Safety-Driven Response Adaptive Randomisation: An Application in Non-inferiority Oncology Trials

Abstract

The majority of response-adaptive randomisation (RAR) designs in the literature rely on efficacy data to guide dynamic patient allocation. However, their applicability becomes limited in settings where efficacy outcomes, such as survival, are observed with a random delay. To address this limitation, we introduce SAFER, a novel RAR design that leverages early-emerging safety data to inform treatment allocation decisions, particularly in oncology trials. The design is broadly applicable to contexts where prioritizing the arm with a superior safety is desirable. This is especially relevant in non-inferiority trials, to demonstrate that an experimental treatment is not inferior to the standard of care, while potentially offering improved tolerability. In such trials, an unavoidable trade-off arises: maintaining statistical efficiency for the efficacy hypothesis while integrating safety-driven adaptations through RAR. The SAFER design addresses this trade-off by dynamically adjusting the allocation proportion based on the observed association between safety and efficacy endpoints. We illustrate the performance of SAFER through a simulation study inspired by the CAPP-IT Phase III oncology trial. Results show that SAFER preserves statistical power, reduces the adverse event rate, and offers flexible adaptation speed depending on the temporal alignment of the endpoints.

Figures (5)

• Figure 1: Target function of the SAFER design for two target allocations, $\hat{\pi}_E = 0.7$ and $\hat{\pi}_E = 0.9$, each evaluated under three shape parameters $\eta = 1, 2, 5$. The SAFER function increases monotonically in $\hat{\Phi}$ and reflects more aggressive allocation as both $\hat{\pi}_E$ and $\eta$ increase.
• Figure 2: Relationship between the ratio of average time-to-dose reduction/drug discontinuation and the ratio of average PFS in experimental (E) vs control (C) arms, at different level of endpoints association: independent, moderate, strong, very strong). Bootstrap (10,000 iterations) 95% confidence intervals.
• Figure 3: $SAFER(\hat{\pi}_E)$ vs. enrolment period, under varying endpoint proximities (A: Median PFS=3 months; B: Median PFS=9 months; C: Median PFS=18 months; D: Median PFS=24 months;). True target allocation value by simulation setting=0.8; IF=0.5; $\eta$=5; Association between endpoints: very strong.
• Figure 4: $SAFER(\hat{\pi}_E)$ vs. enrolment period, under varying endpoint proximities (A: Median PFS=3 months; B: Median PFS=9 months; C: Median PFS=18 months; D: Median PFS=24 months;). True target allocation value by simulation setting=0.8; IF=0.5; $\eta$=5; Association between endpoints: weak.
• Figure S1: Theoretical loss of power (resulting from equations 6 and 7 of the main Manuscript) for the primary endpoint analysis, corresponding to an Information Fractions of 50% and allocation proportions $\pi_E$, from 0.5 to 0.8.