Interim Analysis, Differential Loss to Follow-Up and a Composite Endpoint
An investigator-initiated multicentre open-label trial randomises 240 adults with sepsis-associated acute kidney injury to drug X plus standard care or standard care alone. The primary endpoint is a composite of death or initiation of renal replacement therapy at day 28. At the planned interim analysis, with 50% of events accrued, the composite occurred in 34% of the X arm and 43% of the control arm: hazard ratio 0.76 (95% CI 0.58–0.99), p = 0.045. The pre-specified O'Brien–Fleming boundary at this look is p < 0.005. Twenty-two participants (9%) are lost to follow-up, 15 in the control arm and 7 in the X arm. Six patients in the X arm discontinued for hypotension. The data safety monitoring board asks for your recommendation: stop for efficacy, continue, or stop for futility.
1. Interpret p = 0.045 against the pre-specified boundary and state whether the trial should stop for efficacy.
Read the worked answer
It should not stop for efficacy.
The pre-specified O'Brien–Fleming boundary at this look is p < 0.005, and 0.045 does not cross it. The boundary is not bureaucratic: it exists because repeated hypothesis testing inflates the type I error rate, so the nominal 0.05 is available only at the final analysis. A p of 0.045 at the halfway point carries far weaker evidence than the same p value at the end.
The confidence interval tells the same story. HR 0.76 with a 95% CI of 0.58–0.99 barely excludes 1, and the upper bound is compatible with a relative benefit of about 1% - statistically indistinguishable from no effect and far from a minimally important difference in this population.
Stopping now would also leave the hypotension signal (6 discontinuations in the X arm) poorly characterised. The defensible interim decision is to continue to the planned final analysis, with the DSMB reviewing unblinded data at the next scheduled look.
2. Quantify how the differential loss to follow-up could change the conclusion, using best-case and worst-case assumptions and identifying the tipping point.
Read the worked answer
Approximate per-arm n = 120, with about 41 events in the X arm and 52 in the control arm.
- Worst case against X: all 7 missing X patients had events and all 15 missing control patients did not. Then X = 48/120 = 40% and control = 52/120 = 43% - a 3-point absolute difference, HR near 0.9, with the CI crossing 1.
- Best case for X: the reverse assumption gives X = 34% and control = 56%, which exaggerates the effect.
The plausible range therefore spans 'no meaningful benefit' to 'large benefit', and the truth depends entirely on the missing outcomes. The tipping point is small: fewer than 7 events among the missing X patients would likely push the confidence interval across 1.
Because the trial is open-label, dropout is plausibly related to outcome - patients who deteriorate may transfer or withdraw - so missing-at-random cannot be assumed. Pre-specify multiple imputation plus a pattern-mixture or delta-adjusted sensitivity analysis, and report the range of estimates rather than a single number.
3. Critically assess the composite endpoint and the open-label design, then give a final recommendation to the DSMB.
Read the worked answer
Composite endpoint. Death or initiation of renal replacement therapy is patient-relevant, but the two components are unequal in importance, and RRT initiation is a physician decision rather than an objective event. In an open-label trial, knowledge of allocation can bias the timing of dialysis initiation in either direction. Each component must be reported separately with absolute risk differences and numbers needed to treat; a composite that is driven entirely by RRT initiation would be much weaker evidence than one driven by mortality.
Open-label design. Risks performance bias in co-interventions and ascertainment bias in the subjective component. The objective component (death) can be captured reliably; the RRT component should be adjudicated by a blinded committee.
Recommendation: continue to the planned final analysis. Do not stop for efficacy (boundary not crossed), and do not stop for futility (the point estimate favours X). Require blinded adjudication of the RRT component, pre-specified sensitivity analyses for the missing data, and full reporting of the hypotension discontinuations as a safety outcome.