B0694
Title: A multivariate response model for data with correlation structures
Authors: Yingjuan Zhang - Durham University (United Kingdom) [presenting]
Jochen Einbeck - Durham University (United Kingdom)
Abstract: Multilevel data are common in scientific research. Established tools, such as the variance component model, are widely used for analyzing this type of data, with several functions like lmer() from the lme4 package and allvc() from the npmlreg package available in R. When multilevel data includes multiple response variables, which makes it a multilevel multivariate data, the conventional way of dealing with such data would be to fit separate two-level models each using one of the response variables, however, this approach ignores the correlation of different response variables. A novel approach is proposed for fitting two-level, multivariate response models where correlations between upper-level units are induced through a single, one-dimensional random effect. The random effect represents a latent variable parameterizing a line cutting through the space of predictors. The parameters in the proposed model will be estimated using the EM algorithm. Real data examples are given to illustrate the main applications of the model, including fitting a multivariate response model resulting in reduced standard errors, constructing league tables, and clustering with a specified degree of certainty.