DDcomb: Double demeaning estimator with treatment and outcome...

In youmisuk/CURobustML: Robust Machine Learning Under Cluster-Level Unmeasured Confounding

Double demeaning estimator with treatment and outcome learners

Description

Double demeaning estimator with treatment and outcome learners

Usage

DDcomb(
  Y,
  Z,
  X,
  interZ = formula(~1),
  ID,
  data,
  library = c("glm", "deeplearning"),
  crossFitting = FALSE,
  K = 5,
  mCrossFit = 100
)

Arguments

Y	continuous outcome variable
Z	binary treatment indicator, 1 - treatment, 0 - control
X	vector, matrix, or dataframe containing measured confounders
interZ	formula that contains the variables that "interact" with the treatment. "1" will be always added. The default is no interaction, i.e., formula = formula(~1).
ID	cluster identifier
data	dataframe containing the variables in the model
library	character vector of prediction algorithms. The available methods are glm, deeplearning, gbm, and randomForests. The default methods are glm and deeplearning.
crossFitting	whether to do cross fitting. The default is FALSE, and currently it is not available.
K	number of folds. The default is 5, and currently it is not available.
mCrossFit	number of cross fitting. The default is FALSE, and currently it is not available.

Value

An DDcomb with the following elements:

coef.ER	vector of the coefficients for prediction algorithms in the demeaned treatment model
coef.OR	vector of the coefficients for prediction algorithms in the demeaned outcome model
Estimate	estimates and standard errors of treatment effects
Z.hat	final weighted prediction for the demeaned treatment
Y1.hat	final weighted prediction for the demeaned outcome among treated units
Y0.hat	final weighted prediction for the demeaned outcome among control units
Z.hats	all the predictions for the demeaned treatment from prediction algorithms
Y1.hats	all the predictions for the demeaned outcome among treated units from prediction algorithms
Y0.hats	all the predictions for the demeaned outcome among control units from prediction algorithms

Examples

# two-level data
DDcomb.rlst <- DDcomb(Y=twolevel_data$Y, Z=twolevel_data$Z, interZ=(~ W1),
 X=twolevel_data[, c("X1", "X2", "X3", "W1")], ID=twolevel_data$id,
 library="glm", data=twolevel_data)
summary(DDcomb.rlst)


# two-level data with default algorithms
DDcomb.rlst2 <- DDcomb(Y=twolevel_data$Y, Z=twolevel_data$Z, interZ=(~ W1),
 X=twolevel_data[, c("X1", "X2", "X3", "W1")], ID=twolevel_data$id,
 library=c("glm", "deeplearning"), data=twolevel_data)
summary(DDcomb.rlst2)

# cross-classified data with default algorithms
DDcomb.rlst3 <- DDcomb(Y=crossclassified_data$Y, Z=crossclassified_data$Z, interZ=(~ W1),
 X=crossclassified_data[, c("X1", "X2", "X3", "W1", "Q1")], ID=crossclassified_data$f12id,
 library=c("glm", "deeplearning"), data=crossclassified_data)
summary(DDcomb.rlst3)

youmisuk/CURobustML documentation built on Sept. 11, 2022, 11:04 a.m.