def test_chisquare_prob(self, bin_edges=None, method=None):
"""Moment test for binned probabilites using OPG.
Paramters
---------
binedges : array_like or None
This defines which counts are included in the test on frequencies
and how counts are combined in bins.
The default if bin_edges is None will change in future.
See Notes and Example sections below.
method : str
Currently only `method = "opg"` is available.
If method is None, the OPG will be used, but the default might
change in future versions.
See Notes section below.
Returns
-------
test result
Notes
-----
Warning: The current default can have many empty or nearly empty bins.
The default number of bins is given by max(endog).
Currently it is recommended to limit the nuber of bins explicitly,
see Examples below.
Binning will change in future and automatic binning will be added.
Currently only the outer product of gradient, OPG, method is
implemented. In many case, the OPG version of a specification test
overrejects in small samples.
Specialized tests that use observed or expected information matrix
often have better small sample properties.
The default method will change if better methods are added.
Examples
--------
The following call is a test for the probability of zeros
`test_chisquare_prob(bin_edges=np.arange(3))`
`test_chisquare_prob(bin_edges=np.arange(10))` tests the hypothesis
that the frequences for counts up to 7 correspond to the estimated
Poisson distributions.
In this case, edges are 0, ..., 9 which defines 9 bins for
counts 0 to 8. The last bin is dropped, so the joint test hypothesis is
that the observed aggregated frequencies for counts 0 to 7 correspond
to the model prediction for those frequencies. Predicted probabilites
Prob(y_i = k | x) are aggregated over observations ``i``.
"""
kwds = {}
if bin_edges is not None:
# TODO: verify upper bound, we drop last bin (may be open, inf)
kwds["y_values"] = np.arange(bin_edges[-2] + 1)
probs = self.results.predict(which="prob", **kwds)
res = test_chisquare_prob(self.results,
probs,
bin_edges=bin_edges,
method=method)
return res
def test_probs(self, close_figures):
nobs = self.nobs
probs = self.res.predict_prob()
freq = np.bincount(self.endog) / nobs
tzi = dia.test_chisquare_prob(self.res, probs[:, :2])
# regression numbers
tzi1 = (0.387770845, 0.5334734738)
assert_allclose(tzi[:2], tzi1, rtol=5e-5)
# smoke test for plot
dia.plot_probs(freq, probs.mean(0))
def test_probs(self):
nobs = self.nobs
probs = self.res.predict_prob()
freq = np.bincount(self.endog) / nobs
tzi = dia.test_chisquare_prob(self.res, probs[:, :2])
# regression numbers
tzi1 = (0.387770845, 0.5334734738)
assert_allclose(tzi[:2], tzi1)
# smoke test for plot
try:
import matplotlib.pyplot as plt
except ImportError:
return
fig = dia.plot_probs(freq, probs.mean(0))
plt.close(fig)
def test_probs(self):
nobs = self.nobs
probs = self.res.predict_prob()
freq = np.bincount(self.endog) / nobs
tzi = dia.test_chisquare_prob(self.res, probs[:, :2])
# regression numbers
tzi1 = (0.387770845, 0.5334734738)
assert_allclose(tzi[:2], tzi1, rtol=5e-5)
# smoke test for plot
try:
import matplotlib.pyplot as plt
except ImportError:
return
fig = dia.plot_probs(freq, probs.mean(0))
plt.close(fig)
def test_spec_tests(self):
# regression test, numbers similar to Monte Carlo simulation
res_dispersion = np.array([[0.1396096387543, 0.8889684245877],
[0.1396096387543, 0.8889684245877],
[0.2977840351238, 0.7658680002106],
[0.1307899995877, 0.8959414342111],
[0.1307899995877, 0.8959414342111],
[0.1357101381056, 0.8920504328246],
[0.2776587511235, 0.7812743277372]])
res_zi = np.array([
[00.1389582826821, 0.7093188241734],
[-0.3727710861669, 0.7093188241734],
[-0.2496729648642, 0.8028402670888],
[00.0601651553909, 0.8062350958880],
])
respoi = Poisson(self.endog, self.exog).fit(disp=0)
dia = PoissonDiagnostic(respoi)
t_disp = dia.test_dispersion()[0]
assert_allclose(t_disp, res_dispersion, rtol=1e-8)
nobs = self.endog.shape[0]
t_zi_jh = dia.test_poisson_zeroinflation(method="broek",
exog_infl=np.ones(nobs))
t_zib = dia.test_poisson_zeroinflation(method="broek")
t_zim = dia.test_poisson_zeroinflation(method="prob")
t_zichi2 = dia.test_chisquare_prob(bin_edges=np.arange(3))
t_zi = np.vstack([t_zi_jh[:2], t_zib[:2], t_zim[:2], t_zichi2[:2]])
assert_allclose(t_zi, res_zi, rtol=1e-8)
# test jansakul and hinde with exog_infl
t_zi_ex = dia.test_poisson_zeroinflation(method="broek",
exog_infl=self.exog)
res_zi_ex = np.array([3.7813218150779, 0.1509719973257])
assert_allclose(t_zi_ex[:2], res_zi_ex, rtol=1e-8)
