Logo: to the web site of Uppsala University

The thesis examines the robustness of the Information Component (IC) disproportionality measure. A model used for screening combinations in pharmacovigilance data bases. The focus lies on investigating structural uncertainty in one component of the IC model. This is of interest since as of now this component is assumed to be constant, which could cause the model to overestimate its true precision and ultimately cause faulty screening. To uncover structural uncertainty, a stratified dataset by year is considered. To estimate the structural uncertainty, the residual variance of a generalized additive model, which is fitted on relative reporting rates, is utilized. This estimated structural uncertainty is then incorporated into the IC in two different models. First, in a hierarchical Bayesian model, a prior distribution is included for the previously constant component. The second model follows a parametric bootstrap approach, creating a bootstrap sample of the component in question on which the distribution of the IC is estimated. 

The main finding of this study is that 14\% of combinations exhibit a substantially increased variance in the models with structural uncertainty in comparison to the current IC model. From these combinations, a few screening results were altered due to this discrepancy. Further results indicate complex relationships between multiple parameters of the model, which complicates the interpretation of the impact that the structural uncertainty has.It can be concluded that the results demonstrate that the IC model could be improved by including structural uncertainty. Nevertheless, for the majority of combinations, the current model accurately estimates the disproportionality.