A hazard ratio is the ratio of the hazard rate in one group to the hazard rate in another, where the hazard rate is the instantaneous chance that a person who has survived event-free up to a given moment experiences the event in the next slice of time. In a survival study it summarises how much faster or slower a treatment group reaches the outcome compared with a control group, holding follow-up time in view rather than ignoring it. Because it is built from time-to-event data, the hazard ratio answers a different question from a simple proportion: not just whether more events happened, but how the speed of those events differs between groups across the whole period of observation. Reading the number correctly means knowing what direction it points, what it does and does not say about cumulative risk, and how tightly the data pin it down.
Reading the number relative to one
The whole interpretation pivots on the value 1. A hazard ratio of exactly 1.0 means the two groups have the same instantaneous event rate at every point in follow-up, so the treatment changes nothing about how quickly the outcome arrives. Values above and below 1 then carry opposite meanings depending on whether the outcome is something you want to avoid or something you want to reach, which is the single most common place readers go wrong.
- For a harmful event such as death, relapse, or disease progression, a hazard ratio above 1 means the treatment group reaches that event faster, so a value of 1.5 indicates a 50 percent higher rate of the bad outcome at any instant during follow-up.
- For the same harmful event, a hazard ratio below 1 is protective: a value of 0.7 means the treatment group experiences the event at 70 percent of the control rate, a 30 percent reduction in the instantaneous hazard.
- For a desirable event such as recovery, discharge, or pregnancy, the reading flips: a hazard ratio above 1 is now good because the treatment group reaches the welcome outcome sooner, while a value below 1 means a slower, less favourable rate.
- The further the number sits from 1 in either direction, the larger the difference in event speed; a hazard ratio of 2.0 and one of 0.5 describe effects of equal magnitude pointing in opposite directions.
Always confirm which event is being modelled before you decide whether a number above 1 is bad news or good news, because the arithmetic is identical and only the clinical framing tells you which way to read it.
Why it is a rate ratio, not a cumulative risk
The most important conceptual point is that a hazard ratio is a rate ratio, not a comparison of total accumulated risk by the end of the study. The hazard is a speed, measured per unit of time among those still at risk, so the hazard ratio compares two speeds. A risk ratio or odds ratio, by contrast, compares the share of people who have experienced the outcome by a fixed point, ignoring when those events happened. That timing dimension is exactly what makes the hazard ratio the right tool when follow-up varies between participants or when some people leave the study before the outcome occurs.
This distinction matters when you translate a hazard ratio into plain language. A hazard ratio of 0.7 does not mean 30 percent fewer people will ever have the event; it means the event arrives at 70 percent of the speed throughout follow-up. The eventual difference in cumulative event counts depends on the baseline rate and how long people are followed, so the same hazard ratio can correspond to a large or a modest absolute difference in outcomes. If your design counts whether an event happened rather than when, the contrast between these binary measures is worth reviewing in our explainer on choosing between odds and risk ratios before you report anything.
How it relates to Kaplan-Meier survival curves
Survival studies are usually displayed as Kaplan-Meier curves, the stepped lines that show the proportion of each group still event-free as time passes. The two curves are a description of the data; the hazard ratio is the single number that summarises the gap between them. When a treatment is protective the treatment curve stays higher for longer, falling more slowly, and that slower descent is what a hazard ratio below 1 captures. The wider and more sustained the separation between the curves, the further the hazard ratio sits from 1.
The connection runs deeper than appearance. A log-rank test compares the curves for a difference, and a Cox proportional-hazards model estimates the hazard ratio that best describes the separation. The curves and the ratio are two views of the same data: one shows the shape over time, the other condenses it to a comparable effect measure you can carry into a table or a figure.
Why the confidence interval is not optional
A hazard ratio reported on its own is half a result. The confidence interval around it tells you the range of effects compatible with the data and whether the estimate is precise enough to act on. A hazard ratio of 0.70 with an interval from 0.55 to 0.89 is a clear protective signal, while the same 0.70 with an interval from 0.40 to 1.22 crosses 1 and is consistent with both benefit and harm. The point estimate is identical; only the interval distinguishes a finding you can rely on from one that is still uncertain.
- When the confidence interval lies entirely below 1, the protective effect is statistically significant for a harmful outcome.
- When it lies entirely above 1, the increased hazard is significant for that same harmful outcome.
- When the interval crosses 1, the result is compatible with no difference and should not be reported as a definite effect.
- A wide interval signals an imprecise estimate, usually from few events, and warns against strong conclusions even when the point estimate looks impressive.
For a fuller treatment of how to read these ranges, including the difference between statistical and clinical importance, see our guide to reading a confidence interval correctly. The same logic governs every hazard ratio you encounter in a published survival study.
The proportional-hazards assumption
The hazard ratio earns its meaning from one condition: the proportional-hazards assumption. This is the requirement that the ratio of hazards between the two groups stays constant over the whole follow-up period. If the treatment effect is the same in early follow-up as it is later, a single hazard ratio describes the difference faithfully. The Cox model assumes this constancy, which is why it can report one number for the entire study.
When the assumption fails, a single hazard ratio becomes a misleading average of effects that actually change over time. This happens when Kaplan-Meier curves cross, when a therapy helps early but loses its edge, or when a benefit only emerges after a delay. Analysts check the assumption with Schoenfeld residuals or by inspecting whether the curves diverge proportionally, and when it is violated they may report time-varying effects or split follow-up into intervals. A hazard ratio quoted without any check on proportionality is a result you should question.
What a meta-analysis actually pools
When survival results are combined across studies, the quantity that gets pooled is not the hazard ratio itself but its logarithm. Software takes the natural log of each study's hazard ratio, pairs it with the standard error of that log value, then forms a weighted average and exponentiates the result back to a pooled hazard ratio. Working on the log scale is necessary because ratios are multiplicative and their sampling distribution is far more symmetric in log space, so the weighting and averaging behave properly.
This is why careful extraction matters so much. Each study must yield a log hazard ratio and its standard error, which can be read directly from a reported hazard ratio and confidence interval or reconstructed from a log-rank statistic and event counts when the paper is less generous. Once those two values are in hand for every study, a pooled survival effect estimator will weight each study by its precision and return the combined hazard ratio with its own interval. You can feed those same per-study values into a tool that will render the combined survival evidence as a forest plot so the spread and the pooled estimate are visible at a glance. If you need the per-study figure first, a dedicated time-to-event effect estimator turns raw counts and follow-up into the ratio you will later combine, and our overview of how evidence synthesis works places that pooling step inside the wider review process.