Detection of Attributive Covariates for Heteroscedasticity in Cross-Sectional or Longitudinal Regression Analysis

• Activity Number: 482
• Type: Contributed
• Date/Time: Wednesday, August 7, 2013 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Qian Zhou*+ and Peter X.K. Song and Mary E Thompson
• Companies: Simon Fraser University and University of Michigan and University of Waterloo, Canada
• Keywords: heteroscedasticity ; linear mixed-effects models ; linear regression models ; longitudinal data ; random effects

Abstract

This talk focuses on the development of a new statistical procedure to detect responsible covariates for heteroscedasticity in regression analysis. We propose covariate-specific statistics based on information ratios formed as a contrast between the model-based and sandwich variance estimators. A two-step diagnostic procedure is established, first to test for constant error variances, and if the homoscedasticity is rejected, then to identify attributive covariates. This proposed method enables us to systematically examine heteroscedasticity in either cross-sectional data analysis or longitudinal data analysis. We show that in linear mixed-effects model for longitudinal data, the proposed diagnostic tool provides a powerful statistical approach to identify covariates attributive to random effects or subject-specific covariate effects, which are critical to modelling heteroscedastic variance structures across subjects. The performance of the proposed method is assessed via simulation studies and is illustrated through a data analysis in which it is of interest to detect attributive covariates for subject-specific effects in the application of a linear mixed-effects model.