Abstract
The Cohort Study of Mobile Phone Use and Health (COSMOS) has repeatedly collected self-reported and operator-recorded data on mobile phone use. Assessing health effects using self-reported information is prone to measurement error, but operator data were available prospectively for only part of the study population and did not cover past mobile phone use. To optimize the available data and reduce bias, we evaluated different statistical approaches for constructing mobile phone exposure histories within COSMOS. We evaluated and compared the performance of 4 regression calibration (RC) methods (simple, direct, inverse, and generalized additive model for location, shape, and scale), complete-case analysis, and multiple imputation in a simulation study with a binary health outcome. We used self-reported and operator-recorded mobile phone call data collected at baseline (2007-2012) from participants in Denmark, Finland, the Netherlands, Sweden, and the United Kingdom. Parameter estimates obtained using simple, direct, and inverse RC methods were associated with less bias and lower mean squared error than those obtained with complete-case analysis or multiple imputation. We showed that RC methods resulted in more accurate estimation of the relationship between mobile phone use and health outcomes by combining self-reported data with objective operator-recorded data available for a subset of participants.