15
This specification ignores the parameter “a” from the theoretical model. In theory that
parameter could be estimated, but we do not take the theoretical model that literally. Instead we
imagine that there is a continuum of reactions to increasing inflation with people putting more
and more weight on it until their behavior resembles that of the rational economic actor in the
standard model. The model we estimate here can be thought of as a model where a fraction (1-
0
)
of people are ignoring inflation, or the phi function can be thought of as approximating a more
general function that reflects how much weight the average person is putting on inflation in
making economic decisions.
32
(17)
(18)
thresholds, the third order and higher terms are unimportant. An approximation of the loss
function of the form E
%
2
, where E was chosen so that the approximation was exactly equal to the
loss at 5% inflation, was never off by more than 3% of the loss. One parameter is all that is
necessary to capture the effects of all three parameters from the model (
,
, and
)
) on the
derivative of the argument of the cumulative normal with respect to inflation.
We thus estimate a Phillips curve of the form:
where
%
is the rate of inflation,
0
is the cumulative standard
normal density function,
%
e
is
inflationary expectations, u is a term capturing the effects of current and lagged unemployment on
inflation,
X is a matrix of dummy variables for oil shocks and price controls,
is the error term,
and
d, D, E, e and
g are parameters to be estimated.
15
The term
%
L
represents the effects of past inflation on the likelihood that people will act
rationally towards inflation. Our theory tells us nothing about the way in which inflation should
matter other than the sign of E, so we proxy
%
L
with several different parsimonious specifications.
The first is a geometrically declining weighted moving average of past values of inflation:
33
(19)
where
is a parameter to be estimated.
Alternatively we estimate
%
L
as
where the parameter
is estimated. Our final two specifications for
%
L
treat it as a 4-year moving
average of past inflation with equal weights, or with the relative weights of quarters from each
year are estimated (three additional parameters).
It is standard practice to proxy inflationary expectations with lagged values of inflation in
Phillips Curve estimation. In many specifications discussed below we follow that tradition. When
we do, we use either a 12 quarter unrestricted lag or one of the methods used to construct
%
L
to
construct
%
e
. However, we also want to rule out the possibility that changes in the coefficient on
%
e
might reflect changes in the process by which expectations are formed rather than how they are
used. Thus we also use direct survey measures of inflationary expectations for
%
e
in some
specifications.
Our different specifications include several different measures of unemployment and also
different numbers of lags. The unemployment term, u, is constructed using one of three data
series. The first is the aggregate U.S. unemployment rate from the Current Population Survey.
Because this variable may be influenced by changing demographics, we have also considered two
16
The inclusion of the term for nominal rigidity could be motivated if we included firm
profitability or firm specific labor market considerations into the productivity function. That
would produce heterogeneity in desired wage setting with firms constrained by the floor of no
nominal wage decrease forced to pay a higher wage as in the model in our previous paper.
17
We leave out the term for change in profits, which could not be robustly estimated.
34
alternative measures: the unemployment rate for prime age males and Shimer’s demographically
corrected series. (See Shimer, 1998). We also vary the number of unemployment lags from zero
to 11 quarters.
For the dependent variable we variously use four different measures of inflation: the
annualized percent change in the consumer price index (CPI-UXG), the gross domestic product
deflator, the personal consumption expenditures deflator, and the index of wage and salary
compensation constructed by Brainard and Perry (2000). When we use the percent change in the
compensation index as the dependent variable we subtract off a measure of trend productivity
growth. The three specifications of this trend are: A measure based on Gordon (1998), the
measure we constructed for our 1996 paper, and a 16-quarter moving average.
Since the form of the Phillips curve here is similar in some respects to the one in our
previous paper (Akerlof, Dickens and Perry (1996)) that modelled the implications of downward
nominal wage rigidity, we also examine the question of whether we can succesfully estimate a
Phillips curve which embodies the insights from that model as well as the current one. Below we
estimate a number of specifications that augment equation (17) with the term for nominal rigidity
from that previous paper.
16
When we nest that model we must also estimate its key
parameter—the standard deviation of desired wage changes along with the other parameters from
the current model. (See Akerlof, Dickens and Perry (1996, Appendix A) for its specification.)
17
The model was estimated with quarterly US data from the first quarter of 1954 through the