You are on page 1of 5

Econometrics C

Department of Economics, University of Copenhagen


Heino Bohn Nielsen
PS #5
Cointegration and Error Correction
This problem set deals with cointegration and error correction. Exercise #5.1 is a question
on unit testing and cointegration. Exercise #5.2 takes you through the steps of an Engle-
Granger analysis for Danish aggregate consumption, while exercise #5.3 guide you through
a cointegration analysis based on a dynamic regression model.
#5.1 Unit Roots and Cointegration
(1) Consider the following regression model
r

= c +cr
1
+r
1
+c

, t = 2, 3, ..., T, (5.1)
where the initial values, r
0
and r
1
, are given.
Explain how the presence of a unit root can be tested against the stationary alter-
native.
Let r

denote the eective US Federal funds rate (which is an overnight interest


rate), and let /

denote a 1 year bond rate. Dene the interest rate spread, measuring
the slope of the short end of the yield curve, as
r

= /

.
Imagine that you are informed that both r

and /

are unit root processes, and


that we are interested in testing whether the slope of the yield curve behaves in
a stationary manner. Table 5.1 contains the output of the regression in (5.1) for
the interest rate spread 1988 : 1 2004 : 10. Use the information in the table to
test the hypothesis that r

has a unit root. How is this related to the concept of


cointegration.
1
(2) Now consider the autoregressive distributed lag, ADL, model given by
r

= c +0
1
r
1
+0
2
r
2
+c
0
/

+c
1
/
1
+c
2
/
2
+c

. (5.2)
Derive the corresponding error correction model (ECM), and explain how it is related
to cointegration.
Table 5.1: Modelling r

by OLS for 1988 : 1 2004 : 10


Coecient Std.Error tvalue
Constant 0.0121 0.0142 0.8490
r
1
0.2318 0.0690 3.3600
r
1
0.1060 0.0287 3.7000
b o 0.1966 1oo 7.6919
1
2
0.0940 1(2, 199) 10.3200
No. of observations 202
#5.2 Engle-Granger Analysis for Danish Consumption
In this exercise we want to construct a single equation cointegration model for the Danish
private consumption. We dene the vector of variables
Z

ARBLOS

log(FCP

)
log(FYDP

)
log(REALFOR

)
ARBLOS

,
where FCP is real Danish private consumption, FYDP is real disposable income, REALFOR
is real wealth including the value of owner occupied housing, and ARBLOS is the income
loss from changes in unemployment. The variables are located in ConsumptionData.In7.
(1) Test for unit roots in each of the four variables in Z

. Make sure that you use a


sensible deterministic specication.
What do your conclusions imply for the estimation of a cointegration relation for
the variables in Z

?
(2) Use OLS to estimate the static long-run relation
C

= ,
0
+,
1
1

+,
2
\

+n

, (5.3)
and interpret the signs and the magnitudes of the coecients. Do you think it is
reasonable to use the output to test statistical hypothesis on the coecients?
2
(3) Perform an Engle-Granger residual based test for whether (5.3) is a cointegrating
relation. Explain the test and the outcome.
(4) Irrespective of the conclusions in question (3), dene the error correction term,
ecm

= b n

, corresponding to the relation in (5.3).


Construct a graph of ecm

and try to explain why there is not strong support for


cointegration.
(5) Construct single equation error correction models for C

, 1

, and \

, where
you include also ARBLOS

as an explanatory variable, e.g.


C

= c
0
+c
1
C
1
+c
2
1

+c
3
1
1
+c
4
\

+c
5
\
1
+c
6
ARBLOS

+c
7
ecm
1
+c

. (5.4)
Do any of the variables seem to error correct? Is that in line with your expectations
based on economic theory?
(6) Concentrate on the model for C

. Delete insignicant variables and reestimate the


model (5.4). Perform the usual misspecication tests. Does the model appear well
specied?
(7) Notice that the standardized residuals for 1975 : 4, 1977 : 3 and 1977 : 4 are very
large compared to the rest of the residuals and compared to the expected range for
observations drawn from a standard normal distribution. These observations are
related to two transitory policy measures. Discuss why it might be preferable to try
to remove the eects of these observations.
To remove the eects of the observations, we can construct dummy variables, i.e.
variables of the form 0, ..., 0, 1, 0, ..., 0 with the value 1 located at the particular point
in time. In PcGive you can construct a dummy for 1975 : 4 in the algebra editor as
DUM754=dummy(1975,4,1975,4);
which assigns to DUM754 the value one from the rst to the last entry in the function
and zero otherwise.
Construct dummy variables for the three special episodes, and include them in the
model. What happened to the residuals of the observations with dummy variables,
e.g. c
1975:4
? How do the parameter estimates change? How do the misspecication
tests change?
Now we introduce the concept of recursive estimation, which is very useful in analyzing
the structural stability of an estimated model. Recursive estimation is done by estimating
a model like (5.4) for increasing samples t = 1973 : 2 T
0
, where T
0
takes the values
1973 : 2 +, 1973 : 2 + + 1, 1973 : 2 + + 2, ..., 2003 : 2.
That is, we rst estimate the model with only the rst observations; and then we
successively add a new observation and reestimate the model. For each sample we do an
3
OLS estimation and obtain all the usual statistics. Afterwards we can consider the sample
paths of the dierent statistics calculated for each sample. For example we can consider
the estimated coecients for the expanding samples
c

(T
0
), for T
0
= 1973 : 2 +, ..., 2003 : 2.
The model (5.4) is estimated under the assumption of constant coecients, which implies
that a graph for c

(T
0
), i = 0, ..., 7, should not uctuate too much.
(8) Do a recursive estimation of your error correction model and look at the recursive
parameter estimates. Does the model look stable?
[Hint: To do this, check the recursive estimation box in the estimation dialogue. Set
the shortest sample, , (i.e. the initialization) equal to two times the number of
coecients to be estimated. Select Recursive Graphics in the Test menu.]
#5.3 ADL Analysis for Danish Consumption
This exercise introduces cointegration analysis based on an unrestricted ADL or ECM
model. That is an alternative to the Engle-Granger procedure discussed above.
(1) Explain why the estimates of a cointegration vector based on an unrestricted dy-
namic model (ADL or ECM) are expected to be superior to the estimates derived
from the Engle-Granger two-step approach.
(2) Estimate an unrestricted ADL model
C

= ,
0
+,
1
C
1
+,
2
C
2
+,
3
1

+,
4
1
1
+,
5
1
2
+,
6
\

+,
7
\
1
+,
8
\
2
+,
9
ARBLOS

+,
10
1n:754

+,
11
1n:773

+,
12
1n:774

+c

.
and delete the insignicant terms.
(3) Test the model for misspecications.
(4) Derive the long-run solution and interpret the results.
[Hint: This is done in PcGive by choosing Dynamic Analysis in the [Test] menu
and selecting Static long-run solution.]
(5) Perform the PcGive test for no-cointegration. What do you conclude concerning the
cointegration properties?
[Hint: This is done under Dynamic Analysis by selecting Lag structure analysis.]
(6) Derive and interpret the dynamic multipliers.
[Hint: They are derived with Lag-weights under Dynamic Analysis]
4
(7) Finally, estimate the unrestricted error correction model:
C

= c
0
+c
1
C
1
+c
2
1

+c
3
1
1
+c
4
\

+c
5
\
1
+c
6
C
1
+c
7
1
1
+c
8
\
1
+c
9
ARBLOS

+c
10
1n:754

+c
11
1n:773

+c
12
1n:774

+c

.
Compare the estimates with the estimates from the ADL model.
Derive (manually) the long-run solution and compare with the ADL model.
manually perform the PcGive test for no-cointegration.
5

You might also like