Maximum Likelihood-Based Regression with a Continuous Exposure Variable Assessed in Pools and Subject to Measurement and Processing Errors

• Activity Number: 625 - Environmental Epidemiology and Spatial Statistics
• Type: Contributed
• Date/Time: Thursday, August 3, 2017 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Dane Van Domelen* and Emily Mitchell and Amita Manatunga and Robert H Lyles and Enrique F Schisterman
• Companies: Rollins School of Public Health, Emory University and Agency for Healthcare Research and Quality and Emory University and Rollins School of Public Health, Emory University and Eunice Kennedy Shriver National Institute of Child Health and Human Development
• Keywords: pooling ; measurement error ; maximum likelihood

Abstract

Analyzed appropriately, a study design in which a biomarker is measured in pooled samples from multiple individuals can yield valid estimates of individual-level regression parameters, while often reducing costs and/or improving efficiency. We consider maximum likelihood estimation for linear and logistic regression, with a continuous covariate measured in pools and subject to measurement and/or processing error. We assume (1) a linear model with homoscedastic normal errors for the biomarker given other covariates; (2) normal additive measurement errors and, in pooled measurements, normal additive processing errors irrespective of pool size; and (3) that the two error types are independent. When both error types are present, a hybrid design with multiple different pool sizes is sufficient to consistently estimate regression coefficients. However, a small number of replicate individual measurements substantially improves stability. Using motivating data from the Collaborative Perinatal Project, we apply the proposed approach to assess whether monocyte chemotactic protein 1 is associated with log-odds of spontaneous abortion, controlling for race and smoking status.