def setup_class(cls):
np.random.seed(4321)
n = 20
p = 5
exog = np.empty((n, p))
exog[:, 0] = 1
exog[:, 1] = np.random.randint(low=-5, high=5, size=n)
x = np.repeat(np.array([1, 2, 3, 4]), n / 4)
exog[:, 2:] = get_dummies(x)
beta = np.array([-1, 0.1, -0.05, .2, 0.35])
lin_pred = (exog * beta).sum(axis=1)
family = sm.families.Binomial
link = sm.families.links.log
endog = gen_endog(lin_pred, family, link, binom_version=0)
mod1 = sm.GLM(endog, exog, family=family(link=link()))
cls.res1 = mod1.fit(rtol=1e-10, atol=0, tol_criterion='params',
scaletype='x2')
agg = pd.DataFrame(exog)
agg['endog'] = endog
agg_endog = agg.groupby([0, 1, 2, 3, 4]).sum()[['endog']]
agg_wt = agg.groupby([0, 1, 2, 3, 4]).count()[['endog']]
agg_exog = np.array(agg_endog.index.tolist())
agg_wt = agg_wt['endog']
avg_endog = agg_endog['endog'] / agg_wt
mod2 = sm.GLM(avg_endog, agg_exog,
family=family(link=link()),
var_weights=agg_wt)
cls.res2 = mod2.fit(rtol=1e-10, atol=0, tol_criterion='params')
def setup_class(cls):
np.random.seed(4321)
n = 10000
p = 5
exog = np.empty((n, p))
exog[:, 0] = 1
exog[:, 1] = np.random.randint(low=-5, high=5, size=n)
x = np.repeat(np.array([1, 2, 3, 4]), n / 4)
exog[:, 2:] = get_dummies(x)
beta = np.array([7, 0.1, -0.05, .2, 0.35])
lin_pred = (exog * beta).sum(axis=1)
family = sm.families.Tweedie
link = sm.families.links.log
endog = gen_endog(lin_pred, family, link)
mod1 = sm.GLM(endog, exog, family=family(link=link(), var_power=1.5))
cls.res1 = mod1.fit(rtol=1e-20, atol=0, tol_criterion='params')
agg = pd.DataFrame(exog)
agg['endog'] = endog
agg_endog = agg.groupby([0, 1, 2, 3, 4]).sum()[['endog']]
agg_wt = agg.groupby([0, 1, 2, 3, 4]).count()[['endog']]
agg_exog = np.array(agg_endog.index.tolist())
agg_wt = agg_wt['endog']
agg_endog = agg_endog['endog']
mod2 = sm.GLM(agg_endog, agg_exog,
family=family(link=link(), var_power=1.5),
exposure=agg_wt, var_weights=agg_wt ** 0.5)
cls.res2 = mod2.fit(rtol=1e-20, atol=0, tol_criterion='params')
def setup_class(cls):
np.random.seed(4321)
n = 10000
p = 5
exog = np.empty((n, p))
exog[:, 0] = 1
exog[:, 1] = np.random.randint(low=-5, high=5, size=n)
x = np.repeat(np.array([1, 2, 3, 4]), n / 4)
exog[:, 2:] = get_dummies(x)
beta = np.array([-1, 0.1, -0.05, .2, 0.35])
lin_pred = (exog * beta).sum(axis=1)
family = sm.families.Poisson
link = sm.families.links.log
endog = gen_endog(lin_pred, family, link)
mod1 = sm.GLM(endog, exog, family=family(link=link))
cls.res1 = mod1.fit()
agg = pd.DataFrame(exog)
agg['endog'] = endog
agg_endog = agg.groupby([0, 1, 2, 3, 4]).sum()[['endog']]
agg_wt = agg.groupby([0, 1, 2, 3, 4]).count()[['endog']]
agg_exog = np.array(agg_endog.index.tolist())
agg_wt = agg_wt['endog']
avg_endog = agg_endog['endog'] / agg_wt
mod2 = sm.GLM(avg_endog, agg_exog, family=family(link=link),
var_weights=agg_wt)
cls.res2 = mod2.fit()
def setup_class(cls):
np.random.seed(4321)
n = 10000
p = 5
exog = np.empty((n, p))
exog[:, 0] = 1
exog[:, 1] = np.random.randint(low=-5, high=5, size=n)
x = np.repeat(np.array([1, 2, 3, 4]), n / 4)
exog[:, 2:] = get_dummies(x)
beta = np.array([-1, 0.1, -0.05, .2, 0.35])
lin_pred = (exog * beta).sum(axis=1)
family = sm.families.Binomial
link = sm.families.links.log
endog = gen_endog(lin_pred, family, link, binom_version=0)
wt = np.random.randint(1, 5, n)
mod1 = sm.GLM(endog, exog, family=family(link=link), freq_weights=wt)
cls.res1 = mod1.fit()
exog_dup = np.repeat(exog, wt, axis=0)
endog_dup = np.repeat(endog, wt)
mod2 = sm.GLM(endog_dup, exog_dup, family=family(link=link))
cls.res2 = mod2.fit()
def setup_class(cls):
from .results.results_glm import CancerLog
res2 = CancerLog
endog = res2.endog
exog = res2.exog[:, :-1]
exog = sm.add_constant(exog, prepend=True)
aweights = np.repeat(1, len(endog))
aweights[::5] = 5
aweights[::13] = 3
model = sm.GLM(endog, exog,
family=sm.families.Gamma(link=sm.families.links.log()),
var_weights=aweights)
cls.res1 = model.fit(rtol=1e-25, atol=0)
def setup(self):
# fit for each test, because results will be changed by test
x = self.exog
np.random.seed(987689)
y = x.sum(1) + np.random.randn(x.shape[0])
self.results = sm.GLM(y, self.exog).fit()
def test_gradient_irls():
# Compare the results when using gradient optimization and IRLS.
# TODO: Find working examples for inverse_squared link
np.random.seed(87342)
fams = [(sm.families.Binomial, [links.logit, links.probit, links.cloglog,
links.log, links.cauchy]),
(sm.families.Poisson, [links.log, links.identity, links.sqrt]),
(sm.families.Gamma, [links.log, links.identity,
links.inverse_power]),
(sm.families.Gaussian, [links.identity, links.log,
links.inverse_power]),
(sm.families.InverseGaussian, [links.log, links.identity,
links.inverse_power,
links.inverse_squared]),
(sm.families.NegativeBinomial, [links.log, links.inverse_power,
links.inverse_squared,
links.identity])]
n = 100
p = 3
exog = np.random.normal(size=(n, p))
exog[:, 0] = 1
skip_one = False
for family_class, family_links in fams:
for link in family_links:
for binom_version in [0, 1]:
if family_class != sm.families.Binomial and binom_version == 1:
continue
if (family_class, link) == (sm.families.Poisson,
links.identity):
lin_pred = 20 + exog.sum(1)
elif (family_class, link) == (sm.families.Binomial, links.log):
lin_pred = -1 + exog.sum(1) / 8
elif (family_class, link) == (sm.families.Poisson, links.sqrt):
lin_pred = 2 + exog.sum(1)
elif (family_class, link) == (sm.families.InverseGaussian,
links.log):
#skip_zero = True
lin_pred = -1 + exog.sum(1)
elif (family_class, link) == (sm.families.InverseGaussian,
links.identity):
lin_pred = 20 + 5 * exog.sum(1)
lin_pred = np.clip(lin_pred, 1e-4, np.inf)
elif (family_class, link) == (sm.families.InverseGaussian,
links.inverse_squared):
lin_pred = 0.5 + exog.sum(1) / 5
continue # skip due to non-convergence
elif (family_class, link) == (sm.families.InverseGaussian,
links.inverse_power):
lin_pred = 1 + exog.sum(1) / 5
elif (family_class, link) == (sm.families.NegativeBinomial,
links.identity):
lin_pred = 20 + 5 * exog.sum(1)
lin_pred = np.clip(lin_pred, 1e-4, np.inf)
elif (family_class, link) == (sm.families.NegativeBinomial,
links.inverse_squared):
lin_pred = 0.1 + np.random.uniform(size=exog.shape[0])
continue # skip due to non-convergence
elif (family_class, link) == (sm.families.NegativeBinomial,
links.inverse_power):
lin_pred = 1 + exog.sum(1) / 5
elif (family_class, link) == (sm.families.Gaussian,
links.inverse_power):
# adding skip because of convergence failure
skip_one = True
# GH#4620
# the following fails with identity link, because endog < 0
# elif family_class == fam.Gamma:
# lin_pred = (0.5 * exog.sum(1) +
# np.random.uniform(size=exog.shape[0]))
else:
lin_pred = np.random.uniform(size=exog.shape[0])
endog = gen_endog(lin_pred, family_class, link, binom_version)
with warnings.catch_warnings():
warnings.simplefilter("ignore")
mod_irls = sm.GLM(endog, exog,
family=family_class(link=link()))
rslt_irls = mod_irls.fit(method="IRLS")
if (family_class, link) not in [(sm.families.Poisson,
links.sqrt),
(sm.families.Gamma,
links.inverse_power),
(sm.families.InverseGaussian,
links.identity)]:
# GH#4620
check_score_hessian(rslt_irls)
# Try with and without starting values.
for max_start_irls, start_params in [(0, rslt_irls.params),
(3, None)]:
# TODO: skip convergence failures for now
if max_start_irls > 0 and skip_one:
continue
with warnings.catch_warnings():
warnings.simplefilter("ignore")
mod_gradient = sm.GLM(endog, exog,
family=family_class(link=link()))
rslt_gradient = mod_gradient.fit(
max_start_irls=max_start_irls,
start_params=start_params,
method="newton",
maxiter=300)
assert_allclose(rslt_gradient.params,
rslt_irls.params, rtol=1e-6, atol=5e-5)
assert_allclose(rslt_gradient.llf, rslt_irls.llf,
rtol=1e-6, atol=1e-6)
assert_allclose(rslt_gradient.scale, rslt_irls.scale,
rtol=1e-6, atol=1e-6)
# Get the standard errors using expected information.
gradient_bse = rslt_gradient.bse
ehess = mod_gradient.hessian(rslt_gradient.params,
observed=False)
gradient_bse = np.sqrt(-np.diag(np.linalg.inv(ehess)))
assert_allclose(gradient_bse, rslt_irls.bse,
rtol=1e-6, atol=5e-5)
def test_wtd_gradient_irls():
# Compare the results when using gradient optimization and IRLS.
# TODO: Find working examples for inverse_squared link
np.random.seed(87342)
fam = sm.families
lnk = sm.families.links
families = [(fam.Binomial, [lnk.logit, lnk.probit, lnk.cloglog, lnk.log,
lnk.cauchy]),
(fam.Poisson, [lnk.log, lnk.identity, lnk.sqrt]),
(fam.Gamma, [lnk.log, lnk.identity, lnk.inverse_power]),
(fam.Gaussian, [lnk.identity, lnk.log, lnk.inverse_power]),
(fam.InverseGaussian, [lnk.log, lnk.identity,
lnk.inverse_power,
lnk.inverse_squared]),
(fam.NegativeBinomial, [lnk.log, lnk.inverse_power,
lnk.inverse_squared, lnk.identity])]
n = 100
p = 3
exog = np.random.normal(size=(n, p))
exog[:, 0] = 1
skip_one = False
for family_class, family_links in families:
for link in family_links:
for binom_version in [0, 1]:
method = 'bfgs'
if family_class != fam.Binomial and binom_version == 1:
continue
elif family_class == fam.Binomial and link == lnk.cloglog:
# Can't get gradient to converage with var_weights here
continue
elif family_class == fam.Binomial and link == lnk.log:
# Can't get gradient to converage with var_weights here
continue
elif (family_class, link) == (fam.Poisson, lnk.identity):
lin_pred = 20 + exog.sum(1)
elif (family_class, link) == (fam.Binomial, lnk.log):
lin_pred = -1 + exog.sum(1) / 8
elif (family_class, link) == (fam.Poisson, lnk.sqrt):
lin_pred = -2 + exog.sum(1)
elif (family_class, link) == (fam.Gamma, lnk.log):
# Can't get gradient to converge with var_weights here
continue
elif (family_class, link) == (fam.Gamma, lnk.identity):
# Can't get gradient to converage with var_weights here
continue
elif (family_class, link) == (fam.Gamma, lnk.inverse_power):
# Can't get gradient to converage with var_weights here
continue
elif (family_class, link) == (fam.Gaussian, lnk.log):
# Can't get gradient to converage with var_weights here
continue
elif (family_class, link) == (fam.Gaussian, lnk.inverse_power):
# Can't get gradient to converage with var_weights here
continue
elif (family_class, link) == (fam.InverseGaussian, lnk.log):
# Can't get gradient to converage with var_weights here
lin_pred = -1 + exog.sum(1)
continue
elif (family_class, link) == (fam.InverseGaussian,
lnk.identity):
# Can't get gradient to converage with var_weights here
lin_pred = 20 + 5 * exog.sum(1)
lin_pred = np.clip(lin_pred, 1e-4, np.inf)
continue
elif (family_class, link) == (fam.InverseGaussian,
lnk.inverse_squared):
lin_pred = 0.5 + exog.sum(1) / 5
continue # skip due to non-convergence
elif (family_class, link) == (fam.InverseGaussian,
lnk.inverse_power):
lin_pred = 1 + exog.sum(1) / 5
method = 'newton'
elif (family_class, link) == (fam.NegativeBinomial,
lnk.identity):
lin_pred = 20 + 5 * exog.sum(1)
lin_pred = np.clip(lin_pred, 1e-3, np.inf)
method = 'newton'
elif (family_class, link) == (fam.NegativeBinomial,
lnk.inverse_squared):
lin_pred = 0.1 + np.random.uniform(size=exog.shape[0])
continue # skip due to non-convergence
elif (family_class, link) == (fam.NegativeBinomial,
lnk.inverse_power):
# Can't get gradient to converage with var_weights here
lin_pred = 1 + exog.sum(1) / 5
continue
elif (family_class, link) == (fam.Gaussian, lnk.inverse_power):
# adding skip because of convergence failure
skip_one = True
else:
lin_pred = np.random.uniform(size=exog.shape[0])
endog = gen_endog(lin_pred, family_class, link, binom_version)
if binom_version == 0:
wts = np.ones_like(endog)
tmp = np.random.randint(2,
5,
size=(endog > endog.mean()).sum())
wts[endog > endog.mean()] = tmp
else:
wts = np.ones(shape=endog.shape[0])
y = endog[:, 0] / endog.sum(axis=1)
tmp = np.random.gamma(2, size=(y > y.mean()).sum())
wts[y > y.mean()] = tmp
with warnings.catch_warnings():
warnings.simplefilter("ignore")
mod_irls = sm.GLM(endog, exog, var_weights=wts,
family=family_class(link=link()))
rslt_irls = mod_irls.fit(method="IRLS", atol=1e-10,
tol_criterion='params')
# Try with and without starting values.
for max_start_irls, start_params in ((0, rslt_irls.params),
(3, None)):
# TODO: skip convergence failures for now
if max_start_irls > 0 and skip_one:
continue
with warnings.catch_warnings():
warnings.simplefilter("ignore")
mod_gradient = sm.GLM(endog, exog, var_weights=wts,
family=family_class(link=link()))
rslt_gradient = mod_gradient.fit(
max_start_irls=max_start_irls,
start_params=start_params,
method=method)
assert_allclose(rslt_gradient.params,
rslt_irls.params,
rtol=1e-6, atol=5e-5)
assert_allclose(rslt_gradient.llf, rslt_irls.llf,
rtol=1e-6, atol=1e-6)
assert_allclose(rslt_gradient.scale, rslt_irls.scale,
rtol=1e-6, atol=1e-6)
# Get the standard errors using expected information.
gradient_bse = rslt_gradient.bse
ehess = mod_gradient.hessian(rslt_gradient.params,
observed=False)
gradient_bse = np.sqrt(-np.diag(np.linalg.inv(ehess)))
assert_allclose(gradient_bse, rslt_irls.bse,
rtol=1e-6, atol=5e-5)