# -*- coding: utf-8 -*-
"""
Created on Fri Dec 19 11:29:18 2014
Author: Josef Perktold
License: BSD-3
"""
import numpy as np
from scipy import stats
# this is similar to ContrastResults after t_test, partially copied and adjusted
[docs]class PredictionResults(object):
def __init__(self, predicted_mean, var_pred_mean, var_resid=None,
df=None, dist=None, row_labels=None, linpred=None, link=None):
# TODO: is var_resid used? drop from arguments?
self.predicted_mean = predicted_mean
self.var_pred_mean = var_pred_mean
self.df = df
self.var_resid = var_resid
self.row_labels = row_labels
self.linpred = linpred
self.link = link
if dist is None or dist == 'norm':
self.dist = stats.norm
self.dist_args = ()
elif dist == 't':
self.dist = stats.t
self.dist_args = (self.df,)
else:
self.dist = dist
self.dist_args = ()
@property
def se_obs(self):
raise NotImplementedError
return np.sqrt(self.var_pred_mean + self.var_resid)
@property
def se_mean(self):
return np.sqrt(self.var_pred_mean)
@property
def tvalues(self):
return self.predicted_mean / self.se_mean
[docs] def t_test(self, value=0, alternative='two-sided'):
'''z- or t-test for hypothesis that mean is equal to value
Parameters
----------
value : array_like
value under the null hypothesis
alternative : string
'two-sided', 'larger', 'smaller'
Returns
-------
stat : ndarray
test statistic
pvalue : ndarray
p-value of the hypothesis test, the distribution is given by
the attribute of the instance, specified in `__init__`. Default
if not specified is the normal distribution.
'''
# from statsmodels.stats.weightstats
# assumes symmetric distribution
stat = (self.predicted_mean - value) / self.se_mean
if alternative in ['two-sided', '2-sided', '2s']:
pvalue = self.dist.sf(np.abs(stat), *self.dist_args)*2
elif alternative in ['larger', 'l']:
pvalue = self.dist.sf(stat, *self.dist_args)
elif alternative in ['smaller', 's']:
pvalue = self.dist.cdf(stat, *self.dist_args)
else:
raise ValueError('invalid alternative')
return stat, pvalue
[docs] def conf_int(self, method='endpoint', alpha=0.05, **kwds):
"""
Returns the confidence interval of the value, `effect` of the constraint.
This is currently only available for t and z tests.
Parameters
----------
alpha : float, optional
The significance level for the confidence interval.
ie., The default `alpha` = .05 returns a 95% confidence interval.
kwds : extra keyword arguments
currently ignored, only for compatibility, consistent signature
Returns
-------
ci : ndarray, (k_constraints, 2)
The array has the lower and the upper limit of the confidence
interval in the columns.
"""
tmp = np.linspace(0, 1, 6)
is_linear = (self.link.inverse(tmp) == tmp).all()
if method == 'endpoint' and not is_linear:
ci_linear = self.linpred.conf_int(alpha=alpha, obs=False)
ci = self.link.inverse(ci_linear)
elif method == 'delta' or is_linear:
se = self.se_mean
q = self.dist.ppf(1 - alpha / 2., *self.dist_args)
lower = self.predicted_mean - q * se
upper = self.predicted_mean + q * se
ci = np.column_stack((lower, upper))
# if we want to stack at a new last axis, for lower.ndim > 1
# np.concatenate((lower[..., None], upper[..., None]), axis=-1)
return ci
[docs] def summary_frame(self, what='all', alpha=0.05):
# TODO: finish and cleanup
import pandas as pd
from statsmodels.compat.collections import OrderedDict
#ci_obs = self.conf_int(alpha=alpha, obs=True) # need to split
ci_mean = self.conf_int(alpha=alpha)
to_include = OrderedDict()
to_include['mean'] = self.predicted_mean
to_include['mean_se'] = self.se_mean
to_include['mean_ci_lower'] = ci_mean[:, 0]
to_include['mean_ci_upper'] = ci_mean[:, 1]
self.table = to_include
#OrderedDict doesn't work to preserve sequence
# pandas dict doesn't handle 2d_array
#data = np.column_stack(list(to_include.values()))
#names = ....
res = pd.DataFrame(to_include, index=self.row_labels,
columns=to_include.keys())
return res
def get_prediction_glm(self, exog=None, transform=True, weights=None,
row_labels=None, linpred=None, link=None, pred_kwds=None):
"""
compute prediction results
Parameters
----------
exog : array-like, optional
The values for which you want to predict.
transform : bool, optional
If the model was fit via a formula, do you want to pass
exog through the formula. Default is True. E.g., if you fit
a model y ~ log(x1) + log(x2), and transform is True, then
you can pass a data structure that contains x1 and x2 in
their original form. Otherwise, you'd need to log the data
first.
weights : array_like, optional
Weights interpreted as in WLS, used for the variance of the predicted
residual.
args, kwargs :
Some models can take additional arguments or keywords, see the
predict method of the model for the details.
Returns
-------
prediction_results : generalized_linear_model.PredictionResults
The prediction results instance contains prediction and prediction
variance and can on demand calculate confidence intervals and summary
tables for the prediction of the mean and of new observations.
"""
### prepare exog and row_labels, based on base Results.predict
if transform and hasattr(self.model, 'formula') and exog is not None:
from patsy import dmatrix
exog = dmatrix(self.model.data.design_info.builder,
exog)
if exog is not None:
if row_labels is None:
row_labels = getattr(exog, 'index', None)
if callable(row_labels):
row_labels = None
exog = np.asarray(exog)
if exog.ndim == 1 and (self.model.exog.ndim == 1 or
self.model.exog.shape[1] == 1):
exog = exog[:, None]
exog = np.atleast_2d(exog) # needed in count model shape[1]
else:
exog = self.model.exog
if weights is None:
weights = getattr(self.model, 'weights', None)
if row_labels is None:
row_labels = getattr(self.model.data, 'row_labels', None)
# need to handle other arrays, TODO: is delegating to model possible ?
if weights is not None:
weights = np.asarray(weights)
if (weights.size > 1 and
(weights.ndim != 1 or weights.shape[0] == exog.shape[1])):
raise ValueError('weights has wrong shape')
### end
pred_kwds['linear'] = False
predicted_mean = self.model.predict(self.params, exog, **pred_kwds)
covb = self.cov_params()
link_deriv = self.model.family.link.inverse_deriv(linpred.predicted_mean)
var_pred_mean = link_deriv**2 * (exog * np.dot(covb, exog.T).T).sum(1)
var_resid = self.scale # self.mse_resid / weights
# TODO: check that we have correct scale, Refactor scale #???
# special case for now:
if self.cov_type == 'fixed scale':
var_resid = self.cov_kwds['scale']
if weights is not None:
var_resid /= weights
dist = ['norm', 't'][self.use_t]
return PredictionResults(predicted_mean, var_pred_mean, var_resid,
df=self.df_resid, dist=dist,
row_labels=row_labels, linpred=linpred, link=link)
def params_transform_univariate(params, cov_params, link=None, transform=None,
row_labels=None):
"""
results for univariate, nonlinear, monotonicaly transformed parameters
This provides transformed values, standard errors and confidence interval
for transformations of parameters, for example in calculating rates with
`exp(params)` in the case of Poisson or other models with exponential
mean function.
"""
from statsmodels.genmod.families import links
if link is None and transform is None:
link = links.Log()
if row_labels is None and hasattr(params, 'index'):
row_labels = params.index
params = np.asarray(params)
predicted_mean = link.inverse(params)
link_deriv = link.inverse_deriv(params)
var_pred_mean = link_deriv**2 * np.diag(cov_params)
# TODO: do we want covariance also, or just var/se
dist = stats.norm
# TODO: need ci for linear prediction, method of `lin_pred
linpred = PredictionResults(params, np.diag(cov_params), dist=dist,
row_labels=row_labels, link=links.identity())
res = PredictionResults(predicted_mean, var_pred_mean, dist=dist,
row_labels=row_labels, linpred=linpred, link=link)
return res
