"""
Implementation of class contaning and processing the estimation results.
:author: Michel Bierlaire
:date: Tue Mar 26 16:50:01 2019
.. todo:: rawResults should be a dict and not a class.
"""
# Too constraining
# pylint: disable=invalid-name,
# pylint: disable=too-many-instance-attributes, too-many-locals
# pylint: disable=too-many-branches
# pylint: disable=too-many-statements, too-few-public-methods
# pylint: disable=too-many-lines
#
import pickle
import datetime
from collections import namedtuple
import urllib.request as urlr
import pandas as pd
import numpy as np
from scipy import linalg
from scipy import stats
import biogeme.version as bv
import biogeme.filenames as bf
import biogeme.exceptions as excep
import biogeme.messaging as msg
from biogeme import tools
logger = msg.bioMessage()
GeneralStatistic = namedtuple('GeneralStatistic', 'value format')
[docs]def calcPValue(t):
"""Calculates the p value of a parameter from its t-statistic.
The formula is
.. math:: 2(1-\\Phi(|t|)
where :math:`\\Phi(\\cdot)` is the CDF of a normal distribution.
:param t: t-statistics
:type t: float
:return: p-value
:rtype: float
"""
p = 2.0 * (1.0 - stats.norm.cdf(abs(t)))
return p
[docs]class beta:
"""Class gathering the information related to the parameters
of the model
"""
[docs] def __init__(self, name, value, bounds):
"""
Constructor
:param name: name of the parameter.
:type name: string
:param value: value of the parameter.
:type value: float
:param bounds: tuple (l,b) with lower and upper bounds
:type bounds: float,float
"""
self.name = name #: Name of the parameter
self.value = value #: Current value
self.lb = bounds[0] #: Lower bound
self.ub = bounds[1] #: Upper bound
self.stdErr = None #: Standard error
self.tTest = None #: t-test
self.pValue = None #: p-value
self.robust_stdErr = None #: Robust standard error
self.robust_tTest = None #: Robust t-test
self.robust_pValue = None #: Robust p-value
self.bootstrap_stdErr = None #: Std error calculated from bootstrap
self.bootstrap_tTest = None #: t-test calculated from bootstrap
self.bootstrap_pValue = None #: p-value calculated from bootstrap
[docs] def isBoundActive(self, threshold=1.0e-6):
"""Check if one of the two bound is 'numerically' active. Being
numerically active means that the distance between the value of
the parameter and one of its bounds is below the threshold.
:param threshold: distance below which the bound is considered to be
active. Default: :math:`10^{-6}`
:type threshold: float
:return: True is one of the two bounds is numericall y active.
:rtype: bool
:raise biogemeError: if ``threshold`` is negative.
"""
if threshold < 0:
raise excep.biogemeError(
f'Threshold ({threshold}) must be non negative'
)
if self.lb is not None and np.abs(self.value - self.lb) <= threshold:
return True
if self.ub is not None and np.abs(self.value - self.ub) <= threshold:
return True
return False
[docs] def setStdErr(self, se):
"""Records the standard error, and calculates and records
the corresponding t-statistic and p-value
:param se: standard error.
:type se: float
"""
self.stdErr = se
if se == 0:
self.tTest = np.finfo(float).max
else:
self.tTest = np.nan_to_num(self.value / se)
self.pValue = calcPValue(self.tTest)
[docs] def setRobustStdErr(self, se):
"""Records the robust standard error, and calculates and records
the corresponding t-statistic and p-value
:param se: robust standard error
:type se: float
"""
self.robust_stdErr = se
if se == 0:
self.robust_tTest = np.finfo(float).max
else:
self.robust_tTest = np.nan_to_num(self.value / se)
self.robust_pValue = calcPValue(self.robust_tTest)
[docs] def setBootstrapStdErr(self, se):
"""Records the robust standard error calculated by bootstrap, and
calculates and records the corresponding t-statistic and p-value
:param se: standard error calculated by bootstrap.
:type se: float
"""
self.bootstrap_stdErr = se
if se == 0:
self.bootstrap_tTest = np.finfo(float).max
else:
self.bootstrap_tTest = np.nan_to_num(self.value / se)
self.bootstrap_pValue = calcPValue(self.robust_tTest)
def __str__(self):
s = f'{self.name:15}: {self.value:.3g}'
if self.stdErr is not None:
s += f'[{self.stdErr:.3g} {self.tTest:.3g} {self.pValue:.3g}]'
if self.robust_stdErr is not None:
s += (
f'[{self.robust_stdErr:.3g} {self.robust_tTest:.3g} '
f'{self.robust_pValue:.3g}]'
)
if self.bootstrap_stdErr is not None:
s += (
f'[{self.bootstrap_stdErr:.3g} {self.bootstrap_tTest:.3g} '
f'{self.bootstrap_pValue:.3g}]'
)
return s
[docs]class rawResults:
"""Class containing the raw results from the estimation"""
[docs] def __init__(self, theModel, betaValues, fgHb, bootstrap=None):
"""
Constructor
:param theModel: object with the model
:type theModel: biogeme.BIOGEME
:param betaValues: list containing the estimated values of the
parameters
:type betaValues: list(float)
:param fgHb: tuple f,g,H,bhhh containing
- f: the value of the function,
- g: the gradient,
- H: the second derivative matrix,
- bhhh: the BHHH matrix.
:type fgHb: float,numpy.array, numpy.array, numpy.array
:param bootstrap: output of the bootstrapping. numpy array, of
size B x K, where
- B is the number of bootstrap iterations
- K is the number of parameters to estimate
Default: None.
:type bootstrap: numpy.array
"""
self.modelName = theModel.modelName #: Name of the model
self.userNotes = theModel.userNotes #: User notes
self.nparam = len(betaValues) #: Number of parameters
self.betaValues = betaValues #: Values of the parameters
self.betaNames = (
theModel.id_manager.free_betas.names
) #: Names of the parameters
self.initLogLike = theModel.initLogLike
"""Value of the likelihood function with the initial value of the
parameters
"""
self.nullLogLike = theModel.nullLogLike
"""Value of the likelihood function with equal probability model
"""
self.betas = [] #: List of objects of type results.beta
for b, n in zip(betaValues, self.betaNames):
bounds = theModel.getBoundsOnBeta(n)
self.betas.append(beta(n, b, bounds))
self.logLike = fgHb[0] #: Value of the loglikelihood function
self.g = fgHb[1] #: Value of the gradient of the loglik. function
self.H = fgHb[2] #: Value of the hessian of the loglik. function
self.bhhh = fgHb[3]
"""Value of the BHHH matrix of the loglikelihood function"""
self.dataname = theModel.database.name #: Name of the database
self.sampleSize = theModel.database.getSampleSize()
"""Sample size (number of individuals if panel data)"""
self.numberOfObservations = theModel.database.getNumberOfObservations()
"""Number of observations"""
self.monteCarlo = theModel.monteCarlo
"""True if the model involved Monte Carlo integration"""
self.numberOfDraws = theModel.numberOfDraws
"""Number of draws for Monte Carlo integration"""
self.typesOfDraws = theModel.database.typesOfDraws
"""Types of draws for Monte Carlo integration"""
self.excludedData = theModel.database.excludedData
"""Number of excluded data"""
self.drawsProcessingTime = theModel.drawsProcessingTime
"""Time needed to process the draws"""
self.gradientNorm = linalg.norm(self.g) if self.g is not None else None
"""Norm of the gradient"""
self.optimizationMessages = theModel.optimizationMessages
"""Diagnostics given by the optimization algorithm"""
self.numberOfThreads = theModel.numberOfThreads
"""Number of threads used for parallel computing"""
self.htmlFileName = None #: Name of the HTML output file
self.F12FileName = None #: Name of the F12 output file
self.latexFileName = None #: Name of the LaTeX output file
self.pickleFileName = None #: Name of the pickle outpt file
self.bootstrap = bootstrap
"""output of the bootstrapping. numpy array, of size B x K,
where
- B is the number of bootstrap iterations
- K is the number of parameters to estimate
"""
if bootstrap is not None:
self.bootstrap_time = theModel.bootstrap_time
""" Time needed to perform the bootstrap"""
self.secondOrderTable = None #: Second order statistics
[docs]class bioResults:
"""Class managing the estimation results"""
[docs] def __init__(self, theRawResults=None, pickleFile=None):
"""Constructor
:param theRawResults: object with the results of the estimation.
Default: None.
:type theRawResults: biogeme.results.rawResults
:param pickleFile: name of the file containing the raw results in
pickle format. It can be a URL. Default: None.
:type pickleFile: string
:raise biogeme.exceptions.biogemeError: if no data is provided.
"""
if theRawResults is not None:
self.data = theRawResults
"""Object of type :class:`biogeme.results.rawResults` contaning the
raw estimation results.
"""
elif pickleFile is not None:
try:
with urlr.urlopen(pickleFile) as p:
self.data = pickle.load(p)
except Exception:
pass
try:
with open(pickleFile, 'rb') as f:
self.data = pickle.load(f)
except FileNotFoundError as e:
error_msg = f'File {pickleFile} not found'
raise excep.biogemeError(error_msg) from e
else:
raise excep.biogemeError('No data provided.')
self._calculateStats()
[docs] def writePickle(self):
"""Dump the data in a file in pickle format.
:return: name of the file.
:rtype: string
"""
self.data.pickleFileName = bf.getNewFileName(
self.data.modelName, 'pickle'
)
with open(self.data.pickleFileName, 'wb') as f:
pickle.dump(self.data, f)
logger.general(f'Results saved in file {self.data.pickleFileName}')
return self.data.pickleFileName
def _calculateTest(self, i, j, matrix):
"""Calculates a t-test comparing two coefficients
Args:
i: index of first coefficient \f$\\beta_i\f$.
j: index of second coefficient \f$\\beta_i\f$.
matrix: estimate of the variance-covariance matrix \f$m\f$.
:return: t test
..math:: \f[\\frac{\\beta_i-\\beta_j}
{\\sqrt{m_{ii}+m_{jj} - 2 m_{ij} }}\f]
:rtype: float
"""
vi = self.data.betaValues[i]
vj = self.data.betaValues[j]
varI = matrix[i, i]
varJ = matrix[j, j]
covar = matrix[i, j]
r = varI + varJ - 2.0 * covar
if r <= 0:
test = np.finfo(float).max
else:
test = (vi - vj) / np.sqrt(r)
return test
def _calculateStats(self):
"""Calculates the following statistics:
- likelihood ratio test between the initial and the estimated
models: :math:`-2(L_0-L^*)`
- Rho square: :math:`1 - \\frac{L^*}{L^0}`
- Rho bar square: :math:`1 - \\frac{L^* - K}{L^0}`
- AIC: :math:`2(K - L^*)`
- BIC: :math:`-2 L^* + K \\log(N)`
Estimates for the variance-covariance matrix (Rao-Cramer,
robust, and bootstrap) are also calculated, as well as t-tests and
p value for the comparison of pairs of coefficients.
"""
self.data.likelihoodRatioTestNull = (
-2.0 * (self.data.nullLogLike - self.data.logLike)
if self.data.nullLogLike is not None
else None
)
self.data.likelihoodRatioTest = (
-2.0 * (self.data.initLogLike - self.data.logLike)
if self.data.initLogLike is not None
else None
)
try:
self.data.rhoSquare = (
np.nan_to_num(1.0 - self.data.logLike / self.data.initLogLike)
if self.data.initLogLike is not None
else None
)
except ZeroDivisionError:
self.data.rhoSquare = None
try:
self.data.rhoSquareNull = (
np.nan_to_num(1.0 - self.data.logLike / self.data.nullLogLike)
if self.data.nullLogLike is not None
else None
)
except ZeroDivisionError:
self.data.rhoSquareNull = None
try:
self.data.rhoBarSquare = (
np.nan_to_num(
1.0
- (self.data.logLike - self.data.nparam)
/ self.data.initLogLike
)
if self.data.initLogLike is not None
else None
)
except ZeroDivisionError:
self.data.rhoBarSquare = None
try:
self.data.rhoBarSquareNull = (
np.nan_to_num(
1.0
- (self.data.logLike - self.data.nparam)
/ self.data.nullLogLike
)
if self.data.nullLogLike is not None
else None
)
except ZeroDivisionError:
self.data.rhoBarSquareNull = None
self.data.akaike = 2.0 * self.data.nparam - 2.0 * self.data.logLike
self.data.bayesian = (
-2.0 * self.data.logLike
+ self.data.nparam * np.log(self.data.sampleSize)
)
# We calculate the eigenstructure to report in case of singularity
if self.data.H is not None:
self.data.eigenValues, self.data.eigenVectors = linalg.eigh(
-np.nan_to_num(self.data.H)
)
_, self.data.singularValues, _ = linalg.svd(
-np.nan_to_num(self.data.H)
)
# We use the pseudo inverse in case the matrix is singular
self.data.varCovar = -linalg.pinv(np.nan_to_num(self.data.H))
for i in range(self.data.nparam):
if self.data.varCovar[i, i] < 0:
self.data.betas[i].setStdErr(np.finfo(float).max)
else:
self.data.betas[i].setStdErr(
np.sqrt(self.data.varCovar[i, i])
)
d = np.diag(self.data.varCovar)
if (d > 0).all():
diag = np.diag(np.sqrt(d))
diagInv = linalg.inv(diag)
self.data.correlation = diagInv.dot(
self.data.varCovar.dot(diagInv)
)
else:
self.data.correlation = np.full_like(
self.data.varCovar, np.finfo(float).max
)
# Robust estimator
self.data.robust_varCovar = self.data.varCovar.dot(
self.data.bhhh.dot(self.data.varCovar)
)
for i in range(self.data.nparam):
if self.data.robust_varCovar[i, i] < 0:
self.data.betas[i].setRobustStdErr(np.finfo(float).max)
else:
self.data.betas[i].setRobustStdErr(
np.sqrt(self.data.robust_varCovar[i, i])
)
rd = np.diag(self.data.robust_varCovar)
if (rd > 0).all():
diag = np.diag(np.sqrt(rd))
diagInv = linalg.inv(diag)
self.data.robust_correlation = diagInv.dot(
self.data.robust_varCovar.dot(diagInv)
)
else:
self.data.robust_correlation = np.full_like(
self.data.robust_varCovar, np.finfo(float).max
)
# Bootstrap
if self.data.bootstrap is not None:
self.data.bootstrap_varCovar = np.cov(
self.data.bootstrap, rowvar=False
)
for i in range(self.data.nparam):
if self.data.bootstrap_varCovar[i, i] < 0:
self.data.betas[i].setBootstrapStdErr(
np.finfo(float).max
)
else:
self.data.betas[i].setBootstrapStdErr(
np.sqrt(self.data.bootstrap_varCovar[i, i])
)
rd = np.diag(self.data.bootstrap_varCovar)
if (rd > 0).all():
diag = np.diag(np.sqrt(rd))
diagInv = linalg.inv(diag)
self.data.bootstrap_correlation = diagInv.dot(
self.data.bootstrap_varCovar.dot(diagInv)
)
else:
self.data.bootstrap_correlation = np.full_like(
self.data.bootstrap_varCovar, np.finfo(float).max
)
self.data.secondOrderTable = {}
for i in range(self.data.nparam):
for j in range(i):
t = self._calculateTest(i, j, self.data.varCovar)
p = calcPValue(t)
trob = self._calculateTest(i, j, self.data.robust_varCovar)
prob = calcPValue(trob)
if self.data.bootstrap is not None:
tboot = self._calculateTest(
i, j, self.data.bootstrap_varCovar
)
pboot = calcPValue(tboot)
name = (self.data.betaNames[i], self.data.betaNames[j])
if self.data.bootstrap is not None:
self.data.secondOrderTable[name] = [
self.data.varCovar[i, j],
self.data.correlation[i, j],
t,
p,
self.data.robust_varCovar[i, j],
self.data.robust_correlation[i, j],
trob,
prob,
self.data.bootstrap_varCovar[i, j],
self.data.bootstrap_correlation[i, j],
tboot,
pboot,
]
else:
self.data.secondOrderTable[name] = [
self.data.varCovar[i, j],
self.data.correlation[i, j],
t,
p,
self.data.robust_varCovar[i, j],
self.data.robust_correlation[i, j],
trob,
prob,
]
eigIndex = np.argmin(self.data.eigenValues)
self.data.smallestEigenValue = self.data.eigenValues[eigIndex]
self.data.smallestEigenVector = self.data.eigenVectors[:, eigIndex]
self.data.smallestSingularValue = min(self.data.singularValues)
eigIndex = np.argmax(self.data.eigenValues)
self.data.largestEigenValue = self.data.eigenValues[eigIndex]
self.data.largestEigenVector = self.data.eigenVectors[:, eigIndex]
self.data.largestSingularValue = max(self.data.singularValues)
self.data.conditionNumber = (
self.data.largestEigenValue / self.data.smallestEigenValue
)
[docs] def shortSummary(self):
"""Provides a short summary of the estimation results"""
r = ''
r += f'Results for model {self.data.modelName}\n'
r += f'Nbr of parameters:\t\t{self.data.nparam}\n'
r += f'Sample size:\t\t\t{self.data.sampleSize}\n'
if self.data.sampleSize != self.data.numberOfObservations:
r += f'Observations:\t\t\t{self.data.numberOfObservations}\n'
r += f'Excluded data:\t\t\t{self.data.excludedData}\n'
if self.data.nullLogLike is not None:
r += f'Null log likelihood:\t\t{self.data.nullLogLike:.7g}\n'
r += f'Final log likelihood:\t\t{self.data.logLike:.7g}\n'
if self.data.nullLogLike is not None:
r += (
f'Likelihood ratio test (null):\t\t'
f'{self.data.likelihoodRatioTestNull:.7g}\n'
)
r += f'Rho square (null):\t\t\t{self.data.rhoSquareNull:.3g}\n'
r += (
f'Rho bar square (null):\t\t\t'
f'{self.data.rhoBarSquareNull:.3g}\n'
)
r += f'Akaike Information Criterion:\t{self.data.akaike:.7g}\n'
r += f'Bayesian Information Criterion:\t{self.data.bayesian:.7g}\n'
return r
def __str__(self):
r = '\n'
r += f'Results for model {self.data.modelName}\n'
if self.data.htmlFileName is not None:
r += f'Output file (HTML):\t\t\t{self.data.htmlFileName}\n'
if self.data.latexFileName is not None:
r += f'Output file (LaTeX):\t\t\t{self.data.latexFileName}\n'
r += f'Nbr of parameters:\t\t{self.data.nparam}\n'
r += f'Sample size:\t\t\t{self.data.sampleSize}\n'
if self.data.sampleSize != self.data.numberOfObservations:
r += f'Observations:\t\t\t{self.data.numberOfObservations}\n'
r += f'Excluded data:\t\t\t{self.data.excludedData}\n'
if self.data.nullLogLike is not None:
r += f'Null log likelihood:\t\t{self.data.nullLogLike:.7g}\n'
if self.data.initLogLike is not None:
r += f'Init log likelihood:\t\t{self.data.initLogLike:.7g}\n'
r += f'Final log likelihood:\t\t{self.data.logLike:.7g}\n'
if self.data.nullLogLike is not None:
r += (
f'Likelihood ratio test (null):\t\t'
f'{self.data.likelihoodRatioTestNull:.7g}\n'
)
r += f'Rho square (null):\t\t\t{self.data.rhoSquareNull:.3g}\n'
r += (
f'Rho bar square (null):\t\t\t'
f'{self.data.rhoBarSquareNull:.3g}\n'
)
if self.data.initLogLike is not None:
r += (
f'Likelihood ratio test (init):\t\t'
f'{self.data.likelihoodRatioTest:.7g}\n'
)
r += f'Rho square (init):\t\t\t{self.data.rhoSquare:.3g}\n'
r += f'Rho bar square (init):\t\t\t{self.data.rhoBarSquare:.3g}\n'
r += f'Akaike Information Criterion:\t{self.data.akaike:.7g}\n'
r += f'Bayesian Information Criterion:\t{self.data.bayesian:.7g}\n'
if self.data.gradientNorm is not None:
r += f'Final gradient norm:\t\t{self.data.gradientNorm:.7g}\n'
r += '\n'.join([f'{b}' for b in self.data.betas])
r += '\n'
if self.data.secondOrderTable is not None:
for k, v in self.data.secondOrderTable.items():
r += (
'{}:\t{:.3g}\t{:.3g}\t{:.3g}\t{:.3g}\t'
'{:.3g}\t{:.3g}\t{:.3g}\t{:.3g}\n'
).format(k, *v)
return r
def _getLaTeXHeader(self):
"""Prepare the header for the LaTeX file, containing comments and the
version of Biogeme.
:return: string containing the header.
:rtype: str
"""
h = ''
h += '%% This file is designed to be included into a LaTeX document\n'
h += (
'%% See http://www.latex-project.org for '
'information about LaTeX\n'
)
h += (
f'%% {self.data.modelName} - Report from '
f'biogeme {bv.getVersion()} '
f'[{bv.versionDate}]\n'
)
h += bv.getLaTeX()
return h
[docs] def getLaTeX(self, onlyRobust=True):
"""Get the results coded in LaTeX
:param onlyRobust: if True, only the robust statistics are included
:type onlyRobust: bool
:return: LaTeX code
:rtype: string
"""
now = datetime.datetime.now()
h = self._getLaTeXHeader()
if self.data.latexFileName is not None:
h += '\n%% File ' + self.data.latexFileName + '\n'
h += f'\n%% This file has automatically been generated on {now}</p>\n'
if self.data.dataname is not None:
h += f'\n%%Database name: {self.data.dataname}\n'
if self.data.userNotes is not None:
# User notes
h += f'\n%%{self.data.userNotes}\n'
h += '\n%% General statistics\n'
h += '\\section{General statistics}\n'
d = self.getGeneralStatistics()
h += '\\begin{tabular}{ll}\n'
for k, (v, p) in d.items():
if isinstance(v, bytes):
v = str(v)
if isinstance(v, str):
v = v.replace('_', '\\_')
h += f'{k} & {v:{p}} \\\\\n'
for k, v in self.data.optimizationMessages.items():
if k in ('Relative projected gradient', 'Relative change'):
h += f'{k} & \\verb${v:.7g}$ \\\\\n'
else:
h += f'{k} & \\verb${v}$ \\\\\n'
h += '\\end{tabular}\n'
h += '\n%%Parameter estimates\n'
h += '\\section{Parameter estimates}\n'
table = self.getEstimatedParameters(onlyRobust)
def formatting(x):
"""Defines the formatting for the to_latex function of pandas"""
res = f'{x:.3g}'
if '.' in res:
return res
return f'{res}.0'
## Need to check for old versions of Pandas.
try:
h += table.style.format(formatting).to_latex()
except AttributeError:
h += table.to_latex(float_format=formatting)
h += '\n%%Correlation\n'
h += '\\section{Correlation}\n'
table = self.getCorrelationResults()
## Need to check for old versions of Pandas.
try:
h += table.style.format(formatting).to_latex()
except AttributeError:
h += table.to_latex(float_format=formatting)
return h
[docs] def getGeneralStatistics(self):
"""Format the results in a dict
:return: dict with the results. The keys describe each
content. Each element is a GeneralStatistic tuple,
with the value and its preferred formatting.
Example::
'Init log likelihood': (-115.30029248549191, '.7g')
:rtype: dict(string:float,string)
"""
d = {}
d['Number of estimated parameters'] = GeneralStatistic(
value=self.data.nparam, format=''
)
nf = self.numberOfFreeParameters()
if nf != self.data.nparam:
d['Number of free parameters'] = GeneralStatistic(
value=nf, format=''
)
d['Sample size'] = GeneralStatistic(
value=self.data.sampleSize, format=''
)
if self.data.sampleSize != self.data.numberOfObservations:
d['Observations'] = GeneralStatistic(
value=self.data.numberOfObservations, format=''
)
d['Excluded observations'] = GeneralStatistic(
value=self.data.excludedData, format=''
)
if self.data.nullLogLike is not None:
d['Null log likelihood'] = GeneralStatistic(
value=self.data.nullLogLike, format='.7g'
)
d['Init log likelihood'] = GeneralStatistic(
value=self.data.initLogLike, format='.7g'
)
d['Final log likelihood'] = GeneralStatistic(
value=self.data.logLike, format='.7g'
)
if self.data.nullLogLike is not None:
d['Likelihood ratio test for the null model'] = GeneralStatistic(
value=self.data.likelihoodRatioTestNull,
format='.7g',
)
d['Rho-square for the null model'] = GeneralStatistic(
value=self.data.rhoSquareNull, format='.3g'
)
d['Rho-square-bar for the null model'] = GeneralStatistic(
value=self.data.rhoBarSquareNull,
format='.3g',
)
d['Likelihood ratio test for the init. model'] = GeneralStatistic(
value=self.data.likelihoodRatioTest,
format='.7g',
)
d['Rho-square for the init. model'] = GeneralStatistic(
value=self.data.rhoSquare, format='.3g'
)
d['Rho-square-bar for the init. model'] = GeneralStatistic(
value=self.data.rhoBarSquare, format='.3g'
)
d['Akaike Information Criterion'] = GeneralStatistic(
value=self.data.akaike, format='.7g'
)
d['Bayesian Information Criterion'] = GeneralStatistic(
value=self.data.bayesian, format='.7g'
)
d['Final gradient norm'] = GeneralStatistic(
value=self.data.gradientNorm, format='.4E'
)
if self.data.monteCarlo:
d['Number of draws'] = GeneralStatistic(
value=self.data.numberOfDraws, format=''
)
d['Draws generation time'] = GeneralStatistic(
value=self.data.drawsProcessingTime, format=''
)
d['Types of draws'] = GeneralStatistic(
value=[f'{i}: {k}' for i, k in self.data.typesOfDraws.items()],
format='',
)
if self.data.bootstrap is not None:
d['Bootstrapping time'] = GeneralStatistic(
value=self.data.bootstrap_time, format=''
)
d['Nbr of threads'] = GeneralStatistic(
value=self.data.numberOfThreads, format=''
)
return d
[docs] def printGeneralStatistics(self):
"""Print the general statistics of the estimation.
:return: general statistics
Example::
Number of estimated parameters: 2
Sample size: 5
Excluded observations: 0
Init log likelihood: -67.08858
Final log likelihood: -67.06549
Likelihood ratio test for the init. model: 0.04618175
Rho-square for the init. model: 0.000344
Rho-square-bar for the init. model: -0.0295
Akaike Information Criterion: 138.131
Bayesian Information Criterion: 137.3499
Final gradient norm: 3.9005E-07
Bootstrapping time: 0:00:00.042713
Nbr of threads: 16
:rtype: str
"""
d = self.getGeneralStatistics()
output = ''
for k, (v, p) in d.items():
output += f'{k}:\t{v:{p}}\n'
return output
[docs] def numberOfFreeParameters(self):
"""This is the number of estimated parameters, minus those that are at
their bounds
"""
return sum([not b.isBoundActive() for b in self.data.betas])
[docs] def getEstimatedParameters(self, onlyRobust=True):
"""Gather the estimated parameters and the corresponding statistics in
a Pandas dataframe.
:param onlyRobust: if True, only the robust statistics are included
:type onlyRobust: bool
:return: Pandas dataframe with the results
:rtype: pandas.DataFrame
"""
# There should be a more 'Pythonic' way to do this.
anyActiveBound = False
for b in self.data.betas:
if b.isBoundActive():
anyActiveBound = True
if anyActiveBound:
if onlyRobust:
columns = [
'Value',
'Active bound',
'Rob. Std err',
'Rob. t-test',
'Rob. p-value',
]
else:
columns = [
'Value',
'Active bound',
'Std err',
't-test',
'p-value',
'Rob. Std err',
'Rob. t-test',
'Rob. p-value',
]
else:
if onlyRobust:
columns = [
'Value',
'Rob. Std err',
'Rob. t-test',
'Rob. p-value',
]
else:
columns = [
'Value',
'Std err',
't-test',
'p-value',
'Rob. Std err',
'Rob. t-test',
'Rob. p-value',
]
if self.data.bootstrap is not None and not onlyRobust:
columns += [
f'Bootstrap[{len(self.data.bootstrap)}] Std err',
'Bootstrap t-test',
'Bootstrap p-value',
]
table = pd.DataFrame(columns=columns)
for b in self.data.betas:
if anyActiveBound:
if onlyRobust:
arow = {
'Value': b.value,
'Active bound': {True: 1.0, False: 0.0}[
b.isBoundActive()
],
'Rob. Std err': b.robust_stdErr,
'Rob. t-test': b.robust_tTest,
'Rob. p-value': b.robust_pValue,
}
else:
arow = {
'Value': b.value,
'Active bound': {True: 1.0, False: 0.0}[
b.isBoundActive()
],
'Std err': b.stdErr,
't-test': b.tTest,
'p-value': b.pValue,
'Rob. Std err': b.robust_stdErr,
'Rob. t-test': b.robust_tTest,
'Rob. p-value': b.robust_pValue,
}
else:
if onlyRobust:
arow = {
'Value': b.value,
'Rob. Std err': b.robust_stdErr,
'Rob. t-test': b.robust_tTest,
'Rob. p-value': b.robust_pValue,
}
else:
arow = {
'Value': b.value,
'Std err': b.stdErr,
't-test': b.tTest,
'p-value': b.pValue,
'Rob. Std err': b.robust_stdErr,
'Rob. t-test': b.robust_tTest,
'Rob. p-value': b.robust_pValue,
}
if self.data.bootstrap is not None and not onlyRobust:
arow[
f'Bootstrap[{len(self.data.bootstrap)}] Std err'
] = b.bootstrap_stdErr
arow['Bootstrap t-test'] = b.bootstrap_tTest
arow['Bootstrap p-value'] = b.bootstrap_pValue
table.loc[b.name] = pd.Series(arow)
return table
[docs] def getCorrelationResults(self, subset=None):
"""Get the statistics about pairs of coefficients as a Pandas dataframe
:param subset: produce the results only for a subset of
parameters. If None, all the parameters are involved. Default: None
:type subset: list(str)
:return: Pandas data frame with the correlation results
:rtype: pandas.DataFrame
"""
if subset is not None:
unknown = []
for p in subset:
if p not in self.data.betaNames:
unknown.append(p)
if unknown:
logger.warning(f'Unknown parameters are ignored: {unknown}')
columns = [
'Covariance',
'Correlation',
't-test',
'p-value',
'Rob. cov.',
'Rob. corr.',
'Rob. t-test',
'Rob. p-value',
]
if self.data.bootstrap is not None:
columns += [
'Boot. cov.',
'Boot. corr.',
'Boot. t-test',
'Boot. p-value',
]
table = pd.DataFrame(columns=columns)
for k, v in self.data.secondOrderTable.items():
if subset is None:
include = True
else:
include = k[0] in subset and k[1] in subset
if include:
arow = {
'Covariance': v[0],
'Correlation': v[1],
't-test': v[2],
'p-value': v[3],
'Rob. cov.': v[4],
'Rob. corr.': v[5],
'Rob. t-test': v[6],
'Rob. p-value': v[7],
}
if self.data.bootstrap is not None:
arow['Boot. cov.'] = v[8]
arow['Boot. corr.'] = v[9]
arow['Boot. t-test'] = v[10]
arow['Boot. p-value'] = v[11]
table.loc[f'{k[0]}-{k[1]}'] = pd.Series(arow)
return table
[docs] def getHtml(self):
"""Get the results coded in HTML
:return: HTML code
:rtype: string
"""
now = datetime.datetime.now()
h = self._getHtmlHeader()
h += bv.getHtml()
h += f'<p>This file has automatically been generated on {now}</p>\n'
h += '<table>\n'
h += (
f'<tr class=biostyle><td align=right>'
f'<strong>Report file</strong>: </td>'
f'<td>{self.data.htmlFileName}</td></tr>\n'
)
h += (
f'<tr class=biostyle><td align=right>'
f'<strong>Database name</strong>: </td>'
f'<td>{self.data.dataname}</td></tr>\n'
)
h += '</table>\n'
if self.data.userNotes is not None:
# User notes
h += (
f'<blockquote style="border: 2px solid #666; '
f'padding: 10px; background-color:'
f' #ccc;">{self.data.userNotes}</blockquote>'
)
# Include here the part on statistics
h += '<h1>Estimation report</h1>\n'
h += '<table border="0">\n'
d = self.getGeneralStatistics()
# k is the description of the quantity
# v is the value
# p is the precision to format it
for k, (v, p) in d.items():
if v is not None:
h += (
f'<tr class=biostyle><td align=right >'
f'<strong>{k}</strong>: </td> '
f'<td>{v:{p}}</td></tr>\n'
)
for k, v in self.data.optimizationMessages.items():
if k == 'Relative projected gradient':
h += (
f'<tr class=biostyle><td align=right >'
f'<strong>{k}</strong>: </td> '
f'<td>{v:.7g}</td></tr>\n'
)
else:
h += (
f'<tr class=biostyle><td align=right >'
f'<strong>{k}</strong>: </td> '
f'<td>{v}</td></tr>\n'
)
h += '</table>\n'
table = self.getEstimatedParameters()
h += '<h1>Estimated parameters</h1>\n'
h += '<table border="1">\n'
h += '<tr class=biostyle><th>Name</th>'
for c in table.columns:
h += f'<th>{c}</th>'
h += '</tr>\n'
for name, values in table.iterrows():
h += f'<tr class=biostyle><td>{name}</td>'
for k, v in values.items():
# print(f'values[{k}] = {v}')
h += f'<td>{v:.3g}</td>'
h += '</tr>\n'
h += '</table>\n'
table = self.getCorrelationResults()
h += '<h2>Correlation of coefficients</h2>\n'
h += '<table border="1">\n'
h += '<tr class=biostyle><th>Coefficient1</th><th>Coefficient2</th>'
for c in table.columns:
h += f'<th>{c}</th>'
h += '</tr>\n'
for name, values in table.iterrows():
n = name.split('-')
h += f'<tr class=biostyle><td>{n[0]}</td><td>{n[1]}</td>'
for k, v in values.items():
h += f'<td>{v:.3g}</td>'
h += '</tr>\n'
h += '</table>\n'
h += (
f'<p>Smallest eigenvalue: '
f'{self.data.smallestEigenValue:.6g}</p>\n'
)
h += (
f'<p>Largest eigenvalue: '
f'{self.data.largestEigenValue:.6g}</p>\n'
)
h += f'<p>Condition number: ' f'{self.data.conditionNumber:.6g}</p>\n'
if np.abs(self.data.smallestEigenValue) <= 1.0e-5:
h += (
'<p>The second derivatives matrix is close to singularity. '
'Variables involved:'
)
h += '<table>'
for i, ev in enumerate(self.data.smallestEigenVector):
if np.abs(ev) > 1.0e-5:
h += (
f'<tr><td>{ev:.3g}</td>'
f'<td> *</td>'
f'<td> {self.data.betaNames[i]}</td></tr>\n'
)
h += '</table>'
h += '</p>\n'
h += '</html>'
return h
[docs] def getBetaValues(self, myBetas=None):
"""Retrieve the values of the estimated parameters, by names.
:param myBetas: names of the requested parameters. If None, all
available parameters will be reported. Default: None.
:type myBetas: list(string)
:return: dict containing the values, where the keys are the names.
:rtype: dict(string:float)
:raise biogeme.exceptions.biogemeError: if some requested parameters
are not available.
"""
values = {}
if myBetas is None:
myBetas = self.data.betaNames
for b in myBetas:
try:
index = self.data.betaNames.index(b)
values[b] = self.data.betas[index].value
except KeyError as e:
keys = ''
for k in self.data.betaNames:
keys += f' {k}'
err = (
f'The value of {b} is not available in the results. '
f'The following parameters are available: {keys}'
)
raise excep.biogemeError(err) from e
return values
[docs] def getVarCovar(self):
"""Obtain the Rao-Cramer variance covariance matrix as a
Pandas data frame.
:return: Rao-Cramer variance covariance matrix
:rtype: pandas.DataFrame
"""
names = [b.name for b in self.data.betas]
vc = pd.DataFrame(index=names, columns=names)
for i, betai in enumerate(self.data.betas):
for j, betaj in enumerate(self.data.betas):
vc.at[betai.name, betaj.name] = self.data.varCovar[i, j]
return vc
[docs] def getRobustVarCovar(self):
"""Obtain the robust variance covariance matrix as a Pandas data frame.
:return: robust variance covariance matrix
:rtype: pandas.DataFrame
"""
names = [b.name for b in self.data.betas]
vc = pd.DataFrame(index=names, columns=names)
for i, betai in enumerate(self.data.betas):
for j, betaj in enumerate(self.data.betas):
vc.at[betai.name, betaj.name] = self.data.robust_varCovar[i, j]
return vc
[docs] def getBootstrapVarCovar(self):
"""Obtain the bootstrap variance covariance matrix as
a Pandas data frame.
:return: bootstrap variance covariance matrix, or None if not available
:rtype: pandas.DataFrame
"""
if self.data.bootstrap is None:
return None
names = [b.name for b in self.data.betas]
vc = pd.DataFrame(index=names, columns=names)
for i, betai in enumerate(self.data.betas):
for j, betaj in enumerate(self.data.betas):
vc.at[betai.name, betaj.name] = self.data.bootstrap_varCovar[
i, j
]
return vc
[docs] def writeHtml(self):
"""Write the results in an HTML file."""
self.data.htmlFileName = bf.getNewFileName(self.data.modelName, 'html')
with open(self.data.htmlFileName, 'w') as f:
f.write(self.getHtml())
logger.general(f'Results saved in file {self.data.htmlFileName}')
[docs] def writeLaTeX(self):
"""Write the results in a LaTeX file."""
self.data.latexFileName = bf.getNewFileName(self.data.modelName, 'tex')
with open(self.data.latexFileName, 'w') as f:
f.write(self.getLaTeX())
logger.general(f'Results saved in file {self.data.latexFileName}')
def _getHtmlHeader(self):
"""Prepare the header for the HTML file, containing comments and the
version of Biogeme.
:return: string containing the header.
:rtype: str
"""
h = ''
h += '<html>\n'
h += '<head>\n'
h += (
'<script src="http://transp-or.epfl.ch/biogeme/sorttable.js">'
'</script>\n'
)
h += (
'<meta http-equiv="Content-Type" content="text/html; '
'charset=utf-8" />\n'
)
h += (
f'<title>{self.data.modelName} - Report from '
f'biogeme {bv.getVersion()} '
f'[{bv.versionDate}]</title>\n'
)
h += (
'<meta name="keywords" content="biogeme, discrete choice, '
'random utility">\n'
)
h += (
f'<meta name="description" content="Report from '
f'biogeme {bv.getVersion()} '
f'[{bv.versionDate}]">\n'
)
h += '<meta name="author" content="{bv.author}">\n'
h += '<style type=text/css>\n'
h += '.biostyle\n'
h += ' {font-size:10.0pt;\n'
h += ' font-weight:400;\n'
h += ' font-style:normal;\n'
h += ' font-family:Courier;}\n'
h += '.boundstyle\n'
h += ' {font-size:10.0pt;\n'
h += ' font-weight:400;\n'
h += ' font-style:normal;\n'
h += ' font-family:Courier;\n'
h += ' color:red}\n'
h += '</style>\n'
h += '</head>\n'
h += '<body bgcolor="#ffffff">\n'
return h
[docs] def getBetasForSensitivityAnalysis(
self, myBetas, size=100, useBootstrap=True
):
"""Generate draws from the distribution of the estimates, for
sensitivity analysis.
:param myBetas: names of the parameters for which draws are requested.
:type myBetas: list(string)
:param size: number of draws. If useBootstrap is True, the value is
ignored and a warning is issued. Default: 100.
:type size: int
:param useBootstrap: if True, the bootstrap estimates are
directly used. The advantage is that it does not reyl on the
assumption that the estimates follow a normal
distribution. Default: True.
:type useBootstrap: bool
:raise biogeme.exceptions.biogemeError: if useBootstrap is True and
the bootstrap results are not available
:return: list of dict. Each dict has a many entries as parameters.
The list has as many entries as draws.
:rtype: list(dict)
"""
if useBootstrap and self.data.bootstrap is None:
err = (
'Bootstrap results are not available for simulation. '
'Use useBootstrap=False.'
)
raise excep.biogemeError(err)
index = [self.data.betaNames.index(b) for b in myBetas]
if useBootstrap:
results = [
{myBetas[i]: value for i, value in enumerate(row)}
for row in self.data.bootstrap[:, index]
]
return results
theMatrix = (
self.data.bootstrap_varCovar
if useBootstrap
else self.data.robust_varCovar
)
simulatedBetas = np.random.multivariate_normal(
self.data.betaValues, theMatrix, size
)
index = [self.data.betaNames.index(b) for b in myBetas]
results = [
{myBetas[i]: value for i, value in enumerate(row)}
for row in simulatedBetas[:, index]
]
return results
[docs] def getF12(self, robustStdErr=True):
"""F12 is a format used by the software ALOGIT to
report estimation results.
:param robustStdErr: if True, the robust standard errors are reports.
If False, the Rao-Cramer are.
:type robustStdErr: bool
:return: results in F12 format
:rtype: string
"""
# checkline1 = (
# '0000000001111111111222222222233333333334444444444'
# '5555555555666666666677777777778'
# )
# checkline2 = (
# '1234567890123456789012345678901234567890123456789'
# '0123456789012345678901234567890'
# )
results = ''
# results += f'{checkline1}\n'
# results += f'{checkline2}\n'
# Line 1, title, characters 1-79
results += f'{self.data.modelName[:79]: >79}\n'
# Line 2, subtitle, characters 1-27, and time-date, characters 57-77
t = f'From biogeme {bv.getVersion()}'
d = f'{datetime.datetime.now()}'[:19]
results += f'{t[:27]: <56}{d: <21}\n'
# Line 3, "END" (this is historical!)
results += 'END\n'
# results += f'{checkline1}\n'
# results += f'{checkline2}\n'
# Line 4-(K+3), coefficient values
# characters 1-4, " 0" (again historical)
# characters 6-15, coefficient label, suggest using first 10
# characters of label in R
# characters 16-17, " F" (this indicates whether or not the
# coefficient is constrained)
# characters 19-38, coefficient value 20 chars
# characters 39-58, standard error 20 chars
mystats = self.getGeneralStatistics()
table = self.getEstimatedParameters(onlyRobust=False)
coefNames = list(table.index.values)
for name in coefNames:
values = table.loc[name]
results += ' 0 '
results += f'{name[:10]: >10}'
if 'Active bound' in values:
if values['Active bound'] == 1:
results += ' T'
else:
results += ' F'
else:
results += ' F'
results += ' '
results += f' {values["Value"]: >+19.12e}'
if robustStdErr:
results += f' {values["Rob. Std err"]: >+19.12e}'
else:
results += f' {values["Std err"]: >+19.12e}'
results += '\n'
# Line K+4, " -1" indicates end of coefficients
results += ' -1\n'
# results += f'{checkline1}\n'
# results += f'{checkline2}\n'
# Line K+5, statistics about run
# characters 1-8, number of observations 8 chars
# characters 9-27, likelihood-with-constants 19 chars
# characters 28-47, null likelihood 20 chars
# characters 48-67, final likelihood 20 chars
results += f'{mystats["Sample size"][0]: >8}'
# The cte log likelihood is not available. We put 0 instead.
results += f' {0: >18}'
if self.data.nullLogLike is not None:
results += f' {mystats["Null log likelihood"][0]: >+19.12e}'
else:
results += f' {0: >19}'
results += f' {mystats["Final log likelihood"][0]: >+19.12e}'
results += '\n'
# results += f'{checkline1}\n'
# results += f'{checkline2}\n'
# Line K+6, more statistics
# characters 1-4, number of iterations (suggest use 0) 4 chars
# characters 5-8, error code (please use 0) 4 chars
# characters 9-29, time and date (sugg. repeat from line 2) 21 chars
if "Number of iterations" in mystats:
results += f'{mystats["Number of iterations"][0]: >4}'
else:
results += f'{0: >4}'
results += f'{0: >4}'
results += f'{d: >21}'
results += '\n'
# results += f'{checkline1}\n'
# results += f'{checkline2}\n'
# Lines (K+7)-however many we need, correlations*100000
# 10 per line, fields of width 7
# The order of these is that correlation i,j (i>j) is in position
# (i-1)*(i-2)/2+j, i.e.
# (2,1) (3,1) (3,2) (4,1) etc.
count = 0
for i, coefi in enumerate(coefNames):
for j in range(0, i):
name = (coefi, coefNames[j])
if robustStdErr:
corr = int(100000 * self.data.secondOrderTable[name][5])
else:
corr = int(100000 * self.data.secondOrderTable[name][1])
results += f'{corr:7d}'
count += 1
if count % 10 == 0:
results += '\n'
results += '\n'
return results
[docs] def writeF12(self, robustStdErr=True):
"""Write the results in F12 file."""
self.data.F12FileName = bf.getNewFileName(self.data.modelName, 'F12')
with open(self.data.F12FileName, 'w') as f:
f.write(self.getF12(robustStdErr))
logger.general(f'Results saved in file {self.data.F12FileName}')
[docs] def likelihood_ratio_test(self, other_model, significance_level=0.05):
"""This function performs a likelihood ratio test between a restricted
and an unrestricted model. The "self" model can be either the
restricted or the unrestricted.
:param other_model: other model to perform the test.
:type other_model: biogeme.results.bioResults
:param significance_level: level of significance of the
test. Default: 0.05
:type significance_level: float
:return: a tuple containing:
- a message with the outcome of the test
- the statistic, that is minus two times the difference
between the loglikelihood of the two models
- the threshold of the chi square distribution.
:rtype: LRTuple(str, float, float)
"""
LR = self.data.logLike
LU = other_model.data.logLike
KR = self.data.nparam
KU = other_model.data.nparam
return tools.likelihood_ratio_test(
(LU, KU), (LR, KR), significance_level
)
[docs]def compileEstimationResults(
dict_of_results,
statistics=(
'Number of estimated parameters',
'Sample size',
'Final log likelihood',
'Akaike Information Criterion',
'Bayesian Information Criterion',
),
include_parameter_estimates=True,
):
"""Compile estimation results into a common table
:param dict_of_results: dictionary where the keys are the names of
the models, and the values are either the estimation results,
or the name of the pickle file where to find them.
:type dict_of_results: dict(str:bioResults) or dict(str:str)
:param statistics: list of statistics to include in the summary
table
:type statistics: tuple(str)
:param include_parameter_estimates: if True, the parameter
estimates are included.
:type include_parameter_estimates: bool
:return: pandas dataframe with the requested results.
:rtype: pandas.DataFrame
"""
df = pd.DataFrame(columns=dict_of_results.keys())
for col, res in dict_of_results.items():
if not isinstance(res, bioResults):
try:
res = bioResults(pickleFile=res)
except excep.biogemeError:
warning = f'Impossible to access result file {res}'
logger.warning(warning)
res = None
if res is not None:
stats_results = res.getGeneralStatistics()
for s in statistics:
df.loc[s, col] = stats_results[s][0]
if include_parameter_estimates:
betas = res.getBetaValues()
for name, value in betas.items():
df.loc[name, col] = value
return df.fillna('')