Stepwise Model Selection

Stepwise model selection refers broadly to the use of an iterative procedure for the identification of predictors for inclusion in a model, whereby the utility of predictors is evaluated at each step and the results of this evaluation are used to inform the inclusion of predictors at subsequent steps of the process. Predictive modeling can encompass a number of research goals, such as maximizing accuracy, whittling down predictors to minimize redundancy, testing specific predictive hypotheses, or comparing competing predictive models. In well-researched areas with strong theoretical guidance, the choice of predictors to test in a model may be relatively clear. The focus of such research may be on testing the incremental value of specific predictors or testing competing predictive models. However, in domains lacking strong theoretical or empirical bases, researchers may lack sufficient guidance to choose from myriad candidate predictors relevant to an outcome. Often, this indicates further exploratory research is needed to refine the topic; in other situations, it may still be valuable to develop predictive models on theoretical grounds. For instance, the prediction of suicide completion is an area with limited empirical or theoretical foundation, yet with substantial value in identifying risk factors via predictive modeling.

Stepwise model selection is a procedure for identifying a subset of predictors that maximize explanatory power while culling redundancy in candidate variables. While the specific applications vary, the general [Page 1619]procedure of stepwise model selection is to run a recursive algorithm with an evaluative stage between each repetition of a model to determine the inclusion or exclusion of candidate predictors. This algorithm is run until a prespecified stopping point, resulting in a reduced subset of predictors that are selected into the final model. Stepwise model selection has been criticized for potential bias, misapplications, and its role as a potential data dredging technique, yet proper interpretive guidelines and cross-validation methods can address some of the limitations and applications that have raised such concerns. This entry emphasizes how and under what conditions stepwise techniques can be applied, along with their limitations as stand-alone methods.