Logistic regression models: practical induced prior specification

TL;DR

This paper explores how to specify priors in Bayesian logistic regression models to induce desired prior distributions on the probability parameter, addressing issues of unintentionally informative priors and demonstrating practical applications.

Contribution

It introduces alternative prior specifications for logistic regression coefficients that produce specific induced priors on the probability parameter, including uniform and Beta distributions.

Findings

Induced priors can significantly influence posterior estimates.

Proposed priors achieve targeted induced distributions on probabilities.

Methods are validated through simulations and real data applications.

Abstract

Bayesian inference for statistical models with a hierarchical structure is often characterized by specification of priors for parameters at different levels of the hierarchy. When higher level parameters are functions of the lower level parameters, specifying a prior on the lower level parameters leads to induced priors on the higher level parameters. However, what are deemed uninformative priors for lower level parameters can induce strikingly non-vague priors for higher level parameters. Depending on the sample size and specific model parameterization, these priors can then have unintended effects on the posterior distribution of the higher level parameters. Here we focus on Bayesian inference for the Bernoulli distribution parameter $\theta$ which is modeled as a function of covariates via a logistic regression, where the coefficients are the lower level parameters for which priors are specified. A specific area of interest and application is the modeling of survival probabilities in capture-recapture data and occupancy and detection probabilities in presence-absence data. In particular we propose alternative priors for the coefficients that yield specific induced priors for $\theta$. We address three induced prior cases. The simplest is when the induced prior for $\theta$ is Uniform(0,1). The second case is when the induced prior for $\theta$ is an arbitrary Beta($\alpha$, $\beta$) distribution. The third case is one where the intercept in the logistic model is to be treated distinct from the partial slope coefficients; e.g., $E[\theta]$ equals a specified value on (0,1) when all covariates equal 0. Simulation studies were carried out to evaluate performance of these priors and the methods were applied to a real presence/absence data set and occupancy modelling.