Comparison Of SimulationExtrapolation, Regression Calibration And Instrumental Variable In Reducing The Bias Due To Measurement Error
Abstract
The inability to accurately measure important variables complicates the statistical analysis of data when the purpose of inference is for the model defined in terms of the unobserved variables. In such situation, fitting the model directly to the observed data induces the bias estimates. For reducing the bias, some methods have been proposed such as Simulation-Extrapolation, Regression Calibration and Instrumen1al Variable. In this paper, the effects of measurement error on estimation)n linear and logistic regression will be demonstrated and the methods for bias rcd.uction will be compared. via simulation. The measurement error of predictors reduces the parameter estimate i.e. problem of attenuation. The corre1a1ion in predictors or the correlation in errors influences the magnitude and the direction of the bias. In general, Simulation-Extrapolation, Regression Calibration and Instrumental Variable can reduce the bias except for the structure of complex error in simple linear regression lnstrumental Variable tends more relatively stable on estimation with the increasing of measurement error compared with SimulationExtrapolation or Regression Cahbration estimation.