Hello,

I am estimating a treatment using differences-in-differences. The subjects in my treatment sample were allowed to drop out, which probably induced a bias in my estimate. I've been considering different methods for estimating a treatment effect on the treated, including LATE/CACE, but am having a difficult time finding how to apply these methods in a DiD framework.

Do you think it would be plausible to impute the "missing" observations after the treatment subjects dropped? The assumption would be that the subjects that dropped continued on the same trend they were on, and this would allow me to retain a full sample.

This wouldn't be a perfect estimate, of course, because it's possible that the subjects don't continue on their same path...and I worry that the variability of imputed values would be unrealistically small.

Any comments would be appreciated.

Thanks!