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

TL;DR

This paper introduces RR-BART, a Bayesian inference-based method for variable selection in large healthcare datasets with missing data, offering improved accuracy and computational efficiency over traditional bootstrap imputation techniques.

Contribution

The paper proposes RR-BART, a novel Bayesian approach that effectively handles missing data and improves variable selection accuracy in large-scale healthcare studies.

Findings

RR-BART outperforms bootstrap methods in simulation studies.

RR-BART demonstrates strong detection of discrete predictors.

RR-BART significantly reduces computational time.

Abstract

Prior work has shown that combining bootstrap imputation with tree-based machine learning variable selection methods can provide good performances achievable on fully observed data when covariate and outcome data are missing at random (MAR). This approach however is computationally expensive, especially on large-scale datasets. We propose an inference-based method, called RR-BART, which leverages the likelihood-based Bayesian machine learning technique, Bayesian additive regression trees, and uses Rubin's rule to combine the estimates and variances of the variable importance measures on multiply imputed datasets for variable selection in the presence of MAR data. We conduct a representative simulation study to investigate the practical operating characteristics of RR-BART, and compare it with the bootstrap imputation based methods. We further demonstrate the methods via a case study of risk factors for 3-year incidence of metabolic syndrome among middle-aged women using data from the Study of Women's Health Across the Nation (SWAN). The simulation study suggests that even in complex conditions of nonlinearity and nonadditivity with a large percentage of missingness, RR-BART can reasonably recover both prediction and variable selection performances, achievable on the fully observed data. RR-BART provides the best performance that the bootstrap imputation based methods can achieve with the optimal selection threshold value. In addition, RR-BART demonstrates a substantially stronger ability of detecting discrete predictors. Furthermore, RR-BART offers substantial computational savings. When implemented on the SWAN data, RR-BART adds to the literature by selecting a set of predictors that had been less commonly identified as risk factors but had substantial biological justifications.