topic Re: Countreg nonsense (?) in Statistical Procedures

Countreg nonsense (?)

GVeers — Wed, 12 Aug 2009 15:53:55 GMT

I'm using Proc Countreg to fit a zero inflated poisson distribution. However, I'm either misinterpreting the Inflated statistics or they don't make sense:

Inf_Intercept estimate = 0.525351
Inf_Variable estimate = -0.783426

So now, looking at an individual observation with "Variable" = 0, I'd like to know the odds of that observation producing a zero (without considering the poisson distribution). My understanding is that this probability is

exp(0.525351) * exp(-0.783426) ^ 0

or just exp(0.525351) = 1.69

Obviously 1.69 is not a legitimate probability. What am I missing? By the way, it is fully possible for "Variable" to equal zero in this dataset.

Re: Countreg nonsense (?)

StatDave — Wed, 12 Aug 2009 18:41:54 GMT

The probability of a zero is not a function of only the zero-inflation parameters. There are two processes at play that each contribute to the probability. The probabilities of various counts (including zero) can be computed by the PROBCOUNTS macro available in this sample program:

http://support.sas.com/kb/26161

The Details section of the macro's description discusses further what I'm talking about. See the Long (1997) reference for even more detail.

Re: Countreg nonsense (?)

GVeers — Wed, 12 Aug 2009 21:00:16 GMT

Thank you very much, StatDave. That's just what the doctor ordered.