Skip to lesson
OncoGuideeducationDiana’s wiki
THE EDUCATION LIBRARY

Interim analysis

Ask about this lesson

In one sentence

An interim analysis compares trial arms before the trial's planned end, using prespecified timing and stricter significance thresholds so that looking early does not inflate false-positive findings.

The intuition

If you check a coin-flip contest after every flip, sooner or later one side will look ahead by luck. Peeking many times raises the chance of declaring a false winner. A trial that plans to look early therefore sets a higher bar for each early look. The analogy has a limit: trials also look early for safety, and an early stop can be the right call.

How it works

  1. Plan the looks. The protocol fixes when analyses happen, usually after a set number of events (the information fraction: events so far divided by events planned) [1].
  2. Split the error allowance. The trial's overall false-positive allowance (often a two-sided 5%, called alpha) is divided across the looks, a method called alpha spending. Early looks get small shares, so their thresholds are strict [1, 2].
  3. Apply stopping boundaries. An independent data monitoring committee sees unblinded results. It can recommend stopping for clear benefit, for futility (a real difference has become unlikely) or for harm [1].
  4. Keep everyone else blind. Investigators learn only whether the trial continues [1].

Data maturity is how much follow-up and how many events an analysis has. An endpoint is "immature" when too few events have occurred to estimate it well, which is common for overall survival at early looks.

Real examples in breast cancer

TrialWhat happened at an interim analysis
KEYNOTE-522Pathologic complete response (pCR) was reported at the first interim analysis, among the first 602 randomized; event-free survival at the fourth planned interim analysis [3, 4]
OlympiAAt the first interim analysis, fewer deaths with olaparib (P = 0.02) did not meet that look's threshold of P < 0.01; at the second, overall survival met its threshold of P < 0.015 (P = 0.009) [5, 6]
CREATE-XMet its primary endpoint at a prespecified interim analysis and stopped early [7]
EA1131Stopped for futility at 61% of planned information: platinum was unlikely to prove non-inferior or superior to capecitabine [8]

Why it matters in cancer

  • Results for pCR, event-free survival and overall survival usually arrive years apart, at different looks, each with its own threshold.
  • "Not significant yet" at an early look is not "no effect": OlympiA's survival result crossed its boundary at the next look.
  • Trials stopped early for benefit tend to overestimate the effect. Across 91 such trials, the relative risks averaged 0.71 times those of comparable trials that ran to completion, a more favorable-looking effect [9].

Common confusions

  • An interim analysis is not an unplanned peek; unplanned looks undermine the error control.
  • A P value at an interim look is judged against that look's boundary, not against 0.05.
  • Stopping for futility does not prove the arms are equal; it shows a meaningful difference became unlikely.
  • Median follow-up is not the same as data maturity; event counts matter more.

Try it

A trial plans two survival looks. At the first, P = 0.03 against a boundary of P < 0.01. A headline says "survival benefit not shown." Is that right?

Answer: Partly. The result did not cross that look's boundary, so benefit was not established yet. It is not evidence of no benefit; the next look may cross its own threshold, as OlympiA's did.

Explain it back

"An interim analysis uses a stricter threshold because ___."

One possible answer: "Every extra look is another chance for a false-positive result, so the overall error allowance is split across the looks."

Takeaway

Read an early trial result with its look number, boundary and event count, not just its P value.

Sources and scope

Source-checked November 7, 2026 against ICH E9 and primary abstracts. General statistics education; expert and learner review pending.

References

  1. ICH E9: Statistical principles for clinical trials (FDA) — sections on group sequential designs and interim analysis and early stopping.
  2. O'Brien and Fleming, Biometrics 1979 — a widely used stopping boundary.
  3. Schmid and colleagues, NEJM 2020.
  4. Schmid and colleagues, NEJM 2022.
  5. Tutt and colleagues, NEJM 2021.
  6. Geyer and colleagues, Annals of Oncology 2022.
  7. Masuda and colleagues, NEJM 2017.
  8. Mayer and colleagues, JCO 2021.
  9. Bassler and colleagues, JAMA 2010 — ratio of relative risks 0.71, truncated versus nontruncated trials.

Used in