Age-Period-Cohort Analysis

Description

Separates whether changes in a metric come from user tenure (age), calendar events (period) or sign-up cohort, despite their linear dependence.

In products, a user's tenure (age), the calendar date (period) and sign-up cohort are linked (period = cohort + age), so ordinary regression cannot separate them. Yet the distinction matters: falling engagement could be because users tire with tenure, because of a seasonal or competitive period effect, or because recent cohorts are lower quality. APC methods use Lexis diagrams for visual decomposition, constrained models, intrinsic estimators, or hierarchical APC models (cross-classified random effects) with explicit, stated identifying assumptions. Done carefully, APC prevents misattributing acquisition-mix effects to product changes.

Attributes

Loss type
Unmigrated
Stage
Post-launch
Effort
Days

Steps

  1. Build a dataset of outcomes by tenure interval and calendar period (e.g. weeks), with cohort implied.
  2. Visualise with Lexis surfaces or heat maps of rates by age and period.
  3. Choose an APC approach and state its identifying assumption (e.g. hierarchical APC with random period and cohort effects).
  4. Estimate age, period and cohort effects with uncertainty.
  5. Check robustness across alternative specifications.
  6. Interpret effects in light of known events (launches, seasonality, acquisition campaigns).

Sources

  • Yang & Land, Age-Period-Cohort Analysis: New Models, Methods, and Empirical Applications (2013)
  • Mason et al., Some methodological issues in cohort analysis of archival data (1973), American Sociological Review

Also known as

  • APC analysis
  • Lexis diagram analysis
  • Tenure-calendar-cohort decomposition

Categories

  • Epidemiology & Population Health