Interpreting Hazard Ratios Alongside RMST

Two Measures, Two Questions

When a treatment effect is delayed, a single summary can conceal important features of the survival experience. A hazard ratio describes a comparison of instantaneous event rates at a particular time, whereas restricted mean survival time summarizes survival accumulated over a specified period. Neither measure is a substitute for the other: they answer different questions and use different scales.

This distinction matters in group sequential trials, where treatment effects may be examined at interim as well as final analyses. An estimated hazard ratio at a given time does not, by itself, summarize all earlier or later experience. An RMST difference does summarize survival time up to its chosen horizon, but it does not say when that difference arose. The interpretation should therefore name both the time horizon and the quantity being compared.

Definition: Let $T$ denote event time, and let $a=T$ and $a=C$ label treatment and control. The hazard $h_a(t)$ is the instantaneous event rate at time $t$, conditional on remaining event-free just before $t$. The time-specific hazard ratio is $$ HR(t)=\frac{h_T(t)}{h_C(t)}. $$ When both hazards are positive, $HR(t)<1$ means the instantaneous event rate is lower in the treatment group at time $t$; $HR(t)>1$ means it is higher. The restricted mean survival time through a common horizon $\tau$ is $$ RMST_a(\tau)=\int_0^\tau S_a(t)\,dt, $$ where $S_a(t)=P(T>t\mid a)$ is the survival probability in group $a$. Define the treatment contrast as $\Delta_{RMST}(\tau)=RMST_T(\tau)-RMST_C(\tau)$. A positive contrast means more average event-free time for treatment through $\tau$; a negative contrast means less.

The hazard ratio is dimensionless and time-specific. RMST and its difference have units of time: for example, years of event-free survival accumulated per participant through two years. Under proportional hazards, the hazard ratio is constant over time, so a single ratio can describe the relative hazards throughout follow-up. Under non-proportional hazards, $HR(t)$ may change, and reporting a single fitted hazard ratio can obscure that change. RMST remains a defined summary for a chosen $\tau$, whether or not hazards are proportional.

Why RMST Accumulates Survival Time

The integral definition has a direct interpretation. For any nonnegative event time $T$, the time an individual remains event-free before $\tau$ is $\min(T,\tau)$. For each $t$ between zero and $\tau$, the indicator that this individual is event-free at $t$ is one if $T>t$, and zero otherwise. Thus the time event-free before $\tau$ can be written as the area under that indicator across time. Taking expectations gives

E\{\min(T,\tau)\} =\int_0^\tau P(T>t)\,dt =\int_0^\tau S(t)\,dt =RMST(\tau).

Subtracting the group-specific expectations yields

\Delta_{RMST}(\tau) =E\{\min(T_T,\tau)\}-E\{\min(T_C,\tau)\} =\int_0^\tau \{S_T(t)-S_C(t)\}\,dt.

This also explains why an RMST difference is accumulated rather than instantaneous. It is the signed area between the two survival curves up to $\tau$. Periods with nearly equal survival contribute little area, while a sustained separation contributes more. A late hazard-ratio improvement can therefore coexist with a small RMST difference if it begins near the end of the selected horizon.

Worked Interpretation Under a Delayed Effect

For the calculations below, piecewise exponential models are used as simple hypothetical examples. Within each stated interval, each group has a constant hazard. This assumption gives $S_a(t)=\exp\{-\int_0^t h_a(u)\,du\}$, which is used to calculate the illustrative RMST values. Real trial estimates require appropriate survival methods; the model calculations here demonstrate interpretation, not a claim about an actual study.

Worked Example: Benefit Begins After Six Months

In a hypothetical trial, suppose the control hazard is $0.08$ events per person-year throughout two years. Suppose the treatment hazard is also $0.08$ per person-year during the first $0.5$ years, then falls to $0.02$ per person-year. The assumed hazards are positive and constant within the stated intervals, so the piecewise exponential survival formulas apply. Interpret the time-specific hazard ratios and calculate the RMST contrast through $\tau=2$ years.

During the first $0.5$ years, $HR(t)=0.08/0.08=1$: the assumed instantaneous event rates are equal. After $0.5$ years, $HR(t)=0.02/0.08=0.25$: the treatment event rate is one-quarter of the control event rate at those times, or $75\%$ lower. These are statements about instantaneous rates, not about the proportion of participants who have experienced an event.

For control, the hazard is constant, so

RMST_C(2)=\int_0^2 e^{-0.08t}\,dt =\frac{1-e^{-0.16}}{0.08} =\frac{1-0.8521}{0.08} =1.848\text{ years}.

For treatment, survival is $e^{-0.08t}$ until $0.5$ years. After that time, survival starts at $e^{-0.08(0.5)}=e^{-0.04}$, and declines with hazard $0.02$. Therefore,

\begin{aligned} RMST_T(2) &=\int_0^{0.5}e^{-0.08t}\,dt +e^{-0.04}\int_0^{1.5}e^{-0.02u}\,du\\ &=\frac{1-e^{-0.04}}{0.08} +e^{-0.04}\frac{1-e^{-0.03}}{0.02}\\ &=0.4901+1.4198 =1.9099\text{ years}. \end{aligned}

The contrast is $1.9099-1.8482=0.0617$ years, or $0.0617\times365.25=22.5$ days. In this hypothetical model, treatment adds about 22.5 days of average event-free time through two years. The late hazard ratio of $0.25$ is not a statement that treatment adds 75% more survival time; it describes a relative instantaneous rate after the effect begins.

Interpreting Estimates and Their Uncertainty

In an analysis, the hazard ratio and RMST contrast are estimated quantities. An estimate should be reported with its time reference or horizon, its units where applicable, and an uncertainty measure. A hazard-ratio interval that includes one is compatible with equal instantaneous hazards at the stated time, among other possibilities. An RMST-difference interval that includes zero is compatible with no difference in accumulated event-free time through $\tau$, among other possibilities. Neither interval gives the probability that a hypothesis is true.

The two measures need not have matching inferential conclusions. They can have different precision, and they summarize different features of the data. A hazard ratio at one time may suggest a favorable relative rate while the RMST contrast remains imprecise or small. Conversely, RMST can show a clear accumulated difference even when a late time-specific hazard ratio is close to one, because earlier survival differences remain represented in the area.

Worked Example: Reading Estimates on Their Own Scales

Suppose a hypothetical final analysis reports a hazard ratio at 18 months of $0.72$, with a nominal 95% confidence interval of $0.50$ to $1.03$. It also reports an RMST difference through $2.5$ years of $0.14$ years, with a nominal 95% confidence interval of $0.02$ to $0.26$ years. Assume both quantities were estimated using methods suitable for the trial's censoring and design, and that the intervals are ordinary single-analysis intervals rather than sequentially adjusted intervals.

At 18 months, the point estimate $0.72$ corresponds to an estimated instantaneous event rate $1-0.72=0.28$, or 28%, lower in treatment than control at that time. The interval includes $1$, so it does not rule out equal hazards at that time at the corresponding nominal two-sided 5% level; values above $1$ are also within the interval.

The RMST estimate says that treatment adds $0.14$ years of average event-free time through $2.5$ years. Converting the point estimate and limits to days gives

0.14\times365.25=51.14\text{ days},\qquad (0.02,0.26)\times365.25=(7.305,94.97)\text{ days}.

Thus the estimated gain is about 51.1 days, with a nominal interval of about 7.3 to 95.0 days. This is accumulated time through $2.5$ years, not a rate reduction. In a trial with repeated interim looks, ordinary single-analysis intervals should not be treated as sequentially adjusted: the monitoring design must be respected when making confirmatory claims.

When Effects Change Direction

A delayed effect need not simply switch on and remain beneficial. Effects may weaken or reverse. The hazard-ratio path and the RMST contrast then provide complementary information: the former locates relative rate changes in time, while the latter integrates all survival differences up to its horizon. A common $\tau$ is essential for a direct treatment-control comparison, and the chosen horizon should be clinically interpretable and supported by follow-up in both groups.

Worked Example: A Later Increase in the Treatment Hazard

Suppose, hypothetically, that both groups have hazard $0.10$ events per person-year during the first year. During the second year, the control hazard remains $0.10$, while the treatment hazard rises to $0.15$. Assume these piecewise constant hazards describe the full period and calculate the RMST contrast through two years.

The time-specific hazard ratio is $1$ during year one and $0.15/0.10=1.5$ during year two. Thus the relative rates are equal initially and the treatment event rate is 50% higher in the second year. This does not alone determine the two-year accumulated contrast; that depends on the survival curves over both years.

For control,

RMST_C(2)=\frac{1-e^{-0.20}}{0.10} =\frac{1-0.8187}{0.10} =1.8127\text{ years}.

For treatment, the first-year area is $(1-e^{-0.10})/0.10=0.9516$ years. Survival at one year is $e^{-0.10}=0.9048$; the second-year area is

e^{-0.10}\frac{1-e^{-0.15}}{0.15} =0.9048\frac{1-0.8607}{0.15} =0.8402\text{ years}.

Therefore, using the unrounded integrals, $RMST_T(2)\approx1.7919$ years and $RMST_C(2)\approx1.8127$ years, so $\Delta_{RMST}(2)\approx-0.0208$ years. This is about $-7.61$ days. The negative contrast means about 7.6 fewer average event-free days for treatment through two years in this model. The example shows why a favorable or neutral early comparison cannot be assumed to describe the later period.

Assumptions, Comparisons, and Pitfalls

Interpretation depends on the trial design and the estimation method. Randomization supports a treatment-group comparison, but conclusions still depend on how outcomes, censoring, and the analysis population are handled. Estimating survival curves commonly requires censoring to be independent of event time within the groups, possibly conditional on modeled covariates. The selected RMST horizon should be prespecified where possible, applied to both groups, and lie within follow-up where survival can be estimated reliably. A hazard ratio also requires care when hazards are very small or poorly estimated at some times; in such regions, a ratio may be unstable.

For group sequential monitoring, estimation at interim looks adds a further qualification. Repeatedly inspecting data can affect the operating characteristics of tests and intervals. A pointwise hazard-ratio estimate or ordinary confidence interval does not automatically account for the monitoring plan. The treatment-effect description can still be useful at an interim analysis, but confirmatory inference should use the prespecified sequential procedure. Weighted logrank tests, discussed in the earlier material on weighted logrank monitoring, test a contrast using a chosen weighting scheme; their test statistic is not itself an estimate of days gained. Sequential RMST inference addresses uncertainty for an RMST contrast under the monitoring design, rather than converting a weighted test into a time-specific hazard ratio.

The main alternatives answer distinct questions. A single Cox-model hazard ratio is compact, but under non-proportional hazards it is generally a model-dependent summary rather than the hazard ratio at every time. A plot or estimate of $HR(t)$ makes time variation more visible, but may be noisy and does not directly quantify accumulated survival time. A median survival difference compares one point on each survival curve and may be unavailable or unstable if a curve does not reach 50% during follow-up. RMST uses the area up to $\tau$, but its value depends on that horizon and it does not reveal the timing of the difference without looking at the curves or time-specific measures.

A practical interpretation sequence is to state the hazard-ratio time point and direction, state the RMST horizon and difference in time units, then explain whether the two summaries tell a consistent or complementary story. Also check that the hazard-ratio time point falls within adequate follow-up, that $\tau$ is shared between groups, and that uncertainty statements match the trial's single-look or sequential design.

Pitfall: “The hazard ratio is 0.70, so treatment gives 30% more survival time.” This confuses a relative instantaneous event rate with a time-based accumulated contrast. The correction is to say that the estimated instantaneous rate is 30% lower at the specified time, and report the RMST difference separately for its stated horizon.
Pitfall: Reporting one hazard ratio as if it applied throughout a delayed-effect trial. Under non-proportional hazards, the ratio can change with time. State the time reference or use a model and interpretation appropriate to the changing effect.
Pitfall: Treating a nominal interim interval as if it accounted for repeated monitoring. Use inference calibrated for the sequential design before making a confirmatory error-rate claim.

Key takeaway: A time-specific hazard ratio describes relative instantaneous event rates at a stated time. An RMST difference describes the difference in average event-free time accumulated through a stated horizon. In delayed-effect trials, report and interpret both on their own terms.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

At a particular time, the treatment and control hazards are $0.06$ and $0.10$ events per person-year. What is the hazard ratio, and what does it mean?
A trial estimates an RMST difference of $0.08$ years through a common horizon. Convert the estimate to days and state its interpretation.
Why can a favorable late hazard ratio coexist with a small RMST difference through a horizon that ends soon after the effect begins?
What does a hazard-ratio confidence interval that includes one say, and what additional caution applies if the estimate is from an interim look in a group sequential trial?

Solutions. (1) $HR(t)=0.06/0.10=0.60$. At that time, the estimated instantaneous event rate in treatment is 40% lower than in control; this is not a 40% increase in survival time. (2) $0.08\times365.25=29.22$ days, or about 29.2 days. Treatment is estimated to add that much average event-free time through the stated horizon. (3) A late effect contributes area between the survival curves only for the time it persists before the horizon. If the effect starts late and follow-up ends soon afterward, the accumulated area can remain small. (4) The interval is compatible with equal hazards at the specified time, as well as values on either side of one contained in the interval. At an interim look, an ordinary pointwise interval is not automatically adjusted for repeated monitoring, so confirmatory inference must follow the prespecified sequential procedure.