Mgmt 469
Performing a Chow (partial-F) test
Suppose you have a set of related variables X1-X10 that represent a single
theoretical construct such as a set of industry categories, a set of demographic
variables, or a set of year dummies. You are unsure whether to include the set of
variables in the model. To determine if they belong, you will need to perform a
Chow (or partial-F) test.
In Stata, follow your regression by typing:
test X1 X2 etc. X10
If the F-test statistic is not significant, then the category does not add predictive
power and you should throw it out.
Suppose that you ignore these instructions. Instead, you take a quick peek at the
coefficients on individual predictors and find that X7 is significant at p<.05.
Should you ignore the test procedure and include X7 in the model?
The answer is that unless you had an a priori reason to single out X7, you should
exclude it. The reason has to do with simple statistics. If a set of ten predictors
collectively add no predictive power to a model, there is still a high probability that
at least one will have a significant coefficient, just due to random chance.
Remember, the usual null hypothesis is that a variable is not significant unless
there is evidence that its observed coefficient could not be the result of random
chance. Observing an insignificant group and one significant member of that
group is entirely consistent with random chance.
If the test reveals that the group of variables does add predictive power, this only
means that including some members of the group add power. It does not mean that
all the variables belong in the model. You have two options:
· Keep the entire group if you have sufficient df
· Pick and choose those variable that add the most power. (Don’t forget that
if you have a group of categorical variables, the individual coefficients show
only comparisons with the omitted category. Be sure to include significant
comparisons.)
