Must a cohort study take place in the future?

No. A retrospective cohort reconstructs a defined group's exposure and follow-up from existing records.

Can a cohort study estimate incidence?

Yes, when eligibility, outcome occurrence and follow-up are measured appropriately; the exact measure depends on the design.

Does following the same people remove confounding?

No. Exposure groups can differ in factors that influence outcomes, and those differences need explicit consideration.

AI Glossary / Statistics and probability / Intermediate

Cohort study

Definition

A cohort study follows a defined group of people or other units over time to observe outcomes, often comparing groups with different exposures or characteristics. In the usual observational design, researchers observe exposure rather than assign it randomly. Follow-up can be conducted prospectively or reconstructed from existing records. Cohort studies can estimate outcome occurrence over time, but an observed exposure-outcome association is not automatically a causal effect.

What it means

The study begins conceptually with eligible units and their exposure status, then examines subsequent outcomes. That direction distinguishes it from a case-control design that samples according to outcome status. A retrospective cohort can preserve the same exposure-before-outcome ordering even though investigators assemble the study after the events occurred. For AI research, a cohort might track employees who did or did not begin using a tool. Differences in motivation, workload or access can affect both adoption and outcomes, complicating causal interpretation.

How it works

Specify eligibility, a common time origin, exposure definitions, outcomes and follow-up rules. Ensure that the information used to define exposure is appropriate to the time at which follow-up begins. Identify a comparison group from the same relevant source population and measure potential confounders. Record incident outcomes, time at risk, losses and competing events. Depending on follow-up and the question, analysis may estimate cumulative risks, incidence rates or time-to-event associations. Address missingness, changing exposures and dependence where relevant. A protocol and clear data provenance make the analysis reproducible and reduce opportunities for outcome-driven choices.

Three practical examples

Workplace research: follow employees from a defined eligibility date to examine outcomes associated with adoption of an assistant.
Health research: compare subsequent outcome occurrence among people with different measured exposures without assigning those exposures.
Reliability analysis: follow newly commissioned devices under different operating conditions and record failures and observation time.

Think of it this way

Think of watching two groups begin a journey and recording what happens along the way. You can see which events occur after departure, but the groups may already differ in experience, equipment or route choice. Following them carefully improves the record; it does not make their starting conditions equivalent.

Worked example

Consider a hypothetical closed cohort with complete one-year follow-up. There are 400 exposed and 600 unexposed people, all initially free of the outcome being counted. Forty exposed people and thirty unexposed people experience the outcome. The one-year risks are 40/400 = 10% and 30/600 = 5%. The risk ratio is 2, and the risk difference is 5 percentage points. These counts show an association in this cohort. They do not prove that exposure caused the excess because baseline risk factors may differ. If follow-up lengths were unequal or substantial outcomes were unknown, simply dividing events by initial group size would not automatically estimate comparable one-year risks. Person-time rates or survival methods address different aspects of follow-up and require their own assumptions.

Illustrative case study

Fictional case: Cedar Workplace Research studies whether starting an AI assistant is associated with later task completion. It defines eligibility and baseline workload before classifying adoption and follows both groups from comparable dates. Investigators discover that early adopters have more technical support and different responsibilities, so they document those differences and consider appropriate adjustment. They also avoid labeling someone an adopter based on use that occurs only months after follow-up begins. The final report distinguishes an adjusted association from a randomized intervention effect. This invented case illustrates design concerns, not evidence that assistant use improves productivity.

Common misconceptions and mistakes

Prospective does not mean randomized, and retrospective does not automatically mean invalid. Accurate historical records can support a well-designed cohort, while poor prospective measurement can still bias results. Loss to follow-up can distort estimates if it relates to exposure and outcome. Defining exposure using future survival or participation can introduce immortal-time bias. Adjustment cannot guarantee removal of unmeasured confounding. Risk, rate and hazard are different quantities, so their ratios should not be casually interchanged. A large database can give a very precise estimate of a biased association.

Business use

Businesses use cohort studies when longitudinal evidence is needed and randomization is unavailable or unsuitable. They can reveal outcome patterns, support monitoring and generate hypotheses for stronger evaluations. For AI deployments, document system versions, changing access policies and exposure timing, since use may change during follow-up. Predefine the operational question and the minimum data needed to answer it. Report absolute outcome frequencies alongside relative measures and acknowledge whether the cohort resembles the intended deployment population. Interpretation should remain tied to the measured exposure and observation period.

Check your understanding

Why is the worked example's risk ratio of 2 insufficient to claim a causal doubling?

Answer: Exposure was observed rather than randomly assigned, and the groups may differ in baseline causes of the outcome. The ratio describes observed risks; causal interpretation needs additional design, measurement and identification assumptions.

Related terms

Cohort
Prospective study
Confounding
Survival analysis
retrospective cohort
incidence
risk ratio
rate ratio
loss to follow-up
immortal-time bias

Learn next

compare cohort and case-control sampling, then distinguish cumulative risk from person-time incidence rates. Draw a timeline showing eligibility, exposure assignment and follow-up to uncover inconsistent time origins. Study confounding and informative loss before interpreting adjusted estimates, and examine how time-varying exposure changes the analysis required.

Sources

CDC, Designing and Conducting Analytic Studies in the Field: https://www.cdc.gov/field-epi-manual/php/chapters/design-conduct-analyze-field-studies.html Harrell, Glossary of Statistical Terms, cohort study: https://hbiostat.org/glossary/ CDC, Analyzing and Interpreting Data: https://www.cdc.gov/field-epi-manual/php/chapters/analyze-interpret-data.html

Previous: Cohort Next: Cointegration