drdrtest.superlearner {DRDRtest}R Documentation

The function for performing tests of average treatment effects with SuperLearner

Description

This is the function for testing average treatment effects with user specified nuisance functions.

Usage

drdrtest.superlearner(
  y,
  a,
  l,
  arange,
  pi.sl.lib = c("SL.earth", "SL.glm", "SL.gam", "SL.glmnet"),
  mu.sl.lib = c("SL.earth", "SL.glm", "SL.gam", "SL.glmnet"),
  mu.family = "gaussian",
  h = NULL,
  b = 1000,
  dist = "TwoPoint",
  a.grid.size = 401,
  pi.low = 0.01,
  pi.var.low = 0.01
)

Arguments

y

A vector containing the outcomes for each observation

a

A vector containing the treatment levels (dosage) for each observation

l

A data.frame containing the observations of covariates

arange

A vector of length 2 giving the lower bound and upper bound of treatment levels

pi.sl.lib

Models will be used by SuperLearner to estiamte propensity scores

mu.sl.lib

Models will be used by SuperLearner to estiamte outcome regression function

mu.family

Type of response. Currently only support "gaussian" and "binomial"

h

bandwidth to be used in kernel regression. If not specified, will by default use "rule of thumb" bandwidth selector

b

number of Bootstrap samples to be generated

dist

distibution used to generate residuals for Bootstrap samples. Currently only have two options, "TwoPoint" and "Rademachar"

a.grid.size

size of equally spaced grid points over arange to be generate for numerically evaluating the integral in test statistic

pi.low

Lower bound to truncate propensity scores

pi.var.low

Lower bound to truncate conditional variance of treament (used in propensity score estimation).

Value

A list containing

p.value:

P value of the test result

test.stat:

Value of the observed test statistic

Bootstrap.samples:

A vector containing test statistic values from Bootstrap samples

loc.fit:

A list containg evalution points of average treatment effect and the corresponding values

bandwidth:

Bandwidth used in kernel regression

Examples

mu.mod<-function(a,l,delta,height){
    mu <- as.numeric(l%*%c(0.2,0.2,0.3,-0.1))+triangle(a-2.5,delta,height)+a*(-0.1*l[,1]+0.1*l[,3])
    return(mu)
}
triangle <- function(a,delta,height){
    y <- exp(-a^2/((delta/2)^2))*height
    return(y)
}
set.seed(2000)
n <- 500
d <- 4
sigma <- 0.05
delta <- 1
height <- 0
arange<-c(0.01,4.99)

l <- matrix(rnorm(n*d),ncol=d)
colnames(l) <- paste("l",1:4,sep="")
logit.lambda <- as.numeric(l%*%c(0.1,0.1,-0.1,0.2))
lambda <- exp(logit.lambda)/(1+exp(logit.lambda))
a <- rbeta(n, shape1 = lambda, shape2 =1-lambda)*5

mu <- mu.mod(a,l,delta,height)
residual.list <- rnorm(n,mean=0,sd=sigma)
y <- mu+residual.list

out <- drdrtest.superlearner(y,a,l,arange,pi.sl.lib=c("SL.glm"),mu.sl.lib=c("SL.glm"))

[Package DRDRtest version 0.1 Index]