Monte Carlo Evaluation of the Linearity and Gaussian Error Assumptions for Lasso

• In Event: Advanced Statistical Methods for Complex Data

Abstract

Lasso is a popular machine learning algorithm capable of simultaneously fitting linear models and performing variable selection. The variable selection feature of lasso can be beneficial to educational researchers, particularly when data contains a large number of independent variables. Extensive research has shown that lasso frequently outperforms traditional regression methods (Hindman, 2015). However, gaps remain regarding lasso’s performance under certain data conditions. Specifically, whether lasso's performance is impacted under violations of the linearity or Gaussian error assumptions is unknown. In this study, Monte Carlo simulation was used to assess lasso's performance under these assumption violations. Results indicate that lasso is not robust to the linearity assumption, but lasso's ability to correctly perform variable selection is robust to violation of Gaussian-errors.

Author

• Christina Witta Gillespie, Valencia College