What are most credible references guiding trial reporting when an important interaction is found? We found this in a recent RCT and then reported treatment effects in the subgroups. Review editors not convinced. Hoping to bolster my manuscript w best references on this. Thanks
I think this varies a lot by area. In psychiatry and psychology trials, interactions are very often investigated and reported. I think this is less common elsewhere, but if you’ve gone through all the various procedures to validate it and it makes clinical sense it is hard to see an objection. If it doesn’t make clinical sense, then you may want to do some causal modeling as well. I found Gelman’s vignette on this helpful for trying to validate the interaction in a previous study.
There is a lengthy discussion in Hernan’s book (Ch 5) on what is required from a causal perspective to show an interaction: https://cdn1.sph.harvard.edu/wp-content/uploads/sites/1268/2019/10/ci_hernanrobins_1oct19.pdf
See one analytic strategy laid out in https://hbiostat.org/bbr/md/ancova.html#differential-and-absolute-treatment-effects . Though the norm, it is not appropriate to compute estimates in subgroups. Rather get the best model-based estimates available, with uncertainty bands. This is especially important when one or more of the interacting factors was originally a continuous variable, as subgrouping is entirely improper in that case.
Model-based estimates inherit the appropriate baseline covariate adjustment and handle continuous variables correctly.
The primary graphical display has the interacting factor on the x-axis and the estimated treatment difference on the y-axis, with uncertainty bands. Even better is to use Bayesian information borrowing to estimate the treatment effect with more precision, if you specify a prior distribution for the interaction effect.