Restricted Mean Survival Time (RMST): The Hazard-Ratio Alternative When Proportional Hazards Fails

A trial reports a single hazard ratio for its primary time-to-event outcome, the confidence interval excludes 1.0, and the result gets summarized in one sentence: “the new treatment reduced the hazard by X%.” That sentence is only a valid summary of the whole follow-up period if the two groups’ hazards stay in roughly constant proportion to each other the entire time — the proportional-hazards (PH) assumption a Cox model requires. When that assumption fails, most visibly when the Kaplan-Meier curves for the two groups cross, a single pooled hazard ratio (HR) is not simply “less precise.” It can average together two effects moving in opposite directions and report a number that describes neither phase of the trial. Restricted mean survival time (RMST) answers a different, assumption-free question instead: how much event-free survival time, on average, did a patient in each group actually accumulate by a fixed point in follow-up? This guide works through what RMST measures, how to compute it, how to choose the time horizon it is measured to, and a worked comparison — built from a seeded, reproducible simulation, not a real trial — showing a case where the pooled HR and RMST genuinely disagree.

What RMST Actually Measures

Restricted mean survival time at a truncation time τ is the area under the Kaplan-Meier survival curve from time 0 to τ. Geometrically, plot S(t) — the estimated probability of remaining event-free past time t — on the y-axis against follow-up time on the x-axis; RMST(τ) is the area under that curve, in the same time units as the x-axis (typically months or years). It has a direct clinical reading that a hazard ratio does not: it is the average amount of event-free time a patient in that group experienced, out of a maximum possible τ. Two patients enrolled and event-free for the entire window contribute the maximum (τ each); a patient who has the event at time t contributes only t; a patient censored at time t contributes t as well, with the KM estimator’s usual handling of the remaining uncertainty folded into the curve itself. Unlike the hazard ratio, RMST makes no assumption about the shape of either group’s hazard function, and no assumption that the two groups’ hazards stay proportional to each other — it is computed directly from each group’s own Kaplan-Meier curve, with no model fitted at all.

This is the same estimator behind the site’s own guide to reading a Kaplan-Meier curve: RMST is literally the area under that curve, truncated at τ. The formalization of RMST as a trial-design and reporting statistic, specifically framed as an alternative to the hazard ratio, is due to Royston and Parmar (BMC Medical Research Methodology, 2013); Uno and colleagues (Journal of Clinical Oncology, 2014) separately made the applied case for reporting it alongside, not instead of, the hazard ratio in oncology trials where non-proportional hazards are common.

Why the Hazard Ratio Can Mislead When Proportional Hazards Fails

A Cox model’s hazard ratio is, in effect, a single number that best summarizes the ratio of two hazard functions across the entire observed follow-up, weighted by where the events actually occurred. When the true ratio is roughly constant over time, that single number is an efficient, well-understood summary — see the site’s guides on checking the proportional-hazards assumption and on what a hazard ratio does and doesn’t mean even when PH holds. When the ratio is not constant — most commonly because one arm has a delayed treatment effect (worse or equivalent early, meaningfully better only after some months) — the two groups’ hazard functions genuinely cross, and a single pooled HR is a kind of average of two different, opposing effects. It does not describe the early phase and does not describe the late phase; it describes a follow-up-time-weighted blend of both, and that blend depends on incidental features of the specific trial — enrollment pace, total follow-up duration, censoring pattern — not on the underlying biology alone. A different-length trial of the exact same treatment could report a materially different pooled HR without anything about the treatment having changed.

A Worked Comparison: When RMST and the Hazard Ratio Tell Different Stories

The dataset below is simulated, not from a real trial. It was generated by a seeded, fully reproducible script (PRNG seed 20260829, 300 simulated patients per arm, independent random loss-to-follow-up plus 36-month administrative censoring) built specifically to illustrate crossing hazards — every number in this section is a direct output of that simulation and the standard estimators described below, not an invented or illustrative-only figure. “Arm A” (standard care) has a constant hazard across the full 36 months. “Arm B” (new therapy) has a higher hazard than Arm A for the first 6 months, then a substantially lower hazard from month 6 onward — the delayed-treatment-effect pattern that produces visibly crossing Kaplan-Meier curves.

Metric	Arm A (standard care)	Arm B (new therapy)
Events observed / n	216 / 300	160 / 300
KM survival at 12 months	57.3%	59.7%
KM survival at 24 months	34.8%	52.1%
KM survival at 36 months	22.5%	44.8%

First, the single pooled hazard ratio a naive Cox or log-rank analysis of the full 36 months would report (computed here via the standard closed-form log-rank/Peto hazard-ratio estimator, expressed as Arm A’s hazard relative to Arm B’s): HR 1.46 (95% CI 1.19–1.79, p < 0.001). Read uncritically, that looks like a decisive, fairly substantial benefit for the new therapy, evenly present throughout the trial — equivalent to roughly a 31% relative reduction in Arm B’s hazard versus Arm A. It is not evenly present. Splitting the same data into the two phases the simulation was built around:

Follow-up window	Hazard ratio, Arm A vs. Arm B	95% CI	Reading
0–6 months	0.59	0.44–0.79	Arm A’s hazard is lower — Arm A doing better early
6–36 months	3.34	2.52–4.42	Arm A’s hazard is more than 3× higher — Arm B doing dramatically better late

The pooled 1.46 is neither of these numbers. It is a blend of an early phase favoring Arm A and a late phase favoring Arm B far more strongly, and it happens to land on the “Arm B is better” side mainly because the late-phase hazard difference is large enough to dominate the log-rank test statistic despite fewer patients remaining at risk by then. Now compare what RMST reports at two different, pre-specifiable horizons, using the same simulated dataset and the standard area-under-the-Kaplan-Meier-curve estimator with a bootstrap confidence interval (2,000 resamples):

Truncation time (τ)	RMST, Arm A	RMST, Arm B	Difference (B − A)	95% CI
24 months	14.83 months	15.15 months	+0.31 months	−1.21 to +1.79
36 months	18.15 months	21.04 months	+2.88 months	+0.53 to +5.18

At 24 months — already well past the point where the two Kaplan-Meier curves visibly cross — RMST shows the two arms are not statistically distinguishable in average event-free survival time: a 0.31-month difference whose confidence interval comfortably straddles zero. The pooled hazard ratio computed over the same kind of interim window would already read as a confident, statistically significant result to anyone who hadn’t checked whether PH actually holds. Only by 36 months, the trial’s full observation window, does RMST show a real, statistically supported gain: 2.88 months of additional expected event-free survival for Arm B, a difference whose interval no longer includes zero. In plain terms: judged by “how much more event-free time did patients in this arm actually experience,” Arm B’s advantage is real but had not yet arrived at the 24-month mark, and the single pooled HR obscures that timing entirely because it is not built to represent a moving target in the first place.

Choosing and Justifying the Truncation Time (τ)

RMST is only estimable up to the smallest “largest observed time” (event or censoring) across the groups being compared — the Kaplan-Meier curve for a group is undefined past its last observation, so τ can never exceed that limit in any group. Within that ceiling, three considerations govern a defensible choice, following the practice Royston and Parmar’s original paper recommends:

• Pre-specify it. Choose τ in the statistical analysis plan, before the data are unblinded, for the same reason any other primary-analysis parameter is pre-specified — choosing τ after seeing the results invites the same outcome-selection concern as choosing which endpoint to report after seeing which one came out favorable.
• Anchor it to something clinically meaningful. A common horizon (12 months, the trial’s planned maximum follow-up, a landmark tied to the disease’s natural history) is more interpretable and more defensible than an arbitrary value chosen purely because it maximizes the estimable range.
• Report more than one τ, not a single cherry-picked point. Because RMST is monotonically non-decreasing in τ (the area under a curve can only grow as the window widens), plotting the RMST difference as a function of τ across the whole estimable range — rather than quoting one number — shows a reader exactly where in follow-up a treatment effect appears, the same information the worked comparison above relies on.

Computing RMST in Practice

Mechanically, RMST(τ) is the trapezoidal-integration area under a Kaplan-Meier step function, truncated at τ — the exact calculation performed for the worked comparison above. Several standard survival-analysis packages implement this directly rather than requiring it to be assembled by hand: R’s survRM2 package (built by one of the RMST methodology’s own authors) computes RMST per group, the between-group difference and ratio, variance estimates, and a τ-by-τ plot in one call; Python’s lifelines library and Stata’s strmst2 command offer equivalent functionality. Any of these is preferable to a hand-rolled calculation for a real analysis — the worked example above was built from first principles specifically to make the mechanism transparent, not as a recommended production workflow.

What RMST Doesn’t Solve

RMST is a genuine complement to the hazard ratio, not a strictly superior replacement for it, and it has real limitations of its own:

• It is still a single-number summary of everything up to τ. Two groups can have identical RMST at a given τ while their survival curves take very different shapes to get there — the same information-collapse limitation any single summary statistic has, hazard ratio included.
• The choice of τ genuinely changes the answer, as the worked comparison above demonstrates directly — which is exactly why pre-specification matters, not a reason to avoid RMST.
• It discards information beyond τ. A late, large effect that only emerges after the chosen horizon (as in the 36-month result above, invisible at 24 months) will not show up in an RMST computed at a shorter τ.
• When PH genuinely holds, the hazard ratio remains the more standard, more efficient, more widely understood summary — RMST’s real value is specifically when PH is doubtful or violated, or when a result needs to be communicated in a directly interpretable “months of event-free survival” unit alongside the HR, not as a wholesale substitute for it.
• Neither statistic replaces reporting the actual Kaplan-Meier curves. Both RMST and the HR are summaries of that curve; the curve itself remains the most complete, assumption-free description of what happened in the trial.

Frequently Asked Questions

Is RMST always a better choice than the hazard ratio?

No. When the proportional-hazards assumption genuinely holds, the hazard ratio is a well-understood, efficient single-number summary and there is no strong reason to replace it. RMST earns its place specifically when PH is doubtful or clearly violated — crossing Kaplan-Meier curves being the most visible sign — or when a result needs to be reported in a directly interpretable time unit (“2.9 additional months of event-free survival by month 36”) alongside, not instead of, the HR.

How is RMST different from median survival time?

Median survival is only defined once a group’s Kaplan-Meier curve actually crosses 50% survival within the observed follow-up; if it never does, the median is simply unreported. RMST at any τ within the observed data is always computable regardless of whether either curve has crossed 50%, which is part of why it tends to be more robust for trials with a lower event rate or shorter follow-up.

Does RMST(τ) ever decrease as τ gets larger?

No. Because the survival function S(t) is always non-negative, adding more time to the integration window can only add non-negative area — RMST(τ) is mathematically guaranteed to be non-decreasing in τ.

What software actually computes RMST?

R’s survRM2 package, Python’s lifelines library, and Stata’s strmst2 command all compute it directly, including the between-group difference, ratio, and standard errors — a hand-rolled trapezoidal integration of the Kaplan-Meier curve, as used for illustration in the worked comparison above, is rarely the right tool for a real analysis.

Is RMST accepted as a trial endpoint by regulators?

RMST is increasingly reported as a pre-specified secondary or sensitivity analysis in trials where proportional hazards is doubtful, particularly in oncology settings with delayed treatment effects. Whether it serves as a primary or secondary endpoint for any specific trial remains a protocol-specific decision made with the relevant regulatory body, not a default this guide can generalize.

Related CASRAI Reading

RMST sits alongside a family of time-to-event concepts already covered on CASRAI:

• How to Interpret a Hazard Ratio (and Why It Is Not a Risk Ratio) — what the pooled HR in the worked comparison above actually means, and where it can mislead even when reported correctly.
• The Cox Proportional Hazards Model: Assumptions and How to Check Proportionality — how to test whether PH holds before trusting a single HR in the first place.
• How to Read a Kaplan-Meier Curve: Censoring Marks, Median Survival, and Risk Tables — RMST is literally the area under this exact curve.
• The Log-Rank Test and Its Variants — the significance test underlying the pooled and windowed hazard-ratio estimates used above.