UW-Milwaukee Department of Mathematical Sciences presents,

Dr. Justine Shults;
Associate Professor of Biostatistics
Perelman School of Medicine, University of Pennsylvania

Friday, April 24, 2015
2:00pm in EMS E495

*Refreshments served at 1:30pm in E424A

Title:
Analysis of Discrete Longitudinal Data

Description/Abstract:
Longitudinal studies often consider discrete outcomes such as the occurrence of major depressive episode (yes/no) in a clinical trial to compare two treatments for depression. However, fewer maximum likelihood based approaches are available for the analysis of longitudinal discrete data than are available for continuous outcomes. Semi-parametric approaches such as generalized estimating equations (GEE) can be applied, but the major software implementations of GEE ignore the additional constraints that must be satisfied by the marginal means and correlations for discrete random variables. In this presentation, I describe a class of first-order antedependence (or Markov) models that induce decaying product correlation structures that are plausible for longitudinal studies. I describe the implementation of the decaying product structures for the semi-parametric quasi-least squares approach that is based on GEE. I contrast QLS with the first-order Markov maximum likelihood (MARK1ML) approach.  I also demonstrate how to use a generalization of the MARK1ML models to simulate correlated discrete data, which can be helpful when simulating data to compare methods or to assess power in the planning stages of a clinical trial.