Federal Reserve Bank of New York
Staff Reports
Do Expected Future Marginal Costs Drive Inflation Dynamics?
Argia M. Sbordone
Staff Report no. 204
March 2005
This paper presents preliminary findings and is being distributed to economists
and other interested readers solely to stimulate discussion and elicit comments.
The views expressed in the paper are those of the author and are not necessarily
reflective of views at the Federal Reserve Bank of New York or the Federal
Reserve System. Any errors or omissions are the responsibility of the author.
Do Expected Future Marginal Costs Drive Inflation Dynamics?
Argia M. Sbordone
Federal Reserve Bank of New York Staff Reports, no. 204
March 2005
JEL classification: E31, E32
Abstract
This article discusses a more general interpretation of the two-step minimum distance
estimation procedure proposed in earlier work by Sbordone. The estimator is again applied
to a version of the New Keynesian Phillips curve, in which inflation dynamics are driven by
the expected evolution of marginal costs. The article clarifies econometric issues, addresses
concerns about uncertainty and model misspecification raised in recent studies, and assesses
the robustness of previous results. While confirming the importance of forward-looking
terms in accounting for inflation dynamics, it suggests how the methodology can be applied
to extend the analysis of inflation to a multivariate setting.
Key words: inflation, New Keynesian pricing, marginal costs
Sbordone: Federal Reserve Bank of New York (e-mail: argia.sbordone@ny.frb.org). The views
expressed here are those of the author, and do not necessarily reflect the position of the Federal
Reserve Bank of New York or the Federal Reserve System.
1. Introduction
The standard pricing assumption in real business cycle models implies a constant markup of prices
over marginal cost, and hence an inflation rate equal to the rate of growth of average nominal
marginal cost. These predictions are at odd with the data: in particular, U.S. inflation is less
volatile than marginal costs. However, by introducing nominal price rigidities it is possible to
explain cyclical markup variations, and hence to generate an inflation path whose volatility is like
that observed in the data.
The widely used Calv o model of staggered pricing (Calvo 1983) implies an equilibrium pricing
condition that, in log-linearize d form, links current inflation to expected future inflation and current
real marginal cost
1
π
t
= βE
t
π
t+1
+ ζs
t
+ η
t
(1.1)
Here s
t
is the (log of) av erage real margina l cost in the economy, the parameter β is a discount
factor, and ζ is a nonlinear function of the relevant structural parameters: ζ =
(1−α)(1−αβ)
α(1+θω)
.θis
the elasticity of substitution among different iated goods, ω is the elasticity of firms’ marginal costs
to their own output
2
,andα is the percentag e of prices that are not reset optimally at time t.
The degree of price inertia is measured by 1/1 − α.
3
The error term η
t
is included to account
for fluctuations in the desired mark-up, or for other forms of misspecification of the equation;
4
1
A detailed derivation of this equation can be found in Woodford (2002), ch. 3.
2
The presence of this term is due to the further assumption of firm-specific capital. This term alters the m apping
b etween the param eter ζ and the frequency of price adjustment, a s discussed in Sbordone (2002), making a low
estimate of ζ consistent with a reasonable degrees of price stickiness.
3
The variables are expressed in log deviation from steady state values. If the log-linearization is around a zero
steady state inflation, the log deviation of inflation can be measured by its actual value. Under the assumption that
real wage and productivity share the same long run trend, the log deviation of the labor share can also b e measured
by its actual value. In th e data, we will see below that stationarity may require a slight transformation of the share.
4
This was s uggested by Rotemberg and Woodford (1999). In Steinsson (2002) the error represents exogenous
variations in the elasticity of substitution; in Giannoni (2000) it represents time varying tax distorsions.
1
throughout this article it is assumed to be a mean zero, serially uncorrelated stochastic process.
5
This model has been generalized in a number of ways to be able to generate additional inflation
inertia. Here I follow Christiano et al. (2005) by assuming that firms that are not selected to reset
prices through the Calvo random drawing are nonetheless allowed to index their current price to
past inflation, and I assume that they do so by some fraction [0, 1]. The solution of the model
in this case
6
is
π
t
− π
t−1
= β (E
t
π
t+1
− π
t
)+ζs
t
+ η
t
, (1.2)
which nests eq.(1.1) (the case of =0), and, in the opposite case of full indexation ( =1), as
considered in Christiano et al. (2005), implies an expectational equation in the rate of growth of
inflation. This generalized equation has the same form as the ‘hybrid model’ of Gali and Gertler
(1999), when rewritten as
π
t
=
1+β
π
t−1
+
β
1+β
E
t
π
t+1
+
ζ
1+β
s
t
+ eη
t
, (1.3)
or
π
t
= γ
b
π
t−1
+ γ
f
E
t
π
t+1
+ ζs
t
+ eη
t
. (1.4)
In this expression γ
b
and γ
f
can be interpreted as the weights, respectively, on ‘backward-’ and
‘forward-looking’ components of inflation. Iterating forward, eq. (1.2) gives a present value rela-
tionship, where inflation is a function of lagged inflation and expec ted future real marginal costs:
π
t
= π
t−1
+ ζ
∞
X
j=0
β
j
E
t
s
t+j
+ ν
t
. (1.5)
The empirical evaluation of eq (1.1), or its generalized form (1.2), known as the New Keynesian
Phillips curve (NKPC), has generated a great deal of debate, as the papers presented in this volume
5
In my (2002) paper, I examined the degre e to which the data could be fit by a model with no error term. Here,
instead, an explicit hypothesis about the nature of the error term allows to address various issues such as a possible
simultaneous-equations bias.
6
A detailed derivation of this expression can b e found in Woodford (2003), ch.3.
2
testify. Gali and Gertler (1999) pioneered an approach to estimation based on the Euler equation
(1.4), which raised a lot of discussion about the appropriateness of the use of GMM estimation.
Gali, Gertler and Lopez-Salido respond in this issue to most of the criticisms to their approach.
Sbordone (2002) proposed an alternative two-step procedure based on the empirical evaluation
of the closed form solution (1.5) in its restricted form ( =0), and in this paper I wish to clarify
this methodology, assess the robustness of my previous results, and evaluate some of the criticisms
raised in the other articles in this volume.
2. Estim ating the closed-form solution
My (2002) paper proposed to estimate the basic NKPC specification, eq. (1.1), by matching actual
inflation dynamics to the inflation path predicted by the Calvo model, taking as given the dynamics
of nominal marginal cost, denoted here as mc
t
. I assumed that the model held exactly (η
t
=0),
and solved the model forward to obtain a predicted path of prices as function of expected future
nominal marginal cost:
p
t
= λ
1
p
t−1
+(1− λ
1
)
¡
1 − λ
−1
2
¢
∞
X
j=0
λ
−j
2
E
t
mc
t+j
. (2.1)
The parameters λ
1
and λ
2
in (2.1), roots of the characteristic equation P(λ)=βλ
2
−(1+β+ζ)λ+1 =
0, are non linear combinations of the structural parameters β and ζ. The proposa l was to evaluate
this pricing model along the lines of Campbell and Shiller’s (1987) evaluation of present value
relationships in finance.
For such evaluation I first assumed that appropriate conditions held to guarantee proportion-
ality of average and marginal costs: the unobservable marginal cost could then appropriately be
proxied by a measure of unit labor cost. Furthermore, to express the relationship in terms of sta-
tionary variables, I transformed (2.1) into a relation between the price/unit labor cost ratio and the
rate of growth of unit labor costs (respectively p
t
− u
t
and ∆u
t
, with variables expressed in logs).
3
The empirical evaluation was in two steps: first the estimation of an unrestricted vector autore-
gression model to forecast unit labor costs; then, taking as given this forecast, the estim ation of th e
parameters of the structural model by minimizing the distance between the path of the price/ulc
ratio implied by the model and the actual dynamic path of the data (given the path of unit labor
cost, a predicted path for the price/ulc ratio then implies a path for inflation as well).
More specifically, assuming that all information at time t about current and future values of
therateofgrowthofunitlaborcost∆u
t
could be summarized b y a vector of variables Z
t
, where
{Z
t
} is a stationary Markov process,
Z
t
= AZ
t−1
+ Gε
zt
, (2.2)
and E
t−1
(ε
zt+j
)=0, for all j ≥ 0,
7
the infinite sum of future expected unit labor cost could be
computed as
8
∞
X
j=0
λ
−j
2
E
t
∆u
t+j
=
∞
X
j=0
e
0
u
λ
−j
2
A
j
Z
t
= e
0
u
(I − λ
−j
2
A)
−1
Z
t
.
Letting B =(I − λ
−j
2
A)
−1
, and noting that the price/unit labor cost ratio is the inverse of the
labor share (so that p
t
− u
t
= −s
t
), the solution (2.1) could be written as
s
t
= λ
1
s
t−1
+ ∆u
t
− (1 − λ
1
)e
0
u
BZ
t
− η
t
. (2.3)
Denoting by s
m
t
(ψ, A) the path of the labor share predicted by the model, and by
s
t
= s
t
−s
m
t
(ψ, A)
the distan ce between actual and predicted paths, under the null that the model is true we have
that
E(
s
t
)=E(s
t
− s
m
t
(ψ
0
,A
0
)) = 0, (2.4)
where ψ
0
,A
0
denote true parameter values. With this notation, the proposed two-step estimator
involved first the estimation of the system (2.2), and then the estimation of the vector ψ =(βζ)
0
7
Note that any autoregressive process of order k can be expressed in this form through a suitable definition of the
vector Z
t
and the matrices A and G.
8
The vector e
x
denotes a selection vector for variable x (a unit vector with 1 in the position corresponding to x
and zero otherwise ).
4
by
b
ψ
1
=argmin var(b
s
t
) (2.5)
where b
s
t
= s
t
− s
m
t
(ψ,
b
A), and
b
A is a consistent estimator of the elements of A. The theoretical (or
fundamental) inflation rate was then derived as
π
m
t
= −(s
m
t
(
b
ψ
1
,
b
A) − s
m
t−1
(
b
ψ
1
,
b
A)) + ∆u
t
.
This approach to estimating inflation dynamics provided, I believe, an approach to the empirical
assessment of Phillips curve relationships which was novel in two respects. First, it focused on the
relationship between the dynamics of prices and the dynamics of marginal costs, as opposed to the
relationships between inflation and output gap. This choice was motivated by the observation that
the Calvo model of optimizing firms with staggered prices makes predictions only about the dynamic
relation between prices and marginal cost. In order to get an empirical Phillips curve specification
in terms of output gap one needs further theoretical assumptions, both about how marginal costs
are related to output, and about how to construct a theory-based measure of potential output.
9
The choice of marginal cost as forcing variable was at the same time independently made b y Gali
and Gertler (1999), who similarly proxied marginal costs with unit labor costs.
The second novelty was in the estimation procedure. The paper focused on the estimation of
the present value relationship between prices and marginal costs implied by the optimizing model,
and applied a two-step estimation procedure. As described above, following Campbell and Shiller’s
tests of the present value theory of stock price determination, the first step involved estimating
an auxiliary forecasting model to generate predictions of the future values of the forcing variable
− the growth of nominal marginal costs in my application. The second step involv ed estimating
the parameters of the structural model, conditional on the forecasting model estimated in the first
9
One can obviously simply interpret marginal costs as a particular measure of output gap: given the uncertainty
in the estimation of output gap, and the difficulties of constructing a truly theory-based measure (this is attempted,
however, by Nelson and Neiss, 2001) this is a convenient measure.
5
step, using a distance estimator.
The results w ere quite striking. For eq. (1.1), calibrating the discount parameter β, the estimate
of the coefficien t ζ was positive and statistically significant. Its value was consistent with price
rigidity lasting 3 to 4 quarters, in line with survey-based evidence. Moreover, the dynamics of
predicted inflation were very close to the actual inflation dynamics, and the model allowed to
reproduce quite closely the serial correlation of the data.
These results depended of course upon the correct specification both of the structural mode l and
of the auxiliary forecasting system. The (2002) paper considered only the purely forward-looking
specification of the structural model (the case of =0), but checked the sensitivity of the results
to two modifications of the measure of marginal cost, the presence of adjustment costs for labor,
and the case of a technology with overhead labor. Both modifications altered the specification of
the structural model by adding other forcing variables in the inflation equation.
To address the second problem, I considered alternative forecasting systems for unit labor costs,
both excluding the price/unit labor cost from the system, and including additional variables; in
all cases the qualitative results of the model remained valid. Finally, I showed that the ability of
the model to track inflation dynamics was worsened when excluding the forward looking terms in
(2.1), and concluded that this component appears to be important for explaining the dynamics of
prices.
10
Both specification issues receive a more critical assessment in the contribut ions of this volume,
and in this article I take the opportunity to comment further on them. While I leave the issues
related to the robustness of GMM estimates to the reappraisal by Gali, Gertler and Lopez-Salido
(2005), I try instead to offer some perspective on the issue of uncertainty raised by Kurmann
(2005), which hinges upon the specification of the auxiliary forecasting model; on the issue of
10
My conclusions were in line with those reached by Gali and Gertler, who estimated the mo de l parameters on a
specification of the type (1.4), which explicitely includes backward and forward terms inflation terms.
6
whether the forward-looking component in the inflation dynamics is insignificant, as claimed by
Rudd and Whelan (2005); and on the issue of what should be the appropriate proxy for marginal
cost, as discussed by Batini et al. (2005).
3. Appraisal of the criticism s
As the summary of my approach in the previous section shows, the analysis in m y 2002 paper
treated nominal marginal costs as the forcing variable in a model of price dynamics. Most of the
subsequent literature on the NKPC, and all the papers in this issue of the JME, take instead real
marginal cost as the forcing variable, and estimate directly an inflation equation of the form (1.1)
or (1.5). Since the choice of the most appropriate forcing variable is not the focus of this article,
I will conduct my discussion in terms of this latter version of the model. The application of the
methodology discussed above to this case involves defining a distance function directly in terms of
inflation paths
π
t
= π
t
− π
m
t
(ψ), a nd its minimization is conditional on a forecasting model for real
marginal cost. In what follows the forecasting model has the same form (2.2), and the vector Z
t
includes at least inflation and a measure of real marginal cost (the labor share).
3.1. Kurmann’s critique
Kurmann’s (2005) paper analyzes the robustness of fit of the New Keynesian model to alternative
specifications of the forecasting VAR , and alternative values for the we ight of the forward-looking
component. His criticism is directed to the fit of what Gali and Gertler call ‘fundamental inflation’,
which is the inflation path derived from the solution of the Calvo model. To construct this path
they use the coefficients estimated for the Euler equation (1.4) by a standard GMM procedure, and
the forecast of real marginal costs implied by a separately estimated vector autoregression model.
Kurmann constructs the path of inflation from the restricted form of (1.5),
π
t
= ζ
∞
X
j=0
β
j
E
t
s
t+j
, (3.1)
7
setting ζ = .035 (a ‘benchm a rk’ value among those estimated by Gali and Gertler 1999), and β =1.
He constructs the forecasts s
t+j
from the VAR model estimated by Gali and Gertler, a bivariate
model with four lags of π
t
and s
t
. The goodness of fit is measured by the relative standard deviation
of predicted vs. actual inflation,
σ
π
m
σ
π
, and the correlation between predicted and actual inflation
ρ(π
m
,π).
The criticism raised by Kurmann is that while the point estimates of these statistics indicate an
impressive fit to both the dynamic path and the volatility of inflation, there is a lot of uncertainty
surrounding them.
First, he argues that, even assuming that the forecasting model is correct, in the sense of
containing all the variables that help to forecast the expected future value of the labor share, the
model is merely estimated, and treating the estimated parameters as true population values leads to
underestimate the uncertainty surrounding the estimated inflation statistics. Specifically, he shows
that the confidence interval around the two estimated statistics is quite large, due to the uncertainty
of the estimated VAR coefficients. Second, he shows that the point estimates themselves are highly
sensitive to the specification of the forecasting model that one chooses. Finally, he discusses the
sensitivity of the predicted inflation dynamics to the degree of price stickiness implied by the
assumed value of the coefficient ζ.
Kurmann’s paper aims at sho wing that the evidence provided by graphs of the fundamental
inflation and by point estimates of standard deviation and correlation statistics is misleading,
because it hides the uncertain ty of the estimated forecasting model that is used to construct the
predicted path of inflation.
The question of the uncertainty in the VARestimation is particularly relevant to the estimation
method discussed above, since it uses the auxiliary VARnot just for the construction of the model’s
predicted path, but also as a crucial step of the estimation procedure. Next section addresses
therefore the issue of uncertainty b y revisiting the two-step estimation procedure in a way that
8
shows how to take into account the imprecision of the firststepestimation. Atthesametimeit
tries to clarify the relation between this approach and the instrumental variables approach used by
Gali and Gertler (1999) to estimate the Euler equation of the Calvo model.
3.2.Thedistanceestimator reinterpreted
As noted above, my proposed two-step distance estimator was based on Campbell and Shiller’s pro-
cedure. This analogy can perhaps be better illustrated by giving a slightly different interpretation
to the distance
π
t
, namely by viewing it as a measure of the restrictions imposed by the structural
model on the parameters of the forecasting process.
11
For this interpretation one should observe
that, by definition, the vector of forecasting variables Z
t
includes current inflation and the labor
share, so that we can write, with an appropriate definition of selection vectors e
π
and e
s
,
π
t
= e
0
π
Z
t
, and s
t
= e
0
s
Z
t
. (3.2)
Then, using (2.2), the infinite sum of expected future values of the labor share that appears in the
solution (3.1) is computed as
∞
X
j=0
β
j
E
t
s
t+j
= e
0
s
(I − βA)
−1
Z
t
,
so that the solution (3.1) can be written as
e
0
π
Z
t
= ζe
0
s
(I − βA)
−1
Z
t
.
Under the null that the model is a good representation of the data, this equality must hold for
every Z
t
; hence it must be true that
e
0
π
− ζe
0
s
(I − βA)
−1
=0. (3.3)
This expression defines a (1×2p) vector of restrictions o n the elements of the matrix A that charac-
terizes the process (2.2). These cross-equation restrictions betw een the parameters of the structural
11
The estimator in this form is ap plied to a two-variable model of price and wage dynamics in Sbordone 2003.
9
model and those of the driving process Z
t
represent, in the words of Hansen and Sargent (1979),
the ‘hallmark’ of rational expectations models. For the present value model of stock prices, with
no free parameters, Campbell and Shiller proposed a Wald test of these restrictions. The distance
estimator that I proposed can be interpre ted as an ‘unweighted’ measure of these restrictions. De-
noting the restrictions (3.3) as a vector function z(ψ,A), for a consistent estimate
b
A of the matrix
A, it is the case that
z(ψ,
b
A)
0
≡ e
0
π
− ζe
0
s
³
I − β
b
A
´
−1
(3.4)
converges to 0, and the proposed estimator
b
ψ is the vector that minimizes the square of the function
z(ψ,
b
A). Under this interpretation, one may also modify the proposed estimator to minimize a
‘weighted’ function of the restrictions, by giving a higher weight in the objective function, for
example, to those elements of
b
A which are estimated more precisely. This can be done by weighting
the quadratic function with the covariance matrix of the restricted parameters
b
A. The weighted
estimator is then defined as
b
ψ
2
=argmin
h
z(ψ,
b
A)
0
Σ
−1
A
z(ψ,
b
A)
i
(3.5)
where Σ
A
is a matrix with appropriately selected elements of the estimated variance-covariance
matrix of
b
A.
To summarize, my proposed approach to estimate the present value form of the Calvo model of
inflation dynamics is a t wo-step distance estimator that exploits an ‘auxiliary’ autoregressive repre-
sentation of the data. The estimator may take two forms. In (2.5) the objective function to minimize
is the variance of the distance between model and data, which is an unweighted quadratic form
of this distance, while in (3.5), the objective function is similarly a (possibly weighted) quadratic
form of a distance function representing the restrictions that the model solution imposes on the
parameters of the auxiliary VAR.
The first interpretation emphasizes the role of the auxiliary VAR processasaforecasting
10
process from which to compute the expected future values of the forcing variables. In the second
interpretation, the VAR provides an unrestricted representation of the data, against which to
compare the restrictions imposed by the structural model.
The analogy has thus far been illustrated for the case in which the Calv o model holds exactly.
More generally, when the inflation equation includes an error term, as in specification (1.1) and,
further, when it also includes a term in lagged inflation, as in specification (1.2), the model solution
is
π
m
t
= π
t−1
+ ζe
0
s
(I − βA)
−1
Z
t
+ βη
t
.
In this more general case the vector of structural parameters is redefined as ψ =(, β, ζ)
0
, and
minimizing the distance function
π
t
(ψ) requires some assumption about the stochastic term η
t
. If
one assumes that E(η
t
|Z
t−1
)=0, and furthermore that η
t
is serially uncorrelated, the estimator
b
ψ
1
in (2.5) can be redefined by replacing the moment condition analogous to (2.4) with a conditional
expectation
E(
π
t
|Z
t−1
)=E(π
t
− π
m
t
(ψ
0
,A
0
)|Z
t−1
)=0. (3.6)
Using the auxiliary VARto construct the projection of π
t
and Z
t
on Z
t−1
, (3.6) becomes
e
0
π
AZ
t−1
− e
0
π
Z
t−1
− ζe
0
s
(I − βA)
−1
AZ
t−1
=0, (3.7)
andonecanthendefine a minimum distance estimator for ψ as in (2.5), with the appropriate
redefinition of b
π
t
. Similarly, the estimator
b
ψ
2
in (3.5) w ould be based on an analogous redefinition
of the function z, which is now given by the orthogonality conditions (3.7). Since these conditions
must hold for every Z
t−1
, it must be the case that
e
0
π
A − e
0
π
− ζe
0
s
(I − βA)
−1
A =0. (3.8)
The function z in (3.4) is then replaced by the left hand side of (3.8), with A replaced by its
consistent estimate
b
A, and the estimator of ψ is again defined as
b
ψ
2
in (3.5).
11
3.3. Relation with the GMM approach
Gali and Gertler (1999) estimate the baseline inflation model of (1.1) with a seem ingly different
empirical procedure. Instead of estimating the closed form solution of the model (as discussed
here), they define the error in expectations
ν
t+1
= π
t+1
− E
t
π
t+1
and, substituting actual for expected value of future inflation in the model, obtain
π
t
= β (π
t+1
− ν
t+1
)+ζs
t
,
or
βν
t+1
= βπ
t+1
− π
t
+ ζs
t
.
From the definition of rational expectations, the surprise in inflation at t+1 is unforecastable given
the information set at time t, I
t
: E(ν
t+1
|I
t
)=0, or
E [(π
t
− βπ
t+1
− ζs
t
) |I
t
]=0 (3.9)
Gali and Gertler’s estimation of the parameter vector ψ exploits this orthogonality condition in
a traditional GMM context. They observe that the orthogonality condition implies that any vector
of variables X
t−j
whichisintheinformationsetI
t
should be uncorr elated with the expectational
error: this implies a set of moment conditions based on the unconditional covariance of ν
t+1
and
X
t−j
. They therefore define a vector function
H(ψ, w
t
)=(π
t
− βπ
t+1
− ζs
t
) X
t−j
,
where w
t
=(π
t
,π
t+1
,s
t
,X
t−j
)
0
, and use the orthogonalit y conditions E(H(ψ, w
t
)) = 0 for estima-
tion. They then proceed with textbook GM M estimation: given T observations on the vector of
variables w
t
, the parameter vector ψ is estimated as the vector that minimizes the sample equivalen t
of the orthogonality conditions, for an appropriate weighting matrix.
12
Now suppose that X
t−j
= Z
t−1
; this amounts to choosing as instrumen ts the variables that
optimally forecast Z
t
. Then it is easy to see the relationship between this estimator and the distance
estimators proposed above. Taking conditional expectation of (3.9) one gets
E (π
t
|Z
t−1
) − βE (π
t+1
|Z
t−1
) − ζE (s
t
|Z
t−1
)=0,
which, using the auxiliary VARto compute the projections, gives
e
0
π
AZ
t−1
− βe
0
π
A
2
Z
t−1
− ζe
0
s
AZ
t−1
=0.
Hencewehave
e
0
π
A − βe
0
π
A
2
− ζe
0
s
A =0. (3.10)
A distance estimator of the kind I proposed, but based on restrictions (3.10), would be ex-
ploiting similar orthogonality conditions as a GMM estimator (conditional expectations instead
of unconditional covariances), where the instrumentsetischosentobethesetofpredetermined
variables of t he ‘auxiliary’ VAR.
12
In this context, the issue raised by Kurmann of the uncertainty
in the estimate of the first step VAR would boil down to the issue of the choice of variables in
X
t−j
; insignificant VAR coefficients imply that those variables are weak instrumen ts.
There is an im portan t difference, however, between the restrictions exploited by the GMM
approach described above, and the distance estimator of my formulation. These restrictions are
stated in the infinite horizon form - conditions (3.8), which is based on the projection of all future
values of the forcing variables that appear in the closed form solution. The GMM restrictions are
instead stated in the single period form - conditions (3.10), using the projection of one period-ahead
inflation onto the variables in Z
t−1
; as such, they are a non linear transformation of conditions
12
This is the interpretation that Li (2003) gives to her estim ate of the New Ke ynesian Phillips curve. While the
orthogonality conditions appear the same, however, the distance
h
z(ψ,
e
A) is not a sample mean, as in the metho d of
moments estimation.
13
(3.8), obtained by postmultiplying them by (I − βA) (and using the fact that (I − βA)
−1
A =
A (I − βA)
−1
).Howthisaffects inference is a matter to be explored.
13
3.4. Accounting for the VAR uncertainty
Whichever interpreta tion is given to the two-step distance estimator, a proper account should be
given to the uncertainty associated with the first-step estimate of the autoregressive parameters.
While one can easily derive appropriately corrected asymptotic standard errors,
14
in my application
of this estimator to a two-variable model (Sbordone 2003) I use instead a small sample approach,
which is in the same spirit of Kurmann’s assessment of the significance of the statistics of the Calvo
model.
Specifically, I use the empirical distribution of the parameter vector α
π
≡ e
0
π
A, draw from it N
samples α
πi
(i =1,...N), and for each of those I compute a minimum distance estimate
b
ψ
i
of the
vector of structural parameters ψ. I then compute the sample variance of the estimated
b
ψ
i
,and
report its square root as the standard erro r.
15
Furthermore, for each
b
ψ
i
I compute a value of the distance function
b
z
i
, and from this generated
sample I compute the covariance matrix of
b
z, Σ
z
. IusethelasttocomputeaWaldstatistic,
Q = z(
b
ψ)
0
Σ
−1
z
z(
b
ψ),wherez(
b
ψ) isthevalueofthedistanceevaluatedattheoptimalparameter
values, and use this statistic to evaluate the overall restrictions imposed by the model on the VAR
structure.
13
This issue was raised by Campbell and Shiller, and has been discussed by others as well. See for ex. Lafontaine
and W hite (1986).
14
These involve the derivative of the mo del solution with resp ect to the second stage parameters, and the covariance
matrix of the VAR parameters (an appendix is available from the author).
15
N is set to 500 in this calculation.
14
3.5. Model misspecifications
The other papers in this volume address the issue of potential misspecifications of the basic NKPC
model (1.1), which take the form of o mitted variable problems. I already considered the general
specification (1.5), which allows lagged inflation to affect current inflation directly, beyond its
possible role in forecasting the labor share. Batini et al. (2005) extend the NKPC to the case
of an open economy, and consider the role of material input prices, and of foreign competition.
Furthermore, they allo w for employment adjustment costs, which imply that both current and
future employment en ter the specification of their inflation equation.
Such modifications can be interpreted as corrections to the labor share in order to reach an
appropriate measure of the real marginal cost.
16
For example, when facing labor adjustment costs
employers may vary the effort margin: in this case an appropriate measure of labor input should
include a measure of effort. But if effort depends on how hours are expected to grow, compared to
actual hours, the marginal cost would differ from the average labor cost (or labor share) by such a
difference. In this particular case, the theoretical real marginal cost that drives inflation dynamics
is no more equal to the labor share, but is better approximated as follows
rmc
t
= s
t
+ δ
0
(dh
t
− δ
1
E
t
dh
t+1
) , (3.11)
where the term in brackets represent the expected deviation of future hours growth from current
growth, and the coefficient δ
0
measures the curvature of the adjustment cost function.
17
When
substituted in the pricing equation, this expression leads to an equation similar to the one obtained
by Batini et al. (2005). A closed form solution for inflation dynamics of the form (3.1) is obtained
by computing t he forecast into the infinite future of the deviation of hours from the value expected
16
For an extensive discussion of how to construct suitable measures of marginal cost see Rotemberg and Wo odford
(1999).
17
The model of labor hoarding that generates this result is developed in my previous work (Sbordone 1996). The
parameter δ
1
dep ends on the steady state value of the discount factor, and on the growth rate of hours and wages.
15
one period ahead. In this case the two-step estimation approach requires, in the first step, the
estimation of a VARmodel extended to include hours of work: this allows to construct the forecast
of hours that appears in (3.11). In the second step, the function z () is appropriately redefined to
reflect the modification to the labor share as a measure of marginal costs.
4. Se lecte d results
Table 1 reports some results obtained by applying the described methodology to estimating various
specifications of the pricing model. The baseline unrestricted representation of the data is a VAR
in inflation and labor share, with three lags (p =3). Z
t
is an mp−v ector con taining the curren t
and (p − 1) lags of all elements of y
t
,wherey
t
=[π
t
es
t
]
0
, and es
t
is a measure of the labor share,
transformed to obtain stationarity
18
. The parameters of the matrix A are estimated b y OLS, and
the consistent estimate of the covariance matrix of its relevant elements α
π
(≡ e
0
π
A) is
b
Σ
α
π
. The
weighting matrix in the distance estimator is set equal to diag(
b
Σ
α
π
), which, given the interpretation
of z(ψ, A) as a set of restrictions on the parameters of the inflation equation, do wnw eights the
parameters which are estimated with higher uncertainty. The discount factor β is calibrated in all
specifications to the value of .99.
The data cover the period 1951:1 - 2002:1 (a slightly longer period than that used by Kurmann
2005); both the hypotheses that inflation has no predictive power for the labor share and that labor
share has no predictive power for inflation can be rejected at standard confidence levels.
19
The ‘inertia’ coefficient ζ, is, as we saw, a combination of various stru ctural parameters. Its
18
The labor share is the ratio of real wage to p roductivity. I use instead the variab le hs = w − aq (with a = .9558),
which eliminates t he downward trend of the ratio which characterizes the data in the ‘90s.
19
The F −value of a Granger causality test of the predictive power of inflation for the lab or share has a p−value
of .051, and that for the predictive p ower of labor share for inflation has a p−value of .029. These results contrast
with those of Kurmann, who finds absence of G ran ger causality, but are not due to the differ ent sample length. One
p ossible explanation is Kurmann’s overparametrization.
16
estimate is statistically significant, and corresponds to price rigidity of about 10 to 12 mon ths.
20
π
m
indicates the inflation series predicted by the model, and two statistics measure the approximation
of predicted to actual inflation: the ratio of the standard deviations, σ
π
m
/σ
π
, in column 4, and
the correlation coefficient, corr(π
m
,π), in columns 5. Both measures show a quite high degree of
approximation. Moreov er, as the statistic Q in the last column shows, the restrictions imposed by
the model on the VAR are not rejected at standard significance levels.
The second row considers the role of lagged inflation. I find that π
t−1
enters significantly
the equation, and its inclusion reduces to some extent the size of the estimated coefficient of the
forward-looking compo nent, as would be expected in the case of an omitted variable problem. The
fit in the other dimensions is similar. Note that in models with lagged inflation π
m
t
is constructed
sequentially, starting from the actual value of inflation in period 0 (1951:4 in the sample used
here): π
m
t
= bπ
m
t−1
+
b
ζe
0
s
(I − βA)
−1
Z
t
. Given the initial value of inflation, this series describes
the evolution of inflation implied by the Calvo model, which depends at an y point in time on the
realization in the previous period (according to the model), on the current value of real marginal
cost, and on a forecast of its future realizations. The statistics reported in cols. 4 and 5 measure
how the volatilit y of this implied series compares to the volatility o f actual inflation, and how close
the dynamic evolution of the theoretical and the actual inflation series are.
The results obtained for the generalized model allow to evaluate the relative importance of
backward vs. forward-looking components, an issue addressed by R udd and Whelan (2005). As
expression (1.3) shows, the weight on the backward-looking component is γ
b
= /(1 + β);the
estimates reported in the table imply that γ
b
is approximately .18, while the corresponding weight
for the forward-looking component, γ
f
= β/(1 + β), is approximately .82. These values are
consistent with the results of Gali and Gertler (1999) and Gali, Gertler and Lopez-Salido (2000,
20
As I discuss in Sbordone (2002), the estimated ζ allows inference about the average time between price adjust-
ments, providing one calibrates the capital share a, and the parameter θ which drives the elasticity of deman d. Th e
numbers that I report are o btain ed using inflation measured on a quarterly basis.
17
2005), and show that, even if one may reject the purely forw ard-looking v ersion of the NKPC in fa vor
of an equation including lagged inflation, the forward-looking component remains quantitatively
more relevant. Note that these estimates, like those obtained by Rudd and Whelan (2005), are
computed from the closed-form solution of the model, but differ from theirs in the way in which
expected future values of the labor share are treated. While they proxy expected future values with
realized values, I compute expected values as VARforecast, using two different VARspecifications.
They argue, however, that the forward-looking component doesn’t add much to the explanation of
inflation dynamics.
How do they reach this conclusion? First, it should be noted that they do not provide a
structural interpretation of their lag polynomial, and are therefore not able to map their estimates
of the lagged inflation coefficient into the weight of the bac kward-looking component in a model
such as (1.2). Second, they seem primarily interested in comparing the purely forward-looking
version of the NKPC with a univariate autoregressive model of inflation: although they find that
the forward-looking terms in a generalized equation are statistically significant (table 2, case B of
their paper), they argue that they are quantitatively unimportant, and they do not significantly
reduce the explanatory power of the own-lagged inflation terms.
By contrast, my benchmark is not a purely autoregressiv e model of inflation, but an unrestricted
bivariate representation of inflation and labor share, and I ask whether the New Keynesian Phillips
curve is a structural model that can provide an explanation for the inertial behavior of the data. In
the generalized form of the NKPC model, lagged inflation derives from the assumption of partial
indexation, which can be justified in the context of the micro foundations of the model, typically
by information costs associated with reoptimization. Furhermore, to evaluate the importance of
the forward-looking terms, I ask what would be the fit of the model were the forward looking
component be set to zero. In the context of the closed-form solution of the generalized Calvo
model, this amounts to setting to zero all but the contem poraneous value of the labor share: it is
18
not equivalent to estimate an autoregressive model of inflation.
When estimating the model under this constraint (the results are on the third row of table 1), I
obtain coefficien t estimates on lagged inflation and on labor share both significantly higher than in
the case in which expected future marginal costs are allowed to matter: this can be interpreted as
evidence that there is an omitted variable problem in this specification. More importantly, though,
I obtain in this case a much poorer approximation of the inflation dynamics, as the statistics
presented in the table show. The reverse restriction, which sets the bac kward-looking component
to zero, gives the purely forward-looking pricing equation (3.1) that, as we already saw, provides
instead quite a good approximation of the dynamics of inflation.
Augmenting the model to allow for labor adjustment costs does not improve the fit of the model.
When real marginal cost is specified as in (3.11), the coefficient that measures the curvature of
the adjustment cost function is not statistically significant, whe ther I measure labor by hours or
employment.
21
This difference from the result of Batini et al. (2005), however, may be due to the
tighter specification of the adjustmen t cost adopted here, or may reflect some structural difference
between U.S. and U.K., and it is certainly worth further investigation.
Finally, to show how one can go further with this methodology into the specification of marginal
costs, I report in the last line of the table estimates of the parameters and ζ from my (2003)
study where I analyze both inflation and wage dynamics. The structural model considered in that
study adds to inflation a second equation describing the evolution of the labor share, derived from
a model of wage setting with staggered contracts. The two-equation model therefore imposes a
number of additional restrictions on the time series represen tation of wages and prices which are
exploited for estimation. The parameter vector ψ in this case is six-dimensional, and includes
parameters describing the intertemporal rate of substitution between leisure and consumption and
21
Row 4 in the table rep orts results u sing hours. In the estimation I set δ
1
=1, and report in the table the estimate
of the coefficient of the hours term in t he projected inflation, which is δ = ζδ
0
.
19
the degree of wage indexation (though I report here only the estimates of the parameters of the
inflation equation). The elements of the distance function z include in this case restrictions on the
VAR parameters of both the inflation and the labor share equations. As the reported estimates
show, the endogenization of the wage process allows a sharper estimate of the coefficien ts of the
inflation process, and otherwise confirms the single-equation results.
A final foootnote on another point discussed by Kurmann (2005): the inertia coefficient, in all
of the estimates presented here, when allowing forward looking terms, ranges between .017 and
.026. These results are consistent with the estimates obtained by Gali-Gertler (1999), or Batini et
al.(2005), for example, and they also imply a degree of inertia in prices similar to that reported
by survey estimates. This implies that parametrizations of ζ as high as the value .08 chosen by
Kurmann (2005) do not appear to be supported by the data.
5. Conclusion
In this paper I discuss the two-step estimation procedure used in Sbordone (2002), give a more
general interpretation to it, and presen t some additional results on the estimation of the New
Keynesian Phillips curve.
I show that under this more general interpretation, the auxiliary forecasting model on which
theprocedurerelies(thefirst step of the estimation) is an unrestricted representation of the data,
against which to test the model. While the uncertainty of the first estimation stage, discussed by
Kurmann (2005), can be taken into account within the procedure itself, issues about the VAR
modeling, like the preliminary stationarity-inducing transformations, the size of the model and the
lag length, and the time invariance of the structure still remain to be addressed.
22
And ultimately
only an increase in the precision of the VARestimates can reduce the uncertainty surrounding the
22
The issue of the structural invariance of the Calvo parameters is addressed in joint work with Tim Cogley (2005).
We estima te an unrestricted time series model of inflation with drifting parameters, and investigate the issue of
whether the parameters of the Calvo model are invariant to instability in trend inflation.
20
derived statistics that Kurmann documented so thoroughly.
The partial-information estimation strategy that I discussed has the advantage of relying on a
small number of restrictions (in the case analyzed here, those specifictotheinflation dynamics)
which must hold in every model that incorporates the same form of inflation dynamics. Moreover,
as the application to the model of price and wage dynamics show s, one can sequentially endogenize
variables that are initially modeled only with an unrestricted time series model.
What does all this imply for the empirical assessment of the Calvo model of inflation dynamics?
I would argue that the pricing model explored here is a good representation of the data, and price
stickiness of this kind is a valid hypothesis to incorporate into more complete models for business
cycle and policy analysis. In particular, the forward-looking terms are quite important in explaining
the dynamics of inflation: while it is possible to reproduce the dynamics of inflation fairly well with
a purely forward-looking model, eliminating instead the dependence on expected future values of
the labor share significantly worsens the o verall fit.
The validity of this pricing model, however, does not necessarily imply a relation betw een
inflation and output of the form generally referred to as the NKPC. What has emerged from the
copious empirical research on inflation dynamics, in m y opinion, is that a full understanding of the
Phillips curve can in fact be reached only through an understanding of the dynamics of labor costs,
and how these relate to output dynamics. And this is where future empirical research should be
focused.
References
[1] Batini, N., Jackson, B. and S.Nickell, S. 2005, An Open Economy New Keynesian Phillips
Curve for the U.K., Journal of Monetary Economics, this issue.
[2] Calvo, G., 1983, Staggered Prices in a Utility-Maximizing Framework, Journal of Monetary
Economics, 12 (3), 383-398.
[3] Campbell, J.Y. and R.J. Shiller, 1987, Cointegration and Tests of Present Value Models,
Journal of Political Economy 95, 1062-1088.
21
[4] Christiano, L.,Eichenbaum, M., and C. Evans, 2005, Nominal Rigidities and the Dynamic
Effect of a Shock to Monetary Policy, Journal of Political Economy, forthcoming.
[5] Cogley, T. and A.M. Sbordone, 2005, A Search for a Structural Phillips Curve, Federal Reserv e
Bank of New York Staff Reports n. 203.
[6] Gali, J. and M. Gertler, 1999, Inflation Dynamics: A Structural Econometric Analysis, Journal
of Monetary Economics 44, 195-222.
[7] Gali, J., Gertler, M. and J.D. Lopez-Salido, 2001, European Inflation Dynamics, European
Economic Review 45 (7), 1237-1270.
[8] Gali, J., Gertler, M. and J.D. Lopez-Salido, 2005, Robustness of the Estimates of the Hybrid
New Keynesian Phillips Curve, Journal of Monetary Economics, this issue.
[9] Kurmann, A. 2005, Quantifying the Uncertain ty about a Forward-Looking New Keynesian
Pricing Model, Journal of Monetary Economics, this issue.
[10] Lafont aine, F. and K.J. White, 1986, Obtaining Any Wald Statistic You Want, Economic
Letters 21, 35-40.
[11] Li, H., 2003, Testing Alternative Theories of Aggregate Supply, unpublished, Princeton Uni-
versity .
[12] Nelson, E. and K. Neiss, 2003, Inflation dynamics, marginal cost and the output gap: Evidence
from three countries, Journal of Money, Credit and Banking, forthcoming.
[13] Rotemberg, J.J. and M. Woodford, 1999, The Cyclical Behavior of Prices and Costs, in Ta ylor,
J.B., Woodford, M., eds., Handbook of Macroeconomics, Vol. 1B (North-Holland, Amster-
dam).
[14] Rudd, J. and K. Whelan, 2005, New Tests of the New-Keynesian Phillips Curve, Journa l of
Monetary Economics, this issue.
[15] Sbordone, A.M., 1996, Cyclical Productivity in a Model of Labor Hoarding, Journal of Mone-
tary Economics 38, 331-361.
[16] Sbordone, A.M., 2002, Price and Unit Labor Costs: A New Test of Price Stickiness. Journal
of Monetary Economics 49, 265-292.
[17] Sbordone, A.M., 2003, A Limited Information Approach to the Simultaneous Estimatio n of
Wage and Price Dynamics, Rutgers University, wp # 2003-23.
[18] Woodford, M., 2003, Interest and Prices: Foundations of a Theory of Monetary Policy, Prince-
tonUniversityPress.
22
TABLE 1- Parameter estimates and moments
ζδσ
π
m
/σ
π
corr(π
m
,π) Q − test
Pure forward-looking model .025 .765 .919 2.25
(.013) [.89]
Generalized model .224 .017 .721 .903 3.42
(.15) (.010) [.75]
Excluding forw a rd-looking terms .488 .079 .432 .557 14.96
(.08) (.04) [.02]
Adding labor adjustm ent costs .254 .016 .046 .792 .824 6.27
(.09) (.010) (.07) [.73]
With endogenous labor share .226 .026 .710 .905 23.55
(.103) (.006) [.79]
23
