Multinomial Regression for Correlated Data Using the Bootstrap in R

• Activity Number: 84
• Type: Contributed
• Date/Time: Sunday, August 9, 2015 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistical Computing
• Author(s): Jennifer Thompson* and Timothy Girard and Pratik Pandharipande and E. Wesley Ely and Rameela Chandrasekhar
• Companies: Vanderbilt University and Vanderbilt University Medical Center and Vanderbilt University Medical Center and Vanderbilt University Medical Center and Vanderbilt University
• Keywords: multinomial regression ; bootstrap ; R ; correlated data ; clustered data

Abstract

Multinomial regression is used for circumstances where the outcome of interest is categorical and has no inherent order. In the case where data is correlated - for example, multiple records clustered by subject - ignoring this correlation can result in biased standard errors and invalid inferences. We present a clustered bootstrap method for multinomial regression in the case of multiple measurements per subject. This method estimates coefficients, odds ratios, predicted probabilities and all related confidence intervals in the presence of both main effects and interactions and of restricted cubic splines in covariates. We illustrate the method with an example from a population of critically ill patients which examines the association between a continuous exposure and the patient's mental status on the day following exposure measurement, and compare our results with the multgee package in R, which provides GEE methods for correlated nominal responses.