Extended GEE Methods with Multi-Layer Structured Correlation

Abstract

I will discuss my recent work on Correlated Data Library for Splus. The library is based on the GEE method, and it is expected to be released with the next Splus version 6.2. The focus of this talk is GEE methods for high-dimensional hierarchical data when a correct specification of the correlation matrix is desired. I will also touch on some bases of future R&D in utilizing GEE methods such as mixed models and informative dropout. Our goal is to develop GEE methods for heterogeneous and unbalanced correlated data. This includes the structures for working correlation matrix and variance components. We have implemented GEE0, GEE1, and GEE2 based on multi-layer correlation structures. The approach uses the concept of unbalanced block design for a correlation matrix. There is no limit on the dimension of the correlation matrix or the number of structures, e.g. arbitrary dimensional nested design. The approach is also applicable to intra and inter-cluster heterogeneity. The software library for Splus 6.1 Windows version is available but I don't plan to run a demo.