Must a cohort be followed over time?

The group can be defined by a shared feature without longitudinal analysis, although follow-up is a common use.

Is a cohort the same as a cohort study?

No. A cohort is a group; a cohort study is a research design using follow-up to investigate outcomes.

Can a cohort include organizations rather than people?

Yes. The data units can be businesses, devices or other entities sharing the specified feature.

AI Glossary / Statistics and probability / Beginner

Cohort

Definition

A cohort is a group of people or other data units that share a defined characteristic or experience. Examples include people born during a particular period, customers who joined in the same month or employees who completed a specified training program. Cohorts help organize comparisons and follow outcomes over time. The term describes the group; it does not by itself specify a research design, sampling method or causal relationship.

What it means

The shared feature gives membership a clear meaning. A signup cohort might contain everyone whose first account activation occurred in January, regardless of whether they remain active later. A birth cohort shares a period of birth, while a training cohort shares an enrollment or completion event. Membership and measurement rules must be explicit. A list of people currently active is a changing segment unless the cohort definition deliberately uses that rule. A cohort can consist of businesses, machines or other units as well as people.

How it works

Define the unit, entry condition, relevant dates and any exclusions. Assign each eligible unit using information consistent with that definition. Set a time origin, such as first activation, when comparing subsequent outcomes. Track follow-up, missing records and exits separately from the original membership rule. Distinguish a fixed entry group from an open population in which eligible units may enter over time. For comparisons across cohorts, align the period since entry and check whether calendar conditions, acquisition channels or measurement practices differ. The resulting summaries describe the defined groups and observation windows, not necessarily all future users.

Three practical examples

Product adoption: follow users who first activated an AI assistant during the same month and measure later engagement.
Learning evaluation: compare completion and follow-up outcomes for successive groups entering an AI literacy course.
Equipment monitoring: examine machines commissioned during the same period to compare reliability at similar ages.

Think of it this way

Think of a school graduating class. Members share a milestone, and the group remains meaningful even when people move to different places or stop responding to surveys. A reunion attendance list is not the same as the original class roster. Keeping that distinction prevents misleading denominators.

Worked example

Suppose a fictional assistant has 200 first-time users in January and 150 in February. At 30 days after each person's first activation, 120 January users and 90 February users meet a predefined activity criterion, with complete measurement for both groups. The 30-day activity proportions are 120/200 = 60% and 90/150 = 60%. Combined, 210/350 = 60%. Comparing January users at 60 days with February users at 30 days would answer a different question because follow-up durations differ. If February has not yet reached 60 days, its 60-day value is not zero; it is not yet observed. If ten January records later become unavailable, dropping those people from the original denominator requires an explicit missing-data policy rather than a silent recalculation.

Illustrative case study

Fictional case: Ash Learning launches a course in monthly intakes. Its dashboard initially compares all active learners each month, mixing new entrants with people who have been studying for several months. The team creates entry cohorts and compares completion at the same time since enrollment. It preserves the original cohort sizes and marks follow-up periods that have not matured. Analysts still avoid claiming that a curriculum change caused a difference because recruitment sources and seasonal workloads also changed. This invented case illustrates clearer grouping and comparison, not a measured educational effect.

Common misconceptions and mistakes

A cohort is not automatically a randomized group, representative sample or controlled experiment. Defining membership by a later outcome can create selection bias when studying earlier exposures. Comparing only survivors or active users can hide attrition and exclude people with poor experiences. Calendar time, time since entry and age can be entangled, especially in demographic comparisons. Cohorts may overlap if definitions permit it, but that overlap must be understood when interpreting statistical comparisons. A large cohort does not compensate for an ambiguous membership rule or inconsistent measurement.

Business use

Businesses use cohort analysis to separate changes in customer composition from changes in experience after entry. It can clarify whether apparent growth comes from new acquisition, sustained use or repeated reactivation. For AI services, record the product version and major policy changes alongside cohort dates when those affect interpretation. Use a consistent definition of activity and distinguish missing telemetry from observed inactivity. Small cohorts may need uncertainty intervals or suppression in public reporting, while any interpretation should reflect the populations actually included.

Check your understanding

Why should an unobserved 60-day outcome for the February group not be recorded as zero?

Answer: The required follow-up has not occurred. Zero would claim an observed absence of activity, while the correct status is not yet measurable. Confusing those states would systematically disadvantage newer cohorts.

Related terms

Cohort study
Longitudinal data
Survivorship Bias
Time origin
Population
entry date
retention
attrition
follow-up
birth cohort

Learn next

distinguish fixed cohort membership from a changing user segment. Build a small table that separates calendar month from months since entry, and mark incomplete follow-up explicitly. Then study missing data, censoring and the difference between descriptive cohort comparisons and causal evaluations. Practice writing a membership rule another analyst could reproduce without guessing.

Missing data

Sources

Australian Bureau of Statistics, Statistical terms and concepts glossary, Cohort: https://www.abs.gov.au/statistics/understanding-statistics/statistical-terms-and-concepts/statistical-terms-and-concepts-glossary 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: https://hbiostat.org/glossary/

Previous: Coherent extrapolated volition Next: Cohort study