A flexible approach for variable selection in large-scale healthcare database studies with missing covariate and...

Area of Science:

• Biostatistics
• Machine Learning
• Health Informatics

Background:

• Previous methods combining bootstrap imputation and tree-based variable selection are effective for missing at random (MAR) data but computationally expensive.
• Large-scale datasets pose significant computational challenges for existing MAR data analysis techniques.

Purpose of the Study:

• To propose and evaluate RR-BART, an inference-based variable selection method for MAR data.
• To assess RR-BART's performance against bootstrap imputation methods in simulations and a real-world case study.

Main Methods:

• Leveraged Bayesian additive regression trees (BART) and Rubin's rule for combining estimates from multiply imputed datasets.
• Conducted simulation studies to evaluate prediction and variable selection performance under complex MAR conditions.
• Applied RR-BART to the Study of Women's Health Across the Nation (SWAN) dataset to identify risk factors for metabolic syndrome.

Main Results:

• RR-BART effectively recovers prediction and variable selection performance, even with high missingness and complex data structures.
• Achieved optimal performance comparable to the best bootstrap imputation methods, with superior detection of discrete predictors.
• Demonstrated substantial computational savings and identified biologically plausible risk factors in the SWAN dataset.

Conclusions:

• RR-BART offers a computationally efficient and effective solution for variable selection in MAR data.
• The method exhibits strong operating characteristics, making it suitable for large-scale healthcare database studies.
• RR-BART advances the analysis of incomplete health data, aiding in the identification of significant risk factors.