Why Crossing Hazards Need Careful Interpretation
A treatment can be associated with a lower event rate early in follow-up and a higher event rate later. In that setting, the treatment and control hazards cross. A single hazard ratio may compress these opposing patterns into one summary, obscuring when patients benefit, when they may fare worse, and what the accumulated survival experience looks like.
The key distinction is between an instantaneous event rate among people still event-free and the probability of remaining event-free through a specified time. The first is described by a hazard; the second by a survival probability. Because survival reflects the history of hazards up to that time, crossing hazards do not generally imply that survival curves cross at the same time.
A hazard has units of inverse time, such as events per person-year. It is an instantaneous rate conditional on still being event-free immediately before time \(t\), not the probability that an individual will experience the event during the whole study. The survival function \(S_a(t)\) is the probability of remaining event-free beyond time \(t\) in arm \(a\). Its complement, \(1-S_a(t)\), is the probability of experiencing the event by time \(t\), for a single-event endpoint without competing events.
From Hazards to Survival Probabilities
Assume a continuous event-time distribution with a well-defined hazard. For a given arm, the hazard and survival function satisfy \(dS(t)/dt=-\lambda(t)S(t)\), with \(S(0)=1\). Solving this differential equation shows why survival depends on the entire hazard history rather than only the current hazard.
The integral is the cumulative hazard through time \(t\). If the hazard is piecewise constant, the cumulative hazard is the sum of each interval's hazard multiplied by its duration. Thus, a higher current treatment hazard does not immediately imply lower treatment survival: earlier lower hazards may leave fewer accumulated events by that time.
This interpretation assumes that the hazards describe the event endpoint of interest and that the survival functions are meaningful for the populations being compared. In a randomized trial, the arm-specific survival contrast can be interpreted as a treatment-policy comparison when randomization is preserved and follow-up is appropriately handled. Estimation from censored data additionally requires censoring to be non-informative, at least conditional on variables included in the analysis. If competing events are present, a cause-specific hazard alone does not give the probability of the event in the presence of those competing events.
Worked Example: Hazard Crossing Before Survival Crossing
Suppose, hypothetically, the control hazard is \(0.10\) events per person-year throughout follow-up. The treatment hazard is \(0.08\) per year through year 1 and \(0.12\) per year thereafter. The hazards cross at year 1: the treatment hazard is lower before that time and higher after it. Assume the piecewise hazards are constant within their stated intervals.
For control, the cumulative hazard is \(0.10t\), so \(S_0(t)=e^{-0.10t}\). For treatment, it is \(0.08t\) through year 1, and \(0.08+0.12(t-1)\) after year 1.
At year 1, the survival probabilities are \(S_1(1)=e^{-0.08}=0.923116\) and \(S_0(1)=e^{-0.10}=0.904837\). Their difference is \(e^{-0.08}-e^{-0.10}=0.01828\), or 1.828 percentage points higher survival with treatment at year 1. The treatment hazard is already about 20% lower during this first year, since \(0.08/0.10=0.80\).
At year 3, the treatment cumulative hazard is \(0.08+0.12(3-1)=0.32\), while the control cumulative hazard is \(0.10(3)=0.30\). Therefore, \(S_1(3)=e^{-0.32}=0.726149\) and \(S_0(3)=e^{-0.30}=0.740818\). The survival difference, treatment minus control, is \(-0.01467\), or 1.467 percentage points lower survival with treatment at year 3.
The survival curves cross when their cumulative hazards are equal. After year 1, set \(0.08+0.12(t-1)=0.10t\). This gives \(0.12t-0.04=0.10t\), so \(t=2\) years. At year 2 both survival probabilities equal \(e^{-0.20}=0.818731\). The hazard crossing is at year 1; the survival crossing is at year 2.
What a Single Hazard Ratio Conceals
A constant hazard ratio is a useful summary when the treatment-to-control hazard ratio is reasonably stable over time. Under crossing hazards, however, there is no single number that describes the treatment's relative instantaneous event rate at every time. A fitted constant-hazard-ratio model may still produce a numerical estimate, but its value summarizes a changing pattern rather than identifying a time-invariant effect.
In particular, that estimate is not generally the simple arithmetic average of the early and late hazard ratios. Its value can depend on the event-time pattern, the numbers at risk, censoring, allocation, and the model used. It should not be interpreted as the hazard ratio at every follow-up time, nor directly as a ratio of event probabilities or survival probabilities.
Worked Example: An Average Hazard Ratio Is Not a Full Clinical Summary
Consider a hypothetical two-year trial with a control hazard of \(0.10\) events per person-year in both years. Let the treatment hazard be \(0.05\) in year 1 and \(0.15\) in year 2. The time-specific hazard ratios are \(0.05/0.10=0.50\) in year 1 and \(0.15/0.10=1.50\) in year 2. Their arithmetic average is 1.00, but that average is not a fitted trial hazard ratio and does not describe either year.
At year 1, treatment survival is \(e^{-0.05}=0.951229\); control survival is \(e^{-0.10}=0.904837\). The treatment-minus-control survival difference is \(0.04639\), or 4.639 percentage points. The event probabilities by year 1 are \(1-e^{-0.05}=0.04877\) with treatment and \(1-e^{-0.10}=0.09516\) with control.
At year 2, the treatment cumulative hazard is \(0.05+0.15=0.20\), exactly equal to the control cumulative hazard \(0.10+0.10=0.20\). Both survival probabilities are \(e^{-0.20}=0.818731\), corresponding to a 0.1813 event probability by year 2 in each arm. The same final survival does not erase the earlier separation or the later reversal in hazards. A single number of 1.00 would hide this clinically relevant timing.
A clearer clinical report pairs an explicit time horizon with an absolute survival contrast and, when useful, the arm-specific survival probabilities. For example, “survival at year 1 was 95.12% with treatment and 90.48% with control” is tied to a defined time and scale. A time-specific hazard ratio can add information about instantaneous rates, but it should be labeled with its interval or time point rather than presented as though it applied throughout follow-up.
Comparing Summaries and Recognizing Failure Modes
| Summary | What it describes | Interpretive caution with crossing hazards |
|---|---|---|
| Constant hazard ratio | A model-based single relative-rate summary | May conceal the direction and timing of effects; is not the ratio of survival probabilities |
| Time-specific hazard ratio | The instantaneous relative event rate at a specified time | Needs its time context; it is not a cumulative event probability |
| Survival probability at a specified time | The probability of remaining event-free beyond that time | Depends on accumulated hazards, not just the hazard at that time |
| Absolute survival difference at a specified time | The treatment-minus-control difference in survival probability | Must name the horizon and population; its direction can differ across horizons |
Several checks help prevent overinterpretation. First, inspect whether the estimated hazard ratio appears stable over time or changes direction; uncertainty may be substantial, especially late when few participants remain at risk. Second, compare arm-specific survival estimates at clinically meaningful times, with uncertainty intervals where available. Third, check that the stated event, time origin, analysis population, and censoring assumptions match the clinical question. A visually crossing pattern can arise from random variation, so the curves alone do not establish a true reversal.
The earlier material on delayed treatment effects distinguishes delayed onset from other changes in the treatment effect. Crossing is a different pattern: the relative ordering of hazards reverses. Neither a fitted constant hazard ratio nor a single endpoint survival contrast fully describes when that reversal occurs. A group-sequential design may control its prespecified error rate for its chosen test, but that calibration does not make a constant hazard-ratio interpretation valid when hazards vary over time.
Crossing-hazard summaries also have limitations. A few selected time points cannot describe every feature of the curves, and time-specific estimates may be imprecise. A constant hazard ratio can remain a convenient compact summary, but readers need to know that its interpretation is conditional on proportional hazards or else treat it as a model-dependent average. This tutorial's comparisons do not establish which test or estimand is best for every crossing pattern; that choice depends on the scientific question and prespecified analysis plan.
Worked Example: Interpreting a Clinical Summary
Suppose a hypothetical report gives treatment and control survival estimates of 0.88 and 0.84 at year 2, and 0.70 and 0.75 at year 5. It also reports an overall fitted hazard ratio of 0.95. State what these numbers support, and what they do not.
At year 2, the treatment-minus-control survival difference is \(0.88-0.84=0.04\): estimated survival is 4 percentage points higher with treatment. At year 5, the difference is \(0.70-0.75=-0.05\): estimated survival is 5 percentage points lower with treatment. These time-specific contrasts show that the observed absolute survival ordering differs at the two horizons.
The fitted hazard ratio of 0.95 is a single model-based summary. It does not mean that the treatment hazard was 5% lower at every time, and it does not contradict the later survival difference. Without its model assumptions and uncertainty, it should not be read as a complete description of the time-varying effects. Nor do these estimates alone establish that treatment caused the apparent reversal; study design, uncertainty, follow-up, and analysis assumptions matter.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- If treatment and control hazards are equal at year 4, must their survival probabilities also be equal at year 4? Explain.
- A hypothetical control hazard is 0.08 per year throughout. Treatment hazard is 0.04 in year 1 and 0.10 in year 2. Calculate each arm's survival at year 2 and the treatment-minus-control difference.
- Why is the arithmetic average of early and late hazard ratios not automatically the trial's overall hazard ratio?
- Write one clinically interpretable sentence comparing survival at a named horizon, and state why it is not a hazard-ratio interpretation.
Solutions. (1) No. Equal instantaneous hazards at one time do not imply equal cumulative hazards. Survival is equal only if cumulative hazards through that time are equal. (2) Treatment cumulative hazard is \(0.04+0.10=0.14\), so \(S_1(2)=e^{-0.14}=0.869358\). Control cumulative hazard is \(0.08(2)=0.16\), so \(S_0(2)=e^{-0.16}=0.852144\). The difference is \(0.01721\), or 1.721 percentage points higher treatment survival at year 2. (3) A fitted overall estimate depends on event times, risk sets, censoring, allocation, and model specification; averaging two ratios ignores these features and does not represent a constant effect. (4) For example: “At year 2, estimated survival was 86.9% with treatment and 85.2% with control, a difference of 1.7 percentage points.” This compares probabilities of remaining event-free through a horizon, whereas a hazard ratio compares instantaneous event rates among those still event-free.