Approaches to the problem of nonidentifiability in the age-period-cohort models in the analysis of cancer mortality: a scoping review.

Aiming to detect age, period and cohort effects in cancer mortality, age-period-cohort models (APC) can be applied to distinguish these effects. The main difficulty with adjusting an APC model involving age, period and cohort factors is the linear relationship between them, leading to a condition known as the 'nonidentifiability problem'. Many methods have been developed by statisticians to solve it, but there is not a consensus. All these existing methods, with their advantages and disadvantages, create confusion when choosing which one of them should be implemented. In this context, the present scoping review intends not to show all methods developed to avoid the nonidentifiability problem on APC models but to show which of them are, in fact, applied in the literature, especially in the cancer mortality studies. A search strategy was made to identify evidence on MEDLINE (PubMed), Scopus, EMBASE, Science Direct and Web of Science. A total of 46 papers were analyzed. The main methods found were: Holford's method (n = 14; 30%), ntrinsic estimator (n = 10; 22%), Osmond & Gardner method n = 8; 17%), Carstensen (n = 6;13%), Bayesian approach (n = 6;13%) and others (n = 2; 5%). Even with their limitations, all methods have beneficial applications. However, the decision to use one or another method seemed to be more related to an observed geographic pattern.