Propensity Score Matching

[Mila Marc]

Hi everyone, I am currently working on a database of patients who have all undergone a specific medical procedure. My study aims to investigate the clinical outcomes of this procedure in four different age groups. I am interested in using propensity score matching to account for any potential confounding factors that may influence the results of my study. Therefore, I would greatly appreciate your advice on whether this method is appropriate for my study. If so, could you kindly provide me with the appropriate Stata syntax to carry out this analysis?

Thank you in advance for your help and guidance!

[Matthew Holt]

PSM can only match on observables, so its only advantage, relative to just including those variables in the regression as controls, is that PSM relaxes functional form assumptions. For instance, in an OLS regression using a PS-matched sample, I no longer need to assume a linear relationship between my outcome of interest and an endogenous determinant. (PSM entails its own assumptions.)

My 2 cents: Entropy balancing is arguably a better method for addressing potential functional form misspecification. EB can be more efficient in both a mathematical and a practical sense, and because it involves fewer choices than PSM or similar matching methods, it's arguably less susceptible to "forking path" bias and p-hacking. If I had to use propensity scores, I'd use propensity score weighting instead of matching, for similar reasons.

I can provide some cites related to these points if useful.

To determine whether any of these techniques to address functional form misspecification are appropriate for your study would require institutional and theoretical expertise in your setting, so it may be more useful to ask a colleague than this forum.

Personally, though, I am pragmatic about these things. If a reviewer or senior coauthor wants PSM, I give them PSM. If every other study in my target journal is using PSM, I'll use PSM. I've used the community-contributed command psmatch2 (ssc install psmatch2). The help file for that command demonstrates its syntax.