Discretization Precision and Assessment Error

Continuous probability distributions are often discretized by assigning a weight to each of several percentiles (e.g., the 10th, 50th, and 90th percentiles). Previous work has analyzed the accuracy of various discretization methods. In practice, however, the assessed percentiles may not be precise. In this paper, we compare the performance of several discretization methods when the probability assessments are subject to error. Our results indicate that one should still strive to use the best discretization method even in the face of assessment error. This is particularly true if one is trying to preserve the variance and higher moments of the continuous distribution.