"""
Robust linear models with support for the M-estimators listed under
:ref:`norms <norms>`.
References
----------
PJ Huber. 'Robust Statistics' John Wiley and Sons, Inc., New York. 1981.
PJ Huber. 1973, 'The 1972 Wald Memorial Lectures: Robust Regression:
Asymptotics, Conjectures, and Monte Carlo.' The Annals of Statistics,
1.5, 799-821.
R Venables, B Ripley. 'Modern Applied Statistics in S' Springer, New York,
2002.
"""
import numpy as np
import scipy.stats as stats
import statsmodels.base.model as base
import statsmodels.base.wrapper as wrap
import statsmodels.regression._tools as reg_tools
import statsmodels.regression.linear_model as lm
import statsmodels.robust.norms as norms
import statsmodels.robust.scale as scale
from statsmodels.tools.decorators import cache_readonly
from statsmodels.tools.sm_exceptions import ConvergenceWarning
__all__ = ['RLM']
def _check_convergence(criterion, iteration, tol, maxiter):
cond = np.abs(criterion[iteration] - criterion[iteration - 1])
return not (np.any(cond > tol) and iteration < maxiter)
[docs]class RLM(base.LikelihoodModel):
__doc__ = """
Robust Linear Model
Estimate a robust linear model via iteratively reweighted least squares
given a robust criterion estimator.
%(params)s
M : statsmodels.robust.norms.RobustNorm, optional
The robust criterion function for downweighting outliers.
The current options are LeastSquares, HuberT, RamsayE, AndrewWave,
TrimmedMean, Hampel, and TukeyBiweight. The default is HuberT().
See statsmodels.robust.norms for more information.
%(extra_params)s
Attributes
----------
df_model : float
The degrees of freedom of the model. The number of regressors p less
one for the intercept. Note that the reported model degrees
of freedom does not count the intercept as a regressor, though
the model is assumed to have an intercept.
df_resid : float
The residual degrees of freedom. The number of observations n
less the number of regressors p. Note that here p does include
the intercept as using a degree of freedom.
endog : ndarray
See above. Note that endog is a reference to the data so that if
data is already an array and it is changed, then `endog` changes
as well.
exog : ndarray
See above. Note that endog is a reference to the data so that if
data is already an array and it is changed, then `endog` changes
as well.
M : statsmodels.robust.norms.RobustNorm
See above. Robust estimator instance instantiated.
nobs : float
The number of observations n
pinv_wexog : ndarray
The pseudoinverse of the design / exogenous data array. Note that
RLM has no whiten method, so this is just the pseudo inverse of the
design.
normalized_cov_params : ndarray
The p x p normalized covariance of the design / exogenous data.
This is approximately equal to (X.T X)^(-1)
Examples
--------
>>> import statsmodels.api as sm
>>> data = sm.datasets.stackloss.load()
>>> data.exog = sm.add_constant(data.exog)
>>> rlm_model = sm.RLM(data.endog, data.exog, \
M=sm.robust.norms.HuberT())
>>> rlm_results = rlm_model.fit()
>>> rlm_results.params
array([ 0.82938433, 0.92606597, -0.12784672, -41.02649835])
>>> rlm_results.bse
array([ 0.11100521, 0.30293016, 0.12864961, 9.79189854])
>>> rlm_results_HC2 = rlm_model.fit(cov="H2")
>>> rlm_results_HC2.params
array([ 0.82938433, 0.92606597, -0.12784672, -41.02649835])
>>> rlm_results_HC2.bse
array([ 0.11945975, 0.32235497, 0.11796313, 9.08950419])
>>> mod = sm.RLM(data.endog, data.exog, M=sm.robust.norms.Hampel())
>>> rlm_hamp_hub = mod.fit(scale_est=sm.robust.scale.HuberScale())
>>> rlm_hamp_hub.params
array([ 0.73175452, 1.25082038, -0.14794399, -40.27122257])
""" % {'params': base._model_params_doc,
'extra_params': base._missing_param_doc}
def __init__(self, endog, exog, M=None, missing='none',
**kwargs):
self._check_kwargs(kwargs)
self.M = M if M is not None else norms.HuberT()
super(base.LikelihoodModel, self).__init__(endog, exog,
missing=missing, **kwargs)
self._initialize()
# things to remove_data
self._data_attr.extend(['weights', 'pinv_wexog'])
def _initialize(self):
"""
Initializes the model for the IRLS fit.
Resets the history and number of iterations.
"""
self.pinv_wexog = np.linalg.pinv(self.exog)
self.normalized_cov_params = np.dot(self.pinv_wexog,
np.transpose(self.pinv_wexog))
self.df_resid = (float(self.exog.shape[0] -
np.linalg.matrix_rank(self.exog)))
self.df_model = float(np.linalg.matrix_rank(self.exog) - 1)
self.nobs = float(self.endog.shape[0])
[docs] def score(self, params):
raise NotImplementedError
[docs] def predict(self, params, exog=None):
"""
Return linear predicted values from a design matrix.
Parameters
----------
params : array_like
Parameters of a linear model
exog : array_like, optional.
Design / exogenous data. Model exog is used if None.
Returns
-------
An array of fitted values
"""
# copied from linear_model # TODO: then is it needed?
if exog is None:
exog = self.exog
return np.dot(exog, params)
[docs] def loglike(self, params):
raise NotImplementedError
[docs] def deviance(self, tmp_results):
"""
Returns the (unnormalized) log-likelihood from the M estimator.
"""
tmp_resid = self.endog - tmp_results.fittedvalues
return self.M(tmp_resid / tmp_results.scale).sum()
def _update_history(self, tmp_results, history, conv):
history['params'].append(tmp_results.params)
history['scale'].append(tmp_results.scale)
if conv == 'dev':
history['deviance'].append(self.deviance(tmp_results))
elif conv == 'sresid':
history['sresid'].append(tmp_results.resid / tmp_results.scale)
elif conv == 'weights':
history['weights'].append(tmp_results.model.weights)
return history
def _estimate_scale(self, resid):
"""
Estimates the scale based on the option provided to the fit method.
"""
if isinstance(self.scale_est, str):
if self.scale_est.lower() == 'mad':
return scale.mad(resid, center=0)
else:
raise ValueError("Option %s for scale_est not understood" %
self.scale_est)
elif isinstance(self.scale_est, scale.HuberScale):
return self.scale_est(self.df_resid, self.nobs, resid)
else:
return scale.scale_est(self, resid) ** 2
[docs] def fit(self, maxiter=50, tol=1e-8, scale_est='mad', init=None, cov='H1',
update_scale=True, conv='dev', start_params=None):
"""
Fits the model using iteratively reweighted least squares.
The IRLS routine runs until the specified objective converges to `tol`
or `maxiter` has been reached.
Parameters
----------
conv : str
Indicates the convergence criteria.
Available options are "coefs" (the coefficients), "weights" (the
weights in the iteration), "sresid" (the standardized residuals),
and "dev" (the un-normalized log-likelihood for the M
estimator). The default is "dev".
cov : str, optional
'H1', 'H2', or 'H3'
Indicates how the covariance matrix is estimated. Default is 'H1'.
See rlm.RLMResults for more information.
init : str
Specifies method for the initial estimates of the parameters.
Default is None, which means that the least squares estimate
is used. Currently it is the only available choice.
maxiter : int
The maximum number of iterations to try. Default is 50.
scale_est : str or HuberScale()
'mad' or HuberScale()
Indicates the estimate to use for scaling the weights in the IRLS.
The default is 'mad' (median absolute deviation. Other options are
'HuberScale' for Huber's proposal 2. Huber's proposal 2 has
optional keyword arguments d, tol, and maxiter for specifying the
tuning constant, the convergence tolerance, and the maximum number
of iterations. See statsmodels.robust.scale for more information.
tol : float
The convergence tolerance of the estimate. Default is 1e-8.
update_scale : Bool
If `update_scale` is False then the scale estimate for the
weights is held constant over the iteration. Otherwise, it
is updated for each fit in the iteration. Default is True.
start_params : array_like, optional
Initial guess of the solution of the optimizer. If not provided,
the initial parameters are computed using OLS.
Returns
-------
results : statsmodels.rlm.RLMresults
Results instance
"""
if cov.upper() not in ["H1", "H2", "H3"]:
raise ValueError("Covariance matrix %s not understood" % cov)
else:
self.cov = cov.upper()
conv = conv.lower()
if conv not in ["weights", "coefs", "dev", "sresid"]:
raise ValueError("Convergence argument %s not understood" % conv)
self.scale_est = scale_est
if start_params is None:
wls_results = lm.WLS(self.endog, self.exog).fit()
else:
start_params = np.asarray(start_params, dtype=np.double).squeeze()
if (start_params.shape[0] != self.exog.shape[1] or
start_params.ndim != 1):
raise ValueError('start_params must by a 1-d array with {0} '
'values'.format(self.exog.shape[1]))
fake_wls = reg_tools._MinimalWLS(self.endog, self.exog,
weights=np.ones_like(self.endog),
check_weights=False)
wls_results = fake_wls.results(start_params)
if not init:
self.scale = self._estimate_scale(wls_results.resid)
history = dict(params=[np.inf], scale=[])
if conv == 'coefs':
criterion = history['params']
elif conv == 'dev':
history.update(dict(deviance=[np.inf]))
criterion = history['deviance']
elif conv == 'sresid':
history.update(dict(sresid=[np.inf]))
criterion = history['sresid']
elif conv == 'weights':
history.update(dict(weights=[np.inf]))
criterion = history['weights']
# done one iteration so update
history = self._update_history(wls_results, history, conv)
iteration = 1
converged = 0
while not converged:
if self.scale == 0.0:
import warnings
warnings.warn('Estimated scale is 0.0 indicating that the most'
' last iteration produced a perfect fit of the '
'weighted data.', ConvergenceWarning)
break
self.weights = self.M.weights(wls_results.resid / self.scale)
wls_results = reg_tools._MinimalWLS(self.endog, self.exog,
weights=self.weights,
check_weights=True).fit()
if update_scale is True:
self.scale = self._estimate_scale(wls_results.resid)
history = self._update_history(wls_results, history, conv)
iteration += 1
converged = _check_convergence(criterion, iteration, tol, maxiter)
results = RLMResults(self, wls_results.params,
self.normalized_cov_params, self.scale)
history['iteration'] = iteration
results.fit_history = history
results.fit_options = dict(cov=cov.upper(), scale_est=scale_est,
norm=self.M.__class__.__name__, conv=conv)
# norm is not changed in fit, no old state
# doing the next causes exception
# self.cov = self.scale_est = None #reset for additional fits
# iteration and history could contain wrong state with repeated fit
return RLMResultsWrapper(results)
[docs]class RLMResults(base.LikelihoodModelResults):
"""
Class to contain RLM results
Attributes
----------
bcov_scaled : ndarray
p x p scaled covariance matrix specified in the model fit method.
The default is H1. H1 is defined as
``k**2 * (1/df_resid*sum(M.psi(sresid)**2)*scale**2)/
((1/nobs*sum(M.psi_deriv(sresid)))**2) * (X.T X)^(-1)``
where ``k = 1 + (df_model +1)/nobs * var_psiprime/m**2``
where ``m = mean(M.psi_deriv(sresid))`` and
``var_psiprime = var(M.psi_deriv(sresid))``
H2 is defined as
``k * (1/df_resid) * sum(M.psi(sresid)**2) *scale**2/
((1/nobs)*sum(M.psi_deriv(sresid)))*W_inv``
H3 is defined as
``1/k * (1/df_resid * sum(M.psi(sresid)**2)*scale**2 *
(W_inv X.T X W_inv))``
where `k` is defined as above and
``W_inv = (M.psi_deriv(sresid) exog.T exog)^(-1)``
See the technical documentation for cleaner formulae.
bcov_unscaled : ndarray
The usual p x p covariance matrix with scale set equal to 1. It
is then just equivalent to normalized_cov_params.
bse : ndarray
An array of the standard errors of the parameters. The standard
errors are taken from the robust covariance matrix specified in the
argument to fit.
chisq : ndarray
An array of the chi-squared values of the parameter estimates.
df_model
See RLM.df_model
df_resid
See RLM.df_resid
fit_history : dict
Contains information about the iterations. Its keys are `deviance`,
`params`, `iteration` and the convergence criteria specified in
`RLM.fit`, if different from `deviance` or `params`.
fit_options : dict
Contains the options given to fit.
fittedvalues : ndarray
The linear predicted values. dot(exog, params)
model : statsmodels.rlm.RLM
A reference to the model instance
nobs : float
The number of observations n
normalized_cov_params : ndarray
See RLM.normalized_cov_params
params : ndarray
The coefficients of the fitted model
pinv_wexog : ndarray
See RLM.pinv_wexog
pvalues : ndarray
The p values associated with `tvalues`. Note that `tvalues` are assumed
to be distributed standard normal rather than Student's t.
resid : ndarray
The residuals of the fitted model. endog - fittedvalues
scale : float
The type of scale is determined in the arguments to the fit method in
RLM. The reported scale is taken from the residuals of the weighted
least squares in the last IRLS iteration if update_scale is True. If
update_scale is False, then it is the scale given by the first OLS
fit before the IRLS iterations.
sresid : ndarray
The scaled residuals.
tvalues : ndarray
The "t-statistics" of params. These are defined as params/bse where
bse are taken from the robust covariance matrix specified in the
argument to fit.
weights : ndarray
The reported weights are determined by passing the scaled residuals
from the last weighted least squares fit in the IRLS algorithm.
See Also
--------
statsmodels.base.model.LikelihoodModelResults
"""
def __init__(self, model, params, normalized_cov_params, scale):
super(RLMResults, self).__init__(model, params,
normalized_cov_params, scale)
self.model = model
self.df_model = model.df_model
self.df_resid = model.df_resid
self.nobs = model.nobs
self._cache = {}
# for remove_data
self._data_in_cache.extend(['sresid'])
self.cov_params_default = self.bcov_scaled
# TODO: "pvals" should come from chisq on bse?
@cache_readonly
def fittedvalues(self):
return np.dot(self.model.exog, self.params)
@cache_readonly
def resid(self):
return self.model.endog - self.fittedvalues # before bcov
@cache_readonly
def sresid(self):
if self.scale == 0.0:
sresid = self.resid.copy()
sresid[:] = 0.0
return sresid
return self.resid / self.scale
@cache_readonly
def bcov_unscaled(self):
return self.normalized_cov_params
@cache_readonly
def weights(self):
return self.model.weights
@cache_readonly
def bcov_scaled(self):
model = self.model
m = np.mean(model.M.psi_deriv(self.sresid))
var_psiprime = np.var(model.M.psi_deriv(self.sresid))
k = 1 + (self.df_model + 1) / self.nobs * var_psiprime / m ** 2
if model.cov == "H1":
ss_psi = np.sum(model.M.psi(self.sresid) ** 2)
s_psi_deriv = np.sum(model.M.psi_deriv(self.sresid))
return k ** 2 * (1 / self.df_resid * ss_psi * self.scale ** 2) /\
((1 / self.nobs * s_psi_deriv) ** 2) *\
model.normalized_cov_params
else:
W = np.dot(model.M.psi_deriv(self.sresid) * model.exog.T,
model.exog)
W_inv = np.linalg.inv(W)
# [W_jk]^-1 = [SUM(psi_deriv(Sr_i)*x_ij*x_jk)]^-1
# where Sr are the standardized residuals
if model.cov == "H2":
# These are correct, based on Huber (1973) 8.13
return k * (1 / self.df_resid) * np.sum(
model.M.psi(self.sresid) ** 2) * self.scale ** 2 \
/ ((1 / self.nobs) *
np.sum(model.M.psi_deriv(self.sresid))) * W_inv
elif model.cov == "H3":
return k ** -1 * 1 / self.df_resid * np.sum(
model.M.psi(self.sresid) ** 2) * self.scale ** 2 \
* np.dot(
np.dot(W_inv, np.dot(model.exog.T, model.exog)),
W_inv)
@cache_readonly
def pvalues(self):
return stats.norm.sf(np.abs(self.tvalues)) * 2
@cache_readonly
def bse(self):
return np.sqrt(np.diag(self.bcov_scaled))
@cache_readonly
def chisq(self):
return (self.params / self.bse) ** 2
[docs] def summary(self, yname=None, xname=None, title=0, alpha=.05,
return_fmt='text'):
"""
This is for testing the new summary setup
"""
top_left = [('Dep. Variable:', None),
('Model:', None),
('Method:', ['IRLS']),
('Norm:', [self.fit_options['norm']]),
('Scale Est.:', [self.fit_options['scale_est']]),
('Cov Type:', [self.fit_options['cov']]),
('Date:', None),
('Time:', None),
('No. Iterations:', ["%d" % self.fit_history['iteration']])
]
top_right = [('No. Observations:', None),
('Df Residuals:', None),
('Df Model:', None)
]
if title is not None:
title = "Robust linear Model Regression Results"
# boiler plate
from statsmodels.iolib.summary import Summary
smry = Summary()
smry.add_table_2cols(self, gleft=top_left, gright=top_right,
yname=yname, xname=xname, title=title)
smry.add_table_params(self, yname=yname, xname=xname, alpha=alpha,
use_t=self.use_t)
# add warnings/notes, added to text format only
etext = []
wstr = ("If the model instance has been used for another fit with "
"different fit parameters, then the fit options might not be "
"the correct ones anymore .")
etext.append(wstr)
if etext:
smry.add_extra_txt(etext)
return smry
[docs] def summary2(self, xname=None, yname=None, title=None, alpha=.05,
float_format="%.4f"):
"""Experimental summary function for regression results
Parameters
----------
yname : str
Name of the dependent variable (optional)
xname : list[str], optional
Names for the exogenous variables. Default is `var_##` for ## in
the number of regressors. Must match the number of parameters
in the model
title : str, optional
Title for the top table. If not None, then this replaces the
default title
alpha : float
significance level for the confidence intervals
float_format : str
print format for floats in parameters summary
Returns
-------
smry : Summary instance
this holds the summary tables and text, which can be printed or
converted to various output formats.
See Also
--------
statsmodels.iolib.summary2.Summary : class to hold summary results
"""
from statsmodels.iolib import summary2
smry = summary2.Summary()
smry.add_base(results=self, alpha=alpha, float_format=float_format,
xname=xname, yname=yname, title=title)
return smry
class RLMResultsWrapper(lm.RegressionResultsWrapper):
pass
wrap.populate_wrapper(RLMResultsWrapper, RLMResults) # noqa:E305