Projecting the future burden of cancer: Bayesian age–period–cohort analysis with integrated nested Laplace approximations

Abstract

The projection of age-stratified cancer incidence and mortality rates is of great interest due to demographic changes, but also therapeutical and diagnostic developments. Bayesian age-period-cohort (APC) models are well suited for the analysis of such data, but are not yet used in routine practice of epidemiologists. Reasons may include that Bayesian APC models have been criticized to produce too wide prediction intervals. Furthermore, the fitting of Bayesian APC models is usually done using Markov chain Monte Carlo (MCMC), which introduces complex convergence concerns and may be subject to additional technical problems. In this paper we address both concerns, developing efficient MCMC-free software for routine use in epidemiological applications. We apply Bayesian APC models to annual lung cancer data for females in five different countries, previously analyzed in the literature. To assess the predictive quality, we omit the observations from the last 10 years and compare the projections with the actual observed data based on the absolute error and the continuous ranked probability score. Further, we assess calibration of the one-step-ahead predictive distributions. In our application, the probabilistic forecasts obtained by the Bayesian APC model are well calibrated and not too wide. A comparison to projections obtained by a generalized Lee-Carter model is also given. The methodology is implemented in the user-friendly R-package BAPC using integrated nested Laplace approximations.

Bayesian Computing with INLA: A Review

“…The generality of R-INLA comes with a prize of complexity for the user, hence a simplified interface for a restricted set of models can be useful to improve accessibility for a specific target audience or provide additional tools that are mainly relevant for these models. Examples of such projects, are AnimalINLA (Holand et al, 2013), ShrinkBayes (Van De Wiel et al, 2013a,b, 2014Riebler et al, 2014), meta4diag , BAPC (Riebler and Held, 2016), diseasemapping and geostatp (Brown, 2015), and Bivand et al (2015). Similarly, the excursions package for calculating joint exceedance probabilities in GMRFs Lindgren, 2015, 2016) includes an interface to analyse LGMs estimated by R-INLA.…”

Probabilistic forecasting in infectious disease epidemiology: the 13th Armitage lecture

“…The sign of z indicates if the observations are overdispersed/underdispersed relative to the predictions (+∕− sign of z) [24,42]. A (two-sided) P-value can be computed to quantify the evidence for miscalibration.…”

Calibration tests for multivariate Gaussian forecasts

“…A positive value of z s indicates underdispersed forecasts with prediction intervals too narrow on average, or biased forecasts whose mean parameters are wrongly predicted. A negative value corresponds to overdispersed forecasts whose prediction intervals are too wide; see Riebler and Held [35] for an application in cancer prediction. This feature may provide clues how to further improve the forecasting model.…”