[6415] in s-news-athena
Re: Poisson Regression
daemon@ATHENA.MIT.EDU (Alan M. Zaslavsky)
Wed Feb 8 16:22:39 1995
Date: Wed, 8 Feb 95 15:44:06 EST
From: zaslavsk@hustat.harvard.edu (Alan M. Zaslavsky)
To: s-news@utstat.toronto.edu, ychang@welchlink.welch.jhu.edu
From: ychang@welchlink.welch.jhu.edu (YONG CHANG )
> Is there anyone out there who has encountered an increase of residual
> deviance over null deviance in poisson regression analysis?
> The second model with "yearrisk" variable has an increase of
> residual deviance over null deviance, which is what really puzzles
> me.
>
> Call: glm(formula = death ~ yearrisk + offset(log(pyrs)), family = poisson,data=
> all.process[all.process$pyrs != 0, ])
> Null Deviance: 5990.903 on 47 degrees of freedom
> Residual Deviance: 6571.964 on 46 degrees of freedom
I suspect that what is happening is that implicitly the null model is
death~1, i.e. constant rate for all cases. The correct null model for
testing the effet of yearrisk is death~offset(log(pyrs)), i.e. constant
rate times pyrs. Of course it is interesting that the offset (for
exposure, I suppose) doesn't improve the fit of the model over ignoring
exposure, but that seems to be the case. You should make the
comparison explicitly to test whether including the offset improves the
fit, but clearly it doesn't since the offset plus another variable
(with an estimated coefficient) is still a worse model than the null
model.
