High-dimensional analyses often overlook differences in variable types and treat all covariates as continuous for mathematical convenience. In medical research, however, binary and categorical variables encoding physicians’ diagnoses are informative...
High-dimensional analyses often overlook differences in variable types and treat all covariates as continuous for mathematical convenience. In medical research, however, binary and categorical variables encoding physicians’ diagnoses are informative for disease prediction. In this study, we focus on the Bayesian nonparametric model, Bayesian Additive Regression Trees (BART). BART consists of a decision tree ensemble guided by prior distributions at each split to enhance the detection of relevant covariates. In high-dimensional settings, a Dirichlet prior enables efficient estimation of sparsity through a Markov Chain Monte Carlo (MCMC). We propose a “weighted” BART that assigns variable-specific weights to the concentration parameters of the Dirichlet prior. We provide a theoretical rationale for how prior information is incorporated into statistical models. The proposed model increases the selection probability of informative covariates, regardless of their types, through a weighted Dirichlet prior. The performance of the proposed methods is illustrated through simulation and real data analysis, with evaluation focused on prediction and variable selection based on posterior inclusion probabilities.