Logistic regression, propensity score matching, or IPTW?

[Mohamad El Moheb]

Hello,

I would like to compare the "adjusted" mortality after surgery between two cohorts with very different sample sizes. The first cohort A is compromised of 100,000+ patients while the second cohort Bonly includes 500 patients. After adjusting for certain characteristics, I would like to see if there's a mortality difference between the two groups. Three methods come to mind: logistic regression, PS matching, and IPTW.

I prefer running a propensity score analysis over a multivariate logistic regression to show that the covariates are appropriately balanced between both groups.

I initially favored performing a propensity score matching. However, only a small fraction of patients from cohort A will be included in the final analysis. I am not sure if it is methodologically sound to do that given the huge difference in sample size.

IPTW may therefore be the best method, but wanted to confirm with people more experienced than me first!

[Felix Bittmann]

I had a similar imbalance in my data and I tested several versions (PSM, CEM, EB and IPTW). You can easily do that with kmatch (ssc install kmatch). Then you can check afterwards which method yields the best balancing result.

[Mohamad El Moheb]

Follow-up: I tried calculating the effect size using IPTW, and LR, and the OR of the IPTW method was very large. I found a paper where the authors also had the same problem with IPTW and they dropped observations PS scores <0.05 as they were responsible for the abnormally large OR. Is this statistically sound?

Also, I'm trying to match 1:3 using psmatch2 and neighbor(3) but the ratio of individuals matched is not exactly 1 to 3 (23% vs 77%). Any idea why?

Thanks