Vésale

Types of epidemiological studies

Descriptive, case-control, cohort, randomized trial.

The three main families of bias

Selection bias comes from recruitment: the subjects included do not represent the target population. Information (classification) bias comes from measurement: exposure or disease poorly recorded. Confounding bias comes from a third factor linked to both the exposure and the disease.

Only confounding can be corrected afterward, by adjustment — provided the confounder has been measured. The other two must be prevented at the design stage: once the data are collected, it is too late.

Choosing a study design

The randomized controlled trial is the reference design for evaluating a treatment. Randomization balances confounding factors between groups on average, including unknown ones — which no statistical adjustment method can do. But a harmful exposure cannot be randomized.

A cohort study follows exposed and unexposed subjects over time. It respects the chronology, allows an incidence and a relative risk to be calculated, and suits rare exposures. It is long and expensive.

A case-control study starts from the disease and works back to the exposure. It is fast, inexpensive and suited to rare diseases, and only allows an odds ratio — since the size of each group is set by the investigator, no incidence can be derived from it. It is exposed to recall bias.

Measuring an association

A cohort makes it possible to calculate the risk of disease in exposed and unexposed people. Their ratio is the relative risk: equal to one, it indicates no association; above one, a risk factor; below one, a protective factor. Their difference, the attributable risk, measures the excess of cases due to the exposure.

A case-control study cannot yield these risks, since the proportion of patients is set by the investigator. Instead, the odds ratio is calculated: the ratio of the odds of exposure among cases and among controls. When the disease is rare, it is a good approximation of the relative risk.

An association, even a strong and significant one, is not proof of causation. The Bradford Hill criteria help judge it: strength and consistency of the association, dose-response relationship, exposure preceding the outcome, biological plausibility.

Measuring how common a disease is

Prevalence is the proportion of people affected in a population at a given time: it counts all cases, old and new. Incidence counts new cases occurring over a period, relative to the population at risk; when follow-up times differ, it is expressed per person-years observed, and one speaks of incidence rate.

The two are linked by the duration of the disease: in a stable situation, prevalence is roughly equal to incidence multiplied by average duration. A long-lasting chronic disease, such as diabetes, therefore has a high prevalence even if its incidence is moderate; a brief illness, such as flu, can have a high incidence and a low prevalence at any one moment. A treatment that prolongs survival without curing increases prevalence.

Mortality relates deaths to the whole population; case fatality relates them to patients only. A disease can have high case fatality yet account for little mortality if it is rare.

Calculating an effect, and reading it correctly

In a trial, 10% of patients in the placebo group have a heart attack, compared with 6% in the treated group. The relative risk is 6 / 10 = 0.6: the treatment reduces the risk by 40% in relative terms. The absolute risk reduction is 10 − 6 = 4 percentage points. The number needed to treat to prevent one event is the inverse of this absolute reduction: 1 / 0.04 = 25 patients.

The same relative risk does not carry the same weight depending on the baseline risk. In patients whose initial risk is only 1%, a 40% relative reduction gains only 0.4 points, and 250 people must be treated to prevent one event. Presenting only the relative reduction, which is more impressive, is a common way of overstating a treatment's value.

In a case-control study, risk cannot be calculated, since the investigator chooses the number of cases and controls. An odds ratio is calculated instead, which approaches the relative risk when the disease is rare.

Levels of evidence

Not all studies carry the same weight. At the bottom of the scale, expert opinion and case series suggest hypotheses. Case-control and cohort studies, which are observational, measure associations but remain exposed to bias and confounding. The randomised controlled trial, which allocates treatments at random, is the reference design for demonstrating the effect of an intervention. At the top, systematic reviews and meta-analyses pool all good-quality trials.

In France, clinical practice guidelines state the grade of each recommendation, from A, based on established scientific evidence, to C, based on a low level of evidence, or even mere expert agreement.

The randomised trial is not always possible, however. People cannot be randomised to smoke: the harm of tobacco was established by cohort and case-control studies, whose consistency, strength of association, dose-response relationship and biological plausibility carried the day.

The randomised trial in detail

The protocol sets in advance the inclusion and exclusion criteria, the comparator, the primary endpoint and the number of subjects. Randomisation allocates patients between groups: it makes the groups comparable for all factors, known or unknown, which no adjustment can guarantee. It must be unpredictable for investigators, otherwise they might steer, even unconsciously, certain patients towards one group.

Blinding prevents knowledge of the treatment from influencing the behaviour of patients, carers or assessors. Placebo is used as comparator when no effective treatment exists; otherwise it is generally unethical to deprive patients of the reference treatment.

The main analysis is by intention to treat: each patient is analysed in the group to which randomisation assigned them, whether or not they took the treatment. It preserves the comparability created by randomisation. Finally, the endpoint must matter to patients: a drug that improves a laboratory figure, a surrogate endpoint, does not necessarily improve survival or quality of life, and some have even worsened it.

Monitoring and understanding an epidemic

Surveillance of infectious diseases relies notably on notifiable diseases, which every doctor or laboratory must report to the health authority: tuberculosis, measles, meningococcal meningitis, collective food poisoning, among others. It allows an epidemic to be detected early and action taken around cases.

The basic reproduction number, R0, is the average number of people infected by one case in a fully susceptible population. If the effective reproduction number stays above 1, the epidemic spreads; below it, the epidemic recedes. To stop it by vaccination, a proportion of the population at least equal to 1 − 1/R0 must be immunised: about 50% for R0 = 2, but over 90% for measles, whose R0 exceeds 12.

Investigation of an epidemic begins with a precise case definition, then describes cases by time, place and person. The epidemic curve, number of cases over time, points to the mode of transmission: a single brief peak suggests a point common source, such as a contaminated meal; successive waves suggest person-to-person transmission.

57 answered questions go with this chapter in the practice section.

Sources and further reading

Reference textbooks

  • Jean Bouyer et coll., Épidémiologie. Principes et méthodes quantitatives, Inserm.

Open access

Last revised : . Text licensed under CC BY-SA 4.0.

The 3D atlas, the questions and progress tracking require JavaScript.