On Feb 5, 2004, at 2:33 AM, Stas wrote:
So if heteroskedasticity is related to groups of observations, GQ test
is
likely to be the most powerful once you guessed the groups correctly.
If
it is related to a certain variable, then approaches like Cook-Weisberg
(Stata's -hettest-) would be likely to be the most powerful.
But the GQ test does not involve defining groups of obs: it merely
sorts the data on the basis of a SINGLE variable (which may or may not
be in the model) and then compares the residual variances for large
values and small values of that variable (usually leaving out
middle-sized values). When teaching this subject, I argue that the GQ
test is dominated by the Breusch--Pagan test (findit bpagan), which
allows you to specify a SET of variables that you might expect to have
some relation to the error variances across observations. In other
words, you don't have to "get the groups right"; maybe it's not net
sales, it's total assets, but if you put both in the B-P test, it will
pick up the relationship if it exists.
Also note that the common "White's general test" is a special case of
B-P. It has its own strengths and weaknesses, but I would think that
the two of them would clearly dominate the GQ test for any diagnosis of
H.
Kit
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
