Discussion about this post

User's avatar
Devan V Mehrotra's avatar

Brennan - excellent write-up! Having served as the lead PhRMA rep on the ICH E9(R1) expert working group, your balanced perspective on perceived/actual hits and misses of the addendum is a welcome addition to this forum. I hope a lot of people read and learn from it. -- Devan Mehrotra

Expand full comment
Andrew's avatar

Nice post. I think the issue on time-varying confounding is driven by the fact that FDA only provide guidance on adjusting for baseline covariates, and EMA do so as well, but also explicitly advise against including time-varying confounders part of the primary analysis.

It helps to remember what the definition of a confounder is - related to exposure and outcome, and not on the causal pathway. This is all well and good in an epidemiological study on observational data, but for an RCT, for the variable to be related to treatment, there must be an imbalance between the treatment arms. It is not just that the covariate changes over time, but it does so differentially across arms. And such changes will thus be (likely) caused by the randomised treatment.

These are also trials for drugs which in general have a very specific mechanism of action. So it is difficult to justify how these can not be on the causal pathway, if it is also related to outcome. This argument potentially holds less water if you have more complex,, but for Ph 3 trials aimed at regulators, this is rare.

So the challenge is to find a time-varying confounder (not a covariate), where controlling for it or otherwise impacts the estimation of the treatment effect. It should not be underestimated how incredibly rare this is. Even then, you have to pre-specify and nail your colours to the mast of which analysis you are going to trust, especially as you know you are likely on the casual pathway. Such pre-specification is always going to be more challenging if you don't know what covariates are going to actually be varying, by how much, and if predictive. The EU regulators have done so - they request the unadjusted as primary, and discuss albeit briefly what to do with the adjusted analyses.

Expand full comment
1 more comment...

No posts