Presentation (CTI)

OBJECTIVES: In oncology trials, progression-free survival (PFS) is inherently interval-censored: progression is detected only at scheduled clinic visits, so the true event time lies within a known interval rather than being observed exactly. Despite this, progression is commonly assumed to occur at the scheduled visit. This simulation study quantified the impact of different approaches to interval-censoring on bias, long-term extrapolation and restricted mean survival time (RMST).
METHODS: A single-arm trial (n=200) with maximum follow-up of 36 months was simulated with true progression times following a Weibull distribution and deaths following an exponential distribution. Four non-parametric estimators were compared, assuming 3-monthly scheduled visits: a reference truth estimator (no interval censoring), a right-censored Kaplan-Meier, a midpoint-censored Kaplan-Meier, and a Turnbull non-parametric maximum likelihood estimator. Bias relative to the truth was estimated for each estimator. Parametric extrapolation was undertaken using the seven standard HTA distributions for each estimator, comparing estimated survival at different landmarks and RMST.
RESULTS: All three approaches introduced absolute bias in the survival curve through the observed period, with right-endpoint and midpoint censoring producing larger bias than the Turnbull estimator. This absolute bias also distorted the parametric model selected based on AIC, causing fits to diverge substantially from those based on the truth reference beyond the data cut-off. At a 10-year horizon, survival estimates from the censoring methods ranged from 0.77 (right-censoring) to 7.50 (midpoint-censoring) times the truth reference, corresponding to survival of 1.00% to 9.79%. At 50-year horizons, RMST differed from the truth reference by up to 18 months.
CONCLUSIONS: These findings highlight that the choice of censoring mechanism for progression events is not only a technical detail but can be a major source of bias with direct implications for survival extrapolation and potentially cost-effectiveness conclusions.