Nonparametric empirical bayes estimation through deconvolution in probabilistic risk analysis

The advent of Markov Chain Monte Carlo (MCMC) simulations and availability of open source software made full hierarchical Bayesian analysis tractable and a popular choice in parameter estimation for Probabilistic Risk Assessment (PRA). However, despite its theoretical attractiveness, the hierarchical Bayesian analysis is not without its problems. Two of the most prominent practical difficulties in applying hierarchical Bayes analysis in practice to analyze source-to-source variability, for example, is its sensitivity to the selection of the first-stage prior and dependence of the rate of convergence on the selection of the first-stage prior. The first-stage prior in hierarchical Bayesian analysis is usually assumed having a parametric form, often conjugate to the likelihood function (aleatory model). To facilitate convergence for hierarchical Bayes models, the first-stage prior is frequently set through empirical Bayes estimate, which can use available historical data to find the parameters of the first-stage prior. This paper discusses a new nonparametric empirical Bayes estimation and compares it to several well-known parametric estimates such as method of moments and maximum likelihood. The new nonparametric technique exploits the prior predictive distribution as an integral equation and proceeds to solve it with respect to prior assuming the sampling data distribution and aleatory model are available. The discussion covers topics such as selection of aleatory model for future data, selection of regularization parameter, and density estimation for historical data. For comparison, the paper is using contrived as well as real-world data from utilities. Utilities data represent failure data with Poisson distribution used for aleatory model.