Kernel matching/weighting via psmatch2 followed by a logistic regression

[Rich Otein]

I'm attempting to estimate the ATT of a binary treatment variable (arrest) on a binary outcome variable (college enrollment) for a large observational study with a large number of (measured) covariates. I used propensity score matching, and in particular kernel matching/weighting using the -psmatch2- package. The code may be given as

 psmatch2 arrest, outcome(enroll) pscore(ps) kernel bwidth(.03) common

My understanding is that -psmatch2- will not fit a logistic/probit model, and instead uses a simple linear regression. I'm looking for (1) some guidance on how to properly conduct this analysis outside of -psmatch2-, while correctly employing the given _weights to attain the ATT, and (2) how to obtain proper standard errors that account for error in the estimated propensity scores (I don't think -teffects- accounts for kernel weighting).

From this post (though the question is slightly different) as well as some slides I managed to find online, it seems as though something like this is a recommended approach for (1):

logistic arrest ps if _weight == 1

although I don't understand how this is employing the matched sample if we're restricting it to treated only. I realize this defines the ATT, but I feel like I'm missing something in terms of this command. Should it not be something like

logistic arrest ps [pweight = _weight]

, or is this now the ATE?

Recommendations and guidance are much appreciated!

[Rich Otein]

I left out the response variable in the last two chunks of code. Should be:

logistic enroll arrest ps ...

[David Radwin]

Regrettably, it's not clear what the weights generated by psmatch2 (Leuven and Sianesi, SSC) represent. See http://stephenporter.org/understanding-weight-calculations-in-statas-psmatch2/ .

If you are instead trying to use weights provided with the dataset for the "match first, then perform logistic regression on the matched sample" approach, please see:

DuGoff, E.H., Schuler, M., and Stuart, E.A. (2014). Generalizing Observational Study Results: Applying Propensity Score Methods to Complex Surveys. Health Services Research, 49(1): 284–303.

For a general rationale for this approach, see:

Ho, D.E., Imai, K., King, G., and Stuart, E.A. (2007). Matching as Nonparametric Preprocessing for Reducing Model Dependence in Parametric Causal Inference. Political Analysis, 15(3): 199–236.