4

Is anyone aware of propensity score matching methods for when there are more than 2 treatment groups? I am working on a project with 4 treatment groups:

  1. A
  2. B
  3. A and B
  4. Neither A nor B

Calculating propensity scores using multinomial logistic regression might work, but then I'd get multiple scores for each observation so I'm not sure how I'd match/analyze the matched data.

Ferdi
  • 5,179
Catherine Bulka
  • 91
  • Take a look at my answer here: https://stats.stackexchange.com/questions/405019/matching-with-multiple-treatments/405063#405063 – Noah Apr 25 '19 at 18:07

2 Answers2

4

It is not hard to do simultaneous covariate adjustment for multiple propensity scores. I recommend always using the logit propensity scale, and expanding those into restricted cubic splines. An example paper is Mark et al (1994) Circulation 89:2015-2025 where we analyzed three treatments.

Frank Harrell
  • 91,879
  • 6
  • 178
  • 397
2

Have a look at the mnps function in the twang package, described in this primer from the RAND Corporation. The package optimizes the tuning parameters of gradient-boosted tree models (highly flexible, non-linear, regularized) to user-specified sample-balance criteria. Then you can estimate the sample weights based on the optimized model.

Brash Equilibrium
  • 3,845