Re: st: Measures of fit in clogit

 From "TANIGUCHI, Naoko" To statalist@hsphsun2.harvard.edu Subject Re: st: Measures of fit in clogit Date Tue, 01 Feb 2005 18:26:07 -0500

Dear Clive Nicholas;

perfect, and I agree with your interpretation.

Yes, I just try to compare the fit of 2 models which are very similar
to each other.  Depending on Scott and Freese (2001)  Regression Models
for Categorical Dependent Variables Using STATA, pp.82-87, I searched
for a good measure of for clogit.

The count R square seems good because it is said that it can describe
'the proportion of correct prediction (i.e. the correct predicted vote
/ the observed vote).  But this measure seem contradict to others.

Model A

Maximum Likelihood R2:         0.615     Cragg & Uhler's R2:            0.615
Count R2:                      0.538

Model B

Maximum Likelihood R2:         0.609     Cragg & Uhler's R2:            0.609
Count R2:                      0.557

McFadden's R2, Maximum Likelihood R2,and Cragg & Uhler's R2 in the model A
are better than those in the model B.  But only Count R2 says the model
B is slightly superior to the other.  I know your interpretation that these
two models made little difference still right....

Thank you!

Naoko Taniguchi

||  Visiting Scholar
||  Department of Political Science
||  University of Michigan
||  6642 Haven Hall, 505 S State St
||  Ann Arbor MI 48109-1045

>> // Which model of them does a good prediction?
>
>Assuming that 'utilities' means differing propensities to vote for
>candidate/party X (almost certainly on a Likert scale), the fit statistics
>generated by -fitstat- are almost identical. For instance, the McFadden
>R^2 in Model A = .297; in Model B = .292. Therefore, I would conclude that
>changing your voting propensity scale from one to the other makes very
>little difference to your model's overall prediction of vote choice. (And
>since you're using -clogit-, then presumably your model is seeking to
>predict the _change_ or _stability_ in candidate choice from one election
>to the next for the ith voter.)
>
>> // How do I get the adjusted count Rsquare in clogit?
>
>I'm not entirely clear by what you mean here, but if you simply want an
>adjusted R^2, and you're happy with McFadden's R^2, then the adjusted
>McFadden R^2 displayed to the right of the 'pure' version ought to be good
>enough for you.
>
>> // How do I get a cross table of the observed X the predicted values ?
>> ('lstat' has already been installed, but does not work well...)
>
>I'm afraid I've not used -clogit- in Stata enough to answer this, as I
>couldn't find anything in -whelp clogit- to answer this properly. All I
>can suggest is to -predict- and compare with the observed values. Not as
>nice as having it all in a contingency table, but better than nothing!
>
>> // What is the best measures of fit in clogit?
>
>This has to be your call, but it is worth pointing out that virtually
>every fit statistic across your two models broadcasts the same message:
>changing utility scale makes little difference to your model fit. It is
>interesting that you don't mention whether or not changing utility has any
>impact upon the sign and/or magnitude of the _other_ three variables. If
>they have little or no impact, this indicates that your model is robust to
>such changes.
>
>I hope all this helps.
>
>CLIVE NICHOLAS        |t: 0(044)7903 397793
>Politics              |e: clive.nicholas@ncl.ac.uk
>Newcastle University  |http://www.ncl.ac.uk/geps
>
>*
>*   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/
>

*
*   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/