Multivariate Spatio-temporal Modelling for Completing Cancer Registries and Forecasting Incidence

TL;DR

This paper introduces multivariate spatio-temporal shared component models to forecast cancer incidence using incomplete regional and national data, improving prediction accuracy for cancer registries.

Contribution

It presents a novel modeling approach that jointly analyzes mortality and incidence data to address data incompleteness and regional disparities in cancer forecasting.

Findings

Models accurately predict lung cancer incidence in England from 2001-2019.

Multivariate models outperform univariate approaches in forecasting accuracy.

Shared component models effectively handle incomplete and non-harmonized data series.

Abstract

Cancer data, particularly cancer incidence and mortality, are fundamental to understand the cancer burden, to set targets for cancer control and to evaluate the evolution of the implementation of a cancer control policy. However, the complexity of data collection, classification, validation and processing result in cancer incidence figures often lagging two to three years behind the calendar year. In response, national or regional population-based cancer registries (PBCRs) are increasingly interested in methods for forecasting cancer incidence. However, in many countries there is an additional difficulty in projecting cancer incidence as regional registries are usually not established in the same year and therefore cancer incidence data series between different regions of a country are not harmonised over time. This study addresses the challenge of forecasting cancer incidence with incomplete data at both regional and national levels. To achieve this, we propose the use of multivariate spatio-temporal shared component models that jointly model mortality data and available cancer incidence data. We evaluate the performance of these multivariate models using lung cancer incidence data and the corresponding number of deaths reported in England for the period 2001-2019. Model performance was assessed using different predictive measures to select the best model.