Title stata.

com
irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs

Description Quick start Menu Syntax Options

Remarks and examples Methods and formulas References Also see

Description
irf create estimates multiple sets of impulse–response functions (IRFs), dynamic-multiplier
functions, and forecast-error variance decompositions (FEVDs). All of these estimates and their
standard errors are known collectively as IRF results and are saved in an IRF file under a specified
filename. Once you have created a set of IRF results, you can use the other irf commands to analyze
them.

Quick start
Create impulse–response function myirf with 8 forecast periods in the active IRF file
irf create myirf
Same as above, and use IRF file myirfs.irf
irf create myirf, set(myirfs)
Same as above, but compute the IRF for 12 periods
irf create myirf, set(myirfs) step(12)

Note: irf commands can be used after var, svar, vec, arima, arfima, dsge, or dsgenl; see
[TS] var, [TS] var svar, [TS] vec, [TS] arima, [TS] arfima, [DSGE] dsge, or [DSGE] dsgenl.

Menu
Statistics > Postestimation

1
 2 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs

Syntax
After var
 
irf create irfname , var options

After svar
 
irf create irfname , svar options

After vec, dsge, or dsgenl

 
irf create irfname , vec options

After arima
 
irf create irfname , arima options

After arfima
 
irf create irfname , arfima options

irfname is any valid name that does not exceed 15 characters.

var options Description

Main
 
set(filename , replace ) make filename active
replace replace irfname if it already exists
step(#) set forecast horizon to #; default is step(8)
order(varlist) specify Cholesky ordering of endogenous variables
estimates(estname) use previously stored results estname; default is to use active
results
Std. errors
nose do not calculate standard errors
bs obtain standard errors from bootstrapped residuals
bsp obtain standard errors from parametric bootstrap
nodots do not display “.” for each bootstrap replication
reps(#)   use # bootstrap replications; default is reps(200)
bsaving(filename , replace ) save bootstrap results in filename
 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs 3

svar options Description

Main
 
set(filename , replace ) make filename active
replace replace irfname if it already exists
step(#) set forecast horizon to #; default is step(8)
estimates(estname) use previously stored results estname; default is to use active
results
Std. errors
nose do not calculate standard errors
bs obtain standard errors from bootstrapped residual
bsp obtain standard errors from parametric bootstrap
nodots do not display “.” for each bootstrap replication
reps(#)   use # bootstrap replications; default is reps(200)
bsaving(filename , replace ) save bootstrap results in filename

vec options Description

Main
 
set(filename , replace ) make filename active
replace replace irfname if it already exists
step(#) set forecast horizon to #; default is step(8)
estimates(estname) use previously stored results estname; default is to use active
results

arima options Description

Main
 
set(filename , replace ) make filename active
replace replace irfname if it already exists
step(#) set forecast horizon to #; default is step(8)
estimates(estname) use previously stored results estname; default is to use active
results
Std. errors
nose do not calculate standard errors

arfima options Description

Main
 
set(filename , replace ) make filename active
replace replace irfname if it already exists
step(#) set forecast horizon to #; default is step(8)
smemory calculate short-memory IRFs
estimates(estname) use previously stored results estname; default is to use active
results
Std. errors
nose do not calculate standard errors
 4 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs

The default is to use asymptotic standard errors if no options are specified.

irf create is for use after fitting a model with the var, svar, vec, arima, arfima, dsge, or dsgenl command;
see [TS] var, [TS] var svar, [TS] vec, [TS] arima, [TS] arfima, [DSGE] dsge, or [DSGE] dsgenl.
You must tsset your data before using var, svar, vec, arima, arfima, dsge, or dsgenl and, hence, before using
irf create; see [TS] tsset.

Options

 Main
set(filename[, replace]) specifies the IRF file to be used. If set() is not specified, the active IRF
file is used; see [TS] irf set.
If set() is specified, the specified file becomes the active file, just as if you had issued an irf
set command.
replace specifies that the results saved under irfname may be replaced, if they already exist. IRF
results are saved in files, and one file may contain multiple IRF results.
step(#) specifies the step (forecast) horizon; the default is eight periods.
order(varlist) is allowed only after estimation by var; it specifies the Cholesky ordering of the
endogenous variables to be used when estimating the orthogonalized IRFs. By default, the order
in which the variables were originally specified on the var command is used.
smemory is allowed only after estimation by arfima; it specifies that the IRFs are calculated based
on a short-memory model with the fractional difference parameter d set to zero.
estimates(estname) specifies that estimation results previously estimated by var, svar, or vec,
and stored by estimates, be used. This option is rarely specified; see [R] estimates.

 Std. errors
nose, bs, and bsp are alternatives that specify how (whether) standard errors are to be calculated. If
none of these options is specified, asymptotic standard errors are calculated, except in two cases:
after estimation by vec and after estimation by svar in which long-run constraints were applied.
In those two cases, the default is as if nose were specified, although in the second case, you could
specify bs or bsp. After estimation by vec, standard errors are simply not available.
nose specifies that no standard errors be calculated.
bs specifies that standard errors be calculated by bootstrapping the residuals. bs may not be
specified if there are gaps in the data.
bsp specifies that standard errors be calculated via a multivariate-normal parametric bootstrap.
bsp may not be specified if there are gaps in the data.
 
nodots, reps(#), and bsaving(filename , replace ) are relevant only if bs or bsp is specified.
nodots specifies that dots not be displayed each time irf create performs a bootstrap replication.
reps(#), # > 50, specifies the number of bootstrap replications to be performed. reps(200) is
the default.
 
bsaving(filename , replace ) specifies that file filename be created and that the bootstrap
replications be saved in it. New file filename is just a .dta dataset that can be loaded later
using use; see [D] use. If filename is specified without an extension, .dta is assumed.
 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs 5

Remarks and examples stata.com

If you have not read [TS] irf, please do so. An introductory example using IRFs is presented there.
irf create estimates several types of IRFs, dynamic-multiplier functions, and FEVDs. Which
estimates are saved depends on the estimation method previously used to fit the model, as summarized
in the table below:

Estimation command
dsge/
Saves arima arfima var svar vec dsgenl
simple IRFs x x x x x x
orthogonalized IRFs x x x x x
dynamic multipliers x
cumulative IRFs x x x x x
cumulative orthogonalized IRFs x x x x x
cumulative dynamic multipliers x
structural IRFs x x x
Cholesky FEVDs x x x
structural FEVDs x

Remarks are presented under the following headings:

Introductory examples
Technical aspects of IRF files
IRFs and FEVDs
IRF results for VARs
An introduction to impulse–response functions for VARs
An introduction to dynamic-multiplier functions for VARs
An introduction to forecast-error variance decompositions for VARs
IRF results for VECMs
An introduction to impulse–response functions for VECMs
An introduction to forecast-error variance decompositions for VECMs
IRF results for ARIMA and ARFIMA

Introductory examples

Example 1: After var

Below we compare bootstrap and asymptotic standard errors for a specific FEVD. We begin by
fitting a VAR(2) model to the Lütkepohl data (we use the var command). We next use the irf create
command twice, first to create results with asymptotic standard errors (saved under the name asymp)
and then to re-create the same results, this time with bootstrap standard errors (saved under the name
bs). Because bootstrapping is a random process, we set the random-number seed (set seed 123456)
before using irf create the second time; this makes our results reproducible. Finally, we compare
results by using the IRF analysis command irf ctable.
 6 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs

. use https://www.stata-press.com/data/r18/lutkepohl2
(Quarterly SA West German macro data, Bil DM, from Lutkepohl 1993 Table E.1)
. var dln_inv dln_inc dln_consump if qtr>=tq(1961q2) & qtr<=tq(1978q4), lags(1/2)
(output omitted )
. irf create asymp, step(8) set(results1)
(file results1.irf created)
(file results1.irf now active)
(file results1.irf updated)
. set seed 123456
. irf create bs, step(8) bs reps(250) nodots
(file results1.irf updated)
. irf ctable (asymp dln_inc dln_consump fevd) (bs dln_inc dln_consump fevd),
> noci stderror

(1) (1) (2) (2)

Step fevd S.E. fevd S.E.

0 0 0 0 0
1 .282135 .087373 .282135 .102756
2 .278777 .083782 .278777 .098161
3 .33855 .090006 .33855 .10586
4 .339942 .089207 .339942 .104191
5 .342813 .090494 .342813 .105351
6 .343119 .090517 .343119 .105258
7 .343079 .090499 .343079 .105266
8 .34315 .090569 .34315 .105303

(1) irfname = asymp, impulse = dln_inc, and response = dln_consump.

(2) irfname = bs, impulse = dln_inc, and response = dln_consump.

Point estimates are, of course, the same. The bootstrap estimates of the standard errors, however,
are larger than the asymptotic estimates, which suggests that the sample size of 71 is not large
enough for the distribution of the estimator of the FEVD to be well approximated by the asymptotic
distribution. Here we would expect the bootstrap confidence interval to be more reliable than the
confidence interval that is based on the asymptotic standard error.

Technical note
The details of the bootstrap algorithms are given in Methods and formulas. These algorithms are
conditional on the first p observations, where p is the order of the fitted VAR. (In an SVAR model, p
is the order of the VAR that underlies the SVAR.) The bootstrapped estimates are conditional on the
first p observations, just as the estimators of the coefficients in VAR models are conditional on the
first p observations. With bootstrap standard errors (option bs), the p initial observations are used
with resampling the residuals to produce the bootstrap samples used for estimation. With the more
parametric bootstrap (option bsp), the p initial observations are used with draws from a multivariate
normal distribution with variance–covariance matrix Σ b to generate the bootstrap samples.

Technical note
For var and svar e() results, irf uses Σ,b the estimated variance matrix of the disturbances, in
computing the asymptotic standard errors of all the functions. The point estimates of the orthogo-
nalized impulse–response functions, the structural impulse–response functions, and all the variance
 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs 7

decompositions also depend on Σ. b As discussed in [TS] var, var and svar use the ML estimator of
this matrix by default, but they have option dfk, which will instead use an estimator that includes a
small-sample correction. Specifying dfk when the model is fit—when the var or svar command is
given—changes the estimate of Σ b and will change the IRF results that depend on it.

Example 2: After var with exogenous variables

After fitting a VAR, irf create computes estimates of the dynamic multipliers, which describe
the impact of a unit change in an exogenous variable on each endogenous variable. For instance,
below we estimate and report the cumulative dynamic multipliers from a model in which changes in
investment are exogenous. The results indicate that both of the cumulative dynamic multipliers are
significant.
. var dln_inc dln_consump if qtr>=tq(1961q2) & qtr<=tq(1978q4), lags(1/2)
> exog(L(0/2).dln_inv)
(output omitted )
. irf create dm, step(8)
(file results1.irf updated)
. irf table cdm, impulse(dln_inv) irf(dm)
Results from dm

(1) (1) (1)

Step cdm Lower Upper

0 .032164 -.027215 .091544

1 .096568 .003479 .189656
2 .140107 .022897 .257317
3 .150527 .032116 .268938
4 .148979 .031939 .26602
5 .151247 .033011 .269482
6 .150267 .033202 .267331
7 .150336 .032858 .267813
8 .150525 .033103 .267948

(2) (2) (2)

Step cdm Lower Upper

0 .058681 .012529 .104832

1 .062723 -.005058 .130504
2 .126167 .032497 .219837
3 .136583 .038691 .234476
4 .146482 .04442 .248543
5 .146075 .045201 .24695
6 .145542 .044988 .246096
7 .146309 .045315 .247304
8 .145786 .045206 .246365

95% lower and upper bounds reported.

(1) irfname = dm, impulse = dln_inv, and response = dln_inc.
(2) irfname = dm, impulse = dln_inv, and response = dln_consump.
 8 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs

Example 3: After vec

Although all IRFs and orthogonalized IRFs (OIRFs) from models with stationary variables will taper
off to zero, some of the IRFs and OIRFs from models with first-difference stationary variables will not.
This is the key difference between IRFs and OIRFs from systems of stationary variables fit by var or
svar and those obtained from systems of first-difference stationary variables fit by vec. When the
effect of the innovations dies out over time, the shocks are said to be transitory. In contrast, when
the effect does not taper off, shocks are said to be permanent.
In this example, we look at the OIRF from one of the VECMs fit to the unemployment-rate data
analyzed in example 2 of [TS] vec. We see that an orthogonalized shock to Indiana has a permanent
effect on the unemployment rate in Missouri:
. use https://www.stata-press.com/data/r18/urates
. vec missouri indiana kentucky illinois, trend(rconstant) rank(2) lags(4)
(output omitted )
. irf create vec1, set(vecirfs) step(50)
(file vecirfs.irf created)
(file vecirfs.irf now active)
(file vecirfs.irf updated)

Now we can use irf graph to graph the OIRF of interest:

. irf graph oirf, impulse(indiana) response(missouri)

vec1, indiana, missouri

.3

.2

.1

0
0 50
Step
Graphs by irfname, impulse variable, and response variable

The graph shows that the estimated OIRF converges to a positive asymptote, which indicates that
an orthogonalized innovation to the unemployment rate in Indiana has a permanent effect on the
unemployment rate in Missouri.

Technical aspects of IRF files

This section is included for programmers wishing to extend the irf system.
irf create estimates a series of impulse–response functions and their standard errors. Although
these estimates are saved in an IRF file, most users will never need to look at the contents of this
file. The IRF commands fill in, analyze, present, and manage IRF results.
 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs 9

IRF files are just Stata datasets that have names ending in .irf instead of .dta. The dataset in
the file has a nested panel structure.
Variable irfname contains the irfname specified by the user. Variable impulse records the name
of the endogenous variable whose innovations are the impulse. Variable response records the name
of the endogenous variable that is responding to the innovations. In a model with K endogenous
variables, there are K 2 combinations of impulse and response. Variable step records the periods
for which these estimates were computed.
Below is a catalog of the statistics that irf create estimates and the variable names under which
they are saved in the IRF file.

Statistic Name
impulse–response functions irf
orthogonalized impulse–response functions oirf
dynamic-multiplier functions dm
cumulative impulse–response functions cirf
cumulative orthogonalized impulse–response functions coirf
cumulative dynamic-multiplier functions cdm
Cholesky forecast-error decomposition fevd
structural impulse–response functions sirf
structural forecast-error decomposition sfevd
standard error of the impulse–response functions stdirf
standard error of the orthogonalized impulse–response functions stdoirf
standard error of the cumulative impulse–response functions stdcirf
standard error of the cumulative orthogonalized impulse–response functions stdcoirf
standard error of the Cholesky forecast-error decomposition stdfevd
standard error of the structural impulse–response functions stdsirf
standard error of the structural forecast-error decomposition stdsfevd
 10 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs

In addition to the variables, information is stored in dta characteristics. Much of the following
information is also available in r() after irf describe, where it is often more convenient to obtain
the information. Characteristic dta[version] contains the version number of the IRF file, which
is currently 1.1. Characteristic dta[irfnames] contains a list of all the irfnames in the IRF file.
For each irfname, there are a series of additional characteristics:

Name Contents
dta[irfname model] var, sr var, lr var, vec, arima, arfima, dsge,
or dsgenl
dta[irfname order] Cholesky order used in IRF estimates
dta[irfname exog] exogenous variables, and their lags, in VAR
dta[irfname exogvars] exogenous variables in VAR
dta[irfname constant] constant or noconstant, depending on whether
noconstant was specified in var or svar
dta[irfname lags] lags in model
dta[irfname exlags] lags of exogenous variables in model
dta[irfname tmin] minimum value of timevar in the estimation sample
dta[irfname tmax] maximum value of timevar in the estimation sample
dta[irfname timevar] name of tsset timevar
dta[irfname tsfmt] format of timevar
dta[irfname varcns] constrained or colon-separated list of
constraints placed on VAR coefficients
dta[irfname svarcns] constrained or colon-separated list of
constraints placed on VAR coefficients
dta[irfname step] maximum step in IRF estimates
dta[irfname stderror] asymptotic, bs, bsp, or none,
depending on the type of standard errors requested
dta[irfname reps] number of bootstrap replications performed
dta[irfname version] version of the IRF file that originally
held irfname IRF results
dta[irfname rank] number of cointegrating equations
dta[irfname trend] trend() specified in vec
dta[irfname veccns] constraints placed on VECM parameters
dta[irfname sind] normalized seasonal indicators included in vec
dta[irfname d] fractional difference parameter d in arfima

IRFs and FEVDs

irf create can estimate several types of IRFs and FEVDs for VARs and VECMs. irf create can
also estimate IRFs and cumulative IRFs for ARIMA and ARFIMA models. We first discuss IRF results for
VAR and SVAR models, and then we discuss them in the context of VECMs. Because the cointegrating
VECM is an extension of the stationary VAR framework, the section that discusses the IRF results for
VECMs draws on the earlier VAR material. We conclude our discussion with IRF results for ARIMA
and ARFIMA models.
 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs 11

IRF results for VARs

An introduction to impulse–response functions for VARs

A pth-order vector autoregressive model (VAR) with exogenous variables is given by
yt = v + A1 yt−1 + · · · + Ap yt−p + Bxt + ut
where

yt = (y1t , . . . , yKt )0 is a K × 1 random vector,

the Ai are fixed K × K matrices of parameters,
xt is an R0 × 1 vector of exogenous variables,
B is a K × R0 matrix of coefficients,
v is a K × 1 vector of fixed parameters, and
ut is assumed to be white noise; that is,
E(ut ) = 0
E(ut u0t ) = Σ
E(ut u0s ) = 0 for t 6= s

As discussed in [TS] varstable, a VAR can be rewritten in moving-average form only if it is stable.
Any exogenous variables are assumed to be covariance stationary. Because the functions of interest
in this section depend only on the exogenous variables through their effect on the estimated Ai , we
can simplify the notation by dropping them from the analysis. All the formulas given below still
apply, although the Ai are estimated jointly with B on the exogenous variables.
Below we discuss conditions under which the IRFs and forecast-error variance decompositions have a
causal interpretation. Although estimation requires only that the exogenous variables be predetermined,
that is, that E(xjt uit ) = 0 for all i, j , and t, assigning a causal interpretation to IRFs and FEVDs
requires that the exogenous variables be strictly exogenous, that is, that E(xjs uit ) = 0 for all i, j ,
s, and t.
IRFs describe how the innovations to one variable affect another variable after a given number of
periods. For an example of how IRFs are interpreted, see Stock and Watson (2001). They use IRFs to
investigate the effect of surprise shocks to the Federal Funds rate on inflation and unemployment. In
another example, Christiano, Eichenbaum, and Evans (1999) use IRFs to investigate how shocks to
monetary policy affect other macroeconomic variables.
Consider a VAR without exogenous variables:
yt = v + A1 yt−1 + · · · + Ap yt−p + ut (1)
The VAR represents the variables in yt as functions of its own lags and serially uncorrelated innovations
ut . All the information about contemporaneous correlations among the K variables in yt is contained
in Σ. In fact, as discussed in [TS] var svar, a VAR can be viewed as the reduced form of a dynamic
simultaneous-equation model.
To see how the innovations affect the variables in yt after, say, i periods, rewrite the model in its
moving-average form
X∞
yt = µ + Φi ut−i (2)
i=0
where µ is the K × 1 time-invariant mean of yt , and

IK if i = 0
Φ i = Pi
j=1 Φi−j Aj if i = 1, 2, . . .
12 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs

We can rewrite a VAR in the moving-average form only if it is stable. Essentially, a VAR is stable
if the variables are covariance stationary and none of the autocorrelations are too high (the issue of
stability is discussed in greater detail in [TS] varstable).
The Φi are the simple IRFs. The j, k element of Φi gives the effect of a 1–time unit increase in
the k th element of ut on the j th element of yt after i periods, holding everything else constant.
Unfortunately, these effects have no causal interpretation, which would require us to be able to answer
the question, “How does an innovation to variable k , holding everything else constant, affect variable j
after i periods?” Because the ut are contemporaneously correlated, we cannot assume that everything
else is held constant. Contemporaneous correlation among the ut implies that a shock to one variable
is likely to be accompanied by shocks to some of the other variables, so it does not make sense to
shock one variable and hold everything else constant. For this reason, (2) cannot provide a causal
interpretation.
This shortcoming may be overcome by rewriting (2) in terms of mutually uncorrelated innovations.
Suppose that we had a matrix P, such that Σ = PP0 . If we had such a P, then P−1 ΣP0−1 = IK ,
and
E{P−1 ut (P−1 ut )0 } = P−1 E{(ut u0t )P0−1 } = P−1 ΣP0−1 = IK
We can thus use P−1 to orthogonalize the ut and rewrite (2) as
∞
X
yt = µ + Φi PP−1 ut−i
i=0
∞
X
=µ+ Θi P−1 ut−i
i=0
∞
X
=µ+ Θi wt−i
i=0

where Θi = Φi P and wt = P−1 ut . If we had such a P, the wk would be mutually orthogonal,

and no information would be lost in the holding-everything-else-constant assumption, implying that
the Θi would have the causal interpretation that we seek.
Choosing a P is similar to placing identification restrictions on a system of dynamic simultaneous
equations. The simple IRFs do not identify the causal relationships that we wish to analyze. Thus we
seek at least as many identification restrictions as necessary to identify the causal IRFs.
So, where do we get such a P? Sims (1980) popularized the method of choosing P to be the
Cholesky decomposition of Σ. b The IRFs based on this choice of P are known as the orthogonalized
IRFs. Choosing P to be the Cholesky decomposition of Σ b is equivalent to imposing a recursive
structure for the corresponding dynamic structural equation model. The ordering of the recursive
structure is the same as the ordering imposed in the Cholesky decomposition. Because this choice is
arbitrary, some researchers will look at the OIRFs with different orderings assumed in the Cholesky
decomposition. The order() option available with irf create facilitates this type of analysis.
The SVAR approach integrates the need to identify the causal IRFs into the model specification and
estimation process. Sufficient identification restrictions can be obtained by placing either short-run or
long-run restrictions on the model. The VAR in (1) can be rewritten as

yt − v − A1 yt−1 − · · · − Ap yt−p = ut
Similarly, a short-run SVAR model can be written as

A(yt − v − A1 yt−1 − · · · − Ap yt−p ) = Aut = Bet (3)

 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs 13

where A and B are K × K nonsingular matrices of parameters to be estimated, et is a K × 1 vector

of disturbances with et ∼ N (0, IK ), and E(et e0s ) = 0K for all s 6= t. Sufficient constraints must
be placed on A and B so that P is identified. One way to see the connection is to draw out the
implications of the latter equality in (3). From (3), it can be shown that

Σ = A−1 B(A−1 B)0

As discussed in [TS] var svar, the estimates Ab and Bb are obtained by maximizing the concentrated
log-likelihood function on the basis of the Σ b obtained from the underlying VAR. The short-run
SVAR approach chooses P = A b −1 B
b to identify the causal IRFs. The long-run SVAR approach works
similarly, with P = C b =A b −1 B
b , where Ab −1 is the matrix of estimated long-run or accumulated
effects of the reduced-form VAR shocks.
There is one important difference between long-run and short-run SVAR models. As discussed by
Amisano and Giannini (1997, chap. 6), in the short-run model the constraints are applied directly to
the parameters in A and B. Then A and B interact with the estimated parameters of the underlying
VAR. In contrast, in a long-run model, the constraints are placed on functions of the estimated VAR
parameters. Although estimation and inference of the parameters in C is straightforward, obtaining
the asymptotic standard errors of the structural IRFs requires untenable assumptions. For this reason,
irf create does not estimate the asymptotic standard errors of the structural IRFs generated by
long-run SVAR models. However, bootstrap standard errors are still available.

An introduction to dynamic-multiplier functions for VARs

A dynamic-multiplier function measures the effect of a unit change in an exogenous variable on the
endogenous variables over time. Per Lütkepohl (2005, chap. 10), if the VAR with exogenous variables
is stable, it can be rewritten as

∞
X ∞
X
yt = Di xt−i + Φi ut−i
i=0 i=0

where the Di are the dynamic-multiplier functions. (See Methods and formulas for details.) Some
authors refer to the dynamic-multiplier functions as transfer functions because they specify how a
unit change in an exogenous variable is “transferred” to the endogenous variables.

Technical note
irf create computes dynamic-multiplier functions only after var. After short-run SVAR models,
the dynamic multipliers from the VAR are the same as those from the SVAR. The dynamic multipliers
for long-run SVARs have not yet been worked out.

An introduction to forecast-error variance decompositions for VARs

Another measure of the effect of the innovations in variable k on variable j is the FEVD. This
method, which is also known as innovation accounting, measures the fraction of the error in forecasting
variable j after h periods that is attributable to the orthogonalized innovations in variable k . Because
deriving the FEVD requires orthogonalizing the ut innovations, the FEVD is always predicated upon
a choice of P.
 14 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs

Lütkepohl (2005, sec. 2.2.2) shows that the h-step forecast error can be written as

h−1
X
yt+h − y
bt (h) = Φi ut+h−i (4)
i=0

where yt+h is the value observed at time t + h and ybt (h) is the h-step-ahead predicted value for
yt+h that was made at time t.
Because the ut are contemporaneously correlated, their distinct contributions to the forecast error
cannot be ascertained. However, if we choose a P such that Σ = PP0 , as above, we can orthogonalize
the ut into wt = P−1 ut . We can then ascertain the relative contribution of the distinct elements of
wt . Thus we can rewrite (4) as

h−1
X
yt+h − y
bt (h) = Φi PP−1 ut+h−i
i=0
h−1
X
= Θi wt+h−i
i=0

Because the forecast errors can be written in terms of the orthogonalized errors, the forecast-
error variance can be written in terms of the orthogonalized error variances. Forecast-error variance
decompositions measure the fraction of the total forecast-error variance that is attributable to each
orthogonalized shock.

Technical note
The details in this note are not critical to the discussion that follows. A forecast-error variance
decomposition is derived for a given P. Per Lütkepohl (2005, sec. 2.3.3), letting θmn,i be the m, nth
element of Θi , we can express the h-step forecast error of the j th component of yt as

h−1
X
yj,t+h − y
bj (h) = θj1,1 w1,t+h−i + · · · + θjK,i wK,t+h−i
i=0
K
X
= θjk,0 wk,t+h + · · · + θjk,h−1 wk,t+1
k=1

The wt , which were constructed using P, are mutually orthogonal with unit variance. This allows
us to compute easily the mean squared error (MSE) of the forecast of variable j at horizon h in terms
of the contributions of the components of wt . Specifically,

K
X
2 2 2
E[{yj,t+h − yj,t (h)} ] = (θjk,0 + · · · + θjk,h−1 )
k=1

The k th term in the sum above is interpreted as the contribution of the orthogonalized innovations
in variable k to the h-step forecast error of variable j . Note that the k th element in the sum above
can be rewritten as
h−1
X
2
2
(θjk,0 2
+ · · · + θjk,h−1 )= e0j Θk ek
i=0
 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs 15

where ei is the ith column of IK . Normalizing by the forecast error for variable j at horizon h yields
Ph−1
2
i=0 e0j Θk ek
ωjk,h =
MSE{yj,t (h)}

Ph−1 PK 2
where MSE{yj,t (h)} = i=0 k=1 θjk,i .

Because the FEVD depends on the choice of P, there are different forecast-error variance de-
compositions associated with each distinct P. irf create can estimate the FEVD for a VAR or an
SVAR. For a VAR, P is the Cholesky decomposition of Σ. b For an SVAR, P is the estimated structural
−1 b
decomposition, P = A B for short-run models and P = C
b b for long-run SVAR models. Due to the
same complications that arose with the structural impulse–response functions, the asymptotic standard
errors of the structural FEVD are not available after long-run SVAR models, but bootstrap standard
errors are still available.

IRF results for VECMs

An introduction to impulse–response functions for VECMs

As discussed in [TS] vec intro, the VECM is a reparameterization of the VAR that is especially
useful for fitting VARs with cointegrating variables. This implies that the estimated parameters for
the corresponding VAR model can be backed out from the estimated parameters of the VECM model.
This relationship means we can use the VAR form of the cointegrating VECM to discuss the IRFs for
VECMs.
Consider a cointegrating VAR with one lag with no constant or trend,

yt = Ayt−1 + ut (5)

where yt is a K × 1 vector of endogenous, first-difference stationary variables among which there

are 1 ≤ r < K cointegration equations; A is K × K matrix of parameters; and ut is a K × 1 vector
of i.i.d. disturbances.
We developed intuition for the IRFs from a stationary VAR by rewriting the VAR as an infinite-
order vector moving-average (VMA) process. While the Granger representation theorem establishes
the existence of a VMA formulation of this model, because the cointegrating VAR is not stable, the
inversion is not nearly so intuitive. (See Johansen [1995, chapters 3 and 4] for more details.) For this
reason, we use (5) to develop intuition for the IRFs from a cointegrating VAR.
Suppose that K is 3, that u1 = (1, 0, 0), and that we want to analyze the time paths of the
variables in y conditional on the initial values y0 = 0, A, and the condition that there are no more
shocks to the system, that is, 0 = u2 = u3 = · · · . These assumptions and (5) imply that

y1 = u1

y2 = Ay1 = Au1
y3 = Ay2 = A2 u1
 16 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs

and so on. The ith-row element of the first column of As contains the effect of the unit shock to the
first variable after s periods. The first column of As contains the IRF of a unit impulse to the first
variable after s periods. We could deduce the IRFs of a unit impulse to any of the other variables by
administering the unit shock to one of them instead of to the first variable. Thus we can see that the
(i, j)th element of As contains the unit IRF from variable j to variable i after s periods. By starting
with orthogonalized shocks of the form P−1 ut , we can use the same logic to derive the OIRFs to be
As P.
For the stationary VAR, stability implies that all the eigenvalues of A have moduli strictly less than
one, which in turn implies that all the elements of As → 0 as s → ∞. This implies that all the
IRFs from a stationary VAR taper off to zero as s → ∞. In contrast, in a cointegrating VAR, some of
the eigenvalues of A are 1, while the remaining eigenvalues have moduli strictly less than 1. This
implies that in cointegrating VARs some of the elements of As are not going to zero as s → ∞,
which in turn implies that some of the IRFs and OIRFs are not going to zero as s → ∞. The fact that
the IRFs and OIRFs taper off to zero for stationary VARs but not for cointegrating VARs is one of the
key differences between the two models.
When the IRF or OIRF from the innovation in one variable to another tapers off to zero as time
goes on, the innovation to the first variable is said to have a transitory effect on the second variable.
When the IRF or OIRF does not go to zero, the effect is said to be permanent.
Note that, because some of the IRFs and OIRFs do not taper off to zero, some of the cumulative
IRFs and OIRFs diverge over time.

An introduction to forecast-error variance decompositions for VECMs

The results from An introduction to impulse–response functions for VECMs can be used to show
that the interpretation of FEVDs for a finite number of steps in cointegrating VARs is essentially the
same as in the stationary case. Because the MSE of the forecast is diverging, this interpretation is valid
only for a finite number of steps. (See [TS] vec intro and [TS] fcast compute for more information
on this point.)

IRF results for ARIMA and ARFIMA

A covariance-stationary additive ARMA(p, q) model can be written as

ρ(Lp )(yt − xt β) = θ(Lq )t

where
ρ(Lp ) = 1 − ρ1 L − ρ2 L2 − · · · − ρp Lp
θ(Lq ) = 1 + θ1 L + θ2 L2 + · · · + θq Lq

and Lj yt = yt−j .
We can rewrite the above model as an infinite-order moving-average process

yt = xt β + ψ(L)t

where
θ(L)
ψ(L) = = 1 + ψ1 L + ψ2 L2 + · · · (6)
ρ(L)
 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs 17

This representation shows the impact of the past innovations on the current yt . The ith coefficient
describes the response of yt to a one-time impulse in t−i , holding everything else constant. The ψi
coefficients are collectively referred to as the impulse–response function of the ARMA model. For a
covariance-stationary series, the ψi coefficients decay exponentially.
A covariance-stationary multiplicative seasonal ARMA model, often abbreviated SARMA, of order
(p, q) × (P, Q)s can be written as
ρ(Lp )ρs (LP )(yt − xt β) = θ(Lq )θs (LQ )t
where
ρs (LP ) = (1 − ρs,1 Ls − ρs,2 L2s − · · · − ρs,P LP s )
θs (LQ ) = (1 + θs,1 Ls + θs,2 L2s + · · · + θs,Q LQs )
with ρ(Lp ) and θ(Lq ) defined as above.
We can express this model as an additive ARMA model by multiplying the terms and imposing
nonlinear constraints on multiplied coefficients. For example, consider the SARMA model given by

(1 − ρ1 L)(1 − ρ4,1 L4 )yt = t

Expanding the above equation and solving for yt yields

yt = ρ1 yt−1 + ρ4,1 yt−4 − ρ1 ρ4,1 yt−5 + t

or, in ARMA terms,

yt = ρ1 yt−1 + ρ4 yt−4 + ρ5 yt−5 + t

subject to the constraint ρ5 = −ρ1 ρ4,1 .

Once we have obtained an ARMA representation of a SARMA process, we obtain the IRFs from (6).
An ARFIMA(p, d, q) model can be written as

ρ(Lp )(1 − L)d (yt − xt β) = θ(Lq )t

with (1 − L)d denoting a fractional integration operation.
Solving for yt , we obtain
yt = xt β + (1 − L)−d ψ(L)t

This makes it clear that the impulse–response function for an ARFIMA model corresponds to a
fractionally differenced impulse–response function for an ARIMA model. Because of the fractional
differentiation, the ψi coefficients decay very slowly; see Remarks and examples in [TS] arfima.

Methods and formulas

Methods and formulas are presented under the following headings:
Impulse–response function formulas for VARs
Dynamic-multiplier function formulas for VARs
Forecast-error variance decomposition formulas for VARs
Impulse–response function formulas for VECMs
Algorithms for bootstrapping the VAR IRF and FEVD standard errors
Impulse–response function formulas for ARIMA and ARFIMA
 18 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs

Impulse–response function formulas for VARs

The previous discussion implies that there are three different choices of P that can be used to
obtain distinct Θi . P is the Cholesky decomposition of Σ for the OIRFs. For the structural IRFs,
P = A−1 B for short-run models, and P = C for long-run models. We will distinguish between
the three by defining Θoi to be the OIRFs, Θsr lr
i to be the short-run structural IRFs, and Θi to be the
long-run structural IRFs.
We also define Pb c to be the Cholesky decomposition of Σ, b Pb sr = Ab −1 B
b to be the short-run
structural decomposition, and P
b lr = C
b to be the long-run structural decomposition.
Given estimates of the A
b i and Σ
b from var or svar, the estimates of the simple IRFs and the
OIRFs are, respectively,
i
X
bi =
Φ b i−j A
Φ bj
j=1

and
b oi = Φ
Θ b iP
bc

where A
b j = 0K for j > p.

Given the estimates A

b and B
b , or C
b , from svar, the estimates of the structural IRFs are either

b sr
Θ i = Φi Psr
b b

or
b lr
Θ i = Φi Plr
b b

The estimated structural IRFs stored in an IRF file with the variable name sirf may be from
either a short-run model or a long-run model, depending on the estimation results used to create the
IRFs. As discussed in [TS] irf describe, you can easily determine whether the structural IRFs were
generated from a short-run or a long-run SVAR model using irf describe.
Following Lütkepohl (2005, sec. 3.7), estimates of the cumulative IRFs and the cumulative
orthogonalized impulse–response functions (COIRFs) at period n are, respectively,
n
X
bn =
Ψ Φ
bi
i=0

and
n
X
bn =
Ξ Θ
bi
i=0

The asymptotic standard errors of the different impulse–response functions are obtained by
applications of the delta method. See Lütkepohl (2005, sec. 3.7) and Amisano and Giannini (1997,
chap. 4) for the derivations. See Serfling (1980, sec. 3.3) for a discussion of the delta method. In
presenting the variance–covariance matrix estimators, we make extensive use of the vec() operator,
where vec(X) is the vector obtained by stacking the columns of X.
Lütkepohl (2005, sec. 3.7) derives the asymptotic VCEs of vec(Φi ), vec(Θoi ), vec(Ψ b n ), and
2 2 2
b n ). Because vec(Φi ) is K × 1, the asymptotic VCE of vec(Φi ) is K × K , and it is given by
vec(Ξ
b G0i
Gi Σαb
 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs 19

where Pi−1
Gi = c 0 )(i−1−m) ⊗ Φ
J(M bm Gi is K 2 ×K 2 p
m=0

J = (IK , 0K , . . . , 0K ) J is K×Kp

A1 A b2 ... A b p−1 A
b bp 
 IK 0K . . . 0K 0K 
0 I 0K 0K 
 
M=
c
 .K K . b is Kp×Kp
M
 . .. .. .. 

. . .
0K 0K . . . IK 0K

The Ab i are the estimates of the coefficients on the lagged variables in the VAR, and Σ
b is the VCE
α
b
2 2
b is a K p × K p matrix whose elements come from the VCE
matrix of αb = vec(A b 1, . . . , A
b p ). Σ
α
b
of the VAR coefficient estimator. As such, this VCE is the VCE of the constrained estimator if there
are any constraints placed on the VAR coefficients.
The K 2 × K 2 asymptotic VCE matrix for vec(Ψ
b n ) after n periods is given by

b F0n
Fn Σ α
b
where
n
X
Fn = Gi
i=1

The K 2 × K 2 asymptotic VCE matrix of the vectorized, orthogonalized, IRFs at horizon i, vec(Θoi ),
is
b C0i
b C0i + Ci Σ
Ci Σ α
b bσ
20 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs

where

C0 = 0 C0 is K 2 ×K 2 p

b 0 ⊗ IK )Gi , i = 1, 2, . . .
Ci = (P Ci is K 2 ×K 2 p
c

Ci = (IK ⊗ Φi )H, i = 0, 1, . . . Ci is K 2 ×K 2
n o−1
H = L0K LK NK (P b c ⊗ IK )L0
K H is K 2 ×K
(K+1)
2

(K+1)
LK solves vech(F) = LK vec(F) LK is K 2 ×K 2

for F K × K and symmetric

KK solves KK vec(G) = vec(G0 ) for any K × K matrix G KK is K 2 ×K 2

1
NK = 2 (IK 2 + KK ) NK is K 2 ×K 2

Σb = 2D+ (Σb ⊗Σb )D+ Σ

b is K
(K+1) (K+1)
×K 2
σ
b K K bσ 2

−1
D+K = (D0
D
K K ) D 0
K D+
K
is K (K+1)
2 ×K 2

DK solves DK vech(F) = vec(F) for F K × K and symmetric DK is K 2 ×K

(K+1)
2

x11
 
 x21 
 . 
 . 
 . 
 xK1 
 
x 
 
vech(X) = 
 22 for X K ×K vech(X) is K
(K+1)
2 ×1
 .. 

 . 
 xK2 
 
 . 
 .. 
xKK

Note that Σ
b is the VCE of vech(Σ).b More details about LK , KK , DK and vech() are available in
σ
b
Lütkepohl (2005, sec. A.12). Finally, as Lütkepohl (2005, 113–114) discusses, D+
K is the Moore–
Penrose inverse of DK .
As discussed in Amisano and Giannini (1997, chap. 6), the asymptotic standard errors of the
structural IRFs are available for short-run SVAR models but not for long-run SVAR models. Following
Amisano and Giannini (1997, chap. 5), the asymptotic K 2 × K 2 VCE of the short-run structural IRFs
after i periods, when a maximum of h periods are estimated, is the i, i block of
n o n o0
b (h)ij = G
Σ b G
e iΣ e 0 + IK ⊗ (JM
c i J0 ) Σ(0) IK ⊗ (JM
c j J0 )
αb j
 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs 21

where
G
e 0 = 0K G0 is K 2 ×K 2 p
n  o
e i = Pi−1 P
G b 0 J(M
c 0 )i−1−k ⊗ JM
c k J0 Gi is K 2 ×K 2 p
k=0 sr
b W Q02
b (0) = Q2 Σ
Σ Σ
b (0) is K 2 ×K 2

b AB Q01
b W = Q1 Σ
Σ Σ
b W is K 2 ×K 2

b0 ⊗ P
Q2 = P b sr Q2 is K 2 ×K 2
n sr o
Q1 = (IK ⊗ Bb −1 ), (−P
b 0−1 ⊗ B−1 ) Q1 is K 2 ×2K 2
sr

b AB is the 2K 2 × 2K 2 VCE of the estimator of vec(A, B).

and Σ

Dynamic-multiplier function formulas for VARs

This section provides the details of how irf create estimates the dynamic-multiplier functions
and their asymptotic standard errors.
A pth order vector autoregressive model (VAR) with exogenous variables may be written as

yt = v + A1 yt−1 + · · · + Ap yt−p + B0 xt + B1 xt−1 + · · · + Bs xt−s + ut

where all the notation is the same as above except that the s K × R matrices B1 , B2 , . . . , Bs are
explicitly included and s is the number of lags of the R exogenous variables in the model.
Lütkepohl (2005) shows that the dynamic-multipliers Di are consistently estimated by

D ei B
b i = Jx A i ∈ {0, 1, . . .}
x x
b

where
Jx = (IK , 0K , . . . , 0K ) J is K×(Kp+Rs)
 
M
c B b
Ax = e e
e e x is (Kp+Rs)×(Kp+Rs)
A
0 I
b bs 
B1 B b2 ... B
b =  0̈. 0̈ . . . 0̈ 

B  .. .. .. ..  b is Kp×Rs
B
. . .

0̈ 0̈ . . . 0̈
0R 0R . . . 0R 0R
 
 IR 0R . . . 0R 0R 
eI =  0R IR 0R 0R 
 
 . eI is Rs×Rs
 .. .. .. .. 
. . .

0R 0R . . . IR 0R
0
 
Bx = B0 Ï0
b e b 0x is R×(Kp+Rs)
B

e0 = B
 
B b 0 0̈0 · · · 0̈0 e is R×Kp
B
0

Ï0 = [ IR 0R · · · 0R ] Ï is R×Rs

and 0̈ is a K × R matrix of 0s and 0

e is a Rs × Kp matrix of 0s.
 22 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs

Consistent estimators of the cumulative dynamic-multiplier functions are given by

i
X
Di = D
bj
j=0

Letting βx = vec(A1 A2 · · · Ap B1 B2 · · · Bs B0 ) and letting Σ b be the asymptotic variance–

βx
covariance estimator (VCE) of βx , Lütkepohl shows that an asymptotic VCE of D
b b i is G
e iΣ G e0
bx i
β
where !
i−1
X
0 i−1−j j 0 j
G
ei = Bx Ae
x ⊗ Jx A
e J , IR ⊗ Jx A
x x
e Jx
x
j=0
P  P 
i i e0 .
Similarly, an asymptotic VCE of Di is j=0 G
ej Σ
j=0 Gj
β
bx

Forecast-error variance decomposition formulas for VARs

This section provides details of how irf create estimates the Cholesky FEVD, the structural
FEVD, and their standard errors. Beginning with the Cholesky-based forecast-error decompositions,
the fraction of the h-step-ahead forecast-error variance of variable j that is attributable to the Cholesky
orthogonalized innovations in variable k can be estimated as
Ph−1 0
b i ek )2
(ej Θ
ωbjk,h = i=0
MSE
d j (h)

where MSEj (h) is the j th diagonal element of

h−1
b 0i
X
Φ
b iΣ
bΦ
i=0

(See Lütkepohl [2005, 109] for a discussion of this result.) ω

bjk,h and MSEj (h) are scalars. The square
of the standard error of ω
bjk,h is
b α d0jk,h + djk,h Σ
djk,h Σ b σ djk,h

where
(
Ph−1
djk,h = MSE2 (h)2 i=0 MSEj (h)(e0j Φ b c ek )(e0 P
b iP b0
k c ⊗ e0j )Gi
j
)
b c ek )2 Ph−1 (e0 Φ
−(e0j Φi P 0
m=0 j m Σ ⊗ ej )Gm djk,h is 1×K 2 p
b b
(
Ph−1
djk,h = i=0 MSEj (h)(e0j Φ b i Pc ek )(e0 ⊗ e0j Φ
k
b i )H
)
Ph−1
−(e0j Φ b c ek )2
b iP 0b
m=0 (ej Φm ⊗ ej Φ
b m )DK 1
MSEj (h)2 djk,h is 1×K
(K+1)
2

G0 = 0 G0 is K 2 ×K 2 p

and DK is the K 2 × K{(K + 1)/2} duplication matrix defined previously.

 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs 23

For the structural forecast-error decompositions, we follow Amisano and Giannini (1997, sec. 5.2).
They define the matrix of structural forecast-error decompositions at horizon s, when a maximum of
h periods are estimated, as

W b −1 M
cs = F c
fs for s = 1, . . . , h + 1
s
s−1
!
b sr b sr0
X
F
bs = Θ i Θi IK
i=0
s−1
X sr
M
fs =
c
Θ
bi b sr
Θ i
i=0

where is the Hadamard, or element-by-element, product.

The K 2 × K 2 asymptotic VCE of vec(W
c s ) is given by

Z e0
e s Σ(h)Z
s

b (h) is as derived previously, and

where Σ
( )
c s ) ∂vec(W
∂vec(W c s) ∂vec(W
c s)
Z sr , · · · ,
es = sr ,
∂vec(Θ
b 0 ) ∂vec(Θ
b1 ) b sr
∂vec(Θ h)

∂vec(W
c s) n
b −1 e b sr c 0 b −1 e b sr
o
sr = 2 (IK ⊗ Fs )D(Θj ) − (Ws ⊗ Fs )D(IK )NK (Θj ⊗ IK )
∂vec(Θ
bj )

If X is an n × n matrix, then D(X)

e is the n2 × n2 matrix with vec(X) on the diagonal and zeros
in all the off-diagonal elements, and NK is as defined previously.

Impulse–response function formulas for VECMs

We begin by providing the formulas for backing out the estimates of the Ai from the Γi estimated
by vec. As discussed in [TS] vec intro, the VAR in (1) can be rewritten as a VECM:

∆yt = v + Πyt−1 + Γ1 ∆yt−1 + Γp−1 ∆yp−2 + t

vec estimates Π and the Γi . Johansen (1995, 25) notes that

p
X
Π= A i − IK (6)
i=1

where IK is the K -dimensional identity matrix, and

p
X
Γi = − Aj (7)
j=i+1
 24 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs

Defining
p−1
X
Γ = IK − Γi
i=1

and using (6) and (7) allow us to solve for the Ai as

A 1 = Π + Γ 1 + IK

Ai = Γi − Γi−1 for i = {2, . . . , p − 1}

and
Ap = −Γp−1

Using these formulas, we can back out estimates of Ai from the estimates of the Γi and Π produced
by vec. Then we simply use the formulas for the IRFs and OIRFs presented in Impulse–response
function formulas for VARs.
The running sums of the IRFs and OIRFs over the steps within each impulse–response pair are the
cumulative IRFs and OIRFs.

Algorithms for bootstrapping the VAR IRF and FEVD standard errors
irf create offers two bootstrap algorithms for estimating the standard errors of the various IRFs
and FEVDs. Both var and svar contain estimators for the coefficients in a VAR that are conditional on
the first p observations. The two bootstrap algorithms are also conditional on the first p observations.
Specifying the bs option calculates the standard errors by bootstrapping the residuals. For a
bootstrap with R repetitions, this method uses the following algorithm:
1. Fit the model and save the estimated parameters.
2. Use the estimated coefficients to calculate the residuals.
3. Repeat steps 3a to 3c R times.
3a. Draw a simple random sample of size T with replacement from the residuals. The
random samples are drawn over the K × 1 vectors of residuals. When the tth vector is
drawn, all K residuals are selected. This preserves the contemporaneous correlations
among the residuals.
3b. Use the p initial observations, the sampled residuals, and the estimated coefficients to
construct a new sample dataset.
3c. Fit the model and calculate the different IRFs and FEVDs.
3d. Save these estimates as observation r in the bootstrapped dataset.
4. For each IRF and FEVD, the estimated standard deviation from the R bootstrapped estimates
is the estimated standard error of that impulse–response function or forecast-error variance
decomposition.
Specifying the bsp option estimates the standard errors by a multivariate normal parametric
bootstrap. The algorithm for the multivariate normal parametric bootstrap is identical to the one
above, with the exception that 3a is replaced by 3a(bsp):
3a(bsp). Draw T pseudovariates from a multivariate normal distribution with covariance matrix
Σ.
b
 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs 25

Impulse–response function formulas for ARIMA and ARFIMA

The previous discussion showed that a SARMA process can be rewritten as an ARMA process and
that for an ARMA process, we can express ψ(L) in terms of θ(L) and ρ(L),

θ(L)
ψ(L) =
ρ(L)
Expanding the above, we obtain

1 + θ1 L + θ2 L2 + · · ·
ψ0 + ψ1 L + ψ2 L2 + · · · =
1 − ρ1 L − ρ2 L2 − · · ·

Given the estimate of the autoregressive terms ρb and the moving-average terms b θ, the IRF is
obtained by solving the above equation for the ψ weights. The ψi are calculated using the recursion
p
X
ψbi = θbi + φbj ψbi−j
j=1

with ψ0 = 1 and θi = 0 for i > max(p, q + 1).

The asymptotic standard errors for the IRF for ARMA are calculated using the delta method;
see Serfling (1980, sec. 3.3) for a discussion of the delta method. Let Σb be the estimate of the
variance–covariance matrix for ρb and θ, and let Ψ be a matrix of derivatives of ψi with respect to
b
ρ
b and bθ. Then the standard errors for ψbi are calculated as

b 0
Ψi ΣΨ i

The IRF for the ARFIMA(p, d, q) model is obtained by applying the filter (1 − L)−d to ψ(L). The
filter is given by Hassler and Kokoszka (2010) as
∞
X
(1 − L)−d = bi Li
i=0

with b0 = 1 and subsequent bi calculated by the recursion

bbi = d + i − 1 bbi−1
b
i

The resulting IRF is then given by

i
X
φbi = ψbjbbi−j
j=0

The asymptotic standard errors for the IRF for ARFIMA are calculated using the delta method. Let
Σ be the estimate of the variance–covariance matrix for ρ
b θ, and db, and let Φ be a matrix of
b, b
derivatives of φi with respect to ρ
b, θ, and d. Then the standard errors for φbi are calculated as
b b

b 0
Φi ΣΦi
 26 irf create — Obtain IRFs, dynamic-multiplier functions, and FEVDs

References
Amisano, G., and C. Giannini. 1997. Topics in Structural VAR Econometrics. 2nd ed, revised and enlarged. Heidelberg:
Springer.
Christiano, L. J., M. Eichenbaum, and C. L. Evans. 1999. Monetary policy shocks: What have we learned and to
what end? In Handbook of Macroeconomics: Volume 1A, ed. J. B. Taylor and M. Woodford. New York: Elsevier.
https://doi.org/10.1016/S1574-0048(99)01005-8.
Hamilton, J. D. 1994. Time Series Analysis. Princeton, NJ: Princeton University Press.
Hassler, U., and P. Kokoszka. 2010. Impulse responses of fractionally integrated processes with long memory.
Econometric Theory 26: 1855–1861. https://doi.org/10.1017/S0266466610000216.
Johansen, S. 1995. Likelihood-Based Inference in Cointegrated Vector Autoregressive Models. Oxford: Oxford University
Press.
Lütkepohl, H. 1993. Introduction to Multiple Time Series Analysis. 2nd ed. New York: Springer.
. 2005. New Introduction to Multiple Time Series Analysis. New York: Springer.
Serfling, R. J. 1980. Approximation Theorems of Mathematical Statistics. New York: Wiley.
Sims, C. A. 1980. Macroeconomics and reality. Econometrica 48: 1–48. https://doi.org/10.2307/1912017.
Stock, J. H., and M. W. Watson. 2001. Vector autoregressions. Journal of Economic Perspectives 15: 101–115.
https://doi.org/10.1257/jep.15.4.101.

Also see
[TS] irf — Create and analyze IRFs, dynamic-multiplier functions, and FEVDs
[TS] var intro — Introduction to vector autoregressive models
[TS] vec intro — Introduction to vector error-correction models

Stata, Stata Press, and Mata are registered trademarks of StataCorp LLC. Stata and
®
Stata Press are registered trademarks with the World Intellectual Property Organization
of the United Nations. Other brand and product names are registered trademarks or
trademarks of their respective companies. Copyright c 1985–2023 StataCorp LLC,
College Station, TX, USA. All rights reserved.