Local Covariate Selection for Causal Effect Estimation Without Strong Assumptions

A recent preprint on arXiv (2605.21548) introduces a novel local learning technique aimed at covariate selection for nonparametric causal effect estimation, circumventing the assumptions of pretreatment and causal sufficiency. Traditional methods generally depend on global causal structure learning involving all variables or impose stringent assumptions like causal sufficiency, which presumes no latent confounders among observed variables, or the pretreatment assumption, restricting covariates to those not influenced by treatment or outcome. Such conditions are often impractical, and global learning can be computationally intensive in high-dimensional contexts. The new approach defines a local boundary that includes at least one valid adjustment set for causal effect identification and establishes local identification methods for efficient covariate selection, tackling significant challenges in unbiased total causal effect estimation without unrealistic assumptions.