TMLE when the predictions are provided.

I noticed that this code doesn't work when the user has already the predictions. HI wrote this code and can be added to tmle.py.

This code gives slightly different results. I hope the writers of EconML fix the code and add it:

def tmle_est_given_preds(self, X, a, y, Q1, Q0, g):
n = X.shape[0]

Q = a * Q1 + (1-a) * Q0
Q = Q.reshape(-1, 1)

clever = a / g - (1 - a) / (1 - g)
g = g.reshape(-1, 1)

target_scaler_ = sklearn.preprocessing.MinMaxScaler(feature_range=(0, 1))
target_scaler_.fit(y)
y_rescale = target_scaler_.transform(y)
Q_rescale = target_scaler_.transform(Q)
Q1_rescale = target_scaler_.transform(Q1.reshape(-1, 1))
Q0_rescale = target_scaler_.transform(Q0.reshape(-1, 1))

Q_rescale = _logit(Q_rescale)
Q1_pred = _logit(Q1_rescale)
Q0_pred = _logit(Q0_rescale)

targeted_outcome_model = sm.GLM(
                                endog=y_rescale.flatten(), 
                                exog=clever.flatten(), 
                                offset=Q_rescale.flatten(), 
                                family=sm.families.Binomial(),
                            ).fit()

epsilon = targeted_outcome_model.params[0]

logit_Q1_star = Q1_pred + epsilon / g
logit_Q0_star = Q0_pred - epsilon / (1 - g)

Q1_star = target_scaler_.inverse_transform(_expit(logit_Q1_star))
Q0_star = target_scaler_.inverse_transform(_expit(logit_Q0_star))

est = (Q1_star - Q0_star).mean()
phi = (a * (y - Q1_star) / g) - ((1 - a) * (y - Q0_star) / (1 - g)) + (Q1_star - Q0_star) - est
var =( (phi - phi.mean()) ** 2).mean() / n 

return est, var

Start by locating tmle.py and reading the existing TMLE implementation and its public entry points. Compare the proposed tmle_est_given_preds path with the current behavior, then confirm the intended estimates and variance when predictions are supplied; done means this use case is supported with results matching the project’s expected behavior.