Applied Econometrics-Lecture notes

Sunil Paul

November 2, 2016

Sunil Paul

Cointegration

The components of the vector Xt = (X1t , X2t , ..., Xnt ) are said
to be cointegrated of order d, b, [ Xt CI (d, b) ] if
1. All components of Xt are integrated of order d
2. There exists a vector = (1 , 2 , ..., n ) such that the linear
combination Xt (= 1 X1t + 2 X2t + ... + n Xnt ) is integrated
of order (d b) where b > 0

The vector is called cointegrating vector.

Equilibrium theories involving nonstationary variables

combination of the variables that is stationary (e.g., money
demand, consumption function etc.)

Sunil Paul

An illustration.

The quantity theory of money is defined as:MV = PY

I

where M = stock of money, V =velocity,P = price level and

Y = output

In reality, M, P and Y are I (1) variables; hence the time series

on these variables are not mean reverting.

However, theory asserts that there exists a equilibrium relation

hence a linear combination of these non-stationary variables
may be stationary.

Sunil Paul

An illustration.

Let the lower case letters indicate the log of the variables.

Then the above equation can be written

as:0 + 1 mt 2 pt 3 yt = 0
I

where 0 is velocity and assumed to be a constant.

{0 , 1 , 2 , 3 } is the cointegrating vector.

In regression format:0 + 1 mt 2 pt 3 yt = et
I

where et measures the short-run deviation from the identity.

Sunil Paul

An illustration.

There exists cointegration provided et I (0); hence, the

deviation is temporary.

Also, the theory indicates that the1 = 1; 2 = 1; and

3 = 1

Cointegration vector is not unique,we can normalize the

{0 , 1 , 2 , 3 } with respect to any variable

Let us normalize the vector with respect to mt :

I

mt = a0 + a2 pt + a3 yt + et where ai =

Sunil Paul

i
1

An illustration.

It is possible that when there are n variables there could be

n 1 cointegrating vectors.

Suppose the monetary authority follows a feed-back rule that

whenever there is rise in price level the central bank reduces
money supply and vice versa.

If so, we may have another cointegrating relationship:

{0 , 1 , 2 , 0}; hence, we will have:


0 , 1 , 2 , 3
=
0 , 1 , 2 , 0

Sunil Paul

An illustration.

When there is disequilibrium (et 6= 0), some of the variables

must respond to restore equilibrium.

In that case change in each variable may be expressed a

function of disequilibrium et1 = mt1 a0 a2 pt1 a3 yt1
and lagged values of changes in other explanatory variables

Such representation is known as error correction

representation.

Sunil Paul

An illustration.
I

Error Correction Mechanism (ECM)

Pn
Pn
mt = m (et1 ) +
Pn i=1 11.i mti + i=1 12.i pti
+ P
i=1 13.i yti + mtP
pt = p (et1 ) +P ni=1 21.i mti + ni=1 22.i pti
+ Pni=1 23.i yti + ptP
yt = y (et1 ) +P ni=1 31.i mti + ni=1 32.i pti
+ ni=1 33.i yti + yt

The above model states that if there is any disequilibrium in

the previous period each variables may adjust to equilibrium.

The speed of adjustment is captured by coefficients

since et1 = mt1 a0 a2 pt1 a3 yt1 the

m < 0; p > 0 and y > 0

For a stable LR equilibrium |1|

Sunil Paul

An illustration.

If cointegration exits any one the adjustment. parameters

should be significant

If all s are zero there exits no cointegration

If m = 0; p = 0 and y is statistically significant, only

output responds to restore equilibrium and m and p are weakly
exogenous in the system.

The s are additional source for inferring causality.

For instance, m does not cause p if 21 (i) in the second

equation are zero and p = 0

Sunil Paul

Granger Causality in CI system

In a cointegrated system, Yt does not Granger cause Zt

I
I

if lagged values Yti do not enter the Zti equation and

The speed of adjustment parameter in the Zt equation is zero
(Zt ) should be weekly exogenous

Sunil Paul

Two stage Engle and Granger methodology

First stage
I
I

Pre- test the variables for the order of the variables

Estimate the levels regression and obtain long-run
coefficients(provided Y and X are I (1) )(Yt = b0 + b1 Xt + et )
From the regression form the P
residuals and test for
n
cointegration(b
et = ebt1 + i=1 b
eti + t ) where
H) : = 0 is a test for no cointegration
Usual Dickey Fuller distribution cannot be used to test unit
roots

Sunil Paul

Two stage Engle and Granger methodology

Second stage
I

if the null of no cointegration is rejected in the first stage

estimate the error correction model given as below using the
estimated residuals.
Pn
Pn
Yt = y ebt1 + Pi=1 11 (i)Yti + Pi=1 12 (i)Xti + yt
n
n
Xt = x ebt1 + i=1 21 (i)Yti + i=1 22 (i)Xti + xt
where ebt1 = Yt1 bb0 + bb1 Xt1

Third Stage
I

Conduct diagnostic checks on residuals for model adequacy

(Errors should be white noise)

Sunil Paul

E-G Cointegration test

E-G cointegration test is easy to implement

Single equation method ( Need to specify the relation first).

However cointegration is not unique hence the normalization

should not affect the results

i.e. if we have two variables the residuals from a regression of

X on Y and Y on X should give similar results

But if we have two or more variables it can produce

contradictory results

Sunil Paul

Johansen Cointegration test

Johansen Procedure is based on rank of a matrix and its

characteristic roots

It can be viewed as an multivariate extension of Dickey Fuller

unit root test

Hypothesis test in Johansen method is based on characteristic

roots and rank of matrix

In order to understand the Johansen test we need to first

understand the link between cointegration and rank of a matrix

Johansen test is carried out in a VAR framework

Sunil Paul

VAR in the presence of cointegration

If variables in Xt is C (1, 1), then there exits a error correction

mechanism:

VECM:
Xt = et1 + 1 Xt1 + 2 Xt2 + ... + p1 Xtp + t
I
I
I

et1 error correction

speed of adjustment parameters
i captures the short run adjustments

Note that all variables in ECM are I (0).

Sunil Paul

VAR in the presence of cointegration

I
I

Substituting et1 = 0 Xt1 into the above equation and

replacing 0 = we get:
P
Xt = Xt1 + p1
i=1 i Xti + t
This can be viewed as a multivariate extension of ADF test
(see the similarity with an ADF equation)
P
Rank of the = I pi=1 Ai will indicate the number of
cointegrating vectors

Sunil Paul

Derivation of the VECM for VAR

Consider a system of eqns where Y represents a vector of

variables with number of variables(k)=n and order(p)=4
Yt = A1 Yt1 + A2 Yt2 + A3 Yt3 + A4 Yt4 + t

Add and subtract A4 Yt3 from RHS

Yt = A1 Yt1 + A2 Yt2 + A3 Yt3 + A4 Yt4 + A4 Yt3 A4 Yt3 + t
Yt = A1 Yt1 + A2 Yt2 + (A3 + A4 )Yt3 A4 Yt3 + t

Add and subtract (A3 + A4 )Yt2 from RHS

Yt = A1 Yt1 + A2 Yt2 + (A3 + A4 )Yt3 + (A3 + A4 )Yt2 (A3 + A4 )Yt2 A4 Yt3 + t

Yt = A1 Yt1 + (A2 + A3 + A4 )Yt2 (A3 + A4 )Yt2 A4 Yt3 + t

Sunil Paul

Derivation of the VECM for VAR

I

Add and subtract (A2 + A3 + A4 )Yt1 from RHS

Yt = A1 Yt1 + (A2 + A3 + A4 )Yt2 + (A2 + A3 + A4 )Yt1 (A2 + A3 + A4 )Yt1 + (A3 + A4 )Yt2 A4 Yt3 + t

Yt = (A1 + A2 + A3 + A4 )Yt1 (A2 + A3 + A4 )Yt1 (A3 + A4 )Yt2 A4 Yt3 + t

Subtract Yt1 from LHS and RHS

Yt = (A1 + A2 + A3 + A4 I )Yt1 (A2 + A3 + A4 )Yt1 (A3 + A4 )Yt2 A4 Yt3 + t

Sum the A0i s :

P
P
P
Yt = ( 4j=1 Ai I )Yt1 4j=2 Ai Yt1 4j=3 Ai Yt2 A4 Yt3 + t
Yt = Yt1 1 Yt1 2 Yt2 3 Yt3 + t

Substitute n = 4 and sum the Y :

Yt = Yt1 +

n1
X

i Yti + t

i=1

where =

Pn

j=1 Ai

I , i =
Sunil Paul

Pn

j=i+1 Aj

VECM:Numerical example

Consider two I(1) process Y1t and Y2t

Assume both these variables are contegrated of order CI(1,1)

with the long run relation given as Y1t 1/8Y2t ,

The vector = (1, 1/8) is known as cointegrating space or

vector and 0 Yt I (0)

Given the possibility of short run deviation from equilibrium we

can define the error as et1 = Y1t1 1/8Y2t1

If there is disequilibrium et1 = Y1t1 1/8Y2t1 6= 0 , either

Y1t , Y2t or both should change to correct the
disequilibrium.

Sunil Paul

VECM:Numerical example

This adjustment can be characterized by an ECM as follows:

Y1t = 1/2et1 + 1t
Y2t = 1/2et1 + 2t
in matrix notation this can be writen as
Yt = et1 + t


1/2
where =
1/2

Sunil Paul

VECM:Numerical example

In our example both variables adjust to disequilibrium since

both 6= 0

The speed of adjustment is given by 1/2 in the Y1t

equation and1/2 for the other

is negative in Y1t equation for

et1 > 0 Y1t1 > 1/8Y2t1 (Hence Y1t should decrease to
correct disequilibrium)

Similarly ifY2t must increase to correct disequilibrium hence

is positive in Y2t

We assume lagged values of changes in both variable are zero

for simplicity

Sunil Paul

VECM:Numerical example

I
I
I

Substituting et1 = 0 Yt1 into the ECM we have :

Y1t = 1/2(Y1t1 1/8Y2t1 ) + 1t
Y2t = 1/2(Y1t1 1/8Y2t1 ) + 2t
The
can be written
as follows
 model
 


  

Y1t1
Y1t
1/2

1, 1/8
=
+ 1t
Y2t
Y2t1
2t
1/2
| {z } |
{z
}| {z } | {z }
Yt

Sunil Paul

Yt1

VECM:Numerical example
The above model can be expressed as follows

 

  
Y1t

1/2 1/16 Y1t1

+ 1t
=
Y2t
2t
1/2 1/16 Y2t1
{z
}| {z } | {z }
| {z } |
Yt

Yt1

where = 0
The VECM can also be written as VAR (Expand Yt and take
Yt1 RHS)
Yt = A1 Yt1 + t


1/2 1/16
where A1 = I =
1/2 15/16

Sunil Paul

Cointegration, Rank and Eigen Values

I
I
I

How to detect CI using matrix?



1/2 1/16
Now inspect the matrix =
1/2 1/16
The doesnt have full rank

Remember Rank of a matrix is nothing but number of linearly

independent number of rows or columns

The second row or the second column of is some scaler

times first row or first column, hence its rank is 1

Even though is a 2 2 matrix it has only one independent

row or column hence is said to have reduced Rank

Sunil Paul

Cointegration, Rank and Eigen Values

I

The determinant of a reduced Rank matrix is zero and some

of its eigen values would be zero

Example:|| = (1/2)(1/16) (1/2)(1 16) = 0

To obtain eigen values(i ) solve the characteristic equation

given by

 

1/2 1/16
0
| I | =

=
1/2 1/16
0


1/2
1/16
1/2
1/16
therefore| I | = (1/2 + )(1/16 + ) (1/2)(1/16) =
2 + 9/16 = 0

i.e. ( + 9/6) = 0 hence the roots are 1 = 0 and

2 = 9/16

Sunil Paul

Cointegration, Rank and Eigen Values

The number of independent cointegrating vectors

= r [rank()]

Since rank of matrix equal to number of nonzero characteristic

roots( eigen values)

If rank() = r then r is the column dimension of or row

dimension of 0 .

Note that in our example only one column or 0 has only one
row hence can have one independent row/column and r = 1

Sunil Paul

Cointegration, Rank and Eigen Values

The rank of a matrix= Number of non zero eigen values

Depending on the Rank of we can have the following cases

Rank
Implications
r =k
VAR is stationary in the levels
1r k 1
r cointegrating vectors
r =0
VAR can be reformulated in first difference

Sunil Paul

Cointegration, Rank and Eigen Values

Rank() = 0
I

There are no cointegrating

variables and the system is
Pn
A
=
I
nonstationary.
i=1 i
Pn
By analogy to the univariate case,all the roots of i=1 Ai = 1
and is zero then all the variables as I (1)
First-difference all the variables to remove nonstationarity, then
standard inference applies.
We can thus write the VECM as a simple VAR in first
differences:
n1
X
Yt =
Yti + t
i=1

Sunil Paul

Cointegration, Rank and Eigen Values

Rank() = k (k = n)
I

I
I

Full rank, hence is nonsingular: all rows (columns) are linearly

independent
The system is stationary if all roots are in the unit circle with
modulus<1
All the variables are stationary
Estimating with unrestricted OLS the level VAR and the
VECM will give identical results.

Sunil Paul

Cointegration, Rank and Eigen Values

Rank() = r < k.
I

The system is nonstationary but there are r cointegrating

relations among the variables that are stationary
r rows are linearly independent, thus r linearly independent
combinations of the {Yit } sequence are stationary.
The Y vector may
Pn be I(1) or higher and the CI relation is
determined by i=1 Ai I = = 0
I

=a (k r ) matrix of weights, the loading matrix, which

measures the average speed of convergence towards LR
equilibrium
= a (k r ) matrix of parameters determining the
cointegrating vectors.

Sunil Paul

Johansen test for cointegrating Rank

A test for cointegration can be formulated using eigen values

of matrix

Assume the matrix has n eigen values and are ordered as

follows:1 > 2 > ... > k
Determine number of non zero eigen values

I
I

If Rank()=0, then all i s are zero, hence ln(1 i ) = 0 .

If rank() = 1 then 0 < 1 < 1 ln(1 1 ) < 0 and
ln(1 2 ) = ... = ln(1 k ) = 0

Sunil Paul

Johansen test for cointegrating Rank

I

We can get estimates of eigen values from the estimated

matrix

Using the estimated eigen values we can construct the

following two test statistics: trace and max

For instance consider trace (0|k)

I

Null Hypothesis H0 : r = 0 (At most 0 cointegrating

vector(CI)) against
Alternative Hypothesis H1 : r 1 (One or more than one CI )

max (0, 1)
I

H0 : r = 0
H1 : r = 1

Sunil Paul

Johansen test for cointegrating Rank

trace (r |k) = T
I

Pk

i=r +1 ln(1

bi )

H0 : r = k
H1 : r k + 1

br +1 )
max (r , r + 1) = T ln(1
I

H0 : r = k
H1 : r = k + 1

i = estimated characteristic roots or eigen values

T = the number of usable observations

r = number cointegrating relation and k is number of variables

Sunil Paul

A general structure of hypothesis tests for cointegrating

rank

max tests
H0
H1
r =0
r =1
r 1
r =2
r 2
r =3
..
..
.
.
r k 1 r =k

max
b1 )
T ln(1
b2 )
T ln(1
b3 )
T ln(1
..
.
bk )
T ln(1

Sunil Paul

A general structure of hypothesis tests for cointegrating rank

H0
r =0
r 1
r 2
..
.
r k 1
I

trace tests
H1

Pk trace
bi )
r 1 T i=1 ln(1
Pk
bi )
r 2 T i=2 ln(1
Pk
bi )
r 3 T i=3 ln(1
..
..
.
.
bk )
r =k
T ln(1

As usual if the test statistics is greater than critical value

reject the null

Sunil Paul

Understanding Trace statistics

b1 = 0.9 and
b2 =
b3 = 0
let k = 3, T = 100,

H0 : r = 0; H1 : r 1
I

H0 : r 1; H1 : r 2
I

trace (0|3) = T [ln(1 0.9) + ln(1 0.0) + ln(1 0.0)] =

100(2.302 + 0 + 0) = 230.2 > 34.91( 5% critical value)

trace (1|3) = T [ln(1 0.0) + ln(1 0.0)] = 100(0 + 0) =

0 < 19.96( 5% critical value)

H0 : r 2; H1 : r 3
I

trace (2|3) = T [ln(1 0.0)] = 100(0) = 0 < 9.24( 5%

critical value)

Sunil Paul

Understanding Maximum eigen statistics

H0 : r = 0; H1 : r = 1
I

H0 : r 1; H1 : r = 2
I

max (0|3) = T [ln(10.9] = 100(2.302) = 230.2 > 22.00(

5% critical value)

max (1|3) = T [ln(1 0.0)] = 100(0) = 0 < 15.57( 5%

critical value)

H0 : r 2; H1 : r = 3
I

max (2|3) = T [ln(1 0.0)] = 100(0) = 0 < 9.24( 5%

critical value)

Sunil Paul

Deterministic Trend Specification

I

The test statistics for cointegration is sensitive to the

specification of trend and intercept in the model
I

Case 1: No intercept or trend in CE or VECM

I

Case 2: Intercept (no trend) in CE-no intercept in VECM

I

if all trends are stochastic

case 4: Intercept and trend in CE no intercept in VECM

I

if none of the series appear to have a trend

case 3: Intercept (no trend) in CE and VECM

I

Rarely used -use it when all series have zero mean

if some of the series are trend stationary

case 5: Intercept and trend in CE- intercept in VECM

I

Rarely used .

Sunil Paul

Hypothesis testing

It is possible to test some of the theoretical restriction imposed

on the elements of matrix.

Consider a money demand function as follows:

mtd = 0 + 1 pt + 2 itb + 2 itd + et
I

theoretically 1 = 1, 2 < 0 3 > 0

We can re-estimate the model by imposing these restrictions.

If restrictions are not binding then the number of cointegrating

vectors will remain same.

Sunil Paul

Hypothesis testing

Johansen (1991) suggested the following test statistics that

follows 2 distribution with (n-r) df
T

k
X

b ) ln(1
b1 )]
[ln(1
1

i=r +1

b1 ,
b1 , ...,
bk are the characteristic roots of unrestricted
where
model
b ,
b
b
and
1 1 , ..., k are the characteristic roots of restricted
model

Sunil Paul

In similar fashion, we can impose restriction on speed of

adjustment parameters in ECM specifications.

The cointegration test is also sensitive to choice of lag length

As in the VAR analysis you need to choose the lag length

using LR statistics, AIC and SBC criteria

Post estimation diagnostic checks can be conducted on the

residuals as in the VAR analysis

Sunil Paul

Testing for Granger Causality

I

"X is said to Granger-cause Y if Y can be better predicted

using the histories of both X and Y than it can by using the
history of Y alone."

Consider the following VAR

I
I
I

Yt = a0 + a1 Yt1 + ... + ap Ytp + b1 Xt1 + ... + bp Xtp + ut

Xt = c0 + c1 Yt1 + ... + cp Ytp + d1 Xt1 + ... + dp Xtp + vt
H0 : b1 = b2 = ... = bp = 0 against H1 :0 Not H0
I

I
I

tests that X does not Granger cause Y

H0 : c1 = c2 = ... = cp = 0 against H1 :0 Not H0

tests that Y does not Granger cause X

Usual Granger Causality involves only stationary variables

If some of the data are nonstationary then Wald test won,t

follow standard F or 2 disb under H0

Sunil Paul

Testing for Granger Causality

If variables are not cointegrated it is permissible to test

granger causality in first difference

If variables are cointegrated Granger causality test in first

difference is not appropriate
I

In cointegrated systems we can test weak exogeneity by

restricting EC

Toda and Yamamoto (1995) procedure to test for Granger

causality.

Sunil Paul

Testing for Granger Causality

Steps for the T-Y procedure

I

Determine the order of integration of variable. Note down the

maximum for the group of time-series m.
Set up a VAR model in the levels of the data, regardless of the
orders of integration of the various time-series.
Determine the appropriate maximum lag length for the
variables in the VAR, say p (using AIC, SIC etc).
Make sure that the VAR is well-specified. Add more lagged
variable if needed.

Sunil Paul

Testing for Granger Causality

I

Steps for the T-Y procedure

I

If two or more of the time-series have the same order of

integration, at Step 1, test for cointegration
Cointegration doesnt change the testing proceducre-it is a
possible cross-check.
Use the preferred VAR model and add in m additional lags to
each variable.
Test for Granger non-causality as follows. Consider a bivariate
case
I

Test the hypothesis that the coefficients of (only) the first p

lagged values of X are zero in the Y equation, using a
standard Wald test.
Then do the same thing for the coefficients of the lagged
values of Y in the X equation.

Sunil Paul

Testing for Granger Causality

I

Dont include the coefficients for the extra m lags when you
perform the Wald tests.

Rejection of the null implies a rejection of Granger

non-causality. .

Use cointegration to cross check the results

I

"If two or more time-series are cointegrated, then there must

be Granger causality between them - either one-way or in both
directions. However, the converse is not true."

Dont fit the VAR in the differences of the data when testing
for Granger non-causality.

If you are using a VAR model for other purposes, then you
would use differenced data if the series are I(1), but not CI.

If you are using a VAR model for purposes other than testing
for GC and the series areCI, then estimate a VECM model.
Sunil Paul