Partial Adjustment for Treatment Switching to Represent Expected Switching in Clinical Practice

OBJECTIVES: In some randomized controlled trials (RCTs), adjusting for all participants who switch onto post-study treatments may not represent what would occur in clinical practice (CP). Instead, switching might be anticipated in CP, but at a level different to that observed in an RCT. In such cases, analyses that adjust for only a proportion of switchers may better inform healthcare decision making. We extend existing treatment switching adjustment methods to address the situation where the proportion of switchers in an RCT differs from the proportion expected in CP.

METHODS: Survival data representing RCTs affected by treatment switching were simulated. Data were generated such that ‘true’ survival for a specified switching proportion in CP could be calculated. Scenarios were constructed such that the switching proportion present in the simulated RCTs differed from the CP proportion. We adapted two methods, inverse probability of censoring weights (IPCW) and two-stage estimation (TSE), to adjust the proportion of switching to match the proportion expected in CP. To achieve this, we had to determine which switchers would and would not switch in our counterfactual analysis. We tested two approaches: (i) random allocation with bootstrapping; (ii) allocation based on a modelled probability of switching. Performance of the methods was assessed by comparing estimated restricted mean survival times (RMST) to the CP truth.

RESULTS: The adapted IPCW and TSE methods performed better than intention-to-treat analysis across all scenarios. Deterministic assignment of counterfactual switchers based on modelled switch probabilities produced some bias while random allocation with bootstrapping had negligible bias.

CONCLUSIONS: When the proportion of switchers in an RCT differs from the proportion of switchers expected in CP, a proportion of switching can be adjusted for using adapted versions of adjustment methods combined with techniques for allocating counterfactual switchers and non-switchers.