2019 Volume 34 Issue 1 Pages 3-17
The hierarchical age-period-cohort (HAPC) model, a class of mixed or hierarchical linear models, is now widely used in cohort analysis. This model treats the levels of the period and cohort factors as the groups to which each subject belongs and treats period and cohort effects as random effects. The HAPC model, however, has been criticized as being likely to produce subtle cohort effects even when definite effects are expected. This study elucidates why the HAPC model has such a serious deficiency. To overcome this problem, we revisit the Bayesian age-period-cohort (BAPC) model and focus on its method of taking the first-order differences in successive effect parameters. Like the HAPC model, the BAPC model can be interpreted as a mixed model, but it takes different approaches to the identification problem in cohort analysis. We demonstrate both models’ estimates for trends in the proportion of male college graduates in Japan, which seem to be dominantly affected by cohort. The results show that the HAPC model estimates near-zero cohort effects, whereas the BAPC model clearly detects positive cohort effects.