Weights included in regression after PSMATCH2

[Fabian Becker]

I'm new to this forum, so please excuse if I miss anything I read in the FAQ, it is not on purpose:

I'm using Stata 13 with the current version of PSMATCH2 (downloaded last week at REPEC).

I want to test for the effects of firm characteristics on the labour productivity and one of the core variables is the reception of public support. As this variable is generally not random I implemented a propensity score matching with PSMATCH2 first:

psmatch2 funding var1 var2 var3, outcome(labprod) ps() ate
by _treated, sort: sum funding _treated _pscore var1 var2 var3

Is it then useful to implement the _weights as pweights in the regression or as fweights?

I used

regress labprod funding var1 var2 var3 [pweight = _weight], rob

I hope you can help me whether this is a valid approach to implement the propensity scores, as none of my colleagues here has used PSM earlier.

Best regards and thanks in advance.

[Chelsea Smith]

I believe you are correct, Fabian. Colleagues have told me to use the _weight variable as the pweight in subsequent regressions. One thing to note, however, that I asked about on the forum a few days ago, is that you might have a high number of missing on the _weight variable. To explore this issue, you can double-check the n on your regressions and/or browse the data editor. If you have a lot of missing and can't get around it, Sarah Edgington suggested on my forum thread to use inverse propensity score weighting.

[Sarah Edgington]

I'm not familiar with using the matching weights in regression analysis so I can't comment on this specifically.

One piece of advice that I received from a statistician regarding using propensity score models to estimate treatment effects is this: make sure whatever strategy you're using has literature that formally defines both the treatment effect estimate and its confidence interval. There seem to be a lot of strategies out there that despite being used in the published literature are ad-hoc and cannot be traced back to a methodological literature that lays out the properties of the estimated treatment effects.

From my point of view one major advantage of using built-in Stata commands is that the manual provides detailed references. This means that you can check whether the math behind the method you want to use has been well-justified. If you're new to this kind of analysis I would highly recommend starting with the teffects commands from Stata 13. Or at the very least spend some time reading the treatment effects manual.

I recommend inverse probability weighting because it seems somewhat similar to the match weighting strategy you're suggesting and has a clear literature that can be used to justify it. (I'm not saying the strategy you're proposing does not have a clear literature justifying it, just that if it does I don't know it. As long as you can justify how your estimates and your confidence intervals were created that's fine.) As a bonus since you're using the most recent version of Stata, IPW estimation can easily be implemented without any additional user-written commands.