2014/2015 QMeHSS Workshop (1/9/2015)

Dear Colleagues,

We hope that everyone had a nice holiday and a happy new year. This upcoming Friday, January 9th, Dr. Robert Gibbons and Dr. Don Hedeker will be leading our next Workshop on Quantitative Methods in Education, Health, and the Social Sciences. The title of the presentation is Recent Advances in Generalized Mixed-effects Regression Models. As with workshops that we have held in the past, the seminar will be held from 11am – 12:30pm in the NORC conference room 344. NORC is located at 1155 E. 60th Street. We hope to see you there.


Recent Advances in Generalized Mixed-effects Regression Models

Robert Gibbons & Don Hedeker, University of Chicago


Over the past 35 years, mixed-effects regression models have become an integral part of statistical science and have been applied in the analysis of clustered and/or longitudinal data in virtually every scientific field.  Despite their widespread use, some of the assumptions inherent in mixed-effects models in general and non-linear mixed-effects models in particular have limited their use in certain fields.  In particular, mixed-effects models provide unit-specific (e.g. subject-specific) inferences when we may also be interested in deriving population average inferences.  In addition, the assumption of joint normality of the random effect distributions and the assumption of independence of the random effects and fixed effects in the model have limited their adoption, particularly in the field of economics where “fixed-effects” models have a long history of use.  In this presentation we explore these limitations and provide some solutions.  In particular, we present a general approach for marginalization of the regression parameters in generalized linear mixed-models for any correlated categorical outcome, which support population averaged inferences.   In addition, we explore methods for deriving within-unit (e.g. within-subject) inferences which are robust with respect to the assumption of normality of the random-effect distributions and correlation of the random and fixed-effects, and have increased statistical power over traditional fixed-effects models.  We also show that these “hybrid” mixed models can accommodate treatment heterogeneity which can lead to bias and poor coverage in traditional fixed-effects models.