The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
The Formation of Inflation Expectations DAVID H. RESLER I HE psychology of inflation is often cited as a major barrier in the war against inflation. One cur rently popular view contends that anti-inflation pol icies must not only attack the causes of inflation but also lower inflation expectations. In 1971, the Nixon Administration used this argument to justify the im position of wage and price controls. More recently, President Carter’s March 14 economic policy initiative was intended, in part, to calm financial markets by signaling that inflation expectations should be lowered. Similarly, the objective of lowering inflation expectations has also been used to justify proposed inflation remedies such as the tax-based incomes policy (T IP ).1 Concern over inflation expectations has been motivated primarily by recent developments in mac roeconomic theory that place the expectations of im portant economic variables at the forefront of analysis. For instance, many economists attribute the existence of the so-called Phillips curve trade-off between in flation and unemployment to unrealized inflation ex pectations rather than to inflation per se.2 Although inflation expectations have become crucial to both theoretical and policy analysis, they remain extraordinarily difficult to measure. Generally, 1For a discussion and analysis of the TIP program see Nancy Ammon Jianakoplos, “A Tax-Based Incomes Policy (T IP ): What’s It All About?” this Review (February 1978), pp. 8-12. 2Milton Friedman, in his 1977 Nobel Laureate address, details the intellectual evolution of various theories of the inflationunemployment trade-off. See Friedman, “Nobel Lecture: In flation and Unemployment,” Journal of Political Economy (June 1977), pp. 451-72. 2 economists have relied on various distributed lag models of past inflation rates to estimate inflation expectations. To a lesser extent, they have employed data gathered from surveys of economists, such as those conducted semiannually by Joseph Livingston of the Philadelphia Inquirer, or from surveys of households, such as those conducted by the Institute for Social Research of the University of Michigan.3 Although most studies using these data do so to test alternative hypotheses about economic activity, other scholars have been concerned with the actual process by which price expectations are generated.4 Forecasts of inflation can be modeled in various ways. One simple approach formulates the inflation forecast based solely on the history of inflation. As noted above, variations of such autoregressive schemes have dominated studies that include measures of inflation expectations. If independently determined forecasts of inflation are available, then one could test whether only the history of infla3See Joseph Livingston, (biannual surveys), Philadelphia Sun day Bulletin, June and December, 1948-1971 and Philadelphia Inquirer, June and December, 1972 to the present; and Rich ard T. Curtin, ed., Surveys of Consumers 1974-75, Contribu tions to Behavioral Economics (Ann Arbor: Institute for Social Research, The University of Michigan, 1976). 4Some studies that have focused on this process include James Pesando, “A Note on the Rationality of the Livingston Price Expectations,” Journal of Political Economy (August 1975), pp. 849-58; John A. Carlson, “A Study of Price Forecasts,” Annals of Economic and Social Measurement (June 1977), pp. 27-56; Donald J. Mullineaux, “Inflation Expectations and Money Growth in the United States,” American Economic Review (March 1980), pp. 149-161; and Edward J. Kane and Burton G. Malkiel, “Autoregressive and Nonautoregressive Elements in Cross-Section Forecasts of Inflation,” Econometrica (January 1976), pp. 1-16. APRIL F E D E R A L R E S E R V E B A N K O F ST . L O U I S tion is important in explaining those forecasts. Al ternately, “rational expectations” hypotheses argue that all currently available information relevant to the actual inflation process is considered when fore casts are made. Although such information would not be confined solely to the history of inflation, the set of relevant information may be dominated by it. In this case, again, inflation expectations would be closely approximated by some autoregressive scheme. Related to the process that generates inflation expectations is the mechanism by which revisions in expectations are determined as new information be comes available. Knowledge of whether such revisions are systematically related to recent forecasts is useful in assessing the impact that inflation expectations have on economic activity. The simplest hypothesis is that revisions in forecasts depend on past forecast errors.5 This approach implicitly assumes that all in formation relevant to the forecast revision is contained in the most recent forecast error. As Mincer has noted, such error-leaming models can be interpreted as a reduced form of an autoregressive forecasting model.6 If the actual inflation forecasting process is not de scribed solely by the history of inflation, however, then exclusive reliance on past forecast errors to describe expectations revisions would be inappro priate. This article investigates the process by which in flation expectations are formed and the relevance of error-leaming models for analyzing revisions in these expectations, using the Livingston survey data. The extent of error-learning in the revision of inflation expectations, as well as the process by which these expectations are formed, offers useful clues about the efficacy’ of inflation-reducing strategies. INFLATION EXPECTATIONS AND FORECAST REVISIONS: THE LIVINGSTON DATA Twice each year, Joseph Livingston, a financial re porter for the Philadelphia Inquirer, requests selected business, government, and academic economists to provide forecasts of various measures of economic activity, including levels of the Consumer Price Index 5David Meiselman pioneered the use of the “error-leaming” model in his study, The Term Structure of Interest Rates (Englewood Cliffs: Prentice-Hall, 1962). ®See Jacob Mincer, “Models of Adaptive Forecasting,” in Economic Forecasts and Expectations, Jacob Mincer, ed. ( New York: National Bureau of Economic Research, distributed by Columbia University Press, 1969). 1980 (C P I). John Carlson used Livingston’s data on these price level forecasts to generate a series on inflation expectations for the period from 1947 to 1975.7 For this article, these data have been updated through 1978 using Carlson’s methodology. Although calculations of inflation expectations can be derived directly from price level forecasts, calculations of revisions in these expectations re quire information about inflation forecasts that were made over different time horizons. Since the price ex pectations reported by Livingston are for 6- and 12month horizons, it is easy to calculate a rate of infla tion expected over the 6-month period beginning six months hence.8 Specifically, for the succeeding 6month period, the inflation rate implied by the cur rent 6- (it?) and 12-month (it*) inflation forecasts can be described as: The forecast revision, Rt, is then defined as: (2) Rt = iX —ft.t-1, o.t where the t subscripts identify the moment at which the expectations (or the implied forward inflation rate) are formed. INFLATION EXPECTATIONS AND FORECAST REVISIONS: PREVIOUS STUDIES Pesando and Mullineaux, among others, have used either the predictions published by Livingston or those revised by Carlson to estimate equations that "For details of how the expected rate of inflation is calculated from the Livingston price level forecasts, see Carlson, “A Study of Price Forecasts.” 8Actually, Carlson calculated an implied inflation rate for an 8- and 14-month horizon. An example will clarify this point. The Livingston survey respondents are asked, prior to, say, the June survey, to forecast the level of the CPI for the com ing December and the following June. Since this forecast is made a month or more prior to the survey’s publication, Carl son assumed that most respondents would know only the level of the CPI data two months before (April or October) the survey’s publication. Thus the economists were forecasting the CPI for 8 and 14 months ahead. In this article, these fore casts are described as 6- or 12-month forecasts but are, in fact, identical to those of Carlson. This means that the forecast inflation rates for 6- and 8-month and for 12- and 14-month horizons are assumed to be the same. 9In general, the implied j-month inflation rate for the interval from i to i + 1 is: , *'■ ' _ (l + TtT.m,,)1 *1 (1+T t?.,)1 , This expression is similar to that used in the term-struclure of interest rate literature to describe the implied forward in terest rate. 3 F E D E R A L R E S E R V E BAN K O F ST. L O U IS APRIL 1980 Table 1 Revisions in Inflation Expectations: The Carlson Model (ordinary least squares estimation) R — bo 4" bi E + u Coefficients* Period bo b, R7RJ SEE D.W. 1953-62 -.049 (-.318) .045 (.484) .013/-.042 .568 1.013 1963-71 -.264 (-2.458) .085 (1.227) .086/.029 .327 1.550' 1953-71 -.141 (-1.458) .056 (.926) .023/-.004 .466 1.072 1953-75 -.287 (-2.818) .208 (4.075) .274/.258 .539 1.513' 1953-78 -.305 (-3.319) .212 (4.550) .293/.279 .525 1.569' 1953-78 (omitting 1972-74) -.250 (-2.630) .118 (1.991) .083/.062 .514 1.509 "t-statistics are in parentheses. "Value of Durbin-Watson statistic permits rejection of positive serial correlation. 'Value of Durbin-Watson statistic is in the indeterminate range. “explain” the forecasts in terms of observable eco nomic variables, such as past actual rates of inflation, past rates of money growth, and so on. Most of these studies sought to investigate the rationality and effi ciency of inflation forecasts. None have investigated explicitly the process by which forecasters revise their expectations. Although Carlson did not explicitly examine the process by which inflation forecasts are formed, he did investigate the relevance of a simple error-learning model in explaining forecast revisions. He did not, however, derive his error-learning model from some underlying structural forecasting model. Consequently, his finding of only weak evidence that past forecast errors affected revisions in expectations warrants reexamination. Carlson argued that forecast revisions depend on the most recent forecast error ( E t ) in inflation ex pectations. This error is defined as: (3 ) E, = n,., - n,*,-,, where n 6,t is the most recent 6-month rate of inflation observed at time t. Carlson then investigated the 4 error-learning hypothesis by estimating the equation: (4 ) Rt = b0 + b, Ei + ut, where ut is a random error. This equation states that revisions of previous forecasts depend only on the most recent forecast error. Carlson’s estimations of equation 4 over three dif ferent sample periods (1953-62, 1963-71, and 1953-71) failed to uncover any consistent evidence in support of the error-leaming hypothesis. Although he found some evidence of error-leaming in inflation forecasts based on the Wholesale Price Index (W P I), he found no evidence that error-learning significantly affected revisions of 6-month inflation forecasts based on semi annual observations of the CPI. To test whether this conclusion remains valid when more recent data are included, equation 4 was reestimated over selected time periods. The results, reported in table 1, do not support Carlson’s conclu sion when the sample period is extended to include the 1970s. For example, the estimated coefficient on past errors is positive and significant when the equa tion is estimated through 1975 or 1978. Because the pervasive price controls in effect dur APRIL F E D E R A L R E S E R V E BANK O F ST. L O U IS 1980 Table 2 Revisions in Inflation Expectations: The Carlson Model (Cochrane-Orcutt estimation) R = b0 + bi E + u Coefficients* Period bo 1953-62 -.149 (-.651) 1953-71 R7RJ SEE D.W. Rhob .086 (1.172) .259/.215 .493 1.967 .482 (2.396) -.218 (-1.642) .077 (1.450) .191/.168 .420 1.860 .432 (2.912) 1953-78 -.333 (-3.161) .220 (4.684) .340/.327 .508 2.024 .194 (1.413) 1953-78 (omitting 1972-74) -.228 (-2.238) .114 (2.263) .142/.122 .440 1.759 .245 (1.674) bi *t-statistics are in parentheses. 'The autocorrelation coefficient as estimated by the Cochrane-Orcutt technique; t-statistics are in parentheses. ing the period 1972-74 may have distorted the effect of past errors, equation 4 was estimated through 1978 omitting 1972-74 data. Again, the hypothesis that error-learning has been an important factor in the formation of inflation expectations cannot be rejected. The low Durbin-Watson (D .W .) statistics for several versions of equation 4 suggest the presence of serially correlated residuals (u ,). As a result, estimates of equation 4 may not be efficient. The Cochrane-Orcutt iterative technique was used to cor rect for the presence of autocorrelation for those sam ple periods in which the D.W. statistic indicated that the hypothesis of serially correlated residuals could not be conclusively rejected. Estimates of equation 4, using this method, are reported in table 2 and, like those discussed above, are less conclusive than Carl son’s about the relevance of the error-learning hy pothesis. For example, in the 1953-71 period, the t-statistic for the coefficient on the forecast error, though small, is significent at the 10 percent con fidence level using a one-tailed test.10 Nevertheless, over the two subsamples of the period, 1953-62 and 1963-71, the error-leaming hypothesis still must be rejected. Coefficients for this parameter over all three periods, however, differ by less than .01, suggesting 10The one-tailed test is appropriate for testing the null hypoth esis that bi = 0 against the alternative hypothesis that b, > 0. that estimates from the shorter sample periods may be inefficient but unbiased estimates of the true parameter.11 Since the error-leaming model implies that all information relevant to forecast revisions is contained in the most recent error, the significant negative co efficient on the constant term requires further dis cussion. The significance of these coefficients, together with the low coefficient of determination (R 2), could be interpreted as evidence that important variables have been omitted from the specification. A careful examination of the expectations formation process underlying equation 4 provides additional support for this interpretation. As noted above, the relevant error-leaming model should be derived from and consistent with the under lying structural expectations formation model. Recall ing the definitions for the revision and the fore cast error, it is easy to see that the underlying foreu The estimated coefficients for the three pre-1972 sample pe riods do not differ significantly from each other, suggesting that they are all unbiased estimators of the true parameter. Because the variance of these estimated coefficients tends to decline with increases in the size of the sample, the estimates for the shorter periods cannot be considered efficient (i.e., they are not the minimum variance estimators). For a dis cussion of the efficiency of estimators, see Jan Kmenta, Ele ments of Econometrics (New York: MacMillan Publishing Company, 1971), pp. 157-69. 5 F E D E R A L R E S E R V E B A N K O F ST. L O U I S cast mechanism implied by equation 4 conforms to the following relationships: (5a) ir?,t == a0 + ai (5 b ) fe.t = and oto + a , n?,t, where ai = a , and the t-subscripts identify the period in which the variable is observed.12 (For example, ir6,t is the most recently observed six-month inflation rate.) Equations 5a and 5b imply a highly restrictive ( “naive”) version of the expectations process. These specifications imply that the forecasters’ expectations of inflation for the next period depend only on the most recently observed rate of inflation. No other in formation is incorporated. Note also that both equa tions specify a constant. A nonzero constant in either equation implies that some premium (or discount) is added to the impact of the current 6-month inflation rate to obtain the relevant 6-month inflation forecasts. AN ALTERNATIVE FORECAST MECHANISM If the true underlying expectations formation mechanism is less restrictive or more complex than the one described by equations 5a and 5b, then the revision equation given in equation 4 is misspecified. Recent studies of inflation expectations offer some evi dence for this interpretation. Pesando, in his study of the Livingston data, hy pothesized an autoregressive scheme for the price fore casts.13 In another study of inflation forecasts based on different survey data, Kane and Malkiel empha sized the importance of including some retum-tonormality variable.14 The retum-to-normality model implies that forecasters adjust their forecasts to some notion of the “normal” rate of inflation. Mullineaux also experimented with a variety of variables that could potentially influence inflation expectations and reported . . that inflation forecasts are systematic ally influenced by past inflation rates and past rates of money growth, but not by fiscal-policy-related variables. . . ,”15 Both the Kane-Malkiel and Mulli neaux studies highlight the relevance of other informa 12The revision equation (equation 4) is obtained by lagging each term in equation 5b and subtracting from 5a. 13Pesando, in “A Note on the Rationality,” characterized the Livingston forecast by a highly restrictive autoregressive scheme. 14Kane and Malkiel, “Autoregressive and Nonautoregressive Elements.” 15Mullineaux, “Inflation Expectations and Money Growth,” p. 160. 6 APRIL 1980 tion in addition to the past rate of inflation in the formation of inflation expectations. The remainder of this article explores an alternative inflation ex pectations formation mechanism that is hypothesized to depend on both the time series of past inflation rates and on elements that embody a retum-to-nor mality notion. A Return-to-Normality Model The following mechanism is hypothesized for the expectations formation process: (6a) (6b) 71®, = a> + ai < + a: itj.t + as tt?, and, f j., = etc + a , n*. This process describes the current inflation forecast for the j-month horizon in terms of last period’s fore cast ( T j *-i), the most recently observed j-month T inflation rate (Ttjit), and the currently held expected normal rate of inflation (it?). The implied forward rate of inflation, fJ t — the inflation rate expected to prevail for the j-month period beginning at the com pletion of the current j-month period — is hypothesized to equal the currently held forecast plus some pre mium (or discount), a 0. This specification of inflation expectations em bodies both autoregressive and retum-to-normality elements.16 The presence of a lagged dependent vari able in equation 6a can also be interpreted as cap turing the relevance of any inertia in the forecasts. The larger the coefficient, a1; the more reluctant forecasters are to revise their expectations. The co efficient, a2, which applies to the most recently ob served inflation rate, measures the extent to which new information about inflation is deemed relevant for the current period’s forecast. Finally, the coeffi cient, a3, reflects the dependence of short-run inflation forecasts on the long-run normal rate of inflation. Equation 6a can be rewritten in a form that cap tures the impact of past errors. Adding and sub 16For example, by repeated substitution for the lagged depend ent variable, equation 6a can be shown to be equivalent to Ti:*,t = ao + a2 Ttj.t + a^ n” + n 2 ai (ao + a2 + as Ht-i), i= 1 where the last term embodies primarily autoregressive com ponents as well as the history of the normal rate of inflation. Another point merits attention. Suppose forecasts are made for a minimum j-month horizon and for other horizons that are some multiple of that horizon (for example, a 6-month, a 12-month, and an 18-month horizon). If each of these fore casts are made every j-month, then all forecasts, regardless of the horizon, can be represented as a distributed lag on past j-month inflation rates. F E D E R A L R E S E R V E B A N K O F ST . L O U I S tracting a2 i6,t-i from the right-hand side of 6a T produces: (7) iTj.t = a0 + (ai + a2) n *,-, -f- aiEj.t + a3 n?. Lagging all terms in equation 6b one period and subtracting from equation 7 produces a forecast re vision equation that is consistent with this forecasting process: (8) R j,t = (ao - oto) + (ai + a2 - a t) + a^Ej,, + a3 n?. Clearly, the simple revision equation estimated by equation 4 is not consistent with the forecast mech anism described here. The correct equation for esti mating revisions in such inflation expectations is: (9 ) Rj.t = Po + M T..-. -t" PsEj.t + P T + ut; aT ? where |0 = a0 - a 0, | = (a, + a2 - 0O , |2 = a2, 3 3i 3 p3 r= a3, and ut is a random error. Note that the logic of the model implies that a0 and a 0 should be zero since all relevant information is presumed to be em bodied in the variables E jjt, and Ttj. Conse quently, estimated values of p0 also should not differ significantly from zero. The coefficient for the lagged dependent variable in equation 6a can be interpreted in terms of the speed with which forecasters adjust their expecta tions from one period to the next. Equation 6a de scribes the adjustment of inflation forecasts partly in terms of previously held forecasts. The size of the coefficient, au on the lagged term measures the ex tent to which forecasters maintain previously held forecasts. The speed with which forecasts are adjusted over time, therefore, corresponds to (1 - a ,). Larger values for ax imply that a stronger persistence effect is embedded in the forecast process or, alternatively, that forecasts are revised more slowly when new in formation becomes available. APRIL 1980 For example, Kane and Malkiel found that “. . . return-to-normalitv elements dominate forecasts of future inflation and [show] . . . that developments outside the past history of prices importantly alter respondents’ conceptions of what rate of inflation is ‘normal.’ ”17 In their investigations of the retum-tonormality hypothesis, Kane and Malkiel surveyed large firms and major bond dealers to gather infla tion forecasts over several horizons. They were thereby able to calculate a normal rate forecast as a weighted average of subperiod forecasts which ex tended as far as 10 years into the future. Unfortu nately, the Livingston data do not permit the derivation of any comparable and meaningful normal rate. Because the longest horizon forecast is only 18 months, a test of the retum-to-normality hypothesis comparable to the Kane-Malkiel study is not pos sible.18 Consequently, tests of the retum-to-normality hypothesis must rely on other measures of the nor mal rate of inflation. One way to approximate the normal rate of in flation is to utilize some trend growth of the money stock. Such a proxy introduces a monetarist in terpretation of inflation forecasts into the model — namely, that the trend growth of prices is determined by the trend growth of money.19 In this study, a twenty-quarter moving average of past Ml growth is used as a proxy for the normal rate of inflation. Finally, the degree of persistence evident in the forecasts could vary with the forecast horizon. Be cause information about permanent structural changes in the economy evolves only slowly and is costly to distinguish from transitory phenomena, long-run in flation expectations could be expected to change less from one period to the next than short-run expecta tions. As a result, longer-range forecasts should show greater persistence than shorter-range forecasts. The use of a surrogate for the normal rate of inflation requires some modifications of the forego ing interpretation regarding equations 6a, 6b and 9. Suppose that equation 6a represents the true model that describes inflation expectations over some given short-run time horizon. If currently available information affects the actual rate of inflation only with some lag, then the forecast for the period be ginning one period hence could differ from the forecast made for the period now beginning. To the extent that this currently available information is relevant to the long-run rate of inflation, it should be imbedded somehow in the normal rate of inflation. The proxy for the normal rate of inflation used here does not represent exactly the notion of the normal rate. For example, suppose the Federal Beserve an nounced that it intended to pursue a new money The hypothesized forecasting process described by equations 6a and 6b contains a retum-to-normality variable that reflects the view that forecasters in corporate information about the long-run expected inflation rate. This expected normal rate of inflation embodies relevant information from a wider variety of sources than simply the time series of past prices. 17Kane and Malkiel, “Autoregressive and Nonautoregressive Elements,” p. 3. 18These 18-month forecasts were collected only once each year and were discontinued after 1971. 19See Denis Kamosky, “The Link Between Money and Prices — 1971-76,” this Review (June 1976), pp. 17-23 for a dis cussion of the link between the trend growth of money and inflation. 7 F E D E R A L R E S E R V E B A N K O F ST . L O U I S APRIL 1980 Table 3 Inflation Expectations: 6-Month Forecasts ■rtf — ao + aiti?. t-i + ajTt»lt + ajrc? Coefficients* Period a» a, a? a3 R7R’ SEE Durbin-h Rhob F' 1953-78 -.413 (-1.996) .654 (8.340) .183 (3.832) .186 (2.341) .952/.949 .522 1.119 77.00 1963-78 -.197 (-.557) .577 (6.139) .262 (4.769) .123 (1.064) .947/.941 .488 .070 35.13 1953-71 -.191 (-.952) .773 (7.445) .039 (.660) .156 (1.864) .888/.878 .450 .594 52.82 1963-71 -.171 (-.593) .552 (2.895) .081 (1.101) .260 (1.318) .919/.900 .321 -.701 -.330 (-1.462) .715 (8.317) .100 (1.659) .190 (2.062) .945/.941 .515 .493 1953-78 (omitting 1972-74) -.583 (-2.961) 17.27 73.69 ■t-statistics are in parentheses. bThe autocorrelation coefficient is reported only for that equation which was estimated by the Cochrane-Orcutt technique because of evidence of serial correlation in the OLS estimates; t-statistic is in parentheses. 'F is the F-statistic for assessing the hypothesis that the estimates reported here do not differ from those obtained by estimat ing the more simple model described by equation 5a. The F-statistics permit rejection of this hypothesis at the .01 level for all time periods reported. growth target over the coming six months. If this tar geted growth rate differed from the previous trend growth of money, analysts might expect the trend in money growth to be changing during the current 6-month forecast horizon. Because the proxy for the normal rate used here is entirely “backward-looking,” it omits such additional information. As a result, the implied forward rate equation could be expected to include additional terms. Since these terms are cur rently unmeasurable, however, they are assumed to be imbedded in the constant term; that is, the constant term in the implied forward inflation rate equation reflects the effect of currently available ( but not measurable) information on the future inflation rate. A positive constant would reflect the forecasters’ belief that the net effect of all other currently avail able and inflation-relevant information is to accelerate inflation. Finally, such a positive constant would imply a negative constant term, |0, in equation 9. 3 Empirical Tests of the Alternative Inflation Forecast Model Tables 3 and 4 report the results of estimating equations 6a and 6b, respectively, over various time periods. The first and perhaps most important obser vation is that the coefficients of determination ( R2) in table 3, are greater than 0.90 for four of the five 8 periods. For longer periods, they exceed 0.94. This statistic indicates that over 90 percent of the variance of the inflation forecasts is explained by this relatively simple reduced form. Interestingly, these values for Rz are quite close to those obtained when the forecasts are estimated in terms of more complicated Almon lags on past inflation rates and past money growth.20 More importantly, the coefficients of determination adjusted for degrees of freedom ( R2) are consistently greater than those obtained from estimates (not re ported here) of the “naive” forecast equation given in 5a. The rejection of the naive model in favor of equation 6a is reinforced by F-tests (for the hypoth esis that the two equations do not differ) conducted for the various sample periods. Results of these tests (based on comparisons of ordinary least squares esti mates of the two equations) are reported in the last column of table 3.21 The alternative model is favored 20The adjusted R!s for equation 6a. are similar to those ob tained by Mullineaux. While the R!s reported here generally exceed those of Mullineaux, his sample period differed from those estimated here, making direct comparison inappropri ate. Other estimations by the author of the inflation forecasts based on more complicated Almon lags of past money growth _and past inflation rates did not generate consistently higher R!s than did the equations reported here. - 1F-tests were made on the basis of OLS estimations of the two equations. Cochrane-Orcutt estimations would involve transforming all observations by some coefficient of autocor relation. Unless each equation is characterized by the same degree of serial correlation, the two equations would not be directly comparable. APRI L. 1 9 8 0 F E D E R A L R E S E R V E B A N K O F ST . L O U I S Table 4 Inflation Expectations: The Implied Forward Rate f.,t = cxo + a,n?., Coefficients* Period <o X a, R7R1 SEE D.W. Rhob .978/.978 .338 1.963 .266 (1.974) .525 (3.295) .967 (34.614) QT7 (26.087) .980/.979 .275 2.083 .328 (1.934) 1953-71 .138 (1.530) 1.058 (21.589) .928/.926 .384 1.553 1963-71 .164 (1.486) 1.057 (24.331) .974/.972 .186 2.369 1953-78 (omitting 1972-74) .163 (2.071) 1.037 (39.507) .973/.972 .375 1.624 1953-78 .338 (3.478) 1963-78 *t-statistics are in parentheses. "The autocorrelation coefficient is reported only for those equations which were estimated by the Cochrane-Orcutt technique because of evidence of serial correlation in the OLS estimates; t-statistics are in parentheses. over the naive model at the 0.01 confidence level for all time periods.22 The estimated constants reported in table 3 re inforce the view that this forecast mechanism is more appropriate than the naive model. In esti mates of that model, statistically significant, posi tive constant terms were consistently obtained, sug gesting the importance of omitted variables. In contrast, estimates of the present model produced a significant (though negative) constant term in only one sample period — 1953-78. This constant could capture elements related to the era of the Nixon wage-price controls. When this three-year episode is deleted, the constant is no longer significant at stand ard confidence levels. Finally, it should be noted that, unlike the naive model, the present model shows no evidence of positive serial correlation, though the 1963-71 sample period shows some evidence of nega tive serial correlation.23 The implied forward rate 22In several OLS estimations of the naive model, the DurbinWatson statistic was unacceptably low. While this result is usually interpreted as evidence of positive serial correlation, it may also indicate that important explanatory variables have been omitted. This interpretation seems appropriate here, since by including the two additional variables in equa tion 6a, evidence of positive serial correlation disappears. 23The Durbin-h statistic is appropriate for testing for serial correlation when lagged values of the dependent variable are included. The Durbin-h is normally distributed with a zero mean and a variance of o2. See J. Johnston, Econometric Methods (New York; McGraw-Hill, 1972), pp. 312-13. is also accurately described by equation 6b.24 Taken together, these results provide favorable evidence that the underlying forecast process conforms quite closely to the one hypothesized here. Although the R2s remain high over the various sam ple periods, the variation in the estimated coefficients, especially those for the current inflation rate and the normal rate, suggest that the contribution of these variables in the forecasting process has changed.25 For example, in periods ending with 1971, the cur rent rate of inflation played virtually no independent role in the determination of next period’s forecast. Apparently, current inflationary phenomena was largely discounted — at least until it became embed ded in the past trend of inflation. As the sample period is extended toward the present, however, the most recent inflation rate assumes a dramatically dif ferent role. Both the magnitude and the significance of the a2 term indicate that forecasters viewed the information reflected in the current inflation rate as more relevant. 24Equation 6b implies that the implied forward rate could al ternatively be expressed in a form similar to that given in 6a. Estimating this version of the forward rate did not pro vide as good a fit in terms of R2 as did the more simple form. 25Mullineaux, “Inflation Expectations and Money Growth,” also observed a changing forecast structure over time. Mullineaux’s work gives a thorough and detailed analysis of the behavior of the temporal coefficients on past inflation. 9 F E D E R A L R E S E R V E B A N K O F ST. L O U I S APRIL 1980 Table 5 Inflation Expectations: 12- and 18-Month Forecasts •n*,t = ao + a, itj.t-! + a, Tt«,t + a3 j = 12, 18 ir?; (12-month forecast) Coefficients* Period a0 a, a2 a3 R’/R 1 SEE Durbin-h 1953-78 -.266 (-1.496) .734 (11.231) .143 (3.529) .146 (2.108) .964/.962 .455 .691 1963-78 -.149 (-.451) .641 (7.087) .212 (4.245) .127 (1.135) .952/.947 .455 -.109 1953-71 -.082 (-.513) .837 (11.099) .030 (.619) .113 (1.712) .931/.925 .368 .828 1963-71 -.114 (-.428) .585 (3.315) .080 (1.156) .245 (1.290) .934/.919 .296 -.376 1953-78 (omitting 1972-74) -.178 (-.936) .794 (11.271) .068 (1.307) .141 (1.808) .961/.958 .447 .012 -.151 (-1.341) .809 (11.963) .005 (.100) .282 .604 Rho" -.574 (-2.888) (18-month forecast) 1953-71 .223 (4.841) .955/.946 -.563 (-2.893) ”t-statistics are in parentheses. bThe autocorrelation coefficient is reported only for those equations which were estimated by the Cochrane-Orcutt technique because of evidence of serial correlation in the OLS estimates; t-statistics are in parentheses. Examination of estimates for equation 6a over longer forecast horizons offers an additional perspec tive. Table 5 reports estimates of equation 6a for 12and 18-month horizons.26 Estimated coefficients for the 12-month horizon over various time periods show a pattern similar to that estimated for the 6-month horizon. As expected, all coefficients on the lagged dependent variable are larger for the 12-month horizon than for the 6-month horizon. This suggests that there is greater period-to-period persistence and a slower adjustment speed in the 12-month forecasts than in the 6-month forecasts. For the 18-month fore casts, however, the coefficient on the lagged depend26Note that estimated equations for these forecast horizons differ slightly from those described by equation 6a in that the most recent 6-month inflation rate, rather than the most recent 12- or 18-month inflation rates, is included. The rea son for this is that the 12-month forecast, made six months ago, already incorporated all relevant information from past inflation. Only the most recent 6-month inflation rate is "news.” (As noted in footnote 16, all forecasts, regardless of horizon, can be represented as a distributed lag on past 6-month inflation rates.) If the 12- and 18-month forecasts were made only every 12 and 18 months, then the exact specification given by 6a would be appropriate. (Equa tion 6a was estimated using this latter specification, despite the informational redundancy contained in the 12- or 18month actual inflation rate. Those results did not differ notably from those reported here.) 10 ent variable was slightly (but not significantly) lower than in the 12-month forecast horizon. For those time periods in which the most recently observed 6-month inflation rate significantly affected the forecasts, its impact was greater on the short-run (6-month) forecasts than on the longer-run (12month) forecasts. This observation provides further evidence that the most recently observed inflationary experience is incorporated only slowly into longer-run forecasts. The specification given by equation 6a permits a useful interpretation of the coefficient for the normal rate. Essentially, the long-run tendency for the jth horizon forecast to converge toward the normal rate can be represented by a long-run coefficient on the normal rate described as a3/ ( l - a i ) . 27 Table 6 reports calculations of this parameter for the three forecast horizons over several periods. This long-run retum27The presence of a lagged dependent variable makes equation 6a similar to a stock-adjustment type of equation. The co efficient, ai, in 6a is interpreted as one minus the speed of adjustment of the forecast to the long-run “equilibrium” rate of inflation. The long-run coefficient for any other variable in the equation can then be described as a ratio of the esti mated short-run coefficient to the speed of adjustment, i.e., aj/ (l - a,). F E D E R A L R E S E R V E B A N K O F ST . L O U I S APRIL Table 6 Long-Run Response of Short-Term Expectations to Changes in the “Normal” Rate of Inflation* Period 6-month forecastsb 12-month forecastsb 1953-78 .538 (2.908) .549 (2.614) 1963-78 .291 (1.202) .354 (1.326) 1953-71 .687 (2.095) .693 (1.868) 1963-71 .580 (2.275) .590 (2.235) 1953-78 (omitting 1972-74) .667 (2.585) 18-month forecasts” .684 (2.206) 1.168 (2.649) *Due to rounding, the calculated coefficients may differ slightly from those calculated from results reported in tables 3 and 5. b t-statistics are in parentheses. For a description of the meth odology used to calculate the variance of as/ (1 - ai) used in calculating the t-statistics, see Kmenta Elements of Econometrics, p. 444. to-normality coefficient should be higher for longer forecast horizons, since long-run expectations would tend to converge to the normal rate of inflation. As expected, these long-run coefficients are larger for the 12-month than for the 6-month forecasts. While the differences between the coefficients for these horizons are not great, the long-run coefficient for the 18-month horizon is larger and, in fact, does not differ signifi cantly from unity. Thus it appears that the forecasters do tend to form long-run expectations in a manner consistent with the return-to-normality hypothesis.28 The estimated magnitude of this long-run coeffi cient for both the 6- and 12-month horizons falls dramatically when the sample period includes only the 1960s and the 1970s.29 This observation re inforces the view that the rapid acceleration in infla tion experienced during the 1970s has had an important effect on the way inflation expectations are formed. Throughout this period, rapidly rising inflation may have simultaneously induced forecasters to revise the 28The calculated long-run coefficients for the normal rate are comparable to those obtained by Kane and Malkiel in their estimations based on cross-sectional data. For example, in equations using the CPI, their estimates of the return-tonormality coefficient ranged from about .52 for the 6-month horizon ( in 1969) to about .63 for the 12-month horizon (in 1972). 29Note that the coefficient deteriorates only slightly in the subsample 1963-71. 1980 normal rate of inflation more frequently. Hence, the proxy measure for the normal rate used here may understate the correct value of the normal rate, when inflation is accelerating rapidly.30 This possible meas urement error could distort the evidence reflected in the long-run coefficient for this variable, especially during more recent periods. In summary, several relevant observations emerge from the estimations of equations 6a and 6b. The in flation forecasting process employed by respondents to Livingston’s survey of economists can be described in terms of both autoregressive elements and past money growth ( interpreted here as a proxy for returnto-normality elements). Nevertheless, although the equation performs well over all subsamples of the period 1953-78, the relative roles played by the cur rent and normal rates of inflation appear to have changed. Specifically, during the 1970s when infla tion accelerated sharply, retum-to-normality elements played a less important role while the most recent rate of inflation became more important. Finally, the emergence during the 1970s of a significant, positive constant in the implied forward rate equation pro vides some evidence that forecasters had begun to anticipate accelerating inflation. Implications for Error-Learning Models The relevant equation for examining the errorlearning hypothesis is implied by the underlying expectations formation process. Equation 9 satisfies this criteria. In addition to the forecast error, it in cludes a lagged inflation forecast term and a retum-tonormality element. Table 7 reports statistics obtained from estimating this equation. When the error-leaming hypothesis is examined from the perspective implied by the forecast mecha nism underlying equation 9, evidence of error-learning is clearly present. The coefficient on the forecast error, |2, differs significantly from zero at the 5 percent 3 level (one-tailed test) over all sample periods except 1953-71. Recall that this coefficient reflects the rele vance of the most recently experienced inflation. The results reported above reveal that the current rate of inflation only became important in samples that in cluded the experience of the 1970s, during which in flation was accelerating sharply. Thus, Carlson’s earlier conclusions about the relevance of past errors in ex30This view is reinforced by some results reported by Mullineaux. Using a two-period distributed lag on past 6-month money growth, Mullineaux found that both lagged coeffi cients increased dramatically during the 1970s. Thus, meas ures of “normal” inflation based on a fixed-weight average of past money growth would understate the “true” normal rate. 11 F E D E R A L R E S E R V E B A N K O F ST . L O U I S APRIL 1980 Table 7 Revisions in Inflation Expectations: The Implied Model R*,t = P + Pi U*t-i + Pj Ee.t + Pi Tt* o Coefficients* Period P o P . P * P > R’/R* SEE D.W. 1953-78 -.553 (-2.830) -.174 (-2.932) .215 (4.775) .169 (2.256) .402/.365 .493 1.708 1963-78 -.145 (-.446) -.030 (-.394) .312 (6.178) -.052 (-.489) .585/.541 .448 2.393 1953-71 -.444 (-2.711) -.349 (-4.931) .050 (1.039) .245 (3.595) .431/.381 .366 1.674 1963-71 -.153 (-.731) -.289 (-1.826) .116 (2.137) .130 (.906) .564/.463 .251 2.006 -.544 (-2.860) -.254 (-4.256) .091 (1.790) .238 (3.069) .376/.331 .434 1.622 1953-78 (omitting 1972-74) Rhob -.729 (-4.395) *t-statistics are in parentheses. ''The autocorrelation coefficient is reported only for that equation which was estimated by the Cochrane-Orcutt technique because of evidence of serial correlation in the OLS estimates; t-statistic is in parentheses. plaining forecast revisions is, in one sense, reaffirmed. The error-leaming hypothesis, however, appears to have greater validity when recent, unexpectedly rapid inflation has invalidated prior forecasts. Equation 9 requires that the estimated coefficients conform to restrictions implied by the underlying forecast process. These restrictions, which are listed below equation 9, were confirmed for all sample periods in the estimates of the revision equation. In no case did the coefficients from equation 9 differ significantly from the restricted values for those co efficients derived from the independent estimates of the underlying forecasting process. SUMMARY AND CONCLUSIONS The foregoing analysis and discussion has presented evidence concerning the nature of the inflation fore casting process implicit in the Livingston price ex pectations data. Although earlier conclusions about the relevance of the error-learning hypothesis may have been valid for certain periods over the past 25 years, they do not appear to be valid for the decade of the 1970s. More important, however, is the information re vealed about the nature of the inflation forecasting mechanism. Evidence reported here indicates that 12 when inflation has been accelerating, recent inflation ary experience becomes more important in the expec tations process. This result suggests that policies which can successfully lower current inflation could reap im portant longer-run dividends by simultaneously induc ing a reduction in inflation expectations.31 The results, however, also suggest that once the economy moves from high inflation to lower inflation, retum-to-normality elements may become more important. Under a regime where planned, gradual reductions in the growth rate of money are announced and pursued, inflation expectations would seemingly adapt only slowly. On the other hand, if during periods of de celerating inflation, expectations become more respon sive to current experience — as they were during periods of accelerating inflation — expectations may well adapt more rapidly. Evidence of strong persist ence effects over all time periods suggests that break ing the inflation psychology necessarily involves a long-term commitment by policymakers to an anti inflation policy. Once such a policy is announced and undertaken, any decelerating inflation actually ex perienced should reinforce the adaptation to lower inflation expectations. 31This observation should not be interpreted as supporting incomes-policies since the adoption of price and wage con trols could be expected to alter the structure of expectations formation. Money, Inflation, and Economic Growth: Some Updated Reduced Form Results and Their Implications KEITH M. CARLSON T A HE economic experience of the United States during the 1950s and 1960s provided an opportunity to develop and test a number of hypotheses relating to the performance of the macroeconomy. One such hypothesis that received empirical support during this period held that monetary actions, as measured by movements in the monetary aggregates, have lasting effects on only nominal variables. This proposition is an important element in a body of thought called “monetarism.”1 In contrast to the relative economic tranquility of the 1950s and 1960s, the decade of the 1970s was marked by extensive experimentation with wage and price controls, large supply shocks, proliferation of government regulations, and worldwide inflation. These events and developments prompted economists to question whether or not the performance of the United States economy during this period was con sistent with prior hypotheses relating to the lasting impact of monetary actions. This article is addressed to that question. The article focuses on the magnitude of the re sponse of GNP, output, and the price level to changes in the money stock, defined as currency plus private 'For an extensive discussion of monetarism, see Thomas Mayer, et. al., The Structure of Monetarism (New York: W. W. Norton and Company, 1978). checkable deposits.2 The magnitudes of these responses are derived by estimating reduced form equations; that is, equations in which observations of the rates of change of economic variables are regressed on cur rent and lagged values of the rate of change of money and other suitably chosen exogenous variables. The sum of the coefficients on the money variable is in terpreted as a measure of the magnitude of response during the sample period from which the observations are drawn.3 THE QUANTITY EQUATION OF EXCHANGE AND REDUCED FORMS The underlying framework for the analysis is the quantity equation of exchange. This equation is an identity that states the value of all spending for goods and services in two ways: the product of the stock of money times its velocity of circulation, and the -The regressions were run before data were available for the new definitions of the monetary aggregates. Data for “old” Ml were used, and ATS and NY NOW accounts were added after 111/78. 3Whether or not this magnitude of response can be interpreted statistically as a “long-run” result depends on the length of the lag relative to the number of observations in the sample period. A reliable estimate of the long-run response of a variable that adjusts quickly and completely to an exogenous shock does not require as many observations as does a slowly adjusting variable. 13 APRIL F E D E R A L R E S E R V E B A N K O F ST. L O U I S price level times the quantity of aggregate output. In symbols, this identity is: (1 ) MV = PX = Y, where M = nominal money stock, V — velocity of circulation, P = price level, X = output, and Y = nominal GNP. V P (2 ) e(M,M)M + e(V,M )M + a = e(P,M)M + b X + e(X,M )M + c, The total differential of equation 2 results in an expression that relates the elasticities to each other: 1 + e (V,M) = e (P,M) The empirical analysis of the impact of monetary actions on GNP, output, and the price level uses previous specifications by monetarists as a starting point and modifies these specifications in light of the experience of the 1970s.8 After the equations are sum marized and the variables are defined, the equations are first estimated using data from 1955 through 1969. They are then estimated with data from the 1970s. Of primary interest is the stability of the relationships when data from the 1970s are incorporated into the estimates. Specifications and Definitions of Variables where £ is the elasticity of the first variable in paren theses with respect to the second. A dot over a vari able indicates its compounded annual rate of change. The constants, a, b, and c, represent the effect of non monetary influences on V, P, and X, respectively. (3) U.S. economy have generally concentrated on Y and P, although not always in combination.5 Nelson re cently developed justification for this choice of var iables by testing the hypothesis that the structure of the United States economy is recursive, with disturb ances from GNP flowing to the price level and not the reverse.6 Consequently, this article focuses on re duced form estimates of Y and P.7 REDUCED FORM RESULTS As an identity, the quantity equation of exchange means little. When combined with assumptions re lating to the determination of the variables, however, the equation assumes behavioral content. Writing the equation in rate of change form, where each of the variables is allowed to be influenced by money, yields the following: M 1980 + e (X,M ), (4 ) 1 + e(V,M ) = e(Y,M ). The GNP equation is specified as follows: 5 . 5 . (5 ) Y = ao + Z mi M-i + Z ei E-«, i= 0 i = 0 where Y = compounded annual rate of change of nom inal GNP, M = compounded annual rate of change of Ml (plus ATS deposits and NY NOW accounts after 111/78), and Equations 3 and 4 indicate the constraints that must be considered when attempting to estimate these elasticity parameters. An estimate of either e ( V , M ) or e(Y,M ) implies the other. Given one of these elasticities, only one of the remaining elasticities — e(P,M ) or e(X ,M ) — can be estimated. Alterna tively, estimates of e(P,M ) and e(X ,M ) imply both e(V,M ) and e(Y,M ). This equation is essentially the same as that estimated The elasticity parameters and the constants in equa tion 2 can be estimated in a variety of ways. Reduced form equations could be estimated for Y and P, Y and X, P and X, V and P, or V and X. The choice is arbitrary only if the error terms for each of these reduced form equations have exactly the same serial correlation properties.4 Monetarists researching the 8Charles R. Nelson, “Recursive Structure in U.S. Income, Prices, and Output,” Journal of Political Economy ( Decem ber 1979), pp. 1307-27. 4See Yash P. Mehra, “An Empirical Note on Some Monetarist Propositions,” Southern Economic Journal (July 1978), pp. 154-67. 14 E = compounded annual rate of change of high employment federal expenditures. 5For example, see Leonall C. Andersen and Keith M. Carlson, “A Monetarist Model for Economic Stabilization,” this Review (April 1970), pp. 7-25, and William G. Dewald and Maurice N. Marchon, “A Modified Federal Reserve Bank of St. Louis Spending Equation for Canada, France, Germany, Italy, the United Kingdom and the United States,” Kredit and Kapital (1978), pp. 194-212. 7Additional justification for the Y-P combination is found in Thomas A. Gittings, “A Linear Model of the Long-Run Neu trality of Money,” Staff Memoranda, Federal Reserve Bank of Chicago (1979). 8The specifications summarized here are the “preferred” re sults of estimating a variety of specifications. FE D E R A L RE SE R V E BANK O F ST LOUIS APRIL 1980 Table 1 Estimates of Reduced Form Equations with Pre-1970 Data (Sample Period: I/55-IV/69)1 GNP equation: Y = 3.575 + Zm, M , + Ze. E-, (3.557) m0 m, m2 Uh m4 m e Im i .275 .430 .345 .139 -.067 -.154 .966 (1.653) (5.082) (3.486) (2.038) (.837) (1.679) (4.054) .066 .070 .024 -.039 -.086 -.084 -.051 e0 ei e2 e3 e4 e -s Ze, (1.148) (2.001) (.661) (1.678) (3.265) (2.788) (.561) R! = .438; S.E. = 3.361; and D.W. = 1.934. Price equation: P = -.049 + ZniM^i - .030 (Pp - P) + Zfi (P E - P)-i, (.133) (.509) n0 ni n2 n3 n, n5 n6 nn* ns nJ0 .042 .036 .033 .032 .033 .035 .039 .043 .048 .053 .058 (1.077) (1.365) (1.811) (2.220) ( 2.318) (2.309) (2.400) ( 2.631) (2.995) (3.463) ( 3.954) n„ n,2 n,3 nu n,i n,s n1 7 n,s n,9 naj Znt .062 .065 .068 .068 .067 .063 .057 .048 .036 .020 1.008 (4.321) (4.433) ( 4.293) (4.014) ( 3.703) ( 3.414) (3.164) ( 2.953) ( 2.777) (2.629) (7.420) fo f, h U h f= Zf, -.002 .004 .007 .007 .005 .003 .024 (.062) (.230) (.358) (.381) (.262) (.149) (.279) R! = .559; S.E. = 1.094; and D.W. = 1.996. 1A polynomial distributed lags are third degree with tail constraint only; figures in paren 11 theses are absolute values of t-statistics; a dot over a variable indicates compounded annual rate of change. by Andersen and Jordan in 1968,® but modified so that the coefficients are constrained on a third degree polynomial distributed lag with t — 6 = 0 . The price equation is specified as follows: 20 where = compounded annual rate of change of the GNP deflator, 9Leonall C. Andersen and Jerry L. Jordan, “Monetary and Fiscal Actions: A Test of Their Relative Importance in Economic Stabilization,” this Review (November 1968), pp. 11-24. D2 = decontrol dummy, PF — compounded annual rate of change of the food deflator, • (6 ) P = b0 + b,D, + b2 + b3 (P p - P) + Z mM-i D2 i= 0 5 + Z f , ( P » - P ) - ,, i = 0 P Di = wage and price control dummy, M = compounded annual rate of change of Ml (plus ATS deposits and NY NOW accounts after 111/78), and PE = compounded annual rate of change of pro ducer prices for fuels, related products, and power. This specification builds on one developed by Karnosky except that it introduces variables designed to capture the influence of nonmonetary shocks on the 15 APRIL F E D E R A L R E S E R V E B A N K O F ST. L O U I S price level.10 The polynomial distributed lag is third degree for both money and energy prices with con straints of t — 21 = 0 and t — 6 = 0, respectively. Results Using Pre-1970 Data Table 1 summarizes the estimated equations using data prior to the onset of the shocks of the 1970s. For the 1955-69 period, GNP was dominated by movements in the money stock, and the adjustment to these changes was essentially complete after five quarters. The elasticity of GNP with respect to the money stock, as measured by the sum of the coeffi cients on money, was not significantly different from one at the 5 percent level and implied that the elas ticity of velocity with respect to monev was zero. 10Denis S. Kamosky, “The Link Between Money and Prices — 1971-76,” this Review (June 1976), pp. 17-23. 1980 With the sum of the coefficients on high employment expenditures not significantly different from zero, the constant term was an estimate of the trend growth of velocity. Based on these estimates, the equilibrium growth rate of GNP during the 1955-69 period was equal to the growth rate of money plus the trend rate of change of velocity. According to the estimated price equation, the rate of change of the price level was also dominated by the growth rate of the money stock in the 1955-69 period. Other factors, namely food and energy prices, were not significant in explaining overall price move ments during this period, thus confirming Kamosky’s estimate for essentially the same period. The pattern of the estimated coefficients indicated that prices ad just to a monetary shock very slowly, but the total effect after 20 quarters was an elasticity of the price level equal to one. Since neither the constant term Table 2 Estimates of the GNP Equation1 Y = ao + Zmt M-i + ZeiE-! Coeff. I/55-IV/69 t .275 1.653 Cum.2 Coeff. I/70-IV/79 t Cum.2 Coeff. I/55-IV/79 t .657 2.358 .657 .407 2.968 .407 .815 1.096 1.200 Cum.2 m0 m, .430 5.082 .275 .704 .376 2.279 1.033 .407 5.344 m2 m > .345 .139 3.486 2.038 1.049 1.187 .169 .031 .986 .243 1.201 1.232 .282 .104 3.271 1.782 m4 m5 -.067 -.154 .837 1.120 .966 -.040 -.049 1.144 .242 1.193 1.144 -.052 .752 1.148 1.415 6.115 1.037 .053 .107 1.679 4.054 .282 2.350 Zmi .966 e0 e, e2 .066 1.148 .066 .039 .070 .024 2.001 .136 .068 .544 1.041 .661 .159 .079 e3 -.039 1.678 .120 e4 -.086 3.265 .033 e5 -.084 2.788 -.051 Ze. -.051 a. 3.575 -.111 1.037 .039 .053 1.211 1.248 .106 .186 .055 .023 1.805 .768 .130 .076 1.491 .261 -.021 1.008 .109 .060 1.067 .321 -.054 2.203 .056 .034 .627 .355 -.054 2.064 .001 .561 .355 1.613 .001 .014 3.557 -1.078 .262 3.159 3.437 .410 S.E. 3.361 .272 3.934 3.620 D.W. 1.934 2.331 1.924 R3 .438 1All polynomial distributed lags are third degree with tail constraint only; t-statistics are absolute values; a dot over a vari able indicates compounded annual rate of change. -Numbers are the cumulative sum of coefficients. 16 F E D E R A L R E S E R V E B A N K O F ST . L O U I S APRIL 1980 Table 3 Estimates of the Price Equation1 P = b„ + biDi + b:D, + b, (P F - P) + Z mM-, + Zf, (P . - P)-, Coeff. I/55-IV/69 t Cum.- CoefF. I/70-IV/79 t Cum.2 Coeff. I/55-IV/79 t n0 n, .042 1.077 .042 .056 1.087 .056 .061 .036 1.365 .078 .087 2.183 .143 .064 1.969 3.006 n2 .033 1.811 .112 .110 3.095 .253 .067 4.360 Cum.2 .061 .125 .192 n3 .032 2.220 .144 .126 3.419 .379 .068 5.300 .260 Ui .033 2.318 .177 .134 3.322 .069 5.303 .329 n5 n« .035 .039 2.309 .212 3.076 .251 2.794 .784 .069 .068 4.967 4.717 .397 2.400 .137 .134 .513 .650 n7 .043 2.631 .295 .127 2.505 .910 .066 4.625 .465 .531 n8 n» .048 2.995 3.463 .343 .116 2.211 1.026 .064 4.668 .596 .053 .396 .102 1.909 1.128 .061 4.790 .657 n,0 .058 3.954 .454 .086 1.600 1.215 .058 4.889 nn ni2 .062 4.321 .516 .069 1.284 1.284 .054 4.806 .715 .770 .065 .068 4.433 .582 .052 .967 1.335 .050 4.422 .820 nn .649 .034 .866 .717 .018 1.370 1.388 3.804 .068 .655 .355 .046 n1 4 4.293 4.014 .041 3.139 .907 n« .067 3.703 .784 .004 .074 1.391 .036 2.554 .942 11 1® .063 3.414 .848 -.008 nn .905 -.017 1.383 1.366 .024 2.085 1.718 .972 3.164 .185 .419 .030 .057 n1 8 H is .048 2.953 .953 -.021 .627 1.345 .018 1.432 1.015 .036 .989 no ^ Zn, .020 1.008 2.777 2.629 -.020 -.013 .811 .971 1.326 1.312 .012 .006 1.208 1.030 1.027 1.033 1.312 1.706 1.033 11.459 Io -.002 .062 -.002 -.001 .060 -.001 .008 .909 .008 II fa .004 .007 .007 .230 .358 .019 .025 .021 3.490 4.195 3.916 .019 .044 .064 .031 .058 .078 .021 .012 1.835 .149 .024 .003 .523 .002 2.179 .469 .089 u .076 .079 .024 .026 .020 .011 5.151 5.105 4.971 .005 .003 .381 .262 .003 .009 .016 If. .024 .279 .079 3.064 .091 5.274 bo b, -.049 — .133 — -1.522 .345 3.487 -.018 .052 -2.193 -2.256 3.801 -1.655 1.846 -.316 .407 .130 1.833 .060 1.273 f» u 1.008 7.420 b2 — — b3 -.030 .509 .559 .802 1.094 1.262 1.268 D.W. 1.996 2.180 .091 .819 S.E. .996 1.735 RJ 'All polynomial distributed lags are third degree with tail constraint only; t-statistics are absolute values; a dot over a vari able indicates compounded annual rate of change; D! and D. are wage and price control and decontrol dummies, respectively. ^Numbers are the cumulative sum of coefficients. 17 F E D E R A L R E S E R V E B A N K O F ST . L O U I S APRIL 1980 Table 4 Results of Chow Test (I/55-IV/69 vs. I/70-IV/79) Critical F GNP equation Price equation Calculated F Conclusion F.05 (7,86) ==2. 12 = F.05 (8,60) 2.05 1.20 3.23 Cannot reject Ho1 Reject H o ’ H is the null hypothesis that the regression equations are equal for the two sample periods. o nor the effect of nonmonetary shocks were significantly different from zero, the equilibrium rate of change of the price level during this sample period was equal to the rate of monetary growth. These two estimated equations implied that the equilibrium rate of output growth was independent of the rate of monetary expansion during the 1955-69 period. This implication was derived from an exami nation of the elasticity estimates in conjunction with equations 3 and 4; e(Y,M ) = 1 and e ( P , M ) = 1 together implied that e(X ,M ) = 0. In other words, these estimated reduced form equations substantiated the hypothesis that monetary actions have lasting ef fects on only nominal variables. Updated Results Tables 2 and 3 present the results using data from the 1970s. For purposes of comparison, the pre-1970 estimates are also summarized. The Chow test was used to check the equations for stability. Estimated Equations — Updated estimates for the GNP equation are shown in table 2 and for the price equation in table 3. Results are shown both for the 1970s (I/70-IV /79) and for the extended sample pe riod (I/55-IV /79). The sum of the coefficients on money in the GNP equation was not significantly different from one at the 5 percent level, either for the 1970-79 period or for the fully extended sample period 1955-79. Esti mates of the constant, however, indicated a decline in the trend growth of velocity when data for the 1970s were included. For the 1970-79 sample period, the estimated constant was negative but not signifi cantly different from zero at the 5 percent level. For the fully extended sample period, however, the con stant was positive and significantly different from zero, but the point estimate was less than that for the 1955-69 period. The estimate of the price equation for the 1970-79 period showed an increase in the sum of the coeffi 18 cients on money. However, this sum was not signifi cantly different from zero at the 5 percent level. For the fully extended sample period, the sum of the money coefficients was significantly different from zero, although not significantly different from one at the 5 percent level. Estimates of the remaining coefficients indicate that nonmonetary factors, namely energy prices and wage and price controls, influenced price level movements during the 1970s, and to such an extent that they were also significant over the full sample period. Esti mates of the constant term for both the 1970-79 and 1955-79 periods were not significantly different from zero at the 5 percent level. Tests for stability — The updated results suggest some conflicting conclusions. The Chow-test of sta bility was used to investigate further the appropri ateness of simply extending the sample period to in clude the 1970s.11 Table 4 summarizes the results of applying this test to the GNP and price equations. The test results show that the hypothesis of stabil ity for the GNP equation for the two sample periods was not rejected. However, the hypothesis of stability was rejected for the price equation. The interpreta tion of these results is that the GNP equation, as estimated over the full sample period, can be used to summarize that relationship. However, the choice of the estimated price equation depends on the period that is chosen for analysis.12 Implications of the Results for the Relationship between Money and Output One implication of the reduced form results using data prior to 1970 was that the equilibrium rate of n Gregory Chow, “Tests of Equality Between Sets of Coeffi cients in Two Linear Regressions,” Econometrica (July 1960), pp. 591-605. 12These price equations should not be interpreted as long-run equations, however, because the sample periods are so short. See footnote 3. F E D E R A L R E S E R V E B A N K O F ST . L O U I S output growth was independent of monetary growth. In other words, trend output was determined by real factors: namely, growth of the labor force, capital stock, and technology. When the reduced form equations were updated with data from the 1970s, the implication for equilib rium output was modified. In a strict statistical sense, the hypothesis that monetary actions have lasting effects only on nominal variables was not rejected when data from the 1970s were included in estimating the relationships. However, when the estimated GNP equation for the full sample period was combined with the price equation for the 1970-79 period, the growth of money appeared to influence the rate of growth of output. Although e (Y,M) was still approxi mately one, the point estimate of e(P,M) was 1.31. Consequently, based on the experience of the 1970s, the point estimate of e(X,M) was —.31. The nature of this result, although statistically ten tative, is summarized in table 5. Underlying the cal culations in this table is the assumption that non monetary shocks equal zero. These results, although they do not demonstrate causality, provide indirect support for the view that there is a negative relation between the trend rate of monetary growth (and in flation) and the trend rate of economic growth. This contention that inflation adversely affects out put has received increasing emphasis in the recent literature.13 One view is that inflation slows growth by discouraging investment and saving via the exist ing tax structure.14 The inflation process increases effective tax rates for both individuals and firms and lowers after-tax rates of return, thereby reducing in centives to invest and save. Another argument stresses the uncertainty associ ated with inflation.18 If higher and higher inflation rates also mean greater risks associated with invest ment planning, saving and investment will be dis couraged because a given expected rate of return will be accompanied by a greater variance. 13For general discussions of possible factors contributing to the slowdown of productivity in the 1970s, see Edward F. Denison, “Explanations of Declining Productivity Growth,” Survey of Current Business (August 1979), pp. 1-24; and John A. Tatom, “The Productivity Problem,” this Review (September 1979), pp. 3-16. APRIL 1980 Table 5 Relationship between Trend Output and Money Growth Rate of growth of money 0% 2 4 6 8 Rate of growth of output based on: — --------- —----------------— -------— ----Pre-1970 results1 Updated results1 3.62% 3.54 3.46 3.37 3.29 4.68% 4.13 3.58 3.03 2.48 ■These calculations are based on the point estimates of the parameters in the GNP and price equations and assume that nonmonetary influences are equal to zero except for the constant terms. Still another explanation of the inflation-growth connection is that the inflation process introduces “noise” into the price signals that are transmitted from consumers to producers.16 As a result, the general efficiency of the price system in allocating resources is reduced. Such a reduction must be manifested in a reduced growth rate of output. SUMMARY This article presents updated reduced form results relative to the hypothesis that monetary actions have a lasting impact on only nominal variables. When data from the 1970s were included in the sample, this hypothesis could not be rejected for either the 1970-79 period or the 1955-79 full sample period. When the reduced form equations were tested for stability over the entire period, the hypothesis of sta bility for the GNP equation could not be rejected; but the null hypothesis for the price equation was rejected. When the GNP equation for the fully ex tended sample period was combined with the price equation for the 1970-79 period, the point estimates of the coefficients suggested that the rate of growth of output was negatively related to the growth rate of money during the 1970s. Even though only sugges tive, the results provide tentative evidence to support the notion that real economic gain can be achieved by reducing the trend growth of money.17 14A recent study providing evidence relating to the effect of in flation on corporate rates of return is reported in Martin Feldstein and Lawrence Summers, “Inflation and the Taxa tion of Capital Income in the Corporate Sector,” National Tax Journal (December 1979), pp. 445-70. 16Milton Friedman, “Nobel Lecture: Inflation and Unemploy ment,” Journal of Political Economy (June 1977), pp. 451-72. 16See Stephen L. Able, “Inflation Uncertainty, Investment Spending, and Fiscal Policy,” Federal Reserve Bank of Kan sas City Economic Review (February 1980), pp. 3-13. 17Laurence H. Meyer and Robert H. Rasche, “On the Costs and Benefits of Anti-Inflation Policies,” this Review ( Febru ary 1980), pp. 3-14. 19