rcate.am: Robust estimation of treatment effect using additive b-spline...
In rhli-Hannah/RCATE: Robust Estimation of Heterogeneous Treatment Effect

Description Usage Arguments Details Value Examples

rcate.am fit additive model for robust treatment effect estimation.

rcate.am(
  x,
  y,
  d,
  method = "MCMEA",
  nknots = NA,
  lambda.smooth = 1,
  nlambda = 30,
  nfolds = 5,
  n.trees.p = 40000,
  shrinkage.p = 0.005,
  n.minobsinnode.p = 10,
  interaction.depth.p = 1,
  cv.p = 2,
  n.trees.mu = c(1:50) * 50,
  shrinkage.mu = 0.01,
  n.minobsinnode.mu = 5,
  interaction.depth.mu = 5,
  cv.mu = 5
)

`x`	matrix or a data frame of predictors.
`y`	vector of response values.
`d`	vector of binary treatment assignment (0 or 1).
`method`	character string of CATE estimation method: "MCMEA" - modified co-variate method with efficiency augmentation or "RL" - R-learning.
`nknots`	number of knots for B-spline.
`lambda.smooth`	scalar represent the smoothness penalty lambda2. The default is 2.
`nlambda`	number of lambda1s for cross-validation. The default is 30.
`nfolds`	number of folds for cross-validation. The default is 5.
`n.trees.p`	tuning parameter the number of trees used for estimating propensity score with GBM. the default value is 40000.
`shrinkage.p`	tuning parameter the shrinkage level for estimating propensity score with GBM. the default value is 0.005.
`n.minobsinnode.p`	tuning parameter the minimum node size for estimating propensity score with GBM. the default value is 10.
`interaction.depth.p`	tuning parameter the number of interactions for estimating propensity score with GBM. the default value is 1.
`cv.p`	tuning parameter the number of folds in cross-validation for estimating propensity score with GBM. the default value is 2.
`n.trees.mu`	scalar or vector of the number of trees for estimating mean function with GBM. The default is (1:50)*50.
`shrinkage.mu`	tuning parameter the shrinkage level for estimating mean function with GBM. the default value is 0.01.
`n.minobsinnode.mu`	tuning parameter the minimum node size for estimating mean function with GBM. the default value is 10.
`interaction.depth.mu`	tuning parameter the number of interactions for estimating mean function with GBM. the default value is 5.
`cv.mu`	tuning parameter the folds for cross-validation for estimating mean function with GBM. The default value is 5.

Fits a L_1 additive regression model with the group SCAD penalty.

a list of components

model - the robust estimation model of CATE.
method - estimation method.
algorithm - fitting algorithm.
lambda.smooth
fitted.values - vector of fitted values.
x - matrix of predictors.
y - vector of response values.
d - vector of treatment assignment.
y.tr - transformed outcome.
w.tr - transformed weight.
coef -coefficients.
colnum - column number.
nknots - number of knots of cubic spline.
param - required parameters for utility functions.

n <- 1000; p <- 2; set.seed(2223)
X <- as.data.frame(matrix(runif(n*p,-3,3),nrow=n,ncol=p))
tau = 3*X[,1]-2*X[,2]
p = 1/(1+exp(-X[,1]+X[,2]))
d = rbinom(n,1,p)
t = 2*d-1
y = 100+tau*t/2 + rnorm(n,0,1); set.seed(2223)
x_val = as.data.frame(matrix(rnorm(200*2,0,1),nrow=200,ncol=2))
tau_val = 3*x_val[,1]-2*x_val[,2]

fit <- rcate.am(X,y,d,lambda.smooth = 4, method = 'RL')
y_pred <- predict(fit,x_val)$pred
plot(tau_val,y_pred);abline(0,1)