CANMI: Causal Discovery under Nonstationary Missingness Mechanisms

ICLR 2026 Conference Submission14661 Authors

Keywords: Causality, Missingness mechanism, Nonstationarity

Abstract: Causal discovery from time series data is a typical and fundamental problem across various domains. In real-world scenarios, these data often have missing values occurring under different mechanisms, which limits the applicability of most existing approaches, especially when the missing values do not occur randomly due to the influence of other variables. This challenge is further exacerbated when missingness mechanisms also depend on nonstationarity in time series data. In this paper, we propose CANMI, a novel constraint-based approach designed for CAusal discovery under Nonstationary MIssingness mechanisms. Our proposed method can recover the causal structure using only observed data with different missingness mechanisms, including missing not at random (MNAR). Furthermore, we prove the identifiability of the direct causes of missingness and reveal a formula for recovering the data distribution from nonstationary data with missing values. Extensive experiments on both synthetic and real-world datasets demonstrated that our proposed model outperforms state-of-the-art approaches for causal discovery across various evaluation metrics even under substantial missingness. Our source codes are available at https://anonymous.4open.science/r/CANMI-0CDD.

Primary Area: causal reasoning

Submission Number: 14661