teffects ipw- continuous outcome

[James Meltzer]

Dear all,

I am using the teffects ipw command to estimate the average treatment effect (ATE) of a binary exposure on a continuous outcome.

In the unweighted analysis, the continuous outcome is right skewed and I plan to report group medians and IQRs.

When using the teffects command, I am given the average treatment effect which is the difference in potential outcome means (weighted analysis).

Questions
1. Does STATA take into account that the outcome variable is skewed? Is it appropriate in this case to report the average treatment effect?
2. If not, should I transform the outcome variable (e.g. log transform) so that it is normally distributed in each of the treatment groups before using the teffects commands?
3. If not, is it better to calculate the difference in potential outcome medians using similar methodology to calculating an absolute difference in medians (e.g. cendif command)? This has been previously described on this forum but in my case I would weight the analysis using the ipw.

Any suggestions would be much appreciated.

Thanks

[Isabel Ferreira]

Were there ever answers posted to this question?

[Jeff Wooldridge]

The distribution of the outcome variable doesn't really impact the suitability of IPW as an estimation method. So there is nothing for Stata to "take account of." I think what you should be asking is whether an ATE is suitable for outcomes with skewed distributions. I think the ATE is still valuable. If I'm interested in, say, the effect of a job training program on labor market earnings, and I want to be able to aggregate the overall benefit to some population of workers, then I need the average. If before the program average earnings among the eventual participates was $30,000 and I estimate and ATE on the treated as $2,000, that is a meaningful number that I can even report as a percentage effect. If Y > 0 always, you can approximate a percentage effect by using log(Y).

Of course, one can be interested in effects on other parts of the distribution -- such as the median an quantiles. You can use IPW for that, too. I bet there's a user written command that does the Firpo (2007, Econometrica) method, but it's easy to do using regular Stata commands. But the standard errors are too conservative, so you should bootstrap both steps to get the proper SEs.

logit w x1 x2 ... xK
predict ph
gen ate_wght = w/ph + (1 - w)/(1 - ph)
qreg y w [pweight = ate_wght], q(.5)
qreg y w [pweight = ate_wght], q(.25)
qreg y w [pweight = ate_wght], q(.75)

[Isabel Ferreira]

Thanks Jeff