A1209
Title: Causal DAG identification for count data via Poisson-Thinning structural equation models
Authors: Penggang Gao - Kyoto University (China) [presenting]
Ming Cai - Graduate School of Informatics, Kyoto University (China)
Hisayuki Hara - Kyoto University (Japan)
Abstract: Count data arise frequently in epidemiology, economics, and social sciences, motivating the development of causal models that account for discrete distributions. Existing Poisson branching approaches encode causal structure via binomial thinning operators but face two major limitations: The presence of equivalence classes that hinder full identifiability, and the restriction of exogenous disturbances only to the Poisson distribution. To address these issues, a structural equation model based on the Poisson thinning operator is introduced. The model is proven fully identifiable under exogenous noise following Poisson, negative binomial, or zero-inflated Poisson distributions. The framework includes a parameter estimation procedure derived from the structural equations and a likelihood score-based identification procedure for orienting edges in candidate graphs. Extensive simulations with exogenous noise from these distributions show high identification precision in recovering causal structures, and high accuracy in parameter estimation under finite samples. In summary, the proposed framework advances models and identification theory for count data, providing practical methodology for causal discovery in observational studies.
