The risk ratio (RR) is a measure of association that divides the probability of an event in one group by the probability of the same event in a reference group. Often called relative risk, it is widely used in epidemiology, clinical research, and statistics to compare outcomes between exposed and unexposed populations or between intervention and control groups. A risk ratio concerns a defined event over a specified period; it does not, by itself, establish that the group difference is causal.
Definition and interpretation
Let indicate that an event occurs during a specified observation period, and let and identify the comparison and reference groups. Define
These are conditional probabilities. When , the risk ratio is
Its interpretation depends on which group is placed in the numerator:
- : the two groups have equal risks.
- : the event is more likely in the comparison group.
- : the event is less likely in the comparison group.
For example, means that the comparison-group risk is 1.5 times the reference-group risk, or 50% higher. It does not mean an increase of 50 percentage points. An of indicates a 30% lower risk relative to the reference group.
The word risk here means event probability, not necessarily the probability of harm. The event can be recovery, successful completion of a task, or any other binary outcome. Consequently, a ratio below 1 is not inherently favorable.
When both risks are positive, exchanging the comparison and reference groups replaces with . Replacing the event with its complement generally does not:
Thus, a risk ratio for failure cannot generally be interpreted as the reciprocal of the risk ratio for success.
Calculation from observed data
For two groups with observed binary outcomes, the data can be arranged as follows:
| Group | Event | No event | Total |
|---|---|---|---|
| Comparison | |||
| Reference |
The observed risks and estimated risk ratio are
Suppose an event occurs in 30 of 200 comparison-group participants and 20 of 200 reference-group participants. Then
The corresponding risk difference is
or 5 percentage points.
The risk ratio is dimensionless. It can range from zero upward, but for a fixed positive reference risk its maximum is , because . If , the ordinary ratio is undefined; a positive numerator may be described as producing an infinite ratio in an extended sense. If both risks are zero, their ratio is indeterminate.
Relative and absolute effects
As a relative effect size, the risk ratio describes a proportional contrast rather than an absolute difference. The relationship is
Consequently, the same risk ratio can correspond to very different absolute differences. An of 2 could describe risks of 1% and 2%, or risks of 20% and 40%. The absolute increases are 1 and 20 percentage points, respectively.
When , the quantity
is often called the relative risk reduction, particularly when the outcome is undesirable. For example, corresponds to a relative risk reduction of 25%. This quantity remains a relative comparison; it does not specify how many events differ between groups without information about baseline risk.
Statistical uncertainty
An estimated risk ratio varies across samples. For independent groups with binomial outcomes and sufficiently large event counts, inference often uses the logarithm of the estimated ratio. Under this model, an approximate standard error is
An approximate 95% confidence interval is
Here, denotes the natural logarithm, and the approximation uses the normal distribution for the log estimate. Transforming back by exponentiation produces an interval that is generally asymmetric on the original ratio scale.
The formula is not directly usable when either observed event count is zero, and its approximation can be poor with sparse data. Likelihood-based, score-based, exact, and Bayesian procedures provide alternative approaches. Adding a fixed quantity to cells—a continuity correction—changes the estimate and can be influential when counts are small.
The simple formula also assumes independent observations. Clustered, matched, or repeated observations require methods that account for their dependence.
For hypothesis testing, the usual null value is . Whether an interval excludes 1 does not determine the substantive importance of an association, nor does an interval quantify uncertainty arising from systematic errors such as confounding or selection bias.
Distinction from related measures
Odds ratio
The odds ratio compares odds rather than probabilities:
When both event probabilities are small, odds and risks are numerically similar, so the odds ratio approximates the risk ratio. When events are common, they can differ substantially. If the risks are 60% and 30%, the risk ratio is 2, but the odds ratio is 3.5.
For unequal risks strictly between zero and one, the odds ratio lies farther from 1 than the corresponding risk ratio. This is a mathematical property of the two measures, not a statement that an odds ratio is inherently an incorrect measure.
Ordinary case-control sampling fixes or controls the numbers of cases and noncases, so sampled event proportions do not generally estimate population risks directly. Such data can often identify an odds ratio, but a risk ratio generally requires additional information or assumptions.
Rate ratio
A rate ratio compares event rates, often expressed per unit of person-time. A risk ratio compares probabilities over a specified period. Rates and risks are related but not interchangeable.
Even with constant event rates and , and no competing events, the risks by time are
Their ratio generally differs from , although the quantities can be close when cumulative risks are small.
Hazard ratio
A hazard ratio compares instantaneous event hazards among individuals who have not yet experienced the event. It is commonly used in survival analysis and is not generally equal to a risk ratio at any particular follow-up time.
Under proportional hazards, if the hazard ratio is and the reference-group survival probability is , then
Thus, even a constant hazard ratio can correspond to risk ratios that change over time.
Study design and causal interpretation
Risk ratios can be estimated directly from cohort data and from a randomized controlled trial when group-specific event risks are identifiable. When follow-up is incomplete, simple event counts divided by initial group sizes may not estimate risks adequately; time-to-event methods may be needed.
An observed association is distinct from a causal effect. Confounding, selection bias, measurement error, and differential loss to follow-up can affect a comparison.
In causal inference, a causal risk ratio can be defined using potential outcomes:
where and represent the outcome under two specified interventions or exposure conditions in the same target population.
Identifying this quantity from observed data requires assumptions, commonly including consistency, exchangeability, and positivity. Exchangeability means that the compared groups are sufficiently comparable with respect to their potential outcomes, possibly after conditioning on measured variables. Positivity requires that the relevant exposure conditions occur within the population subgroups needed for the comparison. Random assignment supports exchangeability for assigned treatment, but nonadherence, missing outcomes, and other design features still affect which causal effect is estimated.
Adjusted estimation and synthesis
Adjusted risk ratios can be estimated by stratification, standardization, weighting, or regression.
A log-binomial model is a generalized linear model that relates event probability to explanatory variables through
Without interactions involving , represents a common conditional risk ratio. Because fitted probabilities must remain at most one, the model has parameter constraints and can encounter convergence difficulties.
Another approach, often called modified Poisson regression, fits a Poisson model with a log link to binary outcomes and uses a robust variance estimator. Its uncertainty calculation accounts for the fact that binary outcome variance does not follow the ordinary Poisson variance model.
Logistic regression directly models odds ratios rather than risk ratios. It can nevertheless support estimation of marginal risks by predicting outcomes under each exposure condition and averaging over a target population. The ratio of those standardized risks is a marginal risk ratio, which need not equal a conditional regression contrast.
In meta-analysis, study-specific risk ratios are often combined on the logarithmic scale. Interpretation depends on the comparability of outcome definitions, follow-up periods, study populations, and the underlying effects being estimated. A pooled relative effect does not by itself determine an absolute effect for a population with a different baseline risk.