Question: how to do sensitivity analysis with S_learner

[huanyi89]

I am using s_learner for my causal modeling (BaseSClassifier(learner=XGBClassifier))
However, when I trying to access the sensitivity for the model, it reported an error for the function
ate_new, ate_new_lower, ate_new_upper = self.get_ate_ci(X, p, treatment_new, y)
ValueError: not enough values to unpack (expected 3, got 1)

My main question here is can I do the sensitivity analysis for S_learner? If yes, how to do that. I am not sure if I missed some settings that can make it work.

Thank you

[0trm]

Yes. The not enough values to unpack (expected 3, got 1) was a bug in Sensitivity.get_ate_ci: it called estimate_ate() without return_ci=True, so learners like the S-learner returned only the ATE (#757, fixed in #758, released in 0.15.1). After upgrading, this runs:

from causalml.inference.meta import BaseSClassifier
from causalml.metrics.sensitivity import Sensitivity
from xgboost import XGBClassifier

sens = Sensitivity(
    df=df, inference_features=features, p_col="pihat",
    treatment_col="treatment", outcome_col="y",
    learner=BaseSClassifier(learner=XGBClassifier()),
)
sens.sensitivity_analysis(
    methods=["Placebo Treatment", "Random Cause", "Subset Data", "Random Replace"],
    sample_size=0.5,
)

Leave "Selection Bias" out with a classifier. That method subtracts a confounding function from the outcome (sensitivity.py), so the adjusted y is continuous and the classifier fails with Unknown label type: continuous. Run that one with BaseSRegressor instead. p_col is required even though the S-learner doesn't use the propensity. Checked on causalml 0.17.0.