In population health, the most commonly used measures quantifying the importance of a modifiable risk factor, or the effect of a health intervention, are population attributable risk (PAR) and impact fractions (IF). These metrics are crucial as they enable comparison of competing health interventions and directly inform public policy.However, the standard analytic approaches to estimating these metrics have a number of methodological shortcomings that to be addressed. These include: including incorrect modelling for continuously distributed risk factors, inadequate consideration of the subtleties of direct and indirect causal pathways to disease and overly simplistic use of “plug-in” formulae. Better estimation tools need to be developed as the current estimates may result in sub-optimal utilization of resources and inaccurate public health advice.
Here, I propose a systematic program of research based on Bayesian network models that proposes solutions to all these problems. The networks will initially be developed by combining observational evidence from INTERSTROKE, a large case-control dataset, designed to assess the contributions of known (and potential) risk factors for stroke to global stroke prevalence, with expert medical opinion. Specialized IF and PAR calculations will be designed, respecting causal risk factor pathways leading to disease and appropriately model continuous risk factors. Later, I will incorporate longitudinal data from a number of large cohort studies which study the progression of a variety of risk factors and several diseases, to extend the network to several diseases and better inform inferred causal pathways. These methods will enable accurate predictions for the effect of health interventions for a variety of diseases both nationally and globally. In addition, the constructed Bayesian network will elucidate the causal pathways by which certain risk factors lead to disease and is likely to be of substantial interest in itself.
