topic Re: two-part models in Statistical Procedures

two-part models

proctice — Mon, 08 Aug 2016 01:54:51 GMT

The logit model part of a ZIP model models the extra zeros as the outcome, but my client wants the probability of a count greater than 0 instead. Is it appropriate to do 2 completely separate models - a proc logistic modeling count>0 and a proc genmod with the zeros excluded modeling the count part?

Re: two-part models

Rick_SAS — Mon, 08 Aug 2016 09:49:58 GMT

I think I would stick with the ZIP model. You can use GENMOD or PROC FMM. The model will predict the probability of a zero count, which is P(Y=p | x). The client wants "one minus that predicted probability": 1 - P(Y=0| x)

Re: two-part models

proctice — Mon, 08 Aug 2016 15:45:44 GMT

They are more interested in the odds ratios being in the right direction, can I just invert them?

Re: two-part models

Rick_SAS — Mon, 08 Aug 2016 15:52:25 GMT

You can invert odds ratios by using the EVENT= option on the response variable on the MODEL statement.

Instead of

model y(event="0") = x1 x2 x3;

you can invert the odds ratio by using

model y(event="1") = x1 x2 x3;