New local method estimates causal effects without needing pretreatment or causal sufficiency assumptions

A team of researchers has introduced a novel local learning method for selecting covariates to estimate average causal effects without relying on two problematic assumptions: the pretreatment assumption (that covariates are unaffected by treatment or outcome) and causal sufficiency (that no unobserved confounders exist). Traditional methods depend on these assumptions or require global causal structure learning over all variables, which is computationally impractical for high-dimensional datasets. The new approach, detailed in a paper submitted to arXiv, first identifies a local boundary that is guaranteed to contain a valid adjustment set if one exists, then uses local identification procedures to efficiently search within that boundary. The authors prove the method is sound and complete.

Experiments on multiple synthetic datasets and two real-world applications demonstrate that the method achieves accurate causal effect estimation while substantially improving computational efficiency. This is particularly valuable for fields like epidemiology, social science, and machine learning where practitioners often face high-dimensional data and cannot verify the strict assumptions needed by standard causal inference techniques. By relaxing these assumptions and using a local, rather than global, learning strategy, the method offers a practical tool for unbiased causal analysis in realistic settings.