Causal Inference in Longitudinal Studies Using Causal Bayesian Network with Latent Variables

Longitudinal studies have been broadly used in clinical research to investigate the associations between exposures or treatments and the outcome of the diseases, such as disease onset, subsequent morbidity, and mortality. However, few studies emphasize the causal relationships between observed variables and latent, time-varying confounders. The causal Bayesian network (CBN) shows promise in handling multiple causes and effects. This paper presents an extension of the Bayesian Network for Latent Variable (BN-LV) framework that quantify the causal effects of the latent variables in CBNs by imposing various constraints for the identification of latent structures and the structure learning algorithms. The proposed model employs unit-level causal inference methods that can learn instance-specific causal mechanisms. The proposed model also provides "near" causality inference from the observational data, eliminating causal edges from the traditional BN-LVs framework. The method was validated using a case study: Temporal Associations Between Daytime Napping and Sleep Outcomes. The results showed the quantification for the average causal effects of napping on nocturnal sleep measures and the construction of a learned causal graph involving latent variables.