The interplay between actual and perceived air pollution, physical activity, and stress across mobility-based geographic contexts: A double machine learning approach.

The social ecological framework emphasizes the necessity of integrating individual characteristics, behavioral patterns, environmental factors, and health into a unified, holistic framework, positing that sustainable urban development cannot be achieved if any single facet is neglected. Yet, the interplay of actual and perceived air pollution, physical activity, and stress across mobility-based geographic contexts remains largely unexplored, which impedes precision urban planning. Previous studies have also been constrained by relying on retrospective questionnaires and static, residence-based measurements, often overlooking the uncertain geographic context problem (UGCoP), as well as high dimensionality and non-linearity in the data. To address these gaps, this study employed geographic ecological momentary assessment (GEMA) and developed a DAG-guided multi-relation double machine learning (DAG-MR-DML) framework. Our findings underscore the critical influence of the UGCoP and the necessity of context-based heterogeneity analysis. Specifically, we found that (1) the geographic context strongly moderated the relationships among these variables; (2) perceived air pollution significantly mediated the association between PM2.5 exposure and perceived stress overall and in the context of personal and mixed activities; and (3) physical activity may mitigate the adverse effects of air pollution on perceived stress overall and in work or study and leisure contexts. We emphasized the significance of geographic contexts and advocated for a "context-aware" smart urban air information and lifestyle system to mitigate perception errors and provide timely physical activity recommendations, thereby promoting the mental well-being of urban residents.