Monetary Policy Regime Shifts

and Inflation Persistence

Troy Davig and Taeyoung Doh
January 2009
RWP 08-16

Electronic copy available at: http://ssrn.com/abstract=1335347

Monetary Policy Regime Shifts and

Inflation Persistence
Troy Davig and Taeyoung Doh
December 2008
RWP 08-16
Abstract
This paper reports the results of estimating a Markov-Switching New
Keynesian (MSNK) model using Bayesian methods. The broadest and best fitting
MSNK model is a four-regime model allowing independent changes in the regimes
governing monetary policy and the volatility of the shocks. We use the estimates to
investigate the mechanisms that lead to a decline in the persistence of inflation. We
show that the population moment describing the serial correlation of inflation is a
weighted average of the autocorrelation parameters of the exogenous shocks. Changes
in the monetary or shock volatility regimes shift weight over these serial correlation
parameters and affect the serial correlation properties of inflation. Estimation results
indicate that a shift to a monetary regime that reacts more aggressively to inflation
reduces the weight on the more persistent shocks, so lowers inflation persistence.
Similarly, a shift to the low-volatility regime reduces the weight on the more persistent
shocks and also contributes to reducing inflation persistence. Estimates of modelimplied inflation persistence indicate that it began rising in the late 1960s and peaked
around the Volcker disinflation. The subsequent decline in persistence is due to both a
more aggressive monetary policy regime and less volatile shocks.
Keywords: Inflation persistence, Bayesian estimation, Markov-switching, DSGE
models
JEL Classification: C11, E31, E52

Federal Reserve Bank of Kansas City, 1 Memorial Drive, Kansas City, MO 64198. E-mail:
Troy.Davig@kc.frb.org and Taeyoung.Doh@kc.frb.org. We thank Chris Otrok and Ferre
DeGraeve for very helpful comments. We also thank Marco Del Negro, Jesus FernandezVillaverde, John Geweke, Alejandro Justiniano, Thomas Lubik, Jim Nason, Giorgio Primiceri,
Frank Schorfheide, and Dan Waggoner, including all participants in the Workshop on Methods
and Applications for DSGE Models hosted by the Federal Reserve Bank of Cleveland. The
views expressed herein are those of the authors and do not necessarily represent those of the
Federal Reserve Bank of Kansas City or the Federal Reserve System.

Electronic copy available at: http://ssrn.com/abstract=1335347

Introduction

An important issue for monetary policy is understanding the factors that contribute
to persistence in deviations of inflation from its underlying trend. Several studies address the empirical aspects of U.S. inflation persistence and how it may have changed
over time.1 Motivated by this literature, we asses the role of changes in either monetary policy or shock volatility on inflation persistence in the context of a small-scale
Markov-switching New Keynesian (MSNK) model. We show analytically that a shift
to a monetary regime that responds more aggressively to inflation or a shift to a
regime with less volatile shocks reduces the serial correlation in inflation. To infer
the empirical relevance of these factors in explaining U.S. data, we estimate variants
of the MSNK model using Bayesian methods. The broadest model is a four-regime
model that allows independent regime changes between two monetary regimes and
two shock volatility regimes. Estimation indicates that the combined effects of the
more aggressive monetary regime and low shock volatility regime, which is the configuration during the 1960s and often after 1994, produces an empirically relevant
decline in inflation persistence.
Specifically, we estimate three versions of the MSNK model, with each allowing
different aspects of the model to change regime. First, we estimate a MSNK model
with a switching monetary policy rule that allows policy to shift between active and
passive stances.2 The second MSNK model has a fixed monetary rule, but allows
the volatilities of the shocks to change. Estimates show that the timing of the low
volatility regime corresponds closely to the timing of the active monetary regime
from the model with switching in only in the monetary policy rule. This suggest that
the estimation procedure exploits switching in whatever parameters are available to
capture the high volatility of the 1970s. Also, the coincident timing of regimes across
models suggests that changes in policy and volatility are competing mechanisms in
explaining the large shifts in inflation dynamics in the 1970s. The final model has four
regimes with independent switching between two monetary policy regimes and two
shock volatility regimes. Incorporating independent regime shifts in these elements
provides a framework for untangling the source of the changes in inflation persistence
- that is, whether it is primarily driven by monetary policy or changes in shock
volatility.
Overall, U.S. data favors the four-regime MSNK model. Estimates from this model
indicate that monetary policy was active in responding to inflation throughout the
1

Some recent examples include Cogley and Sargent (2001), Stock (2001), Gadzinski and Orlandi
(2004), Cecchetti and Debelle (2006), Clark (2006), Levin and Piger (2006), Cogley, Primiceri, and
Sargent (2007), Pivetta and Reis (2007), Stock and Watson (2007) and Carlstrom, Fuerst, and
Paustian (2008).
2
Following the language in Leeper (1991), active (passive) monetary policy refers to a policy that
adjusts the nominal interest rate more (less) than one-for-one with movements in inflation.

latter half of the 1950s and most of the 1960s. Beginning in the late 1960s, policy
reverted to passively responding to inflation that lasted throughout the 1970s and
early 1980s. This period coincides with the rise in inflation persistence, which we
illustrate is a manifestation of a passive monetary policy. The Volcker disinflation
period is identified with a period of high volatility, instead of a shift to a policy
regime that responds systematically more aggressively to inflation. Active policy is
in place briefly in the mid 1980s, but then turns to a more persistent active stance
beginning in the mid 1990s. The low-volatility regime is in place throughout most
of the post-1984 period, or Great Moderation era, but is also in place during the
1960s. Also, high-volatility regimes are often associated with recessions as defined by
National Bureau of Economic Research (NBER).
Focusing on the four-regime MSNK model, we investigate how the different monetary and volatility regimes affect inflation persistence. We show that the population
moment describing the serial correlation of inflation is a weighted average of the autocorrelation parameters of the exogenous shocks, which include a technology, monetary
policy and markup shock. The weight on the autocorrelation parameter for each shock
is a function of the monetary policy coefficient and shock volatilities, both of which
depend on the current regime. Changes in regimes then reshuffle weights over the
autocorrelation parameters and alter the serial correlation properties of inflation. A
shift to a monetary regime that more aggressively fights inflation or a shift to a lowvolatility regime reduces the weight on the more persistent shock, which reduces the
serial correlation of inflation. Estimates from the four regime model indicate that a
shift to an active policy or low-volatility regime produces a modest decline in inflation persistence. However, the combined affect of an active policy and low-volatility
regime, a configuration in place after approximately 1994, produces a more substantial decline in persistence that is consistent with empirical work assessing the decline
in U.S. inflation persistence.
One common alternative to the MSNK approach is to exogenously split samples
and use a more conventional fixed-regime framework. The MSNK model, however,
has advantages over a fixed-regime approach with respect to letting the data speak
regarding regime changes and overall logical coherence. In structural estimation of
DSGE models, the informational assumption underlying exogenous sample splitting
is undesirable. For example, the econometrician is assumed to know precisely when
regime changes occur, yet agents behave as if the regimes in place are permanent.3
Also, an information asymmetry arises in that private agents know all the parameter
values, which are objects unknown to the econometrician. The MSNK model resolves
the issues pertaining to informational asymmetry and logical inconsistency, since the
econometrician must infer both the timing of regimes and structural parameters.
Also, private agents incorporate the possibility of future regime change into their
3

See Cooley, LeRoy, and Raymon (1984) for elaboration on the logical issues involved in modelling
regime changes in rational expectations models.

expectation formation.
The general issue of the inflation persistence has generated a large empirical literature. Recently, Cecchetti and Debelle (2006), Clark (2006) and Levin and Piger
(2006) conclude that once allowing for shifts in the mean rate of inflation, persistence
has not declined much over the past few decades. Pivetta and Reis (2007) report
statistical measures indicating that inflation persistence is approximately unchanged
in the U.S. using a sample beginning in 1947. Cogley, Primiceri, and Sargent (2007)
provide evidence that the inflation-gap, measured as the difference between actual inflation and trend inflation, has become less persistent in recent decades. These papers
focus on measurement, where only Cogley, Primiceri, and Sargent (2007) interpret
their findings within a medium-scale DSGE framework. A benefit of using a DSGE
model to address changes in inflation persistence is the ability to disentangle whether
any shifts in inflation persistence are a result of changes in monetary policy or the processes driving the shocks. Our approach is similar to Cogley, Primiceri, and Sargent
(2007), except we use a smaller-scale Markov-switching rational expectations model,
but with the benefit of allowing the data to indicate when shifts occur. Our paper
is also related with the recent literature on estimating MSNK models with switching
coefficients such as Bianchi (2008), Bikbov (2006), and Eo (2008). However, these
papers do not address the link between shifts in policy or volatility and changes in
inflation persistence.

A Markov-Switching New Keynesian Model

This section presents a Markov-Switching New Keynesian model with a relatively

standard private sector specification, following closely the setups in Ireland (2004) and
An and Schorfheide (2007). The primary difference relative to these specifications is
that the parameters in the monetary rule and shock volatilities are subject to regime
shifts.
The basic elements of the model economy include a representative household, a
representative firm that produces a final good and a continuum of monopolistically
competitive firms that each produce an intermediate good indexed by j [0, 1].

2.1

Households

The representative household chooses {Ct , Nt , Bt }

t=0 to maximize lifetime utility
!

X
(Ct /At )1
Ht ,
Et
t
1
t=0
4

where Ct denotes consumption of a composite good, Ht are hours worked, At is a

measure of technology, (0, 1) is the discount factor and > 0 is the coefficient of
relative risk aversion.4 Utility maximization is subject to the intertemporal budget
constraint
Pt Ct + Qt Bt = Bt1 + Wt Ht + Pt Dt Pt Tt ,
where Bt are nominal bond holdings, Dt are real profits from ownership of firms, Tt
are lump-sum taxes, Pt is the aggregate price level, Wt is the nominal wage and Qt
is the inverse of the gross nominal interest rate.

2.2

Firms

Intermediate goods-producing firm j produces output, yjt , according to

yjt = At njt ,
where At is an exogenous measure of productivity that is the same across firms and
njt is the labor input hired by firm j. The labor market is perfectly competitive and
firms are able to hire as much as demanded at the real wage.
The monopolistic intermediate goods-producing firms pay a cost of adjusting their
price, given by

2
pjt

1 Yt ,
(1)
acjt =
2 pjt1
where 0 determines the magnitude of the price adjustment cost, denotes the
central banks inflation target and pjt denotes the nominal price set by firm j [0, 1].
The price adjustment cost is in terms of the final good Yt . Each intermediate goodsproducing firms maximizes the expected present value of profits,
Et

s t+s

s=0

djt+s
,
Pt+s

(2)

where

 
1
Ct+s
At
t+s
Ct
At+s
is the representative households stochastic discount factor and djt are nominal profits
of firm j at time t. Real profits are

2
djt
pjt

pjt
=
yjt t yjt
1 Yt ,
(3)
Pt
Pt
2 pjt1


As we discuss below, technology follows a non-stationary process and induces a stochastic trend
in consumption. Detrending Ct by At is convenient because the model has a well-defined steady
state in terms of detrended variables. Also, an alternative interpretation of At is as a measure of
external habit stock.

where t denotes real marginal cost, where t = (Wt /Pt ) /At .

There is a representative final-goods producing firm that purchases the intermediate inputs at nominal prices pjt and combines them into a final good using the
following constant-returns-to-scale technology
1

Z
Yt =

yt (j)

t 1
t

t
 1
t

dj

(4)

where t > 1 t is the elasticity of substitution between goods. Variations in t

translate into shocks to the desired markup, which is the actual markup in the absence
of price adjustment costs. The steady state markup is
u=

,
1

(5)

and is the steady state elasticity of substitution.

The profit-maximization problem for the final-goods producing firm yields a demand for each intermediate good given by

yjt =

pjt
Pt

t
Yt ,

(6)

where pjt is the nominal price of good j. The zero-profit condition for the final
1
hR
i 1
1 1
goods-producing firm implies Pt 0 pjt dj
is the aggregate price level.

2.3

Policy

The monetary authority sets the short-term nominal rate using the following rule

Rt = r

(st ) 

Yt
At y

(st )
exp(et ),

(7)

where Rt is the gross nominal interest rate, t = Pt /Pt1 , is the target rate of
inflation, r is the steady state real rate, y is the steady state level of the detrended
output and the regime, st , is a discrete-valued random variable that follows a two-state
Markov chain,


p11
1 p11
P1 =
,
(8)
1 p22
p22
where pii = Pr [st = i|st1 = i]. The active, or more aggressive, regime corresponds
to st = 1 and the less active regime, or possibly passive regime, corresponds to st = 2.
This labeling implies (2) < (1).
6

The assumption of a constant inflation target may seem at odds with the empirical
literature that stresses the importance of allowing for mean shifts when measuring
inflation persistence, such as Cecchetti and Debelle (2006), Clark (2006) and Levin
and Piger (2006). Schorfheide (2005) also focuses on a policy rule with a shifting
inflation target using a similar DSGE model. The rationale then for imposing a
constant mean, but shifting reaction coefficients, is to give monetary policy a potential
mechanism to affect inflation persistence.5 A shifting inflation target in the policy rule
of this DSGE model does not affect the model implied serial correlation of inflation,
so is an inadequate framework to address the issue of how changes in monetary policy
affect inflation persistence.
The potential exists, however, that allowing for changes in trend inflation may
better capture changes in U.S. inflation dynamics than shifts in the reaction coefficients in the policy rule or volatilities. To capture this possibility and compare the
results to the MSNK model, we consider an alternative specification that allows trend
inflation to following a driftless random walk,
t = t1 + t ,

(9)

where t N (0, 2 ).6

We assume the fiscal authority passively adjusts lump-sum taxes to satisfy the
governments flow budget constraint and transversality condition on government debt.

2.4

Exogenous Shock Processes

Aggregate productivity follows

ln At = + ln At1 + ln at ,

(10)

ln at = a ln at1 + at ,

(11)

where
with at N (0, a2 (rt )) and |a | < 1 for rt {1, 2}. The process for productivity
imposes that it grows at an average rate of , but is subject to serially correlated
shocks that have varying degrees of volatility depending on the regime.
Shocks to the markup and monetary policy rule follow
ln ut = (1 u ) ln u + u ln ut1 + ut ,
ln et = e ln et1 + et ,
5
6

A subsequent section demonstrates how changes in reaction coefficients affect persistence.

In this version, we do not allow regime changes.

(12)
(13)

where ut N (0, u2 (rt )), et N (0, e2 (rt )), |u | < 1 and |e | < 1 for rt {1, 2}.
The regime governing the volatility of the shock process, rt , also follows a two-state
Markov chain,


q11
1 q11
P2 =
,
(14)
1 q22
q22
where qii = Pr [rt = i|rt1 = i].
In the four-regime MSNK model, the shock volatility regime, rt , is independent
from the monetary regime, st .7 As a result, the shock volatility regime can change
without requiring a change in the monetary regime. This approach allows the data to
indicate whether, say, a period of high inflation volatility is more likely the cause of
monetary policy or higher exogenous shock volatility. Ideally, the volatilties for each
shock could change regime according to their own independent Markov chains. However, the estimation of the four-regime MSNK model is computationally demanding,
so estimation with higher-dimensional chains is left for future research.

2.5

Symmetric Equilibrium

In a symmetric equilibrium, each intermediate goods-producing firm faces the same

marginal cost, so makes the same pricing and production decisions. In equilibrium, we
can then eliminate the j subscripts, yielding yjt = Yt , pjt = Pt , njt = Nt , acjt = ACt
and djt = Dt . In equilibrium, the first-order condition for the firms pricing decision
is
 

t
t
1
(15)
0 = (1 t ) t + t t t t





t+1 Yt+1
t+1
1
.
+Et t+1

Yt
The households first-order conditions are
 
1 Ct
Wt
1 =
,
At At
Pt
"
#
 


Ct
Rt
Ct+1
At
= Et
.
At
t+1 At+1
At+1

(16)
(17)

In addition, the aggregate resource constraint must hold in equilibrium,

Yt = Ct + ACt ,
7

The transition matrix in this case is P1

P2 .

(18)

and also,
Ht = Nt ,
where Nt =

R1
0

(19)

njt dj.

In the case of fully flexible prices, (15) and (16) imply aggregate output is
1

Yt = At ut ,
where ut is the desired markup and is given by ut =

3
3.1

(20)
t
.
t 1

The Conditionally Linear MSNK Model

Equilibrium Relations and Shock Processes

The log-linearized private sector relations (15) and (17) are



bt Et
ybt = Et ybt+1 1 R
bt+1 Etb
at+1 ,

bt = Et
bt+1 + (b
yt + u
bt ) ,

(21)
(22)

where ybt ln (yt /y ) is a measure of the output gap, yt = (Yt /At ) ,

bt = ln (t /)
b
and Rt = ln (Rt /R), where R = r. Conditioning on a given regime, the monetary
rule and shock processes are linear, given by
bt
R
b
at
u
bt
ebt

=
=
=
=

(st )
bt + (st ) ybt + et ,
a b
at1 + at ,
u u
bt1 + ut ,
e ebt1 + et .

(23)
(24)
(25)
(26)

Equations (21) (26) represent the full MSNK model.

As in Davig and Leeper (2007), we compute solutions using the method of undetermined coefficients on the minimum set of state variables (MSV). Solutions have
regime-dependent coefficients as follows

bt (st )
A (st ) B (st ) C (st )
b
at
ybt (st ) = Ay (st ) By (st ) Cy (st ) u
bt .
(27)
bt (st )
AR (st ) BR (st ) CR (st )
ebt
R
since we use a first-order approximation to the equilibrium conditions of households
and firms, the solution coefficients depend only on the monetary regime and not the
shock volatility regime.
9

3.2

Determinacy Restrictions

In a purely forward-looking New Keynesian model with a fixed policy rule, a passive
policy fails to uniquely determine the equilibrium. This results in the possibility that
agents will coordinate on extraneous information, or sunspots, which generate nonfundamental macroeconomic fluctuations. The set of equations and state variables
describing the equilibrium under indeterminacy differ from when the equilibrium is
unique. The different equilibrium representations complicates estimation. One approach to account for indeterminacy in estimation is to follow Lubik and Schorfheide
(2004), where posterior weights apply to the determinate and indeterminate regions
of the parameter space.
The MSNK model is inherently nonlinear, which complicates the conditions for
determinacy. One approach is to follow Davig and Leeper (2007) who compute the
determinacy restrictions of an auxiliary representation, or stacked system, of the
MSNK model given in (21) (23).8 The Davig and Leeper (2007) determinacy restrictions on the stacked system are also the restrictions that render the equilibrium
learnable when agents form expectations by recursively updating a VAR representation of the equilibrium.9 These restrictions permit one monetary regime to be passive
without inducing indeterminacy of the stacked system, but restrict the extent of the
passive behavior. During estimation, we restrict the parameters to lie within the determinate region of the stacked system. This approach has the benefit of allowing a
passive monetary regime, but does not require taking into account the indeterminate
representation of the equilibrium.10
To illustrate how to derive the stacked system, first assume all shocks are i.i.d.
and rewrite expectations as follows
 s
 s

t ],

(28)

]
+
p
E[
]
=
p
E[
Et t+1 = E[t+1 st = i, s
i2
2t+1
i1
1t+1
t
t


 s
s
s
Et xt+1 = E[xt+1 st = i, t ] = pi1 E[x1t+1 t ] + pi2 E[x2t+1 t ],
(29)

0
at u
bt ebt
where it = t (st = i, t ), xit = xt (st = i, t ), and t = b
for i =
s
11
1, 2. The information set, t = {st1 , . . . , rt , rt1 , . . .}, excludes the current regime,
so t = s
t {st }. Distributing probability mass across the different conditional
expectations for inflation, as in (28) (29), is the same approach as in Gordon and
St-Amour (2000) and Bansal and Zhou (2002).
8

Farmer, Waggoner, and Zha (2008) present an example of sunspot solutions to the conditionally
linear system, given by (21) (23), in a region of the determinate parameter space of the stacked
system. In general, determinacy restrictions for nonlinear systems such as (21)(23) are unavailable.
9
See Branch, Davig, and McGough (2007) for details.
10
The determinate and indeterminate representations of the stacked system are tractable, making
estimation along the lines of Lubik and Schorfheide (2004) possible. However, estimation accounting
for indeterminacy is beyond the scope of this paper and left for future work.
11
Whether shocks are i.i.d. or serially correlated does not matter for determinacy, so the assumption of i.i.d. shocks is made here for convenience.

10

Next, define the forecast errors



jt+1
= jt+1 E[jt+1 s
],
 ts
x
jt+1 = xjt+1 E[xjt+1 t ],

(30)
(31)

for j = 1, 2. Substituting expectations, (28) (29), and the policy rule, (23) , into
(21) (22) yields the stacked system
AYt = BYt1 + At + Cut ,
where

1t
2t

Yt =
x1t ,
x2t

1t

2t

t =
x ,
1t
x
2t

(32)

at
ut = ut ,
et

(33)

and A, B and C are conforming matrices consisting of private sector parameters,

policy parameters and the transition matrix. The stacked system is available for any
purely forward-looking rational expectations model with regime-switching. The benefit of the stacked system is that it has constant coefficient matrices, yet captures the
impact regime changes have on expectation formation. Further, standard methods for
solving linear rational expectations systems are applicable to (32), such as Blanchard
and Kahn (1980) or Sims (2002).12
Necessary and sufficient conditions for determinacy, which is the existence of a
unique bounded solution to (32), is that all the generalized eigenvalues of (B, A) lie
inside the unit circle. The determinacy conditions are intuitive. First, the passive
monetary regime cannot be too passive, meaning the response to inflation can be less
than one, but still has to be above some minimum threshold. And second, the passive
regime cannot be too persistent, meaning that the expected duration of the regime
must be below a given threshold. The determinacy conditions are joint restrictions
over both monetary regimes, so the parameters governing the active regime affect the
determinacy restrictions over the passive regime. Therefore, the more persistent or
active the active regime is, the more persistent or passive the passive regime can be.
12

McCallum (2004) proves the equivalence between MSV solutions and determinate (i.e. unique
and non-explosive) solutions from solving a system of linear expectations difference equations for
purely forward looking models. Davig and Leeper (2007) show the equivalence between the MSV
solution and determinate solution of the stacked system in regime-switching rational expectation
models.

11

4
4.1

Estimation
Econometric Methodology

The linear structure of the model solution conditional on the current regime makes the
application of the approximate Kalman filter of Kim and Nelson (1999) feasible. Given
laws of motion for the shock processes and the minimum state variable solutions of
inflation, output, and nominal interest rate, we can write down the following regimedependent state-space representation
Zt = Az + Bz (st )xt + [1, 0, 0]0 ln At , Zt = [ln Yt , t , Rt ]0 ,
xt = xt1 + t , xt = [b
at , u
bt , ebt ]0 , t = [at , ut , et ]0 ,
ln At = + ln At1 + b
at .

(34)
(35)
(36)

xt is a vector of state variables and Zt is a vector of three observed variables consisting of per capita real GDP, inflation (log difference of GDP deflator), and 3 month
Treasury bill rate. Bz (st ) is a conformable state-dependent matrix with elements
given in (32). The sample period is from 1953:Q1 to 2006:Q4. The model has three
observables and three structural shocks, so stochastic singularity is not a problem
and therefore, there is no need to introduce measurement errors. Constructing the
likelihood for the MSNK model requires integrating out latent variables, including
the history of regimes. Kim and Nelson (1999) note that collapsing some paths of
regimes with very small probability is necessary to make the filtering algorithm operable. Otherwise, we have to consider St different paths of regimes to evaluate the
likelihood value at t, where S is the number of possible regimes. We allow 8 (64) different paths of regimes in a two (four) regime case. The likelihood for the four-regime
model is
X
p(Zt |Z t1 , ) =
p(Zt |Z t1 , , st , rt )p(st |Z t1 , )p(rt |Z t1 , ),
(37)
rt, st {1,2}

where is the vector of structural parameters and Z t1 denotes observations up to

time t 1.13 For the models with two regimes, either rt or st is constant and the
corresponding probability density collapses to unity.
Using the Bayesian approach, we combine the likelihood with a prior distribution
of . From the Bayesian perspective, the resulting posterior distribution of reflects
an update to the prior distribution using the information from the likelihood and is
a key tool for inference. Incorporating prior information on provides additional
13

We increased the number of histories that are considered and found little difference in terms of
the likelihood value. To reduce the approximation error, we eliminate or collapse different paths of
regimes based on implied probabilities of them as implemented in Schorfheide (2005). The technical
appendix describes the procedure in detail.

12

curvature for the posterior density and excludes implausible estimates of parameters
which may overfit the sample data.14 The posterior distribution of is hard to
characterize analytically, so we use a random-walk Metropolis-Hastings algorithm to
obtain the posterior draws.15 We initialize the Markov chain at the (local) mode of
the posterior density by using a numerical optimization routine (CSMINWEL provided
by Christopher Sims). The inverse of the negative hessian evaluated at the local mode
is used as the covariance matrix of the proposal density. After obtaining one million
draws from the Markov chain, we compute means and the covariance matrix and
update the covariance matrix of the proposal density. Then, we run another Markov
chain starting at the means of the previous one million draws with the updated
covariance matrix. Trace plots of parameters and tests of the equality of split sample
means confirm that the distribution of Markov Chain Monte Carlo (MCMC) output
converges to the stationary distribution.
In constructing the likelihood, we use the filtered probability for each regime to
integrate out the latent regimes. Since regimes are not directly observable to the
econometrician, we are often interested in computing the estimates of the probability
of different regimes conditional on all the observations available. This approach provides an indication of which history of regimes is most probable given the available
observations. The smoothed probability of each regime can be obtained by applying
the filtering step backwards. In the four-regime model, we compute the smoothed
probabilities as follows
p(Qt |Z T , , rt ) = P

p(Qt |Z t , )p(Qt+1 |Qt )p(Qt+1 |Z T , )

,
t
T
Qt {1,2,3,4} p(Qt |Z , )p(Qt+1 |Qt )p(Qt+1 |Z , )

(38)

where Qt is a composite four-state discrete valued random variable that describes

both the monetary and volatility regimes.16 Either rt or st replaces Qt in (38) for the
two-regime models. Since p(QT |Z T , ) and p(Qt |Z t , ) are obtained as byproducts of
the likelihood evaluation, this is relatively easy to implement.
To identify the sources of the changes in inflation persistence, we need to compare
different regime switching models. The marginal likelihood of each model provides a
coherent framework to compare non-nested models. Conceptually, it is obtained by
integrating the posterior kernel over the entire parameter space in each model Mi
Z
T
p(Z |Mi ) = p(Z T |, Mi )p(|Mi )d.
(39)
14

For further discussion of advantages of Bayesian approach in the estimation of DSGE models,
see An and Schorfheide (2007).
15
We use a mixture of normal distribution and t distribution as a proposal density. The relatively
fat-tailed t distribution makes it more likely for the proposal density to cover the tail area in the
target density and, therefore, facilitates
the convergence of MCMC chain.
N
16
The transition matrix is P1 P2 with elements given by p(Qt+1 |Qt ).

13

The practical computation of this constant is done by the numerical approximation

based on the posterior simulator as in Geweke (1999), for example.17

4.2

Prior Distribution

Table 1 provides information on the prior distribution of the parameters. If possible,

the prior means are calibrated to match the sample moments of observed variables.
For example, the prior mean of the average technology growth rate () is set to match
the average growth rate of per capita real GDP. Similarly, the prior mean of the steady
state inflation rate () in the model is set to match the average inflation rate in the
data. And the prior mean of the discount factor () is then set to match the average
nominal interest rate conditional on the prior means of and . The autocorrelation
of technology growth (a ) and the standard deviation of the technology shock (a )
are set to match the autocorrelation and the standard deviation of per capita real
GDP. Prior distributions of other parameters are mostly set to be consistent with the
existing literature on the estimation of New Keynesian models. For example, the prior
distribution of the slope of the Phillips curve is from Lubik and Schorfheide (2004).
For switching parameters, prior distributions are set to be roughly consistent with
split sample (pre-1983, post-1983) estimates in fixed-regime models. This induces the
natural ordering of regime-dependent parameters and mitigates the potential risk of
the label switching problem as noted in Hamilton, Waggoner, and Zha (2007).
As a check to ensure that the MSNK model can capture empirically relevant
shifts in persistence, Table 2 reports the autocorrelation coefficients for U.S. inflation
over two subsamples - 1953:Q1-1979:Q2 and 1984:Q1-2006:Q4.18 These sub-samples
roughly separate the higher inflation era prior to the Volcker disinflation from the
recent 25 years that has experienced relatively stable and declining inflation rates.
The estimates show that serial correlation in U.S. inflation is lower in the later sample,
dropping from .89 to .57. We compute the model implied inflation persistence for
200 prior draws to ensure that there is positive probability of matching the actual
moments from the split sample data.19
17

Sims, Waggoner, and Zha (2008) argue that Geweke (1999)s method is not robust when the
posterior distribution may be non-Gaussian due to regime switching effects and suggest an alternative
method of computing the marginal likelihood based on a family of elliptical distribution. However,
this turns out to be numerically unstable in our case because the measure is quite sensitive to a few
draws which are somewhat distant from the posterior mode. More discussions on this issue are in
the technical appendix available upon request.
18
These estimates are based on the percentage change in the GDP deflator, which is the same
measure of inflation we using in estimation of the MSNK model.
19
For a more systematic approach of eliciting priors in this way, see Del Negro and Schorfheide
(2008). The technical appendix provides further details.

14

4.3

Posterior Distribution

We estimate three versions of the MSNK model. The first allows switching only in the
monetary policy rule and the second allows switching only in the shock volatilities.
The third model is the four-regime MSNK model that allows independent switching in
monetary policy and the shock volatility regimes. Posterior estimates of parameters
are based on 100,000 draws. For the four-regime model and the model with switching
only in the volatilities, we throw away the first 50,000 draws to induce stationarity of
the MCMC output. Table 3 provides prior and posterior probability intervals for all
the parameters for the models with a constant inflation target.
For the MSNK model with switching only in the monetary policy rule, the different
monetary regimes adjust the nominal interest rate differently in response to inflation.
Table 3 reports the mean and 90% credible posterior interval for each parameter. The
mean of the reaction coefficient to inflation in the active regime is 2.07, which is much
larger than the .89 coefficient in the passive regime. The response to output is similar
across regimes, although the uncertainty associated with the coefficient in the active
regime is much higher. The timing of the different regimes are given by the posterior
expected values of the smoothed probabilities in Figure 1, which shows two clean
shifts during the sample period. The first occurs when the monetary regime changes
from an active stance, which was in place from 1954-1970, to a passive one. The
second shift occurs in 1982 when policy moves back to an active stance and remains
there until the end of the sample.
The timing and nature of monetary regimes is roughly consistent with estimates
from Clarida, Gali`, and Gertler (2000) and Lubik and Schorfheide (2004), where both
find substantial differences in the reaction of the nominal interest rate to inflation
before and after approximately 1980.20 A key difference between the estimates from
the MSNK model and these papers is that policy was active for a significant period
before 1970. Similarly, estimates from Bianchi (2008) and Eo (2008), who allow
regime shifts in policy coefficients, also find that the active policy stance was in place
for much of the time prior to the 1970s.
A key feature of U.S. data is the high and volatile inflation in the 1970s. Since
the steady state level of inflation is constant, the estimation algorithm matches this
shift in inflation volatility using whatever switching parameters it has available, which
in this first case are the monetary reaction coefficients. The reaction coefficient to
inflation plays an important role in determining the volatility of inflation. As this
coefficient increases, the volatility of inflation declines. In the limit, monetary policy
can completely stabilize inflation. Moving in the other direction, inflation becomes
20

There are, of course, alternative interpretations of monetary policy before 1980. Schorfheide
(2005) reports regime switching in the inflation target, while reaction coefficients are left to be time
invariant. Sims and Zha (2006) place emphasis on changes in the volatility of innovations to the
monetary policy rule, instead of changes to the reaction coefficients.

15

more volatile as the reaction to inflation declines. Davig and Leeper (2007) show that
a monetary reaction coefficient less than unity has the affect of amplifying shocks.
Thus, the MSNK model with switching in only the monetary rule uses a passive
monetary regime to generate higher inflation volatility in the 1970s. Alternatively,
the volatile inflation in the 1970s could reflect higher shock volatility, so estimating the
MSNK model with switching only in monetary policy could incorrectly be attributing
the higher volatility in the 1970s to policy.
To address this concern and assess the role of shifting shock volatilities in explaining U.S. data, we estimate the MSNK model with switching only in the variance of
the shocks. Table 3 reports that the standard deviation of each shock roughly doubles in the high volatility state and that the relative volatility of the markup shock,
which is most persistent among three shocks, increases the most. Figure 2 reports the
timing of the low volatility regime. Again, the high inflation periods in the 1970s and
volatile early 1980s stand out as a different regime. However, short-lived transitions
to the high-volatility regime also occur around NBER-recessions, such as the business
cycle peaks that occur in 1953, 1957, and 2001. Less pronounced movements in the
probabilities also occur around the business cycle peaks in 1960 and 1990.
Estimation of two MSNK models suggests that the high and volatile inflation
during the 1970s can be explained by either passive policy or high volatility regime.
To evaluate the relative contribution of each channel, we estimate a broader MSNK
model with switching in both monetary policy and shock volatility regimes. Switching
in monetary policy and volatility regimes is independent, so for example, a change in
the monetary regime does not require a change in the shock volatility regime.
For this four-regime MSNK model, the regime-switching parameters are broadly
similar to the previous estimates with one important exception. Table 3 shows the
monetary policy reaction coefficient in the less aggressive monetary regime is now
substantially greater than unity, implying active policy in both regimes. This reflects
that the model captures higher inflation volatility in the 1970s with a combination of
higher shock volatility and relatively less aggressive monetary policy.
Figure 3 reports the posterior expected values of the smoothed probabilities for
the active monetary policy regime. These estimates indicate that monetary policy
was aggressive in responding to inflation throughout the latter half of the 1950s and
most of the 1960s. Beginning in the late 1960s, however, policy began responding
less aggressively and maintained this stance through the mid-1980s. Policy responds
aggressively for a short-period in the latter part of the 1980s, then turns to a more
persistent active stance in the mid-1990s.
Expected values of the smoothed probabilities for the shock volatility regimes from
the four-regime model are given in Figure 4, which show the timing of fluctuations
between high and low volatility regimes. The high volatility regimes often correspond
to NBER-dated recessions. The low-volatility regime is in place throughout most of
16

post-1984 period, or Great Moderation era, with the exception of the 2001 recession.
Also, estimates characterize the Volcker disinflation period with the high-volatility
regime, rather than the more aggressive monetary regime. Policy is active in both
regimes, so policy still responded systematically more than one-for-one to inflation
during the Volcker disinflation.
Table 4 reports the log marginal likelihood values for each model and indicates
that the data prefers the model with switching shock volatility over the model with
switching monetary policy. However, the model that best fits the data is the fourregime MSNK model. Since the marginal likelihood penalizes overparameterization,
the better fit of the four-regime model is not driven by the increase in parameters.
Also, estimation of the fixed-regime model with an inflation target that follows a
random walk had a substantially lower log marginal likelihood. This suggests that
persistent regime shifts in MSNK models can capture low frequency movements in
inflation dynamics better than models with a drifting inflation target.

Changes in Inflation Persistence

An important issue for monetary policymakers is detecting the sources of changes

in the persistence of inflation. In the MSNK model, monetary policy affects the
contemporaneous effect of shocks - that is, the A (i) , B (i) , and C (i) terms - but
not the rate of decay of a given shock. From this standpoint, monetary policy has
nothing to do with inflation persistence, since the only source of propagation in the
model is exogenous. However, this does not imply that monetary policy is unable to
affect the serial correlation properties of inflation itself.
Typically, empirical studies use a reduced-from specification to measure persistence, where inflation depends on its own lags.21 To more closely link inflation persistence in the MSNK model to the empirical literature, we compute the model implied
inflation persistence as the population moment for the autocorrelation of inflation.
For the four-regime MSNK model, this statistic conditional on a given regime is
AR(t |st = i , rt = j, t) = wa (i, j)a + wu (i, j)u + (1 wa (i, j) wu (i, j))e , (40)
where

a2 (j)
,
wa (i, j) = A (i) W (i, j)
1 2a
 2

u (j)
2
wu (i, j) = B (i) W (i, j)
,
1 2u
2

21

See footnote 1 for a listing of relevant papers.

17

(41)
(42)

and


1
2
2
2
2 a (j)
2 u (j)
2 e (j)
W (i, j) = A (i)
+ B (i)
+ C (i)
,
1 2a
1 2u
1 2e

(43)

for i, j = 1, 2. Equation (40) shows the serial correlation of inflation is a weighted

average of the autocorrelation parameters of the exogenous shocks. A change in the
monetary or shock volatility regime reshuffles the weights across these autocorrelation
parameters. A regime change that shifts weight from more persistent to less persistent
shocks will decrease this measure of inflation persistence. Based on the estimates
given in Table 3, a shift to the more aggressive monetary regime or low-volatility
regime transfers weight to the less persistent shocks and thereby, reduces inflation
persistence.
Table 5 reports the model implied inflation persistence statistic for each MSNK
model across regimes. Focusing on the four-regime MSNK model (i.e. specification P3
in Table 5), the lowest degree of inflation persistence is in the regime with aggressive
monetary policy and low shock volatility (i.e. st = 1 and rt = 2). According to Figure
3, this regime was in place throughout most of the 1960s and after the mid 1990s.
The regimes with the more active policy (i.e. st = 1) have lower persistence than the
regimes with less active policy (i.e. st = 2). The differences across monetary regimes
are substantial and their 90% credible intervals only modestly overlap. For example,
the 90% credible interval for persistence in the more active monetary regime with low
volatility is [.69,.78]. In the high-volatility regime, the interval changes to [.78,.86]. A
similar pattern exists for the less active regimes, except the ranges are a bit higher.
A more complete picture of the changes in model implied inflation persistence
is given in Figure 5, which plots the measure of persistence given in (40) over the
sample period. The more active policy combined with some periods in the lowvolatility regime generates a relatively low degree of persistence early in the sample.
Persistence increases beginning with the shift to less active policy that occurs in the
late 1960s and then peaks around 1980 due to the shift to the high-volatility regime.
The switch to the low-volatility regime decreases persistence in the early 1980s, then
the switch to the more active policy further reduces persistence starting in about
1994.
The pattern of rising persistence during the 1960s and 1970s, followed by a decline
starting around the Volcker disinflation, is consistent with the existing evidence on
changes in inflation persistence. For example, Cogley, Primiceri, and Sargent (2007)
use models with drifting coefficients and stochastic volatility to study inflation dynamics and report a gradual rise in their measure of inflation persistence with it peaking
around 1980, then gradually declining.22 The drop in model-implied persistence from
22

In the context of the MSNK model, the trend, or steady state, rate of inflation is constant, so
there is no distinction between persistence of the inflation gap, which is what Cogley, Primiceri, and
Sargent (2007) study, and persistence of inflation itself.

18

the early-1980s to the mid-1990s also corresponds to the timing in Carlstrom, Fuerst,
and Paustian (2008), who report a decline in the serial correlation in inflation using
40-quarter rolling regressions. Also, Evans and Wachtel (1993) model inflation as following a reduced-form Markov switching process and show a shift occurs to a regime
with high persistence around 1968 that lasts until 1984.23

Conclusion

This paper reports the results of Bayesian estimation of MSNK models with regime
switching in monetary policy and shock volatility. Overall, U.S. data favors the model
with independent switching in both the monetary policy and shock volatility regimes.
We show that the population moment describing the serial correlation of inflation is
a weighted average of the autocorrelation parameters of the exogenous shocks. The
weights depend on the different monetary and shock volatility regimes. Consequently,
changes in either of the regimes reshuffles the weights over these serial correlation parameters and alters the serial correlation properties of inflation. Estimation indicates
that a shift to the more active monetary regime reduces the weight on the more
persistent shocks, so lowers the serial correlation of inflation. Similarly, a shift to
the low volatility regime reduces the weight on the more persistent shocks and also
contributes to reducing inflation persistence. Estimates indicate that inflation persistence began rising in the late 1960s and peaked around the Volcker disinflation.
Our estimates indicate that it is hard to explain the rise in persistence without a less
aggressive monetary policy. The subsequent decline in persistence beginning in the
early 1980s is driven by both a more active monetary policy regime and less volatile
shocks.
23

Kang, Kim, and Morley (2007) report similar findings.

19

References
An, S., and F. Schorfheide (2007): Bayesian Analysis of DSGE Models, Econometric Reviews, 26, 113172.
Bansal, R., and H. Zhou (2002): Term Structure of Interest Rates with Regime
Shifts, Journal of Finance, 57, 19972043.
Bianchi, F. (2008): Regime Switches, Agents Beliefs, and Post-World War II
Macroeconomic Dynamics, Manuscript, Princeton University.
Bikbov, R. (2006): Monetary Policy Regimes and the Term Structure of Interest
Rates, Manuscript, Columbia University.
Blanchard, O. J., and C. Kahn (1980): The Solution of Linear Difference
Models Under Rational Expectations, Econometrica, 48, 13051313.
Branch, W. A., T. Davig, and B. McGough (2007): Adaptive Learning in
Regime-Switching Models, Federal Reserve Bank of Kansas City 07-09.
Carlstrom, C. T., T. S. Fuerst, and M. Paustian (2008): Inflation Persistence, Monetary Policy and the Great Moderation, Journal of Money, Credit and
Banking, forthcoming.
Cecchetti, S. G., and G. Debelle (2006): Has the Inflation Process Changed?,
Economic Policy, pp. 312352.
Clarida, R., J. Gali`, and M. Gertler (2000): Monetary Policy Rules and
Macroeconomic Stability : Evidence and Some Theory, Quarterly Journal of Economics, 115(1), 147180.
Clark, T. E. (2006): Disaggregate Evidence on the Persistence of Consumer Price
Inflation, Journal of Applied Econometrics, 21, 563587.
Cogley, T., G. E. Primiceri, and T. J. Sargent (2007): Inflation-Gap Persistence in the U.S., Working Paper.
Cogley, T., and T. J. Sargent (2001): Evolving Post-World War II U.S. Inflation Dynamics, in NBER Macroeconomics Annual 2001, ed. by B. S. Bernanke,
and J. J. Rotemberg. MIT Press, Cambridge.
Cooley, T. F., S. F. LeRoy, and N. Raymon (1984): Econometric Policy
Evaluation: Note, American Economic Review, 74(3), 467470.
Davig, T., and E. M. Leeper (2007): Generalizing the Taylor Principle, American Economic Review, 97(3), 607635.

20

Del Negro, M., and F. Schorfheide (2008): Forming Priors for DSGE Models
(and How it Affects the Assessment of Nominal Rigidities), Journal of Monetary
Economics, 55(7), 11911208.
Eo, Y. (2008): Bayesian Analysis of DSGE Models with Regime Switching,
Manuscript, Washington University in St. Louis.
Evans, M., and P. Wachtel (1993): Inflation Regimes and the Sources of Inflation Uncertainty, Journal of Money, Credit and Banking, 25, 475511.
Farmer, R. E., D. F. Waggoner, and T. Zha (2008): Generalizing the Taylor
Principle: Comment, University of California - Los Angeles, mimeo.
Gadzinski, G., and F. Orlandi (2004): Inflation Persistence in the European
Union, the Euro Area, and the United States, .
Geweke, J. (1999): Using Simulation Methods for Bayesian Econometric Models:
Inference, Development and Communication, Econometric Reviews, 18(1), 1126.
Gordon, S., and P. St-Amour (2000): A Preference Regime Model of Bear and
Bull Markets, American Economic Review, 90, 10191033.
Hamilton, J., D. F. Waggoner, and T. Zha (2007): Normalization in Econometrics, Econometric Reviews, 26, 221252.
Ireland, P. N. (2004): Technology Shocks in the New Keynesian Model, Review
of Economics and Statistics, 86(4), 923936.
Kang, K. H., C.-J. Kim, and J. Morley (2007): Changes in U.S. Inflation
Persistence, manuscript, Washington University in St. Louis.
Kim, C.-J., and C. R. Nelson (1999): State-Space Models with Regime Switching.
MIT Press, Cambridge, MA.
Leeper, E. M. (1991): Equilibria Under Active and Passive Monetary and Fiscal
Policies, Journal of Monetary Economics, 27, 129147.
Levin, A. T., and J. M. Piger (2006): Is Inflation Persistence Intrinsic in Industrial Economies?, mimeo, University of Oregon.
Lubik, T. A., and F. Schorfheide (2004): Testing for Indeterminacy: An Application to U.S. Monetary Policy, American Economic Review, 94, 190217.
McCallum, B. (2004): On the Relationship Between Determinate and MSV Solutions in Linear RE Models, Economics Letters, 84, 5560.
Pivetta, F., and R. Reis (2007): The Persistence of Inflation in the United
States, Journal of Economic Dynamics and Control, 31, 13261358.
21

Schorfheide, F. (2005): Learning and Monetary Policy Shifts, Review of Economic Dynamics, 8, 392419.
Sims, C. A. (2002): Solving Linear Rational Expectations Models, Journal of
Computational Economics, 20, 120.
Sims, C. A., D. F. Waggoner, and T. Zha (2008): Methods for Inference
in Large Multiple-equation Markov-switching Models, Journal of Econometrics,
forthcoming.
Sims, C. A., and T. Zha (2006): Were There Regime Switching in U.S. Monetary
Policy, American Economic Review, 96, 5481.
Stock, J. H. (2001): Discussion of Cogley and Sargent, Evolving Post-World War
II Inflation Dynamics, in NBER Macroeconomics Annual, ed. by B. S. Bernanke,
and J. J. Rotemberg. MIT Press, Cambridge.
Stock, J. H., and M. Watson (2007): Has Inflation Become Harder to Forecast?, Journal of Money, Credit and Banking, 39, 334.

22

Table 1: Prior Distribution

Parameters

1
2

1
2

a
u
e
a
a,1
a,2
u
u,1
u,2
e
e,1
e,2
ln A0
y
p11
p22
q11
q22

Domain
R+
R+
R+
R+
R+
R+
R+
[0,1)
R+
R+
R+
[0,1)
[0,1)
[0,1)
R+
R+
R+
R+
R+
R+
R+
R+
R+
R
R
[0,1)
[0,1)
[0,1)
[0,1)

Density
Para(1)
Gamma
1.5
Gamma
2
Gamma
1
Gamma
.08
Gamma
.1
Gamma
.1
Gamma
.5
Beta
.998
Gamma
1.5
Gamma
.005
Gamma
.0086
Beta
.3
Beta
.7
Beta
.5
Inverse Gamma
.004
Inverse Gamma
.006
Inverse Gamma
.003
Inverse Gamma
.003
Inverse Gamma
.004
Inverse Gamma
.002
Inverse Gamma
.003
Inverse Gamma
.004
Inverse Gamma
.002
Normal
9.546
Normal
-0.067
Beta
.9
Beta
.9
Beta
.9
Beta
.9

Para(2)
.25
.25
.1
.05
.05
.05
.2
.001
.4
.001
.001
.2
.2
.2
4
4
4
4
4
4
4
4
4
.1
.01
.05
.05
.05
.05

Regime Spec
P2
P1, P3
P1, P3
P2
P1, P3
P1, P3
P1, P2, P3
P1, P2, P3
P1, P2, P3
P1, P2, P3
P1, P2, P3
P1, P2, P3
P1, P2, P2
P1, P2, P3
P1
P2, P3
P2, P3
P1
P2, P3
P2, P3
P1
P2, P3
P2, P3
P1, P2, P3
P1, P2, P3
P1, P2, P3
P3
P3

Notes: Para (1) and Para (2) list the means and the standard deviations for Beta, Gamma,
and Normal distributions; s and for the Inverse Gamma distribution, where pIG (|, s)
2
2
1 es /2 , a and b for the Uniform distribution from a to b. P1 allows switching
only in monetary policy coefficients while P2 allows switching coefficients only in variance
parameters of shocks. P3 allows switching for both policy coefficients and variances.

23

Table 2: Autocorrelation of U.S. Inflation

Data
AR()

1953:Q1 - 1979:Q2
0.8856
(0.0656)

1984:Q1 - 2006:Q4
0.5734
(0.1156)

Notes: For data, numbers in ( ) are standard errors for sample autocorrelation coefficients.
All the standard-errors are Newey-West (1987) adjusted with 4 lags.

24

Table 3: Posterior Distribution

Parameters

Prior 90% Interval

P1

1
2

1
2

a
u
i
a
a,1
a,2
u
u,1
u,2
e
e,1
e,2
ln A0
y
p11
p22
q11
q22

[1.10,1.91]
[1.59,2.41]
[0.84,1.16]
[0.008,0.150]
[0.025,0.174]
[0.024,0.174]
[0.181,0.804]
[0.9965,0.9995]
[0.851,2.134]
[0.0033,0.0066]
[0.0070,0.0103]
[0.001,0.592]
[0.407,0.999]
[0.172,0.828]
[0.0022,0.0080]
[0.0031,0.0112]
[0.0016,0.0060]
[0.0027,0.0099]
[0.0022,0.0079]
[0.0011,0.0039]
[0.0016,0.0059]
[0.0022,0.0079]
[0.0010,0.0039]
[9.381,9.719]
[-0.0832,-0.0504]
[0.82,0.98]
[0.82,0.98]
[0.82,0.98]
[0.82,0.98]

Posterior 90% Interval

P2
[1.58,1.62]

[1.81,2.30]
[0.80,0.98]

P3
[1.74,2.26]
[0.98,1.16]

[0.052,0.076]
[0.029,0.161]
[0.048,0.176]
[0.260,0.551]
[0.9971,0.9990]
[3.149,4.886]
[0.0042,0.0058]
[0.0079,0.0095]
[0.000,0.089]
[0.953,0.982]
[0.629,0.728]
[0.0097,0.0113]

[0.291,0.557]
[0.9982,0.9990]
[3.854,4.089]
[0.0046,0.0050]
[0.0079,0.0087]
[0.005,0.111]
[0.942,0.944]
[0.655,0.738]

[0.013,0.089]
[0.041,0.190]
[0.273,0.562]
[0.9984,0.9997]
[3.041, 4.457]
[0.0048,0.0062]
[0.0067,0.0080]
[0.001,0.065]
[0.934,0.965]
[0.648,0.733]

[0.0136,0.0184]
[0.0067,0.0087]

[0.0156,0.0210]
[0.0067,0.0084]

[0.0042,0.0061]
[0.0014,0.0018]

[0.0027,0.0046]
[0.0009,0.0013]

[0.0057,0.0077]
[0.0030,0.0038]
[9.527,9.528]
[-0.0679,-0.0671]

[0.0050,0.0068]
[0.0031,0.0041]
[9.508,9.538]
[-0.0789,-0.0498]
[0.95,0.98]
[0.95,0.98]
[0.79,0.93]
[0.93,0.98]

[0.0013,0.0020]

[0.0040,0.0050]

[9.509,9.541]
[-0.0826,-0.0503]
[0.94,0.98]
[0.90,0.96]

[0.85,0.96]
[0.91,0.98]

Notes: P1 allows switching only in monetary policy coefficients while P2 allows switching
coefficients only in variance parameters of shocks. P3 allows switching for both policy
coefficients and variances.

Table 4: Log Marginal Data Densities

P1
2,697.2

P2
2,711

P3
2,731.3.

Fixed Regime (Random Walk Inflation Target)

2,652.6

25

Table 5: MSNK Model-Implied Inflation Persistence

Model
AR(|st = 1, rt
AR(|st = 1, rt
AR(|st = 2, rt
AR(|st = 2, rt

P1
P2
P3
= 1t) [.77,.88] [.83,.88] [.78,.86]
= 2t)
[.76,.82] [.69,.78]
= 1t) [.92,.97]
[.89,.94]
= 2t)
[.81,.90]

Notes: The posterior mean of the model implied inflation autocorrelation is reported with
posterior standard deviation in [ ]. st = 1 corresponds to the more active monetary policy
and rt = 1 corresponds to the high-volatility regime.

Figure 1: Posterior Expected Values of the Active Monetary Policy

(2-Regime Model) Regime Probability
1

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

1955

1960

1965

1970

1975

1980

1985

1990

1995

2000

2005

NOTE: The posterior expected value is computed using 2,000 posterior draws.

26

Figure 2: Posterior Expected Values of the High-Volatility Regime (2Regime Model) Probability
1

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

1955

1960

1965

1970

1975

1980

1985

1990

1995

2000

2005

NOTE: The posterior expected value is computed using 2,000 posterior draws.

Figure 3: Posterior Expected Values of the More Active Monetary

Regime Probability (4-regime model)
1

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

1955

1960

1965

1970

1975

1980

1985

1990

1995

2000

2005

NOTE: The posterior expected value is computed using 2,000 posterior draws.

27

Figure 4: Posterior Expected Values of High-Volatility Regime Probability (4-regime model)

1

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

1955

1960

1965

1970

1975

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

1980

1985

1990

1995

2000

2005

NOTE: The posterior expected value is computed using 2,000 posterior draws. Shaded
areas are NBER-defined recessions.

Figure 5: MSNK Model-Implied Inflation Persistence (90% Bands, 4Regime Model)

1

0.9

0.9

0.8

0.8

0.7

0.7

0.6

1955

1960

1965

1970

1975

1980

1985

1990

1995

2000

2005

0.6

NOTE: The posterior expected value is computed using 2,000 posterior draws.

28