Discretization, Simulation, and the Value of Information

Abstract

Summary Most decision analyses include continuous uncertainties (e.g., oil in place, oil price, or porosity). Analysts are frequently concerned with how to best structure, compute, and communicate decision models under these circumstances. While decision trees are well suited for discrete random variables with a few possibilities, they become unmanageable for a large number of outcomes. To address this concern, analysts frequently use discrete approximations such as Swanson's Mean. Previous work has quantified how well differing discretization methods match the moments (e.g., the mean and variance) of the underlying continuous distribution. More specifically, previous work has not included the decision context in which the discretizations are used. In this paper, we begin to address this gap by comparing different discretizations within the context of an information-gathering decision problem. We find that the best discretization is highly dependent on the decision context, which is difficult to specify in advance. In addition, we contrast the use of discrete approximations to Monte Carlo simulation.