Relative Risk & Risk Difference — Compute and interpret...

SECTION 1

Historical Context & Motivation

Epidemiology has always been driven by a deceptively simple question: does a particular exposure increase the risk of disease? Long before the germ theory of disease was established, physicians and public health officials tried to compare outcomes between groups—those exposed to a suspected cause and those who were not. The intellectual scaffolding for modern measures of association grew out of centuries of observational inquiry, from John Snow's cholera investigations to large-scale cohort studies in the twentieth century. Two of the most fundamental quantities that emerged from this tradition are the relative risk (RR) and the risk difference (RD), each capturing a different facet of how exposure relates to outcome.

1854

Snow's Cholera Map

John Snow compared cholera rates between London water-supply districts, implicitly computing risk ratios decades before the formal terminology existed. His work demonstrated the power of comparing incidence across exposure groups.

1951

Doll & Hill Cohort Study

Richard Doll and A. Bradford Hill launched a landmark prospective cohort study of British doctors, computing relative risks for lung cancer among smokers versus non-smokers. Their published ratios crystallized the modern definition of RR.

1970s

Formalization of Measures of Effect

Textbooks by Kleinbaum, Kupper, and Rothman formalized the distinction between ratio measures (RR) and difference measures (RD), emphasizing that each answers a fundamentally different question about public health impact.

2000s–Present

STROBE & Reporting Guidelines

International reporting standards such as STROBE now require authors to present both relative and absolute measures of association so that readers can assess clinical significance alongside statistical significance.

The central gap that these measures fill is the need for a precise, interpretable vocabulary to communicate how much more (or less) disease occurs in an exposed group compared with an unexposed group. Without RR and RD, epidemiologists would be forced to present raw counts—numbers that are virtually meaningless without knowledge of the underlying population sizes. The remainder of this lesson develops both measures from first principles, illustrates their computation, and explores the complementary insights each provides.

SECTION 2

Core Principles & Definitions

Before computing any measure of association, it is essential to understand the data structure that underlies categorical data analysis in epidemiology. We begin with a cohort (or experimental) design in which individuals are classified by their exposure status and then followed to ascertain whether the outcome of interest occurs. The foundational data layout is a 2 × 2 contingency table (also called a fourfold table), which cross-tabulates exposure (yes/no) against outcome (yes/no). From this table, we derive two group-specific risks and then combine them into either a ratio or a difference.

Risk (Incidence Proportion)

The probability of developing the outcome within a defined follow-up period. Computed as the number of new cases divided by the total number of individuals at risk in the group: Risk = a / (a + b) for the exposed group, where 'a' counts events and 'b' counts non-events.

Relative Risk (RR)

The ratio of the risk in the exposed group to the risk in the unexposed group. An RR of 1.0 indicates no association; values above 1.0 suggest increased risk with exposure, while values below 1.0 suggest a protective effect. RR is a multiplicative measure of association.

Risk Difference (RD)

The arithmetic difference between the risk in the exposed group and the risk in the unexposed group. RD quantifies the excess (or deficit) of risk attributable to exposure on the absolute scale. An RD of 0 indicates no association.

The 2 × 2 Table

The standard layout uses cells labeled a, b, c, d. Rows correspond to exposed (a, b) and unexposed (c, d); columns correspond to outcome present (a, c) and outcome absent (b, d). Row totals give n₁ = a + b and n₀ = c + d.

✦ KEY TAKEAWAY

Think of RR and RD as two different lenses on the same photograph. Relative risk is like asking 'How many times brighter is this lamp than that one?'—a ratio. Risk difference is like asking 'How many more lumens does this lamp produce?'—an absolute surplus. In clinical research, a drug might double your risk (RR = 2) of a very rare side effect (baseline risk 0.001), but the risk difference (RD = 0.001) is tiny. You need both numbers to judge whether the finding matters practically.

SECTION 3

Visual Explanation — The 2 × 2 Table and Risk Computation

The 2 × 2 table cross-classifies subjects by exposure and disease status. Cells a and c count diseased individuals in exposed and unexposed groups, respectively. Row totals n₁ and n₀ form the denominators for group-specific risks, which then combine into RR (ratio) and RD (difference).

The diagram above illustrates the logical flow from raw cell counts to the two summary measures. Notice that both RR and RD depend on exactly the same two intermediate quantities—R₁ (risk in the exposed) and R₀ (risk in the unexposed). The only difference is the arithmetic operation applied: division for RR, subtraction for RD. This shared dependence on group-specific risks means that both measures require a cohort-type design (or a randomized trial) in which we can estimate incidence directly. In a case-control study, incidence proportions cannot be computed, so the odds ratio is used instead—a topic treated in a subsequent lesson.

SECTION 4

Mathematical Framework

We now formalize the two measures. Consider a cohort of N individuals followed for a fixed period. Let the exposed group contain n₁ individuals and the unexposed group contain n₀ individuals, with N = n₁ + n₀. Among the exposed, a develop the outcome; among the unexposed, c develop the outcome.

RISK IN EXPOSED GROUP

R₁ = a / n₁ = a / (a + b)

where a = number of exposed individuals who develop the outcome, and n₁ = total number of exposed individuals.

RISK IN UNEXPOSED GROUP

R₀ = c / n₀ = c / (c + d)

where c = number of unexposed individuals who develop the outcome, and n₀ = total number of unexposed individuals.

RELATIVE RISK (RISK RATIO)

RR = R₁ / R₀ = [a / (a + b)] / [c / (c + d)]

RR > 1 indicates the exposure is associated with increased risk; RR < 1 indicates a protective association; RR = 1 means no association. The null value is 1.

RISK DIFFERENCE (ATTRIBUTABLE RISK)

RD = R₁ − R₀ = a/(a + b) − c/(c + d)

RD > 0 indicates excess risk due to exposure; RD < 0 indicates reduced risk. The null value is 0. RD is sometimes called the attributable risk because it quantifies the absolute amount of risk that can be attributed to the exposure under a causal interpretation.

An important derived quantity is the number needed to treat (NNT) or number needed to harm (NNH), obtained as the reciprocal of the absolute value of the risk difference: NNT = 1 / |RD|. This tells a clinician how many patients must be treated (or exposed) for one additional outcome event to occur (or be prevented). For instance, if RD = −0.05, then NNT = 1/0.05 = 20, meaning 20 patients must be treated to prevent one adverse event.

📐 Confidence Intervals

For large samples, a 95% confidence interval for ln(RR) is ln(RR) ± 1.96 × SE, where SE = √(b/(a×n₁) + d/(c×n₀)). Exponentiate both endpoints to obtain the CI for RR. For RD, the CI is RD ± 1.96 × √(R₁(1−R₁)/n₁ + R₀(1−R₀)/n₀). If the CI for RR contains 1, or the CI for RD contains 0, the association is not statistically significant at α = 0.05.

SECTION 5

Interpreting RR and RD — A Comparative View

One of the most important lessons in epidemiology is that a large relative risk can coexist with a small risk difference, and vice versa. Understanding both is critical for sound clinical and policy decisions. The SVG diagram below places three hypothetical scenarios on a common visual scale, showing how baseline risk dramatically influences the relationship between RR and RD.

Three scenarios with identical RR = 2.0 demonstrate that relative risk alone cannot convey the absolute burden of disease. As baseline risk rises from 0.01 to 0.40, the risk difference grows forty-fold, and the number needed to harm drops from 100 to 2.5.

Summary comparison of the two measures of association.

Measure	Scale	Null Value	Range	Interpretation Emphasis
RR	Multiplicative (ratio)	1.0	0 to ∞	Etiologic strength; how many times more likely?
RD	Additive (difference)	0	−1 to +1	Public health impact; how many extra cases per unit population?

SECTION 6

Worked Example — Smoking and Lung Cancer in a Cohort Study

A prospective cohort study follows 5,000 smokers and 5,000 non-smokers for 20 years. By the end of follow-up, 200 smokers and 20 non-smokers have been diagnosed with lung cancer. We wish to compute the relative risk and risk difference for the association between smoking and lung cancer.

Computing RR and RD for Smoking and Lung Cancer

Step 1 — Construct the 2 × 2 TableExposed (smokers): a = 200 lung cancer cases, b = 4,800 non-cases, n₁ = 5,000. Unexposed (non-smokers): c = 20 cases, d = 4,980 non-cases, n₀ = 5,000.

Step 2 — Compute Risk in Each GroupRisk in smokers: R₁ = a / n₁ = 200 / 5,000 = 0.04 (4.0%). Risk in non-smokers: R₀ = c / n₀ = 20 / 5,000 = 0.004 (0.4%).

R₁ = 0.04, R₀ = 0.004

Step 3 — Compute the Relative RiskRR = R₁ / R₀ = 0.04 / 0.004 = 10.0. Smokers are 10 times as likely as non-smokers to develop lung cancer over the 20-year follow-up period.

RR = 10.0

Step 4 — Compute the Risk DifferenceRD = R₁ − R₀ = 0.04 − 0.004 = 0.036 (3.6 percentage points). For every 1,000 smokers, approximately 36 additional lung cancer cases occur compared with 1,000 non-smokers.

RD = 0.036

Step 5 — Compute the Number Needed to HarmNNH = 1 / RD = 1 / 0.036 ≈ 27.8. Roughly 28 smokers must be 'exposed' (i.e., persist in smoking over 20 years) for one additional case of lung cancer to arise compared with non-smokers.

NNH ≈ 28

💡 Interpretation Note

The RR of 10.0 conveys a strong etiologic signal—smoking vastly multiplies the risk. The RD of 0.036, while seemingly small, still has enormous public health significance when multiplied by millions of smokers worldwide. Always consider both the strength of association (RR) and the absolute burden (RD) when evaluating study results.

SECTION 7

Strengths, Limitations & When to Use Each Measure

No single measure of association tells the whole story. Both RR and RD are essential tools, but each has contexts in which it excels and situations in which it may mislead. The table below summarizes the key trade-offs, helping you decide which measure to emphasize depending on the research question and audience.

Comparative profile of RR and RD.

Criterion	Relative Risk (RR)	Risk Difference (RD)
Primary question answered	How many times more likely is the outcome among exposed?	How many additional cases per population unit are due to exposure?
Best for	Etiologic research; understanding causal strength; comparing across studies with different baseline risks	Clinical decision-making; resource allocation; communicating absolute benefit/harm to patients
Key strength	Tends to be stable across populations with differing baseline risks (under certain model assumptions)	Directly translates to NNT/NNH; intuitive for absolute impact
Key limitation	Can exaggerate perceived importance when baseline risk is very low; not estimable from case-control data	Varies with baseline risk, so not generalizable across populations; also not estimable from case-control data
Study designs	Cohort studies, randomized trials, cross-sectional studies (for prevalence ratio)	Same as RR: requires direct estimation of incidence

✦ KEY TAKEAWAY

Imagine two news headlines: 'New drug doubles heart attack risk!' (RR = 2) versus 'New drug increases heart attack risk by 1 in 10,000' (RD = 0.0001). The first sounds terrifying; the second is almost negligible. Neither headline is wrong—they describe the same data from different vantage points. A responsible researcher or clinician always reports both to give decision-makers a complete picture.

SECTION 8

Connections to the Odds Ratio, Attributable Fraction & Regression Models

Relative risk and risk difference are the entry points into a larger family of measures of association and effect. Two natural extensions are the odds ratio (OR) and the attributable fraction (AF). Understanding how RR and RD relate to these quantities deepens your command of epidemiologic reasoning and prepares you for multivariate modeling.

Advanced measures and their links to RR and RD.

Concept	Relationship to RR / RD	When It Appears
Odds Ratio (OR)	OR = (a × d) / (b × c). When the outcome is rare (< 10%), OR ≈ RR (the rare-disease assumption). In case-control studies, OR is the only ratio measure estimable because incidence cannot be computed directly.	Case-control studies; logistic regression output
Attributable Fraction in Exposed (AFₑ)	AFₑ = (RR − 1) / RR = RD / R₁. Gives the proportion of disease among exposed that is attributable to the exposure.	Causal inference; etiologic fraction estimation
Population Attributable Fraction (PAF)	PAF = pₑ × (RR − 1) / [1 + pₑ × (RR − 1)], where pₑ is the prevalence of exposure. Estimates the proportion of total disease burden preventable by eliminating the exposure.	Public health planning; burden-of-disease studies
Log-binomial / Poisson Regression	These generalized linear models directly estimate ln(RR) while adjusting for confounders. The exponentiated regression coefficient gives an adjusted RR.	Multivariable analysis of cohort data with common outcomes

As you advance through biostatistics, you will encounter situations where confounding and effect modification require multivariable adjustment. In such settings, log-binomial regression models the log of the risk directly, yielding adjusted RR estimates, while identity-link binomial regression models the risk on the additive scale to yield adjusted RD estimates. The Mantel-Haenszel method offers a stratified, non-parametric approach to pooling stratum-specific RRs when confounders are categorical. Each of these methods builds directly on the foundational definitions of R₁ and R₀ developed in this lesson.

SECTION 9

Practice Problems

PROBLEM 1 — CONCEPTUAL

A cohort study reports that the risk of developing type 2 diabetes is 0.12 among individuals with obesity and 0.04 among individuals without obesity. Without performing any calculations, explain in your own words what it means for the relative risk to be greater than 1 and for the risk difference to be greater than 0. What distinct aspect of the exposure–disease relationship does each measure capture?

PROBLEM 2 — BASIC CALCULATION

In a randomized controlled trial, 300 patients receive a new vaccine and 300 receive a placebo. Over one year, 15 vaccinated patients and 45 placebo patients develop the infection. Compute R₁ (risk in placebo group, treated here as 'exposed' to no vaccine), R₀ (risk in vaccinated group), the relative risk for infection comparing placebo to vaccine, and the risk difference.

PROBLEM 3 — INTERMEDIATE

A cohort study of 4,000 factory workers (2,000 exposed to a chemical solvent, 2,000 unexposed) finds 80 cases of liver disease among the exposed and 25 among the unexposed. (a) Compute RR and RD. (b) Compute a 95% confidence interval for RR using the log-transformation method: 95% CI for ln(RR) = ln(RR) ± 1.96 × √(b/(a × n₁) + d/(c × n₀)). (c) Does the interval include 1? What do you conclude?

PROBLEM 4 — APPLIED

A public health department reviews two studies of air pollution and childhood asthma. Study 1: In a high-pollution city, risk of asthma = 0.15; in a low-pollution city, risk = 0.05. Study 2: In a different high-pollution setting, risk = 0.30; in the comparison low-pollution setting, risk = 0.10. Compute RR and RD for each study. The health department has budget to intervene in one setting. Using RD and the concept of NNT, argue which setting offers the greatest absolute benefit per intervention dollar if the cost per person is the same.

PROBLEM 5 — CRITICAL THINKING

A colleague claims: 'Because the odds ratio approximates the relative risk when the disease is rare, we can always use OR as a substitute for RR in cohort studies.' Critically evaluate this claim. Under what specific conditions does OR ≈ RR, and what happens to the approximation as the outcome becomes more common? Additionally, explain why the risk difference cannot be approximated from OR data alone without knowledge of the baseline risk.

SUMMARY

Lesson Summary

This lesson introduced two foundational measures for quantifying the association between a binary exposure and a binary outcome in cohort-type designs. The relative risk (RR) is the ratio of the risk in the exposed group (R₁) to the risk in the unexposed group (R₀), with a null value of 1.0 and a range from 0 to infinity. It captures the multiplicative strength of the association and is most useful for etiologic research and causal reasoning. The risk difference (RD) is the arithmetic difference R₁ − R₀, with a null value of 0. It provides the absolute excess risk attributable to exposure, and its reciprocal yields the number needed to treat or harm (NNT/NNH)—a metric of direct clinical utility.

Critically, the same RR can correspond to vastly different values of RD depending on the baseline risk, so responsible reporting always presents both measures. Neither RR nor RD is estimable from case-control data, where the odds ratio serves as the primary measure of association. Extensions such as the attributable fraction, population attributable fraction, and regression-based adjusted estimates all build upon the concepts of R₁, R₀, RR, and RD introduced here.