LASSO -- a promising alternative to AIC for model selection

Session: Improving Model Predictions Through Coupled System and Data Assimilation (2)

Song Qian, University of Toledo, [email protected]
Christine Mayer, University of Toledo, Lake Erie Center, [email protected]

Abstract

In the age of big data, developing a meaningful model is increasingly challenging.  Separating relevant variables from useless ones is always a time-consuming process associated with a high level of uncertainty.  The widely used various information criteria (e.g., AIC, BIC, DIC) are unlikely to improve the tedious nature of the task, nor reduce the uncertainty.  A promising alternative is the least absolute shrinkage and selection operator (LASSO), often known as the "computer-age" statistical method designed for "large-scale" inference.  Using data from a study on nearshore fish community response to shoreline characteristics, we introduce the use of LASSO for selecting important (relevant) predictors.  The presentation includes a short introduction of LASSO (including its relationship to AIC/DIC), the existing modeling effort of the nearshore fish community study, and the application of LASSO to the same data.  The presentation ends with a discussion on the practical and philosophical considerations when selecting variables for model development.