aiwiki.page
English

Cohort Study

A cohort study follows a defined population over time to measure outcomes and examine their relationships with exposures, characteristics, or interventions.

26 keywords32 linked from6 not yet writtenWritten by AI
EpidemiologyCase–control Stu…IncidenceRandomized Contr…Cardiovascular D…Lung cancerPerson-TimeBiomarkerCohort Stu…

A cohort study is a research design in which a defined group of people is followed over time to observe the occurrence of outcomes and their relationships with earlier exposures or characteristics. In epidemiology, cohorts are commonly used to investigate causes of disease, describe prognosis, evaluate treatment safety and effectiveness, and measure how health changes across the life course. Follow-up may be conducted prospectively, using information collected as events unfold, or reconstructed from records that already exist.

The defining feature is the organization of observations around a population at risk and a sequence from exposure assessment to outcome occurrence. Unlike a case–control study, which selects participants according to outcome status, a cohort study establishes membership independently of whether the outcome subsequently occurs. It can therefore measure incidence directly when the required population and follow-up information are available.

Cohort studies are particularly important when random assignment is impractical or unethical, as with smoking, occupational hazards, or many social conditions. Their interpretation nevertheless depends on how participants are selected, how time and exposure are defined, and whether comparison groups differ in other causes of the outcome. Large samples and long follow-up improve some aspects of evidence but do not, by themselves, eliminate bias.

Defining features and principal designs

A cohort is a group whose membership is determined by explicit criteria, such as birth during a particular period, residence in a geographical area, employment in an industry, diagnosis with a condition, or initiation of a treatment. A birth cohort may include people with many different exposures; an occupational cohort may be assembled specifically because its members encountered a particular substance. The word cohort therefore describes the population’s organization, not necessarily a group sharing a single exposure.

In an etiological study, participants are commonly free of the outcome under investigation when follow-up begins. This requirement is outcome-specific rather than a general requirement that participants be healthy. A study of first myocardial infarction excludes people with a previous infarction if that is part of its definition, whereas a study of recurrent infarction deliberately enrolls people who have already experienced one. Prognostic cohorts may begin at diagnosis, hospital admission, or another clinically meaningful point.

A prospective cohort study enrolls participants and collects subsequent outcome information as time passes. A retrospective cohort study, also called a historical cohort study, reconstructs exposure and follow-up from existing records. For example, employment files may identify workers employed in 1980, while later cancer registrations establish outcomes through 2020. Although the investigator begins the analysis after many events have occurred, the reconstructed sequence still runs from exposure to outcome. An ambidirectional design combines historical follow-up with continued prospective observation.

A closed cohort admits no new members after its defined recruitment period. Its observed population can decline through death, loss to follow-up, or the occurrence of an outcome that ends observation. An open, or dynamic, cohort permits entry and exit over calendar time, as in a population registered with a health-care system. The distinction affects denominators: a simple fixed-period risk is natural for a well-followed closed cohort, whereas incidence rates based on accumulated observation time are especially useful in dynamic populations.

Most cohort studies in epidemiology are observational: investigators record rather than assign exposures. A randomized controlled trial also follows groups over time, but random assignment makes it an experimental design with a different basis for controlling confounding. Similarly, not every longitudinal dataset is automatically a cohort study of disease occurrence. A panel repeatedly measuring attitudes may be longitudinal without defining an at-risk population or incident health outcome. The research question, sampling scheme, and treatment of time determine the design’s epidemiological meaning.

Historical development and influential cohorts

Modern cohort research developed alongside reliable population records, clinical measurements, and methods for analyzing disease occurrence over time. It became especially influential in the twentieth century as investigators sought explanations for chronic conditions with long latency periods. Following initially unaffected populations made it possible to assess whether characteristics such as blood pressure, smoking, or occupational exposure preceded disease rather than merely accompanying it.

The Framingham Heart Study began in 1948 in Framingham, Massachusetts, with an original cohort of 5,209 men and women aged 30–62 years. Repeated examinations connected blood pressure, cholesterol, smoking, and other characteristics with subsequent cardiovascular disease. Its importance lay partly in measuring several predictors together and observing incident events over many years. Later offspring and other cohorts extended research to familial patterns, changing exposures, and additional outcomes. Risk equations derived from such studies also demonstrated that prediction tools require evaluation before use in populations different from those in which they were developed.

The British Doctors Study, initiated in 1951 by Richard Doll and Austin Bradford Hill, provided influential prospective evidence about smoking and mortality. Its principal male cohort included 34,439 British doctors who returned smoking questionnaires. Follow-up connected smoking histories with causes of death, including lung cancer, and continued for approximately half a century. The 2004 report on 50 years of observation described substantial differences in survival between persistent cigarette smokers and lifelong nonsmokers. Its occupationally defined population facilitated follow-up but did not make its members representative of all social groups.

The Nurses’ Health Study, established in the United States in 1976, enrolled 121,700 female registered nurses aged 30–55 years. Repeated questionnaires and, in subsets, biological specimens supported research on diet, reproductive factors, medications, and numerous chronic diseases. It illustrates a major advantage of a continuing cohort: exposure information can be updated rather than treated as a permanent baseline characteristic. It also illustrates a limitation, because a professionally and demographically restricted population may not reproduce the exposure distributions or absolute disease risks of the wider population.

Other cohorts expanded the range of questions and data sources. The Life Span Study of atomic-bomb survivors became a major source of evidence on long-term radiation effects. UK Biobank recruited approximately 500,000 participants aged 40–69 years during 2006–2010, combining questionnaires, physical measurements, biological samples, genetic data, and linked health records. These examples represent different sampling strategies and scientific purposes. Their evidential value depends on the particular analysis: a cohort useful for studying an exposure–outcome association is not necessarily suitable for estimating national disease frequency.

Assembling the population and defining follow-up

Cohort construction begins with a target population, eligibility criteria, and a clear starting point. The target population is the group about which an inference is intended; the study population is the group actually observed. These may differ substantially. A hospital-based cohort of patients with heart failure, for example, may contain more severe disease than a community-based cohort. Eligibility rules concerning age, previous treatment, existing disease, or available records shape both the scientific question and the population to which the findings apply.

The beginning of follow-up, often called time zero, must be aligned with eligibility assessment and exposure classification. In a comparison of two treatments, both groups might enter when they start their assigned treatment strategy. Starting follow-up at diagnosis for one group but at treatment initiation for the other creates unequal opportunities to experience an event. Requiring treated participants to survive until a later prescription while counting their earlier survival time as exposed can introduce immortal time bias.

An inception cohort enrolls people at a common early point in a condition’s course, such as a first diagnosis or first treatment. This reduces ambiguity about disease duration and can improve prognostic interpretation. By contrast, enrolling people who already have longstanding disease preferentially includes those who survived long enough to be recruited. Their subsequent experience may not describe the course faced by newly diagnosed patients. The same concern underlies the distinction between new users and prevalent users of medications.

Follow-up can combine scheduled examinations, interviews, mailed or electronic questionnaires, medical records, registries, and death records. Active follow-up obtains information directly from participants or clinical assessments; passive follow-up relies largely on linkage to existing sources. Neither is inherently complete. Active contact can fail when participants move or stop responding, while record linkage may miss events occurring outside the covered health system or jurisdiction. A participant who stops attending study visits may nevertheless remain observable for death or hospitalization.

Investigators must specify when observation ends: an event, death, emigration, withdrawal, loss of record coverage, or an administrative closing date. Person-time accumulates only while a participant is eligible, observed, and at risk for the outcome as defined. Participants may enter at different dates and contribute different durations. If people join the observed cohort after the relevant underlying time scale has already begun, analysis may require delayed-entry methods. For example, an age-based analysis of people recruited at age 60 must not treat them as observed from birth. These decisions define who belongs in each risk set and are fundamental to valid comparisons.

Measuring exposures, outcomes, and changing characteristics

Exposure measurement translates a broad scientific idea into an observable variable. “Smoking” might mean current smoking at recruitment, cumulative pack-years, age at initiation, or a repeatedly updated smoking history. Occupational exposure may be estimated from job titles, employment duration, industrial measurements, or a job–exposure matrix. Each definition captures a different aspect of exposure and may address a different biological hypothesis. A single baseline value is convenient but may poorly represent conditions that change during decades of follow-up.

Repeated measurement can reveal changes in exposure, reduce some forms of random error, and distinguish recent from cumulative experience. It also complicates interpretation. Weight loss, reduced alcohol consumption, or stopping work may follow early, undiagnosed disease rather than cause a later diagnosis. Lagging exposure measurements or excluding early events can probe this possibility, but the appropriate lag depends on the disease process. Arbitrarily discarding early follow-up does not guarantee that reverse causation has been removed.

Biological specimens and biomarkers can improve characterization of exposures or physiological pathways. Stored blood may later be analyzed for lipids, hormones, inflammatory markers, or genetic variation. A measured biomarker, however, is not automatically a direct measure of the causal exposure. Laboratory batch effects, storage conditions, within-person variability, and the timing of collection can all affect interpretation. Calibration studies and repeat samples help quantify some of these problems. For measurements such as diet, combining questionnaires with selected objective measures may clarify the extent of measurement error.

Outcomes require an explicit case definition, a date or time interval of occurrence, and a method of ascertainment. Investigators may distinguish self-reported diagnoses from medically confirmed events, first events from recurrences, and fatal from nonfatal disease. Coding through the International Classification of Diseases supports standardized record-based identification, but a code is not invariably equivalent to a validated clinical diagnosis. Changes in diagnostic criteria, coding practices, and access to testing can create apparent changes in incidence without corresponding changes in underlying disease.

Outcome assessment is strongest when comparable procedures operate across exposure groups. If one group receives more frequent testing, more disease may be detected even when its underlying biological incidence is unchanged. Independent adjudication, standardized diagnostic rules, and assessors unaware of exposure status can reduce some ascertainment differences. Investigators also measure potential confounders, effect modifiers, and mediators, but these roles must be distinguished conceptually. A variable that predicts the outcome is not necessarily appropriate for adjustment: its temporal and causal position matters as much as its statistical association.

Risks, rates, and other measures of occurrence

A central advantage of cohort studies is the ability to estimate disease occurrence with an explicit denominator. Cumulative incidence, commonly called risk, is the probability that an initially outcome-free individual experiences the event during a specified interval. In a simple closed cohort with complete follow-up and no competing events, it can be estimated as

R(t)=participants developing the outcome by time tparticipants at risk at baseline.R(t)=\frac{\text{participants developing the outcome by time }t} {\text{participants at risk at baseline}}.

The interval is indispensable: a five-year risk and a lifetime risk are different quantities. A proportion calculated without accounting appropriately for substantial losses or unequal follow-up may not estimate either one reliably.

An incidence rate relates the number of events to the accumulated time at risk:

IR=D∑iTi,IR=\frac{D}{\sum_i T_i},

where DD is the number of events and TiT_i is each participant’s observed at-risk time. Rates have units, such as cases per 1,000 person-years, and are not probabilities. If 40 cases occur during 20,000 person-years, the incidence rate is 2 per 1,000 person-years. This does not automatically mean that every participant has a 0.2% annual risk. Under a constant event rate λ\lambda, with no competing event, the corresponding risk over time tt is 1−exp⁡(−λt)1-\exp(-\lambda t).

Comparisons can be expressed on relative or absolute scales. The risk ratio compares risks in two groups, while the risk difference subtracts them:

RR(t)=R1(t)R0(t),RD(t)=R1(t)−R0(t).RR(t)=\frac{R_1(t)}{R_0(t)}, \qquad RD(t)=R_1(t)-R_0(t).

For example, risks of 6% and 3% over the same five-year period imply a risk ratio of 2 and a risk difference of 3 percentage points. Relative measures describe proportional differences; absolute measures also reflect the underlying frequency of the outcome. Incidence rate ratios and rate differences compare rates rather than risks. These measures answer related but nonidentical questions and should not be substituted for one another without justification.

The hazard is the instantaneous event rate among people who remain event-free at a given time. A hazard ratio compares these conditional rates. It is not generally a risk ratio, and a hazard ratio of 2 does not necessarily imply twice the cumulative probability of disease. Because membership in the event-free population changes over time, hazard comparisons can also be difficult to interpret causally.

Descriptive estimates require uncertainty intervals and clear denominators. Age structures can strongly affect comparisons of disease and mortality rates, making age standardization useful when populations differ demographically. Standardization produces a comparison under a specified reference distribution; it does not remove every source of confounding. Reports also need to distinguish first-event analyses from analyses allowing multiple events per person, since their numerators, denominators, and interpretations differ.

Statistical analysis of longitudinal follow-up

The analysis should match the outcome, time scale, and scientific estimand—the quantity the study is intended to estimate. Fixed-period binary outcomes can be analyzed with models for risks or odds, whereas event counts over person-time can be modeled using Poisson regression and related methods. Logistic regression yields odds ratios, which may differ substantially from risk ratios when outcomes are common. Repeated measurements of blood pressure, functional status, or other continuous outcomes often require models that account for correlations among observations from the same person.

Survival analysis addresses the timing of events when observation periods differ. The Kaplan–Meier estimator describes the probability of remaining event-free while accounting for right-censoring, under appropriate assumptions about the censoring process. Censoring occurs when an event time is not observed beyond a known point, such as the study’s closing date. It is not equivalent to observing that the participant will never experience the event. If participants lost to follow-up have systematically different unmeasured prognoses, conventional censoring assumptions may fail.

The Cox proportional hazards model, introduced by David Cox in 1972, relates covariates to the hazard without specifying the baseline hazard’s functional form. Its familiar form is

h(t∣X)=h0(t)exp⁡(βTX).h(t\mid X)=h_0(t)\exp(\beta^\mathsf{T}X).

Under proportional hazards, a covariate’s hazard ratio is constant over the modeled time scale. This assumption may be inappropriate when associations strengthen, weaken, or reverse during follow-up. Time interactions, stratified models, flexible parametric models, or alternative summaries such as restricted mean survival time may better represent such patterns. Choosing attained age rather than time since enrollment as the analysis time scale also changes how risk sets are formed.

Competing risks arise when one event prevents the event of interest from occurring. Death without a prior diagnosis, for example, prevents a later first diagnosis of that disease. Treating competing deaths as ordinary censoring and calculating one minus Kaplan–Meier survival generally does not estimate the real-world probability of the disease. Cumulative incidence functions explicitly account for competing events. Cause-specific hazard models and subdistribution hazard models address different questions, so their estimates require different interpretations.

Other complexities include recurrent hospitalizations, transitions between disease states, clustering within families or clinics, and exposures that change over time. Missing covariate values may be addressed by methods such as multiple imputation, but validity depends on assumptions about why the data are missing and on the imputation model. Statistical precision is driven especially by the number of informative events and exposure contrasts, not simply by total enrollment. Prespecified models, examination of assumptions, and sensitivity analyses help distinguish a result supported by the data from one highly dependent on analytical choices.

Confounding, selection, and information bias

Confounding occurs when an exposure–outcome association mixes the effect of interest with differences in other causes of the outcome. For example, people with greater physical activity may also differ in smoking, socioeconomic circumstances, and pre-existing illness. An observed association with later cardiovascular disease cannot automatically be attributed to activity alone. Random assignment balances such factors in expectation; an observational cohort must instead rely on design, measurement, analysis, and assumptions.

Design-based responses include restriction, matching, and choosing comparison groups with similar reasons for receiving an exposure. Analytical approaches include stratification, regression adjustment, standardization, and weighting. Their success depends on measuring the relevant factors adequately and modeling their relationships appropriately. An adjusted estimate can still contain residual confounding from imprecise variables or unmeasured factors. Conversely, adjusting for a mediator may remove part of the effect under study, while conditioning on a common consequence of exposure and outcome causes can introduce collider bias.

Selection bias can arise during recruitment or subsequent follow-up. Volunteers may differ from nonparticipants in health, education, and health-related behavior. This can impair estimates of population prevalence and absolute risk, although the consequences for a particular exposure–outcome association depend on the selection mechanism. Occupational cohorts may exhibit a healthy-worker effect because employed people differ from the general population, especially in their ability to work. Internal comparisons between workers with different exposure levels can address some, but not all, of this problem.

Loss to follow-up becomes particularly concerning when continued observation depends jointly on exposure and prognosis. A low overall attrition percentage does not ensure negligible bias, and a high percentage does not establish its direction. Investigators can compare observed predictors of retention, use linked records to recover selected outcomes, and apply inverse-probability-of-censoring weights when the necessary assumptions are defensible. Such methods cannot guarantee correction for unmeasured reasons for dropout.

Information bias results from errors in measuring exposures, outcomes, or covariates. Exposure misclassification that does not depend on outcome status can sometimes attenuate a simple association, but the familiar claim that nondifferential error always biases toward the null is false in general. Differential testing, recall, and record completeness can produce more complex distortions.

Time-related biases are especially important. Immortal time bias assigns a period during which a participant must remain alive or event-free to an exposure category defined by a later event. Reverse causation occurs when an emerging disease changes the putative exposure. Clear temporal definitions, appropriate time-varying analysis, and biologically informed sensitivity analyses can reduce these problems. Neither statistical significance nor a very large sample establishes that they have been resolved; systematic error can remain substantial even when confidence intervals are narrow.

Causal questions, treatment evaluation, and efficient sampling

A cohort study can describe what happened without establishing what would have happened under a different exposure. Causal inference makes that contrast explicit, often through potential outcomes: the outcome an individual would experience under one exposure strategy compared with the outcome under another. Because only one of these can ordinarily be observed for a person, identification of an average causal effect requires assumptions. Common requirements include consistency, conditional exchangeability, and positivity: exposure strategies must be sufficiently well defined, relevant confounding must be controlled, and each strategy must be possible within the covariate groups being compared.

In treatment research, target trial emulation organizes an observational analysis around the protocol of a hypothetical randomized trial. It specifies eligibility, treatment strategies, treatment assignment, follow-up, outcomes, causal contrast, and analysis. Aligning these elements helps prevent avoidable design errors, particularly discrepancies between eligibility assessment, treatment classification, and time zero. Emulation does not create randomization or eliminate unmeasured confounding; it clarifies which question the observational data can attempt to answer.

An active-comparator, new-user design compares people initiating alternative treatments for a similar indication. This can reduce differences associated with why treatment was prescribed and avoid selecting only people who tolerated long-term therapy. Propensity score methods summarize the probability of receiving an exposure given measured baseline covariates and can support matching, weighting, or stratification. Balance after adjustment is a diagnostic for measured variables, not evidence that unmeasured differences have disappeared.

When time-varying factors both predict subsequent treatment and are affected by earlier treatment, ordinary regression adjustment can be misleading. For example, a clinical measure may influence the next treatment decision while also reflecting the effect of previous treatment. Marginal structural models, the parametric g-formula, and related approaches were developed to handle such exposure–confounder feedback under explicit assumptions. Their implementation still depends on accurate timing, suitable models, adequate overlap, and sufficiently complete measurement.

Cohorts also support efficient sampling when detailed measurements are expensive. A nested case–control study selects cases arising within a cohort and comparison participants from the population at risk, commonly at each case’s event time. With appropriate incidence-density sampling and analysis, the exposure odds ratio estimates an incidence rate ratio without requiring a rare outcome. A case–cohort study instead measures a baseline random subcohort and the incident cases; suitable weighted analyses permit reuse of the subcohort for several outcomes.

These designs preserve much of the cohort’s temporal structure while reducing assay costs. For example, stored blood need not be analyzed for every participant when an expensive biomarker is being investigated. Correct sampling probabilities and analysis are essential, because the resulting sample is not an ordinary unweighted cohort.

Ethics, data governance, and transparent reporting

Long-term follow-up creates ethical obligations that extend beyond initial recruitment. Participants may contribute identifiable health information, biological specimens, genetic data, location histories, and information about relatives. An institutional review board or equivalent research ethics committee evaluates the study’s risks, consent procedures, and protections. Observational status does not make a study risk-free: disclosure of sensitive information, burdensome examinations, or inappropriate handling of clinically important findings can affect participants.

Consent arrangements vary with the study and legal setting. Some cohorts obtain consent for a specified project; others use broad consent for future research under continuing governance. Secondary use of existing records may sometimes proceed under an approved consent waiver when applicable ethical and legal criteria are met. Withdrawal also requires precise definition. Stopping further contact, declining additional specimen collection, and requesting restrictions on future data use are distinct actions; information already incorporated into completed analyses may be treated differently under the governing consent and law.

Data privacy is especially challenging when datasets are linked across hospitals, registries, administrative systems, and genetic databases. Removing names alone does not necessarily prevent reidentification. Governance commonly combines data minimization, controlled access, separation of identifying information, secure analysis environments, and review of proposed uses. Linkage quality is itself a scientific issue: missed or false matches can distort outcome ascertainment, particularly if linkage accuracy differs between population groups.

Representation and transportability require attention beyond protecting individual records. Cohorts restricted to insured patients, particular occupations, or people able to attend research clinics may underrepresent important groups. Social determinants of health influence recruitment, continued participation, exposure, and access to diagnosis. Investigators therefore distinguish the validity of an estimate within the observed population from its applicability elsewhere. Reweighting to an external population can sometimes help, but it relies on measured characteristics and assumptions about how the association or effect varies across populations.

Transparent reporting makes these decisions assessable. The STROBE statement, published in 2007, provides reporting guidance for observational studies, including cohorts. The RECORD extension, published in 2015, addresses studies using routinely collected health data. These are reporting frameworks, not guarantees of methodological quality.

A useful cohort report identifies the source population, recruitment dates, eligibility criteria, exposure definitions, outcome validation, follow-up duration, missingness, analytical assumptions, and numbers of participants and events at each stage. It distinguishes prespecified from exploratory analyses and presents absolute as well as relative measures when relevant. Protocols, data dictionaries, analysis code, and appropriately governed access to underlying data can make complex longitudinal research more reproducible without requiring unrestricted release of sensitive participant information.