Competing Risks Joint Models for Spatio-Temporal Correlated Health Outcomes: Application to Estimate Infectious-Disease-Driven Mortality in Africa
64th ISI World Statistics Congress - Ottawa, Canada
Format: CPS Abstract
Keywords: "bayesian, "competing_risks", correlated outcomes, inla-spde, joint models
Session: CPS 77 - Statistics and mortality
Wednesday 19 July 8:30 a.m. - 9:40 a.m. (Canada/Eastern)
Making inferences about the competing risks of correlated health outcomes is critical for integrating interventions to improve child survival and promote cost-effective approaches. With the increasing availability of data on correlated health outcomes simultaneously collected in national-wide geo-referenced cross-sectional surveys, there is also a growing demand to advance statistical methodologies comparing the competing risks of these outcomes over time and space, and by quantifying the magnitude and nature of their associations in understanding important health indicators. The development of effective statistical models to analyze the spatial context in multivariate outcomes (i.e., “co-regionalisation” or “co-kriging”) from design-weighted health survey data is still emergent. This study aims to establish a Bayesian geostatistical modeling framework that: analyze multivariate correlated response data of varied metrics; examine the correlation; compare the competing risks across spatial clusters; and predict at unobserved locations. Starting from the theoretical construct of categorical (binary) to trivariate outcomes linked to survival data, different forms of joint models are defined and related to each other in a conceptual framework. The established framework is applied to data extracted from Tanzania demographic and health surveys and malaria indicator surveys on illnesses (malaria, diarrhoea, anaemia and respiratory infections) and survival (time-to-death) of children under five years to jointly estimate spatial and temporal patterns of infectious-disease-driven childhood mortality. The approach is extended to data from other African children. These models are implemented in R-INLA.