Statistical Inference After Variable Selection via Penalized Regression: Focusing on Variables Predicting Belonging to School

• In Event: Matching and Selection in Observational Data

Abstract

Despite the popularity of LASSO for variable selection and prediction, statistical inference after variable selection was a challenging task, as LASSO produces biased estimates. Among the recent progress in statistical inference with model selection, this study employed the post-selection inference method by Lee et al. (2016) in analyzing Programme for International Student Assessment (PISA) 2015, social science large-scale data. Notably, this study repeated 100 times of model selection and inference to consider the inconsistency issue of LASSO with respect to variable selection. Among 431 variables explored via LASSO, 14 of them were identified as important to predict students’ belonging to school in every iteration. With post-selection inference, 13 of them were also statistically significant in 75% of the iterations.

Authors

• Jin Eun Yoo, Korea National University of Education

• Minjeong Rho, Korea National University of Education