Estimating models of extreme behavior: A Monte Carlo comparison between SWAT and Quantile regression

We are interested in theoretical explanations of extreme behavior in social and management science situations. For example, in studying organizational performance we theorize that organzations achieve high levels of performance by employing innovative or unique behavioral characteristics. Methodologically, several approaches that provdie tools for modeling extreme behavior: substantively weighted analytical techniques (SWAR) and quantile regression. We evaluate both for their ability to accurately estimate models of extreme behavior when that behavior significantly differs from the average case. Since we attempt to evaluate statistical approaches in situations where standard axiomatic approaches fall short, our strategiy is to use simulation techniques where the underlying data-generating structure is knows and designed to have different underlying mathematical relationships between the middle and the two extremes. We also apply a Monte Carlo approach of repeated simulations to investigate the sampling characteristics of these apporoaches. Finally, we apply standard measure for evaulation of statistical estimators, mean square error, to examine both the bias and the relative efficiency of each approach. The experimental results demonstrate that quantile regression provides a more accurate and reliable estimation of extreme pheneomena.