def gen_endog(lin_pred, family_class, link, binom_version=0):
np.random.seed(872)
fam = sm.families
mu = link().inverse(lin_pred)
if family_class == fam.Binomial:
if binom_version == 0:
endog = 1 * (np.random.uniform(size=len(lin_pred)) < mu)
else:
endog = np.empty((len(lin_pred), 2))
n = 10
endog[:, 0] = (np.random.uniform(size=(len(lin_pred), n)) <
mu[:, None]).sum(1)
endog[:, 1] = n - endog[:, 0]
elif family_class == fam.Poisson:
endog = np.random.poisson(mu)
elif family_class == fam.Gamma:
endog = np.random.gamma(2, mu)
elif family_class == fam.Gaussian:
endog = mu + np.random.normal(size=len(lin_pred))
elif family_class == fam.NegativeBinomial:
from scipy.stats.distributions import nbinom
endog = nbinom.rvs(mu, 0.5)
elif family_class == fam.InverseGaussian:
from scipy.stats.distributions import invgauss
endog = invgauss.rvs(mu)
else:
raise ValueError
return endog
def gen_endog(lin_pred, family_name):
if family_name == "binomial":
probs = 1 / (1 + np.exp(-lin_pred))
endog = 1 * (np.random.uniform(size=len(lin_pred)) < probs)
elif family_name == "poisson":
lam = np.exp(lin_pred)
endog = np.random.poisson(lam)
elif family_name == "gamma":
lam = np.exp(lin_pred / 4)
endog = np.random.gamma(2, lam)
elif family_name == "gaussian":
endog = lin_pred + np.random.normal(size=len(lin_pred))
elif family_name == "negbinom":
from scipy.stats.distributions import nbinom
lam = np.exp(lin_pred)
endog = nbinom.rvs(lam, 0.5)
elif family_name == "inverse_gaussian":
from scipy.stats.distributions import invgauss
lin_pred = 2 + (lin_pred - 2) / 4.
lam = 1 / np.sqrt(lin_pred)
endog = invgauss.rvs(lam)
else:
raise ValueError
return endog
def gen_endog(lin_pred, family_class, link, binom_version=0):
np.random.seed(872)
fam = sm.families
mu = link().inverse(lin_pred)
if family_class == fam.Binomial:
if binom_version == 0:
endog = 1*(np.random.uniform(size=len(lin_pred)) < mu)
else:
endog = np.empty((len(lin_pred), 2))
n = 10
endog[:, 0] = (np.random.uniform(size=(len(lin_pred), n)) < mu[:, None]).sum(1)
endog[:, 1] = n - endog[:, 0]
elif family_class == fam.Poisson:
endog = np.random.poisson(mu)
elif family_class == fam.Gamma:
endog = np.random.gamma(2, mu)
elif family_class == fam.Gaussian:
endog = mu + np.random.normal(size=len(lin_pred))
elif family_class == fam.NegativeBinomial:
from scipy.stats.distributions import nbinom
endog = nbinom.rvs(mu, 0.5)
elif family_class == fam.InverseGaussian:
from scipy.stats.distributions import invgauss
endog = invgauss.rvs(mu)
else:
raise ValueError
return endog
def gen_endog(lin_pred, family_class, link, binom_version=0):
np.random.seed(872)
mu = link().inverse(lin_pred)
if family_class == sm.families.Binomial:
if binom_version == 0:
endog = 1 * (np.random.uniform(size=len(lin_pred)) < mu)
else:
endog = np.empty((len(lin_pred), 2))
n = 10
uni = np.random.uniform(size=(len(lin_pred), n))
endog[:, 0] = (uni < mu[:, None]).sum(1)
endog[:, 1] = n - endog[:, 0]
elif family_class == sm.families.Poisson:
endog = np.random.poisson(mu)
elif family_class == sm.families.Gamma:
endog = np.random.gamma(2, mu)
elif family_class == sm.families.Gaussian:
endog = mu + np.random.normal(size=len(lin_pred))
elif family_class == sm.families.NegativeBinomial:
from scipy.stats.distributions import nbinom
endog = nbinom.rvs(mu, 0.5)
elif family_class == sm.families.InverseGaussian:
from scipy.stats.distributions import invgauss
endog = invgauss.rvs(mu)
elif family_class == sm.families.Tweedie:
# upstream this case wasn't present in test_glm, but there was an
# otherwise identical gen_endog function in test_glm_weights
rate = 1
shape = 1.0
scale = mu / (rate * shape)
endog = (np.random.poisson(rate, size=scale.shape[0]) *
np.random.gamma(shape * scale))
else:
raise ValueError
return endog
