Causal discovery algorithms typically assume complete data, but real-world datasets often contain missing values. A new arXiv preprint addresses this gap by extending identifiability results from complete-data settings to missing-data settings. The authors establish general sufficient conditions under which causal graphs remain identifiable—meaning the underlying causal structure can be uniquely recovered despite incomplete observations.
The work builds on existing causal discovery theory, which has matured for complete data, and relaxes assumptions to accommodate missingness. Rather than focusing on a specific algorithm, the paper provides a theoretical framework that can apply broadly. This is a step toward making causal discovery more robust in practical applications where missing data are common.
The abstract does not detail the exact conditions or compare them with prior work, but the contribution is clear: a general set of sufficient conditions for identifiability in missing-data settings. This opens the door for future algorithms to be designed with these guarantees in mind, though the paper itself does not present new methods or empirical validation.