Ditto. I had a propensity-score matching phase for the past few years because I was collaborating with a lot of surgeons, for whom the answer to any problems of a nonrandomized treatment comparison is “just propensity-score match.” That’s left me somewhat down on the use of PS altogether, but I am sure that it has benefit in some applications, some of which has been described here. I really believe that a large part of PS-matched analyses in medicine is not tied to whether it provides a better assessment of the effect we’re trying to model, but the intuitive feeling that clinicians get of “matching” meaning that the results are now nearly as good as an RCT (or better). I’ve often chuckled at the way PS-matched analyses are described as “We matched patients by age, sex, race…” - no, you created a propensity score using those variables, and then matched the patients by their propensity scores.
Yesterday JAMA published an invited commentary about ‘Studying Drug Safety in the Real World’ https://ja.ma/2E8fL5c
They name 2 example studies where PS matching is used and then state:
Both studies discussed here use propensity-score matching, the aim of which is to mimic, more or less, the random allocation of an RCT by matching exposed to unexposed participants based on the probability of receiving the drug of interest, conditional on measured covariates.
After this they do mention that PS matching only balances on measured variables but I do believe that sentences like this result in people overestimating the amount of confounding control by PS matching.
I’ve interacted with David Juurlink a few times on Twitter (I’m not sure if he posts here - I don’t think so). I have a few bones to pick with the piece, but overall it was fairly reasonable (IMO). The very next sentence after your quoted passage is:
Although this approach can generate a nicely balanced Table 1, it is important to realize that differences are likely to persist between treated and untreated patients, and not all differences are measurable.
And the next paragraph leads with
The bigger issue is how these studies should influence practice, if at all.
So I don’t think this piece was written to be a dismissal of RCT’s nor to be a cheerleader for PSM studies. The authors have a point that, like it or not, we are moving into an era where people will have ready access to lots of EHR data for observational analyses, and so-called “real world evidence” studies testing treatment comps will become increasingly common (whether that is a good idea or not…)