The Prevention and Detection of Differential Measurement Biases in Analyses of Multiply Measured Outcomes

• Activity Number: 154
• Type: Invited
• Date/Time: Monday, August 5, 2013 : 10:30 AM to 12:20 PM
• Sponsor: International Chinese Statistical Association
• Author(s): Karen Bandeen-Roche*+ and Yuxin Zhu
• Companies: Johns Hopkins Bloomberg School of Public Health and Nanjing University
• Keywords: latent class ; multivariate ; aging ; item response ; latent variable ; epidemiology

Abstract

In numerous health fields an outcome of interest cannot be ascertained directly, but is assessed through multiple questionnaire items, symptoms/signs, or other measures. Differential measurement occurs when the accuracy or precision with which the target outcome is approximated by such measures varies by personal or environmental context. Our previous work has studied subsequent biases in estimating the relationship between risk factors and target outcomes (henceforth, "such biases") in a latent class regression framework. This work has yielded findings that motivate the work we now propose: to (i) study bounds on the magnitude of such biases as a function of characteristics of the conditional distribution of the multiple measures given the target outcome; and (ii) develop procedures by which researchers may identify predictor variables for which analysis must accommodate differential measurement if considerable such biases are to be avoided. Work primarily continues in the latent class regression vein. Findings are illustrated using epidemiological data. The work aims to improve practice in the analysis of outcomes assessed through multiple indirect measurements.