Attribute ate_ and method ate() give different results in CausalForestDML

[bart-vanneste]

Attribute ate_ and method ate() do not give the same point estimate and confidence interval. I suspect the former applies a doubly robust correction, whereas the latter may not.

Code to reproduce the issue:
import numpy as np
import pandas as pd
from econml.dml import CausalForestDML

# create synthetic data
n = 1000
np.random.seed(1)
T = np.random.randint(2, size=n)
X = np.random.normal(size=(n, 10))
W = np.random.normal(size=(n, 10))
Y = X**2 * T.reshape(-1, 1) + X * W
Y = np.sum(Y, axis=1)

# train model
m = CausalForestDML(discrete_treatment=True, max_features='sqrt', random_state=1)
m.tune(Y=Y, T=T, X=X, W=W)
m.fit(Y=Y, T=T, X=X, W=W)
m.summary()

# get ate (same as summary())
print(m.ate_)
print(m.ate_stderr_)

# get ate (not the same as summary())
print(m.ate(X=X, T0=0, T1=1))
print(m.ate_interval(X=X, T0=0, T1=1, alpha=.05))

The output is:

Especially, the confidence interval is substantially different.

The use case is to calculate the ATE for data not used for training.

Thanks!

[kbattocchi]

This is expected; as you note the ate_ attribute applies a double-robustness correction to the computation of the ATE itself (on the training data); the ate() method allows you to compute the ATE for any population by averaging the computed CATE values for each individual, so will not provide exactly the same result; however, if your use case is to compute the ATE for a data set that was not used in training then only the ate() method can be used for that.

[bart-vanneste]

Thanks much @kbattocchi

And how can one calculate the ATE with a double-robustness correction on data not used in training?