Towards unique and unbiased causal effect estimation from data with hidden variables

摘要: Causal effect estimation from observational data is a crucial but challenging task. Currently, only limited number of data-driven causal methods are available. These either provide bound the treatment on outcome, or generate unique effect, making strong assumptions and having low efficiency. In this paper, we identify practical problem setting propose an approach to achieving unbiased effects with hidden variables. For approach, have developed theorems support discovery proper covariate sets for confounding adjustment (adjustment sets). Based theorems, two algorithms proposed finding variables obtain estimation. Experiments synthetic datasets generated using five benchmark Bayesian networks four real-world demonstrated efficiency effectiveness algorithms, indicating practicability identified potential in applications.