Assessing Causal Graph Uncertainty and Optimized Multivariate Discretization Strategy

This study investigates the complexities of constructing causal graphs within discretization constraints, focusing on the uncertainty of causal link presence resulting from discretizing continuous variables. Despite discretization’s significant influence on causal graph estimation, it may be necessary, prompting the introduction of a robust index to quantify uncertainty under various discretization strategies. In real-world scenarios, the approach uses an evaluation index tailored to align with estimation characteristics, providing a practical assessment of uncertainty in causal discovery. By examining simulation data and using the true positive mean of causal link presence, the study assesses the impact of discretization. The proposed index offers a realistic evaluation of uncertainty in real-world studies without known truth. The overarching goal is to improve the accuracy and reliability of causal graph discovery by systematically assessing the impact of discretization on Type I error rates. The study explores different discretizations and their effects on independence tests and inferred causal structures.