The Impact of Censoring Assumptions on STC-Adjusted Simulated Time-to-Event Data

OBJECTIVES: Unanchored Simulated Treatment Comparisons (STCs) are used to simulate a control arm, to estimate the survival outcomes and treatment effectiveness estimates for a single-arm trial (SAT) versus a competitor trial. An STC uses a regression model to predict events, however, assumptions are required to predict censoring. Formal guidelines for simulating censoring in STC are lacking, requiring researchers to carefully consider the most suitable approach. We explore how different censoring approaches impact survival outcomes.

METHODS: STCs were conducted using digitized pseudo-IPD for three oncology trials in myeloma, non-small-cell-lung cancer, and renal-cell carcinoma. Dummy baseline data for age and sex were created to perform the STC, ensuring these were nonsignificant treatment effect modifiers. Extreme differences between the SAT and competitor-trial (mean age: 60.1 vs. 20 years; females: 50% vs. 98%) were applied to highlight the adjustment. Following STC adjustment, we applied multiple censoring cases to the pseudo-IPD to observe how these influence survival outcomes. The following cases were explored; at trial-end, evenly spread across the trial-period, centered around the middle, predominantly early or late, and per-trial censoring based on occurrence-probability in the trial. Unadjusted and STC-adjusted Hazard Ratios (HRs) and median Overall-Survival (OS) were compared to assess the impact of each censoring scenario on survival.

RESULTS: Applying various censoring cases to STC-adjusted pseudo-IPD from the myeloma trial showed significant variability in HRs (0.81-2.13) between the unadjusted and STC-adjusted arms, and median OS (12.51-21.72 months). Per-trial-censoring had the closest approximation (HR 0.96), to the unadjusted arm (HR 1.00), with median survival of 20.08 months, most similar to the unadjusted median of 19.09 months. Findings were consistent across datasets.

CONCLUSIONS: Varying the censoring time assumptions in STC-adjusted time-to-event data demonstrated significant variation in survival outcomes. Per-trial censoring demonstrates the smallest error to the unadjusted arm and therefore, is recommended as best option for STC censoring.