def test_logit(simulated_problem: SimulatedProblemFixture, method: str,
center_moments: bool, W_type: str, se_type: str) -> (None):
"""Test that Logit estimates are the same as those from the the linearmodels package."""
_, simulation_results, problem, _, _ = simulated_problem
product_data = simulation_results.product_data
# skip more complicated simulations
if problem.K2 > 0 or problem.K3 > 0 or problem.H > 0:
return
# solve the problem
results1 = problem.solve(method=method,
center_moments=center_moments,
W_type=W_type,
se_type=se_type)
# compute the delta from the logit problem
delta = np.log(product_data.shares)
for t in problem.unique_market_ids:
shares_t = product_data.shares[product_data.market_ids == t]
delta[product_data.market_ids == t] -= np.log(1 - shares_t.sum())
# configure covariance options
W_options = {
'clusters': product_data.clustering_ids
} if W_type == 'clustered' else {}
se_options = {
'clusters': product_data.clustering_ids
} if se_type == 'clustered' else {}
# monkey-patch a problematic linearmodels method that shouldn't be called but is anyways
linearmodels.IVLIML._estimate_kappa = lambda _: 1
# solve the problem with linearmodels
model = linearmodels.IVGMM(delta,
exog=None,
endog=problem.products.X1,
instruments=problem.products.ZD,
center=center_moments,
weight_type=W_type,
**W_options)
results2 = model.fit(iter_limit=1 if method == '1s' else 2,
cov_type=se_type,
**se_options)
# test that results are essentially identical
for key1, key2 in [('beta', 'params'), ('xi', 'resids'),
('beta_se', 'std_errors')]:
values1 = getattr(results1, key1)
values2 = np.c_[getattr(results2, key2)]
np.testing.assert_allclose(values1,
values2,
atol=1e-10,
rtol=1e-8,
err_msg=key1)
def test_logit(
simulated_problem: SimulatedProblemFixture, method: str, center_moments: bool, W_type: str, se_type: str) -> (
None):
"""Test that Logit estimates are the same as those from the the linearmodels package."""
_, simulation_results, problem, _, _ = simulated_problem
product_data = simulation_results.product_data
# skip more complicated simulations
if problem.K2 > 0 or problem.K3 > 0 or problem.H > 0:
return pytest.skip("This simulation cannot be tested against linearmodels.")
# solve the problem
results1 = problem.solve(method=method, center_moments=center_moments, W_type=W_type, se_type=se_type)
# compute the delta from the logit problem
delta = np.log(product_data.shares)
for t in problem.unique_market_ids:
shares_t = product_data.shares[product_data.market_ids == t]
delta[product_data.market_ids == t] -= np.log(1 - shares_t.sum())
# configure covariance options
W_options = {'clusters': product_data.clustering_ids} if W_type == 'clustered' else {}
se_options = {'clusters': product_data.clustering_ids} if se_type == 'clustered' else {}
# monkey-patch a problematic linearmodels method that shouldn't be called but is anyways
linearmodels.IVLIML._estimate_kappa = lambda _: 1
# solve the problem with linearmodels
model = linearmodels.IVGMM(
delta, exog=None, endog=problem.products.X1, instruments=problem.products.ZD, center=center_moments,
weight_type=W_type, **W_options
)
results2 = model.fit(iter_limit=1 if method == '1s' else 2, cov_type=se_type, **se_options)
# test that results are essentially identical (unadjusted second stage standard errors will be different because
# linearmodels still constructs a S matrix)
for key1, key2 in [('beta', 'params'), ('xi', 'resids'), ('beta_se', 'std_errors')]:
if not (se_type == 'unadjusted' and key1 == 'beta_se'):
values1 = getattr(results1, key1)
values2 = np.c_[getattr(results2, key2)]
np.testing.assert_allclose(values1, values2, atol=1e-10, rtol=1e-8, err_msg=key1)
# test that test statistics in the second stage (when they make sense) are essentially identical
if method == '2s' and se_type != 'unadjusted':
nonconstant = (problem.products.X1[0] != problem.products.X1).any(axis=0)
F1 = results1.run_wald_test(results1.parameters[nonconstant], np.eye(results1.parameters.size)[nonconstant])
J1 = results1.run_hansen_test()
F2 = results2.f_statistic.stat
J2 = results2.j_stat.stat
np.testing.assert_allclose(F1, F2, atol=1e-10, rtol=1e-8)
np.testing.assert_allclose(J1, J2, atol=1e-10, rtol=1e-8)