Learning outcomes
By the end of this lesson you should be able to: define epidemiology and state its uses; use the vocabulary of disease occurrence — endemic, sporadic, epidemic, pandemic, common-source and propagated outbreaks; compute and interpret the measures of disease frequency (incidence, prevalence, mortality and attack rates) and the measures of association (risk ratio, odds ratio, attributable risk) with their denominators correct; describe the main study designs with their strengths, weaknesses and the measure of association each yields; apply the Bradford Hill criteria to claim causation and name the principal biases and confounding with their controls; and follow the steps of an outbreak investigation.
Prerequisites
Lesson 5 (Biostatistics): incidence, prevalence and the measures of association are computed here, and interpretation uses the confidence intervals of the previous lesson. Lesson 1 gives the natural-history framework the designs sit on.
The big idea
Medicine asks 'why is this patient ill'; epidemiology asks 'why is this population ill, and not that one'. The power of the second question is that it can be answered without a laboratory: John Snow removed the handle of the Broad Street pump in 1854 on nothing more than a map of cholera deaths around the pump, which is why the plotted map remains the founding image of the discipline. Everything that follows — the cohort, the case-control study, the randomised trial, the outbreak investigation — is a formal version of that act of comparing who fell ill among who did not.
The vocabulary of occurrence
| Term | Meaning | Field example |
|---|---|---|
| Endemic | The constant, usual level of a disease in an area (hyperendemic, holoendemic: intense transmission by age group) | Malaria in the Terai; dengue throughout the monsoon |
| Sporadic | Scattered, occasional cases with no link between them | An imported case of Lassa fever |
| Epidemic | The occurrence of more cases than normally expected in an area, from a common source or by propagation | Hepatitis A after a wedding feast |
| Pandemic | An epidemic that has crossed continents | COVID-19; influenza of 1918 |
| Common-source outbreak | A single exposure contaminates many people at once (single exposure: one point in time; continuous: over days) | Cholera after one contaminated tube well |
| Propagated outbreak | Person-to-person transmission through serial infections | Measles spreading through an unvaccinated class, wave after wave |
Measures of frequency
| Measure | Formula | Units | Interpretation |
|---|---|---|---|
| Incidence (risk) | New cases in a period / population at risk at the start | Per 1000 per year | The chance a risk-free person develops disease; used for cohort studies |
| Incidence density | New cases / person-time at risk | Per 100 person-years | When people are followed for unequal follow-up times |
| Prevalence | All existing cases (old + new) / total population at a point in time | Per 1000 | The burden of disease at a moment; prevalence is roughly incidence x duration |
| Attack rate (secondary attack rate) | Cases among those exposed / number exposed | Per 100 | Outbreak investigation; the secondary attack rate is among household contacts |
| Case fatality | Deaths from the disease / cases of the disease | Per 100 | The deadliness of the disease in the treated population |
| Crude mortality | All deaths / total population | Per 1000 per year | Comparable only between populations with the same age structure, hence age-standardised rates |
NotePrevalence = incidence x average duration. A disease can be stopped by cure (short duration), or by rapid death (short duration); a chronic disease with low incidence accumulates prevalence. This single identity explains why the commonest diseases in clinic are not the commonest causes of death — a long cold lasts weeks, arthritis lasts decades.
Measures of association — the worked tables
Every measure of association comes from the same skeleton: a 2x2 table of exposure against outcome. Fill it correctly and the formula is trivial; fill it wrongly — which the denominator usually does — and the answer is confidently wrong.
Example A — a cohort study of smoking and lung cancer: 1000 smokers followed for a year produce 50 cases, and 1000 non-smokers produce 10 cases.
| Disease developed | No disease | Total | Incidence | |
|---|---|---|---|---|
| Exposed (smokers) | 50 | 950 | 1000 | 50 per 1000 per year (5%) |
| Unexposed (non-smokers) | 10 | 990 | 1000 | 10 per 1000 per year (1%) |
| Total | 60 | 1940 | 2000 | 30 per 1000 per year (3%) |
| Measure | Calculation | Result |
|---|---|---|
| Risk ratio (relative risk) | 5% among the exposed / 1% among the unexposed | 5 — a smoker's risk is five times the non-smoker's |
| Attributable risk (risk difference) | 5% - 1% | 4 percentage points (40 per 1000) attributable to the exposure |
| Population attributable risk | 3% total incidence - 1% incidence if unexposed | 2 percentage points (20 per 1000) removable if the exposure stopped |
| Attributable fraction among the exposed | (5% - 1%) / 5% | 80% of cases among the exposed are attributable |
Example B — a case-control study of the association between eating the wedding chicken and salmonellosis: 200 affected guests (cases) and 200 unaffected guests (controls) are interviewed about what they ate.
| Ate the chicken | Did not eat the chicken | Total | |
|---|---|---|---|
| Cases (affected) | 150 | 50 | 200 |
| Controls (unaffected) | 60 | 140 | 200 |
| Measure | Calculation | Result |
|---|---|---|
| Odds of exposure among cases | 150/50 | 3.0 |
| Odds of exposure among controls | 60/140 | 0.43 |
| Odds ratio | (150 x 140) / (50 x 60) | 21000/3000 = 7.0 |
The cohort gives a risk ratio because the disease is counted after following exposure; the case-control study cannot give a risk, because the investigator fixed the number of cases and controls instead of following a population, and the odds ratio estimates the risk ratio when the disease is rare (under about 10%). This single structural difference between the designs is the most examinable point in the whole lesson.
The study designs
| Design | What it does | Measure it yields | Chief weakness |
|---|---|---|---|
| Ecological (correlational) | Compares disease rates between populations or over time | Correlation between a population exposure and a disease rate | Ecological fallacy — group averages do not describe individuals |
| Cross-sectional | Snapshots disease and exposure together, once | Prevalence and prevalence ratios | Cannot show sequence: the cause may have followed the effect |
| Case-control | Compares past exposure in cases and in controls | Odds ratio | Recall bias, selection of controls; rare disease suits it best |
| Cohort | Follows exposed and unexposed people forward in time | Risk ratio, incidence — the strongest observational evidence | Expensive, loses to follow-up; unsuitable for rare disease |
| Randomised controlled trial | Randomises people to intervention or control and follows them | Effectiveness; relative risk reduction | Ethical constraints, volunteer bias in the population reached |
| Quasi-experimental | Intervention without randomisation, before-after or non-equivalent control | Change attributable to the programme | Secular trend and selection confuse the effect |
Choosing a design — a district wants the burden of hypertension among adults. Cross-sectional (one BP survey). It wants whether a new diet causes weight loss. Randomised controlled trial. It wants whether a rare cancer was caused by a pesticide at a factory. Case-control, because the disease is too rare to follow a cohort for. It wants whether childhood immunisation coverage predicts adult outcomes. Cohort, if the records can be re-traced, otherwise a cross-sectional snapshot with the known limits.
Causality, bias and confounding
| Bradford Hill criterion | The question asked of the association |
|---|---|
| Strength | Is the risk ratio large (RR of 10) rather than marginal (RR of 1.1)? |
| Consistency | Do different studies in different populations find it? |
| Specificity | Does the exposure lead to one effect and the effect follow the exposure? |
| Temporality | Does the cause precede the effect — the only essential criterion |
| Biological gradient | Does more exposure produce more disease (a dose-response curve)? |
| Plausibility and coherence | Does a known mechanism exist, and is it consistent with existing knowledge? |
| Experimental evidence | Does removing the exposure reduce the disease? |
| Analogy | Does a comparable cause-and-effect pair exist (as thalidomide informed rubella)? |
| Error | What happens | Example | The control |
|---|---|---|---|
| Selection bias | Comparison groups are formed differently from the start | A volunteer workforce healthier than the population | Random allocation; same entry criteria; identical sampling frames |
| Information bias (recall, misclassification) | Disease or exposure is measured differently between groups | Mothers of leukaemic children remember infections more keenly than control mothers | Blinding, objective records, identical questioning |
| Loss to follow-up | Those who drop out differ from those who stay | The sickest smokers stop attending | Tracking, intention-to-treat analysis, reporting the fraction lost |
| Confounding | A third factor causes both the exposure and the disease | Alcohol appears to cause lung cancer because it travels with smoking | Randomisation, restriction, matching, stratification, adjustment |
NoteA confounder must be associated with both the exposure and the disease and must not lie on the causal pathway between them. Smoking confounds the alcohol-lung cancer association; oral cancer itself does not confound anything, because it is a consequence — intervening on it would adjust away the very effect being studied.
Outbreak investigation
An outbreak is an epidemiology experiment staged by nature, and the ten steps below are the standing protocol. Attack rate and the epidemic curve do the investigative work: a single-point epidemic curve implies a common source; wave after wave implies person-to-person spread.
| Step | Act | The instrument |
|---|---|---|
| 1. Verify the diagnosis and confirm the outbreak | Is it really the disease, and are there really more cases than usual? | Laboratory confirmation; comparison with baseline |
| 2. Define a case and count cases | Case definition: clinical criteria + place + time | Line listing |
| 3. Describe the epidemic curve | Plot cases by time of onset | Histogram of onset dates; see common source or propagated |
| 4. Describe by person and place | Age, sex, food, water source, map of cases | Spot map; attack rates per food item |
| 5. Calculate attack rates and form hypotheses | Which exposure is shared by the ill and not by the well? | Food-specific attack rates; relative risks |
| 6. Test the hypotheses | Analyse the 2x2 table | Chi-square; relative risk or odds ratio |
| 7. Consider alternative explanations | Diagnostic artefacts, reporting changes, denominator errors | Review the line list carefully |
| 8. Implement control | Stop the exposure — isolate, chlorinate, withdraw the food, recall the vehicle | The control measure, which cannot wait for analysis |
| 9. Communicate | Report to the authorities and the public, timely and accurate | Outbreak report |
| 10. Follow up and evaluate | Monitor for new cases; document lessons | Surveillance and a written evaluation |
Worked example — a village feast reports 45 ill of 60 who ate the chicken, against 5 ill of 40 who did not. Food-specific attack rates: chicken 45/60 = 75%; everything else under 15%. Relative risk for chicken = 75% / 12.5% = 6. The curve is a single point in time — a common source. The control is immediate: withdraw the remaining chicken, hydrate the ill, and take samples. No inferential test is needed to act.
Common mistakes
Watch outUsing a case-control study to compute a relative risk. The investigator fixed the numbers of cases and controls, so incidence cannot be recovered from the table; the correct measure is the odds ratio. Similarly, quoting 'prevalence' when counting new cases over the past month among a population at risk — that total is a cumulative incidence, and its denominator excludes the existing cases.
Watch outDeclaring an association causative on a strong p-value alone. A p-value tests only against chance; it cannot exclude confounding, bias or reverse causality. Temporality, dose-response and consistency carry the causal weight, as the Bradford Hill criteria set out.
Watch outReading an ecological comparison as an individual risk. Counties with more storks have more births — the correlation is real at the county level and means nothing for any household. Individual risk needs an individual design.
Check your understanding
In the cohort worked example, what is the attributable fraction among the exposed, and what does it mean in the district?
Why can a case-control study estimate the odds ratio but not the risk ratio?
An outbreak curve shows a sharp rise and fall on one day. What kind of outbreak does this imply, and what is the next analytic step?
Multiple choice
In the cohort worked example, what is the risk ratio for disease among the exposed?
A case-control study of a rare cancer is most appropriate when:
A study of coffee and myocardial infarction finds the association disappears when cigarette smoking is adjusted for. This illustrates: