Abstract
Cohort Analysis is a method of separating age, period and cohort effects from time-series social survey data classified by age and period. The method is known to be useful in the study of social change. Many authors have pointed out, however, that cohort analysis has an identification problem and that the three effects cannot be uniquely separated without some prior information. In order to overcome the difficulty, this paper describes Bayesian cohort models in which prior information is assumed that the parameters of the effects change gradually and Akaike's Bayesian information criterion, ABIC, is used to select the best model. Four Bayesian cohort models are presented: L-APC and N-APC are ordinary three-effect cohort models for a qualitative and a quantitative response variable, respectively; L-[AP]PC and N-[AP]PC models contain the age-by-period interaction effect instead of the age effect. These models enable us to analyze a data set arranged not only in a standard cohort table but also in a general cohort table. They can also handle a data set with some unobserved values. The SSM (Research on Social Stratification and Social Mobility in Japan) data are analyzed to illustrate the usage of the proposed method.
