MetaLearners and Classification

By default the MetaLearners are .predict() withing .effect() (through const_marginal_effect()). This is appropriate for regressors, but for classification, shouldn't it it be using .predict_proba() to compute the taus? Otherwise the taus are only {-1,0,1} using a base threshold of 0.5 within .predict().

def const_marginal_effect(self, X):
    """Calculate the constant marignal treatment effect on a vector of features for each sample.

    Parameters
    ----------
    X : matrix, shape (m × d_x)
        Matrix of features for each sample.

    Returns
    -------
    τ_hat : matrix, shape (m, d_y, d_t)
        Constant marginal CATE of each treatment on each outcome for each sample X[i].
        Note that when Y is a vector rather than a 2-dimensional array,
        the corresponding singleton dimensions in the output will be collapsed
    """
    # Check inputs
    if 'X' in self._gen_allowed_missing_vars():
        force_all_finite = 'allow-nan'
    else:
        force_all_finite = False
    X = check_array(X, force_all_finite=force_all_finite)
    taus = []
    for ind in range(self._d_t[0]):
        taus.append(self.models[ind + 1].predict(X) - self.models[0].predict(X))
    taus = np.column_stack(taus).reshape((-1,) + self._d_t + self._d_y)  # shape as of m*d_t*d_y
    if self._d_y:
        taus = transpose(taus, (0, 2, 1))  # shape as of m*d_y*d_t
    return taus

Start at the MetaLearners.const_marginal_effect() entry point shown in the issue and trace how its models handle classification versus regression. Run the relevant MetaLearners tests if present; done means classification treatment effects use the intended probability-based behavior without changing regression behavior.