Naoko Taniguchi replied:
> 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
[...]
> Count R2: 0.538
>
> Model B
[...]
> Count R2: 0.557
If you've read pages 93 and 94 of Long and Freese (2001) - which gives the
formula for count R^2, you'll know why. Precisely because "...it is
possible to correctly predict at least 50 percent of the cases by choosing
the outcome category with the largest percentage of cases" is certainly
the main reason that I wouldn't trust this statistic as far as I could
throw it.*
Via -fitstat-, Long and Freese do calculate the adjusted count R^2, so
called because it adjusts for the largest row marginal (formula given on
page 94). However, in all of the fixed-effect logit models I've fitted to
my own electoral data (as well as to play data such as from -webuse
union-), -fitstat- never once offers the adjusted count R^2 statistic
(even when including just one (dummy) independent variable), presumably
because it cannot be calculated after -clogit-. I'd be interested to find
out from others if they have successful experiences of this.
I repeat my advice from earlier: if you value what the R^2 of your model
is telling you, put most of your cash on reporting the adjusted McFadden's
R^2.
CLIVE NICHOLAS |t: 0(044)7903 397793
Politics |e: [email protected]
Newcastle University |http://www.ncl.ac.uk/geps
References:
Long JS and Freese S (2nd ed., 2001) REGRESSION MODELS FOR CATEGORICAL
DEPENDENT VARIABLES USING STATA, College Station: Stata Press.
*Personal best: 11 feet, 2 inches at the All-England Count R^2-Throwing
Championships at Sunningdale last August.
*
* 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/
