Dsge

STATA DYNAMIC STOCHASTIC
GENERAL EQUILIBRIUM MODELS

REFERENCE MANUAL
RELEASE 18
A Stata Press Publication

StataCorp LLC
College Station, Texas
®
Copyright c 1985–2023 StataCorp LLC
All rights reserved
Version 18
Published by Stata Press, 4905 Lakeway Drive, College Station, Texas 77845
Typeset in TEX
ISBN-10: 1-59718-377-6
ISBN-13: 978-1-59718-377-2
This manual is protected by copyright. All rights are reserved. No part of this manual may be reproduced, stored
in a retrieval system, or transcribed, in any form or by any means—electronic, mechanical, photocopy, recording, or
otherwise—without the prior written permission of StataCorp LLC unless permitted subject to the terms and conditions
of a license granted to you by StataCorp LLC to use the software and documentation. No license, express or implied,
by estoppel or otherwise, to any intellectual property rights is granted by this document.
StataCorp provides this manual “as is” without warranty of any kind, either expressed or implied, including, but
not limited to, the implied warranties of merchantability and fitness for a particular purpose. StataCorp may make
improvements and/or changes in the product(s) and the program(s) described in this manual at any time and without
notice.
The software described in this manual is furnished under a license agreement or nondisclosure agreement. The software
may be copied only in accordance with the terms of the agreement. It is against the law to copy the software onto
DVD, CD, disk, diskette, tape, or any other medium for any purpose other than backup or archival purposes.
The automobile dataset appearing on the accompanying media is Copyright c 1979 by Consumers Union of U.S.,
Inc., Yonkers, NY 10703-1057 and is reproduced by permission from CONSUMER REPORTS, April 1979.
Stata, , Stata Press, Mata, , and NetCourse are registered trademarks of StataCorp LLC.
Stata and Stata Press are registered trademarks with the World Intellectual Property Organization of the United Nations.
NetCourseNow is a trademark of StataCorp LLC.
Other brand and product names are registered trademarks or trademarks of their respective companies.
For copyright information about the software, type help copyright within Stata.
The suggested citation for this software is

StataCorp. 2023. Stata: Release 18 . Statistical Software. College Station, TX: StataCorp LLC.
Contents
Intro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Introduction to DSGE manual 1

Intro 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Introduction to DSGEs 4
Intro 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Learning the syntax 20
Intro 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Classic DSGE examples 28
Intro 3a . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . New Keynesian model 29
Intro 3b . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . New Classical model 34
Intro 3c . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Financial frictions model 42
Intro 3d . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Nonlinear New Keynesian model 47
Intro 3e . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Nonlinear New Classical model 55
Intro 3f . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Stochastic growth model 66
Intro 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Writing a DSGE in a solvable form 74
Intro 4a . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Specifying a shock on a control variable 80
Intro 4b . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Including a lag of a control variable 82
Intro 4c . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Including a lag of a state variable 85
Intro 4d . . . . . . . . . . . . . . . . Including an expectation dated by more than one period ahead 88
Intro 4e . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Including a second-order lag of a control 91
Intro 4f . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Including an observed exogenous variable 94
Intro 4g . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Correlated state variables 98
Intro 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Stability conditions 101
Intro 6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Identification 107
Intro 7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Convergence problems 110
Intro 8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Wald tests vary with nonlinear transforms 116
Intro 9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bayesian estimation 123
Intro 9a . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bayesian estimation of a New Keynesian model 130
Intro 9b . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bayesian estimation of stochastic growth model 139
dsge . . . . . . . . . . . . . . . . . . . . . . . . . . Linear dynamic stochastic general equilibrium models 145
dsge postestimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Postestimation tools for dsge 152
dsgenl . . . . . . . . . . . . . . . . . . . . . . . Nonlinear dynamic stochastic general equilibrium models 156
dsgenl postestimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Postestimation tools for dsgenl 162
estat covariance . . . . . . . . . . . . . . . . . . . . . Display estimated covariances of model variables 166
estat policy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Display policy matrix 171
estat stable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Check stability of system 173
estat steady . . . . . . . . . . . . . . . . . . . . . . . . . . Display steady state of nonlinear DSGE model 175
estat transition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Display state transition matrix 177
Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
Subject and author index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
i
Cross-referencing the documentation
When reading this manual, you will find references to other Stata manuals, for example,
[U] 27 Overview of Stata estimation commands; [R] regress; and [D] reshape. The first ex-
ample is a reference to chapter 27, Overview of Stata estimation commands, in the User’s Guide;
the second is a reference to the regress entry in the Base Reference Manual; and the third is a
reference to the reshape entry in the Data Management Reference Manual.
All the manuals in the Stata Documentation have a shorthand notation:
[GSM] Getting Started with Stata for Mac

[GSU] Getting Started with Stata for Unix
[GSW] Getting Started with Stata for Windows
[U] Stata User’s Guide
[R] Stata Base Reference Manual
[ADAPT] Stata Adaptive Designs: Group Sequential Trials Reference Manual
[BAYES] Stata Bayesian Analysis Reference Manual
[BMA] Stata Bayesian Model Averaging Reference Manual
[CAUSAL] Stata Causal Inference and Treatment-Effects Estimation Reference Manual
[CM] Stata Choice Models Reference Manual
[D] Stata Data Management Reference Manual
[DSGE] Stata Dynamic Stochastic General Equilibrium Models Reference Manual
[ERM] Stata Extended Regression Models Reference Manual
[FMM] Stata Finite Mixture Models Reference Manual
[FN] Stata Functions Reference Manual
[G] Stata Graphics Reference Manual
[IRT] Stata Item Response Theory Reference Manual
[LASSO] Stata Lasso Reference Manual
[XT] Stata Longitudinal-Data/Panel-Data Reference Manual
[META] Stata Meta-Analysis Reference Manual
[ME] Stata Multilevel Mixed-Effects Reference Manual
[MI] Stata Multiple-Imputation Reference Manual
[MV] Stata Multivariate Statistics Reference Manual
[PSS] Stata Power, Precision, and Sample-Size Reference Manual
[P] Stata Programming Reference Manual
[RPT] Stata Reporting Reference Manual
[SP] Stata Spatial Autoregressive Models Reference Manual
[SEM] Stata Structural Equation Modeling Reference Manual
[SVY] Stata Survey Data Reference Manual
[ST] Stata Survival Analysis Reference Manual
[TABLES] Stata Customizable Tables and Collected Results Reference Manual
[TS] Stata Time-Series Reference Manual
[I] Stata Index
[M] Mata Reference Manual
ii
Title
Intro — Introduction to DSGE manual
Description Remarks and examples Also see
Description
DSGE stands for dynamic stochastic general equilibrium. DSGE models are multivariate time-series
models that are used in economics, in particular, macroeconomics, for policy analysis and forecasting.
These models are systems of equations that are typically derived from economic theory. As such,
the parameters are often directly interpretable based on theory. DSGE models are unique in that
equations in the system allow current values of variables to depend not only on past values but also
on expectations of future values.
The dsgenl command estimates parameters of nonlinear DSGE models.
The dsge command estimates parameters of linear DSGE models.
Remarks and examples

We recommend that you read this manual beginning with [DSGE] Intro 1 and then continue with
the remaining introductions. In these introductions, we will introduce DSGE models, show you how
to use the dsgenl and dsge commands, walk you through worked examples of classic models, and
present solutions to common stumbling blocks.
[DSGE] Intro 1 and [DSGE] Intro 2 are essential reading. Read them first. Here you will find
an overview of DSGE models, descriptions of concepts used throughout the manual, discussion of
assumptions, a first worked example, and an introduction to the syntax.
[DSGE] Intro 1 Introduction to DSGEs

[DSGE] Intro 2 Learning the syntax
[DSGE] Intro 3 focuses on classical DSGE models. It includes a series of examples that illustrate
model solution, model estimation, and postestimation procedures for simple variants of common
models.
[DSGE] Intro 3 Classic DSGE examples

[DSGE] Intro 3a New Keynesian model
[DSGE] Intro 3b New Classical model
[DSGE] Intro 3c Financial frictions model
[DSGE] Intro 3d Nonlinear New Keynesian model
[DSGE] Intro 3e Nonlinear New Classical model
[DSGE] Intro 3f Stochastic growth model
[DSGE] Intro 4 discusses some features commonly found in DSGE models and how to specify
models with those features to dsge and dsgenl. The structural equations of the DSGE model must
have a specific structure so that the model can be solved. Often, DSGE models are written using
intuitive forms that do not have this structure. These intuitive forms can be rewritten in a logically
equivalent form that has the structure required for solution. [DSGE] Intro 4 provides an overview of
this topic and examples demonstrating solutions.
1
2 Intro — Introduction to DSGE manual
[DSGE] Intro 4 Writing a DSGE in a solvable form

[DSGE] Intro 4a Specifying a shock on a control variable
[DSGE] Intro 4b Including a lag of a control variable
[DSGE] Intro 4c Including a lag of a state variable
[DSGE] Intro 4d Including an expectation dated by more than one period
ahead
[DSGE] Intro 4e Including a second-order lag of a control
[DSGE] Intro 4f Including an observed exogenous variable
[DSGE] Intro 4g Correlated state variables
[DSGE] Intro 5–[DSGE] Intro 8 discuss technical issues. These introductions are essential reading,
even though they are last.
[DSGE] Intro 5 Stability conditions

[DSGE] Intro 6 Identification
[DSGE] Intro 7 Convergence problems
[DSGE] Intro 8 Wald tests vary with nonlinear transforms
[DSGE] Intro 9, [DSGE] Intro 9a, and [DSGE] Intro 9b discuss Bayesian estimation.
[DSGE] Intro 9 Bayesian estimation

[DSGE] Intro 9a Bayesian estimation of a New Keynesian model
[DSGE] Intro 9b Bayesian estimation of stochastic growth model
The main command entries are references for syntax and implementation details. All the examples
are in the introductions discussed above.
[DSGE] dsge Linear dynamic stochastic general equilibrium models

[DSGE] dsge postestimation Postestimation tools for dsge
[DSGE] dsgenl Nonlinear dynamic stochastic general equilibrium models
[DSGE] dsgenl postestimation Postestimation tools for dsgenl
[DSGE] estat covariance Display estimated covariances of model variables
[DSGE] estat policy Display policy matrix
[DSGE] estat stable Check stability of system
[DSGE] estat steady Display steady state of nonlinear DSGE model
[DSGE] estat transition Display state transition matrix
[BAYES] bayes: dsge Bayesian linear dynamic stochastic general equilibrium

models
[BAYES] bayes: dsgenl Bayesian nonlinear dynamic stochastic general equilibrium
models
[BAYES] bayes: dsge postestimation Postestimation tools for bayes: dsge and bayes: dsgenl
[BAYES] bayesirf Bayesian IRFs, dynamic-multiplier functions, and FEVDs
Intro — Introduction to DSGE manual 3
Also see
[DSGE] Intro 1 — Introduction to DSGEs
Title
Intro 1 — Introduction to DSGEs
Description Remarks and examples References Also see
Description
In this entry, we introduce DSGE models and the two commands that estimate the parameters of DSGE
models, dsgenl and dsge. For Bayesian estimation, see [BAYES] bayes: dsge and [BAYES] bayes: ds-
genl. We begin with an overview of DSGE models. We then illustrate the process of DSGE modeling
by working an example. In this example, we describe a model in its original nonlinear form and
estimate its parameters using dsgenl. We also write the model in a corresponding linearized form
and estimate the parameters using dsge. We conclude by showing how this example fits into the
general DSGE modeling framework.

Remarks are presented under the following headings:
Introduction to DSGE models
An example: A nonlinear DSGE model
Writing down nonlinear DSGEs
Data preparation
Specifying the model to dsgenl
Parameter estimation and interpretation of nonlinear DSGEs
An example: A linear DSGE model
Writing down linearized DSGEs
Specifying the model to dsge
Parameter estimation and interpretation of linear DSGEs
Postestimation
Policy and transition matrices
Impulse responses
Forecasts
Structural and reduced forms of DSGE models
Introduction to DSGE models

DSGE models are models for multiple time series used in macroeconomics and finance. These
models are systems of equations that are motivated by economic theory and in which expectations
of future values of variables play an important role. Because these models come from theory, the
parameters of these models can typically be directly interpreted in terms of the motivating theory.
DSGE models are used for macroeconomic policy analysis and forecasting.
In DSGE models, individuals’ actions are summarized by decision rules that take the form of
nonlinear systems of dynamic equations. These decision rules often come from dynamic stochastic
optimization problems. A DSGE model consists of these decision rules, plus any aggregation conditions,
resource or budget constraints, and stochastic processes for exogenous variables.
Because the model’s equations are the solution to dynamic optimization problems, model equations
can feature expectations of future variables. These expectations are endogenous. In DSGE models,
expectations of future values of variables correspond to their conditional mean as implied by the model.
In other words, individuals’ expectations of future values are correct, on average. Such expectations
are said to be model-consistent expectations or rational expectations.
4
Intro 1 — Introduction to DSGEs 5
There are three kinds of variables in DSGE models: control variables, state variables, and shocks.
The terminology is taken from the state-space and optimal control literatures. In DSGE models, the
concepts of exogeneity and endogeneity are understood relative to a time period. A state variable is
fixed, or exogenous, in a given time period. The system of equations then determines the value of
the state variable one period in the future. On the other hand, the system of equations determines
the value of a control variable in the current time period. Control variables in a DSGE model can be
either observed or unobserved. State variables are always unobserved.
DSGE models can be written in multiple forms. The model may consist of equations that are
nonlinear both in the variables and in the parameters. Such a model is said to be nonlinear. A DSGE
model is said to be linear, or linearized, when the model equations are linear in the variables. Linear
models may still be nonlinear in their parameters.
Nonlinear DSGE models are commonly linearized prior to analysis. In Stata, you have two choices
for handling the linearization of a nonlinear model. You can write your model in its nonlinear form
and use the dsgenl command. dsgenl will take a linear approximation to the nonlinear model for
you. Alternatively, you can derive the linear form of the model and use the dsge command. Although
linearizing the equations yourself requires extra effort, there can be advantages. Linearized DSGE
models are easier to manipulate than nonlinear ones. For instance, it is easier to include additional
lag terms in a linearized model than in a nonlinear one.
After linearization, the DSGE model must be solved prior to estimation. In any analysis of
simultaneous equations systems, to solve a model means to write the model’s endogenous variables as
functions of its exogenous variables. In DSGE models, the analogous concept is to write the model’s
control variables in terms of its state variables. The model’s solution consists of a system of equations
relating the control variables to the state variables and a system of equations describing the evolution of
the state variables over time. The solution to a DSGE model thus takes the form of a state-space model.
The solution to a DSGE model is a crucial object for both estimation and analyses after estimation.
Both the likelihood function and the impulse–response functions are formed from the model solution.
dsgenl solves and estimates the parameters of nonlinear DSGE models. It linearizes the model
by taking a first-order approximation to the model’s equations at the model’s steady state, solves the
linearized model, and estimates the parameters of the model using the linearized solution.
dsge solves and estimates the parameters of linearized DSGE models.
For Bayesian estimation, see [DSGE] Intro 9. bayes: dsge fits Bayesian linear DSGE models, and
bayes: dsgenl fits Bayesian nonlinear DSGE models.
General introductions to DSGE modeling are available in Ljungqvist and Sargent (2018) and
Woodford (2003). Canova (2007) and DeJong and Dave (2011) describe parameter estimation using
DSGE models. Fernández-Villaverde, Rubio-Ramı́rez, and Schorfheide (2016) provide an overview of
solution and estimation strategies for nonlinear DSGE models.
6 Intro 1 — Introduction to DSGEs
An example: A nonlinear DSGE model

Writing down nonlinear DSGEs
Consider the following nonlinear model, similar to that in Woodford (2003, chap. 4). The model
consists of equations that describe the behavior of households, firms, and a central bank. Interactions
among these actors produce a model of inflation, output growth, and the interest rate. Models of this
type are popular in academic and policy settings and are used to describe monetary policy.
Household optimization generates an equation that relates current output Yt to the expected value
of a function of tomorrow’s output Yt+1 , tomorrow’s inflation Πt+1 , and the current nominal interest
rate Rt ,

1 1 Rt
= βEt (1)
Yt Yt+1 Πt+1
where β is a parameter that captures households’ willingness to delay consumption.

Optimization by firms generates an equation that links the current deviation of inflation from its
steady state, Πt − Π, to the expected value of the deviation of inflation from its steady state in the
future, Et (Πt+1 − Π), and to the ratio of actual output, Yt , to the natural level of output, Zt ,

1 Yt
(Πt − Π) + = φ + βEt (Πt+1 − Π) (2)
φ Zt
where φ is a parameter linked to the pricing decision of firms. Firms are not affected by inflation
per se; they are affected only by deviations of inflation from its steady-state value.
Finally, there is an equation that describes central bank policy. The central bank adjusts the interest
rate in response to inflation and, perhaps, to other factors that we leave unmodeled. The equation for
the central bank policy is
1/β
Rt Πt
= Ut (3)
R Π
where R is the steady-state value of the interest rate and Ut is a state variable that captures all
movements in the interest rate that are not driven by inflation.
We make a few modifications to the model equations before estimating the parameters. Wood-
ford (2003) rewrites the model in (1)–(3) by defining Xt = Yt /Zt as the output gap. We do the
same. Substituting in Xt gives us the three-equation system

Xt 1 Rt
1 = βEt (4a)
Xt+1 Gt Πt+1
1
(Πt − Π) + = φXt + βEt (Πt+1 − Π) (5a)
φ
1/β
Rt Πt
= Ut (6a)
R Π
where Gt = Zt+1 /Zt is a state variable that captures changes in Zt .

The model is completed by adding equations describing the evolution of the state variables Gt
and Ut . We write them as autoregressive processes in logarithms,
ln(Gt+1 ) = ρg ln(Gt ) + ξt+1 (7a)

ln(Ut+1 ) = ρu ln(Ut ) + et+1 (8a)
The variables ξt+1 and t+1 are shocks to the state variables. We now have a complete nonlinear
DSGE model. To estimate its parameters, we need to get some data.
Data preparation
There are three requirements of the data used with dsgenl and dsge. First, the data must be
tsset so that Stata understands the time structure of your data. Second, the series in your model
must have zero mean. This requirement is actually handled for you. Before solving the models, these
commands first transform variables into differences from their steady state or, with dsgenl, percent
deviations from their steady state. The third requirement is that variables in your data must be weakly
stationary.
We have data on the price level and the nominal interest rate in rates.dta. These data were
obtained from the Federal Reserve Economic Database (FRED), which contains many macroeconomic
and financial time series; see [D] import fred. We can type tsset to see that dateq has already
been set as the quarterly time variable in this dataset.
. use https://www.stata-press.com/data/r18/rates2
(Federal Reserve Economic Data - St. Louis Fed, 2017-02-10)
. tsset
Time variable: dateq, 1947q1 to 2017q1
Delta: 1 quarter
. describe
Contains data from https://www.stata-press.com/data/r18/rates2.dta
Observations: 281 Federal Reserve Economic Data -
St. Louis Fed, 2017-02-10
Variables: 5 26 Apr 2022 21:22
Variable Storage Display Value

name type format label Variable label
datestr str10 %-10s Observation date

daten int %td Numeric (daily) date
gdpdef float %9.0g GDP deflator GDPDEF
r float %9.0g Federal funds rate FEDFUNDS
dateq int %tq Quarterly date
Sorted by: dateq
The dataset includes the price-level variable gdpdef. But the model is written in terms of the
inflation rate. For quarterly data, the inflation rate is conventionally obtained as 400 times the difference
in log of the price level. Therefore, we begin by generating an inflation rate variable p by using the
L. lag operator.
. generate p = 400*(ln(gdpdef) - ln(L.gdpdef))
(2 missing values generated)
. label variable p "Inflation rate"
We now have inflation and interest rates.

Specifying the model to dsgenl
Equations (4a)–(8a) are the nonlinear DSGE model. The command to estimate the parameters of
the system (4a)–(8a) is
. dsgenl (1 = {beta}*(x/F.x)*(1/g)*(r/F.p) ) ///
(1/{phi} + (p-1) = {phi}*x + {beta}*(F.p-1)) ///
({beta}*r = p^(1/{beta})*u) ///
(ln(F.u) = {rhou}*ln(u)) ///
(ln(F.g) = {rhog}*ln(g)), ///
exostate(u g) observed(p r) unobserved(x)
Each equation is bound in parentheses. The equations in the command look similar to the equations
we wrote down. The equations can be written in nearly any form that is logically equivalent to the
model equations. In addition to providing the equations, you must specify which variable plays which
role in the model through the options. The option exostate() lists all exogenous state variables,
those that are subject to shocks. The option observed() lists observed control variables. The option
unobserved() lists unobserved (latent) control variables. A fourth option, endostate(), is available
for endogenous state variables, those that are not subject to shocks. Each variable must be listed in
one and only one of the options. The number of exogenous state variables must be the same as the
number of observed control variables.
Parameters are bound in curly braces; this is how Stata distinguishes between parameters and
variables. Parameters can appear nonlinearly, they can appear more than once, and they can appear
in multiple equations. A parameter repeated across equations is considered the same parameter; for
example, in the above model, beta appears in each of the first three equations, and it is assumed
that there is only one parameter beta to be estimated. These cross-equation restrictions are common
in DSGE models.
The operator F. can be applied to both state and control variables to denote the expected future
value of a variable. For example, we typed F.x to include xt+1 in the model. However, unlike
F. used in other time-series applications, the value of xt+1 is not determined from the value of x at
time t + 1 in your data. Instead, it is an expectation of the value of x at time t + 1 based on your
model. In fact, the entire model can be written in terms of expectations.
Notice that in (4a) and (5a), the Et operator appears. This is the statistical expectation operator.
It refers to expected value of the term in brackets conditional on information known at time t. In
dsgenl, this expectation operator is assumed in front of each equation. Hence, the dsgenl equation
(1 = {beta}*(x/F.x)*(1/g)*(r/F.p))
is interpreted as

Xt 1 Rt
Et 1 − β =0
Xt+1 Gt Πt+1
which is equivalent to what we wrote in (4a).

Parameter estimation and interpretation of nonlinear DSGEs
The model’s parameters are estimated by taking a linear approximation to the model’s equations,
solving the linearized system for its state-space form, and estimating the parameters of the state-space
model by maximum likelihood.
. dsgenl (1 = {beta}*(x/F.x)*(1/g)*(r/F.p))
> (1/{phi} + (p-1) = {phi}*x + {beta}*(F.p-1))
> ({beta}*r = p^(1/{beta})*u)
> (ln(F.u) = {rhou}*ln(u))
> (ln(F.g) = {rhog}*ln(g)),
> exostate(u g) observed(p r) unobserved(x) nolog
Solving at initial parameter vector ...
Checking identification ...
First-order DSGE model
Sample: 1954q3 thru 2016q4 Number of obs = 250
Log likelihood = -768.09383
Coefficient Std. err. z P>|z| [95% conf. interval]
/structural
beta .5112894 .0758003 6.75 0.000 .3627236 .6598553
phi 5.896273 1.653019 3.57 0.000 2.656415 9.136131
rhou .6989551 .0449119 15.56 0.000 .6109294 .7869809
rhog .9556397 .0181344 52.70 0.000 .9200968 .9911825
sd(e.u) 2.317581 .29883 1.731885 2.903277

sd(e.g) .6147482 .0973365 .4239722 .8055242
We see header information describing the sample date range, the number of observations, and the
log likelihood.
In the estimation table, we see estimates of model parameters. Some of these parameters have a
structural interpretation. Parameter beta plays two roles in the model: it is the discount factor in the
household and firm equations, and its inverse captures the degree to which the central bank responds
to inflation in the interest rate equation. Its point estimate is about one half, so a point estimate for
its inverse is about 2.
We use nlcom to obtain an estimate of the parameter 1/β directly.
. nlcom 1/_b[beta]
_nl_1: 1/_b[beta]
_nl_1 1.955839 .2899594 6.75 0.000 1.387529 2.524149
Values for this parameter in the literature are typically around 1.5. Our estimated value is about 2,
but the common value of 1.5 lies within the confidence interval.
The phi parameter is a measure of price adjustment. The remaining four parameters describe the
evolution of the model’s exogenous state variables. Both state variables are highly persistent, with
persistence parameters of 0.70 and 0.96 on u and g, respectively. The estimated standard deviations
of the shocks are also displayed. The shock to the state variable u is denoted e.u and has standard
deviation 2.32. The shock to the state variable g is denoted e.g and has standard deviation 0.61.
An example: A linear DSGE model

Writing down linearized DSGEs
The model in (4a)–(8a) is nonlinear. To use dsge, we first write the model in its corresponding
linearized form, which we show below. Throughout this manual, we use lowercase letters to denote
deviations of variables from the steady state. The linearized versions of (4a)–(8a) are
xt = Et xt+1 − (rt − Et πt+1 − gt ) (4b)

πt = βEt πt+1 + κxt (5b)
1
rt = πt + ut (6b)
β
ut+1 = ρu ut + t+1 (7b)
gt+1 = ρg gt + ξt+1 (8b)
The new parameter κ is a function of the underlying parameter φ from (5a).

We have implicitly constrained some of the coefficients; for example, the coefficient on the interest
rate in the output equation is constrained to −1.
Specifying the model to dsge

The dsge command to estimate the parameters in the system of (4b)–(8b) is
. dsge (p = {beta}*F.p + {kappa}*x) ///
(x = F.x - (r - F.p - g), unobserved) ///
(r = (1/{beta})*p + u) ///
(F.u = {rhou}*u, state) ///
(F.g = {rhog}*g, state)
As with dsgenl, each equation is bound in parentheses. The equations look almost identical to
the system in (4b)–(8b). Because a model has as many variables as it has equations, each variable
will appear on the left-hand side of one and only one equation but can appear on the right-hand side
of as many equations as we like. Equations may be specified in any order.
We identify types of variables differently with dsge than we did with dsgenl. Equation options
that appear within the parentheses indicate which type of variable appears on the left-hand side of
the equation. Equations specified without any options are equations for observed control variables.
We add the unobserved option to specify that a left-hand-side variable is an unobserved control
variable. We use the state option to identify state variables.
In our model, p and r are the observed control variables pt and rt , and their equations are specified
without options. The output gap, xt , is an unobserved control variable, so we specify the equation
for x using the unobserved option.
u and g are the state variables ut and gt , and option state is included when we type their
equations. Recall that state variables are fixed in the current period, so we specify how they evolve
through time by modeling the one-period lead—hence, the F. on the left-hand side of each state
equation. The state equations specify how the state variable evolves as a function of the current state
variables and, possibly, the control variables.
The shocks t and ξt enter the system through the state equations of their corresponding variable.
By default, a shock is attached to each state equation. The number of shocks must be the same as
the number of observed control variables in the model.
When we typed
. dsge ... (F.u = {rhou}*u, state) ...
the underlying equation is what we wrote in (7b). If a state variable is treated as deterministic in
your model, then it will not have a shock. For example, capital accumulation is often treated as
deterministic. To include an equation for a state variable without a shock, we would include the
noshock option within the equation.
As with dsgenl, the operator F. can be applied to both state and control variables to denote the
expected future value of a variable. For example, we can type F.x to specify xt+1 . Unlike F. used
in other time-series applications, the value of xt+1 is not taken from the value of x at time t + 1 in
your data. Instead, it is an expectation of the value of x at time t + 1 based on your model. dsge
interprets equations with F. operators in the following manner. The equation
. dsge (p = {beta}*F.p + {kappa}*x) ...
is interpreted as
Et (pt − βpt+1 − κxt ) = 0
which is equivalent to (4b) above. In a linear DSGE model, it is immaterial whether the expectation
operator is thought of as applying to a variable, or to an equation, or indeed to the entire system of
equations jointly.
The parameters we want to estimate are bound in curly braces.
For more details on the dsge syntax, see [DSGE] Intro 2.
Parameter estimation and interpretation of linear DSGEs
We estimate the parameters of the model in (4b)–(8b). These equations are much discussed in the
monetary economics literature. Equation (4b) is known as the output-gap Euler equation. Equation (5b)
is known as a New Keynesian Phillips Curve, and the parameter κ is known as the slope of the
Phillips curve. In New Keynesian models, prices depend on output, and κ is a measure of that
dependence. Equation (6a) is known as a Taylor rule, after Taylor (1993). The coefficient on inflation
in a Taylor rule is a commonly discussed parameter. β has two roles in the model above. It relates
current inflation deviations to expected future inflation deviations, and it relates interest rate deviations
to inflation deviations.
. dsge (p = {beta}*F.p + {kappa}*x)

> (x = F.x - (r - F.p - g), unobserved)
> (r = (1/{beta})*p + u)
> (F.u = {rhou}*u, state)
> (F.g = {rhog}*g, state)
(setting technique to bfgs)
Iteration 0: Log likelihood = -13931.564
Iteration 1: Log likelihood = -1301.5118 (backed up)
(switching technique to nr)
DSGE model
/structural
beta .5112878 .0757909 6.75 0.000 .3627403 .6598353
kappa .1696301 .0475493 3.57 0.000 .0764353 .2628249
rhou .6989185 .0449192 15.56 0.000 .6108785 .7869585
rhog .9556407 .0181342 52.70 0.000 .9200984 .9911831
sd(e.u) 2.31759 .2988025 1.731948 2.903232

sd(e.g) .614735 .0973277 .4239761 .8054938
Two of the parameters have structural interpretations. The parameter kappa is the slope of the
Phillips curve. Theory predicts that this parameter will be positive, and indeed our estimate is positive.
The parameter beta is the inverse of the coefficient on inflation in the interest rate equation. We
can obtain an estimate of 1/β , which is interpreted as the degree to which the central bank responds
to movements in inflation, by using nlcom.
. nlcom 1/_b[beta]
_nl_1: 1/_b[beta]
_nl_1 1.955846 .2899255 6.75 0.000 1.387602 2.524089
A typical value for 1/β in the literature is 1.5. Our point estimate is around 2.
Postestimation
Several tools are available to help you analyze your model after you have estimated its parameters.
These tools are available after using either dsgenl or dsge. Below, we continue with the model we
fit using dsge. The results would be the same if you used these commands after dsgenl.
The matrix of parameters in the state-space form that specifies how the state variables affect the
control variables is known as the policy matrix. Each policy matrix parameter is the effect of a
one-unit shock to a state variable on a control variable.
We use estat policy to view these results.
. estat policy
Policy matrix
Delta-method
Coefficient std. err. z P>|z| [95% conf. interval]
p
u -.4170858 .0389324 -10.71 0.000 -.4933918 -.3407798
g .8818832 .2330568 3.78 0.000 .4251002 1.338666
x
u -1.580149 .3926325 -4.02 0.000 -2.349694 -.8106032
g 2.658658 .9045254 2.94 0.003 .8858213 4.431496
r
u .1842446 .0567979 3.24 0.001 .0729228 .2955665
g 1.724827 .2210255 7.80 0.000 1.291625 2.158029
Results are listed equation by equation. The first block is the policy equation for inflation p and
writes it as a function of the state variables alone. A unit shock to the state u reduces inflation by
an estimated 0.417, and a unit shock to g raises inflation by an estimated 0.882.
The matrix of parameters that specifies the dynamic process for the state variables is known as
the state transition matrix. The state transition equation relates the future values of the state variables
to their values in the current period. Each state transition matrix parameter is the effect of a one-unit
shock to a state variable on its one-period-ahead mean.
. estat transition
Transition matrix of state variables
Delta-method
F.u
u .6989185 .0449192 15.56 0.000 .6108785 .7869585
g 1.11e-16 . . . . .
F.g
u 0 (omitted)
g .9556407 .0181342 52.70 0.000 .9200984 .9911831
Note: Standard errors reported as missing for constrained transition matrix

values.
Both state variables are modeled as autoregressive processes, so the results in estat transition
repeat the estimates of rhou and rhog from the dsge output. In this case, the other entries in the
state transition matrix are 0 or differ from 0 only because of lack of numerical precision. In more
complicated models, such as those in which a state equation depends on a control variable, the state
transition matrix will contain new information about that state variable.
Impulse responses
The state-space form allows us to trace the path of a control or state in response to a shock to a
state. This path is known as an impulse–response function (IRF). irf after dsge or dsgenl estimates
IRFs, and it puts the named set of estimates into an .irf file, whose results can be displayed using
irf graph or irf table.
To graph the IRF, we first create the nkirf.irf file and set it as the active .irf file using the
irf set command.
. irf set nkirf.irf
(file nkirf.irf created)
(file nkirf.irf now active)
Next, we use irf create to estimate a complete set of impulse responses based on our dsge
command. A complete set of impulse responses is an impulse to each shock and the response of each
state and control variable to that impulse. For the model in this example, irf create generates an
impulse to e.u and e.g, then stores the response to that impulse on p, x, r, g, and u. The results
are stored in the nkirf.irf file.
. irf create model1
(file nkirf.irf updated)
We now use irf graph to plot the impulse responses. The impulse() and response() options
control which impulse and which responses are chosen. To view the response of p, x, r, and u to a
shock to u, we type
. irf graph irf, impulse(u) response(x p r u) byopts(yrescale)
model1, u, p model1, u, r
0 .6
.4
-.5
.2
-1
0
model1, u, u model1, u, x
3 0
2 -2
1 -4
0 -6
0 2 4 6 8 0 2 4 6 8
Step
95% CI Impulse–response function
Graphs by irfname, impulse variable, and response variable
The state variable u models movements in the interest rate that occur for reasons other than the
feedback between inflation and the interest rate. A shock to u is effectively a surprise increase in the
interest rate, and the IRF traces out how this shock causes temporary decreases to inflation (top-left
graph) and to the output gap (bottom-right graph).
Forecasts
The forecast suite of commands produces dynamic forecasts from the fitted model.
We first store the dsge estimation results.
. estimates store dsge_est
We use tsappend() to extend the dataset by 3 years, or 12 quarters.

. tsappend, add(12)
To set up a forecast, we perform three steps. First, we type forecast create to initialize a new
forecasting model, which we name dsgemodel.
. forecast create dsgemodel
Forecast model dsgemodel started.
Next, we add the estimates from dsge to the forecasting model using forecast estimates.
. forecast estimates dsge_est
Added estimation results from dsge.
Forecast model dsgemodel now contains 2 endogenous variables.
This command adds the estimates stored in dsge est to the model dsgemodel. Finally, we produce
dynamic forecasts beginning in the first quarter of 2017 using forecast solve. The prefix(d1 )
option specifies the d1 prefix that will be given to the variables created by forecast. We also
request that dynamic forecasts begin in the first quarter of 2017 with the begin(tq(2017q1)) option.
. forecast solve, prefix(d1_) begin(tq(2017q1))

Computing dynamic forecasts for model dsgemodel.
Starting period: 2017q1

Ending period: 2020q1
Forecast prefix: d1_
2017q1: ..............
2017q2: ..............
2017q3: ..............
2017q4: ..............
2018q1: ..............
2018q2: .............
2018q3: .............
2018q4: .............
2019q1: .............
2019q2: .............
2019q3: .............
2019q4: .............
2020q1: .............
Forecast 2 variables spanning 13 periods.
The dynamic forecast begins in the first quarter of 2017, so all forecasts are out of sample.
We can graph the forecast for inflation d1 p using tsline.
. tsline d1_p if tin(2010q1, 2021q1), tline(2017q1)
3
Inflation rate (dsgemodel d1_)
0
2010q1 2012q3 2015q1 2017q3 2020q1
Quarterly date
The model forecasts that inflation will smoothly return to its long-run value, the sample mean.
We can also begin the forecast during a time period for which observations are available.
Specifying the begin(tq(2014q1)) option produces dynamic forecasts beginning in the first
quarter of 2014, so we can compare the forecast for 2014–2016 with the actual observations over
that period.
. forecast solve, prefix(d2_) begin(tq(2014q1))
Computing dynamic forecasts for model dsgemodel.
Starting period: 2014q1

Ending period: 2020q1
Forecast prefix: d2_
2014q1: ..............
2014q2: ..............
2014q3: ..............
2014q4: ..............
2015q1: ..............
2015q2: ..............
2015q3: ..............
2015q4: .............
2016q1: .............
2016q2: .............
2016q3: .............
2016q4: .............
2017q1: .............
2017q2: .............
2017q3: .............
2017q4: .............
2018q1: .............
2018q2: .............
2018q3: .............
2018q4: .............
2019q1: .............
2019q2: .............
2019q3: .............
2019q4: .............
2020q1: .............
Forecast 2 variables spanning 25 periods.
. tsline p d2_p if tin(2010q1, 2021q1), tline(2014q1)

We plot both the observed inflation and the forecast.

3
0
2010q1 2012q3 2015q1 2017q3 2020q1
Quarterly date
Inflation rate Inflation rate (dsgemodel d2_)
The forecast captures the general upward trend in inflation from 2014–2016, but it does not predict
the variation in inflation around the upward trend.
Structural and reduced forms of DSGE models

Now that we have worked an example, we show how it fits in the more general formulation
of DSGE models. The model in (4a)–(8a) is an example of a nonlinear DSGE model. In general, a
nonlinear DSGE model can be expressed as
Et {f (xt+1 , yt+1 , xt , yt , θ)} = 0 (9)
where f is a vector of equations, xt is a vector of state variables, yt is a vector of control variables, and
θ denotes the vector of structural parameters. To solve a DSGE model means to write it in state-space
form. In general, the state-space form of a nonlinear DSGE model is nonlinear. dsgenl takes a linear
approximation to the model equations to solve for a linear approximation to the state-space form.
The model in (4b)–(8b) is an example of a linearized DSGE model. In general, a linearized DSGE
model can be expressed as
A0 yt = A1 Et (yt+1 ) + A2 yt + A3 xt (10)
B0 xt+1 = B1 Et (yt+1 ) + B2 yt + B3 xt + Ct+1 (11)
where as before yt is a vector of control variables, xt is a vector of state variables, and t is a

vector of shocks. A0 through A3 and B0 through B3 are matrices of parameters. We require that
A0 and B0 be diagonal matrices. The entries in all of these matrices are functions of the structural
parameters, which we denote by vector θ. Economic theory places restrictions on the entries in the
matrices. C is a selection matrix that determines which state variables are subject to shocks.
The reduced form of a DSGE model expresses the control variables as functions of the state variables
alone and specifies how state variables evolve over time. The reduced form is a state-space model.
The state-space form of the model is given by
yt = Gxt (12)
xt+1 = Hxt + Mt+1 (13)
where yt is a vector of control variables, xt is a vector of state variables, and t is a vector of

shocks. G is the policy matrix, and H is the state transition matrix. M is diagonal and contains the
standard deviations of the shocks.
yt is partitioned into observed and unobserved controls, yt = (y1,t , y2,t ). The observed control
variables are related to the control variables by the equation
y1,t = Dyt
where D is a selection matrix. Only observed control variables play a role in estimation. The number
of observed control variables must be the same as the number of state equations that include shocks.
You specify a model of the form (9) to dsgenl and a model of the form (10)–(11) to dsge. Many
models require some manipulation to fit into the structure in (9) and (10); see [DSGE] Intro 4 for
details. Postestimation commands estat policy and estat transition will display the policy
and transition matrices in (12) and (13), respectively.
References
Canova, F. 2007. Methods for Applied Macroeconomic Research. Princeton, NJ: Princeton University Press.
DeJong, D. N., and C. Dave. 2011. Structural Macroeconometrics. 2nd ed. Princeton, NJ: Princeton University Press.
Fernández-Villaverde, J., J. F. Rubio-Ramı́rez, and F. Schorfheide. 2016. Solution and estimation methods for DSGE
models. In Vol. 2A of Handbook of Macroeconomics, ed. J. B. Taylor and H. Uhlig, chap. 9, 527–724. Amsterdam:
North-Holland. https://doi.org/10.1016/bs.hesmac.2016.03.006.
Ljungqvist, L., and T. J. Sargent. 2018. Recursive Macroeconomic Theory. 4th ed. Cambridge, MA: MIT Press.
Taylor, J. B. 1993. Discretion versus policy rules in practice. Carnegie-Rochester Conference Series on Public Policy
39: 195–214. https://doi.org/10.1016/0167-2231(93)90009-L.
Woodford, M. 2003. Interest and Prices: Foundations of a Theory of Monetary Policy. Princeton, NJ: Princeton
University Press.
Also see
[DSGE] dsge — Linear dynamic stochastic general equilibrium models
[DSGE] dsge postestimation — Postestimation tools for dsge
[DSGE] dsgenl — Nonlinear dynamic stochastic general equilibrium models
[DSGE] dsgenl postestimation — Postestimation tools for dsgenl
[DSGE] Intro 9 — Bayesian estimation
[TS] forecast — Econometric model forecasting
[TS] irf — Create and analyze IRFs, dynamic-multiplier functions, and FEVDs
[TS] sspace — State-space models
Title
Intro 2 — Learning the syntax
Description
In this introduction, we demonstrate how to specify a DSGE model using the dsge and dsgenl
commands. We focus on two unique aspects of DSGEs that must be considered when writing the
syntax—writing the system of equations and identifying each type of variable within those equations.

Introduction
Syntax for linear DSGE models
Preview of dsge syntax
Specifying the system of linear equations
Control variables
State variables and shocks
Expectations of future values of control variables
Specifying parameters using dsge’s substitutable expressions
Syntax for nonlinear DSGE models
Preview of dsgenl syntax
Specifying the system of nonlinear equations
State and control variables
Expectations in nonlinear models
Introduction
If you have not read [DSGE] Intro 1, we recommend that you read it first. In particular, see
Structural and reduced forms of DSGE models in [DSGE] Intro 1, where we discuss the structural
forms of a DSGE model that are required by dsge and dsgenl. Below, we assume that your model
has this structural form, and we discuss the syntax for specifying such a model with the dsge and
dsgenl commands.
dsge estimates the parameters of linear DSGE models. dsgenl estimates the parameters of nonlinear
DSGE models. The two commands have elements in common. Both center around a list of equations
and a collection of variables. The two commands differ somewhat in how they expect you to specify
equations and variables. Below, we first describe the syntax for dsge and then for dsgenl.
20
Intro 2 — Learning the syntax 21
Syntax for linear DSGE models
Preview of dsge syntax
As an example, if we wanted to fit the DSGE model
pt = βEt (pt+1 ) + κyt

yt = Et (yt+1 ) − {rt − Et (pt+1 ) − ρzt }
βrt = pt + βut
zt+1 = ρzt + t+1
ut+1 = δut + ξt+1
we would type
. dsge (p = {beta}*F.p + {kappa}*y) ///
(y = F.y -(r - F.p - {rho}*z), unobserved) ///
({beta}*r = p + {beta}*u) ///
(F.z = {rho}*z, state) ///
(F.u = {delta}*u, state)
The syntax looks a lot like the equations. In what follows, we discuss each element of the dsge
syntax that you will need to specify the equations of your DSGE model.
In [DSGE] Intro 1, we told you that three types of variables—control variables, state variables,
and shocks—appear in DSGE models. We will demonstrate how to specify equations involving each
of these types of variables. We will also show you how to specify the required types of equations
that are linear in these variables and nonlinear in the parameters.
Specifying the system of linear equations
For a linear DSGE model, we specify one equation for each control variable and one equation for
each state variable. The equations are bound by parentheses. Therefore, the basic structure of the
dsge command is
. dsge (eq1) (eq2) ..., ...
where eq1 and eq2 will be replaced with the syntax representing one of the equations in our model.
We use . . . to indicate that we can include more than two equations as well as options.
The basic form of an equation within the parentheses is
term = term [ + term [ + term [. . .]]]
where each term includes a variable name. The variable name may be preceded by a parameter or
nonlinear combination of parameters. For instance, a valid equation might look something like
(y = {kappa}*z)
or
(y = {kappa}*z + {kappa}*{beta}*x)
or
(1/{beta}*y = {gamma}*z + x)
22 Intro 2 — Learning the syntax
We will explain this later. For now, simply note that these equations are written in dsge syntax
much like we would write the math, but the parameters we want to estimate are offset with curly
braces, {}. If you have used any other Stata commands that work with substitutable expressions, you
may recognize this notation. In fact, dsge uses a special form of substitutable expressions.
Before we discuss how to use dsge’s substitutable expressions, we will focus on how to specify
each type of variable.
Control variables
Control variables can be modeled as a function of other control variables, expectations of the future
value of control variables, and state variables. The equations for control variables do not include
shocks.
We continue from the basic dsge command in the previous section, which was
. dsge (eq1) (eq2) ..., ...
If we have an observed control variable y , and for eq1, we want to specify that yt be modeled as
a function of zt and xt , we can type this equation in our dsge command as
. dsge (y = pexp*z + pexp*x) ///
(eq2) ///
..., ...
where pexp is a possibly nonlinear expression of parameters in each term.
Control variables can be observed or unobserved. Because we did not include any options within
this set of parentheses, y is assumed to be an observed control variable. If y were unobserved, we
would add the unobserved option as follows:
(y = pexp*z + pexp*x, unobserved)
Note that each control variable in the model must be included on the left-hand side of one, and
only one, equation.
State variables and shocks

To model a state variable, we specify an equation with the one-period lead of that state variable
on the left-hand side. On the right-hand side of the equation, we can include state variables, control
variables, expectations of the future value of control variables, and shocks.
When we specify an equation for a state variable, we include the state option within the
parentheses defining the equation.
Continuing with the syntax above, let’s suppose eq2 is an equation for a state variable x, and we
model xt+1 as a function of xt . We expand our dsge command to
. dsge (y = pexp*z + pexp*x) ///
(F.x = pexp*x, state) ///
..., ...
We used the F. lead operator to specify the one-period lead of x as F.x; see [U] 11.4.4 Time-
series varlists. However, notice that the full list of time-series operators is not available here. We can
specify only equations for one-period leads of state variables; we could not replace F.x with F2.x
in the equation above. Note that this restriction does not limit the types of models we can fit; see
[DSGE] Intro 4c.
By default, the equation for a state variable includes an unobserved shock. However, equations
for state variables are not required to include a shock. Within the system of equations, the number
of shocks should be equal to the number of observed control variables. If we did not wish to include
a shock in the equation for xt+1 , we could add the noshock option,
(F.x = pexp*x, state noshock)
Note that the one-period lead of each state variable must be included on the left-hand side of one,
and only one, equation.
Expectations of future values of control variables

Expectations of the one-period lead of control variables can appear in equations for both control
and state variables. Mathematically, we write these expectations as Et (·). For instance, we write the
expectation of y in time t + 1 as Et (yt+1 ). In dsge, we write this expectation as F.y. Any time
dsge sees the F. operator applied to a control variable, it interprets this as the expectation of that
variable at time t + 1.
If we model yt as a function of Et (yt+1 ) in addition to zt and xt , we can expand our previous
dsge command to
. dsge (y = pexp*F.y + pexp*z + pexp*x) ///
..., ...
Expectations are strictly for one-period leads of control variables. You cannot, for instance, use
F2.y to include the expectation of y in time t + 2 in the model. This does not prevent you, however,
from including such terms in your model. See [DSGE] Intro 4d for details of fitting models, including
expectations of control variables more than one period in the future.
Specifying parameters using dsge’s substitutable expressions

At this point, we know how to specify each type of variable that may arise in our DSGE model.
We now turn to specifying the parameters that we want to estimate.
Recall that the basic form of an equation is
term = term [ + term [ + term [. . . ]]]
where each term includes one variable name. The variable name may be preceded by a parameter
or a nonlinear combination of parameters. For terms on the right-hand side of the equation, variable
names can also be followed by a parameter specification.
We specify the terms using a special type of substitutable expressions that we call scalar substitutable
expressions. In scalar substitutable expressions, parameters are enclosed in curly braces, {}, and may
enter the model either linearly or nonlinearly. The restriction that each term includes only one variable
implies that the variables must enter the equation linearly with these scalar substitutable expressions.
If our model for y in the equation above is
yt = βEt (yt+1 ) + κzt + γxt
we could extend our previous dsge command as follows:

. dsge (y = {beta}*F.y + {kappa}*z + {gamma}*x) ///
..., ...
However, the equation for y need not be linear in the parameters. If instead we wanted to model
y as
yt = (1/β)Et (yt+1 ) + κzt + (γ/β)xt
we would change our command to

. dsge (y = 1/{beta}*F.y + {kappa}*z + ({gamma}/{beta})*x) ///
..., ...
We can even include parameters on the left-hand side of the equation. For instance, we could
model y as
(1/β)yt = Et (yt+1 ) + κzt + (γ/β)xt
and change our dsge command to

. dsge ((1/{beta})*y = F.y + {kappa}*z + ({gamma}/{beta})*x) ///
..., ...
Finally, there is an extension to the basic form of the equation that we should mention. Sometimes,
it is more convenient to write an equation so that a scalar value or a parameter or a nonlinear
combination of parameters is multiplied by a linear combination of terms. The dsge command allows
this. For instance, instead of
yt = (1/β)Et (yt+1 ) + (γ/β)xt + κzt
we can write
yt = (1/β){Et (yt+1 ) + γxt } + κzt
Similarly, we could write this equation in the dsge command as

(y = (1/{beta}) * (F.y + {gamma}*x) + {kappa}*z)
Now you know the general syntax of dsge and are ready to fit your own DSGE model with dsge.
For examples of fitting classic DSGE models using this syntax, see [DSGE] Intro 3a, [DSGE] Intro 3b,
and [DSGE] Intro 3c. For more details on the dsge syntax, see [DSGE] dsge.
Syntax for nonlinear DSGE models

The dsgenl syntax builds on the dsge syntax. However, because of the additional complexity of
nonlinear models, the syntax does differ. As we discuss the dsgenl syntax below, we will highlight
similarities and differences from dsge.
Preview of dsgenl syntax
Suppose we wanted to estimate the parameters of the nonlinear DSGE model,
Yt
Rt = α
Kt

Ct
1 = βEt (1 + Rt+1 − δ)
Ct+1
Yt = Zt Ktα
Kt+1 = Yt − Ct + (1 − δ)Kt
ln(Zt+1 ) = ρ ln(Zt ) + et+1
To estimate the parameters of this model, we would type

. dsgenl (r = {alpha}*y/k) ///
(1 = {beta}*(c/F.c)*(1 + F.r - {delta})) ///
(y = z*k^({alpha})) ///
(F.k = y - c + (1-{delta})*k) ///
(ln(F.z) = {rho}*ln(z)) , ///
observed(y) unobserved(c r) exostate(z) endostate(k)
As with dsge, the dsgenl syntax looks a lot like the equations. Variables are written as normal,
and parameters are written bound by curly braces. We first focus on two syntax elements: the system
of equations that specifies the interrelations among variables and the options that specify how each
variable is treated.
Specifying the system of nonlinear equations
For a nonlinear DSGE model, we specify a system of equations that jointly determines the structure
of the model. There must be as many equations as there are control and state variables. The equations
are bound by parentheses. Therefore, the basic structure of the dsgenl command is
. dsgenl (eq1) (eq2) ..., ...
where eq1 and eq2 will be replaced with the syntax representing one of the equations in our model.
We use . . . to indicate that we can include more than two equations as well as options.
The equations for a nonlinear model fit with dsgenl take on a much more flexible form than
the equations we described for a linear model fit by dsge. For dsgenl, the equation within the
parentheses may include a nonlinear substitutable expression on both sides of the equal sign. An
example of such an expression is the first equation in the model above, which reads
(r = {alpha}*y/k)
The equations written in dsgenl, like those in dsge, use curly braces to identify the parameters such
as alpha that are to be estimated.
This equation could have been expressed in other ways. As written above, it reads like an equation
for r. We could just as well have written it as
(k = {alpha}*y/r)
or
(r*k/y = {alpha})
or even
(r*k/y - {alpha} = 0)
All four of the above Stata equations specify the same mathematical equation, and it does not matter
which parameterization you choose.
The second equation in the model above reads
(1 = {beta}*(c/F.c)*(1 + F.r - {delta}))
We could have written it as

(1/c = {beta}*(1/F.c)*(1 + F.r - {delta}))
or
(1 = {beta}*(F.c/c)^(-1)*(1 + F.r - {delta}))
The three equations just listed specify the same mathematical equation, and it does not matter which
parameterization you choose.
In the second equation of the model, the F. lead operator denotes the future value of a variable,
just as it did in dsge. More information on how dsgenl treats the F. operator is below, under
Expectations in nonlinear models.
State and control variables

So far, we have discussed only the specification of model equations. We have seen that you can
specify the model equations in nearly any form. The price of that freedom is that you must specify a
set of options—exostate(), endostate(), observed(), and unobserved()—that tell Stata what
role each variable plays in the model.
Recall that a DSGE model has control and state variables. Each control variable may be observed
or unobserved, and each state variable may be subject to shocks. Hence, each variable fits into one
of four types: observed controls, unobserved controls, a state variable that is not subject to a shock,
and a state variable that is subject to a shock. The above model has variables of all four types.
Using the jargon common in macroeconomics, we call a state variable that is subject to a shock
an exogenous state variable. We call a state variable that is not subject to a shock an endogenous
state variable. You must have exactly as many exogenous state variables (and thus, as many shocks)
as you have observed control variables.
In the model above, the option list
. dsgenl . . . , . . . observed(y) unobserved(c r) exostate(z) endostate(k) . . .
specifies that c, r, and y are control variables; of those, y is observed. The option list specifies k
and z as state variables, of which z is subject to shocks.
Expectations in nonlinear models

In Expectations of future values of control variables, we saw that terms in models fit with dsge can
include the expectation of the one-period lead of a variable. Such expectations can also be included
in models fit with dsgenl. Because the equations in dsgenl are more flexible, we need to think of
the expectations in a more general way, as we describe below.
Expectations of the one-period lead of variables can appear in any equation. Mathematically, we
write these expectations as Et (·). For instance, we write the expectation of y in time t + 1 as
Et (yt+1 ). As we did in dsge, we write this expectation as F.y.
In a nonlinear model, however, the expectation may not be applied directly to one variable. We might
have

Ct
1 = βEt (1 + Rt+1 − δ)
Ct+1
where the expectation is written as if it is applied to everything in the curly brackets. Really, the
expectation can be thought of as being applied to the entire equation. This means we can use the
F. lead operator and write the equation involving the expectation above as
. dsgenl (1 = {beta}*(c/F.c)*(1 + F.r - {delta})) ..., ...
in dsgenl, and it will be interpreted as

Ct
Et 1 − β (1 + Rt+1 − δ) = 0
Ct+1
The upshot is that you can specify expected future values by using the F.x notation, and both dsgenl
and dsge will take the appropriate action.
Now you have seen all elements of the general syntax dsgenl and are ready to fit your own
model. For examples of fitting classic nonlinear DSGE models, see [DSGE] Intro 3d, [DSGE] Intro 3e,
and [DSGE] Intro 3f. For more details on the dsgenl syntax, see [DSGE] dsgenl.
Also see
[DSGE] Intro 3 — Classic DSGE examples
[DSGE] Intro 4 — Writing a DSGE in a solvable form
Title
Intro 3 — Classic DSGE examples
Description
In this entry, we present several classic DSGE examples. These include linear and nonlinear versions
of a New Keynesian model, linear and nonlinear versions of a New Classical model (Real Business
Cycle model), a linear financial frictions model, and a nonlinear stochastic growth model.

In [DSGE] Intro 3a, [DSGE] Intro 3b, and [DSGE] Intro 3c, we fit simple variants of common
linearized DSGE models. In [DSGE] Intro 3d, [DSGE] Intro 3e, and [DSGE] Intro 3f, we fit simple
variants of common nonlinear DSGE models. Through these examples, we demonstrate model solution,
estimation, and interpretation.
[DSGE] Intro 3a demonstrates how to fit a New Keynesian model. In this example, we interpret
structural parameters, policy matrix parameters, and state transition matrix parameters. We also predict
values of both observed control variables and unobserved states.
[DSGE] Intro 3b illustrates how to solve a New Classical model and plot the IRFs to compare the
model’s theoretical predictions under different sets of parameter values.
[DSGE] Intro 3c fits a financial frictions model. In this example, we also estimate parameters of
the policy matrix and evaluate the IRFs.
[DSGE] Intro 3d revisits the New Keynesian model. In this example, we estimate the parameters
of the nonlinear model, interpret structural parameters, and explain partial identification by parameter
restrictions in a DSGE model.
[DSGE] Intro 3e revisits the New Classical model. We fix some parameters and estimate others,
explain the effect of fixing parameters on postestimation statistics, and explain how to compare models
across different parameter settings.
[DSGE] Intro 3f demonstrates how to solve a nonlinear stochastic growth model. We explain how
dsgenl takes an approximation of the model and how to interpret the steady state and approximations
to the policy and transition equations. This example describes some of the technical differences
between linear and log-linear approximations to nonlinear DSGE models.
Also see
28
Title
Intro 3a — New Keynesian model
Description Remarks and examples Reference Also see
Description
This introduction estimates and interprets the parameters of a simple New Keynesian model. In
this entry, we demonstrate how to constrain parameters in the model and how to interpret structural
parameters, policy matrix parameters, and state transition matrix parameters. We also predict values
of both observed control variables and unobserved states. For Bayesian analysis of this model, see
[DSGE] Intro 9a.

The model
Parameter estimation
One-step-ahead predictions
Estimating an unobserved state
The model
Equations (1)–(5) specify a canonical New Keynesian model of inflation pt , the output gap xt , and
the interest rate rt . These are the linearized equations; the model’s nonlinear equations are similar to
those in Writing down nonlinear DSGEs in [DSGE] Intro 1.
pt = βEt (pt+1 ) + κxt (1)

xt = Et (xt+1 ) − {rt − Et (pt+1 ) − gt } (2)
rt = ψpt + ut (3)
ut+1 = ρu ut + t+1 (4)
gt+1 = ρg gt + ξt+1 (5)
Equation (1) specifies inflation as a linear combination of expected future inflation and the output gap.
Equation (2) specifies the output gap as a linear combination of the expected future output gap, the real
interest rate, and a state variable gt . Equation (3) specifies the interest rate as a linear combination
of inflation and a state variable ut . The state variables are modeled as first-order autoregressive
processes. The state variable ut is the deviation of rt from its equilibrium value of ψpt . The state
variable gt is also the deviation of xt from its equilibrium value.
Three of the parameters have structural interpretation. The parameter κ is known as the slope
of the Phillips curve and is predicted to be positive. The parameter β is the discount factor that
represents the degree to which agents discount the future relative to the current period. The parameter
ψ measures the degree to which interest rates react to movements in inflation.
29
30 Intro 3a — New Keynesian model
Not all model parameters are identified. We constrain β to be 0.96, a common value in the
literature. The remaining parameters are identified.
. use https://www.stata-press.com/data/r18/usmacro2
. constraint 1 _b[beta]=0.96

> (x = F.x - (r - F.p - g), unobserved)
> (r = {psi}*p + u)
> (F.g = {rhog}*g, state),
> from(psi=1.5) constraint(1)
Iteration 6: Log likelihood = -765.73779 (not concave)
DSGE model
( 1) [/structural]beta = .96
/structural
beta .96 (constrained)
kappa .0849631 .0287693 2.95 0.003 .0285763 .14135
psi 1.943004 .2957869 6.57 0.000 1.363272 2.522735
rhou .7005483 .0452604 15.48 0.000 .6118397 .789257
rhog .9545257 .0186424 51.20 0.000 .9179873 .9910641
sd(e.u) 2.318204 .3047434 1.720918 2.91549

sd(e.g) .5689891 .0982975 .3763296 .7616486
The slope of the Phillips curve, kappa, is estimated to be positive. The coefficient on inflation in
the interest rate equation is estimated to be almost 1.94, meaning that interest rates are expected to
rise almost two for one with increases in inflation.
Intro 3a — New Keynesian model 31

Elements of the policy matrix represent the response of a control variable to a one-unit increase
in a state variable.
. estat policy
Policy matrix
Delta-method
p
u -.4172521 .0393609 -10.60 0.000 -.494398 -.3401061
g .9678177 .2777452 3.48 0.000 .4234472 1.512188
x
u -1.608216 .405054 -3.97 0.000 -2.402107 -.8143245
g .9529203 .4813931 1.98 0.048 .0094071 1.896433
r
u .1892776 .059166 3.20 0.001 .0733143 .3052409
g 1.880474 .2616 7.19 0.000 1.367747 2.3932
An increase in u decreases the extent to which the inflation is above its short-run equilibrium
value. This change decreases the output gap and increases interest rate.
An increase in g increases the extent to which the inflation is above its short-run equilibrium value.
This change also increases output gap and interest rates.
Because the states are uncorrelated with each other in this example, the elements of the state
transition matrix are just the persistence parameters in the model.
. estat transition
Delta-method
F.u
u .7005483 .0452604 15.48 0.000 .6118397 .789257
g 3.33e-16 . . . . .
F.g
u 0 (omitted)
g .9545257 .0186424 51.20 0.000 .9179873 .9910641

values.
See [DSGE] Intro 4g for an example in which the state variables depend on each other. The states
can also depend on each other when a state variable is specified to depend on control variables, as in
[DSGE] Intro 3b, or when the state vector includes lagged control variables, as in [DSGE] Intro 4a.
32 Intro 3a — New Keynesian model
One-step-ahead predictions
Predictions after dsge depend on the estimated state-space parameters. Below, we obtain one-
step-ahead predictions for each of the two observed control variables in the model, and we graph the
actual and predicted inflation rates.
. predict dep*
(option xb assumed; fitted values)
. tsline p dep1, legend(col(1))
15
10
1955q3 1970q3 1985q3 2000q3 2015q3

Date (quarters)
Growth rate of prices (GDPDEF)

xb prediction, p, onestep
The graph shows that the one-step-ahead predictions closely follow realized inflation.
Estimating an unobserved state

The observed control variables are driven by two unobserved state variables. We can use predict
with the state option to estimate the state variables.
Intro 3a — New Keynesian model 33
Here we estimate the unobserved state u and plot it.

. predict state1, state
. tsline state1
xb states prediction, u, onestep
-5
-10
1955q3 1970q3 1985q3 2000q3 2015q3
Date (quarters)
The loose monetary policy in the mid-1970s, where predicted values of ut are negative, and the
subsequent Volcker contraction are apparent in this plot.
Reference
Schenck, D. 2017. Estimating the parameters of DSGE models. The Stata Blog: Not Elsewhere Classified.
https://blog.stata.com/2017/07/11/estimating-the-parameters-of-dsge-models/.
Also see
[DSGE] Intro 3b — New Classical model
[DSGE] Intro 3c — Financial frictions model
[DSGE] Intro 3d — Nonlinear New Keynesian model
[DSGE] Intro 9a — Bayesian estimation of a New Keynesian model
Title
Intro 3b — New Classical model
Description
In this example, we solve a New Classical model similar to the one in King and Rebelo (1999). We
also demonstrate how to compare a model’s theoretical predictions under different parameter values
using IRFs.

The model
Solving the model
Impulse responses
Sensitivity analysis
The model
In this model, output, consumption, investment, employment, and other variables are driven by state
variables linked to production and demand. The model is similar to the one in King and Rebelo (1999)
and is referred to as a real business cycle model.
The nonlinear form of the model is

1 1
= βEt (1 + Rt+1 − δ) (1)
Ct Ct+1
Wt
Htη = (2)
Ct
Yt = Ct + Xt + Gt (3)
Yt = Ktα (Zt Ht )1−α (4)
Yt
Wt = (1 − α) (5)
Ht
Yt
Rt = α (6)
Kt
Kt+1 = (1 − δ)Kt + Xt (7)
Equation (1) specifies consumption Ct as a function of expected future consumption and the expected
future interest rate Et (Rt+1 ). Equation (2) specifies labor hours Ht as a function of the wage Wt and
consumption; it is a labor supply equation. Equation (3) is the national income accounting identity for
a closed economy, specifying output Yt as the sum of consumption, investment Xt , and government
spending Gt . Equation (4) is a production function that specifies output as a function of labor input
Ht , capital input Kt , and productivity Zt . Equations (5) and (6) specify labor demand and capital
demand, respectively. Equation (7) specifies the equation for capital accumulation. The model is
completed when we add state transition equations for Zt and Gt . These state transition equations are
conventionally specified after the model has been linearized.
34
Intro 3b — New Classical model 35
The linearized form of the model is
ct = Et (ct+1 ) − (1 − β + βδ)Et (rt+1 )

ηht = wt − ct
φ 1 xt = yt − φ2 ct − gt
yt = (1 − α)(zt + ht ) + αkt
wt = yt − ht
rt = yt − kt
kt+1 = δxt + (1 − δ)kt
zt+1 = ρz zt + t+1
gt+1 = ρg gt + ξt+1
The model has six control variables and three state variables. Two of the state variables, zt+1 and
gt+1 , are modeled as first-order autoregressive processes. The state equation for kt+1 depends on the
current value of a control variable, namely, xt .
Solving the model

The solve option of dsge places the model in state-space form without estimating parameters; it
is similar to iterate(0) but is faster because it does not calculate standard errors. Using solve for
different parameter values of your model is a useful way to explore the model’s theoretical properties.
The parameter values used here are similar to those used in King and Rebelo (1999). Each has
an interpretation. (1 − alpha) is labor’s share of national income. delta is the depreciation rate of
capital. eta is the slope of the labor supply curve. phi1 and phi2 are share parameters related to
investment’s share of national income and consumption’s share of national income, respectively. rhoz
and rhog are autoregressive parameters on the state variables.
36 Intro 3b — New Classical model
. dsge (c = F.c - (1-{beta}+{beta}*{delta})*F.r, unobserved)
> ({eta}*h = w - c, unobserved)
> ({phi1}*x = y - {phi2}*c - g, unobserved)
> (y = (1-{alpha})*(z+h) + {alpha}*k)
> (w = y - h, unobserved)
> (r = y - k, unobserved)
> (F.k = {delta}*x+ (1-{delta})*k, state noshock)
> (F.z = {rhoz}*z, state)
> from(beta=0.96 eta=1 alpha=0.3 delta=0.025 phi1=0.2 phi2=0.6 rhoz=0.8 rhog=0.3)
> solve noidencheck
DSGE model
y Coefficient Std. err. z P>|z| [95% conf. interval]
/structural
beta .96 . . . . .
delta .025 . . . . .
eta 1 . . . . .
phi1 .2 . . . . .
phi2 .6 . . . . .
alpha .3 . . . . .
rhoz .8 . . . . .
rhog .3 . . . . .
sd(e.z) 1 . . .
sd(e.g) 1 . . .
Note: Skipped identification check.

Note: Model solved at specified parameters.
The solve option solves the model at the specified values in from(). We skip the identification
check with noidencheck. Simply solving the model does not involve any reference to the data or
any estimation. Still, we can explore what these parameters imply.

After solving, we can use many of the postestimation commands, though standard errors will be
missing throughout.
The state transition matrix shows how the state vector in the next period is related to the state
vector in the current period. Some state variables are specified as first-order autoregressive processes,
and their transition equations will simply repeat information that is already available in the estimation
table. However, if any state variable equation contains control variables, then that state variable’s
transition equation will depend on the other state variables.
. estat transition
Delta-method
F.k
k .9256785 . . . . .
z .1078282 . . . . .
g -.1070547 . . . . .
F.z
k 0 (omitted)
z .8 . . . . .
g 0 (omitted)
F.g
k 0 (omitted)
z 0 (omitted)
g .3 . . . . .

values.
The value of the state variables z and g in the next period depends only on their value in the
current period, but the value of the capital stock k in the next period depends on the current value
of all three state variables. This feature means that, for example, a shock to the z state variable has
two effects: it increases future values of z, because z is autoregressive, but it also increases future
values of k. Interrelationships among the state variables can generate more interesting patterns in the
IRFs than the AR(1) dynamics that we saw in [DSGE] Intro 3a.
Impulse responses
One way to compare two parameter sets is to graph the impulse response of model variables to a
shock under each parameter set. We first set the impulse–response file with irf set and then add
impulse responses named persistent to the file with irf create.
. irf set rbcirf
(file rbcirf.irf created)
(file rbcirf.irf now active)
. irf create persistent

(file rbcirf.irf updated)
The response of model variables to a shock to z is graphed by typing

. irf graph irf, irf(persistent) impulse(z) response(y c x h w z) noci
> byopts(yrescale)
persistent, z, c persistent, z, h persistent, z, w

.3 .4 .6
.3 .5
.25 .2
.4
.1
.3
.2 0
.2
persistent, z, x persistent, z, y persistent, z, z

1 1
4
.8 .8
3
.6 .6
2
.4 .4
1
.2 .2
0
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8
Step
Each graph is labeled with the IRF name, the impulse variable, and the response variable. For
instance, the top-left graph shows the response of consumption to a shock to the zt state variable. The
bottom-right graph shows the response of the state variable zt itself. The state is persistent, which is
not surprising: we set the autoregressive parameter in the zt equation to 0.8.
In the top-left graph, we see that consumption c rises over time before returning to steady state. The
time unit is quarters, so a value of about 0.27, 4 periods after the shock, indicates that consumption
is 0.27% above its steady-state value one year after the shock. Hours worked h are shown in the
top center graph and rise initially before falling below steady state. The real wage w, output y, and
investment x all rise.
The responses of variables to a shock to z are persistent. Some variables, like consumption and
the wage, show dynamics beyond the simple autoregressive behavior of z itself. To evaluate the role
of persistence in z on the persistence of other model variables, we rerun the dsge command. This
time, we set the persistence of z to a smaller value of 0.6.
. dsge (c = F.c - (1-{beta}+{beta}*{delta})*F.r, unobserved)

> ({eta}*h = w - c, unobserved)
> ({phi1}*x = y - {phi2}*c - g, unobserved)
> (y = (1-{alpha})*(z+h) + {alpha}*k)
> (w = y - h, unobserved)
> (r = y - k, unobserved)
> (F.k = {delta}*x+ (1-{delta})*k, state noshock)
> from(beta=0.96 eta=1 alpha=0.3 delta=0.025 phi1=0.2 phi2=0.6 rhoz=0.6
> rhog=0.3) solve noidencheck
DSGE model
/structural
beta .96 . . . . .
delta .025 . . . . .
eta 1 . . . . .
phi1 .2 . . . . .
phi2 .6 . . . . .
alpha .3 . . . . .
rhoz .6 . . . . .
rhog .3 . . . . .
sd(e.z) 1 . . .
sd(e.g) 1 . . .

The only change in the parameter set is that rhoz has been set to 0.6 from its earlier setting of
0.8. We can add the impulse responses of this model to the irf file with the name transitory,
. irf create transitory, replace
(irfname transitory not found in rbcirf.irf)
and graph them.

. irf graph irf, irf(transitory) impulse(z) response(y c x h w z) noci
> byopts(yrescale)
transitory, z, c transitory, z, h transitory, z, w

.6
.18 .4
.4
.16 .2
.2
0
.14
0
transitory, z, x transitory, z, y transitory, z, z

6 1 1
.5 .5
0
0 0
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8
Step
Model variables are much less persistent. We can use irf ograph to overlay the IRF for a variable
under the two calibrations. This way we can view the differences across calibrations directly.
. irf ograph (persistent z c irf) (transitory z c irf)
.3
.25
.2
.15
0 2 4 6 8
Step
persistent: irf of z -> c

transitory: irf of z -> c
When the shock itself is persistent, consumption responds persistently. When the shock is transitory,
consumption returns to its steady-state value quickly.
Reference
King, R. G., and S. T. Rebelo. 1999. Resuscitating real business cycles. In Handbook of Macroeconomics: Volume 1A,
ed. J. B. Taylor and M. Woodford, 927–1007. New York: Elsevier. https://doi.org/10.1016/S1574-0048(99)10022-3.
Also see
[DSGE] Intro 3a — New Keynesian model
[DSGE] Intro 3c — Financial frictions model
[DSGE] Intro 3e — Nonlinear New Classical model
Title
Intro 3c — Financial frictions model
Description
This introduction estimates and interprets the parameters of a model that incorporates financial
frictions. The model is an extension of the one in [DSGE] Intro 3a.

The model
Impulse responses
The model
Equations (1)–(7) specify a model of financial frictions. The simplest such model places a wedge
between two interest rates: the safe interest rate set by the central bank and the market interest rate
that consumers and producers use.
πt = βEt πt+1 + κxt (1)

xt = Et xt+1 − (it − Et πt+1 − gt ) (2)
rt = ψπt + ut (3)
it = χrt + et (4)
gt+1 = ρg gt + ξt+1 (5)
ut+1 = ρu ut + t+1 (6)
et+1 = ρe et + ηt+1 (7)
These are the linearized equations. The nonlinear equations are similar to those in Writing down
nonlinear DSGEs in [DSGE] Intro 1.
This model is an extension of the one worked in [DSGE] Intro 1. It has an additional equation for
the interest rate spread. As before, (1) specifies the inflation equation (a Phillips curve), (2) specifies
the output gap equation (an Euler equation), and (3) specifies the equation for the safe interest rate (a
Taylor rule). Equation (3) can be thought of as an equation that specifies the safe interest rate rt . The
new element is (4), which specifies an equation for the market interest rate it that enters the output
gap equation. The market interest rate it is a function of the safe interest rate and a state variable et .
The state variable et controls the interest rate spread and can be thought of as representing the state
of the financial system. A large realization of et represents a large interest rate spread, indicating
financial distress.
42
Intro 3c — Financial frictions model 43
We estimate the parameters of the model in (1)–(7) using U.S. data on the Federal funds rate (safe
rate set by the central bank), the high-grade corporate bond interest rate (a measure of market interest
rates), and the inflation rate.
As is typical in these models, we constrain the parameter beta to 0.96.
> (x = F.x - (i - F.p - g), unobserved)
> (i = {chi}*r + e)
> (r = {psi}*p + u)
> (F.e = {rhoe}*e, state)
> (F.g = {rhoz}*g, state),
> from(psi=2 chi=0.8) constraint(1)
44 Intro 3c — Financial frictions model
DSGE model
/structural
kappa .0503257 .0184939 2.72 0.007 .0140783 .086573
chi .9067394 .1249768 7.26 0.000 .6617894 1.151689
psi 6.332377 2.567617 2.47 0.014 1.29994 11.36481
rhoe .8478222 .0301221 28.15 0.000 .788784 .9068605
rhou .815346 .033254 24.52 0.000 .7501694 .8805226
rhoz .9861866 .0099767 98.85 0.000 .9666327 1.00574
sd(e.e) .8857071 .1343023 .6224795 1.148935

sd(e.u) 7.160717 2.933797 1.41058 12.91085
sd(e.g) .3911862 .0358352 .3209506 .4614219
The persistence of the financial shock rhoe is estimated to be 0.85. The slope of the Phillips
curve, kappa, is somewhat flatter in this model than in the one in [DSGE] Intro 1. The coefficient
on inflation in the interest rate equation is 6.3 and indicates that the central bank increases interest
rates much more than one for one in response to movements in inflation.

We can read off the impact effect of shocks using the policy matrix. The response to the financial
shock e will be of most interest.
. estat policy
Policy matrix
Delta-method
p
e -.1832608 .0740949 -2.47 0.013 -.3284842 -.0380374
u -.1584194 .0641353 -2.47 0.014 -.2841223 -.0327165
g .2096327 .1003828 2.09 0.037 .0128861 .4063794
x
e -.6776486 .3498876 -1.94 0.053 -1.363416 .0081185
u -.683934 .2783442 -2.46 0.014 -1.229479 -.1383894
g .2218595 .1284728 1.73 0.084 -.0299427 .4736616
i
e -.0522495 .0403212 -1.30 0.195 -.1312776 .0267785
u -.0028755 .0041244 -0.70 0.486 -.0109591 .0052082
g 1.203672 .0979344 12.29 0.000 1.011724 1.39562
r
e -1.160476 .1361174 -8.53 0.000 -1.427261 -.8936912
u -.0031712 .0045309 -0.70 0.484 -.0120517 .0057092
g 1.327473 .222489 5.97 0.000 .8914029 1.763544
Intro 3c — Financial frictions model 45
Importantly, the financial shock causes r to fall by more than i, indicating an increase in the
interest rate spread. Inflation also falls, as does the output gap. It is the fall in the inflation rate that
causes the central bank’s interest rate to fall on impact of the shock.
Impulse responses
We use the irf set command to set finirf.irf as the active IRF file and then use irf create
to create and store the impulse responses under the name param.
. irf set finirf
(file finirf.irf created)
(file finirf.irf now active)
. irf create param1

(file finirf.irf updated)
Finally, we graph the impulse response to a financial shock.

. irf graph irf, irf(param1) impulse(e) response(e x p i r) byopts(yrescale)
param1, e, e param1, e, i param1, e, p

1.5 .05 0
0
1 -.1
-.05
.5 -.2
-.1
0 -.15 -.3
0 2 4 6 8
param1, e, r param1, e, x
0 0
-.5 -.5
-1 -1
0 2 4 6 8 0 2 4 6 8
Step
The top middle panel shows the effect of the shock to the e equation on e itself. The shock
represents a persistent increase in the spread between the two interest rates. The increased interest rate
spread causes both the output gap and inflation rate to fall. The safe interest rate r falls dramatically,
and the combined effect of the increased spread e and reduced safe interest rate r causes the market
interest rate i to be little changed.
46 Intro 3c — Financial frictions model
Also see
Title
Intro 3d — Nonlinear New Keynesian model
Description
This introduction estimates and interprets the parameters of a nonlinear New Keynesian model.
We demonstrate the effect of a change in constraints on a model that is only partially identified.

The model
Impulse responses
A change in constraints
The model
Equations (1)–(5) specify a simplified nonlinear New Keynesian model of inflation Πt , the output
gap Xt , and the interest rate Rt . These three variables are driven by two state variables, Zt and Mt .
Recall that in this manual, capital letters denote the level of a variable, while lowercase letters denote
deviations from steady state.

Xt 1 Rt
1 = βEt (1)
Xt+1 Zt Πt+1
(θ − 1) + φ(πt − 1)πt = θXt + φβEt {(πt+1 − 1)πt+1 } (2)
βRt = Πψ
t Mt (3)
ln(Mt+1 ) = ρm ln(Mt ) + ut+1 (4)
ln(Zt+1 ) = ρz ln(Zt ) + et+1 (5)
Equation (1) is the Euler equation linking the output gap in the current period to the expected
future output gap and the real interest rate. In this equation, Zt can be interpreted as a state variable
influencing household decisions. In applications, Zt stands in for the natural rate of interest, consumer
spending shocks, or government expenditures. Equation (2) is the Phillips curve linking the current
rate of inflation to expected future inflation and the output gap. Equation (3) is the interest-rate
rule linking the interest rate to the inflation rate; in this equation, the state variable Mt captures
all movements in the interest rate that are due to factors other than inflation. Equation (4) specifies
the stochastic process for the state variable in the interest rate equation. Equation (5) specifies the
stochastic process for the state variable in the Euler equation.
47
48 Intro 3d — Nonlinear New Keynesian model
We estimate the model’s parameters using data on the inflation rate and the nominal interest rate.
Not all model parameters are identified. We constrain β to be 0.96, a common value in the
literature. The parameters θ and φ are collinear and cannot be estimated together. The parameter θ is
set to 5, another common value in the literature. Later, we will explore consequences of the choice
θ = 5. The remaining six parameters are identified.
. constraint 1 _b[theta]=5
To specify this model to dsgenl, we will write each equation and then specify four options. The
first three options tell dsgenl about the model’s structure. The observed() option specifies observed
control variables that are to be used in estimation. The unobserved() option specifies latent control
variables. The exostate() option specifies the collection of state variables. The other option we
use is the standard constraints() option that applies our constraints on the parameters beta and
theta.
. dsgenl (1 = {beta}*(x/F.x)*(1/z)*(r/F.p))
> ({theta}-1 + {phi}*(p -1)*p = {theta}*x + {phi}*{beta}*(F.p-1)*F.p)
> (({beta})*r = (p)^({psi=2})*m)
> (ln(F.m) = {rhom}*ln(m))
> (ln(F.z) = {rhoz}*ln(z)),
> exostate(z m) unobserved(x) observed(p r)
> constraint(1 2) nolog
( 1) [/structural]theta = 5
/structural
theta 5 (constrained)
phi 47.07939 15.9416 2.95 0.003 15.83443 78.32435
psi 1.943008 .2957895 6.57 0.000 1.363271 2.522745
rhom .7005489 .0452605 15.48 0.000 .61184 .7892577
rhoz .9545255 .0186424 51.20 0.000 .9179871 .991064
sd(e.z) .5689908 .0982979 .3763305 .761651

sd(e.m) 2.318208 .3047461 1.720916 2.915499
Because this is the nonlinear version of the model in [DSGE] Intro 3a, it is instructive to compare
the two results. The two models deliver nearly identical estimates of most parameters; the small
differences seen in the fourth and higher decimal places are due to differences in the numerical
techniques used in each command. The only parameter that differs for the two models is the stickiness
parameter phi, which is the counterpart of the Phillips curve slope kappa from [DSGE] Intro 3a. The
parameterization of the Phillips curve in (2) above differs from the one in [DSGE] Intro 3a. However,
the two are linked by the relationship κ = (θ − 1)/φ, and we can use that relationship to show that
the models are equivalent.
Intro 3d — Nonlinear New Keynesian model 49
. nlcom (_b[theta] - 1)/_b[phi]

_nl_1: (_b[theta] - 1)/_b[phi]
_nl_1 .0849629 .0287694 2.95 0.003 .028576 .1413498
The value above is the same as the coefficient for kappa in [DSGE] Intro 3a.

The policy matrix is part of the state-space representation of a DSGE model. It expresses the
control variables of the model as functions of the state variables alone. Elements of the policy matrix
represent the response of a control variable to a one-unit increase in a state variable. In a linear DSGE
model, the policy matrix will also be linear. For a nonlinear DSGE model, the policy function will
typically be nonlinear. The policy matrix reported after dsgenl is a linear approximation of the true,
nonlinear policy matrix evaluated at the estimated parameter vector. For details on the linearization
of the policy function, see DeJong and Dave (2011, sec. 5.3).
. estat policy
Policy matrix
Delta-method
x
z .952921 .4813967 1.98 0.048 .0094009 1.896441
m -1.608216 .405057 -3.97 0.000 -2.402113 -.814319
p
z .9678135 .277745 3.48 0.000 .4234433 1.512184
m -.4172515 .0393611 -10.60 0.000 -.494398 -.3401051
r
z 1.880469 .2615995 7.19 0.000 1.367744 2.393195
m .189277 .0591663 3.20 0.001 .0733133 .3052407
An increase in m decreases inflation, decreases the output gap, and increases interest rates. We
may interpret it as an unexpected contractionary monetary policy. An increase in z increases inflation,
increases the output gap, and increases interest rates.
Next, we can look at the state transition matrix. Because the states are uncorrelated with each
other in this example, the elements of the state transition matrix are the persistence parameters in the
model.
. estat transition
Delta-method
F.z
z .9545255 .0186424 51.20 0.000 .9179871 .991064
m 0 (omitted)
F.m
z 0 (omitted)
m .7005489 .0452605 15.48 0.000 .61184 .7892577

values.
These parameters repeat the results we saw in the estimation table. When the state variables
are correlated, the transition matrix will reveal new information about the correlations among state
variables.
Impulse responses
The impulse–response function traces out the effect of a shock on the model’s state and control
variables. We use two commands to set up the impulse responses: an irf set command to set the
active IRF file and an irf create command to create the impulse–response functions.
From there, we can use irf graph irf to graph the impulse–response functions.
. irf set nlex1irf, replace
(file nlex1irf.irf created)
(file nlex1irf.irf now active)
. irf create irf1, replace
(irfname irf1 not found in nlex1irf.irf)
(file nlex1irf.irf updated)
. irf graph irf, impulse(m z) response(x p r) byopts(yrescale)
irf1, m, p irf1, m, r irf1, m, x

0 0
.6
.4 -2
-.5
.2 -4
-1
0 -6
irf1, z, p irf1, z, r irf1, z, x

.8 1.2 1
1
.6
.8 .5
.4
.6
.2 .4 0
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8
Step
Responses to a monetary shock are displayed in the first row. A shock to monetary policy leads
inflation to fall, the interest rate to rise, and the output gap to fall. From the bottom row, we see that
an increase in the natural rate of interest increases all three control variables.
A change in constraints
In the above example, the parameter θ was set to 5 to achieve identification of the other parameters.
The following example shows the consequences of this decision by examining a case where θ was
fixed to a different value. Suppose that instead of setting θ = 5 above, we set θ = 2. How would
the change in constraints change our estimates? First, we store the current estimates with estimates
store. We call the current estimates theta5 because they represent the case where θ = 5.
. dsgenl
/structural
phi 47.07939 15.9416 2.95 0.003 15.83443 78.32435
psi 1.943008 .2957895 6.57 0.000 1.363271 2.522745
rhom .7005489 .0452605 15.48 0.000 .61184 .7892577
rhoz .9545255 .0186424 51.20 0.000 .9179871 .991064
sd(e.z) .5689908 .0982979 .3763305 .761651

sd(e.m) 2.318208 .3047461 1.720916 2.915499
. estimates store theta5
We can now recall these estimates later as needed. We set up the new constraint with
. constraint 3 _b[theta]=2
and reestimate parameters. The nolog option is used to suppress the iteration log.
. dsgenl (1 = {beta}*(x/F.x)*(1/z)*(r/F.p))
> ({theta}-1 + {phi}*(p -1)*p = {theta}*x + {phi}*{beta}*(F.p-1)*F.p)
> (({beta})*r = (p)^({psi=2})*m)
> (ln(F.m) = {rhom}*ln(m))
> exostate(z m) unobserved(x) observed(p r)
> constraint(2 3) nolog
/structural
phi 11.76979 3.985338 2.95 0.003 3.958675 19.58091
psi 1.943004 .2957869 6.57 0.000 1.363272 2.522736
rhom .7005482 .0452601 15.48 0.000 .6118401 .7892563
rhoz .9545256 .0186424 51.20 0.000 .9179872 .991064
sd(e.z) .5689892 .0982974 .3763299 .7616486

sd(e.m) 2.318204 .3047434 1.720918 2.91549
. estimates store theta2
We saved these estimates in theta2 because they represent the case where θ = 2. estimates
table displays the two results. The stats(ll) option displays the estimated log likelihood with
the estimated structural parameters.
. estimates table theta5 theta2, b(%9.4f) stats(ll)
Variable theta5 theta2
/structural
beta 0.9600 0.9600
theta 5.0000 2.0000
phi 47.0794 11.7698
psi 1.9430 1.9430
rhom 0.7005 0.7005
rhoz 0.9545 0.9545
sd(e.z) 0.5690 0.5690

sd(e.m) 2.3182 2.3182
Statistics
ll -753.5713 -753.5713
The reported log-likelihood values are identical across the two parameterizations. The change in
constraints slides us along the top of a ridge in the likelihood function. The estimated value of phi
differs when the constraint is changed, but all other parameter estimates are identical. Even though
the point estimate of phi has changed, the z statistic and p-value are unchanged.
As discussed above, the underlying parameter that is being estimated here is κ = (θ − 1)/φ. The
restrictions on the parameter θ lead to a change in the estimated value of φ but do not change the
estimated value of κ. We can use nlcom to recover κ first from the model with θ = 5 and then from
the model with θ = 2.
. estimates restore theta5
(results theta5 are active now)
. nlcom (_b[theta]-1) / _b[phi]
_nl_1: (_b[theta]-1) / _b[phi]
_nl_1 .0849629 .0287694 2.95 0.003 .028576 .1413498
. estimates restore theta2

(results theta2 are active now)
. nlcom (_b[theta]-1) / _b[phi]
_nl_1: (_b[theta]-1) / _b[phi]
_nl_1 .0849633 .0287692 2.95 0.003 .0285767 .1413498
The equivalence of the two parameterizations can also be seen through the policy matrix. The
estimated policy matrix when θ = 2 is
. estat policy
Policy matrix
Delta-method
x
z .9529194 .4813921 1.98 0.048 .0094082 1.896431
m -1.608214 .4050521 -3.97 0.000 -2.402102 -.8143265
p
z .9678173 .2777448 3.48 0.000 .4234475 1.512187
m -.4172521 .0393609 -10.60 0.000 -.4943981 -.3401061
r
z 1.880473 .2615995 7.19 0.000 1.367747 2.393199
m .1892774 .059166 3.20 0.001 .0733142 .3052406
This policy matrix is identical to the one shown above, in which the constraint θ = 5 was applied.
The two parameter vectors imply the same reduced-form matrices.
Reference
Also see
[DSGE] Intro 3f — Stochastic growth model
Title
Intro 3e — Nonlinear New Classical model
Description
This introduction describes a nonlinear New Classical model, estimates some of its parameters,
and explores the model’s impulse–response functions under two different parameter settings.

The model
Steady state
Model-implied covariances
Impulse responses
The model
Equations (1)–(8) specify a variant on the New Classical model studied in King and Rebelo (1999).
It includes equations for output Yt , consumption Ct , investment It hours worked Ht , the interest rate
Rt , the real wage Wt , the capital stock Kt , and productivity Zt . Model variables must be weakly
stationary. The model contains six parameters: α, β , χ, δ , ρ, and σ .

1 1
= βEt (1 + Rt+1 − δ) (1)
Ct Ct+1
Wt
χHt = (2)
Ct
Yt = Ct + It (3)
Yt = Zt Ktα Ht1−α (4)
Yt
Rt = α (5)
Kt
Yt
Wt = (1 − α) (6)
Ht
Kt+1 = It + (1 − δ)Kt (7)
ln(Zt+1 ) = ρ ln(Zt ) + et+1 (8)
Equation (1) is a consumption Euler equation that links consumption in the current period to expected
future consumption and the expected future interest rate. Equation (2) is a labor supply equation
that links hours worked to the wage and consumption. Equation (3) is a standard national income
accounting identity, stating that output is split between consumption and investment. Equation (4)
is an output supply equation or production function, stating that output is produced by combining
capital Kt and labor Ht at productivity level Zt . Equation (5) is a capital demand curve. Equation
(6) is a labor demand curve. Equation (7) is the capital accumulation process. Equation (8) specifies
a stochastic process for productivity. The stochastic shock et+1 is i.i.d. normal with mean zero and
standard deviation σz .
55
56 Intro 3e — Nonlinear New Classical model
The model includes six parameters. β is the discount factor reflecting a preference for current
consumption relative to future consumption. α is a production parameter. χ is a preference parameter.
δ is the depreciation rate. ρ measures the persistence of the stochastic productivity process. Finally,
σ is the standard deviation of the innovations to the productivity process.
In this example, we fix some parameters at prespecified values and estimate others. Once the
parameters are estimated, we analyze the model by inspecting its steady state, low-order moments,
policy and transition matrices, and impulse–response functions.
We use the growth rate of output as the empirical counterpart to yt . For the calibrated parameters,
we set α = 0.33, β = 0.99, δ = 0.025, and χ = 2, which follows King and Rebelo (1999). We will
estimate (ρ, σz ). We first bring in the data.
We next set up all the constraints

. constraint 1 _b[alpha] = 0.33
. constraint 2 _b[beta] = 0.99
. constraint 3 _b[delta] = 0.025
. constraint 4 _b[chi] = 2
and estimate the remaining parameters,

. dsgenl (1/c = {beta}*(1/F.c)*(1+F.r-{delta}))
> ({chi}*h = w/c)
> (y = c + i)
> (y = z*k^{alpha}*h^(1-{alpha}))
> (r = {alpha}*y/k)
> (w = (1-{alpha})*y/h)
> (F.k = i + (1-{delta})*k)
> (ln(F.z) = {rho}*ln(z))
> , observed(y) unobserved(c i r w h) exostate(z) endostate(k) constraint(1/4)
BFGS stepping has contracted, resetting BFGS Hessian
Intro 3e — Nonlinear New Classical model 57

( 1) [/structural]alpha = .33
( 3) [/structural]delta = .025
( 4) [/structural]chi = 2
/structural
delta .025 (constrained)
chi 2 (constrained)
alpha .33 (constrained)
rho .3133483 .0614696 5.10 0.000 .1928701 .4338265
sd(e.z) 2.287682 .1036018 2.084627 2.490738
We specified four options related to the model structure. The exostate(z) option lists the
state variable that is subject to shocks, z. The endostate(k) option specifies that k is a state
variable without shocks. The observed(y) option specifies y as an observed control variable. The
unobserved(c i r w h) option specifies those variables as unobserved, or latent, control variables.
The remaining option we specified was constraint(1/4), which applies the constraints we set
up previously. We find that the estimated productivity process is mildly persistent, ρ = 0.31. The
standard deviation of the productivity shock is about 2.3, or about 2%.
Steady state
In the absence of shocks, the variables in a nonlinear DSGE model converge to a steady-state
position in which variables are constant through time. dsgenl finds the steady state of the model at
the estimated parameter values. You can view the steady state with estat steady.
. estat steady
Location of model steady-state
Delta-method
k 18.75991 . . . . .
z 1 . . . . .
c 1.526432 . . . . .
i .4689977 . . . . .
r .035101 . . . . .
w 2.02027 . . . . .
h .6617621 . . . . .
y 1.99543 . . . . .
Note: Standard errors reported as missing for constrained steady-state values.
All the standard errors for the estimated location of the steady state are missing. This is because
of the constraints we placed on some of the model parameters. We estimated the persistence of the
productivity state variable and the standard deviation of the shock to productivity. These dynamic
parameters have no effect on the steady state. Instead, the steady state is a function of the model’s
static parameters. Because we constrained all the static parameters, the steady-state location is not
subject to uncertainty, and the standard errors are all missing. Put another way, the location of the
steady state is determined entirely by parameters that we have constrained in the model and by the
model’s structure itself.
Model-implied covariances
The model’s state-space matrices generate predictions for the variances and covariances of the
model variables.
. estat covariance
Estimated covariances of model variables
Delta-method
c
var(c) .870567 .1701705 5.12 0.000 .537039 1.204095
cov(c,i) 3.39887 .5499555 6.18 0.000 2.320977 4.476763
cov(c,r) .0352466 .0232027 1.52 0.129 -.0102298 .080723
cov(c,w) 1.167688 .2145175 5.44 0.000 .7472416 1.588135
cov(c,h) .2971211 .0448135 6.63 0.000 .2092883 .3849539
cov(c,y) 1.464809 .2590251 5.66 0.000 .9571294 1.972489
i
var(i) 196.9607 19.01966 10.36 0.000 159.6829 234.2386
cov(i,r) 48.87007 4.692839 10.41 0.000 39.67227 58.06786
cov(i,w) 26.14587 2.644813 9.89 0.000 20.96213 31.32961
cov(i,h) 22.747 2.18294 10.42 0.000 18.46852 27.02549
cov(i,y) 48.89287 4.818514 10.15 0.000 39.44876 58.33699
r
var(r) 12.92995 1.273471 10.15 0.000 10.434 15.42591
cov(r,w) 5.774216 .5477555 10.54 0.000 4.700635 6.847797
cov(r,h) 5.73897 .5520461 10.40 0.000 4.656979 6.82096
cov(r,y) 11.51319 1.099565 10.47 0.000 9.358078 13.66829
w
var(w) 4.103074 .4796048 8.56 0.000 3.163066 5.043082
cov(w,h) 2.935386 .2902517 10.11 0.000 2.366503 3.504269
cov(w,y) 7.03846 .7632272 9.22 0.000 5.542562 8.534358
h
var(h) 2.638265 .2521725 10.46 0.000 2.144016 3.132514
cov(h,y) 5.573651 .5419094 10.29 0.000 4.511528 6.635774
y
var(y) 12.61211 1.298179 9.72 0.000 10.06773 15.15649
One object of interest is the relative volatility of a model variable, defined as the standard deviation
of that variable compared with the standard deviation of a reference variable. We can calculate the
volatility of investment (model variable i) relative to output (model variable y) with nlcom.
Before we can use nlcom, we need to add the post option to our estat covariance command.
With this option, the results are stored in e(), where nlcom can access them. First, however, we
save the dsgenl estimates using estimates store.
. estimates store dsgenl
. quietly estat covariance, post
We can now refer to the variances of investment and output as b[i:var(i)] and b[y:var(y)].
We use nlcom to compute the volatility of investment relative to output.
. nlcom sqrt(_b[i:var(i)] / _b[y:var(y)])
_nl_1: sqrt(_b[i:var(i)] / _b[y:var(y)])
_nl_1 3.951809 .0303923 130.03 0.000 3.892241 4.011377
The model predicts that investment is about four times as volatile as output.

The model’s state-space form expresses the control variables as functions of the state variables
alone. It solves the system of equations locally near the steady state, resolving the expectations of
future variables via the rational expectations assumption. The entries in the state-space matrices are
interpretable as the contemporaneous effect of a one-unit change in the state variable on the control
variable.
. estimates restore dsgenl
(results dsgenl are active now)
. estat policy
Policy matrix
Delta-method
c
k .5530748 . . . . .
z .1006471 .0040593 24.79 0.000 .092691 .1086031
i
k -.8741569 . . . . .
z 5.854704 .0219119 267.19 0.000 5.811758 5.897651
r
k -.782376 . . . . .
z 1.453057 .0020449 710.58 0.000 1.44905 1.457065
w
k .3853494 . . . . .
z .7768523 .0010072 771.31 0.000 .7748782 .7788263
h
k -.1677254 . . . . .
z .6762052 .0030521 221.56 0.000 .6702232 .6821872
y
k .217624 . . . . .
z 1.453057 .0020449 710.58 0.000 1.44905 1.457065
Note: Standard errors reported as missing for constrained policy matrix

values.
The standard errors presented here again require some interpretation. The standard errors for the
impact effect of k on all model variables are missing. These impact effects depend entirely on the
parameters we constrained, so their values are fixed entirely by the structure of the model.
The state variables are correlated in this model. The capital equation shown in (7) specifies that
future capital is a function of the current capital stock and current investment. In turn, investment
depends on the current capital stock and on productivity; hence, the future capital stock depends on
both the current capital stock and current productivity. We can see these relationships with estat
transition.
. estat transition
Delta-method
F.k
k .9531461 . . . . .
z .1463676 .0005478 267.19 0.000 .1452939 .1474413
F.z
k 0 (omitted)
z .3133483 .0614696 5.10 0.000 .1928701 .4338265

values.
The “own” effects (k on F.k and z on F.z) are persistence parameters. The other reported effects
show how a change in one state variable contemporaneously affects another. A change in the capital
stock does not affect future productivity, but a change in productivity does affect the future capital
stock. A 1% increase in productivity raises the future capital stock by 0.15%.
Impulse responses
The policy and transition matrices display impact effects. Because of the dynamic structure of
the state, a shock to a state variable has lasting effects on both the state and the controls. An
impulse–response function traces out the full dynamic effect of a shock. When state variables are
correlated, a shock to one state variable will (either immediately or eventually) lead to movements in
another state variable.
We use the irf commands to trace out the effect of a shock to productivity on itself and on each
of the control variables in our model.
. irf set rbcirf, replace
(file rbcirf.irf created)
(file rbcirf.irf now active)
. irf create est, step(20) replace
(irfname est not found in rbcirf.irf)
. irf graph irf, impulse(z) response(y c i h w z) byopts(yrescale)
est, z, c est, z, h est, z, i

.3 15
1.5
.25
10
1
.2
.5 5
.15
.1 0 0
est, z, w est, z, y est, z, z

2 4 3
1.5 3
2
1 2
1
.5 1
0 0 0
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
Step
All model variables rise on a shock to productivity. For most model variables, the increase is
short lived, and they return to their long-run values within five periods. Consumption is the exception
(upper left panel). It rises smoothly for several periods, then gradually falls back to its long-run value.
In this section, we will explore the effect of changing some model parameters on the impulse
responses. We first type the dsgenl command without arguments to replay the current estimates.
. dsgenl
/structural
chi 2 (constrained)
rho .3133483 .0614696 5.10 0.000 .1928701 .4338265
sd(e.z) 2.287682 .1036018 2.084627 2.490738
We can solve the model using different parameter values and analyze how the model behaves
across the two scenarios. It is common to examine a model’s behavior for multiple parameter settings,
for example, for different settings for the persistence of a shock or for different settings of preference,
technology, or policy parameters. The simplest way to solve using a new set of parameters is to use
the from() and solve options. First, store the current parameter vector in a Stata matrix.
. matrix b = e(b)
Next, change the entries in the matrix to the new desired parameters. In this case, we want to look
at the response of the model to a shock when productivity is more persistent than was estimated, so
we set ρ = 0.6.
. matrix b[1,5] = 0.6
Next, we solve the model again. In the following command, note the use of the from(b) and
solve options.
. dsgenl (1/c = {beta}*(1/F.c)*(1+F.r-{delta}))
> ({chi}/(1-h) = w/c)
> (y = c + i)
> (w = (1-{alpha})*y/h)
> (r = {alpha}*y/k)
> (F.k = i + (1-{delta})*k)
> , observed(y) unobserved(c i r w h) exostate(z) endostate(k)
> from(b) solve noidencheck
/structural
beta .99 . . . . .
delta .025 . . . . .
chi 2 . . . . .
alpha .33 . . . . .
rho .6 . . . . .
sd(e.z) 2.287682 . . .

Note: Model solved at specified parameters; maximization options ignored.
We passed the parameter vector in using from(), and we used the solve option to prevent
estimation. Because we specified the solve option, the estimation table simply repeats the values we
fed in with from(). Of course, because we did not estimate, we do not get standard errors. When
we use solve, the point of the analysis is always in the estat and irf commands that come after.
We create a new set of impulse responses.
. irf create alt, step(20) replace
(irfname alt not found in rbcirf.irf)
We graph the response of y to a z impulse across the two models.

. irf ograph (est z y irf) (alt z y irf)
0
0 5 10 15 20
Step
est: irf of z -> y

alt: irf of z -> y
We see that output increases more and takes longer to return to its long-run value when we set ρ
to 0.60 instead of its estimated value of 0.31.
We are not interested only in the response of y to a shock in z. We can also plot the response of
c, h, i, w, y, and z to an impulse to z in panels of a combined graph. We will accomplish this task
with a foreach loop.
. foreach v in c h i w y z {
2. irf ograph (est z ‘v’ irf) (alt z ‘v’ irf), name(‘v’) nodraw
3. }
. graph combine c h i w y z
.5 2.5 15
2
.4
1.5 10
.3
1
5
.2 .5
0 0
.1
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
Step Step Step
est: irf of z -> c est: irf of z -> h est: irf of z -> i

alt: irf of z -> c alt: irf of z -> h alt: irf of z -> i
2 4 2.5
2
1.5 3
1.5
1 2
1
.5 1
.5
0 0 0
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
Step Step Step
est: irf of z -> w est: irf of z -> y est: irf of z -> z

alt: irf of z -> w alt: irf of z -> y alt: irf of z -> z
The model with ρ = 0.60 shows more persistence overall than the estimated value of ρ = 0.31, as
expected. For variables such as consumption and wages, the amplitude of the response is amplified
as well.
Reference
Also see
Title
Intro 3f — Stochastic growth model
Description
In this introduction, we solve a stochastic growth model and interpret objects that are available after
solving. The topics discussed in this introduction are more technical than those in the introductions
describing other classic models. We provide some details about how dsgenl finds the model’s steady
state, we describe how it solves for a first-order approximation to the model’s state-space form,
and we provide guidance on the differences between and interpretations of linear versus log-linear
approximations. For Bayesian analysis of this model, see [DSGE] Intro 9b.

The model
Approximating the solution to a nonlinear DSGE model
Specifying the model to Stata
After solving
The steady state
Approximations to the policy and transition matrices
Linear and log-linear approximations
The model
The model contains equations that jointly determine output Yt , the interest rate Rt , consumption
Ct , capital Kt , and productivity Zt . The model contains four parameters: α, β , δ , and ρ. This model
is a variant on the model used in Schmitt-Grohé and Uribe (2004).
( −1 )
Ct+1
1 = βEt (1 + Rt+1 − δ) (1)
Ct
Yt = Zt Ktα (2)
Rt = αZt Ktα−1 (3)
Kt+1 = Yt − Ct + (1 − δ)Kt (4)
ln(Zt+1 ) = ρ ln(Zt ) + et+1 (5)
Equation (1) defines a relationship between consumption growth Ct+1 /Ct and the real interest
rate Rt+1 . Equation (2) is a production function for output Yt as a function of productivity Zt and
capital Kt . Equation (3) is a model for the interest rate. Equation (4) is the equation for capital
accumulation; capital in the next period is equal to undepreciated capital this period (1 − δ)Kt plus
unconsumed output Yt − Ct . Equation (5) is a law of motion for productivity Zt . The parameter β
is a discount factor in the consumption equation, the parameter α is a production parameter in the
output equation, the parameter δ is a depreciation parameter in the capital equation, and the parameter
ρ is a persistence parameter in the productivity equation.
66
Intro 3f — Stochastic growth model 67
The state variables are the current-period capital stock and the level of productivity, (Kt , Zt ).
The control variables are consumption, the interest rate, and output (Ct , Rt , Yt ). To solve the model
means to write the control variables as functions of the state variables and to write the future values
of the state variables as functions of the current state variables. Equation (2) is already solved, for
example, because it writes output as a function of the state variables alone. But (4) is not a reduced
form, because it writes the future capital stock as a function of the state variable Kt and the control
variables Ct and Yt .
The model equations have two key features. First, the model is nonlinear in its variables. Second,
the model contains expectations of variables at time t + 1, conditional on information at time t.
dsgenl solves the model by taking a linear (or log-linear) approximation to the model equations at
the steady state. Two concepts need to be clarified: the steady state and the approximation. We will
describe these two issues in general, then return to the specific model described above.
Approximating the solution to a nonlinear DSGE model

We can gather up all the model equations into a vector of functions, f . We denote the model’s
vector of state variables by xt and the vector of control variables by yt . Now, the DSGE model can
be written as
Et {f (xt+1 , yt+1 , xt , yt , θ)} = 0
The expectation operator applies to the entire system of equations. In a generic DSGE model, there
are n equations, nx state variables, and ny control variables; it must be the case that nx + ny = n.
The steady state of the model is the vector of values to which model variables converge in the
absence of shocks. This is equivalent to the solution to the model’s system of equations after setting
all future values of variables equal to current values and dropping the expectation operator. This
location is a vector of numbers (x, y) that solves
f (x, y, x, y, θ) = 0
This is a system of n equations in n unknowns, and its solution is the steady state. It is this point
around which we take our approximation. The location of the steady state is influenced by the model
parameters θ. It is possible that the parameter vector you specify does not admit a steady state; in
such a case, dsgenl will halt and issue an error message. You may specify starting values for the
search for the steady state with the steadyinit() option.
We take an approximation to the model equations at the steady state. Differentiate f with respect to
each variable at the point (x, y, x, y). Differentiation generates a linearized model that is described
by four matrices, namely,
F1 E t x
bt+1 + F2 Et y
bt+1 + F3 x
b t + F4 y
bt = 0
The matrix F1 contains partial derivatives of each model equation with respect to one-period-ahead state
variables. The other Fi matrices are defined similarly. Each element of each Fi matrix depends on the
structural parameter vector θ. There are two ways of obtaining the Fi matrices: linear approximation
and log-linear, or percentage, approximation. The “hat” notation denotes the variables of the linear
model. In a linear approximation,
bt = (xt − x)
x
that is, x
bt is the unit difference from steady state. In a log-linear approximation,
bt = (xt − x)/x
x
68 Intro 3f — Stochastic growth model
so that x
bt is measured in percentage difference from the steady state. Which approximation is applied
will affect the Fi matrices, parameter estimation, and the interpretation of all objects available after
estimation.
Hatted variables are measured as deviations from steady state. At this point, the model is in the
form of a linear DSGE model, and the methods of Klein (2000) may be employed to find the solution.
The stability conditions of Klein (2000) apply to the Fi matrices, and dsgenl will not proceed if
the stability conditions are not met. If the stability conditions are met, we exactly solve the linear
model. The solution is of the form
y
bt = Gb
xt
x
bt+1 = Hb
xt + Met+1
Because shocks et are mean zero, it does not matter whether we write et or b
et .
In a nonlinear DSGE model, the linear solution matrices are a linear approximation to the true
nonlinear solution functions,
yt = g(xt )
xt+1 = h(xt ) + Met+1
The matrix G is an approximation to the function g(·), and the matrix H is an approximation to the
function h(·). Once the solution has been obtained, we can view the steady state, the approximate
solution matrices, the impulse responses, and other objects.
Specifying the model to Stata

We import U.S. macro data. Although we are solving the model (hence, the data will play no
role), dsgenl expects there to be data anyway.
Next, we set up a parameter matrix containing the values that we will use for our model. We type
. matrix param_mat = (0.33, 0.96, 0.025, 0.9, 1)
. matrix colnames param_mat = alpha beta delta rho /sd(e.z)
These values are similar to those in King and Rebelo (1999) and Schmitt-Grohé and Uribe (2004).
The elasticity of output with respect to capital α is set to 0.33. The discount factor β is set to 0.96.
The depreciation rate δ is set to 0.025. The persistence of productivity is set to 0.9, and the standard
deviation of the shock to productivity is set to 1.
Next, we specify the model.
. dsgenl (1 = {beta}*(F.c/c)^(-1)*(1+F.r-{delta}))
> (r = {alpha}*z*k^({alpha}-1))
> (y = z*k^{alpha})
> (F.k = y - c + (1-{delta})*k)
> , observed(y) unobserved(c r) exostate(z) endostate(k)
> from(param_mat) solve noidencheck
/structural
beta .96 . . . . .
delta .025 . . . . .
alpha .33 . . . . .
rho .9 . . . . .
sd(e.z) 1 . . .

Each equation is bound in parentheses. Variables appear without any special modification; parameters
are bound in parentheses. After the equation list, we specify seven options. The first four tell dsgenl
how to interpret each variable in the model. The control variables y, r, and c are specified as being
either observed() or unobserved(). Because we are only solving the model, it is inconsequential
how we place control variables into these categories. In an estimation exercise, data on observed
control variables are used to estimate model parameters, and which variables are observable can affect
parameter identification. The state variables are divided into categories according to whether they are
subject to shocks. The state variable k is not subject to shocks; it is an endogenous state variable.
The state variable z is subject to shocks; it is an exogenous state variable.
The from() option specifies the starting values for the parameters. The solve option declares
that we wish to solve the model at the values in from(). The noidencheck option skips the check
for local identification.
The output table repeats the parameter vector that we passed in; it provides no new information.
The notes below the estimation table remind us that we have skipped the identification check and
that we have solved the model rather than estimated the model parameters.
After solving
After solving the model by running dsgenl, you can view the steady state, the approximate policy
and transition matrices, and more using the estat commands.
The steady state
In the process of finding the solution, dsgenl finds the steady state of the model. The steady state
is displayed with
. estat steady, compact
Coefficient
k 10.88
z 1.00
c 1.93
r 0.07
y 2.20
Because we used dsgenl’s solve option, estat steady displays the steady state of the model
under the initial parameter vector. When solve is not specified, estat steady displays the steady
state of the model at the converged parameter values.
The steady-state level of the capital stock is 10.88 units, and the steady-state level of output is 2.20
units. The steady-state interest rate is 0.07, or 7%. Some of these values are more meaningful when
expressed as ratios rather than when they are expressed as raw values. The model’s consumption-to-
output ratio is about 0.88. The model’s capital-to-output ratio is about 4.9.
Approximations to the policy and transition matrices
Next, we approximate the policy and state transition matrices. estat policy provides the ap-
proximate policy matrix.
. estat policy, compact
Policy matrix
k z
c .6077069 .3431366
r -.67 1
y .33 1
By default, dsgenl uses a log-linear approximation, and the results of estat policy are also
based on this approximation. This means that variables are measured as percentage deviations from
steady state. The interpretation of the output equation in the policy matrix is that
Yt − Y Kt − K Zt − Z
≈ 0.33 +
Y K Z
A log-linear approximation was chosen because many DSGE model equations are linear in logs, or
nearly so. For instance, in logs, (2) is
log(Yt ) = α log(Kt ) + log(Zt )
just as the policy matrix describes.

The entries in the policy matrix are impact effects. They display the percentage change in a control
variable that results from a one-percent change in a state variable. From the consumption equation,
a 1% increase in z causes consumption to rise by 34%. From the output equation, a 1% increase in
z causes output to rise by 1%.
estat transition provides the approximate state transition matrix.
. estat transition, compact
k z
k
F1. .9340903 .1412781
z
F1. 0 .9
The state transition matrix displays how state variables in the next period depend on state variables
today. A 1% increase in z causes capital to rise by 0.14% one period in the future. Notice that an
increase in z has an effect on k, but an increase in k has no cross-effect on z.
Linear and log-linear approximations
Specifying option linearapprox causes dsgenl to take a linear approximation rather than a
log-linear approximation. We first re-solve the model.
. dsgenl (1 = {beta}*(F.c/c)^(-1)*(1+F.r-{delta}))
> (r = {alpha}*z*k^({alpha}-1))
> (y = z*k^{alpha})
> (F.k = y - c + (1-{delta})*k)
> , observed(y) unobserved(c r) exostate(z) endostate(k)
> from(param_mat) solve noidencheck linearapprox
/structural
beta .96 . . . . .
delta .025 . . . . .
alpha .33 . . . . .
rho .9 . . . . .
sd(e.z) 1 . . .

We next examine the impact that the linearapprox option has on postsolution objects. The
choice of linear or log-linear approximation does not affect the location of the steady state.
. estat steady, compact
Coefficient
k 10.88
z 1.00
c 1.93
r 0.07
y 2.20
This is to be expected; the steady state is the point around which approximations are taken, but its
location does not depend on the choice of approximation.
The policy matrix has changed.
. estat policy, compact
Policy matrix
k z
c .1075764 .6610193
r -.0041045 .0666667
y .0666667 2.198462
Each entry displays the unit increase in a control variable after a unit increase in a state variable.
As an example, we use the estimates in the last line of this table. The interpretation of the output
equation is
(Yt − Y ) ≈ 0.067(Kt − K) + 2.2(Zt − Z)
so that a one-unit increase in the capital stock leads output to rise by 0.067 units. The other entries
are interpreted analogously.
The state transition matrix is now
. estat transition, compact
k z
k
F1. .9340903 1.537443
z
F1. 0 .9
The approximate capital accumulation equation is
(Kt+1 − K) ≈ 0.93(Kt − K) + 1.54(Zt − Z)
so that a one-unit increase in productivity Zt leads to a 1.54-unit increase in the capital stock in the
next period.
References
Klein, P. 2000. Using the generalized Schur form to solve a multivariate linear rational expectations model. Journal
of Economic Dynamics and Control 24: 1405–1423. https://doi.org/10.1016/S0165-1889(99)00045-7.
Schmitt-Grohé, S., and M. Uribe. 2004. Solving dynamic general equilibrium models using a second-order approximation
to the policy function. Journal of Economic Dynamics and Control 28: 755–775. https://doi.org/10.1016/S0165-
1889(03)00043-5.
Also see
[DSGE] Intro 9b — Bayesian estimation of stochastic growth model
Title
Intro 4 — Writing a DSGE in a solvable form
Description
Many DSGE models, when written in their natural representation based on economic theory, include
problematic terms that do not fit into the algebraic form required to solve a DSGE model. This entry
shows how to accommodate these problematic terms by defining a new state variable or a new control
variable and then rewriting the equations so that they have the required form. We demonstrate how to
make the required changes using dsge, but the advice in this entry applies to both dsge and dsgenl.
We assume that you are already familiar with the elements of a DSGE; see [DSGE] Intro 1.

Introduction
Shocks to a control equation
Including a lag of a control variable
Including a lag of a state variable
Including an expectation of a control dated by more than one period ahead
Including a second-order lag of a control variable
Including an observed exogenous variable
Introduction
We can estimate the parameters of a DSGE model only if we can solve for its state-space form,
and we can solve a DSGE model for its state-space form only if the equations for control variables
and state variables have specific structures. Unfortunately, many DSGE models contain problematic
terms that do not fit into these structures without some manipulation. Fortunately, the manipulation
is easy, once you know the tricks discussed below.
As we discussed in Structural and reduced forms of DSGE models in [DSGE] Intro 1, the structural
form of a DSGE model can be written as
Et [f (xt+1 , yt+1 , xt , yt , θ)] = 0 (1)
and the structural form of a linearized DSGE model can be written as
A0 yt = A1 Et (yt+1 ) + A2 yt + A3 xt (2)
where yt is a vector of control variables, xt is a vector of state variables, and t is a vector of

shocks. A0 through A3 and B0 through B3 are matrices of parameters. The entries in Ai and Bj
are all functions of the structural parameters, which we denote by vector θ. A0 and B0 are diagonal
matrices, and A2 has zeros on its diagonal. C is a selection matrix that determines which state
variables are subject to shocks.
74
Intro 4 — Writing a DSGE in a solvable form 75
For nonlinear DSGE models, (1) specifies the required structure of the equations. Equations can
contain current values of state and control variables and can contain one-step-ahead values of state
and control variables. All other terms, such as future values dated more than one period ahead or
lagged values, are problematic.
For linearized DSGE models, (2) specifies the required structure for control variable equations, and
(3) specifies the required structure for state variable equations. Per (2), a control can depend on only
three types of terms—expected values of control variables in the next period, current values of other
control variables, and current values of state variables. All other terms are problematic. Per (3), the
values of state variables in the next period can depend on only four types of terms—expected values
of control variables in the next period, current values of control variables, current values of state
variables, and shocks. All other terms are problematic.
When written based on economic theory, the structural forms of many DSGE models include
problematic terms that do not fit into the required form. These problematic terms are handled by
defining a new state variable or a new control variable and rewriting the equations to have the required
form. We define a new state variable when the problematic term involves an exogenous, also known
as a predetermined, variable. We define a new control variable when the problematic term includes
a new endogenous variable. The rewriting process replaces the variable in the problematic term with
the new state or the new control variable.
Problematic terms usually violate one of the following four restrictions:
1. Equations for control variables may not contain shocks.

2. Equations may not contain lagged control variables.
3. Equations may not contain lagged state variables.
4. Equations may not contain expectations of control variables dated by more than one period
ahead.
All of these restrictions can be overcome by adding a new state variable or a new control variable
and rewriting the model so that it has the required form. Here are five examples of the four restrictions
and solutions to each one.
1. If you have an equation for a control variable that contains a shock, make this shock a new
state variable; see Shocks to a control equation below.
2. If you have an equation for a control variable that contains a lagged control variable, make
the lagged control variable a new state variable; see Including a lag of a control variable
below.
3. If you have an equation for a state variable that contains a lag of a state variable, make the
lagged state variable a new state variable; see Including a lag of a state variable below.
4. If you have an equation for a control variable that contains an expectation of a control
variable dated by more than one period ahead, make it a new control variable. We use a
new control variable instead of a new state variable because the far-future expectation is
endogenous, not predetermined. See Including an expectation of a control dated by more
than one period ahead below for details.
5. If you have an equation for a control variable that contains a second-order lag of a control
variable, make the second-order lagged control variable a new state variable. You will also
need to create a new state variable for the first-order lag. See Including a second-order lag
of a control variable below for details.
76 Intro 4 — Writing a DSGE in a solvable form
A nuanced issue not covered by restrictions 1–4 is that all the observed variables in a DSGE must
be endogenous control variables. This structure is usually not a problem, because DSGE models are
usually generated by theories that treat everything observed as endogenous. Occasionally, however,
models include an observed variable that is exogenous. For example, a small, open economy might
not be able to affect its real exchange rate. The solution for this problem is to model the observed
variable as a control variable that is equal to an exogenous state variable. See Including an observed
exogenous variable below for details.
For the most part, this entry discusses how to make models with problematic terms fit into the
structure of (1) or (2) and (3). Note that correlated state variables already fit into this structure. This
correlation is modeled by off-diagonal elements in B3 , not by elements in C. See [DSGE] Intro 4g
for an example.
Shocks to a control equation

Sometimes, we observe only a control variable with error. Suppose that the equation for the control
variable consumption is
ct = Et ct+1 − rt + t (4)
where ct is consumption, rt is the interest rate, and t is a shock. Shocks to a control variable are
not allowed, so t is a problematic term.
The solution entails three steps. First, we define a new state variable ut = t . We define a new
state instead of a new control because the shock is exogenous. Second, we write a state equation for
ut . Recall that the state equation specifies tomorrow’s state value in terms of today’s state value and
tomorrow’s shock. The state equation for ut is
ut+1 = t+1
Third, we rewrite (4), substituting ut for t . The consumption equation becomes
ct = Et ct+1 − rt + ut
which is in the required form. In the dsge command, this is expressed as

. dsge ... ///
(c = F.c - r + u) ///
(F.u = 0, state) ///
...
See [DSGE] Intro 4a for a complete example.
Including a lag of a control variable

Models that mix adaptive and rational expectations contain a lag of the control variable and the
expected future value of the control variable. Suppose our model had the following equation for the
control variable yt ,
yt = αyt−1 + (1 − α)Et (yt+1 ) − rt (4)
where yt is output and rt is the interest rate. The problematic term is αyt−1 because it involves a
lag of the control variable.
The solution entails three steps. First, we define a new state variable Lyt = yt−1 . We define a new
state instead of new control because the lagged control is predetermined, which makes it exogenous.
Second, we write its state equation as
Lyt+1 = yt
Third, we rewrite (4), substituting Lyt for yt−1 ,
yt = αLyt + (1 − α)Et (yt+1 ) − rt
which is in the required form. To do this using the dsge command, we would type
. dsge ... ///
(y = {alpha}*Ly + (1-{alpha})*F.y - r) ///
(F.Ly = y, state noshock) ///
...
See [DSGE] Intro 4b for a complete example.
Including a lag of a state variable

Lagged state variables are handled in a manner analogous to lagged control variables. Suppose we
have an equation for a state variable zt that contains a lag of itself. For example, this occurs when zt
is a second-order autoregressive process instead of a first-order autoregressive process. Specifically,
our model contains the equation
zt+1 = ρ1 zt + ρ2 zt−1 + t+1 (5)
The problematic term is ρ2 zt−1 because it violates the form required by (2).
The solution entails three steps. First, we define a new state variable Lzt = zt−1 . We define a
new state instead of new control because the lagged state is exogenous. Second, we write its state
equation as
Lzt+1 = zt
Third, we rewrite (5), substituting Lzt for zt−1 ,
zt+1 = ρ1 zt + ρ2 Lzt + t+1
which is in the required form. In the dsge command, we would type

. dsge ... ///
(F.z = {rho1}*z + {rho2}*Lz, state) ///
(F.Lz = z, state noshock) ///
...
See [DSGE] Intro 4c for a complete example.
Including an expectation of a control dated by more than one period ahead

Many models include an expectation of a control variable dated by more than one period ahead.
For example, suppose that the equation for the control variable consumption growth is
ct = (1 − h)wt + hEt ct+2 + rt (6)

78 Intro 4 — Writing a DSGE in a solvable form
where ct is consumption growth, wt is wage growth, and rt is the interest rate. Only expectations of
control variables dated one period ahead are allowed, so Et (ct+2 ) is a problematic term.
The solution entails three steps. First, we define a new control variable F ct = Et (ct+1 ). The
new variable is a control instead of a state because Et (ct+1 ) is endogenous instead of exogenous.
Second, we include this new control equation,
F ct = Et (ct+1 )
Third, we substitute F ct+1 for ct+2 in (6). The equation becomes
ct = (1 − h)wt + hEt (F ct+1 ) + rt

which is in the required form. Using the dsge command, we would type
. dsge ... ///
(c = (1-{h})*w + {h}*F.Fc + r) ///
(Fc = F.c, unobserved) ///
...
See [DSGE] Intro 4d for a complete example.
Including a second-order lag of a control variable

Higher-order lags require defining more than one new state variable, as we illustrate here.
Equations for the state variable capital frequently involve “time to build”, so that the control
variable investment made in time t becomes available only as new capital several periods in the future.
Suppose it takes three periods to convert investment xt into capital. We denote current capital by kt
and write this three-period relationship as
kt+1 = (1 − δ)kt + xt−2 (7)
which specifies that the capital stock tomorrow is the sum of investment made two periods ago and
the capital today that has not depreciated at rate δ .
The term xt−2 is problematic because it is a second-order lag of the state variable xt , which
violates the form required by (2).
The solution entails three steps. First, we define the new state variables L2xt = xt−2 and
Lxt = xt−1 . Second, we write their state equations as
L2xt+1 = Lxt
and
Lxt+1 = xt
Third, we rewrite (7) as
kt+1 = (1 − δ)kt + L2xt
With the dsge command, we could fit this model as
. dsge ... ///
(F.k = (1-{delta})*k + L2x) ///
(F.L2x = Lx, state noshock) ///
(F.Lx = x, state noshock) ///
...
For a complete example, see [DSGE] Intro 4e.

In general, if the lag t − k appears anywhere in any of your equations, you will need k new state
equations.
Including an observed exogenous variable

Sometimes, we want to treat an observed variable as exogenous. Suppose we want to treat the
changes in an observed exchange rate et as exogenous. The problem is that all the observed variables
in a DSGE model must be modeled as endogenous control variables. The solution is to define a control
variable that is equal to a state variable that models the exogenous process. So we could set the
observed exchange rate et to equal the exchange rate state, est ,
et = est
and specify, say, a first-order autoregressive process for est ,
est+1 = ρe est + ηt+1
To fit this type of model using the dsge command, we would type
. dsge ... ///
(e = es) ///
(F.es = {rhoe}*es, state) ///
...
See [DSGE] Intro 4f for a complete example.
Also see
[DSGE] Intro 2 — Learning the syntax
[DSGE] Intro 3 — Classic DSGE examples
Title
Intro 4a — Specifying a shock on a control variable
Description
A shock to a control variable is problematic because it does not fit into the form of a structural
model that is required to solve for the state-space form. This entry shows how to solve this problem
by defining a new state variable and rewriting the equations.

The model
The model
Equations (1)–(3) specify a model of consumption growth and growth in hours worked. The
equation for consumption growth is subject to an additive shock.
ct = (1 − h)wt + hEt ct+1 + t (1)

nt = wt − γct (2)
wt+1 = ρwt + ξt+1 (3)
Equation (1) specifies that consumption growth ct is a linear combination of wage growth wt ,
expected future consumption growth Et ct+1 , and a consumption shock t . Equation (2) specifies that
the growth rate of hours worked nt depends on wage growth and consumption growth. Equation (3)
specifies an autoregressive process for wage growth. The structural parameter h controls the weights
on wage growth and expected future consumption growth in the consumption equation.
One cannot solve the model in (1)–(3) for the state-space form because the shock t is added
to the control equation ct in (1). Shocks to control variables must be reparameterized as new state
variables so that the model can be solved for its state-space form. We define the shocks as new state
variables instead of new control variables because the shocks are exogenous. We rewrite the model
in (1)–(3) so that the additive shock is a new state variable in (4)–(7). Specifically, we define a new
state variable zt such that zt = t and write the model as
ct = (1 − h)wt + hEt ct+1 + zt (4)

wt+1 = ρwt + ξt+1 (6)
zt+1 = t+1 (7)
That we replaced t with the new state variable zt is the only difference between (4) and (1).
Equations (5) and (6) are identical to their counterparts, (2) and (3). Equation (7) specifies the
evolution of the new state variable zt ; it is similar to (6), except that it lacks an autoregressive
component.
80
Intro 4a — Specifying a shock on a control variable 81
We estimate the parameters of the model in (4)–(7) using U.S. data on consumption growth and
growth in hours worked. We specify w and z as state variables. We specify c and n as control
variables. Both c and n are treated as observed. We specify that w and z are subject to shocks.
. dsge (c = (1-{h})*(w) + {h}*F.c + z)
> (n = w - {gamma}*c)
> (F.w = {rho}*w, state)
> (F.z = , state)
DSGE model
/structural
h .7642505 .0360272 21.21 0.000 .6936386 .8348625
gamma .2726181 .1059468 2.57 0.010 .0649663 .4802699
rho .6545212 .0485627 13.48 0.000 .55934 .7497023
sd(e.w) 2.904762 .1958651 2.520873 3.28865

sd(e.z) 2.077219 .1400659 1.802695 2.351743
The estimate of h is 0.76. Being greater than 0.5, this indicates that consumption growth depends
more strongly on expected future consumption growth than on current wage growth.
Also see
Title
Intro 4b — Including a lag of a control variable
Description
dsge and dsgenl do not allow lags of control variables to be included in the model. The structural
form of the model that is required so that the model can be solved for its state-space form does not
include lags of control variables. This entry shows how to fit a DSGE model that involves a lag of
control variables by defining a new state variable and rewriting the equations.

A model with a lagged endogenous variable
A model with a lagged endogenous variable

The model in (1)–(5) is similar to many in monetary economics in that it includes inertia in the
interest rate because the interest rate depends on its lagged value.
pt = βEt pt+1 + κyt (1)

yt = Et yt+1 − (rt − Et pt+1 − ρz zt ) (2)

1
rt = ρr rt−1 + (1 − ρr ) pt + ut (3)
β
zt+1 = ρz zt + t+1 (4)
ut+1 = ρu ut + ξt+1 (5)
Equation (1) specifies the structural equation for inflation pt . Inflation is a linear combination
of expected future inflation Et (pt+1 ) and output growth yt . Equation (2) specifies the structural
equation for output growth. Output growth is a linear combination of expected future output growth
Et (yt+1 ), the interest rate rt , expected future inflation, and the state zt . The state zt is the first-order
autoregressive process that drives output growth. Equation (3) specifies the structural equation for the
interest rate. The interest rate depends on its own lagged value, the inflation rate, and the state ut .
The state ut is the first-order autoregressive process that drives the interest rate. The control variables
in this model are pt , yt , and rt . The state variables are ut and zt .
The term involving rt−1 in (3) is problematic because the lag of a control variable does not fit
into the structure required to solve the model for the state-space form. We accommodate this term
by defining a new state variable Lrt that equals rt−1 and replacing rt−1 in (3) with this new state
variable. We define a new state instead of new control because the lagged control is predetermined,
which makes it exogenous. These changes yield the model in (6)–(11).
82
Intro 4b — Including a lag of a control variable 83
pt = βEt pt+1 + κyt (6)

yt = Et yt+1 − (rt − Et pt+1 − ρz zt ) (7)

1
rt = ρr Lrt + (1 − ρr ) pt + ut (8)
β
Lrt+1 = rt (9)
zt+1 = ρz zt + t+1 (10)
ut+1 = ρu ut + ξt+1 (11)
We will estimate the parameters in the model in (6)–(11) using data on U.S. interest rate r and
inflation p.
. dsge (p = {beta}*F.p + {kappa}*y)
> (y = F.y - (r - f.p - {rhoz}*z), unobserved)
> (r = {rhor}*lr + (1-{rhor})*((1/{beta})*p + u))
> (F.lr = r, state noshock)
84 Intro 4b — Including a lag of a control variable
DSGE model
/structural
beta .5026374 .0791864 6.35 0.000 .347435 .6578398
kappa .176019 .0511049 3.44 0.001 .0758553 .2761827
rhoz .9591408 .0181341 52.89 0.000 .9235986 .994683
rhor -.0939027 .093966 -1.00 0.318 -.2780727 .0902673
rhou .7094626 .0447808 15.84 0.000 .6216939 .7972313
sd(e.u) 2.324309 .3236223 1.690021 2.958597

sd(e.z) .6111535 .108498 .3985014 .8238056
Ironically, the inertia term for the interest rate rhor was not needed in this model. The autoregressive
state ut is probably modeling both the inertia and the persistence in the shocks to the interest rate.
Also see
Title
Intro 4c — Including a lag of a state variable
Description
Lags of state variables are not allowed to be included in the structural models specified in the
dsge and dsgenl commands. These lags do not fit into the form of a structural model that can be
solved for the state-space form. However, this restriction does not prevent us from fitting models that
involve the lag of a state variable. This entry shows how to define a new state variable and rewrite
the equations in the required structural form.

A model with a lagged state variable
A model with a lagged state variable

The model in (1)–(5) is a standard monetary model in which the state-driving output growth yt
is a second-order autoregressive AR(2) process instead of a first-order autoregressive process.
pt = βEt (pt+1 ) + κyt (1)

yt = Et (yt+1 ) − {rt − Et (pt+1 ) − zt } (2)
1
rt = pt + ut (3)
β
zt+1 = ρz1 zt + ρz2 zt−1 + t+1 (4)
ut+1 = ρu ut + ξt+1 (5)
Equation (1) specifies the structural equation for inflation pt . Inflation is a linear combination of
expected future inflation Et (pt+1 ) and output growth yt . Equation (2) specifies the structural equation
for output growth. Output growth is a linear combination of expected future output growth Et (yt+1 ),
the interest rate rt , expected future inflation, and the state zt . The state zt is an AR(2) process that
drives output growth. Equation (3) specifies the structural equation for the interest rate. The interest
rate depends on the inflation rate and the state ut . The state ut is an AR(1) process that drives the
interest rate. The control variables in this model are pt , yt , and rt . The state variables are ut and zt .
85
86 Intro 4c — Including a lag of a state variable
The AR(2) term in (4) is a problematic term because a lag of a state variable does not fit into the
structure required to solve the model for the state-space form. We accommodate this term by defining
a new state variable Lzt that equals zt−1 and replacing zt−1 in (4) with this new state variable. We
define a new state instead of new control because the lagged state is exogenous. These changes yield
the model in (6)–(11).
pt = βEt (pt+1 ) + κyt (6)

yt = Et (yt+1 ) − {rt − Et (pt+1 ) − zt } (7)
1
rt = pt + ut (8)
β
zt+1 = ρz1 zt + ρz2 Lzt + t+1 (9)
ut+1 = ρu ut + ξt+1 (10)
Lzt+1 = zt (11)
We will estimate the parameters in the model in (6)–(11).

Intro 4c — Including a lag of a state variable 87
We estimate the parameters in the model in (6)–(11) using data on U.S. interest rate r and
inflation p.
> (y = F.y - (r - F.p - z), unobserved)
> (r = (1/{beta})*p + u)
> (F.z = {rhoz1}*z + {rhoz2}*Lz, state)
> (F.Lz = z, state noshock)
DSGE model
/structural
beta .5154869 .0770491 6.69 0.000 .3644734 .6665004
kappa .1662432 .0468473 3.55 0.000 .0744242 .2580622
rhou .6979872 .0452071 15.44 0.000 .6093829 .7865915
rhoz1 .6735503 .2666048 2.53 0.012 .1510145 1.196086
rhoz2 .2709982 .2564615 1.06 0.291 -.2316571 .7736536
sd(e.u) 2.315264 .2992174 1.728809 2.90172

sd(e.z) .7720659 .1716228 .4356915 1.10844
Looking at the confidence interval on rhoz2, we find no evidence that the AR(2) term is needed
in this model.
Also see
Title
Intro 4d — Including an expectation dated by more than one period ahead
Description
Expectations of control variables dated by more than one period ahead are not allowed in the
equations specified in dsge and dsgenl because they do not appear in the required form of a structural
model that can be solved for the state-space form. In this entry, we demonstrate how to fit models
with these types of expectations by defining a new variable and rewriting our equations. Because the
expectation dated more than one period ahead is endogenous, the new variable that we define is a
new control variable.

The model
The model
Equations (1)–(4) specify a model of consumption growth and growth in hours worked.
ct = (1 − h)wt + hEt ct+2 + rt (1)

wt+1 = ρw wt + ξt+1 (3)
rt+1 = ρr rt + t+1 (4)
Equation (1) specifies that consumption growth ct is a linear combination of wage growth wt , the
expected value of consumption growth two periods ahead Et ct+2 , and the interest rate rt . Equation (2)
specifies that the growth rate of hours worked nt depends on wage growth and consumption growth.
Equations (3) and (4) specify a first-order autoregressive process for wage growth and for the interest
rate, respectively. The control variables are ct and nt , and the state variables are wt and rt .
88
Intro 4d — Including an expectation dated by more than one period ahead 89
The expected value of consumption growth two periods ahead in (1) is a problematic term because
it does not fit into the structure required to solve for the state-space form. The structure requires
that expectations be only of one-period ahead values. We accommodate this term by defining a new
control variable F ct that equals Et (ct+1 ) and replacing ct+1 in (1) with this new control variable. We
define a new control variable instead of a new state variable because the expected future consumption
is endogenous instead of exogenous. These changes yield the model in (5)–(9).
ct = (1 − h)wt + hEt (F ct+1 ) + rt (5)

F ct = Et (ct+1 ) (7)
wt+1 = ρw wt + ξt+1 (8)
rt+1 = ρr rt + t+1 (9)
New (7) defines the new control variable F ct as the expected value of next period’s consumption.
Equation (5) is (1) but with ct+2 replaced with the new control F ct+1 .
There are now three control variables and only two shocks, so we treat the new control variable F ct
as unobserved. There is a logic to this process. The one-step-ahead structure of state-space models
makes it impossible to solve for terms like Et (ct+2 ). In contrast, it is possible to solve for terms
like Et {Et (ct+1 )} as long as the unobserved Et (ct+1 ) is determined by another equation, which in
this case is (7). In fact, this example illustrates a part of the recursive dynamic modeling approach
at the heart of the DSGE approach to macroeconomics.
90 Intro 4d — Including an expectation dated by more than one period ahead
We estimate the parameters of the model in (5)–(9) using U.S. data on consumption growth and
growth in hours worked.
. dsge (c = (1-{h})*(w) + {h}*F.fc + r)
> (n = w - {gamma}*c)
> (fc = F.c, unobserved)
> (F.w = {rho_w}*w, state)
> (F.r = {rho_r}*r, state)
DSGE model
/structural
h .6617911 .0427917 15.47 0.000 .5779209 .7456612
gamma .273381 .1120734 2.44 0.015 .0537212 .4930408
rho_w .654523 .0485628 13.48 0.000 .5593417 .7497043
rho_r .1050387 .0638171 1.65 0.100 -.0200406 .230118
sd(e.w) 2.905807 .2023203 2.509267 3.302348

sd(e.r) 2.049915 .1437862 1.7681 2.331731
The estimate of h is greater than 0.5, so we conclude that slightly more of the unobserved wage state
is allocated to the expected growth in consumption two periods ahead than to current consumption.
Also see
Title
Intro 4e — Including a second-order lag of a control
Description
Some DSGE models capture delayed effects by including a second-order lag of a control variable
and excluding the first-order lag. The second-order lag is a problematic term that does not fit into
the form required to solve a structural model for its state-space form. This entry shows how to solve
this problem by defining new state variables and rewriting the equations.

The model
The model
Consider a model in which changes in hours worked take two periods to adjust because next
period’s hours have already been budgeted. In this model, the second-order lag of changes in hours
worked is included, and the first-order lag is excluded. Equations (1)–(4) specify such a model of
growth in hours worked and of consumption growth.
nt = b1 nt−2 + wt − γct (1)

ct = (1 − h)wt + hEt ct+1 + rt (2)
wt+1 = ρwt + ξt+1 (3)
rt+1 = t+1 (4)
Equation (1) specifies that the growth rate of hours worked nt depends on a second-order lag of
itself, wage growth wt , and consumption growth ct . Equation (2) specifies that consumption growth
is a linear combination of wage growth, expected future consumption growth Et ct+1 , and the interest
rate rt . Equation (3) specifies an autoregressive process for wage growth. Equation (4) specifies that
interest rate is just a shock. The control variables are nt and ct . The state variables are wt and rt .
One cannot solve the model in (1)–(4) for the state-space form because the problematic term b1 nt−2
does not fit into the required form. To accommodate this term, we define two new state variables, one
for nt−1 and one for nt−2 . We define new state variables instead of new control variables because
lags of the control are predetermined and thus exogenous. The model with new state variables is
91
92 Intro 4e — Including a second-order lag of a control
nt = b1 L2nt + wt − γct (5)

ct = (1 − h)wt + hEt ct+1 + rt (6)
wt+1 = ρwt + ξt+1 (7)
rt+1 = t+1 (8)
Lnt+1 = nt (9)
L2nt+1 = Lnt (10)
Equation (9) defines the new state for nt−1 , and (10) defines L2nt to be the new state for nt−2 .
The L2nt in (5) replaces nt−2 in (1).
We specify n and c as observed control equations. We specify w, r, Ln, and L2n as state equations.
We specify that w and r are subject to shocks; the new states to accommodate nt−2 are not subject
to shocks.
. dsge (n = {b1}*L2n + w - {gamma}*c)
> (c = (1-{h})*w + {h}*F.c + r)
> (F.w = {rho}*w, state)
> (F.r = , state)
> (F.L2n = Ln, state noshock)
> (F.Ln = n, state noshock)
DSGE model
/structural
b1 .132084 .0608727 2.17 0.030 .0127758 .2513922
gamma .3609224 .1298382 2.78 0.005 .1064442 .6154007
h .7238124 .0406724 17.80 0.000 .6440959 .8035289
rho .6177973 .0533568 11.58 0.000 .5132199 .7223746
sd(e.w) 3.033795 .2423843 2.55873 3.508859

sd(e.r) 1.970291 .1574526 1.66169 2.278893
Intro 4e — Including a second-order lag of a control 93
Looking at the confidence interval for b1, we conclude that the second-order lag of hours’ growth
impacts current hours’ growth.
Also see
Title
Intro 4f — Including an observed exogenous variable
Description
All the observed variables in a DSGE model must be modeled as endogenous control variables.
This requirement implies that there is no reduced form for the endogenous variables as a function of
observed exogenous variables. Any variable that is theoretically exogenous must be modeled.
Mechanically, the solution is to define a control variable that is equal to a state variable that models
the exogenous process. We clarify this issue by discussing a model in which the exchange rate is
theoretically exogenous.

The model
The model
We model an economy in which the growth rate of the trade-weighted exchange rate is exogenous
and in which it affects inflation. Henceforth, we call the trade-weighted exchange rate just the exchange
rate. We begin by adding the growth rate of the exchange rate to the inflation equation in the New
Keynesian model of [DSGE] Intro 1.
xt = Et (xt+1 ) − {rt − Et (pt+1 ) − gt } (1)

1
rt = pt + ut (2)
β
pt = βEt (pt+1 ) + κxt + ψet (3)
ut+1 = ρu ut + t+1 (4)
gt+1 = ρg gt + ξt+1 (5)
xt is the output gap, which is modeled as an unobserved control variable. rt is the interest rate, which
is modeled as an observed control variable. pt is the inflation rate, which is modeled as an observed
control variable. gt and ut are first-order autoregressive state variables. et is the growth rate of the
exchange rate, which we want to model as an observed exogenous variable.
94
Intro 4f — Including an observed exogenous variable 95
The model in (1)–(5) cannot be solved because there is no equation for how the observed et
evolves over time. A first-order autoregressive process [AR(1)] is a standard approximation to an
exogenous variable, like the growth in exchange rate. We complete the model by adding an AR(1)
for the unobserved state variable est and an equation linking the unobserved est to the observed et .
xt = Et (xt+1 ) − {rt − Et (pt+1 ) − gt } (6)

1
rt = pt + ut (7)
β
pt = βEt (pt+1 ) + κxt + ψest (8)
et = est (8a)
ut+1 = ρu ut + t+1 (9)
gt+1 = ρg gt + ξt+1 (10)
est+1 = ρe est + ηt+1 (11)
New (8a) specifies how the unobserved state est is transformed into the observed control variable
et . Equation (11) specifies that the unobserved state est is an AR(1) process. Other equations are
unchanged.
96 Intro 4f — Including an observed exogenous variable
We fit this model using data on the U.S. interest rate r, inflation rate p, and the growth rate of
the exchange rate e. The equation (e = es) links the observed variable e to the unobserved state
variable es.
. dsge (x = F.x - (r - F.p - g), unobserved)
> (r = 1/{beta}*p + u)
> (p = {beta}*F.p + {kappa}*x + {psi}*es)
> (e = es)
> (F.u = {rho_u}*u, state)
> (F.g = {rho_g}*g, state)
> (F.es = {rho_e}*es, state)
DSGE model
/structural
beta .5106726 .102212 5.00 0.000 .3103408 .7110044
kappa .0901559 .0334259 2.70 0.007 .0246424 .1556695
psi .01373 .0037684 3.64 0.000 .006344 .021116
rho_u .7789342 .0473941 16.44 0.000 .6860435 .871825
rho_g .9597848 .0206192 46.55 0.000 .919372 1.000198
rho_e .2372213 .0742225 3.20 0.001 .0917479 .3826948
sd(e.u) 2.206083 .3465985 1.526763 2.885404

sd(e.g) .7039306 .1545726 .4009738 1.006887
sd(e.es) 10.48932 .5671985 9.377629 11.60101
From the estimates for psi, it appears that the growth of the exchange rate impacts inflation.
Intro 4f — Including an observed exogenous variable 97
Also see
Title
Intro 4g — Correlated state variables
Description
Many models include correlated state variables. We illustrate how to specify correlated state
variables in a model of output growth yt and inflation pt .

The model
The model
We model output growth yt and inflation pt as functions of domestic and international factors. The
domestic factor gt that drives output growth is a first-order autoregressive process that is also affected
by the international factor zt . The international factor that drives inflation is a simple first-order
autoregressive process.
Mathematically, the model is
yt = Et yt+1 + αpt + gt (1)

pt = zt (2)
gt+1 = ρg gt + ρgz zt + ξt+1 (3)
zt+1 = ρz zt + t+1 (4)
Equation (1) specifies that output growth depends on expected future output growth Et yt+1 , inflation
pt , and the domestic factor gt . This equation has the form of an aggregate demand curve, and
the parameter α is referred to as the slope of the aggregate demand curve. It should be negative.
Equation (2) specifies that inflation is entirely driven by the international factor zt . Equations (3)
and (4) specify that the factors gt and zt follow a first-order vector autoregressive process with
parameters ρg , ρgz , and ρz and with shocks ξt+1 and t+1 . The factors zt and gt are the state
variables, and pt and yt are the observed control variables.
98
Intro 4g — Correlated state variables 99
From the U.S. macroeconomic data, we use the GDP-growth data in y and the inflation data in p
and estimate the parameters of this model.
. dsge (y = F.y + {alpha}*p + g)
> (p = z)
> (F.g = {rho_g}*g + {rho_gz}*z, state)
> (F.z = {rho_z}*z, state)
DSGE model
/structural
alpha -.1130024 .1554398 -0.73 0.467 -.4176589 .1916541
rho_g .3357768 .0610763 5.50 0.000 .2160694 .4554842
rho_gz .0443504 .09013 0.49 0.623 -.1323012 .221002
rho_z .8626564 .0319007 27.04 0.000 .8001322 .9251806
sd(e.g) 2.184806 .2241106 1.745557 2.624054

sd(e.z) 1.146947 .0519234 1.045179 1.248715
The slope of the aggregate demand curve, α, is estimated to be negative as we expected, but the
confidence interval is wide and includes zero. The imprecision in rho gz has caused imprecision in
the estimate of α.
100 Intro 4g — Correlated state variables
Also see
Title
Intro 5 — Stability conditions
Description
We can estimate only the parameters of models that we can solve, and we can solve only models
that satisfy a stability condition. This entry discusses this condition, how it affects estimation in
practice, and how to find initial values that satisfy it.

Why we care about stability
What if the initial values are not saddle-path stable?
Why we care about stability

DSGE models are dynamic systems subject to random shocks. DSGE models that do not spiral out
of control or converge to a single point when shocked are said to be “saddle-path stable”. We can
solve only saddle-path stable DSGE models, and we can estimate only the parameters of models we
can solve.
The parameter values determine whether a DSGE model is saddle-path stable. Most DSGE models are
saddle-path stable for some parameter values and not for other values. As we discussed in Structural
and reduced forms of DSGE models in [DSGE] Intro 1, the structural form of a linear DSGE model is
A0 yt = A1 Et (yt+1 ) + A2 yt + A3 xt (1)
Given this structural form, the values in the Ai and Bj matrices determine saddle-path stability. For
nonlinear DSGE models, saddle-path stability is assessed on the linear approximation to the nonlinear
model at the steady state.
In the process of solving the structural form in (1) and (2) for the state space, the Klein (2000)
solver computes the generalized eigenvalues of a matrix formed from the values in the Ai and Bj
matrices. An eigenvalue is said to be stable when its absolute value is less than 1. The model is
saddle-path stable when the number of stable eigenvalues equals the number of states in the model.
Once the maximization algorithm gets started, saddle-path stability usually is not an issue. Problems
with saddle-path stability usually occur when trying to get the maximization process started. The
initial values for the maximization algorithm must imply saddle-path stability because we cannot
solve a DSGE model for parameter values that do not imply saddle-path stability.
dsge and dsgenl use a series of small positive numbers as the default initial values for the
structural parameters that appear in the A and B matrices. If you know that your model is not
saddle-path stable when a parameter, say, β , is set to 0.1, you should use the from() option to
specify an initial value that does imply saddle-path stability. Alternatively, you could reparameterize
the model so that β = 0.1 does imply saddle-path stability, but this solution tends to be more work
than changing the initial value.
101
102 Intro 5 — Stability conditions
In rare cases, when the true parameters are close to values that are not saddle-path stable, the
maximum likelihood estimator will suffer from convergence problems. Better initial values can help
find the solution in these cases.
What if the initial values are not saddle-path stable?

If the initial values do not imply a saddle-path stable solution, you will see
. dsge ...
model is not saddle-path stable at current parameter values
The number of stable eigenvalues is greater than the number of state
variables.
(output omitted )
or
. dsge ...
The number of stable eigenvalues is less than the number of state
variables.
(output omitted )
In either case, you should specify the solve and noidencheck options and subsequently use
estat stable to find the problem. solve instructs dsge and dsgenl to only solve the model and
not proceed with estimation, and noidencheck suppresses the check of whether the parameters are
identified. estat stable will display the eigenvalues so that you can determine which ones are
saddle-path stable and which ones are not. The trick is to change the starting values so that the
number of stable eigenvalues equals the number of states in the model. Once you have determined
which parameters are problematic, you can use the from() option to change the initial values to ones
that do not have this problem.
Textbooks like those by Canova (2007), DeJong and Dave (2011), Ljungqvist and Sargent (2018),
and Woodford (2003) are full of examples in which some structural parameters must be either greater
than 1 or less than 1 for the model to be saddle-path stable. Similarly, some parameters must be either
greater than or less than 0 for the model to be saddle-path stable. Moving one structural parameter
over one of these boundaries and looking at how the change affects the number of stable eigenvalues
can help find a vector of parameters that yields a saddle-path stable solution.
That the model that you are trying to fit is likely related to other models in the literature or in the
textbooks suggests another strategy. Any such references probably contain some discussion of the set
of values that yield saddle-path stable solutions.
In especially difficult cases, you can use the strategy of solving the problem for a smaller model
and building back up. You temporarily remove equations from the model, find a parameter vector that
yields a saddle-path stable solution for the smaller model, and then progressively add back equations
as you find initial values that yield saddle-path stable solutions.
Intro 5 — Stability conditions 103
Example 1: Finding saddle-path stable initial values

Equations (1)–(6) model the observed control variable (inflation) pt , the unobserved control variable
(output gap) yt , and the observed control variable (interest rate) rt as functions of the states ut and
zt . Lzt+1 is an auxiliary state that allows zt to be a second-order process instead of a first-order
process.
pt = (1/γ)Et (pt+1 ) + κyt (3)

yt = Et (yt+1 ) − {rt − Et (pt+1 ) − zt } (4)
rt = γpt + ut (5)
ut+1 = ρu ut + ξt+1 (6)
zt+1 = ρz1 zt + ρz2 Lzt + t+1 (7)
Lzt+1 = zt (8)
We try to fit this model using default starting values below.

. dsge (p = (1/{gamma})*F.p + {kappa}*y)
> (r = {gamma}*p + u)
> (F.z = {rho_z1}*z + {rho_z2}*Lz, state)
The number of stable eigenvalues is greater than the number of state
variables.
r(498);
The default initial values specify a model that is not saddle-path stable. We specify options solve
and noidencheck, which cause the command to attempt only to solve the model and to skip the
identification check.

> (r = {gamma}*p + u)
> (F.Lz = z, state noshock),
> solve noidencheck
DSGE model
Log likelihood = .
/structural
gamma .21 . . . . .
kappa .22 . . . . .
rho_u .23 . . . . .
rho_z1 .24 . . . . .
rho_z2 .25 . . . . .
sd(e.u) 1 . . .
sd(e.z) 1 . . .

Warning: Model cannot be solved at current parameter values. Current values
imply a model that is not saddle-path stable.
dsge displays the initial values used in the attempted solution, and it displays a warning message
indicating that the model cannot be solved at these values because they imply a model that is not
saddle-path stable.
We use estat stable to look at the eigenvalues implied by the initial values.
. estat stable
Stability results
Eigenvalues
stable -.3942
stable .6342
stable .21
stable .23
unstable 2.605e+16
unstable 1.046
The process is not saddle-path stable.

The process is saddle-path stable when there are 3 stable eigenvalues and 3
unstable eigenvalues.
The model is saddle-path stable when the number of stable eigenvalues equals the number of states
in the model. The initial values yield four stable eigenvalues, but there are only three states in the
model. So the model is not saddle-path stable at these values.
The initial value for gamma in the output from the failed solution attempt was 0.21. From (3), we
see that the coefficient on expected future inflation is greater than one when gamma is less than one.
We suspect that this relationship could be causing the lack of saddle-path stability. Below, we use
option from() to specify 1.2 as an initial value for gamma and attempt to solve the model.
Intro 5 — Stability conditions 105

> (r = {gamma}*p + u)
> (F.Lz = z, state noshock),
> solve noidencheck from(gamma=1.2)
DSGE model
/structural
gamma 1.2 . . . . .
kappa .22 . . . . .
rho_u .23 . . . . .
rho_z1 .24 . . . . .
rho_z2 .25 . . . . .
sd(e.u) 1 . . .
sd(e.z) 1 . . .

There is no warning message, so the parameter values yielded a saddle-path stable solution. As a
final illustration of this point, we display the eigenvalues below.
. estat stable
Stability results
Eigenvalues
stable -.3942
stable .6342
stable .23
unstable 5.380e+16
unstable 1.2
unstable 1.264
The process is saddle-path stable.
We could now proceed with estimation.
References
Canova, F. 2007. Methods for Applied Macroeconomic Research. Princeton, NJ: Princeton University Press.
Ljungqvist, L., and T. J. Sargent. 2018. Recursive Macroeconomic Theory. 4th ed. Cambridge, MA: MIT Press.
Woodford, M. 2003. Interest and Prices: Foundations of a Theory of Monetary Policy. Princeton, NJ: Princeton
University Press.
Also see
[DSGE] Intro 6 — Identification
[DSGE] Intro 7 — Convergence problems
[DSGE] estat stable — Check stability of system
Title
Intro 6 — Identification
Description
We can estimate only the parameters of a DSGE model when the parameters are identified. We
discuss how to find the identification problems and how to fix them.

The parameters that maximize the likelihood function are the maximum likelihood (ML) estimates
of ML estimators like the one in dsge. When there is a unique vector of parameters that maximizes
the likelihood function, the parameters are identified and we can estimate them. When there is more
than one parameter vector that produces the same maximum value of the likelihood function, the
parameters are not identified, and we cannot estimate them because there is no way to choose among
the sets of equally best parameters.
dsge, dsgenl, and all other commands that implement ML estimators check for identification at
the solution by checking that the matrix of second derivatives, known as the Hessian, is full rank at the
maximum. This test works only when the maximization algorithm arrives at a candidate maximum.
In addition, it provides no information about which parameters are not identified.
For many models, when the parameters are not identified, the only output produced by a standard
ML estimator is an ever growing series of “not concave” messages. This issue is especially relevant
to DSGE models because it is easy to write down theoretical DSGE models whose parameters are not
identified. dsge uses the Iskrev (2010) diagnostic to resolve this issue. The Iskrev diagnostic detects
lack of parameter identification and reports which parameters are not identified at specific parameter
values. By default, dsge performs the Iskrev diagnostic at the initial values and at the solution values.
dsgenl performs the Iskrev diagnostic on the linear approximation to the nonlinear model at the
steady state implied by the initial values and checks again at the solution values.
Instead of looking at the rank of the Hessian, the Iskrev (2010) diagnostic checks that all the
parameters affect the autocovariances of the observed control variables. When there is a one-to-one
relationship between the autocovariances for the observed control variables and the model parameters,
the parameters are identified. Intuitively, this one-to-one relationship allows us to back out the model
parameters from the autocovariances that we can always estimate from the data. When some of the
parameters do not affect the autocovariances of the observed control variables, there is no one-to-one
relationship, and the model parameters are not identified.
Simply adding restrictions to parameters can make the parameters of an unidentified model identified.
This fact is frustrating but intuitive. When the parameters are not identified, they are not uniquely
determined by the data generated by the true model. Only adding restrictions, either by defining
constraints or by changing the equations, can identify the parameters.
107
108 Intro 6 — Identification
Example 1: A model with two unidentified parameters

Consider the following model of inflation pt , output growth yt , and the interest rate rt .
pt = βEt (pt+1 ) + κyt

yt = Et (yt+1 ) − γ{rt − Et (pt+1 ) − ρzt }
βrt = pt + βut
zt+1 = ρzt + t+1
ut+1 = δut + ξt+1
This model specifies how the observed control variables pt and rt depend on the unobserved control
variable yt , on the state variables zt and ut , and on the shocks t and ξt .
Let’s try to estimate the parameters of this model.
> (y = F.y -{gamma}*(r - F.p - {rhoz}*z), unobserved)
> (r = (1/{beta})*p + u)
identification failure at starting values
Constrain some parameters or specify option noidencheck. Likely source of
identification failure: {kappa} {gamma}
r(498);
The error message indicates that kappa and gamma do not influence the autocovariances indepen-
dently; they are linearly dependent. If a list of linearly dependent parameters is supplied, then you
must constrain all but one of them prior to estimation. We constrain gamma.
. constraint 2 _b[gamma]=1
And we apply the constraint to the model:

> (y = F.y -{gamma}*(r - F.p - {rhoz}*z), unobserved)
> (r = (1/{beta})*p + u)
> (F.z = {rhoz}*z, state),
> constraint(2)
Intro 6 — Identification 109
DSGE model
( 1) [/structural]gamma = 1
/structural
beta .5146687 .078349 6.57 0.000 .3611076 .6682298
kappa .1659038 .0474064 3.50 0.000 .0729889 .2588187
gamma 1 (constrained)
rhoz .9545255 .0186424 51.20 0.000 .9179871 .9910639
rhou .7005496 .0452601 15.48 0.000 .6118414 .7892577
sd(e.u) 2.318197 .3047421 1.720914 2.915481

sd(e.z) .650711 .1123848 .4304409 .8709811
The remaining free parameters of the model are identified and can be estimated.
Keep in mind that the Iskrev (2010) diagnostic is performed at specific parameter values. Sometimes,
the parameters of a model cannot be identified at a vector of parameter values but can be identified
at practically any other nearby vector. This problem arises when simple initial values cause terms to
cancel out. If you suspect that the identification diagnostic is finding a problem because of initial
values, you can either specify noidencheck or specify other seemingly random initial values.
Reference
Iskrev, N. 2010. Local identification in DSGE models. Journal of Monetary Economics 57: 189–202.
https://doi.org/10.1016/j.jmoneco.2009.12.007.
Also see
[DSGE] Intro 5 — Stability conditions
[DSGE] Intro 7 — Convergence problems
Title
Intro 7 — Convergence problems
Description
Fitting DSGE models is notoriously difficult. We discuss the causes of convergence problems and
some solutions.

When the iteration process never gets started or when it never ends are two types of convergence
problems. When it never gets started, you will see something like
. dsge ...
identification failure at starting values
(output omitted )
This error occurs when the initial values specify a model that violates the stability conditions required
for DSGE estimation; see [DSGE] Intro 5 for a discussion of this issue and some solutions.
When the iteration process never ends, the cause is either that there is no unique solution or that
the optimizer cannot find the unique solution. When there is no unique solution, the parameters are
said to be not identified. In this case, you see (not concave) after each iteration, like
. dsge ...
(output omitted )
Iteration 50: log likelihood = -337504.44 (not concave)
.
.
.
See [DSGE] Intro 6 for a discussion of this issue and some solutions.
Even when the iterations get started and the parameters are identified, convergence problems are
still common. In these cases, there are usually many parameters.
Changing the optimization technique is the simplest method to overcome a convergence problem,
and it can be surprisingly effective. By default, dsge and dsgenl use the Broyden–Fletcher–Goldfarb–
Shanno (BFGS) algorithm for five iterations before switching to a modified Newton–Raphson (NR)
algorithm. The BFGS algorithm is especially effective at finding a maximum of the DSGE log-likelihood
function from poor initial values. In some cases, simply specifying technique(bfgs 200 nr) can
overcome the convergence problem.
The best way to overcome a convergence problem is to provide better initial values. The optimizer
used by dsge and dsgenl will almost always converge when it is provided with good initial values
and the parameters are identified. Specifying good initial values that come from theory or previous
empirical work can solve your convergence problem. You can specify initial values inside the scalar
substitutable expressions or by using the from() option; see examples 1 and 2 for details.
110
Intro 7 — Convergence problems 111
Sometimes, you have a convergence problem and do not have good initial values. One way to get
initial values is to estimate a series of progressively less constrained models. This process usually
produces convergence when the parameters are identified. See example 3 for details.
Example 1: Specifying starting values in scalar substitutable expressions

Models with lagged state variables and other models that allow for more persistent processes can
exhibit convergence problems. Equations (1)–(6) model the observed control variable (inflation) pt ,
the unobserved control variable (output gap) yt , and the observed control variable (interest rate) rt
as functions of the states ut and zt . Lzt+1 is an auxiliary state that allows zt to be a second-order
process instead of a first-order process.
pt = βEt (pt+1 ) + κyt (1)

yt = Et (yt+1 ) − {rt − Et (pt+1 ) − zt } (2)
1
rt = pt + ut (3)
β
ut+1 = ρu ut + ξt+1 (4)
zt+1 = ρz1 zt + ρz2 Lzt + t+1 (5)
Lzt+1 = zt (6)
Suppose that previous empirical work suggests the initial values of β = 0.5, κ = 0.2, ρu = 0.7,
ρz1 = 0.7, and ρz2 = 0.2. Inside a scalar substitutable expression, we can specify an initial value
by typing {parameter = value}. We specify these initial values for the parameters in the scalar
substitutable expressions in our model specification below.
. dsge (p = {beta=.5}*F.p + {kappa=.2}*y)
> (r = (1/{beta})*p + u)
> (F.u = {rho_u=.7}*u, state)
> (F.z = {rho_z1=.7}*z + {rho_z2=.2}*Lz, state)
112 Intro 7 — Convergence problems
DSGE model
/structural
beta .5154873 .0770495 6.69 0.000 .3644731 .6665015
kappa .1662426 .0468473 3.55 0.000 .0744236 .2580616
rho_u .6979877 .0452071 15.44 0.000 .6093834 .786592
rho_z1 .6735472 .2665984 2.53 0.012 .1510239 1.196071
rho_z2 .2710011 .2564554 1.06 0.291 -.2316422 .7736444
sd(e.u) 2.315263 .2992181 1.728806 2.90172

sd(e.z) .7720675 .1716204 .4356977 1.108437
Note that the output includes results for the standard deviations of the shocks, for which there are
no scalar substitutable expressions. In example 2, we illustrate how to specify initial values for all
the parameters using the from() option.
Example 2: Specifying starting values using from()

In example 1, we specified initial values inside the scalar substitutable expressions. You cannot
specify initial values for the standard deviations of the shocks this way, because the shocks do not
appear in any scalar substitutable expressions. We can specify starting values for any, or all, of the
parameters using the from() option.
from() has several syntaxes, but the vector of starting values is most frequently used in DSGE
applications. In this syntax, we specify a row vector containing the initial values β = 0.5, κ = 0.2,
ρu = 0.7, ρz1 = 0.7, ρz2 = 0.2, σe.u = 2.3, and σe.z = 0.7. We begin by defining and listing the
vector of initial values.
. matrix ivalues = (.5, .2, .7, .7, .2, 2.3, .7)
. matrix list ivalues
ivalues[1,7]
c1 c2 c3 c4 c5 c6 c7
r1 .5 .2 .7 .7 .2 2.3 .7
Although we could specify and use the column names of the vector to assign elements of the vector
to parameters, we use from’s suboption copy to make the assignment based on order of appearance
in the vector. Specifying from(, copy) implies that the first element in the vector specifies the initial
value for the first model parameter, the second element in the vector specifies the initial value for
the second model parameter, and so on.

> (r = (1/{beta})*p + u)
> (F.Lz = z, state noshock), from(ivalues, copy)
DSGE model
/structural
beta .5154875 .0770492 6.69 0.000 .3644738 .6665012
kappa .1662423 .0468472 3.55 0.000 .0744236 .2580611
rho_u .6979879 .0452071 15.44 0.000 .6093836 .7865922
rho_z1 .6735481 .2665998 2.53 0.012 .1510221 1.196074
rho_z2 .2710003 .2564567 1.06 0.291 -.2316456 .7736461
sd(e.u) 2.315262 .2992169 1.728808 2.901716

sd(e.z) .7720669 .171621 .435696 1.108438
Example 3: Starting with a more constrained model to obtain convergence

Suppose that we wanted to estimate the parameters of the model in example 1 but that the only
good initial value was that theory predicted that β should be about 0.5. Further suppose that we were
having convergence problems fitting this model. In this case, one strategy would be to constrain β to
0.5 and to constrain κ to a small, nonzero value. The small value minimizes the interaction between p
and y. Making the value nonzero can prevent problems when solving for the state-space form of the
model. The remaining parameters determine the autoregressive processes for the states. Given that
we suspect that the second-order lag is part of the problem, we also constrain it to a small, nonzero
value. We try to estimate the remaining parameters in the output below.
114 Intro 7 — Convergence problems
. constraint define 1 _b[beta] = 0.5

. constraint define 2 _b[kappa] = 0.1
. constraint define 3 _b[rho_z2] = 0.01
> (r = (1/{beta})*p + u)
> (F.Lz = z, state noshock), constraints(1 2 3)
DSGE model
( 2) [/structural]kappa = .1
( 3) [/structural]rho_z2 = .01
/structural
kappa .1 (constrained)
rho_u .7906612 .0102382 77.23 0.000 .7705947 .8107277
rho_z1 .9440872 .0188166 50.17 0.000 .9072073 .980967
rho_z2 .01 (constrained)
sd(e.u) 2.391372 .1083319 2.179045 2.603698

sd(e.z) .6970129 .0731968 .5535498 .840476
The estimator converged. We will want to use these estimates as initial values, so we put them
into the matrix b and list b to confirm it has the desired values.
. matrix b = e(b)
. matrix list b
b[1,7]
/structural: /structural: /structural: /structural: /structural:
beta kappa rho_u rho_z1 rho_z2
y1 .5 .1 .79066118 .94408716 .01
/: /:
sd(e.u) sd(e.z)
y1 2.3913715 .69701292
We use the previous estimates as initial values, drop the constraints, and try to estimate all the
parameters.
> (r = (1/{beta})*p + u)
> (F.Lz = z, state noshock), from(b, copy)
DSGE model
/structural
beta .5154873 .07705 6.69 0.000 .3644721 .6665024
kappa .1662425 .0468474 3.55 0.000 .0744232 .2580617
rho_u .6979879 .0452071 15.44 0.000 .6093835 .7865922
rho_z1 .6735472 .2666187 2.53 0.012 .1509841 1.19611
rho_z2 .2710012 .2564748 1.06 0.291 -.2316803 .7736826
sd(e.u) 2.315264 .2992199 1.728803 2.901724

sd(e.z) .7720676 .1716278 .4356832 1.108452
The estimator converged.

In larger problems, more steps are sometimes required. In these steps, you use the previous
estimates as initial values, but you drop only some of the constraints at each step.
Also see
[DSGE] Intro 6 — Identification
Title
Intro 8 — Wald tests vary with nonlinear transforms
Description
After fitting a DSGE model, we often perform tests of structural parameters, and these tests often
place nonlinear restrictions on the parameters. The values and rejection rates of a Wald test for
different nonlinear expressions of the same null hypothesis are different. We illustrate this issue, show
that likelihood-ratio (LR) tests do not have this problem, and illustrate that you can parameterize your
model in terms of invertible transforms of each parameter.

Wald tests vary with nonlinear transforms
LR tests do not vary with nonlinear transforms
Wald tests vary with nonlinear transforms

Performing a statistical test of whether a structural parameter in a DSGE has a specific value is one
of the most frequent forms of inference after dsge and dsgenl estimation. The null hypothesis in
one of these tests frequently places nonlinear restrictions on the underlying parameters. Two different
nonlinear expressions of the same null hypothesis produce different Wald test statistics in finite
samples and have different rejection rates. In other words, the Wald test is not invariant to nonlinear
transforms of the null hypothesis. The LR test, on the other hand, is invariant to nonlinear transforms
of the null hypothesis.
Example 1: Different values from logically equivalent Wald tests

Equations (1)–(5) specify how the observed control variable inflation pt , the unobserved control
variable output growth yt , and the observed control variable (interest rate) rt depend on the states zt
and ut , given the shocks t and ξt .
pt = βEt (pt+1 ) + κyt (1)

yt = Et (yt+1 ) − {rt − Et (pt+1 ) − ρzt } (2)
rt = (1/β)pt + ut (3)
zt+1 = ρzt + t+1 (4)
ut+1 = δut + ξt+1 (5)
116
Intro 8 — Wald tests vary with nonlinear transforms 117
We estimate the parameters of this model using the macroeconomic data for the United States in
usmacro2.dta.
> (y = F.y - (r - F.p - {rhoz}*z), unobserved)
> (r = (1/{beta})*p + u)
DSGE model
/structural
beta .5146674 .0783489 6.57 0.000 .3611065 .6682283
kappa .1659057 .0474074 3.50 0.000 .072989 .2588224
rhoz .9545256 .0186424 51.20 0.000 .9179872 .991064
rhou .7005482 .0452604 15.48 0.000 .6118394 .789257
sd(e.u) 2.318202 .3047435 1.720916 2.915489

sd(e.z) .6507117 .1123844 .4304423 .8709811
The interest rate equation shown in (3) links the nominal interest rate to the inflation rate. The
coefficient on inflation is 1/β . We test whether this parameter is 1.5, a common benchmark value in
the literature.
. testnl 1/_b[beta] = 1.5
(1) 1/_b[beta] = 1.5
chi2(1) = 2.24
Prob > chi2 = 0.1342
We do not reject the null hypothesis that 1/β is 1.5.

118 Intro 8 — Wald tests vary with nonlinear transforms
If we test the logically equivalent hypothesis that β = 2/3, the statistic and p-value change.
. test _b[beta] =2/3
chi2( 1) = 3.76
Prob > chi2 = 0.0524
The values of these two logically equivalent Wald tests differ because Wald tests are not invariant
to nonlinear transformation. This issue is well known in the literature; see Gregory and Veall (1985)
and Phillips and Park (1988) for details. In this example, the inference of failing to reject the null
hypothesis remains the same when using a 5% significance level, but this is not true in general.
Different formulations of Wald tests can lead to different inferences.
LR tests do not vary with nonlinear transforms
Example 2: LR tests are invariant to nonlinear transforms

We illustrate this feature by performing LR tests that β = 2/3 and that 1/β = 1.5. The cur-
rent estimates are those of the unconstrained model. We repeat these results and store them as
unconstrained.
. dsge
DSGE model
/structural
beta .5146674 .0783489 6.57 0.000 .3611065 .6682283
kappa .1659057 .0474074 3.50 0.000 .072989 .2588224
rhoz .9545256 .0186424 51.20 0.000 .9179872 .991064
rhou .7005482 .0452604 15.48 0.000 .6118394 .789257
sd(e.u) 2.318202 .3047435 1.720916 2.915489

sd(e.z) .6507117 .1123844 .4304423 .8709811
. estimates store unconstrained

Now, we estimate the parameters of the constrained model in which β = 2/3, store the results as
constrained, and perform an LR test of the null hypothesis that β = 2/3.
. constraint 1 _b[beta] = 2/3
> (r = (1/{beta})*p + u)
> constraint(1)
DSGE model
/structural
kappa .1076811 .0276892 3.89 0.000 .0534113 .1619509
rhoz .9538522 .0187789 50.79 0.000 .9170462 .9906581
rhou .7214328 .0439669 16.41 0.000 .6352593 .8076063
sd(e.u) 1.915459 .0867103 1.74551 2.085408

sd(e.z) .4936797 .080513 .3358771 .6514822
. estimates store constrained

. lrtest unconstrained constrained
Likelihood-ratio test
Assumption: constrained nested within unconstrained
LR chi2(1) = 3.08
Prob > chi2 = 0.0794
Note that the value of the LR statistic is 3.08. We now illustrate an LR of the null hypothesis that
1/β = 1.5 produces the same value.
We cannot impose nonlinear restrictions on parameters, so we must begin by reparameterizing
the unconstrained model by replacing {beta} with 1/{beta}. To avoid having {beta} mean two
different things, we write the model in terms of {gamma}=1/{beta} and estimate the parameters:
. dsge (p = 1/{gamma}*F.p + {kappa}*y)

> (r = ({gamma})*p + u)
> from(gamma=2 kappa=0.15 rhou=0.75 rhoz=0.95)
DSGE model
/structural
gamma 1.943005 .2957867 6.57 0.000 1.363273 2.522736
kappa .1659061 .0474073 3.50 0.000 .0729895 .2588226
rhoz .9545256 .0186424 51.20 0.000 .9179872 .991064
rhou .7005481 .0452604 15.48 0.000 .6118393 .7892568
sd(e.u) 2.318205 .3047433 1.720919 2.915491

sd(e.z) .6507124 .1123842 .4304434 .8709813
. estimates store unconstrained2
The estimates of the parameters other than gamma and the value of the log likelihood are nearly
the same as those for the unconstrained model. The value for gamma = 1.94 is the same as
1/beta = 1/0.514 = 1.95. By tightening the convergence tolerance, we could make these values
exactly the same. These values are nearly the same because this example is an instance of a general
property of maximum likelihood estimators. Transforming a parameter by an invertible function does
not change the log likelihood or the other parameter estimates. In other words, maximum likelihood
estimators are invariant to invertible transformations of the parameters; see Casella and Berger (2002,
319) for details.
Having stored the estimates from the unconstrained model, we now estimate the parameters of the
constrained model and store these results in constrained2.
. constraint 2 _b[gamma] = 1.5
. dsge (p = 1/{gamma}*F.p + {kappa}*y)
> (r = ({gamma})*p + u)
> constraint(2)
DSGE model
( 1) [/structural]gamma = 1.5
/structural
gamma 1.5 (constrained)
kappa .1076811 .0276892 3.89 0.000 .0534113 .1619509
rhoz .9538522 .0187789 50.79 0.000 .9170462 .9906581
rhou .7214328 .0439669 16.41 0.000 .6352593 .8076063
sd(e.u) 1.915459 .0867103 1.74551 2.085408

sd(e.z) .4936797 .080513 .3358771 .6514822
. estimates store constrained2
The estimates of the parameters other than gamma and the value of the log likelihood are the same
as those for the constrained model. This is another instance of the invariance of the maximum
likelihood estimator to invertible transformations of the parameters.
Having stored the log likelihoods from the constrained and unconstrained model, we now perform
an LR of the null hypothesis that γ = 1.5.
. lrtest unconstrained2 constrained2
Likelihood-ratio test
Assumption: constrained2 nested within unconstrained2
LR chi2(1) = 3.08
Prob > chi2 = 0.0794
The LR test statistic and its p-value are the same as those reported for the test against the null
hypothesis that β = 2/3, which illustrates that LR tests are invariant to nonlinear transforms.
References
Casella, G., and R. L. Berger. 2002. Statistical Inference. 2nd ed. Pacific Grove, CA: Duxbury.
Gregory, A. W., and M. R. Veall. 1985. Formulating Wald tests of nonlinear restrictions. Econometrica 53: 1465–1468.
https://doi.org/10.2307/1913221.
Phillips, P. C. B., and J. Y. Park. 1988. On the formulation of Wald tests of nonlinear restrictions. Econometrica 56:
1065–1083. https://doi.org/10.2307/1911359.
Also see
[R] lrtest — Likelihood-ratio test after estimation
[R] test — Test linear hypotheses after estimation
Title
Intro 9 — Bayesian estimation
Description
This entry introduces Bayesian estimation of DSGE models. We discuss the setup of such models,
the specification of priors, diagnostics tests, and other statistical and computational details important
to fit Bayesian models.
We assume that you are already familiar with the elements of a DSGE; see [DSGE] Intro 1. For a
comprehensive discussion of Bayesian analysis and Bayesian methods using Stata, see [BAYES] Intro
and [BAYES] Bayesian commands.

Introduction
Principles of Bayesian DSGE estimation
An uninformative prior
An informative prior
Convergence diagnostics
Introduction
Bayesian estimation provides an alternative to maximum likelihood estimation. Bayesian analysis
combines a likelihood model for the data with a specification of prior knowledge of the distribution of
parameters to arrive at a posterior distribution for the parameters. Incorporation of prior information
in DSGE models takes several forms.
First, the prior incorporates information that would be difficult or impossible to specify through
the likelihood, broadening the amount of information that can be incorporated into the model. The
likelihood is typically specified in terms of aggregate equations. However, some information about
the parameters can be drawn from microeconomic evidence. This evidence can be incorporated into
estimation via the prior. Second, the prior incorporates logical bounds on values that a parameter
might take. Some parameters represent shares, like a capital share of income or depreciation rate of
capital, so that only the (0,1) interval is an appropriate value for the parameter. A prior that takes on
positive value in (0,1) but has a zero probability elsewhere restricts the search to the desired region.
Third, the prior can reflect information from theory, for instance, that a parameter should be positive
or should lie above a critical threshold. All of this makes Bayesian estimation of DSGE models an
appealing alternative to classical frequentist estimation.
Stata provides a full suite of commands for Bayesian estimation. See [BAYES] Intro for an
introduction to Bayesian analysis and [BAYES] Bayesian commands for an introduction to Stata’s
implementation of Bayesian analysis. This entry focuses on Stata’s implementation of Bayesian
estimation for DSGE models.
123
124 Intro 9 — Bayesian estimation
Bayesian estimation of DSGE models is accomplished through use of the bayes: prefix. If your
DSGE model is
. dsge (y = z) (F.z = {rho}*z, state)
then the corresponding Bayesian estimates can be obtained with

. bayes, prior({rho}, uniform(0,1)): dsge (y=z) (F.z={rho}*z, state)
The command has two parts: the DSGE model and the prior specification. The DSGE model is specified
exactly as it is when estimating using maximum likelihood. Priors are specified with the prior()
option of bayes. Most bayes commands come with default priors, intended to be uninformative; see
Remarks and examples in [BAYES] bayes. However, in the case of DSGE models, default priors are
only provided for standard deviation parameters. All other model parameters must be accompanied
by an explicit prior. DSGE parameters are often logically bounded to a certain range by the model
structure, such as the (0,1) interval or greater than some critical value, so that uninformative priors
over the entire real line are inappropriate. Because DSGE parameters are model dependent, informative
priors are readily available and in most cases necessary for estimates to be reasonable.
Principles of Bayesian DSGE estimation
An uninformative prior
We begin with an autoregressive model of order 1, written in its state-space form. This model
is used to provide an introduction to the syntax of bayes: dsge and the specification of priors.
See [TS] sspace for more information on the state-space form of an autoregressive model, and see
[BAYES] bayes: var for a more general approach to the estimation of autoregression and vector
autoregression models using Bayesian methods. We can estimate the parameters of this model by
maximum likelihood using the dsge command (see [DSGE] dsge). We use output growth y as the
dependent variable.
. dsge (y = z) (F.z = {rho}*z, state), nolog
DSGE model
/structural
rho .3392504 .0608551 5.57 0.000 .2199765 .4585243
sd(e.z) 3.321503 .1503577 3.026808 3.616199
The first-order autocorrelation of output growth is 0.339, and the standard deviation of output shock
is 3.32.
To estimate these parameters using Bayesian methods, specify a prior for model parameters, and
use the bayes: prefix. Priors for most model parameters are required. Default priors are provided
for the parameters representing the standard deviation of shocks. For a first example, we specify a
flat prior for the autocorrelation parameter ρ in the range (−1, 1). The command is
. bayes, prior({rho}, uniform(-1, 1)) : dsge (y = z) (F.z = {rho}*z, state)
Intro 9 — Bayesian estimation 125
Because Bayesian estimation is a stochastic procedure, we specify the random-number seed to

ensure reproducibility.
. bayes, rseed(17) prior({rho}, uniform(-1,1)): dsge (y=z) (F.z={rho}*z, state)
note: initial parameter vector set to means of priors.
Burn-in ...
Simulation ...
Model summary
Likelihood:
y ~ dsgell({rho},{sd(e.z)})
Priors:
{rho} ~ uniform(-1,1)
{sd(e.z)} ~ igamma(.01,.01)
Bayesian linear DSGE model MCMC iterations = 12,500

Random-walk Metropolis--Hastings sampling Burn-in = 2,500
MCMC sample size = 10,000
Acceptance rate = .2635
Efficiency: min = .07655
avg = .1075
Log marginal-likelihood = -648.5502 max = .1385
Equal-tailed
Mean Std. dev. MCSE Median [95% cred. interval]
rho .3385195 .0611807 .001644 .3388822 .2251003 .465511

sd(e.z) 3.33619 .1564792 .005656 3.330808 3.062057 3.666267
The model summary shows the likelihood model and the priors used for model parameters. The
likelihood model is dsgell(), meaning the state-space form of the DSGE model. The prior for {rho}
is uniform(−1, 1). The prior for the standard deviation of the state variable, {sd(e.z)}, is by default
an inverse-gamma prior with parameters (0.01, 0.01).
The estimation header reports the total number of MCMC draws, the length of the burn-in period,
the number of MCMC draws used for estimation, the number of observations in the dataset, the
acceptance rate, and the efficiency rates of the draws. The acceptance rate specifies the proportion of
proposed parameter values accepted by the MCMC algorithm. For the Metropolis–Hastings algorithm
used by Bayesian DSGE estimation, this number is rarely above 50% and is often below 30%. A low
acceptance rate—say, below 10%—can indicate convergence problems.
The first two columns of the estimation table report the posterior mean and posterior standard
deviation. Because the prior was uninformative, the posterior mean of the autoregressive parameter
{rho}, 0.339, is the same as the maximum likelihood estimate of 0.339. The estimation table also
reports Monte Carlo standard errors, medians, and equal-tailed credible intervals (CrIs).
By default, the Bayesian estimates are based on an MCMC sample size of 10,000 after a 2,500-
iteration burn-in. It is important to verify that the MCMC simulation has converged; otherwise, the
estimates cannot be trusted. We examine convergence diagnostics later in Convergence diagnostics.
An informative prior
Default priors are not provided for user-defined DSGE model parameters. Reasonable priors are
model dependent and reflect the role the parameter plays in the model. Share parameters are bounded
between 0 and 1; autoregressive parameters are bounded between −1 and 1; some parameters must
be larger than a critical value for model stability. Priors reflect these logical constraints. Additionally,
priors reflect information on the probable location of a parameter within that range. Such information
can be gleaned through prior studies or through different sources of information, such as microdata
or economic theory. Thus priors must be specified for user-defined parameters, and informative priors
can affect estimation results.
For standard deviations of shocks, an inverse-gamma prior with the shape and scale parameters of
0.01 is used. You can change the parameters of this prior by specifying the igammaprior() option,
or you can specify a different prior distribution in the prior() option.
Priors provide several useful functions in DSGE models. Many parameters in DSGE models can be
interpreted as share parameters: the capital share, the depreciation rate, the discount rate, etc. Such
parameters are naturally bounded by the interval (0,1), and a uniform prior on (0,1) guarantees that
the estimated parameter value lies within the economically sensible range. The beta distribution is
frequently useful for share parameters, in which greater weight is to be given to a particular region
of the (0,1) interval.
In addition, many parameters in DSGE models have microeconomic interpretation. As such, mi-
croeconomic information can be brought to bear on the estimation by using informative priors.
In Stata, priors are specified using their statistical parameters as arguments. A uniform prior on
(a, b) is specified with uniform(a, b). A normal prior with mean µ and variance σ 2 is specified
with normal(mu, sigma2). A beta distribution with shape parameters α and β is specified with
beta(alpha, beta). Uniform priors are useful for specifying an uninformative prior on a specified
range. Normal priors are useful for parameters that can take on any real value. Beta priors are useful
for parameters that are naturally bounded between 0 and 1, a case that is especially salient for the
parameters of DSGE models. For a full list of priors, see [BAYES] bayes.
We next use an informative prior for {rho} on (0,1), centered at 0.5. An appropriate distribution
for this task is the beta distribution, and a fairly tight prior is obtained with the shape parameters
(50, 50). A graph of the prior is
. twoway function y = betaden(50, 50, x)
4
y
0
0 .2 .4 .6 .8 1
x
Estimation with the beta(50,50) prior is accomplished using the following command:
. bayes, rseed(17) prior({rho}, beta(50,50)): dsge (y=z) (F.z={rho}*z, state)
Burn-in ...
Simulation ...
Model summary
Likelihood:
y ~ dsgell({rho},{sd(e.z)})
Priors:
{rho} ~ beta(50,50)
{sd(e.z)} ~ igamma(.01,.01)

avg = .0822
Equal-tailed
rho .437131 .0373697 .001183 .4355395 .3616302 .5088862

sd(e.z) 3.35469 .1570423 .006181 3.344853 3.057471 3.684154
The acceptance rate is 15%, somewhat lower than the example with an uninformative prior. Sampling
efficiencies are little changed at around 10%. The informative prior has pulled the point estimate
of {rho} away from its maximum likelihood estimate of 0.34; the posterior mean is now 0.437.
The posterior mean for {rho} lies about halfway between the prior mean of 0.5 and the maximum
likelihood estimate of 0.34. The posterior distribution reflects a balance between prior information
and the data at hand. More informative priors, indicated by a smaller standard deviation around the
prior mean, tip the balance toward the prior mean. Larger sample sizes increase the importance of the
data at hand, though macroeconomic datasets often face practical limitations on the available sample
size.
Convergence diagnostics
Continuing with the previous example, we use bayesgraph diagnostics to plot diagnostic charts
for the autoregressive parameter {rho}. Four plots are provided: the MCMC draws, a histogram of
the posterior distribution, the autocorrelation of the draws, and a density plot.
. bayesgraph diagnostics {rho}
rho
Trace Histogram
.6 15
.5 10
.4
5
.3
0 2000 4000 6000 8000 10000 0
Iteration number .3 .4 .5 .6
Autocorrelation Density
0.80 15
All
0.60 1-half
10 2-half
0.40
0.20 5
0.00
0
0 10 20 30 40
Lag .3 .4 .5 .6
We use these plots to visually assess two aspects of the MCMC chain: convergence and efficiency.
There are no single conclusive convergence criteria for the MCMC chain, but some visual diagnostics
can be used to explore MCMC convergence. (You can also run multiple chains to compute Gelman–
Rubin convergence diagnostics; see Convergence diagnostics using multiple chains.) The trace plot
should not exhibit any time trend and should have constant mean and variance. These properties
can be inspected in the trace plot in the top left. The density of the chain should not vary over the
duration of the MCMC sample. Constancy of the distribution can be assessed in the 1-half and 2-half
density plots in the bottom right; if these plots differ dramatically, then the chain has not converged.
For further discussion of convergence, see Convergence of MCMC.
Efficiency and mixing are terms used to describe how quickly the MCMC chain traverses the
posterior domain. A chain that mixes well has high efficiency, low autocorrelation, and traverses the
posterior quickly. A chain that mixes poorly has low efficiency and traverses the posterior slowly,
exhibiting high autocorrelation in the MCMC chain. A chain that is less efficient requires more MCMC
iterations to obtain the same information as a more efficient chain. Efficiency can also be assessed
numerically; see [BAYES] bayesstats ess.
For our autoregressive model, the trace plot shows no time trend and appears to have constant
mean and variance. The 1-half and 2-half density plots overlap with each other and with the full
chain density. These features are indicators that the chain has converged. The autocorrelation of the
chain dies out quickly, indicating that the chain has acceptable efficiency. To verify convergence more
formally, we could use the nchains() option to run multiple chains and to compute Gelman–Rubin
convergence diagnostics.
In [DSGE] Intro 9a and [DSGE] Intro 9b, we demonstrate Bayesian estimation using some of the
models we have already developed in previous sections.
[DSGE] Intro 9a demonstrates Bayesian estimation of a linear New Keynesian model.
[DSGE] Intro 9b demonstrates Bayesian estimation of a nonlinear stochastic growth model.
Also see
[BAYES] bayes: dsge — Bayesian linear dynamic stochastic general equilibrium models
[BAYES] bayes: dsgenl — Bayesian nonlinear dynamic stochastic general equilibrium models
[BAYES] bayes — Bayesian regression models using the bayes prefix
[BAYES] Intro — Introduction to Bayesian analysis
[BAYES] Glossary
Title
Intro 9a — Bayesian estimation of a New Keynesian model
Description
This entry estimates and interprets the parameters of a simple New Keynesian model using Bayesian
methods. We also discuss some postestimation features, diagnostic tests, and how to improve the
sampling efficiency of parameters.

The model
Posterior diagnostics and plots
Improving sampling efficiency
Impulse responses
The model
Equations (1)–(5) specify a canonical New Keynesian model of inflation pt , the output gap xt ,
and the interest rate rt . The linearized model is
pt = βEt (pt+1 ) + κxt (1)
xt = Et (xt+1 ) − {rt − Et (pt+1 ) − gt } (2)
1
rt = pt + ut (3)
ψ
ut+1 = ρu ut + t+1 (4)
gt+1 = ρg gt + ξt+1 (5)
Equation (1) specifies inflation as a linear combination of expected future inflation and the output gap.
Equation (2) specifies the output gap as a linear combination of the expected future output gap, the real
interest rate, and a state variable gt . Equation (3) specifies the interest rate as a linear combination of
inflation and a state variable ut . The state variables are modeled as first-order autoregressive processes.
The state variable ut is the deviation of rt from its equilibrium value of (1/ψ)pt . The state variable
gt is also the deviation of xt from its equilibrium value.
Three of the parameters have a structural interpretation. The parameter κ is known as the slope
of the Phillips curve and is predicted to be positive. The parameter β is the discount factor that
represents the degree to which agents discount the future relative to the current period. The parameter
1/ψ measures the degree to which interest rates react to movements in inflation.
The choice of notation is intentional and indicates one way to use priors and parameter transfor-
mations. For model stability, the coefficient 1/ψ must be greater than one. For theoretical reasons,
the region (1.5, 2) is of particular interest. As such, the parameter ψ must lie between 0 and 1. By
writing the model in terms of ψ , we use a beta distribution to restrict its range and to put larger
weight on the region that is of interest. For example, a beta distribution for ψ that is centered at 0.67
maps on to a prior for 1/ψ that places much of its mass around 1.5, and this would be a good mix
of logical and theoretical restrictions.
130
Intro 9a — Bayesian estimation of a New Keynesian model 131
We specify priors for the model parameters. The discount rate β must lie between 0 and 1, with
common values in the range (0.90, 0.99). The price-adjustment parameter κ is usually thought to be
small and positive. The autocorrelation parameters must lie in (−1, 1) but are typically believed to
be positive and closer to 1 than 0. The parameter 1/ψ must be greater than 1 for model stability;
hence, ψ must be between 0 and 1, making the beta distribution a good candidate. The parameters
of the beta prior for ψ place its center of mass on 1.5, a value commonly found in the theoretical
literature.
Priors were chosen to match the above theoretical considerations. For the discount rate β , a beta
distribution with shape parameters (95, 5) is used. These shape parameters place the prior mean at
0.95 and place most of the prior mass in the region between 0.9 and 1. For the price-adjustment
parameter κ, a beta distribution with shape parameters (30, 70) is used. These shape parameters place
the prior mean at 0.3 and place most of the prior mass in the region between 0.2 and 0.4. For the
Taylor rule parameter ψ , a beta distribution with shape parameters (0.67, 0.33) is used. These shape
parameters place the prior mean at 0.67, so that its inverse is 1.5, a common value for this parameter
in the literature. For the autoregressive parameters, a beta distribution with shape parameters (75, 25)
is used. This places the prior mean of each state variable’s autoregressive parameter at 0.8, reflecting
a prior belief that the state variables show a fairly high degree of persistence.
. bayes, prior({beta}, beta(95, 5))
> prior({kappa}, beta(30, 70))
> prior({psi}, beta(67, 33))
> prior({rhou}, beta(75, 20))
> prior({rhog}, beta(75, 20))
> rseed(17) :
> dsge (p = {beta}*F.p + {kappa}*x)
> (x = F.x - (r - F.p - g) , unobserved)
> (r = 1/{psi}*p + u)
Burn-in ...
Simulation ...
Model summary
Likelihood:
p r ~ dsgell({beta},{kappa},{psi},{rhou},{rhog},{sd(e.u)},{sd(e.g)})
Priors:
{beta} ~ beta(95,5)
{kappa} ~ beta(30,70)
{psi} ~ beta(67,33)
{rhou rhog} ~ beta(75,20)
{sd(e.u) sd(e.g)} ~ igamma(.01,.01)
132 Intro 9a — Bayesian estimation of a New Keynesian model

avg = .02799
Equal-tailed
beta .9330586 .0273878 .002619 .9354095 .8728009 .975816

kappa .150112 .0247365 .001796 .1485866 .1038851 .203525
psi .5905049 .0405852 .002398 .5895752 .5146669 .6752993
rhou .6275518 .0256116 .001511 .6283089 .5764364 .6770674
rhog .9061189 .0131007 .001059 .9069891 .8800901 .9294626
sd(e.u) 2.114667 .1380783 .006028 2.104597 1.869307 2.400718
sd(e.g) .5579649 .0578754 .002862 .5574834 .4491618 .6778793
Header output repeats the prior specification and reminds us that we are fitting a DSGE model. The
MCMC acceptance rate is 0.2209, with efficiencies ranging from 1% to 5%. An acceptance rate of
20% to 25% is typical for these models. Acceptance rates that are too low indicate that a large portion
of the proposed MCMC iterations were rejected, so that regions of high posterior probability were
not sufficiently explored. Acceptance rates that are too high indicate that the MCMC iterations stayed
in a relatively small area of high probability and did not sufficiently explore the parameter region.
Efficiency is linked to the autocorrelation of the MCMC draws, with higher efficiency indicating lower
autocorrelation.
Turning to parameter estimates, we see the posterior mean for {beta} is 0.93, close to the prior
mean of 0.95. The posterior mean for {kappa} is 0.15, about halfway between the prior mean of 0.3
and the maximum likelihood estimate of 0.08 found in [DSGE] Intro 3a. The autoregressive parameters
for the state variables are positive, with the state variable g showing autocorrelation {rhog} of 0.91
and the state variable u showing autocorrelation {rhou} of 0.63.

We begin by examining the efficiency of the MCMC draws.
. bayesstats ess
Efficiency summaries MCMC sample size = 10,000
avg = .02799
max = .05246
ESS Corr. time Efficiency
beta 109.35 91.45 0.0109

kappa 189.60 52.74 0.0190
psi 286.40 34.92 0.0286
rhou 287.47 34.79 0.0287
rhog 153.01 65.36 0.0153
sd(e.u) 524.63 19.06 0.0525
sd(e.g) 408.84 24.46 0.0409
Some parameters have sampling efficiencies around 1%, indicating poor mixing. Low efficiency
implies that it takes longer for the MCMC chain to explore the posterior distribution. We will come
back to the issue of improving sampling efficiency in Improving sampling efficiency below.
Sometimes, the object of interest is a function of parameters rather than the parameter itself. Such
a situation occurs in this model with the Taylor rule inflation-adjustment parameter 1/ψ . The output
is reported in terms of ψ by default, but what we really want to look at is 1/ψ . We can analyze the
posterior distribution of 1/ψ using the postestimation commands available after the bayes prefix.
The bayesstats summary command produces posterior summary statistics for functions of model
parameters. We investigate the parameter 1/ψ in some detail now.
. bayesstats summary (1/{psi})
Posterior summary statistics MCMC sample size = 10,000
expr1 : 1/{psi}
Equal-tailed
expr1 1.701543 .1181073 .006993 1.696136 1.480825 1.943004
The posterior mean is 1.7, somewhat higher than the prior mean of 1.5.
Next, we investigate the behavior of the MCMC chain using bayesgraph. (Actually, this is something
that we should have done before obtaining summary statistics to verify that the corresponding MCMC
chain converged.) To view the MCMC chain for 1/ψ , bayesgraph diagnostics can be specified
with the desired expression directly, bound in parentheses.
. bayesgraph diagnostics (1/{psi})
expr1
Trace Histogram
2.4 4
2.2
3
2
1.8 2
1.6
1
1.4
0 2000 4000 6000 8000 10000 0
Iteration number 1.4 1.6 1.8 2 2.2
1.00 4
All
0.80 1-half
3
0.60 2-half
0.40 2
0.20
1
0.00
0
0 10 20 30 40
Lag 1.4 1.6 1.8 2 2.2 2.4
expr1: (1/{psi})
The trace plot shows reasonably good mixing, and the autocorrelations decay at a moderate pace. The
density plot shows that the first- and second-half densities are similar to the density of the full MCMC
sample. Densities that differ substantially in their first and second half can indicate nonconvergence.
Next, we compare the prior distribution and the posterior distributions for some model parameters
directly. The bayesgraph kdensity command plots the kernel density of the posterior draws. The
addplot() option can be used to add the prior to the plot. We plot the prior and posterior of the
price-adjustment parameter {kappa} and the discount rate {beta}.
. bayesgraph kdensity {kappa},

> addplot(function Prior = betaden(30,70, x),
> legend(on label(1 "Posterior")))
. bayesgraph kdensity {beta},
> addplot(function Prior = betaden(95, 5, x),
Density of kappa Density of beta
20 20
15 15
Posterior Posterior
10 10
Prior Prior
5 5
0 0
0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1
The price-adjustment parameter {kappa} is now centered on 0.15, rather than the prior mean of 0.3.
By contrast, the discount rate parameter {beta} has shown little updating; the data are uninformative
on this dimension, and the posterior overlaps with the prior.
Next, we turn to the Taylor rule parameter ψ .
. bayesgraph kdensity {psi},
> legend(on label(1 "Posterior"))
Density of psi
10
6
Posterior
Prior
4
0
0 .2 .4 .6 .8 1
The posterior density for ψ places more weight on smaller values compared with the prior, implying
a posterior density for 1/ψ that places more weight on larger values than the prior.
Finally, bayesgraph kdensity all, byparm provides posterior kernel densities for all model
parameters.
. bayesgraph kdensity _all, byparm
Densities
beta kappa psi
15 15 10
10 10 5
5 5
0 0 0
.8 .85 .9 .95 1 .1 .15 .2 .25 .4 .5 .6 .7
rhou rhog sd(e.u)

15 30 3
10 20 2
5 10 1
0 0 0
.5 .55 .6 .65 .7 .85 .9 .95 1.5 2 2.5 3
sd(e.g)
8
6
4
2
0
.4 .6 .8
Graphs by parameter
Such a graph can be useful in summarizing the posterior distribution of all parameters. Roughness
observed in these densities can be smoothed by running a longer MCMC chain.
Improving sampling efficiency

The bayesstats ess output indicates lower sampling efficiency for several model parameters.
Blocking of parameters can improve sampling efficiency. Parameters that are to be sampled indepen-
dently are specified using the block() option of bayes. Efficiency can also be enhanced by running
a longer MCMC chain, accomplished with the mcmcsize() option of bayes. A longer burn-in period
can help with convergence. To demonstrate these procedures, we block some of the parameters and
increase the length of the MCMC chain.
. bayes, prior({beta}, beta(95, 5))

> prior({kappa}, beta(30, 70))
> prior({psi}, beta(67, 33))
> prior({rhou}, beta(70, 20))
> prior({rhog}, beta(70, 20))
> rseed(17)
> block({kappa}) block({rhou}) block({sd(e.u)})
> mcmcsize(20000) dots :
> dsge (p = {beta}*F.p + {kappa}*x)
> (x = F.x - (r - F.p - g) , unobserved)
> (r = 1/{psi}*p + u)
Burn-in 2500 aaaaaaaaa1000aaaaaaaaa2000aaaaa done
Simulation 20000 .........1000.........2000.........3000.........4000.........5
> 000.........6000.........7000.........8000.........9000.........10000........
> .11000.........12000.........13000.........14000.........15000.........16000.
> ........17000.........18000.........19000.........20000 done
Model summary
Likelihood:
p r ~ dsgell({beta},{kappa},{psi},{rhou},{rhog},{sd(e.u)},{sd(e.g)})
Priors:
{beta} ~ beta(95,5)
{kappa} ~ beta(30,70)
{psi} ~ beta(67,33)
{rhou rhog} ~ beta(70,20)
{sd(e.u) sd(e.g)} ~ igamma(.01,.01)

avg = .03764
Equal-tailed
beta .9369982 .0266526 .001015 .9401569 .8745056 .9795873

kappa .1543932 .0267139 .001229 .1525291 .1081603 .2125788
psi .5912198 .0415403 .001646 .5908567 .5109676 .6738902
rhou .6201272 .0272482 .001369 .620959 .5658674 .6728684
rhog .905648 .0145937 .000405 .9062614 .8759466 .9333209
sd(e.u) 2.11736 .1461866 .00511 2.104432 1.85604 2.439449
sd(e.g) .5544951 .0588497 .0019 .5528083 .4448559 .6758059
The above command placed three parameters into their own blocks: {kappa}, {rhog}, and {sd(e.g)}.
(In this example, we specified three separate block() options, but we could have used the shortcut
block({kappa rhog sd(e.g)}, split)). Blocking improves efficiency at the cost of longer run
time. We see that efficiency has improved overall, with the minimum rising to 2% from 1% compared
with the estimation without blocking.
. bayesstats ess
avg = .03764
max = .06486
beta 689.22 29.02 0.0345

kappa 472.47 42.33 0.0236
psi 637.19 31.39 0.0319
rhou 395.87 50.52 0.0198
rhog 1297.26 15.42 0.0649
sd(e.u) 818.49 24.44 0.0409
sd(e.g) 959.44 20.85 0.0480
Effective sample sizes have improved. In particular, the effective sample size for the price-adjustment
parameter {kappa} has more than doubled from 189.60 to 472.47.
Impulse responses
To compute Bayesian impulse–response functions, use bayesirf create after estimation. We
also need to save off our MCMC dataset. To do this, we replay the bayes command with the saving()
option.
. bayes, saving(bdsge_nksim, replace)
With the MCMC results saved, we construct impulse–response functions.

. bayesirf set bayesirf.irf
. bayesirf create nkmodel
Results can be tabulated or graphed with bayesirf table and bayesirf graph, respectively.
. bayesirf graph irf, impulse(g) response(x p r g) byopts(yrescale)
nkmodel, u, p nkmodel, u, r
0 .6
.4
-.5
.2
-1
0
nkmodel, u, u nkmodel, u, x
3 0
2 -1
-2
1
-3
0
0 2 4 6 8 0 2 4 6 8
Step
95% equal-tailed CrI Posterior mean of IRF
Plotted are the responses of the output gap x, inflation p, interest rate r, and the state variable u to
a shock to u. The steps are in the units of time that were used for estimation, so in this case, one
step is one quarter. Eight steps are computed and plotted by default. A rise in the monetary shock
causes inflation and interest rates to fall. The output gap becomes negative, indicating that this is
a contractionary shock. All three endogenous variables return smoothly to their steady-state values
over time, as the effect of the shock dissipates.
Along with point estimates, 95% equal-tailed credible intervals are shown. The point estimates
are comparable with those seen in [DSGE] Intro 1. But the credible intervals are somewhat narrower
than the confidence intervals because of fairly tight priors used.
Also see
[BAYES] bayes: dsge postestimation — Postestimation tools for bayes: dsge and bayes: dsgenl
Title
Intro 9b — Bayesian estimation of stochastic growth model
Description
This introduction estimates and interprets the parameters of a simple stochastic growth model using
Bayesian methods.

The model
Impulse responses
The model
The model contains equations that jointly determine output Yt , the interest rate Rt , consumption
Ct , capital Kt , and productivity Zt . The model contains four parameters: α, β , δ , and ρ. This model
is a variant on the model used in Schmitt-Grohé and Uribe (2004):
( −1 )
Ct+1
1 = βEt (1 + Rt+1 − δ) (1)
Ct
Yt = Zt Ktα (2)
Rt = αZt Ktα−1 (3)
Kt+1 = Yt − Ct + (1 − δ)Kt (4)
ln(Zt+1 ) = ρ ln(Zt ) + et+1 (5)
Equation (1) defines a relationship between consumption growth Ct+1 /Ct and the real interest
rate Rt+1 . Equation (2) is a production function for output Yt as a function of productivity Zt and
capital Kt . Equation (3) is a model for the interest rate. Equation (4) is the equation for capital
accumulation; capital in the next period is equal to underappreciated capital this period (1 − δ)Kt
plus unconsumed output Yt − Ct . Equation (5) is a law of motion for productivity Zt . The parameter
β is a discount factor in the consumption equation, the parameter α is a production parameter in the
output equation, the parameter δ is a depreciation parameter in the capital equation, and the parameter
ρ is a persistence parameter in the productivity equation.
The state variables are the current-period capital stock and the level of productivity, (Kt , Zt ).
The control variables are consumption, the interest rate, and output, (Ct , Rt , Yt ). We estimate the
parameters of the linearized version of the model using Bayesian methods.
139
140 Intro 9b — Bayesian estimation of stochastic growth model
We specify priors for the model parameters. The discount rate β must lie between 0 and 1, with
common values in the range (0.95, 0.99). The capital share α must lie between 0 and 1 and is usually
estimated to be about one-third. The depreciation rate δ must lie between 0 and 1, and a common
value for it is 0.025 or 0.10. The autocorrelation parameter ρ must be less than 1 in absolute value
and is usually thought to be positive and close to 1. Our prior choices for all parameters are driven
by these theoretical considerations. As all four parameters are plausibly restricted to the unit interval,
a beta distribution is chosen for all four priors.
The parameters of the beta distribution were chosen to put the weight of prior mass on theoretically
appropriate values. For the discount factor {beta}, this is the range (0.90, 0.99). For the depreciation
parameter {delta}, this is the range (0.03, 0.05). For the capital share {alpha}, this is the range
(0.3, 0.4). For the autoregressive parameter, this is the range (0.6, 0.99). The prior means for each
parameter are as follows: {beta} prior mean is 0.95; {delta} prior mean is 0.044; {alpha} prior
mean is 0.38; and {rhoz} prior mean is 0.8.
. bayes, prior({beta}, beta( 95, 5))
> prior({delta}, beta( 40, 860))
> prior({alpha}, beta(360, 590))
> prior({rhoz}, beta( 80, 20)) dots rseed(17) :
> dsgenl (1 = {beta}*(c/F.c)*(1 + F.r - {delta}))
> (y = z*k^({alpha}))
> (r = {alpha}*y/k)
> (F.k = y - c + (1-{delta})*k)
> exostate(z) endostate(k) observed(y) unobserved(c r)
Burn-in 2500 aaaaaaaaa1000aaaaaaaaa2000..... done
Simulation 10000 .........1000.........2000.........3000.........4000.........
> 5000.........6000.........7000.........8000.........9000.........10000 done
Model summary
Likelihood:
y ~ dsgell({beta},{delta},{alpha},{rhoz},{sd(e.z)})
Priors:
{beta} ~ beta(95,5)
{delta} ~ beta(40,860)
{alpha} ~ beta(360,590)
{rhoz} ~ beta(80,20)
{sd(e.z)} ~ igamma(.01,.01)
Intro 9b — Bayesian estimation of stochastic growth model 141
Bayesian first-order DSGE model MCMC iterations = 12,500

avg = .05
Equal-tailed
beta .9696788 .0138534 .000918 .9714437 .9379939 .9904246

delta .0412205 .0063456 .000272 .0408097 .029939 .0545771
alpha .3790517 .0154948 .000645 .3792245 .3486508 .4096519
rhoz .6178456 .0442668 .001815 .618141 .5340957 .7060575
sd(e.z) 3.54923 .1751019 .007431 3.544049 3.238064 3.931496
The model summary reports the prior and likelihood specifications, including the default inverse-gamma
prior for the standard deviation of the shock.
The output header reports the burn-in length and MCMC sample size, as well as information about
the efficiency of the Metropolis–Hastings sampler. The overall acceptance rate is 0.25, with sampling
efficiencies between 0.023 and 0.059.
The posterior mean for {beta} is 0.97, close to the prior mean of 0.95. The posterior mean
for {delta} is 0.041, close to its prior mean of 0.044. The posterior mean for {alpha} is 0.38,
identical to its prior mean. The posterior mean for {rhoz} is 0.62, substantially different from its
prior mean of 0.80. Overall, many of the parameters show little updating, indicating that the likelihood
is uninformative along several dimensions of the model’s parameter space. The posterior results for
{beta}, {delta}, and {alpha} are mainly driven by the prior.

We begin by investigating effective sample sizes for each parameter.
. bayesstats ess
avg = .05
max = .05948
beta 227.97 43.87 0.0228

delta 545.58 18.33 0.0546
alpha 576.58 17.34 0.0577
rhoz 594.79 16.81 0.0595
sd(e.z) 555.25 18.01 0.0555
The effective sample size for the discount factor {beta} is somewhat low relative to the other
parameters, which indicates that blocking may improve sampling efficiency.
Because {rhoz} was the only internal parameter to receive substantial updating, we look at its
full set of posterior diagnostic plots.
. bayesgraph diagnostics {rhoz}
rhoz
Trace Histogram
.8 10
8
.7
6
.6 4
2
.5
0 2000 4000 6000 8000 10000 0
Iteration number .5 .6 .7 .8
10
0.80 All
8 1-half
0.60
2-half
6
0.40
4
0.20
2
0.00
0
0 10 20 30 40
Lag .5 .6 .7 .8
Autocorrelations tail off at a moderate pace, the trace plot shows reasonable mixing, and the density
plot shows that the first- and second-half densities do not substantially differ from the full-sample
density.
Next, we generate prior–posterior plots for two parameters, the capital share and the autoregressive
parameter.
. bayesgraph kdensity {alpha},
. bayesgraph kdensity {rhoz},
Density of alpha Density of rhoz
10
25
8
20
6
15
Posterior Posterior
Prior Prior
10 4
5 2
0 0
0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1
Intro 9b — Bayesian estimation of stochastic growth model 143
The posterior density of {alpha} does not differ from its prior density. This situation indicates a flat
likelihood along the {alpha} dimension. By contrast, the posterior density of {rhoz} does differ
from its prior density. The posterior mean has fallen to 0.6 from 0.8.
Prior–posterior plots for the discount rate {beta} and the depreciation parameter {delta} are
. bayesgraph kdensity {beta},
. bayesgraph kdensity {delta},
Density of beta Density of delta
80
30
60
20
Posterior Posterior
40
Prior Prior
10
20
0 0
0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1
The posterior distribution of {beta} lies close to its prior. The posterior distribution of {delta}
lies on top of its prior, indicating that the data provide little information along this dimension of the
model.
Impulse responses
Next, we explore the response of model variables to a shock to the state variable z. We begin by
saving off our MCMC results in a dataset.
. bayes, saving(bayes_dsgenl_sim, replace)
Next, we set up the impulse–response function file and impulse–response functions themselves.
We are using the [BAYES] bayesirf command. We specify step(20) to plot the first 20 periods after
the shock. Because the unit of time in this model is one quarter, 20 periods correspond to 5 years.
. bayesirf set rbcmodel

. bayesirf create model1, step(20)
. bayesirf graph irf, impulse(z) byopts(yrescale)
model1, z, c model1, z, k model1, z, r

1.5 4
1.5
1 2
1
.5 .5 0
0 0 -2
0 5 10 15 20
model1, z, y model1, z, z
4 4
3 3
2 2
1 1
0 0
0 5 10 15 20 0 5 10 15 20
Step
95% equal-tailed CrI Posterior mean of IRF
Each panel displays the response of one model variable to the impulse. Each step is one quarter, so
that four steps are one year after the shock. In the top-left panel, consumption c rises and follows a
mostly flat trajectory for the first eight periods after the shock and then falls to return to its steady
state. In the top-middle panel, the capital stock k does not move in the first period but rises afterward
in a hump-shaped pattern. In the top-right panel, the interest rate r rises on impact, remains elevated
for the first four periods, and then dips below its steady-state value in the fifth period; it returns
to its steady state from below. In the bottom-left panel, output y rises on impact and then declines
monotonically back to its steady state. The bottom-right panel shows the evolution of the state variable
z itself; it rises on a shock and then falls monotonically back to its steady state.
Reference
Schmitt-Grohé, S., and M. Uribe. 2004. Solving dynamic general equilibrium models using a second-order approximation
to the policy function. Journal of Economic Dynamics and Control 28: 755–775. https://doi.org/10.1016/S0165-
1889(03)00043-5.
Also see
[BAYES] bayes: dsge postestimation — Postestimation tools for bayes: dsge and bayes: dsgenl
Title
dsge — Linear dynamic stochastic general equilibrium models
Description Menu Syntax Options Remarks

Stored results Methods and formulas References Also see
Description
dsge fits linear dynamic stochastic general equilibrium (DSGE) models to multiple time series.
DSGE models are systems of equations that are motivated by economic theory. In these systems,
expectations of future values of variables can affect the current values. The parameters of these
models are often directly interpretable in terms of economic theory.
Menu
Statistics > Multivariate time series > Dynamic stochastic general equilibrium (DSGE) models > Linear DSGE
models
Syntax

dsge eqlist if in , options
options Description
Main
constraints(constraints) apply specified linear constraints
noidencheck do not check for parameter identification
solve return model solution at initial values
SE/Robust
vce(vcetype) vcetype may be oim or robust
Reporting
level(#) set confidence level; default is level(95)
nocnsreport do not display constraints
display options control columns and column formats and line width
Maximization
maximize options control the maximization process
Advanced
nodemean do not demean data prior to estimation
post force posting of estimation results in the event of errors
caused by lack of identification or stability
idtolerance(#) set tolerance used for identification check; seldom used
lintolerance(#) set tolerance used for linearity check; seldom used
coeflegend display legend instead of statistics
145
146 dsge — Linear dynamic stochastic general equilibrium models
You must tsset your data before using dsge; see [TS] tsset.
bayes and collect are allowed; see [U] 11.1.10 Prefix commands. For more details, see [BAYES] bayes: dsge.
coeflegend does not appear in the dialog box.
See [U] 20 Estimation and postestimation commands for more capabilities of estimation commands.
Below we present the full specification of eqlist. You may prefer to start with the syntax discussion
in [DSGE] Intro 2.
eqlist is
eq
eq eqlist
eq is
cntrl eq
state eq
cntrl eq contains
(ocntrl var = termlist)
(ucntrl var = termlist, unobserved)
(parm exp * ocntrl var = termlist)
(parm exp * ucntrl var = termlist, unobserved)
state eq contains
(F.state var = , state)

(F.state var = termlist, state noshock )

(parm exp * F.state var = termlist, state noshock )
ocntrl var is Stata variable to be treated as an observed control in the system
ucntrl var is name to be treated as an unobserved control in the system
state var is name to be treated as an unobserved state in the system
If there happens to be a Stata variable with the same name as ucntrl var or state var,
the variable is ignored and plays no role in the estimation.
termlist is
term
term + termlist
term - termlist
term is
rhs var
parm exp * rhs var
parm exp * (termlist)
rhs var is
ocntrl var
ucntrl var
state var
F.ocntrl var
F.ucntrl var
dsge — Linear dynamic stochastic general equilibrium models 147
parm exp is scalar substitutable expression;

parm spec elements are allowed and expected
rhs var are not allowed
parm spec is
{parm name}
{parm name=initial val}
parm name is name used to identify model parameter
initial val is numeric literal; specifies an initial value
Options

Main
constraints(constraints); see [R] Estimation options.
noidencheck skips the check that the parameters are identified at the initial values. Models that
are not structurally identified can still converge, thereby producing meaningless results that only
appear to have meaning; thus care should be taken in specifying this option. See [DSGE] Intro 6
for details.
solve puts the model into state-space form at the initial parameter values. No standard errors are
produced.

SE/Robust
vce(vcetype) specifies the type of standard error reported, which includes types that are derived
from asymptotic theory (oim) and that are robust to some kinds of misspecification (robust); see
[R] vce option.

Reporting
level(#), nocnsreport; see [R] Estimation options.
display options: noci, nopvalues, cformat(% fmt), pformat(% fmt), sformat(% fmt), and nol-
stretch; see [R] Estimation options.

Maximization

maximize options: difficult, technique(algorithm spec), iterate(#), no log, trace,
gradient, showstep, hessian, showtolerance, tolerance(#), ltolerance(#),
nrtolerance(#), and from(init specs); see [R] Maximize.

Advanced
nodemean prevents dsge from removing the mean of the observed control variables prior to estimation.
post causes dsge to post the parameter vector into e(), even in the event of errors that arise from
checking stability conditions or identification.
idtolerance(#) specifies the tolerance used in the identification diagnostic. The default is
idtolerance(1e-6).
lintolerance(#) specifies the tolerance used in the linearity diagnostic. The default is
lintolerance(1e-12).
The following option is available with dsge but is not shown in the dialog box:
coeflegend; see [R] Estimation options.
Remarks
For an introduction to what dsge can do and how to use it, see [DSGE] Intro 1. It is highly
recommended that you read the introduction first.
For examples of dsge, see the examples of classic DSGE models in [DSGE] Intro 3a, [DSGE] Intro 3b,
and [DSGE] Intro 3c. Additional examples are presented in [DSGE] Intro 4a–[DSGE] Intro 4g. See
[DSGE] Intro 4 for an overview of these examples.
Stored results
dsge stores the following in e():
Scalars
e(N) number of observations
e(k) number of parameters
e(k eq) number of equations in e(b)
e(k dv) number of dependent variables
e(k state) number of state equations
e(k control) number of control equations
e(k shock) number of shocks
e(k observed) number of observed control equations
e(k stable) number of stable eigenvalues
e(ll) log likelihood
e(tmin) minimum time in sample
e(tmax) maximum time in sample
e(rank) rank of VCE
e(ic) number of iterations
e(rc) return code
e(converged) 1 if converged, 0 otherwise
Macros
e(cmd) dsge
e(cmdline) command as typed
e(depvar) unoperated names of dependent variables
e(state) unoperated names of state variables
e(control) unoperated names of control variables
e(observed) unoperated names of observed control variables
e(title) title in estimation output
e(tvar) variable denoting time within groups
e(tmins) formatted minimum time
e(tmaxs) formatted maximum time
e(tsfmt) format for the current time variable
e(vce) vcetype specified in vce()
e(vcetype) title used to label Std. err.
e(opt) type of optimization
e(method) likelihood method
e(idencheck) passed, failed, or skipped
e(technique) maximization technique
e(properties) b V
e(estat cmd) program used to implement estat
e(predict) program used to implement predict
e(marginsnotok) predictions disallowed by margins
Matrices
e(b) parameter vector
e(Cns) constraints matrix
e(ilog) iteration log (up to 20 iterations)
e(eigenvalues) generalized eigenvalues
e(gradient) gradient vector
e(shock coeff) estimated shock coefficient matrix
e(transition) estimated state transition matrix
e(policy) estimated policy matrix
e(V) variance–covariance matrix of the estimators
e(V modelbased) model-based variance
Functions
e(sample) marks estimation sample
In addition to the above, the following is stored in r():

Matrices
r(table) matrix containing the coefficients with their standard errors, test statistics, p-values,
and confidence intervals
Note that results stored in r() are updated when the command is replayed and will be replaced when
any r-class command is run after the estimation command.
Methods and formulas

A DSGE model is a system of equations in which the values of the control variables yt and of
the state variables xt depend on their values in previous periods and their one-period-ahead rational
expectations; see [DSGE] Intro 1 for an introduction to control variables and state variables. The
structural form of a DSGE model is
A0 yt = A1 Et (yt+1 ) + A2 yt + A3 xt (1)
where the entries in Ai and Bi are functions of the parameter vector θ and C is a selector matrix of
zeros and ones that puts shocks in the state equations that include them. The A0 and B0 matrices are
diagonal, and A2 has zeros on its diagonal. See [DSGE] Intro 1 for details about how the structural
parameters θ get mapped into entries in Ai and Bi matrices.
dsge estimates the structural parameters θ by maximum likelihood. The remainder of this entry
provides some details about this process.
We can estimate only the parameters of structural models like (1) and (2) that we can solve for the
state-space form that expresses the control variables as functions of the state variables and expresses
the state variables as first-order autoregressive processes, as in (3) and (4),
yt = G(θ)xt (3)
xt+1 = H(θ)xt + M(θ)t+1 (4)
where G is known as the policy matrix and H is known as the state transition matrix. The elements
in the policy and state transition matrices are functions of the structural parameters that appear in the
matrices A and B.
dsge implements the solution technique described in Klein (2000), which obtains the state-space
form in (3) and (4) from the structural model in (1) and (2).
We can only solve a structural model for its state-space form when i) the structural form has
the structure implied by (1) and (2) and when ii) the values of the structural parameters imply that
the model neither spirals out of control nor converges to a point, a condition known as saddle-path
stability.
Condition (i) seems trivial, but some minor manipulation is frequently required to make it true.
Many DSGE models include a problematic term like a shock to a control variable or a lagged state
variable. These problematic terms do not fit into the structure required by (1) and (2), but the model
is easily rewritten to accommodate these problematic terms; see [DSGE] Intro 4 for details.
Condition (ii), the saddle-path stability condition, depends on the values of the parameters. We
can only solve models for parameter values that imply a saddle-path-stable model; see [DSGE] Intro 5
for details.
Assuming that the shocks t are mean-zero, independent, and normally distributed implies that
(3) and (4) form a linear state-space model in which the coefficients are nonlinear functions of the
structural parameters θ. This structure allows dsge to estimate the parameters θ and the standard
deviations of the shocks by maximum likelihood.
The Kalman filter is used to form the log-likelihood function. The Kalman filter is a method for
recursively obtaining linear, least-squares forecasts conditional on past information. These forecasts are
used to construct the log likelihood, assuming normality and stationarity. See Methods and formulas
in [TS] sspace for details of the Kalman filter.
When the shocks are independent and identically distributed, but not normally distributed, the
maximum likelihood estimator is consistent for θ, but vce(robust) standard errors must be used.
(The process of using a maximum likelihood estimator under weaker distributional assumptions and
correcting the standard errors is known as a quasimaximum likelihood estimator in the literature.)
One final caveat is that only a subset y1,t of the controls is treated as observed in some models.
This caveat is handled by augmenting the state-space systems in (3) and (4) with the observation
equation
y1,t = Dyt
where D is a selection matrix that picks out the rows in G corresponding to elements in the control
vector that are observed.
Asymptotic standard errors for postestimation objects such as the entries in the state-space matrices
and the impulse responses are obtained using the delta method. See Serfling (1980, sec. 3.3) for a
discussion of the delta method.
References
Schenck, D. 2017. Estimating the parameters of DSGE models. The Stata Blog: Not Elsewhere Classified.
https://blog.stata.com/2017/07/11/estimating-the-parameters-of-dsge-models/.
. 2018. Dynamic stochastic general equilibrium models for policy analysis. The Stata Blog: Not Elsewhere
Classified. https://blog.stata.com/2018/04/23/dynamic-stochastic-general-equilibrium-models-for-policy-analysis/.
Serfling, R. J. 1980. Approximation Theorems of Mathematical Statistics. New York: Wiley.
Also see
[TS] tsset — Declare data to be time-series data
[TS] var — Vector autoregressive models
[U] 20 Estimation and postestimation commands
Title
dsge postestimation — Postestimation tools for dsge
Postestimation commands predict Remarks and examples Methods and formulas

Also see
Postestimation commands
The following postestimation commands are of special interest after dsge:
Command Description
estat policy display policy matrix of estimated model
estat stable assess stability of the system
estat transition display transition matrix of estimated model
irf create and analyze IRFs and FEVDs
The following standard postestimation commands are also available:
Command Description
estat ic Akaike’s, consistent Akaike’s, corrected Akaike’s, and Schwarz’s Bayesian
information criteria (AIC, CAIC, AICc, and BIC)
estat summarize summary statistics for the estimation sample
estat vce variance–covariance matrix of the estimators (VCE)
estimates cataloging estimation results
etable table of estimation results
forecast dynamic forecasts and simulations
lincom point estimates, standard errors, testing, and inference for linear combinations
of coefficients
lrtest likelihood-ratio test
nlcom point estimates, standard errors, testing, and inference for nonlinear combina-
tions of coefficients
predict one-step-ahead predictions, prediction errors, and other diagnostic measures
predictnl point estimates, standard errors, testing, and inference for generalized predic-
tions
test Wald tests of simple and composite linear hypotheses
testnl Wald tests of nonlinear hypotheses
152
dsge postestimation — Postestimation tools for dsge 153
predict
Description for predict

predict creates new variables containing predictions such as expected values. Predictions are
available as static one-step-ahead predictions or as dynamic multistep predictions, and you can control
when dynamic predictions begin.
Menu for predict

Statistics > Postestimation
Syntax for predict

predict type { stub* | newvarlist } if in , statistic options
statistic Description
Main
xb linear prediction for observed variables
states linear prediction for latent state variables
options Description
Options
rmse(stub* | newvarlist)
put estimated root mean squared errors of predicted statistics in new
variables
dynamic(time constant) begin dynamic forecast at specified time
Advanced
smethod(method) method for predicting unobserved states
method Description
onestep predict using past information
filter predict using past and contemporaneous information
Options for predict

Main
xb, the default, calculates the linear predictions of the observed variables.
states calculates the linear predictions of the latent state variables.
154 dsge postestimation — Postestimation tools for dsge

Options
rmse(stub* | newvarlist) puts the root mean squared errors of the predicted statistics into the specified
new variables. The root mean squared errors measure the variances due to the disturbances but do
not account for estimation error.
dynamic(time constant) specifies when predict starts producing dynamic forecasts. The specified
time constant must be in the scale of the time variable specified in tsset, and the time constant
must be inside a sample for which observations on the dependent variables are available. For
example, dynamic(tq(2008q4)) causes dynamic predictions to begin in the fourth quarter of
2008, assuming that your time variable is quarterly; see [D] Datetime. If the model contains
exogenous variables, they must be present for the whole predicted sample.

Advanced
smethod(method) specifies the method for predicting the unobserved states, smethod(onestep)
and smethod(filter), and causes different amounts of information on the dependent variables
to be used in predicting the states at each time period.
smethod(onestep), the default, causes predict to estimate the states at each time period using
previous information on the dependent variables. The Kalman filter is performed on previous
periods, but only the one-step predictions are made for the current period.
smethod(filter) causes predict to estimate the states at each time period using previous and
contemporaneous data by the Kalman filter. The Kalman filter is performed on previous periods
and the current period. smethod(filter) may be specified only with states.

For examples of estat policy, see [DSGE] Intro 1, [DSGE] Intro 3a, and [DSGE] Intro 3c.
For examples of estat transition, see [DSGE] Intro 1, [DSGE] Intro 3a, and [DSGE] Intro 3b.
For an example of estat stable, see [DSGE] Intro 5.
For examples of irf after dsge, see [DSGE] Intro 1, [DSGE] Intro 3b, and [DSGE] Intro 3c.
For an example of forecast after dsge, see [DSGE] Intro 1.
For examples of predict after dsge, see [DSGE] Intro 3a.

Estimating the unobserved states is the key to predicting the observed variables.
By default and with the smethod(onestep) option, predict estimates the state in each period
by applying the Kalman filter to all previous periods and only making the one-step prediction to the
current period.
With the smethod(filter) option, predict estimates the states in each period by applying
the Kalman filter on all previous periods and the current period. The computational difference be-
tween smethod(onestep) and smethod(filter) is that smethod(filter) performs the update
step on the current period while smethod(onestep) does not. The statistical difference between
smethod(onestep) and smethod(filter) is that smethod(filter) uses contemporaneous in-
formation on the observed variables while smethod(onestep) does not.
The observed control variables are predicted by plugging in the estimated states.
dsge postestimation — Postestimation tools for dsge 155
Also see
[DSGE] estat policy — Display policy matrix
[DSGE] estat transition — Display state transition matrix
Title
dsgenl — Nonlinear dynamic stochastic general equilibrium models
Description Menu Syntax Options Remarks

Stored results Methods and formulas Reference Also see
Description
dsgenl fits nonlinear dynamic stochastic general equilibrium (DSGE) models to multiple time series
via perturbation methods. DSGE models are systems of equations that are motivated by economic
theory. In these systems, expectations of future values of variables can affect the current values. The
parameters of these models are often directly interpretable in terms of economic theory.
Menu
Statistics > Multivariate time series > Dynamic stochastic general equilibrium (DSGE) models > Nonlinear DSGE
models
Syntax

dsgenl (lexp 1 = rexp 1) (lexp 2 = rexp 2) . . . if in ,

observed(varlist) exostate(namelist) options
rexp j and lexp j are substitutable expressions on the right-hand side and left-hand side of equation j.
These are Stata expressions that include scalars, variables, and parameters to be estimated. The
variables may be state variables, observed control variables, and unobserved control variables. The
lead operator, F., can be applied to a variable so that F.varname refers to the expected value of
varname in the next time period. No other time-series operators are allowed; see [DSGE] Intro 4
for information on including additional lags and leads. Parameters to be estimated are enclosed
in curly braces; {beta} is an example of a parameter. Initial values for parameters are given by
including an equal sign and the initial value inside the braces; for example, {beta=2} sets the
initial value for beta to 2. See Specifying the system of nonlinear equations in [DSGE] Intro 2.
156
dsgenl — Nonlinear dynamic stochastic general equilibrium models 157
options Description
Main
∗
observed(varlist) list of observed control variables
unobserved(namelist) list of unobserved control variables
∗
exostate(namelist) list of exogenous state variables
endostate(namelist) list of endogenous state variables
constraints(constraints) apply specified linear constraints
linearapprox take a linear, rather than log-linear, approximation
noidencheck do not check for parameter identification
solve return model solution at initial values
SE/Robust
vce(vcetype) vcetype may be oim or robust
Reporting
nocnsreport do not display constraints
Maximization
maximize options control the maximization process
Advanced
nodemean do not demean data prior to estimation
post force posting of estimation results in the event of errors
caused by lack of identification or stability
idtolerance(#) set tolerance used for identification check; seldom used
steadytolerance(#) set tolerance used for convergence in steady-state calculations;
seldom used
steadyinit(matrix) set initial values for steady state; seldom used
coeflegend display legend instead of statistics
∗
observed() and exostate() are required.
You must tsset your data before using dsgenl; see [TS] tsset.
bayes and collect are allowed; see [U] 11.1.10 Prefix commands. For more details, see [BAYES] bayes: dsgenl.
coeflegend does not appear in the dialog box.
See [U] 20 Estimation and postestimation commands for more capabilities of estimation commands.
Options

Main
observed(varlist) specifies which variables that appear in the model equations are observed control
variables. observed() is required.
unobserved(namelist) specifies which variables that appear in the model equations are unobserved
control variables. The names must be valid Stata names but need not exist in your dataset.
exostate(namelist) specifies which variables that appear in the model equations are exogenous state
variables. Exogenous state variables are subject to shocks. The names must be valid Stata names
but need not exist in your dataset. There must be the same number of exogenous state variables
as there are observed control variables in your model. exostate() is required.
158 dsgenl — Nonlinear dynamic stochastic general equilibrium models
endostate(namelist) specifies which variables that appear in the model equations are endogenous
state variables. Endogenous state variables are not subject to shocks. The names must be valid
Stata names but need not exist in your dataset.
constraints(constraints); see [R] Estimation options.
linearapprox specifies that a linear approximation be applied to the DSGE model rather than a
log-linear approximation. In either case, an approximation is applied. In a log-linear approximation,
variables are interpreted as percentage deviations from steady state. In a linear approximation,
variables are interpreted as unit deviations from steady state.
noidencheck skips the check that the parameters are identified at the initial values. Models that
are not structurally identified can still converge, thereby producing meaningless results that only
appear to have meaning; thus, care should be taken in specifying this option. See [DSGE] Intro 6
for details.
solve puts the model into state-space form at the initial parameter values. No standard errors are
produced.

SE/Robust
vce(vcetype) specifies the type of standard error reported, which includes types that are derived
from asymptotic theory (oim) and that are robust to some kinds of misspecification (robust); see
[R] vce option.

Reporting
level(#), nocnsreport; see [R] Estimation options.

Maximization

maximize options: difficult, technique(algorithm spec), iterate(#), no log, trace,
gradient, showstep, hessian, showtolerance, tolerance(#), ltolerance(#),
nrtolerance(#), and from(init specs); see [R] Maximize.

Advanced
nodemean prevents dsgenl from removing the mean of the observed control variables prior to
estimation.
post causes dsgenl to post the parameter vector into e(), even in the event of errors that arise
from checking stability conditions or identification.
idtolerance(#) specifies the tolerance used in the identification diagnostic. The default is
idtolerance(1e-6).
steadytolerance(#) specifies the tolerance used in calculations of the model’s steady state. The
default is steadytolerance(1e-17).
steadyinit(matrix) specifies a vector of initial values for use in finding the steady state.
The following option is available with dsgenl but is not shown in the dialog box:
coeflegend; see [R] Estimation options.
Remarks
For an introduction to what dsgenl can do and how to use it, see [DSGE] Intro 1. It is highly
recommended that you read the introduction first.
For examples of dsgenl, see the examples of classic DSGE models in [DSGE] Intro 3d, [DSGE] In-
tro 3e, and [DSGE] Intro 3f.
Stored results
dsgenl stores the following in e():
Scalars
e(N) number of observations
e(k) number of parameters
e(k eq) number of equations in e(b)
e(k dv) number of dependent variables
e(k state) number of state equations
e(k control) number of control equations
e(k shock) number of shocks
e(k observed) number of observed control equations
e(k stable) number of stable eigenvalues
e(ll) log likelihood
e(tmin) minimum time in sample
e(tmax) maximum time in sample
e(rank) rank of VCE
e(ic) number of iterations
e(rc) return code
e(converged) 1 if converged, 0 otherwise
Macros
e(cmd) dsgenl
e(cmdline) command as typed
e(depvar) unoperated names of dependent variables
e(state) unoperated names of state variables
e(shock) unoperated names of state variables subject to shocks
e(control) unoperated names of control variables
e(observed) unoperated names of observed control variables
e(title) title in estimation output
e(tvar) variable denoting time within groups
e(tmins) formatted minimum time
e(tmaxs) formatted maximum time
e(tsfmt) format for the current time variable
e(vce) vcetype specified in vce()
e(vcetype) title used to label Std. err.
e(opt) type of optimization
e(method) likelihood method
e(idencheck) passed, failed, or skipped
e(technique) maximization technique
e(properties) b V
e(estat cmd) program used to implement estat
e(predict) program used to implement predict
e(marginsnotok) predictions disallowed by margins
Matrices
e(b) parameter vector
e(Cns) constraints matrix
e(ilog) iteration log (up to 20 iterations)
e(eigenvalues) generalized eigenvalues
e(gradient) gradient vector
e(shock coeff) estimated shock coefficient matrix
e(transition) estimated state transition matrix
160 dsgenl — Nonlinear dynamic stochastic general equilibrium models
e(steady) estimated steady-state vector

e(V) variance–covariance matrix of the estimators
e(V modelbased) model-based variance
Functions
e(sample) marks estimation sample
In addition to the above, the following is stored in r():

Matrices
r(table) matrix containing the coefficients with their standard errors, test statistics, p-values,
and confidence intervals
Note that results stored in r() are updated when the command is replayed and will be replaced when
any r-class command is run after the estimation command.

A DSGE model is a system of equations in which the values of the control variables yt and of
the state variables xt depend on their values in previous periods and their one-period-ahead rational
expectations; see [DSGE] Intro 1 for an introduction to control variables and state variables. The
structural form of a nonlinear DSGE model is
Et {f (xt+1 , yt+1 , xt , yt ; θ)} = 0 (1)
where f is an n × 1 vector of functions, the ny variables yt are control variables, and the nx × 1
variables xt are state variables. The parameter θ contains the model’s structural parameters.
Fernández-Villaverde, Rubio-Ramı́rez, and Schorfheide (2016) contain an introduction to the
solution and estimation of such models. dsgenl proceeds by linearizing (1) around the steady state
and then solving the resulting linear system.
The steady state is found by solving the deterministic system of equations,
f (x, y, x, y; θ) = 0
which is n equations in n unknowns. The location of the steady state is a vector (x, y), each element
of which may depend on the structural parameters.
Once the steady state has been found, the vector of functions f is linearized around that point,
yielding
Et {F1 xt+1 + F2 yt+1 + F3 xt + F4 yt } = 0 (2)
Equation (2) is in the form of a linearized DSGE model, and the remaining solution techniques are
the same as for dsge; see Methods and formulas in [DSGE] dsge. Parameter estimation proceeds on
the basis of the linearized equation shown as (2).
Reference
Fernández-Villaverde, J., J. F. Rubio-Ramı́rez, and F. Schorfheide. 2016. Solution and estimation methods for DSGE
models. In Vol. 2A of Handbook of Macroeconomics, ed. J. B. Taylor and H. Uhlig, chap. 9, 527–724. Amsterdam:
North-Holland. https://doi.org/10.1016/bs.hesmac.2016.03.006.
Also see
[TS] tsset — Declare data to be time-series data
[TS] var — Vector autoregressive models
Title
dsgenl postestimation — Postestimation tools for dsgenl
Postestimation commands predict Remarks and examples Methods and formulas

Also see
Postestimation commands
The following postestimation commands are of special interest after dsgenl:
Command Description
estat covariance display model-implied covariance matrix of model variables
estat policy display policy matrix of estimated model
estat stable assess stability of the system
estat steady display steady state of the system
estat transition display transition matrix of estimated model
irf create and analyze IRFs and FEVDs
The following standard postestimation commands are also available:
Command Description
estat ic Akaike’s, consistent Akaike’s, corrected Akaike’s, and Schwarz’s Bayesian in-
formation criteria (AIC, CAIC, AICc, and BIC)
estat summarize summary statistics for the estimation sample
estat vce variance–covariance matrix of the estimators (VCE)
estimates cataloging estimation results
etable table of estimation results
forecast dynamic forecasts and simulations
lincom point estimates, standard errors, testing, and inference for linear combinations of
coefficients
lrtest likelihood-ratio test
nlcom point estimates, standard errors, testing, and inference for nonlinear combinations
of coefficients
predict one-step-ahead predictions, prediction errors, and other diagnostic measures
predictnl point estimates, standard errors, testing, and inference for generalized predictions
test Wald tests of simple and composite linear hypotheses
testnl Wald tests of nonlinear hypotheses
162
dsgenl postestimation — Postestimation tools for dsgenl 163
predict
Description for predict

predict creates new variables containing predictions such as expected values. Predictions are
available as static one-step-ahead predictions or as dynamic multistep predictions, and you can control
when dynamic predictions begin.
Menu for predict

Syntax for predict

predict type { stub* | newvarlist } if in , statistic options
statistic Description
Main
xb linear prediction for observed variables
states linear prediction for latent state variables
options Description
Options
rmse(stub* | newvarlist)
put estimated root mean squared errors of predicted statistics in new
variables
dynamic(time constant) begin dynamic forecast at specified time
Advanced
smethod(method) method for predicting unobserved states
method Description
onestep predict using past information
filter predict using past and contemporaneous information
Options for predict

Main
xb, the default, calculates the linear predictions of the observed variables.
states calculates the linear predictions of the latent state variables.
164 dsgenl postestimation — Postestimation tools for dsgenl

Options
rmse(stub* | newvarlist) puts the root mean squared errors of the predicted statistics into the specified
new variables. The root mean squared errors measure the variances due to the disturbances but do
not account for estimation error.
dynamic(time constant) specifies when predict starts producing dynamic forecasts. The specified
time constant must be in the scale of the time variable specified in tsset, and the time constant
must be inside a sample for which observations on the dependent variables are available. For
example, dynamic(tq(2008q4)) causes dynamic predictions to begin in the fourth quarter of
2008, assuming that your time variable is quarterly; see [D] Datetime. If the model contains
exogenous variables, they must be present for the whole predicted sample.

Advanced
smethod(method) specifies the method for predicting the unobserved states, smethod(onestep)
and smethod(filter), and causes different amounts of information on the dependent variables
to be used in predicting the states at each time period.
smethod(onestep), the default, causes predict to estimate the states at each time period using
previous information on the dependent variables. The Kalman filter is performed on previous
periods, but only the one-step predictions are made for the current period.
smethod(filter) causes predict to estimate the states at each time period using previous and
contemporaneous data by the Kalman filter. The Kalman filter is performed on previous periods
and the current period. smethod(filter) may be specified only with states.

For examples of estat covariance, see [DSGE] Intro 3e and [DSGE] estat covariance.
For examples of estat policy, see [DSGE] Intro 1, [DSGE] Intro 3d, and [DSGE] Intro 3e.
For examples of estat transition, see [DSGE] Intro 1, [DSGE] Intro 3d, and [DSGE] Intro 3e.
For examples of irf after dsgenl, see [DSGE] Intro 3d and [DSGE] Intro 3e.
For examples of predict and forecast after fitting a DSGE model, see [DSGE] Intro 1. Examples
are shown with estimation results from dsge but apply to results from dsgenl as well.

The most important thing to keep in mind is that postestimation predictions after dsgenl are made
on the basis of the linearized form of the model given to dsgenl.
Estimating the unobserved states is the key to predicting the observed variables.
By default and with the smethod(onestep) option, predict estimates the state in each period
by applying the Kalman filter to all previous periods and only making the one-step prediction to the
current period.
With the smethod(filter) option, predict estimates the states in each period by applying
the Kalman filter on all previous periods and the current period. The computational difference be-
tween smethod(onestep) and smethod(filter) is that smethod(filter) performs the update
step on the current period while smethod(onestep) does not. The statistical difference between
smethod(onestep) and smethod(filter) is that smethod(filter) uses contemporaneous in-
formation on the observed variables while smethod(onestep) does not.
The observed control variables are predicted by plugging in the estimated states.
dsgenl postestimation — Postestimation tools for dsgenl 165
Also see
[DSGE] estat covariance — Display estimated covariances of model variables
[DSGE] estat policy — Display policy matrix
[DSGE] estat steady — Display steady state of nonlinear DSGE model
[DSGE] estat transition — Display state transition matrix
Title
estat covariance — Display estimated covariances of model variables
Description Quick start Menu for estat Syntax

Options Remarks and examples Stored results Methods and formulas
Also see
Description
estat covariance displays model-implied covariances between control variables.
Quick start
Display variances and covariances of all control variables in the model
estat covariance
Include only variances and covariances of x and its lag
estat covariance x L.x
Display the variance of x1 and the covariance of x1 with the lag of x1 and with x2
estat covariance x1, addcovariance(L.x1 x2)
Menu for estat

Syntax

estat covariance varlist , options
varlist may include control variables and their lags. If varlist is not specified, variances and covariances
are reported for all control variables in the model.
options Description
addcovariance(clistc) report additional covariances
nocovariance do not report covariances
post post variances and covariances and their VCE as estimation results
collect is allowed; see [U] 11.1.10 Prefix commands.
Options
addcovariance(clistc) specifies that the covariances between the control variables specified in clistc
and those specified in varlist be displayed. The variances of variables in clistc are not reported.
clistc can contain lags of the control variables in the model.
166
estat covariance — Display estimated covariances of model variables 167
nocovariance specifies that no covariance be displayed. nocovariance may not be specified with
addcovariance().
post causes estat covariance to behave like a Stata estimation (e-class) command. estat
covariance posts the estimated variance–covariance matrix to e(), so you can treat it just as
you would results from any other estimation command.
level(#) specifies the confidence level, as a percentage, for confidence intervals. The default is
level(95) or as set by set level; see [U] 20.8 Specifying the width of confidence intervals.

estat covariance displays covariances between control variables implied by a DSGE model.
We provide examples of the use of estat covariance using the model from [DSGE] Intro 3e.
In that introduction, we estimated some of the parameters of a Real Business Cycle model. Model
setup and estimation were given by
. constraint 1 _b[alpha] = 0.33
. constraint 2 _b[beta] = 0.99
. constraint 3 _b[delta] = 0.025
. constraint 4 _b[chi] = 2
. dsgenl (1/c = {beta}*(1/F.c)*(1+r-{delta}))
> ({chi}*h = w/c)
> (y = c + i)
> (r = {alpha}*y/k)
> (w = (1-{alpha})*y/h)
> (F.k = i + (1-{delta})*k)
> , observed(y) unobserved(c i r w h) exostate(z) endostate(k)
> constraint(1/4) nolog
/structural
chi 2 (constrained)
rho .3132838 .0614709 5.10 0.000 .192803 .4337646
sd(e.z) 2.249022 .101853 2.049394 2.44865

168 estat covariance — Display estimated covariances of model variables
The commands just listed imported data, set up constraints, and estimated the remaining model
parameters. For details of the model, see [DSGE] Intro 3e. The control variables in this model are c,
i, r, w, h, and y.
By default, estat covariance displays the contemporaneous variances and covariances of all
the control variables in the model.
. estat covariance
Delta-method
c
var(c) .8569582 .1674995 5.12 0.000 .5286652 1.185251
cov(c,i) 1.700528 .4329933 3.93 0.000 .8518764 2.549179
cov(c,r) -.4155731 .0553716 -7.51 0.000 -.5240994 -.3070467
cov(c,w) .9560928 .198436 4.82 0.000 .5671653 1.34502
cov(c,h) .0991346 .031423 3.15 0.002 .0375466 .1607226
cov(c,y) 1.055227 .2295048 4.60 0.000 .6054062 1.505049
i
var(i) 208.0512 20.07879 10.36 0.000 168.6975 247.4049
cov(i,r) 50.26689 4.824216 10.42 0.000 40.8116 59.72218
cov(i,w) 25.95045 2.626602 9.88 0.000 20.8024 31.09849
cov(i,h) 24.24992 2.32649 10.42 0.000 19.69008 28.80975
cov(i,y) 50.20036 4.943248 10.16 0.000 40.51178 59.88895
r
var(r) 12.95503 1.275345 10.16 0.000 10.4554 15.45466
cov(r,w) 5.540528 .5242116 10.57 0.000 4.513092 6.567963
cov(r,h) 5.956101 .5726953 10.40 0.000 4.833639 7.078563
cov(r,y) 11.49663 1.096581 10.48 0.000 9.34737 13.64589
w
var(w) 3.893379 .4600228 8.46 0.000 2.991751 4.795007
cov(w,h) 2.937286 .29054 10.11 0.000 2.367838 3.506734
cov(w,y) 6.830665 .7434327 9.19 0.000 5.373564 8.287766
h
var(h) 2.838151 .2712547 10.46 0.000 2.306502 3.369801
cov(h,y) 5.775437 .5612467 10.29 0.000 4.675414 6.875461
y
var(y) 12.6061 1.297192 9.72 0.000 10.06365 15.14855
Applied studies frequently look at the variances and covariances of a subset of the control variables
in the model. Studies also frequently look at the covariances between a subset of the control variables
in the model and the lags of one or two control variables. We illustrate how to obtain these results.
estat covariance — Display estimated covariances of model variables 169
To view only the variance of the consumption c and its covariance with the control variables in
the model, type
. estat covariance c, addcovariance(i r w h y)
Delta-method
c
var(c) .8569582 .1674995 5.12 0.000 .5286652 1.185251
cov(c,i) 1.700528 .4329933 3.93 0.000 .8518764 2.549179
cov(c,r) -.4155731 .0553716 -7.51 0.000 -.5240994 -.3070467
cov(c,w) .9560928 .198436 4.82 0.000 .5671653 1.34502
cov(c,h) .0991346 .031423 3.15 0.002 .0375466 .1607226
cov(c,y) 1.055227 .2295048 4.60 0.000 .6054062 1.505049
Autocovariances are available as well. To view the variance of output and the first-order autoco-
variance of output, type
. estat covariance y, addcovariance(L.y)
Delta-method
y
var(y) 12.6061 1.297192 9.72 0.000 10.06365 15.14855
cov(y,L.y) 4.370091 1.077205 4.06 0.000 2.258808 6.481374
To view the variances and covariance of output and consumption as well as their covariances with
lagged output and lagged consumption, type
. estat covariance y c, addcovariance(L.y L.c)
Delta-method
y
var(y) 12.6061 1.297192 9.72 0.000 10.06365 15.14855
cov(y,c) 1.055227 .2295048 4.60 0.000 .6054062 1.505049
cov(y,L.y) 4.370091 1.077205 4.06 0.000 2.258808 6.481374
cov(y,L.c) 1.000102 .2569841 3.89 0.000 .4964225 1.503782
c
var(c) .8569582 .1674995 5.12 0.000 .5286652 1.185251
cov(y,c) 1.055227 .2295048 4.60 0.000 .6054062 1.505049
cov(c,L.y) 1.000102 .1561667 6.40 0.000 .694021 1.306183
cov(c,L.c) .842517 .1656435 5.09 0.000 .5178617 1.167172
What we typed requested the covariance between y and the lag of c (cov(y,L.c)) and between c
and the lag of y (cov(c,L.y)). In general, these covariances can be different. The structure in this
model causes their estimates to be the same, but their estimated standard errors are different. The
estimated standard errors differ because the derivatives that enter the delta method are different.
170 estat covariance — Display estimated covariances of model variables
Stored results
estat covariance stores the following in r():
Matrices
r(b) estimates
r(V) variance–covariance matrix of the estimates
If post is specified, estat covariance also stores the following in e():

Macros
e(properties) b V
Matrices
e(b) estimates
e(V) variance–covariance matrix of the estimates

Entries in the covariance matrix are functions of the structural parameter vector θ. Standard errors
are calculated using the delta method.
Also see
Title
estat policy — Display policy matrix

Also see
Description
estat policy displays the estimated policy matrix of the state-space form of a DSGE model.
Quick start
Display the estimated policy matrix
estat policy
Same as above, but with 90% confidence intervals
estat policy, level(90)
Menu for estat

Syntax

estat policy , compact post level(#) display options
Options
compact reports only the coefficient values of the estimated policy matrix and displays these coefficients
in matrix form.
post causes estat policy to behave like a Stata estimation (e-class) command. estat policy
posts the policy matrix parameters along with the estimated variance–covariance matrix to e(), so
you can treat the estimated policy matrix as you would results from any other estimation command.
171
172 estat policy — Display policy matrix

The policy matrix is part of the state-space form of a DSGE model. It specifies the model’s control
variables as a function of the model’s state variables.
For examples, see [DSGE] Intro 1, [DSGE] Intro 3a, and [DSGE] Intro 3c.
Stored results
estat policy stores the following in r():
Matrices
r(policy) estimated policy matrix
r(b) estimates
If post is specified, estat policy also stores the following in e():

Macros
e(properties) b V
Matrices
e(b) estimates

Entries in the policy matrix G are functions of the structural parameter vector θ. Standard errors
for G
b are calculated using the delta method.
Also see
Title
estat stable — Check stability of system
Description Menu for estat Syntax Remarks and examples

Stored results Methods and formulas Reference Also see
Description
estat stable displays functions of the model parameters that indicate whether the model is
saddle-path stable at specific parameter values. These results can help you find initial values for
which the model is saddle-path stable. Saddle-path stability is required for solving and estimating the
parameters of DSGE models.
Menu for estat

Syntax
estat stable

DSGE models are dynamic systems that are subject to random shocks. DSGE models that do not
spiral out of control or converge to a single point when shocked are said to be “saddle-path stable”.
We can solve for the state-space form of DSGE models only if they are saddle-path stable, and we
can estimate the parameters of models only if we can solve for the state-space form.
The structural parameter values determine whether a DSGE model is saddle-path stable. In the
process of solving the structural form for the state-space form, the Klein (2000) solver computes the
generalized eigenvalues of a matrix formed from structural parameter values. An eigenvalue is said
to be stable when its absolute value is less than 1. The model is saddle-path stable when the number
of stable eigenvalues equals the number of states in the model.
estat stable displays the generalized eigenvalues implied by the parameter values in e(b),
indicates which are stable and which are unstable, and displays a note indicating whether the model
is saddle-path stable at these parameter values. estat stable can help you find initial values for
the maximization routine when the default values imply a model that is not saddle-path stable; see
[DSGE] Intro 5 for details.
173
174 estat stable — Check stability of system
Stored results
estat stable stores the following in r():
Scalars
r(stable) 1 if stable, 0 otherwise
Matrices
r(eigenvalues) eigenvalues

estat stable displays the generalized eigenvalues computed by the Klein (2000) solver. Values
less than 1 are labeled as stable; values greater than or equal to 1 are labeled as unstable.
Reference
Also see
Title
estat steady — Display steady state of nonlinear DSGE model

Also see
Description
estat steady displays the estimated steady-state values of variables in a nonlinear DSGE model.
Quick start
Display the estimated steady-state values
estat steady
estat steady, level(90)
Menu for estat

Syntax

estat steady , compact level(#) display options
Options
compact reports only the coefficient values of the estimated steady-state vector.

A nonlinear DSGE model is a system of equations describing the relationships among model
variables. These relationships are dynamic and subject to random shocks. When the random shocks
are set to zero, the model variables converge to fixed values over time. This vector of fixed values is
known as the steady-state vector. The steady-state vector solves the system of equations when past
and future values of each variable are replaced with their present value and expectations are dropped.
175
176 estat steady — Display steady state of nonlinear DSGE model
estat steady displays the estimated steady-state value of each variable.

For an example of estat steady, see [DSGE] Intro 3f.
Stored results
estat steady stores the following in r():
Matrices
r(steady) estimated steady-state vector

A nonlinear DSGE model can be written in the structural form
Et {f (xt+1 , yt+1 , xt , yt ; θ)} = 0
The steady state is obtained by solving
f (x, y, x, y; θ) = 0
The resulting (x, y) is a function of the parameter vector θ. Standard errors are calculated using the
delta method.
Also see
Title
estat transition — Display state transition matrix

Also see
Description
estat transition displays the estimated state transition matrix of the state-space form of a
DSGE model.
Quick start
Display the estimated transition matrix
estat transition
estat transition, level(90)
Menu for estat

Syntax

estat transition , compact post level(#) display options
Options
compact reports only the coefficient values of the estimated policy matrix and displays these coefficients
in matrix form.
post causes estat transition to behave like a Stata estimation (e-class) command. estat
transition posts the state transition matrix to e(), so you can treat it as you would results from
any other estimation command.
177
178 estat transition — Display state transition matrix

The state transition matrix is part of the state-space form of a DSGE model. It specifies the transition
matrix of the model’s state variables.
For examples, see [DSGE] Intro 1, [DSGE] Intro 3a, and [DSGE] Intro 3b.
Stored results
estat transition stores the following in r():
Matrices
r(transition) estimated transition matrix
r(b) estimates
If post is specified, estat transition also stores the following in e():

Macros
e(properties) b V
Matrices
e(transition) estimated transition matrix
e(b) estimates

Entries in the state transition matrix H are functions of the structural parameter vector θ. Standard
errors for entries in H
b are calculated using the delta method.
Also see
Glossary
autoregressive process. An autoregressive process is a time series in which the current value of a
variable is a linear function of its own past values and a white-noise error term. A first-order
autoregressive process, denoted as an AR(1) process, is yt = ρyt−1 + t . An AR(p) model contains
p lagged values of the dependent variable.
conditional mean. A conditional mean expresses the mean of one variable as a function of some
other variables. A regression function is a conditional mean.
control variable. A control variable is an endogenous variable. Control variables can be observed or
unobserved.
In a structural DSGE model, the current value of a control variable depends on the current value of
other control variables, on the expected future values of any control variable, and on the current
values of state variables. The current value of a control variable is found by solving the model
for the state-space form.
covariance stationary. A covariance stationary process is a weakly stationary process.
dynamic forecast. A dynamic forecast uses forecasted values wherever lagged values of the endogenous
variables appear in the model, allowing one to forecast multiple periods into the future.
dynamic stochastic general equilibrium model. A dynamic stochastic general equilibrium model is
a multivariate time-series model that specifies the structural relationship between state variables
and control variables and is typically derived from economic theory.
endogenous variable. An endogenous variable is a variable whose values are determined by the
equilibrium of a structural model. The values of an endogenous variable are determined inside the
system.
equilibrium. The equilibrium values of variables in a model are the values that satisfy all the model’s
equations simultaneously.
exogenous variable. An exogenous variable is one whose values change independently of the other
variables in a structural model. The values of an exogenous variable are determined outside the
system. In a time-series model, an exogenous variable is also a predetermined variable.
expected future value. An expected future value is a forecast of the value of a variable in the future
based on current information. In DSGE models, expected future values are computed under rational
expectations.
Under rational expectations, Et (yt+1 ) is the condition mean of yt+1 conditional on the complete
history of all variables in the model and the structure of the model itself.
forward operator. The forward operator F denotes the value of a variable at time t + 1. Formally,
F yt = yt+1 , and F 2 yt = F yt+1 = yt+2 . A forward operator is also called a lead operator.
identified. Identified is a condition required to estimate the parameters of a model. In other words,
only identified parameters can be estimated.
In DSGE models, the parameters are identified when there is a unique parameter vector that
maximizes the likelihood function. For a discussion of identification, see [DSGE] Intro 6.
impulse–response function. An impulse–response function (IRF) measures the effect of a shock to an
endogenous variable on itself or another endogenous variable. The k th impulse–response function
of variable i on variable j measures the effect on variable j in period t + k in response to a
one-unit shock to variable i in period t, holding everything else constant.
179
180 Glossary
independent and identically distributed. A series of observations is independent and identically

distributed (i.i.d.) if each observation is an independent realization from the same underlying
distribution. In some contexts, the definition is relaxed to mean only that the observations are
independent and have identical means and variances; see Davidson and MacKinnon (1993, 42).
initial values. Initial values specify the starting place for the iterative maximization algorithm used
by DSGE.
Kalman filter. The Kalman filter is a recursive procedure for predicting the state vector in a state-space
model.
lag operator. The lag operator L denotes the value of a variable at time t − 1. Formally, Lyt = yt−1 ,
and L2 yt = Lyt−1 = yt−2 .
lead operator. See forward operator.
likelihood-ratio (LR) test. The LR test is a classical testing procedure used to compare the fit of
two models, one of which, the constrained model, is nested within the full (unconstrained) model.
Under the null hypothesis, the constrained model fits the data as well as the full model. The LR test
requires one to determine the maximal value of the log-likelihood function for both the constrained
and the full models.
linearized model. A linearized model is an approximation to a model that is nonlinear in the variables
and nonlinear in the parameters. The approximation is linear in variables but potentially nonlinear
in the parameters. In a linearized model, variables are interpreted as unit deviations from steady
state.
log-linear model. A log-linear model is an approximation to a model that is nonlinear in the variables
and nonlinear in the parameters. In a log-linear model, variables are interpreted as percentage
deviations from steady state.
model solution. A model solution is a function for the endogenous variables in terms of the exogenous
variables. A model solution is also known as the reduced form of a model.
In DSGE terminology, a model solution expresses the control variables as a function of the state
variables alone and expresses the state variables as a function of their values in the previous period
and shocks. The reduced form of a DSGE model is also known as the state-space form of the DSGE
model.
model-consistent expectation. A model-consistent expectation is the conditional mean of a variable
implied by the model under consideration.
For example, under rational expectations the model-consistent expectation of Et (yt+1 ) is the mean
of yt+1 implied by the model, conditional on the realizations of variables dated time t or previously.
nonpredetermined variable. A nonpredetermined variable is a variable whose value at time t is
determined by the system of equations in the model. Contrast with predetermined variable.
null hypothesis. In hypothesis testing, the null hypothesis typically represents the conjecture that one
is attempting to disprove. Often the null hypothesis is that a parameter is zero or that a statistic
is equal across populations.
one-step-ahead forecast. See static forecast.
policy matrix. The policy matrix in the reduced form of a DSGE model is the matrix that expresses
control variables as a function of state variables.
predetermined variable. A predetermined variable is a variable whose value is fixed at time t, given
everything that has occurred previously. More technically, the value of a predetermined variable
is fixed, given the realizations of all observed and unobserved variables at times t − 1, t − 2, . . . .
Glossary 181
rational expectations. A rational expectation of a variable does not deviate from the mean of that
variable in a predictable way. More technically, a rational expectation of a variable is the conditional
mean of the variable implied by the model.
realization. The realization of a random variable is the value it takes on when drawn.
reduced form. The reduced form of a model expresses the endogenous variables as functions of the
exogenous variables.
The reduced form of a DSGE model expresses the control variables as a function of the state
variables alone and expresses the state variables as a function of their values in the previous period
and shocks. The reduced form of a DSGE model is a state-space model.
saddle-path stable. A saddle-path stable model is a structural model that can be solved for its
state-space form. The existence of a saddle-path stable solution depends on the parameter values
of the model. For a discussion of saddle-path stability, see [DSGE] Intro 5.
shock variable. A shock variable is a random variable whose value is specified as an independently
and identically distributed (i.i.d.) random variable. The maximum likelihood estimator is derived
under normally distributed shocks but remains consistent under i.i.d. shocks. Robust standard errors
must be specified when the errors are i.i.d. but not normally distributed.
state transition matrix. The state transition matrix in the reduced form of a DSGE model is the matrix
that expresses how the future values of state variables depend on their current values.
state variable. A state variable is an unobserved exogenous variable.
In DSGE models, a state variable is an unobserved exogenous variable that may depend on its own
previous value, the previous values of other state variables, and shocks.
state-space model. A state-space model describes the relationship between an observed time series
and an unobservable state vector that represents the “state” of the world. The measurement equation
expresses the observed series as a function of the state vector, and the transition equation describes
how the unobserved state vector evolves over time. By defining the parameters of the measurement
and transition equations appropriately, one can write a wide variety of time-series models in the
state-space form.
For DSGE models, the state-space form is the reduced form of the structural model.
The DSGE framework changes the jargon and the structure of state-space models. The measurement
equation is the vector of equations for the control variables, and the transition equation is the
vector of equations for the state variables. In contrast to the standard state-space model, DSGE
models allow a control variable to be unobserved.
static forecast. A static forecast uses actual values wherever lagged values of the endogenous variables
appear in the model. As a result, static forecasts perform at least as well as dynamic forecasts,
but static forecasts cannot produce forecasts into the future when lags of the endogenous variables
appear in the model.
Because actual values will be missing beyond the last historical time period in the dataset, static
forecasts can forecast only one period into the future (assuming only first lags appear in the model);
thus they are often called one-step-ahead forecasts.
steady-state equilibrium. A steady-state equilibrium is a time-invariant rest point of a dynamic
system.
More technically, a steady-state equilibrium is a set of values for the endogenous variables to
which the dynamic system will return after an exogenous variable is changed or a random shock
occurs. This set of values is time invariant in that it does not depend on the time period in which
the change or shock occurs. Multistep dynamic forecasts converge to these values. A steady-state
182 Glossary
equilibrium is also known as a long-run equilibrium because it specifies time-invariant values for
the endogenous variables to which the dynamic system will return, if left unshocked.
stochastic equation. A stochastic equation, in contrast to an identity, is an equation in a forecast model
that includes a random component, most often in the form of an additive error term. Stochastic
equations include parameters that must be estimated from historical data.
stochastic trend. A stochastic trend is a nonstationary random process. Unit-root process and random
coefficients on time are two common stochastic trends. See [TS] ucm for examples and discussions
of more commonly applied stochastic trends.
strict stationarity. A process is strictly stationary if the joint distribution of y1 , . . . , yk is the same
as the joint distribution of y1+τ , . . . , yk+τ for all k and τ . Intuitively, shifting the origin of the
series by τ units has no effect on the joint distributions.
structural model. A structural model specifies the theoretical relationship among a set of variables.
Structural models contain both endogenous variables and exogenous variables. Parameter estimation
and interpretation require that structural models be solved for a reduced form.
trend. The trend specifies the long-run behavior in a time series. The trend can be deterministic or
stochastic. Many economic, biological, health, and social time series have long-run tendencies to
increase or decrease. Before the 1980s, most time-series analysis specified the long-run tendencies
as deterministic functions of time. Since the 1980s, the stochastic trends implied by unit-root
processes have become a standard part of the toolkit.
Wald test. A Wald test is a classical testing procedure used to compare the fit of two models, one
of which, the constrained model, is nested within the full (unconstrained) model. Under the null
hypothesis, the constrained model fits the data as well as the full model. The Wald test requires
one to fit the full model but does not require one to fit the constrained model.
weakly stationary. A process is weakly stationary if the mean of the process is finite and independent
of t, the unconditional variance of the process is finite and independent of t, and the covariance
between periods t and t − s is finite and depends on t − s but not on t or s themselves.
Weakly-stationary processes are also known as covariance stationary processes.
white noise. A variable ut represents a white-noise process if the mean of ut is zero, the variance
of ut is σ 2 , and the covariance between ut and us is zero for all s 6= t.
Reference
Davidson, R., and J. G. MacKinnon. 1993. Estimation and Inference in Econometrics. New York: Oxford University
Press.
Subject and author index
See the combined subject index and the combined author index in the Stata Index.
183

Dsge

Uploaded by

Copyright:

Available Formats

Dsge

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Dsge

Uploaded by

Copyright:

Available Formats

STATA DYNAMIC STOCHASTIC

GENERAL EQUILIBRIUM MODELS

A Stata Press Publication

The suggested citation for this software is

Intro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Introduction to DSGE manual 1

[GSM] Getting Started with Stata for Mac

[M] Mata Reference Manual

Description Remarks and examples Also see

Remarks and examples

[DSGE] Intro 1 Introduction to DSGEs

[DSGE] Intro 3 Classic DSGE examples

[DSGE] Intro 4 Writing a DSGE in a solvable form

[DSGE] Intro 5 Stability conditions

[DSGE] Intro 9 Bayesian estimation

[DSGE] dsge Linear dynamic stochastic general equilibrium models

[BAYES] bayes: dsge Bayesian linear dynamic stochastic general equilibrium

Description Remarks and examples References Also see

Remarks and examples

Introduction to DSGE models

An example: A nonlinear DSGE model

where β is a parameter that captures households’ willingness to delay consumption.

where Gt = Zt+1 /Zt is a state variable that captures changes in Zt .

ln(Gt+1 ) = ρg ln(Gt ) + ξt+1 (7a)

Variable Storage Display Value

datestr str10 %-10s Observation date

Sorted by: dateq

We now have inflation and interest rates.

Specifying the model to dsgenl

which is equivalent to what we wrote in (4a).

Parameter estimation and interpretation of nonlinear DSGEs

Coefficient Std. err. z P>|z| [95% conf. interval]

sd(e.u) 2.317581 .29883 1.731885 2.903277

Coefficient Std. err. z P>|z| [95% conf. interval]

_nl_1 1.955839 .2899594 6.75 0.000 1.387529 2.524149

An example: A linear DSGE model

xt = Et xt+1 − (rt − Et πt+1 − gt ) (4b)

The new parameter κ is a function of the underlying parameter φ from (5a).

Specifying the model to dsge

Parameter estimation and interpretation of linear DSGEs

. dsge (p = {beta}*F.p + {kappa}*x)

Coefficient Std. err. z P>|z| [95% conf. interval]

sd(e.u) 2.31759 .2988025 1.731948 2.903232

Coefficient Std. err. z P>|z| [95% conf. interval]

_nl_1 1.955846 .2899255 6.75 0.000 1.387602 2.524089

Policy and transition matrices

Note: Standard errors reported as missing for constrained transition matrix

. irf graph irf, impulse(u) response(x p r u) byopts(yrescale)

We use tsappend() to extend the dataset by 3 years, or 12 quarters.

. forecast solve, prefix(d1_) begin(tq(2017q1))

Starting period: 2017q1

Starting period: 2014q1

. tsline p d2_p if tin(2010q1, 2021q1), tline(2014q1)

We plot both the observed inflation and the forecast.

Inflation rate Inflation rate (dsgemodel d2_)

Structural and reduced forms of DSGE models

Et {f (xt+1 , yt+1 , xt , yt , θ)} = 0 (9)

where as before yt is a vector of control variables, xt is a vector of state variables, and t is a

where yt is a vector of control variables, xt is a vector of state variables, and t is a vector of

Description Remarks and examples Also see

. dsge (p = {beta}F.p + {kappa}x)

where as before yt is a vector of control variables, xt is a vector of state variables, and t is a

where yt is a vector of control variables, xt is a vector of state variables, and t is a vector of

. dsge (p = {beta}F.p + {kappa}x)

. dsge (c = F.c - (1-{beta}+{beta}{delta})F.r, unobserved)