I'd put it this way: if you buy the Hosmer-Lemeshow
definition -- and in practice this only makes much
difference with categorical or discrete predictors --
then -logit-'s definition is the one to follow. If not,
go for -glm, link(logit)-. The difference may seem
fortuitous, in that different commands follow different
literatures, but there's no reason to use a small random
perturbation except for demonstration purposes, as here.
Nick
[email protected]
Richard Williams
> At 10:13 AM 4/13/2004 -0400, VISINTAINER PAUL wrote:
> >Just as one more example, if you replace MPG with PRICE (a continuous
> >variable with one observation per value) in Nick's program,
> the -logit-
> >command produces the same dev's as -glm-, because each
> observation is a
> >unique covariate pattern. So if your model contains at least one
> >continuous variable, you are likely to get very close agreement among
> >commands, as well as programs.
>
> Also, if you don't mind results being slightly off in the 4th or 5th
> decimal place, I found that something like this works pretty good:
>
> drawnorm e, sd(.001)
> gen educ2 = educ + e
> quietly logit happymar church female educ2
> predict dev, deviance
>
> That is, you add some extremely small random number to one of your
> variables. In this case (and you should of course check this
> out) it made
> virtually no difference in the parameter estimates. But,
> every covariate
> pattern was then unique and I got deviance residual results virtually
> identical to SPSS's. Probably not the "best" way to do it
> but it is easy
> and seems to work. (Like I said before, the fact that such trivial
> differences in covariate values can create such big
> differences in the
> residuals is one of the things that strikes me as odd about
> the covariate pattern approach.)
*
* 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/
