Using the Least Absolute Shrinkage and Selection Operator for Selection and to Assuage Overfitting: A Method Long Ignored in Educational Research

• In Event: 49.061 - Advanced Modeling Strategies for Longitudinal, Multilevel, and/or Survey Data

Abstract

Ordinary least squares and stepwise selection with small or moderate samples are widespread in behavioral science research. However, these methods are well-known to over-fit under such conditions such that R-squared and regression coefficients are inflated while standard errors and p-values are deflated, ultimately reducing the generalizability of models. More optimal methods for selecting predictors and estimating regression coefficients such as regularization methods (e.g., LASSO) have existed for decades and are widely available in mainstream software; yet, educational and psychological researchers have yet to take advantage of these methods. This proposal discusses the merits of LASSO for automatic selection and/or regression applications with moderate samples and will demonstrate the desirable properties through both an applied example and a small simulation study.

Author

• Daniel McNeish, University of Maryland