|
Abstract:
|
In many disciplines, it is important to understand how the effect of an exposure on an outcome is mediated through other variables. Unfortunately, potential mediators may be measured with error. In mediation analysis, regression coefficients obtained by ignoring measurement error in the mediator can be severely biased and thus induce bias in the estimation of causal direct and indirect effects. Using the regression calibration approach, we show how to adjust for measurement error in longitudinal studies with repeated measurements of the mediator. Rather than assuming normality for the random effects in the linear mixed effects calibration model, we correct for measurement error in the mediator allowing for flexibility in the distribution of subject-specific random effects. Simulation results indicate benefits in bias reduction in estimated direct and indirect effects compared to ignoring measurement error or assuming normal random effects. We apply the method to a study among children in which hemoglobin level mediates the effect of cerebral malaria on child cognition.
|
