'''Generalized Method of Moments, GMM, and Two-Stage Least Squares for
instrumental variables IV2SLS
Issues
------
* number of parameters, nparams, and starting values for parameters
Where to put them? start was initially taken from global scope (bug)
* When optimal weighting matrix cannot be calculated numerically
In DistQuantilesGMM, we only have one row of moment conditions, not a
moment condition for each observation, calculation for cov of moments
breaks down. iter=1 works (weights is identity matrix)
-> need method to do one iteration with an identity matrix or an
analytical weighting matrix given as parameter.
-> add result statistics for this case, e.g. cov_params, I have it in the
standalone function (and in calc_covparams which is a copy of it),
but not tested yet.
DONE `fitonce` in DistQuantilesGMM, params are the same as in direct call to fitgmm
move it to GMM class (once it's clearer for which cases I need this.)
* GMM doesn't know anything about the underlying model, e.g. y = X beta + u or panel
data model. It would be good if we can reuse methods from regressions, e.g.
predict, fitted values, calculating the error term, and some result statistics.
What's the best way to do this, multiple inheritance, outsourcing the functions,
mixins or delegation (a model creates a GMM instance just for estimation).
Unclear
-------
* dof in Hausman
- based on rank
- differs between IV2SLS method and function used with GMM or (IV2SLS)
- with GMM, covariance matrix difference has negative eigenvalues in iv example, ???
* jtest/jval
- I'm not sure about the normalization (multiply or divide by nobs) in jtest.
need a test case. Scaling of jval is irrelevant for estimation.
jval in jtest looks to large in example, but I have no idea about the size
* bse for fitonce look too large (no time for checking now)
formula for calc_cov_params for the case without optimal weighting matrix
is wrong. I don't have an estimate for omega in that case. And I'm confusing
between weights and omega, which are *not* the same in this case.
Author: josef-pktd
License: BSD (3-clause)
'''
from __future__ import print_function
from statsmodels.compat.python import lrange
from statsmodels.compat.numpy import np_matrix_rank
import numpy as np
from scipy import optimize, stats
from statsmodels.tools.numdiff import approx_fprime, approx_hess
from statsmodels.base.model import (Model,
LikelihoodModel, LikelihoodModelResults)
from statsmodels.regression.linear_model import (OLS, RegressionResults,
RegressionResultsWrapper)
import statsmodels.stats.sandwich_covariance as smcov
from statsmodels.tools.decorators import (resettable_cache, cache_readonly)
from statsmodels.tools.tools import _ensure_2d
DEBUG = 0
def maxabs(x):
'''just a shortcut to np.abs(x).max()
'''
return np.abs(x).max()
[docs]class IV2SLS(LikelihoodModel):
"""
Instrumental variables estimation using Two-Stage Least-Squares (2SLS)
Parameters
----------
endog: array
Endogenous variable, 1-dimensional or 2-dimensional array nobs by 1
exog : array
Explanatory variables, 1-dimensional or 2-dimensional array nobs by k
instrument : array
Instruments for explanatory variables. Must contain both exog
variables that are not being instrumented and instruments
Notes
-----
All variables in exog are instrumented in the calculations. If variables
in exog are not supposed to be instrumented, then these variables
must also to be included in the instrument array.
Degrees of freedom in the calculation of the standard errors uses
`df_resid = (nobs - k_vars)`.
(This corresponds to the `small` option in Stata's ivreg2.)
"""
def __init__(self, endog, exog, instrument=None):
self.instrument, self.instrument_names = _ensure_2d(instrument, True)
super(IV2SLS, self).__init__(endog, exog)
# where is this supposed to be handled
# Note: Greene p.77/78 dof correction is not necessary (because only
# asy results), but most packages do it anyway
self.df_resid = self.exog.shape[0] - self.exog.shape[1]
#self.df_model = float(self.rank - self.k_constant)
self.df_model = float(self.exog.shape[1] - self.k_constant)
[docs] def initialize(self):
self.wendog = self.endog
self.wexog = self.exog
[docs] def whiten(self, X):
pass
[docs] def fit(self):
'''estimate model using 2SLS IV regression
Returns
-------
results : instance of RegressionResults
regression result
Notes
-----
This returns a generic RegressioResults instance as defined for the
linear models.
Parameter estimates and covariance are correct, but other results
haven't been tested yet, to seee whether they apply without changes.
'''
#Greene 5th edt., p.78 section 5.4
#move this maybe
y,x,z = self.endog, self.exog, self.instrument
# TODO: this uses "textbook" calculation, improve linalg
ztz = np.dot(z.T, z)
ztx = np.dot(z.T, x)
self.xhatparams = xhatparams = np.linalg.solve(ztz, ztx)
#print 'x.T.shape, xhatparams.shape', x.shape, xhatparams.shape
F = xhat = np.dot(z, xhatparams)
FtF = np.dot(F.T, F)
self.xhatprod = FtF #store for Housman specification test
Ftx = np.dot(F.T, x)
Fty = np.dot(F.T, y)
params = np.linalg.solve(FtF, Fty)
Ftxinv = np.linalg.inv(Ftx)
self.normalized_cov_params = np.dot(Ftxinv.T, np.dot(FtF, Ftxinv))
lfit = IVRegressionResults(self, params,
normalized_cov_params=self.normalized_cov_params)
lfit.exog_hat_params = xhatparams
lfit.exog_hat = xhat # TODO: do we want to store this, might be large
self._results_ols2nd = OLS(y, xhat).fit()
return RegressionResultsWrapper(lfit)
# copied from GLS, because I subclass currently LikelihoodModel and not GLS
[docs] def predict(self, params, exog=None):
"""
Return linear predicted values from a design matrix.
Parameters
----------
exog : array-like
Design / exogenous data
params : array-like, optional after fit has been called
Parameters of a linear model
Returns
-------
An array of fitted values
Notes
-----
If the model as not yet been fit, params is not optional.
"""
if exog is None:
exog = self.exog
return np.dot(exog, params)
[docs]class IVRegressionResults(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
"""
[docs] @cache_readonly
def fvalue(self):
const_idx = self.model.data.const_idx
# if constant is implicit or missing, return nan see #2444, #3544
if const_idx is None:
return np.nan
else:
k_vars = len(self.params)
restriction = np.eye(k_vars)
idx_noconstant = lrange(k_vars)
del idx_noconstant[const_idx]
fval = self.f_test(restriction[idx_noconstant]).fvalue # without constant
return fval
[docs] def spec_hausman(self, dof=None):
'''Hausman's specification test
See Also
--------
spec_hausman : generic function for Hausman's specification test
'''
#use normalized cov_params for OLS
endog, exog = self.model.endog, self.model.exog
resols = OLS(endog, exog).fit()
normalized_cov_params_ols = resols.model.normalized_cov_params
# Stata `ivendog` doesn't use df correction for se
#se2 = resols.mse_resid #* resols.df_resid * 1. / len(endog)
se2 = resols.ssr / len(endog)
params_diff = self.params - resols.params
cov_diff = np.linalg.pinv(self.model.xhatprod) - normalized_cov_params_ols
#TODO: the following is very inefficient, solves problem (svd) twice
#use linalg.lstsq or svd directly
#cov_diff will very often be in-definite (singular)
if not dof:
dof = np_matrix_rank(cov_diff)
cov_diffpinv = np.linalg.pinv(cov_diff)
H = np.dot(params_diff, np.dot(cov_diffpinv, params_diff))/se2
pval = stats.chi2.sf(H, dof)
return H, pval, dof
# copied from regression results with small changes, no llf
[docs] 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)
#TODO: reuse condno from somewhere else ?
#condno = np.linalg.cond(np.dot(self.wexog.T, self.wexog))
wexog = self.model.wexog
eigvals = np.linalg.linalg.eigvalsh(np.dot(wexog.T, wexog))
eigvals = np.sort(eigvals) #in increasing order
condno = np.sqrt(eigvals[-1]/eigvals[0])
# TODO: check what is valid.
# box-pierce, breusch-pagan, durbin's h are not with endogenous on rhs
# use Cumby Huizinga 1992 instead
self.diagn = dict(jb=jb, jbpv=jbpv, skew=skew, kurtosis=kurtosis,
omni=omni, omnipv=omnipv, condno=condno,
mineigval=eigvals[0])
#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:', ['Two Stage']),
('', ['Least Squares']),
('Date:', None),
('Time:', None),
('No. Observations:', None),
('Df Residuals:', None), #[self.df_resid]), #TODO: spelling
('Df Model:', None), #[self.df_model])
]
top_right = [('R-squared:', ["%#8.3f" % self.rsquared]),
('Adj. R-squared:', ["%#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=True)
smry.add_table_2cols(self, gleft=diagn_left, gright=diagn_right,
yname=yname, xname=xname,
title="")
return smry
############# classes for Generalized Method of Moments GMM
_gmm_options = '''\
Options for GMM
---------------
Type of GMM
~~~~~~~~~~~
- one-step
- iterated
- CUE : not tested yet
weight matrix
~~~~~~~~~~~~~
- `weights_method` : string, defines method for robust
Options here are similar to :mod:`statsmodels.stats.robust_covariance`
default is heteroscedasticity consistent, HC0
currently available methods are
- `cov` : HC0, optionally with degrees of freedom correction
- `hac` :
- `iid` : untested, only for Z*u case, IV cases with u as error indep of Z
- `ac` : not available yet
- `cluster` : not connected yet
- others from robust_covariance
other arguments:
- `wargs` : tuple or dict, required arguments for weights_method
- `centered` : bool,
indicates whether moments are centered for the calculation of the weights
and covariance matrix, applies to all weight_methods
- `ddof` : int
degrees of freedom correction, applies currently only to `cov`
- maxlag : int
number of lags to include in HAC calculation , applies only to `hac`
- others not yet, e.g. groups for cluster robust
covariance matrix
~~~~~~~~~~~~~~~~~
The same options as for weight matrix also apply to the calculation of the
estimate of the covariance matrix of the parameter estimates.
The additional option is
- `has_optimal_weights`: If true, then the calculation of the covariance
matrix assumes that we have optimal GMM with :math:`W = S^{-1}`.
Default is True.
TODO: do we want to have a different default after `onestep`?
'''
[docs]class GMM(Model):
'''
Class for estimation by Generalized Method of Moments
needs to be subclassed, where the subclass defined the moment conditions
`momcond`
Parameters
----------
endog : array
endogenous variable, see notes
exog : array
array of exogenous variables, see notes
instrument : array
array of instruments, see notes
nmoms : None or int
number of moment conditions, if None then it is set equal to the
number of columns of instruments. Mainly needed to determin the shape
or size of start parameters and starting weighting matrix.
kwds : anything
this is mainly if additional variables need to be stored for the
calculations of the moment conditions
Returns
-------
*Attributes*
results : instance of GMMResults
currently just a storage class for params and cov_params without it's
own methods
bse : property
return bse
Notes
-----
The GMM class only uses the moment conditions and does not use any data
directly. endog, exog, instrument and kwds in the creation of the class
instance are only used to store them for access in the moment conditions.
Which of this are required and how they are used depends on the moment
conditions of the subclass.
Warning:
Options for various methods have not been fully implemented and
are still missing in several methods.
TODO:
currently onestep (maxiter=0) still produces an updated estimate of bse
and cov_params.
'''
results_class = 'GMMResults'
def __init__(self, endog, exog, instrument, k_moms=None, k_params=None,
missing='none', **kwds):
'''
maybe drop and use mixin instead
TODO: GMM doesn't really care about the data, just the moment conditions
'''
instrument = self._check_inputs(instrument, endog) # attaches if needed
super(GMM, self).__init__(endog, exog, missing=missing,
instrument=instrument)
# self.endog = endog
# self.exog = exog
# self.instrument = instrument
self.nobs = endog.shape[0]
if k_moms is not None:
self.nmoms = k_moms
elif instrument is not None:
self.nmoms = instrument.shape[1]
else:
self.nmoms = np.nan
if k_params is not None:
self.k_params = k_params
elif instrument is not None:
self.k_params = exog.shape[1]
else:
self.k_params = np.nan
self.__dict__.update(kwds)
self.epsilon_iter = 1e-6
def _check_inputs(self, instrument, endog):
if instrument is not None:
offset = np.asarray(instrument)
if offset.shape[0] != endog.shape[0]:
raise ValueError("instrument is not the same length as endog")
return instrument
def _fix_param_names(self, params, param_names=None):
# TODO: this is a temporary fix, need
xnames = self.data.xnames
if not param_names is None:
if len(params) == len(param_names):
self.data.xnames = param_names
else:
raise ValueError('param_names has the wrong length')
else:
if len(params) < len(xnames):
# cut in front for poisson multiplicative
self.data.xnames = xnames[-len(params):]
elif len(params) > len(xnames):
# use generic names
self.data.xnames = ['p%2d' % i for i in range(len(params))]
[docs] def set_param_names(self, param_names, k_params=None):
"""set the parameter names in the model
Parameters
----------
param_names : list of strings
param_names should have the same length as the number of params
k_params : None or int
If k_params is None, then the k_params attribute is used, unless
it is None.
If k_params is not None, then it will also set the k_params
attribute.
"""
if k_params is not None:
self.k_params = k_params
else:
k_params = self.k_params
if k_params == len(param_names):
self.data.xnames = param_names
else:
raise ValueError('param_names has the wrong length')
[docs] def fit(self, start_params=None, maxiter=10, inv_weights=None,
weights_method='cov', wargs=(),
has_optimal_weights=True,
optim_method='bfgs', optim_args=None):
'''
Estimate parameters using GMM and return GMMResults
TODO: weight and covariance arguments still need to be made consistent
with similar options in other models,
see RegressionResult.get_robustcov_results
Parameters
----------
start_params : array (optional)
starting value for parameters ub minimization. If None then
fitstart method is called for the starting values.
maxiter : int or 'cue'
Number of iterations in iterated GMM. The onestep estimate can be
obtained with maxiter=0 or 1. If maxiter is large, then the
iteration will stop either at maxiter or on convergence of the
parameters (TODO: no options for convergence criteria yet.)
If `maxiter == 'cue'`, the the continuously updated GMM is
calculated which updates the weight matrix during the minimization
of the GMM objective function. The CUE estimation uses the onestep
parameters as starting values.
inv_weights : None or ndarray
inverse of the starting weighting matrix. If inv_weights are not
given then the method `start_weights` is used which depends on
the subclass, for IV subclasses `inv_weights = z'z` where `z` are
the instruments, otherwise an identity matrix is used.
weights_method : string, defines method for robust
Options here are similar to :mod:`statsmodels.stats.robust_covariance`
default is heteroscedasticity consistent, HC0
currently available methods are
- `cov` : HC0, optionally with degrees of freedom correction
- `hac` :
- `iid` : untested, only for Z*u case, IV cases with u as error indep of Z
- `ac` : not available yet
- `cluster` : not connected yet
- others from robust_covariance
wargs` : tuple or dict,
required and optional arguments for weights_method
- `centered` : bool,
indicates whether moments are centered for the calculation of the weights
and covariance matrix, applies to all weight_methods
- `ddof` : int
degrees of freedom correction, applies currently only to `cov`
- `maxlag` : int
number of lags to include in HAC calculation , applies only to `hac`
- others not yet, e.g. groups for cluster robust
has_optimal_weights: If true, then the calculation of the covariance
matrix assumes that we have optimal GMM with :math:`W = S^{-1}`.
Default is True.
TODO: do we want to have a different default after `onestep`?
optim_method : string, default is 'bfgs'
numerical optimization method. Currently not all optimizers that
are available in LikelihoodModels are connected.
optim_args : dict
keyword arguments for the numerical optimizer.
Returns
-------
results : instance of GMMResults
this is also attached as attribute results
Notes
-----
Warning: One-step estimation, `maxiter` either 0 or 1, still has
problems (at least compared to Stata's gmm).
By default it uses a heteroscedasticity robust covariance matrix, but
uses the assumption that the weight matrix is optimal.
See options for cov_params in the results instance.
The same options as for weight matrix also apply to the calculation of
the estimate of the covariance matrix of the parameter estimates.
'''
# TODO: add check for correct wargs keys
# currently a misspelled key is not detected,
# because I'm still adding options
# TODO: check repeated calls to fit with different options
# arguments are dictionaries, i.e. mutable
# unit test if anything is stale or spilled over.
#bug: where does start come from ???
start = start_params # alias for renaming
if start is None:
start = self.fitstart() #TODO: temporary hack
if inv_weights is None:
inv_weights
if optim_args is None:
optim_args = {}
if not 'disp' in optim_args:
optim_args['disp'] = 1
if maxiter == 0 or maxiter == 'cue':
if inv_weights is not None:
weights = np.linalg.pinv(inv_weights)
else:
# let start_weights handle the inv=False for maxiter=0
weights = self.start_weights(inv=False)
params = self.fitgmm(start, weights=weights,
optim_method=optim_method, optim_args=optim_args)
weights_ = weights # temporary alias used in jval
else:
params, weights = self.fititer(start,
maxiter=maxiter,
start_invweights=inv_weights,
weights_method=weights_method,
wargs=wargs,
optim_method=optim_method,
optim_args=optim_args)
# TODO weights returned by fititer is inv_weights - not true anymore
# weights_ currently not necessary and used anymore
weights_ = np.linalg.pinv(weights)
if maxiter == 'cue':
#we have params from maxiter= 0 as starting value
# TODO: need to give weights options to gmmobjective_cu
params = self.fitgmm_cu(params,
optim_method=optim_method,
optim_args=optim_args)
# weights is stored as attribute
weights = self._weights_cu
#TODO: use Bunch instead ?
options_other = {'weights_method':weights_method,
'has_optimal_weights':has_optimal_weights,
'optim_method':optim_method}
# check that we have the right number of xnames
self._fix_param_names(params, param_names=None)
results = results_class_dict[self.results_class](
model = self,
params = params,
weights = weights,
wargs = wargs,
options_other = options_other,
optim_args = optim_args)
self.results = results # FIXME: remove, still keeping it temporarily
return results
[docs] def fitgmm(self, start, weights=None, optim_method='bfgs', optim_args=None):
'''estimate parameters using GMM
Parameters
----------
start : array_like
starting values for minimization
weights : array
weighting matrix for moment conditions. If weights is None, then
the identity matrix is used
Returns
-------
paramest : array
estimated parameters
Notes
-----
todo: add fixed parameter option, not here ???
uses scipy.optimize.fmin
'''
## if not fixed is None: #fixed not defined in this version
## raise NotImplementedError
# TODO: should start_weights only be in `fit`
if weights is None:
weights = self.start_weights(inv=False)
if optim_args is None:
optim_args = {}
if optim_method == 'nm':
optimizer = optimize.fmin
elif optim_method == 'bfgs':
optimizer = optimize.fmin_bfgs
# TODO: add score
optim_args['fprime'] = self.score #lambda params: self.score(params, weights)
elif optim_method == 'ncg':
optimizer = optimize.fmin_ncg
optim_args['fprime'] = self.score
elif optim_method == 'cg':
optimizer = optimize.fmin_cg
optim_args['fprime'] = self.score
elif optim_method == 'fmin_l_bfgs_b':
optimizer = optimize.fmin_l_bfgs_b
optim_args['fprime'] = self.score
elif optim_method == 'powell':
optimizer = optimize.fmin_powell
elif optim_method == 'slsqp':
optimizer = optimize.fmin_slsqp
else:
raise ValueError('optimizer method not available')
if DEBUG:
print(np.linalg.det(weights))
#TODO: add other optimization options and results
return optimizer(self.gmmobjective, start, args=(weights,),
**optim_args)
[docs] def fitgmm_cu(self, start, optim_method='bfgs', optim_args=None):
'''estimate parameters using continuously updating GMM
Parameters
----------
start : array_like
starting values for minimization
Returns
-------
paramest : array
estimated parameters
Notes
-----
todo: add fixed parameter option, not here ???
uses scipy.optimize.fmin
'''
## if not fixed is None: #fixed not defined in this version
## raise NotImplementedError
if optim_args is None:
optim_args = {}
if optim_method == 'nm':
optimizer = optimize.fmin
elif optim_method == 'bfgs':
optimizer = optimize.fmin_bfgs
optim_args['fprime'] = self.score_cu
elif optim_method == 'ncg':
optimizer = optimize.fmin_ncg
else:
raise ValueError('optimizer method not available')
#TODO: add other optimization options and results
return optimizer(self.gmmobjective_cu, start, args=(), **optim_args)
[docs] def start_weights(self, inv=True):
return np.eye(self.nmoms)
[docs] def gmmobjective(self, params, weights):
'''
objective function for GMM minimization
Parameters
----------
params : array
parameter values at which objective is evaluated
weights : array
weighting matrix
Returns
-------
jval : float
value of objective function
'''
moms = self.momcond_mean(params)
return np.dot(np.dot(moms, weights), moms)
#moms = self.momcond(params)
#return np.dot(np.dot(moms.mean(0),weights), moms.mean(0))
[docs] def gmmobjective_cu(self, params, weights_method='cov',
wargs=()):
'''
objective function for continuously updating GMM minimization
Parameters
----------
params : array
parameter values at which objective is evaluated
Returns
-------
jval : float
value of objective function
'''
moms = self.momcond(params)
inv_weights = self.calc_weightmatrix(moms, weights_method=weights_method,
wargs=wargs)
weights = np.linalg.pinv(inv_weights)
self._weights_cu = weights # store if we need it later
return np.dot(np.dot(moms.mean(0), weights), moms.mean(0))
[docs] def fititer(self, start, maxiter=2, start_invweights=None,
weights_method='cov', wargs=(), optim_method='bfgs',
optim_args=None):
'''iterative estimation with updating of optimal weighting matrix
stopping criteria are maxiter or change in parameter estimate less
than self.epsilon_iter, with default 1e-6.
Parameters
----------
start : array
starting value for parameters
maxiter : int
maximum number of iterations
start_weights : array (nmoms, nmoms)
initial weighting matrix; if None, then the identity matrix
is used
weights_method : {'cov', ...}
method to use to estimate the optimal weighting matrix,
see calc_weightmatrix for details
Returns
-------
params : array
estimated parameters
weights : array
optimal weighting matrix calculated with final parameter
estimates
Notes
-----
'''
self.history = []
momcond = self.momcond
if start_invweights is None:
w = self.start_weights(inv=True)
else:
w = start_invweights
#call fitgmm function
#args = (self.endog, self.exog, self.instrument)
#args is not used in the method version
winv_new = w
for it in range(maxiter):
winv = winv_new
w = np.linalg.pinv(winv)
#this is still calling function not method
## resgmm = fitgmm(momcond, (), start, weights=winv, fixed=None,
## weightsoptimal=False)
resgmm = self.fitgmm(start, weights=w, optim_method=optim_method,
optim_args=optim_args)
moms = momcond(resgmm)
# the following is S = cov_moments
winv_new = self.calc_weightmatrix(moms,
weights_method=weights_method,
wargs=wargs, params=resgmm)
if it > 2 and maxabs(resgmm - start) < self.epsilon_iter:
#check rule for early stopping
# TODO: set has_optimal_weights = True
break
start = resgmm
return resgmm, w
[docs] def calc_weightmatrix(self, moms, weights_method='cov', wargs=(),
params=None):
'''
calculate omega or the weighting matrix
Parameters
----------
moms : array
moment conditions (nobs x nmoms) for all observations evaluated at
a parameter value
weights_method : string 'cov'
If method='cov' is cov then the matrix is calculated as simple
covariance of the moment conditions.
see fit method for available aoptions for the weight and covariance
matrix
wargs : tuple or dict
parameters that are required by some kernel methods to
estimate the long-run covariance. Not used yet.
Returns
-------
w : array (nmoms, nmoms)
estimate for the weighting matrix or covariance of the moment
condition
Notes
-----
currently a constant cutoff window is used
TODO: implement long-run cov estimators, kernel-based
Newey-West
Andrews
Andrews-Moy????
References
----------
Greene
Hansen, Bruce
'''
nobs, k_moms = moms.shape
# TODO: wargs are tuple or dict ?
if DEBUG:
print(' momcov wargs', wargs)
centered = not ('centered' in wargs and not wargs['centered'])
if not centered:
# caller doesn't want centered moment conditions
moms_ = moms
else:
moms_ = moms - moms.mean()
# TODO: store this outside to avoid doing this inside optimization loop
# TODO: subclasses need to be able to add weights_methods, and remove
# IVGMM can have homoscedastic (OLS),
# some options won't make sense in some cases
# possible add all here and allow subclasses to define a list
# TODO: should other weights_methods also have `ddof`
if weights_method == 'cov':
w = np.dot(moms_.T, moms_)
if 'ddof' in wargs:
# caller requests degrees of freedom correction
if wargs['ddof'] == 'k_params':
w /= (nobs - self.k_params)
else:
if DEBUG:
print(' momcov ddof', wargs['ddof'])
w /= (nobs - wargs['ddof'])
else:
# default: divide by nobs
w /= nobs
elif weights_method == 'flatkernel':
#uniform cut-off window
# This was a trial version, can use HAC with flatkernel
if not 'maxlag' in wargs:
raise ValueError('flatkernel requires maxlag')
maxlag = wargs['maxlag']
h = np.ones(maxlag + 1)
w = np.dot(moms_.T, moms_)/nobs
for i in range(1,maxlag+1):
w += (h[i] * np.dot(moms_[i:].T, moms_[:-i]) / (nobs-i))
elif weights_method == 'hac':
maxlag = wargs['maxlag']
if 'kernel' in wargs:
weights_func = wargs['kernel']
else:
weights_func = smcov.weights_bartlett
wargs['kernel'] = weights_func
w = smcov.S_hac_simple(moms_, nlags=maxlag,
weights_func=weights_func)
w /= nobs #(nobs - self.k_params)
elif weights_method == 'iid':
# only when we have instruments and residual mom = Z * u
# TODO: problem we don't have params in argument
# I cannot keep everything in here w/o params as argument
u = self.get_error(params)
if centered:
# Note: I'm not centering instruments,
# shouldn't we always center u? Ok, with centered as default
u -= u.mean(0) #demean inplace, we don't need original u
instrument = self.instrument
w = np.dot(instrument.T, instrument).dot(np.dot(u.T, u)) / nobs
if 'ddof' in wargs:
# caller requests degrees of freedom correction
if wargs['ddof'] == 'k_params':
w /= (nobs - self.k_params)
else:
# assume ddof is a number
if DEBUG:
print(' momcov ddof', wargs['ddof'])
w /= (nobs - wargs['ddof'])
else:
# default: divide by nobs
w /= nobs
else:
raise ValueError('weight method not available')
return w
[docs] def momcond_mean(self, params):
'''
mean of moment conditions,
'''
momcond = self.momcond(params)
self.nobs_moms, self.k_moms = momcond.shape
return momcond.mean(0)
[docs] def gradient_momcond(self, params, epsilon=1e-4, centered=True):
'''gradient of moment conditions
Parameters
----------
params : ndarray
parameter at which the moment conditions are evaluated
epsilon : float
stepsize for finite difference calculation
centered : bool
This refers to the finite difference calculation. If `centered`
is true, then the centered finite difference calculation is
used. Otherwise the one-sided forward differences are used.
TODO: looks like not used yet
missing argument `weights`
'''
momcond = self.momcond_mean
# TODO: approx_fprime has centered keyword
if centered:
gradmoms = (approx_fprime(params, momcond, epsilon=epsilon) +
approx_fprime(params, momcond, epsilon=-epsilon))/2
else:
gradmoms = approx_fprime(params, momcond, epsilon=epsilon)
return gradmoms
[docs] def score(self, params, weights, epsilon=None, centered=True):
deriv = approx_fprime(params, self.gmmobjective, args=(weights,),
centered=centered, epsilon=epsilon)
return deriv
[docs] def score_cu(self, params, epsilon=None, centered=True):
deriv = approx_fprime(params, self.gmmobjective_cu, args=(),
centered=centered, epsilon=epsilon)
return deriv
# TODO: wrong superclass, I want tvalues, ... right now
[docs]class GMMResults(LikelihoodModelResults):
'''just a storage class right now'''
use_t = False
def __init__(self, *args, **kwds):
self.__dict__.update(kwds)
self.nobs = self.model.nobs
self.df_resid = np.inf
self.cov_params_default = self._cov_params()
[docs] @cache_readonly
def q(self):
return self.model.gmmobjective(self.params, self.weights)
[docs] @cache_readonly
def jval(self):
# nobs_moms attached by momcond_mean
return self.q * self.model.nobs_moms
def _cov_params(self, **kwds):
#TODO add options ???)
# this should use by default whatever options have been specified in
# fit
# TODO: don't do this when we want to change options
# if hasattr(self, '_cov_params'):
# #replace with decorator later
# return self._cov_params
# set defaults based on fit arguments
if not 'wargs' in kwds:
# Note: we don't check the keys in wargs, use either all or nothing
kwds['wargs'] = self.wargs
if not 'weights_method' in kwds:
kwds['weights_method'] = self.options_other['weights_method']
if not 'has_optimal_weights' in kwds:
kwds['has_optimal_weights'] = self.options_other['has_optimal_weights']
gradmoms = self.model.gradient_momcond(self.params)
moms = self.model.momcond(self.params)
covparams = self.calc_cov_params(moms, gradmoms, **kwds)
return covparams
[docs] def calc_cov_params(self, moms, gradmoms, weights=None, use_weights=False,
has_optimal_weights=True,
weights_method='cov', wargs=()):
'''calculate covariance of parameter estimates
not all options tried out yet
If weights matrix is given, then the formula use to calculate cov_params
depends on whether has_optimal_weights is true.
If no weights are given, then the weight matrix is calculated with
the given method, and has_optimal_weights is assumed to be true.
(API Note: The latter assumption could be changed if we allow for
has_optimal_weights=None.)
'''
nobs = moms.shape[0]
if weights is None:
#omegahat = self.model.calc_weightmatrix(moms, method=method, wargs=wargs)
#has_optimal_weights = True
#add other options, Barzen, ... longrun var estimators
# TODO: this might still be inv_weights after fititer
weights = self.weights
else:
pass
#omegahat = weights #2 different names used,
#TODO: this is wrong, I need an estimate for omega
if use_weights:
omegahat = weights
else:
omegahat = self.model.calc_weightmatrix(
moms,
weights_method=weights_method,
wargs=wargs,
params=self.params)
if has_optimal_weights: #has_optimal_weights:
# TOD0 make has_optimal_weights depend on convergence or iter >2
cov = np.linalg.inv(np.dot(gradmoms.T,
np.dot(np.linalg.inv(omegahat), gradmoms)))
else:
gw = np.dot(gradmoms.T, weights)
gwginv = np.linalg.inv(np.dot(gw, gradmoms))
cov = np.dot(np.dot(gwginv, np.dot(np.dot(gw, omegahat), gw.T)), gwginv)
#cov /= nobs
return cov/nobs
@property
def bse_(self):
'''standard error of the parameter estimates
'''
return self.get_bse()
[docs] def get_bse(self, **kwds):
'''standard error of the parameter estimates with options
Parameters
----------
kwds : optional keywords
options for calculating cov_params
Returns
-------
bse : ndarray
estimated standard error of parameter estimates
'''
return np.sqrt(np.diag(self.cov_params(**kwds)))
[docs] def jtest(self):
'''overidentification test
I guess this is missing a division by nobs,
what's the normalization in jval ?
'''
jstat = self.jval
nparams = self.params.size #self.nparams
df = self.model.nmoms - nparams
return jstat, stats.chi2.sf(jstat, df), df
[docs] def compare_j(self, other):
'''overidentification test for comparing two nested gmm estimates
This assumes that some moment restrictions have been dropped in one
of the GMM estimates relative to the other.
Not tested yet
We are comparing two separately estimated models, that use different
weighting matrices. It is not guaranteed that the resulting
difference is positive.
TODO: Check in which cases Stata programs use the same weigths
'''
jstat1 = self.jval
k_moms1 = self.model.nmoms
jstat2 = other.jval
k_moms2 = other.model.nmoms
jdiff = jstat1 - jstat2
df = k_moms1 - k_moms2
if df < 0:
# possible nested in other way, TODO allow this or not
# flip sign instead of absolute
df = - df
jdiff = - jdiff
return jdiff, stats.chi2.sf(jdiff, df), df
[docs] 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: add a summary text for options that have been used
jvalue, jpvalue, jdf = self.jtest()
top_left = [('Dep. Variable:', None),
('Model:', None),
('Method:', ['GMM']),
('Date:', None),
('Time:', None),
('No. Observations:', None),
#('Df Residuals:', None), #[self.df_resid]), #TODO: spelling
#('Df Model:', None), #[self.df_model])
]
top_right = [#('R-squared:', ["%#8.3f" % self.rsquared]),
#('Adj. R-squared:', ["%#8.3f" % self.rsquared_adj]),
('Hansen J:', ["%#8.4g" % jvalue] ),
('Prob (Hansen J):', ["%#6.3g" % jpvalue]),
#('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])
]
if title is None:
title = self.model.__class__.__name__ + ' ' + "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=False)
return smry
[docs]class IVGMM(GMM):
'''
Basic class for instrumental variables estimation using GMM
A linear function for the conditional mean is defined as default but the
methods should be overwritten by subclasses, currently `LinearIVGMM` and
`NonlinearIVGMM` are implemented as subclasses.
See Also
--------
LinearIVGMM
NonlinearIVGMM
'''
results_class = 'IVGMMResults'
[docs] def fitstart(self):
return np.zeros(self.exog.shape[1])
[docs] def start_weights(self, inv=True):
zz = np.dot(self.instrument.T, self.instrument)
nobs = self.instrument.shape[0]
if inv:
return zz / nobs
else:
return np.linalg.pinv(zz / nobs)
[docs] def get_error(self, params):
return self.endog - self.predict(params)
[docs] def predict(self, params, exog=None):
if exog is None:
exog = self.exog
return np.dot(exog, params)
[docs] def momcond(self, params):
instrument = self.instrument
return instrument * self.get_error(params)[:,None]
[docs]class LinearIVGMM(IVGMM):
"""class for linear instrumental variables models estimated with GMM
Uses closed form expression instead of nonlinear optimizers for each step
of the iterative GMM.
The model is assumed to have the following moment condition
E( z * (y - x beta)) = 0
Where `y` is the dependent endogenous variable, `x` are the explanatory
variables and `z` are the instruments. Variables in `x` that are exogenous
need also be included in `z`.
Notation Warning: our name `exog` stands for the explanatory variables,
and includes both exogenous and explanatory variables that are endogenous,
i.e. included endogenous variables
Parameters
----------
endog : array_like
dependent endogenous variable
exog : array_like
explanatory, right hand side variables, including explanatory variables
that are endogenous
instrument : array_like
Instrumental variables, variables that are exogenous to the error
in the linear model containing both included and excluded exogenous
variables
"""
[docs] def fitgmm(self, start, weights=None, optim_method=None, **kwds):
'''estimate parameters using GMM for linear model
Uses closed form expression instead of nonlinear optimizers
Parameters
----------
start : not used
starting values for minimization, not used, only for consistency
of method signature
weights : array
weighting matrix for moment conditions. If weights is None, then
the identity matrix is used
optim_method : not used,
optimization method, not used, only for consistency of method
signature
**kwds : keyword arguments
not used, will be silently ignored (for compatibility with generic)
Returns
-------
paramest : array
estimated parameters
'''
## if not fixed is None: #fixed not defined in this version
## raise NotImplementedError
# TODO: should start_weights only be in `fit`
if weights is None:
weights = self.start_weights(inv=False)
y, x, z = self.endog, self.exog, self.instrument
zTx = np.dot(z.T, x)
zTy = np.dot(z.T, y)
# normal equation, solved with pinv
part0 = zTx.T.dot(weights)
part1 = part0.dot(zTx)
part2 = part0.dot(zTy)
params = np.linalg.pinv(part1).dot(part2)
return params
[docs] def predict(self, params, exog=None):
if exog is None:
exog = self.exog
return np.dot(exog, params)
[docs] def gradient_momcond(self, params, **kwds):
# **kwds for compatibility not used
x, z = self.exog, self.instrument
gradmoms = -np.dot(z.T, x) / self.nobs
return gradmoms
[docs] def score(self, params, weights, **kwds):
# **kwds for compatibility, not used
# Note: I coud use general formula with gradient_momcond instead
x, z = self.exog, self.instrument
nobs = z.shape[0]
u = self.get_errors(params)
score = -2 * np.dot(x.T, z).dot(weights.dot(np.dot(z.T, u)))
score /= nobs * nobs
return score
[docs]class NonlinearIVGMM(IVGMM):
"""
Class for non-linear instrumental variables estimation wusing GMM
The model is assumed to have the following moment condition
E[ z * (y - f(X, beta)] = 0
Where `y` is the dependent endogenous variable, `x` are the explanatory
variables and `z` are the instruments. Variables in `x` that are exogenous
need also be included in z. `f` is a nonlinear function.
Notation Warning: our name `exog` stands for the explanatory variables,
and includes both exogenous and explanatory variables that are endogenous,
i.e. included endogenous variables
Parameters
----------
endog : array_like
dependent endogenous variable
exog : array_like
explanatory, right hand side variables, including explanatory variables
that are endogenous.
instruments : array_like
Instrumental variables, variables that are exogenous to the error
in the linear model containing both included and excluded exogenous
variables
func : callable
function for the mean or conditional expectation of the endogenous
variable. The function will be called with parameters and the array of
explanatory, right hand side variables, `func(params, exog)`
Notes
-----
This class uses numerical differences to obtain the derivative of the
objective function. If the jacobian of the conditional mean function, `func`
is available, then it can be used by subclassing this class and defining
a method `jac_func`.
TODO: check required signature of jac_error and jac_func
"""
# This should be reversed:
# NonlinearIVGMM is IVGMM and need LinearIVGMM as special case (fit, predict)
[docs] def fitstart(self):
#might not make sense for more general functions
return np.zeros(self.exog.shape[1])
def __init__(self, endog, exog, instrument, func, **kwds):
self.func = func
super(NonlinearIVGMM, self).__init__(endog, exog, instrument, **kwds)
[docs] def predict(self, params, exog=None):
if exog is None:
exog = self.exog
return self.func(params, exog)
#---------- the following a semi-general versions,
# TODO: move to higher class after testing
[docs] def jac_func(self, params, weights, args=None, centered=True, epsilon=None):
# TODO: Why are ther weights in the signature - copy-paste error?
deriv = approx_fprime(params, self.func, args=(self.exog,),
centered=centered, epsilon=epsilon)
return deriv
[docs] def jac_error(self, params, weights, args=None, centered=True,
epsilon=None):
jac_func = self.jac_func(params, weights, args=None, centered=True,
epsilon=None)
return -jac_func
[docs] def score(self, params, weights, **kwds):
# **kwds for compatibility not used
# Note: I coud use general formula with gradient_momcond instead
z = self.instrument
nobs = z.shape[0]
jac_u = self.jac_error(params, weights, args=None, epsilon=None,
centered=True)
x = -jac_u # alias, plays the same role as X in linear model
u = self.get_error(params)
score = -2 * np.dot(np.dot(x.T, z), weights).dot(np.dot(z.T, u))
score /= nobs * nobs
return score
[docs]class IVGMMResults(GMMResults):
# this assumes that we have an additive error model `(y - f(x, params))`
[docs] @cache_readonly
def fittedvalues(self):
return self.model.predict(self.params)
[docs] @cache_readonly
def resid(self):
return self.model.endog - self.fittedvalues
[docs] @cache_readonly
def ssr(self):
return (self.resid * self.resid).sum(0)
def spec_hausman(params_e, params_i, cov_params_e, cov_params_i, dof=None):
'''Hausmans specification test
Parameters
----------
params_e : array
efficient and consistent under Null hypothesis,
inconsistent under alternative hypothesis
params_i: array
consistent under Null hypothesis,
consistent under alternative hypothesis
cov_params_e : array, 2d
covariance matrix of parameter estimates for params_e
cov_params_i : array, 2d
covariance matrix of parameter estimates for params_i
example instrumental variables OLS estimator is `e`, IV estimator is `i`
Notes
-----
Todos,Issues
- check dof calculations and verify for linear case
- check one-sided hypothesis
References
----------
Greene section 5.5 p.82/83
'''
params_diff = (params_i - params_e)
cov_diff = cov_params_i - cov_params_e
#TODO: the following is very inefficient, solves problem (svd) twice
#use linalg.lstsq or svd directly
#cov_diff will very often be in-definite (singular)
if not dof:
dof = np_matrix_rank(cov_diff)
cov_diffpinv = np.linalg.pinv(cov_diff)
H = np.dot(params_diff, np.dot(cov_diffpinv, params_diff))
pval = stats.chi2.sf(H, dof)
evals = np.linalg.eigvalsh(cov_diff)
return H, pval, dof, evals
###########
class DistQuantilesGMM(GMM):
'''
Estimate distribution parameters by GMM based on matching quantiles
Currently mainly to try out different requirements for GMM when we cannot
calculate the optimal weighting matrix.
'''
def __init__(self, endog, exog, instrument, **kwds):
#TODO: something wrong with super
super(DistQuantilesGMM, self).__init__(endog, exog, instrument)
#self.func = func
self.epsilon_iter = 1e-5
self.distfn = kwds['distfn']
#done by super doesn't work yet
#TypeError: super does not take keyword arguments
self.endog = endog
#make this optional for fit
if not 'pquant' in kwds:
self.pquant = pquant = np.array([0.01, 0.05,0.1,0.4,0.6,0.9,0.95,0.99])
else:
self.pquant = pquant = kwds['pquant']
#TODO: vectorize this: use edf
self.xquant = np.array([stats.scoreatpercentile(endog, p) for p
in pquant*100])
self.nmoms = len(self.pquant)
#TODOcopied from GMM, make super work
self.endog = endog
self.exog = exog
self.instrument = instrument
self.results = GMMResults(model=self)
#self.__dict__.update(kwds)
self.epsilon_iter = 1e-6
def fitstart(self):
#todo: replace with or add call to distfn._fitstart
# added but not used during testing, avoid Travis
distfn = self.distfn
if hasattr(distfn, '_fitstart'):
start = distfn._fitstart(self.endog)
else:
start = [1]*distfn.numargs + [0.,1.]
return np.asarray(start)
def momcond(self, params): #drop distfn as argument
#, mom2, quantile=None, shape=None
'''moment conditions for estimating distribution parameters by matching
quantiles, defines as many moment conditions as quantiles.
Returns
-------
difference : array
difference between theoretical and empirical quantiles
Notes
-----
This can be used for method of moments or for generalized method of
moments.
'''
#this check looks redundant/unused know
if len(params) == 2:
loc, scale = params
elif len(params) == 3:
shape, loc, scale = params
else:
#raise NotImplementedError
pass #see whether this might work, seems to work for beta with 2 shape args
#mom2diff = np.array(distfn.stats(*params)) - mom2
#if not quantile is None:
pq, xq = self.pquant, self.xquant
#ppfdiff = distfn.ppf(pq, alpha)
cdfdiff = self.distfn.cdf(xq, *params) - pq
#return np.concatenate([mom2diff, cdfdiff[:1]])
return np.atleast_2d(cdfdiff)
def fitonce(self, start=None, weights=None, has_optimal_weights=False):
'''fit without estimating an optimal weighting matrix and return results
This is a convenience function that calls fitgmm and covparams with
a given weight matrix or the identity weight matrix.
This is useful if the optimal weight matrix is know (or is analytically
given) or if an optimal weight matrix cannot be calculated.
(Developer Notes: this function could go into GMM, but is needed in this
class, at least at the moment.)
Parameters
----------
Returns
-------
results : GMMResult instance
result instance with params and _cov_params attached
See Also
--------
fitgmm
cov_params
'''
if weights is None:
weights = np.eye(self.nmoms)
params = self.fitgmm(start=start)
# TODO: rewrite this old hack, should use fitgmm or fit maxiter=0
self.results.params = params #required before call to self.cov_params
self.results.wargs = {} #required before call to self.cov_params
self.results.options_other = {'weights_method':'cov'}
# TODO: which weights_method? There shouldn't be any needed ?
_cov_params = self.results.cov_params(weights=weights,
has_optimal_weights=has_optimal_weights)
self.results.weights = weights
self.results.jval = self.gmmobjective(params, weights)
self.results.options_other.update({'has_optimal_weights':has_optimal_weights})
return self.results
results_class_dict = {'GMMResults': GMMResults,
'IVGMMResults': IVGMMResults,
'DistQuantilesGMM': GMMResults} #TODO: should be a default
