Statistical Methods for Multinomial Data with Multilevel Structures

Period of Performance: 08/15/2002 - 02/14/2003


Phase 1 STTR

Recipient Firm

Insightful Corporation
Seattle, WA 98109
Principal Investigator
Firm POC

Research Institution

University of Michigan
3003 South State Street
Ann Arbor, MI 48109
Institution POC


The ultimate objective of our research is the development of methodology and software for the analysis of multi-level binomial and multinomial data in longitudinal studies. Multi-level data is very common in psychological and biomedical studies. Our research will make fundamental contributions to these studies by evaluating, recommending, and developing robust and efficient algorithms for handling such data. We will focus on parametric approach: generalized linear mixed models, and semiparametric approaches: generalized estimating equation models and generalized additive mixed models. The algorithms will be implemented as an object-oriented software library in the S-Plus language and the approaches will be published in peer reviewed journals. Currently, there is no commercial software for handling multinomial data with multiple hierarchical structures. In the proposed work we will evaluate various algorithms and provide recommendations on when each algorithm should be used. A unified library of well-tested algorithms for performing these analyses will be available in Splus and ported to R. In addition, we will incorporate diagnostic techniques and graphical methods into the software, and we will develop a comprehensive case study guidebook using real problems. Generalized linear mixed effects models and generalized estimating equations models are widely used in psychology, social sciences, and other areas of scientific research. The computational algorithms we propose to investigate are aiming at categorical data, particularly, binomial and multinomial responses. The product will be implemented as an S-Plus library and ported to R package. The software will be available to the public and will attract consulting projects and short courses for longitudinal data analysis with S-Plus. The inclusion of a Java graphical user interface and guidebook will make these methods accessible to a wider audience of researchers.