2D-03 - Mind the gap: a simulation study to compare definitions of treatment discontinuation

Background: Treatment discontinuation (TD) is used in oncology studies as both a proxy efficacy endpoint and a censoring event for safety outcomes. In real-world studies, TD is typically operationalized as having a gap in treatment administrations or dispensation that exceeded a maximum allowable gap (max tx gap). However, no study has empirically evaluated how different ways to operationalize TD impact outcome evaluation.


Objectives: We investigate how the choices of max tx gap and TD date (i.e., before or after max tx gap) in real-world settings impact the performance of time to TD and time to a target event (e.g., an adverse event of interest) while on treatment.


Methods: We simulated cohorts of 5,000 patients. For each patient, we randomly generated the timings of treatment administrations, TD, and target event, with base parameter values informed by PD-L1 inhibitor treatment for advanced melanoma (treatment cycle: 21 days, mean time to TD 180 days). We varied the scenarios by lengthening treatment cycles, adding treatment holidays, and varying time to TD and target event. Each scenario was simulated 1000 times, and the bias of the 1-year restricted mean survival time (RMST) for TD and on-treatment target event were estimated using two TD dates - day of last treatment before the max tx gap (PRE) and day after the max tx gap (POST). We used 6 max tx gaps from 15 to 180 days. For each estimate in each scenario, the max tx gap that achieved minimal bias was referred to as the optimal tx gap. RMST estimates +/- 10 days of the truth using the optimal tx gap were considered minimally biased.


Results: For time to TD, with optimally chosen max tx gap length, both the PRE and POST strategies achieved minimal bias in all scenarios. The optimal max tx gap length for the PRE strategy was typically longer than for the POST strategy, and the PRE strategy was generally more robust to choice of max tx gap length, achieving minimal bias with lengths between 90 and 150 days depending on the scenario. Extended tx intervals and tx holidays increased the length of the optimal tx gap. For the on-treatment target event, only the POST strategy achieved minimal bias for RMST across all scenarios, with optimal max tx gap lengths being slightly longer than the tx interval regardless of tx holidays.


Conclusions: For both outcomes, an optimally chosen max tx gap length was the most critical factor for minimizing bias. When TD was the outcome, the PRE strategy is more robust to mis-specification of tx gap length, whereas when TD was used to censor a target event, the POST strategy was preferred to obtain less biased estimates.