linear_model.py

# TODO: Determine which tests are valid for GLSAR, and under what conditions
# TODO: Fix issue with constant and GLS
# TODO: GLS: add options Iterative GLS, for iterative fgls if sigma is None
# TODO: GLS: default if sigma is none should be two-step GLS
# TODO: Check nesting when performing model based tests, lr, wald, lm
"""
This module implements standard regression models:

Generalized Least Squares (GLS)
Ordinary Least Squares (OLS)
Weighted Least Squares (WLS)
Generalized Least Squares with autoregressive error terms GLSAR(p)

Models are specified with an endogenous response variable and an
exogenous design matrix and are fit using their `fit` method.

Subclasses that have more complicated covariance matrices
should write over the 'whiten' method as the fit method
prewhitens the response by calling 'whiten'.

General reference for regression models:

D. C. Montgomery and E.A. Peck. "Introduction to Linear Regression
    Analysis." 2nd. Ed., Wiley, 1992.

Econometrics references for regression models:

R. Davidson and J.G. MacKinnon.  "Econometric Theory and Methods," Oxford,
    2004.

W. Green.  "Econometric Analysis," 5th ed., Pearson, 2003.
"""

from __future__ import print_function

from statsmodels.compat.python import lrange, lzip, range

import numpy as np
from scipy.linalg import toeplitz
from scipy import stats
from scipy import optimize

from statsmodels.compat.numpy import np_matrix_rank
from statsmodels.tools.tools import add_constant, chain_dot, pinv_extended
from statsmodels.tools.decorators import (resettable_cache,
                                          cache_readonly,
                                          cache_writable)
import statsmodels.base.model as base
import statsmodels.base.wrapper as wrap
from statsmodels.emplike.elregress import _ELRegOpts
import warnings
from statsmodels.tools.sm_exceptions import InvalidTestWarning

# need import in module instead of lazily to copy `__doc__`
from statsmodels.regression._prediction import PredictionResults
from . import _prediction as pred

__docformat__ = 'restructuredtext en'

__all__ = ['GLS', 'WLS', 'OLS', 'GLSAR', 'PredictionResults']


_fit_regularized_doc =\
        r"""
        Return a regularized fit to a linear regression model.

        Parameters
        ----------
        method : string
            Only the 'elastic_net' approach is currently implemented.
        alpha : scalar or array-like
            The penalty weight.  If a scalar, the same penalty weight
            applies to all variables in the model.  If a vector, it
            must have the same length as `params`, and contains a
            penalty weight for each coefficient.
        L1_wt: scalar
            The fraction of the penalty given to the L1 penalty term.
            Must be between 0 and 1 (inclusive).  If 0, the fit is a
            ridge fit, if 1 it is a lasso fit.
        start_params : array-like
            Starting values for ``params``.
        profile_scale : bool
            If True the penalized fit is computed using the profile
            (concentrated) log-likelihood for the Gaussian model.
            Otherwise the fit uses the residual sum of squares.
        refit : bool
            If True, the model is refit using only the variables that
            have non-zero coefficients in the regularized fit.  The
            refitted model is not regularized.
        distributed : bool
            If True, the model uses distributed methods for fitting,
            will raise an error if True and partitions is None.
        generator : function
            generator used to partition the model, allows for handling
            of out of memory/parallel computing.
        partitions : scalar
            The number of partitions desired for the distributed
            estimation.
        threshold : scalar or array-like
            The threshold below which coefficients are zeroed out,
            only used for distributed estimation

        Returns
        -------
        A RegularizedResults instance.

        Notes
        -----

        The elastic net approach closely follows that implemented in
        the glmnet package in R.  The penalty is a combination of L1
        and L2 penalties.

        The function that is minimized is:

        .. math::

            0.5*RSS/n + alpha*((1-L1\_wt)*|params|_2^2/2 + L1\_wt*|params|_1)

        where RSS is the usual regression sum of squares, n is the
        sample size, and :math:`|*|_1` and :math:`|*|_2` are the L1 and L2
        norms.

        For WLS and GLS, the RSS is calculated using the whitened endog and
        exog data.

        Post-estimation results are based on the same data used to
        select variables, hence may be subject to overfitting biases.

        The elastic_net method uses the following keyword arguments:

        maxiter : int
            Maximum number of iterations
        cnvrg_tol : float
            Convergence threshold for line searches
        zero_tol : float
            Coefficients below this threshold are treated as zero.

        References
        ----------
        Friedman, Hastie, Tibshirani (2008).  Regularization paths for
        generalized linear models via coordinate descent.  Journal of
        Statistical Software 33(1), 1-22 Feb 2010.
        """


def _get_sigma(sigma, nobs):
    """
    Returns sigma (matrix, nobs by nobs) for GLS and the inverse of its
    Cholesky decomposition.  Handles dimensions and checks integrity.
    If sigma is None, returns None, None. Otherwise returns sigma,
    cholsigmainv.
    """
    if sigma is None:
        return None, None
    sigma = np.asarray(sigma).squeeze()
    if sigma.ndim == 0:
        sigma = np.repeat(sigma, nobs)
    if sigma.ndim == 1:
        if sigma.shape != (nobs,):
            raise ValueError("Sigma must be a scalar, 1d of length %s or a 2d "
                             "array of shape %s x %s" % (nobs, nobs, nobs))
        cholsigmainv = 1/np.sqrt(sigma)
    else:
        if sigma.shape != (nobs, nobs):
            raise ValueError("Sigma must be a scalar, 1d of length %s or a 2d "
                             "array of shape %s x %s" % (nobs, nobs, nobs))
        cholsigmainv = np.linalg.cholesky(np.linalg.inv(sigma)).T
    return sigma, cholsigmainv


class RegressionModel(base.LikelihoodModel):
    """
    Base class for linear regression models. Should not be directly called.

    Intended for subclassing.
    """
    def __init__(self, endog, exog, **kwargs):
        super(RegressionModel, self).__init__(endog, exog, **kwargs)
        self._data_attr.extend(['pinv_wexog', 'wendog', 'wexog', 'weights'])

    def initialize(self):
        self.wexog = self.whiten(self.exog)
        self.wendog = self.whiten(self.endog)
        # overwrite nobs from class Model:
        self.nobs = float(self.wexog.shape[0])

        self._df_model = None
        self._df_resid = None
        self.rank = None

    @property
    def df_model(self):
        """
        The model degree of freedom, defined as the rank of the regressor
        matrix minus 1 if a constant is included.
        """
        if self._df_model is None:
            if self.rank is None:
                self.rank = np_matrix_rank(self.exog)
            self._df_model = float(self.rank - self.k_constant)
        return self._df_model

    @df_model.setter
    def df_model(self, value):
        self._df_model = value

    @property
    def df_resid(self):
        """
        The residual degree of freedom, defined as the number of observations
        minus the rank of the regressor matrix.
        """

        if self._df_resid is None:
            if self.rank is None:
                self.rank = np_matrix_rank(self.exog)
            self._df_resid = self.nobs - self.rank
        return self._df_resid

    @df_resid.setter
    def df_resid(self, value):
        self._df_resid = value

    def whiten(self, X):
        raise NotImplementedError("Subclasses should implement.")

    def fit(self, method="pinv", cov_type='nonrobust', cov_kwds=None,
            use_t=None, **kwargs):
        """
        Full fit of the model.

        The results include an estimate of covariance matrix, (whitened)
        residuals and an estimate of scale.

        Parameters
        ----------
        method : str, optional
            Can be "pinv", "qr".  "pinv" uses the Moore-Penrose pseudoinverse
            to solve the least squares problem. "qr" uses the QR
            factorization.
        cov_type : str, optional
            See `regression.linear_model.RegressionResults` for a description
            of the available covariance estimators
        cov_kwds : list or None, optional
            See `linear_model.RegressionResults.get_robustcov_results` for a
            description required keywords for alternative covariance estimators
        use_t : bool, optional
            Flag indicating to use the Student's t distribution when computing
            p-values.  Default behavior depends on cov_type. See
            `linear_model.RegressionResults.get_robustcov_results` for
            implementation details.

        Returns
        -------
        A RegressionResults class instance.

        See Also
        ---------
        regression.linear_model.RegressionResults
        regression.linear_model.RegressionResults.get_robustcov_results

        Notes
        -----
        The fit method uses the pseudoinverse of the design/exogenous variables
        to solve the least squares minimization.
        """
        if method == "pinv":
            if not (hasattr(self, 'pinv_wexog') and
                    hasattr(self, 'normalized_cov_params') and
                    hasattr(self, 'rank')):

                self.pinv_wexog, singular_values = pinv_extended(self.wexog)
                self.normalized_cov_params = np.dot(
                    self.pinv_wexog, np.transpose(self.pinv_wexog))

                # Cache these singular values for use later.
                self.wexog_singular_values = singular_values
                self.rank = np_matrix_rank(np.diag(singular_values))

            beta = np.dot(self.pinv_wexog, self.wendog)

        elif method == "qr":
            if not (hasattr(self, 'exog_Q') and
                    hasattr(self, 'exog_R') and
                    hasattr(self, 'normalized_cov_params') and
                    hasattr(self, 'rank')):
                Q, R = np.linalg.qr(self.wexog)
                self.exog_Q, self.exog_R = Q, R
                self.normalized_cov_params = np.linalg.inv(np.dot(R.T, R))

                # Cache singular values from R.
                self.wexog_singular_values = np.linalg.svd(R, 0, 0)
                self.rank = np_matrix_rank(R)
            else:
                Q, R = self.exog_Q, self.exog_R

            # used in ANOVA
            self.effects = effects = np.dot(Q.T, self.wendog)
            beta = np.linalg.solve(R, effects)
        else:
            raise ValueError('method has to be "pinv" or "qr"')

        if self._df_model is None:
            self._df_model = float(self.rank - self.k_constant)
        if self._df_resid is None:
            self.df_resid = self.nobs - self.rank

        results_class = getattr(self, '_results_class', None)
        if results_class is not None:
            lfit = results_class(
                self, beta,
                normalized_cov_params=self.normalized_cov_params,
                cov_type=cov_type, cov_kwds=cov_kwds, use_t=use_t,
                **kwargs)
        elif isinstance(self, OLS):
            lfit = OLSResults(
                self, beta,
                normalized_cov_params=self.normalized_cov_params,
                cov_type=cov_type, cov_kwds=cov_kwds, use_t=use_t)
        else:
            lfit = RegressionResults(
                self, beta,
                normalized_cov_params=self.normalized_cov_params,
                cov_type=cov_type, cov_kwds=cov_kwds, use_t=use_t,
                **kwargs)
        return RegressionResultsWrapper(lfit)

    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

        Notes
        -----
        If the model has not yet been fit, params is not optional.
        """
        # JP: this doesn't look correct for GLMAR
        # SS: it needs its own predict method

        if exog is None:
            exog = self.exog

        return np.dot(exog, params)

    def get_distribution(self, params, scale, exog=None, dist_class=None):
        """
        Returns a random number generator for the predictive distribution.

        Parameters
        ----------
        params : array-like
            The model parameters (regression coefficients).
        scale : scalar
            The variance parameter.
        exog : array-like
            The predictor variable matrix.
        dist_class : class
            A random number generator class.  Must take 'loc' and 'scale'
            as arguments and return a random number generator implementing
            an ``rvs`` method for simulating random values. Defaults to Gaussian.

        Returns
        -------
        gen
            Frozen random number generator object with mean and variance
            determined by the fitted linear model.  Use the ``rvs`` method
            to generate random values.

        Notes
        -----
        Due to the behavior of ``scipy.stats.distributions objects``,
        the returned random number generator must be called with
        ``gen.rvs(n)`` where ``n`` is the number of observations in
        the data set used to fit the model.  If any other value is
        used for ``n``, misleading results will be produced.
        """
        fit = self.predict(params, exog)
        if dist_class is None:
            from scipy.stats.distributions import norm
            dist_class = norm
        gen = dist_class(loc=fit, scale=np.sqrt(scale))
        return gen


class GLS(RegressionModel):
    __doc__ = r"""
    Generalized least squares model with a general covariance structure.

    %(params)s
    sigma : scalar or array
        `sigma` is the weighting matrix of the covariance.
        The default is None for no scaling.  If `sigma` is a scalar, it is
        assumed that `sigma` is an n x n diagonal matrix with the given
        scalar, `sigma` as the value of each diagonal element.  If `sigma`
        is an n-length vector, then `sigma` is assumed to be a diagonal
        matrix with the given `sigma` on the diagonal.  This should be the
        same as WLS.
    %(extra_params)s

    **Attributes**

    pinv_wexog : array
        `pinv_wexog` is the p x n Moore-Penrose pseudoinverse of `wexog`.
    cholsimgainv : array
        The transpose of the Cholesky decomposition of the pseudoinverse.
    df_model : float
        p - 1, where p is the number of regressors including the intercept.
        of freedom.
    df_resid : float
        Number of observations n less the number of parameters p.
    llf : float
        The value of the likelihood function of the fitted model.
    nobs : float
        The number of observations n.
    normalized_cov_params : array
        p x p array :math:`(X^{T}\Sigma^{-1}X)^{-1}`
    results : RegressionResults instance
        A property that returns the RegressionResults class if fit.
    sigma : array
        `sigma` is the n x n covariance structure of the error terms.
    wexog : array
        Design matrix whitened by `cholsigmainv`
    wendog : array
        Response variable whitened by `cholsigmainv`

    Notes
    -----
    If sigma is a function of the data making one of the regressors
    a constant, then the current postestimation statistics will not be correct.


    Examples
    --------
    >>> import numpy as np
    >>> import statsmodels.api as sm
    >>> data = sm.datasets.longley.load(as_pandas=False)
    >>> data.exog = sm.add_constant(data.exog)
    >>> ols_resid = sm.OLS(data.endog, data.exog).fit().resid
    >>> res_fit = sm.OLS(ols_resid[1:], ols_resid[:-1]).fit()
    >>> rho = res_fit.params

    `rho` is a consistent estimator of the correlation of the residuals from
    an OLS fit of the longley data.  It is assumed that this is the true rho
    of the AR process data.

    >>> from scipy.linalg import toeplitz
    >>> order = toeplitz(np.arange(16))
    >>> sigma = rho**order

    `sigma` is an n x n matrix of the autocorrelation structure of the
    data.

    >>> gls_model = sm.GLS(data.endog, data.exog, sigma=sigma)
    >>> gls_results = gls_model.fit()
    >>> print(gls_results.summary())

    """ % {'params': base._model_params_doc,
           'extra_params': base._missing_param_doc + base._extra_param_doc}

    def __init__(self, endog, exog, sigma=None, missing='none', hasconst=None,
                 **kwargs):
        # TODO: add options igls, for iterative fgls if sigma is None
        # TODO: default if sigma is none should be two-step GLS
        sigma, cholsigmainv = _get_sigma(sigma, len(endog))

        super(GLS, self).__init__(endog, exog, missing=missing,
                                  hasconst=hasconst, sigma=sigma,
                                  cholsigmainv=cholsigmainv, **kwargs)

        # store attribute names for data arrays
        self._data_attr.extend(['sigma', 'cholsigmainv'])

    def whiten(self, X):
        """
        GLS whiten method.

        Parameters
        -----------
        X : array-like
            Data to be whitened.

        Returns
        -------
        np.dot(cholsigmainv,X)

        See Also
        --------
        regression.GLS
        """
        X = np.asarray(X)
        if self.sigma is None or self.sigma.shape == ():
            return X
        elif self.sigma.ndim == 1:
            if X.ndim == 1:
                return X * self.cholsigmainv
            else:
                return X * self.cholsigmainv[:, None]
        else:
            return np.dot(self.cholsigmainv, X)

    def loglike(self, params):
        """
        Returns the value of the Gaussian log-likelihood function at params.

        Given the whitened design matrix, the log-likelihood is evaluated
        at the parameter vector `params` for the dependent variable `endog`.

        Parameters
        ----------
        params : array-like
            The parameter estimates

        Returns
        -------
        loglike : float
            The value of the log-likelihood function for a GLS Model.


        Notes
        -----
        The log-likelihood function for the normal distribution is

        .. math:: -\\frac{n}{2}\\log\\left(\\left(Y-\\hat{Y}\\right)^{\\prime}\\left(Y-\\hat{Y}\\right)\\right)-\\frac{n}{2}\\left(1+\\log\\left(\\frac{2\\pi}{n}\\right)\\right)-\\frac{1}{2}\\log\\left(\\left|\\Sigma\\right|\\right)

        Y and Y-hat are whitened.

        """
        # TODO: combine this with OLS/WLS loglike and add _det_sigma argument
        nobs2 = self.nobs / 2.0
        SSR = np.sum((self.wendog - np.dot(self.wexog, params))**2, axis=0)
        llf = -np.log(SSR) * nobs2      # concentrated likelihood
        llf -= (1+np.log(np.pi/nobs2))*nobs2  # with likelihood constant
        if np.any(self.sigma):
            # FIXME: robust-enough check? unneeded if _det_sigma gets defined
            if self.sigma.ndim == 2:
                det = np.linalg.slogdet(self.sigma)
                llf -= .5*det[1]
            else:
                llf -= 0.5*np.sum(np.log(self.sigma))
            # with error covariance matrix
        return llf

    def hessian_factor(self, params, scale=None, observed=True):
        """Weights for calculating Hessian

        Parameters
        ----------
        params : ndarray
            parameter at which Hessian is evaluated
        scale : None or float
            If scale is None, then the default scale will be calculated.
            Default scale is defined by `self.scaletype` and set in fit.
            If scale is not None, then it is used as a fixed scale.
        observed : bool
            If True, then the observed Hessian is returned. If false then the
            expected information matrix is returned.

        Returns
        -------
        hessian_factor : ndarray, 1d
            A 1d weight vector used in the calculation of the Hessian.
            The hessian is obtained by `(exog.T * hessian_factor).dot(exog)`
        """

        if self.sigma is None or self.sigma.shape == ():
            return np.ones(self.exog.shape[0])
        elif self.sigma.ndim == 1:
            return self.cholsigmainv
        else:
            return np.diag(self.cholsigmainv)

    def fit_regularized(self, method="elastic_net", alpha=0.,
                        L1_wt=1., start_params=None, profile_scale=False,
                        refit=False, **kwargs):
        # Docstring attached below

        # Need to adjust since RSS/n term in elastic net uses nominal
        # n in denominator
        if self.sigma is not None:
            alpha = alpha * np.sum(1 / np.diag(self.sigma)) / len(self.endog)

        rslt = OLS(self.wendog, self.wexog).fit_regularized(
            method=method, alpha=alpha,
            L1_wt=L1_wt,
            start_params=start_params,
            profile_scale=profile_scale,
            refit=refit, **kwargs)

        from statsmodels.base.elastic_net import (
            RegularizedResults, RegularizedResultsWrapper)
        rrslt = RegularizedResults(self, rslt.params)
        return RegularizedResultsWrapper(rrslt)

    fit_regularized.__doc__ = _fit_regularized_doc


class WLS(RegressionModel):
    __doc__ = """
    A regression model with diagonal but non-identity covariance structure.

    The weights are presumed to be (proportional to) the inverse of
    the variance of the observations.  That is, if the variables are
    to be transformed by 1/sqrt(W) you must supply weights = 1/W.

    %(params)s
    weights : array-like, optional
        1d array of weights.  If you supply 1/W then the variables are
        pre- multiplied by 1/sqrt(W).  If no weights are supplied the
        default value is 1 and WLS results are the same as OLS.
    %(extra_params)s

    Attributes
    ----------
    weights : array
        The stored weights supplied as an argument.

    See regression.GLS

    Examples
    ---------
    >>> import numpy as np
    >>> import statsmodels.api as sm
    >>> Y = [1,3,4,5,2,3,4]
    >>> X = range(1,8)
    >>> X = sm.add_constant(X)
    >>> wls_model = sm.WLS(Y,X, weights=list(range(1,8)))
    >>> results = wls_model.fit()
    >>> results.params
    array([ 2.91666667,  0.0952381 ])
    >>> results.tvalues
    array([ 2.0652652 ,  0.35684428])
    >>> print(results.t_test([1, 0]))
    <T test: effect=array([ 2.91666667]), sd=array([[ 1.41224801]]), t=array([[ 2.0652652]]), p=array([[ 0.04690139]]), df_denom=5>
    >>> print(results.f_test([0, 1]))
    <F test: F=array([[ 0.12733784]]), p=[[ 0.73577409]], df_denom=5, df_num=1>

    Notes
    -----
    If the weights are a function of the data, then the post estimation
    statistics such as fvalue and mse_model might not be correct, as the
    package does not yet support no-constant regression.
    """ % {'params': base._model_params_doc,
           'extra_params': base._missing_param_doc + base._extra_param_doc}

    def __init__(self, endog, exog, weights=1., missing='none', hasconst=None,
                 **kwargs):
        weights = np.array(weights)
        if weights.shape == ():
            if (missing == 'drop' and 'missing_idx' in kwargs and
                    kwargs['missing_idx'] is not None):
                # patsy may have truncated endog
                weights = np.repeat(weights, len(kwargs['missing_idx']))
            else:
                weights = np.repeat(weights, len(endog))
        # handle case that endog might be of len == 1
        if len(weights) == 1:
            weights = np.array([weights.squeeze()])
        else:
            weights = weights.squeeze()
        super(WLS, self).__init__(endog, exog, missing=missing,
                                  weights=weights, hasconst=hasconst, **kwargs)
        nobs = self.exog.shape[0]
        weights = self.weights
        # Experimental normalization of weights
        weights = weights / np.sum(weights) * nobs
        if weights.size != nobs and weights.shape[0] != nobs:
            raise ValueError('Weights must be scalar or same length as design')

    def whiten(self, X):
        """
        Whitener for WLS model, multiplies each column by sqrt(self.weights)

        Parameters
        ----------
        X : array-like
            Data to be whitened

        Returns
        -------
        whitened : array-like
            sqrt(weights)*X
        """

        X = np.asarray(X)
        if X.ndim == 1:
            return X * np.sqrt(self.weights)
        elif X.ndim == 2:
            return np.sqrt(self.weights)[:, None]*X

    def loglike(self, params):
        """
        Returns the value of the gaussian log-likelihood function at params.

        Given the whitened design matrix, the log-likelihood is evaluated
        at the parameter vector `params` for the dependent variable `Y`.

        Parameters
        ----------
        params : array-like
            The parameter estimates.

        Returns
        -------
        llf : float
            The value of the log-likelihood function for a WLS Model.

        Notes
        --------
        .. math:: -\\frac{n}{2}\\log\\left(Y-\\hat{Y}\\right)-\\frac{n}{2}\\left(1+\\log\\left(\\frac{2\\pi}{n}\\right)\\right)-\\frac{1}{2}log\\left(\\left|W\\right|\\right)

        where :math:`W` is a diagonal matrix
        """
        nobs2 = self.nobs / 2.0
        SSR = np.sum((self.wendog - np.dot(self.wexog, params))**2, axis=0)
        llf = -np.log(SSR) * nobs2      # concentrated likelihood
        llf -= (1+np.log(np.pi/nobs2))*nobs2  # with constant
        llf += 0.5 * np.sum(np.log(self.weights))
        return llf

    def hessian_factor(self, params, scale=None, observed=True):
        """Weights for calculating Hessian

        Parameters
        ----------
        params : ndarray
            parameter at which Hessian is evaluated
        scale : None or float
            If scale is None, then the default scale will be calculated.
            Default scale is defined by `self.scaletype` and set in fit.
            If scale is not None, then it is used as a fixed scale.
        observed : bool
            If True, then the observed Hessian is returned. If false then the
            expected information matrix is returned.

        Returns
        -------
        hessian_factor : ndarray, 1d
            A 1d weight vector used in the calculation of the Hessian.
            The hessian is obtained by `(exog.T * hessian_factor).dot(exog)`
        """

        return self.weights

    def fit_regularized(self, method="elastic_net", alpha=0.,
                        L1_wt=1., start_params=None, profile_scale=False,
                        refit=False, **kwargs):
        # Docstring attached below

        # Need to adjust since RSS/n in elastic net uses nominal n in
        # denominator
        alpha = alpha * np.sum(self.weights) / len(self.weights)

        rslt = OLS(self.wendog, self.wexog).fit_regularized(
            method=method, alpha=alpha,
            L1_wt=L1_wt,
            start_params=start_params,
            profile_scale=profile_scale,
            refit=refit, **kwargs)

        from statsmodels.base.elastic_net import (
            RegularizedResults, RegularizedResultsWrapper)
        rrslt = RegularizedResults(self, rslt.params)
        return RegularizedResultsWrapper(rrslt)

    fit_regularized.__doc__ = _fit_regularized_doc


class OLS(WLS):
    __doc__ = """
    A simple ordinary least squares model.

    %(params)s
    %(extra_params)s

    Attributes
    ----------
    weights : scalar
        Has an attribute weights = array(1.0) due to inheritance from WLS.

    See Also
    --------
    GLS

    Examples
    --------
    >>> import numpy as np
    >>>
    >>> import statsmodels.api as sm
    >>>
    >>> Y = [1,3,4,5,2,3,4]
    >>> X = range(1,8)
    >>> X = sm.add_constant(X)
    >>>
    >>> model = sm.OLS(Y,X)
    >>> results = model.fit()
    >>> results.params
    array([ 2.14285714,  0.25      ])
    >>> results.tvalues
    array([ 1.87867287,  0.98019606])
    >>> print(results.t_test([1, 0]))
    <T test: effect=array([ 2.14285714]), sd=array([[ 1.14062282]]), t=array([[ 1.87867287]]), p=array([[ 0.05953974]]), df_denom=5>
    >>> print(results.f_test(np.identity(2)))
    <F test: F=array([[ 19.46078431]]), p=[[ 0.00437251]], df_denom=5, df_num=2>

    Notes
    -----
    No constant is added by the model unless you are using formulas.
    """ % {'params': base._model_params_doc,
           'extra_params': base._missing_param_doc + base._extra_param_doc}

    # TODO: change example to use datasets.  This was the point of datasets!
    def __init__(self, endog, exog=None, missing='none', hasconst=None,
                 **kwargs):
        super(OLS, self).__init__(endog, exog, missing=missing,
                                  hasconst=hasconst, **kwargs)
        if "weights" in self._init_keys:
            self._init_keys.remove("weights")

    def loglike(self, params, scale=None):
        """
        The likelihood function for the OLS model.

        Parameters
        ----------
        params : array-like
            The coefficients with which to estimate the log-likelihood.
        scale : float or None
            If None, return the profile (concentrated) log likelihood
            (profiled over the scale parameter), else return the
            log-likelihood using the given scale value.

        Returns
        -------
        The likelihood function evaluated at params.
        """
        nobs2 = self.nobs / 2.0
        nobs = float(self.nobs)
        resid = self.endog - np.dot(self.exog, params)
        if hasattr(self, 'offset'):
            resid -= self.offset
        ssr = np.sum(resid**2, axis=0)
        if scale is None:
            # profile log likelihood
            llf = -nobs2*np.log(2*np.pi) - nobs2*np.log(ssr / nobs) - nobs2
        else:
            # log-likelihood
            llf = -nobs2 * np.log(2 * np.pi * scale) - ssr / (2*scale)
        return llf

    def whiten(self, Y):
        """
        OLS model whitener does nothing: returns Y.
        """
        return Y

    def score(self, params, scale=None):
        """
        Evaluate the score function at a given point.

        The score corresponds to the profile (concentrated)
        log-likelihood in which the scale parameter has been profiled
        out.

        Parameters
        ----------
        params : array-like
            The parameter vector at which the score function is
            computed.
        scale : float or None
            If None, return the profile (concentrated) log likelihood
            (profiled over the scale parameter), else return the
            log-likelihood using the given scale value.

        Returns
        -------
        The score vector.
        """

        if not hasattr(self, "_wexog_xprod"):
            self._setup_score_hess()

        xtxb = np.dot(self._wexog_xprod, params)
        sdr = -self._wexog_x_wendog + xtxb

        if scale is None:
            ssr = self._wendog_xprod - 2 * np.dot(self._wexog_x_wendog.T,
                                                  params)
            ssr += np.dot(params, xtxb)
            return -self.nobs * sdr / ssr
        else:
            return -sdr / scale

    def _setup_score_hess(self):
        y = self.wendog
        if hasattr(self, 'offset'):
            y = y - self.offset
        self._wendog_xprod = np.sum(y * y)
        self._wexog_xprod = np.dot(self.wexog.T, self.wexog)
        self._wexog_x_wendog = np.dot(self.wexog.T, y)

    def hessian(self, params, scale=None):
        """
        Evaluate the Hessian function at a given point.

        Parameters
        ----------
        params : array-like
            The parameter vector at which the Hessian is computed.
        scale : float or None
            If None, return the profile (concentrated) log likelihood
            (profiled over the scale parameter), else return the
            log-likelihood using the given scale value.

        Returns
        -------
        The Hessian matrix.
        """

        if not hasattr(self, "_wexog_xprod"):
            self._setup_score_hess()

        xtxb = np.dot(self._wexog_xprod, params)

        if scale is None:
            ssr = self._wendog_xprod - 2 * np.dot(self._wexog_x_wendog.T,
                                                  params)
            ssr += np.dot(params, xtxb)
            ssrp = -2*self._wexog_x_wendog + 2*xtxb
            hm = self._wexog_xprod / ssr - np.outer(ssrp, ssrp) / ssr**2
            return -self.nobs * hm / 2
        else:
            return -self._wexog_xprod / scale

    def hessian_factor(self, params, scale=None, observed=True):
        """Weights for calculating Hessian

        Parameters
        ----------
        params : ndarray
            parameter at which Hessian is evaluated
        scale : None or float
            If scale is None, then the default scale will be calculated.
            Default scale is defined by `self.scaletype` and set in fit.
            If scale is not None, then it is used as a fixed scale.
        observed : bool
            If True, then the observed Hessian is returned. If false then the
            expected information matrix is returned.

        Returns
        -------
        hessian_factor : ndarray, 1d
            A 1d weight vector used in the calculation of the Hessian.
            The hessian is obtained by `(exog.T * hessian_factor).dot(exog)`
        """

        return np.ones(self.exog.shape[0])

    def fit_regularized(self, method="elastic_net", alpha=0.,
                        L1_wt=1., start_params=None, profile_scale=False,
                        refit=False, **kwargs):
        # Docstring attached below

        from statsmodels.base.elastic_net import fit_elasticnet

        if L1_wt == 0:
            return self._fit_ridge(alpha)

        # In the future we could add support for other penalties, e.g. SCAD.
        if method != "elastic_net":
            raise ValueError("method for fit_regularized must be elastic_net")

        # Set default parameters.
        defaults = {"maxiter":  50, "cnvrg_tol": 1e-10,
                    "zero_tol": 1e-10}
        defaults.update(kwargs)

        # If a scale parameter is passed in, the non-profile
        # likelihood (residual sum of squares divided by -2) is used,
        # otherwise the profile likelihood is used.
        if profile_scale:
            loglike_kwds = {}
            score_kwds = {}
            hess_kwds = {}
        else:
            loglike_kwds = {"scale": 1}
            score_kwds = {"scale": 1}
            hess_kwds = {"scale": 1}

        return fit_elasticnet(self, method=method,
                              alpha=alpha,
                              L1_wt=L1_wt,
                              start_params=start_params,
                              loglike_kwds=loglike_kwds,
                              score_kwds=score_kwds,
                              hess_kwds=hess_kwds,
                              refit=refit,
                              check_step=False,
                              **defaults)

    fit_regularized.__doc__ = _fit_regularized_doc

    def _fit_ridge(self, alpha):
        """
        Fit a linear model using ridge regression.

        Parameters
        ----------
        alpha : scalar or array-like
            The penalty weight.  If a scalar, the same penalty weight
            applies to all variables in the model.  If a vector, it
            must have the same length as `params`, and contains a
            penalty weight for each coefficient.

        Notes
        -----
        Equivalent to fit_regularized with L1_wt = 0 (but implemented
        more efficiently).
        """

        u, s, vt = np.linalg.svd(self.exog, 0)
        v = vt.T
        q = np.dot(u.T, self.endog) * s
        s2 = s * s
        if np.isscalar(alpha):
            sd = s2 + alpha * self.nobs
            params = q / sd
            params = np.dot(v, params)
        else:
            vtav = self.nobs * np.dot(vt, alpha[:, None] * v)
            d = np.diag(vtav) + s2
            np.fill_diagonal(vtav, d)
            r = np.linalg.solve(vtav, q)
            params = np.dot(v, r)

        from statsmodels.base.elastic_net import RegularizedResults
        return RegularizedResults(self, params)


class GLSAR(GLS):
    __doc__ = """
    A regression model with an AR(p) covariance structure.

    %(params)s
    rho : int
        Order of the autoregressive covariance
    %(extra_params)s

    Examples
    --------
    >>> import statsmodels.api as sm
    >>> X = range(1,8)
    >>> X = sm.add_constant(X)
    >>> Y = [1,3,4,5,8,10,9]
    >>> model = sm.GLSAR(Y, X, rho=2)
    >>> for i in range(6):
    ...     results = model.fit()
    ...     print("AR coefficients: {0}".format(model.rho))
    ...     rho, sigma = sm.regression.yule_walker(results.resid,
    ...                                            order=model.order)
    ...     model = sm.GLSAR(Y, X, rho)
    ...
    AR coefficients: [ 0.  0.]
    AR coefficients: [-0.52571491 -0.84496178]
    AR coefficients: [-0.6104153  -0.86656458]
    AR coefficients: [-0.60439494 -0.857867  ]
    AR coefficients: [-0.6048218  -0.85846157]
    AR coefficients: [-0.60479146 -0.85841922]
    >>> results.params
    array([-0.66661205,  1.60850853])
    >>> results.tvalues
    array([ -2.10304127,  21.8047269 ])
    >>> print(results.t_test([1, 0]))
    <T test: effect=array([-0.66661205]), sd=array([[ 0.31697526]]), t=array([[-2.10304127]]), p=array([[ 0.06309969]]), df_denom=3>
    >>> print(results.f_test(np.identity(2)))
    <F test: F=array([[ 1815.23061844]]), p=[[ 0.00002372]], df_denom=3, df_num=2>

    Or, equivalently

    >>> model2 = sm.GLSAR(Y, X, rho=2)
    >>> res = model2.iterative_fit(maxiter=6)
    >>> model2.rho
    array([-0.60479146, -0.85841922])

    Notes
    -----
    GLSAR is considered to be experimental.
    The linear autoregressive process of order p--AR(p)--is defined as:
    TODO
    """ % {'params': base._model_params_doc,
           'extra_params': base._missing_param_doc + base._extra_param_doc}

    def __init__(self, endog, exog=None, rho=1, missing='none', **kwargs):
        # this looks strange, interpreting rho as order if it is int
        if isinstance(rho, np.int):
            self.order = rho
            self.rho = np.zeros(self.order, np.float64)
        else:
            self.rho = np.squeeze(np.asarray(rho))
            if len(self.rho.shape) not in [0, 1]:
                raise ValueError("AR parameters must be a scalar or a vector")
            if self.rho.shape == ():
                self.rho.shape = (1,)
            self.order = self.rho.shape[0]
        if exog is None:
            # JP this looks wrong, should be a regression on constant
            # results for rho estimate now identical to yule-walker on y
            # super(AR, self).__init__(endog, add_constant(endog))
            super(GLSAR, self).__init__(endog, np.ones((endog.shape[0], 1)),
                                        missing=missing, **kwargs)
        else:
            super(GLSAR, self).__init__(endog, exog, missing=missing,
                                        **kwargs)

    def iterative_fit(self, maxiter=3, rtol=1e-4, **kwds):
        """
        Perform an iterative two-stage procedure to estimate a GLS model.

        The model is assumed to have AR(p) errors, AR(p) parameters and
        regression coefficients are estimated iteratively.

        Parameters
        ----------
        maxiter : integer, optional
            the number of iterations
        rtol : float, optional
            Relative tolerance between estimated coefficients to stop the
            estimation.  Stops if

            max(abs(last - current) / abs(last)) < rtol

        """
        # TODO: update this after going through example.
        converged = False
        i = -1  # need to initialize for maxiter < 1 (skip loop)
        history = {'params': [], 'rho': [self.rho]}
        for i in range(maxiter - 1):
            if hasattr(self, 'pinv_wexog'):
                del self.pinv_wexog
            self.initialize()
            results = self.fit()
            history['params'].append(results.params)
            if i == 0:
                last = results.params
            else:
                diff = np.max(np.abs(last - results.params) / np.abs(last))
                if diff < rtol:
                    converged = True
                    break
                last = results.params
            self.rho, _ = yule_walker(results.resid,
                                      order=self.order, df=None)
            history['rho'].append(self.rho)

        # why not another call to self.initialize
        # Use kwarg to insert history
        if not converged and maxiter > 0:
            # maxiter <= 0 just does OLS
            if hasattr(self, 'pinv_wexog'):
                del self.pinv_wexog
            self.initialize()

        # if converged then this is a duplicate fit, because we didn't
        # update rho
        results = self.fit(history=history, **kwds)
        results.iter = i + 1
        # add last fit to history, not if duplicate fit
        if not converged:
            results.history['params'].append(results.params)
            results.iter += 1

        results.converged = converged

        return results

    def whiten(self, X):
        """
        Whiten a series of columns according to an AR(p)
        covariance structure. This drops initial p observations.

        Parameters
        ----------
        X : array-like
            The data to be whitened,

        Returns
        -------
        whitened array

        """
        # TODO: notation for AR process
        X = np.asarray(X, np.float64)
        _X = X.copy()

        # the following loops over the first axis,  works for 1d and nd
        for i in range(self.order):
            _X[(i+1):] = _X[(i+1):] - self.rho[i] * X[0:-(i+1)]
        return _X[self.order:]


def yule_walker(X, order=1, method="unbiased", df=None, inv=False,
                demean=True):
    """
    Estimate AR(p) parameters from a sequence X using Yule-Walker equation.

    Unbiased or maximum-likelihood estimator (mle)

    See, for example:

    http://en.wikipedia.org/wiki/Autoregressive_moving_average_model

    Parameters
    ----------
    X : array-like
        1d array
    order : integer, optional
        The order of the autoregressive process.  Default is 1.
    method : string, optional
       Method can be 'unbiased' or 'mle' and this determines
       denominator in estimate of autocorrelation function (ACF) at
       lag k. If 'mle', the denominator is n=X.shape[0], if 'unbiased'
       the denominator is n-k.  The default is unbiased.
    df : integer, optional
       Specifies the degrees of freedom. If `df` is supplied, then it
       is assumed the X has `df` degrees of freedom rather than `n`.
       Default is None.
    inv : bool
        If inv is True the inverse of R is also returned.  Default is
        False.
    demean : bool
        True, the mean is subtracted from `X` before estimation.

    Returns
    -------
    rho
        The autoregressive coefficients
    sigma
        TODO

    Examples
    --------
    >>> import statsmodels.api as sm
    >>> from statsmodels.datasets.sunspots import load
    >>> data = load(as_pandas=False)
    >>> rho, sigma = sm.regression.yule_walker(data.endog,
    ...                                        order=4, method="mle")

    >>> rho
    array([ 1.28310031, -0.45240924, -0.20770299,  0.04794365])
    >>> sigma
    16.808022730464351

    """
    # TODO: define R better, look back at notes and technical notes on YW.
    # First link here is useful
    # http://www-stat.wharton.upenn.edu/~steele/Courses/956/ResourceDetails/YuleWalkerAndMore.htm
    method = str(method).lower()
    if method not in ["unbiased", "mle"]:
        raise ValueError("ACF estimation method must be 'unbiased' or 'MLE'")
    X = np.array(X, dtype=np.float64)
    if demean:
        X -= X.mean()                  # automatically demean's X
    n = df or X.shape[0]

    if method == "unbiased":        # this is df_resid ie., n - p
        denom = lambda k: n - k
    else:
        denom = lambda k: n
    if X.ndim > 1 and X.shape[1] != 1:
        raise ValueError("expecting a vector to estimate AR parameters")
    r = np.zeros(order+1, np.float64)
    r[0] = (X**2).sum() / denom(0)
    for k in range(1, order+1):
        r[k] = (X[0:-k] * X[k:]).sum() / denom(k)
    R = toeplitz(r[:-1])

    rho = np.linalg.solve(R, r[1:])
    sigmasq = r[0] - (r[1:]*rho).sum()
    if inv:
        return rho, np.sqrt(sigmasq), np.linalg.inv(R)
    else:
        return rho, np.sqrt(sigmasq)


def burg(endog, order=1, demean=True):
    """
    Burg's AP(p) parameter estimator

    Parameters
    ----------
    endog : array-like
        The endogenous variable
    order : int, optional
        Order of the AR.  Default is 1.
    demean : bool, optional
        Flag indicating to subtract the mean from endog before estimation

    Returns
    -------
    rho : ndarray
        AR(p) coefficients computed using Burg's algorithm
    sigma2 : float
        Estimate of the residual variance

    Notes
    -----
    AR model estimated includes a constant estimated using the sample mean.
    This value is not reported.

    References
    ----------
    Brockwell, P.J. and Davis, R.A., 2016. Introduction to time series and
        forecasting. Springer.
    """
    # Avoid circular imports
    from statsmodels.tsa.stattools import levinson_durbin_pacf, pacf_burg

    endog = np.squeeze(np.asarray(endog))
    if endog.ndim != 1:
        raise ValueError('endog must be 1-d or squeezable to 1-d.')
    order = int(order)
    if order < 1:
        raise ValueError('order must be an integer larger than 1')
    if demean:
        endog = endog - endog.mean()
    pacf, sigma = pacf_burg(endog, order, demean=demean)
    ar, _ = levinson_durbin_pacf(pacf)
    return ar, sigma[-1]


class RegressionResults(base.LikelihoodModelResults):
    r"""
    This class summarizes the fit of a linear regression model.

    It handles the output of contrasts, estimates of covariance, etc.

    Returns
    -------
    **Attributes**

    aic
        Akaike's information criteria. For a model with a constant
        :math:`-2llf + 2(df\_model + 1)`. For a model without a constant
        :math:`-2llf + 2(df\_model)`.
    bic
        Bayes' information criteria. For a model with a constant
        :math:`-2llf + \log(n)(df\_model+1)`. For a model without a constant
        :math:`-2llf + \log(n)(df\_model)`
    bse
        The standard errors of the parameter estimates.
    pinv_wexog
        See specific model class docstring
    centered_tss
        The total (weighted) sum of squares centered about the mean.
    cov_HC0
        Heteroscedasticity robust covariance matrix. See HC0_se below.
    cov_HC1
        Heteroscedasticity robust covariance matrix. See HC1_se below.
    cov_HC2
        Heteroscedasticity robust covariance matrix. See HC2_se below.
    cov_HC3
        Heteroscedasticity robust covariance matrix. See HC3_se below.
    cov_type
        Parameter covariance estimator used for standard errors and t-stats
    df_model
        Model degrees of freedom. The number of regressors `p`. Does not
        include the constant if one is present
    df_resid
        Residual degrees of freedom. `n - p - 1`, if a constant is present.
        `n - p` if a constant is not included.
    ess
        Explained sum of squares.  If a constant is present, the centered
        total sum of squares minus the sum of squared residuals. If there is
        no constant, the uncentered total sum of squares is used.
    fvalue
        F-statistic of the fully specified model.  Calculated as the mean
        squared error of the model divided by the mean squared error of the
        residuals.
    f_pvalue
        p-value of the F-statistic
    fittedvalues
        The predicted values for the original (unwhitened) design.
    het_scale
        adjusted squared residuals for heteroscedasticity robust standard
        errors. Is only available after `HC#_se` or `cov_HC#` is called.
        See HC#_se for more information.
    history
        Estimation history for iterative estimators
    HC0_se
        White's (1980) heteroskedasticity robust standard errors.
        Defined as sqrt(diag(X.T X)^(-1)X.T diag(e_i^(2)) X(X.T X)^(-1)
        where e_i = resid[i]
        HC0_se is a cached property.
        When HC0_se or cov_HC0 is called the RegressionResults instance will
        then have another attribute `het_scale`, which is in this case is just
        resid**2.
    HC1_se
        MacKinnon and White's (1985) alternative heteroskedasticity robust
        standard errors.
        Defined as sqrt(diag(n/(n-p)*HC_0)
        HC1_see is a cached property.
        When HC1_se or cov_HC1 is called the RegressionResults instance will
        then have another attribute `het_scale`, which is in this case is
        n/(n-p)*resid**2.
    HC2_se
        MacKinnon and White's (1985) alternative heteroskedasticity robust
        standard errors.
        Defined as (X.T X)^(-1)X.T diag(e_i^(2)/(1-h_ii)) X(X.T X)^(-1)
        where h_ii = x_i(X.T X)^(-1)x_i.T
        HC2_see is a cached property.
        When HC2_se or cov_HC2 is called the RegressionResults instance will
        then have another attribute `het_scale`, which is in this case is
        resid^(2)/(1-h_ii).
    HC3_se
        MacKinnon and White's (1985) alternative heteroskedasticity robust
        standard errors.
        Defined as (X.T X)^(-1)X.T diag(e_i^(2)/(1-h_ii)^(2)) X(X.T X)^(-1)
        where h_ii = x_i(X.T X)^(-1)x_i.T
        HC3_see is a cached property.
        When HC3_se or cov_HC3 is called the RegressionResults instance will
        then have another attribute `het_scale`, which is in this case is
        resid^(2)/(1-h_ii)^(2).
    model
        A pointer to the model instance that called fit() or results.
    mse_model
        Mean squared error the model. This is the explained sum of
        squares divided by the model degrees of freedom.
    mse_resid
        Mean squared error of the residuals.  The sum of squared
        residuals divided by the residual degrees of freedom.
    mse_total
        Total mean squared error.  Defined as the uncentered total sum
        of squares divided by n the number of observations.
    nobs
        Number of observations n.
    normalized_cov_params
        See specific model class docstring
    params
        The linear coefficients that minimize the least squares
        criterion.  This is usually called Beta for the classical
        linear model.
    pvalues
        The two-tailed p values for the t-stats of the params.
    resid
        The residuals of the model.
    resid_pearson
        `wresid` normalized to have unit variance.
    rsquared
        R-squared of a model with an intercept.  This is defined here
        as 1 - `ssr`/`centered_tss` if the constant is included in the
        model and 1 - `ssr`/`uncentered_tss` if the constant is
        omitted.
    rsquared_adj
        Adjusted R-squared.  This is defined here as 1 -
        (`nobs`-1)/`df_resid` * (1-`rsquared`) if a constant is
        included and 1 - `nobs`/`df_resid` * (1-`rsquared`) if no
        constant is included.
    scale
        A scale factor for the covariance matrix.  Default value is
        ssr/(n-p).  Note that the square root of `scale` is often
        called the standard error of the regression.
    ssr
        Sum of squared (whitened) residuals.
    uncentered_tss
        Uncentered sum of squares.  Sum of the squared values of the
        (whitened) endogenous response variable.
    wresid
        The residuals of the transformed/whitened regressand and
        regressor(s)
    """

    _cache = {}  # needs to be a class attribute for scale setter?

    def __init__(self, model, params, normalized_cov_params=None, scale=1.,
                 cov_type='nonrobust', cov_kwds=None, use_t=None, **kwargs):
        super(RegressionResults, self).__init__(
            model, params, normalized_cov_params, scale)

        self._cache = resettable_cache()
        if hasattr(model, 'wexog_singular_values'):
            self._wexog_singular_values = model.wexog_singular_values
        else:
            self._wexog_singular_values = None

        self.df_model = model.df_model
        self.df_resid = model.df_resid

        if cov_type == 'nonrobust':
            self.cov_type = 'nonrobust'
            self.cov_kwds = {
                'description': 'Standard Errors assume that the ' +
                'covariance matrix of the errors is correctly ' +
                'specified.'}
            if use_t is None:
                self.use_t = True  # TODO: class default
        else:
            if cov_kwds is None:
                cov_kwds = {}
            if 'use_t' in cov_kwds:
                # TODO: we want to get rid of 'use_t' in cov_kwds
                use_t_2 = cov_kwds.pop('use_t')
                if use_t is None:
                    use_t = use_t_2
                # TODO: warn or not?
            self.get_robustcov_results(cov_type=cov_type, use_self=True,
                                       use_t=use_t, **cov_kwds)
        for key in kwargs:
            setattr(self, key, kwargs[key])

    def __str__(self):
        self.summary()

    def conf_int(self, alpha=.05, cols=None):
        """
        Returns the confidence interval of the fitted parameters.

        Parameters
        ----------
        alpha : float, optional
            The `alpha` level for the confidence interval.
            ie., The default `alpha` = .05 returns a 95% confidence interval.
        cols : array-like, optional
            `cols` specifies which confidence intervals to return

        Notes
        -----
        The confidence interval is based on Student's t-distribution.
        """
        # keep method for docstring for now
        ci = super(RegressionResults, self).conf_int(alpha=alpha, cols=cols)
        return ci

    @cache_readonly
    def nobs(self):
        return float(self.model.wexog.shape[0])

    @cache_readonly
    def fittedvalues(self):
        return self.model.predict(self.params, self.model.exog)

    @cache_readonly
    def wresid(self):
        return self.model.wendog - self.model.predict(
            self.params, self.model.wexog)

    @cache_readonly
    def resid(self):
        return self.model.endog - self.model.predict(
            self.params, self.model.exog)

    # TODO: fix writable example
    @cache_writable()
    def scale(self):
        wresid = self.wresid
        return wresid.T.dot(wresid) / self.df_resid

    @cache_readonly
    def ssr(self):
        wresid = self.wresid
        return np.dot(wresid, wresid)

    @cache_readonly
    def centered_tss(self):
        model = self.model
        weights = getattr(model, 'weights', None)
        sigma = getattr(model, 'sigma', None)
        if weights is not None:
            mean = np.average(model.endog, weights=weights, axis=0)
            return np.sum(weights * (model.endog - mean).T**2, axis=-1)
        elif sigma is not None:
            # Exactly matches WLS when sigma is diagonal
            iota = np.ones_like(model.endog)
            iota = model.whiten(iota)
            mean = model.wendog.dot(iota) / iota.dot(iota)
            err = model.endog - mean
            err = model.whiten(err)
            return np.sum(err**2)
        else:
            centered_endog = model.wendog - model.wendog.mean(0)
            return np.dot(centered_endog, centered_endog)

    @cache_readonly
    def uncentered_tss(self):
        wendog = self.model.wendog
        return np.dot(wendog, wendog)

    @cache_readonly
    def ess(self):
        if self.k_constant:
            return self.centered_tss - self.ssr
        else:
            return self.uncentered_tss - self.ssr

    @cache_readonly
    def rsquared(self):
        if self.k_constant:
            return 1 - self.ssr/self.centered_tss
        else:
            return 1 - self.ssr/self.uncentered_tss

    @cache_readonly
    def rsquared_adj(self):
        return 1 - (np.divide(self.nobs - self.k_constant, self.df_resid)
                    * (1 - self.rsquared))

    @cache_readonly
    def mse_model(self):
        return self.ess/self.df_model

    @cache_readonly
    def mse_resid(self):
        return self.ssr/self.df_resid

    @cache_readonly
    def mse_total(self):
        if self.k_constant:
            return self.centered_tss / (self.df_resid + self.df_model)
        else:
            return self.uncentered_tss / (self.df_resid + self.df_model)

    @cache_readonly
    def fvalue(self):
        if hasattr(self, 'cov_type') and self.cov_type != 'nonrobust':
            # with heteroscedasticity or correlation robustness
            k_params = self.normalized_cov_params.shape[0]
            mat = np.eye(k_params)
            const_idx = self.model.data.const_idx
            # TODO: What if model includes implicit constant, e.g. all
            #       dummies but no constant regressor?
            # TODO: Restats as LM test by projecting orthogonalizing
            #       to constant?
            if self.model.data.k_constant == 1:
                # if constant is implicit, return nan see #2444
                if const_idx is None:
                    return np.nan

                idx = lrange(k_params)
                idx.pop(const_idx)
                mat = mat[idx]  # remove constant
                if mat.size == 0:  # see  #3642
                    return np.nan
            ft = self.f_test(mat)
            # using backdoor to set another attribute that we already have
            self._cache['f_pvalue'] = ft.pvalue
            return ft.fvalue
        else:
            # for standard homoscedastic case
            return self.mse_model/self.mse_resid

    @cache_readonly
    def f_pvalue(self):
        return stats.f.sf(self.fvalue, self.df_model, self.df_resid)

    @cache_readonly
    def bse(self):
        return np.sqrt(np.diag(self.cov_params()))

    @cache_readonly
    def aic(self):
        return -2 * self.llf + 2 * (self.df_model + self.k_constant)

    @cache_readonly
    def bic(self):
        return (-2 * self.llf + np.log(self.nobs) * (self.df_model +
                                                     self.k_constant))

    @cache_readonly
    def eigenvals(self):
        """
        Return eigenvalues sorted in decreasing order.
        """
        if self._wexog_singular_values is not None:
            eigvals = self._wexog_singular_values ** 2
        else:
            eigvals = np.linalg.linalg.eigvalsh(np.dot(self.model.wexog.T,
                                                       self.model.wexog))
        return np.sort(eigvals)[::-1]

    @cache_readonly
    def condition_number(self):
        """
        Return condition number of exogenous matrix.

        Calculated as ratio of largest to smallest eigenvalue.
        """
        eigvals = self.eigenvals
        return np.sqrt(eigvals[0]/eigvals[-1])

    # TODO: make these properties reset bse
    def _HCCM(self, scale):
        H = np.dot(self.model.pinv_wexog,
                   scale[:, None] * self.model.pinv_wexog.T)
        return H

    @cache_readonly
    def cov_HC0(self):
        """
        See statsmodels.RegressionResults
        """

        self.het_scale = self.wresid**2
        cov_HC0 = self._HCCM(self.het_scale)
        return cov_HC0

    @cache_readonly
    def cov_HC1(self):
        """
        See statsmodels.RegressionResults
        """

        self.het_scale = self.nobs/(self.df_resid)*(self.wresid**2)
        cov_HC1 = self._HCCM(self.het_scale)
        return cov_HC1

    @cache_readonly
    def cov_HC2(self):
        """
        See statsmodels.RegressionResults
        """

        # probably could be optimized
        h = np.diag(chain_dot(self.model.wexog,
                              self.normalized_cov_params,
                              self.model.wexog.T))
        self.het_scale = self.wresid**2/(1-h)
        cov_HC2 = self._HCCM(self.het_scale)
        return cov_HC2

    @cache_readonly
    def cov_HC3(self):
        """
        See statsmodels.RegressionResults
        """
        h = np.diag(chain_dot(
            self.model.wexog, self.normalized_cov_params, self.model.wexog.T))
        self.het_scale = (self.wresid / (1 - h))**2
        cov_HC3 = self._HCCM(self.het_scale)
        return cov_HC3

    @cache_readonly
    def HC0_se(self):
        """
        See statsmodels.RegressionResults
        """
        return np.sqrt(np.diag(self.cov_HC0))

    @cache_readonly
    def HC1_se(self):
        """
        See statsmodels.RegressionResults
        """
        return np.sqrt(np.diag(self.cov_HC1))

    @cache_readonly
    def HC2_se(self):
        """
        See statsmodels.RegressionResults
        """
        return np.sqrt(np.diag(self.cov_HC2))

    @cache_readonly
    def HC3_se(self):
        """
        See statsmodels.RegressionResults
        """
        return np.sqrt(np.diag(self.cov_HC3))

    @cache_readonly
    def resid_pearson(self):
        """
        Residuals, normalized to have unit variance.

        Returns
        -------
        An array wresid standardized by the sqrt if scale
        """

        if not hasattr(self, 'resid'):
            raise ValueError('Method requires residuals.')
        eps = np.finfo(self.wresid.dtype).eps
        if np.sqrt(self.scale) < 10 * eps * self.model.endog.mean():
            # don't divide if scale is zero close to numerical precision
            from warnings import warn
            warn("All residuals are 0, cannot compute normed residuals.",
                 RuntimeWarning)
            return self.wresid
        else:
            return self.wresid / np.sqrt(self.scale)

    def _is_nested(self, restricted):
        """
        Parameters
        ----------
        restricted : Result instance
            The restricted model is assumed to be nested in the current
            model. The result instance of the restricted model is required to
            have two attributes, residual sum of squares, `ssr`, residual
            degrees of freedom, `df_resid`.

        Returns
        -------
        nested : bool
            True if nested, otherwise false

        Notes
        -----
        A most nests another model if the regressors in the smaller
        model are spanned by the regressors in the larger model and
        the regressand is identical.
        """

        if self.model.nobs != restricted.model.nobs:
            return False

        full_rank = self.model.rank
        restricted_rank = restricted.model.rank
        if full_rank <= restricted_rank:
            return False

        restricted_exog = restricted.model.wexog
        full_wresid = self.wresid

        scores = restricted_exog * full_wresid[:, None]
        score_l2 = np.sqrt(np.mean(scores.mean(0) ** 2))
        # TODO: Could be improved, and may fail depending on scale of
        # regressors
        return np.allclose(score_l2, 0)

    def compare_lm_test(self, restricted, demean=True, use_lr=False):
        """Use Lagrange Multiplier test to test whether restricted model is correct

        Parameters
        ----------
        restricted : Result instance
            The restricted model is assumed to be nested in the
            current model. The result instance of the restricted model
            is required to have two attributes, residual sum of
            squares, `ssr`, residual degrees of freedom, `df_resid`.

        demean : bool
            Flag indicating whether the demean the scores based on the
            residuals from the restricted model.  If True, the
            covariance of the scores are used and the LM test is
            identical to the large sample version of the LR test.

        Returns
        -------
        lm_value : float
            test statistic, chi2 distributed
        p_value : float
            p-value of the test statistic
        df_diff : int
            degrees of freedom of the restriction, i.e. difference in df
            between models

        Notes
        -----
        TODO: explain LM text
        """
        import statsmodels.stats.sandwich_covariance as sw
        from numpy.linalg import inv

        if not self._is_nested(restricted):
            raise ValueError("Restricted model is not nested by full model.")

        wresid = restricted.wresid
        wexog = self.model.wexog
        scores = wexog * wresid[:, None]

        n = self.nobs
        df_full = self.df_resid
        df_restr = restricted.df_resid
        df_diff = (df_restr - df_full)

        s = scores.mean(axis=0)
        if use_lr:
            scores = wexog * self.wresid[:, None]
            demean = False

        if demean:
            scores = scores - scores.mean(0)[None, :]
        # Form matters here.  If homoskedastics can be sigma^2 (X'X)^-1
        # If Heteroskedastic then the form below is fine
        # If HAC then need to use HAC
        # If Cluster, shoudl use cluster

        cov_type = getattr(self, 'cov_type', 'nonrobust')
        if cov_type == 'nonrobust':
            sigma2 = np.mean(wresid**2)
            XpX = np.dot(wexog.T, wexog) / n
            Sinv = inv(sigma2 * XpX)
        elif cov_type in ('HC0', 'HC1', 'HC2', 'HC3'):
            Sinv = inv(np.dot(scores.T, scores) / n)
        elif cov_type == 'HAC':
            maxlags = self.cov_kwds['maxlags']
            Sinv = inv(sw.S_hac_simple(scores, maxlags) / n)
        elif cov_type == 'cluster':
            # cluster robust standard errors
            groups = self.cov_kwds['groups']
            # TODO: Might need demean option in S_crosssection by group?
            Sinv = inv(sw.S_crosssection(scores, groups))
        else:
            raise ValueError('Only nonrobust, HC, HAC and cluster are ' +
                             'currently connected')

        lm_value = n * chain_dot(s, Sinv, s.T)
        p_value = stats.chi2.sf(lm_value, df_diff)
        return lm_value, p_value, df_diff

    def compare_f_test(self, restricted):
        """use F test to test whether restricted model is correct

        Parameters
        ----------
        restricted : Result instance
            The restricted model is assumed to be nested in the
            current model. The result instance of the restricted model
            is required to have two attributes, residual sum of
            squares, `ssr`, residual degrees of freedom, `df_resid`.

        Returns
        -------
        f_value : float
            test statistic, F distributed
        p_value : float
            p-value of the test statistic
        df_diff : int
            degrees of freedom of the restriction, i.e. difference in
            df between models

        Notes
        -----
        See mailing list discussion October 17,

        This test compares the residual sum of squares of the two
        models.  This is not a valid test, if there is unspecified
        heteroscedasticity or correlation. This method will issue a
        warning if this is detected but still return the results under
        the assumption of homoscedasticity and no autocorrelation
        (sphericity).
        """

        has_robust1 = getattr(self, 'cov_type', 'nonrobust') != 'nonrobust'
        has_robust2 = (getattr(restricted, 'cov_type', 'nonrobust') !=
                       'nonrobust')

        if has_robust1 or has_robust2:
            warnings.warn('F test for comparison is likely invalid with ' +
                          'robust covariance, proceeding anyway',
                          InvalidTestWarning)

        ssr_full = self.ssr
        ssr_restr = restricted.ssr
        df_full = self.df_resid
        df_restr = restricted.df_resid

        df_diff = (df_restr - df_full)
        f_value = (ssr_restr - ssr_full) / df_diff / ssr_full * df_full
        p_value = stats.f.sf(f_value, df_diff, df_full)
        return f_value, p_value, df_diff

    def compare_lr_test(self, restricted, large_sample=False):
        """
        Likelihood ratio test to test whether restricted model is correct

        Parameters
        ----------
        restricted : Result instance
            The restricted model is assumed to be nested in the current model.
            The result instance of the restricted model is required to have two
            attributes, residual sum of squares, `ssr`, residual degrees of
            freedom, `df_resid`.

        large_sample : bool
            Flag indicating whether to use a heteroskedasticity robust version
            of the LR test, which is a modified LM test.

        Returns
        -------
        lr_stat : float
            likelihood ratio, chisquare distributed with df_diff degrees of
            freedom
        p_value : float
            p-value of the test statistic
        df_diff : int
            degrees of freedom of the restriction, i.e. difference in df
            between models

        Notes
        -----

        The exact likelihood ratio is valid for homoskedastic data,
        and is defined as

        .. math:: D=-2\\log\\left(\\frac{\\mathcal{L}_{null}}
           {\\mathcal{L}_{alternative}}\\right)

        where :math:`\\mathcal{L}` is the likelihood of the
        model. With :math:`D` distributed as chisquare with df equal
        to difference in number of parameters or equivalently
        difference in residual degrees of freedom.

        The large sample version of the likelihood ratio is defined as

        .. math:: D=n s^{\\prime}S^{-1}s

        where :math:`s=n^{-1}\\sum_{i=1}^{n} s_{i}`

        .. math:: s_{i} = x_{i,alternative} \\epsilon_{i,null}

        is the average score of the model evaluated using the
        residuals from null model and the regressors from the
        alternative model and :math:`S` is the covariance of the
        scores, :math:`s_{i}`.  The covariance of the scores is
        estimated using the same estimator as in the alternative
        model.

        This test compares the loglikelihood of the two models.  This
        may not be a valid test, if there is unspecified
        heteroscedasticity or correlation. This method will issue a
        warning if this is detected but still return the results
        without taking unspecified heteroscedasticity or correlation
        into account.

        This test compares the loglikelihood of the two models.  This
        may not be a valid test, if there is unspecified
        heteroscedasticity or correlation. This method will issue a
        warning if this is detected but still return the results
        without taking unspecified heteroscedasticity or correlation
        into account.

        is the average score of the model evaluated using the
        residuals from null model and the regressors from the
        alternative model and :math:`S` is the covariance of the
        scores, :math:`s_{i}`.  The covariance of the scores is
        estimated using the same estimator as in the alternative
        model.

        TODO: put into separate function, needs tests
        """

        # See mailing list discussion October 17,

        if large_sample:
            return self.compare_lm_test(restricted, use_lr=True)

        has_robust1 = (getattr(self, 'cov_type', 'nonrobust') != 'nonrobust')
        has_robust2 = (
            getattr(restricted, 'cov_type', 'nonrobust') != 'nonrobust')

        if has_robust1 or has_robust2:
            warnings.warn('Likelihood Ratio test is likely invalid with ' +
                          'robust covariance, proceeding anyway',
                          InvalidTestWarning)

        llf_full = self.llf
        llf_restr = restricted.llf
        df_full = self.df_resid
        df_restr = restricted.df_resid

        lrdf = (df_restr - df_full)
        lrstat = -2*(llf_restr - llf_full)
        lr_pvalue = stats.chi2.sf(lrstat, lrdf)

        return lrstat, lr_pvalue, lrdf

    def get_robustcov_results(self, cov_type='HC1', use_t=None, **kwds):
        """create new results instance with robust covariance as default

        Parameters
        ----------
        cov_type : string
            the type of robust sandwich estimator to use. see Notes below
        use_t : bool
            If true, then the t distribution is used for inference.
            If false, then the normal distribution is used.
            If `use_t` is None, then an appropriate default is used, which is
            `true` if the cov_type is nonrobust, and `false` in all other
            cases.
        kwds : depends on cov_type
            Required or optional arguments for robust covariance calculation.
            see Notes below

        Returns
        -------
        results : results instance
            This method creates a new results instance with the
            requested robust covariance as the default covariance of
            the parameters.  Inferential statistics like p-values and
            hypothesis tests will be based on this covariance matrix.

        Notes
        -----
        The following covariance types and required or optional arguments are
        currently available:

        - 'fixed scale' and optional keyword argument 'scale' which uses
            a predefined scale estimate with default equal to one.
        - 'HC0', 'HC1', 'HC2', 'HC3' and no keyword arguments:
            heteroscedasticity robust covariance
        - 'HAC' and keywords

            - `maxlag` integer (required) : number of lags to use
            - `kernel` callable or str (optional) : kernel
                  currently available kernels are ['bartlett', 'uniform'],
                  default is Bartlett
            - `use_correction` bool (optional) : If true, use small sample
                  correction

        - 'cluster' and required keyword `groups`, integer group indicator

            - `groups` array_like, integer (required) :
                  index of clusters or groups
            - `use_correction` bool (optional) :
                  If True the sandwich covariance is calculated with a small
                  sample correction.
                  If False the sandwich covariance is calculated without
                  small sample correction.
            - `df_correction` bool (optional)
                  If True (default), then the degrees of freedom for the
                  inferential statistics and hypothesis tests, such as
                  pvalues, f_pvalue, conf_int, and t_test and f_test, are
                  based on the number of groups minus one instead of the
                  total number of observations minus the number of explanatory
                  variables. `df_resid` of the results instance is adjusted.
                  If False, then `df_resid` of the results instance is not
                  adjusted.

        - 'hac-groupsum' Driscoll and Kraay, heteroscedasticity and
            autocorrelation robust standard errors in panel data
            keywords

            - `time` array_like (required) : index of time periods
            - `maxlag` integer (required) : number of lags to use
            - `kernel` callable or str (optional) : kernel
                  currently available kernels are ['bartlett', 'uniform'],
                  default is Bartlett
            - `use_correction` False or string in ['hac', 'cluster'] (optional) :
                  If False the the sandwich covariance is calulated without
                  small sample correction.
                  If `use_correction = 'cluster'` (default), then the same
                  small sample correction as in the case of 'covtype='cluster''
                  is used.
            - `df_correction` bool (optional)
                  adjustment to df_resid, see cov_type 'cluster' above
                  # TODO: we need more options here

        - 'hac-panel' heteroscedasticity and autocorrelation robust standard
            errors in panel data.
            The data needs to be sorted in this case, the time series
            for each panel unit or cluster need to be stacked. The
            membership to a timeseries of an individual or group can
            be either specified by group indicators or by increasing
            time periods.

            keywords

            - either `groups` or `time` : array_like (required)
              `groups` : indicator for groups
              `time` : index of time periods
            - `maxlag` integer (required) : number of lags to use
            - `kernel` callable or str (optional) : kernel
                  currently available kernels are ['bartlett', 'uniform'],
                  default is Bartlett
            - `use_correction` False or string in ['hac', 'cluster'] (optional) :
                  If False the sandwich covariance is calculated without
                  small sample correction.
            - `df_correction` bool (optional)
                  adjustment to df_resid, see cov_type 'cluster' above
                  # TODO: we need more options here

        Reminder:
        `use_correction` in "hac-groupsum" and "hac-panel" is not bool,
        needs to be in [False, 'hac', 'cluster']

        TODO: Currently there is no check for extra or misspelled keywords,
        except in the case of cov_type `HCx`
        """
        import statsmodels.stats.sandwich_covariance as sw

        # normalize names
        if cov_type == 'nw-panel':
            cov_type = 'hac-panel'
        if cov_type == 'nw-groupsum':
            cov_type = 'hac-groupsum'
        if 'kernel' in kwds:
            kwds['weights_func'] = kwds.pop('kernel')
        if 'weights_func' in kwds and not callable(kwds['weights_func']):
            kwds['weights_func'] = sw.kernel_dict[kwds['weights_func']]

        # TODO: make separate function that returns a robust cov plus info
        use_self = kwds.pop('use_self', False)
        if use_self:
            res = self
        else:
            res = self.__class__(
                self.model, self.params,
                normalized_cov_params=self.normalized_cov_params,
                scale=self.scale)

        res.cov_type = cov_type
        # use_t might already be defined by the class, and already set
        if use_t is None:
            use_t = self.use_t
        res.cov_kwds = {'use_t': use_t}  # store for information
        res.use_t = use_t

        adjust_df = False
        if cov_type in ['cluster', 'hac-panel', 'hac-groupsum']:
            df_correction = kwds.get('df_correction', None)
            # TODO: check also use_correction, do I need all combinations?
            if df_correction is not False:  # i.e. in [None, True]:
                # user didn't explicitely set it to False
                adjust_df = True

        res.cov_kwds['adjust_df'] = adjust_df

        # verify and set kwds, and calculate cov
        # TODO: this should be outsourced in a function so we can reuse it in
        #       other models
        # TODO: make it DRYer   repeated code for checking kwds
        if cov_type in ['fixed scale', 'fixed_scale']:
            res.cov_kwds['description'] = ('Standard Errors are based on ' +
                                           'fixed scale')

            res.cov_kwds['scale'] = scale = kwds.get('scale', 1.)
            res.cov_params_default = scale * res.normalized_cov_params
        elif cov_type.upper() in ('HC0', 'HC1', 'HC2', 'HC3'):
            if kwds:
                raise ValueError('heteroscedasticity robust covarians ' +
                                 'does not use keywords')
            res.cov_kwds['description'] = (
                'Standard Errors are heteroscedasticity ' +
                'robust ' + '(' + cov_type + ')')
            # TODO cannot access cov without calling se first
            getattr(self, cov_type.upper() + '_se')
            res.cov_params_default = getattr(self, 'cov_' + cov_type.upper())
        elif cov_type.lower() == 'hac':
            maxlags = kwds['maxlags']   # required?, default in cov_hac_simple
            res.cov_kwds['maxlags'] = maxlags
            weights_func = kwds.get('weights_func', sw.weights_bartlett)
            res.cov_kwds['weights_func'] = weights_func
            use_correction = kwds.get('use_correction', False)
            res.cov_kwds['use_correction'] = use_correction
            res.cov_kwds['description'] = (
                'Standard Errors are heteroscedasticity and ' +
                'autocorrelation robust (HAC) using %d lags and %s small ' +
                'sample correction') % (maxlags,
                                        ['without', 'with'][use_correction])

            res.cov_params_default = sw.cov_hac_simple(
                self, nlags=maxlags, weights_func=weights_func,
                use_correction=use_correction)
        elif cov_type.lower() == 'cluster':
            # cluster robust standard errors, one- or two-way
            groups = kwds['groups']
            if not hasattr(groups, 'shape'):
                groups = np.asarray(groups).T

            if groups.ndim >= 2:
                groups = groups.squeeze()

            res.cov_kwds['groups'] = groups
            use_correction = kwds.get('use_correction', True)
            res.cov_kwds['use_correction'] = use_correction
            if groups.ndim == 1:
                if adjust_df:
                    # need to find number of groups
                    # duplicate work
                    self.n_groups = n_groups = len(np.unique(groups))
                res.cov_params_default = sw.cov_cluster(
                    self, groups, use_correction=use_correction)

            elif groups.ndim == 2:
                if hasattr(groups, 'values'):
                    groups = groups.values

                if adjust_df:
                    # need to find number of groups
                    # duplicate work
                    n_groups0 = len(np.unique(groups[:, 0]))
                    n_groups1 = len(np.unique(groups[:, 1]))
                    self.n_groups = (n_groups0, n_groups1)
                    n_groups = min(n_groups0, n_groups1)  # use for adjust_df

                # Note: sw.cov_cluster_2groups has 3 returns
                res.cov_params_default = sw.cov_cluster_2groups(
                    self, groups, use_correction=use_correction)[0]
            else:
                raise ValueError('only two groups are supported')
            res.cov_kwds['description'] = (
                'Standard Errors are robust to' +
                'cluster correlation ' + '(' + cov_type + ')')

        elif cov_type.lower() == 'hac-panel':
            # cluster robust standard errors
            res.cov_kwds['time'] = time = kwds.get('time', None)
            res.cov_kwds['groups'] = groups = kwds.get('groups', None)
            # TODO: nlags is currently required
            # nlags = kwds.get('nlags', True)
            # res.cov_kwds['nlags'] = nlags
            # TODO: `nlags` or `maxlags`
            res.cov_kwds['maxlags'] = maxlags = kwds['maxlags']
            use_correction = kwds.get('use_correction', 'hac')
            res.cov_kwds['use_correction'] = use_correction
            weights_func = kwds.get('weights_func', sw.weights_bartlett)
            res.cov_kwds['weights_func'] = weights_func
            if groups is not None:
                groups = np.asarray(groups)
                tt = (np.nonzero(groups[:-1] != groups[1:])[0] + 1).tolist()
                nobs_ = len(groups)
            elif time is not None:
                time = np.asarray(time)
                # TODO: clumsy time index in cov_nw_panel
                tt = (np.nonzero(time[1:] < time[:-1])[0] + 1).tolist()
                nobs_ = len(time)
            else:
                raise ValueError('either time or groups needs to be given')
            groupidx = lzip([0] + tt, tt + [nobs_])
            self.n_groups = n_groups = len(groupidx)
            res.cov_params_default = sw.cov_nw_panel(self, maxlags, groupidx,
                                                     weights_func=weights_func,
                                                     use_correction=use_correction)
            res.cov_kwds['description'] = (
                'Standard Errors are robust to' +
                'cluster correlation ' + '(' + cov_type + ')')
        elif cov_type.lower() == 'hac-groupsum':
            # Driscoll-Kraay standard errors
            res.cov_kwds['time'] = time = kwds['time']
            # TODO: nlags is currently required
            # nlags = kwds.get('nlags', True)
            # res.cov_kwds['nlags'] = nlags
            # TODO: `nlags` or `maxlags`
            res.cov_kwds['maxlags'] = maxlags = kwds['maxlags']
            use_correction = kwds.get('use_correction', 'cluster')
            res.cov_kwds['use_correction'] = use_correction
            weights_func = kwds.get('weights_func', sw.weights_bartlett)
            res.cov_kwds['weights_func'] = weights_func
            if adjust_df:
                # need to find number of groups
                tt = (np.nonzero(time[1:] < time[:-1])[0] + 1)
                self.n_groups = n_groups = len(tt) + 1
            res.cov_params_default = sw.cov_nw_groupsum(
                self, maxlags, time, weights_func=weights_func,
                use_correction=use_correction)
            res.cov_kwds['description'] = (
                        'Driscoll and Kraay Standard Errors are robust to ' +
                        'cluster correlation ' + '(' + cov_type + ')')
        else:
            raise ValueError('cov_type not recognized. See docstring for ' +
                             'available options and spelling')

        if adjust_df:
            # Note: df_resid is used for scale and others, add new attribute
            res.df_resid_inference = n_groups - 1

        return res

    def get_prediction(self, exog=None, transform=True, weights=None,
                       row_labels=None, **kwds):

        return pred.get_prediction(
            self, exog=exog, transform=transform, weights=weights,
            row_labels=row_labels, **kwds)

    get_prediction.__doc__ = pred.get_prediction.__doc__

    def summary(self, yname=None, xname=None, title=None, alpha=.05):
        """Summarize the Regression Results

        Parameters
        -----------
        yname : string, optional
            Default is `y`
        xname : list of strings, optional
            Default is `var_##` for ## in p the number of regressors
        title : string, optional
            Title for the top table. If not None, then this replaces the
            default title
        alpha : float
            significance level for the confidence intervals

        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.summary.Summary : class to hold summary
            results

        """

        # TODO: import where we need it (for now), add as cached attributes
        from statsmodels.stats.stattools import (
            jarque_bera, omni_normtest, durbin_watson)
        jb, jbpv, skew, kurtosis = jarque_bera(self.wresid)
        omni, omnipv = omni_normtest(self.wresid)

        eigvals = self.eigenvals
        condno = self.condition_number

        self.diagn = dict(jb=jb, jbpv=jbpv, skew=skew, kurtosis=kurtosis,
                          omni=omni, omnipv=omnipv, condno=condno,
                          mineigval=eigvals[-1])

        # TODO not used yet
        # diagn_left_header = ['Models stats']
        # diagn_right_header = ['Residual stats']

        # TODO: requiring list/iterable is a bit annoying
        # need more control over formatting
        # TODO: default don't work if it's not identically spelled

        top_left = [('Dep. Variable:', None),
                    ('Model:', None),
                    ('Method:', ['Least Squares']),
                    ('Date:', None),
                    ('Time:', None),
                    ('No. Observations:', None),
                    ('Df Residuals:', None),  # [self.df_resid]), TODO: spelling
                    ('Df Model:', None),  # [self.df_model])
                    ]

        if hasattr(self, 'cov_type'):
            top_left.append(('Covariance Type:', [self.cov_type]))

        rsquared_type = '' if self.k_constant else ' (uncentered)'
        top_right = [('R-squared' + rsquared_type + ':', ["%#8.3f" % self.rsquared]),
                     ('Adj. R-squared' + rsquared_type + ':', ["%#8.3f" % self.rsquared_adj]),
                     ('F-statistic:', ["%#8.4g" % self.fvalue]),
                     ('Prob (F-statistic):', ["%#6.3g" % self.f_pvalue]),
                     ('Log-Likelihood:', None),  # ["%#6.4g" % self.llf]),
                     ('AIC:', ["%#8.4g" % self.aic]),
                     ('BIC:', ["%#8.4g" % self.bic])
                     ]

        diagn_left = [('Omnibus:', ["%#6.3f" % omni]),
                      ('Prob(Omnibus):', ["%#6.3f" % omnipv]),
                      ('Skew:', ["%#6.3f" % skew]),
                      ('Kurtosis:', ["%#6.3f" % kurtosis])
                      ]

        diagn_right = [('Durbin-Watson:',
                        ["%#8.3f" % durbin_watson(self.wresid)]
                        ),
                       ('Jarque-Bera (JB):', ["%#8.3f" % jb]),
                       ('Prob(JB):', ["%#8.3g" % jbpv]),
                       ('Cond. No.', ["%#8.3g" % condno])
                       ]

        if title is None:
            title = self.model.__class__.__name__ + ' ' + "Regression Results"

        # create summary table instance
        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)

        smry.add_table_2cols(self, gleft=diagn_left, gright=diagn_right,
                             yname=yname, xname=xname,
                             title="")

        # add warnings/notes, added to text format only
        etext = []
        if hasattr(self, 'cov_type'):
            etext.append(self.cov_kwds['description'])
        if self.model.exog.shape[0] < self.model.exog.shape[1]:
            wstr = "The input rank is higher than the number of observations."
            etext.append(wstr)
        if eigvals[-1] < 1e-10:
            wstr = "The smallest eigenvalue is %6.3g. This might indicate "
            wstr += "that there are\n"
            wstr += "strong multicollinearity problems or that the design "
            wstr += "matrix is singular."
            wstr = wstr % eigvals[-1]
            etext.append(wstr)
        elif condno > 1000:  # TODO: what is recommended
            wstr = "The condition number is large, %6.3g. This might "
            wstr += "indicate that there are\n"
            wstr += "strong multicollinearity or other numerical "
            wstr += "problems."
            wstr = wstr % condno
            etext.append(wstr)

        if etext:
            etext = ["[{0}] {1}".format(i + 1, text)
                     for i, text in enumerate(etext)]
            etext.insert(0, "Warnings:")
            smry.add_extra_txt(etext)

        return smry

        #  top = summary_top(self, gleft=topleft, gright=diagn_left, #[],
        #                    yname=yname, xname=xname,
        #                    title=self.model.__class__.__name__ + ' ' +
        #                    "Regression Results")
        #  par = summary_params(self, yname=yname, xname=xname, alpha=.05,
        #                       use_t=False)
        #
        #  diagn = summary_top(self, gleft=diagn_left, gright=diagn_right,
        #                      yname=yname, xname=xname,
        #                      title="Linear Model")
        #
        #  return summary_return([top, par, diagn], return_fmt=return_fmt)

    def summary2(self, yname=None, xname=None, title=None, alpha=.05,
                 float_format="%.4f"):
        """Experimental summary function to summarize the regression results

        Parameters
        -----------
        xname : List of strings of length equal to the number of parameters
            Names of the independent variables (optional)
        yname : string
            Name of the dependent variable (optional)
        title : string, 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: string
            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.summary.Summary : class to hold summary
            results

        """
        # Diagnostics
        from statsmodels.stats.stattools import (jarque_bera,
                                                 omni_normtest,
                                                 durbin_watson)

        from collections import OrderedDict
        jb, jbpv, skew, kurtosis = jarque_bera(self.wresid)
        omni, omnipv = omni_normtest(self.wresid)
        dw = durbin_watson(self.wresid)
        eigvals = self.eigenvals
        condno = self.condition_number
        eigvals = np.sort(eigvals)  # in increasing order
        diagnostic = OrderedDict([
            ('Omnibus:',  "%.3f" % omni),
            ('Prob(Omnibus):', "%.3f" % omnipv),
            ('Skew:', "%.3f" % skew),
            ('Kurtosis:', "%.3f" % kurtosis),
            ('Durbin-Watson:', "%.3f" % dw),
            ('Jarque-Bera (JB):', "%.3f" % jb),
            ('Prob(JB):', "%.3f" % jbpv),
            ('Condition No.:', "%.0f" % condno)
            ])

        # Summary
        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)
        smry.add_dict(diagnostic)

        # Warnings
        if eigvals[-1] < 1e-10:
            warn = "The smallest eigenvalue is %6.3g. This might indicate that\
            there are strong multicollinearity problems or that the design\
            matrix is singular." % eigvals[-1]
            smry.add_text(warn)
        if condno > 1000:
            warn = "* The condition number is large (%.g). This might indicate \
            strong multicollinearity or other numerical problems." % condno
            smry.add_text(warn)

        return smry


class OLSResults(RegressionResults):
    """
    Results class for for an OLS model.

    Most of the methods and attributes are inherited from RegressionResults.
    The special methods that are only available for OLS are:

    - get_influence
    - outlier_test
    - el_test
    - conf_int_el

    See Also
    --------
    RegressionResults

    """

    def get_influence(self):
        """
        get an instance of Influence with influence and outlier measures

        Returns
        -------
        infl : Influence instance
            the instance has methods to calculate the main influence and
            outlier measures for the OLS regression

        See also
        --------
        statsmodels.stats.outliers_influence.OLSInfluence
        """
        from statsmodels.stats.outliers_influence import OLSInfluence
        return OLSInfluence(self)

    def outlier_test(self, method='bonf', alpha=.05, labels=None,
                 order=False, cutoff=None):
        """
        Test observations for outliers according to method

        Parameters
        ----------
        method : str

            - `bonferroni` : one-step correction
            - `sidak` : one-step correction
            - `holm-sidak` :
            - `holm` :
            - `simes-hochberg` :
            - `hommel` :
            - `fdr_bh` : Benjamini/Hochberg
            - `fdr_by` : Benjamini/Yekutieli

            See `statsmodels.stats.multitest.multipletests` for details.
        alpha : float
            familywise error rate
        labels : None or array_like
            If `labels` is not None, then it will be used as index to the
            returned pandas DataFrame. See also Returns below
        order : bool
            Whether or not to order the results by the absolute value of the
            studentized residuals. If labels are provided they will also be sorted.
        cutoff : None or float in [0, 1]
            If cutoff is not None, then the return only includes observations with
            multiple testing corrected p-values strictly below the cutoff. The
            returned array or dataframe can be empty if t

        Returns
        -------
        table : ndarray or DataFrame
            Returns either an ndarray or a DataFrame if labels is not None.
            Will attempt to get labels from model_results if available. The
            columns are the Studentized residuals, the unadjusted p-value,
            and the corrected p-value according to method.

        Notes
        -----
        The unadjusted p-value is stats.t.sf(abs(resid), df) where
        df = df_resid - 1.
        """
        from statsmodels.stats.outliers_influence import outlier_test
        return outlier_test(self, method, alpha, labels=labels,
                            order=order, cutoff=cutoff)

    def el_test(self, b0_vals, param_nums, return_weights=0,
                ret_params=0, method='nm',
                stochastic_exog=1, return_params=0):
        """
        Tests single or joint hypotheses of the regression parameters using
        Empirical Likelihood.

        Parameters
        ----------

        b0_vals : 1darray
            The hypothesized value of the parameter to be tested

        param_nums : 1darray
            The parameter number to be tested

        print_weights : bool
            If true, returns the weights that optimize the likelihood
            ratio at b0_vals.  Default is False

        ret_params : bool
            If true, returns the parameter vector that maximizes the likelihood
            ratio at b0_vals.  Also returns the weights.  Default is False

        method : string
            Can either be 'nm' for Nelder-Mead or 'powell' for Powell.  The
            optimization method that optimizes over nuisance parameters.
            Default is 'nm'

        stochastic_exog : bool
            When TRUE, the exogenous variables are assumed to be stochastic.
            When the regressors are nonstochastic, moment conditions are
            placed on the exogenous variables.  Confidence intervals for
            stochastic regressors are at least as large as non-stochastic
            regressors.  Default = TRUE

        Returns
        -------

        res : tuple
            The p-value and -2 times the log-likelihood ratio for the
            hypothesized values.

        Examples
        --------
        >>> import statsmodels.api as sm
        >>> data = sm.datasets.stackloss.load(as_pandas=False)
        >>> endog = data.endog
        >>> exog = sm.add_constant(data.exog)
        >>> model = sm.OLS(endog, exog)
        >>> fitted = model.fit()
        >>> fitted.params
        >>> array([-39.91967442,   0.7156402 ,   1.29528612,  -0.15212252])
        >>> fitted.rsquared
        >>> 0.91357690446068196
        >>> # Test that the slope on the first variable is 0
        >>> fitted.el_test([0], [1])
        >>> (27.248146353888796, 1.7894660442330235e-07)
        """
        params = np.copy(self.params)
        opt_fun_inst = _ELRegOpts()  # to store weights
        if len(param_nums) == len(params):
            llr = opt_fun_inst._opt_nuis_regress(
                [],
                param_nums=param_nums,
                endog=self.model.endog,
                exog=self.model.exog,
                nobs=self.model.nobs,
                nvar=self.model.exog.shape[1],
                params=params,
                b0_vals=b0_vals,
                stochastic_exog=stochastic_exog)
            pval = 1 - stats.chi2.cdf(llr, len(param_nums))
            if return_weights:
                return llr, pval, opt_fun_inst.new_weights
            else:
                return llr, pval
        x0 = np.delete(params, param_nums)
        args = (param_nums, self.model.endog, self.model.exog,
                self.model.nobs, self.model.exog.shape[1], params,
                b0_vals, stochastic_exog)
        if method == 'nm':
            llr = optimize.fmin(opt_fun_inst._opt_nuis_regress, x0,
                                maxfun=10000, maxiter=10000, full_output=1,
                                disp=0, args=args)[1]
        if method == 'powell':
            llr = optimize.fmin_powell(opt_fun_inst._opt_nuis_regress, x0,
                                       full_output=1, disp=0,
                                       args=args)[1]

        pval = 1 - stats.chi2.cdf(llr, len(param_nums))
        if ret_params:
            return llr, pval, opt_fun_inst.new_weights, opt_fun_inst.new_params
        elif return_weights:
            return llr, pval, opt_fun_inst.new_weights
        else:
            return llr, pval

    def conf_int_el(self, param_num, sig=.05, upper_bound=None,
                    lower_bound=None, method='nm', stochastic_exog=1):
        """
        Computes the confidence interval for the parameter given by param_num
        using Empirical Likelihood

        Parameters
        ----------

        param_num : float
            The parameter for which the confidence interval is desired

        sig : float
            The significance level.  Default is .05

        upper_bound : float
            The maximum value the upper limit can be.  Default is the
            99.9% confidence value under OLS assumptions.

        lower_bound : float
            The minimum value the lower limit can be.  Default is the 99.9%
            confidence value under OLS assumptions.

        method : string
            Can either be 'nm' for Nelder-Mead or 'powell' for Powell.  The
            optimization method that optimizes over nuisance parameters.
            Default is 'nm'

        Returns
        -------

        ci : tuple
            The confidence interval

        See Also
        --------

        el_test

        Notes
        -----

        This function uses brentq to find the value of beta where
        test_beta([beta], param_num)[1] is equal to the critical
        value.

        The function returns the results of each iteration of brentq at
        each value of beta.

        The current function value of the last printed optimization
        should be the critical value at the desired significance level.
        For alpha=.05, the value is 3.841459.

        To ensure optimization terminated successfully, it is suggested to
        do el_test([lower_limit], [param_num])

        If the optimization does not terminate successfully, consider switching
        optimization algorithms.

        If optimization is still not successful, try changing the values of
        start_int_params.  If the current function value repeatedly jumps
        from a number between 0 and the critical value and a very large number
        (>50), the starting parameters of the interior minimization need
        to be changed.
        """
        r0 = stats.chi2.ppf(1 - sig, 1)
        if upper_bound is None:
            upper_bound = self.conf_int(.01)[param_num][1]
        if lower_bound is None:
            lower_bound = self.conf_int(.01)[param_num][0]
        f = lambda b0: self.el_test(np.array([b0]), np.array([param_num]),
                                    method=method,
                                    stochastic_exog=stochastic_exog)[0]-r0
        lowerl = optimize.brenth(f, lower_bound,
                                 self.params[param_num])
        upperl = optimize.brenth(f, self.params[param_num],
                                 upper_bound)
        #  ^ Seems to be faster than brentq in most cases
        return (lowerl, upperl)


class RegressionResultsWrapper(wrap.ResultsWrapper):

    _attrs = {
        'chisq': 'columns',
        'sresid': 'rows',
        'weights': 'rows',
        'wresid': 'rows',
        'bcov_unscaled': 'cov',
        'bcov_scaled': 'cov',
        'HC0_se': 'columns',
        'HC1_se': 'columns',
        'HC2_se': 'columns',
        'HC3_se': 'columns',
        'norm_resid': 'rows',
    }

    _wrap_attrs = wrap.union_dicts(base.LikelihoodResultsWrapper._attrs,
                                   _attrs)

    _methods = {}

    _wrap_methods = wrap.union_dicts(
                        base.LikelihoodResultsWrapper._wrap_methods,
                        _methods)

wrap.populate_wrapper(RegressionResultsWrapper,
                      RegressionResults)


if __name__ == "__main__":
    import statsmodels.api as sm
    data = sm.datasets.longley.load(as_pandas=False)
    data.exog = add_constant(data.exog, prepend=False)
    ols_results = OLS(data.endog, data.exog).fit()  # results
    gls_results = GLS(data.endog, data.exog).fit()  # results
    print(ols_results.summary())
    tables = ols_results.summary(returns='tables')
    csv = ols_results.summary(returns='csv')
"""
    Summary of Regression Results
=======================================
| Dependent Variable:            ['y']|
| Model:                           OLS|
| Method:                Least Squares|
| Date:               Tue, 29 Jun 2010|
| Time:                       22:32:21|
| # obs:                          16.0|
| Df residuals:                    9.0|
| Df model:                        6.0|
===========================================================================
|            coefficient       std. error      t-statistic           prob.|
---------------------------------------------------------------------------
| x1             15.0619          84.9149           0.1774          0.8631|
| x2             -0.0358           0.0335          -1.0695          0.3127|
| x3             -2.0202           0.4884          -4.1364        0.002535|
| x4             -1.0332           0.2143          -4.8220       0.0009444|
| x5             -0.0511           0.2261          -0.2261          0.8262|
| x6           1829.1515         455.4785           4.0159        0.003037|
| const    -3482258.6346      890420.3836          -3.9108        0.003560|
===========================================================================
|                        Models stats                      Residual stats |
---------------------------------------------------------------------------
| R-squared:                 0.995479    Durbin-Watson:           2.55949 |
| Adjusted R-squared:        0.992465    Omnibus:                0.748615 |
| F-statistic:                330.285    Prob(Omnibus):          0.687765 |
| Prob (F-statistic):     4.98403e-10    JB:                     0.352773 |
| Log likelihood:            -109.617    Prob(JB):               0.838294 |
| AIC criterion:              233.235    Skew:                   0.419984 |
| BIC criterion:              238.643    Kurtosis:                2.43373 |
---------------------------------------------------------------------------
"""