|
Abstract:
|
Mediation analysis is an important component of environmental epigenetic studies that aim at explaining how an environmental exposure affects health. If the environmental exposure is measured with error the validity of mediation analysis in this context can be severely undermined. We study the bias due to measurement error on a continuous or binary exposure, on direct and indirect causal effects estimated with generalized linear models. Our analytic results and simulation study demonstrate that naïve analyses, that ignore measurement error, underestimate the magnitude of direct effects, while the indirect effects could be biased either away or towards the null. We also show that exposure measurement error can lead to inflated Type I errors of tests for mediation. SIMEX and regression calibration approaches coupled with sensitivity analyses are proposed to mitigate bias due to measurement error in settings without gold standard validation samples. We apply the proposed methods to simulated data from a recent study investigating the effect of maternal smoking during pregnancy, potentially mediated by DNA-methylation, on birth weight.
|
