Fiscal structures and economic growth: international evidence.

Miller, Stephen M. ; Russek, Frank S.

I. INTRODUCTION

A large and growing body of literature searches for the determinants of economic growth employing cross-country regression analysis (e.g., Kormendi and Meguire [1985], Grier and Tullock [1989], Barro [1990; 1991; 1992], Romer [1989; 1990], Levine [1991], and Levine and Renelt [1992]).(1) The cross-country regression approach implicitly assumes that the growth process possesses similar structural properties across the countries in the sample. Structural differences, be they political, economic, social, or other, between countries, therefore, do not condition the growth process. Or if they do, then the effects are randomly distributed with zero mean. If structural differences between countries do matter significantly and non-randomly in the growth process, then the existing cross-country research is potentially flawed. Attempts are sometimes made to control for such differences by including dummy variables for different regions of the world (e.g., dummy variables for African, Latin American, and other groups of countries).

Within the literature on the determinants of economic growth, some papers (e.g., Landau [1983; 1985; 1986], Kormendi and Meguire [1985], Ram [1986], Koester and Kormendi [1989], Barth, Keleher, and Russek [1990], Easterly and Rebelo [1993], and Fischer [1993]) consider the effects, if any, of government expenditure and revenue on economic growth. Overall, the findings of these studies paint a mixed picture; the relationship between growth and government is sometimes significant and positive, sometimes significant and negative, and sometimes not significant. Although these studies consider many different fiscal variables, they do not as a rule examine the effects of these fiscal variables in a systematic way that controls for the mode of financing. They generally overlook, for example, how an increase in government consumption expenditure is financed. Is it with higher taxes, and which taxes? Or is it with lower spending, and which spending? Finally, is it with a larger deficit? Controlling for these factors may lead to different findings about the effect of an increase in government consumption expenditure on economic growth.

Borrowing from some work on the effect of fiscal structure on national economic growth within the United States in Helms [1985], Mofidi and Stone [1990], and Miller and Russek [1993], we introduce the government budget constraint into the regression equations that examine the determinants of economic growth. As a result, we directly address the three financing questions raised in the previous paragraph. In this framework, we consider the growth process using pooled times-series, cross-section data and comparing the results from fixed- and random-effects econometric techniques.(2) These techniques attempt to accommodate structural differences between countries and across time. We also compare the performance of the more-standard ordinary least squares regression approach with that of the fixed- and random-effect models.

II. AN OVERVIEW OF INTERNATIONAL FISCAL STRUCTURES

Before presenting our model and statistical findings, a brief overview of the growth and fiscal characteristics of the countries in our sample provides some additional information for those not familiar with the data, which otherwise would not be revealed by our analysis.(3) This information can contribute in a limited way to the debate about the effects of government on economic growth, although it is not a substitute for modeling the growth process.

For the sample as a whole, the growth rate of real per capita gross domestic product (GDP) averaged 1.35% per year during 1975 to 1984, with a coefficient of variation of 3.52. By comparison, the average growth rate in the United States was 1.45%. The five fastest growth countries were Botswana (6.72), Indonesia (4.19), Korea (6.32), Sri Lanka (3.20), and Thailand (4.31). The five slowest growth countries were El Salvador (-1.74), Iran (-2.16), Liberia (-2.78), Venezuela (-2.21), and Zambia (-3.33).

The average fiscal position of the countries in our sample was a deficit relative to GDP of 2.71%, with a coefficient of variation of 1.72. This average was slightly less than the 2.78% average deficit experienced by the United States. The five countries with the worst fiscal positions were Belgium (7.95), Israel (11.27), Morocco (10.71), Sri Lanka (9.52), and Zambia (9.78). Nine countries had surpluses instead of deficits, with the largest five surpluses relative to GDP appearing in Botswana (5.56), Brazil (3.16), Luxembourg (1.61), Swaziland (2.19), and Venezuela (3.05).

Total central government revenue averaged 28.3% of GDP, with a coefficient of variation of 0.39. The five countries with the highest revenue-GDP ratios were Belgium (43.74), Botswana (44.64), Israel (59.81), Luxembourg (48.92), and the Netherlands (49.81). Eight countries had ratios below 20%. The five with the lowest were Costa Rica (18.69), El Salvador (13.26), Korea (17.26), Paraguay (11.08), and Thailand (14.18).

Finally, total central government expenditure averaged 31.01% of GDP, with a coefficient of variation of 0.39. The five countries with the highest spending-GDP ratios were Belgium (51.69), France (39.89), Israel (71.08), Luxembourg (47.31), and the Netherlands (53.15). The five with the lowest were Brazil (19.93), El Salvador (16.12), Korea (16.58), Paraguay (10.94), and Thailand (17.80). These were also the only countries with spending less than 20% of GDP.

The largest share of total revenue, on average, was collected as revenue from domestic taxes on goods and services (24.55%). In descending order of quantitative importance, the other six sources of revenue are individual income tax revenue (16.48), non-tax revenue (15.34), social security tax revenue (14.57), international trade tax revenue (13.24), corporate income tax revenue (11.17), and all other tax revenue (4.65). Of these, the ones with the largest and smallest coefficients of variation, respectively, were corporate income tax revenue (1.15) and revenue from domestic taxes on goods and services (0.52).

The largest share of total spending, on average, was allocated to social security and welfare (23.61%). In descending order of quantitative importance, the other six categories of spending are other expenditure (20.32), economic affairs and services expenditure (19.12), education expenditure (12.10), defense expenditure (9.32), health expenditure (7.86), and transportation and communication expenditure (7.67). Of these, the ones with the largest and smallest coefficients of variation, respectively, were defense expenditure (0.88), and education (0.50) and economic affairs and services (0.50) expenditures.

III. THE MODEL AND ECONOMETRIC METHOD

Our modeling of national economic growth borrows from some work on state and local economic growth within the United States in Helms [1985], Mofidi and Stone [1990], and Miller and Russek [1993]. We begin by defining the growth rate of gross domestic product per capita (g) as follows:

(1) [g.sub.ct] = ln[y.sub.ct] - ln[y.sub.ct-1],

where y is real gross domestic product per capita, ln is the natural logarithm operator, and c and t indicate the country and time period. Let [X.sub.ct] = ([[x.sup.1].sub.ct], [[x.sup.2].sub.ct], ... [[x.sup.n].sub.ct]) represent those observable factors (e.g., investment, tax and spending patterns, and so on) that can influence national economic growth. Thus, we model national economic growth as follows:

(2) [g.sub.ct] = [Alpha] + [summation over] [[Beta].sub.i] [[x.sup.i].sub.ct] where i = 1 to n + [v.sub.ct].

where [v.sub.ct] is the error term.

The error term [v.sub.ct] incorporates the influences of omitted variables. Classical regression analysis assumes that omitted variables are independent of the included [x.sub.ct] and are independently and identically distributed. When using pooled cross-section, time-series data, however, the omitted variables can be further classified into three groups - country-varying, time-invariant; time-varying, country-invariant; and country- and time-varying variables.(4) The country-varying, time-invariant variables differ across countries, but are constant within a given country over time (i.e., [C.sub.c] give essentially constant country-specific information). Time-varying, country-invariant variables differ over time, but are constant at a point in time across countries (i.e., [T.sub.t] give essentially constant time-specific information). Examples of the former variables include geography and climate, while examples of the latter include world economic conditions such as Euro interest rates. Finally, the country- and time-varying variables differ across both countries and time. Thus, the error term [v.sub.ct] can be written as follows:

(3) [v.sub.ct] = [Delta][C.sub.c] + [Mu][T.sub.t] + [[Phi].sub.ct],

where [Delta] and [Mu] measure the effects of [C.sub.c] and [T.sub.t] on [v.sub.ct].

Substituting equation (3) into (2) gives the following:

(4) [g.sub.ct] = [Alpha] + [summation over] [[Beta].sub.i][[x.sup.i].sub.ct] where i=1 to n + [Delta][C.sub.c] [Mu][T.sub.t] + [[Phi].sub.ct],

Estimation of equation (4) with ordinary least squares and without consideration of possible country-specific or time-specific effects can lead to serious errors. Hsiao [1986, 7] provides illustrations of misleading results. Problems emerge when either the unobservable country-specific or time-specific variables correlate with the included variables [x.sub.ct]. Two alternative, but related, approaches exist for addressing these problems - fixed- and random-effect models.

Fixed-Effect Models

Suppose that omitted country-specific variables that are correlated with the included [x.sub.ct] constitute the problem. Then adjusting the dependent and independent variables by subtracting the mean of each variable over time solves the problem.(5) Since the unobserved country-specific variables and the intercept do not change over time, the subtraction of their respective means over time drops these variables out of the regression equation. If this is the only problem in the estimation of equation (4), then the regression adjusted for the means across time provides unbiased and consistent estimates of [[Beta].sub.1]. Without this adjustment, the ordinary least squares estimates are biased and inconsistent.

Similarly, if time-specific variables correlate with the included [x.sub.ct], then a similar solution adjusts the dependent and independent variables by subtracting the mean of each variable over countries. Since each country faces the same time-specific effects, subtracting the means over countries drops the intercept and the time-specific effects out of the revised regression equation. Once again, the revised regression provides unbiased and consistent estimates of [[Beta].sub.i].

Of course, both country- and time-specific effects may correlate with the included [x.sub.ct]. In this case, we can adjust for the means over countries and time.

Random-Effect Models

The fixed-effect model assumes that the differences across units - either countries or time - reflect parametric shifts in the regression function. Such a view becomes more appropriate when the problem uses the whole population rather than a sample from it. If the problem examines only a sample from a larger population, then the fixed-effect model can be properly interpreted as applying only to the differences within that sample. Our problem considers a sample of countries. Therefore, the random-effect model needs consideration.

Random-effect models treat the country-specific ([e.sub.c]) and time-specific ([e.sub.t]) effects as random variables. Thus, the error term [v.sub.ct] is viewed as having three random components - [e.sub.c], [e.sub.t], and [[Epsilon].sub.ct]. These error terms have the following properties:

(5) E [e.sub.c] = E [e.sub.t] = E [[Epsilon].sub.ct] = 0;

E [e.sub.c] [e.sub.t] = E [e.sub.c] [[Epsilon].sub.ct] = E [e.sub.t] [[Epsilon].sub.ct] = 0;

E [e.sub.c] [e.sub.i] = [[[Sigma].sup.2].sub.C], if c = i; 0 otherwise;

E [e.sub.t] [e.sub.j] = [[[Sigma].sup.2].sub.T], if t = j; 0 otherwise;

E [[Epsilon].sub.ct] [[Epsilon].sub.ij] = [[[Sigma].sup.2].sub.[Epsilon]], if c = i

and t = j; 0 otherwise; and [e.sub.c], [e.sub.t],

and [[Epsilon].sub.ct] are each uncorrelated with [x.sub.ct].

The variance of the growth rate of real per capita GDP conditional on the explanatory variables is given from equation (4) as follows:

(6) [[[Sigma].sup.2].sub.G] = [[[Sigma].sup.2].sub.C] + [[[Sigma].sup.2].sub.T] + [[[Sigma].sup.2].sub.[Epsilon]],

where [[[Sigma].sup.2].sub.G] is the variance of the growth rate of gross domestic product per capita left unexplained by the explanatory variables [x.sub.ct]. As a consequence, this formulation - the random-effect model - is frequently called a variance-components (error-components) model.

If we know the variance components, then the estimation of the random-effect model using generalized least squares (GLS) merely requires the transformation of the dependent and independent variables using the variance components in the appropriate way. Absent knowledge of the variance components, then we must first provide estimates of these components and apply a feasible GLS procedure to estimate the equation.

The Government Budget Constraint and the Interpretation of Results

As discussed above, our method borrows from the empirical literature on the effect of state and local government taxes and spending on state and local economic growth within the United States. That research presents a mixed picture; no consensus exists even on the signs of the effects. Helms [1985, 577] provides a rationale for the divergent results - "... it is not meaningful to evaluate the effects of tax or expenditure changes in isolation: both the sources and uses of funds must be considered." In other words, the regression equations need to include all but one of the possibilities for sources and uses - various revenues, various expenditures, and the surplus. Some uses may raise growth when associated with some sources, but not when associated with others.

We generate a taxonomy of results by excluding, in turn, different revenue and expenditure categories and the surplus from the regression equations. This approach follows the work of Miller and Russek [1993] on state and local economic growth.(6) This more thorough analysis deals explicitly with the overall fiscal structure and may lead to findings not revealed by the more limited approach of including fiscal variables on a more ad hoc basis, as has been the norm in the cross-country growth literature to date.

IV. REGRESSION EQUATIONS AND HYPOTHESES

Our regression equations fall into two distinct categories - equations that do not disaggregate total revenue and expenditure and equations that do. Each of these two sets of regressions includes a set of conditioning variables that have been found to be important in other cross-country growth regressions, including lagged real per capita GDP, the rate of growth of population, the investment share of GDP, the import plus export share of GDP, and the GDP implicit price deflator inflation rate. These two types of regression equations are given as follows:

(7) [g.sub.ct] = [a.sub.1] + [a.sub.2][y.sub.ct-1] + [a.sub.3] [n.sub.ct]

+ [a.sub.4] [inv.sub.ct] + [a.sub.5] [opn.sub.ct] + [a.sub.6] [p.sub.ct]

+ [a.sub.7] [rev.sub.ct] + [a.sub.8] [exp.sub.ct] + [a.sub.9] [sur.sub.ct] + [v.sub.ct],

and

(8) [g.sub.ct] = [b.sub.1] + [b.sub.2][y.sub.ct-1] + [b.sub.3] [n.sub.ct]

+ [b.sub.4] [inv.sub.ct] + [b.sub.5] [opn.sub.ct] + [b.sub.6] [p.sub.ct]

+ [b.sub.7] [rci.sub.ct] + [b.sub.8] [rii.sub.ct] + [b.sub.9] [rss.sub.ct]

+ [b.sub.10] [rdgs.sub.ct] + [b.sub.11] [rtrd.sub.ct] + [b.sub.12] [rot.sub.ct]

+ [b.sub.13] [rnt.sub.ct] + [b.sub.14] [edfs.sub.ct] + [b.sub.15] [eed.sub.ct]

+ [b.sub.16] [ehlh.sub.ct] + [b.sub.17] [ess.sub.ct] + [b.sub.18] [eeas.sub.ct]

+ [b.sub.19] [etc.sub.ct] + [b.sub.20] [eoe.sub.ct] + [b.sub.21] [sur.sub.ct] + [v.sub.ct],

where g is the growth rate of real per capita GDP, y is real per capita GDP, n is the rate of growth of population, inv is the investment share of GDP, opn is the import plus export share of GDP, p is the GDP implicit price deflator rate of inflation, rev is total government revenue to GDP, exp is total government expenditure to GDP, sur is the government surplus to GDP (i.e., rev - exp), tci is corporate income tax revenue to GDP, rii is individual income tax revenue to GDP, rss is social-security tax revenue to GDP, rdgs is domestic goods and services tax revenue to GDP, rtrd is international trade tax revenue to GDP, rot is other tax revenue to GDP, rnt is non-tax revenue to GDP, edfs is defense expenditure to GDP, eed is education expenditure to GDP, ehlh is health expenditure to GDP, ess is social-security and welfare expenditure to GDP, eeas is economic affairs and service expenditure to GDP, etc is transportation and communication expenditure to GDP, and eoe is other expenditure to GDP.

How do we interpret the coefficients on the fiscal variables in these two equations? Consider the revenue, expenditure, and surplus coefficients ([a.sub.7], [a.sub.8], and [a.sub.9]) in equation (7). Because of the underlying identity, only two of these coefficients can be estimated in the same regression, the third is automatically implied. Moreover, at least two of the variables associated with these coefficients must be included to control for the sources and uses of funds. Excluding sur allows the surplus to change freely and the coefficient [a.sub.7] measures the effect on economic growth of an increase in the revenue share of GDP, assuming no change in the expenditure share. In other words, the additional revenue reduces (increases) the deficit (surplus). Similarly, [a.sub.8] measures the effect of an increase in the government expenditure share of GDP, assuming no change in revenue relative to GDP. Of course, this implies that the expenditure was debt-financed. Excluding rev or exp from equation (7) rather than the sur implies that [a.sub.9] measures the effect on economic growth of an increase in the surplus financed by a change in the excluded variable. Thus, it does not matter which of the three variables is excluded, so long as only one is excluded.

Similar interpretations carry over to equation (8), where the components of revenue and expenditure appear instead of the aggregates. Excluding sur causes the corporate income tax coefficient [b.sub.7] to measure the effect on economic growth of an increase in the surplus financed entirely by an increase in corporate tax revenue relative to GDP. Similarly, [b.sub.14] measures the effect on economic growth of a debt-financed increase in defense spending relative to GDP. If this increase in defense spending were instead financed by an increase in corporate taxes relative to GDP, then the effect on economic growth is measured by the sum of [b.sub.7] and [b.sub.14], assuming no change in the government surplus.

In sum, three regression results for equation (7) can be calculated for the cases where rev, exp, and sur are deleted in turn. Fifteen regression results for equation (8) can be calculated for the cases where individual revenue items, individual expenditure items, and sur are deleted in turn. In fact, only two independent regression equations exist - one for equation (7) and one for equation (8). We report only the regression results that eliminate the surplus.(7)

The specifications shown for equations (7) and (8) were modified in several ways before estimation to develop a richer and more robust set of findings. Some concern may emerge that when we use annual data instead of data averaged over time, the fiscal coefficients in equations (7) and (8) capture cyclical rather than long-term trend effects. This issue is difficult to address fully because of the constraints imposed by our sample coverage.(8) Nevertheless, we adopt the strategy of adding the lagged value of the dependent variable as another regressor. With this modification, the long-run effects are gauged by dividing the fiscal coefficient estimates by one minus the coefficient of the lagged dependent variable.

The commingling of information from developed and developing countries may also appear troublesome to some, especially if the determinants of economic growth differ over different stages of development. Consequently, we break our sample into developed and developing countries and redo all the econometric work. These new findings are reported alongside the results for the full (all-country) sample.(9)

Finally, because of the interaction between government budgets and economic conditions, one criticism of these equations - and of most of the existing literature - is that variations in growth may be incorrectly attributed to variations in fiscal variables when correlations exist without causation. As one attempt to address this issue, we estimate equations (7) and (8) using once-lagged values of all fiscal variables.(10)

V. EMPIRICAL RESULTS

All specifications of equations (7) and (8) discussed above were estimated with fixed-and random-effect models, in addition to ordinary least squares (OLS) estimation.(11) Then, three tests were used to determine the appropriate specification in each case. An F-test compares the fixed-effect model and the OLS model as in Greene [1990, 484]. A Lagrange-Multiplier test due to Breusch and Pagan [1980] compares the random-effect model with the OLS model as in Greene [1990, 49192]. And a Wald criterion due to Hausman [1978] compares the random-effect model with the fixed-effect model as in Greene [1990, 495].

The tests of alternative specifications convey a generally consistent story, at least for the fixed- and random-effect models that account for differences between countries and over time. In all cases, the fixed-effect model dominates the OLS model; the random-effect model does not. In addition, the fixed-effect model dominates the random-effect model specified between countries and over time. The one anomaly is that the random-effect model occasionally dominates the fixed-effect [TABULAR DATA FOR TABLE I OMITTED] model when changes over time are not isolated. This occasional inconsistency in the three tests may suggest that the fixed-effect model between countries alone is sometimes misspecified. Thus, we report in Tables I and II only the fixed-effect models between countries and over time.(12)

Several findings stand out.(13) The non-fiscal (conditioning) variables tell a generally consistent story, and a story reasonably consistent with the existing literature. First, the coefficient of lagged real per capita GDP is significantly negative in all regressions, supporting the conditional convergence hypothesis.(14) Second, the investment share of GDP is significantly positive in all cases. Levine and Renelt [1992] report this result as a "robust" finding, appearing consistently across the empirical studies of the determinants of economic [TABULAR DATA FOR TABLE II OMITTED] growth. Third, the inflation rate generally has a significant negative effect, although sometimes this effect is not significant.(15) Fischer [1993] finds a significant negative effect. While Levine and Renelt [1992] find the inflation rate effect to be fragile; they do find it to be consistently negative. Other authors such as Kormendi and Meguire [1985] and Grier and Tullock [1989] report some evidence of a negative effect of inflation on economic growth. Grier and Tullock's strongest evidence is for the African countries, a few of which are in our sample. When spending and taxes are disaggregated, however, the significant negative effect only emerges for developing countries; developed countries' growth rates are not significantly affected by inflation. Finally, the population growth and openness (imports plus exports to GDP) variables have coefficients that are negative and positive, respectively, but generally insignificant. A negative sign for the coefficient of population implies that real output growth adjusts at less than one-to-one with population growth. Levine and Renelt [1992] report a robust positive effect of a country's openness on the investment share of GDP; the effect of openness on real per capita growth was fragile, but positive.

Focusing on the effects of aggregate taxes and spending for the all-country results, several observations emerge. First, the effect of government expenditure on economic growth depends crucially on the method of financing. Tax-financed increases in government expenditure stimulate economic growth (i.e., the positive coefficient of government revenue significantly exceeds in absolute value the negative coefficient of government expenditure), while debt-financed increases in government expenditure retard economic growth. Second, reducing the government deficit stimulates economic growth. Reducing expenditure while holding revenue constant (i.e., a lower government deficit) or increasing revenue while holding expenditure constant (i.e., also a lower government deficit) both stimulate economic growth. Moreover, the tax-financed reduction in the government deficit has a larger effect than the expenditure-financed reduction. Both Easterly and Rebelo [1993] and Fischer [1993] report similar findings.

The effects of aggregate government spending and revenue differ between the developed and developing country regressions. A debt-financed increase in government spending produces a significant decrease in economic growth for developing countries, but an insignificant effect for developed countries. On the other hand, substituting revenue for debt issue significantly increases economic growth in developing countries, but reduces growth in developed countries. Finally, a revenue-financed increase in government spending increases the real per capita growth rate in developing countries, but reduces growth in developed countries. These findings argue for more tax-financed government spending in developing countries, but for less tax-financed spending in developed countries.

When we disaggregate government spending and taxes into their component parts, other interesting findings emerge. First, debt-financed increases in defense spending generally reduce economic growth for developing countries across all specifications, but increase growth for developed countries in those regressions that include lagged fiscal variables. Second, a debt-financed increase in education spending raises economic growth in developed countries across all specifications, but reduces growth in developing countries, according to the specification that uses contemporaneous fiscal variables. Third, a debt-financed increase in health spending or social security spending reduces economic growth in developing countries according to all specifications, but generally has no significant effect in developed countries, although the sign is negative.

VI. CONCLUSION

We examine the effects of national fiscal structures on national economic growth, using an international sample of developed and developing countries and alternative econometric techniques. We adopt the method of Helms [1985], Mofidi and Stone [1990], and Miller and Russek [1993], who considered the determinants of state and local economic growth in the United States. The approach incorporates the government budget constraint into the growth regressions so that we can clearly identify how a particular change in fiscal policy is financed (e.g., the effect of a debt-financed increase in defense spending).

We can succinctly state our findings concerning the effects of fiscal structure on economic growth. First, the method of financing government expenditure plays an important role in determining the effect of that expenditure on economic growth. We find that for developing countries, debt-financed increases in government expenditure retard economic growth and tax-financed increases lead to higher growth, while for developed countries, debt-financed increases in government expenditure do not affect economic growth and tax-financed increases lead to lower growth. The differences between these developed and developing country results may reflect differences in the effect of money-financed and bond-financed spending increases. If developing countries when faced with the need to debt-finance spending more frequently use money rather than bonds, then these differences suggest that money-financed spending retards economic growth while bond-financed spending does not. Of course, we have not examined this issue. Future research may follow this path.

Second, different expenditure categories affect growth differently. Debt-financed increases in defense, health, and social security and welfare expenditures retard growth in developing countries. On the other hand, debt-financed increases in education expenditure stimulate growth in developed countries. Such findings may suggest that defense, health, and social security and welfare expenditures represent too large a share of the government budget in developing countries while education expenditure represents too small a share of the government budget in developed countries, at least when considering economic growth. Future research may want to consider whether for particular expenditure categories optimal shares of government expenditure exist for promoting economic growth.

Before closing, one caveat needs mention. Examination of the literature on the determinants of growth strongly suggests that any one study does not make a case. Results are sensitive to the variables included as noted by Levine and Renelt [1992] as well as to the countries and time periods covered as noted by Clark [1993].

APPENDIX

Our data come from two sources. First, we use information on real and nominal gross domestic product, population, imports and exports of goods and non-financial services, gross domestic investment, and the base year PPP convergence factor from 1975 to 1984, which come from the World Bank data tape. Second, we use information on central government revenue and spending from 1975 to 1984, which was compiled by the International Monetary Fund and distributed in the Government Finance Statistics (GFS) data tape. Revenue categories include total revenue and grants; income, profit, and capital gains tax revenue broken out by corporate and individual classes; social-security tax revenue; domestic taxes on goods and service revenue; international trade tax revenue; and total tax revenue. From these items, we construct as residuals other tax and non-tax revenue. Expenditure categories include total expenditure; defense expenditure; education expenditure; health expenditure; social-security and welfare expenditure; economic affairs and services expenditure; and transportation and communication expenditure. We construct as a residual other expenditure.

Preliminary examination of the GFS data suggested that only 44 countries had the detailed information identified in the previous paragraph. Moreover, this data for these countries had to be restricted to 1975 to 1984. After downloading the data, we discovered that Cyprus, the Solomon Islands, and Uganda were missing one of the needed data series for at least part of the sample period. These countries were deleted from the sample. Finally, after examining the summary statistics, we discovered two additional countries with problems in the other tax and other expenditure variables. Mexico collects taxes on behalf of state governments. This money is rebated to state governments. Thus, our constructed other tax revenue variable turned out to be negative for Mexico. In the Philippines, the data are adjusted to a cash basis between the reporting of expenditure sub-categories and total expenditure. Thus, the other expenditure category in the Philippines was negative. We deleted both Mexico and the Philippines from our final sample.

Finally, as noted by Easterly and Rebelo [1993], who also use the GFS data, this data possess two additional shortcomings. One, few countries provide information on revenue and expenditure of local governments or public enterprises. Thus, we use central government data, as do Easterly and Rebelo. Even here, our country coverage and time series length is restricted because of the break-out of expenditure and revenue categories; many countries do not report sufficient detail for our interest. Two, the IMF's data are sometimes based on budget data.

An earlier version of this paper was presented at the 1993 Western Economic Association meetings in Lake Tahoe. We acknowledge the helpful comments of Mark Wohar, the discussant at the Western meetings. This early version was also presented at the University at Auckland and the University of Sydney. This paper was most recently presented at the 1995 Finance and Macroeconomics Meetings at Taichung, Taiwan. The views expressed are the authors', and do not necessarily reflect those of the Congressional Budget Office or its staff.

1. Typically, the data for each country are averaged over the time-series sample (e.g., the average growth rate of real gross domestic product per capita for a number of years).

2. Grier and Tullock [1989] and Barro [1992] come the closest to our method. Both divide their samples into five-year subperiods and calculate average growth rates over these subperiods. Thus, they have a pooled cross-section, time-series data base. Moreover, for their OECD findings, Grier and Tullock include dummy variables for each time period, save one, producing fixed-effect results across time. For their non-OECD findings, Grier and Tullock include dummy variables for time periods as well as for some geographic regions (i.e., Africa and the Americas), approximating fixed-effect results across time and regions (but not across countries). Barro, on the other hand, includes geographic dummy variables (i.e., sub-Saharan Africa and Latin America), approximating fixed-effect results across regions.

3. We employ a panel of 39 countries with annual data for 1975 to 1984. The data appendix discusses how the countries and sample years were chosen. Summary statistics for all countries and for all variables are not printed to conserve space. Tables are available on request.

4. Since our study uses pooled cross-section, time-series data, we shall be referring to the method associated with pooled estimation. Our discussion draws on Hsiao [1986] and Greene [1990].

5. An alternative procedure is to estimate the first-differenced regression. Lagging equation (4) one period and subtracting the lagged equation from equation (4) causes the intercept and the state-specific terms to drop out.

6. Space limitations preclude the reporting of results that exclude fiscal variables other than the surplus. These other results are available on request.

7. Other results are available on request.

8. Our sample from 1975 to 1984 includes several oil price shocks that the reader needs to keep in mind when evaluating our findings.

9. The developed country subsample includes Australia, Austria, Belgium, Canada, Denmark, Finland, France, Germany, Iceland, Luxembourg, the Netherlands, Spain, Sweden, Switzerland, the United Kingdom, and the United States, 16 countries. The remaining 23 countries in our all-country sample are in the developing country sub-sample.

10. The timing of fiscal data reporting across countries varies slightly in response to differences in fiscal years - some fiscal years end in June, one or two in September, others in December, and still others in March.

11. The transformations for the various fixed- and random-effects models are documented in Judge et al. [1985, 521, 524, 532, and 535].

12. These tables include results for all countries and for countries subdivided into developed and developing categories as well as results that include and exclude the once-lagged dependent variable. Table I and II report the findings for aggregated and disaggregated fiscal variables, respectively. Space limitations prevent the reporting of similar findings for these specifications using once-lagged fiscal variables, both aggregated and disaggregated. Tables that include these additional findings are available on request.

13. Unless otherwise stated, the discussion of findings relate to Tables I and II as well as to the similar results (not reported) when once-lagged fiscal variables are used. See footnote 12 for more details.

14. We also performed several tests (not reported) of conditional and unconditional convergence absent the fiscal variables to place our findings in the context of the existing literature. These results (available on request) support conditional but not unconditional convergence, the standard result in the literature such as in Barro [1991] and Levine and Renelt [1992]. Researchers find evidence of unconditional convergence when the sample includes only developed countries, as in Mankiw, Romer, and Weil [1992]. Our sample includes both developed and developing countries.

15. Clark [1993, 23] considers the relationship between growth and inflation in cross-country regressions and concludes that "the cross-country relationship between long-term growth and inflation is, at best, tenuous." The results are sensitive to the sample of countries as well as the time period.

REFERENCES

Barro, Robert J. "Government Spending in a Simple Model of Endogenous Growth." Journal of Political Economy, January 1990, part 2, 103-25.

-----. "Economic Growth in a Cross Section of Countries." Quarterly Journal of Economics, May 1991, 407-44.

-----. "Human Capital and Economic Growth." Policy for Long-Run Economic Growth, Kansas City: Federal Reserve Bank, 1992, 199-216.

Barth, James R., Robert E. Keleher, and Frank S. Russek. "The Scale of Government and Economic Activity." The Southern Business & Economic Journal, April 1990, 142-83.

Breusch, T. S., and A. R. Pagan. "The Lagrange Multiplier Test and Its Applications to Model Specification in Econometrics." Review of Economic Studies, January 1980, 239-53.

Clark, Todd E. "Cross-Country Evidence on Long-Run Growth and Inflation." Federal Reserve Bank of Kansas City Research Working Paper 93-05, May 1993.

Easterly, William, and Sergio Rebelo. "Fiscal Policy and Economic Growth: An Empirical Investigation." Journal of Monetary Economics, December 1993, 417-58.

Fischer, Stanley. "The Role of Macroeconomic Factors in Growth." Journal of Monetary Economics, December 1993, 485-512.

Greene, William H. Econometric Analysis. New York: Macmillan, 1990.

Grier, Kevin B., and Gordon Tullock. "An Empirical Analysis of Cross-National Economic Growth, 1951-1980." Journal of Monetary Economics, September 1989, 259-76.

Hausman, Jerry A. "Specification Tests in Econometrics." Econometrica, November 1978, 69-85.

Helms, L. Jay. "The Effect of State and Local Taxes on Economic Growth: A Time Series Cross Section Approach." Review of Economics and Statistics, February 1985, 574-82.

Hsiao, Cheng. Analysis of Panel Data. Cambridge, UK: Cambridge University Press, 1986.

Judge, George G., W. E. Griffiths, R. Carter Hill, Helmut Lutkepohl, and Tsoung Chao Lee. The Theory and Practice of Econometrics, 2nd. ed. New York: John Wiley and Sons, 1985.

Koester, Reinhard B., and Roger C. Kormendi. "Taxation, Aggregate Activity and Economic Growth: Cross-Country Evidence on Some Supply-Side Hypotheses." Economic Inquiry, July 1989, 367-86.

Kormendi, Roger, and Philip Meguire. "Macroeconomic Determinants of Growth: Cross-Country Evidence." Journal of Monetary Economics, September 1985, 141-63.

Landau, Daniel. "Government Expenditure and Economic Growth: A Cross-Country Study." Southern Economic Journal, January 1983, 783-92.

-----. "Government Expenditures and Economic Growth in the Developed Countries, 1952-1976." Public Choice, 47(3), 1985, 459-77.

-----. "Government and Economic Growth in the Less Developed Countries: An Empirical Study for 19601980." Economic Development and Cultural Change, 35(1), 1986, 35-75.

Levine, Ross. "Stock Markets, Growth, and Tax Policy." Journal of Finance, September 1991, 1445-65.

Levine, Ross, and David Renelt. "A Sensitivity Analysis of Cross-Country Growth Regressions." American Economic Review, September 1992, 942-63.

Mankiw, N. Gregory, David Romer, and David N. Weil. "A Contribution to the Empirics of Economic Growth." Quarterly Journal of Economics, May 1992, 407-37.

Miller, Stephen M., and Frank S. Russek. "Fiscal Structures and Economic Growth at the State and Local Level." Public Finance Review, March 1997, 213-37.

Mofidi, Alaeddin, and Joe A. Stone. "Do State and Local Taxes Affect Economic Growth?" Review of Economics and Statistics, November 1990, 686-91.

Ram, Rati. "Government Size and Economic Growth: A New Framework and Some Evidence from Cross-Section and Time-Series Data." American Economic Review, March 1986, 191-203.

Romer, Paul M. "Human Capital and Growth: Theory and Evidence." National Bureau of Economic Research Working Paper No. 3173, 1989.

-----. "Capital, Labor, and Productivity." Brookings Papers on Economic Activity, special issue, 1990, 337-420.

Stephen M. Miller: Professor and Head of Economics, University of Connecticut, Storrs, Phone 1-860-486-3853 Fax 1-860-486-4486 E-mail [email protected]

Frank S. Russek: Principal Analyst, Congressional Budget Office Washington, D.C., Phone 1-202-226-2766 Fax 1-202-226-2601, E-mail [email protected]