teffects psmatch and common support with osample()

[Pawel Charasz]

Hello,

I would like to estimate treatment effects in a setting with limited overlap between treatment and control group. Stata command teffects psmatch has the option osample() to identify observations that violate common support, as defined in the command. I have seen other users to do something like this:

teffects psmatch (y) (treatment x1 x2 x3), osample(newvar)
teffects psmatch (y) (treatment x1 x2 x3) if newvar==0

However, if understand it correctly, this is not a very good option as running teffects psmatch the second time while restricting sample will result in propensity scores being estimated again, but using a different sample. What we would like to have is to reuse the same propensity scores estimated in the first command and use them to get treatment effects with the second command.

I am aware that this can be easily solved with the user-written command psmatch2 which has the option pscore() that we can use to "feed it" an already estimated propensity score. However, I would still prefer to use teffects psmatch as it implements the Abadie-Imbens standard errors.

Can anyone suggest some work-around to impose common support while using teffects psmatch?

[Pawel Charasz]

I am still hoping someone may have a good suggestion.

I have found this topic that discusses the osample() option in a different context and I am wondering whether the solution proposed by Jonathan Seiden is correct given the desire to get AI adjusted standard errors: https://www.statalist.org/forums/for...ffects-nnmatch

Basically, the solution proposes to do the following:

logit treatment x1 x2 x3
predict pscore
teffects psmatch (y) (pscore), osample(newvar)
keep if newvar==0
teffects psmatch (y) (treatment pscore)

I am trying to grasp the implications of this, both for point estimates and for standard errors. Any suggestions are welcome.