2

Imagine in this tutorial variable catholic had 3 levels (0 = not Catholic, 1 = semi-catholic, 2 = fully Catholic) instead of 2 levels that are shown in the tutorial.

Under this new condition, how could we estimate propensity scores for matching purposes (not a logistic regression for sure)?

Reza
  • 936
  • You will want to use multinomial logistic regresion. See #4 here: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3710547/#:~:text=Multinomial%20logistic%20regression%20is%20a,treatments%20%5B33%2C%2015%5D. – Alex Jul 31 '20 at 21:23
  • But then, you have two different treatment effects, corresponding to semi-catholic vs. not catholic and fully-catholic vs. not catholic, right? If so, you can just repeat the propensity score estimation on the two subsamples. If not, what treatment effect are you interested in, in fine? – Roland Jul 31 '20 at 22:42
  • See also here: IPTW for multiple treatments – Noah Aug 01 '20 at 04:36
  • @morouszian PS methods are not appropriate for samples that small. Use regression instead. See Leyrat et a. (2019) for missingness in covariates. The answer is always multiple imputation. See my master's for a review of covariate measurement unreliability. You can also look into the SEM literature for small sample sizes, e.g., Rosseel (2020). – Noah Aug 02 '20 at 19:03
  • Looking into this, it seems a few papers indicate that PS can be used with small samples. See Piracchio et al. (2012) and Andrillon et al. (2020) for recommendations. I personally wouldn't even trust an RCT with 20 participants per group, and PS methods only reduce that size. – Noah Aug 02 '20 at 22:36
  • I understand that, but there are methods that are more robust and precise than PS matching, like difference-in-differences. I don't know anything about Bayesian methods, but they exist for PS. If by multi-levels, you mean multiple levels of treatment, the linked post gives a reference. If you mean students nested within classrooms (e.g., where a "multilevel" model would be used), Leite et al. (2015) is my favorite. – Noah Aug 02 '20 at 23:05
  • At this point I'm going to ask that you do your own research, ask a new question on CV, or hire a statistical consultant. – Noah Aug 02 '20 at 23:12

0 Answers0