The Interim Analysis Nobody Pre-Specified
A 240-participant Phase 2b trial in moderate-to-severe plaque psoriasis randomizes 1:1 to an investigational IL-23 inhibitor or placebo, with PASI-75 at Week 12 as the primary endpoint. Historical placebo response is 30%; the sponsor assumed 62% on drug and powered at 90% two-sided alpha 0.05. At 60% enrollment the team proposes an unblinded interim analysis to re-estimate sample size, and separately proposes switching the primary endpoint to IGA 0/1 because "payers prefer it." The protocol describes no alpha-spending function, and the DSMB charter scopes the board to safety review only. One unblinded statistician holds both the unblinding role and a seat on the team drafting the protocol amendment. No approved biologic alternative exists for the enrolled population.
1. What must be in place before any unblinding occurs, and what specifically goes wrong if it is not?
Read the worked answer
The proposal as written cannot proceed; two independent controls are missing.
Alpha control. There is no pre-specified alpha-spending function. A sample-size re-estimation driven by unblinded comparative data inflates the Type I error rate above the stated 0.05, and the eventual p-value becomes uninterpretable to a regulator. A Lan-DeMets O'Brien-Fleming spending function must be written into the protocol and SAP before unblinding, together with the re-estimation rule itself: the timing of the look, the minimum and maximum allowable sample size, and the assumptions used to recompute. Without a floor and ceiling, the "re-estimation" is an open-ended opportunity to keep enrolling until the result is favorable.
Firewall. The same statistician cannot both see unblinded comparative results and shape the amendment. Anyone who knows the treatment effect will influence endpoint definitions, analysis populations, and covariate choices, consciously or not. Move the unblinded analysis to an independent unit with no protocol role and no communication channel to the study team.
DSMB scope. A safety-only charter cannot adjudicate an efficacy-driven interim. Amend the charter, add an efficacy review with a documented recommendation process, and record the recommendation in writing before the sponsor acts.
2. Can the primary endpoint be changed to IGA 0/1 at this point? Set out the conditions under which such a change is defensible and those under which it is not.
Read the worked answer
Defensible only inside narrow limits, and not on the stated rationale.
IGA 0/1 is a different construct from PASI-75: a static global severity rating versus a relative reduction in a continuous score. The two correlate but are not interchangeable, and the assumed effect size for one cannot be borrowed from the other.
Conditions that make the change defensible:
- The decision is made blind to comparative results. Justify it from blinded data only: the baseline severity distribution, IGA-PASI concordance in the pooled dataset, and published placebo IGA 0/1 rates rather than your own unblinded rates.
- The protocol, SAP, and statistical section are amended with a hierarchical testing strategy that protects alpha if IGA becomes primary and PASI-75 becomes a key secondary.
- The sample size is re-powered on an independently sourced effect estimate for the new endpoint.
- IRB approval and, for a registration-enabling study, prior regulatory agreement are obtained.
Conditions that make it indefensible: the change is discussed in the same breath as an unblinded interim, or the stated reason is commercial positioning rather than a scientific one. A reviewer will read a mid-flight endpoint switch made by people who have seen the data as outcome-driven, and no statistical footnote repairs that perception.
3. The statistician proposes counting anyone who stops treatment for lack of efficacy as a non-responder (non-responder imputation) while applying a treatment-policy approach to visits missed during a COVID-19 wave. Is that coherent? Justify using ICH E9(R1).
Read the worked answer
Not coherent as described, although a mixed strategy can be made coherent if it is designed deliberately rather than assembled from convenience.
ICH E9(R1) defines an estimand through five attributes: treatment condition, population, endpoint, handling of intercurrent events, and summary measure. The critical question here is the fourth. Discontinuing treatment for lack of efficacy is an intercurrent event. Under a treatment-policy strategy the participant's post-event data remain in the estimand and the participant is not automatically a failure. Applying non-responder imputation to those participants converts the estimand into a composite of efficacy and tolerability, and it will bias toward the null whenever the drug works but is stopped early by participants who were not responding-an event that is itself correlated with the outcome.
The coherent options:
- Treatment policy for both event types, with observed data analyzed under a pre-specified model and explicit, defensible missing-data assumptions.
- Hypothetical strategy for both, stated up front, with sensitivity analyses exploring departures.
- A genuinely pre-specified composite strategy, defined before unblinding and justified by the clinical question being asked.
Whichever is chosen must be fixed in the SAP before unblinding, and a tipping-point analysis should accompany the primary result to show how robust it is to departures from the missing-data assumption.