lizhengecon / cates

Machine Learning Estimation of Heterogeneous Causal Effects

CATEs

Implementation of all estimators that are applied in the Empirical Monte Carlo Study of Knaus, Lechner and Strittmatter (2018). They are based on the packages grf and glmnet.

Example

We have no permission to share the data used in the study. Therefore, the following example uses the observational data generating process of the example of grf to illustrate how it works. The function IATEs is a wrapper for the underlying functions and uses all the default settings of the packages. If you want to have more control over the settings, use the respective functions in CATEs_utils.

# Download current version from Github
library(devtools)
install_github(repo="MCKnaus/CATEs")
library(CATEs)

# Generate training sample
n = 4000; p = 20
x_tr = matrix(rnorm(n * p), n, p)
tau_tr = 1 / (1 + exp(-x_tr[, 3]))
d_tr = rbinom(n ,1, 1 / (1 + exp(-x_tr[, 1] - x_tr[, 2])))
y_tr = pmax(x_tr[, 2] + x_tr[, 3], 0) + rowMeans(x_tr[, 4:6]) / 2 + d_tr * tau_tr + rnorm(n)

# Generate validation sample of same size
x_val = matrix(rnorm(n * p), n, p)
tau_val = 1 / (1 + exp(-x_val[, 3]))

# Apply all estimators to the training sample and predict IATEs for validation sample
iates_mat = IATEs(y_tr,d_tr,x_tr,tau_tr,x_val)

# Calculate and print mean MSEs
mMSE = colMeans((iates_mat - tau_val)^2)
names(mMSE) = colnames(iates_mat)
mMSE*1000

References

Knaus, Lechner, Strittmatter (2018). Machine Learning Estimation of Heterogeneous Causal Effects: Empirical Monte Carlo Evidence, arXiv