5/6/2016

Post date: May 11, 2016 2:46:48 PM

Time: 5/6/2016, 1:30-2:50pm

Place: Geno Cafe (Room E3609)

Title: Bayesian Nested Partially-Latent Class Models for Multivariate Binary Data; Estimating Disease Etiology

Speaker: Zhenke Wu, Department of Biostatistics, JHU

[abstract] Clinicians routinely use measurements to differentially diagnose a patient’s unknown disease etiology and then choose a treatment from among those available. More often than not, the differential diagnosis is a qualitative process based on judgement and experience. As clinical measurements become more precise and complex and as the number of possible known etiologies grows, such qualitative processes are less likely to be optimal. This talk presents Bayesian hierarchical models developed to estimate from multiple sources of data the probability distribution of unknown disease status and trajectory for a population and for its individuals. The motivating application is the Pneumonia Etiology Research for Child Health (PERCH) study whose goal is to infer the distribution of pneumonia-causing pathogens from multiple peripheral measurements with variable precision. Our models are related to the latent class model but with partially known class membership. In a first model, we assume each child’s data is a draw from a mixture model for which each component represents one pathogen. Conditioned upon the unknown etiology, measurements are assumed independent of one another. In a second model, conditional dependence is induced by nesting latent subclasses within each etiology class. We use stick-breaking priors on the subclass weights to estimate the population and individual etiologic distributions. Finally, I will mention multiple recent extensions to allow for regression of the etiology distribution on covariates and to accommodate the possibility of etiologies involving multiple pathogens. We demonstrate the approach and specific models with simulated and PERCH data analysis.

Research reported in this presentation was partially funded through a Patient-Centered Outcomes Research Institute (PCORI) Award (ME-1408-20318) and a Gates Foundation grant (48968).