Cohort analysis groups customers by a shared starting event or characteristic and compares their outcomes at the same age. Choose a stable cohort assignment, define period zero, build an eligible denominator for each age, and separate calendar effects from lifecycle effects. Incomplete recent cohorts must be masked or labeled rather than compared with mature groups.
Who this is for: Analysts examining how acquisition timing, onboarding changes, or customer attributes relate to behavior over time.
- Assign each entity to a clearly defined cohort that remains stable throughout the analysis.
- Compare cohorts at equal ages so newer groups are not penalized for having less observation time.
- Inspect counts, mix, and external calendar conditions before crediting a launch or campaign.
Choose a cohort that answers the question
Acquisition cohorts group customers by the week or month they first became eligible. Behavioral cohorts group them by a shared action, and attribute cohorts use a stable property such as initial plan. Select the grouping from the business question. A pricing-launch evaluation usually needs signup timing and initial offer, not the customer's current plan.
Define one assignment rule. If a returning customer creates a second account, decide whether the entity is the person, account, or subscription. Avoid changing historical membership when attributes evolve. Store both original and current attributes when each serves a different question, and state which one appears in the analysis.
Create a comparable age axis
Period zero begins at the qualifying event and should mean the same thing for every member. Then calculate age in complete days, weeks, billing cycles, or months. Calendar-month subtraction can give customers different exposure depending on signup day. Choose intervals that fit the customer rhythm and document boundary handling.
Recent cohorts have not reached later ages. Mark those cells unavailable rather than zero. For an age-three-month metric, include only customers whose observation window has fully elapsed, unless the method explicitly handles censoring. Show the eligible count because percentages based on shrinking populations become unstable and easy to overread.
Select outcome and denominator
A cohort table can show retention, cumulative revenue, repeat purchase, feature adoption, or support incidence. Define whether the value is period-specific or cumulative. Cumulative revenue cannot decline, while monthly active rate can. Label units and denominator so a heat map does not encourage comparisons between unlike measures.
Decide how refunds, pauses, account merges, and data deletion affect eligibility. For retention, an original-cohort denominator answers how much of the starting group remains. A surviving-customer denominator answers a different question and can make later rates appear healthier. Reconcile cohort totals to the underlying customer population before interpretation.
Separate lifecycle and calendar effects
A diagonal pattern across a cohort matrix often indicates a calendar event affecting several cohorts at once, such as an outage, holiday, or billing change. A vertical difference near the same age may indicate lifecycle behavior. Annotate releases and market events, then inspect both cohort-age and calendar-date views.
Compare customer mix across cohorts. A campaign may attract smaller customers, a new country, or a different use case. Stratify only with adequate counts and privacy safeguards. Treat observed differences as signals for explanation or testing. Cohort analysis improves comparability, but it does not by itself prove that a product change caused the outcome.
Evaluate a revised onboarding program
A business launched guided setup in April and sees higher activity among April and May signups than among customers acquired earlier.
- Assign accounts to their first paid month and preserve the onboarding version received at entry.
- Define age as completed thirty-day intervals after payment and measure active-team rate for each interval.
- Mask ages that recent cohorts have not completed and display eligible account counts with every rate.
- Compare plan, company size, acquisition channel, and geographic mix before and after the launch.
- Inspect calendar-date activity for outages, then propose a controlled rollout or matched comparison for the remaining difference.
Cohort analysis specification
Fill this template before building a cohort matrix or trend chart.
- Question and entity: decision, customer unit, qualifying event, and unique identifier.
- Assignment: cohort dimension, period boundaries, stable attribute rule, and returning-customer treatment.
- Age: period-zero definition, elapsed-time calculation, timezone, and maturity mask.
- Outcome: event or value, period versus cumulative logic, denominator, exclusions, and revisions.
- Interpretation checks: cohort size, mix, calendar annotations, source reconciliation, and causal limits.
Common mistakes
- Comparing a new cohort's partial third month with an older cohort's complete third month.
- Grouping customers by their current plan when the question concerns the plan they originally purchased.
- Attributing a cohort difference to onboarding while acquisition channel and company size also changed.
Try one
A mobile app reports week-four engagement for this week's signup cohort as zero. Diagnose the error and describe the corrected table.
The cohort has not had four weeks of exposure, so the value is unavailable rather than zero. A correct answer defines the signup event and week boundaries, masks immature cells, reports cohort and eligible counts, and compares equal completed ages. It should also specify the engagement event and check identity, timezone, acquisition mix, and calendar incidents before interpreting differences between cohorts.
Sources
- NIST exploratory data analysis handbookNIST methods for exploring data, checking assumptions, and revealing structure.
- NIST process and production control handbookNIST statistical methods for monitoring processes, time series, and change.