Evaluating the Impact of Multicollinearity on Regression

In empirical regression analysis, the existence of high multicollinearity suggests that predictors may provide redundant information and cause a reduction in statistical power. Meanwhile, dropping correlated variables may result in mis-specified models with biased parameters. Unlike previous studies that are focused on guidelines to diagnose and manage multicollinearity, this paper proposes a practical Monte-Carlo simulation method to determine whether to keep a correlated variable for an empirical model when other factors such as sample size and over-all fitting accuracy could mitigate the effect of multicollinearity.