Propensity Score Matching with a non-binary treatment?

[Andrea Molino]

Hi all,

I am doing an analysis, and am wanting to use propensity score matching (specifically teffects). My treatment/exposure of interest has 3 levels (aka it's non-binary, with levels of 0, 1 and 2).

Is there an option to allow for this using teffects? I'm not seeing anything online, and none of the options seem to allow for this.

Thanks in advance.

Andrea

[FernandoRios]

Hi Andrea
The treatment can be multinomial,
This is what it says in the helpfile

The outcomes can be continuous, binary, count, fractional, or nonnegative. The treatment model can be binary,
or it can be multinomial, allowing for multivalued treatments.

Here the example:

    Estimation of multivalued treatment effects under exogeneity.
        . use http://www.stata-press.com/data/r13/bdsianesi5, clear

    Using teffects ipw.
        . teffects ipw (wage) (ed math7 read7 maed paed)

[Jeff Wooldridge]

I don't think there is that option. In principle, for someone in group 2, one should be able to create a closest match in the other two groups, and then effectively impute the outcomes Y(0) and Y(1). My guess is that there are issues in computing the standard errors that may not have been resolved, but I;m no expert in matching. Have you checked the methodological literature?

In the meantime, I recommend trying ipwra, which is a doubly robust estimate that combines regression adjustment with propensity score weighting. As long as you have sufficient overlap -- which is needed for any approach to be convincing -- the ipwra approach works well. I proposed this version in my 2007 Journal of Econometrics paper, "Inverse Probability Weighted Estimation for General Missing Data Problems," I discuss it in more detail in my 2010 MIT Press book (Chapter 21).

[Jeff Wooldridge]

Fernando: Your command is for inverse probability weighting, not matching. RA, IPW, and IPWRA all allow multiple treatments.

[FernandoRios]

You are absolutely right. teffects psmatch does not allow for multiple treatments.

[Andrea Molino]

Hi all,

Thank you so much! This was all very helpful. I ended up going with IPWRA. Seems I have sufficient overlap (something I checked for before I ran my models just to get a sense of the covariate distributions). The teffects IPWRA also allowed me to use survey weights without have to generate an IPW*survey weights variable and then implement, which is great.

Although I think there is sufficient overlap and balance (after running tebalance summarize), is there a way to trim weights with teffects ipwra? I'm having a hard time finding anything, and it'd be nice to just set a 0.1 to 10 limit just to have peace of mind.

Thanks all,

Andrea