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

1. Problem Statement

In modern oncology trials, particularly non-inferiority studies, there is a critical tension between statistical efficiency and patient safety.

• The Limitation of Traditional RAR: Most Response-Adaptive Randomisation (RAR) designs rely on efficacy data (e.g., Overall Survival or Progression-Free Survival) to dynamically allocate patients. However, efficacy outcomes in oncology are often delayed, making real-time adaptation difficult during patient enrollment.
• The Safety-Efficacy Trade-off: While early safety data (e.g., time to dose reduction or discontinuation due to adverse events) is available early, using it solely to drive allocation carries risks. If a treatment is better tolerated but less effective, a safety-driven RAR might inadvertently assign more patients to an inferior arm, compromising the trial's statistical power and ethical integrity.
• The Gap: Existing methods often fail to integrate early safety data with delayed efficacy data in a way that preserves statistical power for non-inferiority hypotheses while ethically prioritizing patient tolerability.

2. Methodology: The SAFER Design

The authors propose SAFER (Safety-Aware Flexible Elastic Randomization), a novel RAR design that leverages early-emerging safety data to inform allocation while safeguarding against efficacy loss.

Core Components:

1. Endpoints:

   • Efficacy: Time-to-event (e.g., Progression-Free Survival - PFS), modeled using an exponential distribution for simulation and a Cox proportional hazards model for inference.
   • Safety: Time to dose reduction or drug discontinuation due to adverse events (AEs).

2. Target Allocation (Neyman Allocation):

   • The design targets the Neyman allocation, which minimizes the variance of the treatment effect estimator.
   • Under the assumption of exponential survival, the Neyman proportion ($\pi_E$) is driven by the ratio of the mean safety times:
     $$\hat{\pi}_E = \frac{\hat{\theta}^{(s)}_E}{\hat{\theta}^{(s)}_E + \hat{\theta}^{(s)}_C}$$
     Where $\theta^{(s)}$ represents the mean time to dose reduction. This naturally favors the arm with better tolerability.

3. The SAFER Function (The "Elastic" Mechanism):

   • To prevent allocating patients to a safer but less effective arm, the design incorporates a weight parameter ($\hat{\Phi}$) based on the efficacy evidence.
   • $\hat{\Phi}$ is the Cumulative Distribution Function (CDF) of the Wald statistic ($W$) for the efficacy endpoint (PFS). It ranges from 0 to 1, representing the strength of evidence that the experimental arm is non-inferior.
   • The Logic:
     • If $\hat{\Phi} \leq 0.5$ (insufficient efficacy evidence), allocation reverts to 1:1 randomization to preserve statistical power.
     • If $\hat{\Phi} > 0.5$, the allocation shifts toward the safety-driven target ($\hat{\pi}_E$).
   • The Formula:
     $$SAFER(\hat{\pi}_E) = \begin{cases} 0.5 & \text{if } \hat{\Phi} \leq 0.5 \\ 0.5 + (\hat{\pi}_E - 0.5) \cdot \left(1 - \left(1 - \frac{\hat{\Phi} - 0.5}{0.5}\right)^\eta\right) & \text{if } 0.5 < \hat{\Phi} < 1 \\ \hat{\pi}_E & \text{if } \hat{\Phi} = 1 \end{cases}$$
   • Elastic Parameter ($\eta$): Controls the curvature of the adaptation. $\eta=1$ is linear (conservative); $\eta > 1$ is concave (aggressive).

4. Implementation:

   • The design uses a burn-in period (equal allocation) to ensure sufficient data accumulation before adaptation begins.
   • Allocation proportions are updated at pre-specified intervals (e.g., every 3 months) based on accumulated safety and efficacy data.

3. Key Contributions

• Novel Integration of Safety and Efficacy: SAFER is the first design to explicitly model the association between early safety data and delayed efficacy data to drive allocation in non-inferiority trials.
• Protection of Statistical Power: By introducing the $\hat{\Phi}$ threshold, the design ensures that allocation skewing only occurs when there is evidence of non-inferiority, preventing the loss of power associated with unequal allocation in non-inferiority settings.
• Flexibility: The design allows for tuning the "aggressiveness" of adaptation via the $\eta$ parameter and can be adapted to different trial structures (e.g., multiple interim analyses, Bayesian extensions).
• Handling Intercurrent Events: The paper rigorously evaluates the impact of informative drop-outs and under-reporting, proposing composite strategies and highlighting the need for careful estimand definition.

4. Simulation Results

The authors evaluated SAFER using a motivating example based on the CAPP-IT Phase III trial (capecitabine with/without pyridoxine) and conducted extensive simulations (10,000 iterations) across six scenarios:

• Scenario 0 (Independence): Confirmed that safety-driven RAR without the SAFER correction (pure safety) can distort inference if safety and efficacy are independent. SAFER successfully maintained Type I error control.
• Scenarios 1 & 2 (Association Levels):
  • When safety and efficacy are positively associated (better tolerability $\rightarrow$ longer PFS), SAFER successfully increased the allocation to the experimental arm, reducing the Adverse Event (AE) rate per patient-year while increasing or maintaining statistical power.
  • When association was weak, SAFER conservative shrunk allocation toward 1:1, preserving power.
• Scenario 3 (Elasticity): Demonstrated that higher $\eta$ values allow faster adaptation but with a slight trade-off in power if the association is weak.
• Scenario 4 (Timing): Showed that SAFER adapts faster when safety and efficacy endpoints are temporally aligned. When efficacy is significantly delayed, the design naturally slows adaptation, preventing premature allocation shifts.
• Scenarios 5 & 6 (Drop-out & Under-reporting):
  • Informative Drop-out: Using a composite strategy (imputing PFS at discontinuation) increased power but risked inflating Type I error if the safety benefit masked efficacy failure. The authors emphasize that under composite strategies, the estimand shifts to a joint safety-efficacy measure.
  • Under-reporting: As safety data under-reporting increased, the ability of SAFER to detect safety differences and adjust allocation decreased, as expected.

Key Metrics:

• Power: Preserved or enhanced (up to 1.00) when endpoints were strongly associated.
• AE Rate: Significantly reduced (e.g., dropping from 1.28 to <1.00 events/patient-year) compared to fixed allocation.
• Type I Error: Maintained near the nominal 0.05 level under the null hypothesis.

5. Significance and Implications

• Patient-Centric Oncology: SAFER aligns with the FDA's Project Optimus, which advocates for optimizing dosing to balance efficacy and toxicity. It allows trials to ethically assign more patients to better-tolerated regimens without sacrificing the ability to prove non-inferiority.
• Regulatory Relevance: The design offers a robust framework for late-phase trials where long-term tolerability is a key differentiator (e.g., chronic therapies, immunotherapies).
• Methodological Advancement: It resolves the classic trade-off in non-inferiority trials where unequal allocation is usually suboptimal for power. SAFER proves that conditional unequal allocation (based on joint safety-efficacy evidence) can be both ethically superior and statistically efficient.
• Future Directions: The authors suggest extending the framework to multi-arm trials, incorporating digital endpoints for more frequent updates, and exploring non-monotonic safety-efficacy relationships.

In summary, the SAFER design provides a mathematically rigorous and ethically sound solution for integrating early safety signals into the randomization process of oncology trials, ensuring that patient benefit is maximized without compromising the scientific validity of the study.