Modelling Space-time variation of cancer incidence data: a case study

Cancer incidence data are typically available as rates or counts for contiguous geographical regions and are collected over time. Recent methodological developments have moved in the direction of univariate space-time modeling of incidence data especially in a Bayesian context. Based on an example of data on cancer incidence collected between 1988 and 2005 in a specific area of France, this work describes an approach to analyze the space-time evolution of the disease taking into account also of possible non linear effects of other covariates. For this purpose, we consider Generalized Additive Mixed Models (GAMMs) with a Poisson response. The proposed method allows to incorporate a wide range of correlation structures. Besides one dimensional smooth functions accounting for non-linear effects of covariates, the space-time interaction can be modeled using scale invariant tensor product smooths, where the smoothness parameter is estimated and does not depend on the different scales of the covariate axes. Another possibility investigated to account for space-time dependency is to use varying coefficient models. In such case, to explore spatio-temporal patterns, analyzes focused on six time periods, each 3 years in length, between 1988 and 2005. For model implementations we use the R package mgcv