A single relative summary can be useful, but it does not capture every clinically relevant feature of a time-to-event distribution. Delayed effects, crossing curves, and changing hazards can make interpretation difficult. The right response is to clarify the question and assumptions before selecting additional summaries.
Royston and Parmar’s methods paper discusses restricted mean survival time as an alternative measure. It is a methodological reference, not a recommendation that one measure should replace the hazard ratio in every trial.
1. Inspect the time pattern
Begin with the survival estimates, numbers at risk, event counts, and follow-up distribution. A curve’s late tail can look visually dramatic when very few participants remain under observation. Include information that helps the reader distinguish a stable pattern from a sparse estimate.
A hypothetical treatment with a delayed effect may have little early separation and greater later separation. That pattern should prompt a discussion of the planned summary and model assumptions, rather than a retrospective search for whichever statistic gives the most favorable result.
2. Explain the summary’s meaning
A hazard ratio describes a relative hazard under the chosen model. Restricted mean survival time concerns the area under the survival curve up to a specified horizon. A difference in restricted means expresses a difference in expected event-free or survival time within that horizon, depending on the endpoint.
These measures answer different summary questions. The horizon must be justified and supported by follow-up; it should not be selected after inspecting results simply to obtain a preferred conclusion. The ICH E9(R1) framework helps anchor the treatment-effect question.
3. Keep censoring assumptions visible
Censoring rules and the information available after treatment changes affect interpretation. Document why follow-up ended and assess patterns by treatment group. A standard plotting function cannot determine whether the censoring assumptions are plausible in the clinical setting.
For oncology studies, the FDA endpoint guidance is useful additional context. Endpoint definition, measurement, and censoring need to be considered together, especially when follow-up schedules or subsequent treatment differ.
4. Distinguish planned from exploratory work
If additional summaries are examined after the primary analysis, label them accordingly. Explain the reason for exploration and any limitations involving multiplicity, model selection, or limited follow-up. An exploratory estimate can be informative without being presented as confirmatory evidence.
Our proposed review package combines the prespecified primary analysis, transparent risk tables, follow-up diagnostics, and clearly justified supplementary summaries. It should help a clinician describe both the estimated effect and the period of observation over which that estimate is meaningful.
A worked interpretation
For a chosen horizon τ, restricted mean survival time is the integral of the survival function from zero to τ. In a hypothetical analysis with τ equal to 12 months, an estimated restricted-mean difference of 0.8 months describes the difference in expected time before the endpoint within that 12-month period. It does not mean that every participant benefits by 0.8 months or that the treatment extends lifetime survival by that amount.
A Kaplan–Meier-based calculation can sum rectangular areas between event times up to the horizon. Verify the time units and whether the endpoint concerns death or another event. Review uncertainty estimation and ensure the selected horizon has adequate support in both groups.
Report the horizon prominently. A different horizon defines a different summary, so retrospective selection among many horizons requires a transparent account of exploratory analysis and multiplicity.
