Standard errors for ATE with psmatch2

[fishermo]

I'm using the Stata command "psmatch2" to estimate average treatment effects where the treatment is agricultural extension and the outcome is adoption of improved maize varieties. The reason I use "psmatch2" rather than "teffects nnmatch" is I use a mixed logit model to account for clustering on study villages in estimating the propensity score. Unfortunately standard errors for ATE are not provided for psmatch2. Does anyone know if it is possible to (1) account for clustering within communities using the teffects command, or (2) calculate standard errors for ATE with psmatch2?

Here is my Stata code

xtmelogit [outcome] [explanatory variables] || [cluster variable]:, intpoints(10)

predict mypscore

psmatch2 [treatment], pscore(mypscore) n(5) out(outcome variable) logit com ate

Thanks! Monica

[Melissa Garrido]

It's unclear to me why you are accounting for clustering in the creation of your propensity score. Creation of the propensity score and nearest neighbor matching based on the propensity score occurs at the level of your unit of observation.

psmatch2 does provide standard errors for the ATE, but they are not adjusted for the fact that the propensity score was estimated in a separate step from the calculation of the treatment effect. If you are interested in ATE (as your code indicates), the standard errors provided in psmatch2 will be conservative.

With nearest neighbor matching, you need to use the Abadie-Imbens method (see references below) to obtain correctly adjusted standard errors. This is not available within psmatch2.

Abadie, A., and G. Imbens 2008. “On the Failure of the Bootstrap for Matching Estimators." Econometrica 76(6): 1537-1558.

Abadie, A., and G. Imbens. 2012. “Matching on the Estimated Propensity Score.” National Bureau of Economic Research Working Paper No. 15301. Available at: http://www.nber.org/papers/w15301. Last accessed November 8, 2013.

Hope this helps,
Melissa

[fishermo]

Thanks, Melissa. This does help, especially the information about the Abadie-Imbens method.

However, I disagree that psmatch2 provides standard errors for the ATE. When I use psmatch2 it provides standard errors for ATT but not for ATE (or ATU). See also http://www.ssc.wisc.edu/sscc/pubs/stata_psmatch.htm where the output for psmatch2 does not provide standard errors for psmatch2.

Likewise if I try to bootstrap the standard errors for ATE using the command

bs "psmatch2 qpm_aware, pscore(mypscore) out(qpm2007) common" "r(ate)", reps(100) dots

I always get the error message "statistic r(ate) evaluated to missing in full sample". The same error was encountered by another Stata user, see

http://www.stata.com/statalist/archi.../msg00223.html

Any ideas of how to get standard errors for the ATE using psmatch2?

[Melissa Garrido]

I'm sorry, you are right. -psmatch2- does not provide standard errors for the ATE. Also, please make sure you have the most up-to-date version of -psmatch2-. I just double-checked the help file, and they have recently added an option to calculate Abadie-Imbens standard errors for ATT.

Sorry for the confusion.