The Journal of Applied Microeconometrics, vol.4, no.1, pp.31-64, 2024 (Peer-Reviewed Journal)
We address the classical errors-in-variables (EIV) problem in multivariate linear regression with N dependent variables where each left-hand-side variable is a function of a common predictor X subject to measurement error. Our contribution consists in employing the remaining N −1 regressions as extra information to obtain a filtered version of the mismeasured series X. We test the performance of our approach using simulations whereby we control for different cases like low vs. high R2 models, small vs. large sample or small vs. large measurement error variances. The results suggest that the multivariate-Compact Genetic Algorithm (mCGA) approach yields
estimates with lower mean-square-errors (MSEs). The MSEs are decreasing as the number of dependent variables increases. When there is no measurement error, our method gives results similar to those that would have been obtained by ordinary least-squares.