The government spending and private consumption: a panel cointegration analysis

Transcription

1 International Review of Economics and Finance 10 (2001) 95±108 The government spending and private consumption: a panel cointegration analysis Tsung-wu Ho* Department of Economics, Shih Hsin University, No. 1, Lane 17, Section 1, Mu-Cha Road, Taipei, 11602, Taiwan, ROC Received 8 October 1999; revised 14 April 2000; accepted 20 April 2000 Abstract In this article, whether an increase in government spending will crowd out the private consumption is re-examined. This article augments the empirical literature by extending this issue to panel data. The empirical framework applies the panel cointegration model, dynamic OLS (DOLS), proposed by Kao and Chiang [On the estimation and inference of a cointegrated regression in panel data. Working Paper, Economics Department, Syracuse University, 1999.]. Evidence from 24 OECD countries indicates a significant degree of substitutability between government spending and private consumption when the real disposable income is included, which rejects the permanent income hypothesis. The existence of crowding out renders the Keynesian plea for expansionary fiscal policy unconvincing. D 2001 Elsevier Science Inc. All rights reserved. JEL classification: C22; E21 Keywords: Panel cointegration; Dynamic OLS; Fiscal multiplier; Crowding-out 1. Introduction The multiplier process causes an increase in government spending, or any other exogenous increase in spending, to have a greater ultimate effect on the nominal level of income through price increases, real income increases, or both, depending on where the economy is, relative to full employment. This observation makes expansionary fiscal policy attractive to those * Tel.: ext. 542; fax: address: tsungwu@cc.shu.edu.tw (T. Ho) /01/$ ± see front matter D 2001 Elsevier Science Inc. All rights reserved. PII: S (00)

2 96 T-w. Ho / International Review of Economics and Finance 10 (2001) 95±108 who believe in government intervention to control the economy. However, Bailey (1971) first indicates that there may be a degree of substitutability between government spending and private consumption, or the crowding-out effect. Barro (1981) incorporates it into a general model to examine the direct effect of government purchases of goods and services on consumption utility. Aschauer (1985) and Kormendi (1983) employ the permanent-income approach and find a significant degree of substitutability between private consumption and government spending for the United States. Ahmed (1986) estimates the effects of UK government consumption in an intertemporal substitution model and finds that government expenditures tend to crowd out private consumption. Recently, Aiyagari, Rao, Christiano, and Eichenbaum (1992) and Baxter and King (1993) explore the effect of government spending shocks on various economic aggregates in a one-sector neoclassical growth model with constant returns to scale and variable labor supply. They find that increases in government spending significantly led to a decline in private consumption, showing negative relation between government spending and private consumption. Amano and Wirjanto (1997) apply a relative-price approach to estimate the intratemporal elasticity of substitution between government spending and private consumption of US to be about 0.9. In other words, this group of research indicates that the increase in government spending has fiscal crowding-out effect on private consumption. However, some empirical studies have found different results. In terms of a neoclassical model with increasing returns to scale and monopolistic competition, Devereux, Head, and Lapham (1996) examined the impact of government spending shocks and found that an increase in government consumption generates an endogenous rise in aggregate productivity. The increase in productivity raises the real wage sufficiently, that there is a substitution away from leisure and into consumption. Thus, an increase in government expenditures leads to an increase in private consumption. Karras (1994) examines the change of private consumption in response to increases in government spending across a number of countries and finds that public and private consumption are better described as complementary rather as substitutes. The strength of this complementary relationship is shown to be negatively affected by the government size. These findings are robust across all specifications. In other words, in the aggregate, they are best described as complementary goods in the sense that an increase in government spending tends to raise the marginal utility of private consumption. These findings imply that private consumption cannot be responsible for any crowding-out effects that government spending might have on aggregate demand. On the contrary, private consumption is probably ``crowded-in.'' Given the wide range of empirical results, there appears to be no clear consensus among research works on this issue. Methodologically, the statistical inference of the abovementioned research works relies on univariate time series data of single country over a long time span. However, using a long time series data usually involves possible regime changes and structural breaks, which are both economically and empirically relevant, and can severely affect the properties of inferential procedures. For example, Carruth, Hooker, and Oswald (1998) indicate that the two oil crises in the 1970s severely affected the employment level and the national income. Expensive oil raises producers' costs and squeezes their profit margins. To restore these margins, employers strive to cut labor costs. At the aggregate level, and for any given pressure of demand, higher unemployment and the subsequent decrease in

3 T-w. Ho / International Review of Economics and Finance 10 (2001) 95± aggregate consumption are the result. Carruth et al. also point out that the oil shocks in the late 1970s played a stronger and statistically more significant role in driving American unemployment than interest rates. In other words, once the sampling period spans the 1970s, then the inferential result may be spurious. Alternatively, when using smaller samples, applied econometric analysis appeals to panel data where additional information from cross-sectional units helps identify the parameters of concern. The benefits of using panel data model have been discussed extensively by Baltagi (1995) and Hsiao (1996). The fundamental advantage of a panel data set over a cross-section is that, it will provide the researcher far greater flexibility in modeling differences in behavior across individuals. In light of this, this article extends this line of research to panel data. In addition, since our data involves significant nonstationarity, a panel cointegration method is applied as an alternative to traditional time series and cross-sectional regressions. One of the original motivations in applying tests for stationarity in panel data is due to the lack of power of conventional univariate unit root tests against persistent alternatives, typically for sample sizes that occur in practice. Recognizing data of longer span may lead to more reliable inferences, researchers then employ the amount of available information as much as possible in empirical time series works. This has been proven quite satisfactorily in improving the power of unit root tests. Recent developments in panel cointegration methods have sparked a large body of literature. Quah (1994) pioneers the research by proposing the tests that exploit information from cross-sectional dimensions in inferring nonstationarity from panel data. Moreover, Pedroni (1996) proposes a fully modified estimator for heterogeneous panels; Pedroni (1997) derives asymptotic distributions for residual based tests of cointegration for both homogeneous and heterogeneous panels. Im, Pesaran, and Shin (1999) and Levin, Lin, and Chu (1997) constitute further important contributions along the line. McCoskey and Kao (1998) proposed LM-statistic to test the null hypothesis of cointegration in panel data. Kao (1999) studies the asymptotic theory of cointegration in the panel data. The empirical framework of this article applies the dynamic OLS model (henceforth, DOLS) proposed by Kao and Chiang (1999) to examine cointegration in the panel data. The DOLS was originally developed by Phillips and Loretan (1991) and Saikkonen (1991), which is also known as nonlinear single equation ECM; Kao and Chiang extend it to panel data. Data of 24 OECD countries are derived from the AREMOS/OECD, which is a database compiled by the Ministry of Education (Taiwan, ROC) from IMF International Financial Statistics. The sample period extends from 1981 to 1997, which avoids the possible effects of the two oil crises in the 1970s. Although there was another oil crisis in 1990±1991, this one was a not-so-noticed case. The private consumption includes consumer spending on goods and services. The government consumption consists of government spending on goods, services, and investment. The disposable income is calculated by subtracting total government tax revenue from national income. The real variables are calculated according to 1980 current price and exchange rate, and the unit is expressed in millions of US dollars. All per capita variables are obtained by dividing the aggregate measure by total population. This article is organized as follows. Section 2 develops a theoretical model and analyzes basic time series properties. Section 3 conducts the analysis of panel data. Section 4 concludes.

4 98 T-w. Ho / International Review of Economics and Finance 10 (2001) 95± The model 2.1. The theoretical modeling In this section, the theoretical structure of the model is summarized. The standard Keynesian effective consumption C* is assumed to consist of two components, as specified below (Eq. (1)) C ˆ C t ag t 1 where C t is the real per capita private consumption, G t is the real per capita government consumption, and a is the parameter measuring the relationship between them. Assume a representative individual maximizing the lifetime utility given below (Eqs. (2) and (3)) " # X 1 Max E 0 b t UCt 2 tˆ1 fs:t: A t 1 ˆ A t Y t Ct 1 a G t 1 r g 3 where the utility function is concave, E 0 is the expectations operator based on information of period 0, and b is the subjective discount factor. Eq. (3) is the intertemporal budget constraint, where A t is the real financial assets net real government debt at the beginning of period t, and r is a time invariant real rate of interest. Finally, we assume that U is increasing and concave in its arguments, and *!1. Based on Barro (1981) and Christiano and Eichenbaum (1992), a function of G can be added to the utility function so that the government consumption's marginal utility becomes positive. Hence, the Lagrangean function for the optimization problem is given by Eq. (4): " # X 1 E 0 b t U Ct l tfa t 1 1 r A t Y t Ct 1 a G t Šg 4 tˆ1 where l t is the Lagrange multiplier associated with the budget constraint equation above, which measures the marginal utility of wealth. The first order necessary conditions for period t include the following t =@Ct ˆ l t 5 E 0 b 1 r l t 1 Šˆl t 6 for t = 1,2,..., t /@C t *=@U t (C t *)/@C t *. Substituting Eq. (5) for l t and l t +1 into Eq. (6), the Euler equation between periods t and t + 1 can be derived below: E 0 b 1 t 1 =@U t Š ˆ 1: 7 To investigate the empirical implications of the model, we assume that the change in marginal utility is negligibly small over time, so that Eq. (7) can be written as E 0 C t +1 *=[b(1 + r)] s C t *, where s = U 0 C*/{C * U(C*)} is the intertemporal elasticity of substitution. Hence, the econometric relationship below is derived: C t 1 ˆ gc t n t 8

5 T-w. Ho / International Review of Economics and Finance 10 (2001) 95± where n t i.i.d. Eq. (8) can be rewritten as: C t ag t ˆ g C t 1 ag t 1 n t : 9 Rearranging them (Eqs. (8) and (9)), we obtain: C t gc t 1 ˆ a G t gg t 1 n t n t i:i:d: 10 If both C t and G t are I(1), then they are cointegrated in the sense of Engle and Granger (1987) with cointegrating vector A. Hence, Eq. (10) implies an error correction mechanism that can be consistently estimated by the procedures suggested by Park (1992), Phillips (1991), Phillips and Hansen (1990), Phillips and Loretan (1991), and Wickens and Breusch (1988). In other words, it also implies that g = 1 and C t * is integrated of order 1, or I(1). Moreover, Graham (1993) has shown that the robustness of the relationship between government spending and private consumption will be weakened when real disposable income is excluded from the model. Thus, letting Y d denote the real disposable income, we estimate the two models [Eqs. (11A) and (11B)] below: Model 1 : C t ˆ a 0 a 1 G t n t 11A Model 2 : C t ˆ a 0 a 1 G t byt d n t : 11B Eq. (11B) can be easily derived by simply adding Y d to Eq. (1) The time series properties of individual country We first test whether there is a unit root in each series. To this end, we conduct ADF (Dickey & Fuller, 1981; Said & Dickey, 1984) and the KPSS-statistic proposed by Kwiatkowski, Phillips, Schmidt, and Shin (1992). The ADF tests the null hypothesis that there is a unit root and the KPSS-statistic tests the opposite. Table 1 presents the estimation result of individual country. Reading across the rows of the table is individual country by country results for each of the 24 countries, and the presence of a unit root is confirmed. Banerjee, Dolado, Hendry, and Smith (1986), Phillips (1987, 1991), and Phillips and Ouliaris (1990) have shown that conventional tests in multivariate regressions with integrated processes cannot be applied asymptotically. In this case, classical asymmetric theory breaks down and the presence of nuisance parameter dependencies in the limiting distribution theory raises similar issues to panel data with I(1) processes. To expose this problem, assume that the generating mechanism for Y t is a cointegrating system (Eqs. (12) and (13)): Y t ˆ A BX t u 1t ; t ˆ 1; 2;...; T 12 DX t ˆ u 2t : 13 Phillips and Durlauf (1986) showed that under appropriate centering and scaling, OLS estimation of the cointegrating vector in Eq. (12) are asymptotically non-normal. The weak convergence of appropriately scaled sample moments with random matrices rather than constant matrices results in this non-normality. Moreover, OLS leads to estimators that are asymptotically biased, and whose distributions involve unit root asymptotics and nontrivial nuisance parameters (see Phillips and Loretan,

6 100 T-w. Ho / International Review of Economics and Finance 10 (2001) 95±108 Table 1 Unit root tests of individual country Y d C t G t KPSS KPSS KPSS Country ADF Level Trend ADF Level Trend ADF Level Trend Argentina Austria Belgium Canada Denmark Finland France Germany Greece Iceland Ireland Italy Japan Luxembourg Netherlands New Zealand Norway Portugal Spain Sweden Switzerland Turkey UK USA OECD Europe ADF is calculated by one lag, including trend and intercept. The MacKinnon (1991) critical values for ADF at 1%, 5%, and 10% significance level are 4.044, 3.451, and For KPSS, the approximate asymptotic critical values at 10%, 5%, and 1% significance level for the level model are, respectively, 0.347, 0.463, and 0.739, and for the trend model, they are 0.119, 0.146, and 0.216, respectively. 1991, p. 426). Phillips (1991) and Phillips and Ouliaris (1990) have proven that standard tests statistics, such as the Wald test, no longer generate asymptotically distributed c 2 criteria. Because the limiting distribution of the regression coefficients is non-normal, the metric underlying the Wald test is no longer valid. In other words, statistical estimation and inference in these models require a methodology that accounts for the nonstationarity of the underlying time series. Roughly at the same time, Phillips and Loretan (1991) and Saikkonen (1991) propose an estimator that can produce asymptotically efficient estimates of the long-run vector B in the absence of contemporaneous correlation of u 1t and u 2t. It is specified below: Y t ˆ BX t d 1 L Y t BX t d 2 L X t e t 14 where d k L ˆP1 d kj L j, and k = 1,2. Estimation of B in Eq. (14) is achieved through a simple jˆ1

7 T-w. Ho / International Review of Economics and Finance 10 (2001) 95± nonlinear LS regression. For simplicity, we call it NLS estimator. The limiting distribution of the NLS estimate of B is free of nuisance parameters, and asymptotically normal t ratios and asymptotically c 2 criteria constructed in the usual fashion can be used for inferential purposes. Unfortunately, in the NLS modeling, valid conditioning on the regressors always fails. With feedback from u i1 to u i2 (the presence of simultaneous equation bias), the limiting distribution of the estimator of B is biased, asymmetric, and without scaled nuisance parameters. To deal with the long-run endogeneity of X t, Phillips and Loretan (1991) included leads and lags of X t in the regression so that e t is asymptotically orthogonal to the entire history of (DX t ) 1 1. The modified specification has the form below Y t ˆ BX t d 1 L Y t BX t d 2 L DX t d 3 L 1 DX t n t 15 where d 3 (L 1 ) is similarly defined as previous equation. The nonlinear least square estimate of B from Eq. (15) is fully efficient in the limit, asymptotically median unbiased, and asymptotically equivalent to the full system MLE and spectral regression estimator. Conventional c 2 criteria for hypothesis testing, with respect to all coefficients in Eq. (15), can be applied. Parameter estimates in Eq. (15) are obtained by nonlinear least square that is equivalent to maximum likelihood estimation when the errors are normally distributed. The issue of applicability of these asymptotic results to small samples has also been addressed by Phillips and Loretan (1991). Simulations with a small-scale cointegrated system showed that the small-sample properties are consistent with its asymptotic distribution theory. Phillips and Loretan (1991) concluded that the limit distribution theory, with respect to their proposed estimator, provides good approximations in sample sizes that are typical in economic time series data. Therefore, the model we are going to estimate is C t ˆ a 0 a 1 G t by d t d 3 L 1 DG t d 4 L DY d t d 1 C t a 0 a 1 G t byt d d 2 L DG t d 5 L 1 DY d t e t : 16 Because the nonlinear error correction is limited to 1, Eq. (16) is expressed by NLECM( p,q), where p and q denote the number of leads and lags, respectively. The economic mechanism underlying the specification in Eq. (16) is one where countries react to departures from long-run equilibrium through the adjustment coefficient d 1. The disequilibrium correction takes place only if d 1 is negative. Table 2 presents the results of individual country estimated by NLECM(1,1). To save space, we only report four major parameter estimates of Model 2. Examining a 1 and b, not surprisingly, on a country-to-country basis, the data can hardly obtain consistent results of cointegration and substitutability. This is largely consistent with previous studies. It motivates the subsequent study in panel data. 3. The panel DOLS model Pooling time series has traditionally involved a substantial degree of sacrifice in terms of the permissible heterogeneity of the individual time series. In order to ensure broad applicability of any panel cointegration test, it will be important to allow for as much heterogeneity as possible among the individual members of the panel. Table 3 offers a list of

8 102 T-w. Ho / International Review of Economics and Finance 10 (2001) 95±108 Table 2 NLECM(1,1) estimates for individual country Country a 0 a 1 b d AIC ADF Argentina ( 1.48) 0.01 (0.65) (6.9) (0.26) Austria 0.45 (0.56) ( 1.99) 0.9 (14.36) (1.76) Belgium 1.36 ( 1.86) (1.87) 1.05 (18.09) ( 1.38) Canada 2.28 ( 3.45) (0.57) 1.11 (23.66) ( 0.32) Denmark 0.98 ( 1.99) ( 0.35) (27.4) (0.72) Finland ( 5.26) 0.05 ( 4.26) 1.17 (24.9) (3.86) France 0.39 ( 0.98) (0.39) 0.97 (32.87) ( 1.67) Germany 4.27 ( 24.85) (1.143) 1.24 (102.1) 0.74 (4.41) Greece 5.09 ( 7.83) (0.318) (20.7) ( 0.74) Iceland 1.84 (2.12) (0.05) (6.36) ( 0.64) Ireland 2.31 (1.004) ( 0.71) 0.72 (2.92) (0.606) Italy 7.82 ( 1.45) ( 0.98) (3.658) (0.765) Japan (0.356) (3.543) 0.93 (55.0) ( 3.12) Luxembourg (13.26) 0.06 (3.86) (15.67) 0.02 ( 2.95) Netherlands 0.09 ( 0.77) ( 0.57) (85.36) (1.14) New Zealand 4.23 ( 19.2) ( 3.74) 1.35 (55.56) 0.01 (4.88) Norway (42.82) ( 3.3) (107.8) (9.68) Portugal ( 2.87) (1.34) 1.05 (20.8) ( 1.26) Spain 1.34 ( 1.56) 0.01 ( 0.659) 1.05 (13.5) (1.11) Sweden 5.86 ( 1.41) ( 1.76) 1.44 (3.99) 0.02 (2.408) Switzerland 2.34 (5.454) (1.39) 0.74 (19.93) ( 0.14) Turkey 2.72 (1.039) (0.35) (5.16) ( 0.8) UK 5.88 ( 2.23) (0.85) (8.49) 0.02 ( 0.78) USA 1.34 ( 2.18) ( 0.31) (32.3) (0.2) the 24 economies and country annual averages over 1981±1997 for the three variables. It is apparent from Table 3 that these variables vary substantially across economies. If the theoretical predictions are verified empirically, this large difference suggests that the cointegrating relation is expected to differ considerably across countries. To further study the problem in panel data, we test for the presence of a panel unit root. We apply the ADF-based test (t NT, thereafter) proposed by Im et al. (1999) to test the null hypothesis whether there is a panel unit root. Table 4 reports the outcomes of panel unit root test. To avoid the small-sample bias, we calculated the critical values of the panel-based test by using Monte Carlo simulations calibrated for our sample size. These critical values are reported in Table 5. The presence of a panel unit root in three variables is confirmed. Kao and Chiang (1999) show that the DOLS estimator can produce asymptotically efficient estimates of the long-run multipliers vector B in the absence of contemporaneous correlation of u 1,it and u 2,it. They have also shown that the DOLS outperforms both bias-corrected OLS and fully modified OLS. Technically, the panel DOLS is slightly different from the NLS mentioned above; it is specified according to the equation below (Eq. (17)): C it ˆ a i0 a 1 G it by d it Xq jˆ q c 1 ij DG i;t j Xq jˆ q c 2 ij DY d i;t j e it: 17

9 T-w. Ho / International Review of Economics and Finance 10 (2001) 95± Table 3 Sample means: 1981 ±1997, millions of US dollars Countries Y d C t G t Australia Austria Belgium Canada Denmark Finland France Germany Greece Iceland Ireland Italy Japan Luxembourg Norway Netherlands New Zealand Portugal Spain Sweden Switzerland Turkey UK USA Source: AREMOS/OECD Data Set, Ministry of Education, Taiwan, ROC, To test for panel cointegration, we apply McCoskey and Kao's (1998) panel LM-statistic, which is defined below (Eq. (18)): LM ˆ P N P T Sit 2 iˆ1 tˆ1 NT 2 v 2 ; S it ˆ Xt e ij jˆ1 where v 2 is a consistent estimator of s e 2 under the null hypothesis. To avoid the small-sample bias, the critical values of the LM are also calculated by using Monte Carlo simulations calibrated for our sample size; Table 5 reports the simulation results. In addition to LM, we 18 Table 4 Tests for panel unit root Test statistic Y t d t NT for nonstationarity Critical values are listed in Table 5 below. C t G t

10 104 T-w. Ho / International Review of Economics and Finance 10 (2001) 95±108 Table 5 Critical values for the panel unit root tests Significance level (%) t NT LM ,000 replications with N = 25 and T = 17. also employ t NT -statistic to test the presence of a panel unit root in the residuals of cointegrating equation. Table 6 presents the results of the two models and the two tests for panel cointegration estimated by DOLS(1,1). Because the analysis of panel data always removes individual effect by subtracting individual means, the result does not include the intercept term. We first examine Model 1, which does not include the real per capita disposable income. In this, a 1 is 2.035, and the two numbers in the parenthesis are, respectively, the standard deviation and the corresponding t-statistic. The a 1 is significantly positive, which implies that a US$1 increase in government spending will cause a US$2.035 increase in private consumption. That is, the temporary conclusion so far indicates that the government spending multiplier increases private consumption. Although the t NT -statistic can reject the null hypothesis that Eq. (16) is not a cointegrated panel at 5% significance level, the LM-statistic substantially rejects the null hypothesis that Eq. (16) is a cointegrated panel. Examining the goodness-of-fit statistics, the adjusted R 2 is.49 and AIC is 2.3; therefore, Model 1 may not be an appropriately specified model. The underlying reason that causes this result may be model instability indicated by Graham (1993). Unfortunately, conventional stability tests for univariate series cannot directly extend to panel data. We are unable to prove this point here unless its asymptotic properties are proven. Second, let us examine Model 2. In this case, a 1 is and b = Both estimates are significant. In sharp contrast to the previous model, a 1 here implies a crowding-out effect, rather than crowding-in effect. It implies that a US$1 increase in government spending will Table 6 DOLS(1,1) estimates for panel cointegration Parameters Model 1 Model 2 a (0.156, 13.05) (0.2525, 2.134) b 0.77 (0.03, 24.87) Adjusted R AIC t NT LM Critical values are listed in Table 5. The long-run covariance matrix and true covariance of DOLS estimates are available from the author. The numbers in the parenthesis are, respectively, standard errors and corresponding t-statistic.

11 T-w. Ho / International Review of Economics and Finance 10 (2001) 95± cause a US$ decrease in private consumption. b = 0.77 implies that a US$1 increase in disposable income will cause a US$0.77 increase in private consumption. The adjusted R 2 is.96 and AIC is 5.689; both goodness-of-fit statistics and significant b not only show the importance of real disposable income, but also the fact that Model 2 is better specified than Model 1. Most importantly, the t NT -statistic is 4.47, which significantly rejects the null hypothesis that Eq. (16) is not a cointegrated panel; and the LM-statistic is 0.652, which significantly accepts the null hypothesis that Eq. (16) is a cointegrated panel. They consistently show that the government spending and private consumption are cointegrated in this panel. Therefore, Model 2 is better than Model 1. It concludes that there is a significant degree of substitutability between government spending and private consumption in the presence of real disposable income, which also rejects the permanent income hypothesis. 4. Conclusion The crowding-out phenomenon describes the process whereby an increase in government spending decreases other components of aggregate demand, thus reducing the government spending multiplier effect on stimulating aggregate demand. In this article, we employ the Kao and Chiang's (1999) panel DOLS to examine this issue in nonstationary panel data. The panel cointegration test proposed by McCoskey and Kao (1998) is used to test the cointegration hypothesis. In order to have a comparative perspective, we estimate two specifications of the model. We find several interesting results. First, the model with real disposable income exhibits desirable result. It implies the important role of real disposable income in the model specification. When econometricians provide reliable test, we can formally show this point. Current evidences are sufficient to drop the model that excludes the real per capita disposable income. In a nutshell, the permanent income hypothesis is rejected. Second, in sharp contrast to individual country results as shown in Table 2, we find substantial evidence to accept the hypothesis of crowding out. The reason for these findings can be interpreted as: The power of the panel unit-root test is higher than that of conventional tests. The existence of crowding-out effect would make the government spending multiplier smaller than it is anticipated. Besides the level of employment, the crowding-out effect is related to the means used to finance an increase in government spending. If taxes are used to finance an increase in government spending, then this multiplier is called the balanced-budget multiplier, reflecting the fact that the fiscal action has no impact on the size of the government's budget deficit or surplus. In this case, consumers reduce consumption spending to be able to pay the higher taxes. The decrease in consumption demand partially offsets the increase in government spending, reducing the size of the multiplier. The offset is only partial because not all the financing for extra taxes comes from reducing consumption. Some come from reducing saving, which is not a component of aggregate demand. Moreover, the multiplier process usually assumes that government sells bond to finance an increase in its spending, in this case, extra crowding out comes about in two ways.

12 106 T-w. Ho / International Review of Economics and Finance 10 (2001) 95±108 First, it raises the rate of interest. To sell bonds, the government must make them attractive, so it must raise the interest rate. The higher interest rate crowds out all components in aggregate demand. Second, when the bonds mature, interest and principal must be paid to the bondholders. According to the Ricardian equivalence, people would expect that future taxes will be higher because of this, and react by increasing saving to build up a reserve, so that those anticipated higher taxes can be paid without disrupting future consumption levels. During the multiplier process, several factors crowd out aggregate demand. Financing an increase in government spending by increasing taxes or by selling bonds to the public causes crowding-out forces to reduce the magnitude of the multiplier. Although financing by printing money may avert crowding-out effect temporarily, it will cause inflation and raise the interest rate in the future. However, whether the increase in national income is due to the increase in the money supply or the increase in government spending, is still a debatable issue. This article implies that the existence of crowding out renders the Keynesian plea for expansionary fiscal policy unconvincing. Acknowledgments I am deeply indebted to the editor and two anonymous referees for helpful comments and suggestions. Helpful discussions with Dr. Biing-shen Kuo and Dr. River Huang are most appreciated. I am also grateful to Chiang and Kao (2000) who offer the NPT 1.0 as public program. All errors are mine. Appendix A This appendix summarizes the simulation procedure. Interested readers are referred to Im et al. (1999) for detail. In the work of Im et al., the series is specified according to the following time-series representation (Eq. (A1)): z it ˆ d 0 d it z it 1 z it ; i ˆ 1; 2;...; N; t ˆ 1; 2;...; T A1 where x it is assumed to be composed of two random components (Eq. (A2)), x it ˆ q t h it A2 q t is a stationary time-specific common effect that allows for a degree of dependency across groups. h it is independently distributed across groups. To correct the possible serial correlation in h it, we estimate the following model (Eq. (A3)): Dz it ˆ b i b i z it 1 b 2 t Xp jˆ1 d ij Dz it k x it ; A3 where p is selected to make x it uncorrelated over time. The null hypothesis for the presence of unit root is b i =0.

13 T-w. Ho / International Review of Economics and Finance 10 (2001) 95± References Ahmed, S. (1986). Temporary and permanent government spending in an open economy: some evidence for the United Kingdom. Journal of Monetary Economics, 17, 197 ±224. Aiyagari, S., Rao, L., Christiano, J., & Eichenbaum, M. (1992). The output, employment, and interest rates effects of government consumption. Journal of Monetary Economics, 30, 73 ± 86. Amano, R. A., & Wirjanto, T. S. (1997). Intratemporal substitution and government spending. Review of Economics and Statistics, LXXIX, 605 ± 609. Aschauer, D. A. (1985). Fiscal policy and aggregate demand. American Economic Review, 75, 117± 127. Bailey, M. J. (1971). National income and the price level. New York: McGraw-Hill. Baltagi, B. (1995). Econometric analysis of panel data. New York: Wiley. Banerjee, A., Dolado, J. J., Hendry, D. F., & Smith, G. W. (1986). Exploring equilibrium relationships in econometrics through statistical models: some Monte Carlo evidence. Oxford Bulletin of Economics and Statistics, 48, 253 ±277. Barro, R. J. (1981). Output effects of government purchases. Journal of Political Economy, 84, 343 ±350. Baxter, M., & King, R. G. (1993). Fiscal policy in general equilibrium. American Economic Review, 83, 315±334. Carruth, A., Hooker, M., & Oswald, A. J. (1998). Unemployment equilibrium and input prices: theory and evidence from the United States. Review of Economics and Statistics, 4, 621± 628. Chiang, M.-H., & Kao, C. (2000). Nonstationary panel time series using NPT 1.0 Ð a user guide. Center for Policy Research, Syracuse University. Christiano, L., & Eichenbaum, M. (1992). Current real-business-cycle theories and aggregate labor-market. American Economic Review, 82, 430± 450. Devereux, M. B., Head, A. C., & Lapham, B. J. (1996). Monopolistic competition, increasing returns, and the effects of government spending. Journal of Money, Credit, and Banking, 28, 233 ±254. Dickey, D. A., & Fuller, W. A. (1981). Likelihood ratio statistics for autoregressive time series with a unit root. Econometrica, 49, 1057 ± Engle, R. F., & Granger, C. W. J. (1987). Co-integration and error correction: representation, estimation and testing. Econometrica, 55, 251 ± 276. Graham, F. C. (1993). Fiscal policy and aggregate demand: comment. American Economic Review, 83, 659 ±666. Hall, R. E. (1978). Stochastic implications of the life-cycle permanent income hypothesis: theory and evidence. Journal of Political Economy, 86, 971± 987. Hsiao, C. (1996). Analysis of panel data. Cambridge: Cambridge Univ. Press. Im, K. S., Pesaran, M. H., & Shin, Y. (1999). Testing for unit roots in heterogeneous panels. Working paper, University of Cambridge. Kao, C. (1999). Spurious regression and residual-based tests for cointegration in panel data. Journal of Econometrics, 90, 1±44. Kao, C., & Chiang, M. H. (1999). On the estimation and inference of a cointegrated regression in panel data. Working Paper, Economics Department, Syracuse University. Karras, G. (1994). Government spending and private consumption: some international evidence. Journal of Money, Credit, and Banking, 26, 9± 22. Kormendi, R. C. (1983). Government debt, government spending, and private sector behavior. American Economic Review, 73, 994± Kwiatkowski, D., Phillips, P. C. B., Schmidt, P., & Shin, Y. (1992). Testing the null of stationarity against the alternative of a unit root: how sure are we the economic time series have a unit root? Journal of Econometrics, 54, 159±178. Levin, A., Lin, C. F., & Chu, C. S. J. (1997). Unit-root test in panel data: asymptotic and finite sample properties. Discussion paper. University of California, San Diego. MacKinnon, J. C. (1991). Critical values for cointegration tests. In: R. F. Engle, & C. W. J. Granger (Eds.), Longrun economic relationships: readings in cointegration. Cambridge: Oxford Univ. Press (Chapter 13). McCoskey, S., & Kao, C. (1998). A residual-based test of the null of cointegration in panel data. Econometric Reviews, 17, 57 ±84.

14 108 T-w. Ho / International Review of Economics and Finance 10 (2001) 95±108 Park, Joon Y. (1992). Canonical cointegrating regression. Econometrica, 60, 119 ±143. Pedroni, P. (1996). Fully modified OLS for heterogeneous cointegrated panels and the case of purchasing power parity. Working paper, Department of Economics, Indiana University. Pedroni, P. (1997). Panel cointegration: asymptotics and finite sample properties of pooled time series tests with application to the PPP hypothesis. Working paper, Department of Economics, Indiana University. Phillips, P. C. B. (1987). Time series regression with a unit root. Econometrica, 55, 227 ±301. Phillips, P. C. B. (1991). Optimal inference in cointegrated systems. Econometrica, 59, 283± 306. Phillips, P. C. B., & Durlauf, S. N. (1986). Multiple time series regression with integrated processes. Review of Economic Studies, 53, 473 ±495. Phillips, P. C. B., & Hansen, B. E. (1990). Statistical inference in instrumental variables regression with I(1) processes. Review of Economic Studies, 57, 99 ± 125. Phillips, P. C. B., & Loretan, M. (1991). Estimating long-run economic equilibria. Review of Economic Studies, 58, 407±436. Said, S. E., & Dickey, D. A. (1984). Testing for unit roots in autoregressive-moving average models of unknown order. Biometrica, 71, 599 ± 607. Saikkonen, P. (1991). Asymptotically efficient estimation of cointegration regression. Econometric Theory, 7, 1 ±21. Wickens, M. R., & Breusch, T. S. (1988). Dynamic specification, the long run and the estimation of the transformed regression models. Economic Journal, 98, 189 ± 205.