def __predict__(df, model, variable, formula):
"""UNUSED FUNCTION
New hidden predict method to shorten Monte Carlo estimation code. This is slower than statsmodels predict,
but likely can be optimized. I would need to find a faster alternative to np.dot()...
For this to work, I need to do the following:
-need to have each ..._model() store a self.formula
-switch all functions to sm.GLM or sm.GEE
"""
import patsy
from zepid.calc import odds_to_probability
xdata = patsy.dmatrix(formula, df) # , return_type='dataframe')
# pred = xdata.mul(np.array(model.params), axis='columns').sum(axis=1)
pred = xdata.dot(model.params) # TODO optimize this...
if variable == 'binary':
pred = np.random.binomial(1,
odds_to_probability(np.exp(pred)),
size=xdata.shape[0])
elif variable == 'continuous':
pred = np.random.normal(loc=pred,
scale=np.std(model.resid),
size=len(pp))
# TODO add optimization for multinomial (if applicable)
else:
raise ValueError('That option is not supported')
return pred
def test_back_and_forth_conversions(self):
original = 0.12
odd = probability_to_odds(original)
pr = odds_to_probability(odd)
npt.assert_allclose(original, pr)
def test_forth_and_back_conversions(self):
original = 1.1
pr = odds_to_probability(original)
odd = probability_to_odds(pr)
npt.assert_allclose(original, odd)
def test_odds_to_probability(self):
pr = odds_to_probability(1.1)
npt.assert_allclose(pr, 1.1 / 2.1)