Built-in probability distribution in PROC GENMOD

[Quantopic]

Hi all,

I'm trying to understand what is the meaning of 'Built-in probability distribution' and the assumption behind in the PROC GENMOD framework.

Particularly, in the MODEL statement, one can use the DIST= option, where the DIST string can assume the following categories:

• BINOMIAL;
• GAMMA;
• GEOMETRIC;
• ...

as suggested in the SAS Documentation.

In the case one omits such distribution, the PROC assumes automatically that the distribution is normal.

Now, such distribution assumption implies that the data underlying the model are distributed according to the selected distribution?

For instance, in the case one sets DIST=BINOMIAL and LINK=LOGIT one implicitly assumes that the dependent variable is binomial distributed? 

This holds in a bayesian framework too?

Thanks all for the help in advance!

[Rick_SAS · Accepted solution]

No. The assumption is that the residuals (errors) are distributed according to the specified distribution. For a discussion and examples, see the article "On the assumptions (and misconceptions) of linear regression."