A cohort study follows a group of people defined by their exposure status and tracks them forward in time to see who develops the outcome. Because exposure is recorded before the outcome appears, this observational study design can establish temporal order, measure incidence, and estimate relative risk, which is exactly what a single-snapshot design cannot do. That forward direction is the reason cohort studies sit near the top of the observational evidence hierarchy.
In a cohort study you start with people who do not yet have the outcome, classify them as exposed or unexposed, and then wait. Because the exposure is measured first, an association you observe later carries the one thing a cross-sectional study cannot supply: the knowledge that the cause preceded the effect. This is what lets a cohort estimate the incidence rate, the risk ratio, and the hazard ratio, and it is why cohort evidence is so persuasive when a randomized trial is impossible or unethical. You cannot randomize people to smoke; you can follow smokers and non-smokers and count the cancers.
Prospective and retrospective cohorts
The two flavors differ only in when the investigator enters the timeline.
- A prospective cohort study identifies exposure now and follows participants into the future. It gives the cleanest data because the researcher decides in advance what to measure and how, but it is slow and expensive, sometimes running for decades.
- A retrospective cohort study uses existing records to reconstruct an exposed and unexposed group whose outcomes have already occurred. It is faster and cheaper because the follow-up has, in effect, already happened, but it is limited to whatever the historical records captured.
Both are genuine cohorts because both start from exposure and move toward outcome. The retrospective version is not a case-control study; the direction of reasoning is still exposure to outcome, only the calendar is different.
The headline metric is incidence: the rate at which new cases appear in the exposed compared with the unexposed. From that you compute a relative risk, the ratio of those incidences, which has a direct and honest interpretation that the odds ratio only approximates. When you need to turn counts into a comparative risk, our relative risk calculator does the arithmetic and shows the confidence interval. Cohorts also support time-to-event analysis, so a single study can report both whether an outcome happened and how quickly.
The signature threat to a cohort study is attrition, the loss of participants before the outcome is observed. If the people who drop out differ systematically from those who remain, and especially if dropout is related to both exposure and outcome, the result is biased in a way no analysis can fully repair. A well-run cohort plans retention from the start, documents every loss, and reports how the lost participants compared with the retained ones. Differential loss to follow-up is to a cohort what poor sampling is to a survey: the quiet failure that undermines everything downstream.
Confounding and how cohorts handle it
Because exposure is not randomized, the exposed and unexposed groups may differ in ways that also affect the outcome. Confounding is the central analytic problem of any observational comparison. Cohort studies address it by measuring potential confounders at baseline and adjusting for them with multivariable regression, and increasingly with methods such as propensity scoring that balance the groups before the outcome is modeled. The credibility of a cohort's causal claim rests on how completely it anticipated and measured the confounders, which is a study-design decision long before it is a statistics decision.
Appraising and reporting a cohort study
When a cohort enters a systematic review, reviewers judge its internal validity with structured tools. The ROBINS-I risk of bias tool was built specifically for non-randomized studies of interventions and walks through confounding, selection, and measurement domains. For reporting, the STROBE guideline governs cohort manuscripts and asks you to state the source population, the follow-up, the handling of missing data, and the limitations. A cohort whose results are eventually pooled in a meta-analysis has to report these elements consistently, or it cannot be combined with its peers.
Risk, rate, and the right denominator
A cohort can report several distinct measures, and choosing the wrong one quietly misstates the result. Cumulative incidence (risk) is the proportion of an at-risk group that develops the outcome over a fixed period, and its natural contrasts are the risk ratio and the risk difference. It assumes everyone is followed for the same length of time and no one is lost, which is rarely true. When follow-up varies, people enter and leave, or the study runs long enough for competing events, the honest denominator is person-time, and you report an incidence rate (events divided by person-years) and an incidence rate ratio. The risk difference also yields the measures clinicians actually act on, the attributable risk and the number needed to harm, which a ratio alone cannot give. Reporting an absolute measure alongside the ratio is what turns a relative effect into a decision-relevant one.
The moment follow-up is incomplete, with participants entering late, withdrawing, or reaching the end event-free, the analysis should treat time-to-event explicitly rather than collapsing everything into a single proportion. Kaplan-Meier curves describe the unadjusted experience and a Cox proportional hazards model estimates adjusted hazard ratios while using each participant's information up to the moment they are censored. If a participant can experience an event that prevents the outcome of interest, dying of another cause before the studied disease, you have competing risks, and the cumulative incidence function rather than Kaplan-Meier is the correct estimate of absolute risk. A cohort that reports only a crude risk ratio when follow-up was uneven is leaving its strongest analysis on the table.
Adjusting for confounders with multivariable regression is the baseline, but two refinements separate a defensible modern cohort from a dated one. First, choose the adjustment set with a causal diagram (a directed acyclic graph) rather than throwing every available covariate into the model: adjusting for a mediator on the exposure-to-outcome pathway removes part of the real effect, and adjusting for a collider introduces bias, so more covariates is not safer. Second, when a confounder is itself affected by earlier exposure and also predicts later exposure, time-varying confounding, ordinary regression is biased no matter how carefully it is specified, and the correct tools are the g-methods: marginal structural models fitted with inverse probability of treatment weighting, or g-computation. This is the situation that arises whenever exposure and a confounder evolve together over follow-up, for instance a biomarker that both responds to treatment and drives the next treatment decision.
Two structural traps deserve naming because they survive any amount of modelling. The healthy-worker effect is a selection bias in occupational cohorts where employed, exposed people are systematically healthier than the general-population comparison, dragging the apparent effect of exposure toward benefit. Immortal time bias afflicts retrospective cohorts in particular: if group membership is defined by something that can only happen after follow-up has begun (receiving a prescription, surviving to a procedure), then the span before it happened is time the person could not have had the event, and misattributing that span manufactures a spurious survival advantage. The discipline that prevents both is to define the eligible population, the exposure, and time zero at the same instant, the logic of target trial emulation, which asks what randomized trial the cohort is standing in for and aligns the design to it.
library(survival)
library(Epi)
# Incidence rates and the rate ratio on a person-time denominator
pyears_fit <- pyears(Surv(time, event) ~ exposure, data = d, scale = 365.25)
# Adjusted risk ratio (not odds ratio) for a common binary outcome:
# Poisson regression with robust standard errors
library(sandwich); library(lmtest)
pr <- glm(event ~ exposure + age + sex, data = d, family = poisson)
coeftest(pr, vcov = sandwich) # exponentiate for adjusted risk ratios
# Uneven follow-up: hazard ratio from a Cox model
coxph(Surv(time, event) ~ exposure + age + sex, data = d)
Choose by question and constraint. A cohort study is the right tool when you can identify exposure first and afford the follow-up, and when the outcome is common enough to accumulate cases. For a rare outcome, a case-control design reaches an answer far more efficiently. For a pure prevalence question with no causal ambition, a cross-sectional snapshot is faster and cheaper. The design is not a matter of prestige; it is a match between the question you are asking and the time, money, and outcome frequency you are working with.
