BMIselect

Bayesian MI-LASSO for Variable Selection on Multiply-Imputed Data.

BMIselect provides a suite of Bayesian MI-LASSO methods for variable selection on multiply-imputed datasets. It implements four Bayesian MI-LASSO models — three shrinkage priors (Multi-Laplace, Horseshoe, ARD) and one spike-and-slab prior (Spike-and-Laplace) — together with tools for model fitting via MCMC, a four-step projection-predictive variable-selection procedure, and hyperparameter calibration. The methods handle both continuous and binary covariates under MAR or MCAR missingness. The frequentist MI-LASSO of Chen and Wang (2013) is also provided via MI_LASSO().

The MCMC samplers are implemented in C++ (via Rcpp/RcppArmadillo) for speed.

Installation

From CRAN:

install.packages("BMIselect")

Development version from GitHub:

# install.packages("devtools")
devtools::install_github("zjg540066169/BMIselect")

Quick start

library(BMIselect)

# simulate a multiply-imputed dataset (D = 5 imputations)
sim <- sim_A(n = 100, p = 20, type = "MAR", SNP = 1.5,
             low_missing = TRUE, n_imp = 5, seed = 123)
X <- sim$data_MI$X   # D x n x p array of imputed covariates
Y <- sim$data_MI$Y   # D x n matrix of outcomes

# fit a Bayesian MI-LASSO model and run the four-step selection
fit <- BMI_LASSO(X, Y, model = "Horseshoe",
                 nburn = 4000, npost = 4000, nchains = 1, ncores = 1)

fit$best_select   # selected variables

Available model values: "Multi_Laplace", "Horseshoe", "ARD", "Spike_Laplace". See the package vignette (vignette("Introduction", package = "BMIselect")) for a full walkthrough of fitting, projection-predictive selection, and posterior inference.

Citation

Zou, J., Wang, S. and Chen, Q. (2025). Bayesian MI-LASSO for Variable Selection on Multiply-Imputed Data. arXiv:2211.00114. https://doi.org/10.48550/arXiv.2211.00114

The frequentist MI-LASSO is from Chen, Q. and Wang, S. (2013). Variable selection for multiply-imputed data with application to dioxin exposure study. Statistics in Medicine, 32(21): 3646–3659.

License

Apache License (>= 2). See LICENSE.