def _constraints_factor(encoding_matrix,
comparison='pairwise',
k_params=None,
idx_start=None):
"""helper function to create constraints based on encoding matrix
Parameters
----------
encoding_matrix : ndarray
contrast matrix for the encoding of a factor as defined by patsy.
The number of rows should be equal to the number of levels or categories
of the factor, the number of columns should be equal to the number
of parameters for this factor.
comparison : str
Currently only 'pairwise' is implemented. The restriction matrix
can be used for testing the hypothesis that all pairwise differences
are zero.
k_params : int
number of parameters
idx_start : int
Index of the first parameter of this factor. The restrictions on the
factor are inserted as a block in the full restriction matrix starting
at column with index `idx_start`.
Returns
-------
contrast : ndarray
Contrast or restriction matrix that can be used in hypothesis test
of model results. The number of columns is k_params.
"""
cm = encoding_matrix
k_level, k_p = cm.shape
import statsmodels.sandbox.stats.multicomp as mc
if comparison in ['pairwise', 'pw', 'pairs']:
c_all = -mc.contrast_allpairs(k_level)
else:
raise NotImplementedError('currentlyonly pairwise comparison')
contrasts = c_all.dot(cm)
if k_params is not None:
if idx_start is None:
raise ValueError("if k_params is not None, then idx_start is "
"required")
contrasts = _embed_constraints(contrasts, k_params, idx_start)
return contrasts
def _constraints_factor(encoding_matrix, comparison='pairwise', k_params=None,
idx_start=None):
"""helper function to create constraints based on encoding matrix
Parameters
----------
encoding_matrix : ndarray
contrast matrix for the encoding of a factor as defined by patsy.
The number of rows should be equal to the number of levels or categories
of the factor, the number of columns should be equal to the number
of parameters for this factor.
comparison : str
Currently only 'pairwise' is implemented. The restriction matrix
can be used for testing the hypothesis that all pairwise differences
are zero.
k_params : int
number of parameters
idx_start : int
Index of the first parameter of this factor. The restrictions on the
factor are inserted as a block in the full restriction matrix starting
at column with index `idx_start`.
Returns
-------
contrast : ndarray
Contrast or restriction matrix that can be used in hypothesis test
of model results. The number of columns is k_params.
"""
cm = encoding_matrix
k_level, k_p = cm.shape
import statsmodels.sandbox.stats.multicomp as mc
if comparison in ['pairwise', 'pw', 'pairs']:
c_all = -mc.contrast_allpairs(k_level)
else:
raise NotImplementedError('currentlyonly pairwise comparison')
contrasts = c_all.dot(cm)
if k_params is not None:
if idx_start is None:
raise ValueError("if k_params is not None, then idx_start is "
"required")
contrasts = _embed_constraints(contrasts, k_params, idx_start)
return contrasts
def t_test_pairwise(result, term_name, method='hs', alpha=0.05,
factor_labels=None, ignore=False):
"""perform pairwise t_test with multiple testing corrected p-values
This uses the formula design_info encoding contrast matrix and should
work for all encodings of a main effect.
Parameters
----------
result : result instance
The results of an estimated model with a categorical main effect.
term_name : str
name of the term for which pairwise comparisons are computed.
Term names for categorical effects are created by patsy and
correspond to the main part of the exog names.
method : str or list of strings
multiple testing p-value correction, default is 'hs',
see stats.multipletesting
alpha : float
significance level for multiple testing reject decision.
factor_labels : None, list of str
Labels for the factor levels used for pairwise labels. If not
provided, then the labels from the formula design_info are used.
ignore : boolean
Turn off some of the exceptions raised by input checks.
Returns
-------
results : instance of a simple Results class
The results are stored as attributes, the main attributes are the
following two. Other attributes are added for debugging purposes
or as background information.
- result_frame : pandas DataFrame with t_test results and multiple
testing corrected p-values.
- contrasts : matrix of constraints of the null hypothesis in the
t_test.
Notes
-----
Status: experimental. Currently only checked for treatment coding with
and without specified reference level.
Currently there are no multiple testing corrected confidence intervals
available.
"""
desinfo = result.model.data.design_info
term_idx = desinfo.term_names.index(term_name)
term = desinfo.terms[term_idx]
idx_start = desinfo.term_slices[term].start
if not ignore and len(term.factors) > 1:
raise ValueError('interaction effects not yet supported')
factor = term.factors[0]
cat = desinfo.factor_infos[factor].categories
if factor_labels is not None:
if len(factor_labels) == len(cat):
cat = factor_labels
else:
raise ValueError("factor_labels has the wrong length, should be %d" % len(cat))
k_level = len(cat)
cm = desinfo.term_codings[term][0].contrast_matrices[factor].matrix
k_params = len(result.params)
labels = _get_pairs_labels(k_level, cat)
import statsmodels.sandbox.stats.multicomp as mc
c_all_pairs = -mc.contrast_allpairs(k_level)
contrasts_sub = c_all_pairs.dot(cm)
contrasts = _embed_constraints(contrasts_sub, k_params, idx_start)
res_df = t_test_multi(result, contrasts, method=method, ci_method=None,
alpha=alpha, contrast_names=labels)
res = MultiCompResult(result_frame=res_df,
contrasts=contrasts,
term=term,
contrast_labels=labels,
term_encoding_matrix=cm)
return res
def t_test_pairwise(result,
term_name,
method='hs',
alpha=0.05,
factor_labels=None,
ignore=False):
"""
Perform pairwise t_test with multiple testing corrected p-values
This uses the formula design_info encoding contrast matrix and should
work for all encodings of a main effect.
Parameters
----------
result : result instance
The results of an estimated model with a categorical main effect.
term_name : str
name of the term for which pairwise comparisons are computed.
Term names for categorical effects are created by patsy and
correspond to the main part of the exog names.
method : {str, list[str]}
multiple testing p-value correction, default is 'hs',
see stats.multipletesting
alpha : float
significance level for multiple testing reject decision.
factor_labels : {list[str], None}
Labels for the factor levels used for pairwise labels. If not
provided, then the labels from the formula design_info are used.
ignore : bool
Turn off some of the exceptions raised by input checks.
Returns
-------
MultiCompResult
The results are stored as attributes, the main attributes are the
following two. Other attributes are added for debugging purposes
or as background information.
- result_frame : pandas DataFrame with t_test results and multiple
testing corrected p-values.
- contrasts : matrix of constraints of the null hypothesis in the
t_test.
Notes
-----
Status: experimental. Currently only checked for treatment coding with
and without specified reference level.
Currently there are no multiple testing corrected confidence intervals
available.
"""
desinfo = result.model.data.design_info
term_idx = desinfo.term_names.index(term_name)
term = desinfo.terms[term_idx]
idx_start = desinfo.term_slices[term].start
if not ignore and len(term.factors) > 1:
raise ValueError('interaction effects not yet supported')
factor = term.factors[0]
cat = desinfo.factor_infos[factor].categories
if factor_labels is not None:
if len(factor_labels) == len(cat):
cat = factor_labels
else:
raise ValueError(
"factor_labels has the wrong length, should be %d" % len(cat))
k_level = len(cat)
cm = desinfo.term_codings[term][0].contrast_matrices[factor].matrix
k_params = len(result.params)
labels = _get_pairs_labels(k_level, cat)
import statsmodels.sandbox.stats.multicomp as mc
c_all_pairs = -mc.contrast_allpairs(k_level)
contrasts_sub = c_all_pairs.dot(cm)
contrasts = _embed_constraints(contrasts_sub, k_params, idx_start)
res_df = t_test_multi(result,
contrasts,
method=method,
ci_method=None,
alpha=alpha,
contrast_names=labels)
res = MultiCompResult(result_frame=res_df,
contrasts=contrasts,
term=term,
contrast_labels=labels,
term_encoding_matrix=cm)
return res
