Counterfactual dose-response learning of how extracurricular physical activity buffers adolescent smartphone addiction and anxiety

Overview Problematic smartphone use and internalizing anxiety co-occur at high rates during adolescence. Quantifying how extracurricular physical activity buffers this pathway remains difficult, because daily movement durations inhabit a closed 24-hour simplex and both exposures share severe psychosocial confounding. Standard moderated regression and Euclidean neural dose-response estimators treat physical activity volume as an unconstrained scalar, ignoring whether added exercise displaces recreational screen time or nocturnal sleep. Here we formulate PhyBuff-Net, a compositional counterfactual dose-response network coupled with an agentic isotemporal feasibility harness, evaluated across N = 3,600 de-identified adolescents from junior high school (n = 1,600) and senior high school (n = 2,000) cohorts. PhyBuff-Net maps 24-hour movement compositions into orthonormal isometric log-ratio coordinates, decouples joint selection bias via double-continuous Hilbert-Schmidt and contrastive balancing, estimates smooth buffering surfaces through Bernstein splines, and pairs entropic mirror-ascent donor planning with deterministic circadian boundary verification and local-support self-reflection. On the held-out test split (N_test = 720, five seeds), PhyBuff-Net achieves a counterfactual surface coefficient of determination of 0.9564 ± 0.0043 (95% confidence interval [0.9462, 0.9636]), outperforming contrastive balancing representation learning (0.8987 ± 0.0030, Holm-adjusted p = 5.34 × 10^-34) while reducing counterfactual policy regret from 4.5558 to 0.8916 and eliminating circadian feasibility violations (0.00%). Contents This record contains 16 files. The table lists every data table in the archive with its row count. FileRowsSize (KB)data/cohort_participants.csv3,600799data/cohort_reference_surface.csv3,6003209data/cohort_screening_attrition.csv60data/derived/ablation_table.csv504data/derived/ablation_table_summary.csv102data/derived/cohort_descriptive_stats.csv303data/derived/dose_response_curves.csv13211data/derived/error_failure_modes.csv141data/derived/main_table.csv453data/derived/main_table_summary.csv91data/derived/sensitivity_sweep.csv242data/derived/subgroup_consistency.csv635data/predictions/sensitivity_predictions.csv132,30016054data/predictions/test_predictions_ablation.csv36,0005189data/predictions/test_predictions_main.csv32,4003813 Reuse Released under cc-by-4.0. If you use these data, please cite this record together with the associated article.