The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Vol. 25, No. 3 ECONOMIC REVIEW FEDERAL RESERVE BANK OF CLEVELAND 1989 Quarter 3 Vol. 25. No. 3 The Stability of Money Demand, Its Interest 2 Sensitivity, and Some Economic Review Implications for Money as a Policy Guide q uarterly b y the Research is published D e p a rtm e n t o f the Federal R e s e rve B a n k o f C le ve la n d . b y Jo h n B. C arlson C opie s of the Review are availa ble through our Public One common finding of recent empirical research in monetary economics is A ffa irs an d B a n k Relations that the interest elasticity of money demand is substantial, and higher than D e p a rtm e n t, 2 1 6 / 5 7 9 -2 1 5 7 . many economists previously thought. The evidence seems strongest for M 1 demand in the long run. While interest rates appear to have little or no longrun effect on M 2 demand, the short-run interest elasticity seems higher than previously thought. This paper examines the recent findings on money C oordin ating Ec o n o m is t: Randall W . E b e rts demand and discusses their implications for monetary policy and rules. Ed ito rs : Paul J . N ick els Robin Ratliff D esign : M ich ae l G a lk a T y p o g ra p h y : L iz H a n n a Accounting for the Recent 14 Divergence in Regional Wage Differentials O pin ion s s ta te d in Review b y R andall W . Eberts Economic are th o s e o f the a u th ors an d n ot necessarily those of the Federal R e s e rve A fter converging for almost half a century, nominal regional wages have B a n k of C le ve la n d or o f the diverged since 1980. This paper examines the dispersion of individual worker Board o f G o ve rn o rs o f th e F e d wages among the nine census regions. Results show that changes over time eral R e s e rve S y s te m . in the value that each region places on worker characteristics account for much of the switch from wage convergence to divergence. Temporary shocks from the 1980-82 recessions and the fall in oil prices are probably responsible M aterial m a y be reprinted pro for this interruption of the long-term trend of regional wage convergence. vid e d th a t th e source is credited. Ple as e send copies o f reprinted m aterial to the editor. Why We Don’t Know Whether Money 2 7 Causes Output b y C harles T. C arlstrom and E dw ard N. G am ber Most economists and policymakers believe that m oney affects real output. This paper argues, however, that Granger-causality studies that have pur ported to show this relationship are flawed, and that currently we do not have enough information to conclude that changes in the m oney supply cause changes in real G N P . The authors also review different models of bus iness cycles and show how money can be neutral and yet still appear to affect real output. I S S N 0 0 13 -0 2 8 1 The S tability of M oney Dem and, Its Interest S e n s itivity, and Som e Im plications for M oney as a Policy Guide by John B. Carlson John B. Carlson is an economist at the Federal Reserve Bank of Cleve land. The author acknowledges help ful discussions and comments from Charles Carlstrom, William Gavin, and George Tavlas. Thanks also to Christine Dingledine for meticulous research assistance. Introduction The money demand function is one of the most closely studied relationships in economics. One reason is that the question of the stability of money demand has long been central to issues of monetary theory. This largely reflects the influential restatement of the quantity theory of money by Milton Friedman (1956): “The quan tity theory is in the first instance a theory of money demand.” Further, he argued, “The quan tity theorist accepts the empirical hypothesis that the demand for money is highly stable— more stable than functions such as the consumption function that are offered as alternative key relations.” Friedman did not specify precisely the mean ing of “highly stable” or “more stable.” Presum ably, highly stable implies that the parameters of the money demand function do not change over time. Thus, one would expect that any reasona ble specification of money demand might satisfy some sort of in-sample stability test (for exam ple, Chow test) at a minimum. The notion that money demand is more stable than other “key” relationships has been interpreted in the context of a simple IS-LM framework by Poole (1970). In essence, “more stable” implied that the variance relatively of the money demand function was smaller than the variance of the IS curve. For years, the question of stability was simply examined by estimating various specifications of money demand, including both long-run and short-run models. It was commonly affirmed that money demand was a function of relatively few variables, including income and interest rates. By the mid-1970s, a consensus seemed to emerge that money demand was indeed one of the more stable relationships in economics, reliable enough to serve as a basis for formulating monetary policy. Unfortunately, just as a consensus seemed to develop, many of the estimated relationships broke down, first around 1974, and again around 1982. By the mid-1980s, it appeared as though many economists had given up on finding a specification of money demand that might be stable, in either the short or the long run. Recently, however, several researchers have found evidence that some specifications of money demand have remained stable through events of the 1970s and 1980s. One common conclusion of these studies is that money demand is highly interest sensitive— more so than many economists previously thought, particularly in F I G U R E M l Velocity The policy implications of the common finding that money demand is substantially interest sen sitive are analyzed in sections VI and VII. Section VIII offers some concluding thoughts. I. The Demand for M l Ratio Before 1980 SOURCE: Board o f Governors o f the Federal Reserve System. the long run. The magnitude of the interest elas ticity of money demand has important implica tions for the role of money in the economy and hence for the conduct of monetary policy. Much of the early debate about the role of money centered on how interest rates affected the velocity of money. Some analysts argued that interest-rate changes had little effect on velocity in the short or long run. Moreover, some pre sumed that M l velocity had an inherent trend growth rate of about 3 percent. These assump tions now appear to be clearly refuted by the experience of the 1 9 8 0 s. This paper reviews some recent findings of the research on money demand and considers the implications of these findings for monetary policy and rules. Section I reviews briefly a common specification of M l demand that misled many economists about the importance of inter est rates. Section II examines recent evidence that long-run equilibrium demand for the nar row money measure continues to be a stable function of relatively few variables. The implications of these findings for the apparent shift in M l velocity are discussed in section III. Section IV reviews the evidence that M2 demand is stable in the short run. In section V, the findings on M2 demand are reconciled with evidence that M2 velocity is trend stationary. Until the 1 9 8 0 s, most attention in the money demand literature was given to M l— the money measure that then included currency and noninterest-bearing demand deposits. Focus on this measure reflected both theoretical and pragmatic considerations. First, M l was the closest measure of pure transactions balances and hence con formed well to the concepts embodied in the inventory-theoretic model of Baumol (1952) and the portfolio-choice theory of Tobin (1958). These approaches essentially explained why individuals would hold the non-interest-bearing components of M l instead of interest-bearing alternatives. Perhaps more important, the focus on M l seemed justified on empirical grounds. O f the various money measures, M l appeared to be most closely related to economic activity, partic ularly in the short run. Movements in M l served as a relatively useful indicator of current and future changes in economic activity. Moreover, the velocity of M l exhibited a high degree of stability. From 1959 to 1980, M l velocity increased at a trend rate of around 3 percent, deviating only a few tenths of a percent from year to year (see figure 1 ). By the 1970s, a conventional empirical model for M l demand had evolved. 1 Desired real M l balances, were a function of some scale var iable, either real income or wealth; and a measure of the opportunity cost of holding money, the level of interest rates: m *, y, r, (1) m* = a 0 + a xy - a 2r . Earlier studies used annual data (see Meltzer [1963], Laidler [1966], and Chow [1966]). In these studies, the scale variable was typically some measure of wealth, and the opportunity cost was most often a measure of the long-term interest rate. The interest elasticities for M l ranged between -0.7 and -0.9. 2 ■ 1 ■2 See, for example, Goldfeld (1973). For a more complete discussion of earlier studies, see Havrilesky and Boorman (1978), chapters 7 and 8. Later studies in money demand used quarterly data, perhaps motivated by the increasing availa bility of such data and the development of quar terly econometric models (see Goldfeld [1973]). It became more common to use real income as the scale variable and to use a measure of the short-term interest rate as the measure of oppor tunity cost. It was often assumed that in any given quarter, money balances adjusted only par tially to their desired (equilibrium) level. The adjustment process was specified as (2) mt - mt _ j = Disinflation and financial deregulation greatly affected the opportunity cost of Ml. Disinflation resulted in sharply falling interest rates, reversing the secular trend that dated back to the 1950s. Deregulation allowed banks to compete more effectively for funds by offering interest-bearing checking accounts and market rates of interest on savings and time deposits. The opportunity cost of most bank deposits fell markedly after 1982 when market rates fell and when banks priced deposits more competitively. X(m* - mt _ 1 ), II. M l Demand Revisited where A is the speed of adjustment to equili brium. Substituting equation (1) into (2) yields (3) rnt = A.a + k a xy t - X.a2rt + (1 - k)m t _ j . 0 Equation (3) was sometimes estimated in firstdifference form . 3 The speed of adjustment of M l balances to equilibrium levels was typically estimated to be between 0.25 and 0.5 per quarter. The estimates of income elasticities of this specification were typically around 0 . 2 in the short run and less than unity in the long run. Estimates for interestrate elasticities were around -0 . 0 2 in the short run and ranged between -0.05 and -0.15 in the long run . 4 The estimates of interest elasticities seemed lower than the theories predicted and were substantially lower than earlier estimates. Given the absence of any evident interest-rate effects on M l velocity' and the apparent stability of the short-run specifications through the early 1970s, the smaller estimates of interest elasticity appeared to have gained greater acceptance. By the 1980s, however, the quarterly specifica tions for M l demand failed miserably. This was evident in the sharp change in the behavior of M l velocity, which has varied substantially since 1980 and exhibits no clear trend. The break down in the conventional relationship is believed to be largely a consequence of disinfla tion and financial deregulation. 5 long-run ■3 The inclusion of lagged money was also rationalized on an expecta tion^ basis (see Havrilesky and Boorman). ■4 While attempts have been made to rectify M l demand in the short run, no consensus appears to be forming on any particular specification (see Moore, Porter, and Small [1988]). Many analysts now question whether a short-run demand func tion can ever be identified for M l . 6 O n the other hand, recent studies by Poole (1988) and by Hoffman and Rasche (1989) suggest that the long-run (equilibrium) relationship may have endured through the past decade. Their specifi cations find that the long-run equilibrium inter est elasticity of M l demand is substantial. Poole offers an explanation for why some economists may have been misled from models estimated in first-difference form. Such models often included a constant term, which made it equivalent to a linear-time-trend specification in a regression using the levels of the data. He con cludes that in the postwar period, the constant term incorrectly picked up the trend in velocity, which should have been attributed to the post war trend in interest rates. This argument fails to explain, however, why the regressions for M l in levels form (without time-trend variables) also underestimated inter est elasticities. Closer inspection of the conven tional relationships reveals that part of the trend effect of interest rates on M l may have mistakenly been attributed to the trend in income. As noted above, the long-run income elasticity was typically estimated to be less than one— often around one-half. This, in turn, implied that over long periods, velocity would increase at approxi mately half the rate of increase in income, other things being equal. Since the conventional esti mate of income elasticity concurred with the Some specifications included interest paid on passbook savings depos its as an additional measure of opportunity cost. ■5 Some economists believe that the breakdown in the conventional rela tionship w as also a consequence of the change in the Federal Reserve’s oper ■ 6 Poole (1988) discusses the difficulties of identification from a buffer- stock perspective of money demand and concludes that the econometric prob ational procedure in October 1979 and the implications of that regime change lems may well be insurmountable. For a review of the buffer-stock approach on structural coefficients. to money demand, see Laidler (1984). inventory-theoretic models of transactions bal ances, many analysts accepted the low estimate as a confirmation of the theory.7 To estimate long-run money demand, Poole advocates a simple regression of the level of velocity on the level of a long-term interest rate using annual data. By excluding income as an explanatory variable, Poole implicitly constrains the income elasticity to be unitary; hence, any potential trend in velocity must be independent of any trend in income. Poole’s case for using a long-term interest rate is predicated on the assumption that equilibrium money demand would not likely be affected by changes in interest rates in the long run. Investment in cash management techniques is costly and hence only profitable when interestrate increases are sustained. Since long-term rates are believed to embody expectations about future short-term rates, a rise in long-term rates is likely to indicate a more permanent rise in the general level of interest rates. Thus, Poole con cludes, long-term rates better measure the opportunity cost of cash. Finally, Poole argues that adequate estimates of a money-demand function cannot be obtained by using postwar data alone. During this period, both short- and long-term rates rose secularly. Thus, he uses an extensive sample period, 1915-1986, and three different subsamples. He estimates that the interest elasticity is around -0 . 6 for the whole period and for various subsamples, which is substantially larger than conventional estimates. Hoffman and Rasche obtain estimates of a simi lar order of magnitude using a different estima tion and testing method. Unlike Poole, they do not constrain the income elasticity to be unitary. Their approach— based on the notion of cointegration— addresses a potential problem related to the statistical properties of the varia bles included in money demand. As with most economic variables, M l, interest rates, and income are nonstationary in levels. In such variables, there is no tendency to systemati cally return to a unique level or trend over time. It is now well known that standard regression analysis can yield spurious relationships between variables when the variables drift over time. Methods initially developed by Engle and Granger (1987) allow one to examine whether equilibrium relationships exist between nonsta tionary variables. Such variables are said to be temporary cointegrated, if some linear combination of them is stationary. Thus, cointegration implies a longrun equilibrium relationship between variables, and one can obtain long-run elasticities from the cointegrating vector.8 Hoffman and Rasche test for cointegration and find that 1) real M l balances and real income are not cointegrated by themselves; 2) real M l, real income, and the interest rate are cointegrated with one cointegrating vector; and 3 ) one cannot reject the hypothesis that the coefficients of real money and real income in the cointegrating vec tor are equal in value but opposite in sign. 9 The first result is consistent with the common finding that M l velocity is nonstationary. Since both income and money are nonstationary, but not cointegrated, their difference will be nonsta tionary. The second result, however, implies a stable long-run relationship between money, income, and interest rates. The third result implies that it is appropriate to interpret the cointegrating vector as a linear combination of M l velocity and interest rates or, equivalently, that the equilibrium real income elasticity of demand for real balances is unity. To estimate the equilibrium interest-rate elas ticity, Hoffman and Rasche consider both a short term rate (three-month Treasury bill) and a long term rate (10-year Treasury bond). Like Poole, they find that the interest elasticity on the long term rate is about -0 .6 , while somewhat less, -0.4, for the short-term rate. Moreover, they find that cointegration holds for either of the long- or short-term measures. These results are robust across subsample periods investigated. III. M l V elo c ity in the 1 9 8 0 s The Hoffman and Rasche findings imply that any observed drift in the velocity of M l should be proportional to any drift in nominal interest rates. Thus, any shift in the drift of velocity should be the mirror image of any shift in the drift of nom inal interest rates. Rasche (1989) investigates this last property by examining regressions of the changes in the log of M l velocity and changes in the nominal interest rate, each against a constant and a dummy variable, which is zero through December 1981 and 1.0 thereafter. ■ 8 For a more precise description of the concepts of cointegration, see Engle and Granger (1987). ■ 7 Other economic explanations for w hy an income elasticity might be less than one include improvements in cash management technology. ■ 9 All variables are in log form. The results indicate significant shifts in the interest-rate equation and in the velocity equa tion, both in the same direction. Again, the results hold for both long- and short-term rates; but, because of the high variance in the short term rates, the shift is not measured with any precision. Rasche concludes that the abrupt change in the pattern of M l velocity in the early 1980s was indeed associated with a coincidental change in the drift in interest rates. Rasche further investigates the hypothesis that the observed change in velocity behavior is a result of a break in inflationary expectations. He argues that if the postwar period through 1980 is characterized by a steady upward drift in infla tion, then it is reasonable to conjecture that it has been associated with the observed positive drift in nominal interest rates. Moreover, he argues that if inflation expectations stabilized at a lower rate in the early 1980s, it is reasonable to conclude that there has been no drift in interest rates over this period. As evidence for a break in the drift of infla tionary expectations, Rasche notes the general consistency of the Livingston Survey data. These data, which begin in the late 1940s, provide annual inflation forecasts formed at the end of the previous year. The survey reveals a general upward trend through 1980 and then a break sharply downward. Rasche notes that since 1982, the Livingston series has fluctuated without a trend in the 3 percent to 5 percent range. To summarize, the recent evidence of large interest elasticities of M l demand pro vides a basis for understanding the recent shift in the trend in velocity. While the evidence points to a reasonably stable long-run M l demand func tion, no one yet seems to have identified a satis factory short-run model. Without a reliable shortrun model of M l, little can be said about Ml velocity in the short run. long-run IV. The Demand for M 2 Recent research on M2 demand provides evi dence of stable specifications for M2 in the short run, at least in the postwar period. Moore, Porter, and Small (1988) estimate a short-run M2 demand function over the period 1964: IQ to 1986:IIQ.10 The model is specified in two parts. One is an equilibrium money demand function, similar to equation ( 1 ): mt - For further evidence concerning the stability of M 2, see Hetzel and Mehra (1987). yt where log (M2), = log (nominal GNP), and s = log (opportunity cost). Note that the unitary coefficient on nominal GNP assures that this also specifies a velocity relationship. 11 The second component is a dynamic specification based on an error-correction adjustment: U A m t = a + bet_ x + X cjA m t_ i (5) i= 1 v ^ + X dj A s , _ j + X Ji & y t - i + e / ’ / =0 i=0 et where _ x is the deviation of money from its long-run equilibrium value (derived from [4]) and e is white noise. Equation (5) essentially specifies the short-run convergence process of M2 to its equilibrium value. When the coefficient is negative, con vergence is assured. Substituting equation (4) into (5) yields t b (6 ) Am t= a - ba - bfist_ x + b (m t X- y t_ x) U +X i =1 V c i A m t _ , + X d , A s t- i w i =0 + X f i ^ y t- i + v i=0 Moore et al. estimate a version of equation ( 6 ). Simulations, both in-sample and out-of-sample, support the hypothesis that M2 demand has been and continues to be reasonably stable over the whole sample period. One key feature of Moore et al. is the way opportunity cost is measured. By definition, the opportunity cost of money is the forgone interest income of holding a monetary asset. Over the years, it has been common to use a market yield on a relatively risk-free asset, such as a Treasury bill, to measure opportunity cost. For much of the postwar period, this seemed appropriate for the narrow money measures, since holders of currency and demand deposits did not receive explicit interest payments on these instruments. Many instruments in the broader monetary aggregates like M2, however, have yielded explicit interest. Their yields, when not exceed ing interest-rate ceilings, responded at least par tially to market conditions. Moore et al. measure the opportunity cost of these instruments as the ■ ■ 10 mt = a + y t + fis, + e, , (4) 11 Moore et al. include a time index as a regressor to directly estimate any drift in M 2 velocity. While they find the coefficient to be significant, the drift is negligible at around .003 percent per year. difference between their yield and the yield of a Treasury bill. The opportunity cost of M2 then is the weighted average of the opportunity costs of each M2 component, where the weights are equal to the component’s share of M2. The response of money demand to changes in market interest rates in this model requires a specification of the relationship of deposit rates to the market rates. 12 Thus, the interest elasticity of money demand now depends on how rapidly banks adjust their deposit rates in response to changing market rates. To illustrate, consider the extreme case where deposit rates respond instan taneously to changes in market rates so as to maintain a constant spread between them. In such a case, money demand and velocity would be unaffected by changes in market interest rates because the opportunity cost of money would not change. If, on the other hand, deposit rates adjust instantaneously but only partially to a change in interest rates (that is, not point-for-point), then the interest elasticity would be proportional but less than the opportunity cost elasticity. Any trend in interest rates would also be associated with a trend in the opportunity cost of those deposits. Equilibrium money demand would hence be affected, and the trend in velocity would be proportional to the trend in the oppor tunity cost of M2. Finally, consider a case where deposit rates respond sluggishly to changes in open market rates. A permanent increase in market interest rates would initially be associated with an increase in opportunity cost, as market rates moved above deposit rates, followed by a decrease as deposit rates caught up. If the de posit rates ultimately adjusted point-for-point, the long-run equilibrium level of opportunity cost would be unaffected. Moore et al. specify deposit-rate equations to be simple linear functions of the federal funds rate. They assume that competitive forces ulti mately drive the slope coefficients to equal one minus the marginal reserve ratio, and the inter cept to equal some negative value to reflect trans actions costs that are not recovered as fees assessed to the depositor. As with M2 demand, the short run is formulated within an errorcorrection framework. Changes in deposit rates are assumed to be related to deviations of the rates from their long-run equilibrium values, and to changes in the current and past values of interest rates. Moore et al. find that for many components of M2, own rates have been relatively slow to adjust. This is particularly evident for instruments with transactions features such as NOW accounts and, to a lesser extent, money market deposit accounts. On the other hand, some deposit rates, such as those on time deposits, have adjusted relatively quickly and fully to changes in market rates. 13 However, because a significant share of M2 deposit rates adjust sluggishly, changes in market interest rates have substantial short-run effects on the opportunity cost of M2, and con sequently on its demand. Indeed, the model estimated by Moore et al. suggests that the interest elasticity of M2 demand is substantial. What is curious is that some bank deposits appear more interest sensi tive than before deregulation. One might expect just the opposite, as deregulation allows banks to compete more effectively for funds, even if they adjust only slowly. Some analysts have speculated that the in creased sensitivity of some deposits may reflect the increased sophistication of most depositholders and the improved communications technologies that have made funds transfers more convenient. Even if opportunity costs are less affected by changes in interest rates now than before, deposit-holders are much more aware of alternative assets and therefore are more likely to respond to changes in the oppor tunity cost of some deposits. 14 short-run V. M2 Velocity The treatment of opportunity cost as distinct from the market interest rate helps to reconcile why M2 velocity is trendless despite the observed trends in interest rates. This is easiest to understand in the case where deposit rates ultimately adjust point-for-point with changes in market rates. In such a case, opportunity cost is by definition stationary around some trendless differential, and hence would be independent of any trend in interest rates. Thus, the velocity of these deposits would be insulated from chang ing inflationary expectations. ■ 13 Moore et al. also conclude that deposit-rate adjustments are asym metric, adjusting more rapidly to upward movements in market rates than to ■ 12 The advantages of measuring opportunity cost as a differential in downward movements. yields are in principle greater since deregulation than before. Currently, there are no interest-rate ceilings on any of M 2’s noncurrency and non-demand- ■ 14 deposit components, which are 83 percent of the total. ity of the M 2 aggregate after deregulation. However, there appears to be no shift in the opportunity cost elastic F I G U R E Interest Rate and Opportunity Cost of M 2 Percent SOURCE: Board o f Governors o f the Federal Reserve System. F I G U R E M 2 Velocity and Opportunity Cost Ratio Percent However, not all deposits in M2 adjust pointfor-point to changes in interest rates. Reserve requirements assure some wedge preventing complete adjustment. Also, since currency pays no explicit yield, its opportunity cost is essen tially equal to the interest rate. Thus, if the level of interest rates exhibits drift, the opportunity costs of these components of M2 will also exhibit drift in the same direction. M2 velocity would not be independent of the level of interest rates. In practice, however, the drift in the opportu nity cost of M2 has been highly muted relative to the drift in interest rates (see figure 2). The wedge created by reserve requirements is in fact small— 12 percent or less. Moreover, the share of currency and resemble deposits amounts to less than 20 percent of M2; thus, the nonstationary component of the opportunity cost would be small and perhaps negligible. Interest-rate trends, then, would not affect M2 velocity sub stantially in the long run. Some evidence indicates that M2 velocity is, in the long run, independent of interest rates. Engle and Granger (1987) conclude that nominal in come and M2 are cointegrated, implying that M2 velocity is a stationary process and hence is unaf fected by interest-rate trends. Thus, it would appear that M2 velocity is immune to changing inflationary expectations in the long run. This explains why the M2 velocity trend, unlike that of Ml, was unaffected by the rise and fall of inflation in the postwar period. In the short run, however, changes in the opportunity cost of M2 are driven largely by changes in market interest rates; and, as figure 3 illustrates, M2 velocity is quite closely related to the opportunity cost of M2. 8 VI. Money as a Policy Guide During Disinflation SOURCE: Board o f Governors o f the Federal Reserve System. Recent evidence indicating that money demand is substantially interest sensitive has important implications for monetary policy. Interest sensi tivity of money demand poses serious problems for policies that seek to achieve disinflation. Poole (1988) concludes, “There is a serious and probably insurmountable problem to designing a predetermined money growth path to reduce inflation...” (p. 97). Poole offers a clear description of the problem: If policymakers embark on a credible policy of disinflation, they should expect that nominal interest rates will ultimately fall as inflationary expectations subside. Consequently, they should expect velocity growth to decline, and perhaps even become negative, if the policy becomes suc cessful. Under these circumstances, inflation Hypothetical M 2 Demand: Credible Disinflation Annual percent change SOURCE: Author’s calculations. could be reduced without a decline in money growth, at least initially. Indeed, a decline in money growth might have a significant depress ing effect on the economy. He concludes that the gradualist prescription of predetermined reductions in money growth would not be politi cally sustainable, as it would likely be associated with unnecessary weakness in economic activity. Poole further argues that this situation poses a serious dilemma for policymakers. How do they convince markets of their commitment to disin flation without a reduction in money growth rates? Is it not irrational to bet on lower inflation on the basis of a central bank’s promises, with no evidence that the central bank is reducing money growth? Poole concludes that a recession may be necessary to convince markets that the central bank is committed to a disinflationary policy. The problem of targeting money is easy to appreciate in the context of Ml. After all, few analysts anticipated the magnitude of the shift in the drift of Ml velocity. Another reduction in inflation would likely result in another shift in the trend in M l velocity. Moreover, no specifica tion for short-run M l demand seems acceptably stable at present. On the other hand, there is no evidence that the trend of M2 velocity has been affected by the transition to lower inflation in the 1980s. The recent specification by Moore et al. suggests that the short-run demand for M2 may be reasonably stable. A hypothetical example illustrates how the problem applies to a disinflation policy specified as a target path for M2. First, assume that on the basis of a promise alone, markets could be con vinced of a central bank’s commitment to grad ual disinflation from current levels to zero infla tion in 1993- To the extent that disinflation was perfectly anticipated, we might expect that nom inal magnitudes such as interest rates, personal consumption expenditure growth, and nominal GNP growth would decline smoothly to noninflationary trend paths. 15 If the parameters of the M2 demand function estimated by Moore et al. are approximately structural, then we would expect M2 demand to accelerate initially to growth rates above the equilibrium rate of nominal GNP growth and then begin to slow (see figure 4). The additional money growth would not be for the purpose of financing future spending, but would reflect a pure portfolio decision to hold a greater propor tion of wealth as bank deposits in response to a sharply falling opportunity cost; hence, the monetary acceleration could still be associated with a slowing in nominal spending. The pattern of M2 growth reflects two key fea tures of the M2 demand model. First, own rates on deposits adjust slowly enough to changes in market rates that the opportunity cost in the short run is directly related to changes in the level of interest rates. 16 Second, M2 demand is substantially sensitive to changes in opportunity cost. Thus, as interest rates fall with disinflation, so does the opportunity cost of M2. It is this decline in M2’s opportunity cost that induces investors to hold additional bank deposits rela tive to their spending needs. This example is hypothetical, of course. If markets were to maintain an expectation of gradual disinflation, they would need to under stand the consequences of a falling opportunity cost and have confidence that the estimated M2 demand function was reliable. Only then might markets reconcile an accelerat ing money-growth path with a disinflation policy. short-run ■ 15 W e assume here that in noninflationary equilibrium, growth in nominal G N P and personal consumption expenditures equals 3 percent, as does the Treasury bill rate, but that the federal funds rate equals 2 V2 percent. ■ 1 6 This, of course, presumes that banks have a rational basis for adjust ing some deposits more sluggishly than others. Thus, although market interest rates fully anticipate disinflation, bank deposits would respond with some delay. The 22-year estimation period for M2 demand is relatively short, however, and it is not evident that deposit-rate pricing has stabilized since deregulation. It would seem doubtful that markets could be convinced of such a strategy. Nevertheless, the evidence of substantial interest sensitivity of velocity in the short run suggests that policymakers might sometimes prefer to accommodate the effects of interest-rate changes on money demand. During periods of disinflation, one might then expect wide swings in money growth. Once a disinflation strategy becomes credible, velocity could fall substan tially, if only temporarily, and it would be appropriate for policymakers to accommodate the consequent surge in money demand. VII. Interest Sensitivity and Monetary Rules Apart from the problems that arise during disin flation, the evidence that M2 is more interest-rate sensitive than previously thought raises some interesting issues concerning monetary rules. On the one hand, shocks to money demand would have smaller real consequences under a constant-money-growth rule than previously thought. Consider a positive shock to money demand. Given an inelastic money supply, inter est rates would need to rise and output would need to fall. In conventional macroeconomic models, interest rates would respond initially. Higher interest rates would, in turn, tend to slow economic activity. When the interest elasticity of money demand is high, smaller interest-rate changes are required to offset demand shocks, implying smaller adjustments in output. O n the other hand, the consequences of non monetary shocks under a constant-money-growth rule are less clear when the demand for money (and hence velocity) is highly interest-elastic. This longstanding issue is illustrated simply in a debate between Johnson (1965) and Friedman (1966). Johnson argued that interest-sensitive money demand militated against a constantmonetary-growth rule “...because variations in interest rates generated by the real sector would make such a rule automatically destabilizing...” (p. 397). Implicitly, Johnson assumed that varia tions in interest rates would be a natural by product of stable output growth; in turn, these variations would cause procyclical variations in velocity, which, under the assumption of con stant money growth, would produce fluctuations in the rate of nominal income growth. Friedman acknowledged this potential out come, but argued that the conditions assumed by Johnson were highly special. Essentially, Friedman contended that while velocity would tend to move with nominal output, a constantmoney-growth rule would nevertheless dampen output fluctuations relative to “discretionary” policies. Thus, Friedman was not comparing his rule to an ideal rule, but to the existing practice of the central bank. It is useful to separate this debate into two issues. The first is the general issue of rules ver sus discretion. The second is the question of whether monetary rules (or targets) should allow for some kind of systematic (that is, auto matic) feedback to account for interest-rate changes and, hence, shifts in velocity. More spe cifically, should a rule or targeting procedure anticipate changes in interest rates? This first issue is only indirectly relevant to the question of interest-rate sensitivity and therefore is not dealt with here. 1 7 The question of feedback, on the other hand, is relevant whether a policy admits some discretion or not. The feedback issue depends on the kinds of shocks that occur and on the poorly understood dynamics of adjustment in the economy. Specifi cally, it depends on where shocks arise in the economy, what their relative magnitudes are, and how they are propagated through the econ omy. The answers to these questions depend on the particular model one believes is appropriate for characterizing the economy. Unfortunately, no consensus exists or even seems imminent. One large and influential class of empirical models, sharing a common propagation mecha nism, casts some doubt on the efficacy of monetary-growth rules. In these models, the inflation process is characterized by an output-gap accelerationist mechanism: con stant P r P t -i p = ao+ «i (tf* - tf,) + a z t> q where is the inflation rate, is the level of output, is full-employment output, and represents other factors. If is constant, a change in the inflation rate depends on the out put gap. When output exceeds full-employment output (that is, when unemployment is below its natural rate), inflation accelerates. When output is below full-employment output, inflation decelerates. Anderson and Enzler (1987) explain the consequences of such a mechanism for a monetary rule: ■ 17 q* z z For a discussion of the general issue of rules versus discretion, see Carlson (1988). It is easy to se e w h y h o ld in g the m o n e y grow th rate constant might not result in a stable sim ula tion path for a m a crom od el containin g this m e ch anism. T he fixed m o n e y grow th path predeter m ines b o th the rate o f inflation and the price level consistent with the e c o n o m y ’s steady-state path at each point o f time. C onsider what happens if the price level is disturbed upward from the steady-state grow th path. T he dem a n d for m o n e y is increased and interest rates rise. This depresses output and increases un em ploym en t. T he increased un em ploym en t, in turn, d epresses the rate o f change o f prices. As lo n g as the price level remains t o o high, a force is created that tends to k eep u n em p loym en t a b ove its natural rate and the rate o f in fla tion con tin u es to fall. T he d eclin in g rate o f inflation eventually returns the price level to its steady-state value, and this in turn allow s the u n em p loym en t rate to return to the natural rate, but at this p oint in fla tion is to o lo w to b e consistent with the fixed m o n e y grow th path and the price level falls through the steadystate level. This redu ces the d em a n d for m oney, causing interest rates to fall until u n em p loym en t is b e lo w the natural rate. Inflation then acceler ates until at so m e p oint it reaches its steady-state value. But n o w the level o f prices is t o o low . T he mirror im age o f the p reviou s events takes p lace and oversh ootin g occu rs again, (p . 2 9 7 ) While the estimated parameters of these models suggest that the cycle described above eventually converges, the process is generally only slightly dampened . 18 Because the estimated interest elasticity of output in these models is typically relatively small, it is likely that a higher interest elasticity of money demand would only attenuate the cycles of such models. To illustrate this point, consider again the propagation of the upward price dis turbance. The higher the interest elasticity of money demand, the lower the rise in the level of the interest rate that would result as an effect of the price shock on money demanded, given an inelastic supply. However, because the interestrate elasticity of output is low, the consequent effect on output would be even smaller, and would hence slow the process that dampens the shock to inflation. 19 Evidence of a potential for long macroeco nomic cycles is not a unique consequence for models with an output-gap mechanism. Indeed, some simple models linking money and prices also exhibit long cycles. One example is a recent single-equation model estimated by Hallman, Porter, and Small (1989). Theirs is a reducedform model of the relationship between inflation and M2 that does not explicitly include either the current level of output or employment as a variable. 20 While they find rather lengthy adjust ments to simulated shocks (for example, more than 1 0 0 years), the cycles of their model are more damped than those of many large macroeconomic models. From a deterministic point of view, the Hallman et al. results suggest that there is a money-growth path consistent with a relatively smooth to equilibrium. As they note, inflation, could be controlled at any constant rate with constant growth of M2. Notwithstanding the well-known critique of Lucas (1976), the use of deterministic simulations as evidence in the debate about an appropriate policy rule is of only limited value. A critical issue in this debate is how a rule performs in a stochastic framework, one that approximates the of disturbances that have historically affected the various sectors of the economy. In this context, the issue is not the selection of an appropriate policy response to a particular shock, but the robustness of a contractual commitment to a policy rule in responding to a of likely outcomes arising from a typical distribution. One sense of robustness has been stressed by McCallum (1988): that a rule perform well for a variety of models, preferably ones incorporating alternative views of macroeconomic relation ships. It is important to establish robustness (in this sense) because no structural model of the economy enjoys sufficiently wide acceptance; nor does any consensus seem to be evolving. Thus, to gain acceptance for a proposed rule, the rule advocate must demonstrate that the rule would lead to reasonably good outcomes for var iables of interest for a variety of models . 21 nonconstant transition in equilibrium, distribution series and ■ 20 Nevertheless, the model incorporates estimates of full employment output and equilibrium velocity as determinants of the equilibrium price level. In this model, inflation is a function of the gap between the current price level and its equilibrium level. ■ 18 It should be noted that these models typically do not result in a trade-off between inflation and unemployment in the long run. ■ 21 One method of simulation designed to address this issue is suggested by Tinsley and von zur Muehlen (1983). They essentially offer a technique to generate unplanned disturbances consistent with the error structure observed ■ 19 It is perhaps ironic that these models suggest that a constant-money- in historical experience. The robustness of a policy rule is tested by multiple growth rule would result in an interest-rate path that is too smooth to substan simulations of the performance of the rule over multiyear periods, where each tially dampen shocks to inflation over reasonably short horizons. Indeed, these simulation draws a different series or “history” of unplanned disturbances. The models suggest that rather large and sustained increases in interest rates horizons are chosen to be long enough to allow significant differences to would be required to substantially affect the output gap and hence the inflation emerge among the alternative policies and to assure that policies ultimately rate. However, it is uncommon to find antagonists of the money-growth rule stabilize outcomes. who cite this evidence and also publicly advocate the kind of interest-rate variation that large models suggest is required to stabilize the inflation rate. The sum of simulation results provides distributions of outcomes for each of the model’s variables. For instance, one policy m ay be associated with a wide Stochastic simulations, however, are costly to obtain. Moreover, a test for robustness is an open-ended search, encompassing an endless variety of both rules and models. As a conse quence, evidence from this analysis is in only an embryonic state. Preliminary results by Tinsley and von zur Muehlen (1983) and Anderson and Enzler (1987) suggest, however, that monetary rules do not perform as well as alternative rules or intermediate targeting procedures. Neverthe less, the monetary rules and targeting procedures examined were based on older, less interestsensitive estimates of money demand. The ongoing debate over the efficacy of a constant-money-growth rule, when the interest elasticity of money demand is large, is not likely to be resolved without some convincing empiri cal basis. Thus, it would seem appropriate for policymakers to take account of the conse quences of expected interest-rate changes on velocity when choosing target ranges for M2 over a period of a year or less. That is, it may be appropriate for M2 growth to slow substantially when interest rates are rising and expected to rise further, or to accelerate substantially when interest rates fall. VIII. Concluding Comments One common finding of recent empirical research in monetary economics is that the interest elasticity of money demand is estimated to be substantial, and higher than many econo mists previously thought. The evidence seems strongest for M l demand in the long run. While interest rates have little long-term effect on M2, the short-run elasticity seems to be greater than previously thought. When the interest elasticity of money demand is high, velocity can vary widely. This creates a problem for using money as a policy guide. Monetary targets should take into account the consequences of expected changes in interest rates on money demand. This problem is per haps most difficult during periods of disinflation, when changing expectations about inflation result in large swings in interest rates and hence in velocity. range of outcomes for output and interest rates, but with a small range for prices and money for any given simulation horizon. Another policy m ay be associated with small ranges for interest rates and money, but with large ranges for prices and output, or vice versa. Tinsley and von zur Muehlen note, “ ...the essential contribution of stochastic simulation analysis is the empirical premise that while individual unplanned disturbances cannot be predicted (by definition), their ranges of probable outcomes are unlikely to differ significantly from the dispersions observed in historical experience...” (p. 16). Finding that a money-demand function is stable is not a sufficient basis for adopting a constantmoney-growth rule. The rule advocate has the burden of convincing others that the stabilizing effects of the monetary rule would outweigh the potentially destabilizing effects of maintaining constant money growth when velocity varies sys tematically with interest rates. Because no con sensus exists about the best model for the econ omy, the rule advocate must argue his case in the context of a variety of models. The challenge of examining rule robustness has been recognized and addressed by McCallum (1988). It is hoped that others will follow his lead. Recent developments in simulation methods offer promising approaches for examin ing the robustness of alternative policy rules. References Anderson, Robert and Enzler, Jared J., “Toward Realistic Policy Design: Policy Reaction Func tions that Rely on Economic Forecasts,” in Rudiger Dornbusch, Stanley Fischer, and John Bossons, eds., Cam bridge, Mass.: MIT Press, 1987, 291-330. Macroeconomics and Finance: Essays in Honor of Franco Modigliani, Baumol, William J., “The Transactions Demand for Cash: An Inventory Theoretic Approach,” November 1952, 545-56. QuarterlyJournal of Economics, 66, Carlson, John B., “Rules Versus Discretion: Mak ing a Monetary Rule Operational,” Federal Reserve Bank of Cleveland, Quarter 3, 1988, 2-13- Economic Review, Chow, Gregory C., “On the Long-Run and ShortRun Demand for Money,” April 1966, 111-31. Economy, 74, Journal of Political Engle, Robert F. and Granger, C.W.J., “Cointe gration and Error Correction: Representation, Estimation, and Testing,” March 1987, 251-76. 55, Econometrica, Friedman, Milton, “The Quantity Theory of Money—A Restatement,” in Milton Friedman, ed., Chicago: University of Chicago Press, 1956, 3-21. Studies in the Quantity Theory of Money, ------ , “Interest Rates and the Demand for Money,” 1966, 71-85. Journal of Law and Economics, Goldfeld, Steven M., “The Demand for Money Revisited,” in Arthur M. Okun and George L. Perry, eds., 1973, 3, 577-638. Activity, Brookings Papers on Economic Hallman, Jeffrey J., Porter, Richard D., and Small, David H., “M2 per Unit of Potential GNP as an Anchor for the Price Level,” Board of Governors of the Federal Reserve System, April 1989. 157, Staff Study No. Hoflfman, Dennis and Rasche, Robert H., “Long Run Income and Interest Elasticities of Money Demand in the United States,” National Bureau of Economic Research Working Paper No. 2949, April 1989. Johnson, Harry G., “A Quantity Theorist’s Mone tary History of the United States,” June 1965, 388-96. Journal, Economic 75, Laidler, David, “The Rate of Interest and the Demand for Money— Some Empirical Evidenc December 1966, 545-55. e,”Journal of Political Economy, 74, ________, “The Buffer Stock Notion in Monetary Economics,” Supplement, 1984, Economic Journal, 94, 17-33. Lucas, Robert E., “Econometric Policy Evalua tion: A Critique,” Supplementary Series 1, 1976, 19-46. Journal of Monetary Eco nomics, McCallum, Bennett T., “Robustness Properties of a Rule for Monetary Policy,” CarnegieRochester Conference Series on Public Policy, Carnegie Mellon University and National Bureau of Economic Research, February 1988. Meltzer, Allan H., “The Demand for Money: The Evidence for the Time Series June 1963, 219-46. litical Economy, "foum al of Po 71, Moore, George R., Porter, Richard D., and Small, David H., “Modeling the Disaggregated Demands for M2 and M l in the 1980s: The U.S. Experience,” mimeo, Board of Governors of the Federal Reserve System, May 1988. Poole, William, “Optimal Choice of Monetary Policy Instruments in a Simple Stochastic Macro Model,” May 1970, 197-216 . Quarterly Journal of Econom 84, ics, ________, “Monetary Policy Lessons of Recent Inflation and Disinflation,” Summer 1988, 73-100. Economic Perspectives, The foum al of 2, Mone Rasche, Robert H., “Some Evidence on the Elu sive 1982 Shift in Velocity Drift,” Shadow Open Market Committee Policy Statement and Position Papers, No. 89-01, 1989, 37-40. Hetzel, Robert L. and Mehra, Yesh, “The Behav ior of Money Demand in the 1980s,” mimeo, October 1987. Tinsley, P A and von zur Muehlen, P., “The Reliability of Alternative Intermediate Tar gets,” mimeo, Board of Governors of the Fed eral Reserve System, November 8 , 1983. Havrilesky, Thomas and Boorman, John, Arlington Heights, 111.: AMH Publishing Co., 1978. tary Macroeconomics, Tobin, James, “Liquidity Preference as Behavior Towards Risk,” 1958, 65-86. 25, Review of Economic Studies, A ccou nting fo r the Recent Divergence in Regional W age Differentials by Randall W. Eberts Randall W . Eberts is an assistant vice president and economist at the Federal Reserve Bank of Cleveland. Comments and suggestions by Brian Cromwell, Erica Groshen, and Sharon Smith, and computer assistance by Ralph Day and John Swinton, are gratefully acknowledged. Introduction Convergence of regional income differentials is commonly perceived as the natural result of the gradual development and maturation of regional economies. One expects that factors such as improved transportation and communication, enhanced mobility of capital and labor, and the shift away from resource-based activities would lead regions, and their incomes, to look more and more alike. Indeed, since the 1880s, the general trend has been toward convergence of regional per capita income in the United States. Recently, this trend appears to be reversing. Browne (1989) shows that since 1979, regional disparities in per capita personal income have been on the rise. Furthermore, she concludes, “...the key to both the converging per capita incomes of the 1970s and the diverging incomes of the 1980s was changes in industry earnings” (p. 38). According to Nourse (1968), regional income divergence has happened only once in the last century, between 1920 and 1940. After 1940, regional incomes returned to their longer-run path of convergence. Easterlin (1958) concluded from that 2 0 -year disturbance in the longer-run trend that “...it is by no means certain that con vergence of regional income levels is an inevita ble outcome of the process of development. For while migration and trade do appear to exert significant pressure towards convergence, they operate within such a rapidly changing environ ment that dynamic factors may possibly offset their influence” (p. 325). It appears that the con clusion Easterlin drew 30 years ago may be rele vant in today’s situation. This recent deviation from the general ten dency toward convergence raises several ques tions. Why the relatively sudden shift in the direction of regional income differentials after so many years of convergence? What are the sources of this change in regional per capita income? Have the fundamental forces that shape the nation’s economy changed direction during the 1 9 8 0 s, or is this merely a temporary digression from the longer-run trend of convergence? This paper begins with the observation by Browne that earnings account for most of the shift from income convergence to income diver gence among regions. We identify two basic sources of regional wage differentials and exam ine which of them is more responsible for the shift in wage patterns. The two sources are 1) regional differences in the return on various worker attributes and the wage differentials among industries and occupations, and 2 ) regional differences in the level of worker attri butes and the distribution of workers among industries and occupations. These two sources can be distinguished by asking whether earnings per worker differ among regions because of differences in the attributes of workers, or because of differences in the value of worker attributes as determined by the regional labor markets. Explaining con vergence or divergence of regional wages, there fore, rests with the ability to explain convergence or divergence of characteristic prices, levels of characteristics, or both. Several studies have explored the relative size of these two components of wage differentials between regions, primarily in an attempt to explain the difference in wages between the South and other regions of the country. Sahling and Smith (1983) were among the first to look at the wages and attributes of individual workers to examine regional wage differentials over time. They compared the South with four other regions in the country: the Northeast, the North Central, the New York metropolitan area, and the West. They estimated separate real and nominal wage equations using a sample of residents from 2 9 of the largest standard metropolitan statistical areas (SMSAs) found in these five regions. The workerattribute variables included measures of school ing, experience, race, occupation, sex, industry, job status, and union membership. Using two cross sections of data, from the May 1973 and May 1978 Current Population Surveys, they con cluded that much of the variation in wages be tween the South and the other regions examined is a result of substantial variation in the real and nominal rates of return to worker characteristics. Farber and Newman (1987) extended Sahling and Smith’s analysis to look explicitly at changes in characteristic prices over time. In addition to looking at regional wage differentials in two dif ferent years, 1973 and 1979, they estimated the changes in the differentials between the two years for various pairs of regions. They found that more than half of the predicted changes in South/non-South wage ratios can be accounted for by changing relative returns to worker charac teristics between the two areas (p. 2 2 3 ). Other studies, using similar techniques and micro-level data, do not necessarily agree with the conclusion that characteristic prices account for regional wage differentials. Bellante (1979) and Gerking and Weirick (1983) find that regional wage differences are due primarily to differences in the levels of worker characteristics. These results leave open the possibility that both prices and levels are likely sources of regional wage differentials. 1 This paper extends Farber and Newman’s work in two directions. First, it includes three time periods in order to examine the sources of the switch in wage patterns that apparently occurred at the beginning of the 1980s. Each time period is constructed by pooling three years of data: the first period includes the years 1973-75, the second includes 1979-81, and the third includes 1985-87. The interval between the first and second periods is characterized by regional wage convergence, as documented by Farber and Newman (1987) and Browne (1989). The interval between the second and third peri ods exhibits regional wage divergence, as shown by Browne. The second direction is to look at all nine U.S. regions as defined by the U.S. Bureau of the Census, relative to the national average, instead of comparing pairs of selected regions. However, unlike the studies by Sahling and Smith and Farber and Newman, which were con cerned with comparing wage differentials across different regions, our purpose is to see whether the structure that caused a particular region to converge toward the national average during the early periods can also account for the divergence of wages in that same region during the latter periods. Therefore, it appears that using nominal wages is sufficient for an initial look at the sources of the shift in wage patterns. 2 I. Explanations of Regional Wage Differentials One of the longstanding tenets of economics is that efficient markets result in equal prices across regions. Indeed, economists have observed for decades the slow convergence of average wages among the regions of the United States, where goods and factors can flow freely. How, then, can one explain the apparent divergence of wages in recent years? ■ 1 Dickie and Gerking (1988) provide a very comprehensive and insightful critique of the literature. ■ 2 Work by Roback (1982) and Beeson and Eberts (1989) shows that considering nominal wages can be viewed as only a partial-equilibrium analy sis. Household spatial equilibrium includes not only wages, but also the price of housing and nontraded local goods. Therefore, focusing only on nominal wages may introduce estimation bias, especially in the prices of worker char acteristics, for regions in which housing and other local-goods prices have changed significantly from the national trend. International trade theory offers useful insights into conditions that lead to regional wage con vergence and divergence. Much of the relevant literature discusses wage equalization: average wages across regions are equal if both the prices of worker characteristics and the composition of worker characteristics are the same. If the first condition holds, then wages of identical workers will be the same across regions. However, unless the second condition also holds, the average wages of regions will be unequal. Within a regional context, conditions for equal ization of characteristic prices are less stringent than those for equalization of characteristic levels. 3 A well-known theorem in trade theory, the factor-equalization theorem, states that trade in commodities and factor movements are sub stitutes. According to this theorem, free trade of goods leads to equal factor prices among regions, even when factors of production are immobile. Therefore, within the United States, which does not limit trade between regions, one would expect the unimpaired flow of goods to tend to equalize wages. It has been this line of thinking, based on the notion that regions trade because of differences in factor endowments, that has led to expectations of regional wage convergence. Several assumptions, which may or may not be met, are necessary to reach this conclusion, however: a) relative factor endowments are not identi cal across regions, b ) regions have identical technologies, c) regions have identical homothetic demand, d) production is characterized by constant returns to scale, e) production is characterized by perfect competition, and f) there are no domestic distortions in either region. Markusen (1983) demonstrates that the rela tionship between commodity trade and factor trade varies depending on the specific assump tions that are retained. By relaxing each of the assumptions one at a time, he shows that the ini tial trading equilibrium is not characterized by factor-price equalization. In each case, factor prices cannot be equalized between regions until at least one region is specialized. He concludes that the notion that trade in goods and factors are substitutes may be a rather special result, which is generally true only when differences in relative factor endowments are the basis for trade and when no market imperfections exist. ■ 3 Dickie and Gerking (1988) use trade theory to provide a comprehensive Regions may trade goods for reasons other than initial differences in factor endowments. Markusen considers various other bases for trade between regions in which the initial trade equi librium is not characterized by factor-price equal ization. These conditions include a) differences in production technologies, b) production taxes, c) monopolies, d) external economies of scale (increasing returns to scale), and e) factor-market distortions. If these characteristics hold for regions, then factor prices will not be equalized, even though goods may still flow freely among regions. It is easy to envision regional differences in technol ogy, taxes, market share, agglomeration econo mies, and unions— all of which would satisfy one or more of the above conditions. Factor-price equalization can be achieved in these less-specialized cases if factors are mobile. Factors will flow to the region with the higher price, until interregional price differentials dis appear. When trade is based on factors such as those listed above, factor prices will differ in such a way that the price will be higher for the factor that is used intensively in the production of the export good of that region. Consequently, the region will be relatively well endowed with the factor that is more intensively used in the production of the region’s export good. How ever, factor flows, particularly labor migration, are impeded by imperfect information, by mov ing costs (both monetary and psychic) and, in the case of labor, by imperfect matches between labor skills and job requirements. What does this mean for the second compo nent of wage changes— the level or composition of factors? When trade is based on differences in factor endowments, there will be no migration based on wage differentials, for the simple rea son that wages will not differ between regions because of interregional trade in goods.4 When trade is based on differences in production technologies, taxes, or factor-market distortions, factor-price differentials lead to factor flows, but these flows will result in different proportions of factors. Therefore, these models suggest that average wage levels are very unlikely to be the same across regions. 5 Even though interregional ■4 Of course, individuals m ay find regions to be attractive for reasons other than higher wages. Site-specific amenities m ay also influence an individ ual’s preferences. ■5 W ages will also differ across regions because of compensating differen assessment of the necessary and sufficient conditions for regional wage tials for site-specific characteristics, as discussed by Beeson and Eberts equalization. (1989). prices may be equal, as predicted by both m od els, it is most likely that the composition of the characteristics will differ among regions. Dickie and Gerking (1988) summarize the outcomes of trade theory as they pertain to interregional wage differentials. First, equaliza tion of labor-characteristic prices does not depend on geographic mobility of the entire labor force. Rather, equalization occurs if enough markets for goods and factors exist and if those markets are allowed to clear. Second, when a combination of commodity trade and factor mobility guarantees factor-price equalization, then relative factor supplies end up unequal and regions tend not to become homogeneous in factor composition. Third, when labor is hetero geneous, economic efficiency, as evidenced by equal factor prices, does not lead to interregional equality of average wages (pp. 1 0 -1 1 ) . 6 Therefore, it appears that a systematic change in characteristic prices is a likely source of the switch from regional wage convergence before 1980 to regional wage divergence after 1980. The subsequent analysis estimates the two basic components of regional wage changes and examines which of them contributes more to these observed changes. II. Accounting for Regional Wage Differentials Consider the standard hedonic wage equation in which the wage ( ) of individual living in region is a function of the individual’s attri butes ) and job or workplace characteristics j (H'j i (Ctf): (1 ) Wtj= w(Hit, Cjj). Assuming perfectly operating labor markets, prices of each attribute are determined by supply and demand conditions. Under the assumptions of perfect information, costless spatial labor mobility, and zero transactions costs, characteristic prices will be the same across regions. Conse quently, workers with the same characteristics will be paid the same wage regardless of location. The technique used to account for the two sources of wage differentials follows the approach of Oaxaca (1973), with modifications made by Sahling and Smith (1983) and Farber and New man (1987). Writing equation (1) in log-linear form, dropping the individual subscript, and adding a time subscript yields (2) lnwj t = bjtXj t , where j= t= regions, and time periods. 1 1 The parameter vector b -t represents the charac teristic price and vector XJt represents the levels of characteristics, both of which can differ among regions and over time. Using for Inw, we can write the percentage change in wages between two regions ( and ) during one time period as y S (3) N (yst - yNt) = ( bSt - bNt)XSt + ( X Sl - XNt)bNt, t= 1 ,...,7 ! The first term on the right-hand side accounts for the change in characteristic prices between regions and For our purposes, denotes the national average. The second term denotes the change in levels of worker characteristics between the two regions.7 It is clear from equa tion ( 3 ) that wage differences between regions result either from differences in prices or from differences in levels. One can use this frame work to assess which of the two components accounts for the larger share of the regional wage difference.8 The issue of wage convergence or divergence requires examining how these regional wage dif ferences change over time. For wages to con verge toward the national average, the distance between the regional and national wage level must narrow over time. Consequently, if the region starts out with a wage above the national average, convergence requires that the difference, S N. N y y Nt- 0 's, Nt) - Ofc- i " i>> must be greater than zero. The same relationship must be nega tive if the region starts out with a wage below the national average. The condition for divergence, ■7 A residual term, (bs - b N )(Xs - X N ), is omitted for simplicity. Fur thermore, there is an index problem associated with this technique. Changing the base to one region or the other will change the values of the components. Some studies, such as Sahling and Smith (1983), have attempted to avoid the problem by using averages of the two region’s characteristic levels or prices. W e instead choose to follow the technique of Farber and Newm an (1987), which chooses one region as the base. In this w ay, we are better able to compare our results with theirs. ■ 6 Dickie and Gerking also stress a fourth and important point: if data do ■ 8 A s Farber and Newman point out, the accounting framework relies on not adequately distinguish between workers with particular characteristics, the unbiasedness and consistency properties of O L S estimators, and has then estimated returns will be averages and tests of the interregional wage avoided the pre-test biases of imposing implicit restrictions on coefficients equality hypothesis would be biased toward rejection. found to be statistically insignificant (p. 219). obviously, would require the opposite signs.9 The relative change over time in regional wage differentials can be divided into several compo nents using a variation of the same accounting scheme adopted in equation ( 3 ) for the static case. Following Farber and Newman, one can specify equation ( 3 ) for two different time peri ods (in this case, periods 1 and 2 ) and then sub tract one from the other. This technique yields the following accounting framework: (4) ( V.v2 - J ^ 2) - 0 ’5I " JWl) = KX,-2 - * v 2) XN j)] bN2 + ~ j)( by ~ &s ^ + - XN1)(bN2 - bN j) + Xsl [(bs 2- bN2) (bsl - (*S i 2 2 2 The four components can be interpreted in the following way. The first term, referred to as the main effect, reflects how much of the change in the wage differential is due to changes in the differences of wage-determining characteristics between the two regions, evaluated at the national average characteristic prices. Notice that this term may be zero even when characteristic levels differ between region and the national average in each time period, as long as these dif ferences are not the same in each time period. The second term is the price-interaction term and reflects the effects of absolute changes in characteristics of workers in region over time. The third term is the price-interaction effect, which allows for characteristic prices to change over time. The last component, the region-time interaction effect, represents the possibility that the characteristic prices in the two regions may change over time at different rates. These four components of regional wage changes provide the basis for identifying the rel ative contributions of intertemporal changes in characteristic prices and levels to the regional S S wage differentials. To construct these wagechange components, separate hedonic wage equations are estimated for each region in each time period. For nine census regions and three time periods, this requires 27 separate regres sions. The coefficient estimates and the means of the levels of characteristics are then combined according to equation (4 ) . 10 Comparing changes in regional wages relative to the national average partially adjusts for the general nominal wage increases observed over the 15-year period between 1973 and 1987. How ever, any deviations of regional price trends from the national average will be imbedded in the various components, particularly in those related to differences in characteristic prices. Instead of relying on the national trends to capture regional price differentials, it would be ideal to adjust regional wages for differences in the cost of liv ing. Unfortunately, regional indexes are available only for metropolitan areas, and even then, there are no current indexes that can be used to com pare cost-of-living differences across metropoli tan areas. III. Empirical Results Data The data used to estimate the wage differentials are obtained for various years from the Current Population Surveys (CPS) compiled by the U.S. Department of Labor, Bureau of Labor Statistics. The CPS surveys individual workers periodically regarding hours worked, earnings, worker char acteristics, employment status, and so forth. Each time period considered in the analysis consists of a pooled sample of three years. 11 The first ■ 10 One drawback of this approach, as discussed by Dickie and Gerking (1988), is the lack of a confidence interval estimate around these various components, leaving it unclear how the results generalize to the population. ■ 11 Various features of the C P S files have changed over the years covered in this analysis, which introduces several problems when using these data to derive a consistent time series of regional wages. First, the method of collecting wage and worker characteristics has changed. For the years 197378, questions regarding worker wages and characteristics were asked only in the month of M ay. This poses two problems. First, the sample contains only ■ 9 W e have chosen to compare each region with the national average, those individuals who were in the second rotation, which, in addition to being which we feel provides the most clarity when so many regions are being less representative, reduces the number of respondents. Second, annual wage compared. This approach m ay introduce two sources of bias, however. The estimates will reflect wages obtained for only one month of the year. first is because the national sample is not a region separate from the others, Starting in 1979, the wage questions were asked of one-quarter of the indi but is made up of individuals in each region. The second source arises from viduals in each of the 12 monthly surveys conducted each year. Because of the finding that the characteristic prices of each region are significantly differ the difference in the w a y in which information is gathered, the total number of ent. Consequently, the characteristic prices estimated for the nation m ay not workers with sufficiently complete records for analysis is much smaller before represent prices for the national market, but rather the average of prices from 1979 than afterward. Pooling the individual years will ameliorate these prob each distinctly different regional market. lems to some extent. Q T A B L E Regression Estimates for the National Sample of Workers 1 9 7 3 -7 5 Variables Mean Full-time (= 1 ) Race (nonwhite = 1 ) Sex (female = 1 ) Experience Experience squared Schooling Schooling squared (exp) x (sex) Non-SMSA (= 1) 1 9 7 9 -8 1 Coefficient .80 Mean .42 18.72 570.60 1.82 4.03 7.77 .30 Coefficient .141 .81 .1 0 1 -.040 -.192 .026 -.0004 .130 .007 -.004 -.159 .1 0 1 9 8 5 -8 7 .1 2 -.0 3 2 .46 17.76 521.57 1.98 4.57 7.90 .42 -.1 6 8 .0 2 6 -.0004 .131 .0 0 2 -.004 -.083 Mean Coefficient .80 .13 .48 17.59 488.80 2.09 5.00 8.32 .27 .187 -.048 -.133 .029 -.0004 .1 6 2 -.0 0 1 -.005 -.133 occupation dummy variables) (1 3 industry dummy variables) ( 1 2 Dependent variable: log (earnings/hours) R2 Number of observations 1.29 1.74 .91 .95 116,298 554,864 2 .0 2 .96 491,510 NOTE: All coefficients are statistically significant at the 99 percent level. SOURCE: Estimates are derived from the Current Population Surveys. See text for details. period combines the responses from the May survey for the years 1973, 1974, and 1975. The second period pools responses from one-quarter of the individuals in each of the 1 2 monthly sur veys for the years 1979, 1980, and 1981. The third period is derived similarly, except that it includes the years 1985, 1986, and 1987. These time periods were chosen because they correspond to the switch from regional wage convergence to regional wage divergence as documented by Browne (1989). In addition, years were pooled in order that each region con tained enough workers to ensure reliable esti mates. The size of the samples ranges from 7,203 workers for the New England census region in 1973-75 to 84,641 workers for the East North Central region in 1979-81. Following the human-capital specification of Hanoch (1967) and Mincer (1974), individual wages (expressed in logarithms) are specified as a function of various worker attributes. We include education level (entered as a quadratic), poten tial experience (age, minus years of education, minus six, also entered as a quadratic), and the interaction between experience and female. We also include binary dummy variables indicating whether or not the worker is a full-time em ployee, female, and nonwhite. Dummy variables are also used to denote a worker’s occupation, the industry in which he or she is employed, and whether the worker resides in an SMSA. Hourly earnings were computed by dividing average weekly earnings by average weekly hours. 12 Including the industry-dummy variables is somewhat inconsistent with the notion that the human-capital specification captures supply-side aspects of the labor market. These variables are included, as they have been in other studies, to test the popular notion that industrial restructur ing is a primary source of regional wage changes. The changing composition of union membership has also been offered as an explanation for regional wage changes. 13 Unfortunately, the CPS did not ask about union affiliation in the 1979-81 surveys. ■ 1 2 A n interesting extension of the analysis would be to estimate sepa rate regressions for males and females and for whites and nonwhites. Sahling and Smith (1983) found differences in wages between males and females in the South compared with other regions. Changing norms for women and minor ities in the workplace m ay lead to regional differences in the characteristic prices of these groups. ■ 13 However, Farber and Newm an (1987) conclude that while unionization is an important contributor to the change in the wage differential attributable to changes in regional differences in worker characteristics, it is not an impor tant variable in explaining changes in wage ratios between regions (p. 222). F I G U R E Regional Nominal Wage Differentials Relative to the National Average Percent everything else being the same. The full-time wage premium has risen from 1 0 percent in the first period to 19 percent in the most recent period. This fairly sizable increase has occurred even though the percentage of full-time workers in the sample has remained constant. The nonwhite wage gap appears to have nar rowed slightly from 4 percent in 1973-75 to 3-2 percent in 1979-81. However, since that time, the gap has widened, increasing to 4.8 percent in 1985-87. The female wage gap, on the other hand, has steadily narrowed, from 19.2 percent in the first period to 1 3 3 percent in the most recent period. The wage premium placed on additional hours of work experience has risen steadily for both men and women over the three time periods. Taking into account the interaction terms and evaluating at the mean level of expe rience, the elasticity of wages with respect to experience for men, for example, rose from 2 0 . 6 percent in 1973-75 to 26.4 percent in 1985-87. The net effect of schooling on wages fell between the first two periods and then rose in the third period. Patterns of Regional Wage Differentials SOURCE: Author’s calculations from Current Population Surveys. Regression Estimates Separate estimates were obtained for each of the nine census regions for each time period using ordinary least squares. F-tests were performed to test the null hypothesis that the coefficients for each region are equal to the coefficients for the national sample. The null hypothesis was rejected at the 1 percent confidence level for each time period. Even though coefficients differ among regions, estimates from the national sample are displayed and discussed in order to provide an overall perspective of the results. As shown in table 1 , all worker-characteristic varia bles are statistically significant at the 1 percent level and enter with the expected signs. Full-time workers (who work 35 hours or more a week) receive higher wages than part-time workers, Nominal earnings estimates, using the CPS sam ple of workers within nine census regions, reveal a pattern of regional wage convergence followed by divergence, similar to that found by Browne. Figure 1 shows the pattern of regional nominal wage changes relative to the national average. Nominal wages in all regions, except the New England and the Pacific regions, converged toward the national average between 1973-75 and 1979-81. The standard deviation of the relative wage differentials fell from 0.086 to 0 . 0 6 8 during this period. Wages of workers in the Pacific region increased 2 . 6 percentage points faster than the national average between the first two periods, which raised the region’s wage premium to 1 3 - 3 percent. New England, on the other hand, started out below the national average in 1973-75 and continued to lose even more ground by 1979-81, falling from 2 . 1 percent to 5.9 percent below the national average over this time span. Between 1979-81 and 1985-87, wages in most of the regions diverged from the national aver age. The two exceptions were the New England and East North Central regions. Wages in the New England region jumped dramatically during this period, outpacing the national average by 9.1 percentage points. This spurt in wage growth closed New England’s wage gap from the pre vious period and placed its wages 3 . 1 percent above the national average in 1985-87. Wages in the East North Central region also came closer to the national average, but this was achieved by growing slower than the nation by 3 . 0 percent age points. O f the seven regions in which wages diverged from the national average between 1979-81 and 1985-87, five were below the national average. The two regions that lost the most ground were West North Central and East South Central. Wages in the West North Central region fell from 7.0 percent below the national average in 197981 to 10.9 percent below in 1985-87. Wages in the East South Central region, which in the first two periods were the lowest in the country, fell even further, to 15.2 percent below the national average. Wages in the Pacific and Middle Atlantic regions, on the other hand, increased relative to the national average. Overall, six of the nine cen sus regions followed the pattern of wage con vergence before 1979-81 and wage divergence after that period. The relative wage gains and losses across the nine regions combined to in crease the standard deviation from 0 . 0 6 8 in 197981 to 0.086 in 1985-87, which is roughly the same level of dispersion found for the first period. instance, as shown in table 2 , if workers were identical in all regions (or, at least, if the compo sition of worker attributes was the same) and only characteristic prices varied, four of the nine regions would exhibit a convergence/divergence wage pattern. These four cases are consistent with the actual wage patterns of convergence and divergence. The two regions in which price differentials did not yield the desired pattern, even though the actual wage pattern did, were the West South Central and South Atlantic regions. In both cases, differences in the charac teristic levels were consistent with the actual wage patterns and were large enough to bring these patterns into line. Which of the worker characteristics appears to contribute most to these patterns? Three catego ries of variables were considered: human capital variables, industry variables, and occupation vari ables. The most striking result (which is not shown in the tables) is that regional differences in the wage premiums paid in various industries virtually never emerged as the dominant cate gory. Rather, human capital dominated in most cases, being the largest contributor in 1 6 of the 28 cases for the price component, and in 17 of the 28 cases for the level component. Components of Intertemporal Components of Regional Wage Differentials Which of the two components accounts for the switch from convergence to divergence? One way to address this question is to consider the number of cases in which one component or the other dominated the regional wage differential for all three periods. This could be interpreted as indicating that the same “structure” that led to wage convergence also led to wage divergence. Looking only at the cross-sectional results, as shown in table 2, provides a mixed answer. For the six regions that followed the convergence/ divergence pattern, differences in characteristic prices dominated the regional/national wage dif ferential for three regions for all three periods, differences in characteristic levels dominated one region, and the effect was split for the remaining two regions. Tallying up the total number of cases in which differences in charac teristic prices dominated the regional wage dif ferentials results in about the same percentage of cases— about 6 0 percent. Another way to evaluate the importance of each source is to determine the wage patterns gener ated if only one of the components varied. For Regional Wage Changes The previous examination of the sources of regional wage differentials looked at three sepa rate cross sections from different time periods. The next step is to examine how these regional wage differentials changed over time. As men tioned earlier, equation (4) provides a frame work to account for the various components of this wage change. Table 3 displays the components of nominal wage changes for each region between the three time periods. For example, the 1.6 percent reduction between the first two periods in the wages of the East North Central region relative to the national average can be attributed to primar ily two effects. The first is the main effect (column 1 ), which is the change over time in characteristic levels for the region relative to the nation. If all other effects were zero, then these changes in worker characteristics would cause the regional wages to diverge from the national average rather than to converge, as they actually do. The positive sign for this component indi cates that the difference in the characteristic lev els that favored this region over the nation was greater in the second period than in the first. T A B L E 2 Components of Regional Wage Differentials Relative to the National Average (1 ) (2 ) ( b R- b N ) X N ( X R- X N ) b N Census R egion Year (3 ) Actual W age Differe New England 1973-75 1979-81 1985-87 -.0064 -.0469* .0 1 9 8 * -.0096* -.0123 .0138 -.0208 -.0592 .0305 Middle Atlantic 1973-75 1979-81 1985-87 .0547* -.0005 .0429 -.0363* .0504* .0904 .0327 .0758 1973-75 1979-81 1985-87 .0453 .0117 -.0 0 1 0 * .0064* .0490 .0337 .0025 1973-75 1979-81 1985-87 -.0289 -.0137 -.0447 -.0461* -.0483* -.0605* 1973-75 1979-81 1985-87 -.0332* -.0459* -.0396* -.0303 -.0027 -.0076 1973-75 1979-81 1985-87 -.0 8 6 1 * -.0714* -.1047* -.0695 -.0264 -.0544 -.1589 1973-75 1979-81 1985-87 -.0915* -.0496* -.0471* -.0176 -.0 0 2 0 -.1078 -.0502 -.0524 1973-75 1979-81 1985-87 -.0 -.0 0 2 1 -.0270 -.0 2 2 0 * -.0446* -.0436 East North Central West North Central South Atlantic East South Central West South Central Mountain Pacific 1973-75 1979-81 1985-87 .0 2 2 1 3 1 6 * .0 1 0 1 -.0158 .0782* .1280* .1253* .0133* .0007 -.0 0 0 6 .0495 .0256 .0333 -.0 8 6 1 -.0699 -.1089 -.0 6 2 2 -.0448 -.0463 -.1 0 1 1 -.1524 .0 1 0 6 .1063 .1326 .1427 NOTE: Column 1 is the effect o f differences in characteristic price between the region and the nation; colum n 2 is the effect o f differences in characteristic levels between the region and the nation. Columns 1 and 2 d o not add up to colum n 3 because o f a residual com ponent not shown in the table. Asterisks denote the dominant com ponent for each time period and region. SOURCE: Author’s calculations. Offsetting the effect of changes in characteris tic levels are the changes over the time periods in characteristic prices (column 4). If everything else remained the same, these changes in inter temporal prices would result in East North Cen tral wages converging to the national average by 2 . 8 percent. In determining which components contribute most to wage changes, two criteria were used. First, the signs of the components must be con sistent with wage convergence between the first and second periods and with wage divergence between the second and third periods. Second, the components should account for a large share of the total wage change. The asterisks in table 3 indicate the pairs of components that are consistent with the conver gence/divergence wage pattern. For the two com ponents that are based on the intertemporal change in characteristic prices (columns 3 and 4), 1 2 of the possible 18 pairs of estimates are consistent with the convergence/divergence wage pattern. The components related to inter temporal changes in characteristic levels (col umns 1 and 2) contain only five pairs. Further more, the components related to changes in T A B L E 3 Components of Intertemporal Changes in Regional Wage Differentials Time Region Span .0 0 2 -.0 0 1 3-2 .0027 .005 .0 1 0 -.0 2 0 -.015 -.0003 .0 1 0 .003 -.030 .013* -.0 1 2 * .028* -.023* .035* -.033* .0 1 6 ’ -.0 3 8 ' .009 .0 1 0 .017' -.0 0 1 ’ .0 2 0 * -.0 1 1 * .0 2 0 * -.0 1 6 * .0 6 8 * -.0 3 6 * .057' -.050' .003 .0 0 6 * .043 .063 .0 0 2 -.0003* .0 0 1 .0 0 1 .057' -.0 0 1 ’ .037* -.015* -.003 .032 -.0 1 0 * .038* -.028 -.0 2 0 3-2 -.003 .0 1 6 2 -1 -.008 -.0 0 1 * .014* .008* -.0 2 0 * -.0 1 6 * o o VO * 3-2 -.0 0 1 * * -.058 .044' -.0 1 6 2-1 1 1 -.059* .034* -.0 1 0 .0 -.0 0 1 -.039 .091 -.0 2 1 3-2 2 -1 .082 -.0 3 6 -.0 2 0 .005 .008 3-2 (6 ) -.014 .023* -.017* 2 -1 .0 2 0 * -.004* .025* (5) -.003* .007* .007* -.0 0 1 * 3-2 Pacific -.005 -.003 .0 6 1 .0 0 6 2 -1 Mountain -.008* .017* .0 0 2 * -.004* 2 -1 West South Central -.034 .030 3-2 East South Central -.003 -.004 .0 0 2 2 -1 South Atlantic (4) 3-2 2 -1 West North Central (3) 2 -1 3-2 East North Central (2 ) 00 Middle Atlantic (l) i b New England ___________________________________ C om pon ents i b N*J Census - .0 1 1 .0 1 6 .0 1 6 -.0 2 2 .067’ -.053’ .026 .0 1 1 NOTE: Time spans are denoted as 1 (1973-75), 2 (1979-81), and 3 (1985-87). The notation 2-1 represents the difference between the first two periods, and 3-2 represents the difference between the latter two periods. The com ponents are (1 ) main effect, (2 ) interaction effects, (3 ) time-interaction effects, (4 ) regional time-interaction effects, (5 ) the sum o f the four effects, and (6 ) the actual change in the regional wage differential (relative to the national average) between the two time periods. The asterisks indicate the com ponents that are consistent with the convergence/divergence wage pattern. SOURCE: Author’s calculations. prices (again columns 3 and 4) claim the largest share, on average, of the total wage changes. Consequently, it appears (as the trade theory suggests) that differences in characteristic prices account for the larger share of nominal regional wage changes over the three time periods. 14 Therefore, this simple nonparametric test of counting the number of consistent results sug gests that intertemporal changes in worker characteristic prices account for much of the convergence as well as the divergence of wages. ■ 14 Dickie and Gerking (1988) point out that omitted variables, particularly the lack of detailed human-capital variables, could bias the accounting method toward attributing too much importance to characteristic price differences. They find, using another data set that contained unusually detailed measures of worker and workplace characteristics, that they could not reject the hypothesis of equal coefficients across regions. This omission seems less critical for this study, since we look at the change over time in coefficients of the same set of variables within the same regions. It would seem that in order for omittedvariable bias to be significant, the relative contributions of each variable would have to vary substantially over time, which is not supported by the results from the previous section. Consequently, basic changes in the way that worker characteristics were valued by the regional markets must have occurred around the turn of the decade. Trade theory suggests various types of market imperfections as possible candi dates, including differences in production tech nologies and factor-market distortions. The backto-back recessions in 1980-82 and the collapse in oil prices shortly thereafter certainly have taken their toll on regions such as the West South Cen tral, while having little effect on others, such as the Pacific and New England regions. The rela tive effects of these events among regions can be partially explained by the slow adjustment of labor markets and the differential impact of oil prices between energy-using and energyproducing regions. Considering the three categories of worker characteristics defined in the previous section offers further insight. As before, the industry vari ables played very little role in accounting for intertemporal changes in the regional wage dif ferentials (these results are not shown in the tables). However, unlike the cross-section analy sis, occupation variables clearly dominated. For example, with respect to component four (dif ferences in prices), occupation variables were the dominant category in 1 3 of the 18 cases. 15 This result supports some of the speculation made by various authors about possible reasons why worker characteristic prices may not be equal across regions. Farber and Newman (1987) conjecture that characteristic prices may not necessarily converge because of occupationspecific demand disturbances. Topel (1986) shows that disequilibrium in local labor markets results primarily from stochastic disturbances in labor demand. IV. Conclusion After converging for almost half a century, nom i nal regional wages have diverged since 1 9 8 0 . This paper attempts to isolate the source of this switch in direction either as an intertemporal change in the market prices for worker attributes or as an intertemporal change in the levels of worker attributes. For nine census regions between the periods 1973-75, 1979-81, and 198587, results using individual workers from the CPS show that differences in characteristic prices account for a major share of the change in regional wages relative to the national average. Furthermore, virtually all of this intertemporal change in characteristic prices is found in the occupation coefficients; industry and worker characteristic variables account for very little. Theory suggests that the prices of worker char acteristics will converge in the presence of free commodity trade and in the absence of market imperfections. Various types of market imperfec tions were suggested as possible sources of the divergence of regional wages. For example, incomplete information, a mismatch between worker skills and job requirements, and institu tional barriers to mobility can lead to incomplete adjustments to recent changes in the structural demand for labor. A recent study estimates that it takes as much as a decade for local labor markets to adjust fully to such shocks (Eberts and Stone [1989]). Another possibility for nominal wage diver gence is changes in the regional prices of hous ing and other nontraded goods that deviate from ■ 15 Farber and Newm an (1987) also find that the worker characteristics that accounted for much of the cross-sectional accounting of regional wage dif ferences were different from the worker characteristics that accounted for the majority of the intertemporal changes in regional wage differentials. the national average. Because this study did not adjust for regional cost-of-living differences, it may be possible that wage differentials simply compensate workers for higher housing costs. However, this argument runs counter to the pre dicted results of free trade among regions, once equilibrium has been established. If goods are freely traded, then firms would be hard pressed to pay higher wages in some regions than in others, unless employers were compensated by differences in production technologies and worker productivity. Therefore, for cost-of-living differences to explain the results, workers in areas with higher labor costs coincidentally would have to be more productive. There are no compelling reasons why high living costs and high worker productivity should exist concur rently in equilibrium. Two exceptions to this general statement are possible: First, site-specific attributes could enhance firms’ productivity. Firms would move into the more productive region, bidding up the price of land and the price of labor, everything else being equal. The second possibility is that with the slow adjustment to shocks, we are simply observing these effects in disequilibrium. The findings that differences over time in characteristic prices account for a majority of the changes in regional wage differentials does not necessarily diminish the importance of migration in explaining differences in regional growth. Rather, the analysis suggests that these flows have not changed the composition of regional labor forces significantly enough to make them the dominant factor in explaining changes in regional wage differentials. The traditional migra tion patterns of South to North and East to West are less pronounced now than in the past. For merly, the primary migration pattern was toward the West, particularly for college graduates look ing for job opportunities. More recently, the South is receiving many younger persons from the West and North. If stochastic disturbances have changed the course of regional wage differentials, then it is interesting to speculate why these shocks have had such an impact in a relatively short period of time, when for so many decades the workings of efficient markets and equalizing migration flows seemed to prevail in forcing regional wages to converge. Several possibilities come to mind: increased foreign competition, the collapse of oil prices in the early 1980s, and the severe back-toback recessions of 1980-82. These recessions hit some regions harder than others, producing different patterns of change in regional wage differentials. The West South Cen tral states of Texas and Louisiana were particularly hurt as the bottom dropped out of oil prices. This downturn thwarted the sizable gains that region had made in previous years in narrowing its wage gap. The farming states of the West North Central region were also severely affected by the reces sion and the ensuing farm crisis of the early 1980s. After converging toward the national aver age throughout the 1970s, wages in this region diverged significantly, falling from 7.0 percent below the national average at the beginning of the 1980s to 1 0 . 9 percent below the average toward the end of the decade. Wages in some regions continued to grow faster than the national average in spite of the recession. For example, the Pacific region, especially California, was only mildly affected, with its regional wage differential expanding by a percentage point between 1979-81 and 1985-87. Factors other than economic shocks could also contribute to the wage divergence. One possibility is state tax policies. The late 1970s and early 1 9 8 0 s saw the phasing out of substan tial federal grant programs to states and munici palities. Many of these programs were designed to help equalize the fiscal burden across regions. As these funds dried up, many state and local governments found it necessary to raise tax rates to fund the existing programs, while others decided to curtail the programs. These different responses could lead to an increase in regional tax rates, which in turn could affect the location of firms and ultimately the demand for labor. Will these factors persist? If history is any guide, the answer is no. The long-run trend of regional wage convergence has been interrupted only once in the last century. That episode lasted 2 0 years, embracing a postwar period and a much deeper and protracted recession than the one that greeted this decade. Consequently, it appears that shocks of this kind will eventually dissipate as the regions’ economies regain a more equal footing. However, many states and localities are not content to wait the decade or so that it takes for these forces to play themselves out. Many areas have pursued vigorous economic development efforts to help quicken the pace of adjustment. As long as these efforts attempt to remove market inefficiencies and strengthen the region’s comparative advantage, they are socially desira ble. One would expect that as regions continue to develop and mature— and barring further shocks of recent magnitude— the long-run trend of regional wage convergence will return. Beeson, Patricia and Eberts, Randall W., “Identi fying Productivity and Amenity Effects in Interurban Wage Differentials,” August 1989 (forthcoming). Economics and Statistics, Review of Bellante, Don, “The North-South Differential and the Migration of Heterogeneous Labor,” March 1979, 166-175. American Economic Review, 69, Browne, Lynn E., “Shifting Regional Fortunes: The Wheel Turns,” Federal Reserve Bank of Boston, May/June 1989, 27-40. Review, New England Economic Dickie, Mark and Gerking, Shelby, “Interregional Wage Differentials in the United States: A Sur vey,” in Alan Schlottmann, et al., eds., Boston: Kluwer Academic Publishers, 1988. tion and Labor Market Adjustment, Migra Easterlin, Richard A., “Long Term Regional Income Changes: Some Suggested Factors,” in Gerald AP. Carrothers, ed., Papers and Pro ceedings of the Regional Science Association, 1958, 4, 313-25. Eberts, Randall W. and Stone, Joe A., “Wage and Employment Response to Shocks in Local Labor Markets,” presented at the Western Economic Association Meetings, Lake Tahoe, Nev., July 1989. Farber, Stephen C. and Newman, Robert J., “Accounting for South/Non-South Real Wage Differentials and for Changes in Those Differ entials Over Time,” May 1987, 215-23. Statistics, Review of Economics and 69, Gerking, Shelby and Weirick, William, “Com pensating Differences and Interregional Wage Differentials,” August 1983, 483-87. tistics, Review of Economics and Sta 65, Hanoch, Giora, “An Economic Analysis of Earn ings and Schooling,” Summer 1967, 310-29. Resources, Journal of Human 2, Markusen, James R., “Factor Movements and Commodity Trade as Complements,” Mav 1983, 341-56. of International Economics, Journal 14, Schooling, Experience, and Earn Mincer, Jacob, New York: National Bureau of Economic Research, distributed by Columbia University Press, 1974. ings, Regional Economics: A Study in the Economic Structure, Stability, and Growth of Regions, New York: McGraw-Hill, Nourse, Hugh O., References 1968. Oaxaca, Ronald, “Male-Female Wage Differen tials in Urban Labor Markets,” October 1973, 693-709. Economic Review, International 14, Roback, Jennifer, “Wages, Rents and Quality of Life,” December 1982 1257-78. Journal of Political Economy, ,90, Sahling, Leonard G. and Smith, Sharon P., “Regional Wage Differentials: Has the South Risen Again?” February 1983, 131-35. Review of Economics and Statis 65, Topel, Robert H., “Local Labor Markets,” Journal of Political Economy, 1986,94, SI 11-43. tics, W hy We D o n 't K n ow W hether M oney Causes O utp u t by Charles T. Carlstrom and Edward N. Gamber Charles T . Carlstrom is an economist at the Federal Reserve Bank of Cleveland. Edward N. Gamber is an assistant professor of economics at the University of M issouri-St. Louis and is a visiting economist at the Federal Reserve Bank of Cleveland. The authors wish to thank John Carlson, Randall Eberts, William Gavin, William Osterberg, and Alan Stockman for helpful comments. In addition, we wish to thank Susan Black for valuable research assistance. Introduction Macroeconomics has undergone a revolution in the past 2 0 years, in which significant challenges have been made to supposedly well-established theories and facts. Among the most important of these prevailing theories is the positive correla tion between money and real output. Traditionally, most economists and policy makers have interpreted this correlation to imply that Federal Reserve open market operations could affect real output. This interpretation has persisted in spite of weak and sometimes con tradictory empirical evidence. Unfortunately, we cannot attempt to examine all of the existing evidence on the direction of causality between money and output. Instead, this paper examines whether Granger-causality is a valid test for cau sality and what can be inferred from existing tests of Granger-causality. The answers to these ques tions are of paramount importance, since most policymakers assume that money causes output in a consistent and reliable way. This correlation is illustrated in figures 1 , 2 , and 3 using three measures of money: base, M l, and M2. 1 ■ The usual method of distinguishing among competing economic theories involves econo metric testing. However, as is well known (see, for example, Black [1982]), econometric models indicate correlation, but not causality. Even the econometric technique of Granger (1969) does not necessarily identify causality as the term is commonly understood. We will show in the fol lowing section that the concept of Grangercausality is not robust to changes in the underly ing model of the economy .2 In other words, it is impossible to interpret Granger-causality inde pendent of theory. Given this, sections II through IV examine models that try to explain the correlation between money and output. Traditionally, this correlation was explained by assuming some type of nominal rigidity (either prices or wages). Tobin (1970), however, showed that the correlation between money and output could be a result of the Federal Reserve’s operating procedure and that it did not necessar ily imply that changes in money caused output changes. Section III shows that if the Federal Reserve accommodates increases in output with a corresponding increase in the money supply, 1 The series is detrended using a Hodrick and Prescott (1980) filter. Figure 4 illustrates this method as it is applied to real output (G N P ). ■ 2 See also Cooley and LeRoy (1985). F I G U R E Real G N P and Monetary Base Percent deviations 2.0 1.5 - 1 1.0 0.5 — •. • 0.0 V. •*'•. L • V a % -0.5 - 1.0 4 \ \ •/Monetary v m * a* :: I \V A uul a 7* i’ %• # I It\ •/ 1 J •M * %mV • I N %* \J + H *VJ • • • : Real : GNP ••• 2.0 -2.5 / # •• -1.5 - « ••• • *•" base * I ;. .a frt 1* AII* ; / 1 ** /• v • r \ •*!/>/• •/ \l %IVX ' *a -t« a Ayv .'A'. fSr V. 7 V ” _ % *• • a'a •• * w then one would expect to observe a positive correlation between output and money even though money is not causing output. Real business cycle theorists have recently argued that the correlation between money and output could be due to reverse causality; that is, output can cause money independent of the Federal Reserve’s reaction function. Section IV examines a model by King and Plosser (1984) showing that M l and output are correlated because increases in real output cause increases in the demand for financial intermediation. This increased demand leads to the expansion of broader monetary measures, such as M l and M2, even though changes in money have no influ ence on real output. Section V reviews the empirical evidence uncovered in these theories to help ascertain the direction of causality in the money-output corre lation. Section VI concludes with a discussion of policy implications. —l i i i Li i—l i L_i i i i Li i i i Lj i i i 1 L,i .i i i L i 1963 1968 1973 1978 1983 1988 NOTE: Sample period is from 1959: IQ to 1988: IVQ. SOURCES: Data Resources, Inc., and Board of Governors of the Federal Reserve System. I. Granger-Causality Causality is a very elusive concept. In practice, most people define x causing to mean that a change in x leads to a change in As an anal ogy, we would implicitly assume that if we could cause a low-pressure system to appear over a city (all else remaining constant), then there would be a high probability that rain would fall. This causality usually means that if low-pressure sys tems cause rain, then low-pressure systems must precede rain. As can be seen in figure 3, M2 appears to lead GNP. Does this chronology imply that M2 causes GNP? The Granger definition of causality requires two assumptions. As stated by Granger and Newbold (1986, p. 220): a) The future cannot cause the past. Causality can only occur with the past causing the present or future. b) A cause contains unique information about an effect that is not available elsewhere. According to the first assumption, then, if M2 always leads changes in GNP, we can logically infer that GNP does not cause M2. Does this mean that we can conclude the alternative, that M2 causes GNP? Consider the following example. Suppose that a group of individuals always listens to weather forecasts and that these fore casts are always accurate. Further, suppose that these people decide to carry umbrellas on days that rain is forecasted. Clearly, carrying an umbrella and rain will be correlated, and carry ing an umbrella will precede a rainstorm. y F I Real G N P and M l Percent deviations G U R E y. F I G U R E Real G N P and M2 Percent deviations 2.0 1.5 - 1.0 0.5 0.0 -0.5 - *1 , * • • ••• Vf 1.0 — \r* n M* V* •• I j r* v I r mJ\ v r ri* /• u U'l« M* -1.5 “ M2 - 2.0 1 Real GNP #I A' /# IV t* il 1 • 1 A ' M A •:* I •*, l . ; . • 1 , « ®# .* • / a \ • •. IV •• III • IIW'* # • • • • • • t • •• •* I J jt f L i i i i Li -2.5 i i i i Lj i i i Li i i i 1 » ■■* 1 « « ■« 1963 1 9 6 8 1973 1978 1983 1988 NOTE: Sample period is from 1959:IQ to 1988:IVQ. SOURCES: Data Resources, Inc., and Board of Governors of the Federal Reserve System. Logged Real G N P and G N P Trend According to the first assumption of Grangercausality, rain cannot cause umbrella-carrying. Yet, clearly, meteorologists would reject the con clusion that umbrellas cause rain. The problem with our umbrella and rain example is that assumption a) is violated. This assumption is also frequently violated in many econometric tests. A third variable that uniquely causes people to carry umbrellas is omitted. Strictly speaking, rain does not cause umbrellas, but the that rain may occur causes people to carry umbrellas. Expectations are not formed in a vacuum, however; low-pressure sys tems in this example could be shown to cause both umbrella-carrying and rain. Neglecting this third variable would cause one to conclude that carrying an umbrella Granger-causes rain. Because of the importance of expectations in economics, a variable, that precedes another variable, will frequently not cause Variable x may depend on the expected value of caus ing x and to be correlated. Since expectations depend on numerous variables that are, in prin ciple, observable by the econometrician, one could conceivably conduct a Granger-causality test by including all relevant variables. The econ ometrician, however, would need to have a welldefined model of how expectations are formed. It is therefore extremely important that Grangercausality tests be interpreted in light of the theory that one is trying to test. Consider the formal definition of Grangercausality. Let H , be all the information available in the universe at time Let and be two random variables within this universe. Granger says that x causes (does not cause) if expectation x, y y. t. Log y y xt yt y * k (=) ^ O w l n r- **) for > 1 , where F(.|.) is the conditional prob ability density function of given f l o r l) , and is defined to be the universe less Suppose that these conditional distribution functions are equal. If jc and are correlated, it follows that there must exist a third variable in O that causes both x and For example, let denote the occurrence of rain and let de note the occurrence of umbrella-carrying. Leav ing umbrella-carrying out of the information set does not affect the conditional distribution of rain or, in other words, weathermen can accu rately predict rain without seeing whether peo ple are carrying umbrellas. Because the entire universe, including low-pressure systems, is assumed to be in the information set, this exam ple correctly predicts that umbrella-carrying does not Granger-cause rain. x t, Clt - x t xt. y t +k t y y t y. x Sims (1972) showed that Granger-causality is identical to the concept of exogeneity. In other words, jc Granger-causes if x is exogenous to and is not exogenous to x. A variable x is exogenous to if the occurrence of x is inde pendent of the occurrence of Similarly, a var iable is not exogenous to x if the occurrence of is dependent on x occurring. Thus, the occurrence of rain is exogenous to whether people carry umbrellas: rain will fall regardless of whether people carry umbrellas. The converse is not true, however; if it starts to rain, people will tend to carry umbrellas. At first glance, Granger-causality or exogeneity seems to be a reasonable definition of causality. However, it ignores the case of bivariate causal ity, where two variables cause each other. For example, rain causes puddles, and the evapora tion of puddles causes rain to fall at a later date. To make Granger-causality operational, the uni verse of information must be restricted and the moments of the conditional distribution func tions must be tested for equality. The universe of information is restricted by theory. In practice, the distribution functions are said to be equal if their first moments (the means) are equal. Test ing for Granger-causality usually involves the fol is said to Granger-cause lowing: A variable (not Granger-cause) with respect to the information set if y y y y y y y x It E (yt+k l7/) ^ (=) y E (yt+kUt - x t ) f o r k > 1. Because we do not consider all moments of the distribution, and we do not use all of the information set, Granger-causality as practiced is neither a necessary nor a sufficient condition to determine the direction of causation between a* and Consider the case where all the relevant information in the universe is included in a Granger-causality test, but only the means are tested to see if they are equal. If the means were found to be unequal, then one could logically infer that x must cause If the means were found to be equal, however, then one could not infer that a; did not cause Now consider the second assumption in the case where all the moments can be tested, but the universe of information is restricted in an ad hoc manner and an important determinant of is accidentally omitted. Equality between the conditional distribution functions necessarily implies that x does not cause However, if the conditional distributions are not equal, then we cannot infer that jc causes This is the case in our example: umbrellas help to predict rain and y. y y y y. y thus Granger-cause rain if low-pressure systems are excluded from the information set. Since any operational test of causality involves restricting both the moments of the distribution functions to be tested and the information set in the universe relevant to the problem, employing a Granger-causality test exposes one to the risk of incorrectly rejecting causality when it is pres ent and incorrectly rejecting the assumption of no causality when causality is not present. The econometrician can seek the direction of causal ity using a Granger-causality test only by using theory to determine which variables are helpful in predicting However, even after choosing variables based on some theory, a specification test should be conducted to help ensure that important variables have not been omitted. It should be clear from this discussion that Granger-causality is neither a necessary nor a suf ficient test for the existence of true causality. First, if bidirectional causality exists, then Granger-causality cannot indicate the presence of causality. Second, even when bidirectional cau sality is not present, the Granger-causality test may fail to identify whether causality is present if the information set excludes relevant variables or if all moments of the conditional distributions are not tested for equality. In addition, Grangercausality is not a useful test for showing the presence of contemporaneous causality. Sections II and III present representative theo ries that have been developed to explain the money-output correlation. Section IV then inter prets the econometric evidence that has been uncovered in light of these theories and the problems discussed above. y t+k . II. Money Causes Output Most economists currently favor the interpreta tion that money causes output. They believe that some nominal rigidities, or price/wage sluggish ness, allow changes in nominal variables, like money, to have real effects. These rigidities can be motivated by nominal wage contracts (Fischer [1977], Gray [1976]), or by incomplete information (Lucas [1972, 1977]). For expositional ease, we consider the nom i nal wage contracting model as exemplified by Fischer. In his model, agents in the economy have rational (model-consistent) expectations, but wages are “sticky” because of the existence of long-term nominal wage contracts. Further, Fischer assumes that employment is demanddetermined; that is, employment is always chosen so that the real wage is equal to the mar ginal productivity of labor. Thus, changes in the money supply that were unexpected at the time the contract was signed will have real effects. Unanticipated increases in the money supply will cause prices to be higher than expected and will cause the real wage to be lower than expected. The decline in the real wage lowers the marginal cost for firms to hire additional workers, leading to an expansion of employment and thus output. Consider a scaled-down version of the model analyzed by Hoehn (1988). In this example, con tracts will not be overlapping, and the only source of uncertainty will be from the moneysupply process. Assume that the aggregate pro duction function is Cobb-Douglas, that is, = where and are real output and the labor supply, respectively. Because wages are assumed to be demand-determined, we set the real wage equal to the marginal productivity of capital. Taking logarithms gives Nty, (1 ) Yt Yt Nt wt - p t = In ( wt p t 7 ) - ( 1 y )n t , - nt where , , and are the natural logarithms of wages, prices, and employment. Labor supply is assumed to be of the following form: (2 ) n*t = y + wt ~ P / ) for /J0, /? j > 8 0 0 . Setting labor supply equal to labor demand, one can solve for the real wage rate that clears the market. From this equation, it is assumed that wages are chosen so that the labor market clears on average.3 This gives the following equation for nominal wages: (3) u'*, = Et _ xp, + [In ( 7 ) - (1 - 7)/3 0 ] / , \ where / = [ 1 + 0 , ( 1 - 7 )] To close the model, we must posit a form for money demand and money supply. Money de mand is taken to be the simple quantity equation, that is, In logarithmic form, it is Md - KPtYt . (4) m dt = p, + y, + k. For our purposes, this year’s log of money supply is equal to last year’s money supply plus a random shock. That is, = _, + t, where the shock is assumed to be an inde pendently, identically distributed random varia ble over time. With these assumptions, output equals et ■ 3 mst mt t Actually, this assumption is not quite true. W ages in Hoehn’s model are chosen not so that ENd = /Vs, but so that Eln (Nd ) = End = lnNs = ns. (5) y, = A + 7 «t , A = y[(30 B xln where + ( 7 )]/For this simple case, in which contracts do not overlap and there are no shocks other than those to the money supply, changes in output depend only on the shock to this period’s money, If one were to randomly determine different reali zations of t , and were then to graph money supply and output against time (different realiza tions of ), one would obtain a picture very similar to that given in figure 1. In this case, money causes changes in output. However, because changes in money and output occur contemporaneously, money does not Grangercause output. Equation (5) is also the output equation that results from a simple linearized version of the Lucas (1972, 1977) model. Here, workers con fuse nominal and real shocks. Unanticipated increases in money result in higher nominal wages, which workers confuse with higher real wages. They do not know the extent to which higher wages reflect an increase in the relative price of their product or an increase in the gen eral price level. Unanticipated changes in the money supply will cause increases in output as workers rationally mistake this nominal shock for a change in their real wage. Models of the type discussed above were orig inally developed in response to the lack of empirical and theoretical support for traditional Keynesian and monetarist models. Both the Lucas and the Fischer models have recently come under attack. Barro (1977) shows that con tracting models such as Fischer’s are inconsistent with maximizing behavior. He argues that there is no a priori reason why labor should be demand-determined in these models. In addition, economists question why firms have not indexed their wages, because sticky wages result in alleged output swings at both the firm and the macro level. Ahmed (1987) also presents empirical evidence showing that nom i nal wage contracting is not important for explain ing output movements in Canadian data. Although Lucas’ model is consistent with maximizing behavior, it also lacks empirical support. Mishkin (1983) and Boschen and Grossman (1982), for example, find evidence against the equilibrium monetary explanation of the business cycle. The following section shows why the Federal Reserve’s operating procedure may cause money and output to be correlated. et . e et III. Post Hoc: Does The Federal Reserve Cause Christmas? Figure 5 plots a scatter diagram of quarterly changes in the monetary base versus quarterly changes in output. Fourth-quarter points gener ally lie to the northeast of the first- through thirdquarter points. Therefore, money and output are both higher on average in the fourth quarter, or around Christmastime. One could erroneously conclude that Federal Reserve policy causes hol iday spending. Clearly, causality in this case goes the other way. Output increases in the fourth quarter be cause of holiday spending, and the Federal Reserve, attempting to remove the seasonality from the interest-rate series, accommodates this higher output by increasing the money supply. This is an example of a point given by Tobin (1970) in his seminal article, “Money and Income: Post Hoc Ergo Propter Hoc?” meaning “after this therefore because of it.” Tobin’s argument was that a positive correlation between money and output may be the result of the Fed eral Reserve’s operating procedure and not a reflection of the common belief that money causes output. Real G N P and Monetary Base Monetary base, not seasonally adjusted Instead of presenting Tobin’s model, we show how the operating procedure of the Federal Reserve can cause one to incorrectly conclude that the Federal Reserve causes, or at least influ ences, business cycles. Consider the following variation of the model presented in the previous section: Let output be Cobb-Douglas, so that the log of real wages will again be given by equation (1). Further, assume that the log of the labor supply is given by the following equation: (6 ) n st = (7) • IQ O IIQ 2 (8) A • §as$> m °:° • s " a • -1 -2 . . v 00 0 (9 ) L LI____ . 1L _ ,1 J ____ iI____ I____ 1_____ I____ 1 _____L____ I____ I____ -4 - 3 - 2 - 1 0 1 2 )]/ + J P 2 rt . rt = r + rjt . t] . 1V 7 Temporary changes in interest rates, t , can result because of either shifting tastes or tempor ary changes in government expenditures. Incor porating this variable into equation (7), we see that output depends positively on the innovation in real interest rates today. " 0 [p0 + jSj/n ( r ■ IIIQ — n, = Real interest rates in the economy are assumed to fluctuate randomly around a con stant mean : A IVQ 1 O+ (31( w t - p t ) + p 2rt for 0 O , j8 j , /32 > 0. This equation differs from equation (2) be cause the labor supply is also assumed to be influenced by the real interest rate, rt . Equation ( 6 ) assumes that the labor supply depends posi tively on the real interest rate, because of the intertemporal substitution effect. That is, when interest rates are high, workers transfer consump tion from today until tomorrow to take advantage of the high real rate. Consumption is reduced, thus increasing the marginal utility of consump tion in the current period. This, in turn, increases the incentive for agents in the economy to work additional hours in order to consume more today. Instead of assuming that there are long-term nominal wage contracts, this model assumes that wages vary to clear the market continuously so that money does not influence output. By equat ing the real wage in equations ( 1 ) and (5), we solve for the equilibrium amount of labor sup plied (demanded) in this economy: 3 • 0 3 Real GNP, not seasonally adjusted 4 - yt = y[p 0+ p jn (y ) + P2r ] j + y] \32 r j r To close the model, we assume that money demand is given by equation ( 4 ) and that the Federal Reserve follows a nominal interest rate rule: (10) mst = b + k(R t - r ), and A > Rt - rt Etp t 0, pt. where + +j Nominal interest rates are assumed to be the sum of the real rate plus expected inflation over the next period. Using equations (4), ( 8 ), (9), and ( 1 0 ), the reduced form for the nominal interest rate is given by the following equation: (11) / ? , = r + »;, [ ( 1 / ( 1 + A ) ) + 702/ / (l + A)]. Innovations in the real interest rate are assumed to be temporary. An increase in the real interest rate causes policymakers to expand the money supply in order to stabilize nominal interest rates. Prices are then temporarily high and deflation is expected over the next period, which will offset the increase in the real interest rate. When A approaches infinity, the nominal interest rate approaches the long-term real inter est rate, That is, when A approaches infinity, the Federal Reserve is following an interest-rate r. pegFrom equation (11), the reduced form of the money-supply equation is given by (12) The above model illustrates how an interestrate target can produce a positive correlation between money and output. The example was extremely simple and predicted that money and output would move contemporaneously One could likewise construct examples in which money leads changes in output and would thus appear to cause changes in output. For example, consider an economy in which money has no real effects, but in which agents are able to predict future output. The prospect of higher future output will cause agents to borrow (or save less) in an attempt to smooth their con sumption stream over time. This increased bor rowing will boost interest rates. If the effect on output today from an increase in interest rates is negligible, then changes in money will occur be fore changes in output when the Federal Reserve pursues an interest-rate peg. In this economy, money leads, but does not cause, output. The next section discusses another mecha nism in which output can cause changes in money. Unlike the model presented in this sec tion, the mechanism will not come from the Federal Reserve’s operating procedure, but will result from the public’s willingness to hold cur rency versus either demand or time deposits. m st = b + A [(l/ ( 1 + A)) + yfi2J / ( l + A)]t 7,. If one were to randomly determine different realizations of 7 7 ,, and were then to graph money supply and output against time (different realizations of 77 f ), one would again obtain a picture very similar to that given in figure 1. A temporary increase in interest rates causes peo ple to supply more labor today. This occurs since high real interest rates imply that, on the margin, individuals greatly value consumption today, causing them to work longer hours today. The increase in interest rates also causes the Federal Reserve to expand the money supply in order to smooth nominal interest rates, which causes a temporary rise in prices. This example implies that, on average, prices will fall over the next period, leading to a decline in the nominal interest rate. Unlike the example given in the previous section, interest rates in this model cause changes in both output and money. Thus, money and output are posi tively correlated. Like the example given in sec tion III, however, interest rates do not Grangercause output, because interest rates and output occur contemporaneously. IV. Output Causes Money Real business cycle theorists typically assume that the cause of business cycles is either a shock to consumer preferences or a shock to real pro ductivity. 4 Because an indirect measure of these shocks can be obtained through the use of Solow residuals (see Solow [1956]), theorists have tended to concentrate on technology shocks as a source of business cycle fluctuations. Real business cycle theory has been successful in explaining the quantitative aspects of business cycles. These include the standard deviations of— and comovements among— real variables such as output, investment, consumption, and hours worked. In contrast, monetary-driven bus iness cycle models have concentrated on explaining the qualitative aspects of the correla tion between money and output. 5 Because real business cycle models do not include a role for money, they have been criti cized for not explaining the comovements ■ 4 For a thoughtful exposition of real business cycles, see Prescott (1986) or Stockman (1988). ■ 5 A s noted by Stockman (1988, p. 35), “ The large-scale econometric models do not qualify because they are not true structural models in the sense of the Lucas critique of econometric policy evaluation....” Cross-Correlations of Output Variable x x (t - 4) x (t - 3 ) x (t - 2 ) x (t - 1) Real GNP .25 .45 Monetary base .08 .2 0 .29 .39 Ml .1 2 .25 .35 M2 .62 .6 8 .69 .6 8 x (t) x (t+ 1 ) x (t + 2 ) * ( t + 3) x ( t + 4) .87 .6 8 .46 .26 .44 .44 .42 .39 .37 .37 .34 .26 .2 1 .1 6 .62 .48 -.08 -.25 .87 1.0 0 .15 .29 .1 0 Interest rates -.55 -.36 -.15 .1 0 .34 .47 .51 .49 .45 Real interest rates -.57 -.41 -.2 0 .07 .30 .40 .44 .43 .41 CPI -.70 -.75 -.75 -.69 -.56 -.40 -.23 -.05 .13 NOTE: Sample period is from 1959:IQ to 1988:IVQ. SOURCES: Data Resources, Inc., and Board of Governors of the Federal Reserve System. F I G U R E 6 Real G N P and Consumer Price Index Percent deviations NOTE: Sample period is from 1959. IQ to 1988: IVQ. SOURCES: Data Resources, Inc., and Board of Governors of the Federal Reserve System. among nominal variables such as the price level, wages, and money (see Summers [1986]). How ever, as figure 6 and table 1 illustrate, the comovements among interest rates, prices, and real output are qualitatively consistent with real business cycle theory. In particular, interest rates have been contemporaneously procyclical and prices have been countercyclical since 1959.6 ■ 6 Prior to 1953, prices seem to be more procyclical. Procyclical interest rates arise in real business cycle models generated by temporary productiv ity shocks. A temporary increase in productivity today, which is expected to lead to higher out put in the future, causes individuals to borrow money in order to smooth consumption. Coun tercyclical prices arise in these models because the demand for real money balances increases when output increases. Assuming that the Fed eral Reserve does not fully accommodate the increases in interest rates and output, it follows that prices must fall. Table 1 provides further evidence that the Federal Reserve may accommodate increases in output. Note that the strongest correlations between the monetary base and output occur contemporaneously and with money lagging out put by one quarter. Real business cycle theorists argue that the correlation between the monetary base and output is the result of the Federal Reserve’s operating procedure. They point out that this correlation is small relative to the corre lation between output and broader measures of money, such as M l and M2. Table 1 shows that while the contemporaneous correlation between the monetary base (percent deviations from trend) and real GNP is only .44, the correlation between M2 (percent deviations from trend) and real GNP two quarters later is .6 8 . Although table 1 indicates that the correlation between M l and real GNP is similar to the corre lation between the monetary base and real GNP, the correlation between M l (percent deviations from trend) and real GNP is .59 if one ignores the tremendous increase in M l during 1986. While the monetary base is determined solely by the Federal Reserve, components of M l and M2, such as checking accounts, short-term time deposits, money market accounts, and mutual funds, are determined by commercial banks and the public . 7 This suggests an important role for reverse causality. The public appears to respond endogenously to future output changes by shift ing its portfolio from currency to demand and time deposits. Some mechanism must therefore serve to link output and deposits. King and Plosser (1984) develop a model in which individuals demand both currency and financial services (demand deposits). In their model, demand deposits, like other goods, are produced with capital and labor. They derive a demand curve for both inside money (financial services) and outside money (currency). They assume that the cost of making a transaction depends negatively on the real amount of inside and outside money that a person holds. The demand for both financial services and currency increases with real output in this model, explain ing why empirically both real currency and real demand deposits are correlated with real output. However, King and Plosser also show empiri cally that there is a positive correlation between nominal demand deposits and currency with real output. If one restricts their cost of transactions and assumes that with larger purchases (higher output) there is an extra cost associated with currency over demand deposits, one can also generate a positive correlation between nominal demand deposits and output. This assumption seems natural because the demand for highticket durable goods is much more procyclical than for less-expensive purchases such as serv ices. A model like this can explain the positive correlation between nominal bank deposits and real GNP. An example of reverse causality occurred dur ing the Great Depression. The monetary base grew slightly through the period, while the money supply, defined by M l, declined substan tially as depositors shifted out of demand depos its and into currency. The result was a decline in the currency/deposit ratio as output fell and banks failed. The ensuing bank failures were probably both a cause and an effect of the Great Depression. The decline in the money supply, therefore, was partly the effect of factors that caused the Great Depression, although it may also have been a contributing factor in causing the financial collapse. 8 Empirical work has not been able to distinguish this causation. ■ 7 The Federal Reserve currently can control the nonborrowed monetary base with a fair amount of precision. However, to control total monetary base, the Federal Reserve would need to alter the current administrative practices of the discount window and reserve accounting practices. See Laurent (1979). ■ 8 See Friedman and Schwartz (1963). Real business cycle models have generated a resurgence in interest to test for the direction of causality between money and output. The next section reviews this literature in light of the the ories presented in sections I through IV. V. Tests of the MoneyOutput Relationship To determine the direction of causality between money and output, economists since Sims (1972) have employed Granger-causality tests. The results of these tests are not robust to changes in the sample period, to changes in the variables included in the test, or to whether the data are in log-level or first-differenced form. Sims finds that money Granger-causes output in a simple bivariate setting. In a later paper, Sims (1980) determines that money fails to Granger-cause output when the commercial paper rate is included in the test. Litterman and Weiss (1985) replicate this result and also show that the nominal commercial paper rate Grangercauses both money and output. They find that the real interest rate, however, does not Grangercause either output or money. Eichenbaum and Singleton (1986) replace the commercial paper rate with the real rate of return on stocks and the real rate of return on Treasury bills in their Granger-causality tests. They find that while the real rate of return on Treasury bills does not Granger-cause output, the real rate of return on stocks does. Their model allows no explanatory power for money once these variables are included. Stock and Watson (1989) find that money Granger-causes output if the rate of return on stocks is omitted and the nominal rate of return on Treasury bills is included. Friedman and Kuttner (1989), however, find that this result is sensitive to the sample period chosen. They also determine that money fails to Granger-cause output (except for one subsample) when the nominal commercial paper rate is replaced by the spread between the commercial paper rate and the Treasury bill rate. What do these results tell us about the direc tion of causality between money and output? First, the inclusion of interest rates seems to weaken the explanatory power of money. This seems to be inconsistent with a money-driven business cycle. McCallum (1983), however, argues (but does not show) that if the Federal Reserve attempts to peg the interest rate, then interest-rate innovations are a better indication of the influence of money on output than are monetary innovations. This result is obtained because monetary innovations that affect output also cause interest rates to change. There are also nonmonetary shocks that cause interest rates to change, leading to changes in output. Second, different measures of the rate of return yield drastically different results. The rea son is probably that some rates of return are a better proxy for future changes in output than others. As Friedman and Kuttner indicate, the primary determinant of the spread between the Treasury bill rate and the commercial paper rate is the default risk on corporate securities. The primary determinant of the default risk of corpo rate securities is probably the anticipation of future business conditions, that is, future changes in output. The real rate of return on stocks in Eichenbaum and Singleton’s study is probably also a proxy for future changes in output. The issue of whether money is significant in its ability to predict future output when the spread or return on stocks is included in the causality test tells us little about the actual direc tion of causality between money and output. Money will Granger-cause output whether money actually causes output or whether future output causes money, whenever the spread (or the return on stocks) is a proxy, but an imperfect proxy for future output. Money would appear to be significant for both models because it helps to eliminate some of the noise present in the spread. Similarly, money will not Granger-cause output if the spread (or the return on stocks) is a perfect proxy for future output. The two models, money causing output and output causing money, are thus observationally equivalent in their predictions concerning whether money Granger-causes output. This analysis indicates that inferences about the direction of causality between money and output cannot be made from the existing Granger-causality tests. One of the major prob lems with the existing empirical studies is that they use M l as their measure of money. As indi cated in the previous sections, broader measures of money respond to future business conditions more than narrow measures of money, such as the monetary base. It appears that it would be difficult to distinguish between money causing output or output causing money when measures of money containing endogenous components are used. The same caveat holds for narrow meas ures of money like the monetary base. These measures, however, do not seem to respond to future business conditions to the same degree as M l or M2. These results suggest that the use of causality tests should proceed along the lines indicated by Sims (1989). He urges that researchers should concentrate on combining the theoretical tech niques developed by real business cycle theo rists and the empirical technique of vector auto regressions. That is, researchers should proceed along the lines of Prescott (1986), but should compare more than simple correlations when matching simulated data to actual data. Sims recommends that they compare the results of Granger-causality tests run on both simulated data and actual data. This requires models to pass stricter empirical tests before being judged as either successful or unsuccessful. Applying this technique to help determine the direction of causality between money and output would require building a real business cycle model with money and then comparing the vector auto regressions run on simulated data from both models with actual data. VI. Conclusion and Policy Implications This paper has shown that Granger-causality tests alone cannot settle the debate about the direc tion of causality between money and output. One reason is the ever-present problem of a potentially missing third variable. In section I, we showed how umbrellas could Granger-cause rain when a variable proxying for the expecta tion of rain, low-pressure systems, is excluded from the tests. The above studies seem to affirm the notion that leaving out variables that proxy for the expectation of future output could leave money with explanatory power when no causal ity is actually present. It should be clear that this debate is not likely to be settled on the basis of Granger-causality tests alone. Unfortunately, the issue can probably never be completely settled without having the Federal Reserve conduct con trolled experiments with monetary policy that would be infeasible. Causality tests are not necessarily useless, however. They may provide some information about the direction of causality, as long as they are interpreted within the confines of a model. That is, we must start with the null hypothesis that a specific model is correct and attempt to test whether or not we can reject this hypothesis. This approach is in the spirit of Eichenbaum and Singleton (1986); however, the suggestions made by Sims (1989) seem more appropriate. Many policymakers currently assume that money causes output in a consistent and reliable way. Economists have been unable to demon strate this relationship, however. If money does cause output, are policies predicated on such causation benign or harmful? At first glance, not it would seem that the effects of current policy would be benign if money does not cause output. However, by not being able to pin down the direction of causality, we cannot rule out other possibilities. For instance, it may be possible that inflation or monetary growth decreases output. Support for this proposition comes from Kormendi and Meguire (1985). Using cross-country data, they find a negative correlation between inflation and the growth rate of real output. The possibility that inflation may lower output should not be too surprising, given that inflation is a tax on real cash balances. As is the case with any other tax, we would expect increases in this tax to depress output. For example, higher rates of inflation cause people to engage in wasteful activities in order to economize on money hold ings, thus serving to lower output. Because researchers cannot tell whether increases in money cause output to increase— and there is some evidence that increases in the growth rate of money actually depress output— how should policymakers proceed? Policy actions should be analyzed in light of their potential costs and benefits. Traditional Keynes ian analysis assumes that all output fluctuations are inefficient and that policy could improve economic welfare by stabilizing output. How ever, as Lucas (1987) points out, the welfare gains associated with smoothing business-cycle fluctuations are small and are dwarfed by the potential gains associated with increasing longrun economic growth. The costs associated with stabilizing output may not be small. If unanticipated money increases output as described by Lucas (1972, 1977), then the real output effects from money are welfare-reducing. The reason is that the out put effects of money are generated by misper ceptions on the part of the public. As Lucas points out, this analysis prescribes that the Fed eral Reserve should follow a rule when conduct ing monetary policy. In Lucas’s model, any out put changes induced by money are inefficient. Even if his reasons for why money affects output are incorrect, it still may be best for policymakers to follow a rule. Stockman (1988) also makes the point that conducting policy as if output fluctuations are inefficient can be damaging. If the true explana tion of business cycles turns out to require both Keynesian and real business cycle elements, then there may be substantial welfare losses associated with output stabilization. As argued by real business cycle theorists, some output changes are efficient. In addition, it is presently impossible to distinguish inefficient from effi cient movements in output. Using monetary pol icy to offset these shocks could very well leave us worse off. Therefore, even if money has real effects, it is not clear how aggressively, if at all, monetary policy should try to stabilize output. Policymakers should accept the possibility that money does not cause output. Instead of con ducting policy as if money does cause output, they could base monetary policy on what we currently know about its costs and benefits. The preceding analysis leads us to believe that poli cymakers should be more reluctant to fine-tune the economy without understanding the ineffi ciencies present in the economy. Because the costs of economic stabilization are thought to be large, while the potential benefits have been shown to be fairly small, we recommend that monetary policy be predicated on a rule that is easy for policymakers to implement and even easier for the public to monitor. References Ahmed, Shaghil, “Wage Stickiness and the NonNeutrality of Money: A Cross-Industry Analy sis,” July 1987, 25-50. Journal of Monetary Economics. 20, Barro, Robert J., “Long Term Contracting, Sticky Prices and Monetary Policy,” July 1977, 305-16. Journal of Mone 3, tary Economics, Black, Fischer, “The Trouble with Econometric Models,” March/ April 1982, 29-38. Financial Analysts Journal, 38, Boschen, John F. and Grossman, Herschel I., “Tests of Equilibrium Macroeconomics Using Contemporaneous Monetary Data,” November 1982, 309-34. Journal oj 10, Monetary Economics, Hoehn, James G., “Procyclical Real Wages Under Nominal-Wage Contracts with Productivity Variations,” Federal Reserve Bank of Cleveland, Quarter 4, 1988, 11-23. Economic Review, King, Robert G. and Plosser, Charles I., “Money, Credit, and Prices in a Real Business Cycle,” June 1984, 363-80. American Economic Review, 74, Kormendi, Roger C. and Meguire, Philip G., “Macroeconomic Determinants of Growth: Cross-Country Evidence,” September 1985, 141-64. tary Economics, Journal of Mone 16, Laurent, Robert D., “Reserve Requirements: Are They Lagged in the Wrong Direction?” August 1979, 301-10. nal of Money, Credit, and Banking 11, Jour Cooley, Thomas F. and LeRoy, Stephen F., “Atheoretical Macroeconomics: A Critique,” November 1985, 283-308. Litterman, Robert B. and Weiss, Laurence, “Money, Real Interest Rates, and Output: A Reinterpretation of Postwar US. Data,” January 1985, 129-56. Eichenbaum, Martin and Singleton, Kenneth J., “Do Equilibrium Real Business Cycle Theo ries Explain Postwar U.S. Business Cycles?” Cambridge, Mass.: MIT Press, 1986, 91-135. Lucas, Robert E. Jr., “Expectations and the Neu trality of Money,” April 1972, 103-24. Journal of Monetary Economics, 16, NBER Macroeconomics Annual, Fischer, Stanley, “Long-Term Contracts, Rational Expectations, and the Optimal Money Supply Rul February 1977, 191-206. e," Journal of Political Economy, 85, Friedman, Benjamin M. and Kuttner, Kenneth N., “Money, Income and Prices after the 1980s,” National Bureau of Economic Research Work ing Paper No. 2852, February 1989. A Mone tary History of the United States: 1867-1960, Friedman, Milton and Schwartz, Anna, Princeton, N.J.: Princeton University Press, 1963. Granger, C.W.J., “Investigating Causal Relations by Econometric Models and Cross-Spectral Methods,” July 1969, 424-38. Econometrica, 37, Forecasting Eco ________ and Newbold, Paul, second edition, Orlando, Fla.: Academic Press, 1986. nomic Time Series, Gray, Jo Anna, “Wage Indexation: A Macroeco nomic Approach,” s, April 1976, 221-35. nomic Journal of Monetary Eco 2, Hodrick, Robert and Prescott, Edward, “Postwar U.S. Business Cycles: An Empirical Investiga tion,” Working Paper, Carnegie-Mellon Uni versity, 1 9 8 0 . ometrica, Econ 53, 4, Journal of Economic Theory, ------ , “Understanding Business Cycles,” in Stabilization of the Domestic and Interna tional Economy, edited by Karl Brunner and Allan H. Meltzer, Camegie-Rochester Confer ence Series on Public Policy, Amsterdam: North Holland Publishing Co., 1977, 5, 7-30. Models of Business Cycles, ------ , Basil Blackwell, Inc., 1987. New York: Mankiw, Gregory, “Comment: Do Equilibrium Real Business Cycle Theories Explain Postwar U.S. Business Cycles?” Cambridge, Mass.: MIT Press, 1986, 139-45. Annual, NBER Macroeconomics McCallum, Bennett T., “A Reconsideration of Sims’ Evidence Concerning Monetarism,” 1983, 167-71. Economic Letters, 13, Mishkin, Frederic S., A Rational Expectations Approach to Macroeconometrics: Testing Policy IneJJectiveness and Efficient-Markets Models, Chicago: University of Chicago Press, 1983. Prescott, Edward C., “Theory Ahead of Business Cycle Measurement,” Fed eral Reserve Bank of Minneapolis, Fall 1986, Quarterly Review, 9- 22. Sims, Christopher A., “Money, Income and Cau sality,” Sep tember 1972, 540-52. American Economic Review, 62, ------ , “Comparison of Interwar and Postwar Business Cycles: Monetarism Reconsidered,” May 1980, 250-57. American Economic Review, 70, ________, “Models and Their Uses,” Institute for Empirical Macroeconomics, Federal Reserve Bank of Minneapolis, Discussion Paper No. 11, April 1989. Solow, Robert, “A Contribution to the Theory of Quarterly Journal of 70, 65-94. Economic Growth,” February 1956, Economics, Stock, James H. and Watson, Mark W., “Inter preting the Evidence on Money-Income Cau sality,” 1989 (forthcoming). Journal oj Econometrics, Stockman, Alan C., “Real Business Cycle Theory: A Guide, an Evaluation, and New Directions,” Federal Reserve Bank of Cleveland, Quarter 4, 1988, 24-47. Economic Review, Summers, Lawrence H., “Some Skeptical Obser vations on Real Business Cycle Theory,” Federal Reserve Bank of Minneapolis, Fall 1986, 23-27. Quarterly Review, Tobin, James, “Money and Income: Post Hoc Ergo Propter Hoc?” May 1970, nomics, Quarterly Journal oj Eco 84, 301-17. Third Quarter Working Papers Working Paper Notice Current Working Papers of the Cleve land Federal Resen/e Bank are listed in each quarterly issue of the Economic Single copies of individual papers will Institutional subscribers, such as librar be sent free of charge to those who ies and other organizations, will be request them. A mailing list service for placed on a mailing list upon request personal subscribers, however, is not and will automatically receive available. Papers as ■ 8908 ■ 8 9 10 In te rv e n tio n a nd th e R is k P re m iu m in Fo re ig n E x c h a n g e S tr u c tu re , C o n d u c t, and ■ 8 9 12 E n fo rc e m e n t o f P e rfo rm a n c e in th e L o c a l P u b lic S e c to r P o llu tio n R e g u la tio n s in a D e c lin in g In d u s try Review. Copies of specific papers may be requested by completing and mail Working they are published. ing the attached form below. R a te s by Randall W. Eberts by M ary E. Deily by W illiam P. Osterberg and Tim othy J. Gronberg and Wayne B. Gray ■ 8909 ■ 8 9 11 F a c to r-A d ju s tm e n t P u b lic In fra s tru c tu re a nd R e g io n a l Ec o n o m ic D e v e lo p m e n t: C o s ts a t th e In d u s try Le v e l A S im u lta n e o u s E q u a tio n s A p p ro a c h and Dennis W. Jansen by Mary E. Deily by Kevin T. Duffy-Deno and Randall W. Eberts Please com plete and detach the form below and mail to: Federal Reserve B an k of Cleveland R esearch D ep a rtm e n t P .O . B o x 6 3 8 7 C leve lan d , O h io 44101 Check item(s) Please send the fo llo w in g W o rk in g Paper(s): requested □ 8908 □ 8909 Send to: Please print Ad d re s s □ 8910 □ 8911 □ 8912