Continuing Education at JSM 2004: Generalized Linear Latent and Mixed Models, Analysis of Multivariate Survival Data, Introduction to Clinical Trials

Christine E. McLaren, 2004 CE Chair

The Biometrics Section is proud to co-sponsor three short courses during the annual meeting in Toronto. Anders Skrondal, Sophia Rabe-Hesketh, and Andrew Pickles will present a one-day course, "Generalized Linear Latent and Mixed Models." Skrondal, Head of Biostatistics at the Norwegian Institute of Public Health, Oslo, and Rabe-Hesketh, Professor in the area of Educational Statistics at the Graduate School of Education, University of California, Berkeley, are authors of the book “Generalized Latent Variable Modeling: Multilevel, Longitudinal and Structural Equation Models”. Pickles is Professor of Epidemiological & Social Statistics at School of Epidemiology and Health Science, The University of Manchester, UK. The presenters have extensive experience in teaching short-courses on longitudinal, multilevel and latent variable modeling worldwide. Generalized linear mixed (or multilevel) models (GLMMs) are useful for longitudinal data, cluster-randomized trials, surveys with cluster-sampling, genetic studies, meta-analysis and many other applications. The random coefficients in GLMMs are latent variables representing between-cluster variability and inducing within-cluster correlations. Latent variables are also often used to represent true values of variables measured with error, e.g. diet (continuous) or diagnosis (categorical). Measurement models specifying the relationship between measured and latent variables (factor, item response or latent class models) can form part of regression models, giving structural equation models (SEMs), such as covariate measurement error models. SEMS can also be used to model dependence between different processes, for instance the response of a clinical trial and drop-out. Taking a unified view is beneficial since developments for one model-type are often applicable to other model-types and the same software can often be used to estimate seemingly different models. This one-day course will be structured in three parts: (1) generalized linear mixed models, (2) measurement models and (3) structural equation models. The course will benefit statisticians and graduate students in statistics familiar with generalized linear models.

Philip Hougaard will present a one-day course entitled “Analysis of Multivariate Survival Data: An Introduction to Frailty Models”. Hougaard, author of the book “Analysis of Multivariate Survival Data”, is senior specialist at the pharmaceutical company Lundbeck. He has worked with multi-state models and frailty models for more than 20 years and is experienced in teaching courses that have included topics on introductory and advanced multivariate survival, frailty models, multi-state models and marginal models. Noted for his appealing teaching style, Hougaard’s presentations are mathematically solid, comprehensive, and practically useful. This course will start with a brief introduction to survival data. Frailty models for univariate data will be described in detail and examples of multivariate survival data will be put in a common frame. General probability mechanisms for creating dependence will be introduced and discussed. Measures of dependence will be described, starting with the correlation, but emphasizing measures that do not change with the marginal distribution. Discussion of shared frailty models for multivariate data, will include choice of model, the advantages and disadvantages of each model, the interpretation and the applications. A main example will be the survival of twins, but also other applications will be considered. Hougaard will consider a study aim of both finding the effect of covariates and evaluating the degree of dependence. The course will cover parametric as well as non-parametric models and proportional hazards models as well as accelerated failure time models. Software examples will be given showing how gamma shared frailty models can be fitted by means of Splus. Multi-state models will also be described in order to contrast and complement frailty models. Furthermore, frailty models for recurrent events data will be described in detail. Finally extensions to more complex frailty models will be described briefly, in order to illustrate when the shared frailty model is not adequate.

L. Jane Goldsmith will present a one-day course entitled “Introduction to Clinical Trials”. Goldsmith has served as Senior Statistical Consultant and Director of the Health Sciences Biostatistics Center at the University of Louisville School of Medicine. A popular instructor with a sense of humor, Goldsmith emphasizes that clinical trials are the heart of biomedical research. Her teaching experience and extensive collaborative research form the basis for this course. Trials for new drugs and other novel treatment developments comprise a large amount of research activity in the drug and medical equipment industry. Academic research efforts in medicine, nursing, and dentistry also often culminate in clinical trials as the gold standard for the determination of effective treatment methods for humans. Clinical research has an interesting history and nomenclature as well as standards for design, ethics, and reporting. Statistical methods are used in clinical trials for sampling and randomization strategies for treatment assignment, as well as for study design and analysis of data. Specialized statistical methods are sometimes used, but standard statistical theory can often be applied once the statistician gains the right background knowledge. This one day course serves to introduce statisticians to the praxis of clinical trials. It will benefit an experienced statistician who wants to learn about clinical trial research and also a new or student statistician who wishes to learn the role of biostatisticians in clinical research.