Methods/Statistics
Paired censoring in matched cohort studies Alexander Breskin* Alexander Breskin Michael Webster-Clark
Informative censoring remains a threat to internal validity in matched cohort studies even in the absence of uncontrolled confounding. Several methods exist to address informative censoring bias (e.g., inverse probability of censoring weighting (IPCW)). ‘Paired censoring’ analysis, in which both members of a matched pair are censored when the first member is censored, has recently been proposed and implemented, particularly in vaccine safety studies. Conditions under which this approach yields unbiased estimates of risks and incidence rates have yet to be elucidated.
We identified 3 necessary conditions for paired censoring to be valid. First, the outcome and competing events must not be terminal, meaning people remain under follow-up after their first event and can experience multiple events. Second, the treatment effect must be constant across levels of covariates driving informative censoring. Finally, only time-fixed covariates may drive informative censoring.
For demonstration, we conducted a simulation estimating the effect of a two-dose vaccine series on injection site reaction under four scenarios (Figure 1): a) non-terminal outcome, null effect, censoring depending on baseline covariates; b) non-terminal outcome, effect measure modification, censoring depending on baseline covariates; c) non-terminal outcome, null effect, censoring depending on a time-varying covariate; and d) terminal outcome. In each scenario, we estimated effects with a naïve estimate unadjusted for informative censoring and paired censoring. In scenario (a), the unadjusted estimate was biased while the paired censoring approach was not. In scenarios (b), (c), and (d), both estimates were biased, but the paired censoring approach slightly attenuated the bias.
Researchers should consider these conditions when accounting for informative censoring in matched cohort studies. If they are not met, alternatives to paired censoring such as IPCW should be used (given their own conditions are met).
