Relative risk
Relative risk, also known as the risk ratio, is a statistical measure in epidemiology that compares the probability of a specific health event, such as disease onset or death, occurring in one group exposed to a risk factor versus a group not exposed to it.[1][2] It is calculated using the formula: relative risk = (incidence in exposed group) / (incidence in unexposed group), where incidence is typically expressed as the proportion of individuals experiencing the event within a defined period.[1][2] The value of relative risk indicates the strength and direction of the association between exposure and the outcome.[1] A relative risk greater than 1 suggests that the event is more likely in the exposed group, indicating an increased risk; a value less than 1 implies a protective effect or reduced risk; and a value of 1 denotes no difference in risk between groups.[1][2][3] Unlike absolute risk, which measures the actual probability of an event, relative risk focuses solely on the comparative likelihood and does not convey the baseline risk level in the population.[1] Relative risk is primarily employed in prospective cohort studies and clinical trials to quantify the impact of exposures, such as smoking or vaccinations, on health outcomes like cancer or infectious diseases.[1][3] For instance, in a tuberculosis outbreak, the relative risk might show exposed individuals are 6.1 times more likely to develop the disease compared to unexposed ones, highlighting the exposure's role in transmission.[2] Similarly, for varicella, vaccination can yield a relative risk of 0.28, demonstrating substantial risk reduction.[2] It differs from the odds ratio, which approximates relative risk in rare events but is derived from case-control studies, making relative risk more direct for incidence-based comparisons.[1]Fundamentals
Definition
Relative risk (RR), also known as the risk ratio, is a statistical measure used in epidemiology and public health to quantify the association between an exposure and an outcome by comparing the probabilities of the outcome occurring in two groups.[1] Specifically, it is defined as the ratio of the probability of an event occurring in the exposed group to the probability of the same event occurring in the unexposed group.[1] This measure helps assess how much the risk of the event, such as disease onset or adverse health effect, is increased or decreased by the exposure relative to those not exposed.[2] Key terms in this context include the exposed group, which refers to individuals who have experienced a specific risk factor, intervention, or condition under study (e.g., smokers or participants receiving a new treatment), and the unexposed group, which serves as the reference and consists of individuals without that exposure (e.g., non-smokers or those on standard care).[1] The probabilities compared are incidences, representing the risks or likelihoods of the event happening within a defined timeframe in each group.[1] These groups are typically followed prospectively in cohort studies to observe outcomes directly.[2] Prior to understanding relative risk, it is essential to grasp absolute risk, which denotes the actual probability of the event occurring in a given population or subgroup, providing the foundational baseline for any relative comparisons.[1] The term relative risk emerged in the field of epidemiology during the mid-20th century, particularly in the context of cohort studies investigating causal relationships, with Jerome Cornfield and colleagues formalizing its application in their influential 1959 analysis linking cigarette smoking to lung cancer.[4]Calculation
Relative risk (RR) is computed in cohort studies by comparing the incidence of an outcome in exposed and unexposed groups, using data organized in a 2×2 contingency table.[1] The table categorizes participants by exposure status (rows: exposed and unexposed) and outcome status (columns: event and non-event), with cells denoted as follows:| Event | Non-event | Total | |
|---|---|---|---|
| Exposed | a | b | a+b |
| Unexposed | c | d | c+d |
| Total | a+c | b+d | n |
Interpretation
Statistical Meaning
Relative risk (RR) quantifies the association between an exposure and an outcome by comparing the probability of the outcome in the exposed group to that in the unexposed group.[2] An RR greater than 1 indicates an increased risk of the outcome in the exposed group relative to the unexposed group, suggesting a positive association.[2] Conversely, an RR less than 1 signifies a decreased risk in the exposed group, implying a potential protective effect of the exposure.[2] An RR of exactly 1 denotes no association, where the risk is equivalent between the groups.[8] The magnitude of the RR provides insight into the strength of the association, though interpretations must account for potential confounding factors that could inflate or attenuate the estimate.[1] RRs exceeding 3 suggest a strong association, making alternative explanations like bias less likely.[9] These thresholds are not absolute and depend on study context, population characteristics, and adjustment for confounders such as age, sex, or comorbidities.[1] While RR measures directionality—indicating whether the exposure elevates, reduces, or has no effect on risk—it assesses association rather than causation, as temporal relationships, reverse causality, or unmeasured confounders may influence results.[1] Establishing causality requires additional evidence, such as consistency across studies, biological plausibility, and dose-response gradients, as outlined in the Bradford Hill criteria for causal inference in epidemiology. The strength of association criterion within these guidelines emphasizes that larger RRs support but do not prove causality. In scenarios involving rare outcomes (typically with incidence below 10%), the RR closely approximates the odds ratio, facilitating comparisons across study designs.[10] However, for common outcomes, the two measures diverge, with the odds ratio overestimating the RR, which can lead to misinterpretation if not recognized.[11]Reporting Practices
In clinical trials, the CONSORT guidelines recommend reporting relative risks (RR) alongside absolute measures of effect, such as risk differences, for binary outcomes, accompanied by 95% confidence intervals (CIs) to convey precision.[12] This ensures readers can assess both the magnitude and reliability of effects, with covariate-adjusted estimates preferred when applicable to enhance efficiency.[12] For observational studies, the STROBE guidelines similarly advocate presenting relative risks with confidence intervals, alongside unadjusted and adjusted estimates, and translating them to absolute risks where relevant to provide context.[13] These requirements facilitate critical appraisal by highlighting the precision of associations and their practical implications in cohort, case-control, or cross-sectional designs.[13] A common misinterpretation arises from overemphasizing relative risk without absolute risks, which can exaggerate effects; for instance, a 50% relative reduction in heart attack risk (RR = 0.5) from 2% to 1% baseline represents only a 1 percentage point absolute reduction, benefiting few patients overall.[14] Similarly, an RR of 4 on a 1% baseline risk yields a mere 3% absolute increase, yet may appear dramatically elevated without this context.[14] Visualizations like forest plots are standard for presenting multiple relative risks in meta-analyses or systematic reviews, displaying point estimates as squares, CIs as horizontal lines, and a pooled effect as a diamond, with a vertical line at RR=1 indicating no association.[15] Complementing these with risk differences in adjacent plots or tables provides essential context on absolute impacts, aiding interpretation beyond relative magnitudes.[15] Ethically, transparency in subgroup analyses is crucial to prevent cherry-picking significant relative risks while omitting non-significant ones, a form of spin that distorts findings and misleads readers.[16] Adhering to guidelines like CONSORT, which mandate pre-specifying and fully reporting all subgroups, mitigates this by promoting complete and unbiased disclosure of results.[16]Statistical Inference
Frequentist Methods
In frequentist statistics, point estimation of relative risk (RR) from 2×2 contingency table data involves directly applying the RR formula to observed counts, where RR is the ratio of the probability of the outcome in the exposed group to that in the unexposed group.[17] The variance of the natural logarithm of the estimated RR, \ln(\hat{\text{RR}}), is approximated using the delta method as \text{Var}(\ln(\hat{\text{RR}})) \approx \left( \frac{1}{a} - \frac{1}{a+b} \right) + \left( \frac{1}{c} - \frac{1}{c+d} \right), with a and b denoting exposed cases and non-cases, and c and d denoting unexposed cases and non-cases, respectively.[17] This approximation leverages the asymptotic normality of binomial proportions and facilitates subsequent inference by stabilizing the skewed distribution of RR.[18] Confidence intervals for RR are commonly constructed on the logarithmic scale to ensure positivity and symmetry, yielding a $100(1-\alpha)\% interval as \exp\left( \ln(\hat{\text{RR}}) \pm z_{\alpha/2} \sqrt{\text{Var}(\ln(\hat{\text{RR}}))} \right), where z_{\alpha/2} is the critical value from the standard normal distribution.[17] For small samples where asymptotic assumptions may fail, exact methods such as the Katz approach provide more reliable intervals by solving quadratic equations derived from the non-central hypergeometric distribution, particularly useful when expected cell counts are low.[19] Hypothesis testing for RR typically evaluates the null hypothesis that RR = 1 (no association between exposure and outcome) using a z-test on \ln(\hat{\text{RR}}), where the test statistic is z = \frac{\ln(\hat{\text{RR}})}{\sqrt{\text{Var}(\ln(\hat{\text{RR}}))}} under the null, compared to the standard normal distribution.[17] Equivalently, a chi-square test of independence on the 2×2 table assesses overall association, with the test statistic \chi^2 = \sum \frac{(O_i - E_i)^2}{E_i} (Yates' continuity correction optional for small samples) following a chi-square distribution with 1 degree of freedom; rejection indicates RR ≠ 1.[20] To adjust RR estimates for confounders in cohort studies, logistic regression models the log-odds of the outcome, yielding an odds ratio (OR) that approximates RR when the outcome is rare; for common outcomes, the Zhang and Yu correction converts the adjusted OR to an approximate RR via \text{RR} \approx \frac{\text{OR}}{(1 - P_0) + P_0 \cdot \text{OR}}, where P_0 is the outcome incidence in the unexposed group.[21] This method enables multivariable adjustment while providing interpretable risk ratios, though direct RR models like Poisson regression are alternatives for precise estimation in prospective data.[21]Bayesian Methods
In Bayesian estimation of relative risk (RR), the risks in exposed and unexposed groups are typically modeled using a binomial likelihood for the observed events, with conjugate beta priors assigned to the underlying proportions p_1 and p_0. The posterior distributions for p_1 and p_0 are then beta, from which the RR is derived as the ratio p_1 / p_0; for inference on the log RR, a normal approximation to the posterior is often employed due to the ratio's skewness.[22][23] Prior choices for the beta distributions balance informativeness and objectivity: non-informative options like the Jeffreys prior (Beta(0.5, 0.5)) provide weak information equivalent to half an event and non-event, while uniform priors (Beta(1,1)) offer even weaker regularization; informative priors, drawn from historical data such as prior incidence rates in similar populations, incorporate domain knowledge to update beliefs.[22] In contrast to frequentist confidence intervals, which represent long-run coverage probabilities, 95% credible intervals for RR enclose the true value with 95% posterior probability, enabling direct probabilistic statements about the parameter.[23] A key advantage of Bayesian methods emerges in small samples or sparse data, where shrinkage toward the prior mean reduces posterior variability and avoids issues like division by zero in RR calculation. For instance, in an epidemiological study with zero events in the unexposed group (e.g., a small cohort of 20 participants assessing a rare adverse outcome), a Beta(1,1) prior yields a posterior mean for RR of 1.5 with a 95% credible interval of (0.07, 2.47), narrower and more stable than frequentist alternatives, effectively borrowing strength from historical data on similar low-incidence events.[22]Comparisons
Odds Ratio
The odds ratio (OR) is a measure of association between an exposure and an outcome, defined as the ratio of the odds of the outcome occurring in the exposed group to the odds in the unexposed group. In a standard 2×2 contingency table, where a represents exposed cases, b exposed non-cases, c unexposed cases, and d unexposed non-cases, the OR is calculated as: \text{OR} = \frac{a/b}{c/d} = \frac{ad}{bc} This formulation arises from comparing the odds of disease among exposed versus unexposed individuals.[2] Unlike the relative risk (RR), which directly compares probabilities of the outcome, the OR compares odds and thus does not inherently represent a risk ratio. The OR approximates the RR when the outcome is rare (prevalence ≤10% in both groups), as the odds then closely mirror the probabilities; however, for common outcomes, the OR systematically overestimates the RR and is always farther from 1 than the true RR. This discrepancy increases with higher baseline risk (p_0, the probability of the outcome in the unexposed group). The exact relationship is given by: \text{OR} = \frac{\text{RR} \cdot (1 - p_0)}{1 - p_0 \cdot \text{RR}} For instance, if p_0 = 0.05 and RR = 2, then OR ≈ 2.10, a close approximation; but if p_0 = 0.40 and RR = 1.5, then OR ≈ 2.25, showing greater divergence from the RR.[24] In study design, the RR can be directly estimated from cohort studies, where both exposure and outcome incidences are observed prospectively, whereas case-control studies, which select participants based on outcome status, yield the OR as the primary measure since absolute risks are unavailable. To approximate the RR from an OR in such designs (or cohort studies using logistic regression), methods like the Zhang-Yu formula provide a correction: RR ≈ OR / (1 - p_0 + p_0 · OR), assuming knowledge of the baseline risk; this is particularly useful for common outcomes where the rare-disease assumption fails.[21] A common pitfall in interpreting the OR is treating it interchangeably with the RR, leading to exaggerated perceptions of effect sizes, especially in media reports and non-specialist summaries of epidemiological findings. For example, a study on Ebola survivors and safe sexual behavior reported an adjusted OR of 5.59 but interpreted it as "more than five times as likely," overstating the true RR by about 35% (unadjusted RR ≈ 2.71). Similarly, research on point-of-care testing for antibiotics cited an OR of 0.49 as implying a large risk reduction, yet the corresponding RR was 0.82, overstating the effect by 178%; media coverage amplified this as a near-halving of risk. Another case involved alcohol and aggression, where an OR of 6.41 was reported as "six times more likely" in headlines, but the RR was approximately 3.1, doubling the perceived association. Such misinterpretations occur in up to 26% of obstetrics and gynecology studies and underscore the need to specify whether results reflect odds or risks.[25][26]Other Risk Measures
The attributable risk (AR), also referred to as the risk difference, quantifies the excess incidence of an outcome directly attributable to exposure and is calculated as AR = I_e - I_u, where I_e is the incidence rate in the exposed group and I_u is the incidence rate in the unexposed group.[24] This additive measure provides insight into the absolute burden imposed by the exposure, independent of the baseline risk.[24] The population attributable fraction (PAF) extends this concept to the entire population, estimating the proportion of the outcome that would be eliminated if the exposure were removed; it is given by the formula PAF = \frac{P(RR - 1)}{1 + P(RR - 1)}, where P is the prevalence of exposure and RR is the relative risk.[27] For instance, if exposure prevalence is 0.2 and RR is 2, the PAF is approximately 0.167, indicating that 16.7% of cases in the population could be attributed to the exposure.[27] In contrast to the multiplicative scale of relative risk, the risk difference (RD = p_1 - p_0, where p_1 and p_0 are the probabilities of the outcome in the exposed and unexposed groups, respectively) operates on an additive scale and better captures the public health implications of an exposure.[24] While relative risk emphasizes proportional changes, which remain constant across baseline levels, RD highlights absolute increments in cases, making it essential for evaluating intervention impacts in population settings.[28] For example, a relative risk of 2.0 might double the outcome from 1% to 2% in a low-risk scenario (RD = 0.01, minimal public health effect) but from 40% to 80% in a high-risk context (RD = 0.40, substantial effect).[24] The number needed to treat (NNT) or number needed to harm (NNH) derives from the risk difference as NNT/NNH = 1 / |p_1 - p_0|, representing the number of individuals who must be exposed to the intervention (or risk factor) to cause or prevent one additional outcome.[29] These metrics connect directly to AR by translating absolute differences into practical terms, such as requiring 100 individuals to be treated to avert one event if RD = 0.01.[29] Policymakers often prefer NNT/NNH over relative risk for decision-making because they offer tangible estimates of resource needs and net benefits, especially when weighing costs against absolute gains in diverse risk populations.[30] Relative risk has notable limitations, particularly in high-baseline-risk scenarios where it may overstate proportional effects without conveying absolute magnitude; for instance, in cardiovascular disease studies with baseline incidences around 20-30%, a relative risk of 1.5 translates to a 5-15% absolute increase (large public health impact), whereas the same relative risk in rare disease studies (baseline <1%) yields only a 0.5% absolute rise (negligible impact).[24] This discrepancy underscores the need to pair relative risk with additive measures to avoid misjudging intervention value in common versus uncommon conditions.[24]Examples
Numerical Illustration
Consider a hypothetical cohort study examining the association between exposure to a risk factor and the occurrence of an adverse event. The study includes 100 individuals exposed to the risk factor, of whom 20 experience the event, and 100 unexposed individuals, of whom 10 experience the event.[31] This data can be organized into a 2x2 contingency table as follows:| Group | Event | No Event | Total |
|---|---|---|---|
| Exposed | 20 | 80 | 100 |
| Unexposed | 10 | 90 | 100 |
| Total | 30 | 170 | 200 |