Hello. I'm using Stata 15.1.

I have a dataset of 88 observations in the treatment group, and need to find a control group. Since there are some variables wrt which I want to enforce a perfect match, I selected such control group through "calipmatch" (Stepner and Garland, 2017) creating the "match" variable equal to 1 for the first match,..., N for the Nth match. After selecting the control group, I want to estimate effects on repeated binary outcomes ("month" being the count variable). I've thought about this strategy:

1) xtgee outcome i.month group, family(binomial) eform

to get population-averaged effects,

and

2) melogit cont_bin i.month group || match:|| id:, or

to get individual-level effects.

Is this the correct strategy? I've found here:

https://www.statalist.org/forums/for...=1563786760018

the suggestion to introduce a match-level random effect in the case of continuous outcome, where there are no problems about non-collapsibility, so I guess this only has to do with standard error. I've found (with a fake dataset) that such approach leads to an odds ratio away from the null (1), consistently with the behaviour of conditional (vs marginal) odds ratio (OR). Thus, I think 1) is the model to estimate marginal ORs, 2) to estimate conditional ORs. I think that melogit cont_bin i.month group || id:, or would lead to an intermediate estimate between the two, however not corresponding to any relevant quantity because it would ignore that the control has been selected through matching, so that observations are paired. Is this correct?

Michael Stepner & Allan Garland, 2017. "CALIPMATCH: Stata module for caliper matching without replacement," Statistical Software Components S458342, Boston College Department of Economics.