Difference in difference for gamma distribution

[mag_newbie]

Hey all,

I'm pretty new in SAS (version 9.4) and trying to calculate diff-in-diff for cost data.

The question is, if there is a difference in health care expenditures between people participating in a stuctured health care programme (Prog=1) and those who do not participate (Prog=0). Data for both groups is availabe for 1 year before intervention (post=0) and 1 year after intervention (post=1).

proc genmod data= cost;
class prog post / ref= first;
model total_cost= prog post prog*post / dist=gamma link=log;
estimate "diff_in_diff" prog*post  1 -1 -1 1;
lsmeans prog*post  / e linik;
lsestimate prog*post "diff_in_diff" 1 -1 -1 1;
run; 

Using proc genmod for diff-in-diff estimation provides results on means and on the standard errors of the means for all combinations (prog0_post0; prog0_post1; prog1_post0; prog1_post1).

However, the interpretation of the difference contrast of means and estimates is challenging for me.

I have read that a transformation using %margins macro may be usefull. 

I have downloaded the macro, but I found only very little information and no examples for applying it on data with gamma distribution.

Can someone tell me if I am on the right track and how I should proceed?

Grateful for any advice

Maggie

[StatDave]

One issue is whether the before and after groups in each participation group are independent or if each subject is measured both before and after. If subjects are repeatedly measured, implying some correlation, then you might want to include the REPEATED statement to identify the groups of correlated observations (subjects).

For difference in difference analysis with non-normal responses, see this note. You can use any of the Margins, NLMeans, or NLEstimate macros to estimate the difference in difference of the means. You might find using the NLMeans macro to be easier. Note that the NLMeans and NLEstimate macro are already available in your SAS session in the current release of SAS, SAS 9.4 TS1M6.

[mag_newbie]

Thanks for your response. Subjects are measured both before and after. Using NLEstimate macro worked to estimate the difference in difference of the means.