Reproducibility materials for cross-national forecasting of disability and death from longitudinal ageing trajectories

It supports reproduction of conventional machine-learning benchmarks, longitudinal neural-network comparisons, Qwen3-8B QLoRA experiments, temporal ablations, training-scale analyses, missingness and cohort-fingerprint audits, calibration analyses, decision-curve analyses, subgroup evaluations, and participant-clustered paired bootstrap comparisons. Participant-level source data, derived microdata, individual-level predictions, fitted model objects, and model weights are not included because the original cohort data-use agreements do not permit redistribution by the authors.

Researchers must obtain the relevant source data directly from the respective cohort custodians and comply with their registration and data-use requirements. The included reconstruction documentation specifies the required data products, variable harmonisation procedures, participant-level partitioning rules, fixed random seeds, model configurations, and integrity checks. These materials are intended to support transparent methodological evaluation and independent reproduction of the reported analyses.

They are not intended for clinical use or individual-level risk assessment.