Cointegration Testing and Dynamic Simulations of

Autoregressive Distributed Lag Models*

Soren Jordan†
Andrew Q. Philips‡

September 21, 2017

Abstract
In this paper we discuss the bounds cointegration test proposed by Pesaran, Shin and
Smith (2001), which we have adapted into a Stata program, pssbounds. Since the re-
sulting models can be dynamically complex, we follow the advice of Philips (2017) by
introducing dynardl, a flexible program designed to dynamically simulate and plot a va-
riety of types of autoregressive distributed lag models, including error-correction models.
Word Count: 65 (abstract); approx. 6,300 (manuscript)
Keywords: cointegration; dynamic modeling; autoregressive distributed lag; error-correction

* Source code for the programs discussed in this paper can be found at
http://andyphilips.github.io/pssbounds/ and https://andyphilips.github.io/dynardl/.
† sorenjordanpols@gmail.com. Assistant Professor, Department of Political Science, Auburn Univer-

sity, Auburn, AL 36849.

‡ andrew.philips@colorado.edu. Assistant Professor, Department of Political Science, University of

Colorado Boulder, UCB 333, Boulder, CO 80309.

Introduction

Time series models employing an autoregressive distributed lag (ARDL) are commonplace
in the social sciences. Whether the dependent variable is estimated in levels (e.g., yt =
α0 + · · · ) or in first differences (e.g., ∆yt = α0 + · · · ), these models are able to test a host of
theoretically-important theories, from the effect of public opinion on government response
(Jennings and John 2009) to the effect of domestic and international factors on defense
expenditures (Whitten and Williams 2011) or tax rates (Swank and Steinmo 2002), to
analyzing dynamic changes in partisan responsiveness over time (Ura and Ellis 2008).

When employing an error-correction-style ARDL model, it becomes necessary to test

for cointegration.1 In a recent paper, Philips (2017) shows that in small samples common
in the social sciences—typically, when the number of time points is 80 or less—the ARDL
bounds test for cointegration proposed by Pesaran, Shin and Smith (2001) tends to be
more conservative (i.e., does not conclude cointegration when it does not exist) than
either the Engle-Granger “two-step” (Engle and Granger 1987) or the Johansen (1991,
1995) approaches to cointegration testing. In this paper we present pssbounds, which
provides the non-standard critical values from Pesaran, Shin and Smith (2001) needed to
conduct this cointegration test.

In addition to its error-correction form, ARDL models in general may have complex dy-
namic specifications, including multiple lags, first-differences, and lagged first-differences.
This makes it more difficult to interpret the effects of changes—especially short- and
longer-run changes—in the independent variable. To mitigate this we introduce dynardl,
a flexible program that allows users to dynamically simulate a variety of ARDL models,
including the error-correction model. Dynamic simulations offer an alternative to hypoth-
esis testing of model coefficients by instead conveying the substantive significance of the
results through meaningful counterfactual scenarios. Such an approach has been gaining
popularity in the social sciences (e.g., Tomz, Wittenberg and King 2003; Imai, King and
1 Ingeneral, error-correction models regress the first-difference of the dependent variable on a constant,
its own lag in levels, and the contemporaneous first-difference and lagged levels of each of the independent
variables. For instance, in the bivariate case we might estimate ∆yt = α0 + θ0 yt−1 + θ1 xt−1 + ∆xt .

1
Lau 2009; Williams and Whitten 2011, 2012; Philips, Rutherford and Whitten 2016b).

Below, we offer a brief discussion of the ARDL-bounds approach to cointegration

testing. We then present a Stata command that provides the necessary critical values for
the test—pssbounds—along with several applied examples. Next, we discuss dynardl, a
command to produce dynamic simulations of a host of ARDL-style models.

The ARDL-Bounds Cointegration Test

The concept of cointegration has been around for several decades. To understand coin-
tegration, we briefly discuss integrated versus stationary series. Time series may have
“full-memory,” such that current realizations are fully a function of all previous stochas-
tic shocks, plus some new innovation. Such series are said to be integrated of order one
(or I(1)), a form of non-stationarity.2 For instance, the series in Equation 1 is I(1), since
values at time t are a function of the prior value of y at time t − 1, plus innovation εt .3

yt = yt−1 + εt (1)

Two or more I(1) series may have short-run, as well as long-run—or equilibrium—
relationships. That is to say, while short-run perturbations may move the series apart,
over time this disequilibrium is corrected as the series move back towards a stable long-run
relationship. Such series are said to be cointegrating.

While a series may have perfect memory of its own history, it is certainly possible
that that history may not be in equilibrium with another series. Therefore, it becomes
necessary to test for cointegration. The earliest test comes from Engle and Granger (1987),
who show that cointegration between k weakly exogenous I(1) regressors, x1t , x2t , · · · , xkt ,
and an I(1) regressand, yt , exists if the resulting residuals—from a regression of these
2 Another term for an I(1) series is that it contains a unit-root. Series that are stationary are said to
be I(0).
3 This is commonly re-written as (y − y
t t−1 ) = ∆yt = εt .

2
variables entering into the equation in levels—is stationary:

yt = κ0 + κ1 x1t + κ2 x2t + · · · + κk xkt + zt (2)

A number of other cointegration tests have since been proposed. Philips (2017) offers
an in-depth discussion of how to apply the Pesaran, Shin and Smith (2001) ARDL-bounds
test for cointegration. Others include tests by Johansen (1991) and Phillips and Ouliaris
(1990). However, the ARDL-bounds test offers several advantages. Chief among them is
that users do not have make the sharp I(0)/I(1) distinction for the regressors.4 Below, we
briefly summarize this approach.5

First, the analyst must ensure that the dependent variable is I(1). There are many
unit-root tests that can be used to determine the order of integration of a series, including
the Dickey-Fuller, Phillips-Perron, Elliott-Rothenberg-Stock, and Kwiatkowski-Phillips-
Schmidt-Shin tests, among others. Only a non-stationary dependent variable is a potential
candidate for cointegration.

Second, the analyst must ensure that the regressors are not of an order of integration
higher than I(1). While this means that the analyst does not have to make the potentially-
difficult I(0)/I(1) decision, they must ensure that all regressors are not explosive or contain
seasonal unit roots.

Third, the analyst estimates an ARDL model in error-correction form. The model
appears as follows:

p
∆yt =α0 + θ0 yt−1 + θ1 x1,t−1 + · · · + θk xk,t−1 + ∑ αi ∆yt−1 +
i=1
q1 qk (3)
∑ β1 j ∆x1,t− j + · · · + ∑ βk j ∆xk,t− j + εt
j=0 j=0

4 Usersmust ensure, however, that regressors are not of order I(2) or more, and that seasonality has
been removed from the series.
5 A more in-depth discussion can be found in Philips (2017).

3
Where the change in the dependent variable, is a function of a constant, its prior value
(appearing in levels), prior values of all regressors in levels, as well as up to p and qk lags of
the first difference of the dependent variable and regressors, respectively. These may enter
into Equation 3 for theoretical reasons, but also have the added benefit of helping to ensure
white-noise residuals. While it is crucial to have well-behaved residuals in all models, this
is a necessary step before running the ARDL-bounds test for cointegration. Information
criteria such as SBIC and AIC, as well as autocorrelation and heteroskedasticity tests
(e.g., Breusch-Godfrey, Durbin’s Alternative, Cook-Weisberg, or Cumby-Huizinga tests)
can be used to check for white-noise residuals.

While the ARDL-bounds test may relatively easy to implement, users must look up the
special critical values. Pesaran, Shin and Smith (2001) provide asymptotic critical values,
while Narayan (2005) offers them for finite samples. In order to make the bounds test
more accessible to users, below we introduce pssbounds, provides the necessary critical
values through an easy-to-use command.

pssbounds Syntax

pssbounds, observations( ) k( ) fstat( ) [tstat( ) case( )]

Options

observations( ) is the number of observations from the estimated ARDL model in

error correction form. This is a required option.6

k( ) is the number of regressors, k, modeled in levels in the estimated ARDL model.7

The bounds F-test is a test that the k parameters on the regressors appearing in levels
(plus the coefficient on the lagged dependent variable, θ0 ) are jointly equal to zero: H0 =
6 As discussed below, if first using dynardl in error-correction form, users can simply run pssbounds
after without having to specify the required options, since the latter program obtains the necessary stored
values from the former.
7 For instance if the ARDL model was: ∆y = β − θ y
t 0 0 t−1 + β1 ∆x1t + θ1 x1,t−1 + β2 ∆x2t + θ3 x2,t−1 , then
k = 2, since x1,t−1 and x2,t−1 appear in lagged levels.

4
θ0 + θ1 + ... + θk = 0. This option is required, since the critical values differ based on the
number of regressors.

fstat( ) is the value of the F-statistic from the test that all parameters on the
regressors appearing in levels, plus the coefficient on the lagged dependent variable, are
jointly equal to zero: H0 = θ0 + θ1 + ... + θk = 0. After running the ARDL model in error-
correction form, users should use Stata’s test command to obtain the F-statistic.8 This
option is required.

type( ) specifies the potential restrictions on the constant and trend terms, which
in turn lead to different critical values for the bounds test. This option is not required,
however. The default if this option is not specified—by far the most common—is an
unrestricted intercept with no trend term: type(case3). Pesaran, Shin and Smith (2001)
refer to this as “Case III”. Other types that are supported are:

• Case I: No intercept and no trend, type(case1).

• Case II: Restricted intercept and no trend, type(case2).

• Case IV: Unrestricted intercept and restricted trend, type(case4).

• Case V: Unrestricted intercept and unrestricted trend, type(case5).

tstat( ) is the value of the t-statistic for the coefficient on the lagged dependent
variable. This serves as a one-sided auxiliary test to the bounds F-test. Only asymptotic
critical values are available, so this option is not required. Note that critical values do
not currently exist for this test for Cases II, and IV (see below).

Examples

Two example applications of pssbounds are shown below. For the first example, we will
use the Lutkepohl West German quarterly macroeconomic dataset available in Stata:
8 For example, test l.y l.x1 l.x2...

5
. webuse lutkepohl2

(Quarterly SA West German macro data, Bil DM, from Lutkepohl 1993 Table E.1

. tsset

time variable: qtr, 1960q1 to 1982q4

delta: 1 quarter

Next, we will run the following ARDL-bounds model:9

∆ln(Investment)t =α0 + θ0 ln(Investment)t−1 + β1 ∆ln(Income)t + θ1 ln(Income)t−1 +

(4)
β2 ∆ln(Consumption)t + θ2 ln(Consumption)t−1

In other words, we believe that investment is a function of income and consumption,

and that there is a cointegrating relationship between investment and the two regressors.
The results are shown in Model 1 in Table 1.

. regress d.ln_inv l.ln_inv d.ln_inc l.ln_inc d.ln_consump l.ln_consump

We then run an F-test that the coefficients on the variables appearing in levels
(ln_invt−1 , ln_inct−1 , and ln_consumpt−1 ) are jointly equal to zero:

. test l.ln_inv l.ln_inc l.ln_consump

( 1) L.ln_inv = 0

( 2) L.ln_inc = 0

( 3) L.ln_consump = 0

F( 3, 85) = 2.60

Prob > F = 0.0573

Recall that the critical values are non-standard, so we only need the value of the F-
statistic, which is 2.60. In order to test for cointegration, we use pssbounds. For the
required option fstat, we input the F-statistic from above. From the estimated model, we
9 As detailed in Philips (2017), estimation is not the first step of the ARDL-bounds approach; we would
need to conduct unit root test on the dependent variable, and ensure that the independent variables were
I(1) or less. We would also need to ensure that the resulting residuals from the model are white-noise.
This is a simply a stylized example used to showcase the command using readily-available Stata datasets.

6
 Table 1: Lutkepohl Example

(1) (2)
∆ln_invt ∆ln_invt
ln_invt−1 -0.140∗ -0.152∗∗
(0.059) (0.056)
∆ln_inct -0.201 -0.0985
(0.459) (0.423)
ln_inct−1 -0.209
(0.334)
∆ln_consumpt 1.548∗∗ 1.395∗∗
(0.538) (0.459)
ln_consumpt−1 0.336 0.131∗∗
(0.333) (0.048)
Constant -0.008
(0.071)
N 91 91
R2 0.17 0.27
Dependent variable is ∆ln_invt . Standard errors in
parentheses. ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001.

tell pssbounds that the number of observations is 91, the case is case III (i.e., unrestricted
intercept with no trend, as shown in Model 1), and that there are two regressors appearing
in levels (k = 2). The resulting output appears as follows:

. pssbounds, fstat(2.60) obs(91) case(3) k(2)

PESARAN, SHIN AND SMITH (2001) COINTEGRATION TEST

Obs: 91

No. Regressors (k): 2

Case: 3

------------------------------------------------------------

F-test

------------------------------------------------------------

<----- I(0) ---------- I(1) ----->

10% critical value 3.170 4.140

5% critical value 3.790 4.850

1% critical value 5.150 6.360

F-stat. = 2.600

7
------------------------------------------------------------

F-statistic note: Asymptotic critical values used.

For this model, since the F-statistic of 2.60 is below the I(0) critical value—even at the
10 percent level—we can conclude that there is no cointegration, and that all regressors
appearing in levels are stationary.

Purely for illustrative purposes, let us now assume that we wanted to estimate a model
without a constant. These results are shown in Model 2, Table 1. As with Model 1, we
next run an F-test of all variables appearing in levels. Note that the model also has
ln_inc appearing in first differences, but not in levels; thus, k = 1:

. regress d.ln_inv l.ln_inv d.ln_inc d.ln_consump l.ln_consump, noconstant

. test l.ln_inv l.ln_consump

( 1) L.ln_inv = 0

( 2) L.ln_consump = 0

F( 2, 87) = 3.83

Prob > F = 0.0254

We can account for a restricted constant by specifying the case(1) option (i.e., no in-
tercept and no trend). As an additional option, we add tstat(-2.73), which is the
t-statistic on the coefficient on the lagged dependent variable.

. pssbounds, fstat(3.83) obs(91) case(1) k(1) tstat(-2.73)

PESARAN, SHIN AND SMITH (2001) COINTEGRATION TEST

Obs: 91

No. Regressors (k): 1

Case: 1

------------------------------------------------------------

F-test

------------------------------------------------------------

8
 <----- I(0) ---------- I(1) ----->

10% critical value 2.440 3.280

5% critical value 3.150 4.110

1% critical value 4.810 6.020

F-stat. = 3.830

------------------------------------------------------------

t-test

------------------------------------------------------------

<----- I(0) ---------- I(1) ----->

10% critical value -1.620 -2.280

5% critical value -1.950 -2.600

1% critical value -2.580 -3.220

t-stat. = -2.730

------------------------------------------------------------

F-statistic note: Asymptotic critical values used.

t-statistic note: Asymptotic critical values used.

The output of pssbounds now contains critical values for both the F-test and one sided
t-test. Based on the F-statistic, we can conclude cointegration at the 10 percent level
since the F-statistic of 3.83 is above the I(1) critical threshold of 3.28. However, there
is not strong enough evidence to support cointegration at the five percent level. For the
t-test, the t-statistic of -2.73 falls below the critical I(1) threshold of -2.60, supporting
the earlier conclusion of cointegration. Also note that for both tests, pssbounds issued a
warning that asymptotic critical values are used. For all cases, only asymptotic critical
values from Pesaran, Shin and Smith (2001) are provided for the t-statistic test.10 Thus,
interpreting the results of this test should be done with caution in small samples. Small
sample critical values for the F-statistic are not available for case I.

As a second example, we use data from Ura (2014), who examines public mood lib-
10 Nor to critical values for the t-statistic test exist for cases II and IV.

9
eralism in the US. Ura argues that in the short-run, there will be a public “backlash” in
response to liberal Supreme Court decisions, but that in the long-run the sentiments of
the public tend to follow closely to those of the Court. Using the same dataset, Philips
(2017, see supplemental materials p. 110) finds that the dependent variable, public mood
liberalism, is I(1), and that Ura’s ARDL model estimated in error-correction form with
an additional lagged first-difference of unemployment produces the white-noise residuals
needed to conduct the ARDL-bounds test. The model is shown in Table 2.

. use "supreme court mood replication.dta", clear

. tsset

time variable: year, 1955 to 2009

delta: 1 unit

.regress d.mood l.mood d.policy l.policy d.unemployment dl.unemployment ///

l.unemployment d.inflation l.inflation d.caselaw l.caselaw

Next we run the F-test that all variables appearing in levels are jointly equal to zero:

. test l.mood l.policy l.unemployment l.inflation l.caselaw

( 1) L.mood = 0

( 2) L.policy = 0

( 3) L.unemployment = 0

( 4) L.inflation = 0

( 5) L.caselaw = 0

F( 5, 42) = 5.15

Prob > F = 0.0009

We then run pssbounds on the resulting F-statistic of 5.15, where k = 4 (i.e., four regres-
sors), along with the value of the t-statistic of the lagged dependent variable, which is
−3.19:

10
 Table 2: Ura (2014) Example

(1)
∆moodt
moodt−1 -0.241∗∗
(0.076)
∆policyt 0.051
(0.069)
policyt−1 -0.073∗∗∗
(0.020)
∆unemploymentt -0.106
(0.265)
∆unemploymentt−1 -0.538∗
(0.242)
unemploymentt−1 -0.024
(0.198)
∆inflationt -0.306∗
(0.123)
inflationt−1 -0.299∗
(0.120)
∆caselawt -0.093∗
(0.037)
caselawt−1 0.027∗
(0.011)
Constant 15.65∗∗
(5.050)
N 53
R2 0.47
Dependent variable is ∆moodt .
Standard errors in parentheses.
∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001.

11
. pssbounds, fstat(5.15) obs(53) case(3) k(4) tstat(-3.19)

PESARAN, SHIN AND SMITH (2001) COINTEGRATION TEST

Obs: 53

No. Regressors (k): 4

Case: 3

------------------------------------------------------------

F-test

------------------------------------------------------------

<----- I(0) ---------- I(1) ----->

10% critical value 2.578 3.710

5% critical value 3.068 4.334

1% critical value 4.244 5.726

F-stat. = 5.150

------------------------------------------------------------

t-test

------------------------------------------------------------

<----- I(0) ---------- I(1) ----->

10% critical value -2.570 -3.660

5% critical value -2.860 -3.990

1% critical value -3.430 -4.600

t-stat. = -3.190

------------------------------------------------------------

F-statistic note:

t-statistic note: Small-sample critical values not provided

for Case III. Asymptotic critical values used.

We can conclude evidence of cointegration at the 5 percent level for the F-test, since the
F-statistic of 5.15 is above the I(1) critical value of 4.334. However, note that we fail to

12
clear the critical I(1) threshold for the ARDL-bounds t-test. Overall though, given that
the t-test values are asymptotic (as shown by the note at the bottom of the pssbounds
output), we have relatively strong evidence of cointegration.

Dynamic Simulations of ARDL Models

ARDL models may have a fairly complex lag structure, with lags, contemporaneous val-
ues, first differences, and lagged first differences of the independent (and sometimes the
dependent) variable appearing in the model specification. While interpreting short- and
long-run effects may be simple in something like an ARDL(1,1) model (i.e., one lag of
the dependent variable, contemporaneous and one lag of all independent variables), un-
derstanding the short-, medium-, and long-run effects becomes difficult as the model
specification grows in complexity.

To better interpret the substantive significance of our results, below we introduce

dynardl, a command to dynamically simulate a variety of different ARDL models. dynardl
estimates, simulates, stores the results from and automatically plots a variety of ARDL
models. Users can even run pssbounds afterwards as a post-estimation command if sim-
ulating an error-correction model.

The output in dynardl helps us visualize the effect of a counterfactual change in one
weakly exogenous regressor at a single point in time, holding all else equal, using stochastic
simulation techniques. Dynamic simulation approaches are gaining in popularity as a
simple way to show the substantive results of time series models, whose coefficients often
have non-intuitive or “hidden” interpretations (Breunig and Busemeyer 2012; Williams
and Whitten 2011; Philips, Rutherford and Whitten 2016a,b; Gandrud, Williams and
Whitten 2016).11 Before using the command, it is assumed that the user has already
determined the order of integration of the variables through unit-root testing, diagnosed
and addressed other issues such as seasonal unit roots, and used information criteria (and
11 Forinstance, coefficients on regressors in an ARDL(1,0) have both a contemporaneous effect (given
β
by the coefficient) as well as a long-run or cumulative effect, given by 1−θ 0
.

13
theory) to identify the best fitting lagged-difference structure, which is used to purge
autocorrelation and to ensure the residuals are white noise. If an error-correction model
is estimated, users should use the ARDL-bounds test to determine if there is cointegration,
and if there is not, adjust the model accordingly.12

dynardl first runs a regression using OLS. Then, using a self-contained procedure
similar to the popular Clarify program for Stata (Tomz, Wittenberg and King 2003), it
takes 1000 draws (or however many simulations a user desires) of the vector of param-
eters from a multivariate normal distribution. These distributions are assumed to have
means equal to the estimated parameters from the regression. The variance of these dis-
tributions is equal to the estimated variance-covariance matrix from the regression. In
order to re-introduce stochastic uncertainty back into the model when creating predicted
values, dynardl simulates σ̂2∗ by taking draws from a scaled inverse χ2 distribution. The
distribution is scaled by the residual degrees of freedom (n-k), as well as the estimated
σ̂2 from the regression (Gelman et al. 2014, pp. 43, 581).13 Simulated parameters and
sigma-squared values are then used to create predicted values of the dependent variable
over time, Ŷt , for each of the simulations, by setting all covariates to certain values (typ-
ically means). Stochastic uncertainty is introduced into the prediction by taking a draw
from a multivariate normal distribution with mean zero and variance σ̂2 . The program
then averages across the simulations, creating Ŷt∗ (the predicted values plus stochastic un-
certainty) as well as percentile confidence intervals of the distribution of simulated values
at a particular point in time. These are then saved, allowing a user to make a table or
(more commonly) a graph of the results over time.

dynardl Syntax

dynardl depvar indepvars [, options]

12 Ifthe test is inconclusive, “Each regressor should be tested for a unit root. Only I(1) variables can
appear in levels in the error correction model. Stationary variables may still appear in first differences...If
the resulting statistic is still inconclusive, combinations of variables appearing in levels may need to be
tested.” (Philips 2017, pp. 13-14). See Philips (2017) for a step-by-step example of this process for the
ARDL model in general.
13 This ensures that draws of σ2 are bounded by zero and one.

14
 The options below are required:

• lags(numlist) is a numeric list of the number of lags to include for each variable.
The number of desired lags is listed in exactly the same order in which the variables
depvar and indepvars appear. For instance, the command:

dynardl y x1 x2, lags(1 2 3)· · ·

would lag y by t − 1, x1 by t − 2, and x2 by t − 3. Note that the lag on depvar

(always the first entry in lags()) must always be specified. To estimate a model
without a lag for a particular variable, simply replace the number with a “.”; for
instance, if we did not want a lag on the first regressor, and wanted a lag of t − 1 on
the second regressor, we type: lags(1 . 1). dynardl can handle consecutive lags
by specifying the minimum lag, a forward slash, followed by the maximum lag. For
instance, lags(1/3 . .) will introduce lags of yt at t − 1, t − 2, and t − 3 into the
model. To add a single lag of yt at t − 3, specify lags(3 . .).

• shockvar(varname) is a single independent variable from the list of indepvars

that is to be shocked. It will experience a counterfactual shock of size shockval(#)
at time time(#).

• shockval(#) is the amount to shock shockvar(varname) by. A common shock

value is a +/- one standard deviation shock.

The following options are not required:

• diffs(numlist) is a numeric list of the number of contemporaneous first differences

(i.e., t −(t −1)) to include for each variable. Note that the first entry (the placeholder
for the depvar) will always be empty (denoted by “.”), since the first difference of
the dependent variable cannot appear on the right-hand side of the model.14
14 It
can however, appear in lagged first differences, as shown below. Note that only first-differences
can be taken using this option (e.g., diffs(. 1 1)).

15
 • lagdiffs(numlist) is a numeric list of the number of lagged first differences to
include for each variable. The lag syntax is the same as for lags( ). For instance,
to include a lagged first difference at t − 2 for depvar, a lag at t − 1 for the first
weakly exogenous regressor, and none for the second, specify lagdiff(2 1 .). To
include an additional lag for both the first and second lagged first differences of
depvar, specify lagdiff(1/2 1 .).

• level(numlist) is a numeric list of variables to appear in levels (i.e., not lagged

or differenced but appearing contemporaneously at time t).15

• ec if specified, depvar will be estimated in first differences. If estimating an error

correction model, users will need to use this option.

• trend if specified, the program will add a deterministic linear trend to the model.

• noconstant if specified, the constant will be suppressed.

• range(#) is the length of the scenario to simulate. By default, this is t = 20. Note
that the range must be larger than time( ).

• sig(#) specifies the significance level for the percentile confidence intervals. The
default is for 95% confidence intervals.

• time(#) is the scenario time in which the shock occurs to shockvar( ). The default
time is t = 10.

• saving(string) specifies the name of the output file. If no filename is specified,

the program will save the results as “dynardl_results.dta”.

• forceset(numlist) by default, the program will estimate the ARDL model in

equilibrium; all lagged variables and variables appearing in levels are set to their
sample means. All first differences and any lagged first differences are set to zero.
This option allows the user to change the setting of the lagged (or unlagged if using
15 Ifboth level( ) and ec are specified, dynardl will issue a warning message. Of course, users may
have a valid reason to include a variable in levels; for instance, a dummy variable.

16
 level( )) levels of the variables. This could be useful when estimating a dummy
variable; for instance, when we wish to see the effect of a movement from zero to
one.

• sims(#) is the number of simulations (default is 1000). If confidence intervals are

particularly noisy, it may help to increase this number. Note that you may also
need to increase the matsize in Stata.

• burnin(#) allows dynardl to iterate out so starting values are stable. This option
is rarely used. However, if using the option forceset( ), the predicted values will
not be in equilibrium at the start of the simulation, and will take some time to
converge on stable values. To get around this, one can use the burnin option to
specify a number of simulations to “throw away” at the start. By default, this is 20.
Burnins do not change the simulation range or time; to simulate a range of 25 with
a shock time at 10 and a burnin of 30, specify: burnin(30) range(25) time(10).

• graph although dynardl saves the means of the predicted values and user-specified
confidence intervals in saving, users can use this option to automatically plot the
dynamic results using a spikeplot. Two alternative plots are possible:

– By adding the option rarea, the program will automatically create an area
plot. Predicted means along with 75, 90, and 95 percent confidence intervals
are shown with this option.

– By adding the option change, predicted changes (from the sample mean) are
shown across time, starting with the time at which the shock occurs; similar
to an impulse response function.

• expectedval by default, dynardl will calculate predicted values of the dependent

variable for a given number of simulations. For every simulation, the predicted value
comes from a systematic component as well as a single draw from the stochastic
component. With the expectedval option, the program instead calculates expected
values of the dependent variable such that the average of 1000 stochastic draws

17
 now becomes the estimate of the stochastic component for each of the simulations.
This effectively removes the stochastic uncertainty introduced in calculating Ŷt∗ .
Predicted values are more conservative than expected values. Note that dynardl
takes longer to run if calculating expected values. We recommend that unless the
user has a specific theoretical or substantive justification for using expected values,
they instead use the default predicted values that take into account the random
error from the predictions.

Examples

For the first example we will once again use the results from Model 1 in Table 1 using
the Lutkepohl data. We estimate the following model:

∆ln_invt = ln_invt−1 + ∆ln_inct + ln_inct−1 + ∆ln_consumpt + ln_consumpt−1 (5)

Since the dynardl uses Stata’s matrix capabilities, we will increase the maximum
matrix size as well:

. set matsize 5000

. webuse lutkepohl2

(Quarterly SA West German macro data, Bil DM, from Lutkepohl 1993 Table E.1

. tsset

time variable: qtr, 1960q1 to 1982q4

delta: 1 quarter

To estimate the model shown in Equation 5 using dynardl, and see the effect of a −1 shock
to ln_inc (about two standard deviations) we specify the command below. In Equation
5 we have two regressors and the lagged dependent variable, all of which are lagged one
period. Therefore, we specify lags(1 1 1). The first difference of all regressors appear,
so we add diff(. 1 1). There are no lagged differences appearing in Equation 5.

dynardl ln_inv ln_inc ln_consump, ///

18
lags(1 1 1) diffs(. 1 1) ///

shockvar(ln_inc) shockval(-1) ///

time(10) range(30) graph ec

In the command, lags(1 1 1) tells dynardl to add lags (of t −1) for each of the variables,
while diffs(. 1 1) means that the second and third variables (ln_inc and ln_consump)
enter into the model as first-differences as well. In shockvar( ) we include the variable
to be shocked, and specify the amount to shock it by using shockval. Additional options
include the time at which the shock occurs, time(10), the total range of the simulations,
range(30), and that the dependent variable is to be included in first-differences, ec. Last,
since we specified the graph option, dynardl will produce a plot, which is shown in Figure
1. As is clear from the figure, a −1 shock at t = 10 produces a small increase that is not
statistically significant in the short-run, which eventually increases to a predicted value
of about 7.5 over the long-run, an increase that is statistically significant.
8 7
Predicted Value
6
5

0 10 20 30
Time

Figure 1: Plot Produced from dynardl Using the graph Option

Note: Dots show average predicted value. Shaded lines show (from darkest to lightest) the 75, 90, and
95 percent confidence intervals.

19
 In addition to producing figures, dynardl also saved the prediction output, which
can be used to create more customizable figures (e.g., colors, lines, labels) if users desire.
Since we did not specify a filename to save as using the saving() option, the results are
automatically saved as “dynardl_results.dta”.

More complex dynamic specifications are possible using dynardl. For instance, per-
haps we wanted to estimate the following equation:

∆ln_invt =ln_invt−1 + ∆ln_invt−1 + ln_inct−1 + ln_inct−2 + ln_inct−3 +

(6)
∆ln_consumpt + ln_consumpt−1

Now, the dependent variable appears not only in lagged-level form at t − 1, but also as a
lagged first-difference. ln_inc no longer appears contemporaneously but at the first, sec-
ond, and third lags, while ln_consump continues to appear in lagged and first-differenced
forms. Equation 6 would appear as the following in dynardl:

dynardl ln_inv ln_inc ln_consump, ///

lags(1 1/3 1) diffs(. . 1) lagdiffs(1 . .) ///

shockvar(ln_inc) shockval(-1) ///

time(10) range(30) graph ec rarea sims(5000)

Note that, since the first through third lag of ln_inc are desired, we specify lags(1
1/3 1). Since there is a lagged first-difference included for the dependent variable, we
add lagdiffs(1 . .) (the “.” are placeholders for the other variables and must be
included). Last, we add the rarea option to produce an area plot, and increase the
number of simulations to 5000 with sims(5000). The resulting plot is shown in Figure
2. As is clear from the figure, as a result of a negative shock to ln(Income) at time
t = 10, investment decreases over the next few periods, though this does not appear to
be statistically significantly lower than the average predicted value (shown when t < 10).
After about three periods, investment increases in response to the negative income shock,
resulting in a new equilibrium prediction just above 10. Figure 2 is appealing since it shows
75, 90 and 95 percent confidence intervals, and since it is an area plot, its continuous style

20
may allow users to see changes slightly easier than the plot style in Figure 1. Note too
that since we increased the number of simulations to 5000, the confidence intervals in
Figure 2 are more smooth from time point to time point than in Figure 1.

12
10
Predicted Value
8
6
4

0 10 20 30
Time

Figure 2: Plot Produced from dynardl Using the rarea Option

Note: Black dotted line shows average predicted value. Shaded area shows (from darkest to lightest) the
75, 90, and 95 percent confidence intervals.

dynardl can also be used with pssbounds when estimating an error correction model.
For instance, directly after estimating Equation 5, we can run pssbounds as a post-
estimation command to test for cointegration, without having to specify any additional
options; the program automatically obtains the necessary values for the F- and t- statistics,
the number of observations, the case, and the number of regressors, k:

. dynardl ln_inv ln_inc ln_consump, ///

lags(1 1 1) diffs(. 1 1) ///

shockvar(ln_inc) shockval(-1) ///

time(10) range(30) ec

21
. pssbounds

PESARAN, SHIN AND SMITH (2001) COINTEGRATION TEST

Obs: 91

No. Regressors (k): 2

Case: 3

------------------------------------------------------------

F-test

------------------------------------------------------------

<----- I(0) ---------- I(1) ----->

10% critical value 3.170 4.140

5% critical value 3.790 4.850

1% critical value 5.150 6.360

F-stat. = 2.602

------------------------------------------------------------

t-test

------------------------------------------------------------

<----- I(0) ---------- I(1) ----->

10% critical value -2.570 -3.210

5% critical value -2.860 -3.530

1% critical value -3.430 -4.100

t-stat. = -2.372

------------------------------------------------------------

F-statistic note: Asymptotic critical values used.

t-statistic note: Asymptotic critical values used.

In addition to error-correction style models, dynardl can handle ARDL models where
the dependent variable is estimated in levels. For instance, perhaps we wanted to estimate

22
the following ARDL(1,1) model:

ln_invt = ln_invt−1 + ln_inct + ln_inct−1 + ln_consumpt + ln_consumpt−1 (7)

In dynardl, we add the level(. 1 1) to let the program know that the two inde-
pendent varibles are to appear contemporaneously in levels. If we wanted to see the effect
of a change from ln_inc= 6 to ln_inc= 5, while holding ln_consump constant at 7, we
can also use the forceset( ) option to force the program to evaluate the simulations at
these values, not the sample means (by default).16

dynardl ln_inv ln_inc ln_consump, ///

lags(1 1 1) level(. 1 1) forceset(. 6 7) ///

shockvar(ln_inc) shockval(-1) ///

time(10) range(30) graph change sims(5000)

Since we added the change option, the resulting plot is akin to an impulse response
function, as shown in Figure 3. In other words, we are looking not at the level of the
predicted value, but the difference between the predictions at each point in time and the
average predicted value before the shock. Figure 3 shows the change in predicted value,
starting when the shock occurs.17 As is clear from the figure, there is no statistically
significant change in the predicted value in the short-run as a result of the shock. However,
over the long run the change is statistically significant at the 90 percent level of confidence.

Conclusion

In this paper we have introduced two Stata programs to assist time series analysts.
pssbounds is designed to help users test for cointegration by providing critical values
16 Note that while we can set some or all of the independent variables using forceset, the lagged
dependent variable cannot be forced to a fixed value.
17 Since the shock occurred at t = 10, and the total range of the simulation was t = 30, this is why

Figure 3 shows a total range of t = 20

23
 3 2
Change in Predicted Value
0 1-1

0 5 10 15 20
Time

Figure 3: Plot Produced from dynardl Using the change Option

Note: Dots show mean change in predicted value from sample mean. Shaded area shows (from darkest
to lightest) the 75, 90, and 95 percent confidence intervals.

from Pesaran, Shin and Smith (2001) and Narayan (2005) automatically in a tabular for-
mat. The program dynardl helps users dynamically simulate a variety of autoregressive
distributed lag models in order to gain a better understanding of the substantive signifi-
cance of their results. Users can then graph or save their simulated predicted values for
use elsewhere. Both programs make it easier for users to test and interpret their dynamic
models.

24
References

Breunig, Christian and Marius R Busemeyer. 2012. “Fiscal austerity and the trade-off
between public investment and social spending.” Journal of European Public Policy
19(6):921–938.

Engle, Robert F and Clive WJ Granger. 1987. “Co-integration and error correction:
representation, estimation, and testing.” Econometrica 55(2):251–276.

Gandrud, Christopher, Laron K Williams and Guy D Whitten. 2016. “dynsim: Dynamic
simulations of autoregressive relationships.” R package version 1.2.2 .

Gelman, Andrew, John B Carlin, Hal S Stern and Donald B Rubin. 2014. Bayesian data
analysis. Vol. 2 Chapman and Hall/CRC Boca Raton, FL, USA.

Imai, Kosuke, Gary King and Olivia Lau. 2009. “Zelig: Everyone’s statistical software.”
R package version 3(5).

Jennings, Will and Peter John. 2009. “The dynamics of political attention: Public opinion
and the Queen’s Speech in the United Kingdom.” American Journal of Political Science
53(4):838–854.

Johansen, Søren. 1991. “Estimation and hypothesis testing of cointegration vectors in

Gaussian vector autoregressive models.” Econometrica: Journal of the Econometric
Society 59(6):1551–1580.

Johansen, Soren. 1995. Likelihood-based inference in cointegrated vector autoregressive

models. Oxford University Press.

Narayan, Paresh Kumar. 2005. “The saving and investment nexus for China: Evidence
from cointegration tests.” Applied Economics 37(17):1979–1990.

Pesaran, M Hashem, Yongcheol Shin and Richard J Smith. 2001. “Bounds testing
approaches to the analysis of level relationships.” Journal of Applied Econometrics
16(3):289–326.

25
Philips, Andrew Q. 2017. “Have your cake and eat it too? Cointegration and dynamic
inference from autoregressive distributed lag models.” American Journal of Political
Science .

Philips, Andrew Q, Amanda Rutherford and Guy D Whitten. 2016a. “Dynamic pie: A
strategy for modeling trade-offs in compositional variables over time.” American Journal
of Political Science 60(1):268–283.

Philips, Andrew Q, Amanda Rutherford and Guy D Whitten. 2016b. “dynsimpie: A com-
mand to examine dynamic compositional dependent variables.” Stata Journal 16(3):662–
677.

Phillips, Peter CB and Sam Ouliaris. 1990. “Asymptotic properties of residual based tests
for cointegration.” Econometrica: Journal of the Econometric Society pp. 165–193.

Swank, Duane and Sven Steinmo. 2002. “The new political economy of taxation in ad-
vanced capitalist democracies.” American Journal of Political Science pp. 642–655.

Tomz, Michael, Jason Wittenberg and Gary King. 2003. “CLARIFY: Software for inter-
preting and presenting statistical results.” Journal of Statistical Software 8(1):1–30.

Ura, Joseph Daniel. 2014. “Backlash and legitimation: Macro political responses to
supreme court decisions.” American Journal of Political Science 58(1):110–126.

Ura, Joseph Daniel and Christopher R Ellis. 2008. “Income, preferences, and the dynamics
of policy responsiveness.” PS: Political Science & Politics 41(4):785–794.

Whitten, Guy D and Laron K Williams. 2011. “Buttery guns and welfare hawks: The
politics of defense spending in advanced industrial democracies.” American Journal of
Political Science 55(1):117–134.

Williams, Laron K and Guy D Whitten. 2011. “Dynamic simulations of autoregressive

relationships.” Stata Journal 11(4):577–588.

Williams, Laron K and Guy D Whitten. 2012. “But wait, there’s more! Maximizing
substantive inferences from TSCS models.” The Journal of Politics 74(03):685–693.

26