# Fiscal policy and trade adjustment: are the deficits really twins?

I. INTRODUCTION

The unprecedented increase of both the fiscal and trade deficits in the United States during the 1980s led many observers to posit a causal linkage. A sizable contingent of economists believe that the "twin deficits" concept aptly characterizes the recent relationship between federal budgetary policy and the U.S. trade balance. In this view, increased federal fiscal deficits during the 1980s were a primary cause of the massive U.S. trade deficits experienced as the decade unfolded. However, this posited causal relationship does not command universal acceptance. In this paper, we examine data on U.S. trade and federal fiscal measures along with exchange and interest rates to uncover the empirical relationship between them and to discover whether fiscal policy plays an essential role in the adjustment of the massive U.S. trade imbalance.

The twin deficits concept asserts that the deficits are "twins" because they are related in the following manner. Increased government deficits are thought to put incipient upward pressure on real interest rates. This induces incipient capital in-flows that lift or appreciate the real exchange value of the currency, which erodes competitiveness and, after a time lag, results in increased trade deficits.

A number of recent papers examine the relationship between budget and trade deficits. Each contributes important insights, but nothing approaching a consensus has yet emerged. Miller and Russek |1989~ find evidence of a secular relationship between the deficits under two of the three statistical techniques they employ. Bernheim |1988~ concludes that the government deficit is a prime determinant of U.S. trade deficits, whereas Evans |1989~, Enders and Lee |1990~, and Dewald and Ulan |1990~ find no such causal impact of fiscal deficits. Darrat |1988~ and Abell |1990~ report somewhat mixed results.

The conflicting results in the existing literature stem from wide differences in the empirical techniques, data measures and samples employed. Further, we have found in the literature and in our work reported below that the use of differenced data versus data in level form has a major impact on the results produced and inferences drawn. Recent empirical literature suggests that stationarity of data is a crucial element in model specification. Thus, the question of whether or not to difference the data should be addressed before estimation and inference.

Classical inference about stationarity of data suffers from what Sims |1988~ refers to as a "discontinuity" in the asymptotic theory. Hence, the classical tests used in much of the previous literature have low power to reject the null hypothesis of nonstationarity in levels. Bayesian inference does not suffer from this problem; therefore, we generate posterior probabilities using Monte Carlo integration that estimate probabilities that suggest whether the data are better analyzed in level or difference form. Despite somewhat contrasting results, the classical augmented Dickey-Fuller tests and the Bayesian probability estimates both suggest that a difference specification is inappropriate.

We analyze empirically a five-variable nonstructural vector autoregressive (VAR) model. The model includes both a federal government balance and a government purchases measure to summarize fiscal policy. Along with these two variables and a trade balance measure, we constructed real interest rate and real exchange rate variables spanning a quarterly data sample from 1961 to 1989. We estimate the model in both levels and difference form. Using the level specification, we find strong support for the twin deficit notion that government deficits are a cause of trade deficits, whereas in differences we find no such relationship. Furthermore, the results from our Monte Carlo work imply that the model specified in levels is likely most appropriate for estimation and inference.

II. CONTRASTING THEORETICAL FRAMEWORKS

The presence of both deficits in a variant of the basic savings/investment identity may provide a useful reference, as in Bernheim |1988~ and Branson |1985~, for discussing the hypothesized causal relationship embodied in popular twin deficits stories:

(1) TDEF = GDEF + (I - |S.sub.p~),

where TDEF is the broad trade deficit, GDEF is the government budget deficit, and I and |S.sub.p~ are private sector investment and savings, respectively.(1) However, the presence of both deficits in this identity does not imply that the movements in one deficit cause commensurate changes in the other deficit because the term (I-|S.sub.p~) can change as well.

The above identity provides at best casual motivation for the twin deficits relationship. However, the twin deficits story receives some theoretical justification from various frameworks, including the traditional Mundell-Fleming model.(2) The Mundell-Fleming model assumes that an increased government deficit policy (for example, a tax cut) raises aggregate demand (essentially the fiscal shock is not fully offset by increased private savings).

This increased demand places incipient upward pressure on the domestic real interest rate. This interest rate pressure induces an incipient net capital inflow from abroad, appreciating the domestic currency's real foreign exchange value.(3) After a lag, this relatively more expensive domestic currency pushes the nation's trade balance towards deficit. Thus, this model implies that increased government deficits ultimately widen the trade deficit (under both fixed and flexible exchange rate regimes, although the transmission mechanisms differ).(4)

This model is not universally accepted. A key reason is that there exists a major controversy on the link between government deficits and exchange rates. Branson |1988~ suggests that increased U.S. government deficits are the prime determinant of the dollar's appreciation. Feldstein |1986~ also supports a linkage whereby increased government deficits appreciate the dollar. Conversely, Evans |1986~ finds evidence that increased deficits depreciate the dollar.(5)

Unlike the Mundell-Fleming model, one major school of thought expects no causal impact from government budget deficits to trade deficits. A model incorporating Ricardian Equivalence would suggest that a substitution of debt for taxes by the government that increases the fiscal deficit would be offset by increased private savings, rather than increased net foreign borrowing (trade deficit). However, changes in (temporary) government spending behavior can affect trade deficits in this framework. Thus, we include a measure of federal government purchases in our empirical study to distinguish between government purchases and government balances as the potential sources of the twin deficits correlation.

Our five-variable model, by including two measures of fiscal policy, should help us distinguish between a Mundell-Fleming and a Ricardian interpretation of the role of fiscal policy in trade adjustment. The former approach argues that the fiscal balance affects subsequent trade balances, whereas the latter implies that (temporary) government purchases, not government balances, impact the trade balance.

Given these contrasting views and the mixed results from prior research, our study aims to provide additional empirical evidence on the relationship between fiscal deficits, the dollar, and trade deficits.

III. DATA SAMPLE AND EMPIRICAL TECHNIQUES

Our data selection follows directly from the two contrasting theoretical models summarized above. Our model includes the five variables that span the two theoretical frameworks summarized above. We examine an extensive sample (1961:I to 1989:IV) that includes quarterly data points from the recent record-sized federal budget and U.S. trade deficits period, as well as recent apparent turning points in these series.(6) We construct the federal deficit measure as the change in the real par value of privately held federal debt. Our constructed deficit measure accounts for the inflation tax on government debt by measuring the change in the real value of the debt before renominalizing it.(7,8) We include federal government purchases as the companion fiscal policy measure. Our trade balance measure is net foreign investment (NFI), the series consistent with NIPA accounting.

Theoretical considerations suggest using real rather than nominal variables for analysis. By constructing ratios to nominal GNP, we have created trade balance (TBAL) and federal government fiscal policy variables (GBAL for federal balances and GPURCH for federal purchases) that are essentially real measures. We also constructed real interest and real exchange rate series for our sample. The real interest rate (IRATE) is the difference between the nominal rate on a three-month Treasury bill and the ex post CPI inflation rate. We constructed a multilateral trade-weighted real exchange rate index of the U.S. dollar's foreign exchange value (ERATE). We paralleled the standard Federal Reserve Board index, using its weights comprising ten major currencies, but constructed a real index by multiplying the nominal trade-weighted value of the dollar by the U.S. CPI divided by the trade-weighted foreign CPI. Increases in this index signify a real appreciation of the dollar.

The theoretical frameworks discussed above do not imply a specific dynamic structure. Consequently, we choose an empirical methodology that places minimal restrictions on the explicit structure of the relationship. The VAR techniques we employ provide a convenient estimation procedure that investigates data relationships while imposing little structure.(9)

Our VAR system employs the five variables suggested by the two theoretical frameworks: GPURCH, GBAL, TBAL, IRATE, and ERATE. Prior to estimating the system, we determine the appropriate lag length with likelihood ratio tests for models using levels as well as first differences. The recent literature on U.S. trade adjustment, for example Hooper and Mann |1989~, finds that trade balances react to factors such as exchange rate fluctuations with lags that span years rather than quarters. This prior evidence motivated us to test a twelve- versus a six-quarter lag VAR specification, as opposed to any shorter lag structures. Test results suggested that we could not reject the null hypothesis that six lags are sufficient.(10)

Much attention in recent macroeconomic empirical literature focuses on questions regarding the stationarity of time-series data. Nonstationary or integrated data series (i.e., those containing unit roots) render statistical inference from traditional output invalid. Classical statistical methods, such as augmented Dickey-Fuller tests, are used most often to detect the possible existence of unit roots in data series. Table I reports results using these techniques on the data series. At the 5 percent significance level, we could reject the null hypothesis of nonstationarity for only one of the five variables in level form.(11) Analyzing first-differenced data, test statistics imply rejection at 5 percent significance of the null hypothesis of nonstationarity for all five series. Overall, using the 5 percent significance level, the results suggest that four of the five variables are integrated of order one and the remaining variable is level stationary.(12) To justify a first-difference specification, there should be evidence of five unit roots in the system. Thus, at the 5 percent significance level, the classical Augmented Dickey-Fuller tests do not dictate a first-difference specification.

Inferences based on classical statistical methods, however, have recently been challenged. A growing body of literature indicates that: "the power of Dickey-Fuller type integration tests can be quite low (less than 50 percent) against plausible trend stationary alternatives."(13) Hence, a number of researchers, notably Sims |1988~, have suggested approaching this issue from a Bayesian perspective, asking the question: which data representation, difference or level stationarity, is more probable given the data?(14) Following this suggestion, we use Monte Carlo integration of the VAR specification in levels to generate posterior probabilities, which indicate the probability that the system is better specified as level or difference stationary.(15)

The Monte Carlo strategy for generating the posterior probability odds ratios involves relatively few steps but numerous iterations of the main procedures. Essentially, we generate a sequence of independent drawings from the posterior distribution of the parameters. Before we begin the iterations, we estimate the model in level form to get the posterior means of the model parameters. By assuming a diffuse prior on the model coefficients and the variance-covariance matrix of the model error terms, the posterior distribution of the variance-covariance matrix is an inverted Wishart distribution, and the posterior distribution of the coefficients is normal, conditional on the variance-covariance matrix of the model error terms. For each iteration of the integration procedure, we make a draw of the variance-covariance matrix, and use this draw to specify the conditional normal distribution for the parameters of interest. From this conditional distribution, we make a drawing of the parameter matrix and determine its roots.

The next step requires that we distinguish between roots that are unit or greater (implying nonstationarity) and those that are close to but less than one (implying level stationarity). The iterations performed using Monte Carlo integration are divided into sets of observations: for example, draws consistent with level stationarity versus those consistent with nonstationarity in levels. The number of draws consistent with level stationarity relative to the total number of simulations represents an estimate of the probability that the model is level stationary. Conversely, the relative frequency of iterations with one or more roots greater than or equal to unity reflects an estimate of the posterior probability that the model is nonstationary in levels. After 20,000 iterations of the program, we find a 64.6 percent posterior probability that the model is stationary in levels, implying a 35.4 percent probability that the largest root is unit or greater (implying nonstationarity).

These probability estimates provide some evidence that unit roots may exist in this system. The model would need at least one unit root in order to have cointegration of the series, and there is a nontrivial probability of one unit root.(16) Yet, for our purposes, the relevant issue is whether there is evidence supporting the existence of five unit roots, which would indicate difference stationarity. In our Monte Carlo work, the estimate of the probability that the system has five unit roots is only .00125. In contrast to the results from the classical tests that find four unit roots, the estimated probability of four or more unit roots in the system is less than 1 percent. Thus, we find Monte Carlo evidence against a difference stationary specification.

The results of both diagnostic techniques suggest possible cointegration of the series; however, this does not invalidate an unrestricted level VAR specification. Although an unrestricted level specification may be inefficient, by ignoring the long-run restrictions that cointegration imposes, the estimation is consistent. On the other hand, an unrestricted VAR in first differences when the true model is cointegrated is misspecified because it ignores lagged equilibrium errors as regressors, which renders it inconsistent.(17) In applications with limited data samples, cointegrating relationships are hard to identify precisely. If restrictions implied by cointegration are misspecified, the estimation is inconsistent. Furthermore, Watson |1987~ finds that cointegration restrictions at best add marginally to the interpretation of the results.

Results from an unrestricted level specification of the system are robust to the potential existence of cointegration, and are least likely to be subject to misspecification error if there are less than five unit roots. The posterior probabilities generated from the Monte Carlo integration indicate that the model likely is best specified in level form. Thus, our primary set of results are from the level specification. We also report a secondary set of results, those from a model specified in first differences. We perform this exercise to emphasize how the choice of a level or a first-difference specification affects the empirical evidence and the resulting inferences.

IV.EMPIRICAL RESULTS

Our primary specification employs levels of the five variables. We examine decompositions of the forecast variance and analyze the relative contribution of orthogonalized innovations associated with each variable in the system toward explaining the forecast variance of each series.(18) We derive the orthogonalized innovations from a Choleski decomposition in which we specify the causal ordering of the variables in a recursive structure.

Our primary results, reported in Table II, were generated with an ordering that is the reverse of the causal chain in the Mundell-Fleming model to ensure that the choice of ordering does not favor the finding of "twin deficits." The ordering runs from trade balance (TBAL) to exchange rate (ERATE) to interest rates (IRATE) to government balances (GBAL) to government purchases (GPURCH). The variable GBAL in the primary specification is our constructed (from changes in real debt) government balance as a ratio to GNP.

The results from the primary five-variable VAR in levels show the variance decompositions at a sixteen-quarter forecast horizon. We employ a method similar to that developed by Runkle |1987~ to generate standard errors for our system and produce 90 percent posterior probability ranges for each decomposition.(19) These ranges are listed below the posterior mean variance decompositions in Table II (and all subsequent tables). Despite the ability to generate posterior probability ranges, there is no consensus on explicit criteria for significance in VAR analysis. For example, Runkle |1987~ argues that a useful criterion for significance is a probability range completely above 10 TABULAR DATA OMITTED percent in variance decompositions, but 10 percent is an arbitrary threshold. On the other hand, Sims |1987~ refers to the impulse response functions and attributes significance to whether responses to an innovation are large and bounded away from zero. Consequently, we present both forms of evidence to investigate significance.

Variance Decompositions

The notable result in Table II is that government balance (GBAL) innovations are associated with 42.2 percent of broad trade balance (TBAL) variance at a sixteen-quarter horizon. In addition, the 90 percent posterior probability range far exceeds 10 percent, satisfying Runkle's criterion. The result is consistent with the twin deficits story, implying a significant impact of fiscal balance innovations on the trade balance, but we require impulse response function results to determine the sign of the interaction. In contrast, TBAL innovations are associated with only 8.0 percent of GBAL variance, and the probability range includes zero.(20) Thus, the evidence is consistent with the twin deficits notion that the direction of causality runs from government fiscal balances to the trade balance.

The postulated relationship between the trade and fiscal balances is the primary element of the twin deficits story. The variance decompositions show that, except for the impact of GBAL on TBAL, each variable explains an insignificant proportion (using Runkle's criterion) of any other variable's variance.

Table II shows that innovations in GBAL, but not in federal purchases (GPURCH), provide significant predictive power for future trade balances. This distinction is consistent with a twin deficit or Mundell-Fleming, as opposed to Ricardian, notion of the impact of fiscal policy on the trade balance.

We examine the robustness of these results with regard to the chosen measure of the fiscal deficit variable and with regard to the ordering of the variables in the Choleski decomposition.(21) Table III reports results for a model that differs from the primary one by substituting the conventionally measured NIPA federal government deficit series for the nominalized 'real' deficit that we constructed from changes in real federal debt. A comparison of Tables II and III indicates that our results are qualitatively similar for either measure of GBAL.(22) The major finding that the only significant explanatory linkage is the one running from GBAL innovations to TBAL appears robust to the particular measure of government balances employed.(23)

All specifications using the data in level form yield similar results. To highlight the role of differencing data, a factor that appears crucial in the inconclusive existing literature, Table IV presents evidence from a VAR specified in first differences. Differencing clearly has a profound effect, because Table IV shows that in this specification no variable's innovations explain a significant proportion of any other variable's variance. Specifically, the impact of GBAL innovations on TBAL falls to 12.4 percent. No variance decompositions satisfy the Runkle significance criterion. The results in Table IV provide little support of the twin deficits notion.(24) Consequently, the contrast between Table IV and Table II emphasizes the importance of evidence against the first-difference specification provided by the Bayesian posterior probability odds estimates as well as the classical augmented Dickey-Fuller tests.

Impulse Response Functions. Variance decompositions do not indicate the sign and dynamic pattern of the relationship between the variables of interest. However, theory postulates unambiguous signs for the impulse response functions of the variables. To examine the dynamic pattern and sign of impacts, we present the impulse response functions over a sixteen-quarter horizon. We generate 90 percent posterior probability bands for the impulse responses in a further attempt to assess significance, following the suggestion of Sims |1987~. The figures show impulse response functions generated from our primary specification in levels.

Figure 1 shows that a positive shock to the federal government balance to GNP ratio ultimately leads to unambiguous increases in the trade balance to GNP ratio. The probability bounds are clearly above TABULAR DATA OMITTED zero for the entire period four to sixteen quarters beyond the shock. This impulse response result, in addition to the variance decompositions, provides strong support for the central necessary relationship in the twin deficits story.

Given the support for the twin deficits story, we investigate other proposed linkages as diagnostics to be sure that the interactions among variables have the predicted sign. We find in Figure 2 that the innovation in GBAL has a negative effect on the real exchange rate (ERATE). The effect is immediate and the response function is bounded away from zero for the first ten periods. This result is robust across alternative fiscal measures and sample periods. Thus, budget deficits are associated with an appreciating real currency, consistent with a twin deficit story. This result adds evidence to the controversial issue regarding the impact of U.S. fiscal deficits on the dollar.

In Figure 3, we find further evidence consistent with the twin deficit story. Innovations in the real exchange value of the dollar (ERATE) are associated with subsequent declines in the trade balance (TBAL).(25) The negative relationship is bounded away from zero for a period from eight to fifteen quarters out, depicting a TABULAR DATA OMITTED lag similar to that found in the recent literature on the link between trade balances and the dollar, notably Hooper and Mann |1989~.(26)

In sum, the impulse response functions and the variance decompositions for the level specifications provide evidence consistent with the predictive role of government deficits in the typical twin deficits explanation.

V. SUMMARY AND IMPLICATIONS

Spurred by the dramatic surge in both U.S. trade and fiscal deficits during the 1980s, a particular causal story ("twin deficits") asserting a relationship whereby government deficits lead to trade deficits gained many adherents. Previous research investigating the empirical relevance of the twin deficits story has not produced a consensus. We add to this literature by testing for the various possible relationships between fiscal policy, exchange rates, and trade adjustment, employing VAR techniques with relatively few prior restrictions.

The relationship between fiscal and trade variables reported in the existing empirical literature appears crucially dependent on whether the model is specified in levels or first differences. We use posterior probability estimates generated by Monte Carlo methods to determine whether the system is best specified in levels. We find, despite a slight contrast between inferences from classical Dickey-Fuller tests and the Monte Carlo methods, that the VAR system is probably best specified in levels. For completeness and comparison purposes, we estimate the VAR system in both level and difference specifications. The level specification generates results that are consistent with the twin deficits notion embedded in a Mundell-Fleming framework. Overall, the results suggest that increased U.S. government deficits may have contributed to dollar appreciation and the large U.S. trade deficits in the 1980s. Thus, fiscal policy appears to have a significant role to play in U.S. trade balance adjustment.

1. We employ net foreign investment (NFI) from the National Income and Product Accounts (NIPA) as our trade balance measure, and a federal balance series constructed from federal debt data (described below) as the government balance measure. A few studies employ the current account from the Balance of Payments accounts as the trade balance measure. Although this is very close to the net foreign investment measure of the trade balance that we use, historic U.S. current account data unfortunately include capital gains and losses reflecting the foreign exchange translation of direct foreign investment. Thus, we use net foreign investment as our trade balance measure, consistent with NIPA accounting.

2. For a clear and extensive discussion of the Mundell-Fleming framework as well as other theoretical approaches, see Miller and Russek |1989~. Branson |1985; 1988~ offers an alternative framework for this linkage.

3. The real appreciation stems from an appreciation in the nominal exchange rate in a flexible exchange rate regime, and from an increased domestic price level under fixed rates. The appreciation relies on an assumption in the basic model of sufficient capital mobility to outweigh direct domestic demand effects worsening the current account.

4. Under a fixed rate regime, the trade deficit deteriorates because of an income effect as well as the real appreciation.

5. For a clear and comprehensive discussion of this controversial linkage, highlighting the role played by the degree of capital mobility, see Willett and Wihlborg |1990~.

6. The sample is the longest continuous one we could create--our starting date was limited by the price data for the real exchange rate series we constructed. The ending dates are limited by the various measures of the fiscal deficits.

7. We thank W. Michael Cox for providing the series on the par value of privately held federal debt. To construct the government balance series, we use the following identity. Normalized real government balances (GBAL) are

GBAL = - |P.sub.t~|(|D.sub.t~/|P.sub.t~) - (|D.sub.t-1~/|P.sub.t-1~)~,

where P is the consumer price index and D is the par value of privately held federal debt outstanding. We thank a referee for this suggestion.

8. We estimate a system that employs the standard NIPA nominal federal deficit measure as well as one that includes a cyclically adjusted federal government deficit series (provided by the Bureau of Economic Analysis based on mid-expansion GNP trend). Results from these systems allow us to compare them with some of the existing literature. The results from both systems are qualitatively similar to those in our primary specification, indicating robustness.

9. The Choleski decomposition of the variance-covariance matrix does impose a recursive structure on the residuals. However, we have examined various alternative orderings, and find qualitatively similar results (available on request).

10. The likelihood ratio test generated a |X.sup.2~(150) of 156.86 for the level specification, and 137.13 for the difference specification. At the 5 percent significance level, the test statistic must exceed the critical value of 179.58. Thus, neither test could reject the null.

11. On theoretical grounds, we did not include a trend in the augmented Dickey-Fuller tests. Because the trade and government variables are measured as ratios to nominal GNP, we should not expect unbounded trending of the series. We reject the nonstationarity hypothesis for ERATE due to the presence of a significant drift term that necessitated using the critical values in Schmidt |1988~.

12. Note, however, the choice of a 1 percent significance level would imply difference stationarity. The potential ambiguity of the implications from the augmented Dickey-Fuller tests provide further motivation for exploring an alternative diagnostic procedure.

13. DeJong and Whiteman |1989, 63~.

14. In our case, since the main variables analyzed are ratios to nominal GNP and are thus strictly bounded, we assume a priori that these variables cannot trend.

15. Our method is similar to Roberds and Whiteman |1992~.

16. See Engle and Granger |1987~ and DeJong |1992~.

17. See Campbell and Perron |1991~. We thank a referee for valuable suggestions on these issues.

18. See Sims |1980~ and Litterman and Weiss |1985~ for more detail on the techniques and usefulness of variance decompositions.

19. Previous work in the literature on government and trade balances perform VARs without explicitly examining the significance of the results.

20. The normal approximation method for generating posterior probability bounds does not restrict the variance decompositions above zero.

21. We generated results using an ordering reverse from our primary specification. The results are qualitatively the same as those from the primary model reported in Table II. GBAL now explains 44 percent of TBAL's forecast variance (full results available on request).

22. The appendix reports results for a system that uses the Bureau of Economic Analysis's cyclically adjusted federal deficit series. Recessions could mask an underlying relationship between the deficits by increasing the actual government deficit while possibly reducing imports and thus the trade deficit. Despite this, our results using the cyclically adjusted series in the appendix are qualitatively similar to those reported above for the two other deficit measures. The impulse response function showing the impact of (a cyclically adjusted) GBAL on TBAL closely mimics that in Figure 1 from the primary model.

23. As a further test of robustness, we estimate the system over a shortened sample comprising only the floating rate era commencing in 1971. Results (available on request) are qualitatively similar to those for the full sample used in this paper. In the variance decompositions, slight quantitative differences exist. However, the impulse response functions for the major relationships under study (government fiscal balances on the trade balance and exchange rate) display nearly identical patterns in both the shortened and full data sample.

24. The absence of significant relationships when using differenced data is robust to the use of the NIPA as well as the cyclically adjusted fiscal deficit.

25. This lagged negative relationship from exchange rates to trade balances appears in most trade models. Thus, this empirical finding may be more useful as a diagnostic for our methodology than as specific support for any given theory.

26. We have impulse response functions that suggest the data do not support the Ricardian notion that the government purchases variable (GPURCH) directly affects the trade balance. There appears to be no significant relationship from innovations in GPURCH to TBAL. Further, another impulse response function shows that innovations in TBAL are not associated with significant subsequent movements in GBAL. Thus, the results are not consistent with a direction of causality that is the reverse of the one implied by the twin deficits story. These graphs are available on request.

REFERENCES

Abell, John D. "Twin Deficits During the 1980s: An Empirical Investigation." Journal of Macroeconomics, Winter 1990, 81-96.

Bernheim, B. Douglas. "Budget Deficits and the Balance of Trade," in Tax Policy and the Economy. Cambridge: MIT Press, 1988, 1-31.

Branson, William H. "Causes of Appreciation and Volatility of the Dollar," in The U.S. Dollar--Recent Developments, Outlook, and Policy Options, Federal Reserve Bank of Kansas City Conference Proceedings, 1985, 33-52.

-----. "Sources of Misalignment in the 1980s," in Misalignment of Exchange Rates: Effects of Trade and Industry, edited by Richard C. Marston. Chicago: University of Chicago Press, 1988, 9-38.

Campbell, John Y., and Pierre Perron. "Pitfalls and Opportunities: What Macroeconomists Should Know About Unit Roots," in National Bureau of Economic Research Macroeconomics Annual 1991, edited by Olivier J. Blanchard and Stanley Fischer. Cambridge: MIT Press, 1991, 141-201.

Darrat, Ali F. "Have Large Budget Deficits Caused Rising Trade Deficits?" Southern Economic Journal, April 1988, 879-87.

DeJong, David N. "Co-integration and Trend-Stationarity in Macroeconomic Time Series: Evidence from the Likelihood Function." Journal of Econometrics, December 1992, 347-70.

DeJong, David N., and Charles H. Whiteman. "Trends and Cycles as Unobserved Components in U.S. Real GNP: A Bayesian Perspective." Proceedings of the American Statistical Association, Business and Economic Statistics Section, August 1989, 63-70.

Dewald, William G., and Michael Ulan. "The Twin-Deficit Illusion." Cato Journal, Winter 1990, 689-707.

Enders, Walter, and Bong-Soo Lee. "Current Account and Budget Deficits: Twins or Distant Cousins?" Review of Economics and Statistics, August 1990, 373-81.

Engle, Robert F., and C. W. J. Granger. "Co-integration and Error Correction: Representation, Estimation, and Testing." Econometrica, March 1987, 251-76.

Evans, Paul. "Is the Dollar High Because of Large Budget Deficits?" Journal of Monetary Economics, November 1986, 227-49.

-----. "Do Budget Deficits Affect the Current Account?" Working Paper, Ohio State University, 1989.

Feldstein, Martin. "The Budget Deficit and the Dollar," in National Bureau of Economic Research Macroeconomics Annual 1986, edited by Stanley Fischer. Cambridge: MIT Press, 1986, 355-92.

Hooper, Peter, and Catherine Mann. "The Emergence and Persistence of the U.S. External Imbalance." Princeton Studies in International Finance No. 65, October 1989.

Litterman, Robert B., and Laurence Weiss. "Money, Real Interest Rates, and Output: A Reinterpretation of Postwar U.S. Data." Econometrica, January 1985, 129-56.

Miller, Stephen M., and Frank S. Russek. "Are the Deficits Really Related?" Contemporary Policy Issues, October 1989, 91-115.

Roberds, William, and Charles H. Whiteman. "Monetary Aggregates as Monetary Targets: A Statistical Investigation." Journal of Money, Credit, and Banking, May 1992, 142-61.

Runkle, David E. "Vector Autoregressions and Reality." Journal of Business and Economic Statistics, October 1987, 437-42.

Schmidt, Peter. "Dickey-Fuller Tests with Drift." Michigan State University Working Paper No. 8717, June 1988.

Sims, Christopher A. "Macroeconomics and Reality." Econometrica, January 1980, 1-48.

-----. "Vector Autoregressions and Reality: Comment." Journal of Business and Economic Statistics, October 1987, 443-51.

-----. "Bayesian Skepticism on Unit Root Econometrics." Discussion Paper No. 3, Institute for Empirical Macroeconomics, Federal Reserve Bank of Minneapolis and University of Minnesota, May 1988.

U.S. Department of Commerce. Survey of Current Business, various issues, 1973-1990.

Watson, Mark W. "Comment: On Vector Autoregressions and Reality." Journal of Business and Economic Statistics, October 1987, 451-53.

Willett, Thomas D., and Clas G. Wihlborg. "International Capital Flows, the Dollar, and U.S. Financial Policies," in Monetary Policy for a Volatile Global Economy, edited by William S. Haraf and Thomas D. Willett. Lanham: American Enterprise Institute Press, 1990, 51-88.

Jeffrey A. Rosensweig and Ellis W. Tallman, Assistant Professor of Finance, Emory University Business School and Economist, Federal Reserve Bank of Atlanta, respectively. The contents of this paper do not reflect the views of the Federal Reserve System or of the Federal Reserve Bank of Atlanta. We are particularly indebted to William Roberds. We thank the participants at the WEA International 1990 Annual Conference and Penn State University macro workshop for useful comments. We also thank Shaghil Ahmed, George Benston, Joseph Gagnon, Curt Hunter, Chuck Whiteman, the editor (Thomas Willett), and two anonymous referees for helpful criticism. The authors are fully responsible for any errors contained in this paper. Rosensweig acknowledges support from the Emory University Research Committee.

The unprecedented increase of both the fiscal and trade deficits in the United States during the 1980s led many observers to posit a causal linkage. A sizable contingent of economists believe that the "twin deficits" concept aptly characterizes the recent relationship between federal budgetary policy and the U.S. trade balance. In this view, increased federal fiscal deficits during the 1980s were a primary cause of the massive U.S. trade deficits experienced as the decade unfolded. However, this posited causal relationship does not command universal acceptance. In this paper, we examine data on U.S. trade and federal fiscal measures along with exchange and interest rates to uncover the empirical relationship between them and to discover whether fiscal policy plays an essential role in the adjustment of the massive U.S. trade imbalance.

The twin deficits concept asserts that the deficits are "twins" because they are related in the following manner. Increased government deficits are thought to put incipient upward pressure on real interest rates. This induces incipient capital in-flows that lift or appreciate the real exchange value of the currency, which erodes competitiveness and, after a time lag, results in increased trade deficits.

A number of recent papers examine the relationship between budget and trade deficits. Each contributes important insights, but nothing approaching a consensus has yet emerged. Miller and Russek |1989~ find evidence of a secular relationship between the deficits under two of the three statistical techniques they employ. Bernheim |1988~ concludes that the government deficit is a prime determinant of U.S. trade deficits, whereas Evans |1989~, Enders and Lee |1990~, and Dewald and Ulan |1990~ find no such causal impact of fiscal deficits. Darrat |1988~ and Abell |1990~ report somewhat mixed results.

The conflicting results in the existing literature stem from wide differences in the empirical techniques, data measures and samples employed. Further, we have found in the literature and in our work reported below that the use of differenced data versus data in level form has a major impact on the results produced and inferences drawn. Recent empirical literature suggests that stationarity of data is a crucial element in model specification. Thus, the question of whether or not to difference the data should be addressed before estimation and inference.

Classical inference about stationarity of data suffers from what Sims |1988~ refers to as a "discontinuity" in the asymptotic theory. Hence, the classical tests used in much of the previous literature have low power to reject the null hypothesis of nonstationarity in levels. Bayesian inference does not suffer from this problem; therefore, we generate posterior probabilities using Monte Carlo integration that estimate probabilities that suggest whether the data are better analyzed in level or difference form. Despite somewhat contrasting results, the classical augmented Dickey-Fuller tests and the Bayesian probability estimates both suggest that a difference specification is inappropriate.

We analyze empirically a five-variable nonstructural vector autoregressive (VAR) model. The model includes both a federal government balance and a government purchases measure to summarize fiscal policy. Along with these two variables and a trade balance measure, we constructed real interest rate and real exchange rate variables spanning a quarterly data sample from 1961 to 1989. We estimate the model in both levels and difference form. Using the level specification, we find strong support for the twin deficit notion that government deficits are a cause of trade deficits, whereas in differences we find no such relationship. Furthermore, the results from our Monte Carlo work imply that the model specified in levels is likely most appropriate for estimation and inference.

II. CONTRASTING THEORETICAL FRAMEWORKS

The presence of both deficits in a variant of the basic savings/investment identity may provide a useful reference, as in Bernheim |1988~ and Branson |1985~, for discussing the hypothesized causal relationship embodied in popular twin deficits stories:

(1) TDEF = GDEF + (I - |S.sub.p~),

where TDEF is the broad trade deficit, GDEF is the government budget deficit, and I and |S.sub.p~ are private sector investment and savings, respectively.(1) However, the presence of both deficits in this identity does not imply that the movements in one deficit cause commensurate changes in the other deficit because the term (I-|S.sub.p~) can change as well.

The above identity provides at best casual motivation for the twin deficits relationship. However, the twin deficits story receives some theoretical justification from various frameworks, including the traditional Mundell-Fleming model.(2) The Mundell-Fleming model assumes that an increased government deficit policy (for example, a tax cut) raises aggregate demand (essentially the fiscal shock is not fully offset by increased private savings).

This increased demand places incipient upward pressure on the domestic real interest rate. This interest rate pressure induces an incipient net capital inflow from abroad, appreciating the domestic currency's real foreign exchange value.(3) After a lag, this relatively more expensive domestic currency pushes the nation's trade balance towards deficit. Thus, this model implies that increased government deficits ultimately widen the trade deficit (under both fixed and flexible exchange rate regimes, although the transmission mechanisms differ).(4)

This model is not universally accepted. A key reason is that there exists a major controversy on the link between government deficits and exchange rates. Branson |1988~ suggests that increased U.S. government deficits are the prime determinant of the dollar's appreciation. Feldstein |1986~ also supports a linkage whereby increased government deficits appreciate the dollar. Conversely, Evans |1986~ finds evidence that increased deficits depreciate the dollar.(5)

Unlike the Mundell-Fleming model, one major school of thought expects no causal impact from government budget deficits to trade deficits. A model incorporating Ricardian Equivalence would suggest that a substitution of debt for taxes by the government that increases the fiscal deficit would be offset by increased private savings, rather than increased net foreign borrowing (trade deficit). However, changes in (temporary) government spending behavior can affect trade deficits in this framework. Thus, we include a measure of federal government purchases in our empirical study to distinguish between government purchases and government balances as the potential sources of the twin deficits correlation.

Our five-variable model, by including two measures of fiscal policy, should help us distinguish between a Mundell-Fleming and a Ricardian interpretation of the role of fiscal policy in trade adjustment. The former approach argues that the fiscal balance affects subsequent trade balances, whereas the latter implies that (temporary) government purchases, not government balances, impact the trade balance.

Given these contrasting views and the mixed results from prior research, our study aims to provide additional empirical evidence on the relationship between fiscal deficits, the dollar, and trade deficits.

III. DATA SAMPLE AND EMPIRICAL TECHNIQUES

Our data selection follows directly from the two contrasting theoretical models summarized above. Our model includes the five variables that span the two theoretical frameworks summarized above. We examine an extensive sample (1961:I to 1989:IV) that includes quarterly data points from the recent record-sized federal budget and U.S. trade deficits period, as well as recent apparent turning points in these series.(6) We construct the federal deficit measure as the change in the real par value of privately held federal debt. Our constructed deficit measure accounts for the inflation tax on government debt by measuring the change in the real value of the debt before renominalizing it.(7,8) We include federal government purchases as the companion fiscal policy measure. Our trade balance measure is net foreign investment (NFI), the series consistent with NIPA accounting.

Theoretical considerations suggest using real rather than nominal variables for analysis. By constructing ratios to nominal GNP, we have created trade balance (TBAL) and federal government fiscal policy variables (GBAL for federal balances and GPURCH for federal purchases) that are essentially real measures. We also constructed real interest and real exchange rate series for our sample. The real interest rate (IRATE) is the difference between the nominal rate on a three-month Treasury bill and the ex post CPI inflation rate. We constructed a multilateral trade-weighted real exchange rate index of the U.S. dollar's foreign exchange value (ERATE). We paralleled the standard Federal Reserve Board index, using its weights comprising ten major currencies, but constructed a real index by multiplying the nominal trade-weighted value of the dollar by the U.S. CPI divided by the trade-weighted foreign CPI. Increases in this index signify a real appreciation of the dollar.

The theoretical frameworks discussed above do not imply a specific dynamic structure. Consequently, we choose an empirical methodology that places minimal restrictions on the explicit structure of the relationship. The VAR techniques we employ provide a convenient estimation procedure that investigates data relationships while imposing little structure.(9)

Our VAR system employs the five variables suggested by the two theoretical frameworks: GPURCH, GBAL, TBAL, IRATE, and ERATE. Prior to estimating the system, we determine the appropriate lag length with likelihood ratio tests for models using levels as well as first differences. The recent literature on U.S. trade adjustment, for example Hooper and Mann |1989~, finds that trade balances react to factors such as exchange rate fluctuations with lags that span years rather than quarters. This prior evidence motivated us to test a twelve- versus a six-quarter lag VAR specification, as opposed to any shorter lag structures. Test results suggested that we could not reject the null hypothesis that six lags are sufficient.(10)

Much attention in recent macroeconomic empirical literature focuses on questions regarding the stationarity of time-series data. Nonstationary or integrated data series (i.e., those containing unit roots) render statistical inference from traditional output invalid. Classical statistical methods, such as augmented Dickey-Fuller tests, are used most often to detect the possible existence of unit roots in data series. Table I reports results using these techniques on the data series. At the 5 percent significance level, we could reject the null hypothesis of nonstationarity for only one of the five variables in level form.(11) Analyzing first-differenced data, test statistics imply rejection at 5 percent significance of the null hypothesis of nonstationarity for all five series. Overall, using the 5 percent significance level, the results suggest that four of the five variables are integrated of order one and the remaining variable is level stationary.(12) To justify a first-difference specification, there should be evidence of five unit roots in the system. Thus, at the 5 percent significance level, the classical Augmented Dickey-Fuller tests do not dictate a first-difference specification.

TABLE I Stationarity Tests for Individual Series Augmented Dickey-Fuller Tests Variable Levels First Differences GBAL (constructed) -2.702 -5.432(***) GBAL (NIPA) -2.162 -5.133(***) IRATE -2.391 -5.811(***) ERATE -2.132 @(**) -3.816(***) TBAL -1.554 -3.948(***) G PURCH -1.853 -4.043(***) The numbers above are augmented Dickey-Fuller statistics, including four lags of the (differenced) dependent variable and a constant. Autocorrelation tests suggest that our specification is sufficient to remove residual autocorrelation. Inferences are made based on the Schmidt |1988~ critical values where applicable (denoted by @). In the case of ERATE, a s ignificant drift term required using Schmidt critical values (for a standardized drift of approximately 1.3 with nearly 100 observations, the 5 percent critical value is roughly 1.8). Significant test statistics indicate rejection of the null hypothesis of nonstationarity. For each table: * denotes significance at the 10 percent level. ** denotes significance at the 5 percent level. *** denotes significance at the 1 percent level.

Inferences based on classical statistical methods, however, have recently been challenged. A growing body of literature indicates that: "the power of Dickey-Fuller type integration tests can be quite low (less than 50 percent) against plausible trend stationary alternatives."(13) Hence, a number of researchers, notably Sims |1988~, have suggested approaching this issue from a Bayesian perspective, asking the question: which data representation, difference or level stationarity, is more probable given the data?(14) Following this suggestion, we use Monte Carlo integration of the VAR specification in levels to generate posterior probabilities, which indicate the probability that the system is better specified as level or difference stationary.(15)

The Monte Carlo strategy for generating the posterior probability odds ratios involves relatively few steps but numerous iterations of the main procedures. Essentially, we generate a sequence of independent drawings from the posterior distribution of the parameters. Before we begin the iterations, we estimate the model in level form to get the posterior means of the model parameters. By assuming a diffuse prior on the model coefficients and the variance-covariance matrix of the model error terms, the posterior distribution of the variance-covariance matrix is an inverted Wishart distribution, and the posterior distribution of the coefficients is normal, conditional on the variance-covariance matrix of the model error terms. For each iteration of the integration procedure, we make a draw of the variance-covariance matrix, and use this draw to specify the conditional normal distribution for the parameters of interest. From this conditional distribution, we make a drawing of the parameter matrix and determine its roots.

The next step requires that we distinguish between roots that are unit or greater (implying nonstationarity) and those that are close to but less than one (implying level stationarity). The iterations performed using Monte Carlo integration are divided into sets of observations: for example, draws consistent with level stationarity versus those consistent with nonstationarity in levels. The number of draws consistent with level stationarity relative to the total number of simulations represents an estimate of the probability that the model is level stationary. Conversely, the relative frequency of iterations with one or more roots greater than or equal to unity reflects an estimate of the posterior probability that the model is nonstationary in levels. After 20,000 iterations of the program, we find a 64.6 percent posterior probability that the model is stationary in levels, implying a 35.4 percent probability that the largest root is unit or greater (implying nonstationarity).

These probability estimates provide some evidence that unit roots may exist in this system. The model would need at least one unit root in order to have cointegration of the series, and there is a nontrivial probability of one unit root.(16) Yet, for our purposes, the relevant issue is whether there is evidence supporting the existence of five unit roots, which would indicate difference stationarity. In our Monte Carlo work, the estimate of the probability that the system has five unit roots is only .00125. In contrast to the results from the classical tests that find four unit roots, the estimated probability of four or more unit roots in the system is less than 1 percent. Thus, we find Monte Carlo evidence against a difference stationary specification.

The results of both diagnostic techniques suggest possible cointegration of the series; however, this does not invalidate an unrestricted level VAR specification. Although an unrestricted level specification may be inefficient, by ignoring the long-run restrictions that cointegration imposes, the estimation is consistent. On the other hand, an unrestricted VAR in first differences when the true model is cointegrated is misspecified because it ignores lagged equilibrium errors as regressors, which renders it inconsistent.(17) In applications with limited data samples, cointegrating relationships are hard to identify precisely. If restrictions implied by cointegration are misspecified, the estimation is inconsistent. Furthermore, Watson |1987~ finds that cointegration restrictions at best add marginally to the interpretation of the results.

Results from an unrestricted level specification of the system are robust to the potential existence of cointegration, and are least likely to be subject to misspecification error if there are less than five unit roots. The posterior probabilities generated from the Monte Carlo integration indicate that the model likely is best specified in level form. Thus, our primary set of results are from the level specification. We also report a secondary set of results, those from a model specified in first differences. We perform this exercise to emphasize how the choice of a level or a first-difference specification affects the empirical evidence and the resulting inferences.

IV.EMPIRICAL RESULTS

Our primary specification employs levels of the five variables. We examine decompositions of the forecast variance and analyze the relative contribution of orthogonalized innovations associated with each variable in the system toward explaining the forecast variance of each series.(18) We derive the orthogonalized innovations from a Choleski decomposition in which we specify the causal ordering of the variables in a recursive structure.

Our primary results, reported in Table II, were generated with an ordering that is the reverse of the causal chain in the Mundell-Fleming model to ensure that the choice of ordering does not favor the finding of "twin deficits." The ordering runs from trade balance (TBAL) to exchange rate (ERATE) to interest rates (IRATE) to government balances (GBAL) to government purchases (GPURCH). The variable GBAL in the primary specification is our constructed (from changes in real debt) government balance as a ratio to GNP.

The results from the primary five-variable VAR in levels show the variance decompositions at a sixteen-quarter forecast horizon. We employ a method similar to that developed by Runkle |1987~ to generate standard errors for our system and produce 90 percent posterior probability ranges for each decomposition.(19) These ranges are listed below the posterior mean variance decompositions in Table II (and all subsequent tables). Despite the ability to generate posterior probability ranges, there is no consensus on explicit criteria for significance in VAR analysis. For example, Runkle |1987~ argues that a useful criterion for significance is a probability range completely above 10 TABULAR DATA OMITTED percent in variance decompositions, but 10 percent is an arbitrary threshold. On the other hand, Sims |1987~ refers to the impulse response functions and attributes significance to whether responses to an innovation are large and bounded away from zero. Consequently, we present both forms of evidence to investigate significance.

Variance Decompositions

The notable result in Table II is that government balance (GBAL) innovations are associated with 42.2 percent of broad trade balance (TBAL) variance at a sixteen-quarter horizon. In addition, the 90 percent posterior probability range far exceeds 10 percent, satisfying Runkle's criterion. The result is consistent with the twin deficits story, implying a significant impact of fiscal balance innovations on the trade balance, but we require impulse response function results to determine the sign of the interaction. In contrast, TBAL innovations are associated with only 8.0 percent of GBAL variance, and the probability range includes zero.(20) Thus, the evidence is consistent with the twin deficits notion that the direction of causality runs from government fiscal balances to the trade balance.

The postulated relationship between the trade and fiscal balances is the primary element of the twin deficits story. The variance decompositions show that, except for the impact of GBAL on TBAL, each variable explains an insignificant proportion (using Runkle's criterion) of any other variable's variance.

Table II shows that innovations in GBAL, but not in federal purchases (GPURCH), provide significant predictive power for future trade balances. This distinction is consistent with a twin deficit or Mundell-Fleming, as opposed to Ricardian, notion of the impact of fiscal policy on the trade balance.

We examine the robustness of these results with regard to the chosen measure of the fiscal deficit variable and with regard to the ordering of the variables in the Choleski decomposition.(21) Table III reports results for a model that differs from the primary one by substituting the conventionally measured NIPA federal government deficit series for the nominalized 'real' deficit that we constructed from changes in real federal debt. A comparison of Tables II and III indicates that our results are qualitatively similar for either measure of GBAL.(22) The major finding that the only significant explanatory linkage is the one running from GBAL innovations to TBAL appears robust to the particular measure of government balances employed.(23)

All specifications using the data in level form yield similar results. To highlight the role of differencing data, a factor that appears crucial in the inconclusive existing literature, Table IV presents evidence from a VAR specified in first differences. Differencing clearly has a profound effect, because Table IV shows that in this specification no variable's innovations explain a significant proportion of any other variable's variance. Specifically, the impact of GBAL innovations on TBAL falls to 12.4 percent. No variance decompositions satisfy the Runkle significance criterion. The results in Table IV provide little support of the twin deficits notion.(24) Consequently, the contrast between Table IV and Table II emphasizes the importance of evidence against the first-difference specification provided by the Bayesian posterior probability odds estimates as well as the classical augmented Dickey-Fuller tests.

Impulse Response Functions. Variance decompositions do not indicate the sign and dynamic pattern of the relationship between the variables of interest. However, theory postulates unambiguous signs for the impulse response functions of the variables. To examine the dynamic pattern and sign of impacts, we present the impulse response functions over a sixteen-quarter horizon. We generate 90 percent posterior probability bands for the impulse responses in a further attempt to assess significance, following the suggestion of Sims |1987~. The figures show impulse response functions generated from our primary specification in levels.

Figure 1 shows that a positive shock to the federal government balance to GNP ratio ultimately leads to unambiguous increases in the trade balance to GNP ratio. The probability bounds are clearly above TABULAR DATA OMITTED zero for the entire period four to sixteen quarters beyond the shock. This impulse response result, in addition to the variance decompositions, provides strong support for the central necessary relationship in the twin deficits story.

Given the support for the twin deficits story, we investigate other proposed linkages as diagnostics to be sure that the interactions among variables have the predicted sign. We find in Figure 2 that the innovation in GBAL has a negative effect on the real exchange rate (ERATE). The effect is immediate and the response function is bounded away from zero for the first ten periods. This result is robust across alternative fiscal measures and sample periods. Thus, budget deficits are associated with an appreciating real currency, consistent with a twin deficit story. This result adds evidence to the controversial issue regarding the impact of U.S. fiscal deficits on the dollar.

In Figure 3, we find further evidence consistent with the twin deficit story. Innovations in the real exchange value of the dollar (ERATE) are associated with subsequent declines in the trade balance (TBAL).(25) The negative relationship is bounded away from zero for a period from eight to fifteen quarters out, depicting a TABULAR DATA OMITTED lag similar to that found in the recent literature on the link between trade balances and the dollar, notably Hooper and Mann |1989~.(26)

In sum, the impulse response functions and the variance decompositions for the level specifications provide evidence consistent with the predictive role of government deficits in the typical twin deficits explanation.

V. SUMMARY AND IMPLICATIONS

Spurred by the dramatic surge in both U.S. trade and fiscal deficits during the 1980s, a particular causal story ("twin deficits") asserting a relationship whereby government deficits lead to trade deficits gained many adherents. Previous research investigating the empirical relevance of the twin deficits story has not produced a consensus. We add to this literature by testing for the various possible relationships between fiscal policy, exchange rates, and trade adjustment, employing VAR techniques with relatively few prior restrictions.

The relationship between fiscal and trade variables reported in the existing empirical literature appears crucially dependent on whether the model is specified in levels or first differences. We use posterior probability estimates generated by Monte Carlo methods to determine whether the system is best specified in levels. We find, despite a slight contrast between inferences from classical Dickey-Fuller tests and the Monte Carlo methods, that the VAR system is probably best specified in levels. For completeness and comparison purposes, we estimate the VAR system in both level and difference specifications. The level specification generates results that are consistent with the twin deficits notion embedded in a Mundell-Fleming framework. Overall, the results suggest that increased U.S. government deficits may have contributed to dollar appreciation and the large U.S. trade deficits in the 1980s. Thus, fiscal policy appears to have a significant role to play in U.S. trade balance adjustment.

1. We employ net foreign investment (NFI) from the National Income and Product Accounts (NIPA) as our trade balance measure, and a federal balance series constructed from federal debt data (described below) as the government balance measure. A few studies employ the current account from the Balance of Payments accounts as the trade balance measure. Although this is very close to the net foreign investment measure of the trade balance that we use, historic U.S. current account data unfortunately include capital gains and losses reflecting the foreign exchange translation of direct foreign investment. Thus, we use net foreign investment as our trade balance measure, consistent with NIPA accounting.

2. For a clear and extensive discussion of the Mundell-Fleming framework as well as other theoretical approaches, see Miller and Russek |1989~. Branson |1985; 1988~ offers an alternative framework for this linkage.

3. The real appreciation stems from an appreciation in the nominal exchange rate in a flexible exchange rate regime, and from an increased domestic price level under fixed rates. The appreciation relies on an assumption in the basic model of sufficient capital mobility to outweigh direct domestic demand effects worsening the current account.

4. Under a fixed rate regime, the trade deficit deteriorates because of an income effect as well as the real appreciation.

5. For a clear and comprehensive discussion of this controversial linkage, highlighting the role played by the degree of capital mobility, see Willett and Wihlborg |1990~.

6. The sample is the longest continuous one we could create--our starting date was limited by the price data for the real exchange rate series we constructed. The ending dates are limited by the various measures of the fiscal deficits.

7. We thank W. Michael Cox for providing the series on the par value of privately held federal debt. To construct the government balance series, we use the following identity. Normalized real government balances (GBAL) are

GBAL = - |P.sub.t~|(|D.sub.t~/|P.sub.t~) - (|D.sub.t-1~/|P.sub.t-1~)~,

where P is the consumer price index and D is the par value of privately held federal debt outstanding. We thank a referee for this suggestion.

8. We estimate a system that employs the standard NIPA nominal federal deficit measure as well as one that includes a cyclically adjusted federal government deficit series (provided by the Bureau of Economic Analysis based on mid-expansion GNP trend). Results from these systems allow us to compare them with some of the existing literature. The results from both systems are qualitatively similar to those in our primary specification, indicating robustness.

9. The Choleski decomposition of the variance-covariance matrix does impose a recursive structure on the residuals. However, we have examined various alternative orderings, and find qualitatively similar results (available on request).

10. The likelihood ratio test generated a |X.sup.2~(150) of 156.86 for the level specification, and 137.13 for the difference specification. At the 5 percent significance level, the test statistic must exceed the critical value of 179.58. Thus, neither test could reject the null.

11. On theoretical grounds, we did not include a trend in the augmented Dickey-Fuller tests. Because the trade and government variables are measured as ratios to nominal GNP, we should not expect unbounded trending of the series. We reject the nonstationarity hypothesis for ERATE due to the presence of a significant drift term that necessitated using the critical values in Schmidt |1988~.

12. Note, however, the choice of a 1 percent significance level would imply difference stationarity. The potential ambiguity of the implications from the augmented Dickey-Fuller tests provide further motivation for exploring an alternative diagnostic procedure.

13. DeJong and Whiteman |1989, 63~.

14. In our case, since the main variables analyzed are ratios to nominal GNP and are thus strictly bounded, we assume a priori that these variables cannot trend.

15. Our method is similar to Roberds and Whiteman |1992~.

16. See Engle and Granger |1987~ and DeJong |1992~.

17. See Campbell and Perron |1991~. We thank a referee for valuable suggestions on these issues.

18. See Sims |1980~ and Litterman and Weiss |1985~ for more detail on the techniques and usefulness of variance decompositions.

19. Previous work in the literature on government and trade balances perform VARs without explicitly examining the significance of the results.

20. The normal approximation method for generating posterior probability bounds does not restrict the variance decompositions above zero.

21. We generated results using an ordering reverse from our primary specification. The results are qualitatively the same as those from the primary model reported in Table II. GBAL now explains 44 percent of TBAL's forecast variance (full results available on request).

22. The appendix reports results for a system that uses the Bureau of Economic Analysis's cyclically adjusted federal deficit series. Recessions could mask an underlying relationship between the deficits by increasing the actual government deficit while possibly reducing imports and thus the trade deficit. Despite this, our results using the cyclically adjusted series in the appendix are qualitatively similar to those reported above for the two other deficit measures. The impulse response function showing the impact of (a cyclically adjusted) GBAL on TBAL closely mimics that in Figure 1 from the primary model.

23. As a further test of robustness, we estimate the system over a shortened sample comprising only the floating rate era commencing in 1971. Results (available on request) are qualitatively similar to those for the full sample used in this paper. In the variance decompositions, slight quantitative differences exist. However, the impulse response functions for the major relationships under study (government fiscal balances on the trade balance and exchange rate) display nearly identical patterns in both the shortened and full data sample.

24. The absence of significant relationships when using differenced data is robust to the use of the NIPA as well as the cyclically adjusted fiscal deficit.

25. This lagged negative relationship from exchange rates to trade balances appears in most trade models. Thus, this empirical finding may be more useful as a diagnostic for our methodology than as specific support for any given theory.

26. We have impulse response functions that suggest the data do not support the Ricardian notion that the government purchases variable (GPURCH) directly affects the trade balance. There appears to be no significant relationship from innovations in GPURCH to TBAL. Further, another impulse response function shows that innovations in TBAL are not associated with significant subsequent movements in GBAL. Thus, the results are not consistent with a direction of causality that is the reverse of the one implied by the twin deficits story. These graphs are available on request.

REFERENCES

Abell, John D. "Twin Deficits During the 1980s: An Empirical Investigation." Journal of Macroeconomics, Winter 1990, 81-96.

Bernheim, B. Douglas. "Budget Deficits and the Balance of Trade," in Tax Policy and the Economy. Cambridge: MIT Press, 1988, 1-31.

Branson, William H. "Causes of Appreciation and Volatility of the Dollar," in The U.S. Dollar--Recent Developments, Outlook, and Policy Options, Federal Reserve Bank of Kansas City Conference Proceedings, 1985, 33-52.

-----. "Sources of Misalignment in the 1980s," in Misalignment of Exchange Rates: Effects of Trade and Industry, edited by Richard C. Marston. Chicago: University of Chicago Press, 1988, 9-38.

Campbell, John Y., and Pierre Perron. "Pitfalls and Opportunities: What Macroeconomists Should Know About Unit Roots," in National Bureau of Economic Research Macroeconomics Annual 1991, edited by Olivier J. Blanchard and Stanley Fischer. Cambridge: MIT Press, 1991, 141-201.

Darrat, Ali F. "Have Large Budget Deficits Caused Rising Trade Deficits?" Southern Economic Journal, April 1988, 879-87.

DeJong, David N. "Co-integration and Trend-Stationarity in Macroeconomic Time Series: Evidence from the Likelihood Function." Journal of Econometrics, December 1992, 347-70.

DeJong, David N., and Charles H. Whiteman. "Trends and Cycles as Unobserved Components in U.S. Real GNP: A Bayesian Perspective." Proceedings of the American Statistical Association, Business and Economic Statistics Section, August 1989, 63-70.

Dewald, William G., and Michael Ulan. "The Twin-Deficit Illusion." Cato Journal, Winter 1990, 689-707.

Enders, Walter, and Bong-Soo Lee. "Current Account and Budget Deficits: Twins or Distant Cousins?" Review of Economics and Statistics, August 1990, 373-81.

Engle, Robert F., and C. W. J. Granger. "Co-integration and Error Correction: Representation, Estimation, and Testing." Econometrica, March 1987, 251-76.

Evans, Paul. "Is the Dollar High Because of Large Budget Deficits?" Journal of Monetary Economics, November 1986, 227-49.

-----. "Do Budget Deficits Affect the Current Account?" Working Paper, Ohio State University, 1989.

Feldstein, Martin. "The Budget Deficit and the Dollar," in National Bureau of Economic Research Macroeconomics Annual 1986, edited by Stanley Fischer. Cambridge: MIT Press, 1986, 355-92.

Hooper, Peter, and Catherine Mann. "The Emergence and Persistence of the U.S. External Imbalance." Princeton Studies in International Finance No. 65, October 1989.

Litterman, Robert B., and Laurence Weiss. "Money, Real Interest Rates, and Output: A Reinterpretation of Postwar U.S. Data." Econometrica, January 1985, 129-56.

Miller, Stephen M., and Frank S. Russek. "Are the Deficits Really Related?" Contemporary Policy Issues, October 1989, 91-115.

Roberds, William, and Charles H. Whiteman. "Monetary Aggregates as Monetary Targets: A Statistical Investigation." Journal of Money, Credit, and Banking, May 1992, 142-61.

Runkle, David E. "Vector Autoregressions and Reality." Journal of Business and Economic Statistics, October 1987, 437-42.

Schmidt, Peter. "Dickey-Fuller Tests with Drift." Michigan State University Working Paper No. 8717, June 1988.

Sims, Christopher A. "Macroeconomics and Reality." Econometrica, January 1980, 1-48.

-----. "Vector Autoregressions and Reality: Comment." Journal of Business and Economic Statistics, October 1987, 443-51.

-----. "Bayesian Skepticism on Unit Root Econometrics." Discussion Paper No. 3, Institute for Empirical Macroeconomics, Federal Reserve Bank of Minneapolis and University of Minnesota, May 1988.

U.S. Department of Commerce. Survey of Current Business, various issues, 1973-1990.

Watson, Mark W. "Comment: On Vector Autoregressions and Reality." Journal of Business and Economic Statistics, October 1987, 451-53.

Willett, Thomas D., and Clas G. Wihlborg. "International Capital Flows, the Dollar, and U.S. Financial Policies," in Monetary Policy for a Volatile Global Economy, edited by William S. Haraf and Thomas D. Willett. Lanham: American Enterprise Institute Press, 1990, 51-88.

Jeffrey A. Rosensweig and Ellis W. Tallman, Assistant Professor of Finance, Emory University Business School and Economist, Federal Reserve Bank of Atlanta, respectively. The contents of this paper do not reflect the views of the Federal Reserve System or of the Federal Reserve Bank of Atlanta. We are particularly indebted to William Roberds. We thank the participants at the WEA International 1990 Annual Conference and Penn State University macro workshop for useful comments. We also thank Shaghil Ahmed, George Benston, Joseph Gagnon, Curt Hunter, Chuck Whiteman, the editor (Thomas Willett), and two anonymous referees for helpful criticism. The authors are fully responsible for any errors contained in this paper. Rosensweig acknowledges support from the Emory University Research Committee.

Printer friendly Cite/link Email Feedback | |

Author: | Rosenssweig, Jeffrey A.; Tallman, Ellis W. |
---|---|

Publication: | Economic Inquiry |

Date: | Oct 1, 1993 |

Words: | 5826 |

Previous Article: | Discount window borrowing and Federal Reserve operating regimes. |

Next Article: | Toward an economic theory of fashion. |

Topics: |