Time-to-event studies use the timing of outcomes, not only whether they happened. Their most familiar summary, the hazard ratio, is easy to misread as an ordinary risk ratio. A useful interpretation keeps the event, the clock and the people still under observation in view.
- A hazard ratio compares instantaneous event rates among those still at risk.
- Read absolute risks at a stated time alongside a hazard ratio.
- Censoring and changes in effects over time matter to the interpretation.
Five-year risks in the invented constant-hazard model
Percentage with a first symptom recurrence by five years
Find the event and the start of the clock
A survival analysis might concern death, first hospital admission, symptom recurrence or recovery. The name of the statistical method does not tell you whether the event is desirable. A higher event rate can mean faster recovery in one study and earlier deterioration in another.
Write down the event definition, time origin and follow-up period. In our original hypothetical model, the event is the first recurrence of a fictional symptom after study entry. Assume no competing event and constant hazards of 0.10 per year in the comparison group and 0.08 per year in the intervention group.
These are model assumptions, not estimates from participants. They let us calculate a transparent example without borrowing a clinical dataset.
Compare the hazard, then calculate the risk
The ratio of the two assumed hazards is 0.08/0.10=0.80: a 20% lower hazard. For a constant hazard h, the probability of at least one event by time t is 1−exp(−h×t).
At five years, the model gives approximately 39.3% with a first recurrence in the comparison group and 33.0% in the intervention group. The absolute difference is about 6.4 percentage points using unrounded values. The relative reduction in five-year risk is approximately 16.2%, not 20%.
The calculation illustrates why a hazard ratio should not be relabelled as a fixed-time risk ratio. Cochrane distinguishes time-to-event measures from ordinary binary risk measures.
Source 1 ↗Do not convert a hazard ratio without the needed information
Our calculation works because we specified both constant hazards and a time horizon. A published hazard ratio of 0.80 alone does not provide those ingredients. It cannot tell you the absolute benefit, the median survival gain or the expected extra years of life.
Ask the report for event probabilities or survival estimates at a clinically relevant time, their uncertainty and the numbers still at risk. If the report provides only a relative summary, keep that limitation in the article instead of inventing a conversion.
In our model, changing the follow-up from one year to five changes the cumulative risks even though the hazard ratio stays fixed. Time is part of the result.
Read the shrinking number at risk
A person may enter late, reach the study’s end without the event, or stop being observed. Censoring records that their event time is not fully observed; it does not mean they will never have the event.
Standard methods need assumptions about the relationship between censoring and the event process. Losing contact because someone becomes seriously unwell could be different from observation ending on a planned calendar date. A report should describe the reasons and methods.
For a plotted curve, inspect the numbers at risk beneath it. A dramatic separation near the end may rest on very few remaining observations. The visual line alone does not tell you how much information supports its tail.
Source 1 ↗A single ratio can hide a changing pattern
The constant-hazard exercise is deliberately simple. In real studies, an intervention’s relative effect can change over time. Curves may separate late or cross. A single proportional-hazards summary can then conceal clinically important timing.
Restricted mean survival time offers another summary: the area under a survival curve up to a specified horizon. Its difference compares average event-free or survival time within that window, depending on the event. The methodological work by Royston and Parmar discusses this approach when hazards are not proportional.
The horizon must travel with the number. A difference in average survival through three years is not a claim about lifetime survival. Read why that horizon was chosen.
Source 2 ↗Connect follow-up papers without changing the endpoint
A longer follow-up can show whether an early separation persists, narrows or changes direction. It may also involve treatment switching or other changes after the original phase. Those developments belong beside the new estimate.
For our fictional recurrence model, a future study would need observed event times and information about follow-up before anyone could evaluate whether the constant-hazard assumption was sensible. A better graph could show uncertainty and numbers at risk; it would not turn assumed values into evidence.
When summarising an actual report, use a sentence with four anchors: the population, event, effect measure and time context. Then add the absolute result if available. That prevents a relative timing statistic from becoming a promise of extra life.
One model, different summaries
| Quantity | Comparison group | Intervention group |
|---|---|---|
| Assumed constant hazard per year | 0.10 | 0.08 |
| One-year event risk | 9.5% | 7.7% |
| Five-year event risk | 39.3% | 33.0% |
| Five-year absolute difference | About 6.4 percentage points lower with intervention | Calculated before rounding |
| Hazard ratio | Reference | 0.80 |
Original exponential model: risk = 1 − exp(−hazard × years). Rounded percentages; complete model assumptions, no competing event, not a clinical dataset.
Your questions, answered
Does a hazard ratio of 0.80 mean 20% fewer events?
Not necessarily. It compares hazards, not simply the proportions with an event by a fixed time. In our constant-hazard example, the five-year risk reduction is approximately 16.2%, although the hazard ratio is 0.80.
Can I calculate extra years of life from the hazard ratio?
No, not from the ratio alone. You need additional information about the event-time distributions and the relevant horizon. A study about symptom recurrence does not measure extra life at all.
Does censored mean event-free forever?
No. It means the event time is not fully observed beyond a point. Interpretation depends on the follow-up process and the analysis assumptions, not on treating every censored person as permanently event-free.
What if the survival curves cross?
Look for discussion of time-varying effects and suitable analyses. A single summary may hide early disadvantage and later advantage. Crossing curves alone do not identify the best alternative analysis.
What does restricted mean survival time measure?
It measures the area under the survival curve through a specified time. With death as the event, it describes average survival within that window; with another event, the interpretation changes accordingly.
Are the chart’s numbers from a real trial?
No. They are calculated from assumed constant hazards of 0.10 and 0.08 per year, with no competing event. No confidence intervals or claims of treatment efficacy are attached to this teaching model.
Limits of this interpretation
- This is a selected educational explanation, not a systematic review, validated appraisal instrument or personal care recommendation.
- Numerical examples are hypothetical. Their deliberately simplified assumptions must not be transferred to a real study without checking its methods.
- An AI source check can miss errors; source access and the absence of independent human review are stated explicitly.
Sources & transparency
- Cochrane Handbook, chapter 6: Choosing effect measures and computing estimates of effect
Public HTML sections 6.7–6.8 on counts, person-time and time-to-event outcomes checked by AI. · Accessed 27 Sep 2026
- Royston and Parmar (2017): Life expectancy difference and life expectancy ratio
Publisher-indexed methodological description of restricted mean survival time and non-proportional hazards checked by AI. Direct full-text retrieval returned 403; original study examples are not reproduced. · Accessed 27 Sep 2026
Prepared and source-checked with AI on 27 September 2026. Press-news Team is the publication’s collective byline, not a claim of medical credentials or human review. No human editorial or clinical review has taken place. Source access is described below each reference. Worked examples are invented for education and do not report a clinical trial or predict an individual outcome.
Source check: AI source check — selected methods references and worked examples
Clinical review: No human editorial or clinical review
Suggest a correction