Geert Molenberghs is Professor of Biostatistics at UHasselt and KU Leuven. He received a degree in mathematics (1988) and a Ph.D. in biostatistics (1993) from UAntwerpen. He published on surrogate markers in clinical trials, and categorical, longitudinal, and missing data. He was Editor for Applied Statistics, Biometrics, and Biostatistics, and is currently Executive Editor of Biometrics. He was President of the International Biometric Society. He is Fellow of the American Statistical Association, received the Guy Medal in Bronze from the Royal Statistical Society, and held visiting positions at Harvard. He is founding director of the Center for Statistics at UHasselt and of the Interuniversity Institute for Biostatistics and statistical Bioinformatics (UHasselt and KU Leuven). He received research funding from FWO, IWT, the EU (FP7), U.S. NIH, U.S. NSF, UHasselt, KU Leuven, ECDC, and EMA. He is member of the Belgian Royal Academy of Medicine. He has been active (as advisor, researcher, and communicator) in the SARS-CoV-2 pandemic response.
Talk: Handling Negative Correlation and/or Over/Underdisperson in Gaussian and Non-Gaussian Hierarchical Data
Abstract: The occurrence and interpretation of negative variance components in the context of linear mixed models is well understood at this point, even though the issue is surrounded by subtle issues for estimation and testing (Verbeke and Molenberghs 2003, Molenberghs and Verbeke 2007). Broadly, negative variance components often point to negative within-cluster correlation. It is even possible to give such linear mixed models a meaningful hierarchical interpretation (Molenberghs and Verbeke 2011). Matters are more complicated when the outcomes are non-Gaussian, either in the context of the generalized linear mixed model, or extensions thereof that allow for flexible modeling of both within-unit correlation as well as overdispersion (Molenberghs et al. 2010). An additional complication is that, in practice, not only negative variance components due to negative correlation, but also underdispersion instead of overdispersion can occur, sometimes even jointly. With focus on both continuous and count data, we describe how models can be made sufficiently flexible and, in a number of cases, interpreted hierarchically (Luyts et al. 2019).