Identifying informative censoring from censoring patterns across successive follow-ups.

In time-to-event analyses of clinical trials, some data will inevitably be missing at the time of data cut-off. The Kaplan-Meier estimator aims to address these "incomplete observations", which has become a cornerstone of statistical methods in oncology trials. However, for the Kaplan-Meier principles to apply, censoring must occur randomly.

If patients drop out for reasons related to treatment, this may lead to a remaining patient population with a different underlying risk of experiencing the event, producing biased estimates referred to as informative censoring. Here, we introduce the concept of an "informative censoring area", defined as a time period over which informative censoring more likely occurred. We demonstrate how examining the evolution of censoring patterns over time can help distinguish between informative and non-informative censoring.

Using two clinical trials as examples, we show that comparing data from different follow-up periods reveals distinct patterns: NADINA trial showed early censored patients progressively disappearing with longer follow-up, indicating non-informative censoring, while NATALEE trial demonstrated stable early censoring patterns over time, reinforcing concerns of informative censoring. This approach helps identify when early censoring remains significant even with longer follow-up. We conclude that studying the evolution of censoring patterns over time may help differentiate between informative and non-informative censoring, reinforcing the need for systematic data sharing including reasons for censoring in trials seeking regulatory approval.