Assessing methods for dealing with treatment crossover in clinical trials: A follow-up simulation study

Background: Treatment switching commonly occurs in clinical trials of novel interventions, particularly in the advanced or metastatic cancer setting, which causes important problems for health technology assessment. Previous research has demonstrated which adjustment methods are suitable in specific scenarios, but scenarios considered have been limited. Objectives: We aimed to assess statistical approaches for adjusting survival estimates in the presence of treatment switching in order to determine which methods are most appropriate in a new range of realistic scenarios, building upon previous research. In particular we consider smaller sample sizes, reduced switching proportions, increased levels of censoring, and alternative data generating models. Methods: We conducted a simulation study to assess the bias, mean squared error and coverage associated with alternative switching adjustment methods across a wide range of realistic scenarios. Results: Our results generally supported those found in previous research, but the novel scenarios considered meant that we could make conclusions based upon a more robust evidence base. Simple methods such as censoring or excluding patients that switch again resulted in high levels of bias. More complex randomisation-based methods (e.g. Rank Preserving Structural Failure Time Models (RPSFTM)) were unbiased when the “common treatment effect” held. Observational-based methods (e.g. inverse probability of censoring weights (IPCW)) coped better with time-dependent treatment effects but are heavily data reliant, and generally led to higher levels of bias in our simulations. Novel “two stage” methods produced relatively low bias across all simulated scenarios. All methods generally produced higher bias when the simulated sample size was smaller and when the censoring proportion was higher. All methods generally produced lower bias when switching proportions were lower. We find that the size of the treatment effect in terms of an acceleration factor has an important bearing on the levels of bias associated with the adjustment methods. Conclusions: Randomisation-based methods can accurately adjust for treatment switching when the treatment effect received by patients that switch is the same as that received by patients randomised to the experimental group. When this is not the case observational-based methods or simple twostage methods should be considered, although the IPCW is prone to substantial bias when the proportion of patients that switch is greater than approximately 90%. Simple methods such as censoring or excluding patients that switch should not be used.