What you likely are looking for is a Poisson regresion with the response, dollars spent, considered as a discrete variable. Use the log link, which is the default (and the canonical) and health status as a binary predictor. Fitted values (predict mu, mu) will provide you with expected dollars for a certain covariate structure. If there is only one perdictor, health status, then its pretty easy. If you use a offset, remember to log it (natural log) in order to put everything into the same parameterization. Using -eform- as an option provides incident rate ratios (analogically to eform with the logit link providing odds ratios). The interpretation of the predictor, health status, should be fairly straighforward. A quick look under -glm- in the reference manual will help.
Note: check the dispersion statistic for the Poisson model. If it is overdispersed, consider the negative binomial.
There are several good books (I think so) published by Stata Press that deal with these types of models.
Joe Hilbe
> Hi,
>
> I'm predicting health care expenditures as a function of health status.
> Since the ependiture data are rightly skewed I am using the generalized
> liniear model (GLM) with a Log link. My question is: can I transform the
> coefficients on my regressors into actual dollars? Meaning, can I say that a
> poor health status cost so many dollars more compared to a
> good health
> status? Pl., advise.
>
> Thanks.
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
