Background: Progression-free survival (PFS) is a common endpoint in oncology, where a patient either has a PFS event (progression or death) or is censored (e.g., treatment discontinuation due to adverse events [AEs]). In common dose/exposure-PFS analyses using Kaplan-Meier plots or Cox proportional hazards models, non-informative censoring is assumed, i.e., censored patients would experience PFS events at the same rate as those not censored. However, when this assumption is violated, as potentially in cancer treatments, where frailer patients are more likely to discontinue treatment due to AEs as well as to progress, informative censoring bias occurs. This simulation study illustrates how this bias develops and can affect dose selection decisions. Methods: A dose-finding study comparing low and high doses was simulated using a competing risk multistate model developed to describe PFS in patients with renal cancer receiving sunitinib. The dataset included 750 patients per dose. The model was adjusted to meet the following conditions: (1) an equal progression rate between doses, (2) a higher rate of treatment discontinuation in the high-dose group due to AEs, and (3) a positive correlation between the hazard of progression and treatment discontinuation due to AEs (i.e., informative censoring). Non-informative censoring was simulated by omitting the third condition. The Cox proportional hazards model was used to evaluate the dose-PFS relationships. Results: With informative censoring, the Cox proportional hazards model indicated a lower risk of PFS events in the high-dose group than in the low-dose group, with a hazard ratio of 0.85 (95% CI: 0.75-0.97) and a p-value of 0.01. However, no relationship between dose and PFS was observed when the third condition was removed, resulting in non-informative censoring. The hazard ratio was 0.97 (95% CI: 0.85-1.10), with a p-value of 0.63. Conclusion: In the presence of informative censoring bias, the traditional analysis suggested greater efficacy for the higher dose even when the progression rate was equivalent in the simulations, which could mislead dose selection decisions. This study highlights the importance of carefully assessing whether the assumption of non-informative censoring holds before applying methods like the Cox model to PFS data.
