Sie sind auf Seite 1von 10

EC306

Michael P Clements
Handout 5
1 Spurious regressions versus cointegration
Even completely unrelated economic time series may appear to be related using conventional
testing procedures.
For example, suppose:
j
t
= jj
t1
+
t
where
t
~ iid
_
0, o
2
v
_
(1)
and:
r
t
= jr
t1
+-
t
where -
t
~ iid
_
0, o
2
"
_
. (2)
and the two shocks
t
and -
t
are independent at all points in time, i.e.,
1 [
t
-
s
] = 0 \t, :. (3)
When j = 1, both series are said to have a unit root, specically, they are random walks.
j
t
= j
t1
+
t
(4)
r
t
= r
t1
+-
t
Suppose we run an OLS regression to nd out if there is a linear relationship between j
t
and
r
t
:
j
t
= c +,r
t
+n
t
(5)
Despite lack of causal relationship between the two variables, we are likely to nd a signicant
t-ratio for the null H
0
: , = 0.
1.1 Simulation experiment.
Simulate data from (4) on the computer using pseudo-random numbers for
t
and -
t
for
a given sample size, T, to give r
t
, j
t

T
t=1
. Run regression (5) calculate the usual t-ratio,
t =
^
,,

:c(
^
,). Record whether [t[ 2, where 2 are the critical points for a test at the 5% level
(i.e., with a probability of a Type 1 error of 5%).
Simulate another set of data for j
t
and r
t
, r
t
, j
t

T
t=1
. Again calculate the t-ratio, and
record whether it lies within 2.
Repeat 25,000 times.
Calculate the percentage of replications on which the null hypothesis (of no relationship)
was rejected.
Find that Pr(reject H
0
using 5% level test) 75%. See Figure 1.
0
60 80 100 120 140 160 180 200 220 240
10
20
30
40
50
60
70
80
90
= 0
= 0 . 5
= 1
Figure 1: Rejection frequencies for t-test of , = 0 in equation (5) for three values of j, j = 0,
0.5 and 1, and for all sample sizes from T = 50 to 250
1.2 Testing for a relationship in a dynamic model.
One explanation is that the regression model (5) is false under the null. Under the null, (5)
implies that:
j
t
= c +n
t
i.e., j
t
is not related to its past values, when in fact it equals its previous value plus a random
disturbance term. The alternative hypothesis (j
t
and r
t
are related) is false, but so is the null,
so it is not clear what one might expect to nd.
Suppose we augment (5) with a lagged dependent variable:
j
t
= c +,r
t
+j
t1
+c
t
. (6)
Now the regression model embeds the true model (e.g., when , = 0 and = 1). Now nd
Pr(reject H
0
using 5% level test) - 45%, so there is nearly an evens chance that we will
conclude the variables are related. So that the over-rejection is not solely due to the regression
model being incorrectly specied. See Figure 2.
The problem is that the t-test of , = 0 is not (0, 1) even asymptotically the standard
asymptotic distribution theory does not apply when variables have unit roots. So we want to
discriminate between two situations.
1 Spurious regressions. Apparently signicant relationships between genuinely unrelated se-
ries. Using standard critical values for tests of signicance we will reject the null of no
relationship far too often (actual size of the test exceeds the nominal size (prob. of Type
1 error) ).
2 Genuine relationships between series. These arise when the time series are cointegrated.
1
60 80 100 120 140 160 180 200 220 240
5
10
15
20
25
30
35
40
45
= 0
= 0 . 5
= 1
Figure 2: Rejection frequencies for t-test of , = 0 in equation (6) for three values of j, j = 0,
0.5 and 1, and for all sample sizes from T = 50 to 250
2 Cointegration
2.1 Unit root accounting.
Suppose r
t
~ 1 (1), j
t
~ 1 (1) and .
t
~ 1 (0). c
1
, c
2
are scalars.
1. r
t
+.
t
~ 1 (1). (1 (1) is the dominant property)
2. c
1
r
t
~ 1 (1)
3. c
1
.
t
~ 1 (0)
4. Generally, r
t
= c
1
j
t
+ c
2
r
t
~ 1 (1) for arbitrary c
1
, c
2
. However, if there are values of
c
1
and c
2
, say, c

1
and c

2
, such that r
t
= c

1
j
t
+c

2
r
t
~ 1 (0), then the series are said to be
cointegrated.
Denition.
(General). Two or more series which are individually integrated of order d (1 (d)) are
said to be cointegrated when a linear combination of the series is integrated of order d /
(1 (d /)) for / 0.
(Specic). When d = 1, the series are cointegrated when there is an 1 (0) linear com-
bination (/ = 1).
The weights c
1
and c
2
conventionally written as a column vector [c

1
: c

2
]
0
, which is known
as the cointegrating vector, and r
t
is the cointegrating combination, since:
[c

1
: c

2
]
_
j
t
r
t
_
= r
t
~ 1 (0)
2
Because if r
t
is 1(0), then :r
t
is 1(0), let : =
1
c

1
:
:r
t
=
1
c

1
[c

1
: c

2
]
_
j
t
r
t
_
=
_
1 :
c

2
c

1
_ _
j
t
r
t
_
~ 1 (0)
so we can always normalise the cointegrating combination as:
j
t
,r
t
~ 1 (0)
where , =
c

2
c

1
, so , is unique up to multiplication by a scalar.
For two series to be cointegrated, they must have a common 1 (1) component, e.g., suppose:
j
t
= ,A
t
+ j
t
r
t
= A
t
+ r
t
where A
t
is 1 (1), and r
t
and j
t
are 1 (0).
Then r
t
and j
t
are 1 (1), but because their 1 (1) component is common, it can be eliminated
so that:
j
t
,r
t
= j
t
, r
t
~ 1 (0)
A
t
is sometimes known as the common stochastic trend, e.g.:
A
t
= A
t1
+-
t
,
A
t
=
t

s=1
-
s
.
if A
0
= 0. [c.f. t =

t
s=1
1]
To see uniqueness, calculate the linear combination for

, = , +c, c ,= 0.
2.2 Cointegration and equilibrium
Variables that cointegrate move together through time. Thus, if two variables are cointegrated,
they will remain more or less in the relationship to each other given by the cointegrating vector,
i.e., the dierence between j
t
and ,r
t
will be an 1 (0) term for all t. The error, r
t
= j
t
,r
t
is 1(0). It has a mean of zero, and since r
t
is 1(0), we know 1[r
2
t
] is constant over t. So j
t
can
not drift increasingly far away from ,r
t
: cointegration is a statistical notion of equilibrium
Co-integrating combinations dene equilibrium relationships.
When we model relationships between variables (e.g., consumption and income; or exchange
rates and interest rates, we can exploit these equilibrium relationships.
An alternative approach is simply to dierence all the 1 (1) variables at the beginning of
the modelling exercise (thus, r
t
, j
t
) to make them 1 (0) and then model the relationship
between the dierenced variables, where standard distribution theory applies. This approach is
often associated with Box and Jenkins (1970). But if the series are cointegrated their approach
throws away useful information, and often economic theory has something to say about the
specication of the long run, if not the process of dynamic adjustment.
3
3 Cointegration: Testing & Estimation
3.1 Static Regressions
Simplest method. To discover if there is a cointegrating relationship between two 1(1) variables:
j
t
, r
t
.
Run the OLS regression:
j
t
= ,
0
+,
1
r
t
+
t
The estimated residual is:
^
t
= j
t

^
,
0

^
,
1
r
t
The linear combination given by ^
t
is the linear combination most likely to be a cointe-
grating relationship, i.e., the most likely cointegrating vector is
_
1,
^
,
1
_
: 1 j
t

^
,
1
r
t
(
^
,
0
).
Why?
OLS will estimate the cointegrating linear combination (if there is one), since it minimizes
the sum of squares of the error, i.e.,
arg minT
1

0
;
1

2
t
If the cointegrating relationship is given by

t
= j
t
,

0
,

1
r
t
then \ [

t
] = ' a nite value, for all t. For

t
:
1
_
T
1

2
t
_
= T
1

1
_

2
t

= T
1

\ [

t
] = '
But for all combinations other than ,

0
, ,

1
,
t
~ 1 (1), so that \ [
t
] is increasing in t, say,
\ [
t
] =

't.
1
_
T
1

2
t
_
= T
1

1
_

2
t

= T
1

'

t ~ O(T)
and the Sum of Squares grows without bound. [

T
t=1
t =
1
2
T (T + 1)]
Hence OLS (i.e., minimizing T
1

2
t
) should pick out the cointegrating combination, if
there is one, i.e.,
_
^
,
0
,
^
,
1
_
[,

0
, ,

1
].
3.2 Testing whether the
_
^

0
;
^

1
_
combination is a cointegrating combination.
If j
t
and r
t
are cointegrated, then
t
~ 1 (0).
Test for cointegration is whether ^
t
is 1 (1) (no cointegration) or 1 (0) (there is cointegration).
Use ordinary unit root tests, which we previously applied to test the order of integration of
a variable.
^
t
= c^
t1
+-
t
H
0
: c = 0 versus H
A
: c < 0. It may be necessary to add lagged values of ^
ti
, i 0 as
regressors to ensure that the error (-
t
) is serially uncorrelated.
4
Critical values dier from those for testing an individual series for integration. Critical
values are larger negative numbers - more dicult to reject the null. This is because by
choosing the best possible candidate for cointegration, OLS biases the test toward nding
cointegration (i.e., rejecting the unit root in ^
t
). Consequently, we have to make it harder to
reject the unit root null. Put another way, the standard Dickey-Fuller critical values would
apply if we were testing
t
(with ,
0
, ,
1
known).
Critical values depend on the number of regressors. See table 1. : is the total number of
variables, so : = 2 is relevant for a single regressor.
Table 1: Cointegration unit root test critical values
Lower-tail critical values
Test statistic 1% 5% 10%
: = 2
t
c
3.90 3.34 3.04
t
ct
4.32 3.78 3.50
: = 3
t
c
4.29 3.74 3.45
t
ct
4.66 4.12 3.84
: = 4
t
c
4.64 4.10 3.81
t
ct
4.97 4.43 4.15
: = 5
t
c
4.96 4.42 4.13
t
ct
5.25 4.72 4.43
: = 6
t
c
5.25 4.71 4.42
t
ct
5.52 4.98 4.70
Critical values for (A)DF tests of cointegration. Table 20.2 from Davidson and Mackinnon,
Estimation and Inference in Econometrics, OUP, (1993).
t
c
is appropriate when a constant is included in the cointegrating regression, t
ct
when a constant
and trend are included.
Example.
To test whether j
t
and r
t
are cointegrated allowing for a constant term, i.e.:
j
t
= ,
0
+,
1
r
t
+
t
compare the unit root DF test of ^
t
to the : = 2, t
c
critical value of 3.34 for a 5% level
test. Need a larger negative test statistic to reject the null of 1(1) (no cointegration).
Advantages and disadvantages of static regression approach to cointegration:
simple - requires only LS estimation
5

^
,
1
is a consistent estimator, in fact it is super-consistent.
But
t
will typically be autocorrelated if there is a dynamic relationship between j and
r, as the dynamics are ignored. This means that in practice there may be signicant
small-sample biases in
^
,
1
.
If there are more than 2 variables, there may be more than one cointegrating relationship.
Only a linear combination of these can be estimated.
^
, is not normally distributed -
cannot test hypothesis about , based on the static regression estimator.
A VAR-based cointegration approach will be shown to x some of these problems.
4 Cointegration and structural breaks
When there has been a structural break, the null of no cointegration will not be rejected often
enough. Tests that do not make an allowance for the break will have low power.
Rather than basing the test on:
j
t
= c
1
+,
1
r
t
+
t
we can allow for various types of strucural break by allowing:
j
t
= c
1
+c
2
1 (t T
b
) +`t +,
1
r
t
+,
2
1 (t T
b
) r
t
+c
t
This allows for the possibility that j
t
and r
t
cointegrate with
i) a change in the intercept but no change in the slope, if we set ` = 0, ,
2
= 0 in the model
ii) a change in the intercept allowing for a time trend, ,
2
= 0
iii) a change in the slope and intercept, ` = 0.
But the critical values of the ADF test will increase (ADF needs to be more negative to
reject the null) and will depend on which of i), ii) and iii) we estimate.
Assuming T
b
is unknown, can compute the ADF test for ^ c
t
based on every possible value
of T
b
, and select the ADF test value with the largest negative value across all possible break
points T
b
.
1
5 Error-correction models
It can be shown that if two variables are cointegrated, then an error-correction formulation
exists.
Suppose we have a dynamic relationship between consumption and income of the form:
c
t
= ,
1
j
t
+,
2
j
t1
+,
3
c
t1
+-
t
. (7)
1
See Table A.7 of Harris and Sollis, p. 276 for critical values, and their discussion and illustration on p.85-6.
6
Suppose c
t
and j
t
are 1(1), which is reasonable empirically. What can we say about the status
of (7)?
Firstly, subtract c
t1
from both sides:
c
t
= ,
1
j
t
+,
2
j
t1
+ (,
3
1) c
t1
+-
t
. (8)
Add and subtract ,
1
j
t1
from the RHS:
c
t
= ,
1
j
t
+ (,
1
+,
2
) j
t1
+ (,
3
1) c
t1
+-
t
. (9)
And then rewrite in error-correction formulation:
c
t
= ,
1
j
t
+ (,
3
1)
_
c
t1

(,
1
+,
2
)
(1 ,
3
)
j
t1
_
+-
t
In this equation, c
t
and j
t
are 1(0) as they are the dierences of 1(1) variables. The term
in squared brackets is 1(0) i [1 ,] is a cointegrating vector, where , = (,
1
+,
2
)(1,
3
)
1
.
When all the terms are integrated of the same order, the equation is said to be balanced.
.
t1
= [1 ,]
_
c
t1
j
t1
_
= c
t1
,j
t1
~ 1(0)
If two variables are cointegrated, then an error-correction (EC) formulation always exists.
If an error-correction formulation exists for two 1(1) variables, then they are cointegrated.
Cointegration implies error-correction, and error-correction implies cointegration.
If the variables were not cointegrated, then ,
3
1 = 0, and there is no equilibrium in the
equation. The equation would be a relationship between the rates of change (if variables are in
logs) of c
t
and j
t
: c
t
and j
t
. In that case, c
t
(or c
t
) would not respond to the gap between
the levels of c and j.
When there is an EC term, then c
t
responds to j
t
, but also to .
t1
. .
t1
is the equilibrium
error last period. If .
t1
0, so that c
t1
,j
t1
, then c
t
was above its equilibrium value at
t 1. Hence, c.p. with ,
3
1 < 0, then c
t
< 0. So c
t
is reduced. In this way, deviations from
equilibrium are corrected the next period, so the equation traces out an equilibrium trajectory.
The error-correction term is the mechanism which ensures the cointegrated variables move
together in the long run.
5.1 Two-step procedure
How can we estimate the error-correction model?
c
t
= c
1
j
t
+c
2
[c
t1
,j
t1
] +-
t
The two step-procedure of Engle and Granger (1987)
2
suggests rst running OLS on:
c
t
= j
t
+c
t
2
Engle, R.F. and Granger, C.W.J. (1987) Cointegration and error correction: Representation, estimation and
testing Econometrica, 55, 251276.
7
and taking the residual ^ c
t
= c
t
^ j
t
as the error-correction term.
The second stage estimates by OLS:
c
t
= c
1
j
t
+c
2
^ c
t1
+
t
= c
1
j
t
+c
2
(c
t1
^ j
t1
) +
t
This approach is feasible because ^ , as T gets large. So the estimator ^ is consistent.
This is the case regardless of whether j
t
, the explanatory variable, and c
t
, the disturbance, are
correlated. Usually, if the regressor(s) and disturbance term are correlated, OLS estimators are
inconsistent. Not so, here, as long as c
t
is 1(0), i.e., c
t
and j
t
are cointegrated. However, there
are other ways of estimating error-correction models.
Empirical illustration. Real wages (rn
t
) and productivity (jr
t
). 1964:2 to 1999:3.
rn
t
= 0.23
(0:078)
rn
t1
+ 0.00155
(0:0018)
+ 0.357
(0:11)
jr
t
0.306
(0:11)
jr
t3
0.109
(0:034)
^-
t1
0.00551
(0:0025)
Seasonal
t1
+ 0.0128
(0:0026)
Seasonal
t2
where ^-
t
is the OLS residual from the regression of rn
t
on 1, jr
t
.
The t-statistic on ^-
t1
is 3.26. The EC term has the right sign, in that nearly 11% of
the disequilibria is corrected each period.
6 Properties of error-correction models
Illustrate with consumers expenditure. Basis of the DHSY model (Davidson, Hendry, Srba
and Yeo, 1978
3
) is (lower-case denotes logs):
c
t
=
1
j
t
+
2
(j
t1
c
t1
) +
t
,
1
0,
2
0 (10)
1. This is a restricted version of a general dynamic model:
c
t
= a
1
c
t1
+a
2
j
t
+a
3
j
t1
+
t
(11)
where:
a
1
+a
2
+a
3
= 1. 1
2
= a
1
,
1
= a
2
,
2

1
= a
3
.
Unrestricted if:
c
t
=
1
j
t
+
2
(j
t1
c
t1
) +
3
j
t1
+
t
,
1
0,
2
0 (12)
or:
c
t
=
1
j
t
+
2
(,j
t1
c
t1
) +
t
,
1
0,
2
0 (13)
3
Davidson, J.E.H., Hendry, D.F., Srba, F. and Yeo, J.S. (1978) Econometric modelling of the aggregate time-
series relationship between consumers expenditure and income in the United Kingdom Economic Journal, 88,
661692. Reprinted in Hendry, D.F. (1993), Econometrics: Alchemy or Science? Oxford: Blackwell Publishers.
8
2. Along a steady state growth path, (10) implies the APC (C,1 ) will depend on the growth
rates of the variables but not on the level of the variables. Steady state growth path charac-
terised by j
t
= q and c
t
= q, \t, then:
q =
1
q +
2
(j
t1
c
t1
),
q(1
1
) =
2
(ln1
t
lnC
t
)
since:
j
t1
= j
t
q, c
t1
= c
t
q.
Thus:
ln C
t
= ln1
t
+q

1
1

2
= ln
_
1
t
exp
_
q

1
1

2
__
C
t
= 11
t
, 1 = exp
_
q

1
1

2
_
Therefore, the EC model implies C proportional to 1 in equilibrium. A basic tenet of the LC
or PIH is this proportionality, e.g, the PIH is C
t
= /1
p
t
. PIH predicts proportionality between
C and 1
p
, rather than C and 1 , but over a suciently long span of time, 1
p
and 1 must be
about the same, so the PIH also predicts proportionality between C and 1 in the long-run.
Note that the EC model has the APC depending on the growth rate of income, q.
For the unrestricted model:
c
t
=
1
j
t
+
2
(,j
t1
c
t1
) +
t
where we assume 0 < , < 1, along a steady-state growth path the APC depends on the level
of income:
C
t
1
t
= exp
_
q

1
1

2
q(, 1)
_
1
'1
t
.
3. If c and j are 1(1) variables, then the restricted EC model is only a valid representation
of the data if c and j are cointegrated with a cointegrating parameter of unity. Then the
dierence terms in the EC model are stationary by virtue of being dierenced, and the EC
term is stationary by virtue of cointegration. For the unrestricted model, the cointegrating
combination needs to be [1 : ,] [c
t
: j
t
]
0
.
9

Das könnte Ihnen auch gefallen