Fix #1060: make the uplift multiplier bootstrap seedable

[Hari20032005]

Fixes #1060.

calc_uplift drew the multiplier bootstrap behind the uniform confidence bands from the global numpy stream, so DRTester.evaluate_uplift() and evaluate_all() returned different bands on every call for identical inputs — and no caller could pin them.

The change

econml/validate/utils.py:

w = check_random_state(random_state).normal(0, 1, size=(n, n_bootstrap))

random_state is threaded through calc_uplift, evaluate_uplift and evaluate_all, normalized with sklearn.utils.check_random_state — the convention already used in _ortho_learner.py:741 and orf/_ortho_forest.py:234.

evaluate_uplift builds one RandomState and passes that instance into each per-treatment calc_uplift call, rather than passing the raw seed down. That keeps each treatment's multipliers independent — as they are today — while making the whole call reproducible from a single seed. evaluate_all does the same across its qini and toc calls.

The parameter defaults to None, so existing callers see no behavior change.

Effect

Three calc_uplift calls on identical inputs, before:

run 0: coeff=0.2579597804  stderr=0.0137873008  uniform_crit=2.845789  one_side_crit=2.633597
run 1: coeff=0.2579597804  stderr=0.0137873008  uniform_crit=2.808732  one_side_crit=2.609976
run 2: coeff=0.2579597804  stderr=0.0137873008  uniform_crit=2.884547  one_side_crit=2.595785

The point estimate and standard error were already deterministic — they never touch the bootstrap — which isolates the multiplier draws as the only source of the drift. With a seed passed, all three lines are identical.

Tests

test_uplift_random_state in econml/tests/test_drtester.py asserts that:

• the same seed gives identical bands,
• a different seed gives different bands, so the test cannot be satisfied by hard-coding a constant,
• the point estimates are unchanged across seeds,
• evaluate_all is reproducible end to end.

econml/tests/test_drtester.py is 6 passed locally on Python 3.13 (numpy 2.5.2, scikit-learn 1.9.0), and ruff check is clean on the three changed files.

🤖 Generated with Claude Code