The Chow test is the standard method to
test for differences in regression response across groups. In some cases, the
groups being tested are composed of a time series of cross sections. For
example, when testing for differences across industries, each industry may be
composed of several observations on several individual firms. If the
individuals themselves have systematic differences, the Chow test will be
compromised: the individual and group effects become confounded. This can cause
rejections in the absence of the group effect of interest. We illustrate the
problem with a Monte Carlo analysis, and show that the effects cannot be
separated. We propose a bootstrap-like testing procedure that can eliminate
excessive Type I errors, and when used with the standard Chow test can help to
arrive at an appropriate conclusion when both effects are present.
JEL classification numbers: C01, C10, C12, C15, C18.
Chow test, Bootstrap, Empirical distribution, Statistical