Causal discovery and inference in COVID-19: A multiscale longitudinal analysis using mixtures of DAGs

Real-world longitudinal clinical datasets present major challenges for causal discovery due to mixed variable types, temporal dependencies, treatment-selection mechanisms, missingness, and patient heterogeneity. We extended the Causal Inference over Mixtures (CIM) framework to accommodate mixed-type longitudinal data and applied it to 1,937 hospitalized patients from the Khorshid COVID Cohort. The analysis integrated admission characteristics, symptom trajectories, in-hospital treatments, laboratory measures, and short-term outcomes up to 12 weeks after discharge using mixed conditional-independence testing, bootstrap stability selection, graph-based interpretation, and follow-up outcome assessment. The recovered ancestral structures revealed clinically coherent symptom clusters at admission and persistent associations among several post-discharge symptoms, including cough, fatigue/weakness, shortness of breath, weight loss, and loss of smell. Treatment-related associations were interpreted cautiously because cumulative treatment exposure may reflect disease severity, hospitalization duration, physician decision-making, and other selection mechanisms. In particular, the observed associations involving ceftriaxone and interferon-β should be regarded as exploratory and require independent validation. Simulation studies demonstrated that CIM consistently outperformed PC, FCI, RFCI, and CCI, achieving the highest recovery accuracy and balanced accuracy while maintaining low false-positive rates and competitive computational efficiency, even in the presence of latent confounding and selection bias. These results support the feasibility and utility of the proposed framework for identifying robust structural patterns in complex longitudinal clinical data. Given the observational nature of the study, the single-cohort design, and the absence of external replication, the findings should be viewed primarily as hypothesis-generating rather than confirmatory evidence of causal effects.