Proceedings of the Annual Meeting of the Cognitive Science Society

Cognitive impairment is a major public health challenge in aging populations. Identifying stable, actionable drivers is crucial for targeted intervention. Causal discovery has been leveraged for its interpretability to explore these factors. However, causal modeling approaches that integrate long-span, multi-wave data to specifically address the heterogeneity of cognitive impairment remain underdeveloped, thus inherently failing to reveal domain-specific causal pathways from long-term, sparse longitudinal data. To address this, we introduce a framework that decomposes cognition into distinct domains and then applies time-slice causal discovery to sparse, multi-wave data. Applied to 8-year CHARLS data, the framework decomposed cognition into Mental State and Contextual Memory, identifying both domain-specific and shared causal drivers. These drivers were validated through robust predictive modeling, external generalization (CFPS), and interpretability analyses, demonstrating their utility. Our findings provide new insights for heterogeneous prevention and targeted early intervention.