Content uploaded by Jeffrey C Fuhrer
Author content
All content in this area was uploaded by Jeffrey C Fuhrer on Nov 12, 2014
Content may be subject to copyright.
1
No.09‐14
InflationPersistence
JeffreyC.Fuhrer
Abstract:
Thispaperexaminestheconceptofinflationpersistenceinmacroeconomictheory.Itbegins
withadefinitionofpersistence,emphasizingthedifferencebetweenreduced‐formand
structuralpersistence.Itthenexaminesanumberofempiricalmeasuresofreduced‐form
persistence,consideringthepossibilitythatpersistencemayhavechangedovertime.The
paperthenexaminesthetheoreticalsourcesofpersistence,distinguishing“intrinsic”from
“inherited”persistence,andderivinganumberofanalyticalresultsonpersistence.It
summarizestheimplicationsforpersistencefromtheliteratureson“sticky‐information”
models,learningmodels,andso‐called“trendinflationmodels,”providingsomenewresults
throughout.
Keywords:inflationpersistence,Phillipscurve,autocorrelation
JELClassifications:E31,E52
JeffreyC.Fuhreristhedirectorofresearchandanexecutivevicepresidentintheresearchdepartmentofthe
FederalReserveBankofBoston.Hise‐mailaddressisjeff.fuhrer@bos.frb.org.
Thispaper,whichmayberevised,isavailableonthewebsiteoftheFederalReserveBankofBostonat
http://www.bos.frb.org/economic/wp/index.htm.
TheauthorthanksFabiàGumbau‐Brisa,TimothyCogley,DennyLie,GiovanniOlivei,ScottSchuh,OzShy,Raf
Wouters,andparticipantsattheFederalReserveBankofBoston’slunchtimeworkshopforhelpfulcomments.
Theviewsandopinionsexpressedinthispaperarethoseoftheauthoranddonotnecessarilyrepresenttheviews
oftheFederalReserveBankofBostonortheFederalReserveSystem.
Thisversion:November30,2009
1 Introduction
What is meant by “persistence” and why is it important to macroeconomists
and policymakers? In broad terms, persistence is the economic analogue of in-
ertia in physics. Inertia may be defined as the resistance of a body to changing
its velocity (direction and rate of speed) unless acted upon by an external force.
This law is often paraphrased as “a body at rest will remain at rest unless acted
upon by an external force,” which is one example of the principle. Newton’s
second law of motion, F=ma or a=F
m, captures this idea algebraically: The
magnitude of the force required to produce a given change of velocity (acceler-
ation) is proportional to the body’s mass.
While no analogy is perfect, this one works reasonably well at an intuitive
level. An economic variable is said to be persistent if, other things being equal, it
shows a tendency to stay near where it has been recently, absent other economic
forces that move it elsewhere. In the case of inflation, the rate of change of the
price level tends to remain constant (inflation tends to be persistent) in the
absence of an economic “force” to move it from its current level. This paper
provides more precise definitions of inflation persistence in the sections below,
but this physical analogy may provide motivating intuition.
1.1 Early inflation models and the empirical necessity of
lagged inflation
For many decades, economists assumed that inflation was an inertial or per-
sistent economic variable. The concept of the sacrifice ratio—the number of
point-years of elevated unemployment required to reduce inflation by a per-
centage point—implies that inflation does not move freely, requiring significant
economic effort in the form of lost output to reduce its level.1The early in-
carnations of the accelerationist Phillips curve modeled the apparent inertia in
inflation by including lags of inflation. A canonical example of such specifica-
tions is Gordon’s “triangle model” of inflation, here replicated in simplified form
1See Gordon, King, and Modigliani (1982) for the first study that uses the term “sacrifice
ratio.” This study followed on the work of Arthur Okun (1977).
2
(Gordon 1982):2
πt=
k
X
i=1
aiπt−i−b(Ut−¯
U) + cxt+t.(1.1)
Inflation, πt, depends on its own lags (normally constrained to sum to one to
reflect the Friedman-Phelps accelerationist principle), a measure of real activity
(here the deviation of unemployment, Ut, from the non-accelerating inflation
rate of unemployment or “NAIRU”), and supply-shifters such as key relative
price shifts, summarized in xt. In such a model, inflation moves gradually,
partially anchored by its recent history, in response to real activity and supply
shocks. These variables may themselves be persistent, in which case inflation
will“inherit” some of their persistence. A key question is whether and why
inflation has its own or “intrinsic” persistence, beyond that inherited from Ut
and xt(or perhaps t, if it is also serially correlated). If inflation exhibits
intrinsic persistence, then a model of inflation may require the equivalent of the
lags in equation 1.1 above.
In this early literature, the theoretical justification for including lags of in-
flation was as a proxy for expected inflation and a proxy for contracting and
other price-setting frictions. As an empirical matter, the lags helped the model
fit the data. To see this last point simply, consider the R2for estimates of
Gordon-style Phillips curves with and without the lags of inflation in table 1.
The specification is πt=P4
i=1 αiπt−i+P2
j=1 βjUt−j+P2
k=1 γk∆rpo
t−k+C,
where πtis the quarterly percentage change in the core CPI, Uis the civilian
unemployment rate, and rpois the relative price of oil. The sum of the αi’s is
constrained to one in some of the estimates.3
The points to take from this table are that (1) the lags of inflation are
empirically critical, whatever they may represent structurally; and (2) as a
consequence, it is critical to understand what these lags represent structurally.4
2See Friedman (1968) and Phelps (1968) for the earliest explications of the accelerationist
Phillips curve.
3Note that this constraint is not statistically significant in these regressions—the R2’s are
identical in the constrained and unconstrained cases, and the p-value for the test of these
restrictions exceeds 0.8 for both samples.
4The results in the table are completely invariant to the choice of the lag length for lagged
inflation. For lag lengths up to 24, we obtain nearly identical coefficient sums, R2’s, and
non-binding unit sum constraints. However, if we begin the later sample in 1997, results
3
Table 1: R2for Gordon-style Phillips curves
Model R2
Core CPI, 1966:Q1-1984:Q4
with lags, Pαi= 1 0.74
with lags, Pαi6= 1 0.74
without lags 0.24
Core CPI 1985:Q1-2008:Q4
with lags, Pαi= 1 0.79
with lags, Pαi6= 1 0.79
without lags 0.09
Core PCE, 1966:Q1-1984:Q4
with lags, Pαi= 1 0.76
with lags, Pαi6= 1 0.77
without lags 0.39
Core PCE 1985:Q1-2008:Q4
with lags, Pαi= 1 0.72
with lags, Pαi6= 1 0.72
without lags 0.16
1.2 Rational expectations and inflation persistence: An
introduction to some of the issues
The introduction of Muth’s (1961) theory of rational expectations into the
macroeconomics literature and the consequent move toward explicit modeling
of expectations posed considerable challenges in modeling prices and inflation.
In the earliest rational expectations models of Lucas (1972) and Sargent and
Wallace (1975), the price level was a purely forward-looking or expectations-
based variable like an asset price, which in these models implied that prices
were flexible, and could “jump” in response to shocks. It was difficult at first
to reconcile the very smooth, continuous behavior of measured aggregate price
indexes such as the consumer price index with the flexible-price implications of
these early rational expectations models. Figure 1 displays the consumer price
index for the post-war period.
A number of economists recognized the tension between the obvious persis-
tence in the price level data and the implications of these early rational expec-
tations models. Fischer (1977), Gray (1977), Taylor (1980), Calvo (1983), and
change dramatically. The constraint on the sum of the lag coefficients is now binding, and
the R2’s drop to 0.25 or below, a result highlighted in Williams (2006). This apparent shift
in reduced-form persistence is discussed in more detail below.
4
Figure 1: The consumer price index
Rotemberg (1982, 1983) developed a sequence of models that rely on nominal
price contracting in attempts to impart a data-consistent degree of inertia to
the price level in a rational expectations setting. The overlapping contracts
of Taylor and Calvo/Rotemberg were successful in doing so, allowing contracts
negotiated in period tto be affected by contracts set in neighboring periods,
which would remain in effect during the term of the current contract.5The
subsequent trajectory of macroeconomic research drew heavily on these seminal
contributors, who had neatly reconciled rational expectations with inertial (or
persistent) macroeconomic time series.
However, in the early 1990s, a number of authors discovered that these ratio-
nal expectations formulations had less satisfying implications for the change in
the price level, that is, the rate of inflation. Ball (1994) demonstrated that such
models could imply a counter-factual “disinflationary boom”—the central bank
could engineer a disinflation that caused output to rise rather than contract.
Fuhrer and Moore (1992, 1995) showed that Taylor-type contracting models
implied a degree of inflation persistence that was far lower than was apparent in
5For example, a four-period contract negotiated in period twould be influenced by the
contracts negotiated in the previous three periods, as well as by the contracts expected to be
negotiated in the following three periods.
5
inflation data of the post-war period to that point. To build intuition, compare
equation 1.1 with a Calvo- or Taylor-type inflation equation6
πt=Etπt+1 +γ˜yt+t.(1.2)
Implicitly, this makes inflation, πt, a function of all expected future output gaps,
˜yt, (assuming that Ett+i= 0 ∀i > 0, or equivalently that thas no autocor-
relation).7Inflation will indeed inherit the persistence in output, but nothing
beyond that. The lags in the Gordon triangle model are gone. Inflation is com-
pletely forward-looking: Following a shock to output, inflation can jump imme-
diately in response. If output exhibits no persistence (that is, there are no “real
rigidities”), then neither will inflation. In contrast, equation 1.1 implies that
inflation depends on all past output gaps and shocks. As in the forward-looking
model, inflation inherits the persistence in output, but the lagged inflation terms
mean that inflation cannot jump in response to a shock to output—inflation ex-
hibits additional persistence (or inertia) in the sense that in response to shocks,
inflation has a tendency to remain near its most recent values.
To explore the dynamic implications of this forward-looking specification,
we embed equation 1.2 in a skeletal macro model. To complete the model, we
include a rudimentary I–S curve and a simplified policy rule
˜yt=ρ˜yt−1−a(rt−Etπt+1) (1.3)
rt=b(πt−¯π),
where rtis the short-term policy rate controlled by the central bank, the inflation
target is denoted ¯π, and the equilibrium real rate is suppressed for convenience.8
We will consider cases with zero and non-zero output persistence, that is, ρ= 0
or ρ6= 0. The policy response to inflation gaps is calibrated by parameter b,
which is set to 1.5 throughout.
Consider first the case in which output is not persistent. Inflation is at its
target and the output gap is zero. For comparison with the large literature that
6See Roberts (1997) for a derivation that shows the approximate equivalence of these
formulations.
7This is made explicit below in equation 3.2, but one can obtain the result by inspection
through repetitive substitution of the definition of future inflation into equation 1.2.
8Note that the derivation of the Calvo Phillips curve of equation 1.2 assumes a zero steady-
state inflation rate, whereas equations 1.3 allow for the possibility of a non-zero steady-state
inflation rate. Cogley and Sbordone (2009) derive the appropriate Phillips curve for the case
in which the central bank pursues a non-zero inflation rate. For the purposes of this exercise,
the inaccuracy in the approximation is likely small.
6
attempts to measure the economy’s response to an identified monetary policy
shock, consider a one unit positive shock to the policy rate rt. With no inertia in
output or inflation, both output and inflation are perturbed below their steady
state for one period. In period 2, both return to their steady state. The solid
lines in figure 2 display the results of this rather uninteresting exercise.
When output is persistent (here ρ= 0.9), the dynamics are a bit more
complex. Starting from the same steady state, consider a one unit positive shock
to the policy rate. The results of this simulation are depicted in the dashed lines
in the same figure. Output is depressed below its steady state in the first period,
and because of its own persistence, remains below its steady state for some
time. In order for equation 1.2 to hold, it must be that expected inflation lies
above current inflation whenever the output gap is negative. Because output is
persistent, the expected change in inflation must be positive for as long as output
remains negative. But ultimately, inflation will have to return to its original
steady state. Thus, inflation must immediately jump down, and then rise to its
equilibrium from below. Inflation exhibits dynamics that are reminiscent of the
exchange rate in the famous Dornbusch (1976) overshooting model.9
For the sake of comparison, the dotted line in figure 2 shows the outcome
when the inflation equation includes a lag of inflation. Here, the equation is a
“hybrid” equation that mixes both forward- and backward-looking elements, of
the form
πt=µπt−1+ (1 −µ)Etπt+1 +γ˜yt+t(1.4)
with µ= 0.5 in the figure. Now, in response to the monetary policy shock,
output behaves approximately as before, but inflation declines gradually over
several quarters, exhibiting the more typical “hump-shaped” response found in
the literature on monetary VAR’s.10 It is no longer the case that expected
inflation must exceed current inflation for as long as the output gap is negative.
Perhaps as striking are the differences in behavior in a fully credible disinfla-
tion. In the solid and heavy dashed lines of figure 3, inflation is forward-looking
(µ= 0). The central bank announces a permanent and fully credible reduc-
tion in its target inflation rate from 2 percent to 0 at time 1. As the figure
9This overshooting property is examined in Fuhrer and Moore (1995), and in Estrella and
Fuhrer (2002).
10See Christiano, Eichenbaum, and Evans (2005) for representative results. Note that the
output response differs modestly from the dashed line because the path of inflation, and thus
the path of the short-term real rate, differs.
7
Figure 2: Response of inflation to a monetary policy shock in the simple Calvo
model
indicates, regardless of how persistent is output, the inflation rate jumps to its
new equilibrium in the period after the announcement, with no disruption to
output, so that the solid and heavy dashed lines coincide in the figure.11 When
lagged inflation is added to the inflation equation, inflation declines gradually
to its new long-run equilibrium, with a concomitant decline in output during
the transition.
Many would view the dynamics of the purely forward-looking specification as
strikingly counter-factual. Counter-factual or not, one needs to understand the
dynamics of inflation to pursue appropriate monetary policy. A knowledge of
the reduced-form behavior of inflation is not sufficient. The central bank needs
to understand the sources of inflation dynamics—in these examples, whether it
arises from the persistence of output, which may in turn arise from the behav-
ior of monetary policymakers, or from persistence intrinsic to the price-setting
process. This is one of the reasons that the issue of persistence is of more than
passing interest to macroeconomists and policymakers.
To be sure, understanding why and when inflation may be persistent can
11If the inflation target were pre-announced, the inflation rate would jump to its new target
in the first period, again with no disruption to output.
8
Figure 3: Credible disinflations in the simple Calvo model
be more complicated than these simple examples suggest. For instance, the
examples leave out the possibility that the policy shock is caused by a mon-
etary authority with imperfect credibility; they also abstract from imperfect
knowledge of the economy that might lead to learning on the part of private
agents. Both of these situations would alter the dynamic implications of the
simple examples. We return to some of these complications below.
2 Defining and measuring reduced-form infla-
tion persistence
The discussion above suggests that it will be useful to distinguish reduced-
form from structural inflation persistence. Reduced-form persistence will refer
to an empirical property of an observed inflation measure, without economic
interpretation. Structural persistence will refer to persistence that arises from
identified economic sources. A key objective of recent inflation research has
been to map observed or reduced-form persistence into the underlying economic
9
structures that produce it. To a significant extent, this challenge remains.
2.1 Defining reduced-form persistence
There is no single definitive measure of reduced-form persistence. As the sec-
tions below discuss, researchers have employed a variety of measures to capture
the idea that inflation responds gradually to shocks, or remains close to its recent
history.12 Most of the measures of inflation persistence derive from the auto-
correlation function for inflation, so it will be useful to define it here. The ith
autocorrelation, ρi, of a stationary variable, xt—the correlation of the variable
with its own ith lag, xt−i—may be expressed as13
ρi=E(xtxt−i)
V(x),(2.1)
where V(x) is the variance of x. The correlations are of course bounded between
−1 and 1. The variable’s autocorrelation function is correspondingly defined as
the vector of correlations of current period xwith each of its own lags xt−ifrom
i= 1 to k:
A= [ρ1, ..., ρk].(2.2)
A time series will be said to be relatively persistent if its correlations with its
own past decay slowly. Thus, a graphical depiction of this measure of a series’
reduced-form persistence is provided by a plot of the variable’s autocorrelation
function. In figure 4 below, the correlation of the persistent series, Pt, with its
own lags declines gradually from 0.9 to 0.3 over 12 periods. In contrast, the
less persistent time series, Nt, shows a more rapid decline in its autocorrelation
function from 0.5 to 0 in about eight periods. To a large extent, the alternative
measures of persistence surveyed in section 2.2 below provide alternative ways
of quantifying the rate at which inflation’s autocorrelations decay.
The analytical representation of the autocorrelation function will also be
useful in this discussion. For example, if the variable xtis defined as a first-
order autoregressive process
xt=axt−1+t(2.3)
12This definition implicitly assumes that inflation is positively correlated with its own lags,
an assumption that holds up well over most of post-war history. More generally, a time series
may be deemed persistent if the absolute value of its autocorrelations is high, so that a strongly
negatively autocorrelated series would also be characterized as persistent.
13For convenience, this definition assumes that xtis a mean zero series.
10
Figure 4: Hypothetical autocorrelation functions
−1< a < 1, then its autocorrelation function is simply
A= [a, a2, ..., ak].(2.4)
From equation 2.4 it is clear that the autocorrelations of xtdie out geometrically
at the rate determined by the autoregressive parameter a. The analytical ex-
pressions for the autocorrelations of inflation in more complex structural models
with rational expectations are derived below.
Some researchers define persistence as the extent to which shocks in the past
have an effect on current inflation. This concept is related to the autocorrelation
function. The more correlated is inflation with its distant past, the more shocks
that perturb inflation in the distant past will be reflected in current inflation.
More formally, adopting a simple first-order autoregressive model for inflation
from equation 2.3, one can iterate the equation backward to obtain the moving-
average representation of inflation
πt=t+a(aπt−2+t−1) (2.5)
=t+at−1+a2t−2+a3t−3+. . . ,
which shows that the larger is a, and thus the more slowly inflation’s autocor-
11
relations decay, the larger is the influence of past shocks on current inflation—
equivalently, shocks have a more persistent effect on inflation.14
2.2 Measuring reduced-form inflation persistence
There is little agreement in the extant literature on how best to measure per-
sistence. Thus, we examine a battery of measures that attempt to capture the
persistence in inflation:
•Conventional unit root tests;
•The autocorrelation function of the inflation series (as defined in equa-
tion 2.4 above);
•The first autocorrelation of the inflation series;
•The dominant root of the univariate autoregressive inflation process (de-
fined below);
•The sum of the autoregressive coefficients for inflation;
•Unobserved components decompositions of inflation that estimate the rel-
ative contributions of “permanent” and “transitory” components of infla-
tion (for example, the IMA(1,1) and related models proposed by Stock
and Watson (2007)).
Because the autocorrelation function summarizes much of the information in a
time series, it may be the best overall measure of persistence. But researchers
often desire a single number that captures the overall persistence implied by the
full autocorrelation function; hence, many of the other scalar measures itemized
above. One can find overlap across the results from these tests, but, to be sure,
the results are neither uniform nor unambiguous. Throughout, we examine
these measures of inflation persistence across a number of subsamples, providing
suggestive evidence as to whether persistence has changed over time.
14This algebra provides a more rigorous justification for the argument that old-style Phillips
curves like those in equation 1.1 add persistence to inflation beyond that inherited from
output—the lags of inflation imply persistent effects of past shocks on current inflation.
12
2.2.1 The data
For the purposes of this paper, we will focus on three key inflation measures.
Each will be defined as 400 times the log change in the corresponding price index.
The three indexes are the GDP deflator, the consumer price index (CPI), and
the personal consumption expenditures chain-type price index (PCE). In some
cases, we will examine the so-called “core” versions of the CPI and PCE, that
is, the price indexes that exclude food and energy prices. We denote these
series throughout as “CPI-X” and “PCE-X.” These series abstract from the
high-frequency noise that can be introduced by volatile food and energy price
series.15
Figure 5 displays the three overall series in the top panel, with corresponding
core series in the bottom panel. Table 2 provides summary statistics for the three
inflation series over several sample periods. Evident in the figure and echoed in
table 2 are the drop in the level of all inflation measures since the early 1980s
(the top panel of the table); the decline in the variance of all measures, consistent
with the so-called “Great Moderation” (the second panel of the table); the high
correlation among the three series (the third panel of the table); and the lesser
volatility of the core series. Less evident in the figure but clear in the bottom
panel of table 2 is the decline in the correlations across the series. In particular,
while the CPI and PCE remain highly correlated, the correlations between the
core and overall measures have declined, as have the correlations between the
GDP deflator and the consumer price measures. Despite the relatively high
correlations among these series, the measures of persistence presented below
will show some noticeable differences across the series.
2.2.2 Unit root tests
Perhaps the first test of persistence should be a unit root test. If inflation
contains a unit root, its persistence is unquestionably large (infinite) and its
variance is unbounded.16 Many papers test for a unit root in inflation (see
Barsky (1987), Ball and Cecchetti (1990)); prior to the 1990s, the results tend
to suggest a unit root in inflation. In more recent years, researchers are more
15As indicated in figure 5, the core price series are available only starting in the late 1950s.
16A series with a unit root has infinite “memory,” in the sense that a shock in period thas
influence on all periods t+k, k > 0. More formally, if xt=xt−1+t, then xt=P∞
i=0 t−i.
Thus, any shock to a series with a unit root persists forever.
13
Figure 5: Inflation data
14
Table 2: Summary statistics, inflation measures
Series Means
59-08 59-84 85-08 95-08
CPI 4.0 5.0 3.0 2.5
PCE 3.6 4.5 2.6 2.1
GDP 3.6 4.6 2.5 2.2
CPI-X 4.0 4.9 3.0 2.3
PCE-X 3.5 4.4 2.5 1.8
Series Variances
59-08 59-84 85-08 95-08
CPI 9.4 12.7 3.7 4.3
PCE 6.4 8.6 2.2 2.2
GDP 5.7 7.8 0.95 0.83
CPI-X 7.2 10.9 1.2 0.31
PCE-X 4.7 6.4 1.2 0.25
Correlation matrix, 1959-2008
CPI PCE GDP CPI-X PCE-X
CPI 1.00
PCE 0.96 1.00
GDP 0.86 0.92 1.00
CPI-X 0.86 0.85 0.86 1.00
PCE-X 0.81 0.90 0.92 0.91 1.00
Correlation matrix, 1985-2008
CPI PCE GDP CPI-X PCE-X
CPI 1.00
PCE 0.94 1.00
GDP 0.57 0.70 1.00
CPI-X 0.48 0.56 0.53 1.00
PCE-X 0.40 0.62 0.65 0.87 1.00
15
likely to be unable to reject stationarity. Most monetary models would suggest
that the more vigorous attention to inflation on the part of central banks around
the world in recent decades is responsible for this change.
Table 3 provides univariate tests for the null that the inflation series contains
a unit root, for a long sample (1966–2008), and for a “Great Moderation” sample
(1985–2008). The results in the table are somewhat ambiguous. For the most
recent decades, one can develop strong rejections of the null of a unit root,
although this varies somewhat depending on the inflation series and on the
test employed. For the longer sample, the Phillips-Perron test rejects the null,
although not always very strongly. The ADF test fails to reject for all three
inflation series.
Table 3: Unit Root Tests for Inflation Measures
(p-values, null = series has unit root)
1966-2008
Series ADF Phillips-Perron
CPI 0.14 0.0069
PCE 0.37 0.045
GDP 0.33 0.025
1985-2008
Series ADF Phillips-Perron
CPI 0.00 0.00
PCE 0.07 0.00
GDP 0.31 0.00
What should one make of these tests? Certainly for the past 25 years, the
U.S. central bank has behaved as if it had a specific, low inflation goal, although
it has so far chosen not to announce that goal. If it does indeed have such a goal,
then theory suggests that the U.S. inflation rate will not have a unit root—or
that if it does have a unit root, the variance of that component of inflation
is quite small.17 This reasoning, combined with the more frequent rejections
of the unit root null in recent decades, suggests that one may safely assume
that inflation does not contain a unit root, at least not under current monetary
policy.
17This statement is somewhat too strong, as a very gradual drift downward in the target
inflation rate could manifest itself as a unit root component of inflation, albeit with a relatively
small variance. See Stock and Watson (2007) for an inflation model that formalizes this
reasoning.
16
2.2.3 First-order autocorrelations
If we can proceed under the assumption that inflation does not contain a unit
root, then section 2 above suggests the first-order autocorrelation coefficient
for the series as a simple measure of persistence, a measure used, for example,
in Pivetta and Reis (2007). Figure 6 extends and expands their figure 2a,
presenting rolling-sample estimates of the first-order autocorrelation coefficient
for three aggregate inflation measures.18 The figure shows that all three inflation
Figure 6: Rolling-sample estimates of the first-order autocorrelation coefficient
measures display similar time-variation in their autocorrelation. All rise in the
1970s to 0.8 or higher, and remain there until the mid-1990s, at which point
the correlation drops to 0.5 or 0.6. A third decline is evident in the mid-2000s,
with first-order autocorrelation dropping to very low levels indeed, between 0
and 0.4. These simple measures support the conclusion that inflation is not
currently well-characterized as a process with a unit root. The autocorrelations
also suggest that there has been noticeable time-variation in inflation’s reduced-
form persistence over time.
18The rolling-sample estimates presented here employ a 14-year window, the same as in
Pivetta and Reis.
17
2.2.4 Autocorrelation functions
Extending the previous section’s results, figure 7 displays the full autocorrelation
function for five of the key measures of inflation (CPI-X, PCE-X, CPI, PCE
and the GDP deflator) for two sample periods, 1966:Q1–1984:Q4 and 1985:Q1–
2008:Q4.19 As the figure suggests, over the first half of the past 43 years,
Figure 7: Autocorrelations of inflation data, various measures and samples
inflation exhibited considerable persistence in this reduced-form sense, according
to all measures. Since the mid-1980s, roughly corresponding to the onset of the
“Great Moderation,” the persistence of inflation has declined for some, but not
all, measures.20 The ambiguity of these results and the source of any reduction
19I use both core and total measures here because, as discussed in the next subsection, the
influence of key relative prices on the reduced-form persistence of inflation in the latter half of
the sample can be significant. The beginning of the early sample is chosen to coincide with the
first use of the federal funds rate as the Federal Reserve’s policy instrument. Extending the
sample back to 1959:Q2, the earliest date for which the PCE price series is available, produces
results that are qualitatively the same.
20See Benati (2008) for an empirical investigation into changing inflation persistence.
18
that may have occurred is of considerable interest, and section 2.4 discusses this
in some detail.
2.2.5 Dominant root of the univariate timeseries process
An alternative measure of inflation persistence is the dominant root implied by
the univariate autoregressive process for inflation. In particular, if the autore-
gressive representation of inflation is of lag length k
πt=c1πt−1+. . . +ckπt−k+t , (2.6)
then the companion matrix for the state-space representation of πtis
C≡c1. . . ck
Ik−1×k−10k−1×1,(2.7)
and the root of Cwith the largest magnitude is the (dominant) root of interest.
Table 4 summarizes the results for our three measures of inflation for a va-
riety of samples. The table shows a high degree of persistence over the past
25 years, with a modest decline as the earlier decades are dropped from the
sample. The results are somewhat dependent on the inflation measure; studies
that focus on the GDP deflator (such as Pivetta and Reis 2007) may well un-
cover less evidence of a decline in reduced-form persistence. Both the CPI and
the PCE measures show a more pronounced decline in persistence, particularly
in the post-1995 subsample.21 Another widely used measure of reduced-form
persistence is the sum of the autoregressive coefficients, which approximates
the long-run impulse response to a unit shock c(1) ≡Pk
i=1 ci. This measure
is provided in the right-hand-most column of table 4. As the table shows, the
mapping from the sum of the ci’s to the dominant root is not perfect, particu-
larly when the coefficients imply a dominant complex pair (indicated by the ∗),
as is the case for the CPI and the PCE in the latter subsample. In this case,
even though the sum of autoregressive coefficients may be small, the oscillatory
behavior implied by the complex pair of roots may die out only slowly, implying
significant persistence.
Because persistent relative price shifts can influence the persistence of mea-
sures of overall inflation, particularly in shorter samples, the bottom panel of
21Standard errors are computed from a Monte Carlo that draws 100,000 permutations of
the estimated residuals for each inflation measure and sample, creating a new inflation series
for each such residual permutation given the original estimated autoregression coefficients,
and re-estimating the autoregression and the dominant root for each permutation.
19
Table 4: Dominant root of autoregressive process for inflation
Measure Sample Dominant Root Sum of AR coeff.’s (c(1))
(Std. Error) (Std. Error)
CPI 66-08 0.94 0.89
(0.032) (0.055)
66-84 0.92 0.88
(0.034) (0.068)
85-08 0.70 0.41
(0.12) (0.18)
95-08 0.64∗-0.039
(0.11) (0.32)
PCE 66-08 0.95 0.91
(0.029) (0.046)
66-84 0.90 0.87
(0.034) (0.065)
85-08 0.81 0.61
(0.094) (0.14)
95-08 0.68∗0.22
(0.11) (0.26)
GDP 66-08 0.95 0.92
(0.027) (0.039)
66-84 0.87 0.84
(0.033) (0.075)
85-08 0.86 0.72
(0.074) (0.11)
95-08 0.83 0.64
(0.12) (0.17)
CPI-X 66-08 0.93 0.90
(0.027) (0.042)
66-84 0.86 0.81
(0.041) (0.081)
85-08 0.96 0.92
(0.039) (0.054)
95-08 0.75 0.62
(0.0641) (0.16)
PCE-X 66-08 0.95 0.93
(0.023) (0.033)
66-84 0.86 0.84
(0.032) (0.066)
85-08 0.95 0.91
(0.044) (0.061)
95-08 0.72 0.49
(0.097) (0.18)
∗denotes complex roots
20
table 4 presents the same results for the core CPI and PCE inflation measures.
The core-based measures suggest less evidence of a decline in persistence in
recent years. Neither CPI nor PCE inflation measures show a decline in the
dominant root or the sum of the AR coefficients for the past 25 years. The
dominant root estimate declines modestly for the period since 1995, but the
standard error of this estimate is correspondingly larger. For the more recent
and relatively short subsamples, these results suggest that large movements in
key relative prices may well distort the extent to which underlying inflation
persistence has changed, particularly if persistence is measured over a relatively
short sample.
2.3 Evidence of changing reduced-form persistence in the
United States
Recognizing the reduced-form nature of inflation persistence, a number of au-
thors have looked for evidence that changes in the underlying determinants of
inflation may have given rise to a change in reduced-form persistence. The
leading cause of a change in inflation behavior is thought to be a change in the
systematic behavior of the central bank. To make matters simple, consider the
stylized, backward-looking model of inflation below
πt=πt−1+axt
xt=−bft
ft=cπt.
(2.8)
The first equation is a skeletal “Phillips curve,” in which the change in inflation
is positively related to a variable xt, which we will take here to be the output
gap. The output gap in turn depends negatively on the short-term policy rate
ft(for federal funds rate), and the policy rate is a positive function of inflation
(with an implicit target inflation rate of 0). The solution for inflation is
πt=απt−1
α≡1
1+abc .(2.9)
Inflation will follow a first-order autoregression, and will be less persistent—the
coefficient αwill be smaller—the larger is the policy response to inflation (c), the
more responsive is the output gap to the policy rate (b), and the more responsive
is inflation to the output gap (a). In this simple framework, a central bank that
behaves more aggressively in moving inflation towards its target will reduce the
21
persistence of inflation. The intuition from this skeletal model generalizes to a
number of more sophisticated models that include rational expectations and a
richer description of the key elements sketched above.22
2.3.1 Pre-War inflation persistence under the gold standard
If persistence can be expected to change as a result of changes in the monetary
regime, a natural question is the extent to which inflation persistence changed in
the United States under considerably different monetary regimes. The contrast
in the United States between the post-World War II fiat money system and the
pre-World War I gold standard provides a natural experiment. Barsky (1987)
finds that while U.S. inflation persistence was very high from 1960 to 1979,
it was virtually non-existent prior to World War I—indeed, Barsky’s ARIMA
modeling of inflation during this period suggests that inflation is white noise.23
2.3.2 A parameterized characterization of reduced-form persistence
Stock and Watson (2007) posit a relatively straightforward time series model
of inflation that captures many of the features of reduced-form persistence dis-
cussed so far. The model can be expressed as an integrated moving-average
process of order one, or IMA(1,1)
∆πt=at−Θat−1(2.10)
or equivalently as an unobserved components model with stochastic trend and
stationary components24
πt=τt+ηt(2.11)
τt=τt−1+t,(2.12)
22A separate line of research examines the contribution of a time-varying inflation target
to measures of inflation persistence. The paper will return to this topic in more detail in
section 3.5 below.
23Barsky uses wholesale price data prior to World War I. To the extent that such data reflect
movements of commodity prices rather than consumer goods and services prices, this result
may be confounded by an inherent difference between commodities and final goods prices,
rather than a difference in persistence across the monetary regimes. Note that Barsky also
makes the link between persistence and forecastability, a link exploited in Cogley, Primiceri,
and Sargent (2007), discussed in section 2.3.3 below.
24One can see the equivalence by substituting the definition for τtτt=t
(1−L)into the
equation for inflation to obtain πt(1 −L) = t+ηt(1 −L), which, after re-arrangement, yields
an equation like equation 2.10, with inflation an integrated process with a moving average
error term.
22
where ηtand tare uncorrelated with one another, and are serially uncorrelated
with mean zero and variances σ2
ηand σ2
, respectively. One can think of τtas
reflecting the “permanent” or trend component of inflation, and ηtas capturing
the stationary component.
Interestingly, Stock and Watson find that the decline in the variance of in-
flation is largely due to a marked decline in the variance of the permanent
component, that is, a reduction in σ2
. It is still not possible, in their method-
ology, to reject the null of a unit root in inflation, but the variance of the shock
that drives τtis currently at historic lows.
2.3.3 Multivariate evidence of changes in reduced-form inflation per-
sistence
For the most part, the measures of persistence discussed so far are univari-
ate measures; that is, they use only the information in an inflation time series
to draw inferences about its persistence. However, some authors have argued
that one can draw more accurate inferences using a multivariate approach. In
part, the intuition behind this claim is connected to the “trend inflation” mod-
els discussed in section 3.5 below. In those models, the slow-moving or trend
component of inflation accounts for much of the persistence in inflation. That
trend in turn is most commonly associated with the central bank’s target rate of
inflation. As a consequence, including variables that reflect the central bank’s
inflation-targeting behavior, such as short-term policy rates, may help to iden-
tify both trend inflation and its persistence, and thus the persistence of inflation.
Cogley, Primiceri, and Sargent (2007) use a time-varying VAR to estimate
the trend component of inflation, which they associate with the central bank’s
inflation target. They find continued persistence in inflation, but they associate
it strongly with trend inflation. This implies that the Federal Reserve’s implicit
target for inflation continues to have a unit root (or near-unit root), although
Cogley et al. estimate the variance of that component to have declined, con-
sistent with the findings in Stock and Watson (2007). The persistence of the
“inflation gap”—the difference between actual inflation and its trend—appears
to have declined in recent years.
Methodologically, Cogley et al. introduce a new measure of persistence that
is related to the predictability of near-term movements in the variable of interest.
Formally, persistence is calibrated by the R2of the j-step-ahead forecast of the
23
variable. The higher is the R2, the more predictable it is, and thus the more
persistent, precisely because past shocks have a persistent influence on future
inflation. They examine the R2s for 1-, 4- and 8-quarter-ahead forecasts.
They find economically and statistically significant changes in the j-step-
ahead R2s for the inflation gap from 1960 to 2006, with the 1-quarter ahead
R2peaking at over 90 percent in the 1970s and early 1980s, falling to about 50
percent by the mid-1980s and through the end of their sample. The 4-quarter
ahead R2s peak at 50 to 75 percent during the Great Inflation, and decline
to about 15 percent more recently; the 8-quarter ahead R2s peak at 20 to 35
percent in the same period, falling to 10 percent more recently. All of these
changes appear to be quite significant statistically, judged by the estimated
joint posterior distribution of the R2s in earlier and later periods.
2.4 International evidence of changing reduced-form per-
sistence
A number of authors have developed empirical evidence on changes in the
reduced-form persistence of inflation for samples of developed countries. Be-
nati (2008) surveys the evidence from a broad array of developed countries over
long samples. His empirical work focuses on differences in estimated persistence
across different monetary regimes, and his key hypothesis is that regimes that
clearly anchor inflation (or the price level, as in the gold standard) induce less
persistence in the inflation rate. Benati examines a number of European coun-
tries pre- and post-EMU, the United Kingdom, Canada and Australia pre- and
post-inflation targeting, and the United States pre- and post-Volcker disinfla-
tion.
A key finding in the paper, summarized in table 5 reproduced from Benati
(2008), is that reduced-form persistence has declined in recent years for all of
the aforementioned countries that have adopted an inflation-targeting regime.
The stand-out is the United States, which has not formally adopted such a mon-
etary regime.25 The conclusion from Benati’s results is that all of the inflation-
targeting countries exhibit a marked decline in inflation persistence during their
inflation-targeting period, and that lower persistence is statistically quite dif-
ferent from the persistence exhibited prior to inflation targeting. The United
25See Benati (2008), tables I–VIII.
24
Table 5: Estimates of reduced-form inflation persistence (from Benati 2008)
Country Early sample Late sample Test of difference
U.K. Bretton Woods to inflation targeting Inflation targeting
(RPI) 0.95 -0.07 0.00
Canada 1971 to inflation targeting Inflation targeting
(CPI) 0.90 -0.33 0.00
Euro area Bretton Woods to EMU EMU
(GDP defl.) 1.01 0.35 0.00
U.S. Great Inflation post-Volcker
(CPI) 0.77 0.49 0.046
(PCE) 0.74 0.81 0.59
States and Japan (not displayed in the table) are stand-outs, and they are also
the countries that have not adopted a formal inflation-targeting regime. Be-
nati also reports similar results using the degree of “indexation” in a structural
New-Keynesian model of the economy, that is, the estimates of µin equation 3.7
below.26
Levin and Piger (2004) also examine inflation persistence across a number of
countries, focusing on the possibility that the reduced-form process for inflation
has changed in recent years, perhaps owing to changes in the central bank’s
inflation objective. Their results, which employ unknown breakpoint methods
in both the classical and Bayesian traditions, show that simply allowing for a
change in the mean of inflation appears to reduce estimated reduced-form per-
sistence for many of the countries in their sample. Other international evidence
develops mixed results about changing inflation persistence. Ravenna (2000)
documents a large post-1990 drop in Canadian inflation persistence. O’Reilly
and Whelan (2005) employ methods very similar to those in Levin and Piger,
but find that for the Euro-area price indexes (as compared with Levin and
Piger’s individual-country price indexes) there has been no discernible change
in inflation persistence.
2.5 Conclusions from the reduced-form evidence
From both theoretical and empirical perspectives, it seems likely that the con-
tribution to inflation from its unit root component has diminished significantly
26Benati (2009) explores the stability of parameters that reflect intrinsic persistence across
developed countries. In general, he finds that these parameters, whether explicit or implicit,
are not stable across changes in monetary regime.
25
in recent decades. In most macroeconomic models, inflation would contain an
important unit root (in terms of contribution to variance) if the central bank
were not acting to keep inflation low and stable, consistent with either an im-
plicit or explicit inflation target. From a practical perspective, this suggests that
most macroeconomists can think of inflation as a stationary series that will (in
normal times) return to the central bank’s inflation goal in finite time, and that
the central bank’s inflation goal, while not written in stone (or anywhere else at
present in the United States) is unlikely to vary significantly over time. Minor
time-variation in the inflation goal could add a small unit root component to
inflation, but its contribution to the variance of inflation will likely be small.
With regard to the specific autocorrelation properties of a stationary infla-
tion rate, the picture is considerably murkier. All authors agree that in the
United States and many other developed countries, inflation exhibited consid-
erable persistence from the 1960s through the mid-1980s. After that time, the
statistical evidence is mixed. For both the United States and other countries,
studies fall on both sides of the argument about the possibility of declining
reduced-form persistence. On a methodological note, for the United States, the
evidence on changing persistence from so-called “core” measures appears to dif-
fer substantially from the evidence from so-called “headline” or total inflation
measures. As a rule, the evidence of a change in persistence from core measures
is less compelling than the evidence from headline measures.
Weighing all of the evidence, it seems reasonable to conclude that the per-
sistence of inflation has declined somewhat in recent years. But how much it
has declined and whether in the extreme case inflation is now a non-persistent
series, remain issues for further study. At the time of writing, we are in the
midst of an economic environment characterized by large relative price swings
and significant changes in common estimates of the output gap and marginal
cost, factors that are commonly thought to influence inflation. These condi-
tions should provide data that will help economists to test a number of aspects
of inflation dynamics, including its persistence.
3 Structural sources of persistence
While establishing the degree of reduced-form persistence in inflation is an im-
portant first step, knowledge of the degree of reduced-form persistence is of
26
limited use to a policymaker unless she can understand the underlying sources
of reduced-form persistence. As she contemplates potential policy actions, the
policymaker must be able to determine whether or not the persistence is struc-
tural and thus may be taken as a stable feature of the economic landscape. In
order to know this, she must be able to parse the sources of persistence into
three types: (1) those generated by the driving process, (2) those that are a
part of the inflation process “intrinsic” to inflation (that is, persistence that is
imparted to inflation independent of the driving process), and (3) those that are
induced by her own actions or communications. With respect to the last source,
the research cited above suggests that central banks that are more explicit about
their inflation goal—and act in accordance with that commitment—may enjoy
less persistence in their nations’ inflation rates.
Disentangling these sources of inflation persistence is extraordinarily difficult
in relatively short aggregate time series. The paper will return to this issue
later. To begin, it is important to distinguish theoretically among the potential
sources of persistence in inflation. Significant differences in theoretical models
will imply somewhat different ways of dissecting inflation persistence. We begin
with the most widely used model of price-setting.
3.1 Persistence in the Calvo/Rotemberg model
Many modern models of inflation derive from the seminal contributions of Calvo
(1983) and Rotemberg (1982,1983) and typically imply an Euler equation for
inflation πtthat takes the form
πt=βEtπt+1 +γxt+t.(3.1)
Etdenotes the mathematical expectation using information available in period
t, and βdenotes the discount rate. The variable xtrepresents a measure of the
output gap or marginal cost, depending on details of the model. The role of the
shock term, denoted t, will be explored in greater detail below. The parameter
γis a function of the underlying frequency of price adjustment and the discount
factor.27 By iterating expectations forward and assuming that the expectation
27As Gal´ı and Gertler (1999) demonstrate, γ=(1−θ)(1−βθ)
θ. The more frequent is price
adjustment, the larger is θ, and the smaller is the coefficient on the driving process.
27
at time tof future shocks t+i= 0, equation 3.1 can be expressed as
πt=γ
∞
X
i=0
βiEtxt+i+t.(3.2)
As this rendering makes clear, inflation is the sum of two components, the dis-
counted sum of expected marginal cost (say) and a shock that is by assumption
iid, but that can in principle be serially correlated. This formulation clarifies
the motivation for this inflation specification: In a Calvo world in which prices
are expected to be fixed for some time, price-setters who can re-set their prices
set them equal to the discounted average of marginal cost that is expected to
prevail over the expected life of the contract price.
3.2 The analytics of inflation persistence: “inherited” and
“intrinsic” persistence
The expression above for inflation affords a natural taxonomy of the sources of
inflation’s persistence. First, equation 3.2 implies that the inflation rate directly
“inherits” the persistence in the variable xt. If the output gap is a persistent
series in the sense defined in section 2 above, then other things equal, inflation
will inherit some of that persistence.28
3.2.1 The baseline case
In the simplest case, inflation is given by equation 3.1 above, and the process
for xtis a univariate first-order autoregression with autoregressive parameter ρ:
πt=βEtπt+1 +γxt
xt=ρxt−1+ut
V ar(ut) = σu.
(3.3)
For simplicity, no shock perturbs the inflation Euler equation (t= 0∀t). In this
case, one can show that the autocorrelation function for inflation is29
Ai=ρi.
That is, inflation inherits exactly the autocorrelations of the first-order autore-
gressive process describing xt. In this simple version of the model, the effects of
28The reasons for persistence in output gap or marginal cost series, that is, the source of so-
called “real rigidities,” is the subject of a number of papers. See, for example, Blanchard and
Gal´ı (2007), Fuhrer (2000), Smets and Wouters (2003). Section 3.2.6 provides some empirical
results on the persistence of widely used driving processes for the canonical inflation models.
29See Fuhrer (2006) for derivations.
28
monetary policy, real rigidities in consumption, or real wages—in short, the be-
havior of inflation arising from any aspect of the economy—must enter through
their effects on xt.30
3.2.2 More complex cases
The autocorrelations of inflation become more complex when one allows for
•Non-zero shocks to the Euler equation (t6= 0)
•Variation in the size of γ, given non-zero shocks, and
•The possibility of some “backward-looking” element to inflation, as in
Fuhrer and Moore (1995) or Christiano, Eichenbaum, and Evans (2005).
These added complexities make the identification of underlying sources of per-
sistence correspondingly complex, both theoretically and empirically. The fol-
lowing subsections derive the analytical results for these cases.
3.2.3 Shocks to the Euler equation
The augmentation of the Euler equation with an iid shock changes the interpre-
tation of inflation persistence quite significantly, and in a way that is not well
recognized in much of the literature on this subject. Modifying equations 3.3 to
include this disturbance
πt=βEtπt+1 +γxt+t
xt=ρxt−1+ut
V ar(et, ut) = Σ ≡σ2
e0
0σ2
u(3.4)
makes a subtle but important difference in the autocorrelation function for in-
flation. The presence of the iid shock tnow makes the behavior of inflation a
mixture of its inherited persistence from current and expected future xt, with
weight γ, and the nonpersistent shock process (with implicit weight of one).
The larger is the variance of the shock process, the more inflation looks like
white noise, with zero persistence. The smaller is γ, the smaller will be the
importance of xtin determining the autocorrelation of inflation.
30Models in which price changes are state- rather than time-dependent allow for inherited
persistence as well. Dotsey, King, and Wolman (1999) and Burstein (2006) examine cases in
which variations in the size and persistence of money shocks result in more or less persistent
inflation responses to monetary shocks.
29
Normalizing the variance of the shock to the xtprocess to 1 for convenience,
one can express the inflation autocorrelations in this case as
Ai=ρiγ2
aσ2
e+γ2
a= (1−ρ2)(1 −ρβ)2.(3.5)
Note that equation 3.5 indicates that the autocorrelations decay at rate ρ(the
expression is premultiplied by ρi, and this is the only expression that varies
with i). More generally, the expression suggests that inflation will be more
autocorrelated
•The higher is ρ—that is, the greater is the persistence of the real driving
variable xt;
•The higher is γ—the larger is the coefficient on the driving variable xt,
and thus the more of xt’s persistence is inherited by inflation;
•The smaller is the variance of the shock etthat disturbs inflation from the
Euler equation.31
Table 6 below provides the first autocorrelation of inflation for various values
of the key parameters in the simple inflation model.32 For values of γthat
correspond to those estimated in the literature (generally below 0.1), a modest
relative variance for tcan imply a fairly low first autocorrelation. For example,
if γis estimated to be 0.05, and the variances of the two shocks are the same,
the first autocorrelation is 0.44. As the autocorrelation of the driving process
approaches 1, the persistence of the driving process begins to dominate the
white noise of the error term, as shown in the bottom panels of the table. Thus
even in this very simple model, it is clear that a persistent driving process need
not impart any persistence to inflation, depending on the sizes of γan σe.
3.2.4 The pivotal role of the coefficient on xt
As the analytical results of the preceding subsection suggest, the influence of
xton inflation—the size of the parameter γ—is pivotal both in interpreting
the sources of inflation persistence and in identifying equation 3.4 as a Phillips
curve or aggregate supply relation. If γ= 0, then (a) the Euler equation can no
31Because we have normalized σuto 1, this should be interpreted as the smaller is σerelative
to σu.
32The full derivation for the results in the table is provided in Fuhrer (2006).
30
Table 6: Value of A1for selected values of σ2
eand γ
γ
σ2
e.01 .03 .05 .1 .2
ρ= 0.9
0 0.90 0.90 0.90 0.90 0.90
0.1 0.25 0.70 0.81 0.88 0.89
0.3 0.10 0.48 0.68 0.83 0.88
0.5 0.06 0.36 0.59 0.79 0.87
1 0.03 0.23 0.44 0.71 0.84
3 0.01 0.09 0.22 0.50 0.75
5 0.01 0.06 0.14 0.39 0.68
ρ= 0.95
.5 0.29 0.76 0.87 0.93 0.94
3 0.06 0.37 0.61 0.83 0.92
ρ= 0.99
0.5 0.91 0.98 0.99 0.99 0.99
3 0.65 0.93 0.97 0.98 0.99
longer be interpreted as a Phillips curve; (b) the equation becomes decoupled
from marginal cost or the output gap, and thus from monetary policy; (c)
in many models, this decoupling will lead to indeterminacy for inflation, as
monetary policy can no longer determine the steady-state value of inflation;
and (d) inflation no longer inherits any persistence from xt.33
Given the centrality of the parameter γ, it is of interest to determine how
well identified this parameter is in the data. The answer varies from study
to study, but a brief empirical exercise may help to illuminate the potential
problems. Consider generalized method of moments (GMM) estimates of an
Euler equation that follows the format of Gal´ı and Gertler (1999) in allowing
for “rule of thumb” price-setters in addition to the Calvo price setters:
πt=λbπt−1+λfEtπt+1 +γxt+t(3.6)
The parameters λband λfcalibrate the amount of backward- and forward-
looking price-setting behavior that influences inflation. Following Gal´ı and
Gertler (1999), we employ an instrument set that consists of four lags each
of inflation, real marginal cost, a measure of the output gap, wage inflation, the
33To see point (c), consider the simplified model in equations 2.8, and its solution in equa-
tion 2.9. When the Phillips parameter a→0, the solution for πtbecomes πt=πt−1. Inflation
is a random walk, and thus fails to converge to any value in particular. This logic transfers
to more complex specifications with explicit expectations.
31
spread of the 10-year Treasury constant maturity rate over the federal funds
rate, and oil prices. Table 7 summarizes the results.
Only when allowing for 12th-order correlation in the weighting matrix does the
Table 7: Estimates of parameters in equation 3.4
Sample period: 1960–1997
GMM estimates
(HAC Standard errors in parentheses)
Number of terms
in weight matrix λbλfγ
ma=4 - 0.99 -0.00067
- (0.012) (0.0060)
0.34 0.65 0.0042
(0.044) (0.045) (0.0042)
ma = 12 - 0.99 0.0020
- (0.0088) (0.0046)
0.36 0.63 0.0065∗
(0.024) (0.025) (0.0033)
ML estimates
(BHHH Standard errors in parentheses)
λbλfγ
0.51 0.48 0.0047
(0.027) (0.031) (0.0028)
Bayesian estimates
(Max. of posterior, estimated posterior sd’s in parentheses)
λbλfγ
0.51 0.48 0.0184
(0.044 ) (0.044 ) (0.014)
coefficient on marginal cost enter with the correct sign and significantly at the
5 percent level. These results are provided as suggestive of the difficulties in
identifying γ; they are broadly consistent with the aggregated results found in
Gal´ı and Gertler (1999) and Rudd and Whelan (2006).34
For comparison, the lower panel of the table provides maximum likelihood
and Bayesian estimates of the same model, augmenting equation 3.6 with vector
autoregressive equations for the variables employed as instruments above.35
34See Mavroeidis (2005) for a careful treatment of the difficulties in identifying New-
Keynesian Phillips curves.
35The Bayesian priors for the three parameters are conventional, with generalized beta
densities for the λ’s, and a gamma density for γ. The prior distributions for the three param-
32
The VAR coefficients are taken as fixed from OLS estimates over the same
sample, and the ML estimates of the λ’s (the weights on lagged and expected
inflation in the hybrid model) and γare presented in the table, along with
BHHH standard errors. Once again, the estimates of γare quite small and not
significantly different from zero. Note that the ML estimate of the backward-
looking component is somewhat larger than the GMM estimate, and is quite
precisely estimated. The likelihood-ratio test for the restriction that λband λf
take the GMM values (with γfreely estimated) has p-value 0.0000. As we will
see in the section on “trend inflation” models below (section 3.5), a pattern is
emerging in which more tightly constrained models provide larger estimates of
intrinsic inflation persistence.
3.2.5 Hybrid models of inflation and “intrinsic” persistence: Includ-
ing lagged inflation
The debate over the empirical success of the basic specification summarized
in equation 3.1 continues, with the ability of the specification to replicate the
reduced-form persistence of inflation an important focus. A number of au-
thors have proposed rationales for the presence of a lagged inflation term in
their aggregate supply relation (aka intrinsic persistence), through indexation
of price contracts (see Christiano, Eichenbaum, and Evans (2005)),“rule-of-
thumb” behavior (see Gal´ı and Gertler (1999)), alternative contract assump-
tions (see Fuhrer and Moore (1995)), alterations to the Calvo framework that
assume a rising, rather than a constant hazard for the ability to reset prices
(see Mash (2004) and Sheedy (2007)), or alternatives to rational expectations
(see section 3.7, Orphanides and Williams (2004) and Roberts (1997)). Wood-
ford (2007) provides a very helpful summary of the state of modeling intrinsic
inflation persistence.36
An augmented Phillips curve specification that allows for the influence of
eters are centered on [0.5,0.5,0.05], respectively, with standard deviations of [0.2,0.2,0.02],
respectively. The posterior distributions are estimated using a Markov-Chain Monte Carlo
algorithm, with four simulation blocks of 200,000 draws each.
36Disaggregated price data provide limited support for many of these theoretical rationales
for intrinsic persistence. The data examined in Bils and Klenow (2004) and Nakamura and
Steinsson (2008), for example, provide little evidence that prices rise at a roughly constant
rate between more significant resets, as might be implied by firms following a rule of thumb,
or indexation.
33
lagged inflation takes the form
πt=µπt−1+ (β−µ)Etπt+1 +γxt+t,(3.7)
with the rest of the specification as detailed in equation 3.4. In a sense that
is central to this paper, the presence of lagged inflation provides an augmented
channel for what might be called “intrinsic” inflation persistence, that is, per-
sistence that is not inherited from the driving process xt. In the simpler model
of equation 3.4, the iid shock tprovided a trivial source of intrinsic persistence;
depending on the relative variance of that shock and the coefficient on xt, the
persistence inherent in xtwould be more or less inherited by inflation.
With a lag of inflation added in equation 3.7, any shock to the Euler equa-
tion will persist for longer, other things equal, independent of the evolution of
the driving variable. A shock to inflation becomes part of the history of infla-
tion, independent of shocks to xt. In addition, the forward-looking component
of the model incorporates the direct dependence on history, augmenting the
direct effect on persistence. One can think of the model as comprising both
rule-of-thumb and sophisticated forward-looking price-setters. The forward-
looking price-setters, in forming an expectation for future inflation, must take
into account the behavior of the rule-of-thumb price setters, who set current
prices based on lagged inflation. Thus the sophisticated price-setters’ behavior
reinforces the behavior of the rule-of-thumb price-setters.
To see this algebraically and graphically, consider the analytic expression for
the autocorrelations of the model augmented with a lag:
A1=a
bσ2
e−cρµd+λs,(3.8)
where [a, b, c, d] are functions of the stable root λs(in turn a function of β
and µ) and the other underlying structural parameters of the model. Fuhrer
(2006) shows that the additive term in λsdominates A1, and that λsand thus
A1rise monotonically with µ. Figure 8 plots the stable root and the first
autocorrelation of inflation as a function of µ.37 In generating this figure, we
set β= 0.98, γ= 0.05, ρ= 0.9, and σ2
e= 1. The figure shows that the first
autocorrelation rises rapidly from about 0.4 to above 0.9 as µincreases from 0
37Taken from Fuhrer (2006), Figure 2. Note that in this case, while the algebra is a bit
messier, it can also be shown that the autocorrelations decay at rate ρ, so ρand the first
autocorrelation are sufficient statistics for the autocorrelation function.
34
Figure 8: Dependence of stable root and first autocorrelation on µ
35
to 0.6. This figure thus emphasizes the role that forward-looking behavior plays
in augmenting the direct effect of lagged inflation in the model.
Table 8 provides the first autocorrelation for this model for a variety of
parameter settings. In particular, the size of µvaries from 0.1 to 0.9, and the
relative variance of varies from 0 to 5. In addition, the table displays the
sensitivity of A1to the value of ρ.
Table 8: Value of A1for selected values of σ2
eand µ
γ=0.05, β=0.98, ρ=0.9
µ
σ2
e0.0 0.1 0.3 0.5 0.7 0.9
0 0.90 0.92 0.96 0.99 1.00 1.00
0.3 0.68 0.74 0.86 0.96 0.98 0.99
0.5 0.59 0.66 0.82 0.94 0.97 0.98
1 0.44 0.53 0.74 0.92 0.97 0.98
2 0.29 0.39 0.64 0.89 0.96 0.98
3 0.22 0.32 0.59 0.88 0.96 0.98
5 0.14 0.25 0.54 0.86 0.96 0.98
ρ= 0.5
µ
σ2
e0.0 0.1 0.3 0.5 0.7 0.9
0 0.5 0.58 0.76 0.94 0.99 0.99
0.3 0.02 0.13 0.44 0.84 0.96 0.98
0.5 0.012 0.12 0.43 0.84 0.96 0.98
1 0.0063 0.12 0.43 0.84 0.96 0.98
2 0.0032 0.11 0.42 0.83 0.96 0.98
3 0.0021 0.11 0.42 0.83 0.96 0.98
5 0.0013 0.11 0.42 0.83 0.96 0.98
As shown in the top panel of the table, for relatively high values of σ2
ethat
significantly lower the first autocorrelation in the purely forward-looking model
(the first column of the table, µ= 0), a modest lag coefficient dramatically raises
the persistence of inflation. The lower panel of the table suggests that even with
very little inherited persistence—ρ= 0.5—a moderate value of µimplies a high
degree of inflation persistence. An important implication of this table is that
it may be difficult to distinguish among sources of inflation persistence, as the
reduced-form implications of an inflation process that inherits a highly persistent
driving process can be nearly the same as those of an inflation process that
inherits little persistence from the driving process, but has a modest amount of
36
intrinsic persistence.
3.2.6 The persistence of the driving process
Most researchers will agree that the observed persistence of inflation is deter-
mined at least in part by the inherited persistence of the driving process. Thus,
in thinking about potential structural causes of a change in reduced-form persis-
tence, a natural (but so far largely unexplored) question is to what extent there
have been changes in the persistence of the driving process.38 In this section,
we employ many of the same persistence measures that are used above for in-
flation. We consider three candidates for the driving process: a measure of real
marginal cost, proxied by the labor share (or equivalently, real unit labor costs)
of the nonfarm business sector; and two measures of the output gap, the first
a Hodrick-Prescott detrended log GDP gap, and the second the log deviation
between real GDP and the Congressional Budget Office’s estimate of poten-
tial GDP. We look at several subsamples. The first autocorrelation, the sum
of the autoregressive coefficients, and the dominant root of the autoregressive
process are displayed for each driving process and each subsample. The results
are summarized in table 9. As the table indicates, there is remarkable stability
Table 9: Estimated persistence of driving variables
Driving variable 1966–2008 66–83 84–08 95–08
First autocorrelation
Real mc 0.92 0.80 0.92 0.90
HP gap 0.85 0.86 0.84 0.80
CBO gap 0.92 0.91 0.92 0.90
Sum of the autoregressive coefficients
Real mc 0.94 0.78 0.96 0.94
HP gap 0.78 0.77 0.80 0.80
CBO gap 0.91 0.90 0.95 0.97
Dominant Root of the autoregressive process
Real mc 0.95 0.69 0.97 0.96
HP gap 0.81 0.82 0.79 0.79
CBO gap 0.78 0.78 0.82 0.87
in persistence measures across subsamples in all three proxies for the driving
38As demonstrated above, a change in the coefficient on the driving process or in the relative
variances of the inflation shock and the shock to the driving process may also affect the
persistence of inflation.
37
process, for all three measures of persistence. This table suggests little evidence
of a change in persistence for the driving variables most commonly associated
with inflation.39 It also suggests that the stronger hypothesis that inflation has
lost all its autocorrelation—that is, inflation is an iid time series—is hard to
justify. Unless the driving process is utterly decoupled from inflation, inflation
must inherit some of its persistence. If inflation were completely decoupled from
output or marginal cost, we would need an entirely new theory of inflation that
steps outside the historical tradition of Phillips, Gordon, Calvo, and Rotemberg.
This de-coupling between the (still ambiguous) evidence of declining reduced-
form inflation persistence and a relatively stable and persistent driving process
may help guide the search for structural interpretations of possibly changing in-
flation persistence. One simple interpretation is that the evidence for changes in
reduced-form persistence is weak, and the stable, high persistence of the driving
variables is consistent with that observation. Another is that while the inherited
persistence from the driving process may not have changed much, the impor-
tance of the lagged inflation term in Phillips curves has diminished, leading to
diminished intrinsic persistence in the face of unchanged inherited persistence.
Of course, it may also be that the Phillips slope parameter and relative error
variances have changed, which will affect the extent to which the unchanged
persistence in the driving process is inherited by inflation. The next section
examines a number of these possibilities from the perspective of a standard
dynamic stochastic general equilibrium (DSGE) model.
3.3 Using a DSGE model to interpret structural sources
of persistence
The skeletal model summarized in equations 3.4 and 3.7 provides important in-
sights into some of the structural sources of inflation persistence. However, the
model leaves implicit the determination of real output and the role of monetary
policy in influencing output and inflation. In theory, both the systematic com-
ponent of monetary policy and the nature of the transmission of policy through
the real side can have significant effects on the dynamic properties of inflation.
39Of course, a number of authors have suggested that this standard proxy for real marginal
cost is imperfect, and others have derived model-based measures of the output gap that can
differ significantly from the simple measures used here. The results presented above are
suggestive, and further research is warranted.
38
In this section, we explore the quantitative effects on inflation of various aspects
of an articulated dynamic stochastic general equilibrium (DSGE) model.
The model remains relatively simple. It comprises the “hybrid” inflation
model discussed above; an optimizing I–S curve that links real output to ex-
pected short-term real interest rates, allowing for real rigidity in the form of
a lagged output term that can be motivated by the presence of habits in the
consumer’s utility function;40 and a canonical policy or Taylor (1993) rule that
makes the short-term policy interest rate a function of deviations of inflation
and output from their desired levels. The last also allows for the possiblity of
interest-rate smoothing.
The model can be summarized in the three equations41
πt=µπt−1+ (1 −µ)Etπt+1 +γ˜yt+t
˜yt=µy˜yt−1+ (1 −µy)Et˜yt+1 −yρ(rt−Etπt+1) + ut
rt=ρrt−1+ (1 −ρ)[aππt+ay˜yt].(3.9)
The two shocks to the system, tand ut, are assumed to be independent and iid
with diagonal covariance matrix σ2
0
0σ2
uThe baseline parameters for the
model are displayed in the second column of table 10.
Table 10: Baseline and alternative parameter sets for DSGE model
Parameter Baseline value Alternate value
µ0.50000 0
πy0.10000 0.025
µy0.50000 0
yρ0.10000
ρ0.80000 0
aπ1.50000 5
ay0.50000
σ2
0.5 0.1
σ2
u0.5
We vary the values of the key parameters in the model to gauge the effect of
changes in the behavior of monetary policy, as well as changes in the price-setting
40See, for example, Fuhrer (2000).
41While the model affords a more structural decomposition of inflation persistence than
is possible using the model in equations 3.4 and 3.7, it still abstracts from some potentially
important influences on inflation dynamics. The model does not allow for “trend inflation”;
see section 3.5 below. The model uses the output gap, rather than the more current marginal
cost measure, and thus ignores the role of wages and productivity in the inflation process.
39
and output sectors, on the autocorrelations of inflation. The goal is to determine
the extent to which the persistence of inflation—whether fixed or changing over
time—can plausibly be attributed to specific underlying structural features, or
to changes in those features.
Figure 9 displays inflation’s autocorrelation function for a variety of param-
eter configurations. The autocorrelation function corresponding to the baseline
parameters in table 10 is displayed in the solid line in both panels. It mirrors
the properties of the autocorrelation functions displayed in figure 7: The au-
tocorrelations are high for the first several quarters, decaying gradually toward
zero and turning negative for several quarters thereafter.
The top panel displays the autocorrelation function when the parameters
governing the central bank’s behavior are altered. A dramatic shift in the
emphasis on inflation, aπ—from the conventional 1.5 to 5.0—reduces the au-
tocorrelations of inflation noticeably (the dotted line in the figure).42 The first
autocorrelation of inflation is reduced from about 0.65 to 0.5, and subsequent
autocorrelations drop a bit more rapidly below zero. But the difference aris-
ing from more aggressive inflation-fighting is not dramatic, especially given the
threefold increase in the policy response to inflation. Thus, while a dramatic
change in monetary policy might well account for some of the reduction in
reduced-form inflation persistence, one would not want to overstate this struc-
tural source of changes in inflation persistence.
A somewhat larger change is induced by removing interest-rate smoothing
in the policy rule (ρ= 0, the dashed line in the top panel of the figure.) Now
the autocorrelations decay smoothly toward zero after about six quarters. The
first autocorrelation of inflation is above that of the baseline case, but the effect
on inflation of shocks after about five quarters is essentially zero. This suggests
that a less inertial central bank could have removed some of the longer-lived,
second-order oscillations that characterized inflation in the 1960s and 1970s.
The bottom panel of the figure displays inflation autocorrelations when other
aspects of the model economy are altered from the baseline. The most strik-
ing change occurs when the backward-looking components of the inflation and
output equations are eliminated, as shown in the dotted line. With purely
forward-looking output and inflation equations (µ=µy= 0), the inflation au-
42A similar result, not shown, is obtained for a large shift in the emphasis on output in the
policy rule.
40
Figure 9: Effect of key structural parameters on inflation persistence, DSGE
model
41
tocorrelations jump quickly to zero (reminiscent of the disinflation simulations
in section 1.2), hovering near zero for all periods after the first. The results
are almost identical if we only shut off the backward-looking component of the
inflation equation (µ= 0), as shown in the dashed line in the bottom panel.
Smaller effects are achieved by reducing the size of the output effect on inflation
(πy= 0.025), and smaller still is the effect of reducing the relative size of the
inflation shock (“small inflation shock”), the circles in the bottom panel.
Overall, these simulations suggest that while all of the aspects of the econ-
omy captured in the stylized model likely contribute to the persistence of in-
flation, the most potent effects arise from the lag coefficients in the inflation
equation. More inflation-responsive monetary policy has some effect, as does
the slope of the Phillips curve (πy), but these effects are small compared to
the effect from eliminating the intrinsic persistence in the Phillips curve. While
these conclusions may well vary somewhat depending on details of the DSGE
specification, the results suggest that it may be inappropriate to attribute too
much persistence—or too much of the change in persistence—to the behavior of
monetary policy. The largest effects may be attributed to the “hybrid” portion
of the aggregate supply equation.
3.4 Persistence in state-dependent models of inflation
Most of the models in this paper employ a time-dependent pricing convention,
that is, the probability that a firm will adjust its price is solely a function of
time, not of economic conditions. An important and appealing alternative is
that the time for adjusting prices is endogenously chosen by firms in response
to economic conditions. In general, the literature on state-dependent pricing
(SDP) has focused relatively little attention on the issue of inflation persistence.
While the early models of Caplin and Leahy (1991) and Dotsey, King, and
Wolman (1999) focused on the search for “persistence mechanisms,” that focus
centered largely on the difficulty in developing a non-zero and persistent output
response to a one-time monetary shock. The intuition behind this difficulty
is straightforward. In a classic (S,s) model of SDP motivated by a fixed cost
to changing prices, modest to large monetary shocks could push all firms to
their (S,s) boundary, and consequently all prices would adjust one-for-one with
money: Money is neutral. For smaller monetary shocks, a smaller fraction of
42
firms would adjust prices, so money could have a small aggregate effect. Thus, it
can be difficult to avoid monetary neutrality in these models; persistent effects
of monetary shocks on prices or inflation are even harder to come by.
Burstein (2006) suggests a variant of the SDP paradigm in which firms choose
a price path, rather than a fixed price level, when they hit their (S,s) boundary.
Equivalently, the firm faces a fixed cost to adjusting its price path rather than
the level of prices.43 Thus, a monetary shock that forces a firm to its boundary
will lead the firm to set a sequence of price changes that in turn implies a per-
sistent response of inflation and output to a monetary shock. Burstein’s model
is capable of producing inflation responses to changes in the money growth rate
that exhibit the hump shape typically found in the VAR literature.44
A recent paper by Bakhshi, Khan, and Rudolf (2007) develops a Phillips
curve from the Dotsey, King, and Wolman SDP framework, and examines the
dynamics of inflation. Interestingly, that Phillips curve includes lagged inflation
terms, reflecting the fact that optimal relative prices set in period tdepend
on optimal relative prices set in previous price vintages (see their equations
9–11). They find that the persistence of inflation implied by the SDP-based
Phillips curve is significantly lower than that implied by the time-dependent
New-Keynesian Phillips curve, largely because the persistence induced by lagged
inflation in the SDP Phillips curve is more than offset by the number of price-
setters who reset following a shock.
While recent theoretical developments in SDP pricing models such as those
in Burstein (2006) and Bakhshi et al. are promising, the empirical literature
based on state-dependent models is considerably less well developed. As a con-
sequence, there are few empirical results on inflation persistence that map well
into the reduced-form and structural persistence concepts developed in this pa-
per. Overall, it seems fair to conclude that the theoretical results suggest that
SDP models are likely to provide a less compelling explanation of reduced-form
persistence than hybrid versions of the time-dependent models.
43Calvo, Celasun, and Kumhoff (2002) develop a related, time-dependent pricing model in
which firms also choose price paths when they are allowed to reset prices.
44A fuller empirical examination of Burstein’s specification would be of interest. A key
question in this regard is how the price paths set by resetting firms are allowed to respond to
different shocks—mark-up shocks, shocks to the central bank’s inflation target, and so on. In
the model of Calvo, Celasun, and Kumhoff (2002), it can be difficult to obtain data-consistent
impulse responses to all shocks without making different assumptions about when price paths
can and cannot respond to shocks. For more on this issue, see Fuhrer (2008).
43
3.5 Persistence in models of “trend inflation”
A series of papers beginning with Cogley, Primiceri, and Sargent (2007) and
Cogley and Sbordone (2009) emphasizes the importance in modeling inflation
of recognizing the slowly moving component of inflation that they dub “trend in-
flation.” The Cogley-Sbordone paper introduces two innovations. First, because
long-run inflation is not a constant, the typical simplifications that give rise to
the log-linearized Calvo model of equation 3.1 no longer apply. The standard
log-linearization depends on the constancy of the long-run value of inflation.45
Cogley and Sbordone derive the log-linear approximation that is appropriate
when the long-run value of inflation has a trend. Hatted variables in the next
equation denote deviations from steady-state values; for inflation, this implies
the deviation of inflation from trend inflation
ˆπt= ˜ρt(ˆπt−1−ˆgπ
t)+ζtcmct+b1t˜
Etˆπt+1+b2t˜
Et
∞
X
j=2
ϕj−1
1tˆπt+j+b3t˜
Et
∞
X
j=0
ϕj
1t(ˆ
Qt+j,t+j+1+ˆgy
t+j+1)+ut
(3.10)
where ˆ
g¯π
tand ˆgyare the innovation to trend inflation and the growth rate of real
output, respectively; cmc is the deviation of real marginal cost from its steady
state; and ˆ
Qt+j,t+j+1 is the one-period discount factor between periods t+j
and t+j+ 1. A key parameter in this specification is ˜ρt, which calibrates the
degree to which a lagged inflation term is required to match the autoregressive
properties of inflation, once trend inflation is accounted for. Regardless of the
estimate of ˜ρ, this specification is useful for researchers who wish to allow for
time variation in the steady-state value of inflation (such as a time-varying
inflation target).
Second, Cogley and Sbordone find that the point estimate for ˜ρt, the coeffi-
cient on lagged inflation in this specification, centers on zero. That is, once the
model has accounted for the slow-moving variation in trend inflation, there is no
need for a lag of inflation to account for the reduced-form persistence of infla-
tion. While this empirical finding is controversial, the concept of trend inflation,
which the authors associate with the central bank’s time-varying inflation goal,
is an important contribution to the inflation literature.
45See, for example, Woodford (2003) for the derivation.
44
3.5.1 Cogley and Sbordone’s measure of trend inflation
Table 11 displays the first autocorrelation for actual and de-trended inflation,
using Cogley and Sbordone’s measure of trend inflation, for the subsamples in
table 1 of their paper.46 The table shows that the autocorrelations of detrended
Table 11: First autocorrelation of inflation
Sample Detrended πRaw Data
60:Q3–03:Q4 0.81 0.89
60:Q3–83:Q4 0.83 0.88
84:Q1–03:Q4 0.36 0.56
inflation are somewhat lower than those of the raw inflation data. However,
in contrast to the table in Cogley and Sbordone (2009), the differences are
small for the first two samples. For the most recent sample, the autocorrelation
declines both for the detrended and for the raw series. Most of the decline in the
detrended data’s autocorrelation can be explained by a corresponding decline
in that of the raw data. This suggests that there are other, equally important
factors at work in explaining the persistence of inflation, and in explaining
changes in persistence over time.
While not presented in Cogley and Sbordone (2009), it is of interest to
examine their model’s implication for the first autocorrelation of inflation. Using
values of their key parameters that center on the median estimates over time
presented in figure 4 of their paper, that is, ˜ρt≈0, ζt= 0.03, b1= 0.9, b2= 0.02,
and b3= 0, and assuming a diagonal covariance matrix with variances estimated
from a VAR over their sample period, I obtain a first autocorrelation for inflation
of 0.22.47 This estimate differs markedly from their data’s first autocorrelation
over the same sample, 1960–2003. Matching the autocorrelation of inflation
from the data requires a significantly different set of parameters—for example,
46Inflation is measured as four times the log change in the GDP deflator, as in Cogley and
Sbordone (2009). The measure of trend inflation was kindly provided by the authors and
replicates that in Figure 1 of their paper.
47Note that setting b1to 1.05 as in their figure 4 implies multiple solutions. I reduce the
value of b1to 0.9 to keep it as high as possible, while still obtaining a unique solution to the
model. The VAR employed in the exercise includes four lags each of the inflation rate from the
GDP deflator, real marginal cost defined as in Cogley-Sbordone, the federal funds rate, and
an output gap defined as the log difference between real GDP and Hodrick-Prescott filtered
real GDP. The VAR equations for marginal cost, the funds rate, and the output gap are used
in conjunction with equation 3.10 in order to compute stability conditions, and to compute
autocorrelations given the estimated covariance matrix of the shocks.
45
setting ˜ρnear unity raises the first autocorrelation toward 0.8.
A recent paper by Barnes, Gumbau-Brisa, Lie, and Olivei (2009) examines
the robustness of the finding that the detrended inflation model implies a value
for ˜ρtof 0. Details of Cogley and Sbordone’s estimation procedure make a
significant difference to the estimation results. Barnes et al. show that simply
changing the form of the Euler equation—which implicitly imposes an additional
constraint that is implied by the model—completely reverses the finding on ˜ρt.
They develop a precise estimate of this key “intrinsic persistence” parameter
of about 0.8. This finding suggests caution in interpreting the rather striking
implications of trend inflation models for inflation persistence.
3.6 Persistence in sticky-information models
Mankiw and Reis (2002) propose a model in which information, rather than
prices, is sticky. In essence, the model applies the Calvo machinery to the
updating of information, rather than that of prices. In the Mankiw and Reis
(2002) model, a fraction of price setters get to update their information in each
period in a manner analogous to that of the Calvo model. As a consequence, the
age or vintage of price-setters’ information sets will be described by a geometric
distribution, as is the case for the duration of price contracts in the Calvo setting.
The model implies a Phillips curve that links inflation to output and a ge-
ometric weighted average of lagged expectations of inflation and output (see
Mankiw and Reis (2002), p. 1300):
πt=αλ
1−λyt+λ
∞
X
j=0
(1 −λ)jEt−1−j(πt+α∆yt).(3.11)
As the authors emphasize, in this Phillips curve it is past expectations of current
conditions, rather than current expectations of future conditions, that determine
inflation. The Mankiw-Reis Phillips curve is thus a close cousin to those of Fis-
cher (1977) and Koenig (1996). The Mankiw-Reis paper documents a number
of desirable features of their model, based on simulations in response to a per-
manent change in the level of demand, the growth rate of demand (a stylized
“disinflation”), and an anticipated drop in the growth rate of demand.
In this model, one can see by inspection (and the authors verify) that infla-
tion will inherit the persistence of the output process. The authors compute the
46
autocorrelations of inflation under the assumption that a simple quantity equa-
tion holds, trivially linking money to output, and that money growth follows an
autoregressive process
mt=pt+yt
∆mt= 0.5∆mt−1+εt.(3.12)
Under these assumptions, the autocorrelations for inflation are indeed quite
high. Figure 10 displays the autocorrelations taken from table I of their paper
(the solid line). The dashed line in the figure displays the autocorrelations
Figure 10: Autocorrelations of inflation, Mankiw-Reis model
when one allows for supply shocks (shocks to the Phillips curve). As discussed
in section 3.2.3 above, the autocorrelation in the presence of supply shocks will
depend on the variance of supply shocks relative to the shocks to the driving
process. Here, we have set the ratio of supply shocks to money growth shocks
at 0.25. As the figure indicates, supply shocks with relatively modest variance
dramatically alter the implications for inflation’s autocorrelations in this model,
as in the Calvo/Rotemberg model.
47
3.7 Persistence in learning models
When the agents that “populate” the models discussed above know the structure
and parameters of the model, they can use that knowledge to form expectations
by taking the mathematical expectation of the variable of interest. That is the
essence of rational expectations, and up to this point, the structural models in
this section have assumed rational expectations.
However, a long tradition dating back at least to Bray (1982) and Marcet
and Sargent (1989) examines the dynamics of macroeconomic variables when the
agents in the model lack perfect knowledge of the model and consequently must
learn about their environment. Orphanides and Williams (2004) and Williams
(2006) explore monetary economies in which agents must learn about their eco-
nomic environment. Learning can significantly alter the dynamics of an other-
wise standard macroeconomic model. Consider the simple case in which inflation
is governed by a two-equation model comprising a Calvo-like Phillips curve and
a reduced-form equation for output, as in equations 3.3 in section 3.2.1 above:
πt=Ft−1πt+1 +γxt
xt=ρxt−1+ut.(3.13)
(3.14)
The key difference is that expectations of inflation are not the mathematical
expectation, denoted by the Etoperator, but are reasonable forecasts given the
information available to the agents at time t−1, denoted by Ft−1. One plausible
way to formalize learning, posited in Williams (2006), is for agents to estimate
the reduced-form of the model using an adaptive estimation rule. The unique
and stable reduced-form solution of this model under rational expectations is
given by πt
xt=0ργ
1−ρ
0ρ πt−1
xt−1.(3.15)
As discussed earlier, the model implies no dependence on lagged inflation, al-
though inflation will indeed exhibit persistence to the extent that xtdoes. But if
agents are not endowed with knowledge of the solution coefficients under ratio-
nal expectations, they may instead attempt to estimate a reduced-form equation
for inflation and xtsuch as
πt
xt=ˆ
Atπt−1
xt−1+et.(3.16)
48
If their initial estimate for ˆ
Atcoincides with the solution in equation 3.15, then
the model will behave as under rational expectations. But in general, the model
will exhibit different dynamics, governed in part by agents’ current estimate
ˆ
Atin equation 3.16. If etis unforecastable from period t, then equation 3.16
implies that
Ft−1πt+1 ≡eπˆ
A2πt−1
xt−1,(3.17)
where eπselects the row of ˆ
Acorresponding to the inflation equation in the
reduced form. Substituting this expectation for inflation in equation 3.13, we
obtain48
πt= ˆa11 πt−1+ ˆa12 xt−1+γxt.(3.18)
Depending on the current estimated values of ˆaij , inflation may now exhibit
some intrinsic persistence, and will inherit—more or less, depending on the
value of ˆa12—the persistence of xt.49
This very stylized example makes it clear that learning can add another layer
of dynamics to inflation. If agents use a forecasting rule that differs from the
rational expectations solution, they can add intrinsic persistence to an otherwise
forward-looking model that implies no such persistence.
4 Inference about persistence in small samples:
“Anchored expectations” and their implica-
tions for inflation persistence
Implicit in the analysis of Benati (2008) and explicit in a paper by Williams
(2006) is the suggestion that inflation expectations that are “well anchored” by
the central bank’s explicit commitment to an inflation target may have altered
the persistence and, more generally, the overall dynamics of inflation. Theoret-
ically, this can, of course, be true to a degree. Consider once again the simple
model of equations 2.8, modified slightly to make the central bank’s inflation
target explicit:
πt=πt−1+axt
xt=−b(ft−¯π)
ft= ¯π+c(πt−¯π).
(4.19)
48The time subscripts on the elements of ˆ
Atare dropped for notational convenience.
49Additional dynamics would, in principle, be added as the agents’ estimates of the ˆaij
evolve over time. The specifics will vary depending on the estimation rule assumed, and a
detailed investigation of this issue lies outside the scope of this paper.
49
As long as c6= 0, the steady state for this model is given by50
πt= ¯π
xt= 0
ft= ¯π .
(4.20)
But when c= 0, the central bank does nothing to move inflation toward a
particular target, the steady-state for inflation becomes indeterminate, and the
solution for inflation becomes
πt=πt−1.(4.21)
In contrast to the stable autoregressive process in equation 2.9, the inflation
rate in this case follows a random walk. This simple model demonstrates the
importance in this class of models of keeping central bank actions consistent with
pursuing an inflation target. When central banks do so, inflation is stationary.
When they do not, it is not. In most models with explicit expectations, a similar
proposition holds, as is well known. The intuition for a model with rational
expectations is relatively straightforward. Consider a simple model based on
the Calvo (1983) specification of inflation
πt=βEtπt+1 +γxt
xt=−a(rt−Etπt+1)
rt= ¯π+b(πt−¯π).(4.22)
As demonstrated in equation 3.2 above, the solution for πtis the weighted sum
of future xt, which in turn depends on the future short-term real interest rates,
rt−Etπt+1. The central bank sets the short-term nominal interest rate rt.
Well-anchored inflation expectations require two things from the central bank.
First, the central bank must have an inflation goal that is known to the private
agents in the economy, and second, the central bank must move its policy rate
in a way that systematically pushes the inflation rate toward that goal. Put
differently, the “Taylor Principle” (Taylor 1999) operates in this model: As long
as the central bank moves the policy rate by more than the inflation rate, so
that it is increasing the short-term real rate when inflation is above its target
and conversely when it is below, then the expected path of xtwill be consistent
with returning inflation to its target under arbitrary initial conditions. In this
sense, inflation expectations will be well anchored, and inflation will have a
50The equilibrium real rate of interest is set to 0 for convenience.
50
Figure 11: Errors for Williams (2006) and random walk inflation models in
recent years
determinate solution and be stationary. As is discussed in the introduction and
in more detail below, in a purely forward-looking model such as this one, well-
anchored expectations can imply not only determinacy and stationarity, but an
inflation rate that follows a white noise process.51
Williams (2006) suggests that in recent years expectations may have become
so well anchored that inflation may be well characterized by random deviations
around a constant
πt=c+εt.
As a description of the data, this simple model does reasonably well in recent
years, as figure 11 suggests. The figure shows the errors made by a model that
forecasts the four-quarter log change in prices as equal to its sample mean,
versus one that sets the forecast equal to the previous four-quarter change.52
The random walk model has been advocated by Atkeson and Ohanian (2001) as
51This implication also relies on the assumption that all the shocks disturbing the equations
above are iid.
52Of course, the sample mean forecast uses information not available in real time to the
forecaster.
51
an alternative to poorly performing Phillips curves. The errors over Williams’s
sample are smaller for the constant-based forecast, with a root-mean-squared
error of 0.29 versus 0.37 for the random walk model.53
The diversity of results on reduced-form inflation persistence presented in
the preceding subsections together should suggest some caution in arriving at
conclusions about a possible change in inflation persistence in the past decade
or two. We examine a simple Monte Carlo exercise to highlight the difficulty in
inferring changes in inflation dynamics and their implications for persistence in
relatively short samples.
Inflation in the exercise is generated by three simple equations: A backward-
looking Phillips curve with lagged coefficients that sum to 1, an output equation
that makes output a function of its own lag and the real interest rate, and a
simple policy rule in which the policy rate responds only to the gap between
inflation and the inflation target.
πt=
4
X
i=1
aiπt−i+b˜yt+εt
˜yt=c˜yt−1−d(rt−1−πt) + ut
rt=ert−1+ (1 −e)(f(πt−¯π)) .(4.23)
The variances of the two shocks are calibrated from vector autoregressive equa-
tions estimated from 1966 to 2006, or split at 1984 to allow for changes in the
estimated variance due to the so-called Great Moderation.54 The baseline val-
ues for band dare 0.1, for c0.8, for e0.8, and for f1.5. The inflation target is
set to 2.0. Fifty thousand samples of length 40 quarters are created using draws
of the shocks to the inflation and output equations, assuming the individual
shocks are random normal draws scaled by the estimated variances of εand u.
Initial conditions are randomized for each draw, rather than using a “burn-in”
period.55
53Adding the last three years to the sample diminishes the advantage of the constant-based
model. The RMSEs for the constant-based and random-walk models for the extended sample
are 0.32 and 0.35, respectively. The constant-based model, with an estimated mean of about
1.8, consistently under-forecasts inflation over the past three years.
54The equations estimate a bivariate VAR in Hodrick-Prescott-filtered real GDP and the
annualized rate of inflation in the core consumer price index. The three samples are considered
at 1966:Q1 to 2006:Q4, 1966:Q1 to 1984:Q4, and 1985:Q1 to 2006:Q4.
55Tests for convergence of the distribution of estimated parameters suggest that 50,000
samples are more than adequate to assure convergence for this exercise.
52
Figure 12: Distribution of estimated coefficients
For each simulated sample, the following simple regression models are esti-
mated. The first is designed as a simple test that inflation exhibits no persis-
tence, perhaps because inflation expectations are well anchored at the central
bank’s target, and can thus be modeled well as white noise around a constant
c. The null for this model is that α= 0, which would replicate Williams’s
(2006) result. The second allows for the influence of output. Both are clearly
misspecified, but the issue is whether one could be misled with a relatively short
sample into believing that the properties of inflation had changed when in fact
the underlying inflation process still exhibits persistence and a correlation with
output.
πt=απt−1+c
πt=απt−1+byt−1+c . (4.24)
(4.25)
The resulting regression estimates are displayed in the summary table below
and the histograms in the panels of figure 12. For the first model, figure 12
and table 12 show that the standard error on the estimated lag coefficient is
large; the median t-statistic for the hypothesis that α= 0 has a p-value above
53
Table 12: Distribution of estimated coefficients
Model 1
Coefficient Median Std. deviation
α0.47 0.24
c1.06 0.16
Model 2
Coefficient Median Std. deviation
α0.41 0.25
b0.069 0.33
c1.2 0.17
0.05. The estimate of cimplies a mean for πtof almost exactly 2.56 One could
be forgiven for inferring from these estimates that inflation had little or no
persistence, and was well anchored around the central bank’s inflation target.57
The results of the second regression model, displayed in the bottom panel of
table 12, suggest that matters are even worse when one tests for the presence
of a Phillips-like relationship in the misspecified model for a small sample. Now
the median estimate of αis smaller still, it is imprecisely estimated (the median
p-value for the test of α= 0 is greater than 0.10), and the effect of output is
biased downward and very imprecisely estimated. Once again, the mean of the
estimated inflation target matches the true value quite precisely. One might
conclude from simple regressions such as these that inflation has little or no
persistence and that the Phillips correlation is absent. These small samples
allow one to recover the average value of inflation over the period, but the rest
of inflation dynamics may be very poorly estimated.
5 Microeconomic evidence on persistence
5.1 Persistence in micro data: U.S. evidence
Up to this point, the paper has focused on macroeconomic, aggregate evidence
bearing on inflation persistence. Yet the dominant models in the literature aim
56The standard error of the intercept, at 0.16, is quite small. The distribution of the mean
of inflation, which is a nonlinear function of αand c(¯π=α
(1−c)), implies a relatively large
standard error for the mean of 2.3.
57A good portion of the downward bias evident in the estimates in figure 12 arises from the
truncation of the lags of inflation compared with the lags in the data-generating process. Still,
in large samples, the regression should retrieve the sum of the underlying AR coefficients in
the single autoregressive coefficient. As the sample size in the exercise above increases from
10 years to 25 to 50, the median estimate of αin Model 1 increases from 0.47 to 0.70 to 0.84.
54
to provide microeconomic foundations for inflation, based on the price-setting
decisions of individual firms. In this regard, it is striking that the now large
literature that examines micro-price data has emerged relatively recently, led by
the work of Bils and Klenow (2004). Bils and Klenow employ an unpublished
dataset of about 350 CPI expenditure categories that account for about 70
percent of consumer expenditures. In examining the persistence and volatility
of inflation, they use a subset of 123 spending categories that account for about
63 percent of spending.
For each of the CPI components that they examine pi, they estimate a sim-
ple AR(1) process in order to assess the degree of persistence and volatility for
its inflation rate dpi(defined as the log change)
dpi,t =ρdpi,t−1+εi,t .
They find that the average persistence of inflation, which they define as the
arithmetic average of the ρi’s, is slightly negative (−0.05) for their shorter
sample (1995–2000) and only modestly positive (0.26) for their longer sam-
ple (1959–2000). Interestingly, for the shorter sample, they find that the degree
of persistence across price categories is positively correlated with the frequency
of price change, clearly contrary to predictions from the Calvo/Rotemberg and
Taylor models. Over the longer sample period, the correlation is positive but
not statistically different from zero.
The authors estimate that adjusting for the presence of “temporary sales”—
explicitly transitory reductions in prices that revert within days or weeks and
would likely lead to an understatement of the persistence of non-sale prices—
would have only a small impact on their estimates of persistence. Of course, the
fact that temporary sales occur regularly casts some doubt on the underlying
assumption of time-dependent pricing models, in which prices are simply fixed
for long periods. It also raises the question as to whether the relevant price
measures should exclude or include temporary sales. This issue is discussed in
more detail in Nakamura and Steinsson (2008), who employ a dataset that allows
them to study prices at a more disaggregated level. They find that temporary
sales have a larger effect on the measured frequency of price changes than in
the Bils and Klenow data; excluding temporary sales reduces the frequency of
price changes by about one-half. They do not estimate the effect of temporary
55
sales on inflation persistence in their paper.
The results from these seminal papers suggest that the micro data exhibit
behavior that is at odds with the prevailing time-dependent pricing models’
description of underlying price behavior. Nakamura and Steinsson’s paper sug-
gests that several key features are also at odds with a menu-cost model. But
whether the micro data are consistent with the estimates of persistence in ag-
gregate data is less clear, for two reasons. First, aggregate price series may
exhibit quite different properties from the individual price series. Second, what
one observes in individual price changes likely reflects the combined influence of
firm- or industry-specific shocks and macro shocks. The two points are related,
and the relative importance of the two is an empirical question, but if indi-
vidual prices respond differently to macro versus micro shocks, then it will be
important to sort out these influences in evaluating the relevance of micro-data
evidence for aggregate inflation persistence.
Section 5.3 discusses some recent work bearing on the first point. Boivin,
Giannoni, and Mihov (2009) provide results bearing on the second point. They
estimate a small number of common factors (principal components) from a large
number of macroeconomic variables. They then relate individual price changes
to these common factors in order to decompose individual inflation rates into
idiosyncratic, sector-specific fluctuations and macroeconomic fluctuations. De-
noting the matrix of common factors by Ctand the log change in the individual
price series pit as dpit, they estimate regressions
dpit =λiCt+it .
The R2of each regression indicates the fraction of individual price changes that
may be attributed to the common macroeconomic factors; one minus this R2is
the fraction of variation attributable to sector-specific sources. Their baseline
results suggest that about 85 percent of the variation in the individual price
changes may be attributed to sector-specific shocks.
Using the decomposed individual inflation series, Boivin et al. estimate
simple autoregressions for each of the series and their two components, mea-
suring persistence by the sum of the autoregressive coefficients.58 They find,
like Bils and Klenow (2004), that individual inflation series exhibit relatively
58Their data are observed at the monthly frequency, and they estimate autoregressions with
13 lags.
56
little persistence. The idiosyncratic components of the individual inflation se-
ries it exhibit essentially no persistence, while the common components (λiCt)
vary in persistence from negative for some series to above 0.95 for some health-
care components and tenant room and board. These findings imply that the
aggregate inflation measures, which are quite persistent in their data, inherit
their persistence from the common macroeconomic components of the individual
price series, particularly from those that exhibit the highest persistence. The
non-persistent idiosyncratic components essentially wash out in aggregation.
5.2 Persistence in micro data: Euro-area evidence
Altissimo, Ehrmann, and Smets (2006) examine both aggregate and disaggre-
gated data to explore the properties of Euro-area inflation. Their conclusions on
aggregate data echo those of others—inflation has been persistent; its reduced-
form persistence has declined in recent years so that it now exhibits moderate
persistence, although how moderate depends on how it is estimated; its decline
may be attributable to a stable and well-focused monetary regime that anchors
long-run inflation expectations. Their disaggregated sectoral data suggest that
individual price series exhibit less persistence on average than their correspond-
ing aggregates.
The studies of Angeloni, Aucremanne, Ehrmann, Gal´ı, Levin, and Smets
(2006) and Alvarez, Dhyne, et al. (2006) find ample evidence of infrequent price
changes at the micro level. The former estimates that while there is substantial
heterogeneity across sectors, prices on average are quite sticky, exhibiting a
four to six quarter duration. Seen through the lens of the Calvo model, these
estimates suggest that the “Calvo parameter,” which indexes the frequency of
price changes, is large, implying a small effect of marginal cost on inflation.
From equation 3.5 above, the smaller is the effect of marginal cost on inflation,
the less of the persistence in marginal cost is inherited by inflation. Neither
of these studies examines the persistence of disaggregated price series, nor do
the studies explore the complications in aggregating disaggregated series with
heterogeneous dynamics. But to the extent that they find relatively infrequent
price changes, and to the extent that one feels comfortable mapping these into
aggregate Calvo parameters, the studies imply less inherited persistence, other
things equal.
57
5.3 More on aggregation and persistence
Bils and Klenow find some difference between the persistence properties of in-
flation based on individual price series and their expenditure-share-weighted
aggregate inflation measure. For their longer sample, they estimate an autore-
gressive coefficient of 0.63, with a standard error of 0.03, in contrast to the
much smaller average of the autoregressive coefficients for the individual price
series. Boivin et al. find that the low-persistence, idiosyncratic components of
disaggregated inflation series appear to wash out in aggregate inflation mea-
sures, leaving the persistent common macroeconomic components to dominate
the persistence of aggregate inflation. These observations suggest that aggre-
gation of price series may play an important role in determining the degree of
aggregate inflation persistence.59
Several recent papers examine the role of aggregation in inflation persis-
tence. Mumtaz, Zabczyk, and Ellis (2009) employ a methodology that draws
importantly on Boivin et al., studying disaggregated price data for the United
Kingdom. They also find that the persistence of the aggregate inflation measure
is biased upwards relative to the persistence of the underlying price series and
that the bias is driven by the macro components of those series. Altissimo, Mo-
jon, and Zaffaroni (2009) delve more deeply into the aggregation process, using
existing results on aggregation of time series (see, for example, Granger (1980)
for parametric results and Zaffaroni (2004) for non-parametric results that are
more closely related to the work in Altissimo et al.). Similarly to Boivin et al.,
they assume that individual price series may be characterized by an unobserved
components model that makes each price change series dpi,t a function of its
own idiosyncratic persistence, a common shock ut, and an idiosyncratic shock
εi,t:
dpi,t =αidpi,t−1+ut+εi,t .
By assuming a particular form for the distribution of the persistence parameters
f(α), as in Granger (1980), one can derive results for the persistence of the
simple aggregate of the individual price changes dPt= (1/n)Pn
i=1 dpi,t. The
relative contributions of the common and idiosyncratic shocks will be important
59On a note of theoretical counterpoint, Carvalho (2006) develops a multisector version of
the Calvo model with heterogeneity in price stickiness that implies an aggregate inflation rate
with less persistence than that of the standard Calvo model.
58
in understanding the relationship between individual and aggregate prices, but
so too will be the differences in the way in which the common shock is propagated
in the individual prices—that is, the differences in the αi’s. Price series that
perpetuate the common shock through a larger αiwill have a larger effect on
the persistence of the aggregate than series with a smaller αi.
They find that disaggregated price changes exhibit significantly less persis-
tence than the aggregate and that the preponderance of the variance of the
individual price series is accounted for by idiosyncratic volatility, in agreement
with all of the studies cited above. Like Boivin et al., they find that a single
principal component of the individual price series dominates in explaining the
low-frequency variation in the individual price changes. It follows that this fac-
tor must account for the common persistence among the disaggregated series.
Estimation of a model like 5.3 reveals in addition that the propagation of the
common shock in individual series is indeed quite varied. The high persistence of
the common shock in services prices, in particular, combined with the relatively
high weight of services in the aggregate price index, accounts for a substantial
part of the persistence in aggregate inflation.
6 Conclusions
It may be early to draw firm conclusions about the structural sources of inflation
persistence, or about the extent to which these sources have changed and man-
ifested themselves in changes in reduced-form inflation persistence. In the first
case, it may be premature because there is not yet widespread agreement about
the appropriate mapping between micro data or reduced-form aggregate data
and our structural models. In the second case, we have a fairly short sample
from which to draw inferences about potential changes (see section 4 for more
details).
Still, the research to date allows one to draw some conclusions. First, to
the extent that reduced-form persistence has changed, policymakers need to
gain clarity about the source of the change. This paper discusses a number of
structural channels through which persistence may have changed. There may
have been a change in the “intrinsic” persistence in inflation—the importance
of lagged inflation in the structural Phillips curves. Alternatively, the amount
of inherited persistence may have changed. In principle, this could arise because
59
the persistence of the driving process has changed, or because the coefficient on
the driving process has changed, or because the relative variances of the shocks
to inflation and the driving process have changed.
The analysis in this paper suggests that it is unlikely that any change in
persistence has arisen from a change in the persistence of the driving process,
as this has remained remarkably stable throughout the period. In addition,
a DSGE model-based analysis suggests that while changes in the systematic
component of monetary policy likely have led to less-persistent inflation, the
largest changes in persistence are most likely due to changes in the so-called
intrinsic sources of inflation persistence—whether those arise from indexation,
rule-of-thumb price-setters, or a rising price reset hazard. Finally, the models
that depart from the standard Calvo framework suggest that other aspects of the
economy that impinge upon inflation persistence may be responsible for changes
in inflation persistence. These may include smaller or less-frequent changes in
“trend inflation” or a smaller role for learning, as central bank transparency
about its goals has increased.
Second, we have now accumulated an impressive and growing body of evi-
dence on the behavior of price- (and wage-)setting at the disaggregated level.
This evidence strongly suggests that some of the inferences drawn from micro
data about the frequency of price changes, as well as the degree of inflation
persistence, may pertain largely to price responses to industry- or firm-specific
shocks. The response to aggregate shocks by the aggregate component common
to the individual price series may well have quite different properties from the re-
sponses of individual firms to idiosyncratic shocks. Integrating this evidence into
our structural models, perhaps along the lines of models of “rational inatten-
tion” (see Sims (2003), Gorodnichenko (2008) and Ma´ckowiak and Wiederholt
(2009)) seems a promising avenue for research.
Finally, we are currently accumulating additional evidence that should allow
us to take a firmer stance on whether reduced-form persistence has changed, and
to discern the structural sources of any such changes. The upheaval created by
the 2007–2009 financial crisis and recession, with the concomitant prospect of a
prolonged period of elevated unemployment, suggest that over the next decade
we will have accumulated evidence that will allow us to test more fully the hy-
pothesis of a decline in reduced-form inflation persistence and to test competing
60
theories that attribute the structural sources of persistence.
References
Altissimo, Filippo, Benoit Mojon, and Paolo Zaffaroni (2009). “Can
aggregation explain the persistence of inflation?” Journal of Monetary
Economics 56: 231–241.
Altissimo, Filippo, Michael Ehrmann, and Frank Smets (2006). “Inflation
persistence and price-setting behaviour in the Euro area: A summary of the
IPN evidence.” National Bank of Belgium working papers No. 95, October.
Alvarez, Luis, E. Dhyne, M. Hoeberichts, C. Kwapil, H. Le Bihan, P.
Lunnemann, F. Martins, R. Sabbatini, H. Stahl, P. Vermeulen, and J.
Vilmu, (2006). “Sticky prices in the euro area: a summary of new
micro-evidence.” Journal of the European Economic Association 4:
575–584.
Angeloni, Ignazio, Luc Aucremanne, Michael Ehrmann, Jordi Gal´ı, Andrew
Levin, and Frank Smets (2006). “New evidence on inflation persistence and
price stickiness in the Euro area: Implications for macro modeling.”
Journal of the European Economic Association,4: 562–574.
Atkeson, Andrew and Lee Ohanian (2001). “Are Phillips curves useful for
forecasting inflation?” Federal Reserve Bank of Minneapolis Quarterly
Review 25, No. 1: 2–11, Winter.
Bakhshi, Hasan, Hashmat Khan, and Barbara Rudolf (2007). “The Phillips
curve under state-dependent pricing.” Journal of Monetary Economics 54:
2321–2345.
Ball, Laurence (1994). “Credible disinflation with staggered price-setting.”
American Economic Review 84(1): 282–289, March.
Ball, Laurence, and Stephen G. Cecchetti (1990). “Inflation and uncertainty at
short and long horizons.” Brookings Papers on Economic Activity: 215–54.
Barnes, Michelle, Fabi`a Gumbau-Brisa, Denny Lie, and Giovanni Olivei
(2009). “Closed-form estimates of the New Keynesian Phillips curve with
time-varying trend inflation.” Working paper, Federal Reserve Bank of
Boston.
Barsky, Robert B. (1987). “The Fisher hypothesis and the forecastibility and
persistence of inflation.” Journal of Monetary Economics 19: 3–24.
Benati, Luca (2008). “Investigating inflation persistence across monetary
regimes.” Quarterly Journal of Economics 123(3): 1005–1060, August.
Benati, Luca (2009). “Are intrinsic inflation persistence models structural in
the sense of Lucas (1976)?” European Central Bank Working paper series
No. 1038, March.
Bils, M. and P. Klenow (2004). “Some evidence on the importance of sticky
prices.” Journal of Political Economy 112(5).
61
Blanchard, Olivier and Jordi Gal´ı (2007). “Real wage rigidities and the new
Keynesian model.” Journal of Money, Credit and Banking, Supplement:
Conference on “Quantitative Evidence on Price Determination,” September
29 and 30, 2005, co-sponsored by the JMCB and the Federal Reserve
Board, 39(1): 35–65.
Boivin, Jean, Marc P. Giannoni and Ilian Mihov (2009). “Sticky prices and
monetary policy: Evidence from disaggregated U.S. data.” American
Economic Review 99(1): 350–384.
Bray, Margaret (1982). “Learning, estimation, and the stability of rational
expectations.” Journal of Economic Theory 26: 318–39, April.
Burstein, Ariel T. (2006). “Inflation and output dynamics with
state-dependent pricing decisions.” Journal of Monetary Economics 53:
1235–1257.
Calvo, Guillermo (1983). “Staggered prices in a utility-maximizing
framework.” Journal of Monetary Economics 12(3).
Calvo, G.A., Celasun, O., and M. Kumhof (2002). “A theory of rational
inflationary inertia.” in: P. Aghion, R. Frydman, J. Stiglitz, and M.
Woodford, eds., Knowledge, Information and Expectations in Modern
Macroeconomics: In Honor of Edmund S. Phelps. Princeton: Princeton
University Press.
Caplin, Andrew and John Leahy (1991). “State-dependent pricing and the
dynamics of money and output.” Quarterly Journal of Economics 106(3):
683–708.
Carvalho, Carlos (2006). “Heterogeneity in price stickiness and the real effects
of monetary shocks.” Frontiers of Macroeconomics Volume 2, Issue 1,
Article 1
Christiano, Lawrence, Martin Eichenbaum, and Charles Evans (2005).
“Nominal rigidities and the dynamic effects of a shock to monetary policy.”
Journal of Political Economy.
Cogley, T. and A. Sbordone (2009). “Trend inflation, indexation, and inflation
persistence in New Keynesian Philips curve.” American Economic Review.
Cogley,Timothy, Giorgio E. Primiceri, and Thomas J. Sargent (2007).
“Inflation-gap persistence in the U.S.” Working paper, December.
Dornbusch, Rudiger (1976). “Expectations and exchange rate dynamics.”
Journal of Political Economy 84(6): 1161–1176, December.
Dotsey, Michael, Robert King, and Alexander Wolman (1999).
“State-dependent pricing and the general equilibrium dynamics of money
and output.” Quarterly Journal of Economics,114: 655–690, May.
Estrella, Arturo and Jeffrey Fuhrer (2002). “Dynamic inconsistencies:
Counterfactual implications of a class of rational expectations models.”
American Economic Review 92(4): 1013–1028.
Fischer, Stanley (1977). “Long-term contracts, rational expectations, and the
optimal money supply rule.” Journal of Political Economy 85(1): 191–205.
Friedman, Milton (1968). “The role of monetary policy.” American Economic
Review 58(1): 1–17.
62
Fuhrer, Jeffrey (2000). “Habit formation in consumption and its implications
for monetary-policy models.” American Economic Review 90(3): 367–390.
Fuhrer, Jeffrey (2006). “Intrinsic and inherited inflation persistence.”
International Journal of Central Banking 2(3): 49–86.
Fuhrer, Jeffrey (2008). “Special issue comment on optimal price setting and
inflation inertia in a rational expectations model.” Journal of Economic
Dynamics and Control 32: 2536–2542.
Fuhrer, Jeffrey and George Moore (1992). “Monetary policy rules and the
indicator properties of asset prices.” Journal of Monetary Economics 29:
303–336.
Fuhrer, Jeffrey and George Moore (1995). “Inflation persistence.” Quarterly
Journal of Economics 110(1): 127–159, February.
Gal´ı, Jordi and Mark Gertler (1999). “Inflation dynamics: A structural
econometric analysis.” Journal of Monetary Economics 44: 195–222.
Gordon, Robert J. (1982). “Price inertia and policy ineffectiveness in the
United States, 1890–1980.” Journal of Political Economy 90(6): 1087–1117.
Gordon, Robert J., Stephen R. King, and Franco Modigliani (1982). “The
output cost of disinflation in traditional and vector autoregressive models.”
Brookings Papers on Economic Activity 1982(1): 205–244.
Gorodnichenko, Yuri (2008). “Endogenous information, menu costs and
inflation persistence.” NBER Working Paper 14183.
Granger, Clive (1980). “Long memory relationships and the aggregation of
dynamic models.” Journal of Econometrics 14: 227–238.
Gray, Jo Anna (1977). “Wage indexation: A macroeconomic approach.
Journal of Monetary Economics 2(2): 221–235.
Koenig, Evan (1996). “Aggregate price adjustment: The Fischerian
alternative.” Federal Reserve Bank of Dallas Working Paper No. 9615.
Levin, Andrew and Jeremy Piger (2004). “Is inflation persistence intrinsic in
industrial economies?” European Central Bank working paper No. 334,
April.
Levin, A. and J. Williams (2003). “Robust monetary policy with competing
reference models.” Journal of Monetary Economics 50(5).
Lucas, Robert E. Jr. (1972). “Expectations and the neutrality of money.”
Journal of Economic Theory 4: 103–124.
Mackowiak, Bartosz and Mirko Wiederholt (2009). “Optimal sticky prices
under rational inattention.” American Economic Review 99(3): 769–803.
Mankiw, N.G. and R. Reis (2002). “Sticky information versus sticky prices: A
proposal to replace the New Keynesian Phillips curve.” Quarterly Journal
of Economics 117(4).
Marcet, Albert and Thomas J. Sargent (1989). “Convergence of least-squares
learning in environments with hidden state variables and private
information.” Journal of Political Economy 97(6): 1306–1322, December.
Mash, Richard, (2004). “Optimizing microfoundations for inflation
persistence.” Oxford University Dept. of Economics Discussion Paper No.
183, January.
63
Mavroeidis, Sophocles (2005). “Identification issues in forward-looking models
estimated by GMM, with an application to the Phillips curve.” Journal of
Money, Credit, and Banking 37(3): 421–448, June.
Mumtaz, Haroon, Pawel Zabczyk and Colin Ellis (2009). “What lies beneath:
What can disaggregated data tell us about the behaviour of prices?” Bank
of England Working Paper No. 364, March.
Muth, John (1961). “Rational expectations and the theory of price
movements.” Econometrica 29(3): 315–335.
Nakamura, Emi and Jon Steinsson (2008). “Five facts about prices: A
reevaluation of menu cost models.” Quarterly Journal of Economics 112:
1415–64, November.
Okun, Arthur M. (1977). “Efficient disinflationary policies.” American
Economic Review 68 (May 1978, Papers and Proceedings (1977)): 348–52.
O’Reilly, G. and K. Whelan (2005). “Has Euro-area inflation persistence
changed over time?” Review of Economics and Statistics 87(4): 709–720.
Orphanides, A. and J.C. Williams (2004). “Imperfect knowledge, inflation
expectations, and monetary policy.” in B. Bernanke and M. Woodford,
eds., The Inflation Targeting Debate.
Phelps, Edmund S. (1968). “Money-wage dynamics and labor-market
equilibrium.” Journal of Political Economy 76(4, Part 2): 678–711.
Pivetta, Frederic and Ricardo Reis (2007). “The persistence of inflation in the
United States.” Journal of Economic Dynamics and Control 31:
1326–1358.
Ravenna, F. (2000). “The impact of inflation targeting in Canada: A
structural analysis.” manuscript, New York University.
Roberts, J. (2006). “Monetary policy and inflation dynamics.” International
Journal of Central Banking 2.
Roberts, J. (1997). “Is inflation sticky?” Journal of Monetary Economics
39(2).
Rotemberg, Julio J. (1982). “Sticky prices in the United States.” Journal of
Political Economy 90(6): 1187–1211.
Rotemberg, Julio J. (1983). “Aggregate consequences of fixed costs of price
adjustment.” American Economic Review 73(3): 433–436.
Rudd, J. and K. Whelan (2006). “Can rational expectations sticky-price
models explain inflation dynamics?” American Economic Review 96.
Sargent, Thomas J. and Neil Wallace (1975). “’Rational’ expectations, the
optimal monetary instrument, and the optimal money supply rule.” The
Journal of Political Economy 83(2): 241–254.
Sheedy, Kevin D. (2007). “Intrinsic inflation persistence.” Centre for
Economic Performance Discussion Paper No. 837, November.
Sims, Christopher A. (2003). “Implications of rational inattention.” Journal of
Monetary Economics 50: 665–690.
Smets, F., and R. Wouters (2003). “Monetary policy in an estimated
stochastic dynamic general equilibrium model of the Euro area.” Journal of
European Economic Association 1(5).
64
Smets, F. and R. Wouters (2003). “An estimated dynamic stochastic general
equilibrium model of the Euro area.” Journal of European Economic
Association 1(5).
Stock, J.H. and M. Watson (2007). “Has inflation become harder to forecast?
Journal of Money, Credit, and Banking 39: 3–34.
Taylor, John (1980). “Aggregate dynamics and staggered contracts.” Journal
of Political Economy 88.
Taylor, John (1993). “Discretion versus policy rules in practice.”
Carnegie-Rochester Conference Series on Public Policy 39: 195–214.
Taylor, John (1999). “A historical analysis of monetary policy rules.” in John
B. Taylor, ed., Monetary Policy Rules. Chicago: University of Chicago
Press: 319–41.
Williams, John (2006). “The Phillips curve in an era of well-anchored inflation
expectations.” Federal Reserve Bank of San Francisco, unpublished
working paper, September.
Woodford, Michael (2003). Interest and Prices: Foundations of a Theory of
Monetary Policy. Princeton: Princeton University Press.
Woodford, Michael (2007). “Interpreting inflation persistence: Comments on
the conference on Quantitative Evidence on Price Determination.” Journal
of Money, Credit and Banking, Supplement to 39(1): 203–210, February.
Zaffaroni, Paolo (2004). “Contemporaneous aggregation of linear dynamic
models in large economies.” Journal of Econometrics 120: 75–102.
65
