OBJECTIVES: Treatment switching commonly occurs in randomised controlled trials (RCTs). Participants may switch onto the comparator, or another treatment. Statistical adjustment methods have been used in health technology assessment to estimate outcomes that would have been observed in the absence of switching. The performance (e.g., bias, accuracy) of adjustment methods has been assessed in simulation studies, but these have focused on switching between randomised treatments. In practise, patients could switch onto a range of alternative treatments with different treatment effects. This study assesses the performance of adjustment methods in the presence of switching to multiple different treatments.

METHODS: Survival data was simulated with patients in the control group able to switch onto multiple treatments. Inverse probability of censoring weights (IPCW) and Two-stage estimation (TSE) adjustment methods were applied, testing methods that (i) considered and adjusted for each type of switching separately; (ii) combined switchers and adjusted for switching without differentiating by treatment switched to. Adjusted results were compared to the simulated truth in 40 scenarios. Switch proportions, treatment effects, censoring proportions, and sample size were varied.

RESULTS: IPCW and TSE applications that distinguished between the type of switch produced similar bias to applications which grouped all switches together. Modelling each type of switch separately did not result in more accurate adjustment analyses. TSE analyses performed well across all scenarios, while IPCW resulted in substantial bias in scenarios with high switching proportions.

CONCLUSIONS: In situations where it is appropriate to adjust for switches to multiple treatments in RCTs, adjustment methods can be applied in different ways. In the scenarios tested, there was little advantage associated with adjusting for each type of switch separately, compared with applications that combined types of switch together. TSE produced low bias across all scenarios, whereas IPCW performed less consistently.