The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
W a g e G r o w t h a n d Sectoral Shifts: Phillips C u r v e R e d u x Ellen R. Rissman LIBRARY F E B 11. 199 3 FEDERAL r e s e r v e BANK OF CHICAGO Working Papers Series Macroeconomic Issues Research Department Federal Reserve Bank of Chicago December 1992 (W P-92-23) FEDERAL RESERVE BANK OF CHICAGO W a g e G r o w t h a n d Sectoral Shifts: Phillips C u r v e R e d u x E llen R. R issm an 1 Federal Reserve Bank of Chicago December 3, 1992 1All correspondence should be sent to the author at the Federal Reserve Bank of Chicago, 230 S. LaSalle St., Chicago, IL 60690. This paper is a revised version of a 1987 working paper titled “Wage Growth and Sectoral Shifts: New Evidence on the Stability of the Phillips Curve”. The author would like to thank Prakash Loungani and Steve Strongin for their helpful comments, suggestions, and encouragement. I alone am responsible for any remaining errors. The views expressed here are not necessarily those of the Federal Reserve Bank of Chicago or the Federal Reserve System. Abstract Standard Phillips curve analyses of nominal wage growth performed poorly over the Seventies. Because wage pressures arising from sectoral shocks largely balance out, only that portion of unemployment that is net of the effects of compositional shifts in the structure of labor demand should influence nominal wage growth. Sectoral disturbances in the Seventies caused the unemployment rate to rise independently of cyclical activity. Once unemployment has been corrected to reflect the effects of sectoral shifts, the apparent instability of the Phillips curve is resolved. This result holds for both an employment and a stock price dispersion measure of sectoral shifts. 1 In trod u ction Advocates of the Phillips curve approach to modeling the inflation process typically relate wage or price inflation to some measure of current economic activity and an inflationary expectations variable. Economic activity is usually proxied by the unemployment rate or some function of the unemployment rate, and expected inflation is specified as a function of lagged values of actual inflation. This traditional approach to the inflation process appears to have dwindled in popularity in recent years as many researchers have found that the parameters of the Phillips curve changed over the Seventies. Cagan (1975), Sachs (1980), and O ’Brien (1985) argue that the trade-off between inflation and unemployment has worsened in that a given decline in the rate of wage inflation requires a larger increase in the rate of unemployment than previously needed, i.e. the short run Phillips curve has become flatter. In contrast, Wachter (1976) and Schultze (1981) suggest that the opposite is true with wage inflation becoming more cyclically sensitive, implying a steeper short run Phillips curve. Attempts to resurrect the Phillips curve have typically fallen along two lines of reasoning. First, unstable parameters may be the result of an initial misspecification of the inflation ary expectations process. Noting this, several researchers including Sachs (1980) and more recently Neumark and Leonard (1991) have experimented with alternative expectations for mulations and more sophisticated models of wage and price dynamics. However, as Sachs (1980) observes, the evidence indicates that more complicated models of expected price in flation do not account for what appears to be a significant change in the short run trade-off between inflation and unemployment. Omitted variables offer an alternative explanation for the parameter instability. Several researchers, including Gordon (1977), Perry (1980), and Hamilton (1983), have suggested that demographic changes combined with oil price shocks, the acceleration and termination of the Vietnam War, and the implementation of wage and price controls account for the instability of the Phillips curve. These theories have met with only modest success. Others, including Sachs (1980), Barro (1977), and Taylor (1980), argue that there has been a change 1 in the underlying structure of the inflation process. For example, an increase in the use of longer term labor contracts and a belief held by the public that monetary and fiscal policy will be used so as to promote high employment and stable prices have worsened the trade-off. Although there issome theoreticaljustification for this hypothesis, empirical evidence on this point islacking. Oddly, in light of research on sectoral shifts and their relation to the unemployment rate, the search for alternative specifications of the Phillips curve has led to relatively little consideration ofthe appropriate measure ofeconomic activity to be employed in the analysis.1 Typically, the unemployment rate or some function of the unemployment rate is used to capture the disequilibrium aspect although some researchers, including Sachs (1980) and Gordon (1977), have chosen instead to use alternative measures such as deviations of real Gross National Product from potential. These variables are allmeant to measure the current level ofeconomic activity. While rigorous theoretical underpinnings of the Phillips curve phenomenon are conspic uously absent in these works, discussion of wage inflation is typically expressed in terms of labor market tightness. Controlling for inflationary expectations, the negative coefficient on the unemployment rate in a reduced form regression of wage inflation is interpreted as indicating that tight labor markets, as signaled by lower unemployment rates, put upward pressure on wages and, therefore, on prices through some sort of markup equation. According to the sectoral shiftshypothesis ofunemployment [Lilien (1982)], changes in the distribution ofdemand, given itslevel, cause the unemployment rate to increase as itiscostly both for the employer and employees to adjust instantaneously. Therefore, the sectoral shifts hypothesis suggests that higher unemployment is not necessarily indicative of a weakening economy, while similarly lower unemployment need not signify stronger aggregate demand. In the case of structural realignment, the unemployment rate is no longer an accurate measure 1R ecent works on the topic of sectoral shifts and the unem ploym ent rate include Lilien (1982, 1990), N eum ann and Topel (1991), Loungani (1985), Loungani, R ush, and Tave (1990), D avis (1985), Abraham and K atz (1986), and Brainard and Cutler (1990). A braham and M edoff (1982) exam ine different m easures of econom ic activity but do not relate this to the sectoral shifts hypothesis. 2 oflabor market activity. Unemployment resultingfrom sectoral reallocations does not affect the wage inflation pro cess in the same way as cyclical unemployment since any upward wage pressure that results in one sector will be tempered by downward wage pressure in another. Ifthe sectoral shifts hypothesis is correct and sectoral reallocation was responsible for the higher unemployment rates in the Seventies, the Phillips curve would then appear to be unstable over this time pe riod. Specifically, a higher unemployment rate would be required to induce a given reduction in wage inflation than needed when sectoral shifts were relatively unimportant. The purpose of this paper is to analyze the effect of permanent sectoral shifts on the wage inflation process. The remainder consists of four sections. In Section 2 two different measures ofsectoral shifts are examined, namely Neumann and Topel’s(1991) construction of permanent changes in the distribution ofemployment across industries and the stock market dispersion index of Loungani, Rush, and Tave (1990). These measures are then used in Section 3 to construct estimates of the natural rate of unemployment. In Section 4 the stability ofthe Phillips curve isexamined with particular emphasis given to the role of sectoral shifts. A test of the sectoral shifts hypothesis in its strictest form is proposed that tests whether the coefficients on the natural and actual rates ofunemployment are of equal and opposite signs as would be expected if sectoral shifts are noninflationary. Conclusions and suggestions for further research are found in Section 5. 2 M easuring S ectoral Shifts Generally, two distinct ways have been proposed in the literature formeasuring sectoral shifts. One is based upon dispersion in employment growth across industries and the other upon stock price dispersion.2 Abraham and Katz (1986) argue that employment-based dispersion measures poorly capture sectoral shifts since, ifa declining industry is also relatively more cyclically sensitive, a positive correlation between an employment-based dispersion index 2Lilien (1982, 1990) and N eum ann and Topel (1991) have focused on the former while Loungani, Rush and Tave (1990) and Brainard and Cutler (1990) have exam ined the latter. 3 and the unemployment rate can result purely from aggregate disturbances. They suggest that these two factors have been present in post-War U.S. data so that a sectoral shifts interpretation ofemployment growth dispersion isinappropriate. As a result oftheir criticism, researchers have attempted to both refine their employment-based measures ofsectoral shifts and create new measures that do not suffer from the Abraham and Katz critique. Ifchanges in the distribution of employment across industries are fundamentally related to the business cycle, then a given industry’s employment response should be consistent over the various cycles after controlling for the intensity of the cyclical shock. Define gi as the growth rate of employment in the industry from business cycle peak to subsequent trough. Similarly, define g as aggregate employment growth. Relative employment growth, <7*/<7, for the 20 two-digit manufacturing industries over the eight post-War business cycles is found in Table 1. The industry response varies greatly from cycle to cycle. For example, Non-Electrical Equipment lost over twice as much as the industry average in the recession of 1949 but in 1974 the industry lost almost no workers, even though the ’74 recession was more severe. Similarly, Electrical Equipment was particularly hard-hit in 1970 but fared relatively well in the 1960 downturn. Ifthe business cycle hypothesis is correct, then the cross-cycle variance ofg ijg should be relatively small. In testing this hypothesis, a fundamental problem arises in determining how small the variance should be under the null. A somewhat weaker variant ofthe business cycle hypothesis isthat the underlying distributions ofthe observations are the same across cycles. A likelihood ratio test of the equality of the variances of the observations across cycles can be easily constructed from the sample within-cycle variances.3 The hypothesis of constant 3T he com puted test sta tistic ,—2 log A, equals 33.73 and is asym ptotically distributed as degrees o f freedom . T h e likelihood ratio is given by: A = n < * ; > ‘v ;=i where 20 / 4 \ 2 w ith (8-1) variance across cycles is rejected. Although some industries may fare relatively better than others in terms of employment over the course ofthe business cycle, there issufficient variation in the employment response of a given industry across cycles to suggest that changes in the sectoral composition of employment over time are driven by more than simply general cyclical fluctuations. While a cyclical interpretation is not precluded, any cyclical explanation must necessarily explain the differing character of these cycles. 2.1 D ispersion M easures In the analysis that follows, two different measures ofsectoral shifts are examined: Neumann and Topel’s (1991) employment-based measure, A p,and Loungani, Rush, and Tave’s (1990) stock price dispersion measure SQ . Unlike Lilien’s (1982) original measure, each of these is constructed in such as way as to capture perm anent shifts in the relevant distribution. Neumann and Topel (N-T) focus on decomposing deviations inemployment shares from trend into two components— one predictable from observations on future employment shares and one which is orthogonal to it. Loungani, Rush, and Tave (LRT) suggest that a stock market dispersion index is weighted towards capturing permanent shifts in an industry’s expected profitability. Since stock prices reflect the present discounted value of expected industry profits over an infinitely long horizon, the impact of a current innovation in profits on stock price depends on whether that innovation is expected to be temporary or permanent.* 4 The Abraham and Katz critique suggests that an appropriate measure of sectoral shifts should be independent of past unemployment, i.e. that the unemployment rate should not Granger-cause sectoral shifts. The advantage to the stock price dispersion measure is that movements in stock prices precede changes in unemployment.5 However, reallocation timing arguments such as proposed by Hamilton (1983), Davis (1985), and Rogerson (1987) suggest and j = 1 , . . . , 8 indexes the cycles. T he sm all sam ple properties of the test sta tistic are unknown. 4First differences in A p and S Q have a correlation coefficient of -0.11. 5T o sum m arize the results of Granger-causality tests, unem ploym ent appears to Granger-cause A p but does not Granger-cause S Q . 5 that reallocation in response to a sectoral disturbance will occur when the economy is in a low state and the opportunity cost issmallest. In light ofthe reallocation timing arguments, one would expect to see unemployment Granger-causing dispersion to the extent that the unemployment rate also reflects cyclical disturbances. Both Ap and SQ have the benefit of being forward looking and capturing permanent changes. However, the interpretation ofS Q isunclear ifmarkets are not efficient. In addition, there may be factors that affect the return on capital that do not necessarily influence labor demand. To that extent, stock market dispersion would be a poor measure of those sectoral shifts that entail some reallocation oflabor. For this reason an employment-based measure of sectoral shifts seems better able to capture those disturbances that do influence the pattern of labor demand and presumably have some bearing on the unemployment rate. 3 T h e N atu ral R a te 3.1 S ectoral Shifts and U nem ploym ent There are many factors which may affect the unemployment rate. According to the standard sectoral shifts theory of unemployment as suggested in Lilien (1982), permanent changes in the distribution of employment across industries should increase the unemployment rate in the short run as a mismatch between workers and employers results. Over time the long term effects of a sectoral shift disappear as agents adjust to the disturbance. The demographic composition of the labor force has been cited as one cause for the rise in the unemployment rate over the Seventies as a larger proportion of those groups having traditionally higher levels ofunemployment, namely women, nonwhites, and youths, entered the labor market. In the analysis that follows a Perry-weighted unemployment rate ( W U R ) that has been adjusted to reflect the age-race-sex composition of the labor force in 65Q1 is used to capture this demographic effect. Traditional macroeconomic models focus on cyclical factors as the driving force behind movements in the unemployment rate. Following Lilien (1982), this cyclical effect isassumed 6 to be captured by ‘unanticipated money growth’,D M t,and deviations ofreal Gross National Product from trend, D G N P t.e Finally, unemployment insurance and other social welfare programs are thought to have an effect on the unemployment rate. On the one hand, such programs subsidize job search with the effect of encouraging a longer period of unemployment. On the other hand, the extended length ofjob search that results should lead to better employee/employer matches and a reduction in future unemployment. The net effect is ambiguous. The effect of social welfare expenditures is assumed to be measured by social insurance expenditures expressed as a percentage of Gross National Product, S it - 6 7 Because social insurance expenditures typically rise during economic downturns, the ratio ofsuch expenditures to G N P iscounter cyclical. Thus, S I may simply proxy for additional cyclical information independent of any hypothesized effect it may have upon job search or job matching. Unemployment rate regressions were estimated of the form: [1- b0( L ) L ] W U R t = c + bx{ L ) D G N P t + b2( L ) D M t + b3( L ) S I t + b4( L ) a t + et (1) where the bi(L) are polynomials in the lag operator, L, and a t is some dispersion measure of sectoral shifts.8 For a = A p,the preferred regression includes bo(L) and b4(L) as first order polynomials, with b\(L) and b2( L )as second order polynomials, and b3(L) as a constant. The preferred regression for a = SQ is of the same form as for the employment-based dispersion measure with the exception that b4(L) has a lag length of 16 quarters. 6 A s defined in Barro (1978), unanticipated m oney growth is constructed as the least squares residuals from a regression o f M l growth on a vector of explanatory variables. T he construction varies slightly from that described in Barro since the d ata employed here are quarterly while Barro’s estim ates were com puted from annual data. A side from a constant, the vector of explanatory variables includes eight lags of m oney growth, current and three lags o f the log of real federal government expenditures calculated as d eviations from a one sided m oving average of past values, and four lags of the ratio of unem ploym ent to em ploym ent expressed in logarithm s. Current and eight lags o f the log of real federal governm ent expenditures are used to calculate the moving average w ith geom etrically declining weights that sum to unity and adaptive param eter o f 0.2. T he variable D G N P t is the deviation in the log of real G N P from a linear trend. 7T he data are reported annually in the Social Security A dm inistration’s S o c i a l S e c u r i t y B u l l e t i n : A n n u a l S t a t i s t i c a l S u p p l e m e n t . Linear interpolation was used to obtain quarterly figures. 8 D etailed results are available from the author. 7 3,2 C o n stru c tio n of th e N a tu ra l R a te The “natural rate ofunemployment”,according to Lilien (1982), isthe rate ofunemployment that would occur in the absence of cyclical fluctuations. In Lilien’s formulation it reflects only the unemployment that occurs as a consequence of sectoral shifts and the resulting temporary mismatch of workers and employers. The terminology is somewhat misleading as the “natural rate” has typically referred to the rate of unemployment compatible with nonaccelerating or nondecelerating wage growth and has usually been derived from a relation between unemployment and nominal wage growth. The natural rate that Lilien refers to does not necessarily have any bearing on the question of nominal wage growth. In fact, one can view equation (1) above as derived from a production function or Okun’s Law with sectoral shifts affecting the production possibility frontier. In this context the “natural rate of unemployment” is unrelated to the Phillips curve phenomenon. Perhaps a more accurate expression for the construct that Lilien proposed is ‘structural unemployment’as it refers to that unemployment which isthe result ofthe changing structure oflabor demand. However, in keeping with the literature on sectoral shifts, I will continue to use the ‘natural rate’ terminology. The natural rate of unemployment, in this context, is simply calculated as the rate of unemployment that occurs when the cyclical variables, namely DGNP and Z?M, are set identically equal to zero over the entire time period analyzed. How one treats other variables in constructing the natural rate isopen to some debate. Ifone isconcerned exclusively with a measure ofunemployment associated with sectoral reallocation, then the natural rate should be constructed holding these other variables equal to their mean values as in Lilien (1982) for example. However, a broader measure that is consistent with an interpretation of the natural rate as non-cyclical or structural unemployment would permit these other variables to vary over time. The latter approach is taken here.9 Figure 1 presents the actual age-weighted unemployment rate and estimates of the nat 9T o im plem ent these com putations, it is necessary to specify initial values o f the natural rate, NUR. Initial values of the natural rate are taken to be the actual values o f the dem ographically-w eighted unem ploym ent rate. T h e effect o f these initial values on the calculations decreases over time. 8 ural rate of unemployment calculated from the estimated parameter values of the preferred equations (1). The natural rate series calculated from A p and S Q are denoted as N U R jv t and N U R l r t respectively. It should be noted that because the dependent variable in the re gression analysis, W U iZ,isa demographically fixed-weight unemployment rate, the estimates of the natural rate constructed are also independent of demographic effects. Generally, it is presumed that low unemployment rates signal a healthy economy while high unemployment rates indicate low aggregate demand. Such an interpretation is valid only ifother non-cyclical factors influencing unemployment are stable. In this case, move ments in the unemployment rate correspond to changes only in the cyclical component of unemployment. The estimates of the natural rate constructed above indicate that the im portance of such other factors has varied considerably over the time period examined. The series N U R l r t ismore or less levelin the Sixties at about 5.5% and climbs over 2 percentage points during the Seventies to a peak of 7.5% in the fourth quarter of 1982. This measure of the natural rate falls to 6.4% by the end of the sample period. The natural rate series constructed from SQ is much smoother than that constructed from A p . However, the time series patterns are more or less consistent. The two natural rate series have a correlation of 0.87 in the levels but only 0.17 in first differences. The natural rate constructed from A p has varied considerably being as high as 8.0% in ’83 and as low as 4.7% in ’66. The relation between the natural rate and the actual rate ofunemployment has changed considerably over time. N U R n t remained fairly stable until the late Sixties while the actual unemployment rate fluctuated around it. The natural rate exceeded the actual rate ofunem ployment for a brief period from ’55 to ’57 and then for a much longer period of time from 65Q2 until 74Q3. However, between 69Q1 and 71Q3 the natural rate of unemployment rose over 1.5 percentage points in response to underlying sectoral shifts. Between 74Q3 and 77Q3 actual unemployment exceeded the natural rate. This was followed by a briefperiod oftight labor markets in the late Seventies. Both the natural rate and actual rate of unemployment continued to rise during the early Eighties with labor markets appearing slack. From 85Q4 9 until the end of the sample period labor markets again appear to be tight with the natural rate exceeding the actual rate by a small amount. 4 W age In flation and th e N atu ral R a te 4.1 S ta n d a rd A nalysis Traditional approaches to modeling the inflation process describe the rate of change ofwages or alternative price variable in terms ofitsequilibrium and disequilibrium components. Thus, itishypothesized that the rate ofgrowth ofwages, w ,isa function ofexpected price inflation and the difference between labor demand and labor supply. Assuming that wage inflation is a linear function of these variables, then wage growth is expressed as: wt — + 6\ ( L t — L\)+ 62Pt (2) where w is the logarithmic rate of change of wages at time £,and L * and L \ are respectively labor demand and labor supply at time t,pt isthe inflation rate at time /,and the superscript 4e’indicates that the variable is in expectations form. The parameter <$1 is hypothesized to be positive. In estimating equation (2) the labor market conditions variable, L*}— L*,isusually proxied by the actual unemployment rate. It has been suggested that the wage response is larger as labor market conditions become tighter. Therefore, the unemployment rate isusually entered in inverse form to capture this nonlinearity. Implementation of equation (2) also requires that the expectations process be specified. Various forms have been investigated in the literature. Typically itisassumed that expected inflation depends upon lagged realizations of actual inflation so that p\ = b(L)pt, where b(L) is a polynomial in the lag operator L. In the analysis that follows it is assumed that expected inflation isdescribed by a second order polynomial distributed lag model so that the coefficient on the lag, isexpressed as (a0+ aiz+ a2 *2)- The inclusion ofbeginning and endpoint constraints changes this three parameter model to a one parameter model. Finally, 10 for purposes ofestimation itis assumed that an additive error term isincluded at the end of equation (2).10 Table 2 provides OLS parameter estimates and associated t-statisticsforthe wage inflation model described above using quarterly observations for two sample periods. The restricted sample extends from 60Q2 to 81Q3 while the full sample includes data through 87Q1. There are two reasons for separating the data in this way. First, it is widely accepted that the parameters ofthe Phillips curve shifted over the Seventies.11 The restricted sample facilitates comparison with previous results. Second, as noted by Neumark and Leonard (1991), the rapid decline in nominal wage growth over the Eighties suggests that the parameters of the Phillips curve may have changed yet again. For this reason, the full sample is analyzed. The dependent variable in the analysis is the difference in logarithms of average hourly earnings for production workers in manufacturing and nonsupervisory workers in nonmanu facturing. The labor market activity variable isthe demographically-weighted unemployment rate, W U R ,entered linearly. Price inflation is calculated as the difference in logarithms of the Consumer Price Index for urban workers. The coefficient on p e reported is the estimate of 820,2 in an eight quarter polynomial distributed lag on past inflation with beginning and endpoint constraints. A negative sign on the coefficient estimate indicates that the 6t-are all positive and concave in i assuming that 82 > 0. In column (1) of Panel A the coefficient estimates ofthe standard specification are found. The sign of the coefficient on the unemployment rate is negative and clearly significant at traditional confidence levels. In addition, higher expected price inflation is associated with higher wage growth as indicated by the parameter estimate on pe.12 The relatively low 10Any measure o f expected price inflation that may be used in the regression analysis is m easured with error. It is well known that such a classical errors in variables problem biases the OLS param eter estim ates. Specifically, the coefficient on the expected inflation variable will be biased towards zero w ith the m agnitude depending upon the variance of the m easurem ent error relative to the true series. Furthermore, the other param eters of the m odel are also biased with the direction depending upon the variance-covariance m atrix of the observations. W ithout prior knowledge of the variance o f the errors, it is difficult to correct for the problem . However, the estim ates o f the coefficient on expected inflation are surprisingly robust to alternative specifications. Therefore, it is hoped that the biases introduced are small. 11See Cagan (1975), Sachs(1980), Gordon (1977), W achter (1976), and Schultze (1981). 12T he estim ation was carried out using a one-step procedure. Several other lag structures were tested that 11 Durbin-Watson statistic suggests that some underlying factors have not been properly in cluded in the analysis and that the estimated standard errors are incorrect. Various alternative specifications were estimated to investigate the idea that the wage inflation process was somehow different over the Seventies. The results are presented in columns (2) through (4) of Panel A. A dummy variable D is defined equaling 1 after the second quarter of 1971, when wage and price controls were introduced, and 0 otherwise. The regression reported in column (2) tests ifthe trade-offbetween wage inflation and unemploy ment worsened in recent years, holding all other parameters constant over the entire time period, by checking the significance ofthe interaction term D X W U R . The results appear to support the contention of Cagan (1975) and Sachs (1980) that the trade-offbetween inflation and unemployment worsened over the latter partof the sample. The Durbin-Watson statistic improves with this specification. The regressions in columns (3) and (4) test whether a change in the effect of expected inflation on wage growth occurred post-’71. There isno significant change in the effect of the expected price variable in the latter part ofthe sample when the coefficient on W U R is held constant over the entire period. However, when both W U R and p e are permitted to change in the latterpart ofthe sample, both interaction terms are significant and the Durbin-Watson statistic increases to 1.9. Various other specifications were tried testing whether the shift in the parameters could be explained by the wage and price controls of the early Seventies, changes in the union composition of the labor force, and changes in the competitive position of the economy as are less restrictive. However, the regression results did not im prove significantly and are unreported here due to space lim itation s. In addition, the estim ation was also performed with the unem ploym ent rate variable entered in inverse form. T h ese regressions in general had lower Durbin-W atson sta tistics and lower values of R 2. A tw o-step procedure which is unreported in the tex t was also performed. First, the param eters of the polynom ial b ( L ) were estim ated by an unrestricted regression o f current price inflation on eight lagged values. T h e sum of the coefficients was 0.95. An F -test of whether the coefficients sum to unity produced an F-value of 2.70 which is below the 5% critical value o f 3.91. T h e predicted values from this m odel were then used as the expected inflation regressor in estim ating equation (2). T h e estim ate of 62 calculated in this way is 0.50 w ith a standard error o f 0.05. Because o f the errors in variables problem that occurs, this estim ate o f 62 is more properly thought o f as a lower bound. By calculating the variance-covariance m atrix o f the observations, the coefficient estim ate on the unem ploym ent rate is also found to be biased towards zero. 12 measured by the ratio of imports to Gross National Product. Although inclusion of these variables helped explain some ofthe apparent instability ofthe Phillips curve, the coefficient on D X W U R remained positive and significant in the restricted sample period. The remaining columns in Table 2 reproduce the analysis for the full sample period. The only difference in specification is that a dummy variable, D U M M Y ,has been included that takes on a value of 1 from 81Q4 to 87Q1. It appears that post-’81 nominal wage growth dropped independently ofany changes in the coefficients on W U R and p e. The similarity of the results between the two sample periods once a change in the intercept has been permitted isquite striking. The estimated parameters are ofsimilar magnitudes and are estimated with similar accuracy as those estimated on the restricted sample. Various other specifications were tried that permitted the coefficients on W U R and p e to change post-’71 while keeping the intercept constant. In general, these regressions did not perform well. 4.2 A d ju stm e n t for S ectoral Shifts Unemployment due to sectoral reallocation does not affect the wage inflation process in the same way as cyclical unemployment since any upward wage pressure that occurs in one sector would be tempered by downward wage pressure in another. These tendencies for wages to rise or fall in response to the the shifting composition of labor demand would net out on average. Unemployment could be high as a result of sectoral shifts while aggregate demand is also high and wage inflation would result. This suggests that the Phillips curve as typically estimated may well be misspecified if the unemployment rate is not a good indicator of the performance of the economy overall. A more appropriate measure of labor market tightness would filter out the effects of factors such as sectoral shifts. A natural way to measure only cyclical fluctuations is to calculate the difference between the actual and natural rates of unemployment, a measure which by construction reflects only cyclical variations provided that the measure of sectoral shifts is independent of the cycle. The regression results reported in Table 3 include the age-weighted unemployment rate 13 and the natural rate of unemployment, N URn t, as separate explanatory variables with tests of the hypothesis that the estimated coefficients are of equal and opposite signs. Ifthe hypothesis is accepted, then it suggests that the difference between the actual rate and the constructed natural rate is a good indicator of general economic performance and supports the sectoral shifts hypothesis.13 Panel A reports regression estimates from the restricted data set. From column (1) it appears that the natural rate provides additional information independent of the level of unemployment that is relevant to the modeling of the wage inflation process. Increases in the natural rate relative to the actual signal a tightening oflabor market conditions. A one percentage point increase in the natural rate relative to the actual causes quarterly wage growth to rise by 0.3 percent. Similarly, given the natural rate, increases in the actual unemployment rate are associated with a weakening economy and, therefore, wage inflation islower. An F-test of equal and opposite signs on the actual and natural rates cannot reject the hypothesis, suggesting that the difference between actual and natural unemployment is a good measure of general economic conditions. Both the R 2 and Durbin-Watson statistic improve with this specification. Inclusion of N U R n t clearly adds to the performance of the estimating equation. The parameter instabilityofthe modified Phillips curve isexamined in the regression results found in columns (2) through (4). As in the preceding analysis, a dummy variable D is defined taking on the value of 1 for the period from the third quarter of 1971 through the end of the sample period, and 0 otherwise. From column (2),once compositional changes in employment are accounted for via the natural rate series, there is no evidence that the trade-off between inflation and unemployment worsened over the latter part of the sample given that the other parameters, namely the coefficient on inflationary expectations and the intercept, have been constrained to be constant throughout the sample period. Stability of the coefficient estimate on expected inflation, assuming constant parameters 13T h e errors in variables problem is more com plicated for this m odel than for the previous one. Both expected price inflation and the natural rate o f unem ploym ent are measured with error and OLS may either underestim ate or overestim ate the true param eters o f the m odel. 14 on the other variables, istested in column (3). The results suggest that over the earlier part of the sample, inflationary expectations was not an important determinant ofthe wage inflation process. In fact,over this period the parameter estimate on p e isinsignificantly different from zero. However, the latter part of the sample shows a significant positive effect of expected inflation on wage growth. Because inflationary expectations exhibit relatively little variation in the earlier part of the sample, the effect is picked up in large part by the constant term. In column (4) the coefficients on both the labor market tightness variables and the expected inflation variable are permitted to vary over the sample period. Due to multicollinearity, most of the coefficient estimates are insignificant at traditional confidence levels with the exception of the intercept, expected inflation, and the interaction term D x N URn t • The basic regression of column (1) was reestimated with the added constraint that the coefficients on the unemployment rate and the natural rate be of equal and opposite signs. The results are reported in column (5). In summary, once sectoral shifts have been included in the analysis, there is very little support for the claim that wage growth has become less cyclically sensitive. It would appear that much of the debate about the stagflation of the Seventies is in large part attributable to a failure to distinguish among the sources of the underlying disturbances to the economy. Regression results for the full sample are found in Panel B. Even after the inclusion ofthe natural rate of unemployment as an explanatory variable, the Phillips curve still appears to have an unstable intercept, ceteris paribus, as wage growth declined by between 0.6 and 0.8 percent post-’81.14 It has been argued that inflationary expectations were somehow different over the Eighties as monetary policy shifted its focus. However, even after inclusion of the inflationary expectations variable in columns (3) and (4), the intercept still seems to have changed significantly since ’81.15 14T he coefficient estim ates on W U R , N U R n t , and p e are similar to those found in the restricted sam ple. To sum m arize the results, inclusion of the natural rate as an explanatory variable seem s to improve the fit o f the equation. T h e coefficient on the unem ploym ent rate no longer rises significantly over the latter half of the sam ple. T h e expected inflation variable appears to differ p o st-’71. T he sam e caveats apply here as in the above discussion. ^ Inflationary ex p ectation s are calculated here from past inflation rates. If there were a fully anticipated regime change th at affected expected future inflation, it would only be incorporated in the expected inflation 15 The decline in wage growth observed over the Eighties has many potential explanations including the decline inunion strength, changes in “wage norms,”import penetration, deregu lation, and a change in the inflationary expectations mechanism. Union strength, as measured by the percentage of the labor force claiming union membership, has been declining continu ously since the Sixties. As an empirical matter, itisunlikely that unionism can explain only a relatively recent decline in wage growth. In addition, the decline in union membership is predominantly a phenomenon encountered in manufacturing industries and other traditional union strongholds. These industries have allbeen undergoing a long term decline in employ ment share. Thus, itisquite possible that the shiftin employment captured by the measure of sectoral redistribution and the decline in union membership are simply two different aspects of the same phenomenon. The relation between union wages and membership is difficult to disentangle. Rissman (1987) suggests that an optimizing monopolistic union facing deregulation and reductions in barriers to trade may initially increase wages and then permit wages to decline to their new long run equilibrium level as unions weigh short term wage gains against long term employment losses. In terms ofwage growth, import penetration results in higher short term wage growth immediately after a change in trade barriers with lower wage growth thereafter. This explanation ofdeclining wage growth focuses on the effect ofimports on predominantly unionized industries. However, itisunlikely that the broader-based phenomenon that appears to have occurred can be explained by industry-specific factors. This criticism isalso relevant to the deregulation explanation. The idea of a shifting wage norm is consistent with the data. However, the hypothesis offers little insight into wage determination and is more a description of the data than an explanation of it. Finally, the evidence supplied in Neumark and Leonard (1991) does not support the hypothesis that a change in the inflationary expectations mechanism was respon sible for the declining wage growth. In fact, they suggest that there was a true structural shift in the Phillips curve in the Eighties. In accord with their findings, there isno significant measure used here with some lag, thereby giving misleading results. 16 change in the coefficient on expected inflation over the Eighties and the intercept stillappears to have shifted downwards. The coefficient estimates of expected inflation and the actual and natural rates of unem ployment are surprisingly robust to alternative specifications. In addition to modifying the Phillips curve specification by concentrating on an alternative measure oflabor market tight ness, the role ofdemographics, unionization, and wage and price controls were also examined. Gordon (1977) and Perry (1980) suggest that the changing demographic composition of the labor force has had an effect upon wage growth. However, the proportion of females, non whites and youths in thelabor force were insignificant in the examined regressions, suggesting that demographics has an impact on wage growth only via its affect on unemployment. Gor don (1977) argues that wage and price controls were responsible for the parameter instability of the Seventies. A test of this hypothesis was performed by creating a dummy variable for when the controls were in effect. After controlling for sectoral unemployment, wage and price controls did not significantly affect wage growth, suggesting that the impact ofwage and price controls may be on the distribution of employment. Finally, union strength, measured as a percentage of employment, did not significantly affect wage growth. 4,3 T h e P hillips C urve and A ltern ativ e N a tu ra l R a te Series Regression results for an alternative natural rate series constructed from the stock price dispersion index N U R l r t are found in Table 4. These results are surprisingly similar to those based upon N U R n t • Coefficient estimates are generally of the same sign and magnitude as those found in Table 3 and their standard errors are similar. The natural rate enters positively in this set of regressions and is generally significant except when the interaction term D X N U R l r t is included. This occurs because this particular estimate of the natural rate varied little in the earlier part of the sample. F-tests of the strict sectoral shifts hypothesis, i.e. that the coefficients on the actual rate and the natural rate are equal and opposite in sign, are somewhat less strong than those based upon employment dispersion, A p . However, with a 5% critical value of approximately 17 3.96 the hypothesis cannot be rejected. 5 C onclusions Standard specifications of the Phillips curve employ the unemployment rate as a proxy for the general level of economic activity. However, unemployment can be high either because of sectoral realignment or because of low aggregate activity. Because wage pressures arising from sectoral shocks largely balance out, only that portion of unemployment that is net of the effects of compositional shifts in the structure oflabor demand should influence nominal wage growth. The worsening of the trade-off between inflation and unemployment over the Seventies is illusory. Much of the nominal wage growth over this period can be explained by sectoral shifts that caused the natural rate of unemployment to rise relative to the actual. Rather than having slack labor markets, as indicated by the high levelofunemployment, the opposite occurred with sectoral realignment attenuating the traditional linkage of unemployment to the generallevelofeconomic activity. Once thenatural rateisexplicitly considered, itappears that the cyclical sensitivity of wages has not changed over time. The results reported here are consistent with the sectoral shifts hypothesis of unem ployment. Sectoral shifts should not give rise to persistent inflationary pressures as wage movements are averaged out across sectors. Only increases in the general level of economic activity are inflationary. Ifsectoral shifts are truly independent of the cycle, then the differ ence between the actual rate of unemployment and the constructed natural rate series is the appropriate explanatory variable to use in a nominal wage growth regression. Tests of the hypothesis that the actual and natural rates ofunemployment enter with equal and opposite signs are generally supported by the data. The experience ofthe Eighties poses a somewhat different problem. The Phillips curve ap pears to be unstable over this period even aftersectoral shiftshave been explicitly considered. Although the natural rate stillenters the wage growth regressions with the hypothesized sign and is of the appropriate magnitude, there appears to be a downward shift in the intercept 18 term that is as yet unexplained. These results, specifically the stability of the trade-off, the shifting of the intercept, and the coefficient restrictions, are generally robust to markedly different measures of the natural rate. Finally, the sectoral shifts hypothesis and its relation to the wage inflation process has potentially important insights to yield for the role of government policy. First, and perhaps most obviously, the sectoralshiftshypothesis suggests that a countercyclical policy may be not be implementable because of the difficulty in interpreting movements in the unemployment rate. An expansionary policy adopted at the wrong time in the cycle may serve only to increase inflation rather than decrease the unemployment rate. Second, traditional monetary and fiscalpolicy effectiveness may depend in large part on the composition ofunemployment. Such policies seem better suited to dealing with unemployment that iscyclical in nature but may prove to be largely ineffective in reducing unemployment that is attributable to long term sectoral shifts. Therefore, a combination of monetary and fiscal stimuli combined with an industrial and jobs policy may be a more effective means of dealing with the problem of unemployment. Third, although the Phillips curve relations estimated here are reduced form expressions, they suggest that there may be a trade-off that policymakers can exploit between cyclical unemployment and the inflation rate. Before running monetary policy as if the relation is stable, a better understanding of the downward shift in wage growth in the Eighties is warranted. Finally, monetary policy may impact industries differentially, having a larger effect upon those industries that are relatively more interest-sensitive. Changes in policy may therefore affect the natural rate of unemployment and, hence, the nature of the trade-offbetween unemployment and inflation. 19 R eferences A braham , K atharine G. and Lawrence F. Katz (1986), “Cyclical Unemployment: Sectoral Shifts or Aggregate Disturbances?” Journal of Political E c o n o m y , 94, June, 507-522. Abraham , K atharine G. and James L. Medoff (1982), “Unemployment, Unsatisfied Demand for Labor, and Compensation Growth, 1956-80” in M artin N. Baily, (ed.), Workers, Jobs, a n d Inflation. Brookings Institution: W ashington, D.C., 49-88. Barro, Robert J. (1978), “Unanticipated Money, O utput, and the Price Level in the United States” Jou r n a l of Political E c o n o m y , 8 6 , August, 549-80. Barro, Robert J. (1977), “Long Term Contracting, Sticky Prices, and M onetary Policy” Journal of M o n e t a r y E c o n o m i c s , 3, July, 306-316. Brainard, S. Lael and David M. Cutler (1990), “Sectoral Shifts and Cyclical Unemployment Reconsidered” NBER Working Paper #3491, Cambridge: National Bureau of Economic Research. Cagan, Philip (1975), “Changes in the Recession Behavior of Wholesale Prices in the 1920’s and Post-World War II” Explorations in E c o n o m i c Resources , 2 , 54-104. Davis, Steven J. (1985), “Sectoral Shifts and the Dynamic Behavior of Unemployment: A Theoretical Analysis” University of Chicago: Working Paper No. 86-35. Gordon, Robert J. (1977), “Can the Inflation of the 1970’s Be Explained?” P a p e r s o n E c o n o m i c Activity, 1 , 253-277. Brookings Hamilton, James D. (1983), “Oil and the Macroeconomy since World War II” Political E c o n o m y , 91, April, 228-248. Lilien, David (1982), “Sectoral Shifts and Sectoral Unemployment” E c o n o m y , 90, August, 777-793. Journal of Journal of Political Lilien, David (1990), “Labor M arket Dispersion and the N atural Rate of Unemployment” University of California, Irvine. Loungani, Prakash (1985), “Sectoral Shifts and Business Cycles: A Fresh Look at the Evidence” University of Rochester. Loungani, Prakash, Mark Rush, and William Tave (1990), “Stock M arket Dispersion and Unemployment” Journal of M o n e t a r y E c o n o m i c s , 25, June, 367-388. 20 Neumann, George R. and Robert H. Topel (1991), “Employment Risk, Diversification, and Unemployment” Quarterly Journal of E c o n o m i c s , 106, November, 1341-1366. Neumark, David and Jonathan S. Leonard (1991), “Inflation Expectations and the Structural Shift in Aggregate Labor-Cost Determination in the 1980s” Working Paper. O’Brien, Anthony (1985), “The Cyclical Sensitivity of Wages” 75, December, 1124-1132. Perry, George L. (1980), “Inflation in Theory and Practice” Activity, 1 , 207-41. Rissman, Ellen R. (1987), Imp o r t Penetration a n d University: Unpublished doctoral thesis. American E c o n o m i c Review, Brookings P a p e r s o n E c o n o m i c Union W a g e D y n a m i c s . Northwestern Rogerson, Richard (1987), “An Equilibrium Model of Sectoral Reallocation” Political E c o n o m y , 15, May, 309-321. Journal of Sachs, Jeffrey (1980), “The Changing Cyclical Behavior of Wages and Prices: 1890-1976” A m e r i c a n E c o n o m i c R e v i e w , 70, March, 78-80. Schultze, Charles L. (1981), “Some Macro Foundations for Micro Theory” o n E c o n o m i c Activity, 2 , 521-592. Taylor, John B. (1980), “Aggregate Dynamics and Staggered C ontracts” Political E c o n o m y , 8 8 , February, 1-23. Brookings P a p e r s Journal of Wachter, Michael J. (1976), “The Changing Cyclical Responsiveness of Wage Inflation” Brookings Pap e r s o n E c o n o m i c Activity, 1 , 115-168. 21 22 Table 1: R elative E m ploym ent G row th R ates, Peak to Trough M anufacturing Industries Trough Food and Kindred Tobacco Textiles Apparel Lumber and Wood Furniture and Fixtures Paper Printing and Publishing Chemicals Petroleum Rubber and Plastic Leather Stone, Clay, and Glass Primary Metals Fabricated Metals Non-Electical Equipment Electical Equipment Transportation Equipment Instruments Miscellaneous 49Q4 54Q2 58Q2 61Q1 70Q4 75Q1 80Q3 82Q4 0.1870 0.7301 0.5858 -0.0084 0.7495 0.3643 0.3448 0.0854 0.8005 0.2425 0.9235 0.5777 0.9833 3.9298 1.5214 2.2828 1.3763 1.2793 1.1461 0.5261 0.1481 -0.0392 1.3147 0.7475 0.7356 1.1439 0.1369 -0.1044 0.4067 0.2026 1.3431 0.5432 0.7543 1.7375 1.3775 1.1431 1.7443 1.7906 1.1825 1.0771 0.2037 0.1913 0.8313 0.4685 0.9208 0.7816 0.1940 -0.0249 0.2831 0.3704 1.2619 0.6333 0.7304 2.2966 1.1944 1.8257 1.1522 2.1732 0.8194 0.5253 0.1395 0.7588 1.1760 0.6880 2.0548 1.6467 0.2241 -0.0837 0.2480 0.8369 1.3716 0.3554 1.2942 3.0539 1.4187 1.1531 0.4529 1.5374 0.7971 1.1513 0.1189 0.0512 0.5892 0.5779 0.4155 0.7176 0.5202 0.1656 0.3589 -0.0663 0.0000 1.1004 1.1528 1.2381 1.4564 1.0608 3.5602 3.6392 -3.8166 0.8592 0.4533 0.3845 2.0448 1.8173 2.5514 2.2149 0.8473 0.1732 0.2349 0.2537 1.4406 1.6489 1.2675 0.6968 1.1608 0.0864 1.5333 1.2951 0.6101 1.2008 0.1130 -0.2017 1.2078 0.5494 2.4544 1.9006 0.6003 0.1383 0.3076 -3.4369 2.0872 0.9319 1.8818 2.6628 1.6891 0.7189 0.9491 1.3834 0.2858 1.3883 0.1779 0.3661 1.1537 0.8533 0.9440 0.8869 0.5021 0.0110 0.4425 0.5965 0.8544 1.3616 1.2228 3.1209 1.5431 1.8902 0.6776 1.2114 0.5228 0.9879 23 Table 2 Param eter S tab ility o f th e Standard P hillips Curve Specification Panel A: 60Q2 to 81Q3 WUR ve D x WUR (1) (2 ) (3) (4) (1) (2 ) (3) (4) -0.0013 -0.0017 -0.0014 -0.0015 - 0.0010 -0.0014 - 0.0010 -0.0013 (3.519) (4.216) (3.281) (3.810) (3.225) (4.062) (2.949) (3.981) -0.0050 -0.0039 -0.0047 -0.0060 -0.0048 -0.0038 -0.0048 -0.0063 (10.065) (5.516) (4.612) (5.890) (10.795) (6.467) (5.150) (6.773) - 0.0005 - 0.0015 - 0.0005 - 0.0014 (2.247) D x pe c DUMMY Panel B: 60Q2 to 87Q1 - - (2.512) (3.619) -0.0003 0.0041 (0.331) (2.785) - - (4.282) 0.0000 0.0041 (0 .0 12 ) (3.375) 0.0145 0.0168 0.0150 0.0142 0.0134 0.0154 0.0133 0.0130 (8.235) (8.427) (6.617) (6.655) (8.915) (9.235) (7.158) (7.540) - - - - -0.0036 -0.0053 -0.0036 -0.0082 (3.010) (3.939) (2.996) (5.311) R? 0.569 0.594 0.570 0.630 0.636 0.657 0.636 0.692 D -W 1.505 1.693 1.516 1.883 1.443 1.614 1.443 1.812 NOTE: The dependent variable is the first difference in the logarithm of Average Hourly Earnings reported in various issues of E m p lo y m e n t the BLS. T-statistics are in parentheses. and E a rn in g s by 24 Table 3: T he P h illip s Curve and Sectoral Shifts Panel A: 60Q2 to 81Q3 WUR N U Rnt Pe x D x D x D WUR (5) (1 ) (2 ) (3) - 0.0020 - 0.0022 -0.0031 -0.0004 - 0.0020 -0.0016 -0.0023 -0.0019 - 0.0002 -0.0015 c (4) (4.403) (3.482) (4.944) (0.331) (4.366) (4.128) (4.029) (4.105) (0.204) (4.046) 0.0027 0.0027 0.0046 - 0.0011 0.0020 0.0022 0.0026 0.0027 -0.0028 0.0015 (2.517) (2.397) (3.534) (0.423) (4.366) (2.476) (2.639) (2.759) (1.189) (4.046) -0.0041 -0.0031 -0.0009 -0.0080 -0.0044 -0.0041 -0.0028 -0.0029 -0.0098 -0.0044 (6.758) (3.230) (0.606) (2.521) (10.966) (8.228) (3.851) (2.593) (3.365) (12.018) — -0.0009 -0.0017 — (0.652) NURnt pe (3) (5) (4) (1 ) (2 ) Panel B: 60Q2 to 87Q1 0.0014 - 0.0002 (1.160) (1.416) (0.207) 0.0014 0.0031 -0.0008 0.0020 (1.157) (1.946) (0.914) (1.463) — -0.0026 0.0055 0.0011 0.0076 2.429 (1.620) (1.208) (2.479) 0.0047 0.0055 0.0017 0.0132 0.0079 0.0049 0.0071 0.0043 0.0177 0.0080 ( 1 . 10 2 ) (1.342) (0.391) (2 . 1 1 2 ) (10.801) (1.316) (1.891) (1.149) (3.145) (11.922) -0.0057 -0.0080 -0.0062 -0.0073 -0.0047 (3.952) (4.699) (4.140) (4.292) (5.294) DUMMY R? 0.600 0.644 0.627 0.655 0.597 0.657 0.680 0.662 0.698 0.654 D -W 1.654 1.991 1.902 2.004 1.631 1.554 1.768 1.616 1.814 1.532 0.563 0.381 2.781 1.047 - 0.715 0.117 1.235 3.687 - F —value NOTE: The dependent variable is the log first difference in the BLS’s Average Hourly Earnings. T statistics are in parentheses. The F statistic reports tests for equal and opposite signs on the coefficients of W U R and the natural rate, NURn t • 25 T a b l e 4: T h e Phillips C u r v e a n d S e c t o r a l Shifts Panel B: 60Q2 to 87Q1 Panel A: 60Q2 to 81Q3 (1) lrt (4) (2) (3) (4) (5) (2.617) (4.208) (1.753) (4.529) (4.302) (3.137) (3.826) (1.964) (4.169) 0.0036 0.0018 0.0045 0.0011 0.0018 0.0031 0.0028 0.0040 0.0016 0.0014 (3.216) (0.984) (3.479) (0.600) (4.529) (3.127) (1.752) (3.604) (0.999) (4.169) x W U R (5.150) (4.378) (4.558) (5.624) (11.061) (6.268) (4.695) (5.402) (5.842) (12.093) - -0.0008 - -0.0004 - - 0.0005 - 0.0005 - xN U R lrt - 0.0010 (0.435) - 0.0017 (1.052) D (1) (4.728) (0.827) D (5) -0.0035 -0.0038 -0.0044 -0.0066 -0.0044 -0.0036 -0.0033 -0.0047 -0.0065 -0.0044 pe D (3) -0.0018 -0.0015 -0.0017 -0.0010 -0.0018 -0.0014 -0.0016 -0.0013 -0.0010 -0.0014 W U R N U R (2) x pe c DUMMY - - (0.643) - - -0.0003 (1.831) 0.0013 0.0048 (1.337) (3.328) (0.651) - (0.377) - - - 0.0007 - (0.928) 0.0015 0.0046 (1.737) (3.604) - -0.0023 0.0056 -0.0081 0.0052 0.0073 - 0.0011 0.0013 -0.0075 0.0028 0.0076 (0.420) (0.646) (1.159) (0.637) (9.458) (0.230) (0.175) (1.243) (0.392) (10.761) - - - - - -0.0053 -0.0060 -0.0059 -0.0081 -0.0039 (4.204) (3.728) (4.542) (4.958) (3.906) R? 0.617 0.625 0.626 0.671 0.603 0.668 0.741 0.677 0.709 0.657 D - W 1.702 1.752 1.703 2.017 1.629 1.596 1.645 1.610 1.860 1.532 3.096 0.052 4.907 0.0042 - 3.293 0.741 6.207 0.167 - F —value NOTE: The dependent variable is the first difference in the logarithm of the BLS’s Average Hourly Earnings. T-statistics are in parentheses. The F statistic reports tests for equal and opposite signs of the parameters on W U R and the natural rate, N U R l r t •