This paper outlines a new class of shrinkage priors for Bayesian isotonic regression modeling a binary outcome against a predictor, where the probability of the outcome is assumed to be monotonically non-decreasing with the predictor. The predictor is categorized into a large number of groups, and the set of differences between outcome probabilities in consecutive categories is equipped with a multivariate prior having support over the set of simplexes. The Dirichlet distribution, which can be derived from a normalized cumulative sum of gamma-distributed random variables, is a natural choice of prior, but using mathematical and simulation-based arguments, we show that the resulting posterior can be numerically unstable, even under simple data configurations. We propose an alternative prior motivated by horseshoe-type shrinkage that is numerically more stable. We show that this horseshoe-based prior is not subject to the numerical instability seen in the Dirichlet/gamma-based prior and that the posterior can estimate the underlying true curve more efficiently than the Dirichlet distribution. We demonstrate the use of this prior in a model predicting the occurrence of radiation-induced lung toxicity in lung cancer patients as a function of dose delivered to normal lung tissue.
Biostatistics | Probability | Statistical Methodology | Statistical Models
Boonstra, Philip S.; Owen, Daniel R.; and Kang, Jian, "Shrinkage Priors for Isotonic Probability Vectors and Binary Data Modeling" (January 2020). The University of Michigan Department of Biostatistics Working Paper Series. Working Paper 128.