Cohort Study: Prospective, Retrospective, and Examples

What is a cohort study?

The word cohort comes from the Latin for a Roman military unit — a group that moves together. In research, a cohort is a group defined by a shared experience (an exposure to a risk factor, enrollment in a program, or birth in a particular period) and followed forward in time to observe outcomes.

In a cohort study, participants are classified at baseline into two groups: those exposed to the factor of interest and those unexposed. Both groups are then followed for a defined period, and the incidence of the outcome (disease, death, recovery, or any other measurable event) is compared between groups. Because exposure is measured before the outcome develops, cohort studies can establish the temporal sequence required for causal inference — a feature that cross-sectional and case-control designs cannot provide as cleanly.

Cohort studies are observational: the researcher does not assign exposures. This distinguishes them from randomized controlled trials (RCTs), where random assignment controls confounding. In a cohort study, confounding must be addressed at the design stage (matching, restriction) or the analysis stage (multivariable regression, propensity score methods).

Prospective cohort studies

In a prospective cohort study, exposure status is assessed at the start of the study and participants are followed forward in time until the outcome occurs or the follow-up period ends. Data are collected as events unfold, which has several advantages.

• Exposure measurement quality — because exposure is measured before the outcome, there is no recall bias from participants trying to explain a disease they already have.
• Multiple outcomes — a single prospective cohort can examine many different outcomes from the same baseline exposure, making it highly efficient.
• Incidence data — the researcher can calculate true incidence rates and absolute risks, not just odds ratios.
• Temporal clarity — the time sequence (exposure precedes outcome) is established by design.

The primary disadvantages are cost and time. Prospective studies can run for decades, require sustained funding, and demand ongoing participant engagement. Rare outcomes require very large cohorts to accumulate enough cases for analysis. The Nurses' Health Study, launched in 1976, enrolled 121,700 US nurses and has generated hundreds of publications on diet, lifestyle, and chronic disease — but it required decades of follow-up to do so.

Retrospective cohort studies

A retrospective cohort study (also called a historical cohort study) assembles the cohort using records that already exist. The researcher identifies a past exposure, then looks back through historical data to determine who was exposed and forward through records to determine who developed the outcome — all without the waiting time of a prospective design.

This approach is much faster and cheaper than prospective designs because it relies on pre-existing data such as employment records, medical charts, insurance databases, or national registries. It is particularly useful for rare diseases with long latency periods, such as occupational cancers, because a group of exposed workers from decades ago can be traced through death certificates or tumor registries.

The tradeoff is data quality. Because records were created for purposes other than research, exposure information may be incomplete, inconsistently recorded, or measured with instruments that would not meet current standards. Confounding variables that were not recorded at the time cannot be controlled after the fact. The researcher is constrained by what was measured, not what they would ideally want to know.

Cohort vs. case-control studies

Cohort and case-control designs are both observational, but they start from different points and answer questions differently.

Cohort studies are generally considered stronger evidence than case-control studies for establishing causation — particularly prospective designs where exposure is measured before the outcome. However, for rare diseases or those with long latency, case-control designs are more practical.

Calculating relative risk

The primary effect measure in a cohort study is the relative risk (RR), also called the risk ratio. It compares the incidence of an outcome in exposed individuals to the incidence in unexposed individuals.

An RR of 1.0 means no association; RR > 1 indicates increased risk in exposed individuals; RR < 1 indicates a protective effect. Confidence intervals and p-values are essential for judging the precision and statistical significance of the estimate. Multivariable regression models (Cox proportional hazards for time-to-event data, Poisson regression for incidence rates) are used to adjust for confounders.

Limitations and threats to validity

Loss to follow-up

The greatest methodological threat to prospective cohort studies is loss to follow-up — participants who drop out, move away, or die from causes unrelated to the outcome of interest. If attrition is differential (exposed participants are more or less likely to drop out than unexposed), the observed association will be biased. Researchers track follow-up rates, compare baseline characteristics of those who stayed versus those lost, and use sensitivity analyses to assess the potential impact of missing data.

Confounding

Because exposure is not randomly assigned, exposed and unexposed groups often differ on other characteristics that also predict the outcome. For example, smokers may also drink more alcohol, exercise less, and have lower incomes — all of which are independent risk factors for various diseases. Multivariable adjustment can control for measured confounders, but unmeasured confounding remains an irreducible limitation of all observational designs.

Selection bias

The healthy worker effect is a classic example: people who are employed tend to be healthier than the general population, so occupational cohorts may underestimate disease risk relative to a general-population comparison group. Careful selection of the comparison group and restriction criteria are needed to minimize this bias.

Information bias

In retrospective designs, exposure information extracted from historical records may be inaccurate or incomplete. In prospective designs, self-reported exposures (diet, physical activity, alcohol intake) are subject to measurement error and social desirability bias. Objective biomarkers or repeated measurements over time can improve exposure assessment.

Notable examples in epidemiology

Framingham Heart Study

Launched in 1948, the Framingham Heart Study enrolled 5,209 adults from Framingham, Massachusetts, and began following them every two years with clinical examinations. It identified the major cardiovascular risk factors — elevated cholesterol, high blood pressure, smoking, obesity, and physical inactivity — and gave the world the concept of the "risk factor" itself. The study's second and third generations are still being followed today, making it one of the longest-running cohort studies in history.

British Doctors' Study

Doll and Hill (1954) enrolled 40,000 British physicians and followed them for decades. By comparing smoking rates and lung cancer incidence between smokers and non-smokers, they provided the first convincing prospective evidence that cigarette smoking causes lung cancer — evidence that eventually led to landmark public health interventions worldwide.

Nurses' Health Study

Beginning in 1976 with 121,700 married female nurses aged 30–55, the Nurses' Health Study collected detailed dietary and lifestyle information every two years. It has produced influential findings on oral contraceptives and breast cancer risk, dietary fat and heart disease, and the health effects of hormone replacement therapy — often overturning earlier case-control findings.

UK Biobank

The UK Biobank enrolled approximately 500,000 participants aged 40–69 between 2006 and 2010, collecting biological samples, imaging data, and linked electronic health records. It exemplifies the modern biobank-based prospective cohort: massive scale, deep molecular phenotyping, and linkage to national health registries, enabling genome-wide association studies (GWAS) and large-scale epidemiological analyses across hundreds of outcomes simultaneously.

Quick summary