Dear Statalisters,
I'm using Statalist 15.1. By using the "teffects" command in a case of binary outcome, I face the problem of a confidence interval for the proportion of events in the control group that's not admissible (in particular, I get a lower bound below 0). I've tried different models (logit, probit, linear probability model for the outcome only), different standard errors (robust, jackknife, bootstrap), different weighting systems (ipw and augmented ipw) and even the option "nls" and "wnls" (that I confess I don't understand) but nothing changes. I was hopeful in the bootstrap, but then I understood that it's not used to estimate confidence intervals directly, but just the standard errors (thus, for probability close to 0 or 1, the risk of not admissible values is always there).
The only solution I've found is to proceed with two-steps: using "teffects" to estimate propensity score, using it to create weights, and then perform a logit regression. But I'm afraid I'm losing efficiency because, in this way, I'm not taking into account that weights themselves are estimated.
Code:
teffects aipw (outcome, logit) (txcond varlist, logit)
predict ps, ps
generate weights=txcond/ps if txcond==1
replace weights=1-ps if txcond==0
logit outcome txcond [pw=weights]
Do you have any advise and how keep the efficiency of simultaneous estimation, by enforcing the constraint of confidence intervals restraint between 0 and 1?
Federico