Size,complexity,and organization

Abstract

Bayesian networks and influence diagrams are powerful methods for modeling complex relationships between variables, but the usefulness of these networks is dependent on the way that an analyst arranges the model’s structure. Model misspecification is perhaps most likely to occur when the model starts to become large or the relationships complex. The current study was designed to empirically evaluate three qualities of influence diagrams that might be potential sources of confusion for model authors and consumers. We found that the size of the network was the greatest predictor of user confusion and error, followed by the complexity of the causal chains within it, followed by the length of those causal chains. Furthermore, we found that network size exacerbates the effects of the other factors. These data suggest that analysts seeking to minimize user error within Bayesian networks should pay special attention to the size of their networks.