level: research
estimating total causal effects from observational data often requires selecting the right set of covariates to adjust for. existing methods usually depend on learning the full causal graph among all variables, which is hard in high-dimensional settings. they also commonly assume causal sufficiency, meaning no unobserved confounders, or the pretreatment assumption, where covariates must not be affected by treatment or outcome. these assumptions are frequently violated in real-world data.
the paper introduces a local learning approach that sidesteps both the pretreatment and causal sufficiency assumptions. it first identifies a local boundary around the treatment and outcome that is guaranteed to contain a valid adjustment set if one exists. then, it applies local identification procedures to select covariates from this boundary. this avoids the need to learn the entire causal structure, making the method computationally feasible even with many variables.
the method works for nonparametric causal effect estimation and does not require knowledge of the full causal graph. by focusing only on variables in the local boundary, it reduces the search space and relaxes common but restrictive assumptions. this makes it more practical for applied researchers who face unmeasured confounders or variables that are affected by treatment. the approach is shown to correctly identify valid adjustment sets under weaker conditions than prior work.
why it matters: it enables more reliable causal effect estimation in real-world data science tasks where unobserved confounders or post-treatment variables are common, without needing full causal graph discovery.