APPLIED ECONOMETRIC
TIME SERIES 4TH ED.
WALTER ENDERS
Chapter 6
WALTER ENDERS, UNIVERSITY OF ALABAMA
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
Example of Cointegration and Money
Demand
In logarithms, an econometric specification for
such an equation can be written as:
mt = b0 + b1pt + b2yt + b3rt + et
where: mt = demand for money
pt = price level
yt = real income
rt = interest rate
et = stationary disturbance term
bi = parameters to be estimated
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
Other Examples
Consumption function theory.
Unbiased forward rate hypothesis.
Commodity market arbitrage and purchasing power parity.
The formal analysis begins by considering a set of economic
variables in long-run equilibrium when
1x1t + 2x2t + + nxnt = 0
Letting and xt denote the vectors (1, 2, , n) and (x1t,
x2t, , xnt)', the system is in long-run equilibrium when bxt =
0. The deviation from long-run equilibriumcalled the
equilibrium erroris et, so that
et = xt
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
Generalization
Letting and xt denote the vectors (1, 2, ..., n)
and (x1t, x2t, ..., xnt), the system is in long-run
equilibrium when xt' = 0. The deviation from
long-run equilibrium--called the equilibrium
error--is et, so that:
et = x't
If the equilibrium is meaningful, it must be the
case that the equilibrium error process is
stationary.
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
Figure 6.1: Scatter Plot of Cointegrated Variables
0
-0.5
Values of z
-1
-1.5
-2
-2.5
-3
-3
-2.5
-2
-1.5
-1
-0.5
Values of y
The scatter plot was drawn using the {y} and {z} sequences from Case 1 of Worksheet 6.1.
Since both series decline over time, there appears to be a positive relationship between the
two. The equilibrium regression line is shown.
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
Figure 6.2: Three Cointegrated Series
2
-2
-4
-6
-8
-10
-12
0
10
20
30
40
50
y
60
z
70
80
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
90
100
Three important points
1. Cointegration refers to a linear combination of non-stationary
variables.
If (1, 2, ... , n) is a cointegrating vector, then for any nonzero value of , (1, 2, ... , n) is also a cointegrating
vector.
Typically, one of the variables is used to normalize the
cointegrating vector by fixing its coefficient at unity.
To normalize the cointegrating vector with respect to x1t,
simply select = 1/1.
2. The equation must be balanced in that the order of integration
of the two sides must be equal
3. If xt has m components, there may be as many as m-1 linearly
independent cointegrating vectors
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
Example of Multiple Cointegrating Vectors
Let the money supply rule be:
mt = 0 - 1(yt + pt) + e1t
(1.3)
= 0 - 1yt - 1 pt + e1t
where: {e1t} is a stationary error in the money supply
feedback rule.
Given the money demand function in (1.1), there are two
cointegrating vectors for the money supply, price level, real
income, and the interest rate. Let be the (5 x 2) matrix:
1 0 1 2 3
=
0
1
1
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
COINTEGRATION AND COMMON TRENDS
yt = yt + eyt
zt = zt + ezt
where it = a random walk process representing the trend in
variable i
eit = the stationary (irregular) component of variable i
If {yt} and {zt} are cointegrated of order (1,1), there must be
nonzero values of 1 and 2 for which the linear combination
1yt + 2zt is stationary. Consider the sum
1yt + 2zt = 1(yt + eyt) + 2(zt + ezt)
= (1yt + 2zt) + (1eyt + 2ezt)(6.6)
For 1yt + 2zt to be stationary, the term (1yt + 2zt) must
vanish.
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
Granger Representation Theorem
In an error-correction model, the short-term
dynamics of the variables in the system are
influenced by the deviation from equilibrium.
rSt = S(rLt1 rSt1) + St
S > 0
rLt = L(rLt1 rSt1) + Lt
L > 0
This finding illustrates the Granger representation theorem
stating that for any set of I(1) variables, error correction and
cointegration are equivalent representations.
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
The Engle-Granger Methodology
Step 1: Pretest the variables for their order of integration.
Step 2: Estimate the long-run equilibrium relationship.
If the results of Step 1 indicate that both {yt} and {zt} are
I(1), the next step is to estimate the long-run equilibrium
relationship in the form:
yt = 0 + 1 zt + et
Consider the autoregression of the residuals:
n
=
et a1et 1 + ai +1et i + t
i =1
Test a1 = 0?
Step 3: Estimate the error-correction model
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
The Error Correction Model
yt = 1 + y [ yt 1 1 zt 1 ] + a11 (i )yt i + a12 (i )zt i + yt
=i 1 =i 1
zt = 2 + z [ yt 1 1 zt 1 ] + a21 (i )yt i + a22 (i )zt i + zt
=i 1 =i 1
Instead of a cross-equation restriction, use
yt = 1 + y et 1 + a11 (i )yt i + a12 (i )zt i + yt
=i 1 =i 1
zt = 2 + z et 1 + a21 (i )yt i + a22 (i )zt i + zt
=i 1 =i 1
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
Speed of adjustment coefficients
The speed of adjustment coefficients y and z are of
particular interest in that they have important implications
for the dynamics of the system.
Direct convergence necessitates that be negative and z be
positive. If we focus on (6.36) it is clear that for any given
value of the deviation from long-run equilibrium, a large
value of z is associated with a large value of zt.
If one of these coefficients is (say y) is zero,the {zt}
sequence does all of the correction to eliminate any
deviation from long-run equilibrium. Since {yt} does not
do any of the error-correcting, {yt} is said to be weakly
exogenous.
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
Problems with the EG-Method
1. In practice, it is possible to find that one regression indicates the
variables are cointegrated whereas reversing the order indicates no
cointegration. This is a very undesirable feature of the procedure since
the test for cointegration should be invariant to the choice of the
variable selected for normalization. The problem is obviously
compounded using three or more variables since any of the variables
can be selected as the left-hand-side variable.
2. Moreover, in tests using three or more variables, we know that there
may be more than one cointegrating vector. The method has no
systematic procedure for the separate estimation of the multiple
cointegrating vectors.
3. Another serious defect of the Engle-Granger procedure is that it
relies on a two-step estimator.
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
Johansen Methodology
Reconsider the n-variable first-order VAR given by
(6.3): xt = A1xt-1 + t. Subtract xt-1 from each side to obtain:
xt = A1xt1 xt1 + t
= (A1 I)xt1 + t
= xt1 + t
The rank of (A1 I) equals the number of cointegrating
vectors.
If (A1 I) consists of all zeroesso that rank() = 0all of the {xit}
sequences are unit root processes.
If (A1 I) is of full rankso that rank() = neach of the {xit}
sequences converges to a point.
The process can be modified to include a drift and
seasonal dummy variables.
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
Consider xt = xt-1:
11 12
21 22
* =
.
.
n1 n 2
... 1n 10
... 2 n 20
...
.
.
... nn n 0
If rank * = 2,
11x1t + 12x2t + 13x3t + ... + 1nx1n + 10 = 0
21x1t + 22x2t + 23x3t + ... + 2nx1n + 20 = 0
Note: Adding a column of constants still means that
rank(*) cannot exceed n
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
The number of distinct cointegrating vectors can be obtained
by checking the significance of the characteristic roots of .
We know that the rank of a matrix is equal to the number of
its characteristic roots that differ from zero. Suppose we
obtained the matrix and ordered the n characteristic roots
such that 1 > 2 > ... > n. 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 - i) will equal zero if the variables are
not cointegrated. Similarly, if the rank of is to unity, the
first expression ln(1 - 1) will be negative and all the other
expressions are such that ln(1 - 2) = ln(1 - 3) = ... = ln(1 n) = 0.
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
trace( r ) = T ln(1 i )
i = r +1
The null hypothesis that the number of distinct cointegrating vectors is less
than or equal to r against a general alternative. From the previous discussion,
it should be clear that trace equals zero when all i = 0.
T ln(1 r +1 )
max ( r, r + 1) =
The null that the number of cointegrating vectors is r against the
alternative of r+1 cointegrating vectors. Again, if the estimated
value of the characteristic root is close to zero, max will be small.
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
Null
Hypothesis
Alternative
Hypothesis
trace tests:
95%
Critical
Value
90%
Critical
Value
trace value
r=0
r>0
44.94926
29.68
26.79
r <= 1
r>1
14.80894
15.41
13.33
r <= 2
r>2
3.60231
3.76
2.69
max tests:
max value
r=0
r=1
30.14032
20.97
18.60
r=1
r=2
11.2066
14.07
12.07
r=2
r=3
3.60231
3.76
2.69
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
In order to test other restrictions on the cointegrating vector, Johansen
defines the two matrices and both of dimension (n x r) where r is the
rank of . The properties of and are such that:
= '
In essence, we can normalize to obtain '
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
Hypothesis Testing
n
*
[ln(1
i ) ln(1 i )]
i = r +1
Asymptotically, the statistic has a 2 distribution with (n - r) degrees of
freedom.
The value of this statistic should be zero if the restriction is not binding.
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
Lag Length and Causality Tests
p 1
xt 1 + i xt i + t
x=
t
i =1
Estimate the models with p and p 1 lags. Let c denote the
maximum number of regressors contained in the longest
equation. The test statistic
(Tc)(logr logu)
can be compared to a 2 distribution with degrees of freedom
equal to the number of restrictions in the system.
Alternatively, you can use the multivariate AIC or SBC to
determine the lag length.
If you want to test the lag lengths for a single equation, an Ftest is appropriate.
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
To difference or not to difference?
Difference
Tests lose power if you do not
difference: you estimate n2 more
parameters (one extra lag of each
variable in each equation).
If you use first differences, you can use
the standard F distribution to test for
Granger causality.
When the VAR has I(1) variables, the
impulse responses at long forecast
horizons are inconsistent estimates of
the true responses. Since the impulse
responses need not decay, any
imprecision in the coefficient estimates
will have a permanent effect on the
impulse responses.
Do not difference
If the system contains a
cointegrating relationship,
the system in differences is
misspecified since it
excludes the long-run
equilibrium relationships
among the variables that are
contained in xt1.
All of the coefficient
estimates, t-tests, F-tests,
tests of cross-equation
restrictions, impulse
responses and variance
decompositions are not
representative of the true
process.
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
Restrictions on the cointegrating vectors
Testing coefficient restrictions: As in the previous section, once you select
the number of cointegrating vectors, you can test restrictions on the resulting
values of and/or . Suppose you want to test the restriction that the
intercept is zero. From the menu, you select Restrictions on subsets of .
1 1
2 0
=
3 0
0 0
0
1
0
0
0
11
0
21
1
31
0
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
Instead, suppose you want to test the three restrictions: 1 = 2, 1 = -3, and
3 = 0 (so that the normalized cointegrating vector has the form yt + zt - wt =
0). In matrix form, the
1 1
1
2 = [ 11]
3 -1
4 0
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
Linear vs Threshold Cointegration
In the simplest case, the two-step methodology
entails using OLS to estimate the long-run
equilibrium relationship as:
x1t = 0 + 2x2t + 3x3t + ... + nxnt + et
where: xit are the individual I(1) components of xt, i
are the estimated parameters, and et is the
disturbance term which may be serially correlated.
The second-step focuses on the OLS estimate of
in the regression equation:
et = et-1 + t
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
The TAR Specification
Let the error process have the form
et = It 1et-1 + (1 - It )2et-1 + t
where: It is the Heaviside indicator function
such that:
1 if e t-1
It =
0 if e t-1 <
The Momentum Specification
1 if e t-1 0
It =
0 if e t-1 < 0
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
TABLE 7: Estimates of the Interest Rate Differential
From Enders and Siklos (JBES)
Engle-Granger
Threshold
Momentum
MomentumConsistent
1a
-0.068
(-2.858 )
-0.085
(-2.522)
-0.021
(-0.628)
-0.020
(-0.680)
2a
NA
-0.020
(-1.582)
-0.117
(-3.526)
-0.141
(-3.842)
1a
0.188
(-2.782)
0.190
(2.787)
0.183
(2.730)
0.186
(2.790)
2a
-0.149
(-2.197)
-0.147
(-2.153)
-0.161
(-2.376)
-0.155
(-2.312)
AIC b
11.74
13.24
9.285
7.022
NA
4.32
6.363
7.548
1 = 2 d
NA
0.495
(0.482)
4.418
(0.037)
6.698
(0.010)
Q(4)e
Q(8)
Q(12)
0.65
0.60
0.75
0.64
0.58
0.73
0.64
0.52
0.68
0.48
0.51
0.70
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
xit = 1.iItet-1 + 2.i(1 - It)et-1 + ... + vit
where: 1.i and 2.i are the speed of adjustment
coefficients of xit.
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
10. Error-Correction and ADL Tests
yt = 1(yt1 zt1) + e1t
zt = 2(yt1 zt1) + e2t
where: e1t = e2t + vt
As such, we can always write
yt = (yt1 zt1) + zt + vt
(6.67)
The general problem is that zt will be correlated with the
error term vt so that there is a simultaneity problem.
However, if zt is weakly exogenous and causally prior to yet
we can estimate (6.67)
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
The ADL Test
yt = 1yt1 1zt1 + zt + vt
Table F uses the work of Ericsson and MacKinnon
(2002) to calculate the appropriate critical values
necessary to determine whether 1 < 0.
Given that the variables are cointegrated:
If zt is unaffected by innovations in yt, it is
appropriate to conduct inference on (6.69) using a
standard t-tests and F-tests.
Copyright 2015 John, Wiley & Sons, Inc. All rights reserved.
