A simulation study comparing backfill and randomized expansion in oncology dose optimization

Two commonly used dose-optimization designs are the two-stage randomized expansion design, in which dose escalation is completed at the first stage and then patients are randomized across selected dose levels for further evaluation, and the one-stage backfill design, which allows enrollment at previously explored doses while escalation is ongoing, integrating both efficacy and toxicity data to support a more robust selection of optimal dose. Although multiple methods have been proposed for each design, there has not been much comparison between the two designs in terms of the selection of the optimal dose. We conduct comprehensive simulation studies to compare these two designs in terms of optimal dose selection and other key evaluation metrics, both with and without the presence of imbalance factors. We assess and compare the performance of three dose selection methods - rule-based, utility-based, and model-based - under various scenarios. Additionally, we investigate the impact of futility monitoring on backfill designs and the consequences of excluding backfill patients from dose-escalation decisions. The simulations show that backfill designs can shorten study duration and reduce sample size while maintaining comparable dose selection performance to two-stage expansion designs in relatively homogeneous settings. Futility monitoring further improves efficiency with little loss of power, though backfill design is more sensitive to delayed efficacy or heterogeneity. Overall, backfill design is suitable in homogeneous populations, while two-stage expansion design may be more robust for heterogeneous settings.