Tax and spend, or spend and tax? An inquiry into the Turkish budgetary process.

Darrat, Ali F.

1. Introduction

Contemporary economies, perhaps without exception, have been plagued with huge and escalating government budget deficits. These deficits are expected to have adverse economic consequences including high real interest rates, slow capital formation, and high unemployment rates. Moreover, to the extent that the deficit is financed through the issuance of government bonds, the recurrent large deficits have further worsened the public debt problem, which threatens the well-being of numerous countries, both developed and developing.

Therefore, researchers and policymakers have expended enormous efforts attempting to analyze the deficit problem and to suggest ways to resolve it. Some, for example, advocate cuts in government spending rather than tax increases as the optimal solution to the deficit dilemma. They reason that governments often spend all that they receive in taxes and perhaps much more. Under this line of reasoning, raising taxes would simply induce more spending, leaving the deficit unchanged (or even higher). Others, however, deny this implied tax-and-spend nexus, and argue that it is taxes that adjust gradually to spending. Under this latter scenario, tax increases will not lead to higher spending, and thus, could be used as an effective deficit-cutting measure along with spending cuts. Still, other researchers posit that changes in spending and taxes could occur simultaneously. Therefore, focusing on one component of the government budget while ignoring the interdependence with the other component would have an ambiguous overall effect on the deficit.

Clearly, then, the optimal approach to solving the problem of government budget deficits depends to a large extent on the intertemporal relationship between government spending and taxation. Indeed, there have been numerous studies that empirically gauge this relationship (e.g., Anderson, Wallace, and Warner 1986; Manage and Marlow 1986; von Furstenberg, Green, and Jeong 1986; Ram 1988; Ahiakpor and Amirkhalkhali 1989; Joulfain and Mookerjee 1990; Miller and Russek 1990; Hoover and Sheffrin 1992; Lee and Vedder 1992; Owoye 1995).

Interestingly, the bulk of this voluminous empirical literature has focused almost exclusively on the U.S. experience, with few studies examining the case of other large industrialized major OECD countries. By contrast, there has been, thus far, no attempt to study the interrelationship between government spending and taxation for developing countries; yet, the problem may be more acute in these countries with huge and escalating deficits in recent years. Investigating the government spending/taxation nexus in these countries can provide useful information pertaining to the optimal solution to their deficits. Note also that developing countries have often been required to curtail budget deficits as a precondition to receiving financial aid and/or loans from international organizations like the International Monetary fund (IMF) or the World Bank.

The main purpose of this paper is to fill this gap in the literature and examine the intertemporal relationship between government spending and taxation in the case of Turkey. Since the 1960s, Turkey has had a large and growing government sector. Measured by the share of government expenditures in Gross National Product (GNP), the government sector in Turkey expanded from less than 18% in the 1960s to claim more than 25% of the economy in the 1990s. Taxes have also risen during the last three decades but at a slower pace, resulting in persistent budget deficits. Many analysts and popular commentators (Bugra 1994; Adaman and Sertel 1995; Pope 1996) have warned that such deficits, if they continue, can seriously stifle economic activities and worsen the already bleak international credit-worthiness of the country. So far in the 1990s, inflation in Turkey has been running at about 75% annually, interest rate on three-month government securities is more than 80% and the public debt rose nearly 25% in the first quarter of 1996. Such bad economic conditions have already threatened a new IMF aid package for Turkey and also prompted the Standard & Poor's Ratings Group to rank Turkey's long-term government securities among the lowest of all countries. Although many elements, including perhaps political instability, may be behind Turkey's economic and financial woes, the large budget deficits have undoubtedly been a major contributing factor. Indeed, there is a near consensus among economists and public-policy observers in the region that it is indispensable for Turkey to aggressively resolve the budget deficit dilemma and put the public sector in order.

Of course, the contrasting viewpoints with respect to government spending and taxation in Turkey presume the endogeneity of the Turkish budgetary process.(1) Budget decisions in an endogenous framework are determined in a large political body whose agents often exhibit conflicting objectives. Despite several military interventions in Turkey since the 1960s, the multiparty political system has remained largely a democratic system with fundamental rights and freedoms guaranteed to all citizens. The 1961 Constitution created a bicameral legislature comprised of a National Assembly and a Senate with an elaborate system of checks and balances on government authority. Among many constitutional responsibilities, the legislators have exclusive powers to make law, and they freely debate and amend the government's proposed budget. The president, who is appointed by the parliament, can veto legislation passed by the parliament. Since 1983, the country has been ruled by a new Constitution, which was approved by a national referendum in 1982. This Constitution established a popularly elected, single-chambered parliament and continued to preserve the civil and political rights of all citizens, but made these rights subordinate to "national security," "national unity," and "public morality."(2) Therefore, it appears that the Turkish budgetary process is sufficiently endogenous without substantial regime shifts, a setting that seems appropriate for the tests I perform in this study.

The remainder of the paper is organized as follows. Section 2 provides a brief account of the literature on the government spending/taxation nexus with emphasis on four alternative hypotheses pertaining to this relationship. Section 3 describes the data and the empirical methodology of the paper. The methodology is based on the recent cointegration and error correction modeling in bivariate and multivariate contexts. Section 4 reports the empirical results. Section 5 concludes the paper.

2. Alternative Hypotheses

Researchers have advanced four alternative hypotheses regarding the government spending/taxation nexus. First, some researchers (e.g., Wildavsky 1988) have argued that government's decisions to spend are independent from its decisions to tax. Owing to the institutional separation between spending allocation and taxation in the United States, Hoover and Sheffrin (1992) report empirical results for the U.S. that are consistent with an independent determination of the two sides of the budget, especially since the 1960s.

Most of the literature, however, suggests that spending and taxes are interrelated, giving rise to three additional possibilities. Several economists, led by Milton Friedman (1982), contend that raising taxes would likely fall to lower budget deficits because they would instead invite more spending. Because of this positive causal impact of taxes on spending, Friedman has long proposed tax cuts as a means to reducing budget deficits. He reasons that the larger budget deficits resulting from tax reductions should exert mounting pressures on the government to curtail its spending. Interestingly, like Friedman, Buchanan and Wagner (1977, 1978) have also suggested that causality runs from taxes to spending. However, unlike him, Buchanan and Wagner hypothesize a negative causal relationship. They argue that reducing taxes would lower the cost of government programs as perceived by the public. Hence, voters tend to accept or even demand further government programs, resulting in higher government spending. The tax cut, in conjunction with the resultant government spending increase, would lead to higher (not lower) budget deficits. Thus, Buchanan and Wagner advocate, instead, tax increases which would raise the cost of government spending as perceived by voters, thus resulting in lower spending. The tax increase, combined with spending cuts, could drastically curtail budget deficits.

A third group of economists, most notably Barro (1979) and Peacock and Wiseman (1979), has challenged the above views and argues that governments spend first and then tax later. They contend that temporary increases in government spending (perhaps easily justified by natural crises and/or severe humanitarian needs) tend to become enduring and lead to permanent tax increases to finance the excessive spending. Under this causal pattern from spending to taxation, the optimal solution to controlling the budget deficit is clearly spending cuts. Proposals like these will become particularly attractive in the absence of the severe crises that initially justified the spending hikes.

The fourth and final causative link between government spending and taxation suggests a mutual change. This fiscal synchronization hypothesis of Musgrave (1966) and Meltzer and Richard (1981) contends that the public simultaneously determine the levels of government spending and taxation by contrasting the benefits of government services with their costs. Therefore, these economists maintain that spending and taxes are determined concurrently.

In summary, there are four alternative hypotheses pertaining to the causal relationship between government spending and taxation. These are (i) taxes and spending are causally independent, (ii) taxes cause spending, (iii) spending causes taxes, and (iv) taxes and spending are mutually causal. Theoretical debate also exists within hypothesis (ii); namely, that higher taxes cause higher spending (Friedman's view), or higher taxes cause lower spending (Buchanan-Wagner view).

As I stated, numerous studies have examined the empirical validity of the above hypotheses, primarily for the United States, with remarkably mixed results. For example, Blackley (1986), Manage and Marlow (1986), Ahiakpor and Amirkhalkhali (1989), and to some extent Ram (1988) have reported results showing that taxes cause spending.(3) On the other hand, results consistent with the opposite view that spending causes taxes have emerged from many other studies including Anderson, Wallace, and Warner (1986) and von Furstenberg, Green, and Jeong (1986). Still, researchers like Joulfain and Mookerjee (1990), Miller and Russek (1990), and Owoye (1995) have found a bidirectional causality between government spending and taxation.

It is important to observe that these apparently contradictory studies have used a variety of empirical procedures that are capable of yielding different and conflicting results. Equally important is the fact that the majority of these empirical procedures are deficient and are known to yield biased inferences. For instance, Manage and Marlow (1986) and Joulfain and Mookerjee (1990) employ bivariate Granger causality models; yet, it is widely recognized that such models are suspect due to the omission-of-variables bias discussed in Lutkepohl (1982, 1993), among others. If a variable is found not to cause another variable in a bivariate setting, this does not necessarily imply that such an inference holds in the context of a larger economic system which includes other germane variables. Lutkepohl (1982, p. 367) writes, "this conclusion is a consequence of the well-known problem that a low dimensional subprocess contains little information about the structure of a higher dimensional system." In order to avoid these potential biases, Anderson, Wallace, and Warner (1986), von Furstenberg, Green, and Jeong (1986), and Ahiakpor and Amirkhalkhali (1989) have incorporated other relevant variables and examined multivariate Granger causality models.

However, another source of inadequacy in such standard Granger causality tests is their neglect of other sources of causality stemming from the underlying equilibrium (long-run) relationships among the variables. Such cointegration relationships are taken into account by Miller and Russek (1990) and Owoye (1995). Unfortunately, their error correction causal models are bivariate in nature, ignoring other relevant variables that may influence government spending and/or taxes. Recent literature (Miller 1991; Darrat 1994; Darrat and Arize 1996) has shown that the problem of omission-of-variables bias is not unique with standard Granger causality tests, but it also distorts inferences from cointegration and error correction models.

In summary, a better way to discriminate among the alternative causal hypotheses linking government spending and taxation is to use multivariate error correction models. I perform this task below for Turkey.

3. Methodology and Data Used

I test for Granger causality between government spending (G) and taxes (T) in the context of the Turkish economy. For examining the direction of causality between any two variables, the Granger test has gained a lot of popularity, partly due to its simplicity. This testing procedure further saves degrees of freedom which, in relatively small samples, is an important advantage. Briefly, a stationary time series {[Z.sub.t]} is said to Granger-cause(4) another stationary time series {[V.sub.t]} if the prediction error from regressing V on Z significantly declines by using past values of Z in addition to using past values of V.

Granger-causality tests require stationary variables (e.g., whose stochastic properties are time invariant). Granger (1986) demonstrates that a nonstationary time series (Y) can achieve stationarity if differenced appropriately. To determine the proper order of differencing for any variable, I use three alternative testing procedures. They are the augmented Dickey-Fuller (ADF), Phillips-Perron (PP), and weighted symmetric (WS) tests. Although the ADF procedure is perhaps the most common test, it requires that the errors in the testing equation be homoscedastic and serially uncorrelated. The PP test generalizes the ADF procedure allowing for less restrictive assumptions regarding the error terms. Finally, Pantula, Gonzales-Farias, and Fuller (1994) report extensive Monte Carlo evidence supporting the empirical power of the WS test against several alternative unit root testing procedures.

To avoid the omission-of-variables bias, I incorporate two additional variables that theory suggests as potentially relevant for the determination of G and/or T; namely, real GNP and interest rates.(5) Both fiscal variables are potentially sensitive to changes in the level of economic activity as measured by real GNP. Such automatic stabilizing effects of business cycles are usually removed by using cyclically adjusted or full-employment data on government spending and taxes. Since these data are usually unavailable (except for the U.S.), it is important to include real GNP in the testing equations for Turkey to prevent contaminating the data with nonpolicy (business cycles) effects.(6) On the other hand, interest rates are often considered a key controlling variable in any macroeconomic model (see Sims 1980; Fackler 1985). Moreover, government spending is particularly sensitive to changes in interest rates, because a significant portion of government outlays is related to interest payments on public debt.(7) In light of the above considerations, it can be argued that important information would be lost by excluding real GNP and interest rates from the analysis. To provide some empirical insight into this issue, I report, in the next section, results from bivariate models followed by the results from multivariate extensions.

In a multivariate context, Granger causality running from, say, taxes (T) to spending (G) can be tested by estimating:

[G.sub.t] = [[Psi].sub.0] + [summation of] [[Psi].sub.1i][G.sub.t-i] where i = 1 to [n.sub.1] + [summation of] [[Psi].sub.2i][T.sub.t-i] where i = 1 to [n.sub.2] + [summation of] [[Psi].sub.3i][X.sub.t-i] where i = 1 to [n.sub.3] + [summation of] [[Psi].sub.4i][R.sub.t-i] where i = 1 to [n.sub.4] + [[Mu].sub.t] (1)

where X is real GNP, R is the interest rate on three-month T-bills, [Mu] is a white-noise error term, and the summations of [Psi]'s are polynomials of appropriate orders ([n.sub.1], [n.sub.2], [n.sub.3], [n.sub.4]) for the four explanatory variables. I use the Akaike's final prediction error (FPE) criterion to determine the proper lags for each variable within a maximum lag of three years for each variable.(8) A longer lag profile could seriously deplete degrees of freedom, particularly damaging in small samples. The null hypothesis that T does not Granger-cause G is rejected if the summation of [[Psi].sub.2i] is significant as a group. To test for the reverse hypothesis that G does not Granger-cause T, I estimate as in Equation 1, except I use T as the left-side variable.

[T.sub.t] = [[Phi].sub.0] + [summation of] [[Phi].sub.1j][T.sub.t-j] where j = 1 to [m.sub.1] + [summation of] [[Psi].sub.2j] [G.sub.t-j] where j = 1 to [m.sub.2] + [summation of] [[Psi].sub.3j][X.sub.t-j] where j = 1 to [m.sub.3] + [summation of] [[Phi].sub.4j][R.sub.t-j] where j = 1 to [m.sub.4] + [[Xi].sub.t] (2)

That is, the hypothesis that G does not Granger-cause T is rejected if the summation [[Phi].sub.2j] is significant as a group.

Observe that the above standard Granger causality tests between G and T are valid only if the variables are not cointegrated.(9) Otherwise, as Hendry (1986) and Engle and Granger (1987), among others, have demonstrated, inferences from such tests will be biased because they overlook valuable long-run (low-frequency) information. Therefore, it is important to examine the cointegrating properties of the variables before testing for Granger causality. To that end, I use two alternative testing methodologies.

Perhaps the most popular and simple procedure to test for cointegration is the two-step approach suggested by Engle and Granger (1987), hereafter EG. Suppose that the four variables are integrated of order one. The first step is to estimate the underlying cointegrating equations using the variables in their nonstationary (level) form. With three or more variables, the choice of the left-side conditioning variable is arbitrary (Banerjee et al. 1993). I follow Miller (1991) and examine all possible cointegrating regressions and choose that which yields the highest adjusted [R.sup.2]. In the second step of the EG procedure, the estimated residuals from the optimal cointegrating equation are recovered and checked for nonstationarity using the ADF test. To provide further evidence, I also use the PP and the WS tests, in addition to the Cointegration Regression Durbin-Watson (CRDW) test recommended by EG. If the null hypothesis of residual nonstationarity (noncointegratedness) is rejected, then the variables are said to be cointegrated. Engle and Yoo (1987) provide the critical values for the cointegration test for finite sample sizes, and Davidson and MacKinnon (1993) calculate the corresponding asymptotic values.

However, recent literature (e.g., Kramers, Ericson, and Dolado 1992; Inder 1993) has shown that the EG test suffers from poor finite sample properties that may result in large estimation biases. In addition, even moderate departures from the presumed Gaussian errors (especially in regard to normality) can significantly impair the reliability of the inferences, as pointed out by Davidson and MacKinnon (1993) and Noriega-Muro (1993). Moreover, the EG procedure is essentially limited to testing for a single (unique) cointegrating vector, which is proper only in bivariate models. With three or more variables, the model could exhibit multiple cointegrating vectors, in which case the EG test will not be useful.

Compared to the EG test, Johansen (1988) provides a better and more efficient approach to test for cointegration based on the well-accepted likelihood ratio principle. Although most of the advantages of the Johansen test are realized in multivariate models, Enders (1995) presents arguments favoring the Johansen approach over the EG test even in bivariate models. Moreover, Cheung and Lai (1993), Gonzalo (1994), and Johansen (1995) have all demonstrated that, unlike the EG test, the Johansen method has the additional advantage of not requiring Gaussian errors. Although I use both the EG and the Johansen methods to test for cointegration, I place more emphasis on the results from the Johansen robust test.

If the variables are found to be cointegrated, then they possess a long-run (equilibrium) relationship. Such a long-run (low-frequency) relationship would be filtered out if the variables are expressed in first-differences as the stationarity requirement dictates, but without due consideration to the underlying long-run comovements among the variables. Thus, to satisfy the stationarity requirement and, at the same time, avoid the loss of low-frequency information, I follow Granger's (1986) Representation Theorem and construct an error correction model (ECM) to analyze the underlying causal relationships among the variables. In the ECM, all variables are expressed in their stationary form, but a once-lagged error correction term is added as another regressor.(10) This error correction term is the stationary residuals estimated from the associated cointegrating equation. Granger discusses some interesting causality implications from error correction models. An independent variable is said to Granger-cause the dependent variable if the lagged coefficients on the independent variable are jointly significant (the standard Granger causality test) and/or the coefficient on the once-lagged error correction term is significant. Moreover, the former can be interpreted as short-run Granger causality, whereas the latter reflects long-ran Granger causality.(11)

I apply the above empirical procedure to detect the direction of Granger causality between government spending and taxes in Turkey in the context of bivariate as well as multivariate models that include real GNP and interest rates. The estimation period covers the annual period 1967-1994, the maximum amount of data available from the IMF data tape, International Financial Statistics. The unavailability of quarterly data dictated the use of annual figures. This period of three decades appears sufficiently long to gauge the cointegrating properties of the system. It is also not too long to contaminate the analysis with quite different episodes of fiscal policy regimes that might introduce structural shifts.

For the purpose at hand, the proper budgetary measures are those commonly discussed by the public and popular press; namely, total government expenditures and total government revenues from taxation. Given the extremely high rates of inflation in Turkey during most of the [TABULAR DATA FOR TABLE 1 OMITTED] estimation period, I express the figures on total government expenditures and total revenues in real terms. Many previous studies in this area have also used real values of the budgetary measures to abstract from the inflationary effects. Business cycles are measured by real GNP (X), and interest rates (R) are measured by the annualized yield on three-month government securities.

4. Empirical Results

I now discuss the empirical results obtained for Turkey from applying the methodology described in the previous section.

Unit Root and Cointegration Test Results

A key step in testing for cointegration is to determine the degree of integration of each of the four variables using the ADF, PP, and WS procedures. I assemble the unit root test results in Table 1. It is clear from the table that the log-level of each variable is nonstationary. However, according to all three tests, the variables are stationary when expressed in first-differences (of the logs). Thus, each variable is integrated of the first order. Following Dickey, Bell, and Miller [TABULAR DATA FOR TABLE 2 OMITTED] [TABULAR DATA FOR TABLE 3 OMITTED] (1986), I do not include a deterministic time trend in the unit root testing equations. Nevertheless, similar inferences emerge when the trend is included.

Next, I check whether the variables are indeed cointegrated. Table 2 reports the results from the EG test in two separate panels. Panel A contains the results for the simple bivariate model (G and T only), and panel B constitutes the results for the multivariate model (G, T, X, and R). The EG test results for both bivariate and multivariate systems suggest the presence of significant cointegration among the variables.

However, in light of the well-known problems with the EG test discussed earlier, I now turn attention to Table 3, where I report the results from the Johansen test. The evidence there for the bivariate model (panel A) clearly indicates no cointegration between government spending and taxes on the basis of both the maximal eigenvalue and the trace tests, even at the relatively weak 10% level of significance. In contrast, the Johansen test results in panel B for the multivariate (four-variable) system soundly reject the null hypothesis of no cointegration at better than the 5% level. Moreover, both the maximal eigenvalue and the trace tests imply that there is one nonzero cointegrating vector in the multivariate model which is reported in panel B after being normalized on G.(12) Consequently, there exists a stationary long-run relationship in Turkey between government spending, taxes, real GNP, and interest rates.(13)

Further Tests of the Cointegrating Relationship

I have performed additional tests to check whether I have been misled by the above statistics. In particular, how robust are the Johansen results to different lag specifications? Cheung and Lai (1993) and Gonzalo (1994) demonstrate through extensive Monte Carlo experiments that overparameterized (longer lagged) tests exhibit higher empirical power than those that are underparameterized (with shorter lags). In congruence with this evidence, the results in Table 3 are based on VARs with three annual lags, a lag length that is sufficiently long given the relative brevity of the sample. Nonetheless, to further check the sensitivity of the Johansen results to this issue of lag lengths, I performed the Johansen approach using shorter lags (one and two) as well as a longer lag (four). It is encouraging that the results (available upon request) are very similar to those reported in the table on the basis of three lags.

Note that the Johansen test reported in Table 3 suggests the presence of cointegration in multivariate but not in bivariate models. What rationale can be provided for this finding? A possible explanation may lie in omitting relevant variables from the bivariate model (in this case, real GNP and/or interest rates). As I mentioned earlier, Lutkepohl (1982, 1993) has shown that noncausality inferences drawn from bivariate models are often misleading due to the omission of relevant variables that affect either or both of the two included variables. Interestingly, Granger (1988) and Perman (1991) theorize that this omission-of-variables phenomenon is not unique with causality testing but also extends to hamper cointegration inferences. Such a theoretical contention finds empirical support in Marin (1991), Miller (1991), Darrat (1994), and Choudhry (1996). The results in this paper provide yet another piece of evidence confirming the sensitivity of cointegration tests to the omission-of-variables phenomenon.

Observe also that across the two testing methodologies, the results for the bivariate model are remarkably different. Whereas the EG test finds cointegration, the Johansen test detects none. In light of the fragility of the EG procedure, it is of course reasonable to dismiss its conclusion. And more specifically, the residual regression in the bivariate model is plagued with severe nonnormality, rendering the EG results unreliable (the Jarque-Bera [[Chi].sup.2] = 13.16; the 5% critical value = 5.99).(14)

In addition to testing for the presence of cointegration, Table 3, at the bottom of panel B also reports the results from testing various hypotheses regarding the cointegrating relationship among the four variables. In particular, the first column tests and rejects the null hypothesis that the individual series within the nonzero cointegrating vector are stationary by themselves. The test statistics in the second column checks whether any variable within the four-variable system does not belong in the cointegration space and thus can be excluded. The results from this exclusion test suggest that only real GNP appears redundant and can be safely omitted from the cointegration relationship. However, interest rates cannot be excluded because the test displays a highly significant long-run statistic. Accordingly, a more parsimonious system, which I examine below, would be to estimate a trivariate model that contains government spending, taxes, and interest rates. Finally, panel B also provides results of testing whether any variable in the four-variable system can be considered weakly exogenous. Except for government spending, each of the remaining three variables appears weakly exogenous. This implies that, among the four variables, only government spending should be considered endogenous.(15) Moreover, as Hams (1995) noted, these results also suggest that omitting the error correction term from the ECM of government spending equation would entail significant loss of pertinent information. The same, however, cannot be said regarding the error correction term in the tax equation because its role in the corresponding ECM equation appears trivial.

Considering the exclusion test results discussed above, I report in Panel C of Table 3 the cointegration results from the parsimonious trivariate system (G, T, R) after excluding real GNP. Both the maximal eigenvalue and the trace tests of the Johansen approach continue to indicate the presence of one nonzero cointegration vector in the parsimonious system. Consistent with the findings from the four-variable model, the results in the trivariate system strongly suggest the need to keep all three variables in the cointegrating space (exclusion tests) and that only government spending can be considered an endogenous variable within the system.

Granger Causality Test Results

I turn now to discussing Granger causality tests in light of the above cointegration findings. In Table 4, panels A and B, I report the Granger causality results from the bivariate and multivariate systems, alternatively using taxes and government spending as dependent variables.(16) For the multivariate system with a significant cointegrating vector, I specify an error correction [TABULAR DATA FOR TABLE 4 OMITTED] representation to analyze Granger causality.(17) However, in the context of the bivariate model, no cointegration was found and the standard Granger causality tests should be sufficient for that purpose.

As was the case for cointegration, the causality results from the bivariate and multivariate models are also strikingly different. Within the bivariate model, I could not reject the null hypothesis of no Granger causality running from government spending to taxes or vice versa (i.e., the two variables are deemed causally independent). This inference of no causality refers both to the short-run (as represented by the insignificant coefficients on lagged differences of the independent variable), as well as to the long-run (since no cointegration was revealed). However, as I discussed earlier, noncausality and noncointegration inferences drawn in a bivariate setting are suspect due to the omission-of-variables bias.

Therefore, attention should be focused instead on the Granger causality results from the multivariate system, which I report in panel B of Table 4. The results there quite decisively suggest that Granger causality does exist between government spending and taxes in Turkey. Moreover, the pattern of causality is consistent with the tax-and-spend hypothesis rather than with the spend-and-tax thesis. More specifically, based on the statistical significance of the error correction terms, the null hypothesis of no Granger causality running from taxes to government spending is soundly rejected at better than the 1% level of significance ([[Chi].sup.2] = 7.49, the 1% critical value = 6.63). However, the reverse hypothesis of no Granger causality from spending to taxes through the error term channel could not be rejected at any conventional level as indicated by the statistical insignificance of the associated error correction term ([[Chi].sup.2] = 2.03).

Besides these long-run unidirectional Granger-causal impacts, the multivariate results in panel B also suggest the presence of significant short-run unidirectional causal effects from taxes to government spending. Specifically, according to the likelihood ratio tests of the joint significance of lagged dynamic terms, the null hypothesis that taxes do not Granger-cause government spending is easily rejected ([[Chi].sup.2] = 9.20, the 5% critical value [[Chi].sup.2] = 7.81), while the reverse hypothesis is not rejected ([[Chi].sup.2] = 3.65). These results, taken together, suggest that taxes exert significant unidirectional causal effects upon government spending in Turkey in both the short-run as well as in the long-run.(18) The inferences from the multivariate ECMs of Table 4 are consistent with the earlier finding from the weak-exogeneity tests of the cointegration relationship reported in Table 3.

Before proceeding any further, it seems advisable to check the sensitivity of these causality results to alternative model specifications. For example, to what extent are the results sensitive to using different lag specifications? The results in Table 4 are based on three lags that were deemed appropriate by the FPE criterion (and were also confirmed by the asymptotically equivalent procedure, Akaike Information Criterion). However, imposing shorter or longer lags yielded very similar results (available upon request). It may also be important to check whether the results are sample specific. This, of course, requires evidence on the structural stability of the estimated models. Given the brevity of my sample, I was unable to extensively experiment with different sample periods, especially in the context of multivariate models with many parameters. This notwithstanding, I inspected the sensitivity of my conclusions to sample selection by omitting the period before the oil price hike of 1973, by using the periods before and after the 1980 coup, and alternatively, by omitting the years after the approval of the new Turkish Constitution of 1982. The results (available upon request) continued to reveal inferences very similar to those reported in Table 4. As a further check, I applied the Chow and the custom-of squares tests of structural stability. The results (available upon request) corroborated parameter constancy in the estimated models.

Note also that, as Fischer (1989) and Lutkepohl (1989) point out, reliable policy inferences from causality tests hinge crucially on their invariance to policy regime changes. Recently, Engle and Hendry (1993) demonstrate that this invariance proposition is closely associated with the concept of superexogeneity. I use the methodology proposed by them and test for superexogeneity of the relevant variables. In so doing, the auxiliary equations include a dummy variable for the purpose of testing the superexogeneity hypotheses for spending and taxes in the bivariate and multivariate models. This dummy variable takes the value of one for the post-1980 period, and zero otherwise. Compared to all other political turbulence in modern Turkey, the 1980 coup is perhaps the worst in terms of its relative lack of tolerance of civil liberties and partisan politics (Hooglund 1996). Indeed, if the estimated models prove invariant to the structural change introduced by the 1980 coup, a case can be easily made for their stability throughout the sample period. In the bivariate equations and the multivariate ECMs of Table 4, the coefficient vectors of the appropriate variables are proven invariant to the policy regime shift of the 1980s. (The F-values are 0.28 and 0.44 for the two bivariate equations and 0.41 and 0.29 for the two multivariate ECMs. None of these values is significant at any conventional level.)(19)

In summary, the empirical results lend strong support to the contention that taxes have unidirectionally Granger-caused government spending in Turkey. Moreover, this Granger causality is potent, occurring both in the short-run and in the long-run. These results, taken together, are consistent with the tax-and-spend hypothesis of Friedman and Buchanan-Wagner. What implications for the deficit solution debate emerge from such results? In other words, should Turkish authorities cut taxes in order to control budget deficits a la Friedman, or should they instead raise taxes to accomplish that same goal a la Buchanan and Wagner? The empirical evidence from the multivariate ECMs are more in line with the Buchanan and Wagner policy recommendation. This is because the (short-run) impact of tax changes on government spending proves negative (-1.86) and statistically significant at better than the 5% level.

Therefore, consistent with the Buchanan and Wagner view, raising taxes in Turkey should raise the perceived price of government goods and services, which in turn exert pressures on the government to reduce its spending, leading to significant reductions in budget deficits.

5. Concluding Remarks

This paper investigates the causal (lead/lag) relationship between government spending and taxes for Turkey. In contrast to many previous studies for the U.S. and other large industrialized countries, my empirical analysis incorporates the cointegrating properties of the variables and, moreover, expands the common restrictive bivariate model to include other theoretically relevant variables. In particular, the multivariate model includes real GNP and interest rates as macroeconomic control variables. The bivariate and the multivariate models come to dramatically different conclusions regarding cointegration and causality between the two fiscal variables. This finding underscores the need to continue with efforts to specify and test more complete (broader) models of the budgetary process. Owing perhaps to the familiar omission-of-variables bias, I place more emphasis on inferences drawn from multivariate models.

The results that consistently emerge from the multivariate cointegration analysis for Turkey support the existence of one nonzero cointegrating vector representing a long-run equilibrium relationship between the two fiscal variables. Moreover, evidence from the multivariate error correction models suggests that taxes unidirectionally and significantly Granger-cause government spending, both in the short-run and in the long-run. These results imply rejection of the spend-and-tax hypothesis in favor of the tax-and-spend proposition in the case of Turkey. The results further reveal that the unidirectional causal impact of taxes on spending is significantly negative, as hypothesized by Buchanan and Wagner. Therefore, from the perspective of policy making and the deficit solution debate, it appears that raising taxes in Turkey should prove an optimal solution to the current budget deficit predicament.

I thank, without implicating, D.C. Anderson, O. W. Gilley, and an anonymous reviewer for several useful comments and suggestions.

1 I am indebted to an anonymous reviewer for pointing out this and many other important aspects of the paper.

2 For more on these and other relevant political issues in the Turkish context, see Hooglund (1996).

3 Observe, however, that none of these studies investigates whether higher taxes cause higher spending a la Friedman or cause lower spending a la Buchanan-Wagner.

4 Throughout the paper, I attach "Granger" to "cause" because controversy still surrounds the Granger concept of causality, which somewhat differs from the definition of causality in the strict philosophical sense. Indeed, tests of Granger causality are essentially tests of the incremental predictive content of economic time series. See Bishop (1979) and Zellner (1979) for a fruitful discussion.

5 Of course, the addition of the two variables may not be fully adequate, and any expanded model runs the risk of omitting other important variables. Yet, it was felt that these two additional variables are reasonably grounded in theory and seem sufficient for the purpose of illustrating the susceptibility of bivariate models to the omission-of-variables bias.

6 Ahiakpor and Amirkhalkali (1989) have also included GNP in their models for Canada.

7 McCallum's (1984) work further implies the need to adjust fiscal variables for interest payments.

8 The FPE is based on the minimization of the one-step-ahead prediction error. It is a compromise between the predictive power of a model and its complexity as measured by its lagged order. For further details on the FPE criterion and its application, see Darrat (1988). Thornton and Batten (1985) report results showing the superiority of the FPE criterion over other alternative lag length selection procedures.

9 Cointegrated variables, if disturbed, will not drift away from each other and thus possess a long-run equilibrium relationship.

10 Additional lags of the error correction term are unnecessary since they are already reflected in the distributed lags of the first-differences of the variables. See Miller (1991).

11 Jones and Joulfaian (1991) suggest a similar interpretation.

12 As Dickey, Jansen, and Thornton (1991) and Alogoskonfis and Smith (1991) pointed out, cointegrating vectors are difficult to interpret because they do not reflect structural equations. Nonetheless, the estimated cointegrating vector implies, as it should, that there is a positive long-run relationship between spending and taxes. I do not report the remaining three eigenvectors since they are statistically insignificant (nonstationary).

13 Other studies have also found a long-run relation between government spending and taxes in the U.S. See, for example, Miller and Russek (1990), Bohn (1991), and Jones and Joulfaian (1991).

14 Interestingly, the residual regression based on the multivariate system does not reveal any evidence of nonnormality (the Jarque-Bera [[Chi].sup.2] = 3.24).

15 In itself, this finding may be taken to imply support for the tax-and-spend hypothesis. More on this below.

16 Since real GNP and interest rates are likely related to factors other than the fiscal variables (e.g., monetary policy and international developments), separate equations for X and R are not reported here.

17 The quadrivariate ECMs (DG, DT, DX, DR) are based on residuals obtained from the Johansen efficient maximum likelihood estimations within the parsimonious trivariate system (G, T, R).

18 The results in panel B of Table 4 further indicate that interest rates (and not real GNP) exert important causal impacts on both fiscal variables and particularly on tax revenues.

19 I performed the superexogeneity test using alternative dummy variables. The results, available upon request, do not suggest different conclusions.

References

Adaman, E, and M. R. Sertel. 1995. The changing economic role of the state from a Turkish perspective. Economic Research Forum Working Paper No. 9510.

Ahiakpor, J. W., and S. Amirkhalkhali. 1989. On the difficulty of eliminating deficits with higher taxes: Some Canadian evidence. Southern Economic Journal 56:24-31.

Alogoskoufis, G., and R. Smith. 1991. On error-correction models: specification, interpretation, and estimation. Journal of Economic Survey 5:97-128.

Anderson, W., M. S. Wallace, and J. T. Warner. 1986. Government spending and taxation: What causes what? Southern Economic Journal 52:630-39.

Banerjee, A., J. J. Dolado, J. W. Galbraith, and D. F. Hendry. 1993. Cointegration, error-correction and the econometric analysis of non-stationary data. London: Oxford University Press.

Barro, R. J. 1979. On the determination of the public debt. Journal of Political Economy 81:940-71.

Bishop, R. V. 1979. The construction and use of causality tests. Agricultural Economics Research 31:1-6.

Blackley, P. R. 1986. Causality between revenues and expenditures and the size of the federal budget. Public Finance Quarterly 14:139-56.

Bohn, H. 1991. Budget balance through revenue or spending adjustments? Some historical evidence for the United States. Journal of Monetary Economics 27:333-59.

Buchanan, J. M., and R. W. Wagner. 1977. Democracy in deficit. New York: Academic Press.

Buchanan, J. M, and R. W. Wagner. 1978. Dialogues concerning fiscal religion. Journal of Monetary Economics 4:627-36.

Bugra, A. 1994. State and business in modern Turkey. New York: Suny Publications.

Cheung, Y. W., and K. S. Lai. 1993. Finite sample sizes of Johansen's likelihood ratio tests for cointegration. Oxford Bulletin of Economics and Statistics 55:313-28.

Choudhry, T. 1996. Real stock prices and the long-run money demand function: Evidence from Canada and the USA. Journal of International Money and Finance 15:1-17.

Darrat, A. E 1988. Have large budget deficits caused raising trade deficits? Southern Economic Journal 54:879-87.

Darrat, A. E 1994. Wage growth and the inflationary process: A reexamination. Southern Economic Journal 61:181-90.

Darrat, A. E. and A. Arize. 1996. Budget deficits and the value of the dollar: Further results from cointegration and error-correction modeling. Louisiana Tech University, Working Paper.

Davidson, R., and J. G. MacKinnon. 1993. Estimation and inference in econometrics. Oxford: Oxford University Press.

Dickey, D. A., W. R. Bell, and R. B. Miller. 1986. Unit roots in time series models: Tests and implications. The American Statistician 40:12-26.

Dickey, D. A., D. W. Jansen, and D. L. Thornton. 1991. A primer on cointegration with an application to money and income. Federal Reserve Bank of St. Louis Review 73:58-78.

Enders, W. 1995. Applied econometric time series. New York: John Wiley & Sons.

Engle, R., and D. F. Hendry. 1993. Testing super-exogeneity and covariance in regression models. Journal of Econometrics 57:119-39.

Engle, R. F. and C. W. J. Granger. 1987. Co-integration and error-correction representation, estimation, and testing. Econometrica 55:251-76.

Engle, R. F., and S. Yoo. 1987. Forecasting and testing in cointegrated systems. Journal of Econometrics 35:143-59.

Fackler, J. 1985. An empirical analysis of the markets for goods, money and credit. Journal of Money, Credit and Banking 17:28-42.

Fischer, A.M. 1989. Policy regime changes and monetary expectations: Testing for superexogeneity. Journal of Monetary Economics 6:423-36.

Friedman, M. 1982. Supply-side policies: Where do we go from here in supply-side economics in the 1980s: Conference proceedings. Westport, CT: Quorum Books.

Furstenbert, G. M. von, R. J. Green, and J. Jeong. 1986. Tax and spend, or spend and tax? Review of Economics and Statistics 68:179-88.

Gonzalo, J. 1994. Five alternative methods of estimating long-run equilibrium relationships. Journal of Econometrics 60:203-33.

Granger, C. W. J. 1986. Developments in the study of cointegrated economic variables. Oxford Bulletin of Economics and Statistics 48:213-28.

Granger, C. W. J. 1988. Some recent developments in the concept of causality. Journal of Econometrics 16:199-211.

Harris, R. I. D. 1995. Using cointegration analysis in econometric modelling. London: Prentice Hall.

Hendry, D. F. 1986. Econometric modelling with cointegrated variables: An overview. Oxford Bulletin of Economics and Statistics 48:201-12.

Hooglund, E. 1996. Government and politics. In Turkey: A country study. 5th Edition. Washington, DC: Federal Research Division of Library of Congress, pp. 233-301.

Hoover, K. D., and S. M. Sheffrin. 1992. Causation, spending, and taxes: Sand in the sandbox or tax collector for the welfare state? American Economic Review 82: 225-48.

Inder, B. 1983. Estimating long-run relationships in economics: A comparison of different approaches. Journal of Econometrics 57:53-68.

Johansen, D. 1988. Statistical analysis of cointegration vectors. Journal of Economic Dynamics and Control 12:231-54.

Johansen, D. 1995. Likelihood-based inference in cointegrated vector autoregressive models. New York: Oxford University Press.

Jones, J. D., and D. Joulfaian. 1991. Federal government expenditures and revenues in the early years of the American republic: Evidence from 1792 to 1860. Journal of Macroeconomics 14:133-55.

Joulfaian, D., and R. Mookerjee. 1990. The intertemporal relationship between state and local government revenues and expenditures: Evidence from OECD countries. Public Finance 45:109-17.

Kramers, J. J. M., N. R. Ericsson, and J. Dolado. 1992. The power of cointegration tests. Oxford Bulletin of Economics and Statistics 54:325-48.

Lee, D. R., and R. K. Vedder. 1992. Friedman tax cuts vs. Buchanan deficit reduction as the best way of constraining government. Economic Inquiry 30:722-32.

Lutkepohl, H. 1982. Non-causality due to omitted variables. Journal of Econometrics 19:367-78.

Lutkepohl, H. 1989. The stability assumption. In Tests of causality between money and income, edited by W. Kramer. Heidelberg, Germany: Physica-Verlag, pp. 75-86.

Lutkepohl, H. 1993. Introduction to multiple time series analysis. 2nd edition. Berlin: Springer-Verlag.

McCallum, B. T. 1984. Are bond-financed deficits inflationary? A Ricardian analysis. Journal of Political Economy 92: 123-35.

Manage, N., and M. L. Marlow. 1986. The causal relation between federal expenditures and receipts. Southern Economic Journal 52:617-29.

Marin, D. 1992. Is the export-led growth hypothesis valid for industrialized countries? Review of Economics and Statistics 74:678-88.

Meltzer, A. H., and S. E Richard. 1981. A rational theory of the size of government. Journal of Political Economy 89: 914-27.

Miller, S. M. 1991. Monetary dynamics: An application of cointegration and error-correction modelling. Journal of Money, Credit and Banking 23:139-54.

Miller, S. M., and F. S. Russek. 1990. Co-integration and error-correction models: The temporal causality between government taxes and spending. Southern Economic Journal 57:221-29.

Musgrave, R. 1966. Principles of budget determination. In Public Finance: Selected readings, edited by H. Cameron and W. Henderson. New York: Random House, pp. 15-27.

Noriega-Muro, A. E. 1993. Nonstationarity and structural breaks in economic time series. New York: Avebury Publishers.

Owoye, O. 1995. The causal relationship between taxes and expenditures in the g-7 countries: Cointegration and error-correction models. Applied Economics Letters 2:19-22.

Pantula, S. G., G. Gonzales-Farias, and W. A. Fuller. 1994. A comparison of unit-root test criteria. Journal of Business and Economic Statistics 12:449-59.

Peacock, A. T, and J. Wiseman. 1979. Approaches to the analysis of government expenditure growth. Public Finance Quarterly 7:3-23.

Perman, R. 1991. Cointegration: An introduction to the literature. Journal of Economic Studies 18:3-30.

Pope, K. 1996. Turkey's in turmoil - And the bulls cheer. Wall Street Journal June 26:11.

Ram, R. 1988. Additional evidence on causality between government revenue and government expenditure. Southern Economic Journal 54:763-69.

Sims, C. A. 1980. Macroeconomics and reality. Econometrica 48:1-48.

Thornton, D. L., and D. S. Batten. 1985. Lag-length selection and tests of granger-causality between money and income. Journal of Money, Credit and Banking 17:164-78.

Wildavsky, A. 1988. The new politics of the budgetary process. Glenview, IL: Scott Foresman.

Zellner, A. 1979. Causality and econometrics. In Three aspects of policy and policy-making: Knowledge, data and institutions, edited by K. Brunner and A. H. Meltzer. Amsterdam: North-Holland Publishing Company, pp. 9-54.