Stat991 Spring 1998

Methods for over-dispersed data

Overview

This seminar course will provide students with exposure to statistical methods designed to be used when analyzing ``over-dispersed data''. Over-dispersion is rather a catch-all term, but the analyst can expect to deal with such data when the data structure is of one of the following types;

• Repeated measures
• Longitudinal
• Hierarchical

Topics covered will include

• Classical Random Effects Models for continuous data/REML estimation.
• Parametric over-dispersion models for discrete data, Beta/Binomial, Poisson/Gamma.
• Semi-parametric mixtures.
• Conditional inference.
• Generalized Estimating Equations.
• Bayesian viewpoints

Approach: I will present this course using a thirds approach: In class time will be equally divided between the the three components ``Theory'', ``Algorithms'' and ``Examples''.

Students with their own datasets are encouraged to bring them. The datasets would be up for discussion in the second half of the course.

Expectations: active in class participation, coding up algorithms, analyzing example data together with an in-class presentation of a research paper from the literature.

Pre-requisites: PhD level mathematical statistics and linear models.

Computing: the computing environment will be S-Plus. This will be used for both the data analysis and algorithm development. Students must be willing to spend time learning the computing environment.

Implementing procedures will give students experience with the following algorithms; IRLS, VDM, EM and MCMC.

Books. See the WEB site for a complete list, but we will start with the required one:

Variance Components. Searle, Casella and McCulloch. Wiley. (1992)