Generalized propensity score matching for multilevel treatment

[Nadia Ahm]

Hi,

I am a first-time user of the generalized propensity score. I have 3 treatment levels that are qualitatively different. I used mlogit and predict commands to estimate three sets of gpscore. But I am not sure how to match the pairs of treatments. I have read in several papers that the researchers used nearest neighbour matching to match the treated and control units. Can I use the same if I have a vector of propensity scores, and can I do it in Stata? If not, is there any other way to match the units based on the gpscores in Stata?

Will appreciate any help on this matter.

Thanks,
Nadia

[Fei Wang]

For multiple treatment levels, you may directly use teffects ipwra which integrates inverse probability (p-score) weighting and regression adjustment, and is doubly robust.

[Nadia Ahm]

Great, thank you so much!

[Jeff Wooldridge]

Like Fei, I'm a fan of ipwra. What is your outcome variable? ipwra supports binary, fractional, count, and other nonnegative responses with multiple treatment levels.

I've always been curious about how matching is done for multiple treatment levels. I guess one can always match each treatment level with a control unit to get the effects of going from untreated to the particular treatment level. Then, one would just use a binary comparison. So, if Treat = 1 is the control, then one could separately analyze the Treat = 1, Treat = 2 and Treat = 1, Treat = 3 samples separately. It seems pretty inefficient (beyond the usual inefficiency of matching).

[Nadia Ahm]

My outcome variable is binary. I am also wondering how matching can be done for three distinct treatment levels. I have read one paper where the authors mentioned three-way matched cohorts (Rassen, J. A., Shelat, A. A., Franklin, J. M., Glynn, R. J., Solomon, D. H., & Schneeweiss, S. (2013)—Matching by Propensity Score in Cohort Studies with Three Treatment Groups. Epidemiology, 24(3), 401–409. http://www.jstor.org/stable/23486755). But I am not sure how to do this in Stata.


For now, I think I will go ahead with propensity score weighting. Thank you very much for your help and suggestions.

[Jeff Wooldridge]

Theoretically, you can do better than PS weighting by using a logit outcome model for your Y and multinomial logit for the treatment in ipwra.

[Nadia Ahm]

Thank you very much Dr. Wooldridge, using ipwra for my analysis seems more meaningful.