29
As in Brainard and Perry (2000) the results just described cast doubt on conventional
estimates with the NAIRU model. However they both treat expectations as adaptive and so
cannot refute Sargent's (1971) criticism that rational expectations are formed differently and that
the coefficient on properly measured expectations might be 1.0. We now address this issue by
using direct measures of expected inflation as explanatory variables in place of distributed lags of
actual inflation rates, while maintaining our division of the sample into periods of high and low
inflation. The other explanatory variables are the same as those used in the regressions behind
figures 3 and 4. We used the two direct measures of expected rates of inflation that are available
over our sample period: one from the Survey of Consumer Finances and the other the Federal
Reserve’s Livingston Surveys. Figures 5 and 6 plot the estimated coefficients on expected
inflation for the variously specified wage and price regressions respectively. As with the results
using adaptive expectations, the coefficients on expected inflation are substantially different in the
low- and high-inflation periods. For 288 wage equations the low- and high-period means are 0.29
and 0.85. For 144 price equations the means are 0.25 and 1.00.
Figures 5 and 6 about here
These results support our general hypothesis even more convincingly than the results with
adaptive expectations. Not only do they address the point that the relevant coefficient for natural
rate theory is not necessarily the coefficient estimated with adaptive expectations, but the results
are as clear about price inflation as they are about wage inflation.
One possible objection to the results presented here and in the next section is that the
30
lower coefficients on inflationary expectations during periods of low inflation are an artifact of
measurement error. For example, if the variance of measurement error is constant while the
variance of true inflationary expectations are higher in times of high inflation, then the coefficient
on expectations could be biased towards zero more in times of low inflation than high inflation.
We investigated this possibility. While it is true that the variance of expectations is higher in
periods of high inflation, it is also true that the sampling error in both the SCF and the Livingston
surveys are also higher. In fact, the sampling error is so much higher that the computed bias is
higher in the low inflation periods imparting a bias
against our finding that the coefficient on
expectations is lower in periods of low inflation. Sampling error may not be the only source of
error in the survey expectations. Neither survey may be asking the right people with the right
weights. In an attempt to approximate how much error this problem might introduce we computed
the bias that would be caused if the measurement error variance in expectations was equal to the
variance of the residual of a regression of one of our survey expectations on the other. Again we
found that the “measurement error” variance grew faster than the conditional variance of the
expectations so that the bias caused would work against our finding that the coefficient on
expectations was lower when inflation was low.
Estimating The Model
Previously we showed how a Phillips Curve type relation can be derived from our
theoretical model (equation 15). In this section we present estimates of the model and of the
optimal rate of inflation and the gain in employment that is possible from moving to the optimal
rate. This section will first discuss the specification of the model we estimate, then our benchmark
31
results, and finally, an analysis of their robustness.
Specifications
In theory, with a large enough sample, it would be possible to estimate the full model
presented above. The elasticity of demand (
), the parameter for the curvature of the unit cost
function (
), and the parameters of the distribution of rationality thresholds (
and
)
), all have
different effects on the objective function. However, in practice, it was impossible to estimate
more than the mean of the distribution of rationality thresholds, and one of the other parameters
because all three of them— the elasticity of demand, the curvature of the unit cost function and
the standard deviation of the distribution of rationality thresholds—act in much the same way to
determine the impact of past rates of inflation on the cumulative normal term. (See equation (15)
above).
The lack of identification in practice can be understood if we
consider a Taylor series
approximation to the argument of the cumulative normal in equation (15) expanded around a
value of zero inflation. There is no reason to expect that the argument will be exactly zero at zero
inflation so the constant term will likely be present. As we have shown above, the first derivative
of the firm’s loss function with respect to inflation is zero at zero inflation and very small at most
rates of inflation less than 10 percent. Thus the first order term of the Taylor series expansion of
the argument of the cumulative normal will also be zero. Second and higher order terms will be
present, but analysis we have conducted of the loss function suggests that with inflation between
zero and ten percent, with elasticity of demand between 2 and 10, with curvature of the unit cost
function from .05 to .95, and any value of the standard deviation of the distribution of rationality