itrFit: Fitting Individualized Treatment Recommendation (ITR)
In menghaomiao/itrrnr: Alternative Individual Treatment Recommendation

Description Usage Arguments Details Value Author(s) Examples

View source: R/itrFit.R

Fit alternative individualized treatment recommendation (A-ITR) or ITR given a loss function and a kernel.

1	itrFit(x, a, y, p = 1/k, c = 1.2, method = "svm", kernel = "linear", epsilon = 1/2/median(d)^2, d = 2, lambda = 5^(-5:5), cv = FALSE, tunning = TRUE)

`x`	a data frame or a matrix.
`a`	a factor or an integer vector indicating the treatment assigned to patients.
`y`	a numeric vector indicating the outcome after taking the treatment.
`p`	a numeric vector with each element between 0 to 1, indicating the probability of the treatment been assigned to the patient.
`c`	a numeric value no less than 1, indicating the bent slope of the loss.
`method`	the method to be used. Must be one of `"svm"`, `"dwd"`, `"exp"`, `"log"` and `"sq"`.
`kernel`	kernel type to be used. Currently only `linear`, `gaussian`, and `polynomial` are supported.
`epsilon`	a positive numeric value used only when `kernel = "gaussian"`.
`d`	a positive integer used only when `kernel = "polynomial"`.
`lambda`	a numeric vector indicating the penalty for tunning.
`cv`	logical. If `TRUE` then a 5-fold cross validation is performed.
`tunning`	logical. If `TRUE` then a tunning procedure for `delta` and `fence` (used for predicting A-ITR) is performed.

This function fits an individualized treatment recommendation under a bent loss. The bent slope is controled by the parameter c. When c = 1, it degenerates to the traditional ITR fit. For the purpose of saving computation time, it is suggested to set cv = TRUE for bent svm loss and cv = FALSE for other bent losses. It is also suggested to set tunning = TRUE for fitting A-ITR.

itrFit returns an object of class c("itrfit.svm", "itrfit") or c("itrfit.dif", "itrfit"). An object of class "itrfit" is a list containing at least the following components:

`obj_value`	the expected outcome under the traditional ITR. If `tunning = TRUE`, the weighted expected outcome (for evaluating an A-ITR) is also returned.
`optlambda`	the optimal lambda.
`c`	the bent slope.
`level`	the treatment levels.
`x`	the input covariate `x` used for fitting.
`a`	the input treatment `a` used for fitting.
`y`	the input outcome `y` used for fitting.
`method`	the method used.
`kernel`	the kernel matrix computed.
`predict`	an object of class `"ITR"`. If `tunning = TRUE`, return A-ITR, otherwise return ITR.
`refine_par`	if `tunning = TRUE`, the tunning parameter used for predicting the A-ITR.

In addition, the fitted coefficients are returned depending on the object class. Specifically, theta_s_gamma for "itrfit.svm" object, and coef for "itrfit.dif" object.

Haomiao Meng

data(sim1)
data = sim1[1:1000, ]

res = itrFit(data[, 1:2], data[, 3], data[, 4], cv = TRUE) #using bent svm loss
res = itrFit(data[, 1:2], data[, 3], data[, 4], kernel = 'polynomial', d = 4, lambda = 5^(4:11)) #using bent svm loss with polynomial kernel
res = itrFit(data[, 1:2], data[, 3], data[, 4], method = 'square', kernel = 'polynomial', d = 2, lambda = 5^(0:6)) #using bent square loss