OBJECTIVES: Studies are commonly conducted to estimate effect of treatment switching by comparing outcomes between patients who initiated and continued one treatment (continuers) and patients who initiated one treatment but switched to another treatment (switchers). Analytical challenges may rise when the risk of clinical event of interest varies over time, which can be mitigated by assigning pseudo-switch date among continuers.

METHODS: Monte Carlo simulation was conducted to randomly draw the time from initiation to switch from switchers and assign it randomly to continuers to generate pseudo-switch date. Switch or pseudo-switch date was defined as start of follow-up. To achieve balance of switch and pseudo-switch time, time from initiation to switching were divided into 100 quantiles, and randomly assigned to continuers quantile-wise. In case the assigned pseudo-switch time fall beyond the follow up period of the continuers, assignment was repeatedly run until success to retain maximal samples. A simulation was conducted to evaluate impact of unbalanced distribution in pre-switch time between two groups on survival outcome. Application of the method to a study that assessed clinical impact of switching or continuation of apixaban or rivaroxaban among patients with non-valvular atrial fibrillation was assessed.

RESULTS: The simulation showed that using initiation/switching date as start of follow-up for continuers/switchers led to the estimated hazard ratio (HR) (0.82) to deviate from the true HR (0.7). After randomly assigning pseudo-switch time to make the distribution of pre-switch time balanced between continuers and switchers, estimated HR (0.68) was close to the true HR (0.7). Use of the method in the real-world study improved the balance of the distribution of pre-switch time between apixaban continuers and apixaban-to-rivaroxaban switchers, and between rivaroxaban continuers and rivaroxaban-to-apixaban switchers.

CONCLUSIONS: The proposed method for assigning random switching time to continuers reduced bias in estimating the effect of treatment switching.