Understanding Hazard Ratios, and How They Differ From Risk Ratios

A hazard ratio compares how fast an event is arriving in one group against another, moment by moment, across the whole stretch of follow-up. It is a ratio of rates, not a ratio of plain chances, and that one distinction explains most of what people get wrong when they read one. A hazard ratio of 0.7 does not mean 30 percent fewer people had the event. It means that at any given instant while the study was running, the treated group was reaching the event at about 70 percent of the speed of the comparison group. Timing is built into the number, which is why a hazard ratio answers a different question than a risk ratio does. This is general education for reading evidence, not medical advice, and any decision about your own care belongs in a conversation with a qualified clinician.

Why trust this walk-through? Dr. Damon Tojjar is a physician-scientist (M.D., Lund University) whose peer-reviewed research reports effect sizes and pools them across studies, including a meta-analysis of ethnic differences in insulin sensitivity and response published in Diabetes Care. Pooling forces precision about which measure is on the table, because a hazard ratio and a risk ratio cannot be averaged as if they were one thing.

What a hazard actually is

Start with the word hazard, because it is doing quiet work. A hazard is an instantaneous rate, the chance that an event happens in the next sliver of time given that it has not happened yet. It is a speed, not a count. Picture the reading on a speedometer rather than the total distance on the odometer.

That conditional clause, given that it has not happened yet, sits at the center of the idea. At every point in a study, the people still at risk are only those who have not yet had the event. The hazard describes how quickly events are landing among exactly those people, right then.

A hazard ratio takes that rate in one group and divides it by the rate in another. At one, the two groups accumulate events at the same speed. Above one, the first group is faster. Below one, it is slower. The number compares velocities.

Why a hazard ratio is not a risk ratio

A risk ratio compares plain chances measured over a fixed window. By the end of two years, what fraction of each group had the event, and how do those fractions compare? It is a snapshot taken at one chosen moment, and it ignores when within the window each event landed.

A hazard ratio refuses to ignore the timing. It uses every event and the exact time it occurred, along with how long event-free people were followed before the study lost track of them. Survival methods can therefore handle people who drop out partway through or who simply have not had the event when the study ends, a situation called censoring that a simple risk ratio handles poorly.

The practical gap shows up in interpretation. Two treatments can post the same two-year risk ratio while having very different hazard ratios, because one front-loads its events early and the other spreads them late. The risk ratio sees only the totals at the finish line. The hazard ratio watches the whole race.

The proportional-hazards assumption, and why it matters

Here is the assumption hiding inside almost every hazard ratio you will read. The standard model assumes the ratio of the two hazards stays constant for the entire follow-up. Early, middle, and late, the treated group is supposed to run at the same fraction of the comparison group's speed. That single constant is what the reported hazard ratio represents.

When the assumption holds, one number fairly summarizes the whole study. When it fails, the single number becomes an average of a relationship that was actually changing, and an average can describe a curve that no real patient ever experienced.

Failure is common, and often clinically interesting. An intervention might do nothing for a year and then pull ahead, so the early hazard ratio sits near one while the late ratio drops well below it. Two survival curves that cross are the loudest signal that the assumption has broken, because crossing means the advantage reversed direction partway through.

How to tell whether the assumption was respected

You do not need to refit the model to sense trouble. The first move is to look at the Kaplan Meier curves, the step plots of event-free survival in each group. If they separate steadily and never touch after splitting, a constant ratio is plausible. If they cross, converge, or fan apart unevenly, treat the single hazard ratio with caution.

A careful report will say it checked. Good papers test the proportional-hazards assumption directly and mention the result. Silence on this point does not prove a problem, yet it is a gap worth noting, especially for a study whose whole headline rides on one hazard ratio.

The honest fix, when the assumption fails, is to report the effect over separate time periods rather than collapsing it into one figure. A treatment that helps only after the first year is a real and useful finding. Flatten it into a single averaged hazard ratio and you bury the very pattern a reader most needs to see.

Reading a hazard ratio without overinterpreting it

A few habits keep the number in its lane. Resist translating the ratio into a percentage of people, because it describes a rate, not a head count, and the share affected depends on how long the study ran and how common the event was. A hazard ratio of 0.5 over a short study with rare events may correspond to a very small absolute difference in people.

Pair the ratio with its confidence interval and with the absolute numbers underneath. Ask how many events actually occurred, since a hazard ratio built on a handful of events is fragile no matter how clean it looks. Ask what the plain survival figures were in each group.

The model to keep is compact. A hazard ratio compares the speed of events between groups across time. It assumes that speed ratio held steady unless the study shows otherwise. It tells you about rates, not about how many people were touched. Hold those three ideas together, and a survival result becomes something you can weigh.