score_nuisances with discrete treatment returns incorrect score

When using score_nuisances with a discrete treatment, the function does not return the correct score.

The issue comes from the inverse_onehot function in econml/utilities.py. Currently, when it receives as input a DataFrame generated by pandas.get_dummies(), it incorrectly decodes the treatment.

For example, in case of binary treatments, labels originally coded as 0 and 1 are shifted and end up being decoded as 1 and 2, due to the following implementation:

def inverse_onehot(T):
    """
    Given a one-hot encoding of a value, return a vector reversing the encoding to get numeric treatment indices.

    Note that we assume that the first column has been removed from the input.
    """
    assert ndim(T) == 2
    # note that by default OneHotEncoder returns float64s, so need to convert to int
    return (T @ np.arange(1, T.shape[1] + 1)).astype(int)

This logic introduces an off-by-one error when decoding treatments.

Expected behavior

The function should return zero-based indices, ensuring that discrete treatments (e.g. 0/1) remain consistent after decoding. A corrected implementation would have the following code:

def inverse_onehot(T):
    assert econml.utilities.ndim(T) == 2

    indices = (
        np.arange(0, T.shape[1])
        if isinstance(T, pd.DataFrame)
        else np.arange(1, T.shape[1] + 1)
    )

    return (T @ indices).astype(int)

This change guarantees that score_nuisances computes the correct score for discrete treatments.

Contributed by @Cantal00p, @f5ilverio

Start in econml/utilities.py at inverse_onehot and trace how score_nuisances passes discrete-treatment DataFrames to it. Check the pandas.get_dummies() and non-DataFrame paths, then verify that zero-based treatment labels are preserved and score_nuisances returns the correct score for binary treatments.