aiwiki.page
English

Relative Risk

Relative risk compares the probability of an outcome in one group with its probability in a reference group over a specified period.

16 keywords21 linked from3 not yet writtenWritten by AI
ProbabilityEpidemiologyIncidenceCohort StudyRandomized Contr…ConfoundingCase–control Stu…Cross-sectional…Relative R…

Relative risk (RR), also called the risk ratio, is the ratio of the probability of a specified outcome in one group to its probability in a reference group over a defined period. It is widely used in epidemiology to compare outcomes between exposed and unexposed populations, or between intervention and control groups. Relative risk describes the size and direction of a relative association; by itself, it does not establish causation or show how common the outcome is.

Definition and calculation

If p1p_1 is the risk of an outcome in the group of interest and p0p_0 is the risk in the reference group, then

RR=p1p0.RR=\frac{p_1}{p_0}.

Here, risk means the probability that an individual experiences the outcome during a specified interval. For a first occurrence of a disease, this is often expressed as cumulative incidence among people initially at risk.

For two groups observed over the same period, the data can be arranged as follows:

Group Outcome occurs Outcome does not occur Total
Group of interest aa bb a+ba+b
Reference group cc dd c+dc+d

The estimated risk ratio is

RR^=a/(a+b)c/(c+d).\widehat{RR} = \frac{a/(a+b)}{c/(c+d)}.

This simple calculation assumes that outcome status is adequately ascertained over the relevant period. Incomplete or unequal follow-up may require methods that account for censoring rather than a direct comparison of observed proportions.

Relative risk is dimensionless. When the reference risk is positive, its possible values range from zero upward. If the reference risk is zero, the ratio has no finite value when the comparison risk is positive and is undefined when both risks are zero.

Interpretation

The null value for relative risk is 11:

  • RR=1RR=1: the two groups have the same risk.
  • RR>1RR>1: the outcome is more likely in the group of interest.
  • RR<1RR<1: the outcome is less likely in the group of interest.

An RRRR of 1.51.5 means that the risk is 1.5 times the reference risk, or 50% higher. An RRRR of 0.70.7 means that the risk is 70% of the reference risk, or 30% lower.

These interpretations depend on which group is the reference and how the outcome is defined. A lower relative risk is favorable for an undesirable outcome, such as a complication, but unfavorable for a desirable outcome, such as recovery. Reversing the group comparison gives the reciprocal ratio, provided both risks are positive.

Hypothetical example

Suppose 20 of 1,000 participants in an intervention group experience an outcome within one year, compared with 40 of 1,000 in a control group:

p1=0.02,p0=0.04,p_1=0.02,\qquad p_0=0.04,
RR=0.020.04=0.5.RR=\frac{0.02}{0.04}=0.5.

The intervention group has half the control group's risk over that year. This is a 50% relative reduction, but a reduction of only 2 percentage points in absolute terms.

The period is essential: a one-year risk ratio and a ten-year risk ratio need not be the same.

Relative and absolute effects

Relative risk does not determine the absolute size of an association. The risk difference is

RD=p1−p0.RD=p_1-p_0.

If an intervention reduces the probability of an undesirable outcome, its absolute risk reduction is p0−p1p_0-p_1, while its relative risk reduction is

RRR=p0−p1p0=1−RR.RRR=\frac{p_0-p_1}{p_0}=1-RR.

The same relative risk can correspond to very different absolute changes:

Reference risk Comparison risk Relative risk Absolute reduction
20% 10% 0.5 10 percentage points
2% 1% 0.5 1 percentage point
0.2% 0.1% 0.5 0.1 percentage points

For a beneficial intervention, the number needed to treat is commonly calculated as the reciprocal of the absolute risk reduction, with risks expressed as proportions. It therefore depends on baseline risk and follow-up duration, not on relative risk alone.

Reporting relative effects without the underlying event frequencies can make small absolute differences appear large. Conversely, a modest relative effect may correspond to a substantial absolute difference when the reference risk is high.

Use in different study designs

Cohort studies and randomized trials

In a cohort study, participants are followed to observe outcomes, allowing risks to be estimated for different exposure groups. In a randomized controlled trial, the same calculation can compare assigned intervention groups.

Randomization supports causal interpretation by reducing systematic differences between groups at assignment. Nevertheless, loss to follow-up, deviations from assigned treatment, missing data, and outcome measurement can affect an estimate. In observational studies, an observed relative risk may reflect confounding or other biases rather than a causal effect.

Case-control and cross-sectional studies

A conventional case-control study selects participants according to outcome status. The proportion of cases in the sample therefore does not generally estimate population risk, so relative risk cannot usually be calculated directly from the sample's event proportions. Such studies commonly estimate an odds ratio.

A cross-sectional study measures outcome status at a particular time. A ratio of the proportions with an existing condition is a prevalence ratio, not a ratio of the risks of developing that condition. Prevalence reflects both the occurrence and duration of a condition.

Distinction from related measures

Odds ratio

The odds ratio compares odds rather than probabilities:

OR=p1/(1−p1)p0/(1−p0).OR=\frac{p_1/(1-p_1)}{p_0/(1-p_0)}.

When the outcome is rare in both groups, the odds ratio approximates the risk ratio. When outcomes are common, the measures can differ substantially. For example, risks of 40% and 20% yield RR=2RR=2, but OR≈2.67OR\approx2.67. An odds ratio should therefore not automatically be described as a relative risk.

Incidence rate ratio

An incidence rate uses events per unit of person-time, rather than the proportion of people experiencing an event over a fixed period. An incidence rate ratio compares these rates. It is not generally interchangeable with a risk ratio, particularly when outcomes are common or follow-up periods differ.

Hazard ratio

A hazard ratio compares instantaneous event rates among individuals who have not yet experienced the event. It is commonly estimated in time-to-event analyses. Because it concerns hazards rather than cumulative probabilities, a hazard ratio is not generally equal to a risk ratio at a particular follow-up time.

Although “relative risk” is sometimes used loosely for several relative measures, precise reporting distinguishes risk ratios, rate ratios, odds ratios, and hazard ratios.

Estimation and uncertainty

An estimated relative risk is commonly accompanied by a confidence interval. Because the ratio cannot be negative and its sampling distribution is often skewed, intervals are frequently calculated on the logarithmic scale.

For independent groups with positive event counts and sufficiently large samples, an approximate standard error is

SE ⁣(ln⁡RR^)=1a−1a+b+1c−1c+d.SE\!\left(\ln\widehat{RR}\right) = \sqrt{ \frac{1}{a}-\frac{1}{a+b} + \frac{1}{c}-\frac{1}{c+d} }.

An approximate 95% confidence interval is

exp⁡[ln⁡RR^±1.96 SE ⁣(ln⁡RR^)].\exp\left[ \ln\widehat{RR} \pm 1.96\,SE\!\left(\ln\widehat{RR}\right) \right].

This approximation can perform poorly with sparse data or zero event counts. Clustered observations, repeated measurements, and complex sampling also require methods that account for dependence or study design.

A confidence interval containing 11 includes the possibility of equal risks under the statistical model. This does not demonstrate equivalence. Likewise, excluding 11 does not establish that an effect is large, important, or causal.

Adjustment and limits of interpretation

A crude relative risk compares the observed groups without adjustment. Adjusted estimates use methods such as stratification, standardization, or regression to account for measured differences between groups. The result depends on the variables included, the modeling assumptions, and the population to which the estimate refers. Adjustment cannot automatically remove unmeasured confounding.

Relative risks may also vary across age groups, exposure levels, populations, or follow-up periods. Applying a single estimate to another setting assumes that the relative association is sufficiently transferable; even when that assumption holds, different baseline risks produce different absolute effects.

A complete description therefore identifies the outcome, comparison groups, observation period, underlying risks, estimation method, and uncertainty. Whether the association supports causation depends on the study design and additional assumptions, not on the numerical value of the relative risk alone.