Comparative evaluation of time-to-event analysis methods for mobility limitation in older adults: A multi-cohort benchmarking study.

Time-to-event models are central to aging research, quantifying both whether and when an event occurs. However, guidance on selecting among traditional statistical and machine learning (ML) methods for prediction modeling remains limited, particularly for epidemiologic aging cohorts with wide assessment intervals. Mobility limitation, as a clinically important aging outcome with such complexities, provides a relevant test case for methodological benchmarking. We conducted a systematic benchmark comparison of 14 statistical and ML time-to-event methods using harmonized data from four longitudinal U.S. aging cohorts (N = 14,821). The outcome was time to incident mobility limitation, defined by self-reported walking difficulty or gait speed < 0.8 m/s. Models were evaluated using 4-fold cross-validation repeated 100 times within cohorts. Performance metrics captured complementary dimensions of predictive quality: discrimination (C-index), accuracy and calibration (Brier score, integrated log loss, calibration index), and computational efficiency (model training time). Prespecified decision rules synthesize results across metrics to identify preferable methods. Across pooled analyses, multiple methods demonstrated comparable discrimination, including penalized Cox models, tree-based ensembles, and boosting approaches. Oblique random survival forest achieved the most favorable overall trade-off, combining high discrimination (C-index 0.673), good calibration (Brier score 0.147), and moderate training time (9.2 min). When computational efficiency was prioritized, penalized regression methods may be preferable, with good discrimination and Brier score. These findings support random forest and penalized regression as strong defaults for time-to-event prediction modeling in gerontologic research and highlight the importance of joint evaluation of discrimination, calibration, and feasibility.