Adjusting for differential proportions of second-line treatment in cancer clinical trials. Part I: Structural nested models and marginal structural models to test and estimates treatment arm effects

In randomized trials, post-randomization variables such as compliance, prescription of alternative treatments and so on are usually ignored to compare treatment arms. Intent-to-treat (ITT) analysis is a standard approach but it does not adjust for those variables. However, we may need to evaluate treatment arm effects that have the desired causal interpretation. Previously proposed methods such as time-dependent Cox model may not properly adjust for post-randomization variables and may produce biased results. Alternatively, we propose to use two causal models, structural nested models and marginal structural models. The two models appropriately adjust for such variables. We apply these models to adjust for differential proportions of post-randomization second-line treatment in cancer clinical trials. With sufficient care to several assumptions, these methods, especially structural nested failure time models with randomized analyses, are useful to take the influence of second-line treatment into account and to test and estimate the direct treatment arm effect.