Correction for Bias Due to Measurement Error in Assessing Association Between Baseline and Slope in Longitudinal Studies

• Activity Number: 87
• Type: Contributed
• Date/Time: Sunday, July 31, 2011 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Anny Hui Xiang*+ and Jun Xiao and Thomas A. Buchanan and Steven J. Jacobsen
• Companies: Kaiser Permanente Southern California and Forest Laboratories, Inc. and University of Southern California and Kaiser Permanente Southern California
• Keywords: Bias correction formula ; Measurement error ; baseline and longitudinal slope of change ; random effect mixed models

Abstract

Regressing observed change against observed baseline value can produce a very biased estimate of the true relationship when the observed values contain random measurement error. Blomqvist proposed a simple formula to correct for such bias under linear regression framework for one follow-up visit. We conducted simulation studies to investigate the bias under random effect mixed modeling framework where the paradigm involves assessing the relationship between baseline and the slope of change with multiple follow-up visits. We derived a very simple empirical formula in the format similar to the Blomqvist formula. Our results showed that bias is approximately linearly associated with the inverse of number of follow-ups. Our formula is reduced to the Blomqvist formula under one follow-up. Validity of our formula was also evaluated through simulation studies. Bias corrected standard errors and 95% confidence intervals can be estimated using bootstrapping method. A real application example was provided to demonstrate its use and highlight the importance of correcting measurement error in assessing the true relationship between baseline and change over time.