def test_score_test(self):
res1 = self.res1
modr = self.klass(endog, exog.values[:, :-1])
resr = modr.fit(method="newton", maxiter=300)
params_restr = np.concatenate(
[resr.params[:-1], [0], resr.params[-1:]])
r_matrix = np.zeros((1, len(params_restr)))
r_matrix[0, -2] = 1
exog_extra = res1.model.exog[:, -1:]
from statsmodels.base._parameter_inference import score_test
sc1 = score_test(res1,
params_constrained=params_restr,
k_constraints=1)
sc2 = score_test(resr, exog_extra=(exog_extra, None))
assert_allclose(sc2[:2], sc1[:2])
sc1_hc = score_test(res1,
params_constrained=params_restr,
k_constraints=1,
r_matrix=r_matrix,
cov_type="HC0")
sc2_hc = score_test(resr,
exog_extra=(exog_extra, None),
cov_type="HC0")
assert_allclose(sc2_hc[:2], sc1_hc[:2])
def test_wald_score(self):
mod_full = self.model_full
mod_drop = self.model_drop
restriction = 'x5=0, x6=0'
res_full = mod_full.fit()
res_constr = mod_full.fit_constrained('x5=0, x6=0')
res_drop = mod_drop.fit()
wald = res_full.wald_test(restriction)
lm_constr = np.hstack(score_test(res_constr))
lm_extra = np.hstack(score_test(res_drop, exog_extra=self.exog_extra))
res_wald = np.hstack([wald.statistic.squeeze(), wald.pvalue, [wald.df_denom]])
assert_allclose(lm_constr, res_wald, rtol=self.rtol_ws, atol=self.atol_ws)
assert_allclose(lm_extra, res_wald, rtol=self.rtol_ws, atol=self.atol_ws)
assert_allclose(lm_constr, lm_extra, rtol=1e-12, atol=1e-14)
# regression number
assert_allclose(lm_constr[1], self.res_pvalue[0], rtol=1e-12, atol=1e-14)
cov_type='HC0'
res_full_hc = mod_full.fit(cov_type=cov_type, start_params=res_full.params)
wald = res_full_hc.wald_test(restriction)
lm_constr = np.hstack(score_test(res_constr, cov_type=cov_type))
lm_extra = np.hstack(score_test(res_drop, exog_extra=self.exog_extra,
cov_type=cov_type))
res_wald = np.hstack([wald.statistic.squeeze(), wald.pvalue, [wald.df_denom]])
assert_allclose(lm_constr, res_wald, rtol=self.rtol_ws, atol=self.atol_ws)
assert_allclose(lm_extra, res_wald, rtol=self.rtol_ws, atol=self.atol_ws)
assert_allclose(lm_constr, lm_extra, rtol=1e-13)
# regression number
assert_allclose(lm_constr[1], self.res_pvalue[1], rtol=1e-12, atol=1e-14)
if not self.skip_wooldridge:
# compare with Wooldridge auxiliary regression
# does not work for Poisson, even with family attribute
# diao.lm_test_glm assumes fittedvalues is mean (not linear pred)
lm_wooldridge = diao.lm_test_glm(res_drop, self.exog_extra)
assert_allclose(lm_wooldridge.pval1, self.res_pvalue[0],
rtol=1e-12, atol=1e-14)
assert_allclose(lm_wooldridge.pval3, self.res_pvalue[1],
rtol=self.rtol_wooldridge)
# smoke test
lm_wooldridge.summary()
def test_score_test(self):
res1 = self.res1
resr = self.resr
params_restr = np.concatenate([resr.params, [0]])
r_matrix = np.zeros((1, len(params_restr)))
r_matrix[0, -1] = 1
exog_prec_extra = res1.model.exog_precision[:, 1:]
from statsmodels.base._parameter_inference import score_test
sc1 = score_test(res1, params_constrained=params_restr,
k_constraints=1)
sc2 = score_test(resr, exog_extra=(None, exog_prec_extra))
assert_allclose(sc2[:2], sc1[:2])
sc1_hc = score_test(res1, params_constrained=params_restr,
k_constraints=1, r_matrix=r_matrix, cov_type="HC0")
sc2_hc = score_test(resr, exog_extra=(None, exog_prec_extra),
cov_type="HC0")
assert_allclose(sc2_hc[:2], sc1_hc[:2])
def test_score_test_alpha(self):
# this is mostly sanity check
# we need heteroscedastic model, i.e. exog_dispersion as comparison
# res1 = self.res1
modr = self.klass(endog, exog.values[:, :-1])
resr = modr.fit(method="newton", maxiter=300)
params_restr = np.concatenate([resr.params[:], [0]])
r_matrix = np.zeros((1, len(params_restr)))
r_matrix[0, -1] = 1
# use random expg_extra, then we don't reject null
# exog_extra = res1.model.exog[:, 1:2]
np.random.seed(987125643)
exog_extra = 0.01 * np.random.randn(endog.shape[0])
from statsmodels.base._parameter_inference import (score_test,
_scorehess_extra)
# note: we need params for the restricted model here
# if params is not given, then it will be taked from results instance
sh = _scorehess_extra(resr,
exog_extra=None,
exog2_extra=exog_extra,
hess_kwds=None)
assert not np.isnan(sh[0]).any()
# sc1 = score_test(res1, params_constrained=params_restr,
# k_constraints=1)
sc2 = score_test(resr, exog_extra=(None, exog_extra))
assert sc2[1] > 0.01
# sc1_hc = score_test(res1, params_constrained=params_restr,
# k_constraints=1, r_matrix=r_matrix, cov_type="HC0")
sc2_hc = score_test(resr,
exog_extra=(None, exog_extra),
cov_type="HC0")
assert sc2_hc[1] > 0.01