On Jan 5, 2004, at 2:33 AM, Stephen wrote:
I follow your analysis. But I have a question: Are you implying that
if I use say "test" in place of the "lrtest" command it will produce
a Wald chi-square test statistic instead of Likelihood-ratio test
statistic?
Now concerning my comment on "mvreg" and Wald test, I was making a
reference to Stata manuals' statement that the "testparm" (a varaint of
the "test" command) can be used after all estimation commands except
"anova" for single equations, and "mlogit" or "mvreg" for
multiple-equations.
With regard to the first q, yes and no; if you use an estimator that
generates t-statistics in the coefficient listings, -test- will generate
a Wald F-test. If you use an estimator that generates z-statistics in
the coefficient listings (e.g. ivreg2 by default generates z-stats),
-test- will generate a Wald Chi-squared statistic. Stata considers that
small-sample standard errors should provide an F, and large-sample
standard errors should provide a Chi^2.
-whelp testparm- shows that testparm may indeed be used for -mvreg- (see
the description of equation(eqno) and that -testanova- is appropriate
after -anova-. That help file also specifies that -test- and -testparm-
perform Wald tests, and -help testnl- indicate that "Wald-type" tests
are performed via the delta method; it follows the same rule regarding
F- and Chi-squared statistics.
Kit
I followed the syntax you previously suggested in estimating the
constrained and unconstrained models, but used the "test" command
