Abstract
Age-Period-Cohort (APC) models in gerontology can be used to understand time-varying events that account for variability in patient care and health outcome trajectories. Age effects have been thought of as the consequences of growing older; period effects as the consequences of general influences that vary through time or epoch; and cohort effects as the consequences of being born in a given year or birth cohort. APC models are often criticized for inability to independently estimate age, period, and cohort effects, since age = period + cohort presents a model identification problem. Bell and Jones (2014) have criticized the efficiency of current proposed solutions and recommended that an investigator performing APC analyses choose which two among the three APC effects are most interesting to study. We propose a novel method for APC analyses for including functions of all three effects. We model period and cohort along with an internal risk score based on age and other risk factors relevant to predicting the outcome of interest. We exclude period and cohort in the internal risk score model. The resulting risk index contains information about an age effect, but is not linearly dependent with period and cohort. A necessary condition is 0.05 ≤ |ρ| ≤ 0.95, where ρ is the correlation between age and the risk index. We illustrate these “risk-period-cohort” methods using individual level data in electronic health records over a 15-year period in a sample of 49,609 adults with type 2 diabetes with an outcome of time until an ASCVD event.