When: Friday, October 27th, from 4:00 PM to 5:00 PM
Where: ENGR 040
Who: Professor Ming-Hui Chen, Department of Statistics, University of Connecticut
Abstract: In this paper, we introduce a novel application of the linear mixed-effects model (LMM) repurposed for statistical monitoring. We develop an efficient EM algorithm to handle rapid estimation, especially in scenarios with a large number of parameters. The proposed approach provides two modes of prediction, namely marginal and conditional, derived from the LMM. Additionally, we incorporate predictive test statistics for anomaly detection in future observations. These predictive test statistics are accompanied by predictive p-values, enabling hypothesis testing to identify anomalous values in the observed data. To control the familywise error rate, we employ an F-approximation technique that mitigates inflation associated with the chi-squared test. The performance of our methodology is evaluated through extensive simulation studies, which encompass an assessment of familywise error rate, statistical power, and false discovery rates. The simulation results demonstrate that the proposed methods provide consistent statistical guarantees aligned with our theoretical findings. Furthermore, we apply the developed methodology to real-world gas consumption data, illustrating its practical utility in monitoring and detecting anomalies. This is a joint work with Daeyoung Lim, Nalini Ravishanker, Haiwei Zhou, and Max Sun (UConn Statistics); and Mark Bolduc, Brian McKeon, and Stanley Nolan (UConn Facilities Operations).