Reference Effect Measures for Quantifying, Comparing and Visualizing Variation from Random and Fixed Effects in Non-Normal Multilevel Models

• Activity Number: 522 - Contributed Poster Presentations: Biometrics Section
• Type: Contributed
• Date/Time: Wednesday, July 31, 2019 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Gary Grunwald and Thomas J Glorioso* and Michael Ho and Thomas M Maddox
• Companies: University of Colorado Anschutz Medical Campus and US Veterans Administration and US Veterans Administration and Washington University School of Medicine
• Keywords: Generalized Linear Mixed Model; Hierarchical model; Median Odds Ratio; Random effects

Abstract

Multilevel models for non-normal outcomes are widely used in medical and health sciences research. A common case is patients in hospitals. Easily interpretable methods to quantify, interpret and visualize random cluster variation (RCV) and compare it with other sources of variation are needed. We propose Reference Effect Measures (REM) to quantify and compare RCV to 1) individual subject and cluster covariate effects, and 2) variation from sets of covariates, e.g. all patient or all hospital covariates. REM is based on percentiles of random effect distributions, transformed to the effect scale. As an example, we used REM to show that for initiation of rhythm control for atrial fibrillation (AF) patients in the Veterans Affairs (VA), RCV across hospitals is substantially greater than that due to most individual patient factors, and explains at least as much variation in treatment initiation as do all patient factors combined. These results contrast with small RCV compared with patient factors for one-year mortality for AF patients. We also apply REM to overdispersed clustered counts and joint longitudinal/survival models. Results are easily visualized in forest or other plots.