Variable Selection for Decision Making.

Abstract

In decision making research, scientists collect a large number of variables that may be useful in deciding which action is best. Researchers might use a combination of theory, clinical expertise and statistical variable selection methods to choose betweenthese variables. Most variable selection techniques, however, were developed for use in a supervised learning setting where the goal is optimal prediction of the response. While variables selected by these methods may be useful in predicting the outcome variable, they may not affect the choice of action. This thesis discusses the necessary characteristics of variables that are useful for decision making. It presents multiple techniques designed specifically to select variables that aid in decision making. Simulation analysis is used to assess the proposed methods' ability to find good decision making variables and limit the selection of spurious variables. The methods are applied to data from a randomized controlled trial for the treatment of depression.