Topic 1.4 — Hazard Ratios

In one sentence: A hazard ratio is a compact, widely reported, and widely misread summary of how fast events happen in one group compared with another — and reading it well means always asking "over what follow-up, in which population, and what does it buy in absolute terms?" before you let it drive a commissioning, evaluation, or reporting decision.

Why this matters in health and care#

When a vendor's slide says "our intervention achieved a hazard ratio of 0.75," a leader is being invited to feel a 25% reduction in something bad. The feeling is usually stronger than the fact warrants. A hazard ratio (HR) is one of the most common statistics in clinical and public-health evidence, and one of the most reliably misunderstood — not by statisticians, but by the boards, commissioners, and communications teams who act on it. In health and care, that gap between what an HR says and what people think it says is not an academic nuisance. It changes what gets bought, what gets scaled, and what patients are told to expect.

This topic is a practical companion to Topic 1.3 — Evidence-Driven Decisions. That topic argued you should match the strength of evidence to the level of risk and keep clinical and delivery evidence distinct. This one goes one layer deeper into a single statistic that recurs in the clinical evidence you will be asked to weigh: how to read it, where it misleads, and what to demand alongside it. You do not need to compute a hazard ratio. You do need to know that the same headline HR can correspond to a life-changing benefit or a barely detectable one depending on facts the slide left out — and that a leader who cannot ask for those facts is at the mercy of whoever chose the framing. Because these numbers shape public messaging and clinical pathways, getting them wrong is a matter of trust and equity, not merely of accuracy.

Core concepts#

The hazard, and the hazard ratio. A hazard is the instantaneous rate at which an event (death, hospital admission, relapse) occurs among those still at risk at a given moment. A hazard ratio is the ratio of that rate in a treatment group to the rate in a control group. An HR below 1 means the event is happening more slowly in the treatment group; above 1, faster; exactly 1, no difference. It emerges from survival analysis — the branch of statistics for "time to event" data — and is usually estimated with the proportional hazards model, known after its originator as the Cox model. Crucially, an HR describes a rate, not a probability, and not a quantity of time gained.

The proportional-hazards assumption. The Cox model's convenience rests on an assumption: that the hazard ratio is constant — the same at three months as at three years, the same for a 50-year-old as for an 80-year-old. This is the proportional-hazards assumption, and it is frequently false. Real effects vary over follow-up time (a surgical intervention may raise risk early and lower it later) and across subgroups (age, sex, comorbidity, deprivation). When the assumption fails, the single reported HR is an average over the trial's particular follow-up window and population mix — a weighted blend, not a universal constant you can lift into a different setting.

Relatives, odds, and absolutes. An HR is easily confused with three neighbors. The relative risk (or risk ratio) compares the cumulative probability of an event by some fixed time, not the rate; it is close to the HR when events are rare but diverges when they are common. The odds ratio compares odds rather than probabilities and can look dramatic when the underlying risk is high. All three are relative measures. What patients and payers actually care about is absolute: the absolute risk reduction (risk difference) — how many fewer events per hundred people — and its reciprocal, the number needed to treat (NNT), the count of people you must treat to prevent one event. A relative measure without its absolute partner is half a sentence.

Censoring and the survival curve. Trials end; people move away, withdraw, or simply have not had the event when the data are locked. Censoring (statistics) is how survival analysis handles these partially observed lives, and it is a strength — but heavy or unbalanced censoring weakens what an HR can tell you. The honest visual companion to any HR is the Kaplan–Meier estimator curve, the stepped plot of the proportion event-free over time. Two curves that separate steadily tell a different story from two that cross or converge — yet all three can yield the same headline HR.

Confounding and non-collapsibility. An HR from an observational study can be distorted by confounding: some third factor drives both who gets the treatment and who has the event. Randomized trials guard against this; real-world comparisons do not, unless carefully adjusted. There is also a subtler trap specific to ratios like the HR, called non-collapsibility: even with no confounding, the ratio you get after adjusting for a risk factor can differ from the unadjusted one, purely as an artifact of how ratios behave. The lesson is the same — an HR is tied to the exact model and population that produced it, and does not travel as freely as its clean single number suggests.

What an HR buys in years. People translate "HR 0.75" into "I'll live longer," but the size of that gain depends entirely on baseline risk and how compressed a population's mortality already is. Think of it as loosening one chamber in a game of chance: reducing a rate buys a great deal of extra time when the baseline risk is high and spread out, and very little when most people already survive to old age and deaths are packed into a narrow late window. The same relative reduction converts into different amounts of life expectancy depending on that baseline — which is why an HR cannot be read off as a fixed number of added years without knowing the population it came from.

Statistical versus clinical significance. Finally, an HR can be statistically significant — its confidence interval excluding 1 — while the effect it points to is clinically trivial, especially in a very large trial. Significance answers "is this distinguishable from chance?"; it does not answer "is this worth changing a pathway for?" Both questions matter, and only the second should decide investment.

Best practices#

1. Always ask "over what follow-up, and in which population?" An HR is an average over a specific trial window and a specific enrolled group. Before you act on it, establish the follow-up length and the inclusion and exclusion criteria — trials that excluded the very old, the very young, or people with common comorbidities may not describe your population at all. Treat an HR without this context as an incomplete figure, not a headline.

2. Demand absolute numbers beside every relative one. For any HR or relative risk, ask for the absolute risk reduction and the number needed to treat in the relevant population. "HR 0.75" feels large; "one avoided event per 200 people treated over three years" tells you whether it is worth the cost, workflow change, and side effects. Relative and absolute measures should always appear together, and communications to the public should lead with the absolute.

3. Insist on seeing the Kaplan–Meier curves. A single HR compresses a whole story into one number and hides whether the effect is early, late, growing, shrinking, or reversing. Ask for the survival curves and look at their shape: do they separate cleanly, cross, or converge? Curves that cross are a visible sign the proportional-hazards assumption has failed and the single HR is misleading.

4. Do not extrapolate a short trial to a lifetime. A hazard ratio measured over three years does not license a claim about thirty. Effects can wane, reverse, or accumulate, and the proportional-hazards assumption almost never holds over a lifespan. Be especially wary of economic models and press lines that project a short-term HR forward to dramatic gains in life-years; ask what assumption is doing that work, as covered in Topic 6.1 — Health Economics.

5. Separate statistical from clinical significance. A confidence interval that excludes 1 tells you an effect is probably real; it says nothing about whether the effect is large enough to matter. In big trials, trivial differences become "significant." Judge the magnitude — the absolute benefit — against the burden and cost of the intervention, and let clinicians and patients define what counts as worthwhile.

6. Treat observational HRs with extra suspicion. Outside a randomized trial, an HR may reflect confounding — sicker or healthier people selecting into treatment — rather than a true effect. Ask what was adjusted for and whether unmeasured confounders are plausible. Real-world evidence is valuable (see Topic 8.6 — Research & Real-World Evaluation) but its hazard ratios need more scrutiny, not less.

7. Check whether the HR is reused out of context. An HR from one population, dose, or follow-up is routinely quoted as if universal. When a figure is cited, trace it to its source study and confirm the population and endpoint match your decision. Non-collapsibility and differing baselines mean a number that was true there may be quietly wrong here.

8. Report honestly, and lead with what people can act on. When your own organization publishes results, pair every HR with absolute figures and plain-language framing, state the follow-up and population, and avoid implying lifetime benefit from short-term data. This connects directly to how you choose measures in Topic 10.1 — Key Performance Indicators (KPIs): the metric you headline shapes what people believe.

Questions to discuss with your team#

1. When a hazard ratio lands on our board, do we know how to ask for the follow-up, the population, and the absolute benefit — or do we act on the relative figure alone? This question tests whether your leadership can interrogate the single most misread statistic they will encounter. It matters because an HR is an average over a particular trial window and a particular enrolled group, and the same figure can represent a transformative benefit or a negligible one depending on the baseline risk and follow-up that the slide omitted. The tension is that the relative number is emotionally compelling and easy to repeat, while the absolute number — the number needed to treat, the events avoided per hundred people — is duller and harder to obtain, so it is routinely dropped. Work through a real recent example: take an HR that influenced a decision and reconstruct what follow-up it covered, who was excluded, and what absolute benefit it implied for your population. Ask whether anyone requested those facts at the time, and whether the answer would have changed the decision. A good discussion ends with a standing expectation that no HR is acted on without its follow-up, population, and absolute partner alongside it.

2. Where might we be extrapolating a short-term hazard ratio into a long-term or lifetime claim — in a business case, a benefits model, or a public message? The topic warns that an HR measured over a few years does not license claims about decades, because effects wane, reverse, or accumulate and the proportional-hazards assumption rarely holds over a lifespan. This matters because health-economic models and communications teams both have strong incentives to project short-term relative reductions forward into headline gains in life-years or cost savings, and those projections can quietly become the justification for large investments. The trade-off is real: you must plan and model beyond the trial horizon, yet every year you project past the evidence is an assumption, not a finding. Surface a live business case or benefits realization model and ask what time horizon its HR-derived benefits assume, and what would happen to the case if the effect flattened after the trial period. Consider too whether any patient-facing message implies a lifetime of added years from short-term data. An honest answer names where extrapolation is doing heavy lifting and commits to stating the assumption openly rather than burying it in a spreadsheet.

3. Do we consistently separate "is this real?" from "is this big enough to matter?" — or does statistical significance alone drive our decisions? This question confronts the confusion between statistical and clinical significance, where a confidence interval excluding 1 is treated as a green light regardless of the effect's size. It matters because large trials and large real-world datasets make trivial differences statistically significant, and a digital-health intervention can post a "significant" hazard ratio while delivering a benefit too small to justify its cost, workflow disruption, or opportunity cost against other uses of the same money. The deeper tension is that magnitude is a clinical and value judgment — what counts as worthwhile depends on the condition, the burden of treatment, and patient preference — so it cannot be delegated to a p-value or automated. Take a decision your team made partly on the strength of a significant result and ask what the absolute effect size actually was, and who decided it was large enough to act on. Explore whether clinicians and, ideally, patients were involved in judging that the benefit was worth having. A strong answer shows the team routinely asks for magnitude, not just significance, and treats the two as separate gates that both must be passed.

4. When a hazard ratio comes from routine data or a vendor's real-world dataset rather than a randomized trial, do we ask what was adjusted for and whether unmeasured confounding could explain the effect? Increasingly the hazard ratios that reach a decision are not from clean trials but from observational sources — a supplier's before-and-after analysis, a registry, or your own routinely collected records — where the people who received an intervention differ systematically from those who did not. This matters because confounding by indication is the norm rather than the exception: healthier or sicker patients select into treatment, and unless that is carefully adjusted for, the resulting HR measures who was treated as much as what the treatment did. The tension is that real-world evidence is genuinely valuable and often the only evidence available for a digital service deployed at scale, yet its convenience tempts teams to read it as causal when it is merely associational. Interrogate a real observational HR your team has relied on: ask what variables were adjusted for, what plausible confounders were left unmeasured, and whether the direction of the likely bias flatters or deflates the effect. Consider whether non-collapsibility — the way adjusting for a risk factor can shift a ratio even absent confounding — has been accounted for in how the number is quoted. An honest answer treats observational hazard ratios as hypotheses needing more scrutiny, not less, and can state plainly what would have to be true for the effect to be real.

5. Do we routinely look at the Kaplan–Meier curves behind a hazard ratio, and would we recognize when the proportional-hazards assumption has failed? A single HR silently averages the whole shape of an effect into one number, concealing whether the benefit is early, late, growing, shrinking, or reversing over the follow-up. This matters because the Cox model that produces most hazard ratios assumes the ratio is constant over time, and that assumption is frequently false — a surgical or procedural intervention can raise risk early and lower it later, and curves that cross or converge can yield the same tidy headline HR as curves that separate cleanly. The trade-off is that reading survival curves takes more effort and more expertise than repeating a single figure, and the curves are often absent from the summary slide precisely because they complicate the story. Take a hazard ratio your team acted on and ask whether anyone saw the Kaplan–Meier plot, what the curves' shape implied about when the benefit accrued, and whether the proportional-hazards assumption was ever tested. Discuss who on the team has the statistical literacy to spot crossing or converging lines and to say what they mean for your population. A good answer makes seeing the curves a standing expectation, not a specialist favor, and treats a missing curve as a reason to ask rather than a detail to skip.

6. Have we checked whether a single hazard ratio hides an uneven distribution of benefit across age, comorbidity, and deprivation in the people we actually serve? One headline HR presents an effect as if it lands uniformly, but a benefit averaged across a trial's enrolled group can fall very differently across the older, sicker, and more deprived people who make up much of a real health and care population. This matters because acting on the aggregate figure can widen inequalities while appearing to improve outcomes overall: an intervention that helps a fitter subgroup and does little for a frailer one still posts an attractive average HR. The tension is that disaggregation demands data, statistical power, and a willingness to see uncomfortable variation, while the single number is simpler to report and easier to champion. Look at a decision driven by one hazard ratio and ask whether the effect was ever broken down by age band, comorbidity, or deprivation, and whether the trial even enrolled enough of your population to support such a breakdown. Consider how you would monitor for unequal benefit once a service is live, connecting to your real-world evaluation and equity duties. An honest answer accepts that a uniform-looking HR may mask an unequal reality, and commits to looking at the disaggregated, absolute picture before scaling.

In practice: a health & care example#

An integrated care system (ICS) is offered a remote-monitoring service for people with heart failure. The vendor's central claim is a hazard ratio of 0.72 for the composite outcome of death or unplanned admission, described as "a 28% reduction" and paired with a projection of substantial life-years gained across the population.

The digital and clinical leads treat the figure as a starting question, not a conclusion. They ask for the source study and find a single randomized trial with a median follow-up of eighteen months, enrolling patients aged under 75 with a recent admission and no significant cognitive impairment — a fitter, younger group than the ICS's real heart-failure population, which skews older and more comorbid. They ask for the Kaplan–Meier estimator curves and see the two lines separate mainly in the first six months, then run roughly parallel — a hint that the benefit is concentrated early and that the constant-HR assumption is shaky over a longer horizon.

Then they ask for absolutes. Translated into the trial population, the HR corresponds to a modest absolute risk reduction and a number needed to treat of around 20 over eighteen months to prevent one composite event — meaningful, but far less dramatic than "28% reduction" suggested, and quite possibly smaller still in their older population. The health-economics team is asked to rebuild its model without extrapolating the eighteen-month HR across a lifetime, and to test what happens if the early benefit does not persist. The clinical-safety and information-governance leads assess the service on its own merits (Topic 1.6 — Clinical Safety & Risk Management). The ICS proceeds — but as a monitored, staged rollout with its own real-world evaluation, and it briefs its communications team to lead with the absolute benefit and the honest limits, not the relative headline. When a local paper later asks about "the 28% figure," the ICS can explain exactly what it means and does not mean.

Sector lenses#

Everyone reading an HR owes the same duty of interpretation, but the pressures and safeguards differ by organization.

Startup#

A startup's incentive is to present the most flattering true number, and a relative hazard ratio is often the most flattering true number available — so "HR 0.7" appears on the pitch deck without the follow-up, population, or absolute benefit that would right-size it. The honest and ultimately more credible move is to lead with the absolute effect and the number needed to treat, state the trial's follow-up and exclusions plainly, and resist projecting short-term results across a lifetime in the business case. Diligence-minded NHS assessors and serious investors increasingly know to ask these questions, so a team that volunteers the context looks stronger, not weaker. The discipline costs almost nothing and protects the founders from a later, more damaging correction.

Small business#

An established small provider — a GP practice, a community pharmacy, a care home, or a single-site clinic weighing a monitoring tool or a supplier's claim — has real patients on the line but no in-house statistician and little time to read a trial. The practical risk is that the vendor's relative hazard ratio is taken at face value because there is no one to reconstruct the follow-up, population, and absolute benefit, and no analytics team to test it against the practice's own list. The proportionate response is not to master survival analysis but to ask three plain questions of any supplier — over what follow-up, in which patients, and how many people would we treat to prevent one event — and to lean on shared resources such as NICE guidance, a CCG or ICS pharmacist, or a peer network rather than judging the number alone. A small provider that insists on the absolute figure and the trial's exclusions before signing protects its patients and its budget without needing the statistical firepower of a large trust.

Enterprise#

An NHS trust or ICS has the analytics capability and the patient volumes to interrogate an HR properly — and the accountability that demands it, because deploying at scale on a misread statistic is a clinical-safety and public-money risk at once. The characteristic failure is institutional enthusiasm: a compelling relative figure championed by a respected clinician, adopted before anyone reconstructs the absolute benefit for the trust's actual, often older and sicker, population. The corrective is a standing habit — every HR in a board paper arrives with its follow-up, population, absolute risk reduction, and Kaplan–Meier curves, and every extrapolation names its assumption. Large enterprises also generate their own observational HRs from routine data, where confounding is the ever-present danger and adjustment must be explicit.

Government#

A national body — NICE, NHS England, the UK Health Security Agency, DHSC — often decides how evidence is framed for the whole system, so its handling of hazard ratios shapes public understanding at population scale. Decisions here attract regulatory, parliamentary, and media scrutiny, and a relative figure quoted without its absolute partner can mislead millions and erode trust when later corrected. The duty is to publish absolutes alongside relatives, state population and follow-up, and be candid about where the proportional-hazards assumption and short trial horizons limit what can be claimed. At this scale, honest framing is not a courtesy but an instrument of equity: a benefit that looks uniform in a single HR may fall unevenly across age and deprivation, and only the disaggregated, absolute picture reveals it.

Common failure modes#

• Relative-only reporting. Quoting "28% reduction" with no absolute risk reduction or number needed to treat. Antidote: pair every relative measure with its absolute partner and lead public messaging with the absolute.
• Lifetime extrapolation. Projecting a short-trial HR across decades in an economic model or press line. Antidote: model only to the evidence horizon, and name every assumption beyond it.
• Ignoring the population. Applying a trial's HR to older, sicker, or more deprived groups the trial excluded. Antidote: check inclusion and exclusion criteria against your real population.
• Assuming constancy. Treating one HR as valid at every age and time point. Antidote: inspect the Kaplan–Meier curves for crossing or converging lines.
• Significance mistaken for importance. Acting on a statistically significant HR whose effect is clinically trivial. Antidote: judge magnitude separately, with clinical and patient input.
• Trusting observational HRs uncritically. Reading a real-world HR as causal without asking about confounding. Antidote: ask what was adjusted for and whether unmeasured confounders are plausible.
• Number laundering. Reusing an HR out of its original context as if universal. Antidote: trace every cited figure to its source study and endpoint.

Maturity model#

Checklist#

• Every hazard ratio arrives with its follow-up length and the trial's inclusion and exclusion criteria.
• Each relative measure (HR, relative risk) is paired with absolute risk reduction and number needed to treat.
• Kaplan–Meier curves are reviewed for the shape of the effect and any crossing or convergence.
• The trial population is checked against your real, local population before acting.
• No short-term HR is extrapolated to a lifetime without an explicitly stated assumption.
• Statistical significance and clinical significance are judged as separate gates.
• Observational HRs are scrutinized for confounding and what was adjusted for.
• Every cited HR is traced to its source study, endpoint, and follow-up.
• Public and board communications lead with absolute benefit and honest limits.

Key sources#

• Wikipedia — Hazard ratio, Proportional hazards model, Survival analysis (foundational definitions).
• BMJ / The BMJ Statistics Notes series — Douglas Altman and J. Martin Bland — plain explanations of hazard ratios, survival curves, and absolute versus relative risk.
• David Spiegelhalter — The Art of Statistics: Learning from Data (Pelican, 2019) — on communicating relative and absolute risk honestly.
• Gerd Gigerenzer et al. — Helping Doctors and Patients Make Sense of Health Statistics (Psychological Science in the Public Interest, 2007) — on why relative figures mislead.
• Cochrane — Handbook for Systematic Reviews of Interventions — on interpreting effect measures.
• dynomight — essay on hazard ratios and life expectancy (dynomight.net) — the Russian-roulette intuition for why relative reductions do not equal life-years, and why interventions have different hazard ratios at different ages.
• Equator Network — CONSORT reporting guideline — on how trial effects, including time-to-event outcomes, should be reported.

References#

1. Hazard ratio — Wikipedia — https://en.wikipedia.org/wiki/Hazard_ratio
2. Proportional hazards model — Wikipedia — https://en.wikipedia.org/wiki/Proportional_hazards_model
3. Survival analysis — Wikipedia — https://en.wikipedia.org/wiki/Survival_analysis
4. Relative risk — Wikipedia — https://en.wikipedia.org/wiki/Relative_risk
5. Odds ratio — Wikipedia — https://en.wikipedia.org/wiki/Odds_ratio
6. Risk difference (absolute risk reduction) — Wikipedia — https://en.wikipedia.org/wiki/Risk_difference
7. Number needed to treat — Wikipedia — https://en.wikipedia.org/wiki/Number_needed_to_treat
8. Censoring (statistics) — Wikipedia — https://en.wikipedia.org/wiki/Censoring_(statistics)
9. Kaplan–Meier estimator — Wikipedia — https://en.wikipedia.org/wiki/Kaplan%E2%80%93Meier_estimator
10. Confounding — Wikipedia — https://en.wikipedia.org/wiki/Confounding
11. Life expectancy — Wikipedia — https://en.wikipedia.org/wiki/Life_expectancy
12. Statistical significance — Wikipedia — https://en.wikipedia.org/wiki/Statistical_significance
13. The Art of Statistics: Learning from Data — David Spiegelhalter, Pelican — https://www.penguin.co.uk/books/294857/the-art-of-statistics-by-spiegelhalter-david/9780241398630
14. Helping Doctors and Patients Make Sense of Health Statistics — Gerd Gigerenzer et al., Psychological Science in the Public Interest — https://doi.org/10.1111/j.1539-6053.2008.00033.x
15. Statistics Notes (survival and risk) — Douglas Altman & J. Martin Bland, The BMJ — https://www.bmj.com/specialties/statistics-notes
16. Cochrane Handbook for Systematic Reviews of Interventions — Cochrane — https://training.cochrane.org/handbook
17. Hazard ratios and life expectancy — dynomight — https://dynomight.net/
18. CONSORT reporting guideline — Equator Network — https://www.equator-network.org/reporting-guidelines/consort/