roger webb wrote:
> >I'm generating Poisson regression models with an aggregated data
> >set (i.e. each record in the data set represents a stratum of
> >aggregated numbers of deaths and person-years of observation).
> >
> >I wish to check that the models are not over-dispersed. The manual
> >tells me that I can use either the 'poisgof' or the
> 'poisgof, pearson'
> >command. These produce the following contradictory results:
> >
> >poisgof
> >
> > Goodness-of-fit chi2 = 1191.579
> > Prob > chi2 (5304) = 1.0000
> >
> >poisgof, pearson
> >
> > Goodness-of-fit chi2 = 29207.21
> > Prob > chi2 (5304) = 0.0000
> >
> >A colleague has told me that these results have no meaning for my
> >data set, because the degrees of freedon are incorrect (I
> think). He
> >says that I should instead apply the Breslow adjusted score test
> >(Breslow NE. Generalized linear models: checking assumptions
> >and strengthening conclusion. Statistica Applicata 1996;
> 8: 23-41).
Irrespective of whether it's the correct test -- I am happy
to understand that it's not -- the difference in results is
a nice (or rather nasty) _numerical_ example of how which
chi-square measure you use can be crucial. That is, the two
sample statistics differ by a factor of 29207.21 / 1191.579 ~ 25.
Presumably lots and lots of small expected frequencies
are blowing up the Pearson measure.
And, as far as P-values are concerned, partly because
of the incorrect number of df, you flip from one tail to
the other...
Nick
[email protected]
*
* 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/
