filter5STAR: Filter covariates for 5STAR

In rmarceauwest/fiveSTAR: Run 5-STAR Algorithm

Filter covariates for 5STAR

Description

Performs Step 2 of the 5-STAR Algorithm: Forms dummy variable matrix for factors, and fits an elastic net (ENET) or random forest (RF) model to determine which covariates to keep for building trees in 5-STAR

Usage

filter5STAR(yy, X, family = "cox", plot = FALSE, verbose = 0,
  filter.hyper = filter_control(method = "ENET", lambdatype = "min",
  mixparm = NULL, vimpalpha = 0.05, nfolds = 10, filterseed = 2019, ...),
  vars2keep = NULL)

Arguments

yy	Response - either Surv() object for time to event data or 1 column matrix of case/control status for binary data
X	Data frame of all possible stratification covariates
family	Trait family, current options: "cox", "binomial", or "gaussian"
plot	Whether to make plots for filter results (e.g., variable importance plot for RF-based filtering or solution path for ENET-based filtering)
verbose	Numeric variable indicating amount of information to print to the terminal (0 = nothing, 1 = notes only, 2+ = notes and intermediate output)
filter.hyper	List of control parameters for filtering step (see also filter_control); key agruments include: method: filtering method; current options are: "ENET" (Elastic Net), "RF" (Random Forest with default or input parameters, using delete-d jackknife confidence intervals for VIMP for variable selection), "RFbest" (Random Survival Forest, performing tuning for mtry and nodesize, and using double bootstrap confidence intervals for VIMP for variable selection; more accurate but slower than RF option). See glmnet for more details on the elastic net and rfsrc for more details on the random forest filtering lambdatype: Optional elastic net parameter; whether to use the tuning parameter lambda that minimizes cross validation error (lambdatype="min", default) or the largest lambda that gives error within 1 standard error of the minimum error (lambdatype="1se"). Ignored when method is "RF" or "RFbest" mixparm: Optional elastic net mixing parameter alpha or grid of alpha values to search over. If nothing is entered, the algorithm will perform a grid search for the best mixing parameter between 0.05 and 0.95. Ignored when method is "RF" or "RFbest" vimpalpha: Optional significance level for RF VIMP confidence intervals nfolds: number of folds to use for cross validation tuning of lambda parameter for elastic net filtering. Default = 10. Ignored when method="RF" or "RFbest" filterseed: optional seed for filtering step ... : Optional arguments for glmnet or rfsrc
vars2keep:	List of variable names (matching column names in X) of variables to be passed through filtering step without penalization, etc. This is ignored in the main 5-STAR algorithm but may be used when filtering step is used as a stand alone function. Currently only used when method = ENET

Value

cov2keep: List of all covariates kept after filtering step, selected by elastic net or random forest

For method=ENET, additionally outputs

• cvout: vector containing fraction of null deviance explained, mean cross validation error, and optimal tuning parameter values (see cv.glmnet and glmnet for more details)

• beta: the coefficients of glmnet fit given tuned parameters

• ENETplot: plot containing the deviance over different combinations of the mixing and tuning parameters alpha and lambda, with optimal alpha shown in blue and minimum deviance point for each alpha in black (left panel), and solution path for the best mixing parameter (right panel). Output when plot = TRUE

For method=RF or RFbest, additionally outputs:

• varselectmat: matrix of VIMP confidence intervals, p-values, and selection decision for each variable

• VIMPplot: default variable importance CI plot, output if plot = TRUE

• varselect2plot: variable selection information from subsample.rfsrc for making customizable VIMP CI plots

• forest: rfsrc object

rmarceauwest/fiveSTAR documentation built on June 30, 2023, 7:38 a.m.