[carbocation]

Not having prescribed either since I was a PCP years ago, I would just say that I agree with this line of inquiry.

If both drugs were equally good and were pretty much used at random, then they could be good instruments. But if there is confounding by indication, then it’s harder to get value out of an analysis that compares them.

It doesn’t mean the conclusions are wrong (and nor does my comment above), it just tempers my interpretation.

[elcritch]

You could try and reduce the possible effect by controlling for disease severity. If it reduced the effect a large amount it'd strengthen the case that the effect is actually due to RA disease severity. Actually I'd be a bit surprised if the authors didn't do that for an observational study.

[carbocation]

Agreed. I didn't see it described in their methods, although it seemed that the main focus of this paper (based on length dedicated to the material) was the biological/model data rather than the EHR data.