12
We are grateful to a seminar participant at the Bank of Canada for suggesting this approach.
13
By sorting our sample on the basis of long lags of the endogenous variable we considerably reduce concern about
sample selection on the basis of an endogenous variable.
26
contrast, our model allows the possibility that the coefficient on expected inflation will be lower
in extended periods of low inflation than in extended periods of high inflation. Absent estimation
biases, we would expect the coefficient to approach 1.0 in a sufficiently inflationary environment.
We first look at the empirical evidence using the conventional adaptive expectations framework.
We then provide evidence using direct measures of inflationary expectations that address
Sargent’s [1971] criticism of the assumption that the coefficient on lagged inflation must equal
one in an accelerationist model. Sargent argued that a coefficient of less than one on lagged
inflation may not reflect incomplete projection of inflation but rather forecasters' views that the
process generating inflation does not have a unit root. By using direct measures of inflationary
expectations we can rule out the possibility that our results reflect differences in how people form
expectations rather than how they use them.
12
In order to separately estimate wage and price Phillips curves for periods of low and high
inflation, we sorted the quarters since the Korean War according to the average CPI inflation rate
in the five-year period ending each quarter. We first classified quarters with average inflation
rates below 3 percent as low inflation and quarters with average inflation rates above 4 percent as
high inflation.
13
By this sorting, the low inflation quarters run from 1954:1 through 1969:1 and
from 1995:3 through 1999:4, the end of our sample period. The high inflation quarters run from
1970:2 through 1986:1 and from 1990:4 through 1993:2. There are 77 quarters in the high
inflation sample and 77 quarters in the low inflation sample. The mean CPI inflation rates in the
two samples are 2.0 percent and 6.3 percent. This separation was used in half the wage and price
27
inflation regressions. In the other half we limited the low inflation sample to quarters with
inflation rates below 2.5 percent, which brought the sample size down to 62 quarters and reduced
the mean CPI inflation rate in the low inflation sample to 1.9 percent.
Estimates with Adaptive Expectations
The quarterly Phillips curve equations we estimated were intended to span the
specifications that analysts have used in conventional estimation of NAIRU models except for the
fact that we did not constrain the coefficients on lagged inflation. To this end, we tried a large
number of data combinations and specifications on both wage and price Phillips curves, and ran
each separately for the low and high inflation samples just described. In all cases the dependent
variable was an annualized inflation rate in either wages or prices, and the explanatory variables
were current or lagged values of unemployment, price inflation and, for the wage equations, trend
productivity growth. For price inflation we used the CPI, the GDP deflator and the PCE deflator
and estimated price Phillips curves with each. Twelve values of lagged inflation were used as
explanatory variables. For wage inflation we used the best series available for any time period,
linking private ECI wages and salaries for 1980-1999 to the adjusted hourly earnings index for the
nonfarm economy for 1961-1980 and to adjusted hourly earnings in manufacturing for 1954-
1961. Twelve lagged values of CPI inflation were used as explanatory variables. For
unemployment we used the total rate, the 25-to-54 year old male rate, and Robert Shimer’s
demographically adjusted series. We used the current and three lagged values of unemployment
and, alternatively, the current and eleven lagged values. For the wage Phillips curves, we used
two estimates of trend productivity growth, one being the series created by Robert Gordon and the
14
All equations also used the customary dummy variables for the guidepost period in the 1960s and the price control period of
the 1970s, and used the difference between inflation with and without oil prices in 1979-1980 as an additional variable.
28
other a smoothed version of that series. We ran regressions with the productivity coefficient both
freely estimated and constrained to be 1.0 (for the wage inflation equations), and with just the
current trend and with the current plus seven lagged values of the trend.
14
Figures 3 and 4 about here
The key results are summarized in figure 3 for equations explaining wages and in figure 4
for equations explaining prices. The figures present the results of 144 and 72 specifications
respectively. Each point represents the sum of the coefficients on lagged inflation estimated for
the low and high inflation samples for one specification. If the sum of coefficients were similar
for the two samples, the points would cluster along the forty-five degree line. If they were similar
and near 1.0, the points would cluster near the upper right corner. In fact, for both wages and
prices, and over the wide range of specifications and data we used, the points cluster near 1.0 on
the high inflation axis, but on the low-inflation axis, they range from around zero to around 0.5
for the wage equations. This is consistent with the predictions of our model. The range on the
price equations is broader and less conclusive. The third of the observations at the highest end of
the range are from equations using the PCE deflator. The mean values of the coefficients on the
high and low inflation axes respectively are 0.25 and 0.82 for the wage equations and 0.60 and
0.95 for the price equations.
Direct Measures of Inflationary Expectations
Dostları ilə paylaş: |