Logistics of treatment effect estimation with interim analyses

These are probably simple concepts for you guys but I am trying to understand the practicalities of analysing data from a group-sequential design where the outcome is continuous and you plan to measure each participant at multiple time points (longitudinal). Say you plan a trial with 3 interim analyses and stopping rules for efficacy. At the 2nd look (50% of the complete data available), you have evidence of efficacy and terminate the study.

In a fixed sample design one might analyse the treatment effect for the primary endpoint using a mixed model with a treatment x time interaction term. I read somewhere that naive treatment effects are biased in group sequential designs if terminating early (or even at completion)? Does this mean that the typical approach of using said mixed model with treatment by time interaction won’t work and some other method needs to be used to estimate an unbiased treatment effect? (If so, how is this done in R)

Hope that makes sense and I am not confusing some basic concepts.

I think the two issues need to be kept separate. For frequentist designs, any method that stops early for evidence of an effect in the right direction will overestimate the magnitude of the effect if the effective sample size at the point of stopping is not very large. The correction for this bias is complicated and investigators always conveniently forget to do it. For Bayesian designs the prior perfectly pulls back the posterior mean at the stopping point so there is no issue.

Whether the output is univariate or longitudinal doesn’t affect this very much, except that if the follow-up period is very long secular trends may get in the way a bit. A bigger issue with the longitudinal design is that mixed models that allow only random slopes/intercepts and no serial correlation pattern in addition are unlikely to fit the correlation structure. See here.

1 Like