Investigating omitted variable bias in regression parameter estimation: A genetic algorithm approach

Abstract

Bias in regression estimates resulting from the omission of a correlated relevant variable is a well-known phenomenon. In this study, we apply a genetic algorithm to estimate the missing variable and, using that estimated variable, demonstrate that significant bias in regression estimates can be substantially corrected with relatively high confidence in effective models. Our interest is restricted to the case of a missing binary indicator variable and the analytical properties of bias and MSE dominance of the resulting dependent error generated vector process. These findings are compared to prior results for the independent error proxy process. Simulations are run for medium sample sizes and the method is shown to produce substantial reduction in estimation bias and often renders useful estimates of the missing vector. Limited simulations for the continuous variable case are reported and indicate some potential for the method and future research.