[Some methodological aspects of age-period-cohort models. Application to cancer mortality trends]

Age-period-cohort models are usually used in descriptive epidemiological studies to analyze time trends in incidence or mortality. The exact linear relationship between the three effects of these models has the effect of making the parameters of the full model impossible to estimate, which is called non-identifiability. In these notes two of the most frequently used methods to analyze age-period-cohort models will be explained. One is based on penalty functions and the other on estimable functions (drift and curvatures or deviation from linearity). Both methods will be illustrated with two examples in which temporal trends of breast and lung cancer mortality in women from Catalonia in Spain will be studied. These examples show how the methods based on penalty functions tend to attribute the trend exclusively to a cohort effect. Consequently, the use of methods based on estimate functions is recommended.