Age-period-cohort analysis with a constant-relative-variation constraint for an apportionment of period and cohort slopes

• Published in: PloS one 2019-12, Vol.14 (12), p.e0226678
• Main Authors: Su, Shih-Yung, Lee, Wen-Chung
• Format: Article
• Language: English
• Subjects: Age; Age Factors; Analysis; Bias; Biology and Life Sciences; Cancer; Cohort analysis; Cohort Studies; Computer Simulation; Constraint modelling; Curvature; Disease; Epidemiology; Health care; Humans; Incidence; Male; Medical technology; Medicine and Health Sciences; Monte Carlo Method; Monte Carlo methods; Monte Carlo simulation; Mortality; Parameter estimation; People and Places; Physical sciences; Preventive medicine; Prostate; Prostate cancer; Prostatic Neoplasms - epidemiology; Public health; Research and Analysis Methods; Research Design - standards; Slopes; Stochasticity; Survival Rate; Taiwan; Variation; White People - statistics & numerical data
• ISSN: 1932-6203; 1932-6203
• DOI: 10.1371/journal.pone.0226678

Age-period-cohort analysis of incidence and/or mortality data has received much attention in the literature. To circumvent the non-identifiability problem inherent in the age-period-cohort model, additional constraints are necessary on the parameters estimates. We propose setting the constraint to reflect the different nature of the three temporal variables: age, period, and birth cohort. There are two assumptions in our method. Recognizing age effects to be deterministic (first assumption), we do not explicitly incorporate the age parameters into constraint. For the stochastic period and cohort effects, we set a constant-relative-variation constraint on their trends (second assumption). The constant-relative-variation constraint dictates that between two stochastic effects, one with a larger curvature gets a larger (absolute) slope, and one with zero curvature gets no slope. We conducted Monte-Carlo simulations to examine the statistical properties of the proposed method and analyzed the data of prostate cancer incidence for whites from 1973-2012 to illustrate the methodology. A driver for the period and/or cohort effect may be lacking in some populations. In that case, the CRV method automatically produces an unbiased age effect and no period and/or cohort effect, thereby addressing the situation properly. However, the method proposed in this paper is not a general purpose model and will produce biased results in many other real-life data scenarios. It is only useful in situations when the age effects are deterministic and dominant, and the period and cohort effects are stochastic and minor.