Hazard Ratio Explained: What Time-to-Event Studies Are Actually Comparing

TLDR: A hazard ratio compares the event rates of two groups over follow-up among participants who are still event-free and being observed. An HR of 0.75 can mean the estimated hazard was 25% lower in one group than another, depending on the group order and endpoint. It does not mean that 25% fewer people experienced the event, that absolute risk fell by 25 percentage points, or that every participant’s risk was reduced by 25%. Read the survival curves, fixed-time event rates, confidence interval, numbers at risk, censoring patterns, and proportional-hazards assessment alongside the HR.

The key to hazard ratio explained medical research is recognizing that time matters. A study of death, recurrence, hospitalization, recovery, or another time-to-event outcome does not merely ask whether an event occurred. It also considers when it occurred and how long each participant remained under observation.

That makes a hazard ratio useful, but also easy to misstate. It compresses a potentially changing relationship between two groups into one relative estimate. To understand the practical effect, readers usually need absolute event probabilities at meaningful times and a view of how the groups’ survival curves behave.

What is a hazard ratio in plain English?

A hazard is the event rate at a particular point in follow-up among participants who have not yet experienced the event and remain under observation. It is conditional on having reached that point event-free; it is not the cumulative probability that an event has happened since enrollment. A hazard can also change over time. The event rate after surgery, for example, may be different in the first month than in the second year.

A hazard ratio, or HR, compares the hazards in two groups. If a paper defines the ratio as treatment divided by control, an HR of 0.75 means the estimated hazard in the treatment group is 0.75 times the hazard in the control group. Under a proportional-hazards interpretation, this is commonly described as a 25% lower hazard during follow-up.

That sentence needs two checks. First, confirm which group is in the numerator. Reversing the groups turns 0.75 into approximately 1.33. Second, identify the event. A lower hazard of death or recurrence is generally favorable, but a lower hazard of recovery or hospital discharge is generally unfavorable. “Lower HR” does not automatically mean “better.”

Why HR 0.75 is not a 25% absolute risk reduction

Calculating 1 minus 0.75 describes the relative difference between the estimated hazards: 25%. It does not provide a conventional cumulative risk reduction. Hazards concern the event rate among the changing set of people still at risk, whereas cumulative risk concerns the proportion who have experienced the event by a stated time.

Consider two invented examples. In Study A, a paper reports HR 0.75 and five-year event risks of 20% in the comparison group and 16% in the intervention group. The absolute difference at five years is 4 percentage points. In Study B, the same HR is reported alongside five-year risks of 4% and 3%, an absolute difference of 1 percentage point. These figures are illustrative rather than values calculated from the HR. The point is that the HR alone does not determine either absolute difference.

Fixed-time survival estimates answer a more concrete question: what proportion is estimated to remain event-free through one, three, or five years? If five-year event-free survival is 84%, the corresponding cumulative event risk is 16%, assuming the survival analysis and endpoint make that complement appropriate.

Measure	Question it answers	What it does not tell you alone
Hazard ratio	How do the groups’ conditional event rates compare over follow-up?	The absolute number or proportion experiencing the event by a fixed time
Risk ratio	How does cumulative event risk by a specified time compare between groups?	The absolute risk difference
Absolute risk difference	How many percentage points apart are the groups at a specified time?	How event timing differed within the interval
Odds ratio	How do the event odds compare?	The risk ratio, especially when the event is common
Restricted mean survival time difference	How much average event-free time through a chosen horizon separates the groups?	What happens after that time horizon

How to interpret HR = 1 and the confidence interval

An HR of 1 indicates equal estimated hazards under the fitted model. Values below or above 1 indicate direction, but the endpoint and numerator determine whether that direction is favorable.

The confidence interval describes uncertainty around the estimate. If a study reports HR 0.75 with a 95% confidence interval of 0.60 to 0.93, the interval excludes 1. If it reports 0.75 with an interval of 0.52 to 1.08, the interval includes values corresponding to both a lower hazard and a higher hazard.

When a 95% confidence interval crosses 1, the result is generally not statistically significant at a two-sided 0.05 threshold. That does not prove the groups are equivalent or that there is no effect. The interval may be wide because the study observed few events. Conversely, an extremely narrow interval can make a small difference statistically significant without making it clinically important.

Reading a Kaplan–Meier curve with the hazard ratio

A Kaplan–Meier curve typically places time on the horizontal axis and estimated survival or event-free probability on the vertical axis. Each observed event causes a downward step. Short marks on a curve often indicate censoring: that participant’s event-free status was known up to that point, but the later event time was not observed.

The curve provides context that a single HR hides. Look for when the curves separate, whether the gap grows or narrows, whether they cross, and how much follow-up remains after separation. A late difference supported by only a small number of participants is less stable than a difference observed while most of the cohort remains under observation.

Always inspect the numbers-at-risk table below the graph. It reports how many participants remain event-free and observed at selected times. When only a small fraction of the original groups remains at risk, the far tail of the curve becomes uncertain and may be driven by very few events.

What censoring means—and when it becomes a problem

A participant is censored when the study does not observe the event after a certain point. This may happen because the study reaches its administrative cutoff, a participant withdraws, or contact is lost. The participant contributes information until the censoring time and then leaves later risk sets.

Standard Kaplan–Meier and Cox analyses rely on an independent or non-informative censoring assumption. In practical terms, censoring should not selectively remove people with systematically different future event risks after accounting for relevant information. Heavy loss to follow-up can bias results if, for example, sicker participants are more likely to disappear from one group.

Administrative censoring at a planned data cutoff is often less concerning because it is driven by the calendar rather than an individual’s prognosis. Even so, readers should compare the amount, timing, and reasons for censoring between groups instead of treating every censoring mark as harmless.

The proportional-hazards assumption

Many reported HRs come from the Cox proportional-hazards model, introduced in David Cox’s 1972 paper on regression models and life tables. The model is designed to estimate a relative hazard without requiring investigators to specify the full baseline hazard over time.

A central assumption is that the hazard ratio is approximately constant over time. The groups’ underlying hazards may both rise or fall, but their ratio should remain reasonably stable. Under that condition, one HR can summarize the relationship across follow-up more coherently.

Researchers can assess this assumption using survival plots, time-by-treatment interactions, log-minus-log plots, and statistical procedures based on Schoenfeld residuals. No single diagnostic is perfect, and a formal test can have low power when few events occur. The observed shape of the curves and the clinical timing of effects still matter.

When one hazard ratio is not enough

Non-proportional hazards occur when the groups’ relative event rates change over time. The curves may separate only after a delay, converge after an early benefit, or cross because one group has early harm followed by later benefit. In those situations, one average HR can obscure the pattern readers most need to understand.

Suppose an intervention raises postoperative complications during the first month but reduces recurrence after the first year. A single HR over three years may look favorable, neutral, or unfavorable depending on event timing and follow-up. None of those summaries communicates the early-versus-late tradeoff adequately.

Useful companion results include survival probabilities and absolute differences at prespecified times, interval-specific effects, median event-free survival when estimable, and restricted mean survival time. The last measure is the area under a survival curve through a defined horizon. A difference in restricted mean survival time can be expressed as additional average event-free time through that horizon, making it a useful complement or alternative when proportional hazards is doubtful.

The time horizon must still be clinically sensible and supported by follow-up. An average gain through three years does not establish what happens after three years.

A practical checklist for reading a reported hazard ratio

Define the endpoint precisely. Overall survival, disease-free survival, hospitalization, recovery, and composite outcomes answer different questions.
Find the group order. Determine which group is in the numerator and which is the reference.
Check whether the event is unfavorable or favorable. This determines whether an HR below 1 points in the preferred direction.
Read the estimate and confidence interval together. Note whether the interval excludes 1 and whether it remains compatible with effects that would be clinically trivial or important.
Look for event counts and follow-up duration. Enrollment size alone does not reveal how much time-to-event information was available.
Inspect the Kaplan–Meier curves and numbers at risk. Note delayed separation, convergence, crossing, censoring patterns, and sparse tails.
Find fixed-time absolute results. Compare survival or cumulative event risk at clinically meaningful, preferably prespecified times.
Check the proportional-hazards assessment. If hazards vary over time, look for time-specific estimates or restricted mean survival time.
Review how intercurrent events were handled. Treatment switching, discontinuation, rescue therapy, and competing events can change what the estimate means. The ICH E9(R1) estimand guideline emphasizes aligning the analysis with the clinical question.
Identify the study design. Randomization supports a causal interpretation when the trial is well conducted; adjustment in observational data does not recreate randomization.

Does an adjusted hazard ratio establish causation?

No. In an observational study, an adjusted HR is an association conditional on the variables and model used. Adjustment can reduce measured confounding, but unmeasured confounding, selection bias, exposure misclassification, informative censoring, and model misspecification may remain.

A randomized trial offers stronger protection against confounding, but its HR must still be interpreted in light of adherence, missing data, treatment switching, endpoint definitions, competing events, and proportional hazards. A hazard ratio for a surrogate endpoint also does not automatically establish improvement in survival, symptoms, function, or quality of life.

Frequently asked questions

Does HR 0.75 mean a treatment prevents 25% of events?

No. It indicates a 25% lower estimated hazard when the treatment group is the numerator, subject to the model and time pattern. The number of events prevented depends on baseline risk, follow-up duration, censoring, and how effects vary over time. Use fixed-time absolute risks to estimate the event difference.

Is an HR of 0.90 clinically meaningful?

It could be, but the HR alone cannot answer. A 10% relative hazard reduction may correspond to a meaningful absolute difference for a common, serious event or a tiny difference for a rare event. Consider absolute outcomes, uncertainty, adverse effects, follow-up, and patient-relevant importance.

What if the confidence interval crosses 1?

The data do not exclude equal hazards at the conventional confidence level. They may also remain compatible with clinically important benefit or harm, especially when the interval is wide. Treat the interval as a range of estimates reasonably compatible with the data and model, not as a binary verdict.

What should I do when survival curves cross?

Be cautious about summarizing the result with one HR. Look for prespecified time-specific survival estimates, absolute differences, restricted mean survival time, and an analysis designed for non-proportional hazards. Also check whether the crossing occurs while substantial numbers remain at risk.

Can I calculate number needed to treat from a hazard ratio?

Not from the HR alone. Number needed to treat requires an absolute risk difference over a stated period and depends on the chosen outcome and time horizon. A study must report or support estimation of those absolute risks.

The most defensible takeaway

Treat a hazard ratio as a relative comparison of conditional event rates over follow-up—not as a percentage-point change in risk or a complete description of what happened. Confirm the endpoint and group order, then pair the HR with its confidence interval, Kaplan–Meier curves, numbers at risk, fixed-time absolute probabilities, censoring information, and proportional-hazards assessment.

If those supporting details are missing, the HR may answer a narrower question than the headline suggests. For personal medical decisions, discuss the absolute benefits, harms, and relevant timeframe with a qualified clinician rather than applying a study’s HR directly to an individual prognosis.