Bayesian Benefit Risk Analysis

Abstract: The process of approving and assessing new drugs is often quite complicated, mainly due to the fact that multiple criteria need to be considered. A standard way to proceed is with benefit risk analysis, often under the Bayesian paradigm to account for uncertainty and combine data with expert judgement, which is operationalised via multi-criteria decision analysis (MCDA) scores. The procedure is based on a suitable model to accommodate key features of the data, which are typically of mixed type and potentially depended, with factor models providing a standard choice. The contribution of this paper is threefold: first, we extend the family of existing structured factor models. Second, we provide a framework for choosing between them, which combines fit and out-of-sample predictive performance. Third, we present a sequential estimation framework that can offer multiple benefits: (i) it allows us to efficiently re-estimate MCDA scores of different drugs each time new data become available, thus getting an idea on potential fluctuations in differences between them, (ii) it can provide information on potential early stopping in cases of evident conclusions, thus reducing unnecessary further exposure to undesirable treatments; (iii) it can potentially allow to assign treatment groups dynamically based on research objectives. A drawback of sequential estimation is the increased computational time, but this can be mitigated by efficient sequential Monte Carlo schemes which we tailor in this paper to the context of Bayesian benefit risk analysis. The developed methodology is illustrated on real data on Type II diabetes patients who were administered Metformin (MET), Rosiglitazone (RSG) and a combination of the two (AVM).