The geographical distribution of disease risk is known to vary over time and the space-time (ST) analysis of disease maps is now a frequent focus of analysis with Epidemiology or Public Health. One important aspect of the risk profile that can change over time is the existence of underlying (latent) components of risk that can yield common disease outcomes. Conventional risk models usually examine the spatial structure by modeling the mean level of risk. This can lead to quite sophisticated Bayesian hierarchical models. These models can account for clustering of risk and smooth out noise from disease data. However they are not designed to find out about underlying latent components of risk. Our proposal has four components that address latent structure for ST analysis of disease risk: The development of spatio-temporal Poisson mixture models. It is planned to extend spatial mixture models into the spatio-temporal disease mapping context. The primary concern will be to extend the spatial Poisson mixture hidden component model to the spatio-temporal domain. This extension will allow the use of spatially- and temporally- structured weights with a random number of latent components. In this way we expect to be able to produce a flexible modeling strategy for spatio-temporal disease incidence data. Development and evaluation of a mean mixture model approach to latent structure. It is planned to change the approach to examine the use of mean mixtures in a ST context. In this approach, we will focus on temporally-varying underlying components with spatial structure confined to the weights that are available for each region. This separation leads to interesting issues concerning continuity of the components and the appropriate design of prior distributions, including the use of multivariate CAR prior distributions for weights. The development of Dirichlet process mixture models for the spatio-temporal domain. It is proposed that Dirichlet process models for univariate geo-referenced disease incidence data be extended for the situation where spatio-temporal data is observed. The extension will be focused on the use of time-referenced basis functions and the idea of latent spatio-temporal risk profiles will be considered. Comparative evaluation and software development there is a need to make a comparison of the above methods in application to real and simulated data scenarios. This will be achieved by application of the methods to credible temporally-varying simulated risk component scenarios as well as readily available real ST datasets for childhood outcomes (asthma, birth anomalies and birth weight).There is also a need for flexible software to be made available that can allow researchers and public health workers to be able to use advanced modeling approaches, with their ability to flexibly build appropriate descriptions of the observed disease data.
This aim will be achieved by the development of software within the R and WinBUGS programming environment.
Underlying structure in disease map evolution may be hidden but its estimation could be potentially useful in predicting future disease outcomes. Our methods address the estimation of these components and can help to establish hitherto unseen etiological factors or effects that are important.
Showing the most recent 10 out of 12 publications
