F-tests continued

Introduction
Discuss the problems associated with
structural breaks in the data.
Examine the Chow test for a structural
break.
Assess an example of the use of the
Chow test and ways to solve the problem
of structural breaks.
Introduce the problem of multicollinearity.

Structural Breaks
Structural breaks can occur in time series data
or cross sectional data, when there is a sudden
change in the relationship being examined.
Examples include sudden policy changes such
as a change in government or sudden move in
asset prices (1987) or serious international
disaster such as a civil war
We then need to decide whether 2 separate
regression lines are more efficient than a single
regression.

Structural Break in 1997

Structural Break
In this example a single regression line is
not a good fit of the data due to the
obvious structural break in 1997.
We need to test if a structural break has
occurred in 1997, usually the break is not
as obvious as this.
We will use the Chow test, which is a
variation of the F-test for a restriction

Chow Test (stages in using test)

Run the regression using all the
observations, before and after the
structural break, collect the RSS
Run 2 separate regressions, one before,
RSS(1) and one after, RSS(2) the
structural break.
Calculate the test statistic using the
following formulae:

Chow Test
RSS c ( RSS1 RSS 2 ) / k
F
RSS1 RSS 2 / n 2k
RSS c combined _ RSS
RSS1 pre break _ RSS
RSS 2 post break _ RSS

Chow Test
The final stage of the Chow Test is to
compare the test statistic with the critical
value from the F-Tables.
The null hypothesis in this case is
structural stability, if we reject the null
hypothesis, it means we have a structural
break in the data
We then need to decide how to overcome
this break.

Chow Test
If there is evidence of a structural break, it
may mean we need to split the data into 2
samples and run separate regressions.
Another method to overcome this problem
is to use dummy variables (To be covered
later in term), the benefit of this approach
is that we do not lose any degrees of
freedom through a loss of observations.

Chow Test Example

The following model is regressed using data
in quarterly form from 1990 to 2005 (64
observations)for Malaysian stock prices
against output (structural break in 1997).

st 0 1 yt ut

Chow Test
The first regression using all the data produced a
RSS( c) of 0.56, then 2 regressions were run on a
sub-sample of the data from 1990-1997, giving a
RSS(1) of 0.23. The final regression was on the
sample from 1998 to 2005, producing a RSS(2) of
0.17, n=64, k=2.

0.56 (0.23 0.17) / 2

0.08
F

12
0.23 0.17 / 64 2 * 2 0.00667

Chow Test
As the critical value for F(2,60) =3.15(5%)
As 12> 3.15, we reject the null hypothesis
of structural stability.
We conclude that there is a structural
break in this model, we need to split the
data into 2 sub-samples or use another
method to overcome the break.

Problems with Chow Test

The test may suggest splitting the data,
this may mean fewer degrees of freedom
When should the cut off point be for the
test, usually there should be a theoretical
basis for this.
There is the potential for structural
instability across the whole data range. It
is possible to test every observation for a
structural break.

Multicollinearity
Multicollinearity occurs when two explanatory
variables are strongly correlated with each other.
In all multiple regression models there is some
degree of collinearity between the explanatory
variables, however not enough to cause a
serious problem.
However in some cases the collinearity between
variables is so high, it affects the regression,
producing coefficients with high standard errors.

Multicollinearity
It may be that multicollinearity is not a
problem if the other conditions are
favourable:
- High number of observations
- Sample variance of explanatory variables
is high
- Variance of the residual is low

Models which can have

multicollinearity
Models with large numbers of lags.
Models which use asset returns or interest
rates, i.e. 3 month and 10 year interest
rates. (This can be overcome by using a
term structure of interest rates variable,
i.e. one rate minus the other)
Demand models which include different
prices of goods.

Measuring Multicollinearity
The main way of testing for multicollinearity is to
check the t-statistics and R-squared statistic.
If the regression produces a high R-squared
statistic (>0.9) but low t-statistics which are not
significant, then multicollinearity could be a
problem
We could then produce a pair-wise correlation
coefficient to determine if the variables are
suffering from high levels of multicollinearity.
The problem with this approach is to decide on
when the correlation is so large that
multicollinearity is present.

Remedies for Multicollinearity

Remove one of the variables from the
regression which is causing the multicollinearity
or alternatively replace it with a variable that is
not collinear (This can cause omitted variable
bias).
Find data that has more observations.
Transform the variables, i.e. put data into ratio
form or take logarithms of the data
Ignore the problem, after all the estimators are
still BLUE.

Increasing Observations
To overcome multicollinearity, it may be
necessary to increase the number of
observations by:
Extending data series.
Increasing the frequency of the data, with
financial data it is often possible to get daily
data.
Pooling the data, it could be that cross section
and time series data could be combined.

Multicollinearity Example
st 0.4 0.8 yt 0.2lit 0.1sit
(0.9) (1.2) (0.4) (0.1)
2

R 0.98, (standard errors in parentheses)

(n 60). where :
st stock prices
yt output
lit long - run interest rates
sit - short - run interest rates

Multicollinearity
The example in the previous slide shows
evidence of multicollinearity, with a high Rsquared statistic and low t-statistics. T-statistics
are: y-0.667, li-0.5 and si-1, where the critical
value is 2.00.
The long and short interest rates are probably
highly collinear.
The remedy in this case is to use a different
variable and combine both interest rates into a
single variable, which is the long rate minus the
short rate. (called the term structure of interest
rates)

Dummy Variables
Dummy Variables are a common way of solving
structural breaks, as it does not involve splitting
the data.
These variables consist of 1s and 0s and are
often termed on-off variables.
They can be used to determine the importance
of policy actions on models and are often used
to account for qualitative effects.
Their coefficients and t-statistics can then be
interpreted in the usual way.

Conclusion
The F-test can be used to test a specific
restriction on a model, such as constant returns
to scale.
The Chow test is used to determine if the data is
structurally stable.
If there is a structural break, we need to split the
data or use dummy variables
Multicollinearity occurs when the explanatory
variables are closely correlated.