Question: In the context of designing and analyzing clinical trials, particularly in the phase of evaluating the efficacy and safety of new drugs, how do you integrate complex statistical models like mixed-effects models or Bayesian hierarchical models to account for the inherent variability between individual patients, treatment groups, and potential confounding factors?

Connor Fitchett answered on 28 May 2025:

The answer is with lots of trial and improvement!

If we want to use complicated models, we have to be sure that they’re adding something of value over a simple model (Occam’s razor). For example, if we expect there’s lots of variability between patients or unknown confounding, we might use one of those model you talked about to account for this.

What we would do is look at past papers to see whether the method we would to apply is suitable for this situation. Once we’re convinced it could be interesting, we would go into simulations (or standard analysis if it’s tractable). Making different assumptions on how the patient data is generated (high variance, low variance, unknown confounding) we compare the different methods. If the method, ex mixed-effect modelling, explains the data a lot better than a simple model, ex linear modelling, then we’re justified to carry on with that and plan to use it on the actual data.

A lot of medical stats comes down to you have an idea, and you have to see if it works better (or similarly) than the current standard. With the examples you’ve given, you’d A) confirm if the data you’re expecting to see has confounding and B) confirm if the methods you want to use perform well on the confounded data (probably through simulations). If both of these are true, you’re good to go!