How to Interpret a Hazard Ratio

A hazard ratio (HR) is one of the most consistently misread numbers in the clinical and epidemiological literature. It gets reported alongside a 95% confidence interval from a Cox proportional-hazards model, and then gets talked about as if it were a risk ratio — “patients on the new drug were 30% more likely to die.” That sentence is usually wrong, or at least imprecise in a way that changes what a reader takes away. A hazard ratio is a ratio of instantaneous event rates, summarized across the follow-up period under a proportionality assumption — it is not a ratio of the probability that an event ever happens to a given patient. This guide works through what a reported HR is actually telling you, gives a lookup table for translating a value into a plain-language statement, and covers the assumption check that determines whether a single HR is even a valid summary of your data in the first place.

A Worked Example First

Suppose a randomized trial compares a new maintenance therapy against standard care for time to disease progression, and the Cox proportional-hazards model produces this line in the results table:

Comparison	Hazard Ratio	95% CI	p-value
New therapy vs. standard care	0.68	0.51 – 0.90	0.007

The sentence a statistician would write for the results section: “The new therapy was associated with a 32% reduction in the instantaneous hazard of disease progression relative to standard care (HR 0.68, 95% CI 0.51–0.90, p = 0.007), consistent with the proportional-hazards assumption over the observed follow-up period.” Notice what is present and what is deliberately absent. The HR describes a rate reduction, not “32% of patients were spared progression,” and not a claim about how much longer any individual patient will remain progression-free — that is a separate question answered by median survival times or restricted mean survival time, not by the HR itself. The confidence interval excluding 1.0 is what licenses calling the reduction statistically significant at the conventional alpha = 0.05 threshold; the point estimate of 0.68 alone does not.

What a Hazard Ratio Actually Measures

The hazard function, h(t), is the instantaneous rate at which an event occurs at time t, given that a subject has survived event-free up to that point. A Cox proportional-hazards model assumes the hazard in one group is a constant multiple of the hazard in a comparison group at every point in time — that constant multiple is the hazard ratio. An HR of 1.0 means the two groups have the same instantaneous event rate throughout follow-up. An HR of 2.0 means one group’s instantaneous rate is twice the other’s at any given moment, not that twice as many people in that group eventually experienced the event.

Because the HR is estimated from the full follow-up period and from everyone still at risk at each event time, it is a summary measure, not a snapshot at one moment. That is precisely why the proportional-hazards assumption (the ratio staying roughly constant over time) matters so much: if it does not hold, a single reported HR is averaging over a relationship that is actually changing shape, and can misrepresent both the early and late parts of the follow-up.

Why an HR Is Not a Risk Ratio (or an Odds Ratio)

A risk ratio (relative risk) compares the cumulative probability of an event by a fixed point in time between two groups — a simple proportion. A hazard ratio compares instantaneous rates among those still event-free, re-evaluated continuously across follow-up. When an event is rare and follow-up is short, the two converge and an HR can be read loosely as an approximate risk ratio. As cumulative incidence rises, they diverge, and the HR is typically the more extreme number of the two (further from 1 than the corresponding risk ratio) — not because the treatment effect is more extreme, but because of how the two statistics are constructed.

Part of that divergence is a documented statistical property, not a rounding quirk: hazard ratios are non-collapsible and subject to a form of built-in selection bias from the depletion of susceptible individuals in the higher-hazard group over time, as the higher-risk members of that group experience the event (and leave the risk set) earlier. This is discussed at length in the biostatistics literature, notably Miguel Hernán’s widely cited note “The Hazards of Hazard Ratios” (Epidemiology, 2010) — worth reading directly if the HR in your own analysis is being used to support a causal claim rather than reported as a descriptive association. The practical takeaway for interpretation: never convert a reported HR into a “times more likely” or “% more likely” statement about outcomes without stating explicitly that this refers to the instantaneous hazard, not cumulative risk.

Translating an HR Into a Plain-Language Statement

The following is a reading guide, not a set of fixed cutoffs — unlike some effect-size measures, there is no widely adopted convention (comparable to, say, Cohen’s d thresholds) for calling a hazard ratio “small” or “large.” What counts as a meaningful HR depends entirely on the outcome, the baseline event rate, and the clinical or research context.

Hazard Ratio	Plain-language reading
1.00	No difference in instantaneous event rate between groups.
0.90 – 0.99	Slightly lower hazard in the comparison group; check whether the CI excludes 1.0 before treating this as a real effect.
0.50 – 0.89	Meaningfully lower instantaneous hazard (roughly an 11–50% relative reduction in event rate at any given moment).
< 0.50	Hazard reduced by more than half at any given moment — a large effect if the CI is precise and the proportional-hazards assumption holds.
1.01 – 1.10	Slightly higher hazard; again, check the CI before treating this as a real effect.
1.10 – 2.00	Meaningfully higher instantaneous hazard (10–100% relative increase in event rate at any given moment).
> 2.00	Hazard more than doubled at any given moment — report alongside absolute event counts, since a large HR on a rare outcome can still mean a small absolute difference.

Every row in that table is a statement about instantaneous rate, and every row should be paired with the confidence interval before it is described as an effect at all, not just a point estimate.

Reading the Confidence Interval Together With the HR

A hazard ratio without its confidence interval is close to unreportable. The 95% CI reflects the precision of the estimate: a narrow interval close to the point estimate (e.g., HR 0.68, 95% CI 0.60–0.77) supports a confident reading; a wide interval (e.g., HR 0.68, 95% CI 0.32–1.45) means the data are consistent with anything from a large protective effect to a modestly harmful one, and the point estimate alone overstates what is actually known. If the interval crosses 1.0, the result is not statistically significant at conventional thresholds regardless of how far the point estimate itself sits from 1.0 — see CASRAI’s guide to confidence intervals for how the interval is constructed and what it does and does not claim about repeated sampling.

The Assumption That Makes a Single HR Valid: Proportional Hazards

Reporting one HR for the entire follow-up period is only defensible if the hazard ratio is genuinely stable over time — the proportional-hazards assumption. Before reporting a Cox model’s HR as a single summary number, check it:

Kaplan–Meier curves for the groups should not cross. Crossing curves are a visual red flag that the direction of effect itself may be reversing over time, which a single HR cannot represent.
Log(−log(survival)) plots against log(time) should be roughly parallel across groups; converging, diverging, or crossing lines indicate non-proportionality.
Schoenfeld residuals, tested formally (e.g., via `cox.zph()` in R), should show no significant trend over time for a covariate. A significant test result is direct statistical evidence against the proportionality assumption for that covariate.

When the assumption fails, options include reporting time-varying hazard ratios (splitting follow-up into intervals with a separate HR per interval), adding a time interaction term to the model, or switching to a method that does not require the assumption (e.g., restricted mean survival time, which summarizes an absolute time difference rather than a rate ratio). Reporting a single HR without checking this assumption — and without saying whether it was checked — is one of the most common and consequential gaps in published survival analyses.

Other Conditions Worth Checking

Non-informative censoring. The Cox model assumes that being censored (lost to follow-up, study ends) is unrelated to a subject’s underlying risk of the event. If sicker patients are disproportionately lost to follow-up, the HR can be biased in either direction.
Enough events, not just enough subjects. Precision in a Cox model is driven by the number of events observed, not the number of subjects enrolled. A trial with many enrolled patients but few events will still produce a wide, imprecise confidence interval around the HR.
Competing risks. If subjects can experience a different, competing event that removes them from being able to experience the event of interest (e.g., death from another cause before disease progression), a standard Cox HR can be inflated or biased; a subdistribution hazard model (Fine–Gray) or cause-specific hazard framing may be more appropriate, and the choice should be stated explicitly.
Multiple covariates. An adjusted HR from a multivariable Cox model answers a different question than an unadjusted (crude) HR — state which one is being reported and what was adjusted for. See CASRAI’s confounding variable entry for what adjustment is and is not able to control for.

Frequently Asked Questions

Is a hazard ratio of 1.5 the same as “50% more likely to die”?

No. It means the instantaneous rate of the event at any given moment during follow-up is 1.5 times higher in one group than the other, under the proportional-hazards assumption. It is not a statement that 50% more people in that group experienced the event by the end of the study — that is a cumulative-risk claim, and a risk ratio (not a hazard ratio) is the correct statistic for it.

Can a hazard ratio be negative?

No. A hazard ratio is a ratio of two non-negative rates, so it is always greater than or equal to zero. Values below 1.0 indicate a lower hazard in the numerator group; values above 1.0 indicate a higher hazard; a value of exactly 1.0 indicates no difference.

What sample size or number of events does a Cox model need?

There is no single universal minimum, but the relevant quantity is the number of events, not enrolled subjects — a widely used rule of thumb in applied survival analysis is roughly 10 events per covariate included in the model, though this is a heuristic rather than a formal requirement, and simulation-based power calculations are preferable when feasible.

Why do published hazard ratios sometimes look more extreme than the corresponding risk ratio in the same paper?

This is the non-collapsibility and built-in selection-bias property described above: because a Cox model conditions on the risk set surviving to each event time, and because susceptible individuals are depleted from the higher-hazard group earlier, the HR is generally pulled further from 1.0 than a risk ratio computed over the same data, even absent any modeling error.

How to Interpret a Hazard Ratio (and Why It Is Not a Risk Ratio)

Ask about How to Interpret a Hazard Ratio (and Why It Is Not a Risk Ratio)