A1044
Title: An EM algorithm for fitting matrix-variate Student's-t distributions on interval-censored and missing data
Authors: Victor Hugo Lachos Davila - University of Connecticut (United States) [presenting]
Abstract: Matrix-variate normal distributions are powerful tools for modeling three-way data structures frequently encountered in longitudinal studies and multidimensional spatiotemporal analyses. However, in practical applications, such datasets often exhibit incomplete information, including censored values reported as being above or below detection limits and missing observations. Furthermore, deviations from normality, such as skewness and heavy tails, introduce new challenges. An efficient EM-type algorithm is proposed for maximum likelihood estimation under interval-censored and/or missing data, utilizing the matrix-variate Student's-t distribution framework, which offers increased robustness to outliers and heavy-tailed observations. The algorithm provides closed-form expressions based on the truncated moments of multivariate Student's-t distributions, which can be efficiently computed using existing software tools. Simulation studies highlight the limitations of matrix-variate normal distributions when applied to non-normal data, while demonstrating the increased robustness of the matrix-variate Student's-t distribution in such settings. Finally, the practical utility of the proposed approach is illustrated through two real-world case studies involving water quality monitoring.
