Accounting for treatment switching by pairwise censoring: Is this needed, sufficient to mitigate attrition bias, or should more advanced methodology be used?

Background: In matched time-to-event analyses of treatment effectiveness, censoring rules for unexposed participants subsequently receiving treatment have been inconsistently applied. While some studies have used pairwise censoring, others have only censored the unexposed participant, understanding the degree of bias associated with these approaches would facilitate consistent application in observational studies and robust effect estimates.

Objectives: To simulate and compare the impact of five censoring strategies in time-to-event analyses.

Methods: We simulated 10000 matched observations- 5000 exposed participants matched to 5000 unexposed. We generated survival data using a piecewise exponential distribution over three time periods with varying baseline hazards. A hazard ratio of 0 was assumed. Of the unexposed group, 30% were randomly determined to have subsequently received treatment at random time points during follow-up. Five scenarios were evaluated: 1) no artificial censoring (base case); 2) only censoring the unexposed participant at the point of subsequently receiving treatment and allowing the treated counterpart to continue accruing follow-up; 3) pairwise censoring, i.e., both the unexposed participant and their matched treated counterpart are censored at the point of the unexposed subsequently receiving treatment; 4) scenario 2 with inverse probability of censoring weighting (IPCW), and 5) scenario 3 with IPCW. For scenarios 4 and 5, an additional continuous and a categorical covariate were generated. Cumulative incidence (Cin) plots and hazards ratios (HR) associated with the effectiveness estimate, and their 95% confidence intervals (CI) were then compared between the scenarios.

Results: The HRs (95% CI) for the three main scenarios were 1.01 (0.92 – 1.10), 1.19 (1.08 – 1.31), and 0.998 (0.90 to 1.11) respectively. Cin plots for Scenarios 1 and 3 showed similar event incidence in the unexposed and treated groups, but Scenario 2 showed a consistently lower cumulative event incidence in the unexposed group when compared to the treated group. The impact of the additional use of IPCW was found to vary depending on the strength of effect modification.

Conclusions: Not censoring the matched counterpart (of a censored participant) will yield biased estimates even with IPCW, due to differential follow-up accrual in the presence of changing baseline hazards (as seen during a pandemic), time of censoring, and attrition bias due to informative censoring. Pairwise censoring balances both follow-up accrual and the time of censoring between the two arms. Pairwise censoring with IPCW can help address the problem of informative censoring in the presence of strong effect modification.