Epidemiological Methods

Epidemiological methods are the tools epidemiologists use to learn about health from populations rather than from individual patients.

9 frameworks, from 1662 to today · Epidemiology · Medicine & Health

still a working view ended · each leads to its entry below

Overview

Epidemiological methods are the tools epidemiologists use to learn about health from populations rather than from individual patients1. The central difficulty is that a population cannot be put in a laboratory: you cannot randomize a city to clean water and a second city to dirty water just to see what happens2. So the field has had to invent ways of making fair comparisons out of the world as it is—counting deaths across parishes, comparing people who happened to be exposed with people who happened not to be, and, eventually, building formal logic for when a comparison can support a causal claim. The history of these methods is a history of successive answers to one question: how can we tell whether something really causes disease, when we can only observe?

Counting the dead: sanitary statistics and the mortality comparison

The earliest methods grew out of the "political arithmetic" of the seventeenth century. In London, John Graunt's Natural and Political Observations (1662) analyzed the weekly Bills of Mortality—parish tallies of christenings and burials—and produced the first systematic estimates of death rates, including the striking observation that a large share of deaths occurred in infancy. Graunt's work established the basic move of the field: aggregate individual events into rates so that populations of different sizes can be compared.

The method matured in the nineteenth century. In England, William Farr, appointed to the General Register Office after civil registration of births and deaths began in 1837, built a national system of mortality statistics and devised standardized measures—notably the comparison of observed deaths with those expected from age-specific rates in a reference population, an early form of what became the standardized mortality ratio. Farr's work made it possible to compare the health of one district, occupation, or year with another on a common footing.

The most famous demonstration came from John Snow's investigation of the 1854 cholera outbreak in the Soho district of London34. Snow mapped deaths and traced them to the Broad Street water pump, but his decisive argument was comparative: he contrasted mortality among customers of the Southwark and Vauxhall Water Company, which drew from a sewage-contaminated stretch of the Thames, with customers of the Lambeth Company, which had moved its intake upstream. The two companies' pipes ran intermingled through the same districts, so their customers differed in water supply but not much else—a comparison Snow himself described as approximating an experiment3. That reasoning, comparing groups that differ in exposure but are otherwise similar, is the seed of nearly everything that follows.

The natural experiment and the observational comparison

Snow's water comparison became the model for what epidemiologists call a natural experiment: a situation in which circumstances—not the investigator—assign exposure in a way that plausibly separates it from other differences3. The classic nineteenth-century example is John Snow's contemporary setting again, but the logic recurs wherever a policy, a disaster, or a geographic accident creates a clean contrast. In the twentieth century the most celebrated case was the 1854 London comparison's intellectual descendant: the study of British doctors' smoking habits by Richard Doll and Austin Bradford Hill, begun in 1951, which followed tens of thousands of physicians over decades and compared lung cancer mortality in smokers and non-smokers. The doctors differed in smoking but were otherwise a relatively homogeneous professional group, which strengthened the comparison.

Natural experiments remain a workhorse today—studies of what happens when a country raises the minimum drinking age, tightens air-pollution rules, or introduces a vaccine on a staggered schedule. Their appeal is that the exposure was not chosen by the researcher, reducing some kinds of bias; their weakness is that the "assignment" is rarely as clean as a true experiment, so the investigator must argue that the comparison groups really are comparable.

Two designs for studying who gets sick: cohort and case-control

By the early twentieth century, epidemiologists had settled on two basic ways to structure an observational comparison. The cohort study follows a group of people forward in time, measuring exposure at the start and counting who develops disease56. It yields directly interpretable measures—incidence in the exposed versus the unexposed—and can capture multiple outcomes from one exposure. The British doctors study is the archetype: a defined cohort, a baseline exposure measure, and decades of follow-up.

The case-control study works backward: it takes people who already have the disease (cases) and comparable people who do not (controls), then looks back to compare their past exposures7. Its great advantage is efficiency—rare diseases can be studied without waiting for them to accumulate in a cohort. The design was formalized in the mid-twentieth century; Janet Lane-Claypon's 1926 comparison of the histories of women with and without breast cancer is an early example, and the design became central to the postwar explosion of cancer epidemiology, including the studies that linked smoking to lung cancer. The trade-off is vulnerability to bias: recall of past exposures may differ between cases and controls, and choosing truly comparable controls is hard.

The randomized controlled trial

The decisive methodological innovation of the twentieth century was the randomized controlled trial (RCT), in which the investigator assigns participants to treatment or control by chance8. Randomization solves the comparability problem at its root: on average, known and unknown differences between groups balance out, so a difference in outcome can be attributed to the intervention with far more confidence than in any observational comparison8.

The intellectual foundations were laid in the 1920s and 1930s by Ronald Fisher, whose work on agricultural experiments established randomization and statistical inference as the basis of experimental design89. The first well-known clinical application was the British Medical Research Council's trial of streptomycin for pulmonary tuberculosis, reported in 1948 and designed with Austin Bradford Hill's statistical guidance; it randomized patients to streptomycin plus bed rest or bed rest alone and assessed outcomes under blinded conditions. The trial became the template for the modern RCT, and randomization, allocation concealment, and blinding became standard requirements. From the 1960s onward the RCT spread from infectious disease and cancer into cardiology, mental health, and prevention, and by the 1980s large multicenter trials had become the accepted standard for establishing whether a treatment works.

Bias and confounding: naming what goes wrong

As observational studies proliferated, so did awareness of how they can mislead. Two concepts became central. Confounding is the mixing of the effect of the exposure of interest with the effect of another factor associated with both exposure and outcome—drinking coffee and smoking, say, if smokers drink more coffee and smoking causes the disease10. Bias is a systematic error in how subjects are selected, measured, or followed—for example, selecting hospital controls who are hospitalized for conditions related to the exposure, or measuring exposure differently in cases and controls11.

The systematic study of these errors is largely a mid- to late-twentieth-century development. The 1959 U.S. Surgeon General's report on smoking drew heavily on the need to rule out confounding and bias as explanations for the smoking–lung cancer association, and the methodological literature of the 1970s and 1980s—much of it associated with Abraham Lilienfeld, Brian MacMahon, and later Kenneth Rothman—codified the taxonomy of selection bias, information bias, and confounding, and the design and analytic strategies for controlling them. The case-control study in particular was subjected to intense scrutiny and defense in this period, emerging with clearer rules about control selection and exposure measurement.

Causal criteria: the Hill viewpoints

How much evidence is enough to call something a cause? In 1965 Austin Bradford Hill, addressing the Royal Society of Medicine, set out a list of "viewpoints"—often called criteria—for judging whether an observed association is causal: strength, consistency across studies, specificity, temporality (cause before effect), biological gradient (dose–response), plausibility, coherence, experiment, and analogy. Hill was careful to say these were aids to judgment, not a checklist that could prove causation; temporality is the only one he treated as essential. The viewpoints became enormously influential because they gave investigators a shared vocabulary for arguing about causation, and they remain in use—often invoked, sometimes criticized for encouraging mechanical reasoning—in textbooks and regulatory assessments12.

Multivariable regression: adjusting for many factors at once

Controlling confounding one factor at a time—stratifying, say, by age and sex—becomes impractical when many confounders must be handled simultaneously. The solution was multivariable regression: a statistical model that estimates the association between exposure and outcome while holding other measured variables constant13. The technique rests on the least-squares regression developed in the nineteenth century and on the logistic regression model, whose use in epidemiology spread from the 1960s and 1970s as computing became available. By the 1980s, regression models—logistic for binary outcomes, Cox proportional hazards models (introduced by David Cox in 1972) for time-to-event data—were standard tools, and epidemiology textbooks devoted substantial space to model building, interaction, and the interpretation of adjusted estimates. The method made large observational datasets analyzable but also raised new questions: what if important confounders were never measured, and what does "adjustment" mean when variables are entangled?

The counterfactual turn

The most recent methodological shift reframed causation itself. The counterfactual or potential-outcomes framework asks what would have happened to the same people under a different exposure—the outcome that was not observed14. Causal effects are defined as contrasts between potential outcomes, and confounding is defined precisely as a difference between the groups that would remain even absent the exposure. This framework, developed in statistics and econometrics from the 1970s onward (with contributions from Donald Rubin and others) and imported into epidemiology in the 1980s and 1990s, gave the field a formal language for old intuitions. It clarified what regression adjustment actually estimates, motivated methods such as propensity scores (Rosenbaum and Rubin, 1983) and inverse-probability weighting, and underpinned the modern toolkit of causal inference—directed acyclic graphs for representing assumptions, instrumental variables, and sensitivity analysis for unmeasured confounding. It also made explicit that every causal claim rests on assumptions that must be stated and defended, not hidden inside a model.

Where the methods stand now

Contemporary epidemiological methods combine all of these strands. Randomized trials remain the gold standard where feasible, but much of the field's work is observational, and the counterfactual framework now organizes how observational studies are designed and analyzed: target-trial emulation, where an observational dataset is analyzed as if it were a randomized trial, has become a prominent approach. Natural experiments and quasi-experimental designs—difference-in-differences, regression discontinuity, instrumental variables—are increasingly used to approximate randomization. Multivariable modeling has been extended by machine learning and high-dimensional data methods, while the study of bias has grown into formal quantitative bias analysis and the reporting of assumptions through causal diagrams. Open questions persist: how to handle unmeasured confounding credibly, how to combine evidence across heterogeneous studies, how to make causal claims from routinely collected data such as electronic health records, and how to weigh the assumptions of any method against the practical constraints of studying real populations. The field's history is, in effect, a long argument that careful comparison—properly designed, honestly analyzed, and explicitly reasoned—can extract reliable knowledge from populations that can never be fully controlled.

References

  1. Epidemiology, Wikipedia
  2. Epidemiological method (Outline of the process of an epidemiological study), Wikipedia
  3. Natural experiment (History, introduction), Wikipedia
  4. Epizootiology (History), Wikipedia
  5. Cohort study, Wikipedia
  6. Prospective cohort study, Wikipedia
  7. Case–control study, Wikipedia
  8. Randomized controlled trial (introduction, History), Wikipedia
  9. Randomized experiment (History), Wikipedia
  10. Confounding, Wikipedia
  11. Selection bias, Wikipedia
  12. Bradford Hill criteria, Wikipedia
  13. Logistic regression, Wikipedia
  14. Multiple treatments, Wikipedia

The frameworks of Epidemiological Methods

Each framework is a way of working the field has used: when it held, and the question it answers.

  1. 1662-1900

    Political Arithmetic and Sanitary Statistics

    How can death rates be compared across populations?

    Fly to its tower in Nousfarer

  2. 1854-Present

    Natural Experiment

    When does a natural contrast reveal a cause?

    Fly to its tower in Nousfarer

  3. 1900-Present

    Cohort Study

    Who develops disease after a given exposure?

    Fly to its tower in Nousfarer

  4. 1926-Present

    Case-Control Study

    What did people with disease do differently?

    Fly to its tower in Nousfarer

  5. 1948-Present
  6. 1959-Present

    Bias and Confounding Analysis

    What else could explain an observed association?

    Fly to its tower in Nousfarer

  7. 1960-Present
  8. 1965-Present
  9. 1980-Present

How the ideas connect

Each arrow runs from an earlier idea to the later one it shaped. The list below says what passed between them.

Political Arithmetic and Sanitary Sta…Natural ExperimentCohort StudyCase-Control StudyRandomized Controlled TrialBias and Confounding AnalysisMultivariable Regression ModelingBradford Hill criteriaCounterfactual (Potential Outcomes) F…

Beyond the field

Where Epidemiological Methods's ideas came from in other fields, and what they led to there: the strongest 12 of 80 links.

A city in Nousfarer: a red-carpeted stair climbing terraces to a crowned tower on an acropolis, more towers round it, forest and snowy mountains beyond.
A city in Nousfarer, the game played in the browser.

Epidemiological Methods is a planet in Nousfarer

Its city has a tower for each of these 9 frameworks. Land, walk in, and climb one concept at a time.

Land on Epidemiological Methods

Related fields

Fields its ideas are linked to

More in Epidemiology

See where Epidemiological Methods sits on the Map of Human Knowledge