Bayesian models for evidence-based medicine using PyMC3

Dec. 6, 2017 from 3:30 to 4:20 p.m. — Physics/Astronomy Auditorium, room A102

Chris Fonnesbeck
Assistant professor of biostatistics and director of VUMC Database Analysis Core, Vanderbilt University

Abstract

Bayesian hierarchical models have established themselves as one of the most effective means for obtaining inference from structured data, such as geographical information, multi-center studies, and meta-analysis. Moreover, with the advent of easy-to-use probabilistic programming tools, this approach is available to a wider audience.

I will introduce the application of Bayesian modeling to quantitative meta-analysis using PyMC3, a statistical library for the Python programming language. With increased public scrutiny on the reliability of biomedical findings, the use of systematic reviews for synthesizing scientific evidence for the effectiveness of interventions has never been more important.

Meta-analysis aims to generalize estimates of treatment efficacy (and other parameters) across a set of research outcomes, while accounting for confounding issues related to the heterogeneity among individual studies. Properly applied, meta-analysis can provide more complete evidence regarding the effectiveness of candidate biomedical interventions by increasing the sample size with which to draw conclusions, and by overcoming biases related to the idiosyncrasies of individual studies.

A photo of Chris Fonnesbeck

Bayesian hierarchical methods are particularly useful in this setting, allowing analyses to be based on more realistic assumptions and more fully account for uncertainty related to estimates of underlying effects due to heterogeneity. I will show how meta-analytic models can be specified and fit in PyMC3 using its concise, easy-to-read syntax, and how the results of these models can be readily summarized and checked by users.

Bio

Chris Fonnesbeck is an assistant professor in the Department of Biostatistics at the Vanderbilt University School of Medicine. He specializes in computational statistics, Bayesian methods, meta-analysis, and applied decision analysis. He originally hails from Vancouver, BC, and received his Ph.D. from the University of Georgia.