Age, Period, and Cohort Effects in Longevity Studies
Three dimensions of time
Age describes time since birth. An age-associated pattern may reflect biological change, accumulated exposures or changing circumstances; it does not establish a single ageing mechanism.
Period describes calendar time when an outcome is observed. Changes in healthcare, environmental conditions or measurement practices may affect several age groups during the same period, although their effects need not be equal.
Cohort usually describes birth year or a range of birth years. People born around the same time may share historical exposures, but members of a cohort are not interchangeable.
These distinctions concern the interpretation of human population data. They are not evidence that an intervention extends lifespan. See the methodological discussion by Bell and Jones.
A hypothetical comparison
Suppose researchers compare people turning 70 in 2020 with people turning 70 in 2000. Age is held constant, but both calendar period and birth cohort differ: these groups were born in 1950 and 1930 respectively. A difference in physical function cannot be attributed uniquely to cohort history or contemporary conditions.
Following the 1950 cohort from age 70 in 2020 to age 80 in 2030 changes both age and period. Comparing 50-year-olds and 70-year-olds in 2020 instead changes both age and cohort. These are illustrative comparisons, not reported study findings.
Why statistical adjustment cannot separate everything
For consistently measured time variables, age = calendar period − birth cohort. Knowing two determines the third. Consequently, unrestricted linear age, period and cohort effects cannot all be uniquely estimated from the same observations.
A larger sample can improve precision but cannot remove this mathematical dependency. A fitted model must introduce restrictions or other identifying assumptions. Its apparent separation of the three effects therefore depends on more than the observed data.
This does not make every temporal comparison useless. It means that a claim to have isolated an ageing effect needs an explanation of the assumptions that made that estimate possible. Bell and Jones discuss this identification problem.
Assess the assumptions behind the model
A model’s name does not guarantee that it has resolved identification. Simulation research examining hierarchical age–period–cohort models shows why their results require scrutiny of the data structure and modelling assumptions. See Bell and Jones’s simulation study.
When reading a study, ask:
- What outcome, ages, calendar years and birth cohorts were observed?
- Which constraints allowed the model to distinguish temporal effects?
- Why are those constraints plausible for this population and outcome?
- Do conclusions change under alternative defensible assumptions?
- Does the conclusion describe an association, or make a causal claim requiring additional evidence?
Keep the evidence level clear
Human temporal comparisons address population patterns. Animal experiments and laboratory studies address different questions and should be identified separately when used to discuss possible mechanisms. Neither an animal survival result nor a cellular measurement, by itself, establishes the explanation for a human age–period–cohort pattern.
For complementary design questions, see longitudinal versus cross-sectional biomarker studies and missing data and attrition in long-term ageing studies.