Discussion about this post

User's avatar
Tristin Beckman's avatar

The wide variability of bodies and injuries is a great illustration of heterogenous treatment effects, and why estimating these "scientifically" is probably impossible in this case.

Body types and injury types plausibly interact with the treatment, but this Gelman post from years back (https://statmodeling.stat.columbia.edu/2018/03/15/need16/) shows, with some assumptions, that you need about 16 times the sample size to estimate an interaction than the main effect. Higher order interactions would blow this up even more. Given the already tiny sample sizes (31 people on the study you cited on Monday's post), there probably won't ever be enough data to know if something works given the widely accepted practices, even assuming a perfectly run experiment.

Steph's avatar

I'm loving your discussion of comparing the general (RCT) with the personal. I think about a lot of the same issues with trying to reconcile the fact that my PhD topic (causal modelling, causal inference) which seems like it should universally help reasoning is actually hard to apply to my own health

4 more comments...

No posts

Ready for more?