Add capability to predict the outcomes to causal tree/forest

[winston-zillow]

While we use CausalML to predict the effects, one often wants to know the outcome values of the control and/or treatment given the covariates at the same time. Even though one could build separate prediction tree/forest for this purpose, not only that approach is more inconvenient and expensive, but it is hard to ensure the prediction model agrees with the causal model. (It seems that the nodes of CausalTree/CausalRandomForest already contain the necessary values, e.g. ct_y_sum and ct_count etc. It currently lack ways to aggregate them at the API level.)

[jeongyoonlee]

Hi @winston-zillow, once you train a causalml model, you can predict for both the control and treatment units with the same covariates. Is is different from what you describe here?

If so, can you elaborate more on what you'd like to achieve? I'd appreciate it if you could provide a pseudo code with the APIs you have in mind.

[winston-zillow]

@jeongyoonlee I meant the predict method output the effects, i.e. the delta of the outcome between control and treatment, correct?

tree1_ite_pred = tree1.predict(df_test[feature_names].values)
tree2_ite_pred = tree2.predict(df_test[feature_names].values)

df_result = pd.DataFrame(
    {
        'tree_mse_ite': tree1_ite_pred,
        'tree_causal_mse_ite': tree2_ite_pred,
        'outcome': df_test['outcome'], # <== at inference, we also want to estimate this
        'is_treated': df_test['treatment'],
        'treatment_effect': df_test['treatment_effect']
    }
)

But during inference, given a unit with covariates, we also want the estimated outcome using the same trained model.

The GRF in EconML has the predict_full() method that also estimate the counterfactual outcomes along with the effects, as shown in the attached screenshot for the model built as following:

# Code for EconML predict_full()
from econml.grf import CausalForest
est = CausalForest(criterion='het', n_estimators=400, min_samples_leaf=5, max_depth=None,
                   min_var_fraction_leaf=None, min_var_leaf_on_val=True,
                   min_impurity_decrease = 0.0, max_samples=0.45, min_balancedness_tol=.45,
                   warm_start=False, inference=True, fit_intercept=True, subforest_size=4,
                   honest=True, verbose=0, n_jobs=-1, random_state=1235)
est.fit(X, T, y)
effect_and_Y0 = est.predict_full(X_test, alpha=0.01)

Is this clear? Is there a way to do the same in CausalML already?

[jeongyoonlee]

I'm closing this issue as it has been addressed in #623.