topic Re: Poisson Distribution in Statistical Procedures

Poisson Distribution

moealmo91 — Tue, 17 Jul 2018 05:10:17 GMT

Hi all,

I am working on a crash count data, using GENMOD... it sounds that I only get significant results if I use Poisson distribution. However, I am getting Devianc/DF= 0.64 which is away from 1 >>>> that means poission is not suitable. I have also tried Negative binomial... nothing worked --- please I need help?

Re: Poisson Distribution

Ksharp — Tue, 17 Jul 2018 09:56:54 GMT

I think it is OK.

Devianc/DF= 0.64 which means your model is not overdisperse.

While Devianc/DF>1 which means your model is overdisperse. that is not good.

Re: Poisson Distribution

StatDave — Tue, 17 Jul 2018 13:50:48 GMT

That is as much an indicator of fit as over- or under-dispersion, so you might want to investigate alternative models such as models including interactions and such. If you are still concerned about dispersion, there are alternative models such as the Conway-Maxwell Poisson, the generalized Poisson, or GEE models as mentioned near the end of this note. As also noted there, you can explicitly model the dispersion using the DISPMODEL option in PROC COUNTREG.

Re: Poisson Distribution

moealmo91 — Tue, 17 Jul 2018 16:58:28 GMT

Hi,

Thanks for your prompt response. Yes, I am really concerned about dispersion. However, the only models I know are NB and Poisson. would you please help me with the SAS code since I couldn't find any source for those models you mentioned. Here is the SAS code.

proc genmod data=odot.crossover;
model C_F_P = lnaadt Installation_length__miles_ Number_of_Lanes / dist=poisson link=log; run;