An Exploratory Stability Selection (ESS) Framework for Robust Predictor Discovery: An Application to Physical...

Identifying biological and clinical signals that consistently predict physical resilience, defined as one's ability to maintain or regain function following a health stressor, is essential for advancing precision approaches to aging and recovery. High-dimensional datasets hold tremendous promise but pose analytic challenges due to correlation, distributed signals, instability, and sensitivity to analytic choices. The complexity of these data requires strategies that prioritize transparency and stability in variable selection. We present a resampling-based statistical framework, the Exploratory Stability Selection (ESS) framework, designed for hypothesis-generating predictor discovery. ESS is an ensemble-style variable selection technique that integrates multiple resampling strategies and sparsity levels, enabling exploration of robustness and context-dependence across diverse data perturbations. We demonstrate the utilization of the ESS framework with a data example using the PRIME-KNEE study, which examines physical resilience in older adults undergoing elective total knee arthroplasty. ESS analyses were applied to clinical-only, plasma biomarker-only, and combined predictor sets to evaluate the stability and competitiveness of candidate variables associated with the probability of having a highly resilient recovery trajectory for pain interference. The data example highlights how ESS distinguishes highly stable predictors from context-dependent signals whose selection varies with predictor competition and analytic configuration. ESS retains configuration-level results and summarizes stability metrics across configurations to provide insight into the subsequent prioritization and validation of candidate predictors. This framework is well-suited for hypothesis-generating variable selection problems common to exploratory resilience research and other aging-related applications that involve complex, multi-domain predictor sets.