Answers are grounded in this paper and cite their sources.
Tree boosting accommodates high-dimensional predictors but typically treats repeated measurements as independent, overlooking within-subject dependence. Existing mixed-effects tree methods alternate between fitting a tree learner and predicting random effects. With a sufficiently flexible learner, this procedure can absorb subject-specific variation into the fixed-effect component and shrink the estimated random-effects covariance. We propose mixed-effects extreme gradient boosting (MEXGB) for continuous longitudinal outcomes. The method combines a relaxed lasso with second-order boosted trees, uses cross-fitted fixed-effect predictions to update the random effects and variance components, and selects model complexity using out-of-fold prediction error. We derive an oracle inequality for candidate selection, establish consistency of variance-component estimation under consistency of the fixed-effect estimator, and obtain prediction-error lower bounds for methods that ignore subject specific effects, misspecified linear mixed models, and tree-only fixed effects. MEXGB had the lowest mean prediction error across 13 simulation scenarios with up to 2,000 predictors and ten competing methods. In the six main scenarios, it reduced mean squared error by 13--83% relative to mixed-effects gradient boosting and by 42--95% relative to extreme gradient boosting (XGBoost). In addition, MEXGB estimated the fixed effect more accurately than mixed-effects gradient boosting in every replicate. Across three applications, MEXGB was the most accurate method on one dataset and the most accurate mixed-effects method on a microbiome dataset with 1,879 predictors. The method is implemented in the R package MEXGB.
No takes yet. Share an insight, caveat, or question.
Olaniran et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: