Bayesian inference for comorbid disease state risks using marginal disease risks and correlation information from a separate source

Background. Public health interventions are increasingly being evaluated for their cost-effectiveness. Such interventions act `upstream' on the determinants of ill health and therefore may reduce the incidence of several diseases. In this case the risks of the separate diseases are likely to be correlated at the individual level, and considerable comorbidity may be present. An economic evaluation should take this comorbidity into account, but estimates of the risks and intervention effects may only be available separately for each disease. This paper proposes a method for combining marginal disease risks and treatment effects with correlation information from a separate source in order to estimate comorbid disease risks and treatment effects.

Method. A case study is presented based on a physical activity cost-effectiveness model. The correlation between the risk of coronary heart disease, stroke and diabetes is estimated from cross sectional data using a Bayesian multivariate probit model. This information is then combined with disease specific marginal baseline risks and intervention effects to give comorbid disease state risks. The expected numbers of QALYs gained through avoiding the comorbid states is estimated from disease specific utility data under a range of assumptions. Finally, the incremental benefit of physical activity is calculated under these utility assumptions. The difference in effectiveness of the intervention due to its impact on reducing or increasing the disease risk correlations is explored in a sensitivity analysis.

Results. If comorbidity is not taken into account, total benefit is overestimated compared with all scenarios in which the comorbidity is included in the model. The overestimation is greatest when physical activity is assumed to reduce disease state co-occurrence as well as overall disease incidence.