RICWL: RICWL, robust confidence index weighted learning functions.

In sambiostat/RICWL:

Description

This is confidence index weighted learning that takes patient's characteristic X, treatment A and outcome Y and estimate an optimal individual treament rule.

Usage

RICWL(
  H,
  A,
  R2,
  pi = rep(0.5, n),
  kernel = "rbf",
  pentype = "lasso",
  XS = NULL,
  sigma,
  cpar,
  theta,
  frac.par = 0.5,
  method = "cos",
  m = 10,
  e = 1e-05
)

Arguments

H:	n by P feature matrix
A:	treatment, takes value -1 and +1 with size n
R2:	outcome or residual vector with length n
pi:	propensity score with length n, when RCT, it shoud be set as 0.5 for each subject
kernel:	kernel used in SVM
pentype:	penalty in the residual estimation process, useful or RWL only
XS:	variable used in similarity calculation, if not provided, use all features
sigma:	hyper-parameter in SVM rbf kernel
cpar:	hyper-parameter in SVM rbf kernel
theta:	hyper-parameter governing the definition of neighborhood. It's the quantile probability for similarity
frac.par:	if "frac" similarity is used, then provide this fraction parameter, ranges from 0-1
method:	either "cos" or "frac" for similarity definition
m:	number of fold for cross validation in parameter tunning, default is 10
e:	least tolerated error, default is 1e-5

Examples

n = 200
p = 5
H = matrix(rnorm(n*p), nrow=n)
A = rbinom(n=n, size=1, prob=0.5)*2 - 1
R2 = 4*H[,1]^2 + A*(H[,2]^2 + H[,4]^2 < 1.5) + rcauchy(n)
fit = RICWL(H, A, R2)
optimalITR_pred = predict(fit, H)

sambiostat/RICWL documentation built on Dec. 31, 2020, 3:17 a.m.