An Exploration of Regression-Based Data Mining Techniques Using Super Computation

Regression analysis is intended to be used when the researcher seeks to test a given hypothesis against a data set. Unfortunately, in many applications it is either not possible to specify a hypothesis, typically because the research is in a very early stage, or it is not desirable to form a hypothesis, typically because the number of potential explanatory variables is very large. In these cases, researchers have resorted either to overt data mining techniques such as stepwise regression, or covert data mining techniques such as running variations on regression models prior to running the final model (also known as “data peeking”). While data mining side-steps the need to form a hypothesis, it is highly susceptible to generating spurious results. This paper draws on the known properties of OLS estimators in the presence of omitted and extraneous variable models to propose a procedure for data mining that attempts to distinguish between parameter estimates that are significant due to an underlying structural relationship and those that are significant due to random chance.