Predict after poisson

[Jacob Petersen]

Dear All

I'm about to perform an analysis that aims to predict how much money is used for elderly care. $ 0 is used for the majority of the elderly (97% of the population). However 3% of elderly between 100 $ and 1 million $ are used. I am using poisson-regression rather than OLS, as the distribution of expenses is very skewed (many zeros).
My question: How do I get predicted values after poisson which indicate the expected amount of money used for each of the elderly? Had I used OLS I would have used the following command:

reg expense sex
predict pr_expenses

Should I just use the following command?:

poisson expense sex, vc(robust)
predict pr_expenses

Best regards
Jacob

[Carlo Lazzaro]

Jacob:
welcome to this forum.
With such a relevant proportion of zeros, it is difficult to get any useful inference from your data. If I were you, I probably stop at descriptive stats.
That said, if you really want to regress, you should consider a zero inflated poisson (see -help zip-).

[Jacob Petersen]

Thank you Carlo
I will try zero inflated poisson (zip).
However, I'm still not sure which of the options to use (if any) when using predict. The STATA manual says that the defualt for predict (after the zip comande) "calculates the predicted number of events, which is(1−pj) exp(xjβ), where pj is the predicted probability of a zero outcome." I assume this is as close as it gets to predict (default) after regress, but please correct me if I am wrong.

[Carlo Lazzaro]

Jacob:
I would consider Example 1 in -zip postestimation- entry, Stata .pdf manual.

[Marcos Almeida]

I'd also consider a glm model with gamma family and log link. Another option: 2 models. One for "zero" expense (logistic) and the other for the expenses.

[Jacob Petersen]

Thanks a lot.