causalDML

This package implements the Double Machine Learning based methods reviewed in Knaus (2022) for binary and multiple treatment effect estimation. Further, it implements the Heiler and Knaus (2021) decomposition of treatment effects with heterogeneous treatments.

For examples of applications check out the replication notebook for Knaus (2022), replication notebook 1 for Heiler and Knaus (2021) or replication notebook 2 for Heiler and Knaus (2021)

The current version can be installed via devtools:

library(devtools)
install_github(repo="MCKnaus/causalDML")

The following code creates synthetic data to showcase a comprehensive analysis based on the package:

library(causalDML)
library(policytree)
library(lmtest)
library(sandwich)

set.seed(1234)
## Generate random data
n = 2000
p = 10
X = matrix(rnorm(n * p), n, p)
W = sample(c("A", "B", "C"), n, replace = TRUE)
Y = X[, 1] + X[, 2] * (W == "B") + X[, 3] * (W == "C") + rnorm(n)

## Create components of ensemble
# General methods
mean = create_method("mean")
forest =  create_method("forest_grf",args=list(tune.parameters = "all"))

# Pscore specific components
ridge_bin = create_method("ridge",args=list(family = "binomial"))
lasso_bin = create_method("lasso",args=list(family = "binomial"))

# Outcome specific components
ols = create_method("ols")
ridge = create_method("ridge")
lasso = create_method("lasso")

## Run the main function that outputs nuisance parameters, APO and ATE
# Can take some minutes, set quiet=F to see progress or remove forest to speed up
DML = DML_aipw(Y,W,X,ml_w=list(mean,forest,ridge_bin,lasso_bin),
                ml_y=list(mean,forest,ols,ridge,lasso),quiet=T)

## Show and plot estimated APOs
summary(DML$APO)
plot(DML$APO)

## Show estimated ATEs
summary(DML$ATE)

## Estimate all possible ATETs
for (i in 1:3) {
  APO_atet = APO_dml_atet(Y,DML$APO$m_mat,DML$APO$w_mat,DML$APO$e_mat,DML$APO$cf_mat,treated=i)
  ATET = ATE_dml(APO_atet)
  print(summary(ATET))
}

## Best linear prediction of CATEs along variable X1
for (i in 1:3) {
  temp_ols = lm(DML$ATE$delta[,i] ~ X[, 1])
  print(coeftest(temp_ols, vcov = vcovHC(temp_ols)))
}

## Kernel regression CATEs along variable X1
# Might also take some minutes
for (i in 1:3) {
  temp_kr = kr_cate(DML$ATE$delta[,i],X[, 1])
  print(plot(temp_kr))
}

## Spline regression CATEs along variable X1
for (i in 1:3) {
  temp_sr = spline_cate(DML$ATE$delta[,i],X[, 1])
  print(plot(temp_sr))
}

## Plain DR-learner for all observations in the sample
# Can take some minutes, set quiet=F to see progress or remove forest to speed up
# Be sure to specify only linear smoothers for tau
dr = dr_learner(Y,W,X,ml_w=list(mean,lasso_bin),
                  ml_y = list(mean,lasso),
                  ml_tau = list(mean,lasso),quiet=T)

# DR-learner distribution of B-A
hist(dr$cates[,1])

## (N)DR-learner for all observations in the sample
# Can take some minutes, set quiet=F to see progress or remove forest to speed up
# Here we only compare B and C to the "baseline" treatment A
ndr = ndr_learner(Y,W,X,ml_w=list(mean,forest,ridge_bin,lasso_bin),
                  ml_y = list(mean,forest,ols,ridge,lasso),
                  ml_tau = list(mean,forest),quiet=T,
                  compare_all = FALSE)

# Plot DR-learner distribution of B-A
hist(ndr$cates[1,,1])

# Plot NDR-learner distribution of C-A
hist(ndr$cates[2,,2])

## (N)DR-learner for out-of-sample prediction in test set
# Can take some minutes, set quiet=F to see progress or remove forest to speed up
# Be sure to specify only linear smoothers for tau
X_test = matrix(rnorm(n/2 * p), n/2, p)
ndr.oos = ndr_oos(Y,W,X,xnew=X_test,ml_w=list(mean,forest,ridge_bin,lasso_bin),
                  ml_y = list(mean,forest,ols,ridge,lasso),
                  ml_tau = list(mean,forest),quiet=T)

# Plot DR-learner distribution of B-A
hist(ndr.oos$cates[[1]][,1])

# Plot NDR-learner distribution of C-A
hist(ndr.oos$cates[[2]][,2])

## Policy tree
tree1 = policy_tree(X,DML$APO$gamma,depth=1)
plot(tree1)
tree2 = policy_tree(X,DML$APO$gamma,depth=2)
plot(tree2)

References

Heiler, P., Knaus, M. C. (2021). Effect or Treatment Heterogeneity? Policy Evaluation with Aggregated and Disaggregated Treatments, arXiv

Knaus, M. C. (2022). Double machine learning based program evaluation under unconfoundedness, The Econometrics Journal, forthcoming, arXiv