HSCI 341, Lesson 6

Measures of
Association

Fundamental Epidemiological Concepts and Approaches

Learning objectives for this lesson:

  1. Calculate and interpret the risk ratio, the incidence rate ratio and the odds ratio from data set out in the course’s 2×2 layout, and the prevalence ratio from a cross-sectional survey.
  2. Explain why the odds ratio is the measure of association for case-control studies and when it approximates the risk ratio, and identify which measures each study design can supply.
  3. Calculate and interpret the risk difference, the incidence rate difference and the number needed to treat or harm, and distinguish the questions that ratio (relative) and difference (absolute) measures answer.
  4. Calculate and interpret the attributable fraction in the exposed and vaccine efficacy, stating the causal assumption that each interpretation needs.
  5. Calculate and interpret the population attributable risk and the population attributable fraction, and explain how the strength of the association, the prevalence of exposure and confounding affect them.
  6. Interpret confidence intervals, P-values and test results for measures of association, distinguish the strength of an association from its statistical significance and precision, and choose an appropriate test for a 2×2 table or for matched pairs.

This course was developed by Dr. Kiffer G. Card based on Dohoo, I. R., Martin, S. W., & Stryhn, H. (2012). Methods in Epidemiologic Research. VER Inc.

Reference

Glossary: Key Terms, People & Concepts

📚 Reference page, available throughout the lesson

This glossary collects the key concepts, people, and ideas you will meet in this lesson. Use it as a reference while you work through the material, or as a review before assessments. Type in the search box to filter entries.

Ratio Measures of Association
Risk Ratio (Relative Risk, RR) The ratio of the cumulative incidence (risk) in the exposed group to that in the unexposed group: RR = R₁ / R₀. Used in cohort studies and RCTs over a defined follow-up period.
Incidence Rate Ratio (IRR) The ratio of incidence rates (events per person-time) in the exposed and unexposed groups. Appropriate when follow-up time varies across individuals.
Odds Ratio (OR) The ratio of the odds of disease in the exposed group to the odds in the unexposed group (Cornfield, 1951). The standard measure in case-control studies and the natural output of logistic regression.
Prevalence Ratio (PR) The ratio of prevalence in the exposed group to prevalence in the unexposed group. Preferred over the prevalence odds ratio when the outcome is not rare in cross-sectional analyses.
Hazard Ratio (HR) The ratio of instantaneous event rates (hazards) between groups, typically from Cox proportional hazards regression. Approximates the rate ratio under proportional hazards.
Rare-Disease Assumption When disease prevalence is low (commonly < 5%), the odds ratio approximates the risk ratio. Important for interpreting case-control results.
Difference Measures of Association
Risk Difference (RD, Attributable Risk) The absolute difference in risk between exposed and unexposed groups: RD = R₁ − R₀. Captures the public-health impact in absolute terms.
Rate Difference The absolute difference in incidence rates between exposed and unexposed groups, in events per person-time.
Attributable Fraction in the Exposed (AFe) Among the exposed, the proportion of disease attributable to the exposure: AFe = (R₁ − R₀) / R₁ = (RR − 1) / RR.
Population Attributable Risk (PAR) The excess risk in the total population attributable to the exposure: PAR = Rpop − R₀. Reflects both effect size and exposure prevalence.
Population Attributable Fraction (AFp) The proportion of disease in the total population attributable to the exposure (Northridge, 1995). Useful for prioritising public-health interventions.
Number Needed to Treat (NNT) 1 / |risk difference|: the average number of patients who must receive a beneficial intervention for one to avoid the bad outcome (Laupacis, Sackett, & Roberts, 1988).
Number Needed to Harm (NNH) 1 / |risk difference| for harmful exposures: the average number exposed for one additional case of harm.
Inference & Interpretation
Null Hypothesis (H₀) A statement of no effect or no difference (e.g., RR = 1, RD = 0). The hypothesis tested by significance tests.
p-Value The probability of observing data as extreme or more extreme than the observed, assuming the null hypothesis is true. Not the probability that the null is true.
Confidence Interval A range of values consistent with the data at a chosen confidence level. For a ratio measure, an interval that excludes 1 implies statistical significance at the corresponding level.
Strength of Association vs. Statistical Significance The size of an effect (e.g., RR) is conceptually distinct from how confident we are that it differs from chance (p-value, CI width). Large samples can yield significant findings for trivial effects, and vice versa.
Effect Modification (Interaction) When the magnitude of an exposure-outcome association differs across levels of a third variable. Reported as stratum-specific estimates rather than adjusted away. Additive and multiplicative scales can give different interaction conclusions; reporting both is recommended (Knol & VanderWeele, 2012; VanderWeele & Knol, 2014).
No matching entries. Try a different search term.
Section 1

Introduction & Ratio Measures of Association

⏱ Estimated reading time: 15 minutes

Lesson 6 · HSCI 341

Comparison Is the Core Operation

A single frequency tells you how common a disease is. A measure of association tells you how much the exposure changes it.

Section 1 of 4

Ratio Measures of Association

Risk ratio, incidence rate ratio, and odds ratio: formulas, relationships, and worked examples.

The core idea

Why ratios, and why not just P-values

Measure of association

How strongly is the exposure linked to disease? Independent of sample size. The number that answers the biological or causal question.

P-value

How likely are these data if there were truly no association? Sensitive to sample size. Does not describe effect magnitude.

Null value for ratio measures: 1. Values > 1 indicate increased risk; values < 1 indicate protection.

Three formulas

RR, IR, and OR from the 2x2 table

Risk Ratio (RR)
\[ \color{#0B7B6B}{RR} = \frac{\color{#C2410C}{a_1/(a_1+b_1)}}{\color{#1D4ED8}{a_0/(a_0+b_0)}} = \frac{\color{#C2410C}{R_{E+}}}{\color{#1D4ED8}{R_{E-}}} \]
RR risk ratioRE+ risk in the exposedRE− risk in the unexposed
Incidence Rate Ratio (IR)
\[ \color{#0B7B6B}{IR} = \frac{\color{#C2410C}{a_1 / T_1}}{\color{#1D4ED8}{a_0 / T_0}} \]
IR incidence rate ratioa1/T1 rate in the exposeda0/T0 rate in the unexposed
Odds Ratio (OR)
\[ \color{#0B7B6B}{OR} = \frac{\color{#C2410C}{a_1 / b_1}}{\color{#1D4ED8}{a_0 / b_0}} = \frac{\color{#C2410C}{a_1} \color{#6D28D9}{b_0}}{\color{#1D4ED8}{a_0} \color{#BE185D}{b_1}} \]
OR odds ratioa1 exposed casesb1 exposed non-casesa0 unexposed casesb0 unexposed non-cases
Worked example 1

Brazil water cistern study

Exposure: household water cistern
Outcome: diarrhea

RR
\[ RR = \frac{194/1782}{303/1617} = \frac{0.109}{0.187} = 0.58 \]
OR
\[ OR = \frac{194 \times 1314}{303 \times 1588} = 0.53 \]

Interpretation

Both measures are below 1: cistern access is protective. The RR of 0.58 means the risk of diarrhea is 42% lower in the cistern group. OR is further from the null than RR, as the general relationship predicts.

Worked example 2

Migraine incidence rates

Exposure: female sex (vs. male)
Outcome: migraine onset, ages 30-40 (hypothetical data)

Incidence Rate Ratio
\[ IR = \frac{131/250}{44/236} = \frac{0.524}{0.186} = 2.81 \]

Interpretation

The migraine rate is 2.81 times higher in females than males aged 30-40. Person-months in the denominator account for variable follow-up time across participants.

The OR's special role

Why OR is the case-control measure

OR is symmetric: the odds of disease given exposure equal the odds of exposure given disease. Case-control studies sample on outcome, so only OR can be computed directly.

Cornfield approximation (rare disease)
\[ \text{When } \color{#C2410C}{p(D+)} < 0.05:\quad \color{#0B7B6B}{OR} \approx \color{#1D4ED8}{RR} \]
p(D+) disease frequencyOR odds ratioRR risk ratio

As disease prevalence rises, OR diverges from RR and moves further from the null. IR sits between RR and OR. This ordering holds on both sides of one.

Carry forward

Three anchors from this section

  • RR and IR come from cohort studies where the population at risk is observed directly.
  • OR is the case-control measure; it approximates RR when disease is rare (under 5%).
  • Ordering from the null: IR further than RR, OR furthest of all, on both sides of 1.

Overview

How often disease occurs is described by prevalence, incidence risk, incidence rate and odds, and the 2×2 table is the basic way of cross-classifying people by exposure and disease. This lesson brings the two together to compare the frequency of disease in exposed and non-exposed groups, which is what a measure of association does and why such measures are the central output of analytic epidemiology. The lesson builds the toolkit in four sections: the ratio measures (this section), the difference measures and attributable fractions in the exposed group (Section 2), the population-level measures and the designs that supply each measure (Section 3), and the standard errors, tests and confidence intervals that turn each point estimate into a defensible inference (Section 4). This section first sets out the 2×2 layouts used throughout the lesson, then defines and calculates the risk ratio, the odds ratio and the incidence rate ratio, explains how the three are related and why the odds ratio is the measure for case-control studies, and closes with what a measure of association tells you that a P-value does not.

Learning Objectives

  • 1.1 Set up incidence risk and incidence rate data in the course’s 2×2 layout, with exposure in the columns and disease in the rows.
  • 1.2 Calculate and interpret the risk ratio (RR), incidence rate ratio (IR), and odds ratio (OR).
  • 2.1 Describe the relationships among RR, IR, and OR, and explain why the OR is the measure for case-control studies and when it approximates the RR.
  • 6.1 Explain why measures of association are used in epidemiology, and what a measure of association tells you that a P-value does not.

1.1 Setting Out the Data in a 2×2 Table

Depending on study design, disease frequency can be expressed as incidence risk, incidence rate, prevalence or odds. Every formula in this lesson is written in terms of the cell labels of two layouts, and setting the data out in the same way each time is the simplest guard against putting a count in the wrong place. For risk data the course uses the 2×2 table in Table 6.1, with exposure status in the columns (exposed, then non-exposed) and disease status in the rows (diseased, then non-diseased). The cells a1 and a0 count the exposed and non-exposed people who developed the disease, b1 and b0 count those who did not, the column totals n1 and n0 are the numbers of exposed and non-exposed people, and the row totals m1 and m0 are the numbers with and without the disease.

Table 6.1. Layout of the 2×2 table for incidence risk data, with exposure status in the columns and disease status in the rows.

ExposedNon-exposedTotal
Diseaseda1a0m1
Non-diseasedb1b0m0
Totaln1n0n

For rate data the denominator is the person-time at risk rather than the number of individuals. Person-time is the total amount of time that the people in a group spent at risk of the disease while they were being observed. By this we mean that each person contributes the time between entering the study and either developing the disease, leaving the study or reaching its end, and these contributions are added up. For example, one person followed for two years and another followed for six months together contribute 2.5 person-years. Counting time rather than people matters because, in many studies, people are followed for different lengths of time, and someone watched for longer has more opportunity to become a case. Because the denominator is time, the second row of the table holds the person-time contributed by each exposure group, t1 and t0, in place of the non-diseased people, as in Table 6.2.

Table 6.2. Layout of the table for incidence rate data, giving the number of cases and the person-time at risk in each exposure group.

ExposedNon-exposedTotal
Number of casesa1a0m1
Person-time at riskt1t0t

A case-control study uses the same columns, but its rows hold cases and controls. The cases fill the first row, a1 exposed and a0 non-exposed, with row total m1; the controls fill the second row, b1 exposed and b0 non-exposed, with row total m0; and the column totals n1 and n0 and the grand total n are formed as before. The difference is that the investigator fixes the row totals m1 and m0 by deciding how many cases and controls to recruit, so the column totals n1 and n0 are not groups of people at risk. By this we mean that the column totals are partly a product of the investigator’s choices rather than a picture of how many exposed and non-exposed people exist in the population. For example, if the investigator had recruited three controls per case instead of two, the number of people in each column would change even though nothing about the disease or the exposure had changed. Because of this, dividing the cases in a column by the column total does not give a risk of disease. Worked Example 6.3 in 2.1.2 sets out a case-control table in this layout, and 2.1 explains the consequences. With the data in one of these layouts, each ratio measure is a single division of one group’s disease frequency by the other’s.

1.2 The Risk Ratio, the Odds Ratio and the Incidence Rate Ratio

The risk ratio, the odds ratio and the incidence rate ratio each divide a measure of disease frequency in the exposed group by the same measure in the non-exposed group. All three were defined in HSCI 230, and the recall panel below summarises those definitions and poses a short retrieval question; try to answer it before you open the answer.

Recall: Risk, Rate and Odds Ratios

HSCI 230 Lesson 5, Sections 1 and 2 (Introduction and Cohort Study Design; Risk-Based and Rate-Based Designs), defined the risk ratio as the risk of disease in the exposed divided by the risk in the unexposed, and the incidence rate ratio as the same comparison of incidence rates, each a count of new cases over person-time. HSCI 230 Lesson 4, Section 1 (Introduction and the Study Base), defined the odds ratio as the cross-product of the two-by-two table, the measure that a case-control study estimates. For all three, a value of 1 means no association, values above 1 mean more disease in the exposed and values below 1 mean less. This section adds which designs supply each measure and how the three relate to each other, and Section 4 adds their confidence intervals.

Retrieval question. In a cohort, 30 of 200 exposed and 10 of 200 unexposed people develop the disease. What are the risk ratio and the odds ratio?

Show the answer▼

RR = (30/200) ÷ (10/200) = 0.15 ÷ 0.05 = 3.0, and OR = (30 × 190) ÷ (10 × 170) = 5,700 ÷ 1,700 = 3.35. The odds ratio lies further from 1, as the relationships later in this section explain.

The three measures differ in the data they require, and so in the study designs that can supply them, a point Section 3 returns to when it maps every measure onto the designs. Click each card below to see the formula for each measure and the designs from which it can be computed.

Risk Ratio (RR)Click to learn more
Incidence Rate Ratio (IR)Click to learn more
Odds Ratio (OR)Click to learn more

The cards show that only the odds ratio can be computed from a case-control study, which is why 2.1 gives it particular attention. The formulae themselves are set out next, each with a worked example and a calculator.

1.2.1 The Risk Ratio and the Odds Ratio

The risk ratio (RR) divides the risk of disease in the exposed group, a1/n1, by the risk in the non-exposed group, a0/n0, as Equation 6.1 shows.

Risk ratio
RR = (a1/n1) / (a0/n0)Eq 6.1
The risk ratio is the risk in the exposed group divided by the risk in the unexposed group.

The odds ratio (OR) is built from odds rather than risks. The risk of an outcome is the number of cases divided by everyone at risk, meaning cases plus non-cases, whereas the odds of the same outcome is the number of cases divided by the non-cases alone. If 20 of 100 people get sick, the risk is 20/100 = 0.20, but the odds are 20/80 = 0.25. Risk and odds stay close while an outcome is uncommon and pull apart as it becomes common, which is the same fact that lets the odds ratio stand in for the risk ratio only when disease is rare (2.1). Equation 6.2 divides the odds of disease in the exposed group, a1/b1, by the odds in the non-exposed group, a0/b0, and the result simplifies to the cross-product of the table.

Odds ratio
OR = (a1/b1) / (a0/b0) = (a1 × b0) / (a0 × b1)Eq 6.2
The odds ratio is the odds of disease in the exposed group divided by the odds of disease in the unexposed group, which simplifies to the cross-product of the table.

Worked Example 6.1 applies Equation 6.1 and Equation 6.2 to the same table, from a study in Brazil that compared diarrhea among those with and without a water cistern (the counts are a teaching version of the survey data of Marcynuk and colleagues, 2013), set out in the layout of Table 6.1. Watch how the two measures compare when the outcome is fairly common.

Worked Example 6.1: Brazil Water Cistern Study

Diarrhea & Water Cistern Presence
Water CisternNo CisternTotal
Diarrhea Present194303497
Diarrhea Absent1,5881,3142,902
Total1,7821,6173,399
  • RR = (194/1782) / (303/1617) = 0.109 / 0.187 = 0.58
  • OR = (194 × 1314) / (303 × 1588) = 0.53

Both measures indicate that having a water cistern is protective against diarrhea (values < 1). The RR of 0.58 means the risk is 42% lower among those with cisterns.

The two measures agree on the direction of the association but not on its size: the OR of 0.53 lies further from 1 than the RR of 0.58, because diarrhea was common in this population, a pattern that 2.1 explains. The example also shows the easiest way to express a protective ratio, as a percentage reduction, 1 − RR. Calculator 6.1 and Calculator 6.2 start from the cistern data, so entering the same counts into both shows how far the two measures separate when the outcome is common.

The risk ratio and the odds ratio both count people. When people are followed for different lengths of time, the comparison is made with incidence rates instead.

1.2.2 The Incidence Rate Ratio

The incidence rate ratio (IR) divides the incidence rate in the exposed group, a1/t1, by the incidence rate in the non-exposed group, a0/t0, each a count of new cases over person-time at risk, as Equation 6.3 shows.

Incidence rate ratio
IR = (a1/t1) / (a0/t0)Eq 6.3
The incidence rate ratio is the incidence rate in the exposed group divided by the incidence rate in the unexposed group, each a count of cases over person-time.

Worked Example 6.2 applies Equation 6.3 to hypothetical rate data in the layout of Table 6.2, comparing migraine in females and males aged 30 to 40; note that each rate divides by person-months, not by people.

Worked Example 6.2: Migraine Incidence Rates

Gender and Migraine (Ages 30–40)
FemaleMaleTotal
Cases of migraine13144175
Person-months250236486

IR = (131/250) / (44/236) = 0.524 / 0.186 = 2.81

The rate of migraine is 2.81 times higher in females than males aged 30–40.

The IR is read against the same null value of 1 as the other ratio measures, but its interpretation compares rates rather than risks. A risk tells you what proportion of a group became ill over a period, whereas a rate tells you how quickly new cases arose for each unit of time that people spent at risk. This changes the wording of the result. For example, an IR of 2.81 is described as a rate of migraine 2.81 times as high in females, that is, almost three times as many new cases per person-month of follow-up, and not as 2.81 times the risk. Calculator 6.3 starts from the migraine data of the worked example.

The cistern example has already shown that the ratio measures can give different values for the same data. The next subsection explains how the three measures are related and when one of them can stand in for another.

2.1 How the Ratio Measures Relate, and the Odds Ratio in Case-Control Studies

2.1.1 When the Three Ratios Agree

In general, IR values are further from the null (1) than RR values, and OR values are even further away (Cornfield, 1951; Knol et al., 2008). By this we mean that, when an exposure is harmful, the OR is usually the largest of the three and the RR the smallest, and when an exposure is protective, the OR is usually the smallest and the RR closest to 1. The reason lies in the denominators. The odds divide cases by non-cases only, and in the group with more disease the number of non-cases is smaller, so its odds are pushed up more than its risk is. For example, a risk of 0.40 corresponds to odds of 0.40/0.60 = 0.67, while a risk of 0.20 corresponds to odds of 0.20/0.80 = 0.25, so a risk ratio of 2.0 becomes an odds ratio of about 2.7. In a similar way, people in the group with more disease leave the population at risk sooner when they become cases, so that group contributes less person-time, and its rate is pushed a little further from the other group’s than its risk is. Figure 6.1 shows this ordering on a number line, for a protective exposure to the left of 1 and a harmful exposure to the right of it.

1 (null) 0 ∞ OR IR RR RR IR OR

Figure 6.1. General relationships among RR, IR, and OR. OR is always furthest from the null value of 1.

The rule to remember is that the OR approximates the RR only when the disease is rare, with a risk below about 5 per cent; for a common outcome, reading an OR as if it were an RR overstates the effect. Expand each panel below to see why the OR approximates the RR for a rare disease, when the RR and the IR are close, and how the way controls are sampled decides whether the OR from a case-control study estimates the IR.

When is OR ≈ RR?▼

When the disease is rare (prevalence or incidence risk < 5%), OR approximates RR, the classic Cornfield (1951) approximation, which showed how case-control studies such as the landmark study of smoking and lung cancer by Doll and Hill (1950) could be interpreted in terms of relative risk. This is because when a1 is small relative to n1, the denominator of the odds (b1) is approximately equal to n1, and similarly for the non-exposed group. By this we mean that when almost nobody becomes ill, the non-cases make up nearly the whole group, so dividing the cases by the non-cases gives almost the same answer as dividing them by everyone. For example, 2 cases among 200 people give a risk of 2/200 = 0.010 and odds of 2/198 = 0.0101, which are practically identical, and so the ratio of odds is practically the same as the ratio of risks. As the disease becomes more common, the non-cases shrink as a share of the group, and the two measures pull apart. In the cistern example, the overall risk was 14.6%, so OR (0.53) was more extreme than RR (0.58); when outcomes are not rare, treating OR as RR can substantially overstate the effect (Knol et al., 2008, 2012).

When is RR ≈ IR?▼

RR and IR will be close to each other if the exposure has a negligible impact on the total time at risk in the study population. By this we mean that the exposed and non-exposed groups contribute roughly the person-time you would expect from their sizes, because few people in either group stop being at risk early by becoming cases. This occurs when the disease is rare, since then almost everyone stays at risk for the whole follow-up, or when IR is close to the null value (IR ≈ 1), since then cases remove person-time from the two groups at about the same pace. For example, if only a handful of people in a large cohort become ill, the person-time in each group is almost the number of people multiplied by the length of follow-up, so dividing by person-time and dividing by people give nearly the same comparison.

OR as an Estimator of IR▼

OR is a good estimator of IR under certain conditions in case-control studies. If controls are selected using cumulative or risk-based sampling (all non-cases after cases have occurred), then OR estimates IR only if the disease is rare. If controls are selected using density sampling (a control selected from non-cases each time a case occurs), then OR is a direct estimate of IR regardless of disease rarity.

The panels share one theme: the measures converge when the outcome is rare and diverge as it becomes common. Interactive 6.1 lets you see this directly. It computes the RR and the OR from one 2×2 table and plots how far the OR moves from the RR as the outcome becomes more common; move the prevalence shifter and try the presets to see when the rare-disease condition holds.

⚖ Interactive 6.1: Risk Ratio vs. Odds Ratio

This tool calculates the risk ratio (RR) and the odds ratio (OR) from the same 2×2 table and plots how far apart they are as the outcome becomes more common. It is meant to show you why the OR can stand in for the RR only when the outcome is rare. As you move the outcome prevalence shifter to the right, notice that the RR stays about the same (the flat RR line on the chart) while the OR moves away from it: above the RR for a harmful exposure and below it for a protective one. You can also click any cell of the table to type your own counts, or start from one of the presets.
2×2 table (click cells to edit)
Y+Y−Total
E+4060100
E−2080100
Scales the table to give this overall prevalence (preserving RR).
Presets:
RR vs. OR as outcome prevalence climbs
Risk in E+
–
Risk in E−
–
Outcome prevalence
–
Risk Ratio
–
Odds Ratio
–
OR / RR
–
Try the Common outcome preset: a true RR of 2.0 produces an OR around 3.0+. Reporting the OR as if it were a "risk ratio" overstates the harm by 50%. The "rare disease assumption" is what justifies treating OR as RR, and it should not be assumed without checking.

The rare-disease condition matters most in the one design that can supply only the odds ratio, the case-control study.

2.1.2 The Odds Ratio in Case-Control Studies

In a case-control study the investigator chooses how many cases and controls to include, so the column totals of the table are not groups of people at risk, and neither the risk of disease in each exposure group nor the risk ratio can be calculated. The OR still works because of its symmetry: it gives the same value whether you compute the odds of disease given exposure or the odds of exposure given disease. By this we mean that the table can be read in two directions. Reading down the columns compares the odds of disease in exposed and non-exposed people, which is the question we care about, while reading across the rows compares the odds of exposure in cases and controls, which is the only comparison a case-control study measures directly. Both readings reduce to the same cross-product, (a1 × b0) / (a0 × b1), so the exposure odds ratio that a case-control study measures equals the disease odds ratio. In addition, the cross-product is not altered by the investigator’s choice of how many controls to recruit per case, because doubling the controls, drawn from the same source population, would be expected to double both b1 and b0, and the two changes cancel. Worked Example 6.3 applies Equation 6.2 to a case-control study and shows the symmetry at work, how the result is worded, and what the table cannot tell you.

Worked Example 6.3: The Odds Ratio from a Case-Control Study

A health unit compares 200 people with laboratory-confirmed salmonellosis (cases) with 400 people from the same community without it (controls), and asks each about eating raw or undercooked eggs in the week before illness (hypothetical data). The table keeps the course layout: exposure in the columns, cases and controls in the rows. The row totals, 200 cases and 400 controls, were fixed by the investigator.

Ate raw eggsDid notTotal
Casesa1 = 60a0 = 140m1 = 200
Controlsb1 = 60b0 = 340m0 = 400
Totaln1 = 120n0 = 480n = 600
  • OR = (a1 × b0) ÷ (a0 × b1) = (60 × 340) ÷ (140 × 60) = 20,400 ÷ 8,400 = 2.43.
  • Read along the rows, the odds of exposure are 60 ÷ 140 = 0.429 among cases and 60 ÷ 340 = 0.176 among controls, and 0.429 ÷ 0.176 = 2.43. This is the symmetry of the odds ratio: the exposure odds ratio, which the study measures, equals the disease odds ratio.

Interpretation. The odds of having eaten raw eggs were 2.43 times as high among people with salmonellosis as among controls; equivalently, the odds of salmonellosis were 2.43 times as high among people who ate raw eggs.

What it cannot say. The investigator chose one case for every two controls, so 60 ÷ 120 is not the risk of salmonellosis among people who ate raw eggs, and no risk ratio or risk difference can be calculated from this table. Salmonellosis is rare in the community (far below 5 per cent), so the OR of 2.43 also approximates the risk ratio and may be described, with that caveat, as about two and a half times the risk. For a common outcome it could not be.

The salmonellosis example brings together both points of this subsection: the cross-product gives the same answer whichever way the table is read, and it may be described as approximately a risk ratio only because the disease is rare.

What the OR estimates also depends on how the controls were chosen, as the third panel in 2.1.1 shows. In cumulative (risk-based) sampling, controls are chosen from the people still free of the disease at the end of the risk period, and the OR then approximates the risk ratio (and the IR) only if the disease is rare. In density (incidence-density) sampling, a control is chosen from the people still at risk at the moment each case occurs, and the OR then estimates the incidence rate ratio directly, whether or not the disease is rare. For example, if a study of a common workplace infection picks, each time a worker falls ill, a control from the workers who are still well that day, its OR can be read as an IR even though the infection is not rare. Whichever ratio is reported, its size answers a different question from the P-value that usually accompanies it.

6.1 What a Measure of Association Tells You

Measures of association assess the magnitude of the relationship between an exposure (a potential cause) and a disease. A single frequency, such as 10 per cent of exposed people developing the disease, cannot be judged on its own; it becomes informative only when it is set beside the frequency in people who were not exposed, and a measure of association makes that comparison precise. Unlike measures of statistical significance, which are heavily dependent on sample size, measures of association indicate the strength of the effect, that is, how much more (or less) likely disease is in exposed compared with non-exposed groups (Tripepi et al., 2007). The distinction is worth committing to memory.

Strength Versus Significance

A measure of association tells you how strongly an exposure is linked to disease. A P-value tells you how likely the observed data would be under the null hypothesis of no association. A strong association can be non-significant (small sample), and a weak association can be highly significant (large sample). Always report both.

Section 4 returns to this distinction when it introduces hypothesis tests and confidence intervals for each of the measures in this lesson.

This section has set out the layouts for risk and rate data, defined the risk ratio, the odds ratio and the incidence rate ratio, shown that the OR lies furthest from the null, approximates the RR only when the outcome is rare and is the measure for case-control studies, and separated the strength of an association from its statistical significance. The key takeaways below summarise these points, the knowledge check tests them, and Section 2 turns from ratio measures to difference measures.

Key Takeaways

  • 1.1 Every calculation in this lesson starts from the course’s 2×2 layout, so set your table up the same way each time: exposure in the columns (exposed, then non-exposed) and disease in the rows (diseased, then non-diseased). The cells are a1 and a0 for exposed and non-exposed people with the disease and b1 and b0 for those without it, with column totals n1 and n0 and row totals m1 and m0 (Table 6.1). For rate data the second row holds the person-time at risk, t1 and t0, instead of non-diseased people (Table 6.2), because a rate divides cases by time rather than by people; in a case-control study the rows are cases and controls. Before you calculate anything, check whether the question gives you counts of people or person-time, and check the order of the groups, since a question may describe them in a different order from the layout. Writing the symbol beside each number is the quickest way to make sure the right count goes into each formula.
  • 1.2 The three ratio measures share one idea: divide a measure of disease frequency in the exposed group by the same measure in the non-exposed group. The risk ratio divides risks, RR = (a1/n1) / (a0/n0); the incidence rate ratio divides rates, IR = (a1/t1) / (a0/t0); and the odds ratio divides odds, which simplifies to the cross-product OR = (a1 × b0) / (a0 × b1). Each is read against the null value of 1: above 1 means more disease in the exposed group and below 1 means less. A good interpretation names the measure, its size, both groups and the time frame, for example “the risk of a low-birth-weight baby was 1.60 times as high among women who smoked as among women who did not.” A protective ratio is easiest to explain as a percentage reduction, 1 − RR, so the cistern RR of 0.58 means the risk of diarrhea was 42 per cent lower among those with a cistern. Check that the exposed group is in the numerator, and say “times the odds” rather than “times the risk” for an OR unless the rare-disease condition (2.1) applies.
  • 2.1 The odds ratio has a special place because of how case-control studies are built. The investigator chooses how many cases and controls to include, so the table cannot give the risk of disease in either exposure group, and no risk ratio or risk difference can be calculated (in the salmonellosis example, 60 ÷ 120 is not a risk). The OR still works because it is symmetric: the odds of exposure in cases compared with controls equal the odds of disease in exposed compared with non-exposed people, so one cross-product answers both questions. For the same data, IR values generally lie further from the null than RR values, and OR values further still (Figure 6.1), which is why the cistern OR (0.53) is more extreme than its RR (0.58) when the outcome is common. The OR approximates the RR only when the disease is rare in the source population (a risk below about 5 per cent), and if controls are chosen by density sampling it estimates the IR directly. So before you describe an OR as “about X times the risk”, check how common the outcome is and say that you are relying on the rare-disease assumption.
  • 6.1 A measure of association tells you how strongly an exposure and a disease are linked, that is, how much more (or less) common the disease is in the exposed group than in the non-exposed group. A P-value answers a different question: how likely data like yours would be if there were no association at all, and it depends heavily on the size of the study. This is why a strong association can be non-significant in a small study, while a weak one can be highly significant in a very large study. When you work through an example, report the measure of association as the answer to “how strong is the link?”, and treat the P-value as the answer to a separate question about chance. A small P-value never, on its own, tells you that an association is strong.
Knowledge Check: this section

1. A hospital follows its 1,000 kitchen staff through one summer. Of the 250 staff who handled raw poultry, 20 developed Campylobacter infection; of the 750 who did not handle raw poultry, 15 did. The data are set out in the course’s 2×2 layout (exposure in the columns, disease in the rows). Which set of cells and risk ratio is correct?

This reads exposure along the rows. In Table 6.1 exposure runs down the columns, so a0 is the 15 cases among non-handlers, and b1 and b0 hold the non-cases in each column (230 and 735).
This puts the column totals in the Non-diseased row. b1 counts the 230 handlers who stayed well, not all 250 handlers; the column totals n1 = 250 and n0 = 750 sit below the table.
Exposure runs down the columns and disease along the rows, so the cases are a1 = 20 and a0 = 15 and the non-cases b1 = 230 and b0 = 735. RR = (20/250) ÷ (15/750) = 0.08 ÷ 0.02 = 4.0.
The cells are right, but 0.25 divides the non-exposed risk by the exposed risk, which reverses the reference group. The RR puts the exposed risk on top: 0.08 ÷ 0.02 = 4.0.
Correct answer: a1 = 20, a0 = 15, b1 = 230, b0 = 735 and RR = 4.0. In Table 6.1 the exposed staff (poultry handlers) form the first column and the non-exposed staff the second; the Diseased row holds the cases (a1 = 20, a0 = 15) and the Non-diseased row holds everyone else in each column (b1 = 250 − 20 = 230, b0 = 750 − 15 = 735), so the column totals are n1 = 250 and n0 = 750. Then RR = (a1/n1) ÷ (a0/n0) = (20/250) ÷ (15/750) = 0.08 ÷ 0.02 = 4.0: the summer risk of infection was four times as high among staff who handled raw poultry. The wrong options put the column totals in the Non-diseased row (b1 counts the 230 non-cases, not all 250 handlers), swap a0 and b1 by reading exposure along the rows, or divide the non-exposed risk by the exposed risk (0.25), which reverses the reference group. (Section 1, 1.1 Setting Out the Data in a 2×2 Table, Table 6.1, and 1.2.1 The Risk Ratio and the Odds Ratio, Eq 6.1; Objectives 1.1 and 1.2)

2. In a cohort of adults with type 2 diabetes, current smokers had 18 new foot ulcers during 1,200 person-years of follow-up, and non-smokers had 20 new foot ulcers during 4,000 person-years. Which answer gives the incidence rate ratio (IR) and interprets it correctly?

Dividing the case counts (18 ÷ 20) ignores person-time: non-smokers were observed for more than three times as long. Each group's cases must first be divided by its own person-time.
Each rate divides new cases by person-time at risk: 18 ÷ 1,200 = 0.015 and 20 ÷ 4,000 = 0.005 per person-year. IR = 0.015 ÷ 0.005 = 3.0, a comparison of rates.
The value is right, but the IR compares rates per unit of person-time, not numbers of people. In fact fewer smokers (18) than non-smokers (20) developed an ulcer.
0.010 per person-year is the incidence rate difference, the absolute measure, not the IR. The rate ratio is 0.015 ÷ 0.005 = 3.0 and has no units.
Correct answer: IR = 3.0, read as a comparison of rates. Each rate divides the new cases by the person-time at risk (Table 6.2): smokers 18 ÷ 1,200 = 0.015 per person-year (15 per 1,000 person-years) and non-smokers 20 ÷ 4,000 = 0.005 per person-year (5 per 1,000 person-years), so IR = (a1/t1) ÷ (a0/t0) = 0.015 ÷ 0.005 = 3.0. Dividing the case counts (18 ÷ 20 = 0.90) ignores that the non-smokers were observed for more than three times as much person-time, and the IR does not say that more smokers had ulcers (fewer did). The value 0.010 per person-year is the incidence rate difference, the absolute measure of Section 2, not the IR. (Section 1, 1.2.2 The Incidence Rate Ratio, Eq 6.3 and Worked Example 6.2; Objective 1.2)

3. A health unit compares 100 adults with Legionnaires’ disease (cases) with 200 adults from the same community without it (controls), and asks each whether they used a hotel hot tub in the two weeks before illness (or before the interview, for controls). Legionnaires’ disease is rare in the community. The results are in the table below. Which statement about the measure of association is correct?

Used a hot tubDid notTotal
Cases4060100
Controls50150200
Total90210300
OR = (a1 × b0) ÷ (a0 × b1) = 6,000 ÷ 3,000 = 2.0. The investigator fixed the numbers of cases and controls, so no risk can be calculated; the OR approximates the RR here only because the disease is rare.
The column totals are not groups of people at risk, because the investigator chose two controls for every case. 40 ÷ 90 is therefore not a risk, and a case-control study cannot give a risk ratio.
The OR lies further from 1 than the RR, not closer. Because Legionnaires' disease is rare, the OR of 2.0 approximates the RR, so the RR would be close to 2.0, not larger.
The inverted cross-product, 0.50, is the OR for not using a hot tub. With hot-tub use as the exposure, the OR is (a1 × b0) ÷ (a0 × b1) = 2.0.
Correct answer: the first statement. In the course layout a1 = 40, a0 = 60, b1 = 50 and b0 = 150, so OR = (a1 × b0) ÷ (a0 × b1) = 6,000 ÷ 3,000 = 2.0; read along the rows, the odds of hot-tub use were 40/60 = 0.67 among cases and 50/150 = 0.33 among controls, the same ratio, which is the symmetry that lets a case-control study estimate the OR. The tempting “risk” 40 ÷ 90 is meaningless, because the investigator chose two controls for every case, so the column totals are not groups of people at risk. The OR lies further from 1 than the RR, not closer, and it approximates the RR here only because the disease is rare (well under 5%); the inverted cross-product, 0.50, is the OR for not using a hot tub. (Section 1, the slide “Why OR is the case-control measure”, 2.1.1 When the Three Ratios Agree and the panel “When is OR ≈ RR?”, and 2.1.2 The Odds Ratio in Case-Control Studies and Worked Example 6.3; Objective 2.1)

4. A study of 250,000 adults finds that daily energy-drink use is associated with frequent headaches, with a risk ratio of 1.08 (P < 0.001). A news release says: “The link was highly significant, so energy drinks have a strong effect on headaches.” Which response is best?

Strength is read from the measure of association, not the P-value. An RR of 1.08 is weak; P < 0.001 reflects the very large sample and shows only that chance is an unlikely explanation.
A weak association can be statistically significant when the sample is large, because the P-value shrinks as sample size grows. With 250,000 people an RR of 1.08 can easily give P < 0.001.
A P-value is how surprising the data would be if there were no association. It is not the probability that the exposure causes the outcome.
RR = 1.08 means an 8% higher risk, a weak association. The P-value answers a separate question, how surprising the data would be with no association, and it shrinks as the sample grows.
Correct answer: the last option. The strength of an association is read from the measure of association: RR = 1.08 means that the risk of frequent headaches was only 1.08 times as high (8% higher) among daily users. The P-value answers a different question, how surprising the data would be if there were no association, and it shrinks as the sample grows, so with 250,000 people even a weak association gives P < 0.001. That is why the release is wrong: a small P-value is not evidence of a strong effect, and it is not the probability that the exposure causes the outcome. (Section 1, 6.1 What a Measure of Association Tells You and the callout “Strength Versus Significance”; Objective 6.1)

✦ Pass the knowledge check with 100% to continue

Section 2

Measures of Effect in the Exposed Group

⏱ Estimated reading time: 15 minutes

Section 2 of 4

Measures of Effect in the Exposed Group

Risk difference, attributable fraction in the exposed, and vaccine efficacy as a special case.

Risk difference

RD: the absolute increase in risk

Risk Difference (RD) / Attributable Risk
\[ \color{#0B7B6B}{RD} = \color{#C2410C}{R_{E+}} - \color{#1D4ED8}{R_{E-}} = \frac{\color{#C2410C}{a_1}}{a_1+b_1} - \frac{\color{#1D4ED8}{a_0}}{a_0+b_0} \]
RD risk differenceRE+ risk in the exposedRE− risk in the unexposed
Incidence Rate Difference (ID)
\[ \color{#0B7B6B}{ID} = \frac{\color{#C2410C}{a_1}}{\color{#C2410C}{T_1}} - \frac{\color{#1D4ED8}{a_0}}{\color{#1D4ED8}{T_0}} \]
ID rate differencea1/T1 rate in the exposeda0/T0 rate in the unexposed

Null value: 0. RD > 0 means excess risk in the exposed; RD < 0 means a protective effect.

Worked example

Smoking and low birth weight

Smokers: 40/351 = 0.114
Non-smokers: 331/4,649 = 0.071

RD
\[ RD = 0.114 - 0.071 = 0.043 \]

Per 100 women who smoked, approximately 4.3 additional low-birth-weight babies attributable to smoking.

Why RD matters

A risk ratio of 1.60 sounds substantial. An RD of 4.3 per 100 tells you the actual scale of the problem when planning an intervention.

Attributable fraction

AFe: proportion of exposed cases due to exposure

Attributable Fraction in the Exposed
\[ \color{#0B7B6B}{AF_e} = \frac{\color{#1D4ED8}{RR} - 1}{\color{#1D4ED8}{RR}} = \frac{\color{#C2410C}{RD}}{\color{#6D28D9}{R_{E+}}} \]
AFe attributable fraction in the exposedRR risk ratioRD risk differenceRE+ risk in the exposed

Ranges from 0 (when RR = 1, no excess) to 1 (when all disease in the exposed is due to the exposure).

From the smoking example:

\[ AF_e = \frac{1.60 - 1}{1.60} = \frac{0.60}{1.60} \approx 0.375 \]

37.5% of low-birth-weight cases among smoking women were attributable to their smoking.

Special case

Vaccine efficacy as AFe

Unvaccinated (exposure+): 20% develop disease
Vaccinated (exposure-): 5% develop disease

Vaccine Efficacy
\[ \color{#0B7B6B}{VE} = AF_e = \frac{\color{#C2410C}{RD}}{\color{#6D28D9}{R_{E+}}} = \frac{0.15}{0.20} = 0.75 \]
VE vaccine efficacyRD risk difference (unvaccinated vs vaccinated)RE+ risk in the unvaccinated

Plain language

The vaccine prevented 75% of the cases that would have occurred in vaccinated individuals had they remained unvaccinated.

Carry forward

Relative vs. absolute measures

Ratio measures (section 1)

Relative strength of association. Tend to be stable across different baseline risks. Used in clinical risk communication.

Difference measures (section 2)

Absolute excess risk and proportional causal share. Needed for resource allocation and policy prioritisation.

A later section takes these ideas and asks: what is the impact on the whole population, not just on those who are exposed?

Overview

Section 1 covered the three ratio measures (RR, IR, OR), which say how many times more likely disease is in the exposed group than in the unexposed group. This section turns to the parallel set of difference measures, which answer a different question: not how many times more, but how many extra cases occur because of the exposure. It first contrasts the two kinds of measure and the scales on which they work, then defines and calculates the risk difference and the incidence rate difference and turns the risk difference into the numbers needed to harm and to treat. It closes by expressing the excess risk as a share of the disease in the exposed group, the attributable fraction in the exposed, and by showing that vaccine efficacy is a special case of that fraction.

Learning Objectives

  • 3.1 Distinguish between “ratio” (relative) and “difference” (absolute) measures of association, and the question each answers.
  • 3.2 Calculate and interpret the risk difference (RD) and incidence rate difference (ID).
  • 3.3 Calculate and interpret the number needed to harm (NNH) and the number needed to treat (NNT).
  • 4.1 Calculate and interpret the attributable fraction in the exposed (AFe), stating the causal assumption.
  • 4.2 Calculate vaccine efficacy and explain it as a special case of AFe.

3.1 Ratio and Difference Measures

The ratio measures from Section 1 (RR, IR, OR) tell us the relative strength of association, but they do not indicate the absolute number of cases attributable to the exposure. By this we mean that a ratio tells you how many times larger one risk is than another, but not how large either risk is. For example, a risk ratio of 2 could come from risks of 2 per 100,000 and 1 per 100,000, or from risks of 20 per 100 and 10 per 100; the ratio is identical, yet the second situation produces ten thousand times as many extra cases. Difference (absolute effect) measures address this gap by subtracting one risk from the other, which shows how many additional cases occur because of the exposure. Because the two kinds of measure answer different questions, the choice between ratio and difference measures is itself a substantive scientific decision rather than a statistical convenience (Greenland & Pearce, 2015; Tripepi et al., 2007).

It helps to name the two scales this contrast lives on. Ratio measures work on a multiplicative scale: a risk ratio of 2 says the exposed risk is the baseline risk multiplied by two. Difference measures work on an additive scale: a risk difference of 0.04 says the exposed risk is the baseline plus four percentage points. The same association can look striking on one scale and slight on the other, so the choice of scale is a real part of the scientific question rather than a matter of presentation. The callout below states why public health needs both.

Why Both Matter

Even when an exposure is very strongly associated with disease (high RR), if the exposure is rare in a population, it may contribute very few cases. Conversely, a relatively weak risk factor (modest RR) that is common can be responsible for many cases. Difference measures capture this “public health impact.”

Ratio measures describe the relative strength that a clinician leans on when explaining risk to a patient, while difference measures give the case counts that a planner needs when deciding where to spend resources. The smoking cohort used throughout this section shows the contrast: a risk ratio of 1.60 sounds substantial, but a risk difference of 4.3 extra low-birth-weight babies per 100 women who smoked shows the actual scale of the problem when an intervention is planned. The first difference measure, the risk difference, is computed from the same 2×2 table as the risk ratio.

3.2 The Risk Difference and the Incidence Rate Difference

3.2.1 The Risk Difference

The risk difference (RD), also called attributable risk (Walter, 1976), is simply the risk in the exposed group minus the risk in the non-exposed group. Both risks are read from the 2×2 table introduced in Section 1, which is repeated below as Table 6.1. In the repeated table the exposed column is shown in orange and the non-exposed column in blue, the same colours that Equation 6.4 uses for the two risks, and a final row gives the risk of disease computed from each column.

Table 6.1 (repeated from Section 1). Layout of the 2×2 table for incidence risk data, with a row added that gives the risk of disease in each exposure group. Each cell shows its symbol and what it counts.

Exposed (E+)Non-exposed (E−)Total
Diseased (D+)a1exposed people with the diseasea0non-exposed people with the diseasem1all people with the disease
Non-diseased (D−)b1exposed people without the diseaseb0non-exposed people without the diseasem0all people without the disease
Totaln1 = a1 + b1all exposed peoplen0 = a0 + b0all non-exposed peopleneveryone in the study
Risk of diseasea1/n1 = p(D+|E+)a0/n0 = p(D+|E−)m1/n = p(D+)

Equation 6.4 writes the risk difference first in probability notation and then in terms of the cells of Table 6.1. The key beneath the equation defines each term and identifies the cell of the table from which it comes.

Risk difference
RD = p(D+|E+) − p(D+|E−) = (a1/n1) − (a0/n0) Eq 6.4
The risk difference is the risk in the exposed group minus the risk in the unexposed group.

where

RD
is the risk difference, the risk of disease in the exposed group minus the risk in the non-exposed group. Because it is a difference between two proportions it has no units, and it is usually reported in percentage points or as extra cases per 100 exposed people over the follow-up period of the study.
p(D+|E+)
is the probability of disease (D+) given exposure (E+), which is the risk of disease in the exposed group. The vertical bar is read as “given”. Worked Example 6.4 writes this risk as RE+.
p(D+|E−)
is the probability of disease given no exposure (E−), which is the risk of disease in the non-exposed group. Worked Example 6.4 writes this risk as RE−.
a1
is the number of exposed people who developed the disease, the cell in the Diseased row and Exposed column of Table 6.1.
n1
is the total number of exposed people, the cell in the Total row and Exposed column. It equals a1 + b1, where b1 is the number of exposed people who did not develop the disease.
a0
is the number of non-exposed people who developed the disease, the cell in the Diseased row and Non-exposed column.
n0
is the total number of non-exposed people, the cell in the Total row and Non-exposed column. It equals a0 + b0, where b0 is the number of non-exposed people who did not develop the disease.

In the cell labels, the subscript 1 marks the exposed group and the subscript 0 marks the non-exposed group. The fraction a1/n1 is therefore the risk in the exposed group, and a0/n0 is the risk in the non-exposed group.

Because the risk difference subtracts rather than divides, its null value is 0, not 1, and it is usually reported as extra cases per 100 exposed people over the follow-up period. When the groups contribute different amounts of follow-up time, the same subtraction is made with incidence rates.

3.2.2 The Incidence Rate Difference

The incidence rate difference (ID) is the difference between two incidence rates. Each rate divides the number of new cases in an exposure group by the person-time at risk that the group contributed. The counts and the person-time are read from the rate layout introduced in Section 1, which is repeated below as Table 6.2 with the same colour coding and with a row added for the incidence rate in each group.

Table 6.2 (repeated from Section 1). Layout of the table for incidence rate data, with a row added that gives the incidence rate in each exposure group. Each cell shows its symbol and what it measures.

Exposed (E+)Non-exposed (E−)Total
Number of casesa1new cases among exposed peoplea0new cases among non-exposed peoplem1all new cases
Person-time at riskt1person-time contributed by exposed peoplet0person-time contributed by non-exposed peopletall person-time
Incidence ratea1/t1a0/t0m1/t

Equation 6.5 subtracts the incidence rate in the non-exposed group from the incidence rate in the exposed group. Its null value is also 0, but unlike the RD it keeps the units of the two rates. The key beneath the equation defines each term in the same way as the key for Equation 6.4.

Incidence rate difference
ID = (a1/t1) − (a0/t0) Eq 6.5
The rate difference is the rate in the exposed minus the rate in the unexposed, each a count of cases over person-time.

where

ID
is the incidence rate difference, the incidence rate in the exposed group minus the incidence rate in the non-exposed group. It keeps the units of the two rates, so it is reported as extra cases per unit of person-time, for example cases per 100 person-months.
a1/t1
is the incidence rate in the exposed group, the bottom row of the Exposed column in Table 6.2.
a1
is the number of new cases of disease that occurred among exposed people during follow-up, the cell in the Number of cases row and Exposed column.
t1
is the person-time at risk contributed by the exposed group, the sum of the time that each exposed person was observed while still free of the disease, measured in a unit such as person-months or person-years.
a0/t0
is the incidence rate in the non-exposed group, the bottom row of the Non-exposed column.
a0
is the number of new cases of disease that occurred among non-exposed people during follow-up.
t0
is the person-time at risk contributed by the non-exposed group, measured in the same unit as t1.

The two rates must be expressed in the same unit of person-time before they are subtracted. The slide summary at the start of this section writes person-time with a capital letter, as T1 and T0, and these symbols have the same meaning as t1 and t0.

With both formulae in place, the next step is to read and apply them.

3.2.3 Interpreting and Applying the Two Differences

The three tabs below interpret and apply Equation 6.4 and Equation 6.5; open each one in turn. The first tab explains how the sign of a difference measure is read. The second holds Worked Example 6.4, which applies Equation 6.4 to a cohort of 5,000 women followed through pregnancy, with smoking as the exposure and low birth weight as the outcome; watch how the unitless RD is turned into extra cases per 100 women. The third holds Worked Example 6.5, which applies Equation 6.5 to the migraine data of Worked Example 6.2 and keeps the units of the rates.

Interpretation of Difference Measures

  • RD or ID < 0 → Exposure is protective
  • RD or ID = 0 → No effect of exposure
  • RD or ID > 0 → Exposure is positively associated with disease

RD indicates the increase (or decrease) in the probability of disease in the exposed group, beyond the baseline risk. The baseline risk is the risk in the non-exposed group, which is the risk the exposed people would be expected to have if they had not been exposed. By this we mean that the RD separates the cases that would have happened anyway from the extra cases that come with the exposure. For example, if 7 in 100 non-exposed people and 11 in 100 exposed people become ill, then about 7 of the 11 exposed cases would have occurred regardless, and the RD of 0.04 describes the remaining 4. It tells you: “For every X exposed individuals, how many additional cases occur because of the exposure?”

Worked Example 6.4: Smoking & Low Birth Weight

From a cohort of 5,000 women followed through pregnancy:

SmokerNon-smokerTotal
Low birth weight40331371
Normal birth weight3114,3184,629
Total3514,6495,000
  • Risk in exposed: RE+ = 40/351 = 0.114
  • Risk in non-exposed: RE− = 331/4649 = 0.071
  • RD = 0.114 − 0.071 = 0.043

For every 100 women who smoked, approximately 4.3 had a low-birth-weight baby due to the fact that they smoked (assuming causal relationship).

Worked Example 6.5: Migraine Incidence Rates

Using the migraine data from the previous section, 131 cases occurred among women over 250 person-months and 44 cases among men over 236 person-months.

  • Rate in women: 131/250 = 0.524 cases per person-month.
  • Rate in men: 44/236 = 0.186 cases per person-month.
  • ID = 0.524 − 0.186 = 0.338 cases per person-month.

Women aged 30 to 40 experienced about 34 more cases of migraine per 100 person-months of observation than men of the same age. The rate ratio of 2.81 from the previous section describes the same data on a relative scale.

Together the tabs show the RD as an excess of 4.3 cases per 100 exposed women and the ID as an excess of about 34 cases per 100 person-months, each carrying the assumption that the association is causal; the ID and the IR of 2.81 describe the same migraine data on the additive and the multiplicative scales. Calculator 6.4 and Calculator 6.5 reproduce the two worked examples, so that the effect of changing the counts on each difference can be seen.

The risk difference can also be turned upside down, which gives a number that is often easier to communicate.

3.3 Numbers Needed to Harm and to Treat

The reciprocal of a risk difference, 1 ÷ RD, is the number of people who must be exposed for one additional case to occur, the number needed to harm (NNH). For a protective exposure or intervention the same calculation, with the minus sign of the RD dropped, gives the number needed to treat (NNT), the number of people who must be treated for one case to be avoided. The reciprocal works because the RD is a number of extra cases per person. For example, an RD of 0.05 means 5 extra cases for every 100 people exposed, which is the same as 1 extra case for every 20 people, and 1 ÷ 0.05 = 20. Turning the RD around in this way changes a small decimal that is hard to picture into a whole number of people, which is why clinicians and patients often find it easier to understand. Worked Example 6.6 takes the reciprocal of the risk difference from Worked Example 6.4 to obtain the number needed to harm, and applies the same calculation to a protective intervention to obtain the number needed to treat; look for the two qualifications that every such number needs.

Worked Example 6.6: Number Needed to Harm and Number Needed to Treat

The reciprocal of a risk difference gives the number of people who must be exposed for one additional case to occur. In the smoking example, RD = 0.043, so the number needed to harm is 1 ÷ 0.043 = 23.3: about 23 women who smoke for each additional low-birth-weight baby, assuming a causal relationship.

For a protective intervention the same calculation gives the number needed to treat. If, for example, an intervention lowered the risk of an outcome from 0.20 to 0.15, RD = −0.05 and NNT = 1 ÷ 0.05 = 20 people treated for one outcome avoided.

Both numbers apply to the follow-up period of the study from which the risks come, here the pregnancy, so each is quoted with its time frame.

Both numbers also assume that the risk difference is caused by the exposure or the intervention. When people choose for themselves whether to be exposed or treated, as smokers do or as volunteers for a programme do, the two groups may differ in other causes of the outcome. Part of the risk difference may then be confounding. If confounding has exaggerated the risk difference, the NNT (or NNH) is too small, and the benefit (or harm) is overstated.

The worked example shows those two qualifications: the time frame of the risks from which the number was calculated, and the assumption that the risk difference was caused by the exposure or the intervention. The risk difference and its reciprocal describe the excess risk in absolute terms, and the next subsection expresses the same excess as a proportion of the risk in the exposed group.

4.1 The Attributable Fraction in the Exposed

The AFe (also called the attributable fraction among the exposed) expresses the proportion of disease in exposed individuals that is due to the exposure, assuming the relationship is causal (Walter, 1976). It can be viewed as the proportion of disease in the exposed group that would be avoided if the exposure were removed. By this we mean that the AFe starts with all the cases among exposed people and asks what share of them are the extra cases described by the risk difference, rather than the cases that would have occurred anyway at the baseline risk. For example, if the risk is 0.10 in the exposed group and 0.04 in the non-exposed group, then 0.06 of the 0.10 is excess, and the AFe is 0.06 ÷ 0.10 = 0.60, or 60 per cent of the exposed cases. The causal assumption matters because the calculation simply divides one number by another; only if the exposure really causes the excess can that excess be described as cases that removing the exposure would prevent. Equation 6.6 gives the AFe in three forms: from the risk difference and the risk in the exposed, from the risk ratio, and approximately from the odds ratio.

Attributable fraction in the exposed
AFe = RD / p(D+|E+) = (RR − 1) / RR ≈ (OR − 1) / OR Eq 6.6
The attributable fraction in the exposed is the risk difference over the risk in the exposed; equivalently it can be written from the risk ratio, or approximated from the odds ratio.

AFe ranges from 0 (where risk is equal, RR = 1) to 1 (where all disease in the exposed group is due to the exposure, RR = ∞). In case-control studies, which cannot supply a risk ratio, AFe can be approximated by substituting OR for RR. This substitution works only when the disease is rare (a risk below about 5 per cent); for a common outcome the OR lies further from 1 than the RR, so an AFe calculated from the OR is too large. Worked Example 6.7 applies Equation 6.6 to the smoking cohort of Worked Example 6.4, first through the risk ratio and then through the risk difference.

Worked Example 6.7: AFe for Smoking

From the smoking example above:

  • RR = 0.114 / 0.071 = 1.60
  • AFe = (1.60 − 1) / 1.60 = 0.60 / 1.60 = 0.375 (37.5%)

Among women who smoked, 37.5% of the low-birth-weight cases were attributable to smoking. Alternatively: 0.043 / 0.114 = 0.377 ≈ 37.7% (slight rounding difference).

The two routes agree apart from rounding, so one makes a useful check on the other, and the interpretation describes cases among the exposed women only. Calculator 6.6 reproduces the calculation from the two risks, from the RR or from the OR.

A related quantity, the etiologic fraction, is easily confused with the AFe; the callout below sets out the difference.

AFe Versus the Etiologic Fraction

The etiologic fraction is the proportion of cases in the exposed group for which exposure was a component of the sufficient cause (Rothman, 1976). While AFe measures the excess fraction, that is, the share of exposed cases that would not have occurred at all without the exposure, the etiologic fraction can be higher because exposure may contribute to cases even when the baseline risk would have produced them eventually. By this we mean that an exposure can bring a case forward in time as well as create a new one. For example, a person who would have developed the disease at age 70 might, because of the exposure, develop it at age 60. That person counts as a case in both groups by the end of a long follow-up, so they add nothing to the excess, yet the exposure was still part of the cause of their disease. Because the AFe counts only the excess cases and misses cases like this one, in general, AFe provides a lower bound for the etiologic fraction (Greenland & Robins, 1988).

The same fraction, with the exposure reframed, also measures how well a vaccine works.

4.2 Vaccine Efficacy

Vaccine efficacy is a special form of AFe, where “not vaccinated” is the exposure (factor positive) and “vaccinated” is the comparison group. The reason is that the AFe describes the share of cases in the higher-risk group that a factor accounts for. Without the vaccine the risk is higher, so treating “not vaccinated” as the exposure gives a risk ratio above 1 and lets the AFe formula be used unchanged, with the risk in the unvaccinated group as the denominator; the result then reads as the share of cases that vaccination prevents among the vaccinated. Equation 6.7 writes it with Ru, the risk in the unvaccinated group, and Rv, the risk in the vaccinated group. Its first form is the AFe of Equation 6.6, the risk difference divided by the risk in the “exposed” (unvaccinated) group, and its second form shows that it equals one minus the risk ratio comparing vaccinated with unvaccinated people.

Vaccine efficacy
VE = (Ru − Rv) / Ru = 1 − Rv / RuEq 6.7
Vaccine efficacy is the risk in the unvaccinated group minus the risk in the vaccinated group, divided by the risk in the unvaccinated group; equivalently, it is one minus the risk ratio comparing vaccinated with unvaccinated people.

Worked Example 6.8 applies Equation 6.7 to a simple case in which 20% of unvaccinated individuals develop disease versus 5% of vaccinated individuals; notice which risk goes in the denominator.

Worked Example 6.8: Vaccine Efficacy Calculation

  • RD = 0.20 − 0.05 = 0.15
  • AFe = 0.15 / 0.20 = 0.75 (75%)

The vaccine has prevented 75% of the cases of disease that would have occurred in the vaccinated group if the vaccine had not been used.

A vaccine efficacy of 75% is a proportional reduction in risk, the share of cases prevented among the vaccinated, and not the share of vaccinated people who are immune; a common error is to read it the second way. Calculator 6.7 reproduces the worked example.

This section has defined the risk difference and the incidence rate difference, the numbers needed to harm and to treat, and the attributable fraction in the exposed, with vaccine efficacy as a special case. The key takeaways and the knowledge check below consolidate the section, including the calculation of these measures for a single exposure. Section 3 moves from the exposed group to the whole population.

Key Takeaways

  • 3.1 Ratio and difference measures describe the same data on two different scales, and each answers a different question. Ratio measures (RR, IR, OR) work on a multiplicative scale and tell you how many times more likely disease is in the exposed group, the relative strength that a clinician leans on when explaining risk to a patient. Difference measures (RD, ID) work on an additive scale and tell you how many extra cases the exposure produces, which is what a planner needs when deciding where a health system should spend its money. The same association can look striking on one scale and slight on the other: in the smoking cohort an RR of 1.60 sounds substantial, while an RD of 4.3 per 100 shows the actual scale of the problem. When a question asks which measure suits a decision-maker, match relative measures to questions about strength and individual risk, match absolute measures to questions about burden, case counts and resources, and give your reason.
  • 3.2 The risk difference, which this lesson also calls the attributable risk, is the risk in the exposed minus the risk in the non-exposed, RD = (a1/n1) − (a0/n0), and the incidence rate difference does the same with rates, ID = (a1/t1) − (a0/t0). Because you are subtracting rather than dividing, the null value is 0, not 1: a positive value means extra cases in the exposed group and a negative value means the exposure is protective. An RD has no units, so translate it into extra cases per 100 (or per 1,000) exposed people over the follow-up period, as in “for every 100 women who smoked, about 4.3 additional low-birth-weight babies occurred during pregnancy, assuming the association is causal.” An ID keeps the units of the rates, so it must carry them, as in about 34 extra migraine cases per 100 person-months. Convert both rates to the same person-time unit before you subtract, and do not confuse the attributable risk (a difference) with the attributable fraction (a proportion).
  • 3.3 Turning the risk difference upside down gives a number that is often easier to communicate. The number needed to harm, NNH = 1 ÷ RD, is how many people must be exposed for one additional case to occur; for a protective exposure or intervention, the same calculation gives the number needed to treat (NNT), how many people must be treated for one case to be avoided. In the smoking cohort 1 ÷ 0.043 ≈ 23 women who smoke for each additional low-birth-weight baby, and an intervention that lowers the risk from 0.20 to 0.15 (RD = −0.05) has an NNT of 1 ÷ 0.05 = 20. Drop the minus sign of a protective RD before taking the reciprocal, quote the result with the study’s time frame, and state that it assumes the risk difference is caused by the exposure or treatment. If people chose for themselves whether to be exposed or treated, part of the RD may be confounding, which makes the NNT (or NNH) too small and overstates the benefit (or harm).
  • 4.1 The attributable fraction in the exposed takes the extra cases in the exposed group and expresses them as a share of all the cases in that group: of the exposed people who developed the disease, what proportion would not have developed it without the exposure? It can be calculated as AFe = RD / RE+ or, equivalently, as (RR − 1) / RR, and in a case-control study, where no RR is available, you substitute the OR, (OR − 1) / OR. For the smoking cohort (1.60 − 1) / 1.60 = 0.375, so about 37.5 per cent of the low-birth-weight cases among women who smoked were attributable to smoking, and the other route, 0.043 / 0.114 = 0.377, makes a useful check (small rounding differences are expected). Your interpretation must include “assuming the association is causal”, because the AFe can be read as the share of cases due to the exposure only if the association is not explained by bias or confounding. Keep the denominator in mind as well: the AFe describes cases among the exposed only, not in the whole population, and it is a lower bound for the etiologic fraction.
  • 4.2 Vaccine efficacy is the attributable fraction in the exposed with the exposure reframed: being unvaccinated is treated as the exposure, and the vaccinated group is the comparison. The formula is VE = (Ru − Rv) / Ru = 1 − Rv/Ru, so if 20 per cent of unvaccinated and 5 per cent of vaccinated people develop the disease, VE = 0.15 / 0.20 = 0.75. Interpret it as the share of cases the vaccine prevented: 75 per cent of the cases that would have occurred among the vaccinated people had they not been vaccinated. A common trap is the claim that a VE of 75 per cent means 75 per cent of vaccinated people are immune; VE is a proportional reduction in risk between the two groups, not the share of vaccinated people who are immune. When you check your working, make sure the risk in the unvaccinated group is the denominator.
Knowledge Check: this section

1. A cohort study finds that the one-year risk of a gastrointestinal bleed is 0.025 among adults who take a certain pain medication every day and 0.010 among adults who do not. A physician asks how much the medication multiplies her patient’s risk; a health planner asks how many extra bleeds to expect per 1,000 daily users in a year. Which pair of answers is correct?

This swaps the two measures. The physician asks a relative question, answered by RR = 2.5; the planner asks an absolute one, answered by RD = 0.015, which is 15 extra bleeds per 1,000 users, not 2.5.
The physician's question is relative, answered by RR = 0.025 ÷ 0.010 = 2.5. The planner's is absolute, answered by RD = 0.015, or 15 extra bleeds per 1,000 users in a year if the medication causes them.
The RR is right, but 25 per 1,000 is the whole risk among users. Extra bleeds subtract the 10 per 1,000 that would have happened anyway, giving 15 per 1,000.
The RD is right, but 1.5 is RR − 1, not the risk ratio. The RR is 0.025 ÷ 0.010 = 2.5.
Correct answer: the RR for the physician and the RD for the planner. The physician’s question is relative (how many times), answered on the multiplicative scale by RR = 0.025 ÷ 0.010 = 2.5; the planner’s question is absolute (how many extra cases), answered on the additive scale by RD = 0.025 − 0.010 = 0.015, and 0.015 × 1,000 = 15 extra bleeds per 1,000 daily users over one year, if the medication causes them. The figure of 25 per 1,000 is the whole risk among users and forgets to subtract the 10 per 1,000 bleeds that would have happened anyway, and 1.5 is RR − 1, not the RR. (Section 2, 3.1 Ratio and Difference Measures and the callout “Why Both Matter”, and 3.2.1 The Risk Difference, Eq 6.4; Objectives 3.1 and 3.2)

2. A city follows its letter carriers for dog bites. Carriers on suburban routes had 28 bites during 7,000 person-months of work, and carriers on downtown routes had 10 bites during 10,000 person-months. Which answer gives the incidence rate difference (ID) and interprets it correctly?

Subtracting counts ignores the different person-time in the two groups (7,000 against 10,000 person-months). The ID subtracts the rates: 0.004 − 0.001 = 0.003 per person-month.
4.0 is the rate ratio, 0.004 ÷ 0.001, a relative measure. The ID is the absolute difference between the rates, 3 bites per 1,000 person-months.
The rates are 28 ÷ 7,000 = 0.004 and 10 ÷ 10,000 = 0.001 per person-month, so ID = 0.003 per person-month. An ID keeps the rates' units and counts excess cases only if the association is causal.
The value 0.003 is right, but an ID keeps the units of the rates: it is 0.003 bites per person-month, an excess per unit of person-time, not a percentage of carriers.
Correct answer: 3 bites per 1,000 person-months. The rate on suburban routes is a1/t1 = 28 ÷ 7,000 = 0.004 bites per person-month (4 per 1,000) and the rate downtown is a0/t0 = 10 ÷ 10,000 = 0.001 (1 per 1,000), so ID = 0.004 − 0.001 = 0.003 per person-month, or 3 per 1,000 person-months. An ID keeps the units of the two rates, so it is an excess per unit of person-time, not a percentage of carriers. Subtracting the counts (28 − 10 = 18) ignores the different person-time of the two groups, and 4.0 is the rate ratio, the relative measure of Section 1. (Section 2, 3.2.2 The Incidence Rate Difference, Eq 6.5, and 3.2.3 Interpreting and Applying the Two Differences, Worked Example 6.5; Objective 3.2)

3. Over one winter, 8% of older adults who received a home safety visit and 13% of older adults who did not had a fall that needed medical care. How many older adults would need a home safety visit to prevent one such fall?

Five percentage points is the risk difference itself. The NNT is its reciprocal: 1 ÷ 0.05 = 20.
This uses only the risk in the visited group. The NNT is the reciprocal of the risk difference between the groups, 1 ÷ (0.13 − 0.08) = 20.
0.13 ÷ 0.08 compares the risks as a ratio. The NNT is based on the absolute difference, 1 ÷ 0.05 = 20.
NNT = 1 ÷ |RD| = 1 ÷ 0.05 = 20, quoted with the time frame of the risks. It assumes the visits caused the difference; confounding would make the true NNT larger.
Correct answer: 20 adults over one winter. The number needed to treat is the reciprocal of the risk difference: RD = 0.08 − 0.13 = −0.05, so NNT = 1 ÷ 0.05 = 20 adults visited for one fall avoided, quoted with the time frame of the risks (one winter). Five is the risk difference itself, in percentage points, not its reciprocal; 1 ÷ 0.08 uses only the risk in the visited group; and 1.6 is a risk ratio. The NNT also assumes that the visits caused the difference: if the adults who accepted a visit were healthier to begin with, part of the difference is confounding and the true NNT is larger than 20. (Section 2, 3.3 Numbers Needed to Harm and to Treat and Worked Example 6.6; Objective 3.3)

4. In a 10-year cohort of quarry workers, the risk of chronic obstructive lung disease was 0.15 among workers exposed to silica dust and 0.06 among unexposed workers. Which statement gives the attributable fraction in the exposed (AFe) and interprets it correctly?

AFe = RD ÷ risk in the exposed = 0.09 ÷ 0.15 = 0.60, or (RR − 1) ÷ RR with RR = 2.5. It is a share of the exposed workers' cases, due to silica only if the association is causal.
The value is right, but the AFe is a share of the cases among exposed workers only. A share of all cases in the workforce is the population attributable fraction, which also depends on how many were exposed.
0.09 is the risk difference, an absolute excess of 9 cases per 100 exposed workers. The AFe divides it by the risk in the exposed: 0.09 ÷ 0.15 = 0.60.
0.40 is 1 ÷ RR, not the AFe. The AFe is (RR − 1) ÷ RR = 1.5 ÷ 2.5 = 0.60.
Correct answer: AFe = 0.60, a share of the exposed workers’ cases. AFe = RD ÷ p(D+|E+) = (0.15 − 0.06) ÷ 0.15 = 0.09 ÷ 0.15 = 0.60, or from the risk ratio, RR = 0.15 ÷ 0.06 = 2.5 and AFe = (RR − 1) ÷ RR = 1.5 ÷ 2.5 = 0.60; reading it as the share of exposed cases that would not have occurred without the exposure assumes the association is causal. Applying it to all cases in the workforce confuses it with the population attributable fraction of Section 3, which also depends on how many workers were exposed. The value 0.09 is the risk difference, and 0.40 is 1 ÷ RR. (Section 2, 4.1 The Attributable Fraction in the Exposed, Eq 6.6 and Worked Example 6.7; Objective 4.1)

5. During a mumps outbreak on a university campus, 3% of students who had two doses of mumps vaccine and 10% of unvaccinated students developed mumps. A student newspaper writes that “the vaccine is 70% effective, so 70% of vaccinated students are immune.” Which statement is correct?

VE = 70% is right, but it is a share of cases prevented among the vaccinated, not a share of people who are immune. The second clause misreads what VE measures.
0.30 is the risk ratio Rv/Ru. VE is 1 minus that ratio, (0.10 − 0.03) ÷ 0.10 = 0.70.
VE = (Ru − Rv) ÷ Ru = 0.07 ÷ 0.10 = 0.70, the AFe with being unvaccinated as the exposure. It is the share of expected cases prevented among the vaccinated, not a share of students immune.
Seven percentage points is the risk difference, an absolute measure. VE divides that difference by the risk in the unvaccinated: 0.07 ÷ 0.10 = 0.70.
Correct answer: the third statement. Vaccine efficacy treats being unvaccinated as the exposure: VE = (Ru − Rv) ÷ Ru = (0.10 − 0.03) ÷ 0.10 = 0.70, equivalently 1 − Rv/Ru = 1 − 0.30, which is the AFe with the unvaccinated as the exposed group. It is a proportion of cases prevented among the vaccinated, not a proportion of people protected, so the newspaper’s second clause is wrong. The value 0.30 is the risk ratio Rv/Ru, and 7 percentage points is the risk difference. (Section 2, 4.2 Vaccine Efficacy, Eq 6.7 and Worked Example 6.8; Objective 4.2)

6. A health authority followed 4,000 children under 5 years of age through one winter (hypothetical data). The risk of admission to hospital with a lower respiratory infection was 40 ÷ 800 = 0.050 among children in homes with visible mould or damp and 64 ÷ 3,200 = 0.020 among children in homes without mould or damp. Which set of results is correct?

RR and RD are right, but 0.40 is 1 ÷ RR, not the AFe. The AFe is (RR − 1) ÷ RR = 1.5 ÷ 2.5 = 0.60, or RD ÷ risk in the exposed = 0.030 ÷ 0.050.
An RR of 0.40 and a negative RD come from putting the non-exposed group first. With mould or damp as the exposure, RR = 0.050 ÷ 0.020 = 2.5 and RD = +0.030.
0.070 adds the two risks instead of subtracting them. The risk difference is 0.050 − 0.020 = 0.030.
RR = 0.050 ÷ 0.020 = 2.5, RD = 0.030 and AFe = 0.030 ÷ 0.050 = 0.60. Reading them as extra or attributable admissions assumes that the association is causal.
Correct answer: RR = 2.5, RD = 0.030 and AFe = 0.60. The information needed is the two risks given in the stem: RR = 0.050 ÷ 0.020 = 2.5, RD = 0.050 − 0.020 = 0.030, and AFe = (RR − 1) ÷ RR = 1.5 ÷ 2.5 = 0.60 (check: RD ÷ risk in the exposed = 0.030 ÷ 0.050 = 0.60), as in Worked Example 6.4 and Worked Example 6.7; each interpretation assumes that the association is causal. The value 0.40 is 1 ÷ RR, not the AFe; an RR of 0.40 and a negative RD come from putting the non-exposed group in the numerator; and 0.070 adds the two risks instead of subtracting them. (Section 2, 3.2.1 The Risk Difference and 4.1 The Attributable Fraction in the Exposed; Objectives 3.2 and 4.1)

7. In the same study of 4,000 children, the winter risk of admission was 0.050 among children in homes with mould or damp and 0.020 among the other children. A housing manager uses the risk difference to estimate that a home-repair programme would prevent 30 admissions per 1,000 children in affected homes each winter. Families in damp homes are also more likely to live in crowded housing and to have someone who smokes at home. Which statement about the estimate is correct?

The 30 per 1,000 counts as prevented only if mould or damp causes the whole difference and repairs remove it. If crowding or smoking explains part of it, the true effect of repairs is smaller.
This reverses the effect of confounding. If crowding and smoking inflate the observed difference, the part due to mould or damp is smaller, so repairs would prevent fewer than 30 per 1,000.
Difference measures are as open to confounding as ratio measures. Whether a measure is absolute or relative does not protect it from a third factor linked to both exposure and outcome.
The AFe is a share of the admissions among exposed children, not a count per 1,000. 0.60 of the 50 admissions per 1,000 exposed children is 30, the same as the RD.
Correct answer: the estimate rests on a causal assumption and is too high if the association is partly confounded. The information needed is what the risk difference counts: 0.050 − 0.020 = 0.030, or 30 extra admissions per 1,000 children per winter, is the number a programme would prevent only if the mould or damp causes the whole difference and repairs remove it. Crowding and smoking are linked both to damp homes and to respiratory infection, so if they explain part of the difference the true effect of repairs is smaller and fewer than 30 admissions per 1,000 would be prevented, for the same reason that an NNH or NNT is too small when confounding inflates the risk difference. Difference measures are as open to confounding as ratio measures, and the AFe is a share of the 50 admissions per 1,000 exposed children (0.60 × 50 = 30), not of 1,000 children. (Section 2, 3.1 Ratio and Difference Measures and 3.3 Numbers Needed to Harm and to Treat; Objectives 3.1 and 3.3)

✦ Pass the knowledge check with 100% to continue

Section 3

Population-Level Measures & Study Design

⏱ Estimated reading time: 12 minutes

Section 3 of 4

Population-Level Measures & Study Design

PAR and AFp: scaling attributable risk to the whole population, and matching measures to study designs.

The two population measures

PAR and AFp

Population Attributable Risk (PAR)
\[ \color{#0B7B6B}{PAR} = \color{#C2410C}{p(D+)_{\text{pop}}} - \color{#1D4ED8}{p(D+)_{E-}} \]
PAR population attributable riskp(D+)pop risk in the whole populationp(D+)E− risk in the unexposed
Population Attributable Fraction (Levin, 1953)
\[ \color{#0B7B6B}{AF_p} = \frac{\color{#C2410C}{PAR}}{\color{#1D4ED8}{p(D+)_{\text{pop}}}} = \frac{\color{#6D28D9}{p_e}(\color{#BE185D}{RR}-1)}{\color{#6D28D9}{p_e}(\color{#BE185D}{RR}-1)+1} \]
AFp population attributable fractionPAR population attributable riskp(D+)pop total population riskpe exposure prevalenceRR risk ratio

where \(p_e\) is the prevalence of exposure. Both association strength and exposure prevalence determine the result.

Worked example

Same cohort, population perspective

Overall risk: 371/5,000 = 0.074
Risk in non-smokers: 331/4,649 = 0.071
Exposure prevalence: 351/5,000 = 7%

\[ PAR = 0.074 - 0.071 = 0.003 \]
\[ AF_p = 0.003/0.074 = 4.1\% \]

The key insight

AFe = 37.5% but AFp = 4.1%. A strong association with a rare exposure produces a small population fraction.

A critical contrast

Strong and rare vs. weak and common

High RR, rare exposure

Example: intravenous drug use and HIV. Very strong association, but small AFp because the exposure is uncommon in the full population.

Modest RR, common exposure

Example: poor diet and chronic disease. Weaker association, but large AFp because most of the population is exposed.

Both the strength of association and the prevalence of exposure drive public-health impact.

Study design constraints

Which measures each design delivers

Design Direct measures Constraints Cohort RR, IR, OR, RD, ID,AFe, PAR, AFp All available Case-control OR directly RR, PAR need external data Cross-sectional Prevalence ratio (PR), OR No incidence
Carry forward

Two things into the next section

  • Population impact depends on exposure prevalence, not only on association strength. A common weak factor can outweigh a rare strong one.
  • Study design constrains measures: cohort gives everything; case-control gives OR directly; cross-sectional gives prevalence ratios.

A later section adds the last piece: standard errors and confidence intervals to quantify how precise these estimates really are.

Overview

Sections 1 and 2 computed their measures mainly from cohort data, at the level of the exposed group. This section first completes the ratio measures with the prevalence ratio from cross-sectional surveys and sets out which measures each study design can supply. It then zooms out from the exposed group to the whole population: even if an exposure powerfully causes disease in the exposed, its public health importance also depends on how common the exposure is. The population attributable risk (PAR) and the population attributable fraction (AFp) capture this combination, and the section closes by showing how the prevalence of exposure and confounding shape them.

Learning Objectives

  • 1.3 Calculate and interpret the prevalence ratio from a cross-sectional survey, and explain why it is preferred to the prevalence odds ratio when the outcome is common.
  • 2.2 Identify which measures of association can be computed from each study design.
  • 5.1 Calculate and interpret the population attributable risk (PAR) and population attributable fraction (AFp), stating the assumptions the interpretation needs.
  • 5.2 Explain how the prevalence of exposure and confounding affect population-level measures.

1.3 The Prevalence Ratio in Cross-Sectional Surveys

A cross-sectional survey measures exposure and outcome at the same time, so it counts the cases that exist at one point in time (prevalence) rather than new cases over a follow-up period. Its ratio measure is therefore a prevalence ratio (PR): the prevalence of the outcome in the exposed group divided by the prevalence in the non-exposed group. It is calculated from the same cells as a risk ratio (Equation 6.1) and read against the same null value of 1, and the same cells also give a prevalence odds ratio (POR), calculated as the cross-product of Equation 6.2. Worked Example 6.9 calculates both from one survey; watch how far apart they are when the outcome is common, and what the design cannot show.

Worked Example 6.9: Prevalence Ratio from a Cross-Sectional Survey

A survey of 2,000 adult employees asks about shift work and current insomnia (hypothetical data). Of 500 shift workers, 90 report insomnia; of 1,500 day workers, 180 do.

  • Prevalence in shift workers = 90 ÷ 500 = 0.18; in day workers = 180 ÷ 1,500 = 0.12.
  • PR = 0.18 ÷ 0.12 = 1.50: insomnia was 1.5 times as common among shift workers as among day workers.
  • POR = (90 × 1,320) ÷ (180 × 410) = 118,800 ÷ 73,800 = 1.61.

The outcome is common (270 of 2,000, or 13.5 per cent), so the POR lies further from 1 than the PR, as Figure 6.1 predicts for the OR and the RR, and the PR is the measure to report. Because shift work and insomnia were measured at the same time, the survey cannot show that shift work came first; people with insomnia may also have moved into or out of shift work.

The example applies the rule of 2.1 in Section 1 to prevalence data: when the outcome is common, the odds-based measure drifts away from the ratio of proportions, so the PR is the measure to report, worded in terms of prevalence (“as common as”) rather than risk. The cross-sectional survey is one of three designs met in this lesson, and each can supply a different set of measures.

2.2 Which Measures Each Study Design Can Supply

Not all measures can be computed from all study designs. Table 6.3 summarises which measures are available from cross-sectional, cohort and case-control studies, and its footnotes name the extra information that some of them need; in the cross-sectional column, the “RR” is the prevalence ratio described in 1.3, because a cross-sectional study measures prevalence rather than risk.

Table 6.3. Measures of association that can be computed from cross-sectional, cohort and case-control studies.

MeasureCross-sectionalCohortCase-control
RR✓✓
IR✓
OR✓✓✓
RD✓✓
AFe✓✓✓b
PAR✓✓a
AFp✓✓a✓c

a From the cohort’s own data only if the cohort is a representative sample of the population, so that its exposure prevalence and overall risk reflect the population; otherwise requires an independent estimate of p(D+) or p(E+). b Estimated using OR. c Requires the OR and the proportion of cases who were exposed, p(E+|D+), which the study’s own cases can supply if they represent all the cases in the population (Equation 6.10 with the OR in place of the risk ratio).

The pattern in the table follows from what each design fixes. A cohort study measures disease as it develops in exposed and non-exposed people, so it can supply every measure in this lesson, although it can supply the population measures from its own table only when it is a representative sample of the population; when it recruits fixed numbers of exposed and non-exposed people, its exposure prevalence and overall risk say nothing about the population, and an outside estimate is needed (footnote a). A case-control study fixes the numbers of cases and controls, so, as 2.1 in Section 1 explained, it gives the OR directly and the AFe and AFp only through the OR (the AFp also needs the proportion of cases who were exposed, which the study’s own cases supply if they represent all cases, footnote c), but it cannot give risks, the RR, the RD or the PAR. The last two rows of the table, the population measures, are the subject of the rest of this section.

5.1 The Population Attributable Risk and Fraction

While RD and AFe describe the effect of exposure among exposed individuals, public health decisions often require understanding the impact of an exposure on the entire population. The population attributable risk (PAR) is the increase in overall population risk attributable to the exposure. By this we mean that it compares the risk in the whole population, exposed and non-exposed people together, with the risk the population would have if everyone had the risk of the non-exposed group. The difference is the extra risk that the exposure adds to the population as a whole. Because only the exposed people carry the extra risk, the PAR is the risk difference spread across the whole population: if 10 per cent of people are exposed and each of them carries an extra risk of 0.05, the population as a whole carries an extra risk of 0.10 × 0.05 = 0.005. It therefore reflects both the strength of the association and the frequency of the exposure in the population, an idea originally developed by Levin (1953) for lung cancer and reviewed by Northridge (1995) as a link between causal inference and public-health action. Equation 6.8 gives the PAR in three equivalent forms, the last of which shows that it equals the risk difference multiplied by the proportion of the population that is exposed.

Population attributable risk
PAR = p(D+) − p(D+|E−) = (m1/n) − (a0/n0) = RD × p(E+) Eq 6.8
The population attributable risk is the disease risk in the whole population minus the risk in the unexposed.

The population attributable fraction is the proportion of all disease in the population that is attributable to the exposure. It was introduced in HSCI 341 Lesson 1, and the recall panel below restates Levin's formula and poses a retrieval question; work it out before opening the answer.

Recall: The Population Attributable Fraction

HSCI 341 Lesson 1, Section 3 (Seeking Causes and Models of Causation), introduced the population attributable fraction with Levin's formula (1953), AFp = pe(RR − 1) ÷ [pe(RR − 1) + 1], where pe is the proportion of the population exposed. It is the proportion of disease in the whole population that is attributable to the exposure, and that would be avoided if the exposure were removed, assuming causation and no confounding. HSCI 230 Lesson 5, Section 3 (The Exposure), showed a published estimate of it for alcohol and cancer in the EPIC cohort.

Retrieval question. If 20% of a population is exposed and the risk ratio is 3, what is the population attributable fraction?

Show the answer▼

AFp = 0.20 × 2 ÷ (0.20 × 2 + 1) = 0.40 ÷ 1.40 = 0.29, so about 29% of cases would be attributable to the exposure.

This section adds a second route to the same quantity: AFp is the population attributable risk divided by the total population risk, which links it to Equation 6.8. Equation 6.9 gives both routes, the second of which is Levin's formula with the exposure prevalence written as p(E+).

Population attributable fraction
AFp = PAR / p(D+) = p(E+)(RR − 1) / [p(E+)(RR − 1) + 1] Eq 6.9
The population attributable fraction is the population attributable risk as a share of the total population risk; it rises with both the exposure prevalence and the risk ratio.

Worked Example 6.10 returns to the smoking and low birth weight cohort of Worked Example 6.4, computes the PAR with Equation 6.8, and divides it by the population risk, as in the first form of Equation 6.9. Because the 5,000 women were followed as one cohort, rather than recruited in fixed numbers of smokers and non-smokers, and are treated as representative of the population, the table itself supplies the exposure prevalence (351 ÷ 5,000, about 7 per cent) and the overall risk, as footnote a of Table 6.3 requires. Take care to use the overall risk from the whole table, not the risk in the exposed.

Worked Example 6.10: Smoking & Low Birth Weight (Population Level)

From the cohort of 5,000 women (351 smokers, 4,649 non-smokers):

  • Overall risk: p(D+) = 371/5000 = 0.074
  • Risk in non-exposed: 331/4649 = 0.071
  • PAR = 0.074 − 0.071 = 0.003
  • AFp = 0.003 / 0.074 = 0.041 (4.1%)

Only 4.1% of all low-birth-weight babies in the population were attributable to smoking. The low AFp is because very few women (351/5000 = 7%) smoked during the 2nd trimester, despite the relatively strong association (RR = 1.60).

The same cohort gave an AFe of 37.5% in Section 2 but an AFp of only 4.1% here, because the AFe describes the cases among the exposed women whereas the AFp describes all the cases in the population. Calculator 6.8 and Calculator 6.9 reproduce the two results, and Calculator 6.9 also offers presets that compare a common, weak exposure with a rare, strong one, the contrast that 5.2 explains.

5.2 Exposure Prevalence, Confounding and the AFp

The second form of Equation 6.9 contains both the exposure prevalence p(E+) and the risk ratio, so the population impact of a risk factor depends on the two together. Worked Example 6.10 showed the consequence: smoking had a relatively strong association with low birth weight (RR = 1.60), yet so few women smoked that it accounted for only about 4% of all low-birth-weight babies. The general point, set out in the callout below, is a frequent source of exam errors.

Why Exposure Prevalence Matters

A strong risk factor (high RR) that is rare in the population will have a small AFp. A weaker risk factor (modest RR) that is common may have a large AFp. For example, intravenous drug use has a very high RR for HIV, but if it is rare in the population, eliminating it would prevent few total cases. A modestly elevated risk factor like poor diet, affecting millions, may account for more total cases.

Worked Example 6.10 used the unadjusted risks of a single cohort, so the AFp it gives assumes that the association is not confounded. Confounding occurs when a third factor is linked to both the exposure and the disease, so that part of the crude association belongs to that factor rather than to the exposure. For example, if women who smoked were also more likely to be poorly nourished, some of the extra low-birth-weight babies among smokers might be due to nutrition, and the crude RR would overstate the effect of smoking. If confounding is present, adjusted estimates of RR should be used, that is, risk ratios that have been corrected for the confounding factors. The AFp can then be estimated with Equation 6.10, where pd is the proportion of cases exposed to the risk factor and aRR is the adjusted risk ratio. This form is preferred because it starts from the cases themselves: it takes the share of cases who were exposed and multiplies it by the share of those exposed cases that are due to the exposure. Placing an adjusted RR into Levin’s formula, which uses the exposure prevalence in the whole population, does not in general give a correct answer when confounding is present. For multiple exposure categories, such as never, former and current smokers, the same calculation is made for each exposed category and the results are added together, which is why a summation formula is used.

Population attributable fraction (adjusted)
AFp = pd × (aRR − 1) / aRR Eq 6.10
The population attributable fraction can be estimated from the proportion of cases exposed and the confounding-adjusted risk ratio.

Worked Example 6.11 applies Equation 6.10 to a hypothetical cohort study of physical inactivity and type 2 diabetes; note that pd is the share of cases who were exposed, not the share of the population.

Worked Example 6.11: AFp with an Adjusted Risk Ratio

Suppose a cohort study of type 2 diabetes finds that 45% of the people who developed diabetes were physically inactive at baseline (pd = 0.45), and that the risk ratio for inactivity, adjusted for age and body mass index, is aRR = 1.6 (hypothetical values).

AFp = 0.45 × (1.6 − 1) / 1.6 = 0.45 × 0.375 = 0.169, or about 17%.

About 17% of the diabetes cases in this population would be attributable to physical inactivity, assuming a causal relationship and that the adjustment has removed the confounding. The factor (aRR − 1)/aRR is the attributable fraction among the exposed, and multiplying it by the proportion of cases who were exposed scales it to the whole population.

The example shows why the adjustment matters: when confounding has inflated the crude risk ratio, an AFp calculated from the crude RR overstates the share of cases due to the exposure, and the adjusted RR removes that part of the excess. Calculator 6.10 reproduces the worked example.

This section has applied the ratio logic to cross-sectional surveys, matched each measure to the designs that can supply it, extended the measures of Section 2 from the exposed group to the whole population, and shown how exposure prevalence and a confounding-adjusted risk ratio shape the AFp. The key takeaways and the knowledge check below consolidate the section, including a comparison of two risk factors that differ in strength and prevalence. Section 4 turns to the precision of these estimates.

Key Takeaways

  • 1.3 A cross-sectional survey measures exposure and outcome at the same time, so it counts existing cases (prevalence) rather than new ones. The prevalence ratio is calculated just like a risk ratio, the prevalence in the exposed divided by the prevalence in the non-exposed, and it is read against the same null value of 1: in the shift-work survey PR = 0.18 ÷ 0.12 = 1.50, so insomnia was 1.5 times as common among shift workers as among day workers. You can calculate a prevalence odds ratio from the same cells (here 1.61), but when the outcome is common, as insomnia was at 13.5 per cent, the POR lies further from 1 than the PR, so the PR is the measure to report. Word your interpretation in terms of prevalence (“as common as”, not “as likely to develop”), and add the design’s main limitation: because exposure and outcome were measured together, the survey cannot show which came first.
  • 2.2 Not every measure can be calculated from every design, and Table 6.3 is your map. A cohort study measures disease as it develops in exposed and non-exposed people, so it can supply every measure in this lesson, although the population measures need an estimate of how common the exposure (or the disease) is in the population. A case-control study fixes the numbers of cases and controls, so it gives the OR directly and the AFe and AFp only through the OR (the AFp also needs the proportion of cases who were exposed), but it cannot give risks, the RR, the RD or the PAR. A cross-sectional survey gives prevalences, so its ratio measure is a prevalence ratio. When a question asks what a study can report, first identify the design from how participants were selected, then read across the table, and explain any gap by saying what the design fixes or cannot measure.
  • 5.1 The population measures move the question from the exposed group to everyone. The population attributable risk is the extra risk in the whole population that is due to the exposure, PAR = p(D+) − p(D+|E−) = (m1/n) − (a0/n0), which also equals RD × p(E+). The population attributable fraction expresses it as a share of all the disease in the population, AFp = PAR / p(D+), or with Levin’s formula, p(E+)(RR − 1) / [p(E+)(RR − 1) + 1]. In the smoking cohort PAR = 0.074 − 0.071 = 0.003 and AFp = 0.003 / 0.074 = 0.041, so about 4 per cent of all low-birth-weight babies were attributable to smoking. A common slip is to use the risk in the exposed in place of the overall risk, which comes from the whole table (m1/n), and working the AFp both ways is a good check. A complete interpretation describes the proportion of all cases in the population that would be avoided if the exposure were removed, and it states two assumptions: that the association is causal and that it is not confounded.
  • 5.2 The population impact of a risk factor depends on two things together: how strong the association is (the RR) and how common the exposure is, p(E+). This is why a strong but rare risk factor can have a small AFp while a weak but common one has a large AFp; smoking had an RR of 1.60 in the cohort but an AFp of only about 4 per cent, because only 7 per cent of the women smoked. When confounding is present, the AFp should use a confounding-adjusted risk ratio, as in Equation 6.10, AFp = pd × (aRR − 1) / aRR, where pd is the proportion of cases who were exposed. When a question asks you to rank exposures for a public health programme, look at both the RR and the prevalence of exposure before you decide, and be ready to explain how an exposure can have a large AFe but a small AFp.
Knowledge Check: this section

1. A town followed 2,000 adults aged 60 and over for five years. Of the 600 with untreated high blood pressure, 30 had a stroke; of the 1,400 without it, 28 had a stroke. Which answer gives the population attributable fraction (AFp) and interprets it correctly?

0.60 is the AFe, (2.5 − 1) ÷ 2.5, which describes only the exposed adults' strokes. Because only 30% of adults were exposed, the share of all strokes is smaller, 0.31.
0.009 is the PAR, a difference in risks (0.029 − 0.020), not a fraction. The AFp divides it by the overall risk: 0.009 ÷ 0.029 = 0.31.
The value is right, but the AFp is a share of all strokes in the group, not of strokes among the exposed. The share among the exposed is the AFe, 0.60.
AFp = PAR ÷ p(D+) = 0.009 ÷ 0.029 = 0.31, matching Levin's formula. Reading it as strokes avoided assumes the association is causal and not confounded.
Correct answer: AFp = 0.31, a share of all strokes in the group. PAR = p(D+) − p(D+|E−) = 58/2,000 − 28/1,400 = 0.029 − 0.020 = 0.009, and AFp = PAR ÷ p(D+) = 0.009 ÷ 0.029 = 0.31; Levin’s formula, with p(E+) = 600/2,000 = 0.30 and RR = 0.05 ÷ 0.02 = 2.5, gives the same value, 0.30 × 1.5 ÷ (0.30 × 1.5 + 1) = 0.45 ÷ 1.45 = 0.31. Reading it as strokes avoided needs two assumptions: the association is causal, and it is not confounded. The tempting 0.60 is the AFe, (2.5 − 1) ÷ 2.5, which describes only the exposed adults’ strokes and is larger because only 30% of the adults were exposed; 0.009 is the PAR, a difference in risks, not a fraction. (Section 3, 5.1 The Population Attributable Risk and Fraction, Eq 6.8, Eq 6.9 and Worked Example 6.10; Objective 5.1)

2. In a cohort study of falls among older adults, 50% of the adults who fell had been taking a sedative (pd = 0.50). The crude risk ratio for sedative use is 2.0; after adjustment for age and frailty, which are linked both to sedative use and to falls, the adjusted risk ratio is aRR = 1.25. Which AFp should the team report?

Using all the data does not remove confounding. Part of the crude RR of 2.0 reflects age and frailty, so an AFp based on it overstates the share of falls due to sedatives.
With confounding present, Eq 6.10 uses the adjusted RR: 0.50 × (1.25 − 1) ÷ 1.25 = 0.10. Even this assumes the adjusted association is causal and the confounding fully removed.
0.20 is the AFe from the adjusted RR, a share of the exposed cases. The AFp scales it by the proportion of cases exposed: 0.50 × 0.20 = 0.10.
pd is the share of cases who were exposed, not the share caused by the exposure. Some exposed adults would have fallen anyway, so pd is multiplied by (aRR − 1) ÷ aRR.
Correct answer: 10%. When confounding is present, the AFp is calculated from the confounding-adjusted risk ratio with Eq 6.10: AFp = pd × (aRR − 1) ÷ aRR = 0.50 × 0.25 ÷ 1.25 = 0.50 × 0.20 = 0.10. The crude RR would give 0.50 × (2.0 − 1) ÷ 2.0 = 0.25, which overstates the share of falls due to sedatives because part of the crude association reflects age and frailty. The value 0.20 is the AFe from the adjusted RR, not yet scaled by the proportion of cases exposed, and pd alone is the share of cases exposed, not the share caused by the exposure. Even the 10% figure assumes that the adjusted association is causal and that adjustment has removed the confounding. (Section 3, 5.2 Exposure Prevalence, Confounding and the AFp, Eq 6.10 and Worked Example 6.11; Objective 5.2)

3. A case-control study compares 300 people with oral cancer with 600 controls without it. Of the cases, 120 had regularly chewed betel quid; of the controls, 120 had. Which of the following can be calculated from this study’s table alone?

The investigator chose two controls for every case, so the 240 chewers are not a group of people at risk. 120 ÷ 240 reflects that design choice, not a risk of oral cancer.
A risk difference needs the risk of disease in each exposure group, which a case-control study cannot give because the numbers of cases and controls were fixed by design (Table 6.3).
The OR can be calculated because of its symmetry: (120 × 480) ÷ (180 × 120) = 2.67. Substituting it for the RR gives an approximate AFe of 0.63, as Table 6.3 allows.
The 900 participants are not a population: the investigator set the ratio of controls to cases. The PAR needs the overall risk of disease, which a case-control table cannot supply.
Correct answer: the odds ratio and, through it, the AFe. The table has a1 = 120 and a0 = 180 (cases) and b1 = 120 and b0 = 480 (controls); the investigator chose two controls for every case, so the column totals are not groups of people at risk, 120 ÷ 240 is not a risk, and neither the risk difference nor the population attributable risk, which both need risks of disease, can be calculated (Table 6.3). The OR can be calculated because of its symmetry, OR = 57,600 ÷ 21,600 = 2.67, and substituting it for the RR in Eq 6.6 gives AFe ≈ (OR − 1) ÷ OR = 1.67 ÷ 2.67 ≈ 0.63 (footnote b of Table 6.3; oral cancer is rare, so the OR also approximates the RR). (Section 3, 2.2 Which Measures Each Study Design Can Supply, Table 6.3; Objective 2.2)

4. A cross-sectional survey of 1,000 secondary-school students finds that 60 of the 200 students who vape and 120 of the 800 who do not vape report a persistent cough. Which answer gives the measure to report and the right caution?

PR = 0.30 ÷ 0.15 = 2.0. With 18% of students affected the POR (2.43) lies noticeably further from 1, so the PR is reported, and a survey cannot show whether vaping came first.
Lying further from 1 is not clearer evidence; it is a distortion of the PR when the outcome is common. With 18% affected the POR overstates the PR, so the PR is the measure to report.
A survey counts existing cases at one time, so it measures prevalence, not the risk of developing a cough. It also cannot show whether vaping began before the cough.
Dividing 0.15 by 0.30 puts the non-vapers' prevalence on top, which reverses the reference group. With vaping as the exposure, PR = 0.30 ÷ 0.15 = 2.0.
Correct answer: PR = 2.0, with the caution about timing. The prevalence of cough is 60 ÷ 200 = 0.30 among students who vape and 120 ÷ 800 = 0.15 among those who do not, so PR = 0.30 ÷ 0.15 = 2.0, while POR = (60 × 680) ÷ (120 × 140) = 40,800 ÷ 16,800 = 2.43. Because 180 of the 1,000 students (18%) have the outcome, the POR lies further from 1 than the PR without being any “clearer”, so the PR is the measure to report. A survey counts existing cases at one time, so it gives a prevalence, not the risk of developing a cough, and it cannot show which came first; dividing 0.15 by 0.30 reverses the reference group. (Section 3, 1.3 The Prevalence Ratio in Cross-Sectional Surveys and Worked Example 6.9; Objective 1.3)

5. A dental public health team followed 4,000 children aged 2 to 5 years for three years (hypothetical data). Of the 2,000 children who had a sugary drink every day, 60 needed dental surgery for tooth decay; of the 2,000 who did not, 40 did. Which answer gives the population attributable fraction (AFp) for daily sugary drinks and interprets it correctly?

0.33 is the AFe, (1.5 − 1) ÷ 1.5, which describes only the cases among daily drinkers. The share of all cases is smaller because half the children were not exposed.
0.005 is the PAR, a difference in risks (0.025 − 0.020), not a fraction of cases. The AFp divides it by the overall risk: 0.005 ÷ 0.025 = 0.20.
PAR = 0.025 − 0.020 = 0.005 and AFp = 0.005 ÷ 0.025 = 0.20, matching Levin's formula. Reading it as cases avoided assumes the association is causal and not confounded.
This divides the PAR by the risk among daily drinkers (0.030). The AFp divides by the overall risk in the whole group, 0.025, giving 0.20.
Correct answer: AFp = 0.20. The information needed is the overall risk from the whole table, p(D+) = 100 ÷ 4,000 = 0.025, and the risk in the non-exposed, 40 ÷ 2,000 = 0.020. Then PAR = 0.025 − 0.020 = 0.005 (check: RD × p(E+) = 0.010 × 0.50 = 0.005) and AFp = PAR ÷ p(D+) = 0.005 ÷ 0.025 = 0.20; Levin’s formula gives the same value, 0.50 × (1.5 − 1) ÷ [0.50 × (1.5 − 1) + 1] = 0.25 ÷ 1.25 = 0.20. Reading it as cases avoided assumes that the association is causal and not confounded. The value 0.33 is the AFe, (1.5 − 1) ÷ 1.5, which describes only the cases among daily drinkers; 0.005 is the PAR, a difference in risks; and dividing the PAR by the risk among daily drinkers (0.030) instead of the overall risk gives 0.17. (Section 3, 5.1 The Population Attributable Risk and Fraction, Eq 6.8, Eq 6.9 and Worked Example 6.10; Objective 5.1)

6. In the same cohort of 4,000 children, daily sugary drinks (50% of children) had a risk ratio of 1.5 for tooth decay needing surgery and an AFp of 0.20. The 2.5% of children who were put to bed with a bottle of juice or milk every night had a risk ratio of 5.0 compared with the other children. A board member argues that the programme should tackle bedtime bottles first “because their association is much stronger.” Which response is best?

A strong association does not guarantee a large share of all cases: the AFp depends on how common the exposure is as well as on the RR. Bedtime bottles affect only 2.5% of children.
Levin's formula gives 0.10 ÷ 1.10, about 0.09, for bedtime bottles against 0.20 for sugary drinks. A strong but rare exposure can account for fewer cases than a weaker, common one.
0.80 is the AFe for bedtime bottles, a share of the cases among exposed children only. The share of all cases also depends on the 2.5% exposure prevalence, giving about 0.09.
The AFp does depend on how common an exposure is, which is exactly what makes the comparison possible. Levin's formula combines exposure prevalence with the RR.
Correct answer: bedtime bottles account for about 9% of all cases against 20% for daily sugary drinks. The information needed is both the strength of each association and how common each exposure is, because the AFp depends on both: Levin’s formula gives 0.025 × (5.0 − 1) ÷ [0.025 × (5.0 − 1) + 1] = 0.10 ÷ 1.10 = 0.091 for bedtime bottles. A strong but rare risk factor can have a small AFp, while a weaker but common one has a larger AFp. The value 0.80 is the AFe for bedtime bottles, a share of the cases among the exposed children only; for an individual child, stopping bedtime bottles matters more, which is why targeted advice for those families still makes sense alongside a programme on sugary drinks. (Section 3, 5.2 Exposure Prevalence, Confounding and the AFp and the callout “Why Exposure Prevalence Matters”; Objective 5.2)

✦ Pass the knowledge check with 100% to continue

Section 4

Hypothesis Testing & Confidence Intervals

⏱ Estimated reading time: 15 minutes

Section 4 of 4

Hypothesis Testing & Confidence Intervals

Standard errors, P-values, confidence intervals, and a guide to choosing the right statistical test.

Standard error

Precision of a point estimate

Difference measures (RD, ID)

Variance computed directly from cell counts.

Var(RD)
\[ \frac{\color{#C2410C}{R_{E+}}(1-\color{#C2410C}{R_{E+}})}{\color{#C2410C}{n_1}} + \frac{\color{#1D4ED8}{R_{E-}}(1-\color{#1D4ED8}{R_{E-}})}{\color{#1D4ED8}{n_0}} \]
RE+, n1 risk and size, exposed groupRE−, n0 risk and size, unexposed group

Ratio measures (RR, IR, OR)

Variance on the log scale using Taylor series approximations.

Var(ln OR)
\[ \frac{1}{\color{#C2410C}{a_1}} + \frac{1}{\color{#BE185D}{b_1}} + \frac{1}{\color{#1D4ED8}{a_0}} + \frac{1}{\color{#6D28D9}{b_0}} \]
a1 exposed casesb1 exposed non-casesa0 unexposed casesb0 unexposed non-cases
Confidence intervals

Uncertainty around the point estimate

CI for difference measures (symmetric)
\[ \color{#0B7B6B}{\hat{\theta}} \pm \color{#1D4ED8}{z_{\alpha/2}} \cdot \color{#C2410C}{SE(\hat{\theta})} \]
θ̂ point estimatezα/2 critical valueSE standard error
CI for ratio measures (log scale, then exponentiate)
\[ \exp\!\left[\color{#0B7B6B}{\ln\hat{\theta}} \pm \color{#1D4ED8}{z_{\alpha/2}} \cdot \color{#C2410C}{SE(\ln\hat{\theta})}\right] \]
ln θ̂ estimate on the log scalezα/2 critical valueSE standard error of the log estimate

Significance criterion: the CI for OR, RR, or IR must exclude 1; the CI for RD or ID must exclude 0.

From the lesson's examples

Computed confidence intervals

Measure Point estimate 95% CI RD (smoking)0.043(0.009, 0.077) RR (smoking)1.60(1.174, 2.182) OR (smoking)1.68(1.154, 2.387) ID (migraine)0.338(0.232, 0.443) IR (migraine)2.81(1.983, 4.050)
Forest plot of the three ratio-measure confidence intervals from the lesson examples, on a log axis, all sitting to the right of the null value of 1.
The three ratio measures plotted on a log axis. Each interval lies entirely to the right of the null (1), so each is significant at the 5% level; the differing interval widths show how much more the data constrain some estimates than others.
Four test statistics

Tests for 2x2 tables and regression

Pearson chi-squared

For categorical outcomes with expected cell counts all above 5. Compares observed to expected counts under independence.

Fisher's exact

Small expected counts (below 5). Exact P-value, no large-sample assumption. Often reported with the OR and its exact CI.

Wald test

Estimate divided by its standard error. Fast, but less accurate in small samples.

Likelihood ratio test

Generally preferred in regression settings; more accurate than Wald in small samples.

Choosing the right test

A decision framework

Situation Parametric Non-parametric Continuous, 2 indep. groupsTwo-sample t-testMann-Whitney U Continuous, 2 paired groupsPaired t-testWilcoxon signed-rank Continuous, 3+ groupsOne-way ANOVAKruskal-Wallis Categorical, independentPearson chi-sq / Fisher-- Categorical, pairedMcNemar's test-- Two continuous variablesPearson rSpearman rho
Carry forward

Point estimates need uncertainty bands

  • Standard error quantifies precision; computed differently for difference vs. ratio measures.
  • Confidence intervals show the range of plausible effects, not just a binary signal.
  • P-values are not the only verdict: a P-value of 0.049 and one of 0.051 carry almost identical information.

The final assessment is just below. Take the synthesis materials slowly before attempting the seventeen questions.

Overview

Sections 1 to 3 produced point estimates of association, single numbers such as RR = 2.5 or AFp = 30%. A point estimate means little until its uncertainty has been quantified, and this section closes the lesson with the statistical-inference machinery applied directly to the measures computed in this lesson. It defines the standard error of each measure and the null hypothesis and P-value on which a test is built, describes the four test statistics in common use and how a test is chosen for other kinds of outcome and comparison, calculates and interprets confidence intervals, and finally separates statistical significance from the strength of an association.

Learning Objectives

  • 6.2 Explain the concepts of standard error, null hypothesis, and P-value.
  • 6.3 Describe the four common test statistics for evaluating associations, and choose a test for two independent groups, for small expected counts and for matched pairs.
  • 6.4 Interpret confidence intervals for measures of association, including why intervals for ratio measures are not symmetric.
  • 6.5 Distinguish between statistical significance and the strength of association.

6.2 Standard Errors, Null Hypotheses and P-Values

6.2.1 The Standard Error of a Measure of Association

The standard error (SE) provides a measure of the precision of a point estimate, that is, how much uncertainty exists in the estimate, and it is the square root of the estimate’s variance. By this we mean that a study measures only one sample of people, and a different sample drawn from the same population would give a somewhat different estimate. If the same study could be repeated many times, the estimates would scatter around the true value, and the standard error describes how widely they would typically scatter. A small standard error means the estimates would cluster tightly, so any one of them is likely to be close to the truth; a large standard error means they would spread widely. Because small groups give less stable estimates than large ones, the standard error shrinks as the numbers in the table grow. For difference measures (RD, ID), the variance can be computed directly from the cell counts, as Equation 6.11 shows for the risk difference.

Variance of the risk difference
var(RD) = [(a1/n1)(1 − a1/n1)] / n1 + [(a0/n0)(1 − a0/n0)] / n0 Eq 6.11
The variance of the risk difference adds a contribution from the exposed group and one from the unexposed group, each computed straight from the cell counts.

Worked Example 6.12 applies Equation 6.11 to the smoking and low birth weight cohort; watch which of the two groups contributes most of the variance.

Worked Example 6.12: Variance and Standard Error of the RD

For the smoking and low birth weight cohort, a1/n1 = 40/351 = 0.1140 and a0/n0 = 331/4,649 = 0.0712.

  • Exposed contribution: (0.1140 × 0.8860) / 351 = 0.000288.
  • Unexposed contribution: (0.0712 × 0.9288) / 4,649 = 0.0000142.
  • var(RD) = 0.000288 + 0.0000142 = 0.000302, so SE(RD) = √0.000302 = 0.0174.

Almost all of the variance comes from the exposed group, because it contains only 351 women. The estimate would become more precise mainly through a larger exposed group.

The example shows that precision improves mainly by enlarging the smallest group. Calculator 6.11 reproduces the calculation so that the contributions of the two groups to the variance can be compared.

For ratio measures (RR, IR, OR), the variance is computed on the log scale using Taylor series approximations, because the sampling distribution of the log of a ratio is approximately symmetric whereas that of the ratio itself is skewed. By this we mean that a ratio is squeezed between 0 and 1 on the protective side but can run from 1 to infinity on the harmful side, so repeated estimates of a ratio bunch up below and stretch out above. For example, a halving of risk gives an RR of 0.5, which lies only 0.5 below the null value, while a doubling gives an RR of 2, which lies 1 above it, even though the two effects are equally large in opposite directions. Taking the natural logarithm fixes this, because ln(0.5) = −0.69 and ln(2) = 0.69 sit at equal distances on either side of ln(1) = 0. On the log scale the estimates therefore spread out evenly, and the usual methods for symmetric distributions can be applied. A Taylor series approximation is simply the mathematical tool used to work out the variance of this transformed quantity from the cell counts. The two tabs below give the variance of ln(RR) (Equation 6.12) and of ln(OR) (Equation 6.13), each followed by a worked example on the smoking cohort and a calculator. Open both tabs and compare the two variances for the same data.

Variance of ln(RR)
var(ln RR) = 1/a1 − 1/n1 + 1/a0 − 1/n0 Eq 6.12
The variance of the log risk ratio is built from the diseased counts and group totals of the exposed group and the unexposed group; the variance is calculated for ln(RR) because its sampling distribution is approximately symmetric, whereas that of RR is skewed.

Worked Example 6.13: var(ln RR)

For the smoking cohort (a1 = 40, n1 = 351, a0 = 331, n0 = 4,649):

var(ln RR) = 1/40 − 1/351 + 1/331 − 1/4,649 = 0.025000 − 0.002849 + 0.003021 − 0.000215 = 0.024957, so SE(ln RR) = √0.024957 = 0.158.

The term 1/40 is by far the largest, because the exposed group has only 40 cases.

Variance of ln(OR)
var(ln OR) = 1/a1 + 1/a0 + 1/b1 + 1/b0 Eq 6.13
The variance of the log odds ratio is the sum of the reciprocals of all four cells of the table, two from the exposed column and two from the unexposed column.

Worked Example 6.14: var(ln OR)

For the smoking cohort (a1 = 40, a0 = 331, b1 = 311, b0 = 4,318):

var(ln OR) = 1/40 + 1/331 + 1/311 + 1/4,318 = 0.025000 + 0.003021 + 0.003215 + 0.000232 = 0.031468, so SE(ln OR) = √0.031468 = 0.177.

The variance of ln(OR) is larger than that of ln(RR) for the same data, so the odds ratio is estimated slightly less precisely.

The tabs show that the smallest cell dominates both variances and that ln(OR) has the larger variance, so the odds ratio is estimated slightly less precisely than the risk ratio. The standard errors computed in this subsection are used in the Wald statistic in 6.3 and in the confidence intervals in 6.4.

6.2.2 Null Hypotheses and P-Values

A hypothesis test asks how likely the observed data would be if there were no association between exposure and disease. HSCI 230 defined the P-value and separated statistical from practical significance; the recall panel below restates those points and poses a retrieval question to try before opening the answer.

Recall: P-values and Practical Significance

The glossary of HSCI 230 Lesson 11 defined the p-value as the probability, assuming the null hypothesis is true, of observing data at least as extreme as those obtained. It says nothing directly about the probability that the null hypothesis is true or about the size of an effect. HSCI 230 Lesson 12, Section 2 (Stepwise Critical Appraisal), separated statistical significance from practical, clinical or public health significance: a very large study can detect, with a small p-value, an effect too small to change any decision.

Retrieval question. A study of 200,000 people finds a risk ratio of 1.03 with p = 0.001. Is the association statistically significant, and is it practically important?

Show the answer▼

It is statistically significant at the 5% level, because p is below 0.05. A 3% relative increase is small, so its practical importance depends on how common the outcome and the exposure are; the small p-value reflects the precision of a very large study, and the risk ratio with its confidence interval carries the practical message.

For a measure of association, the null hypothesis states that there is no association, so it sets the measure at its null value:

  • For difference measures (RD, ID): H0: θ = 0
  • For ratio measures (RR, IR, OR): H0: θ = 1

An alternative hypothesis can be 1-tailed or 2-tailed. The alternative hypothesis is the statement that is accepted if the null hypothesis is rejected. A 2-tailed alternative states only that the measure differs from its null value, in either direction, so for a risk ratio it allows RR > 1 or RR < 1. A 1-tailed alternative states the direction in advance, for example that RR > 1, and ignores the possibility of an effect in the other direction. In general, 2-tailed hypotheses are preferred because 1-tailed hypotheses are harder to justify. This is because researchers rarely know for certain that an exposure can only increase disease, and an exposure expected to be harmful sometimes turns out to be protective. In addition, a 1-tailed test gives a smaller P-value for the same data in the expected direction, so choosing it can look like an attempt to make a result appear significant.

The P-value is too often reduced to a verdict, and the callout below sets out what that costs.

Limitations of P-values

P-values are often dichotomised into “significant” or “non-significant” at α = 0.05, but this entails a huge loss of information (Wasserstein & Lazar, 2016; Greenland et al., 2016). A P-value of 0.049 and 0.051 lead to different conclusions despite being virtually identical. Always report the actual P-value and a confidence interval, which conveys both significance and precision.

A P-value is produced by a test statistic, and the next subsection sets out the statistics in common use and how to choose among them.

6.3 Test Statistics and Choosing a Test

6.3.1 Four Test Statistics for Measures of Association

Four test statistics are in common use for measures of association: the Pearson χ², exact tests such as Fisher’s exact test, the Wald statistic and the likelihood ratio test. They differ in how they are computed and in the sample sizes for which they are reliable. Click each card below to see how each statistic is calculated and when it applies.

Pearson χ²Click to explore
Exact TestsClick to explore
Wald StatisticClick to explore
Likelihood Ratio TestClick to explore

Of the four, the Pearson χ² and the Wald statistic can both be calculated by hand. The two tabs below give each formula, the Pearson χ² as Equation 6.14 and the Wald statistic as Equation 6.15, each with a worked example on the smoking and low birth weight cohort and a calculator. Open both tabs; the Wald example uses the standard errors from Worked Example 6.12 and Worked Example 6.13.

Pearson χ² statistic
χ² = Σ (O − E)² / E,  with E = (row total × column total) / grand totalEq 6.14
The Pearson χ² statistic sums, over the four cells, the squared difference between the observed count and the expected count, divided by the expected count; each expected count is the row total times the column total divided by the grand total.

Worked Example 6.15: Pearson χ²

For the smoking and low birth weight table, the expected number of low-birth-weight babies among smokers, if smoking and birth weight were unrelated, is (371 × 351) / 5,000 = 26.0442, compared with 40 observed. The four expected counts are 26.0442 (a1), 344.9558 (a0), 324.9558 (b1), and 4,304.0442 (b0).

χ² = (40 − 26.0442)²/26.0442 + (331 − 344.9558)²/344.9558 + (311 − 324.9558)²/324.9558 + (4,318 − 4,304.0442)²/4,304.0442 = 7.478 + 0.565 + 0.599 + 0.045 = 8.69.

With 1 degree of freedom, χ² = 8.69 corresponds to P = 0.003, so an association this strong would be unlikely if smoking and low birth weight were unrelated.

Wald statistic
ZWald = (θ − θ0) / SE(θ)Eq 6.15
The Wald statistic is the distance between the point estimate and its null value, measured in standard errors; for ratio measures, θ and θ0 are replaced by their natural logarithms and the standard error is that of ln θ.

Worked Example 6.16: Wald Statistic

For the risk difference in the smoking cohort, RD = 0.0428 and SE(RD) = 0.0174. Under the null hypothesis θ0 = 0, Z = (0.0428 − 0) / 0.0174 = 2.46, which corresponds to a two-sided P-value of 0.014.

For the risk ratio, the statistic is computed on the log scale with θ0 = 1: Z = (ln 1.6006 − ln 1) / 0.158 = 0.4704 / 0.158 = 2.98, with a two-sided P-value of 0.003.

For the risk ratio the two tests give the same P-value (0.003), and both say only that chance is an unlikely explanation, not how strong the association is. The smallest expected count in the smoking table is about 26, so the χ² approximation is valid there; had any expected count fallen below 5, Fisher’s exact test would have been the better choice. The χ² statistic also grows with the size of the study: for the same proportions, doubling every cell of the table doubles χ² and shrinks the P-value, which Calculator 6.14 lets you check. A large χ² therefore shows that an association is unlikely to be due to chance, not that it is strong.

6.3.2 Choosing a Test for Other Outcomes and Comparisons

The Pearson χ², Fisher’s exact, Wald, and likelihood ratio tests above are the workhorses for 2×2 tables and the regression-based measures of association. But epidemiological analyses often require comparing means, proportions, or whole distributions across groups, sometimes paired, sometimes not, sometimes badly skewed. The right test then depends on three structural questions about your data: the outcome type, whether continuous (means or medians) or categorical (counts or proportions); the group structure, whether one group, two groups, or three or more, and whether independent or paired (matched); and the distributional assumptions, whether you can defend approximate normality (or appeal to a large-sample central limit theorem argument) or need a non-parametric (rank-based) alternative.

The matrix in Table 6.4 answers those three questions together. Find the row whose outcome type and comparison structure match your data, then read across to the test. The parametric column assumes approximate normality (or a large enough sample to appeal to the central limit theorem), and the rank-based column is the fallback when that assumption cannot be defended, when the sample is small, or when the outcome is ordinal and the spacing between its categories is not meaningful. For the 2×2 comparisons of this lesson, the binary rows give the three choices to remember: the Pearson χ² for two independent groups, Fisher’s exact test when any expected count is below 5, and McNemar’s test on the discordant pairs for matched data, reported with the matched odds ratio b/c.

Table 6.4. Choosing a statistical test by outcome type, comparison structure and distributional assumptions.

Outcome variableComparisonParametric / standard testRank-based or exact alternativeWhat to report alongside the p-value
ContinuousOne group against a hypothesised valueOne-sample t-testWilcoxon signed-rankMean (or median) difference with 95% CI
ContinuousTwo independent groupsTwo-sample t-test (usually the Welch version)Mann–Whitney U (Wilcoxon rank-sum)Difference in means with 95% CI
ContinuousTwo paired observationsPaired t-testWilcoxon signed-rank on the differencesMean within-pair change with 95% CI
ContinuousThree or more independent groupsOne-way ANOVAKruskal–WallisGroup means, then pairwise contrasts with a multiplicity adjustment
ContinuousOrdered exposure categories (dose, quartile, stage)Linear contrast within ANOVA, or regression on the category scoreJonckheere–Terpstra trend testChange in mean per category step with 95% CI
OrdinalTwo independent groupsNot recommended; the numeric codes do not carry equal spacingMann–Whitney UMedian and IQR by group, or the proportional-odds OR from an ordinal model
OrdinalThree or more independent groupsNot recommended, as aboveKruskal–WallisMedian and IQR by group
OrdinalTwo paired observationsNot recommended, as aboveWilcoxon signed-rankMedian of the within-pair differences
BinaryTwo independent groupsPearson χ²Fisher’s exact test when any expected count is below 5Risk difference, risk ratio, or odds ratio with 95% CI
BinaryPaired or matchedMcNemar’s χ² on the discordant pairsExact binomial on the discordant pairsMatched odds ratio b/c with 95% CI
BinaryOrdered exposure categoriesCochran–Armitage test for trendExact trend testOdds ratio per category step with 95% CI
Nominal, 3+ categoriesAny categorical exposurePearson χ² on the r × c tableFisher–Freeman–Halton exact testCramér’s V, or category-specific odds ratios against a reference
Two continuousAssociation between themPearson r (linear)Spearman ρ (monotonic)Coefficient with 95% CI
Ordinal with ordinal or continuousAssociation between themNot recommended, as aboveSpearman ρ, or Kendall’s τb when ties are commonCoefficient with 95% CI

The McNemar row of Table 6.4 needs a table of its own. In a pair-matched case-control study each case is matched to one control, and the pair, not the person, is the unit of analysis. The pairs are classified by the exposure of both members: pairs in which both were exposed, pairs in which neither was, pairs in which only the case was exposed (b) and pairs in which only the control was exposed (c). These b and c count pairs, so they are not the cells b1 and b0 of Table 6.1. Pairs in which the case and the control had the same exposure, the concordant pairs, carry no information about the association, because within such a pair exposure cannot tell the case from the control. Only the discordant pairs are used. By this we mean that each discordant pair acts as a small piece of evidence pointing one way or the other. A pair in which only the case was exposed suggests that exposure goes with disease, while a pair in which only the control was exposed suggests the opposite. If the exposure had no association with the disease, the two kinds of discordant pair would be expected to occur about equally often, so the question becomes whether one kind clearly outnumbers the other. For this reason the matched odds ratio is b/c, and McNemar’s test compares them with χ² = (b − c)² ÷ (b + c) on 1 degree of freedom. Worked Example 6.17 sets out a matched-pairs table and applies both formulae; notice that two of its four cells are never used.

Worked Example 6.17: A Matched Odds Ratio and McNemar’s Test

A pair-matched case-control study enrols 120 adults admitted with a first hip fracture, each matched by age and sex to one control from the same community, and asks whether each person had been taking a sleeping medication (hypothetical data). The table classifies the 120 pairs by the exposure of the case (columns) and of the matched control (rows).

Case exposedCase not exposedTotal
Control exposed18 (concordant)c = 1230
Control not exposedb = 3060 (concordant)90
Total4872120 pairs
  • The 18 pairs in which both members were exposed and the 60 in which neither was are concordant and are set aside.
  • Matched OR = b ÷ c = 30 ÷ 12 = 2.5.
  • McNemar’s χ² = (30 − 12)² ÷ (30 + 12) = 324 ÷ 42 = 7.71 on 1 degree of freedom, P ≈ 0.005.

Interpretation. The odds of having taken a sleeping medication were 2.5 times as high among people with a hip fracture as among their matched controls, and an imbalance this large between the two kinds of discordant pair would be unlikely if there were no association. Analysing the 240 people as two independent groups would ignore the matching.

The example shows that a matched analysis uses only the 42 discordant pairs: the matched OR is their ratio, and McNemar’s test asks whether they are more unbalanced than chance would allow.

The matrix narrows the choice, but it cannot make every decision, and the callout below names three judgements that stay with the analyst.

Three Things the Matrix Does Not Decide for You

  • The measurement scale is a claim the analyst makes about the data, and storing a variable as a number does not settle it. A 1-to-5 satisfaction item stored as a number is ordinal, and averaging it assumes the step from 1 to 2 is the same size as the step from 4 to 5.
  • A non-significant test is not evidence that no association exists. Every row of the matrix names an effect estimate for this reason, because the estimate and its confidence interval answer the question the study actually asked.
  • Rank-based tests answer a different question from their parametric counterparts. The Mann–Whitney U compares whole distributions, and it can be read as a comparison of medians only when the two distributions have a similar shape.

Table 6.5 summarises, for the tests most commonly reported alongside measures of association, when each is used, how it is calculated and how it is interpreted.

Table 6.5. When each test is used, how it is calculated and how it is interpreted.

TestWhen to useHow calculatedHow to interpret
One-sample t-test Compare a single sample mean to a known/hypothesised value μ0; continuous, approximately normal (or n large). t = (x̄ − μ0) / (s/√n); df = n−1. If p < α (or 95% CI for the mean excludes μ0), the population mean differs from μ0.
Two-sample (independent) t-test Compare means of 2 independent groups; continuous, approximately normal. Welch’s version does not assume equal variances. t = (x̄1 − x̄2) / SEdiff; df = n1+n2−2 (Student) or Welch–Satterthwaite df (Welch). p < α → the two group means differ. Always report the mean difference and its 95% CI.
Paired t-test Two related observations on the same unit (before/after, twin pairs, matched cases-controls); continuous differences approximately normal. Compute within-pair differences di; t = d̄ / (sd/√n); df = n−1. p < α → mean within-pair change ≠ 0. Reduces between-subject variability, and is usually more powerful than treating data as unpaired.
One-way ANOVA Compare means across 3+ independent groups; continuous, approximately normal, roughly equal variances. F = MSbetween / MSwithin; df1 = k−1, df2 = N−k. Significant F → at least one group mean differs. Follow with post-hoc pairwise comparisons (Tukey HSD, Bonferroni) to see which.
Pearson χ² Test independence of two categorical variables (any r×c table). Assumption: all expected counts > 1 and ≥ 80% > 5. χ² = Σ(O−E)²/E; df = (r−1)(c−1). Expected = (row total × column total) / N. p < α → row and column variables are associated. The test signals whether there is association; report a measure of association (OR, RR) for strength.
Fisher’s exact test 2×2 (or larger) tables with small expected counts where the χ² approximation is suspect. Hypergeometric: enumerates every table with the same margins, sums probabilities of tables as extreme or more extreme than observed. Exact p-value, with no large-sample assumption. With modern computing, fine to use even when χ² would also be valid.
McNemar’s test Paired/matched binary outcomes (before/after on the same person; matched case-control on exposure; agreement of two diagnostic tests). Look only at discordant pairs b and c: χ² = (b−c)² / (b+c); df = 1. Concordant pairs ignored. p < α → the discordant pairs are unbalanced, so a real change/effect exists. The matched OR is b/c.
Wilcoxon signed-rank Non-parametric alternative to one-sample / paired t-test. Use when differences are skewed, ordinal, or have outliers. Rank |di|, attach signs, sum positive (or negative) ranks; compare to its null distribution (or large-sample z). p < α → the median difference is non-zero (under symmetry). Robust to outliers.
Mann–Whitney U (Wilcoxon rank-sum) Non-parametric alternative to two-sample t-test. Two independent groups, continuous or ordinal. Pool all observations, rank them, sum ranks in one group; U = R1 − n1(n1+1)/2. p < α → the two distributions differ. If shapes are similar, this is a test of medians; otherwise it tests stochastic dominance.
Kruskal–Wallis Non-parametric alternative to one-way ANOVA. 3+ independent groups; continuous or ordinal. H from rank sums; approximately χ²k−1 under H0. p < α → at least one group’s distribution differs. Follow with pairwise rank-sum tests (Dunn’s test or pairwise Wilcoxon with adjustment).
Pearson correlation (r) Linear association between two continuous variables; assumes approximate bivariate normality, sensitive to outliers. r = Σ(x−x̄)(y−ȳ) / √[Σ(x−x̄)² Σ(y−ȳ)²]; tested via t = r√[(n−2)/(1−r²)]. Range −1 to +1: sign = direction, magnitude = strength of linear association. Always inspect a scatterplot first.
Spearman ρ Monotonic (not necessarily linear) association between two ordinal or non-normally distributed continuous variables. Pearson r applied to the ranks of x and y. Same −1 to +1 interpretation, but for monotonic association. Robust to outliers and non-linearity.

Table 6.5 recommends pairwise comparisons after a significant one-way ANOVA or Kruskal–Wallis test, and such comparisons need an adjusted significance threshold. The recall panel below restates the Bonferroni correction from HSCI 230 and its alternatives; try the retrieval question before opening the answer.

Recall: Multiple Comparisons

HSCI 230 Lesson 5, Section 7 (Randomised Controlled Trials: Conduct, Analysis, Vaccine Trials and Reporting), introduced the Bonferroni correction for multiple comparisons: when k comparisons are tested, each is tested at α ÷ k, so that the chance of at least one false positive across all of them stays at or below α. After a significant one-way ANOVA the comparisons are usually all pairs of groups. The Holm procedure is a less conservative alternative to Bonferroni: it orders the p-values from smallest to largest, tests them in turn against α ÷ k, α ÷ (k − 1) and so on, and stops at the first one that is not significant. Tukey's honestly significant difference is the usual method for all pairwise comparisons of means after ANOVA. A single primary outcome and comparison, named before the data are analyzed, avoids most multiplicity problems.

Retrieval question. A study makes four comparisons with an overall α of 0.05. What threshold does the Bonferroni correction set for each comparison?

Show the answer▼

0.05 ÷ 4 = 0.0125.

Every test in Table 6.4 and Table 6.5 yields a P-value, and the confidence interval carries more information than the P-value alone.

6.4 Confidence Intervals

6.4.1 Computing Confidence Intervals

A confidence interval (CI) reflects the level of uncertainty in a point estimate. A 95% CI means that if the study were repeated many times under identical conditions, 95% of the computed CIs would contain the true parameter value. This is a property of the procedure, not a probability statement about the parameter (Greenland et al., 2016). By this we mean that the true value is fixed, even though we do not know it, and it is the intervals that vary from one study to the next because each study draws a different sample. The 95 per cent describes how often the method succeeds over many repetitions, much as a fishing net that catches the fish 95 times out of 100 casts. Once a particular interval has been calculated, it either contains the true value or it does not, so it is not correct to say that there is a 95 per cent probability that the true value lies inside that one interval. In practice, the interval is best read as the range of values that are reasonably compatible with the data. The method of computing a confidence interval depends on whether the measure is a difference or a ratio. For difference measures, the CI is computed directly from the standard error, as Equation 6.16 shows.

Confidence interval, difference measures
θ ± Zα × √var(θ) Eq 6.16
A symmetric interval places the point estimate at the centre, then steps out by a critical value times the standard error.

Worked Example 6.18 applies Equation 6.16 to the risk difference of the smoking cohort, using the standard error from Worked Example 6.12.

Worked Example 6.18: CI for the Risk Difference

For the smoking cohort, RD = 0.0428 and var(RD) = 0.000302, so SE(RD) = √0.000302 = 0.0174.

0.0428 ± 1.96 × 0.0174 = 0.0428 ± 0.0341, giving a 95% CI from 0.0087 to 0.0769 (0.009 to 0.077 after rounding, as in Table 6.6).

The interval excludes 0. It indicates that smoking is compatible with anywhere from about 1 to 8 additional low-birth-weight babies per 100 women who smoke.

The interval steps the same distance below and above the estimate, as every interval for a difference measure does. Calculator 6.16 reproduces it.

For ratio measures, the CI is computed on the log scale, where the standard errors of 6.2.1 apply, and then exponentiated, as Equation 6.17 shows. The CI is symmetrical about lnθ but not about θ itself, which is why confidence intervals for ratio measures appear asymmetric. By this we mean that the limits are first placed an equal distance below and above ln θ, and exponentiating then turns that equal step on the log scale into multiplication by the same factor. Dividing by a factor moves the estimate down less than multiplying by it moves the estimate up. For example, an RR of 2 with a factor of 1.5 gives limits of 2 ÷ 1.5 = 1.33 and 2 × 1.5 = 3, so the upper limit lies 1 above the estimate while the lower limit lies only 0.67 below it.

Confidence interval, ratio measures
θ × exp(± Zα × √var(ln θ)) Eq 6.17
For a ratio measure the interval is built on the log scale, then converted back by exponentiating, so it ends up asymmetric around the point estimate.

Worked Example 6.19 applies Equation 6.17 to the risk ratio and compares the result with the interval for the odds ratio; notice how much further the upper limit lies from the estimate than the lower limit.

Worked Example 6.19: CI for the Risk Ratio

For the smoking cohort, RR = 1.6006 and var(ln RR) = 0.024957, so SE(ln RR) = 0.15798 and the margin on the log scale is 1.96 × 0.15798 = 0.30964.

  • Lower limit = 1.6006 × e−0.30964 = 1.6006 × 0.73371 = 1.174.
  • Upper limit = 1.6006 × e0.30964 = 1.6006 × 1.36293 = 2.182.

The interval runs from about 0.43 below the estimate to about 0.58 above it, the asymmetry described above. Applying the same formula to the odds ratio (OR = 1.678, var(ln OR) = 0.031468) gives 1.185 to 2.375. Table 6.6 lists 1.154 to 2.387 for the OR because that interval was computed by the exact method, and small differences of this kind between interval methods are expected.

Calculator 6.17 reproduces both intervals. Once an interval has been computed, the remaining task is to read it correctly.

6.4.2 Interpreting Confidence Intervals

A computed interval can be read in two ways: as a test of the null value and as a range of plausible effect sizes. The callout below gives both readings, and the first is a rule to commit to memory.

Interpreting CIs for Measures of Association

  • For RR, IR, OR: if the 95% CI includes 1, the association is not statistically significant at α = 0.05.
  • For RD, ID: if the 95% CI includes 0, the association is not statistically significant.

However, this “surrogate significance test” is an under-use of the CI. The CI also shows the range of plausible effect sizes, which is far more informative than a binary significant/non-significant classification. For example, an RR of 1.1 (95% CI 1.02 to 1.19) and an RR of 3.0 (95% CI 1.05 to 8.6) both exclude 1, yet the first describes a small effect estimated precisely and the second a possibly large effect estimated very imprecisely. Reading only whether 1 is excluded would treat these two results as the same, whereas reading the whole interval shows how different they are.

When a study reports an interval, or an exam question gives you one, a complete interpretation takes four steps. First, state the point estimate in words, naming the comparison and the time frame. Second, give the range of values the data are compatible with, translated into the same words. Third, say whether the interval includes the null value (1 for a ratio, 0 for a difference), and therefore whether the result is statistically significant at the 5 per cent level. Fourth, comment on precision: a wide interval means the estimate is imprecise. Table 6.6 lists the point estimates and 95% confidence intervals for the smoking and migraine examples, with the odds ratio interval computed by the exact method, and supplies the numbers for an example.

Table 6.6. Point estimates and 95% confidence intervals for the smoking and migraine examples (odds ratio interval by the exact method).

MeasurePoint Estimate95% CI
RD (smoking)0.043(0.009, 0.077)
RR (smoking)1.601(1.174, 2.182)
OR (smoking)1.678(1.154, 2.387)
ID (migraine)0.338(0.232, 0.443)
IR (migraine)2.811(1.983, 4.050)

Using Table 6.6: “The risk of a low-birth-weight baby was 1.60 times as high among women who smoked as among women who did not. The data are compatible with risk ratios from about 1.17 to 2.18, so the interval excludes 1 and the association is statistically significant at the 5 per cent level.” None of the intervals in the table for ratio measures includes 1, and none for difference measures includes 0, confirming statistical significance for all the associations. Remember that a confidence interval describes random error only. It cannot show that the association is free of selection bias, information bias or confounding. Keeping the width of an interval separate from the size of the estimate leads to the last topic of the lesson.

6.5 Statistical Significance and Strength of Association

Strength, significance and precision are three different properties of a result. Strength is the size of the measure of association, significance is whether chance alone is an unlikely explanation, and precision is how narrow the confidence interval is. Section 1 drew the first distinction in 6.1, and the tools of this section show why it holds. Because standard errors shrink as a study grows, a very large study can make a weak association highly significant, as in the retrieval question of 6.2.2, where a risk ratio of 1.03 had p = 0.001 in a study of 200,000 people. The same holds for the Pearson χ², which 6.3.1 showed grows in proportion to the sample size when the proportions stay the same. Conversely, a small study can leave a strong association non-significant, which reflects imprecision rather than evidence of no effect. The callout below states the rule that follows for every test in this section.

Pair Every Test with an Effect Estimate

None of the tests above are themselves measures of effect size. A statistically significant chi-square confirms that some association exists, but the magnitude must come from the OR, RR, mean difference, or correlation coefficient. Likewise, a non-significant t-test in a small study is not evidence of no effect. Always pair every test with the corresponding effect estimate and its 95% CI.

When two studies are compared, each of the three properties should therefore be discussed separately, and an interval that includes the null value should be described as compatible both with no effect and with an effect of meaningful size, rather than as showing that there is no effect.

This section has defined the standard error of each measure, set out the test statistics in common use for measures of association, used the standard error in the Wald statistic and in confidence intervals, shown how a test is chosen for other kinds of outcome and comparison, and separated statistical significance from the strength of an association. The key takeaways and the knowledge check below consolidate the section, including a comparison of two studies by strength of association, statistical significance and precision.

Key Takeaways

  • 6.2 The standard error describes how precise a point estimate is, that is, how much uncertainty surrounds it; precision improves mainly by enlarging the smallest group, as the smoking cohort shows, where almost all the variance came from the 351 smokers. For difference measures the variance is calculated directly, while for ratio measures it is calculated on the log scale, because ln(RR) and ln(OR) have roughly symmetric sampling distributions. A hypothesis test starts from the null hypothesis of no association, θ = 0 for a difference and θ = 1 for a ratio, and the P-value is the probability of data at least as extreme as those observed if that null hypothesis were true. Two traps come up again and again: the P-value is not the probability that the null hypothesis is true, and it is not a measure of how large the effect is. When you use a P-value in an answer, state the null value being tested, report the actual P-value rather than only “significant” or “not significant”, and pair it with the effect estimate.
  • 6.3 Four test statistics are in common use for measures of association: the Pearson χ², exact tests such as Fisher’s exact test, the Wald statistic and the likelihood ratio test. For a 2×2 table, the choice usually comes down to three situations. Use the Pearson χ² for two independent groups when every expected count (row total × column total ÷ grand total) is 5 or more, use Fisher’s exact test when any expected count is below 5, and use McNemar’s test for matched pairs, because it uses only the discordant pairs, b and c in the matched-pairs table, and gives the matched OR as b/c. Base the choice on the expected counts, not the observed ones, and check first whether the observations are independent or paired. Beyond 2×2 tables, Table 6.4 guides the choice through three questions: the type of outcome, the number of groups and whether they are independent or paired, and whether approximate normality can be defended.
  • 6.4 A 95 per cent confidence interval gives the range of effect sizes that are compatible with the data; strictly, if the study were repeated many times, 95 per cent of the intervals calculated would contain the true value. Interpret a given interval in four steps: state the point estimate in words, give the range in the same words, say whether the interval includes the null value (1 for a ratio, 0 for a difference) and so whether the result is statistically significant at the 5 per cent level, and comment on its width, which shows how precise the estimate is. For example, the smoking RR of 1.60 (95% CI 1.17 to 2.18) means the data are compatible with a risk anywhere from about 1.2 to 2.2 times as high among women who smoked, and because the interval excludes 1 the association is statistically significant. Ratio intervals are built on the log scale and then exponentiated, so they stretch further above the estimate than below it, and this asymmetry is expected rather than an error. Finish by remembering that an interval reflects random error only: it cannot show that a result is free of selection bias, information bias or confounding.
  • 6.5 Strength, significance and precision are three different things, and keeping them apart is one of the most useful habits you can take from this lesson. Strength is the size of the measure of association, significance is whether chance alone is an unlikely explanation, and precision is how narrow the confidence interval is, which is why an interval tells you more than a P-value alone. A very large study can make a weak association highly significant (an RR of 1.03 with P = 0.001 in a study of 200,000 people), and a small study can leave a strong association non-significant, which reflects imprecision, not evidence of no effect. When a question asks you to compare two studies, comment on each of the three separately, and never conclude that an exposure or treatment “has no effect” just because its interval includes the null value; when the interval is wide, say instead that the data are compatible both with no effect and with an effect of meaningful size.
Knowledge Check: this section

1. A cohort study of construction workers reports that the 10-year risk of hearing loss was 1.8 times as high among workers who did not wear hearing protection as among those who did (95% CI 1.2 to 2.7). Which interpretation is correct?

The 95% describes the procedure: 95% of intervals computed this way would contain the true value. A confidence interval reflects random error only, so it cannot show causation.
This states the estimate, gives the compatible range, checks the null value of 1 and avoids a causal claim. A confidence interval describes random error only, not bias or confounding.
Significance is judged against the null value for a ratio, which is 1, not 2. The interval 1.2 to 2.7 excludes 1, so the result is significant at the 5% level.
A risk ratio compares two groups; it does not belong to individual workers. The interval gives values of the group-level RR compatible with the data.
Correct answer: the second option, which follows the four steps given in 6.4.2. State the estimate in words (a 10-year risk 1.8 times as high), give the range the data are compatible with (1.2 to 2.7), check the null value (the interval excludes 1, so the result is significant at the 5% level) and comment on precision (a moderately wide interval). A confidence interval describes random error only, so it cannot show causation or rule out bias or confounding, and the 95% describes the procedure (95% of intervals computed this way would contain the true value), not a probability for this one interval. Significance is judged against the null value 1, not against 2, and a risk ratio compares two groups, so it does not belong to individual workers. (Section 4, 6.4.2 Interpreting Confidence Intervals, the callout “Interpreting CIs for Measures of Association” and the four steps after it; Objective 6.4)

2. A case-control study reports OR = 2.0 with a 95% CI of 1.25 to 3.20 (the standard error of ln OR is 0.24). A classmate says the interval must contain an error, because the upper limit is 1.20 above the estimate but the lower limit is only 0.75 below it. Which reply is correct?

Intervals for ratio measures are built symmetrically for ln θ and then exponentiated (Eq 6.17). Equal steps on the log scale become a shorter step below the OR and a longer one above it.
Only intervals for difference measures such as RD and ID are symmetric on the original scale. Ratio intervals are symmetric on the log scale, so they are asymmetric once converted back.
The number of controls per case affects the size of the standard error, not the shape of the interval. The asymmetry comes from building the interval on the log scale.
An interval reflects random error only and contains no correction for bias. Its asymmetry arises because it is calculated for ln OR and then exponentiated.
Correct answer: the first reply. Intervals for RR, IR and OR are calculated for ln θ, whose sampling distribution is roughly symmetric, and then converted back with Eq 6.17, θ × exp(± Z × √var(ln θ)). Here ln 2.0 = 0.693 and the limits on the log scale are 0.693 ± 1.96 × 0.24 = 0.693 ± 0.470, that is 0.223 and 1.163, which exponentiate to 1.25 and 3.20: equal steps on the log scale become a shorter step below the OR and a longer step above it. Only intervals for difference measures (RD, ID) are symmetric on the original scale, and an interval reflects random error, so its shape has nothing to do with bias or with the number of controls per case. (Section 4, 6.4.1 Computing Confidence Intervals, Eq 6.17 and Worked Example 6.19; Objective 6.4)

3. An analyst has three 2×2 comparisons to test: (i) a cohort of 2,000 workers in which the smallest expected cell count is 40; (ii) an outbreak among 24 wedding guests in which one expected cell count is 2.5; (iii) a case-control study of 150 cases, each pair-matched to one control. Which tests should she use?

This reverses (i) and (ii). Pearson χ² suits independent groups with every expected count at least 5, as in (i); with an expected count of 2.5, as in (ii), Fisher's exact test is used.
McNemar's test is for matched pairs, not for small samples, so it belongs to (iii). The small wedding outbreak in (ii) calls for Fisher's exact test.
Pearson χ² is unreliable when an expected count falls below 5, as in (ii), and it assumes independent groups, which matched pairs in (iii) are not.
Pearson χ² suits independent groups with expected counts of 5 or more; Fisher's exact test handles small expected counts; and matched pairs need McNemar's test on the discordant pairs.
Correct answer: the last option. The information needed is whether the groups are independent and how large the expected counts are. The Pearson χ² suits two independent groups when every expected count is 5 or more, as in (i); when an expected count is below 5, as in (ii), the χ² approximation is unreliable and Fisher’s exact test gives an exact P-value instead. Matched pairs, as in (iii), are not independent groups, so they are analysed with McNemar’s test on the discordant pairs and reported with the matched OR b/c; a Pearson χ² on all 300 people ignores the matching. (Section 4, the slide “Tests for 2x2 tables and regression”, and 6.3.2 Choosing a Test for Other Outcomes and Comparisons, Table 6.4 and Table 6.5; Objective 6.3)

4. A small cohort study of a workplace solvent and kidney disease reports RR = 2.0 (95% CI 0.8 to 5.0), P = 0.14. A manager concludes that “the solvent has no effect on kidney disease.” Which response is best?

Not rejecting the null hypothesis is not evidence that it is true. The interval 0.8 to 5.0 is compatible with no effect and also with a five-fold increase.
A P-value is the probability of data at least as extreme as these if there were no association. It is not the probability that the null hypothesis is true.
Strength, significance and precision are separate: an RR of 2.0 suggests a doubling, but the wide interval including 1 shows the small study is imprecise and inconclusive.
A point estimate of 2.0 does not establish causation, and the P-value is not wrong. The interval from 0.8 to 5.0 shows the data are also compatible with no association.
Correct answer: the third response. Strength, significance and precision are separate things: the RR of 2.0 describes a fairly strong association; the P-value of 0.14, like the interval that includes 1, means the data are also compatible with no association at the 5% level; and the width of the interval (0.8 to 5.0) shows that the small study estimated the RR imprecisely. Not rejecting H0 (RR = 1) is not evidence that H0 is true, and a P-value is the probability of data at least as extreme as these if there were no association, not the probability that the null hypothesis is true. (Section 4, 6.2.2 Null Hypotheses and P-Values and the callouts “Recall: P-values and Practical Significance” and “Limitations of P-values”, the callout “Three Things the Matrix Does Not Decide for You” in 6.3.2, and 6.5 Statistical Significance and Strength of Association; Objectives 6.2 and 6.5)

5. Two cohort studies examined regular use of a proton-pump inhibitor (a heartburn medication) and hip fracture over five years in adults aged 65 and over (hypothetical data). Study 1 followed 900 adults and reported RR = 1.80 (95% CI 0.90 to 3.60), P = 0.10. Study 2 used the health records of 60,000 adults and reported RR = 1.20 (95% CI 1.10 to 1.31), P < 0.001. Which comparison is correct?

A smaller P-value does not mean a stronger association. Strength is read from the point estimates, and Study 1's RR of 1.80 is larger than Study 2's 1.20; Study 2's tiny P-value mainly reflects its size.
Ratio intervals are built on the log scale and then exponentiated, so they reach further above the estimate than below it. Study 1's interval is asymmetric by construction, not by error.
A non-significant result from a small study is inconclusive, not evidence of no effect. Study 1's interval, 0.90 to 3.60, includes values well above 1.
Strength comes from the point estimates, significance from whether each interval includes 1, and precision from interval width. Study 1 is stronger but imprecise; Study 2 is weaker but precise.
Correct answer: the last comparison. The information needed is the three separate properties of each result. Strength is read from the point estimates (1.80 against 1.20), significance from whether each interval includes 1 (Study 1’s does, P = 0.10; Study 2’s does not, P < 0.001), and precision from the width of the intervals (an upper limit four times the lower in Study 1 and about 1.2 times the lower in Study 2, mainly because Study 2 is so much larger). Study 1’s interval is not an error: a ratio interval is built on the log scale and then exponentiated, so it reaches further above the estimate than below it. A smaller P-value does not mean a stronger association, and a non-significant result from a small study is inconclusive rather than evidence of no effect. (Section 4, 6.4.1 Computing Confidence Intervals and 6.5 Statistical Significance and Strength of Association; Objectives 6.4 and 6.5)

6. A cohort study of 60,000 adults aged 65 and over found a five-year risk ratio for hip fracture of 1.20 (95% CI 1.10 to 1.31, P < 0.001) among regular users of a heartburn medication compared with non-users. A pharmacist says the study “proves the medication causes hip fractures.” Which response is best?

A P-value is not the probability that the exposure causes the outcome. It shows only that data like these would be unlikely if there were no association.
P < 0.001 and a narrow interval make chance unlikely and show precision, but both describe random error only. Confounding, such as by frailty, could still explain the association.
A narrow interval shows precision, which reflects random error only. No amount of precision removes bias or confounding, such as frailer people being both more likely to take the medication and to fall.
A weak association can still matter when the exposure is common, because a small increase in risk across many people adds up to many extra cases.
Correct answer: the second response. The information needed is what a P-value and a confidence interval can and cannot show. P < 0.001 and an interval that excludes 1 make chance an unlikely explanation, and the narrow interval shows a precise estimate, but both describe random error only: frailer people with more illnesses may be both more likely to take the medication and more likely to fall, which is confounding that no amount of precision removes. The P-value is not the probability that the medication causes fractures, and a weak association can still matter for public health when the exposure is common, because a small increase in risk applied to many people adds up to many extra cases. (Section 4, 6.2.2 Null Hypotheses and P-Values, 6.4.2 Interpreting Confidence Intervals and 6.5 Statistical Significance and Strength of Association; Objectives 6.2, 6.4 and 6.5)

✦ Pass the knowledge check with 100% to continue

Section 5

Final Review & Assessment

⏱ Estimated time: 20 minutes

Bringing It All Together

This lesson translated measures of disease frequency into measures of association between an exposure and an outcome. You worked through the three ratio measures (RR, IR, OR), the difference measures of effect in the exposed (RD, AFe), the population-level extensions (PAR, AFp), and the inferential machinery (standard errors, hypothesis tests, confidence intervals) that surrounds every estimate.

The measures you computed here are the outputs that study designs deliver, the quantities in which bias and confounding are judged, and the estimates that systematic reviews eventually pool. For deeper treatment of these measures and their statistical foundations, see Greenland and Pearce (2015) and the historical landmark papers by Cornfield (1951) and Doll and Hill (1950).

Key Takeaways from this lesson

  1. You can now set out risk, rate and case-control data in the course’s 2×2 layout, with exposure in the columns and disease in the rows, and calculate and interpret the RR, the IR and the OR, as well as the prevalence ratio from a cross-sectional survey. Each one divides disease frequency in the exposed group by that in the non-exposed group and is read against the null value of 1. The most useful exam habit is to label every cell with its symbol before you calculate, then write an interpretation that names the measure, both groups and the time frame, turning a protective ratio into a percentage reduction (1 − RR).
  2. You can explain why a case-control study, which fixes the numbers of cases and controls, cannot give risks and so relies on the symmetric odds ratio, and why the OR approximates the RR only when the disease is rare (a risk below about 5 per cent). You can also use Table 6.3 to say which measures cohort, case-control and cross-sectional designs can each supply. In an exam, identify the design from how participants were selected before you choose a measure, and call an OR “times the risk” only after you have checked the rare-disease condition.
  3. You can calculate and interpret the RD and the ID, whose null value is 0, and turn an RD into an NNH or NNT, quoting each with its units, its time frame and the causal assumption. You also know that ratio measures describe relative strength, the clinician’s question, while difference measures count extra cases, the planner’s question. The key reminder is to translate a difference into extra cases per 100 (or per 1,000) people, or per unit of person-time, so that your reader can see the real size of the problem.
  4. You can calculate the AFe from the risks, from the RR or, in a case-control study, from the OR, and calculate vaccine efficacy as the special case in which being unvaccinated is the exposure. Both describe the share of cases among the exposed (or the vaccinated) that the exposure caused (or the vaccine prevented). Always add “assuming the association is causal”, and remember that a vaccine efficacy of 75 per cent does not mean that 75 per cent of vaccinated people are immune.
  5. You can calculate the PAR and the AFp, interpret the AFp as the share of all cases in the population that would be avoided if the exposure were removed, and explain why a common weak risk factor can outweigh a rare strong one. That interpretation rests on two assumptions, causation and no confounding, and when confounding is present you use an adjusted RR (Equation 6.10). In an exam, state both assumptions and check whether the question is about the exposed group (AFe) or the whole population (AFp).
  6. You can interpret a given confidence interval, P-value or test result for a measure of association, explain why ratio intervals are asymmetric, and choose the Pearson χ², Fisher’s exact test or McNemar’s test for a 2×2 comparison. You keep strength, significance and precision apart, and you know that a non-significant result is not evidence of no effect. For any interval you are given, work through the four steps (estimate, range, null value, precision) and finish by noting that the interval says nothing about bias or confounding.

Final Knowledge Assessment

Complete all 17 questions below with 100% accuracy to finish this lesson. The Submit button unlocks once you have passed every section knowledge check.

Final Assessment: Measures of Association

1. During one month, 18 of 300 children in daycare centres with hand-washing stations and 50 of 500 children in centres without them had gastroenteritis. Which answer gives the risk ratio for centres with stations (compared with centres without) and reads it correctly?

Dividing 0.10 by 0.06 makes centres without stations the exposed group. With stations as the exposure, RR = 0.06 ÷ 0.10 = 0.60.
RR = (18/300) ÷ (50/500) = 0.06 ÷ 0.10 = 0.60. A ratio below 1 is protective and is read as a reduction of 1 − RR = 0.40, so the risk was 40% lower.
This reads the RR itself as the reduction. An RR of 0.60 means the risk was 60% of the comparison risk, which is 40% lower, not 60% lower.
0.04 is the size of the risk difference (0.06 − 0.10 = −0.04), an absolute measure, not a ratio. The RR is 0.06 ÷ 0.10 = 0.60.
Correct answer: RR = 0.60, a 40% lower risk. With the centres with stations as the exposed group, RR = (a1/n1) ÷ (a0/n0) = (18/300) ÷ (50/500) = 0.06 ÷ 0.10 = 0.60. A ratio below 1 is protective and is read as a percentage reduction, 1 − RR = 0.40, so the risk was 40% lower; “60% lower” wrongly reads the RR itself as the reduction (the risk is 60% of the comparison risk, not 60% below it). Dividing 0.10 by 0.06 (1.67) makes the centres without stations the exposed group, and 0.04 is the size of the risk difference (RD = −0.04), not a ratio. (Section 1, 1.2.1 The Risk Ratio and the Odds Ratio, Eq 6.1 and Worked Example 6.1; Objective 1.2)

2. Over five years, surveillance recorded 36 cases of invasive pneumococcal disease among adults aged 65 and over living in long-term care (40,000 person-years) and 150 cases among adults aged 65 and over living in the community (500,000 person-years). Which answer is correct?

Each rate is cases ÷ person-time: 36 ÷ 40,000 gives 90 and 150 ÷ 500,000 gives 30 per 100,000 person-years. IR = 90 ÷ 30 = 3.0, a ratio of rates.
Comparing case counts ignores person-time: the community group contributed 12.5 times as much. Once each count is divided by its person-time, the long-term care rate is three times as high.
Dividing the community rate by the long-term care rate reverses the reference group. With long-term care as the exposed group, IR = 90 ÷ 30 = 3.0.
The rates are right, but 60 per 100,000 person-years is the incidence rate difference, an absolute measure. The ratio is 90 ÷ 30 = 3.0, with no units.
Correct answer: rates of 90 and 30 per 100,000 person-years and IR = 3.0. Each rate is cases ÷ person-time: 36 ÷ 40,000 = 0.0009 per person-year, and 0.0009 × 100,000 = 90 per 100,000 person-years; 150 ÷ 500,000 = 0.0003, or 30 per 100,000 person-years; so IR = 90 ÷ 30 = 3.0. Comparing the case counts (36 ÷ 150) ignores that the community group contributed 12.5 times as much person-time, and 60 per 100,000 person-years is the incidence rate difference, an absolute measure, not the ratio. An IR of 0.33 divides the community rate by the long-term care rate, which reverses the reference group. (Section 1, 1.2.2 The Incidence Rate Ratio, Eq 6.3 and Worked Example 6.2; Objective 1.2)

3. A cross-sectional survey of 1,500 farm workers finds that 100 of the 400 workers who spray pesticides and 110 of the 1,100 who do not report frequent headaches. Which answer is correct?

A cross-sectional survey can give a prevalence ratio from the same cells as a risk ratio; it does not have to use an odds ratio. With 14% affected, the POR of 3.0 lies further from 1 than the PR.
The PR is right, but a survey measures prevalence at one time, not the risk of developing headaches. It also cannot show whether spraying came before the headaches.
PR = 0.25 ÷ 0.10 = 2.5. With 14% of workers affected the POR (3.0) lies noticeably further from 1, so the PR is reported, describing prevalence rather than risk of developing headaches.
0.15 is the prevalence difference, 0.25 − 0.10, an absolute measure, not a ratio. The prevalence ratio is 0.25 ÷ 0.10 = 2.5.
Correct answer: PR = 2.5, reported instead of the POR. The prevalence of frequent headaches is 100 ÷ 400 = 0.25 among workers who spray and 110 ÷ 1,100 = 0.10 among those who do not, so PR = 0.25 ÷ 0.10 = 2.5, and POR = (100 × 990) ÷ (110 × 300) = 99,000 ÷ 33,000 = 3.0. With 210 of 1,500 workers (14%) affected the outcome is not rare, so the POR is noticeably further from 1 than the PR, and the PR, calculated from the same cells as a risk ratio, is the measure to report. Because spraying and headaches were measured at the same time, the survey describes prevalence, not the risk of developing headaches, and it cannot show which came first; 0.15 is the prevalence difference, not a ratio. (Section 3, 2.2 Which Measures Each Study Design Can Supply, Table 6.3, and 1.3 The Prevalence Ratio in Cross-Sectional Surveys, Worked Example 6.9; Objective 1.3)

4. A case-control study of 150 never-smoking adults with lung cancer and 300 never-smoking controls measured radon in each person’s home (table below). Which answer gives the odds ratio for living in a high-radon home and interprets it correctly?

High radonLow radonTotal
Cases50100150
Controls50250300
Total100350450
The column totals reflect the investigator's choice of two controls per case, so 50/100 is not a risk. A case-control study cannot give a risk ratio.
The inverted cross-product, 0.40, is the OR for low radon. With high radon as the exposure, OR = (a1 × b0) ÷ (a0 × b1) = 12,500 ÷ 5,000 = 2.5.
An odds ratio is a ratio of odds, not a percentage of cases. Saying what share of cases is due to radon would need an attributable fraction and a causal assumption.
OR = (50 × 250) ÷ (100 × 50) = 2.5. By the symmetry of the OR, the ratio of the odds of exposure among cases and controls equals the ratio of the odds of disease.
Correct answer: OR = 2.5, with both readings. In the course layout a1 = 50 and a0 = 100 (cases) and b1 = 50 and b0 = 250 (controls), so OR = (a1 × b0) ÷ (a0 × b1) = 12,500 ÷ 5,000 = 2.5. Along the rows, the odds of high radon were 50/100 = 0.50 among cases and 50/250 = 0.20 among controls, and 0.50 ÷ 0.20 = 2.5; by the symmetry of the OR this equals the ratio of the odds of disease, which is why a case-control study can estimate it. The tempting “risk ratio” of 1.75 uses column totals that the investigator fixed by choosing two controls per case, so it is not a valid measure; the inverted cross-product, 0.40, is the OR for low radon, and an OR is not a percentage of cases. (Section 1, 1.2.1 The Risk Ratio and the Odds Ratio, Eq 6.2, and 2.1.2 The Odds Ratio in Case-Control Studies, Worked Example 6.3; Objective 2.1)

5. In a cohort of 800 patients followed for 30 days after knee surgery, 80 of the 200 patients who smoked and 120 of the 600 who did not developed a wound complication. A report gives the odds ratio, 2.67, and says that smokers had “2.67 times the risk” of a complication. Which comment is correct?

The OR approximates the RR only when the outcome is rare, below about 5%. Here 25% of patients had a complication, so the OR exaggerates the RR, whatever the design.
Risks of 0.40 and 0.20 give RR = 2.0. With a common outcome (25%) the OR of 2.67 lies further from 1, and a cohort, which can calculate risks directly, should report the RR.
Reporting the RR is right, but the RR lies closer to 1 than the OR, not further. The OR exaggerates the RR when the outcome is common.
The OR is calculated correctly. The OR and RR from the same table are expected to differ when the outcome is common; they agree closely only when it is rare.
Correct answer: report the RR of 2.0. The risks are 80 ÷ 200 = 0.40 and 120 ÷ 600 = 0.20, so RR = 2.0, whereas OR = (80 × 480) ÷ (120 × 120) = 38,400 ÷ 14,400 = 2.67. The OR approximates the RR only when the outcome is rare (below about 5%); here 200 of 800 patients (25%) had a complication, so the odds (0.40/0.60 = 0.67 and 0.20/0.80 = 0.25) are far from the risks and the OR exaggerates the RR. A cohort can calculate risks directly, so it should report the RR and must not describe the OR as “times the risk”; the OR lies further from 1 than the RR, not closer, and the two are expected to differ when the outcome is common. (Section 1, 2.1.1 When the Three Ratios Agree, Figure 6.1 and the panel “When is OR ≈ RR?”; Objective 2.1)

6. A cohort of adult commuters who cycle was followed for head injuries. Cyclists who rarely wore a helmet had 30 head injuries during 15,000 person-years of cycling, and cyclists who always wore one had 20 during 40,000 person-years. Which answer gives the incidence rate difference (ID) with its units and interprets it correctly?

Subtracting counts ignores that the two groups cycled for very different person-time (15,000 against 40,000 person-years). The ID subtracts the rates and carries their units.
2.0 ÷ 0.5 = 4.0 is the rate ratio, which has no units. The ID subtracts the rates: 2.0 − 0.5 = 1.5 per 1,000 person-years.
The rates are 30 ÷ 15,000 = 2.0 and 20 ÷ 40,000 = 0.5 per 1,000 person-years, so ID = 1.5 per 1,000 person-years, an excess rate if the association is causal.
The value 0.0015 is right, but an ID keeps the units of the rates: it is 0.0015 per person-year, an excess rate, not a percentage of cyclists.
Correct answer: 1.5 head injuries per 1,000 person-years. The rates are a1/t1 = 30 ÷ 15,000 = 0.002 per person-year (2.0 per 1,000) and a0/t0 = 20 ÷ 40,000 = 0.0005 per person-year (0.5 per 1,000), so ID = 0.002 − 0.0005 = 0.0015 per person-year, which is 1.5 per 1,000 person-years. The ID keeps the person-time units of the rates, so it is an excess rate, not a percentage of cyclists; subtracting the counts ignores the very different person-time of the two groups, and 2.0 ÷ 0.5 = 4.0 is the rate ratio, which has no units. (Section 2, 3.2.2 The Incidence Rate Difference, Eq 6.5, and 3.2.3 Interpreting and Applying the Two Differences, Worked Example 6.5; Objective 3.2)

7. In a randomised trial at a large workplace, 1,000 employees received a new ergonomic chair and 1,000 kept a standard chair. Over one year, 6% of the first group and 9% of the second reported a back injury. The risk difference is −0.03 (95% CI −0.053 to −0.007). Which interpretation is correct?

RD = 0.06 − 0.09 = −0.03, about 3 fewer injuries per 100 employees over one year. Both limits lie below 0, the null value for a difference, so the result is significant at the 5% level.
This reads the sign backwards. RD is the risk with the new chair minus the risk with the standard chair, so a negative value means fewer injuries with the new chair.
The verdict is right but the rule is wrong. 1 is the null value only for ratio measures; for a risk difference the question is whether the interval contains 0.
The interval −0.053 to −0.007 lies entirely below 0, so it excludes 0. The result is statistically significant at the 5% level.
Correct answer: the first interpretation. RD = 0.06 − 0.09 = −0.03, and the negative sign means fewer injuries among employees with the new chair: about 3 per 100 employees over one year. The interval −0.053 to −0.007 means the data are compatible with a reduction of anywhere from about 0.7 to 5.3 injuries per 100 employees, and since both limits lie below 0 the interval excludes the null value for a difference measure, so the result is statistically significant at the 5% level. The tempting “excludes 1” option reaches the right verdict with the wrong rule: 1 is the null value only for ratio measures, and for an RD the question is always whether the interval contains 0. (Section 2, 3.2.3 Interpreting and Applying the Two Differences, the tab “Interpreting RD/ID”; Section 4, 6.4.2 Interpreting Confidence Intervals; Objectives 3.2 and 6.4)

8. Community pharmacies offered a free smoking-cessation programme, and smokers chose whether to join. After one year, 25% of the 400 smokers who joined and 15% of the 400 who did not had quit. Which answer gives the number needed to treat (NNT) and a valid caution?

This uses only the joiners' quit proportion. The NNT is the reciprocal of the risk difference: 1 ÷ (0.25 − 0.15) = 10.
The value is right, but an NNT from observed risks is not certain to be accurate. Smokers chose whether to join, so part of the difference may be confounding, making the true NNT larger.
0.25 ÷ 0.15 is the risk ratio. The NNT is based on the absolute difference, 1 ÷ 0.10 = 10.
RD = 0.25 − 0.15 = 0.10, so NNT = 10 over one year. Because smokers chose to join, more motivated smokers may have enrolled, which would inflate the RD and make the NNT look better than it is.
Correct answer: NNT = 10, with the self-selection caution. Quitting is the good outcome here, so RD = 0.25 − 0.15 = 0.10 and NNT = 1 ÷ 0.10 = 10 smokers enrolled for one extra quitter over one year. Because smokers chose whether to join, those who joined may have been more motivated or had other advantages, so part of the RD may be confounding rather than an effect of the programme; if the RD is exaggerated, the true NNT is larger than 10 and the benefit is overstated, which is why the NNT is not “certain to be accurate”. 1 ÷ 0.25 uses only the joiners’ quit proportion, and 1.67 is the risk ratio. (Section 2, 3.3 Numbers Needed to Harm and to Treat and Worked Example 6.6; Objective 3.3)

9. A case-control study of a rare cancer finds OR = 4.0 for work with a particular industrial solvent. A colleague writes: “Workers exposed to the solvent had four times the risk of this cancer, so the solvent caused three-quarters of the cancers among exposed workers.” Which assessment is correct?

Table 6.3 lists the AFe as estimable in a case-control study through the OR, and for a rare disease the OR approximates the RR. Only the causal claim goes too far.
The arithmetic holds, but an OR of 4.0 does not by itself establish causation. The word caused assumes the association is causal, which requires considering selection bias, information bias and confounding.
For a rare cancer the OR approximates the RR, and (OR − 1) ÷ OR = 0.75 approximates the AFe. Saying the solvent caused those cases needs the causal assumption, after bias and confounding are considered.
1 ÷ OR is not the AFe. Substituting the OR for the RR gives AFe ≈ (OR − 1) ÷ OR = 3 ÷ 4 = 0.75.
Correct answer: the third assessment. The arithmetic is defensible: because the cancer is rare, the case-control OR approximates the RR, so “four times the risk” is acceptable with that caveat, and substituting the OR for the RR in Eq 6.6 gives AFe ≈ (OR − 1) ÷ OR = (4.0 − 1) ÷ 4.0 = 0.75. The word “caused” is not defensible: an attributable fraction is the share of exposed cases due to the exposure only if the association is causal, and an observed OR may also reflect selection bias, information bias or confounding. A case-control design does not rule out attributable fractions, since Table 6.3 lists the AFe estimated through the OR, and 1 ÷ OR is not the AFe. (Section 1, 2.1.1 When the Three Ratios Agree and the panel “When is OR ≈ RR?”; Section 2, 4.1 The Attributable Fraction in the Exposed, Eq 6.6; Objectives 2.1 and 4.1)

10. During a chickenpox outbreak at a group of daycare centres, 60 of 400 unvaccinated children and 30 of 1,000 vaccinated children developed chickenpox. Which answer gives the vaccine efficacy (VE) and the number of children who would need to be vaccinated to prevent one case during the outbreak?

VE is right, but the reciprocal of the VE is not the number needed to vaccinate. That number is the reciprocal of the risk difference, 1 ÷ (0.15 − 0.03) = 8.3.
VE = (0.15 − 0.03) ÷ 0.15 = 0.80, and the number needed to vaccinate is 1 ÷ RD = 1 ÷ 0.12 = 8.3, about 8 children for each case prevented during the outbreak.
0.20 is the risk ratio Rv/Ru = 0.03 ÷ 0.15, not the VE. VE is 1 minus that ratio, 0.80.
1 ÷ 0.03 uses only the vaccinated children's risk. The number needed to vaccinate uses the risk difference between the groups: 1 ÷ 0.12 = 8.3.
Correct answer: VE = 80% and about 8 children per case prevented. With Ru = 60 ÷ 400 = 0.15 and Rv = 30 ÷ 1,000 = 0.03, VE = (Ru − Rv) ÷ Ru = 0.12 ÷ 0.15 = 0.80: the vaccine prevented 80% of the cases that would have occurred among the vaccinated children had they not been vaccinated (not “80% of children are immune”). The number needed to vaccinate is the reciprocal of the risk difference, 1 ÷ (0.15 − 0.03) = 1 ÷ 0.12 = 8.3, so about 8 children vaccinated for each case prevented during this outbreak. Taking the reciprocal of the VE (1.25) or of the vaccinated children’s risk (33) uses the wrong quantity, and 0.20 is the risk ratio Rv/Ru, not the VE. (Section 2, 4.2 Vaccine Efficacy, Eq 6.7 and Worked Example 6.8, and 3.3 Numbers Needed to Harm and to Treat, Worked Example 6.6; Objectives 4.2 and 3.3)

11. In a city, 25% of children live in a home where someone smokes indoors, and the risk ratio for an asthma attack needing emergency care within a year is 3.0 for these children (assume the association is not confounded). Using Levin’s formula, which answer gives the AFp and interprets it correctly?

Levin's formula gives 0.25 × 2 ÷ (0.25 × 2 + 1) = 0.50 ÷ 1.50 = 0.33, a share of all such attacks among the city's children. It counts as avoidable only if the association is causal.
0.67 is (RR − 1) ÷ RR, the AFe, which describes only the exposed children's attacks. It ignores the 75% of children who are not exposed.
0.50 is only the numerator of Levin's formula. It must be divided by 0.50 + 1, giving 0.33.
The value is right, but an attributable fraction is a share of cases, not the risk of an attack among exposed children.
Correct answer: AFp = 0.33. Levin’s formula combines the exposure prevalence and the RR: AFp = p(E+)(RR − 1) ÷ [p(E+)(RR − 1) + 1] = 0.25 × 2 ÷ (0.25 × 2 + 1) = 0.50 ÷ 1.50 = 0.33, the share of all such attacks among the city’s children that is attributable to indoor smoking, which would be avoided by removing it only if the association is causal and not confounded. The tempting 0.67 is (RR − 1) ÷ RR, the AFe, which describes only the exposed children’s attacks and ignores the 75% of children who are not exposed; 0.50 forgets the denominator of the formula. An attributable fraction is a share of cases, not the risk of an attack among exposed children. (Section 3, 5.1 The Population Attributable Risk and Fraction, the panel “Recall: The Population Attributable Fraction” and Eq 6.9; Objective 5.1)

12. A cohort study finds that an occupational exposure has a risk ratio of 5.0, so the AFe is 0.80. Only 1% of the population is exposed, and the AFp for the same disease is 0.04. A colleague says one of the two calculations must be wrong. Which reply is correct?

This reverses the relationship: the AFp is always at most the AFe, not at least. The AFp is a share of all cases, most of which occur among the unexposed when exposure is rare.
Nothing in the stem points to confounding, and confounding can move an estimate in either direction. The AFp is small simply because only 1% of the population is exposed.
A rare exposure can have a strong association; rarity and strength are separate. Intravenous drug use and HIV is the lesson's example of a strong but rare exposure.
The AFe is a share of cases among exposed people, while the AFp is a share of all cases and depends on exposure prevalence. With 1% exposed, even an RR of 5.0 accounts for few cases.
Correct answer: the last reply. The two fractions have different denominators: AFe = (RR − 1) ÷ RR = 4 ÷ 5 = 0.80 is a share of the cases among exposed people, whereas AFp = p(E+)(RR − 1) ÷ [p(E+)(RR − 1) + 1] = 0.01 × 4 ÷ (0.04 + 1) = 0.04 ÷ 1.04 = 0.04 is a share of all cases, and it depends on the prevalence of exposure as well as on the RR. This is the strong-but-rare pattern of intravenous drug use and HIV; the AFp is always at most the AFe, not at least, and nothing in the stem points to confounding, which in any case can move an estimate in either direction. (Section 3, 5.2 Exposure Prevalence, Confounding and the AFp, the callout “Why Exposure Prevalence Matters”, Worked Example 6.10 and the slide “Strong and rare vs. weak and common”; Objective 5.2)

13. A case-control study of hepatitis A reports OR = 2.6 (95% CI 1.5 to 4.5) for eating raw shellfish in the six weeks before illness. Which interpretation is correct?

Excluding 1 makes chance an unlikely explanation, but a confidence interval reflects random error only. It says nothing about selection bias, information bias or confounding.
This states the estimate in words, gives the compatible range, checks that 1 lies outside it and notes that the interval covers random error only, not bias or confounding.
A confidence interval is not a probability that the exposure is causal. The 95% describes the procedure, and the interval reflects random error only.
A wide interval shows imprecision, but it still supports conclusions: it excludes 1, and even its lower limit of 1.5 shows raised odds.
Correct answer: the second interpretation. A complete reading of a given interval states the estimate in words (for a case-control study, the odds of exposure among cases compared with controls, equivalent to the odds of disease among the exposed compared with the unexposed), gives the range of values compatible with the data (1.5 to 4.5), checks the null value (1 lies outside the interval, so P < 0.05) and comments on precision. A confidence interval reflects random error only: excluding 1 makes chance an unlikely explanation but says nothing about selection bias, information bias or confounding, and it is not a probability that the exposure is causal. The interval is fairly wide, so the estimate is imprecise, but even its lower limit shows raised odds. (Section 4, 6.4.2 Interpreting Confidence Intervals; Objective 6.4)

14. A registry study of 80,000 births compares mothers who did and did not work night shifts during pregnancy. Pearson’s χ² for the 2×2 table is 12.6 on 1 degree of freedom (P < 0.001), and the risk ratio for preterm birth is 1.12 (95% CI 1.05 to 1.19). Which statement is correct?

A χ² test shows whether exposure and outcome are associated, not how strongly. Strength comes from the RR of 1.12, a weak association.
Pearson χ² is appropriate when every expected count is large, as it is with 80,000 births. A large sample makes the test more powerful, not inappropriate.
A weak association can be highly significant in a very large sample. Significance and strength are separate properties, so the χ² result and the RR do not disagree.
The χ² test makes chance an unlikely explanation, the RR of 1.12 shows a weak association, and the narrow interval shows precision. Strength, significance and precision are reported separately.
Correct answer: the last statement. The Pearson χ², which is appropriate here because every expected count is large, tests whether exposure and outcome are associated: P < 0.001 says data like these would be very unlikely if there were no association, but it says nothing about how strong the association is. Strength comes from the effect estimate, RR = 1.12, a weak association (a 12% higher risk). The very large sample makes even a weak association highly significant and gives a narrow, precise interval (1.05 to 1.19), so strength, significance and precision are reported separately, and every test is paired with an effect estimate and its CI. (Section 4, 6.3.2 Choosing a Test for Other Outcomes and Comparisons, Table 6.5, and 6.5 Statistical Significance and Strength of Association, the callout “Pair Every Test with an Effect Estimate”; Objectives 6.3 and 6.5)

15. In a pair-matched case-control study of 100 pairs, both the case and the control were exposed in 20 pairs, neither was exposed in 48 pairs, only the case was exposed in 24 pairs, and only the control was exposed in 8 pairs. Which analysis is correct?

In a matched design the pair is the unit, and only discordant pairs carry information. Matched OR = b/c = 24 ÷ 8 = 3.0, and McNemar's χ² = 16² ÷ 32 = 8.0.
The 68 concordant pairs, in which case and control had the same exposure, carry no information about the association. The matched OR uses only the discordant pairs: 24 ÷ 8 = 3.0.
McNemar's test is used for matched case-control pairs as well as before-and-after data (Table 6.5). Treating the 200 people as independent groups ignores the matching.
The matched OR puts the pairs in which only the case was exposed on top: b/c = 24 ÷ 8 = 3.0. Inverting it gives the OR for non-exposure.
Correct answer: the matched OR from the discordant pairs, with McNemar’s test. In a matched design the pair is the unit of analysis, and the 68 pairs in which the case and the control had the same exposure (20 + 48) carry no information about the association. Of the 32 discordant pairs, the case was the exposed member in b = 24 and the control in c = 8, so the matched OR = b/c = 24 ÷ 8 = 3.0, and McNemar’s χ² = (b − c)² ÷ (b + c) = 16² ÷ 32 = 8.0 on 1 degree of freedom (P ≈ 0.005), so the imbalance is unlikely to be due to chance. Analysing the 200 people as unmatched ignores the pairing (Table 6.5 lists matched case-control pairs among the uses of McNemar’s test), and putting the concordant pairs into the OR, or the control-only pairs on top, gives a wrong value. (Section 4, 6.3.2 Choosing a Test for Other Outcomes and Comparisons, Table 6.4 and Table 6.5; Objective 6.3)

16. A provincial health agency carried out a case-control study of invasive listeriosis in adults aged 65 and over (hypothetical data). Of the 90 cases, 30 had eaten ready-to-eat deli meats in the four weeks before illness and 60 had not; of the 180 controls, 20 had and 160 had not. The 95% confidence interval for the odds ratio is 2.11 to 7.58. Which answer is correct?

The inverted cross-product, 0.25, is the OR for not eating deli meats. With deli meats as the exposure, OR = (30 × 160) ÷ (60 × 20) = 4.0.
The investigator chose two controls for every case, so 30/50 is not a risk and no risk ratio can be calculated from a case-control table.
OR = 4,800 ÷ 1,200 = 4.0. The interval excludes 1, so the association is significant at the 5% level; its width shows imprecision, and it reflects random error only.
Width and asymmetry do not decide significance; only whether the interval includes 1 does. The asymmetry comes from building the interval on the log scale, and 2.11 to 7.58 excludes 1.
Correct answer: OR = 4.0 with that reading of the interval. The information needed is the four cells in the course layout: a1 = 30, a0 = 60, b1 = 20 and b0 = 160, so OR = 4,800 ÷ 1,200 = 4.0 (check: the odds of exposure were 30 ÷ 60 = 0.50 among cases and 20 ÷ 160 = 0.125 among controls, and 0.50 ÷ 0.125 = 4.0). The interval excludes 1, so the association is statistically significant at the 5% level; its width shows imprecision and its asymmetry comes from building it on the log scale, but neither makes it non-significant, and it reflects random error only. The investigator chose two controls for every case, so 30/50 is not a risk and no risk ratio can be calculated, and the inverted cross-product, 0.25, is the OR for not eating deli meats. (Section 1, 1.2.1 The Risk Ratio and the Odds Ratio and 2.1.2 The Odds Ratio in Case-Control Studies; Section 4, 6.4.2 Interpreting Confidence Intervals; Objectives 1.2, 2.1 and 6.4)

17. In the listeriosis case-control study (OR = 4.0; 30 of the 90 cases and 20 of the 180 controls had eaten deli meats), the disease is rare: the province records fewer than 2 cases per 100,000 older adults each year. In a case-control study the AFp can be estimated as pd × (OR − 1) ÷ OR, where pd is the proportion of cases who were exposed. A draft news release says: “Older adults who ate deli meats had four times the risk of listeriosis, so deli meats caused 75 per cent of the listeriosis cases in the province.” Which correction is best?

(OR − 1) ÷ OR = 0.75 approximates the AFe, a share of the exposed cases only. For all cases in the province the AFp is pd × 0.75 = 0.33 × 0.75 = 0.25, and caused needs the causal assumption.
For a rare disease the OR approximates the RR, so about four times the risk is defensible. But 0.75 is the AFe; the share of all cases is pd × 0.75 = 0.25, and caused assumes a causal association.
When the disease is rare, the case-control OR approximates the RR, so describing it in terms of risk is defensible with that caveat. And 0.75 is the AFe, not the AFp.
30 ÷ 50 reflects the investigator's choice of two controls per case, not a quantity of interest. pd is the proportion of cases who were exposed, 30 ÷ 90 = 0.33, giving AFp = 0.25.
Correct answer: the second correction. The information needed is how rare the disease is, the proportion of cases who were exposed, and the difference between the AFe and the AFp. Because listeriosis is far rarer than the 5% guide, the OR approximates the risk ratio, so “about four times the risk” is defensible with that caveat, and (OR − 1) ÷ OR = 3 ÷ 4 = 0.75 approximates the AFe, the share of cases among people who ate deli meats. For all the province’s cases, Equation 6.10 with the OR in place of the risk ratio gives AFp = pd × (OR − 1) ÷ OR = (30 ÷ 90) × 0.75 = 0.25. The 30/50 figure reflects the investigator’s choice of two controls per case, not a risk, and “caused” needs the causal assumption: selection bias, information bias or confounding could inflate the OR. (Section 1, 2.1.1 When the Three Ratios Agree; Section 2, 4.1 The Attributable Fraction in the Exposed; Section 3, 5.2 Exposure Prevalence, Confounding and the AFp; Objectives 2.1, 4.1 and 5.1)

✦ Pass every section knowledge check with 100% before submitting