Epidemiology is the study of the distribution and determinants of disease in populations, and the use of that knowledge to control it
Studies are observational (the investigator records what happens) or experimental (the investigator assigns the exposure or treatment)
Observational studies are descriptive (who, where, when) or analytic (testing whether an exposure is linked to an outcome)

## Hierarchy of evidence

The best design depends on the question: an RCT for treatment, a cohort for prognosis, a cross-sectional study with a reference standard for diagnostic accuracy, and a case-control study for a rare disease

## Descriptive studies

Case report and case series: describe unusual presentations or new diseases and generate hypotheses. There is no comparison group
Cross-sectional (prevalence) study: exposure and outcome are measured at one point in time

Gives prevalence, not incidence. It cannot show which came first, exposure or disease
Example: population surveys of blindness such as the Rapid Assessment of Avoidable Blindness (see Global Epidemiology of Blindness)

Ecological study: compares groups (countries, regions), not individuals. Prone to the ecological fallacy: an association between groups need not hold for individuals

## Analytic observational studies

Feature	Cohort	Case-control
Starts with	People grouped by exposure, free of the disease	People grouped by outcome: cases with the disease and controls without
Direction	Forwards from exposure to outcome (prospective), or using past records (retrospective)	Backwards from outcome to past exposure
Measure	Incidence, relative risk	Odds ratio only (incidence cannot be measured)
Best for	Rare exposures; several outcomes of one exposure; establishing that exposure came first	Rare diseases and those with a long latent period; several exposures for one disease
Weaknesses	Large, slow and expensive; loss to follow-up	Recall bias; choosing suitable controls; selection bias
Ophthalmic example	Smoking and the incidence of AMD over 10 years	Past use of a drug among patients with uveal melanoma and matched controls

A nested case-control study takes its cases and controls from within an existing cohort, which saves cost and reduces recall bias
Matching controls on age and sex controls for those confounders, but a matched variable can no longer be studied as a risk factor

## Randomised controlled trials

Randomisation gives each participant a known chance of each treatment. It balances known and unknown confounders between the groups, so any difference in outcome can be attributed to the treatment
Allocation concealment: the person enrolling a patient cannot know the next allocation. It prevents selection bias at enrolment and is different from blinding
Blinding (masking) of the participant, the treating clinician and the outcome assessor reduces performance and detection bias. A sham injection blinds the patient in anti-VEGF trials
Intention-to-treat (ITT) analysis: participants are analysed in the group they were randomised to, whatever happened afterwards. It keeps the benefit of randomisation and reflects real-world effectiveness
Per-protocol analysis includes only those who completed the treatment as assigned. It can overestimate benefit; it is often reported alongside ITT, especially in non-inferiority trials
Trial phases are summarised in Clinical Trials Summary Table; trials are reported to the CONSORT standard

Trial design	Principle	Example
Parallel group	Each group receives one treatment throughout	Most drug trials
Crossover	Each participant receives both treatments in turn, after a washout period. Needs a stable, chronic condition	Comparing two glaucoma drops
Factorial	Two or more treatments tested at once in every combination (e.g. A, B, A + B, placebo)	AREDS: antioxidants, zinc, both, or placebo
Cluster randomised	Groups (villages, clinics) are randomised, not individuals. Needs a larger sample	Mass azithromycin distribution for trachoma
Paired-eye	One eye is randomised to treatment and the fellow eye acts as control	Early Treatment Diabetic Retinopathy Study (one eye early laser, the other deferred)
Non-inferiority	Shows a new treatment is not worse than the standard by more than a pre-set margin	CATT: bevacizumab compared with ranibizumab for neovascular AMD

The two eyes of one person are not independent: their measurements are correlated. Counting both eyes as separate observations inflates the sample size and gives falsely small p-values. Use one eye per person, or a method that allows for the correlation
Rare disease: case-control. Rare exposure: cohort. Treatment effect: RCT

## Systematic reviews and meta-analysis

A systematic review uses a pre-specified protocol to find, appraise and summarise all the studies on a question. It is reported to the PRISMA standard. Cochrane reviews are an example
A meta-analysis combines the results statistically into a pooled estimate

Forest plot: each study is a square (its size shows its weight) with a horizontal line for its confidence interval. The pooled result is a diamond. The vertical line is no effect (1 for a ratio, 0 for a difference)
Heterogeneity: variation between study results beyond chance, measured by the I² statistic (about 25% low, 50% moderate, 75% high)
A fixed-effect model assumes one true effect; a random-effects model allows the true effect to vary between studies and gives a wider confidence interval
Funnel plot: effect size against study size. Asymmetry suggests publication bias, as small negative studies tend to go unpublished

Qualitative research (interviews, focus groups) explores why people do or do not use services, for example barriers to cataract surgery
Reporting standards: CONSORT (trials), STROBE (observational studies), STARD (diagnostic accuracy), PRISMA (systematic reviews)
Bias, confounding and causation are covered in Bias, Confounding and Causation