dear s-plus users,
thanks to brian ripley, hugh jones, spencer graves, pedram sendi, and harry
southworth for taking time to answer my question.
prof. ripley wrote:
There is a very good reason! t-values for logistic regression can be
highly misleading: small values indicate either a small effect or a very
large effect (the `Hauck-Donner' phenomenon). The distribution theory is
far too complicated, as the asymptotic approximations are not good enough.
(the t-values are not t-distributed).
Use dropterm() in library MASS to obtain likelihood-ratio tests for each
term in the model: those are more reliable. Also use profiling (?profile)
to obtain confidence intervals.
harry, spencer, and hugh argue similarly, but suggest to use
anova(glm.object, test = "Chisq") instead of dropterm():
The tests you mean are based on the asymptotic normality of the parameter
estimates and that assumption is frequently very poor. You should use
likelihood ratio tests instead....
If you specify anova(glm.object, test = "Chisq") then you will obtain an
analysis of deviance table with probabilities for Chi-square distribution.
best regards
bernd puschner
On Wed, 16 Jul 2003, Bernd Puschner wrote:
> when computing a logistic regression model, either via glm or
> menuBinomialGlm, the summary of the model provides coefficients and their
> standard errors and t-values, but for some reason no significance tests
> (p-values) which i would like to have. any ideas?
|