A SIMEX-Based Approach for Measurement Error Depending on Several Covariates
Measurement error (ME) is a challenge in nutritional epidemiology. Simulation-Extrapolation (SIMEX) methods have been developed for generalized linear, nonparametric, and generalized partially linear regression models. However, existing approaches generally assume that the error has mean zero, is independent of the true value, and, when covariate-dependent, depends on only one variable (SIMEX-CC). Data from the Generation XXI birth-cohort at 13 years-old suggest that these assumptions may not hold. Disagreement between reported nutrient intake and biomarkers was associated with socioeconomic status, obesity, and body-image perceptions, while energy misreporting was associated with obesity and body-image perceptions. Moreover, ME was correlated with the true value. These findings motivate extending SIMEX-CC to a non-parametric extrapolation and/or to allow ME to depend on several covariates. Simulation studies and applications to data from the Generation XXI birth-cohort are presented.
Keywords: measurement error SIMEX generalized linear regression