Hi Statalist,

My colleagues and I are using propensity score matching to evaluate the impact of an educational intervention. We’re intending to use propensity-score matching, and estimate the impact using a regression model on the matched sample. We are unsure whether we should use teffects psmatch or psmatch2.

We understand that teffects psmatch produces more efficient standard errors than psmatch2 as the former accounts for propensity scores being estimated (Abadie and Imbens, 2016). We are using a many-to-one caliper matching and we want to retain the weights to use with our regression model. However, teffects psmatch doesn’t create these (whereas psmatch2 creates a _weight variable). We have found syntax online that can be used to estimate the weights for a 1:1 matching:

https://www.ssc.wisc.edu/sscc/pubs/stata_psmatch.htm

We have been able to replicate for a one-to-one matching, but despite trying to adapt it for a many-to-one match we can’t produce weights that would be consistent with psmatch2.

Do we need to use teffects psmatch given that we are going to estimate the effect size from a regression model using the matched sample? Are the efficiency gains going to be lost anyway, in which case, we should use psmatch2?

If it does make a difference in this scenario, how can we estimate the matching weights?

We have been using the example dataset from Stata16 for teffects psmatch:

webuse cattaneo2

Thanks