Potential Outcome Framework - Dichotomous Dependent Variables

[John Larsson]

Hello,
I am using Stata/SE 13.1 and am testing my data using the new teffects command.
This has five different estimators: regression adjustment, inverse-probability-weighted (IPW), augmented IPW, nearest-neighbor matching, and propensity-score matching.
My dependent variable is dichotomous. My basic question is this: are all the above estimators suitable for dichotomous dependent variables?
The reason I ask is that, according to the documentation, only with regression adjustment does one explicitly model a dichotomous outcome, such as with the "logit" option.
With some of the other estimators, one can specify "logit" but this appears to be to model the treatment, rather than the outcome. And in the case of the nonparametric option -- nearest neighbor matching -- there is no such option.
With any of these estimators, is the estimated Average Treatment Effect (ATE) a difference in probabilities of the outcome given the treatment as opposed to not having the treatment?
Thanks,
John

[Jeff Wooldridge]

John: All of the estimators you mentioned are valid with any kind of response variable, including dichotomous. One of the appealing features of the IPW and PS matching approaches is that only the propensity score needs to be modeled; the same method can be used for any kind of response variable. With matching on covariates, no functional forms are needed.

The two approaches where the nature of Y0, Y1 should be recognized are the ones that use regression adjustment. These are RA and IPWRA, and these are where Stata offers an option for the conditional mean specification (via "omodel"). The IPWRA estimator is a "doubly robust" estimator as discussed, say, in Imbens and Wooldridge, 2009, "Recent Developments in the Econometrics of Program Evaluation," Journal of Economic Literature, 2009.

Jeff

[John Larsson]

Hello Jeff, Thanks for the clarification on this. When I use teffects I am finding appreciable effect sizes in some cases where no such effect was found with logistic regression. I had assumed that generally the opposite would be the case -- that statistically significant effects found during a logistic regression might then be found to be zero when using these more robust effects estimators. Is this a red flag or are both phenomena plausible? Thanks, John