Table 1 Candidate methods
Abbreviation | Description | Selection rule | R packages |
|---|---|---|---|
UNIV-BFN | Univariable models with Bonferroni-adjusted p-val | P < 5*10−5 | base R [23] |
LASSO-MIN | Lasso with λ chosen at the minimum prediction error | β ≠ 0 | glmnet [24] |
LASSO-1SE | Lasso with λ chosen at 1 SE above the minimum error | β ≠ 0 | glmnet |
ELNET-MIN | Elastic net, grid search for α (0.05–0.95 by 0.05), λ at min | β ≠ 0 | glmnet |
ELNET-1SE | Elastic net, grid search α (0.05–0.95 by 0.05), λ at 1 SE | β ≠ 0 | glmnet |
HCLST-CORR-SGL | Hierarchical clustering, groups with corr > 0.8, sparse group lasso | β ≠ 0 | SGL [25] |
HCLST-BOOT-SGL | Hierarchical clustering, groups from bootstrap, sparse group lasso | β ≠ 0 | SGL, pvclust [16] |
RF | Random Forests algorithm with bootstrap-based confidence intervals for the variable importance scores | 99.995% CI > 0 | randomForestSRC [26] |
BAGGING | Similar to Random Forests, but with all variables considered candidates for splitting at each node | 99.995% CI > 0 | randomForestSRC |
BART-LOCAL | Bayesian Additive Regression Trees, local criteria for Inclusion Proportion (IP) | IP > 0.95 quantile of local distribution | bartMachine [27] |
BART-GLOBALSE | Bayesian Additive Regression Trees, global SE criteria for IP | IP > threshold from local distribution with global multiplier | bartMachine |
BART-GLOBALMAX | Bayesian Additive Regression Trees, global Max criteria for IP | IP > 0.95 quantile of global max distribution | bartMachine |