
Re: Discrepancy in estimating goodness of Poisson GLM
Posted:
Jan 18, 2012 8:17 PM


OK, more progress!
I had originally expressed a doubt about one of my chisquaredbased tests, but I think I was suspecting the wrong one. The R function pchisq() gives the probability of getting a value less than its first argument, given the degrees of freedom specified in its second argument. So to calculate the probability of getting a value greater than that given, you subtract the output of pchisq() from 1.0
So, if I correct my likelihood ratio test, I get fantastic results for each model I've built. Some better than others, but all better than the null model. That's one discrepancy out of the way. On to figure out the others.
The goodness of fit test, based on chisquared test of residual deviance relative to degrees of freedom, is much better if I use a gaussian family rather than poisson, on transformed count data. So that's the other and last discrepancy out of they way!
Also, I'm less worried about the multiple tests issue now. The pruning of the original ten to the offered four independent variables (as mentioned in my second post), was in fact done by VIF (tolerance) analysis, removing significantly collinear variables until acceptable VIF (tolerance) scores remained. The other five offered were rationally (expertly, subjectively) selected from a much larger list, with no testing of their predictive power. So I'd hope that the significance of relationships I'm finding can more be seen as more reliable.
Thankyou gentlemen for your guidance on this one. I'm sure there are prickles left in my socks that could sting me on a warm day, but for now I'm going to hang up my pencils and call it a day.
Eric

