teffects: How to truncate ("trim") inverse propensity weights within teffects

[Joshua Thorpe]

We noticed a peculiar finding when we used the "brute-force" inverse propensity weights (IPWs) via [pw=IPW] compared to the results from teffects. Results (average treatment effects) would flip from positive to negative. We noticed some very extreme IPWs...at both extremes.

We would like to consider various recommendations in the literature for "truncating" IPWs. While easy to do using the "brute-force" svy-weighted approach above, we'd like to top-code (and not exclude) several unusually high IPWs.

However, it is not clear how IPWs can be modified within the teffects command framework.

Thoughts/suggestions would be greatly appreciated.

Thank you,
Josh Thorpe

[Philip Jones]

I also have this question. Truncation would seem to be a useful addition to the teffects commands that use ipw.

[Joerg Luedicke (StataCorp)]

Hi everybody, a quick comment on this: if the inverse-probability weights are getting too large, it means that the propensity score is too close to zero or one. This in turn means that there is a problem with the overlap assumption to a point where the treatment effect may not be empirically identifiable. In that case, one may need to improve the treatment effects model or obtain better data, or perhaps use a different estimator. At any rate, the practice of propensity score trimming bears no statistically sound derivation and doesn't solve the problem. At times, it may even be considered an ad-hoc way of just sweeping it under the rug. If the propensity score is very close to zero or one but the researcher considers the overlap assumption to still hold, it might be best to avoid IPW estimators in the first place and use a different estimator such as a matching estimator.

Best,
Joerg