COINTEGRATION AND ERROR CORRECTION MODELS

(Enders, 2014: Chapter 4 & 6)

INTRODUCTION

Error Correction Mechanism

Non-stationary time series give rise to spurious regression. Traditional approaches to

handling non-stationary time series:

• Detrending, but:-not possible with stochastic trends (I(1), I(2), …)

• Differencing, but:- sensitive to short-term noise components; &

biased estimates when series have long-run equilibrium
relationship  error correction mechanism

With differencing yt = xt + t becomes yt = xt + t.

* But t = t - t-1 and t and t-1 are not independent when a long-run equilibrium
between y and x exists.

* Future y‟s cannot merely „grow along‟ with x (yt+i = xt+i) when an equilibrium
relationship in levels exists. The gap between y and x must be closed over time: error
correction mechanism. Engle/Granger (1987)

Cointegration

Cointegration can be viewed as the statistical expression of the nature of long-run

equilibrium relationships. If y and x are linked by some long-run relationship, from
which they can deviate in the short run but must return to in the long run, residuals will
be stationary. If variables diverge without bound (i.e. non-stationary residuals) we must
assume no equilibrium relationship exists

If a linear combination of I (1) variables is a stationary process of I (0), then the

variables are said to be cointegrated. The concept of cointegration links relationships
between integrated processes and the concept of (steady state) equilibrium (Mills,
1990). It was originally introduced by Granger (1981) and extended by Granger and
Weiss (1983) and Engle and Granger (1987). Because the type of model to be
1
estimated might depend on whether a “dependent” variable may be cointegrated with
an “independent” variable, it is important to test whether two or more variable are
cointegrated.

In general, a linear combination of two or more time series will be non-stationary if one
or more of them is non-stationary, and the degree of integration of the combination will
be equal to that of the most highly integrated individual series. Hence, for example, a
linear combination of an I(1) series and an I(0) series will be I(1), that of two I(1)
series will also be I(1), and that of an I(1) series and an I(2) series will be I(2). We can
consider I (0) and I (1) processes.
NB: If economic series are non stationary and their linear combination is stationary,
then the variables are said to be co- integrated. Economic theory always suggests that
certain group or pair of variables are linked by a long run relationship. The long run
equilibrium relationship is referred to as cointegration. Cointegration implies that the
variables may drift away from each other in the short run but may not divert from each
other in the long run-long run co-movements.

• The deviation from long-run equilibrium is called the equilibrium error (et).
• Cointegration implies that equilibrium error process is stationary.
• In econometrics, “equilibrium” means long run relationship among “non-
stationary” variables.
Cointegration always implies an error-correction model (Granger representation
theorem) ECM is superior to modeling integrated data in first-differences or in levels
(Engle/Granger, 1987).
• For example, simple dynamic (short-run) model of y and x in levels
yt = 0 + 1 yt-1 + 0 xt + 1 xt-1 + t
• (1) Long-run or steady-state solution (yt=yt-1, xt=xt-1) is
• yt =  0 +  1 x t with 0 = 0/(1 – 1), 1 = (0+1)/(1 – 1)
• (2) Alternatively, rearranging dynamic model yields
• yt = 0 xt – (1 – 1) [yt-1 – 0 – 1 xt-1] + t

2
 * y, x, [yt-1 – 0 – 1 xt-1] are all stationary variables
* clear long-run component [y – 0 – 1 x] if cointegration exists
* (1- 1) measures speed of adjustment to long-run equilibrium

General vector-error-correction model (VECM)

yt = 1(L) yt-1 + 1(L) xt-1 + 1 zt-1 + 1t
xt = 2(L) yt-1 + 2(L) xt-1 + 2 zt-1 + 2t
y, x represent vectors of possibly more than 1 variables i.e. x = [x1, x2, … xk]
* (short run) dynamic structure is captured by the difference terms and ECM.
* Error-correction term zt-1 captures the levels (long run) information.
* Appropriate number of lags (L) of the variables is unknown, but can be determined by
examining the data.

TESTING COINTEGRATION/COINTEGRATION APPROACHES

SINGLE EQUATION APPROACHES
A)Engle Granger Approach
To test for cointegration between two variables, Engle and Granger (1987)
recommended the following steps;
1) Step 1: Pre-test the variables of their order of integration. By definition,
cointegration necessitates that the two variables be integrated of the same order
(Differenced). As a prerequisite, it must first be determined whether Xt and Yt
are I(0) or I(1) processes, that is we check whether Xt and Yt contain unit roots.
2) Step 2: If the variables are not stationary in level but are I(1), then they may be
cointegrated. If all are integrated of different order, then they cannot be
cointegrated. However, one may want to determine if they are multicointegrated.
If the variables are I(1), then we can estimate the long run equilibrium
relationship („cointegrating regression‟).
Yt = β0 + β1Xt + et (a)

3
In order to determine if the variables are cointegrated, we then test if the residual from

equation (a) ( êt ) appears to be I (0) or not using critical values for Engle-Granger

cointegration test. If the residual is stationary, then we conclude that Xt and Yt are
cointegrated.

êt  êt 1  ut (b)

NB:

 Since êt sequence is a residual from a regression equation, there is no need to

include an intercept term; the parameter of interest is  . If we cannot reject the

hypothesis ρ = 0, then the series contains a unit root. Hence, the variables are not
cointegrated.
 In most applied studies, it is not possible to use Dickey Fuller tables themselves. The

problem is êt series is generated from a regression equation and the actual error is

not known to the researcher, only the estimate. The methodology of fitting the
regression in equation (a) selects values of the beta parameters that minimize the
sum of squared residuals. Since the residual variance is made as small as possible,
the procedure is prejudiced toward finding a stationary error process in (26a). Hence
the test statistic used to test the magnitude of  must reflect this. On estimating

the cointegrating vector, it is advisable then to use the Engle granger cointegration
test tables (see Enders 2004, pg 441). If calculated value is less than the critical,
reject the null hypothesis of unit root and conclude the parameters are cointegrated.
 If residuals for (b) do not appear to be white noise/ diagnostic tests indicate that the
(et) in equation (b) exhibits serial correlation, an augmented form of the test can be
used.
k
êt  êt 1    i êt  i  ut (c)
i 1

4
 Where k is the pre-selected order of lags for the white noise residuals. The test
statistic is the t-statistic for  . If we reject null hypothesis that ρ = 0, then the

series is stationary and the variables are cointegrated.

3) Step 3: If Yt and Xt are both I (1) and cointegrated, then they have a long-run
relationship.If the variables are co – integrated, There exists an error correction
representation. Then an ECM may be used to model their relationships. ECM
combines both the short run and the long run characteristics of the variables/
models
The Error Correction Model is of the form;
ΔYt = α1 + αy(Yt-1 – β0 – β1Xt-1) + ∑ α11(i) ΔYt-i + ∑ α12(i) ΔXt-i + vyt (d)
We should notice that from equation (a)
et = Yt - β0 - β1Xt ~ I(0) (e)
so that et-1 = Yt-1 - β0 - β1Xt-1 is the term in the brackets.
et-1 represents the extent of disequilibrium (i.e departure from equilibrium in the
previous period). αy tells us the speed of adjustment towards the equilibrium. If it is
100 %, it tells us that the long-run adjustments are instantaneous. It means that β0 and
β1 are exact long run variables. If it is 55 %, it means 55 % of the adjustment takes
place in the first year.
Basic ECM Cointegration Test
ECM test equation: yt = 1(L) yt-1 + 1(L) xt + 1 zt-1 + 1t,
zt (= t) = yt –  –  xt
* Estimate using OLS
*test significance of ECM variable zt-1: i.e. test 1<0
Standard t-test allowed as approx. using normal distribution (Banerjee et al.,
1993)
Even though the Engle-Granger approach has attractive properties, it does have a
number of important weaknesses. In particular:

• All X variables are assumed exogenous.

You can do the cointegrating regression in different ways (LHS, RHS variables)!
Different results for each alternative. Should we test all combinations?

5
 • No statistical tests possible on cointegration vector (long-run model) coefficients
because standard errors are unreliable.
• Long-run model estimate suffers from small-sample bias.
• Rules out multiple cointegration vectors between more than 2 variables.
• Step-wise procedure of testing implies the compounding of errors.
• Univariate cointegration test with residuals imposes possibly invalid restrictions
on short run behavior

Difference between Granger Causality with the structural model – cause and effect
relationship (related with theory). Granger causality has nothing to do with cause and
effect relationship

Eviews Practicals For Engle Granger Using Tsa.Xls, Data Analysis Original
STEPS: Estimate the long-run equation first
Make residuals: Prog/make residual series/
Test for stationarity of the residuals using Engle granger
cointegration test tables
Ho: ‫׀‬ρ‫ = ׀‬0 (non stationarity), H1: ‫׀‬ρ‫ < ׀‬0, If ‫׀‬Calculated‫׀ < ׀‬tabulated‫ ׀‬,
reject Ho.
Estimation of the error correction model
 Use OLS on static regression in the levels of variables
 This is Engle- Granger two step procedure
 All estimates of this procedure are consistent.

B) Fully Modified OLS approach (FMOLS) of Phillips and Hansen (1990)

Fully modified least squares (FM-OLS) regression was originally designed in work by
Phillips and Hansen (1990) to provide optimal estimates of cointegrating regressions.
The method modifies least squares to account for serial correlation effects and for the
endogeneity in the regressors that results from the existence of a cointegrating
relationship.

6
d) Dynamic OLS-Stock and Watson, (1993)
This method improves on OLS by coping with small sample and dynamic sources of
bias. The Johanson method is exposed to the problem that parameter estimates I one
equation are affected by any misspecification in other equations. The dynamic-OLS by
contrast is a single equation which corrects for regressor endogeneity by inclusion of
leads an lags of first differences of the regressors, and for serially correlated errors by a
GLS procedure. Nevertheless, it has the same asymptotic optimality properties as the
Johanson distributionT
Steps:
 Test the variables to see if they are non-stationary (e.g. use ADF)
 Given that 2 or more are non-stationary, test whether variables are cointegrated
Example: Modeling demand for energy:
Aggregate energy consumption (Q), Real price of energy(P), Real income(GDP(Y)),
Total area constructed in square metres(A).

Where

And m, n, and l are the lengths of leads(futures) and lags(past) of the regressors.
Suppose Q has been found to be I(1) and at least some of the right hand side variables
I(1) or I(), then DOLS estimates are obtained by regression analysis of the equation.

7
Estimating the general model:
n n k 2
yt   0    i xi ,t    i, j xi ,t  j   t
i 1 i 1 j   k 1

OLS estimates of long-run coefficients  are superconsistent

* Correct standard errors for serial correlation in : use Newey-West, HAC-errors or
DGLS (Cochrane-Orcutt)
* Normal t-tests allowed on cointegration model yt = 0 + i=1n i xi,t. for example to
determine which variables are important
Step 2: Test for cointegration as in Engle-Grander, using one of residual based
cointegration tests (see earlier)
zt = yt - 0 - i=1n i xi,t
* Univariate residual unit root test
* ECM cointegration test
Notes:
* Some studies employ first EG-OLS to test for cointegration and then second apply
DOLS to estimate and test final cointegration coefficients.
* Stock-Watson (1988) developed a special cointegration test. Other studies use the
Johansen test for cointegration first.
Diagnostic tests:
Since a parametric econometric model is described by its parameters, model stability
can be equivalent to parameter stability. The Cumulative sum of recursive residual
(CUSM) and the Cumulative sum of squared recursive residual (CUSUMSQ) tests ate
conducted to investigate the model parameters stability. In general if CUSM or
CUSUMSQ move out of the critical lines of 5% significance level, the null hypothesis will
be rejected. Meaning the model is unstable. Other tests include serial correlation,
Functional form, Normality and heteroskedasticity.

D) ARDL Approach to Cointegration Analysis

ARDL model was introduced by Pesaran, shin & schmidt (2001) in order to incorporate
I(0) and I(1) variables in same estimation. If your variables are stationary I(0) then OLS

8
is appropriate. If all are non stationary-I(1) then it is advisable to do VECM (Johanson
Approach) as it is much simple model.
We cannot estimate conventional OLS on the variables if any one of them or all of them
are I(1) because these variable will not behave like constants which is required in OLS.
most of then are changing in time so OLS will mistakenly show high t-values and
significant results but in reality it would be inflated because of common time
component, in econometric it is called spurious results where R square of the model
becomes higher than the Durban Watson Statistic. So we move to a new set of models
which can work on I(1) variables.
Testing Procedure
Pesaran, Shin and Smith (PSS 2001) developed a new approach to co-integration
testing which is applicable irrespective of whether the regressor variables are I(0), I(1)
or mutually co-integrated. The starting point of their test is a data generating process
represented by a general VAR of order p which is rewritten in vector ECM form involving
a vector z of variables.
They focus on the conditional modeling of the dependent scalar variable y. To that end,
the vector z is partitioned into the scalar y and vector x of dependent variables. Under
the assumption that there is no feedback from y to x, the model can be written as the
following conditional ECM model for ∆yt
= ∑ ∑

Where, w is a set of deterministic variables like the constant term, trend, seasonal
dummies, etc.. c is a vector of coefficients of deterministic variables u t is the residual
term. To test the absence of a level relationship between y and x, the approach uses a
Wald or F-statistic to test for the joint hypothesis that all coefficients of all (lagged)
levels in the ECM equation are zero.
The null hypothesis of no co-integration is H0: = = 0 and the alternative
hypothesis of co-integration is H1: ≠ ≠0

9
The F test has a non-standard distribution which depends upon: (a) whether variables
included in the ARDL model are I(0) or I(1); (b) the number of regressors, (c) whether
the ARDL model contains an intercept and/or a trend; and (c) the sample size.

In the case of co-integration based on the bounds test, the Granger causality tests
should be done under vector error correction model (VECM) when the variables under
consideration are co-integrated. By doing so, the short-run deviations of series from
their long-run equilibrium path are also captured by including an error correction term
(See also Narayan and Smyth, 2004). Therefore, error correction models of co-
integration can be specified as follows:
=
where
=∑

ECTt-1 is the lagged error correction term derived from the long-run co-integration
model. According to the VECM for causality tests, having statistically significant F and t
ratios for ECTt-1 would be enough condition to have causation from X to Y and from Y to
X respectively.
PSS distinguishes five cases according to how the deterministic components are
specified:
 no intercepts, no trends;
 restricted intercepts, no trends;
 unrestricted intercepts, no trends;
 unrestricted intercepts, restricted trends,
 unrestricted intercepts, unrestricted trends.
The resulting conditional ECM‟s may be interpreted as autoregressive of orders (p, p,…
p), i.e. ARDL (p, p,…p) models. PSS publishes tabulated asymptotic critical value
bounds for the F-statistic for all 5 conditional ECM models. The two sets of critical
values (CV) provided in PSS are generated for sample sizes of 500 and 1000
observations and 20,000 and 40,000 replications respectively. Narayan(2005) argues

10
 that exiting CVs, because they arebased on large sample sizes, cannot be used for small
sample sizes. critical values for sample sizes ranging from 30–80 observations are
calculated in Narayan(2005)
 If the computed F-statistic from exclusion of levels in the conditional ECS‟s fall
outside the critical value bounds, the test allows a conclusive inference without
needing to know the integration/cointegration status of the underlying
regressors.
 If the F-statistic falls inside the bounds, inference is inconclusive and knowledge
of the order of integration of the underlying variables is required before
conclusive inferences can be made.
 If the computed F-statistic lies below the 0.05 lower bound, the hypothesis that
there is no level relationship is not rejected at the 5 percent level.
 If the statistic falls within the 0.05 bounds, the test is inconclusive
 When the F-statistic lies above the 0.05 upper bound, the hypothesis of no level
relationship is conclusively rejected.
In addition to the F-test, PSS also tabulates asymptotic critical value bounds of the t-
statistic for testing the significance of the coefficient on the lagged dependent variable
in the conditional ECM. Concerning the use of the F and t-statistics, PSS suggests the
following procedure:
 Test H0 using the bounds procedure based on the Wald or F-statistic. If H0 is not
rejected, proceed no further. If H0 is rejected test the coefficient of the lagged
dependent variables using the bounds procedure based on the t-statistic. A large
value of t confirms the existence of a level relationship between y and x;
How to Estimate ARDL model
In order to run ARDL some preconditions needed to be checked
 None of the variable should be I(2) in normal conditions (ADF test)
 none of the variable should be I(2) in structural break
 Step 1 Check Optimal Lag order: First we need to check the lag order to see
what lag we use to the ADF test for each variable which is being used in the
model. This is done using VARSOC Table (Vector Auto Regressive Specification

11
 Order Criterion) which is available in STATA that can be quickly applied, in
EVIEWS you have to do it after VAR model and check the Lag length criterion
 Estimate ARDL

In this table you have to note the highlighted things first and for most you have to see
if the F-statistic (of 5.85) is higher than the upper bound at 95% (of 4.82) or
90% (of 3.87) any one will work.

12
  If it comes higher, then you can say that there is cointegration among the set of
(I(0) & I(1)) variables so we can assume that there can be at least long run or
short run relation among these variables.
 If this F statistic is not higher than any of the upper bound critical values for first
try to modify the lag order so that F statistic might go above the critical value of
upper bound if it is not possible see the guide in following passage.
 If the F statistic is higher than the lower bound but lower than the upper bound
then you must check all of your I(1) variables if there is any useless (to remove)
or you have missed any one (to add) .
 if F statistic is lower than the lower bound then all of your variables I(0) and
I(1) are not appropriate. They are not cointegrated-try to change the model add
more variables modify their specification if possible other wise you can report
that these variables do not relate with each other their relationship is spurious.
So if you found the F statistic larger than the upper bound like in the image above then
you have to see the diagnostics there are four provided at the end of the table (which
are auto-correlation, normality, specification and hetroskedasticity test) first is serial
correlation which is insignificant as per F version but significant at 5% in LM version so
we can assume that there is no auto-correlation at 1% or according to F version.
Similarly the functional form is insignificant (no issue); normality is insignificant (no
issue) and heteroskedasticity is insignificant (no issue) too hence there is no apparent
issue which with this model. So we can proceed to next step.
If you find any of the issue try to change the lag order increase sample or variable
specification in order to correct them. if your sample is higher than the 30 then you
can ignore the normality issue if it exists as per central limit theorem. There is another
information on the top of the table which is the lag order of the estimation which is to
be reported in the regression analysis.
NB: The ARDL estimation procedure is based on two basic steps. First, the existence of
a level relationship is tested using an unrestricted ARDL specification. Secondly, when
the existence of such a level relationship cannot be rejected, a more parsimonious
ARDL lag model is selected using information criteria to determine the optimal lag
13
orders of the independent variables in the ARDL equation. The second step involves the
estimation of the conditional ECM model from which the co-integration vector and the
short term dynamics can be obtained.

Non stationary Multivariate Linear Models

Multivariate VAR Model (More Than One Cointegrating Equations Expected)
Testing of the cointegrating vector
A) Johansen procedure:
The concepts that we have learned on unit roots, cointegration can also be extended to
the VAR models and the general multivariate linear models.
As noted earlier, Engle granger has some limitations:
1) The estimation of long run equilibrium regression requires that the researcher
places one variable on the left hand side and use the others as regressors. In
tests using more than one variables, there may be more than one cointegrating
relationship.
2) Since it is a two step approach, if you make an error in the first step, it will be
carried in second step
Ways to resolve such problems has been suggested. One such is the Johansen(1988)
and stock and Watson(1988) maximum likelihood estimators. They can estimate and
test in presence of multiple cointegrating vetors

Johansen procedure is nothing more than a multivariate generalization of the dickey

fuller test. As with the augmented Dickey–Fuller test, the multivariate model can also be
generalized to allow for a higher-order autoregressive process
=
= (a)
Putting the equation in a more useful form ,add and subtract on the right
hand side
=

14
 = ( )
Next add and subtract ( )
= ( )
( ) ( )
= ( )
=
( )
=
If = 0,

Lag = = = =
Just as in ADF, we can continue with this fashion, we obtain:
= ∑ (b)
=( ∑ ) and = ∑

= ∑
where 0 = an (n ⋅ 1) vector of intercept terms with elements i0

i = (n ⋅ n) coefficient matrices with elements jk(i)

= a matrix with elements jk such that one or more of the jk ≠0

t= an (n ⋅ 1) vector with elements it

Note that the disturbance terms are such that it may be correlated with jt. Let all
variables in xt be I(1). Now, if there is an error-correction representation of these
variables as in (a), there is necessarily a linear combination of the I(1) variables that is
stationary.
Alternatively:
VAR yt = A(L) yt-1 + t
A(L) = A1 + A2 L +… + AkLk-1 a series of coefficient matrices
for all the lagged variables t-1 to t-k

Rewritten in VECM
yt = (L) yt-1 +  yt-k + t

15
 i = – (1 – A1 –…– Ai), i = 1,…, k-1
 = – (1 – A1 –…– Ak)
or  =  ´ where
 = speeds of adjustment matrix, and
 is matrix of long-run coefficients(matrix of parameters which
determines the cointegrating relationship such that
´ yt-k represent the multiple cointegration relationships. It is interpreted
as long-run equilibrium relationships between the variables. Values of 
close to zero imply close convergence

The key feature to note in (b) is rank of the matrix ; the rank of is equal to the
number of independent cointegrating vectors.
Ranks: Three cases:
(a) Rank =0
If rank( ) = 0, the matrix is null and (b) is the usual VAR model in first differences.
If Rank = 0, all of are unit root processes. If each of are zero, it implies that
= implying =
Also means VAR is non stationary with no cointegrating equations. We estimate the
VAR in first difference

(b) Rank =
if is of rank n, the vector process is stationary. In intermediate cases, if rank( ) = 1,
there is a single cointegrating vector and the expression xt−1 is the error-correction
term
 No roots
 All variables are stationary – represents a convergent systems of difference
equations. Here VAR is estimated at levels

(c)
There are multiple cointegrating vectors.
the number of distinct cointegrating vectors can be obtained by checking the
significance of the characteristic roots of . The rank of a matrix is equal to the number
of its characteristic roots that differ from zero.

16
Suppose we obtained the matrix and ordered the n characteristic roots such that 𝜆1 >
𝜆2 > · · · > 𝜆n. If the variables in xt are not cointegrated, the rank of is zero and all of
these characteristic roots will equal zero. Since ln(1) = 0, each of the expressions ln(1
− 𝜆i) will equal zero if the variables are not cointegrated. Similarly, if the rank of is
unity, 0 < 𝜆1 < 1 so the first expression ln(1 − 𝜆1) will be negative and all the other 𝜆i
= 0 so that ln(1 − 𝜆2) = ln(1 − 𝜆3) = · · · = ln(1 − 𝜆n) = 0.
In practice, we can obtain only estimates of and its characteristic roots. The test for
the number of characteristic roots that are insignificantly different from unity can be
conducted using the following two test statistics:
The trace statistic is determined using the following formula:
n
λtrace (r)=  T  ln(1  i ) r = 0,1,2, …, n-1
i  r 1

T = number of observations, n number of x components,

λi =the estimated value of the characteristic root (is the ith eigenvalue), obtained from
the estimated matrix.
H0: The number of disntict cointegrating vectors is less than or equal to r
H1: > than r
H0: r = 0, H1: r >0; H0: r ≤ 1, H1: r > 1; H0: r ≤ 2, H1: r > 2;
Large value of λtrace rejects the null-hypothesis

The maximum eigenvalue statistic is determined using the following formula:

λmax(r, r+1) =  T ln(1  r 1 ) r = 0,1,2, …, n-2, n-1 H0: r = 0,

H0: The number of cointegrating vectors is r

H1: The number of cointegrating vectors is r+1
H1: r = 1; H0: r = 1, H1: r = 2; etc
Large value of λmax rejects the null-hypothesis
To make inferences regarding the number of cointegrating relationships, the trace and
maximum eigenvalue statistics are compared with the critical values tabulated in
Osterwald-Lenum (1992). The critical values are reproduced in Table E Enders 2004

17
If calculated > Critical, Reject H0

Sometimes the results may be conflicting but λmax has sharper alternative hypothesis
and usually preffered for trying to pin down the number of cointegrating vectors

Normalization
Each sequence can be written in error correction form. Eg, we can write as:
=
Normalizing with respect to
Set and to obtain
=
In the long run the will satisfy the relationship:
Hence the normalized cointegrating vector is: and the speed of
adjustment parameter is
Performing Johansen test procedure-Steps

Step 1: Pretest all variables to assess their order of integration. Plot the data to see if a
linear time trend is likely to be present in the data-generating process. In most
instances you will have variables that are integrated of the same order. In other cases,
you can check for multicointegration. The results of the test can be quite sensitive to
the lag length, so it is important to be careful. The most common procedure is to
estimate a vector autoregression using the undifferenced data. Then use the same lag-
length tests as in a traditional VAR. Begin with the longest lag length deemed
reasonable and test whether it can be shortened.
STEP 2: Given that 2 or more variables are I(1), consider whether they are
cointegrated.
Estimate the model and determine the rank of . Here, it suffices to say that OLS is not
appropriate because it is necessary to impose cross-equation restrictions on the
matrix. In most circumstances, you may choose to estimate the model in three forms:
(1) with all elements of A0 set equal to zero, (2) with a drift, or (3) with a constant term
in the cointegrating vector. Analyze the properties of the residuals of the estimated

18
model. Any evidence that the errors are not white noise usually means that lag lengths
are too short.
STEP 3: Analyze the normalized cointegrating vector(s) and speed of adjustment
coefficients.
STEP 4: Finally, innovation accounting and causality tests on the error-correction model
(if used-to be covered later) could help to identify a structural model and determine
whether the estimated model appears to be reasonable. Since the simulated data have
no economic meaning, innovation accounting is not performed.

Performing Johansen Test in Eviews/ Estimation of the Cointegrating Vector

To carry out the Johansen test,

 Select View/Cointegration Test… on the group or VAR toolbar. Note that

since this is a test for cointegration, this test is only valid when you are working
with series that are known to be non-stationary. You may wish first to apply unit
root tests to each series in the VAR (see Unit Root Tests).
 Provide information on the Cointegration Test dialog.

19
  Choose one of the options specifying the type of deterministic trends that are
present in the data. The first five options provide particular alternatives for
whether an intercept or trend term should be included in the specification of the
cointegrating equations. These five cases are nested from the most restrictive to
the least restrictive, given any particular cointegrating rank r:
 The sixth option runs the Johansen procedure under all five sets of assumptions.
A summary table of results is displayed with this option. To see the full output of
the test; select one of the other options.
NB. There are two dimensions you can vary while performing tests within this
framework. You can assume one of the five cases listed above, and carry out tests
for the cointegrating rank. Alternatively, you can fix the rank, and test which of the
five cases describes the data best. EViews provides you with the option of
summarizing all five cases, so you can look at all possible combinations of rank and
intercept-trend.

 The dialog also allows you to specify exogenous variables, such as seasonal
dummies, to include in the test VARs. The constant and linear trend should not
be listed in this dialog; these should be chosen from the specification options.
 Specify the lags of the test VAR as pairs of intervals. Note that in contrast to
some other statistical packages, the lags are specified as lags of the first
differenced terms, not in terms of the levels. To run a cointegration test with
zero lags in the levels, type 0 0 in the field.
The results give you the number of cointegrating equations for the integrated variables.
The estimated cointegrating relation is given under the normalized cointegrating
coefficients. Only look for the table(s) that has normalized cointegrating coefficients, in
which the coefficient of one of the six variables is normalized to one. There may be
more than one table with normalized coefficients. If the above-mentioned LR test
indicates one cointegrating equation, look at the first normalized coefficient table only.
If the test indicates two cointegrating equations, you need to also look at the second
normalized coefficient table, and so on. The cointegrating vector is not identified

20
unless we impose some arbitrary normalization. EViews adopts a normalization
such that the first r series in the vector are normalized to an identity matrix. The
number in parentheses under the estimated coefficients are the asymptotic standard
errors. Some of the normalized coefficients will be shown without standard errors. This
will be the case for coefficients that are normalized to 1 and for coefficients that are not
identified. A normalized coefficient table presents the estimate of the model
(cointegrating equation) with all variables taken to the left hand side.

The Vector Error Correction Model (VECM)

According to the Granger Representation theorem, when variables are cointegrated,
there must also be an error correction model (ECM) that describes the short-run
dynamics or adjustments of the cointegrated variables towards their equilibrium values.
ECM consists of one-period lagged cointegrating equation and the lagged first
differences of the endogenous variables. Using the Vector Autoregression (VAR)
method, we can estimate the ECM.

A vector error correction (VEC) model is a restricted VAR that has cointegration
restrictions built into the specification, so that it is designed for use with nonstationary
series that are known to be cointegrated. The VEC specification restricts the long-run
behavior of the endogenous variables to converge to their cointegrating relationships
while allowing a wide range of short-run dynamics. The cointegration term is known as
the error correction term since the deviation from long-run equilibrium is corrected
gradually through a series of partial short-run adjustments.

Estimation of Vector Error Correction Model

Before estimation of VECM model with associated cointegrating vector it is necessary to

select optimal lag length of initial VAR( see criterion explained earlier)
 Click the Estimate button in the VAR toolbar and choose the Vector Error
Correction specification. Or When the workfile is open, from the main EViews
menu bar, click on Quick/ Estimate VAR. A new window on Vector
21
 Autoregression appears. Under VAR specification, click on Vector Error
Correction
 provide the same information as for an unrestricted VAR specify exogenous
variables (this is optional). However, the specification of the exogenous
intercepts and trends should be chosen from the five models discussed above
(see Johansen test discussed). This choice should be the same as in the
cointegration test.
 It is important to note that the lag specification prompted by EViews refers to
lags of the first difference terms in the VEC. For example, 1 1 specifies a model
involving a regression of the first differences on one lag of the first difference.
 Specify the number of cointegrating equations in the VEC model. This number
should be determined from the cointegration test. The maximum number of
cointegrating equations is one less than the number of endogenous variables in
the VAR.
 Estimation of a VEC model is carried out in two steps. In the first step, the
cointegrating relations from the Johansen procedure as used in the cointegration
test is estimated. Then the error correction terms from the estimated
cointegrating relations is constructed and a VAR in first differences including the
error correction terms as regressors estimated/. In other words the first
difference of each endogenous variable is then regressed on a one period lag of
the cointegrating equation(s) and lagged first differences of all of the
endogenous variables in the system.
 The first table presents the estimates of the cointegrating equation , and the
second table presents the rest of the ECM. The first row in the second table
presents the estimates of the speed of adjustment coefficient for each variable ,
their standard errors and the t-statistics.
NB: Working with a VEC is analogous to working with a VAR(see the discussion in
Working with a VAR). Note that standard errors for the impulse response functions are
not available for the VEC

22
Weak Exogeneity

In a cointegrated system, if the variable does not correspond to the discrepancy from
the long run equilibrium relationship, it is weakly exogenous, Hence if the speed of
adjustment parameter αi is zero, the variable in question is weakly exogenous. Each
speed of adjustment coefficient measures the degree to which the variable in question
responds to the deviation from the long run equilibrium relationship. Coefficients that
are small in absolute value imply a slow adjustment. The practical importance is that a
weakly exogenous variable does not experience the type of feedback used in VAR.

Imposing Restrictions

Restrictions can be imposed on the cointegrating vector (elements of the β matrix) and/or on
the adjustment coefficients (elements of the α matrix). To impose restrictions in estimation,
open the test, select Vector Error Correction in the main VAR estimation dialog, then click on
the VEC Restrictions tab. You will enter your restrictions in the edit box that appears when
you check the Impose Restrictions box:

Restrictions on the Cointegrating Vector

To impose restrictions on the cointegrating vector β, you must refer to the (i,j)-th
element of the transpose of the β matrix by B(i,j). The i-th cointegrating relation has the
representation:
B(i,1)*y1 + B(i,2)*y2 + ... + B(i,k)*yk
where y1, y2, ... are the (lagged) endogenous variable. Then, if you want to impose the
restriction that the coefficient on y1 for the second cointegrating equation is 1, you
would type the following in the edit box:
B(2,1) = 1
You can impose multiple restrictions by separating each restriction with a comma on the
same line or typing each restriction on a separate line. For example, if you want to
impose the restriction that the coefficients on y1 for the first and second cointegrating
equations are 1, you would type:
B(1,1) = 1

23
B(2,1) = 1

Restrictions on the Adjustment Coefficients

To impose restrictions on the adjustment coefficients, you must refer to the (i,j)-th
elements of the α matrix by A(i,j). The error correction terms in the i-th VEC equation
will have the representation:
A(i,1)*CointEq1 + A(i,2)*CointEq2 + ... + A(i,r)*CointEqr
One restriction of particular interest is whether the i-th row of the α matrix is all zero. If
this is the case, then the i-th endogenous variable is said to be weakly exogenous with
respect to the β parameters. For example, if we assume that there is only one
cointegrating relation in the VEC, to test whether the second endogenous variable is
weakly exogenous with respect to β you would enter:
A(2,1) = 0
To impose multiple restrictions, you may either separate each restriction with a comma
on the same line or type each restriction on a separate line. For example, to test
whether the second endogenous variable is weakly exogenous with respect to β in a
VEC with two cointegrating relations, you can type:
A(2,1) = 0
A(2,2) = 0
You may also impose restrictions on both β and α. However, the restrictions on β and α
must be independent. for example,
A(1,1) = 0
B(1,1) = 1
Findings and Interpretations (Include all relevant computer printouts)

1. What are the orders of integration of the time series variables? Are they integrated of the
same order? …to be done on original data
2. Run a Johansen test on the I(1) variables
3. According to the likelihood ratio test of the Johansen test, how many cointegrating
equations are possible?

24
4. Write down the estimated KF model. Are the signs of estimated coefficients the same as
anticipated? Comment which coefficients are significant.
5. Write down the ECM model estimated by VAR. Interpret the short run and long run
equilibrium estimations. Do you get the same estimate of cointegrating equation?
6. Test weak exogeneity for each of the variables and interprete your results
7. Evaluate the model for serial correlation, normality and Heteroscedasticity

25