IFAS Statistics

Applied Generalized Linear Mixed Models

Course Description

Traditional statistical methods largely assume that data are normally distributed. However, not all data are normally distributed. Generalized linear models are simply linear models that have been extended for the analysis of non-normal data with only fixed effects. Generalized linear mixed models are designed to analyze data that are not normally distributed and have both fixed and random effects. This course begins by motivating the move from linear models to generalized linear models and includes a discusssion of the comparison of the analysis of data that have been transformed so that the assumption of normality is more nearly met and the analysis of data using generalized linear models. Logistic and Poisson regression will be discussed in the context of generalized linear models. Then after a quick review of the difference in fixed and random effects, generalized linear mixed models will be presented. All generalized linear mixed models have an underlying covariance structure associated with the response variable. This covariance of the responses is partitioned into the covariance matrix G associated with the modeled random effects and the covariance matrix R associated with the errors. How to determine an approrpiate covariance structure for a model is both challenging and important. The various covariance structures and how to choose from among them when analyzing data will be reviewed. The challenges that arise with generalized linear mixed models that are not present with linear mixed models will be discussed, including the differences in marginal and conditional models (and why you should care), prediction on the data scale versus the scale of the analysis, and assessing the fit of the model. The use of spatial covariance functions and radial smoothing to account for the covariance structure will be presented. The ability to have multiple response variables with possibly different distributions is another aspect of GLIMMIX that will be explored. Students will have the opportunity to practice the analysis of generalized linear mixed models. Extensive use will be made of real-data sets throughout the course.

Who should come

This workshop is targeted toward those who are familiar with PROC MIXED and who want to learn how to extend this to use mixed models when the data are non-normal. After this workshop, participants should be able to analyze non-normal data arising from studies with mixed effects.

Instructor

Dr. Linda J. Young, Professor in UF’s Department of Statistics, has consulted with researchers on the faculties of Oklahoma State University, University of Nebraska, and University of Florida. Her research interests are in sampling and modeling of ecological and environmental data.

Registration

To register for a workshop please contact Allison Pipkin at apipkin@stat.ufl.edu or 352.392.1946. Cash or check (payable to University of Florida) is accepted; or Department Fund Transfer (E2R). Click here to register for the this workshop online.

Fee

The cost of the workshop is $500 - accept cash or check (payable to University of Florida); or Department Fund Transfer (E2R).

Dates for workshop

The workshop will take place on the following dates

Date	Location	Time
11, March 2010	Conference room 426 of McCarty Hall, Building C	1 day workshop (8am - 5pm)