A spatiotemporal back-calculation approach to estimate cancer incidence measures.

Abstract

Incidence rates are an important population-level disease risk measure. Cancer incidence data in the United States, which are collected by disease registries, have been spatiotemporally sparse. Back-calculation methods can yield incidence estimates for a spatial domain by solving a convolution equation that relates mortality to incidence through survival estimates. We propose a novel back-calculation approach that uses spatiotemporal age-period-cohort (APC) modeling to estimate incidence for spatial units within a region. The method is applied to state-specific lung cancer data in the United States for males ages 30 to 84 in years 1975-2004. SEER data are used to model cancer progression from incidence to mortality along three timescales (APC) and across four regions. Using mortality data from National Vital Statistics System, incidence is back-calculated for unique birth cohorts in 49 states. A conditionally autoregressive model with cubic splines for temporal effects is used to smooth back-calculated estimates. Bayesian estimates of model parameters are obtained using integrated nested Laplace approximation. Results show varying time trends in lung cancer risk across states, which may quantify effects of state policies.