Adaptive Trial Design

An alpha-spending function is a curve that allocates the total Type I error (α, typically 0.05) across multiple interim and final analyses. At each interim look, the cumulative alpha allocated up to that point determines the p-value threshold for stopping. For example, the O'Brien-Fleming function allocates very little alpha to early interim looks (high p-value thresholds, hard to stop early), and more alpha to later looks and final analysis. This preserves Type I error at 0.05 overall while allowing multiple statistical tests. If you conduct 2 interim + 1 final analysis and use O'Brien-Fleming spending, the p-value thresholds might be 0.001, 0.01, and 0.05 at interim 1, interim 2, and final, respectively, cumulatively spending 0.05.

In fixed randomization (standard RCT), all participants are randomized 1:1 (or by pre-specified ratio) regardless of interim outcomes. In RAR, the randomization ratio adjusts based on interim outcome data. Example: if interim analysis shows Treatment A has 70% response rate and Treatment B has 50%, RAR might randomize new participants with 60% to A and 40% to B (favoring the better-performing arm). Rationale: it minimizes exposure to inferior treatment, ethical benefit. Challenge: RAR complicates inference (sequential testing adjustment required) and can bias effect estimates if not handled carefully. RAR is most useful when: (1) outcomes are observed quickly (not delayed), (2) goal is to maximize individual benefit (not just minimize sample size), (3) ethical concerns about exposure to inferior arms are high.

Pre-specify efficacy and futility stopping boundaries in the statistical analysis plan before trial initiation. Efficacy stopping: stop if interim data show a p-value (adjusted for sequential testing) that crosses the efficacy boundary—strong evidence of superiority. Example: if O'Brien-Fleming boundary crossed at first interim, stop and declare efficacy. Futility stopping: stop if interim data show the effect size is so small that, even enrolling all remaining planned participants, you are unlikely to achieve statistical significance. Futility is assessed via conditional power: 'Given the effect observed so far, what is the probability of reaching significance by final analysis?' If conditional power is <20%, consider stopping for futility. Always require independent Data Monitoring Committee (DMC) review before stopping decisions to avoid bias.

If an efficacy boundary is crossed at an interim look and enrollment stops early, the observed effect size at that point is likely inflated due to random variation. Imagine flipping a coin: on average, you expect 50 heads per 100 flips. But if you stop when you reach 70 heads (a rare occurrence), and count flips, you'll observe a biased ratio (70/100 = 70%). Similarly, in trials, if you stop when the interim effect estimate crosses a high threshold, that estimate is biased upward. For clinical decision-making, adjust interim estimates downward (using methods like p-value function analysis or conditional power calculations). The sequential testing p-value still controls Type I error, but the point estimate is optimistic.