Build.RF.ITR: Builds an rcRF model for risk controlled optimization

In kdoub5ha/mvITR: Discovery of risk controlled individualized treatment rules (ITRs)

Description

This function constructs a random forest of rcDT trees method. A forest object can be an argument into 'predict.ITR()' along with data in order to obtain treatment predictions. An output from this function can also be given to the 'Variable.Importance.ITR()' function to estimate predictor importance.

Usage

Build.RF.ITR(dat, split.var, efficacy = "y", risk = "r", col.trt = "trt",
  col.prtx = "prtx", risk.control = TRUE, risk.threshold = NA,
  lambda = 0, stabilize = TRUE, stabilize.type = c("linear", "rf"),
  test = NULL, ctg = NULL, N0 = 20, n0 = 5, max.depth = 10,
  ntree = 500, mtry = max(floor(length(split.var)/3), 1),
  avoid.nul.tree = FALSE, AIPWE = FALSE, verbose = FALSE,
  use.other.nodes = TRUE, extremeRandomized = FALSE)

Arguments

dat	data.frame. Data used to construct rcRF model. Must contain efficacy variable (y), risk variable (r), binary treatment indicator coded as 0 / 1 (trt), propensity score (prtx), candidate splitting covariates.
split.var	numeric vector. Columns of spliting variables.
efficacy	char. Efficacy outcome column. Defaults to 'y'.
risk	char. Risk outcome column. Defaults to 'r'.
col.trt	char. Treatment column name
risk.control	logical. Should risk be controlled? Defaults to TRUE.
risk.threshold	numeric. Desired level of risk control.
lambda	numeric. Penalty parameter for risk scores. Defaults to 0, i.e. no constraint. Optional arguments
stabilize	logical indicating if efficacy should be modeled using residuals. Defaults to TRUE.
stabilize.type	character specifying method used for estimating residuals. Current options are 'linear' for linear model (default) and 'rf' for random forest.
test	data.frame of testing observations. Should be formatted the same as 'data'.
ctg	numeric vector corresponding to the categorical input columns. Defaults to NULL. Not available yet.
N0	numeric specifying minimum number of observations required to call a node terminal. Defaults to 20.
n0	numeric specifying minimum number of treatment/control observations needed in a split to declare a node terminal. Defaults to 5.
max.depth	numeric specifying maximum depth of the tree. Defaults to 15 levels.
ntree	numeric. Number of trees generated. Defaults to 500.
mtry	numeric specifying the number of randomly selected splitting variables to be included. Defaults to number of splitting variables.
avoid.nul.tree	logical. Should null trees be discarded?
AIPWE	logical. Should AIPWE (TRUE) or IPWE (FALSE) be used. Not available yet.
verbose	logical. Give updates about forest progression?
use.other.nodes	logical. Should global estimator of objective function be used. Defaults to TRUE.
extremeRandomized	logical. Experimental for randomly selecting cutpoints in a random forest model. Defaults to FALSE and users should change this at their own peril.
col.ptrx	char. Propensity score column name.

Value

A list of characteristics of the forest.

ID.Boots.Samples	list of bootstrap sample IDs
TREES	list of trees
Model.Specification	information about the input parameters of the forest
...	Summaries for in and out of bag samples

Examples

dat <- generateData(n = 500)
# Generates rcRF model using simualated data with splitting variables located in columns 1-10.
fit <- Build.RF.ITR(dat = dat, split.var = 1:10, ntree = 200,
                    risk.control = TRUE, risk.threshold = 2.75, 
                    lambda = 1)

kdoub5ha/mvITR documentation built on April 7, 2020, 3:59 a.m.