def test_invalid_max_weight(self):
# Should fail when the max_weight is smaller than the uniform weight.
with self.assertRaises(ValueError):
ec.calibrate(
covariates=self.covariates,
target_covariates=self.target_covariates,
max_weight=1.0 / (_SIZE + 1000))
def test_invalid_l2_norm(self):
# Should fail with negative l2_norm.
with self.assertRaises(ValueError):
ec.calibrate(
covariates=self.covariates,
target_covariates=self.target_covariates,
l2_norm=-0.01)
def test_autoscale(self, objective):
weights_raw = ec.calibrate(covariates=self.covariates,
target_covariates=self.target_covariates,
autoscale=False,
objective=objective)[0]
weights_scaled = ec.calibrate(covariates=self.covariates,
target_covariates=self.target_covariates,
autoscale=True,
objective=objective)[0]
# With and without autoscale should give almost identical weight.
self.assertAlmostEqual(0, np.linalg.norm(weights_raw - weights_scaled))
def test_quadratic_balancing_approximate(self):
weights_exact = ec.calibrate(covariates=self.covariates,
target_covariates=self.target_covariates,
objective=ec.Objective.QUADRATIC,
l2_norm=0.0)[0]
weights_approximate = ec.calibrate(
covariates=self.covariates,
target_covariates=self.target_covariates,
objective=ec.Objective.QUADRATIC,
l2_norm=0.1)[0]
# Approximate match should lead to a better objective.
self.assertGreaterEqual(np.sum(np.square(weights_exact)),
np.sum(np.square(weights_approximate)))
def test_entropy_balancing_approximate(self):
weights_exact = ec.calibrate(covariates=self.covariates,
target_covariates=self.target_covariates,
objective=ec.Objective.ENTROPY,
l2_norm=0.0)[0]
weights_approximate = ec.calibrate(
covariates=self.covariates,
target_covariates=self.target_covariates,
objective=ec.Objective.ENTROPY,
l2_norm=0.1)[0]
# Approximate match should lead to a better objective.
self.assertGreaterEqual(-scipy.stats.entropy(weights_exact),
-scipy.stats.entropy(weights_approximate))
def test_entropy_balancing_bounded(self):
weights_unbounded = ec.calibrate(
covariates=self.covariates,
target_covariates=self.target_covariates,
objective=ec.Objective.ENTROPY,
max_weight=1.0)[0]
weights_bounded = ec.calibrate(
covariates=self.covariates,
target_covariates=self.target_covariates,
objective=ec.Objective.ENTROPY,
max_weight=0.005)[0]
# Imposing an upper bound on weights should lead to a worse objective.
self.assertLessEqual(-scipy.stats.entropy(weights_unbounded),
-scipy.stats.entropy(weights_bounded))
def test_quadratic_balancing_bounded(self):
weights_unbounded = ec.calibrate(
covariates=self.covariates,
target_covariates=self.target_covariates,
objective=ec.Objective.QUADRATIC,
max_weight=1.0)[0]
weights_bounded = ec.calibrate(
covariates=self.covariates,
target_covariates=self.target_covariates,
objective=ec.Objective.QUADRATIC,
max_weight=0.005)[0]
# Imposing an upper bound on weights should lead to a worse objective.
self.assertLessEqual(np.sum(np.square(weights_unbounded)),
np.sum(np.square(weights_bounded)))
def test_uniform_weights(self, objective):
# when covariates == target_covariates, uniform weights should be returned.
covariates = target_covariates = np.random.normal(size=(100, 4))
weights, success = ec.calibrate(covariates=covariates,
target_covariates=target_covariates,
objective=objective)
self.assertTrue(all(np.isclose(weights, 1.0 / 100)))
def test_calibrate(self, objective, max_weight, l2_norm):
weights, success = ec.calibrate(
covariates=self.covariates,
target_covariates=self.target_covariates,
objective=objective,
max_weight=max_weight,
l2_norm=l2_norm)
self.assertTrue(success)
self.assert_weights_constraints(weights, max_weight)
self.assert_balancing_constraint(weights, l2_norm)
def test_baseline_weights(self, objective):
covariates = target_covariates = np.random.normal(size=(100, 2))
repeats = np.repeat([2, 1], 50)
weights, success = ec.calibrate(covariates=covariates,
target_covariates=target_covariates,
baseline_weights=repeats,
target_weights=repeats,
objective=ec.Objective.QUADRATIC)
# First 50 rows should have 2 / 3 of the weights.
expected = repeats * 2 / 3 / 100
self.assertTrue(all(np.isclose(weights, expected)))
