This talk is sponsored in part by the Pacific Institute of Mathematical Sciences (PIMS).
The study of the influence of environmental and socio-economic risk factors on health often requires the use of multiple data sources, such as small-area aggregated data, routinely collected administrative data as well as data coming from surveys and cohort studies. Some data sources will contain detailed information on a small sample of individuals; in contrast, others will have a limited number of variables for the whole population. Building models that can link various sources of data presents a number of challenges, for example dealing with model mis-specification and ecological bias, or having to impute missing confounders.
Bayesian graphical models provide a coherent way to connect a series of local sub-models based on different datasets into a global unified analysis. In particular, hierarchically related regression (HRR) models can be built to carry out simultaneous regression on aggregated and individual level data, control ecological bias and study the relative contribution of contextual and individual level risk factors. In a similar manner, Bayesian graphical models can be used to fit a common regression model to a combination of data sets with different sets of covariates, correctly propagating information between the model components.
We will discuss the application of Bayesian graphical models in different contexts. In the first one, the aim is to use these to help separate the effects of place of residence and personal circumstances on two contrasting health outcomes: perceived "limiting long-term illness" and hospitalization for cardiovascular disease. The second example is concerned with a case study on the effect of water disinfection by-products on low birth weight in the U.K, where the analysis uses a combination of register, survey and small area data.
Joint work with Nicky Best, Chris Jackson and Jassy Molitor
