A hazard ratio compares the rate at which events (like death or disease progression) happen in one group versus another, and the confidence interval around it tells you how precise that estimate is. Together, they form the backbone of how clinical trials report whether a treatment works. But reading them correctly requires more than glancing at a single number, because the most common mistakes people make when interpreting these statistics can flip their meaning entirely.
What a Hazard Ratio Actually Tells You
The word “hazard” in statistics refers to the instantaneous rate of an event happening at any given moment for someone who has survived up to that point. It is not the same as overall risk or total probability. The hazard ratio is the ratio of that instantaneous event rate in one group compared to another.1Journal of Investigative Dermatology. Research Techniques Made Simple: Interpreting Measures of Association in Clinical Research If you see an HR of 0.75 for a new drug versus a placebo in a cancer trial, it means that at any given point during the study, people taking the drug were experiencing the event (say, death) at 75% of the rate of the placebo group. In plain terms, the drug slowed things down by about 25%.
Three reference points make interpretation straightforward. An HR of 1.0 means there is no difference between groups: both experience the event at the same rate. An HR below 1.0 favors the treatment group, meaning the event happens more slowly for them. An HR above 1.0 means the treatment group fared worse, experiencing the event faster than the comparison group. The further the number is from 1.0 in either direction, the larger the difference between the two groups.
One crucial nuance: the hazard ratio describes speed, not destination. An HR of 0.70 does not mean 30% fewer people will ultimately experience the event. It means the event is happening 30% more slowly in the treatment group at any given moment during the observation period. Over a long enough time frame, both groups might eventually reach similar total event counts. This distinction between rate and cumulative probability is where most casual misreadings begin.
What the Confidence Interval Adds
A hazard ratio by itself is a point estimate, a single best guess derived from the study data. The 95% confidence interval surrounding it gives you a range that, based on the study’s methods, is expected to contain the true value 95% of the time if the study were repeated many times under the same conditions. A trial reporting HR 0.80 with a 95% CI of 0.65 to 0.98 is telling you the best estimate is a 20% reduction in the event rate, but the true effect could plausibly be as large as a 35% reduction or as small as a 2% reduction.
The key diagnostic question is whether the confidence interval crosses 1.0. If the entire interval sits below 1.0, the result is statistically significant in favor of the treatment. If it crosses 1.0, the data are compatible with the treatment having no effect at all, and the result is not statistically significant at the 95% level. A confidence interval of 0.72 to 1.15 around an HR of 0.91 looks like the treatment might help, but the interval includes 1.0, so the study cannot rule out that the treatment does nothing.
Width matters too. A narrow interval like 0.78 to 0.82 signals a precise estimate, usually from a large study with many events. A wide interval like 0.45 to 1.30 signals substantial uncertainty, often from a small trial or one with few observed events. Two studies can report the same HR of 0.75 but mean very different things if one has a tight interval and the other has a wide one.
There is an occasional technical wrinkle worth knowing about. The standard method for building a confidence interval around a hazard ratio can sometimes disagree with the log-rank test, which is the other common way of testing whether two survival curves differ. A study might report a statistically significant log-rank p-value while the confidence interval for the hazard ratio still includes 1.0.2PubMed Central. On Confidence Intervals for the Hazard Ratio in Randomized Clinical Trials When this happens, it creates genuine ambiguity about whether the trial should be considered positive or negative. These discrepancies are uncommon but not rare, and they tend to arise in trials where the treatment effect is borderline.
Why a Hazard Ratio Is Not the Same as a Risk Reduction
This is the single most consequential misinterpretation in how trial results reach the public. When a study reports an HR of 0.70, it is tempting to say the treatment “reduces the risk of death by 30%.” That phrasing conflates two different things. An HR reduction of 30% means the rate of events slowed by 30%. A relative risk reduction of 30% would mean 30% fewer people experienced the event. These are not the same, because hazard ratios describe the speed at which events occur, not the binary chance of the event happening at all.3Qeios. Risk of death derived from Hazard Ratio should not be communicated as Relative Risk reductions for death in cancer clinical trials. Intentional or inadvertent?
The practical problem is that press releases, drug advertisements, and even some medical journal discussions routinely present HR reductions as though they were relative risk reductions. A patient hearing “30% reduction in the risk of death” imagines that out of every 100 people, 30 fewer will die. But a 30% HR reduction might translate to only a modest difference in actual survival rates by the end of the trial. The confusion can inflate expectations about what a treatment will accomplish for any individual patient.
To get a clearer picture, you need the absolute numbers. If a control group had a 10% death rate over five years and the treatment group had a 7% death rate, the absolute risk reduction is 3 percentage points, regardless of what the hazard ratio says. Reporting guidelines from major oncology journals explicitly call for including absolute differences alongside hazard ratios for exactly this reason.4Journal of Clinical Oncology. Statistical Guidelines When you are reading a study and the only number you see is an HR with no absolute survival figures, treat that as an incomplete picture.
The Proportional Hazards Assumption
The hazard ratio as most studies report it comes from a statistical model that assumes the ratio between the two groups stays constant over time. If the treatment group has half the event rate of the control group at month six, the model assumes it still has half the rate at month 24 and month 60. This is called the proportional hazards assumption, and when it holds, the HR is a clean, interpretable summary of the treatment effect across the entire study.
The problem is that it frequently does not hold, and researchers do not always check. A review of over 300 studies in joint-replacement research found that only about 36% even mentioned the proportional hazards assumption. Among studies that plotted survival curves, roughly 45% showed the curves crossing at some point, a visual red flag that the hazard ratio between groups was not constant over time.5BMC Musculoskeletal Disorders. Testing the proportional hazards assumption in cox regression and dealing with possible non-proportionality in total joint arthroplasty research: methodological perspectives and review Only about 20% of the reviewed studies met all the criteria for properly testing the assumption, reporting the result of the test, and correcting the model if needed.
When the assumption fails, a single hazard ratio can be misleading or even meaningless. Imagine a new cancer drug that performs worse than standard treatment for the first few months (because of side effects) but then dramatically outperforms it after six months. A single HR averaged over the full study period might come out near 1.0, suggesting no benefit, even though the drug clearly helped people who survived the early phase. Alternatively, it could be pulled toward a favorable value by the late benefit while hiding the early harm.
This is especially relevant in immunotherapy trials, where delayed treatment effects are common. Checkpoint inhibitor drugs often show little or no benefit in the first weeks or months before the immune response kicks in, which violates the proportional hazards assumption. Researchers have argued that alternative measures like restricted mean survival time should be reported in such trials instead of relying solely on the hazard ratio.6Annals of Oncology. Treatment effects measured by restricted mean survival time in trials of immune checkpoint inhibitors for cancer Restricted mean survival time sidesteps the proportional hazards assumption entirely by asking a simpler question: on average, how much longer did patients in the treatment group survive up to a specific time point?
The Subgroup Analysis Trap
Clinical trials often break their results down by subgroups: men versus women, older versus younger participants, patients with a specific genetic marker versus those without. You will commonly see forest plots with a hazard ratio and confidence interval for each subgroup. A natural instinct is to look at which subgroups have a statistically significant HR and conclude the treatment works for those groups but not the others. That instinct is wrong.
The correct way to determine whether a treatment truly works differently in two subgroups is to test the interaction between the subgroup and the treatment effect. Simply observing that the HR is significant in men but not in women does not mean the treatment works differently by sex. The confidence intervals for the two subgroups might overlap substantially. Only the interaction test tells you whether the difference between the two hazard ratios is itself statistically significant.7Journal of Thoracic Oncology. Biostatistics Primer: What a Clinician Ought to Know: Subgroup Analyses
This mistake appears constantly, even in high-profile trials. A subgroup with a smaller sample size will naturally have a wider confidence interval, making it easier for the interval to cross 1.0 and look “non-significant,” even if the treatment effect is the same size as in a larger subgroup. When you encounter a forest plot in a trial publication, look for the interaction p-value. If it is not reported, be skeptical of any claim that the treatment helps one group more than another.
Immortal Time Bias Can Fake a Treatment Effect
When hazard ratios come from observational studies rather than randomized trials, a specific type of design flaw can produce results that look real but are artifacts. Immortal time bias occurs when the period between the start of observation and when a person receives the treatment is counted as “time on treatment” even though, by definition, the person could not have experienced the outcome during that window. Patients who receive the treatment had to survive long enough to get it, while patients in the comparison group could have died at any time.8PubMed Central. Immortal Time Bias in Cohort Studies: A Concept Simply Explained
The effect on the hazard ratio can be dramatic. In one analysis of postmastectomy radiation therapy using a national cancer database, a standard analysis showed a statistically significant survival benefit for radiation, with an HR of 0.93 (95% CI 0.88 to 0.99). But when the analysis was corrected for immortal time bias using methods that properly account for the timing of treatment, the apparent benefit vanished: the HR shifted to 0.97 (95% CI 0.92 to 1.03), which was no longer statistically significant.9PubMed Central. Immortal Time Bias in Observational Studies of Time-to-Event Outcomes: Assessing Effects of Postmastectomy Radiation Therapy Using the National Cancer Database A treatment that looked beneficial was, after proper analysis, indistinguishable from no effect.
A broader analysis of studies at high risk for this bias found that about a quarter showed statistically significant differences between the original and bias-corrected estimates, and another quarter reached different conclusions about whether the treatment effect was significant at all.10PubMed Central. Identifying, handling and impact of immortal time bias on addressing treatment effects in observational studies using routinely collected data When you encounter an observational study reporting a favorable hazard ratio for a treatment, check whether the authors used landmark analysis or time-dependent models. If they used a standard approach without discussing immortal time, the result deserves extra skepticism.
When the Hazard Ratio Does Not Match the Median Survival Ratio
Another common shortcut is to assume the hazard ratio corresponds neatly to the ratio of median survival times between two groups. If the control group’s median survival is 12 months and the treatment group’s is 16 months, the ratio is 0.75 (12 divided by 16). People sometimes treat this as interchangeable with an HR of 0.75. It is not.
The hazard ratio summarizes the entire time-to-event curve, while median survival is a single point on that curve. When the proportional hazards assumption holds and both survival curves are well-behaved, the two numbers can be close. But research comparing them directly has found the discordance is too large to consider them interchangeable.11PubMed. A hazard ratio was estimated by a ratio of median survival times, but with considerable uncertainty The gap becomes especially problematic when the proportional hazards assumption fails, in which case the hazard ratio from a Cox regression model and the median survival time ratio can diverge substantially and point to different conclusions about the treatment’s value.12PubMed. The relationship between hazard ratio and median survival time
This matters in practice because news coverage of cancer trials frequently leads with the median survival improvement in months, and then cites the HR alongside it as though the two figures are simply different ways of saying the same thing. They can diverge enough to give you a meaningfully different impression of how much the treatment helped.
Clinical Importance Versus Statistical Significance
A confidence interval that excludes 1.0 tells you the result is statistically significant. It does not tell you the result matters clinically. A massive trial with tens of thousands of participants can detect an HR of 0.97 as statistically significant, meaning the confidence interval might be 0.95 to 0.99, entirely below 1.0. But a 3% reduction in the event rate is unlikely to change how anyone practices medicine. The confidence interval is answering a narrow question: is the effect distinguishable from zero? It is not answering the question patients and doctors actually care about, which is: is the effect large enough to matter?13PubMed Central. Using the confidence interval confidently
This is where the width and position of the confidence interval become more informative than the p-value. An HR of 0.80 with a CI of 0.65 to 0.98 is significant and the lower bound suggests the effect could be quite meaningful. An HR of 0.97 with a CI of 0.95 to 0.99 is also significant, but even the most optimistic end of the interval represents a tiny difference. Looking at where the confidence interval falls relative to a pre-specified threshold of clinical meaningfulness (not just relative to 1.0) gives you a more honest read of the data. Some trial designs, particularly equivalence and non-inferiority studies, are built around exactly this idea: defining in advance what size of difference would actually matter and then checking whether the confidence interval falls within that window.
How Hazard Ratios Should Be Reported
Major medical journals have guidelines for how hazard ratios and their confidence intervals should appear in a paper. Hazard ratios are typically reported to two decimal places, and authors are expected to include a 95% confidence interval around the estimates related to their main research questions.14PubMed Central. Guidelines for reporting of statistics for clinical research in urology More than that, leading oncology journals require that hazard ratios be accompanied by absolute differences, such as the difference in survival rates at a specific time point.4Journal of Clinical Oncology. Statistical Guidelines
When you are reading a trial report, look for three things together: the hazard ratio, its 95% confidence interval, and an absolute measure of the treatment effect. If the paper provides only the HR and a p-value without a confidence interval, you are missing the information that tells you how uncertain the estimate is. If it provides the HR and CI but no absolute risk difference, you are missing the information that tells you how much the treatment actually moved outcomes in real terms. A well-reported trial gives you all three.
Bayesian Credible Intervals and How They Differ
Most hazard ratios and confidence intervals you encounter in medical literature use the frequentist statistical framework. But a growing number of trials, particularly in hematology and oncology, are incorporating Bayesian methods. In a Bayesian analysis, you do not get a confidence interval; you get a credible interval, and the interpretation is subtly but meaningfully different.
A 95% frequentist confidence interval says: if we repeated this study many times, 95% of the resulting intervals would contain the true value. A 95% Bayesian credible interval says something more intuitive: given the data we observed and our prior assumptions, there is a 95% probability the true value falls in this range. For most practical purposes, the two intervals can look similar, but the Bayesian version also allows researchers to make probability statements directly about the treatment effect, such as estimating the probability that the hazard ratio is below a specific threshold.15Blood Cancer Journal. Bayesian interim analysis for prospective randomized studies: reanalysis of the acute myeloid leukemia HOVON 132 clinical trial
For a reader trying to understand a clinical trial result, the most useful takeaway is that a Bayesian credible interval can be read more directly as “there is a 95% chance the true answer is between these values,” whereas the frequentist confidence interval is a statement about the long-run behavior of the method, not about the specific interval in front of you. If you encounter a trial reporting Bayesian credible intervals around a hazard ratio, the number reads the same way (below 1.0 favors treatment, above 1.0 favors control), but the probability interpretation attached to the interval is more straightforward.
Hazard Ratios Beyond Clinical Trials
Although most people encounter hazard ratios in the context of medical research, the underlying method of survival analysis applies wherever researchers track time until an event occurs. Demographers use it to study the time between births or changes in marital status. Labor economists use it to analyze how long people remain unemployed. Environmental scientists apply it to the duration between pollution episodes.16Demographic Applications OF Event History Analysis. Incomplete Data in Event History Analysis In all these settings, the hazard ratio means the same thing: the rate of the event in one group relative to another, at any given moment in time. The confidence interval interpretation is identical as well.
Understanding the logic in any one of these fields transfers directly to the others. If you can read a cancer trial’s HR of 0.70 (95% CI 0.58 to 0.85) and understand that the treatment group experienced the event about 30% more slowly, with enough precision to be confident the effect is real, you can also read that an HR of 1.40 (95% CI 1.15 to 1.70) for re-employment in a labor study means one group found jobs about 40% faster than the other, with solid statistical confidence. The vocabulary differs across disciplines, but the math and the interpretation do not.