Hazard Ratio Interpretation for Clinicians

where \(T\) is the event time.

The conditioning on \(T\ge t\) is important. The hazard is calculated among individuals who have remained at risk up to time \(t\).

What Is a Hazard Ratio?

Suppose there are two treatment groups:

• Experimental treatment
• Control treatment

Let their hazard functions be:

The hazard ratio is:

If the proportional-hazards assumption is reasonable, this ratio is treated as approximately constant over time:

for the relevant follow-up period.

Thus, a hazard ratio of 0.70 means that, at a given time among patients who remain at risk, the estimated event hazard in the experimental group is approximately 70% of that in the control group.

The Three Fundamental Hazard-Ratio Values

The direction always depends on which group is in the numerator.

If the hazard ratio is defined as:

then:

• HR < 1 favors treatment when the event is undesirable.
• HR = 1 indicates equal hazards.
• HR > 1 favors control when the event is undesirable.

How to Interpret HR = 0.70

Suppose a trial reports:

If treatment is compared with control and the event is undesirable, the usual clinical interpretation is:

The treatment group has an estimated hazard that is 30% lower than the control group.

The corresponding relative reduction in hazard is:

or:

This is commonly called the relative reduction in hazard.

How to Interpret HR = 1.30

Suppose instead:

The treatment group has an estimated hazard 30% higher than the control group.

Thus the relative increase in hazard is approximately 30%.

For an undesirable event, an HR above 1 generally indicates a worse outcome.

However, if the event is desirable, such as recovery or remission, an HR above 1 may represent a favorable treatment effect.

The Event Definition Matters

The same numerical hazard ratio can have completely different clinical meanings depending on the endpoint.

Therefore, saying "an HR of 0.75 is good" is incomplete.

You must know:

• Which group is in the numerator
• Which group is the reference
• What constitutes the event
• Whether a lower event hazard is clinically desirable

Hazard Ratio Is Not Risk Ratio

This is one of the most important distinctions for clinicians.

A risk ratio compares probabilities over a specified period:

A hazard ratio instead compares event hazards.

These quantities are not generally interchangeable.

A Simple Example

Suppose that after two years:

The two-year risk ratio is:

The treatment reduces the observed two-year risk by 25% relative to control.

But this does not mean that the hazard ratio must equal 0.75.

The hazard ratio incorporates the timing of events and censoring throughout follow-up.

Why Time Matters

Consider two treatments with exactly the same number of events by five years.

The clinical experience could still be very different if events occur earlier in one group.

For example:

• Control events occur primarily during the first year.
• Treatment events occur primarily during years four and five.

A simple five-year event proportion does not describe this difference well.

Survival analysis incorporates the timing of events.

That is why hazard ratios are particularly useful for time-to-event endpoints.

Hazard Ratio and the Kaplan-Meier Curve

The Kaplan-Meier curve estimates the probability of remaining event-free over time.

For a survival endpoint, the survival function is:

The hazard function and survival function are related.

Thus, the hazard contributes to the cumulative probability of remaining event-free.

The hazard ratio compares the underlying hazard functions, whereas the Kaplan-Meier curve shows the resulting survival experience over time.

What Does HR = 0.50 Mean?

An HR of 0.50 is often described as a 50% reduction in hazard.

That statement is appropriate as a relative hazard interpretation.

But it does not mean:

• 50% fewer patients will experience the event.
• Patients will live twice as long.
• The median survival will double.
• Absolute risk is reduced by 50 percentage points.

For example, suppose the event probability at a clinically important time is:

The absolute risk reduction is:

or 15 percentage points.

The relative risk reduction is:

or 37.5%.

Neither quantity is necessarily equal to the hazard-ratio reduction.

Relative Effect vs. Absolute Effect

Clinicians generally need both relative and absolute measures.

Confidence Intervals

A hazard ratio is an estimate rather than an exact population quantity.

Therefore, it is usually reported with a confidence interval.

Suppose:

The point estimate suggests a 28% lower hazard.

The confidence interval indicates that values from 0.58 to 0.90 are compatible with the statistical uncertainty represented by the interval under the specified confidence procedure.

Because the entire interval is below 1, the result is statistically incompatible with a null hazard ratio of 1 at the corresponding two-sided 5% significance level.

What If the Confidence Interval Crosses 1?

Suppose:

The point estimate favors treatment.

However, the confidence interval includes 1.

Therefore, the study does not provide conventionally statistically significant evidence of a difference at the 5% two-sided level.

The Confidence Interval Is Often More Informative Than the P-Value

Suppose two studies produce:

Both point estimates are identical.

But the evidence is very different.

Study A provides a highly precise estimate. Study B is much more uncertain and is compatible with substantial benefit, little effect, or potential harm.

Converting a Hazard Ratio to a Percentage Reduction

When HR < 1, a common summary is:

Examples:

For HR > 1, the relative increase is:

For example:

Do Not Say "40% Higher Risk" Automatically

An HR of 1.40 is not automatically equivalent to a 40% higher probability of experiencing the event by the end of follow-up.

The safer wording is: "The estimated hazard was 40% higher."

If you want to discuss risk, report a risk estimate at a specified time.

Hazard Ratios From the Cox Proportional Hazards Model

One of the most common ways to estimate hazard ratios is the Cox proportional hazards model.

The model is:

Here:

• \(h_0(t)\) is the baseline hazard.
• \(X_1,\ldots,X_p\) are covariates.
• \(\beta_1,\ldots,\beta_p\) are regression coefficients.

For a one-unit increase in \(X_j\), the hazard ratio is:

Thus, the coefficient is on the log-hazard scale while the hazard ratio is on the multiplicative hazard scale.

Example: Treatment Indicator

Suppose treatment is coded:

If:

then:

The treatment is therefore associated with an estimated 30% lower hazard than the reference group, assuming the model and proportional-hazards interpretation are appropriate.

Adjusted vs. Unadjusted Hazard Ratios

Clinical trials may report both unadjusted and adjusted hazard ratios.

An unadjusted HR compares the treatment groups without adjusting for additional covariates in the model.

An adjusted HR comes from a multivariable model containing additional covariates.

In a randomized trial, the primary treatment effect is often reported from a prespecified analysis model, which may be stratified or adjusted for important design factors.

Continuous Covariates

Hazard ratios can also describe continuous predictors.

Suppose age is included as a continuous covariate and:

If age is measured in years, the interpretation is: Each additional year of age is associated with a 12% higher hazard, holding the other model covariates constant.

But this interpretation depends on the model's functional form. If age was modeled linearly, it assumes the log-hazard relationship with age is linear unless nonlinear terms are included.

Categorical Covariates

Suppose disease stage has three categories:

• Stage I
• Stage II
• Stage III

If Stage I is the reference category and Stage III has:

the interpretation is: Patients in Stage III have an estimated 2.10-fold hazard compared with Stage I, conditional on the other model terms.

Equivalently, the hazard is estimated to be approximately 110% higher:

What Does "Holding Other Variables Constant" Mean?

For an adjusted Cox model, the hazard ratio for a covariate describes the multiplicative hazard comparison associated with that covariate while the other model covariates are held fixed.

This is a model-based conditional interpretation.

It should not automatically be interpreted as proof that the covariates are causally responsible for the observed hazard difference.

The Proportional-Hazards Assumption

The classic Cox model assumes that hazard ratios are constant over time.

For two groups, this means:

does not systematically change with \(t\).

This is the proportional-hazards assumption.

What Happens When Hazards Are Not Proportional?

Suppose a treatment has little effect during the first six months but produces a substantial benefit later.

The hazard ratio might therefore be:

with a meaningful dependence on time.

In that situation, the Cox model's single HR may represent an overall summary of a time-varying treatment effect rather than a literal constant hazard ratio at every time point.

Crossing Kaplan-Meier Curves

One warning sign for nonproportional hazards is crossing survival curves.

For example:

• Treatment performs worse early.
• The treatment effect improves later.
• The survival curves eventually cross.

A single HR can obscure this clinically important pattern.

Hazard Ratio and Median Survival

Clinical reports often contain both a hazard ratio and median survival.

For example:

The HR describes the relative hazard over the modeled follow-up period.

The median survival compares the time at which the estimated survival probability reaches 50%.

These are different summaries.

A Common Special Case: Exponential Survival

Under a simple exponential survival model with a constant hazard, the median survival is related to the hazard by:

In that very restrictive setting, a constant hazard ratio can correspond to an inverse relationship between median survival.

But clinical survival data rarely justify assuming such a simple exponential model automatically.

Therefore, clinicians should not generally calculate a median-survival ratio from a hazard ratio.

Hazard Ratio Does Not Mean "Patients Live 30% Longer"

Suppose:

The correct relative interpretation is approximately a 30% lower hazard.

It is not correct to conclude: "Patients live 30% longer."

Time-to-event distributions do not generally transform that way.

Hazard Ratio Does Not Give the Number Needed to Treat

The number needed to treat depends on an absolute risk difference over a specified time horizon.

If treatment reduces a two-year risk from 20% to 15%, then:

and:

The NNT therefore requires an absolute risk difference.

An HR alone does not determine the NNT.

Why Baseline Risk Matters

Consider two clinical populations with the same hazard ratio:

Suppose one population has a low baseline event risk and another has a high baseline event risk.

The relative treatment effect may be similar, but the absolute benefit can be very different.

This is why clinicians should combine relative treatment effects with absolute event probabilities whenever possible.

Reading a Clinical Trial Hazard-Ratio Table

Consider:

For overall survival, the treatment is associated with an estimated 28% lower hazard of death.

The confidence interval is entirely below 1, supporting evidence of a difference at the conventional 5% level.

For progression-free survival, the estimated hazard is 35% lower.

The confidence interval is also entirely below 1.

However, a clinician should still ask:

• What are the absolute survival probabilities?
• How large is the median survival difference, if estimable?
• How long was follow-up?
• Were the curves proportional?
• What is the confidence interval for clinically relevant survival milestones?
• Is the magnitude of benefit clinically meaningful?

Statistical Significance vs. Clinical Significance

A hazard ratio can be statistically significant without representing a large clinical benefit.

For example:

With a very large sample size, this can be highly statistically significant.

But a 5% relative reduction in hazard may or may not be clinically important.

Conversely, an HR of 0.70 with a very wide confidence interval may be clinically exciting but statistically uncertain.

Hazard Ratio and Censoring

Survival analysis commonly involves censoring.

For example, a patient may:

• Complete follow-up without experiencing the event.
• Withdraw from the study.
• Be lost to follow-up.
• Reach administrative study end without an event.

The exact treatment of censoring depends on the study design and analysis assumptions.

Censoring allows information from patients with incomplete event histories to contribute to the analysis rather than simply discarding them.

Competing Risks Can Change Interpretation

Some clinical settings contain competing events.

For example, when studying cardiovascular death, a noncardiovascular death may prevent the patient from subsequently experiencing cardiovascular death.

In such settings, the cause-specific hazard and the subdistribution hazard answer different questions.

Therefore, clinicians should determine exactly which survival-analysis method was used before interpreting an HR.

Cause-Specific Hazard Ratio

A cause-specific hazard treats other competing events as censoring for the cause-specific hazard analysis.

The interpretation is about the instantaneous rate of the event of interest among individuals who remain free of any event that prevents that event from occurring.

This is not automatically the same as the treatment effect on the cumulative incidence probability of that event.

Subdistribution Hazard Ratio

A competing-risks analysis may instead use a subdistribution hazard model, such as the Fine-Gray approach.

The resulting subdistribution hazard ratio has a different interpretation from a cause-specific HR.

Stratified Cox Models

Clinical trials sometimes use stratified Cox models.

For example, randomization may be stratified by:

• Disease stage
• Geographic region
• Prior treatment
• Risk group

A stratified Cox model allows the baseline hazard to differ across strata while estimating a common treatment hazard ratio under the model.

The exact model specification matters when interpreting the reported HR.

Subgroup Hazard Ratios

Clinical trials often report HRs for subgroups.

It is tempting to conclude that the treatment works in younger patients but not older patients.

That conclusion is not justified simply because one confidence interval excludes 1 and the other does not.

Different HRs Do Not Automatically Mean Different Effects

Suppose two subgroups have:

These point estimates differ numerically.

But the observed difference could easily be due to sampling variation.

A formal interaction test is needed to assess whether the treatment effect differs statistically between subgroups.

Hazard Ratio vs. Odds Ratio

Hazard ratios are also frequently confused with odds ratios.

An odds ratio compares odds:

It is not the same as a hazard ratio.

Common Hazard-Ratio Interpretation Errors

1. Calling the HR a risk ratio

An HR of 0.70 does not automatically mean that 30% of patients were prevented from having the event.

2. Ignoring the reference group

The same HR can have the opposite clinical meaning if treatment and control are reversed.

3. Ignoring the event definition

An HR below 1 is favorable for death but may be unfavorable for an event such as recovery.

4. Treating HR = 0.50 as "twice the survival"

A hazard ratio does not imply that survival time doubles.

5. Reporting only the point estimate

The confidence interval communicates statistical uncertainty and should usually accompany the HR.

6. Assuming proportional hazards without checking

A single HR may be misleading when treatment effects change substantially over time.

7. Ignoring absolute benefit

A relative hazard reduction does not tell clinicians how many additional patients avoid the event.

8. Equating statistical significance with clinical importance

A small but highly precise effect may be statistically significant without being clinically meaningful.

A Clinician's Four-Step Interpretation

A Complete Clinical Example

Suppose a randomized oncology trial reports:

The first interpretation is:

The treatment is associated with an estimated 30% lower hazard of death relative to control.

The confidence interval ranges from 0.56 to 0.88.

In relative terms, the interval corresponds approximately to a hazard reduction between:

and:

or approximately 12% to 44% lower hazard.

The interval is entirely below 1, providing evidence of a difference.

The median survival estimates provide additional clinical context:

The observed median difference is 7 months.

But Is a 7-Month Median Difference Always the Best Summary?

Not necessarily.

Median survival can be unstable or impossible to estimate when the survival curve does not fall below 50%.

Furthermore, two survival curves can have the same median but differ substantially at other clinically important time points.

Therefore, clinicians may also want:

• 1-year survival
• 2-year survival
• 3-year survival
• 5-year survival
• Restricted mean survival time
• Absolute survival differences

The appropriate time points depend on the disease and clinical question.

Milestone Survival

Suppose:

These milestone estimates can be much easier for clinicians and patients to understand than a hazard ratio alone.

What If the Hazard Ratio Is Exactly 1?

An HR of:

means that the estimated hazards are equal under the model.

It does not mean that every patient has the same outcome.

It also does not prove that the treatments are equivalent.

The confidence interval is important.

For example:

This indicates considerable uncertainty around the treatment effect.

Hazard Ratio and Equivalence

A nonsignificant hazard ratio does not establish equivalence or noninferiority.

Equivalence and noninferiority require prespecified margins and appropriate statistical designs.

For example, if a noninferiority margin is specified on the hazard-ratio scale, the interpretation depends on whether the confidence interval remains within that margin.

Hazard Ratio in Noninferiority Trials

Suppose a noninferiority trial defines an upper HR margin of 1.30.

If the estimated HR is:

the upper confidence limit of 1.24 remains below the prespecified margin of 1.30.

Under the specified analysis and assumptions, this can support a noninferiority conclusion.

The precise interpretation depends on the direction of the endpoint, the analysis population, and the prespecified noninferiority framework.

Why the Hazard Ratio Is Log-Transformed

Cox regression estimates coefficients on the log-hazard scale.

If:

then:

This transformation ensures that the hazard ratio is positive.

For example:

gives:

Why HRs Are Multiplicative

Suppose two independent covariates have hazard ratios:

In a standard multiplicative Cox model, the combined relative hazard associated with both covariate changes is:

This illustrates why Cox-model coefficients are additive on the log-hazard scale but multiplicative on the hazard-ratio scale.

Interaction Terms

The simple multiplication above changes when the model contains interactions.

For example:

The effect of \(X_1\) then depends on the value of \(X_2\).

Therefore, clinicians should not interpret a main-effect HR in isolation when a treatment-by-covariate interaction is included.

Time-Varying Treatment Effects

When the treatment effect changes over time, the model may include a time-dependent treatment effect.

Conceptually:

The treatment effect can therefore be different during different periods of follow-up.

In such settings, reporting only one overall HR can hide important clinical information.

Alternatives When Proportional Hazards Is Questionable

Several approaches can complement or replace a single proportional-hazards summary.

• Time-specific survival probabilities
• Restricted mean survival time (RMST)
• Milestone survival analysis
• Time-varying hazard ratios
• Piecewise hazard models
• Weighted log-rank procedures

Restricted Mean Survival Time

Restricted mean survival time summarizes the average event-free survival up to a prespecified time horizon \(\tau\).

It can be defined as:

The treatment effect can then be summarized as:

For example, if:

library(survival)

fit <- coxph(
  Surv(time, event) ~ treatment,
  data = trial_data
)

summary(fit)

coef = -0.357

exp(-0.357)
# approximately 0.70

exp(coef(fit))

exp(confint(fit))

cox.zph(fit)

plot(cox.zph(fit))