Primary Submission Category: Sensitivity Analysis
Model Selection for Causal Inference with Generalized Information Criteria
Authors: Yuchen Xiao, Stephen Walker,
Presenting Author: Yuchen Xiao*
Model selection plays a critical role in ensuring reliable and reproducible statistical inference in linear regression. The use of Generalized Information Criteria (GIC) to construct regression estimators corresponds to the OLS estimator under (ell_0) constraint. One key advantage of using information criteria lies in its ability to simultaneously integrate model estimation and model evaluation, which strikes a balanced parsimony between predictive performance and model complexity. Although traditionally viewed as computationally infeasible in high-dimensional settings due to its NP-hard nature, we demonstrate that applying information criteria for model selection and estimation can be efficiently solved in polynomial time using Hopfield network optimization. This rejuvenation of information criteria for model selection in high dimensions is applied to select covariates for inclusion into the propensity score to reduce the bias and improve the statistical efficiency of propensity score estimator. Extensive simulation results are presented to validate the efficiency and effectiveness of the proposed approach with the objective of rejuvenating the application of GIC for modeling the propensity score model and the outcome model. The simulation results show that the GIC includes all true confounders and predictors of outcome while maintaining minimum false positives.