Title stata.

com
vec — Vector error-correction models

Description Quick start Menu Syntax

Options Remarks and examples Stored results Methods and formulas
References Also see

Description
vec fits a type of vector autoregression in which some of the variables are cointegrated by using
Johansen’s (1995) maximum likelihood method. Constraints may be placed on the parameters in the
cointegrating equations or on the adjustment terms. See [TS] vec intro for a list of commands that
are used in conjunction with vec.

Quick start
Vector error-correction model for y1, y2, and y3 using tsset data
vec y1 y2 y3
Use 4 lags for the underlying vector autoregressive (VAR) model
vec y1 y2 y3, lags(4)
Use 2 cointegrating equations
vec y1 y2 y3, lags(4) rank(2)
Add a linear trend in the cointegrating equations and a quadratic trend in the undifferenced data
vec y1 y2 y3, lags(4) rank(2) trend(trend)
Same as above, but without a trend or a constant
vec y1 y2 y3, lags(4) rank(2) trend(none)

Menu
Statistics > Multivariate time series > Vector error-correction model (VECM)

1
 2 vec — Vector error-correction models

Syntax
     
vec depvarlist if in , options

options Description
Model
rank(#) use # cointegrating equations; default is rank(1)
lags(#) use # for the maximum lag in underlying VAR model
trend(constant) include an unrestricted constant in model; the default
trend(rconstant) include a restricted constant in model
trend(trend) include a linear trend in the cointegrating equations and a
quadratic trend in the undifferenced data
trend(rtrend) include a restricted trend in model
trend(none) do not include a trend or a constant
bconstraints(constraintsbc ) place constraintsbc on cointegrating vectors
aconstraints(constraintsac ) place constraintsac on adjustment parameters
Adv. model
sindicators(varlistsi ) include normalized seasonal indicator variables varlistsi
noreduce do not perform checks and corrections for collinearity among
lags of dependent variables
Reporting
level(#) set confidence level; default is level(95)
nobtable do not report parameters in the cointegrating equations
noidtest do not report the likelihood-ratio test of overidentifying
restrictions
alpha report adjustment parameters in separate table
pi report parameters in Π = αβ0
noptable do not report elements of Π matrix
mai report parameters in the moving-average impact matrix
noetable do not report adjustment and short-run parameters
dforce force reporting of short-run, beta, and alpha parameters when
the parameters in beta are not identified; advanced option
nocnsreport do not display constraints
display options control columns and column formats, row spacing, and line width
Maximization
maximize options control the maximization process; seldom used
coeflegend display legend instead of statistics

vec does not allow gaps in the data.

You must tsset your data before using vec; see [TS] tsset.
varlist must contain at least two variables and may contain time-series operators; see [U] 11.4.4 Time-series varlists.
by, collect, fp, rolling, statsby, and xi are allowed; see [U] 11.1.10 Prefix commands.
coeflegend does not appear in the dialog box.
See [U] 20 Estimation and postestimation commands for more capabilities of estimation commands.
 vec — Vector error-correction models 3

Options

 Model

rank(#) specifies the number of cointegrating equations; rank(1) is the default.

lags(#) specifies the maximum lag to be included in the underlying VAR model. The maximum lag
in a vector error-correction (VEC) model is one smaller than the maximum lag in the corresponding
VAR model in levels; the number of lags must be greater than zero but small enough so that the
degrees of freedom used up by the model are fewer than the number of observations. The default
is lags(2).
trend(trend spec) specifies which of Johansen’s five trend specifications to include in the model.
These specifications are discussed in Specification of constants and trends below. The default is
trend(constant).
bconstraints(constraintsbc ) specifies the constraints to be placed on the parameters of the coin-
tegrating equations. When no constraints are placed on the adjustment parameters—that is, when
the aconstraints() option is not specified—the default is to place the constraints defined by
Johansen’s normalization on the parameters of the cointegrating equations. When constraints are
placed on the adjustment parameters, the default is not to place constraints on the parameters in
the cointegrating equations.
aconstraints(constraintsac ) specifies the constraints to be placed on the adjustment parameters.
By default, no constraints are placed on the adjustment parameters.


 Adv. model

sindicators(varlistsi ) specifies the normalized seasonal indicator variables to include in the model.
The indicator variables specified in this option must be normalized as discussed in Johansen (1995).
If the indicators are not properly normalized, the estimator of the cointegrating vector does not
converge to the asymptotic distribution derived by Johansen (1995). More details about how these
variables are handled are provided in Methods and formulas. sindicators() cannot be specified
with trend(none) or with trend(rconstant).
noreduce causes vec to skip the checks and corrections for collinearity among the lags of the
dependent variables. By default, vec checks to see whether the current lag specification causes
some of the regressions performed by vec to contain perfectly collinear variables; if so, it reduces
the maximum lag until the perfect collinearity is removed.

 Reporting

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

nobtable suppresses the estimation table for the parameters in the cointegrating equations. By default,
vec displays the estimation table for the parameters in the cointegrating equations.
noidtest suppresses the likelihood-ratio test of the overidentifying restrictions, which is reported
by default when the model is overidentified.
alpha displays a separate estimation table for the adjustment parameters, which is not displayed by
default.
pi displays a separate estimation table for the parameters in Π = αβ0 , which is not displayed by
default.
noptable suppresses the estimation table for the elements of the Π matrix, which is displayed by
default when the parameters in the cointegrating equations are not identified.
 4 vec — Vector error-correction models

mai displays a separate estimation table for the parameters in the moving-average impact matrix,
which is not displayed by default.
noetable suppresses the main estimation table that contains information about the estimated adjustment
parameters and the short-run parameters, which is displayed by default.
dforce displays the estimation tables for the short-run parameters and α and β—if the last two are
requested—when the parameters in β are not identified. By default, when the specified constraints
do not identify the parameters in the cointegrating equations, estimation tables are displayed only
for Π and the MAI.
nocnsreport; see [R] Estimation options.
display options: noci, nopvalues, vsquish, cformat(% fmt), pformat(% fmt), sformat(% fmt),
and nolstretch; see [R] Estimation options.

 Maximization
 
maximize options: iterate(#), no log, trace, toltrace, tolerance(#), ltolerance(#),
afrom(matrixa ), and bfrom(matrixb ); see [R] Maximize.
toltrace displays the relative differences for the log likelihood and the coefficient vector at every
iteration. This option cannot be specified if no constraints are defined or if nolog is specified.
afrom(matrixa ) specifies a 1 ×(K ∗r) row vector with starting values for the adjustment parameters,
where K is the number of endogenous variables and r is the number of cointegrating equations
specified in the rank() option. The starting values should be ordered as they are reported in
e(alpha). This option cannot be specified if no constraints are defined.
bfrom(matrixb ) specifies a 1 × (m1 ∗ r) row vector with starting values for the parameters of the
cointegrating equations, where m1 is the number of variables in the trend-augmented system and
r is the number of cointegrating equations specified in the rank() option. (See Methods and
formulas for more details about m1 .) The starting values should be ordered as they are reported
in e(betavec). As discussed in Methods and formulas, for some trend specifications, e(beta)
contains parameter estimates that are not obtained directly from the optimization algorithm.
bfrom() should specify only starting values for the parameters reported in e(betavec). This
option cannot be specified if no constraints are defined.

The following option is available with vec but is not shown in the dialog box:
coeflegend; see [R] Estimation options.

Remarks and examples stata.com

Remarks are presented under the following headings:
Introduction
Specification of constants and trends
Collinearity
 vec — Vector error-correction models 5

Introduction
VEC models are used to model the stationary relationships between multiple time series that contain
unit roots. vec implements Johansen’s approach for estimating the parameters of a VEC model.
[TS] vec intro reviews the basics of integration and cointegration and highlights why we need
special methods for modeling the relationships between processes that contain unit roots. This manual
entry assumes familiarity with the material in [TS] vec intro and provides examples illustrating how to
use the vec command. See Johansen (1995), Hamilton (1994), and Becketti (2020) for more in-depth
introductions to cointegration analysis.

Example 1
This example uses annual data on the average per-capita disposable personal income in the eight
U.S. Bureau of Economic Analysis (BEA) regions of the United States. We use data from 1948–2002
in logarithms. Unit-root tests on these series fail to reject the null hypothesis that per-capita disposable
income in each region contains a unit root. Because capital and labor can move easily between the
different regions of the United States, we would expect that no one series will diverge from all the
remaining series and that cointegrating relationships exist.
Below we graph the natural logs of average disposal income in the New England and the Southeast
regions.
. use https://www.stata-press.com/data/r18/rdinc
. line ln_ne ln_se year

11

10

9 ln(New England)
ln(Southeast)

1950 1960 1970 1980 1990 2000

Year

The graph indicates a differential between the two series that shrinks between 1960 and about
1980 and then grows until it stabilizes around 1990. We next estimate the parameters of a bivariate
VEC model with one cointegrating relationship.
6 vec — Vector error-correction models

. vec ln_ne ln_se

Vector error-correction model
Sample: 1950 thru 2002 Number of obs = 53
AIC = -11.00462
Log likelihood = 300.6224 HQIC = -10.87595
Det(Sigma_ml) = 4.06e-08 SBIC = -10.67004
Equation Parms RMSE R-sq chi2 P>chi2

D_ln_ne 4 .017896 0.9313 664.4668 0.0000

D_ln_se 4 .018723 0.9292 642.7179 0.0000

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

D_ln_ne
_ce1
L1. -.4337524 .0721365 -6.01 0.000 -.5751373 -.2923675

ln_ne
LD. .7168658 .1889085 3.79 0.000 .3466119 1.08712

ln_se
LD. -.6748754 .2117975 -3.19 0.001 -1.089991 -.2597599

_cons -.0019846 .0080291 -0.25 0.805 -.0177214 .0137521

D_ln_se
_ce1
L1. -.3543935 .0754725 -4.70 0.000 -.5023168 -.2064701

ln_ne
LD. .3366786 .1976448 1.70 0.088 -.050698 .7240553

ln_se
LD. -.1605811 .2215922 -0.72 0.469 -.5948939 .2737317

_cons .002429 .0084004 0.29 0.772 -.0140355 .0188936

Cointegrating equations
Equation Parms chi2 P>chi2

_ce1 1 29805.02 0.0000

Identification: beta is exactly identified

Johansen normalization restriction imposed

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

_ce1
ln_ne 1 . . . . .
ln_se -.9433708 .0054643 -172.64 0.000 -.9540807 -.9326609
_cons -.8964065 . . . . .

The default output has three parts. The header provides information about the sample, the model
fit, and the identification of the parameters in the cointegrating equation. The main estimation table
contains the estimates of the short-run parameters, along with their standard errors and confidence
intervals. The second estimation table reports the estimates of the parameters in the cointegrating
equation, along with their standard errors and confidence intervals.
 vec — Vector error-correction models 7

The results indicate strong support for a cointegrating equation such that

ln ne − 0.943 ln se − 0.896

should be a stationary series. Identification of the parameters in the cointegrating equation is achieved
by constraining some of them to be fixed, and fixed parameters do not have standard errors. In this
example, the coefficient on ln ne has been normalized to 1, so its standard error is missing. As
discussed in Methods and formulas, the constant term in the cointegrating equation is not directly
estimated in this trend specification but rather is backed out from other estimates. Not all the elements
of the VCE that correspond to this parameter are readily available, so the standard error for the cons
parameter is missing.
To get a better idea of how our model fits, we predict the cointegrating equation and graph it over
time:
. predict ce, ce
. tsline ce

-.05
Predicted cointegrated equation

-.1

-.15

-.2

-.25
1950 1960 1970 1980 1990 2000
Year

Although the predicted cointegrating equation has the right appearance for the time before the
mid-1960s, afterward the predicted cointegrating equation does not look like a stationary series. A
better model would account for the trends in the size of the differential.

As discussed in [TS] vec intro, simply normalizing one of the coefficients to be one is sufficient to
identify the parameters of the single cointegrating vector. When there is more than one cointegrating
equation, more restrictions are required.
 8 vec — Vector error-correction models

Example 2
We have data on monthly unemployment rates in Indiana, Illinois, Kentucky, and Missouri from
January 1978 through December 2003. We suspect that factor mobility will keep the unemployment
rates in equilibrium. The following graph plots the data.
. use https://www.stata-press.com/data/r18/urates, clear
. line missouri indiana kentucky illinois t

12

10

Missouri
8 Indiana
Kentucky
Illinois
6

2
1980m1 1985m1 1990m1 1995m1 2000m1 2005m1
Time

The graph shows that although the series do appear to move together, the relationship is not as clear
as in the previous example. There are periods when Indiana has the highest rate and others when
Indiana has the lowest rate. Although the Kentucky rate moves closely with the other series for most
of the sample, there is a period in the mid-1980s when the unemployment rate in Kentucky does not
fall at the same rate as the other series.
We will model the series with two cointegrating equations and no linear or quadratic time trends
in the original series. Because we are focusing on the cointegrating vectors, we use the noetable
option to suppress displaying the short-run estimation table.
 vec — Vector error-correction models 9

. vec missouri indiana kentucky illinois, trend(rconstant) rank(2) lags(4)

> noetable
Vector error-correction model
Sample: 1978m5 thru 2003m12 Number of obs = 308
AIC = -2.306048
Log likelihood = 417.1314 HQIC = -2.005818
Det(Sigma_ml) = 7.83e-07 SBIC = -1.555184
Cointegrating equations
Equation Parms chi2 P>chi2

_ce1 2 133.3885 0.0000

_ce2 2 195.6324 0.0000

Identification: beta is exactly identified

Johansen normalization restrictions imposed

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

_ce1
missouri 1 . . . . .
indiana 0 (omitted)
kentucky .3493902 .2005537 1.74 0.081 -.0436879 .7424683
illinois -1.135152 .2069063 -5.49 0.000 -1.540681 -.7296235
_cons -.3880707 .4974323 -0.78 0.435 -1.36302 .5868787

_ce2
missouri -1.11e-16 . . . . .
indiana 1 . . . . .
kentucky .2059473 .2718678 0.76 0.449 -.3269038 .7387985
illinois -1.51962 .2804792 -5.42 0.000 -2.069349 -.9698907
_cons 2.92857 .6743122 4.34 0.000 1.606942 4.250197

Except for the coefficients on kentucky in the two cointegrating equations and the constant
term in the first, all the parameters are significant at the 5% level. We can refit the model with the
Johansen normalization and the overidentifying constraint that the coefficient on kentucky in the
second cointegrating equation is zero.
. constraint define 1 [_ce1]missouri = 1
. constraint define 2 [_ce1]indiana = 0
. constraint define 3 [_ce2]missouri = 0
. constraint define 4 [_ce2]indiana = 1
. constraint define 5 [_ce2]kentucky = 0
10 vec — Vector error-correction models

. vec missouri indiana kentucky illinois, trend(rconstant) rank(2)

> lags(4) noetable bconstraints(1/5)
Iteration 1: Log likelihood = 416.9718
(output omitted )
Iteration 20: Log likelihood = 416.9744
Vector error-correction model
Sample: 1978m5 thru 2003m12 Number of obs = 308
AIC = -2.311522
Log likelihood = 416.9744 HQIC = -2.016134
Det(Sigma_ml) = 7.84e-07 SBIC = -1.572769
Cointegrating equations
Equation Parms chi2 P>chi2

_ce1 2 145.233 0.0000

_ce2 1 209.9344 0.0000

Identification: beta is overidentified

( 1) [_ce1]missouri = 1
( 2) [_ce1]indiana = 0
( 3) [_ce2]missouri = 0
( 4) [_ce2]indiana = 1
( 5) [_ce2]kentucky = 0

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

_ce1
missouri 1 . . . . .
indiana 0 (omitted)
kentucky .2521685 .1649653 1.53 0.126 -.0711576 .5754946
illinois -1.037453 .1734165 -5.98 0.000 -1.377343 -.6975626
_cons -.3891102 .4726968 -0.82 0.410 -1.315579 .5373586

_ce2
missouri 0 (omitted)
indiana 1 . . . . .
kentucky 0 (omitted)
illinois -1.314265 .0907071 -14.49 0.000 -1.492048 -1.136483
_cons 2.937016 .6448924 4.55 0.000 1.67305 4.200982

LR test of identifying restrictions: chi2(1) = .3139 Prob > chi2 = 0.575

The test of the overidentifying restriction does not reject the null hypothesis that the restriction
is valid, and the p-value on the coefficient on kentucky in the first cointegrating equation indicates
that it is not significant. We will leave the variable in the model and attribute the lack of significance
to whatever caused the kentucky series to temporarily rise above the others from 1985 until 1990,
though we could instead consider removing kentucky from the model.
Next, we look at the estimates of the adjustment parameters. In the output below, we replay
the previous results. We specify the alpha option so that vec will display an estimation table for
the estimates of the adjustment parameters, and we specify nobtable to suppress the table for the
parameters of the cointegrating equations because we have already looked at those.
 vec — Vector error-correction models 11

. vec, alpha nobtable noetable

Vector error-correction model
Sample: 1978m5 thru 2003m12 Number of obs = 308
AIC = -2.311522
Log likelihood = 416.9744 HQIC = -2.016134
Det(Sigma_ml) = 7.84e-07 SBIC = -1.572769
Adjustment parameters
Equation Parms chi2 P>chi2

D_missouri 2 19.39607 0.0001

D_indiana 2 6.426086 0.0402
D_kentucky 2 8.524901 0.0141
D_illinois 2 22.32893 0.0000

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

D_missouri
_ce1
L1. -.0683152 .0185763 -3.68 0.000 -.1047242 -.0319063

_ce2
L1. .0405613 .0112417 3.61 0.000 .018528 .0625946

D_indiana
_ce1
L1. -.0342096 .0220955 -1.55 0.122 -.0775159 .0090967

_ce2
L1. .0325804 .0133713 2.44 0.015 .0063732 .0587877

D_kentucky
_ce1
L1. -.0482012 .0231633 -2.08 0.037 -.0936004 -.0028021

_ce2
L1. .0374395 .0140175 2.67 0.008 .0099657 .0649133

D_illinois
_ce1
L1. .0138224 .0227041 0.61 0.543 -.0306768 .0583215

_ce2
L1. .0567664 .0137396 4.13 0.000 .0298373 .0836955

LR test of identifying restrictions: chi2(1) = .3139 Prob > chi2 = 0.575

All the coefficients are significant at the 5% level, except those on Indiana and Illinois in the first
cointegrating equation. From an economic perspective, the issue is whether the unemployment rates
in Indiana and Illinois adjust when the first cointegrating equation is out of equilibrium. We could
impose restrictions on one or both of those parameters and refit the model, or we could just decide
to use the current results.
 12 vec — Vector error-correction models

Technical note
vec can be used to fit models in which the parameters in β are not identified, in which case only
the parameters in Π and the moving-average impact matrix C are identified. When the parameters in
β are not identified, the values of β
b and α
b can vary depending on the starting values. However, the
estimates of Π and C are identified and have known asymptotic distributions. This method is valid
because these additional normalization restrictions impose no restriction on Π or C.

Specification of constants and trends

As discussed in [TS] vec intro, allowing for a constant term and linear time trend allow us to
write the VEC model as
p−1
X
∆yt = α(βyt−1 + µ + ρt) + Γi ∆yt−i + γ + τ t + t
i=1

Five different trend specifications are available:

Option in trend() Parameter restrictions Johansen (1995) notation

trend none H(r)
rtrend τ=0 H ∗ (r)
constant ρ = 0, and τ = 0 H1 (r)
rconstant ρ = 0, γ = 0 and τ = 0 H1∗ (r)
none µ = 0, ρ = 0, γ = 0, and τ = 0 H2 (r)

trend(trend) allows for a linear trend in the cointegrating equations and a quadratic trend in
the undifferenced data. A linear trend in the cointegrating equations implies that the cointegrating
equations are assumed to be trend stationary.
trend(rtrend) defines a restricted trend model that excludes linear trends in the differenced data
but allows for linear trends in the cointegrating equations. As in the previous case, a linear trend in
a cointegrating equation implies that the cointegrating equation is trend stationary.
trend(constant) defines a model with an unrestricted constant. This allows for a linear trend
in the undifferenced data and cointegrating equations that are stationary around a nonzero mean. This
is the default.
trend(rconstant) defines a model with a restricted constant in which there is no linear or
quadratic trend in the undifferenced data. A nonzero µ allows for the cointegrating equations to be
stationary around nonzero means, which provide the only intercepts for differenced data. Seasonal
indicators are not allowed with this specification.
trend(none) defines a model that does not include a trend or a constant. When there is no trend
or constant, the cointegrating equations are restricted to being stationary with zero means. Also, after
adjusting for the effects of lagged endogenous variables, the differenced data are modeled as having
mean zero. Seasonal indicators are not allowed with this specification.
 vec — Vector error-correction models 13

Technical note
vec uses a switching algorithm developed by Boswijk (1995) to maximize the log-likelihood
function when constraints are placed on the parameters. The starting values affect both the ability of
the algorithm to find a maximum and its speed in finding that maximum. By default, vec uses the
parameter estimates that correspond to Johansen’s normalization. Sometimes, other starting values
will cause the algorithm to find a maximum faster.
To specify starting values for the parameters in α, we specify a 1 × (K ∗ r) matrix in the afrom()
option. Specifying starting values for the parameters in β is slightly more complicated. As explained
in Methods and formulas, specifying trend(constant), trend(rtrend), or trend(trend) causes
some of the estimates of the trend parameters appearing in β b to be “backed out”. The switching
algorithm estimates only the parameters of the cointegrating equations whose estimates are stored in
e(betavec). For this reason, only the parameters stored in e(betavec) can have their initial values
set via bfrom().
The table below describes which trend parameters in the cointegrating equations are estimated by
the switching algorithm for each of the five specifications.
Trend specification Trend parameters in Trend parameter estimated
cointegrating equations via switching algorithm
none none none
rconstant cons cons
constant cons none
rtrend cons, trend trend
trend cons, trend none

Collinearity
As expected, collinearity among variables causes some parameters to be unidentified numerically.
If vec encounters perfect collinearity among the dependent variables, it exits with an error.
In contrast, if vec encounters perfect collinearity that appears to be due to too many lags in the
model, vec displays a warning message and reduces the maximum lag included in the model in an
effort to find a model with fewer lags in which all the parameters are identified by the data. Specifying
the noreduce option causes vec to skip over these additional checks and corrections for collinearity.
Thus the noreduce option can be used to force the estimation to proceed when not all the parameters
are identified by the data. When some parameters are not identified because of collinearity, the results
cannot be interpreted but can be used to find the source of the collinearity.
 14 vec — Vector error-correction models

Stored results
vec stores the following in e():
Scalars
e(N) number of observations
e(k rank) number of unconstrained parameters
e(k eq) number of equations in e(b)
e(k dv) number of dependent variables
e(k ce) number of cointegrating equations
e(n lags) number of lags
e(df m) model degrees of freedom
e(ll) log likelihood
e(chi2 res) value of test of overidentifying restrictions
e(df lr) degrees of freedom of the test of overidentifying restrictions
e(beta iden) 1 if the parameters in β are identified and 0 otherwise
e(beta icnt) number of independent restrictions placed on β
e(k #) number of variables in equation #
e(df m#) model degrees of freedom in equation #
e(r2 #) R2 of equation #
e(chi2 #) χ2 statistic for equation #
e(rmse #) RMSE of equation #
e(aic) value of AIC
e(hqic) value of HQIC
e(sbic) value of BIC
e(tmin) minimum time
e(tmax) maximum time
e(detsig ml) determinant of the estimated covariance matrix
e(rank) rank of e(V)
e(converge) 1 if the switching algorithm converged, 0 if it did not converge
Macros
e(cmd) vec
e(cmdline) command as typed
e(trend) trend specified
e(tsfmt) format of the time variable
e(tvar) variable denoting time within groups
e(endog) endogenous variables
e(covariates) list of covariates
e(eqnames) equation names
e(cenames) names of cointegrating equations
e(reduce opt) noreduce, if noreduce is specified
e(reduce lags) list of maximum lags to which the model has been reduced
e(title) title in estimation output
e(aconstraints) constraints placed on α
e(bconstraints) constraints placed on β
e(sindicators) seasonal indicator variables
e(properties) b V
e(predict) program used to implement predict
e(marginsok) predictions allowed by margins
e(marginsnotok) predictions disallowed by margins
e(marginsdefault) default predict() specification for margins
Matrices
e(b) estimates of short-run parameters
e(V) VCE of short-run parameter estimates
e(beta) estimates of β
e(V beta) VCE of β b
e(betavec) directly obtained estimates of β
e(pi) estimates of Π
b
e(V pi) VCE of Π b
e(alpha) estimates of α
e(V alpha) VCE of α b
 vec — Vector error-correction models 15

e(omega) estimates of Ω
b
e(mai) estimates of C
e(V mai) VCE of C b
Functions
e(sample) marks estimation sample

In addition to the above, the following is stored in r():

Matrices
r(table) matrix containing the coefficients with their standard errors, test statistics, p-values,
and confidence intervals

Note that results stored in r() are updated when the command is replayed and will be replaced when
any r-class command is run after the estimation command.

Methods and formulas

Methods and formulas are presented under the following headings:
General specification of the VEC model
The log-likelihood function
Unrestricted trend
Restricted trend
Unrestricted constant
Restricted constant
No trend
Estimation with Johansen identification
Estimation with constraints: β identified
Estimation with constraints: β not identified
Formulas for the information criteria
Formulas for predict

General specification of the VEC model

vec estimates the parameters of a VEC model that can be written as

p−1
X
∆yt = αβ0 yt−1 + Γi ∆yt−i + v + δt + w1 s1 + · · · + wm sm + t (1)
i=1

where
yt is a K × 1 vector of endogenous variables,
α is a K × r matrix of parameters,
β is a K × r matrix of parameters,
Γ1 , . . . , Γp−1 are K × K matrices of parameters,
v is a K × 1 vector of parameters,
δ is a K × 1 vector of trend coefficients,
t is a linear time trend,
s1 , . . . , sm are orthogonalized seasonal indicators specified in the sindicators() option, and
w1 , . . . , wm are K × 1 vectors of coefficients on the orthogonalized seasonal indicators.
 16 vec — Vector error-correction models

There are two types of deterministic elements in (1): the trend, v + δt, and the orthogonalized
seasonal terms, w1 s1 + · · · + wm sm . Johansen (1995, chap. 11) shows that inference about the
number of cointegrating equations is based on nonstandard distributions and that the addition of any
term that generalizes the deterministic specification in (1) changes the asymptotic distributions of the
statistics used for inference on the number of cointegrating equations and the asymptotic distribution
of the ML estimator of the cointegrating equations. In fact, Johansen (1995, 84) notes that including
event indicators causes the statistics used for inference on the number of cointegrating equations to
have asymptotic distributions that must be computed case by case. For this reason, event indicators
may not be specified in the present version of vec.
If seasonal indicators are included in the model, they cannot be collinear with a constant term. If
they are collinear with a constant term, one of the indicator variables is omitted.
As discussed in Specification of constants and trends, we can reparameterize the model as
p−1
X
∆yt = α(βyt−1 + µ + ρt) + Γi ∆yt−i + γ + τ t + t (2)
i=1

The log-likelihood function

We can maximize the log-likelihood function much more easily by writing it in concentrated
form. In fact, as discussed below, in the simple case with the Johansen normalization on β and no
constraints on α, concentrating the log-likelihood function produces an analytical solution for the
parameter estimates.
To concentrate the log likelihood, rewrite (2) as

e 0 Z1t + ΨZ2t + t
Z0t = αβ (3)

where Z0t is a K × 1 vector of variables ∆yt , α is the K × r matrix of adjustment coefficients,

and t is a K × 1 vector of independent and identically distributed normal vectors with mean 0 and
contemporaneous covariance matrix Ω. Z1t , Z2t , β,
e and Ψ depend on the trend specification and are
defined below.
The log-likelihood function for the model in (3) is
1n
L = − T K ln(2π) + T ln(|Ω|)
2
T
e 0 Z1t − ΨZ2t )0 Ω−1 (Z0t − αβ
e 0 Z1t − ΨZ2t )
X o
+ (Z0t − αβ (4)
t=1

e have rank r.
with the constraints that α and β
Johansen (1995, chap. 6), building on Anderson (1951), shows how the Ψ parameters can be
e , and the data, yielding the concentrated log-likelihood function
expressed as analytic functions of α, β

1n
Lc = − T K ln(2π) + T ln(|Ω|)
2
T
e 0 R1t )0 Ω−1 (R0t − αβ
e 0 R1t )
X o
+ (R0t − αβ (5)
t=1
 vec — Vector error-correction models 17

where
PT
Mij = T −1 t=1 Zit Z0jt , i, j ∈ {0, 1, 2};
R0t = Z0t − M02 M−1
22 Z2t ; and
R1t = Z1t − M12 M−1
22 Z2t .

The definitions of Z1t , Z2t , β,

e and Ψ change with the trend specifications, although some of their
components stay the same.

Unrestricted trend

When the trend in the VEC model is unrestricted, we can define the variables in (3) directly in
terms of the variables in (1):
Z1t = yt−1 is K × 1
0 0
Z2t = (∆yt−1 , . . . , ∆yt−p+1 , 1, t, s1 , . . . , sm )0 is {K(p − 1) + 2 + m} × 1;
Ψ = (Γ1 , . . . , Γp−1 , v, δ, w1 , . . . , wm ) is K × {K(p − 1) + 2 + m}
e = β is the K × r matrix composed of the r cointegrating vectors.
β
In the unrestricted trend specification, m1 = K, m2 = K(p − 1) + 2 + m, and there are
nparms = Kr + Kr + K{K(p − 1) + 2 + m} parameters in (3).

Restricted trend

When there is a restricted trend in the VEC model in (2), τ = 0, but the intercept v = αµ + γ is
unrestricted. The VEC model with the restricted trend can be written as
  p−1
0 yt−1 X
∆yt = α(β , ρ) + Γi ∆yt−i + v + w1 s1 + · · · + wm sm + t
t
i=1

This equation can be written in the form of (3) by defining

0

0
Z1t = yt−1 , t is (K + 1) × 1
0 0
Z2t = (∆yt−1 , . . . , ∆yt−p+1 , 1, s1 , . . . , sm )0 is {K(p − 1) + 1 + m} × 1
Ψ = (Γ1 , . . . , Γp−1 , v, w1 , . . . , wm ) is K × {K(p − 1) + 1 + m}
e = β0 , ρ 0 is the (K + 1) × r matrix composed of the r cointegrating vectors and the r


β
trend coefficients ρ
In the restricted trend specification, m1 = K + 1, m2 = {K(p − 1) + 1 + m}, and there are
nparms = Kr + (K + 1)r + K{K(p − 1) + 1 + m} parameters in (3).

Unrestricted constant

An unrestricted constant in the VEC model in (2) is equivalent to setting δ = 0 in (1), which can
be written in the form of (3) by defining
Z1t = yt−1 is (K × 1)
0 0
Z2t = (∆yt−1 , . . . , ∆yt−p+1 , 1, s1 , . . . , sm )0 is {K(p − 1) + 1 + m} × 1;
Ψ = (Γ1 , . . . , Γp−1 , v, w1 , . . . , wm ) is K × {K(p − 1) + 1 + m}
e = β is the K × r matrix composed of the r cointegrating vectors
β
 18 vec — Vector error-correction models

In the unrestricted constant specification, m1 = K, m2 = {K(p − 1) + 1 + m}, and there are

nparms = Kr + Kr + K{K(p − 1) + 1 + m} parameters in (3).

Restricted constant

When there is a restricted constant in the VEC model in (2), it can be written in the form of (3)
by defining
0

0
Z1t = yt−1 , 1 is (K + 1) × 1
0 0
Z2t = (∆yt−1 , . . . , ∆yt−p+1 )0 is K(p − 1) × 1
Ψ = (Γ1 , . . . , Γp−1 ) is K × K(p − 1)
e = β0 , µ 0 is the (K + 1) × r matrix composed of the r cointegrating vectors and the r


β
constants in the cointegrating relations.
In the restricted trend specification, m1 = K + 1, m2 = K(p − 1), and there are nparms =
Kr + (K + 1)r + K{K(p − 1)} parameters in (3).

No trend

When there is no trend in the VEC model in (2), it can be written in the form of (3) by defining
Z1t = yt−1 is K × 1
0 0
Z2t = (∆yt−1 , . . . , ∆yt−p+1 )0 is K(p − 1) + m × 1
Ψ = (Γ1 , . . . , Γp−1 ) is K × K(p − 1)
e = β is K × r matrix of r cointegrating vectors
β
In the no-trend specification, m1 = K , m2 = K(p − 1), and there are nparms = Kr + Kr +
K{K(p − 1)} parameters in (3).

Estimation with Johansen identification

Not all the parameters in α and β e is K × r
e are identified. Consider the simple case in which β
and let Q be a nonsingular r × r matrix. Then
0 0 0 0
e = αQQ−1 β
αβ e Q −1 )0 = α̇β̇
e = αQ(β
0
Substituting α̇β̇ into the log likelihood in (5) for αβ e 0 would not change the value of the log
likelihood, so some a priori identification restrictions must be found to identify α and β.
e As discussed
in Johansen (1995, chap. 5 and 6) and Boswijk (1995), if the restrictions exactly identify or overidentify
β,
e the estimates of the unconstrained parameters in β e will be superconsistent, meaning that the estimates
of the free parameters in β will converge at a faster rate than estimates of the short-run parameters
e
in α and Γi . This allows the distribution of the estimator of the short-run parameters to be derived
conditional on the estimated β.e
Johansen (1995, chap. 6) has proposed a normalization method for use when theory does not
provide sufficient a priori restrictions to identify the cointegrating vector. This method has become
widely adopted by researchers. Johansen’s identification scheme is
e 0 = (Ir , β̆0 )
β (6)
where Ir is the r × r identity matrix and β̆ is a (m1 − r) × r matrix of identified parameters.
 vec — Vector error-correction models 19

Johansen’s identification method places r2 linearly independent constraints on the parameters in

β,
e thereby defining an exactly identified model. The total number of freely estimated parameters is
nparms − r2 = {K + m2 + (K + m1 − r)r}, and the degrees of freedom d is calculated as the
integer part of (nparms − r2 )/K .
When only the rank and the Johansen identification restrictions are placed on the model, we can
e α,
further manipulate the log likelihood in (5) to obtain analytic formulas for the parameters in β,
0
and Ω. For a given value of β, α and Ω can be found by regressing R0t on β R1t . This allows a
e e
further simplification of the problem in which
α(β
e ) = S01 β e 0 S11 β
e (β e )−1

Ω(βe ) = S00 − S01 β e 0 S11 β

e (β e 0 S10
e )−1 β
PT 0
Sij = (1/T ) t=1 Rit Rjt i, j ∈ {0, 1}
Johansen (1995) shows that by inserting these solutions into equation (5), β b is given by the r
eigenvectors v1 , . . . , vr corresponding to the r largest eigenvalues λ1 , . . . , λr that solve the generalized
eigenvalue problem
|λi S11 − S10 S−100 S01 | = 0 (7)
The eigenvectors corresponding to λ1 , . . . , λr that solve (7) are the unidentified parameter estimates.
To impose the identification restrictions in (6), we normalize the eigenvectors such that

λi S11 vi = S01 S−1

00 S01 vi (8)

and
1 if i = j
n
vi0 S11 vj = (9)
0 otherwise

At the optimum the log-likelihood function with the Johansen identification restrictions can be expressed
in terms of T, K, S00 , and the r largest eigenvalues

r
1 n X o
Lc = − T K ln(2π) + K + ln(|S00 |) + ln(1 − λ
bi )
2 i=1

where the λ
bi are the eigenvalues that solve (7), (8), and (9).
Using the normalized β,
b we can then obtain the estimates

b = S01 β
α b 0 S11 β
b (β b )−1 (10)

and
b = S00 − α
Ω bβb 0 S10

Let βb y be a K × r matrix that contains the estimates of the parameters in β in (1). β

b y differs
from βb in that any trend parameter estimates are omitted from β.b We can then use β b y to obtain
predicted values for the r nondemeaned cointegrating equations

E b 0y yt
et = β
b
20 vec — Vector error-correction models

The r series in Ee t are called the predicted, nondemeaned cointegrating equations because they still
b
contain the terms µ and ρ. We want to work with the predicted, demeaned cointegrating equations.
Thus we need estimates of µ and ρ. In the trend(rconstant) specification, the algorithm directly
produces the estimator µ b . Similarly, in the trend(rtrend) specification, the algorithm directly
produces the estimator ρb. In the remaining cases, to back out estimates of µ and ρ, we need estimates
of v and δ, which we can obtain by estimating the parameters of the following VAR model:

X p−1
∆yt = αE
e t−1 + Γi ∆yt−i + v + δt + w1 s1 + · · · + wm sm + t (11)
b
i=1

Depending on the trend specification, we use α

b to back out the estimates of

α0 α
b = (b
µ b 0v
b )−1 α b (12)

α0 α
b = (b
ρ b 0b
b )−1 α δ (13)
if they are not already in β
b and are included in the trend specification.
We then augment β
b y to
b 0f = (β
β b 0y , µ
b, ρ
b)

where the estimates of µb and ρb are either obtained from β

b or backed out using (12) and (13). We
b f to obtain the r predicted, demeaned cointegrating equations, E
next use β b t , via

E b 0f (yt0 , 1, t)0
bt = β

We last obtain estimates of all the short-run parameters from the VAR model:
p−1
X
∆yt = αE
b t−1 + Γi ∆yt−i + γ + τt + w1 s1 + · · · + wm sm + t (14)
i=1

Because the estimator β b f converges in probability to its true value at a rate faster than T −1/2 , we
can take our estimated E b t−1 as given data in (14). This allows us to estimate the variance–covariance
(VCE) matrix of the estimates of the parameters in (14) by using the standard VAR VCE estimator.
Equation (11) can be used to obtain consistent estimates of all the parameters and of the VCE of all
the parameters, except v and δ. The standard VAR VCE of v b and bδ is incorrect because these estimates
converge at a faster rate. This is why it is important to use the predicted, demeaned cointegrating
equations, Eb t−1 , when estimating the short-run parameters and trend terms. In keeping with the
cointegration literature, vec makes a small-sample adjustment to the VCE estimator so that the divisor
is (T − d) instead of T , where d represents the degrees of freedom of the model. d is calculated as
the integer part of nparms /K, where nparms is the total number of freely estimated parameters in
the model.
In the trend(rconstant) specification, the estimation procedure directly estimates µ. For
trend(constant), trend(rtrend), and trend(trend), the estimates of µ are backed out us-
ing (12). In the trend(rtrend) specification, the estimation procedure directly estimates ρ. In the
trend(trend) specification, the estimates of ρ are backed out using (13). Because the elements of
the estimated VCE are readily available only when the estimates are obtained directly, when the trend
parameter estimates are backed out, their elements in the VCE for βb f are missing.
 vec — Vector error-correction models 21

Under the Johansen identification restrictions, vec obtains β,

b the estimates of the parameters in
0
the r × m1 matrix β in (5). The VCE of vec(β) is rm1 × rm1 . Per Johansen (1995), the asymptotic
e b
distribution of β
b is mixed Gaussian, and its VCE is consistently estimated by
 
1  0 −1 −1
(Ir ⊗ HJ ) (bαΩ α b ) ⊗ (H0J S11 HJ ) (Ir ⊗ HJ )0 (15)
T −d
where HJ is the m1 × (m1 − r) matrix given by HJ = (00r×(m1 −r) , Im1 −r )0 . The VCE reported in
e(V beta) is the estimated VCE in (15) augmented with missing values to account for any backed-out
estimates of µ or ρ.
The parameter estimates α b , using (10) or from the VAR
b can be found either as a function of β
model in (14). The estimated VCE model of αb reported in e(V alpha) is given by
1 b ⊗Σ
Ω bB
(T − d)
0
b B = (β
where Σ b )−1 .
b S11 β
As we would expect, the estimator of Π = αβ0 is

Π
b =α
bβb0
and its estimated VCE is given by
1 b ⊗ (β b0)
Ω bΣbBβ
(T − d)
The moving-average impact matrix C is estimated by
C
b =β
b ⊥ (b
α⊥ Γ
bβ b 0⊥
b ⊥ )−1 α

where βb ⊥ is the orthogonal complement of β by , α

b ⊥ is the orthogonal complement of α b , and
Pp−1
Γ = IK − i=1 Γi . The orthogonal complement of a K × r matrix Q that has rank r is a matrix
b
Q⊥ of rank K − r, such that Q0 Q⊥ = 0. Although this operation is not uniquely defined, the results
used by vec do not depend on the method of obtaining the orthogonal complement. vec uses the
following method: the orthogonal complement of Q is given by the r eigenvectors with the highest
eigenvalues from the matrix Q0 (Q0 Q)−1 Q0 .
Per Johansen (1995, chap. 13) and Drukker (2004), the VCE of C
b is estimated by
T − d b b b0
Sq Vb
ν Sq (16)
T
where
S b ⊗b
bq = C ξ

(b
ξ1 , b
ξ2 ) if p > 1
ξ=
b
ξ1
b if p = 1
ξ1 = (C
b b0 − IK )α
b 0Γ
α=α α0 α
b (b b )−1
ξ2 = ιp−1 ⊗ C
b b
ιp−1 is a (p − 1) × 1 vector of ones
V ν = (b
b is the estimated VCE of b
bν α, Γ
b1 , . . . Γ
bp−1 )
 22 vec — Vector error-correction models

Estimation with constraints: β identified

vec can also fit models in which the adjustment parameters are subject to homogeneous linear
constraints and the cointegrating vectors are subject to general linear restrictions. Mathematically,
vec allows for constraints of the form

R0α vec(α) = 0 (17)

where Rα is a known Kr × nα constraint matrix, and

R0evec(β
e) = b (18)
β

where Rβe is a known m1 r × nβ constraint matrix and b is a known nβ × 1 vector of constants.

Although (17) and (18) are intuitive, they can be rewritten in a form to facilitate computation.
Specifically, (17) can be written as
vec(α0 ) = Ga (19)
where G is Kr × nα and a is nα × 1. Equation (18) can be rewritten as

vec(β
e ) = Hb + h0 (20)

where H is a known n1 r × nβ matrix, b is an nβ × 1 matrix of parameters, and h0 is a known

n1 r × 1 matrix. See [P] makecns for a discussion of the different ways of specifying the constraints.
When constraints are specified via the aconstraints() and bconstraints() options, the
Boswijk (1995) rank method determines whether the parameters in β e are underidentified, exactly
identified, or overidentified.
Boswijk (1995) uses the Rothenberg (1971) method to determine whether the parameters in β e are
e are exactly identified if ρβ = r2 , and the parameters in β
identified. Thus the parameters in β e are
overidentified if ρβ > r2 , where
n o
ρβ = rank Rβe(Ir ⊗ β̈)

e The computed ρβ is stored in e(beta icnt).

and β̈ is a full-rank matrix with the same dimensions as β.
e is given by ρjacob , where
Similarly, the number of freely estimated parameters in α and β
n o
α ⊗ Im1 )H, (IK ⊗ β
ρjacob = rank (b b )G

Using ρjacob , we can calculate several other parameter counts of interest. In particular, the degrees of
freedom of the overidentifying test are given by (K + m1 − r)r − ρjacob , and the number of freely
estimated parameters in the model is nparms = Km2 + ρjacob .
Although the problem of maximizing the log-likelihood function in (4), subject to the constraints in
(17) and (18), could be handled by the algorithms in [R] ml, the switching algorithm of Boswijk (1995)
has proven to be more convergent. For this reason, vec uses the Boswijk (1995) switching algorithm
to perform the optimization.
 vec — Vector error-correction models 23

Given starting values (b

b0, a
b0 , Ω
b 0 ), the algorithm iteratively updates the estimates until convergence
is achieved, as follows:
b j is constructed from (19) and a
α bj
b j is constructed from (20) and b
β bj

b b −1
α0j Ω
b j+1 = {H0 (b
j αb j ⊗ S11 )H}−1 H0 (b b −1
αj Ω j ⊗ S11 ){vec(P) − (b
b αj ⊗ InZ1 )h0 }

a b −1
bj+1 = {G(Ω j ⊗ βj S11 βj )G}
b b −1 0 b −1
G (Ωj ⊗ β
b j S11 )vec(P)
b

b j+1 = S00 − S01 β

Ω b 0j − α
bj α b 0j S10 + α
bjβ b 0j S11 β
bjβ b 0j
bj α

The estimated VCE of β

b is given by

1
H{H0 (W ⊗ S11 )H}−1 H0
(T − d)

−1
where W is α b 0Ω
b αb . As in the case without constraints, the estimated VCE of αb can be obtained
either from the VCE of the short-run parameters, as described below, or via the formula
" #
1 n −1 0
o−1
0 b
b = (T − d) G G Ω ⊗ (β S11 β)G
Vbα b b G0

Boswijk (1995) notes that, as long as the parameters of the cointegrating equations are exactly
identified or overidentified, the constrained ML estimator produces superconsistent estimates of β.
e
This implies that the method of estimating the short-run parameters described above applies in the
presence of constraints, as well, albeit with a caveat: when there are constraints placed on α, the
VAR models must be estimated subject to these constraints.

With these estimates and the estimated VCE of the short-run parameter matrix V
b , Drukker (2004)
bν
shows that the estimated VCE for Π is given by
b

b ⊗ IK )Vb (β 0
(β b ⊗ IK )
b
α

Drukker (2004) also shows that the estimated VCE of C b can be obtained from (16) with the extension
that Vb
b
ν is the estimated VCE of νb that takes into account any constraints on α
b.
 24 vec — Vector error-correction models

Estimation with constraints: β not identified

When the parameters in β are not identified, only the parameters in Π = αβ and C are identified.
The estimates of Π and C would not change if more identification restrictions were imposed to
achieve exact identification. Thus the VCE matrices for Π b and Cb can be derived as if the model
exactly identified β.

Formulas for the information criteria

The AIC, BIC, and HQIC are calculated according to their standard definitions, which include the
constant term from the log likelihood; that is,
 
L 2nparms
AIC = − 2 +
T T
 
L ln(T )
BIC = − 2 + nparms
T T
  
L 2ln ln(T )
HQIC = − 2 + nparms
T T

where nparms is the total number of parameters in the model and L is the value of the log likelihood
at the optimum.

Formulas for predict

xb, residuals and stdp are standard and are documented in [R] predict. ce causes predict to
compute E
bt = β
b f yt for the requested cointegrating equation.

levels causes predict to compute the predictions for the levels of the data. Let ybtd be the
predicted value of ∆yt . Because the computations are performed for a given equation, yt is a scalar.
Using ybtd , we can predict the level by ybt = ybtd + yt−1 .
Because the residuals from the VEC model for the differences and the residuals from the corresponding
VAR model in levels are identical, there is no need for an option for predicting the residuals in levels.

References
Anderson, T. W. 1951. Estimating linear restrictions on regression coefficients for multivariate normal distributions.
Annals of Mathematical Statistics 22: 327–351. https://doi.org/10.1214/aoms/1177729580.
Baum, C. F., and S. Hurn. 2021. Environmental Econometrics Using Stata. College Station, TX: Stata Press.
Becketti, S. 2020. Introduction to Time Series Using Stata. Rev. ed. College Station, TX: Stata Press.
Boswijk, H. P. 1995. Identifiability of cointegrated systems. Discussion Paper #95-78, Tinbergen Institute.
http://hdl.handle.net/11245/1.112017.
Boswijk, H. P., and J. A. Doornik. 2004. Identifying, estimating and testing restricted cointegrating systems: An
overview. Statistica Neerlandica 58: 440–465. https://doi.org/10.1111/j.1467-9574.2004.00270.x.
Box-Steffensmeier, J. M., J. R. Freeman, M. P. Hitt, and J. C. W. Pevehouse. 2014. Time Series Analysis for the
Social Sciences. New York: Cambridge University Press.
Drukker, D. M. 2004. Some further results on estimation and inference in the presence of constraints on alpha in a
cointegrating VECM. Working paper, StataCorp.
 vec — Vector error-correction models 25

Du, K. 2017. Econometric convergence test and club clustering using Stata. Stata Journal 17: 882–900.
Engle, R. F., and C. W. J. Granger. 1987. Co-integration and error correction: Representation, estimation, and testing.
Econometrica 55: 251–276. https://doi.org/10.2307/1913236.
Hamilton, J. D. 1994. Time Series Analysis. Princeton, NJ: Princeton University Press.
Johansen, S. 1988. Statistical analysis of cointegration vectors. Journal of Economic Dynamics and Control 12:
231–254. https://doi.org/10.1016/0165-1889(88)90041-3.
. 1991. Estimation and hypothesis testing of cointegration vectors in Gaussian vector autoregressive models.
Econometrica 59: 1551–1580. https://doi.org/10.2307/2938278.
. 1995. Likelihood-Based Inference in Cointegrated Vector Autoregressive Models. Oxford: Oxford University
Press.
Jordan, S., and A. Q. Philips. 2018. Cointegration testing and dynamic simulations of autoregressive distributed lag
models. Stata Journal 18: 902–923.
Maddala, G. S., and I.-M. Kim. 1998. Unit Roots, Cointegration, and Structural Change. Cambridge: Cambridge
University Press.
Park, J. Y., and P. C. B. Phillips. 1988. Statistical inference in regressions with integrated processes: Part I. Econometric
Theory 4: 468–497. https://doi.org/10.1017/S0266466600013402.
. 1989. Statistical inference in regressions with integrated processes: Part II. Econometric Theory 5: 95–131.
https://doi.org/10.1017/S0266466600012287.
Phillips, P. C. B. 1986. Understanding spurious regressions in econometrics. Journal of Econometrics 33: 311–340.
https://doi.org/10.1016/0304-4076(86)90001-1.
Phillips, P. C. B., and S. N. Durlauf. 1986. Multiple time series regressions with integrated processes. Review of
Economic Studies 53: 473–495. https://doi.org/10.2307/2297602.
Pickup, M. 2015. Introduction to Time Series Analysis. Thousand Oaks, CA: Sage.
Rothenberg, T. J. 1971. Identification in parametric models. Econometrica 39: 577–591. https://doi.org/10.2307/1913267.
Sims, C. A., J. H. Stock, and M. W. Watson. 1990. Inference in linear time series models with some unit roots.
Econometrica 58: 113–144. https://doi.org/10.2307/2938337.
Stock, J. H. 1987. Asymptotic properties of least squares estimators of cointegrating vectors. Econometrica 55:
1035–1056. https://doi.org/10.2307/1911260.
Stock, J. H., and M. W. Watson. 1988. Testing for common trends. Journal of the American Statistical Association
83: 1097–1107. https://doi.org/10.1080/01621459.1988.10478707.
Watson, M. W. 1994. Vector autoregressions and cointegration. In Vol. 4 of Handbook of Econometrics, ed. R. F.
Engle and D. L. McFadden. Amsterdam: Elsevier. https://doi.org/10.1016/S1573-4412(05)80016-9.

Also see
[TS] vec postestimation — Postestimation tools for vec
[TS] tsset — Declare data to be time-series data
[TS] var — Vector autoregressive models+
[TS] var svar — Structural vector autoregressive models
[TS] vec intro — Introduction to vector error-correction models
[U] 20 Estimation and postestimation commands
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. StataNow and NetCourseNow are trademarks of StataCorp
LLC. 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.
For suggested citations, see the FAQ on citing Stata documentation.