Hazard Ratio: What It Means, How to Interpret It, and How It Differs From Odds Ratio and Relative Risk

By Dr. Zubair Khalid, DVM, MS, PhD · 2026-10-02

A hazard ratio compares how fast an event happens in two groups over time. If you follow patients until they relapse, die, heal, or leave the study, the hazard ratio tells you how the instantaneous event rate in one group relates to the other at any point during follow-up. It is the standard effect measure in time-to-event research, and it appears in almost every oncology trial, cardiology cohort study, and registry analysis you will read.

You need to understand it properly because the hazard ratio is easy to misread. It is not a risk ratio, it is not an odds ratio, and it does not tell you how much longer people survive. Journals report it alongside Kaplan-Meier curves and logrank tests, and clinical decisions often hinge on whether the confidence interval crosses 1. Getting the interpretation right matters for reading papers, designing analyses, and explaining results to colleagues.

Quick Answer

The hazard h(t) is the instantaneous event rate at time t among those still at risk, meaning people who have survived event-free up to that point [2].
The hazard ratio (HR) is the hazard in the treated or exposed group divided by the hazard in the control group [1].
HR = 1 means no difference in hazard. HR < 1 means lower event rate in the treated group. HR > 1 means higher event rate [2].
In a Cox proportional hazards model, HR = exp(b), where b is the regression coefficient for the group variable [3].
A 95% confidence interval that excludes 1 indicates a statistically significant difference at the 5% level.
The HR describes relative rates, not absolute risks or survival times. An HR of 0.70 does not mean 30% fewer events or 30% longer survival.

What the Hazard Actually Is

The hazard is a rate, not a probability in the usual sense. It answers a conditional question: given that a person has not had the event yet, how likely is the event right now? Formally, the hazard is the probability that the event occurs in a very short interval, given it has not already occurred, divided by the length of that interval [1]. As the interval shrinks toward zero, you get the instantaneous rate.

NIST defines the failure or hazard rate as:

$$h(t) = \frac{f(t)}{1 - F(t)} = \frac{f(t)}{R(t)}$$

where f(t) is the event-time density, F(t) is the cumulative distribution of event times, and R(t) = 1 - F(t) is the survival or reliability function [6]. The denominator R(t) is the key: only people still at risk contribute to the hazard at time t. Units that have already failed drop out of the calculation [6].

This is why the hazard is not the same as the survivor function. The survivor function S(t) tracks cumulative non-occurrence, the probability of making it past time t without the event. The hazard tracks the current event rate among survivors [2]. One is a stock, the other is a flow.

A constant hazard over time corresponds to an exponential distribution of survival times [2]. That special case makes the math clean, but real hazards often change: surgical risk is high early and falls, cancer recurrence risk can peak months after treatment, and drug toxicity may rise with cumulative exposure.

How to Calculate a Hazard Ratio

In practice you rarely compute a hazard ratio by hand. You fit a Cox proportional hazards model, introduced by D.R. Cox in 1972 [4]. The model is:

$$h(t) = h_0(t) \times \exp(b_1 x_1 + b_2 x_2 + \dots + b_p x_p)$$

where h₀(t) is the baseline hazard (the hazard when all covariates equal zero), x₁ through xₚ are the covariates, and b₁ through bₚ are the coefficients [3]. The baseline hazard is estimated nonparametrically, so the model does not assume survival times follow any particular distribution [3]. Covariates act multiplicatively on the hazard, and the model is essentially a multiple linear regression of the log hazard on the covariates, with the baseline hazard playing the role of a time-varying intercept [3].

The quantities exp(bᵢ) are hazard ratios. A coefficient above 0 gives an HR above 1, meaning the hazard rises and survival shortens as the covariate increases [3].

With two groups and no covariates, the logrank test tests the null hypothesis that the ratio of hazard rates equals 1 [2]. From a logrank analysis, you can estimate the HR as:

$$\text{HR} = \frac{O_1 / E_1}{O_2 / E_2}$$

where O₁ and O₂ are the observed event counts in each group and E₁ and E₂ are the expected counts under the null [2]. This is a useful approximation, though Cox regression gives you a confidence interval and handles covariates.

A 95% confidence interval for the HR is computed on the log scale:

$$\exp\left(b \pm 1.96 \times \text{SE}(b)\right)$$

where b is the Cox coefficient and SE(b) is its standard error. Because the transformation is exponential, the interval is asymmetric around the point estimate.

How to Interpret a Hazard Ratio

Start with the direction. An HR of 1 means no difference in hazard between groups [2]. An HR below 1 means the treated group has a lower event rate at any given time, assuming proportional hazards. An HR above 1 means the treated group has a higher event rate.

The magnitude is a ratio of rates. An HR of 0.70 means the treated group's event rate is 30% lower than the control group's at any point during follow-up, again assuming proportional hazards. This is similar in spirit to a risk ratio, but it applies to rates, not cumulative probabilities [2]. Hernan notes that hazards can be thought of as incidence rates for practical purposes, so a hazard ratio can be roughly read as an incidence rate ratio [5].

Now the confidence interval. Suppose a trial reports HR 0.70 (95% CI 0.55 to 0.89). The point estimate says the treated group's event rate is about 30% lower. Because the interval excludes 1, the result is statistically significant at the 5% level. The data are compatible with reductions as small as 11% and as large as 45%. That range matters for judging clinical importance.

One common heuristic: if the HR is for a good event such as healing, the probability that a treated patient heals before a control patient is P = HR / (1 + HR) [1]. An HR of 2 corresponds to a 67% chance. For a harmful event with HR 0.70, HR / (1 + HR) = 0.41, so in a random treated-control pair the treated patient has the event first 41% of the time and the control patient 59% of the time. This interpretation assumes proportional hazards and ignores ties and censoring, so treat it as a rough guide.

The HR does not translate directly into the duration of time until events [1]. Its magnitude may be greater or smaller than the benefit seen in median survival times. Spruance and colleagues compare hazard-based and time-based measures to the odds of winning a race versus the margin of victory [1]. You can have a large HR with a small median difference, or a modest HR with a large median difference, depending on the shape of the hazard over time.

Worked Example

Consider a simulated two-arm trial with 400 patients per arm. Event times are exponential: the control hazard is 0.03 per month and the treated hazard is 0.021 per month, giving a true HR of 0.70. Administrative censoring occurs at 24 months.

Observed events: 199 in the control arm, 140 in the treated arm. A Cox proportional hazards fit with treatment as the only covariate gives a coefficient b = -0.4799 with SE 0.1104. The hazard ratio is exp(-0.4799) = 0.619, with a 95% CI of exp(-0.4799 ± 1.96 × 0.1104) = 0.498 to 0.768, and p = 1.4e-5. A second implementation using Breslow ties gives the identical HR of 0.619 (0.498 to 0.768). The logrank test gives chi-square 19.26, p = 1.1e-5. A proportional hazards check using rank-transformed Schoenfeld residuals gives p = 0.54, showing no evidence against proportionality, which is expected because the data were simulated with proportional hazards.

Kaplan-Meier survival at 24 months is 0.503 in the control arm and 0.650 in the treated arm. At 12 months it is 0.728 versus 0.825. Median survival was not reached in either arm by 24 months because S(24) is still above 0.5. This is a common situation where an HR is reported but medians are not.

The estimate of 0.619 differs from the true 0.70 because of sampling error. The true value lies inside the confidence interval. From the same hazards, the true-model contrasts are: 12-month risks of 30.2% versus 22.3% (risk ratio 0.737, absolute risk reduction 7.96 percentage points); 24-month risks of 51.3% versus 39.6% (risk ratio 0.771, odds ratio 0.621); true medians of 23.1 versus 33.0 months (ratio 1.43, which equals 1/0.70 and is valid only because the hazards are constant); and a probability of 0.70/1.70 = 0.41 that the treated member of a random pair has the event first.

Notice that the HR (0.70), the 24-month risk ratio (0.771), and the 24-month odds ratio (0.621) are three different numbers describing one data-generating process. They answer different questions.

Hazard Ratio vs Odds Ratio vs Relative Risk

These three measures get confused constantly. Here is how they differ.

Measure	What it compares	Scale	Typical use
Hazard ratio	Instantaneous event rates over time	Rate ratio	Time-to-event data with censoring
Relative risk	Cumulative risk (proportion with event)	Risk ratio	Fixed follow-up, no censoring
Odds ratio	Odds of event (p / (1 - p))	Odds ratio	Case-control studies, logistic regression

The relative risk is the ratio of the probability of an event in the exposed group to the probability in the unexposed group. It requires a defined follow-up period and works best when everyone is followed for the same length of time. The odds ratio compares the odds of the event, where odds = probability / (1 - probability). It is the natural output of logistic regression and the usual measure reported from a case-control study. You can compute one from a 2x2 table with the Odds Ratio Calculator.

The hazard ratio is different because it accounts for time. It uses all the information in the follow-up period, including when events occur, and it handles censoring. Censoring happens when a patient has not had the event by study close, is lost to follow-up, or has a different event that stops follow-up. Censored times underestimate the true time to event, which is called right censoring [2]. The HR is designed for this setting.

In the worked example, the true HR is 0.70, the 24-month risk ratio is 0.771, and the 24-month odds ratio is 0.621. The HR is not interchangeable with either. If you reported the odds ratio as if it were the hazard ratio, you would overstate the effect in this example. If you reported the risk ratio as if it were the hazard ratio, you would understate it slightly in this case, but the direction and size of the discrepancy depend on the baseline risk and the follow-up time.

For a related discussion of when relative risk and odds ratio diverge, see the article on relative risk vs odds ratio interpretation.

Common Mistakes

Treating the HR as a risk ratio. An HR of 0.70 does not mean 30% fewer patients have the event. It means the event rate at any point in time is 30% lower. The cumulative risk difference depends on follow-up length and baseline hazard.
Treating the HR as a survival time ratio. An HR of 0.70 does not mean 30% longer survival. Only under a constant hazard does the ratio of medians equal 1/HR, and even then the median may not be reached.
Ignoring the proportional hazards assumption. The HR is only meaningful as a single summary if the hazard ratio is constant over time [1][3]. If hazards cross or diverge, a single HR hides the pattern.
Reading a confidence interval that includes 1 as proof of no effect. A wide interval that crosses 1 may be compatible with a meaningful benefit or harm. The width reflects sample size and event count.
Confusing the HR with the odds ratio from logistic regression. They are different measures from different models. Logistic regression ignores time and censoring.
Reporting the HR without the confidence interval. The point estimate alone does not convey precision. Always report the interval.
Assuming the HR applies to every patient. The HR is an average over the study population and follow-up period. Subgroups may differ.

Limitations

The proportional hazards assumption is the big one. It says the hazard in one group is a constant multiple of the hazard in the other, so the hazard curves are proportional and cannot cross [3]. Spruance and colleagues state the same assumption as the hazard ratio being constant over time [1]. When this fails, a single HR is misleading.

Hernan identifies two problems with causal interpretation of HRs. First, the HR may change over time. Second, period-specific HRs have a built-in selection bias [5]. The hazard in period t is computed only among people who reached period t free of the outcome, and those survivors may differ between arms. In the Women's Health Initiative hormone trial, the overall HR for coronary heart disease was 1.24, while year-specific HRs were 1.81, 1.34, 1.27, 1.25, 1.45, and 0.70 for years 1 to 5 and 6 or more [5]. The overall HR averages over a pattern that changes direction.

Because a single HR averages over follow-up, conclusions can depend on how long the study ran [5]. Hernan recommends summarizing results with appropriately adjusted survival curves and notes accelerated failure time models as an alternative to Cox models [5].

Proportional hazards tests have low power in small trials and high power to detect trivial departures in large ones. Plots of Schoenfeld residuals or log(-log S) curves are commonly used, but no single diagnostic is definitive. The HR/(1 + HR) interpretation assumes proportional hazards and ignores ties and censoring. The median ratio = 1/HR relation holds only for exponential hazards. With other distributions, such as Weibull with shape not equal to 1, the relation differs.

Frequently Asked Questions

What is a hazard ratio in plain language?

A hazard ratio compares how fast an event happens in two groups. If the HR is 0.70, the treated group's event rate at any point in time is about 30% lower than the control group's, assuming the ratio stays constant. It is not a risk ratio and not a survival time ratio.

How do I interpret a hazard ratio confidence interval?

The 95% CI gives the range of HR values compatible with the data. If the interval excludes 1, the result is statistically significant at the 5% level. If it includes 1, the data are compatible with no difference, though the interval may still include meaningful effects.

What is the difference between a hazard ratio and an odds ratio?

A hazard ratio compares event rates over time and accounts for censoring. An odds ratio compares the odds of an event and ignores time. In the worked example, the true HR is 0.70 while the 24-month odds ratio is 0.621. They are not interchangeable.

What is the difference between a hazard ratio and relative risk?

Relative risk compares cumulative probabilities over a fixed period. The hazard ratio compares instantaneous rates over time. In the worked example, the true HR is 0.70 while the 24-month risk ratio is 0.771. The HR uses more information and handles censoring.

How is a hazard ratio calculated?

In practice, you fit a Cox proportional hazards model and exponentiate the coefficient for the group variable. With two groups and no covariates, you can approximate the HR from a logrank analysis as (O₁/E₁)/(O₂/E₂), the ratio of observed to expected events in each group [2].