978) The limitations of tax limitations, Policy Review, 5, Hasan, M. and Lincoln, I. (1997) Tax then spend or spend then tax? Experience in the Uk , Applied Economics Letters, 4, Hansen, H. and Rand, J. (2006) On the causal links between FDI and growth in developing countries, The World Economy, 29(1), Ho, Y-H. and Huang, C. J. (2009) Tax-spend, spend-tax or fiscal synchronization: a panel analysis of the Chinese provincial real data, Journal of Economics and Management, 5(2), Joulfain, D. and Mookerjee, R. (1990) The government revenue-expenditures nexus: Evidence from a state, Public Finance Quarterly, 18(1), Kollias, C. and Paleologou, S. M. (2006) Fiscal policy in the European Union. Tax and spend, spend and tax, fiscal synchronization or institutional separation?, Journal of Economic Studies, 33(2), Marloe, M. and Manage, L. (1987) Expenditures and receipts: testing for causality in state and local government finances, Public Choice, 53, Marloe, M. and Manage, L. (1988), Expenditures and receipts in state and local government finances: reply, Public Choice, 59, Meltzer, A. and Richard, S. (1981) A rational theory of the size of government, Journal of Political Economy, 89, Miller, S. M. and Russek, F. S. (1990) Cointegration and error correction models: temporal causality between government taxes and spending, Southern Economic Journal, 57, Moisio, A. (2000) Spend and tax or tax and spend?. Panel data evidence from Finnish municipalities during , VATT Discussion Papers No. 242, pp

11 Musgrave, R. (1966) Principles of budget determination, in Public Finance: Selected Readings, (Eds) H. Cameron and W. Henderson, Random House, New York, pp Payne, J. E. (1997), The tax-spend debate: the case of Canada, Applied Economics Letters, 4, Payne, J. (2003) A survey of the international empirical evidence on the tax-spend debate, Public Finance Review, 31, Peacock, A. and Wiseman, J. (1979) Approaches to the analysis of government expenditures growth, Public Finance Review, 31, Poterba, J. (1994), State responses to fiscal crises: the effects of budgetary institutions and politics, Journal of Political Economy, 102, Ram, R. (1988a) A multicountry perspective on causality between government revenue and government expenditure, Public Finance, 43, Ram, R. (1988b) Additional evidence on causality between government revenues and government expenditure, Southern Economic Journal, 54, Reddick, C. and Hassan, S. (2003) Long run and short run budgeting: empirical evidence for Canada, UK and USA, Journal of Public Budgeting, Accounting and Financial Management, 15, Roberts, P. (1978) Idealism in public choice theory, Journal of Monetary Economics, August, Ross, K. and Payne, J. E. (1998) A reexamination of budgetary disequilibria, Public Finance Review, 26, Von Furstensberg, G. M., Green, R. J. and Jeong, J. H. (1986) Tax and spend or spend and tax?, Review of Economics and Statistics, 68, Sorensen, B., Wu, L. and Yosha, O. (2001) Output fluctuations and fiscal policy: U.S. state and local governments , European Economic Review, 45, Wagner, R.E. (1976) Revenue structure, fiscal illusion and budgetary choice, Public Choice, Spring, Westerlund, J., Mahdavi, S. and Firoozi, F. (2009) The tax-spending nexus: evidence from a panel of US State-Local governments, Working Papers in Economics No. 378, University of Gothenburg. Wildavsky, A. (1988) The new politics of budgetary process, Glenview, Scott Foresman. Zapf, M. and Payne, J. (2009) Asymmetric modelling of the revenue-expenditure nexus. Evidence from aggregate state and local government in the US, Applied Economics letters, 16,

12 Table 1. Stationary tests for the series LnEXP LnEXP a LnREV LnREV LnGDP lngdp H o = unit root (common process) Constant and trend Levin, Lin and Chu (0.3521) b (0.0305) (0.9256) (0.0003) (0.2078) (0.0000) Breitung (0.8134) (0.0000) (0.9040) (0.000) (0.0919) (0.0000) H o = unit root (individual process) Constant and trend Im, Pesaran and Shin (0.7539) (0.0061) ( ) (0.0002) (0.2909) (0.0085) ADF- Fisher * (0.5571) (0.0106) (0.9921) (0.0004) (0.4595) (0.0223) PP-Fisher * (0.0845) (0.000) (0.1409) (0.0000) (0.9337) (0.000) Hadri,H o = there is not unit root Constant and trend (0.0000) (0.000) (0.0000) (0.000) (0.000) (0.000) Notes: a First differences of the variables. b Numbers in parentheses are p-values. * In Fisher s test, the p-values are computed using a chi-square asymptotic distribution. For the rest of the tests, asymptotic normality is assumed. 35

13 Table 2. Cointegration test Pedroni residual cointegration LnEXP, LnREV, LnGDP LnEXP, LnREV test H 0 = No cointegration H 1 = common AR coefficients (within-dimension) Panel v (0.0174) (0.0000) Panel ρ (0.0781) (0.0000) Panel PP (0.0000) (0.0000) Panel ADF (0.0008) (0.0000) H 1 = individual AR coefficients (between-dimension) Group ρ (0.4776) (0.0051) Group PP (0.0000) (0.0000) Group ADF (0.0000) (0.0000) Kao ADF (0.0000) (0.0000) Johansen Fisher Maximum trace test Own value Maximum trace test Own value test test H 0 (No cointegration vector) (0.0000) (0.0000) (0.0000) (0.0000) At most one (0.0002) (0.0000) (0.0004) (0.0000) At most two (0.2342) (0.2342) 36

14 Table 3. Causality using LnEXP, LnREV PMG estimation MG estimation FE estimation Dep variable LnEXP LnREV LnEXP LnREV LnEXP LnREV LnEXP(t-1) (0.9534) ) (2.0396) (0.5347) (0.6564) (0.3965) LnEXP(t-2) (0.8343) (0.3360) (1.0128) (0.5659) (1.1967) (1.6803) LnREV(t-1) ) (0.7554) ) ) ) ) LnREV(t-2) ) (1.3329) ) ) ) ) LnEXP(t-1) ) (0.3378) (2.2188) (1.3980) LnREV(t-1) (2.5861) ) ) ) EC ) Granger causality Neutrality

15 Table 4. Granger-causality with three variables PMG estimation MG estimation FE estimation Dep variable LnEXP LnREV LnEXP LnREV LnEXP LnREV LnEXP(t-1) (1.0354) (0.1258) (2.2556) (0.6679) (0.6610) (0.3640) LnEXP(t-2) (0.5165) (0.2398) (1.0149) (1.3712) LnREV(t-1) ) LnREV(t-2) (0.0240) LnGDP (t-1) (0.9747) LnGDP(t-2) (3.3182) ) (0.2070) ) ) ) ) (0.1836) ) LnEXP(t-1) ) LnREV(t-1) (1.4030) LnGDP(t-1) ) EC ) ) ) ) (0.8487) (0.8595) ) (0.7564) (0.3317) (0.5252) ) ) (0.3032) (3.3504) ) (2.2140) (1.0605) ) ) ) (3.6899) ) ) (1.6732) Granger causality Neutrality