Interim Analysis and Stopping Rules

An interim analysis is an examination of trial data before the planned end, so a study can be stopped early for harm, for convincing benefit or for futility. Repeated testing creates the problem: each look is another chance to cross a significance threshold, so several unadjusted looks at the conventional 5 percent level push the overall false-positive rate well above it. Stopping rules spend the total error allowance across the planned looks, with the alpha-spending framework of Lan and DeMets keeping the schedule flexible while the total stays fixed.

The two classical boundary shapes behave differently. Pocock boundaries apply the same threshold at every look, making an early stop relatively achievable but leaving a reduced threshold for the final analysis. O'Brien-Fleming boundaries are extremely demanding early and relax to near the nominal level at the end, which is why they dominate in practice. Futility boundaries work in the opposite direction, stopping when continuing could not plausibly produce a positive result.

For a reader the consequence is that trials stopped early for benefit tend to overstate the effect. Stopping happens at a moment when random variation is running in the treatment's favour, and meta-epidemiological comparisons of truncated with completed trials on the same question have repeatedly found larger effects in the truncated ones.

The failure to recognise is an interim look with no prespecified boundary. Peeking at unblinded data and continuing until the p-value cooperates produces false positives regardless of intent, and it is invisible in a paper that never states when the looks occurred or what threshold applied. Equally, a trial stopped early for harm has demonstrated a signal but rarely has enough follow-up to characterise it, so reporting the halt as a quantified risk overstates what was observed.