Professor and Head of the Division of Biostatistics
The University of California, San Francisco
Improving Predictions When Interest Focuses on the Identification of Unusual Random
Effects in Multi-Level Models
Predicted random effects are widely used to evaluate the performance of and rank clusters such as doctors and hospitals using longitudinal and multilevel data. “Best” predicted random effects using mixed models perform better than simply treating clusters as fixed effects and are optimal in the sense of minimum mean square error of prediction. However, predicted random effects are often used to identify extreme values such as poorly performing hospitals and the performance of best predicted values has not been systematically evaluated in this context. Using theoretical calculations, simulation and an example, I motivate new methods to predict extreme clusters and evaluate their performance. I show that the new methods can produce predicted values with smaller mean square or absolute error of prediction than standard best predicted values when interest focuses on extreme clusters. The methods are illustrated using data on length of stay in hospitals.
Dr. Charles McCulloch is a Professor and Head of the Division of Biostatistics at the University of California, San Francisco, as well as, Vice Chair of the Department of Epidemiology and Biostatistics. He is an expert on the development and use of statistical methods for longitudinal data analysis, mixed models and latent class models. Dr. McCulloch is the co-author of the Wiley texts Variance Components and Generalized, Linear, and Mixed Models (now in its second edition), the Springer text Regression Methods in Biostatistics (now in its second edition) and the author of the Institute of Mathematical Statistics monograph, Generalized Linear Mixed Models. He is a fellow of the American Statistical Association and an elected member of the International Statistical Institute. He has over 35 years of statistical consulting experience and has authored or co-authored over 500 peer-reviewed publications on a wide variety of topics including mathematical statistics, nephrology, toilet seats, women’s clothes, radioactive caterpillars, sugar-sweetened sodas, smoking in the movies, and arthritis.
Friday, November 15, 2019 11:30 a.m. BLOC 113