def _handle_constant(self, hasconst):
if hasconst is not None:
if hasconst:
self.k_constant = 1
self.const_idx = None
else:
self.k_constant = 0
self.const_idx = None
elif self.exog is None:
self.const_idx = None
self.k_constant = 0
else:
# detect where the constant is
check_implicit = False
const_idx = np.where(self.exog.ptp(axis=0) == 0)[0].squeeze()
self.k_constant = const_idx.size
if self.k_constant == 1:
if self.exog[:, const_idx].mean() != 0:
self.const_idx = const_idx
else:
# we only have a zero column and no other constant
check_implicit = True
elif self.k_constant > 1:
# we have more than one constant column
# look for ones
values = [] # keep values if we need != 0
for idx in const_idx:
value = self.exog[:, idx].mean()
if value == 1:
self.k_constant = 1
self.const_idx = idx
break
values.append(value)
else:
# we didn't break, no column of ones
pos = (np.array(values) != 0)
if pos.any():
# take the first nonzero column
self.k_constant = 1
self.const_idx = const_idx[pos.argmax()]
else:
# only zero columns
check_implicit = True
elif self.k_constant == 0:
check_implicit = True
else:
# shouldn't be here
pass
if check_implicit:
# look for implicit constant
# Compute rank of augmented matrix
augmented_exog = np.column_stack(
(np.ones(self.exog.shape[0]), self.exog))
rank_augm = np_matrix_rank(augmented_exog)
rank_orig = np_matrix_rank(self.exog)
self.k_constant = int(rank_orig == rank_augm)
self.const_idx = None
def _handle_constant(self, hasconst):
if hasconst is not None:
if hasconst:
self.k_constant = 1
self.const_idx = None
else:
self.k_constant = 0
self.const_idx = None
elif self.exog is None:
self.const_idx = None
self.k_constant = 0
else:
# detect where the constant is
check_implicit = False
const_idx = np.where(self.exog.ptp(axis=0) == 0)[0].squeeze()
self.k_constant = const_idx.size
if self.k_constant == 1:
if self.exog[:, const_idx].mean() != 0:
self.const_idx = const_idx
else:
# we only have a zero column and no other constant
check_implicit = True
elif self.k_constant > 1:
# we have more than one constant column
# look for ones
values = [] # keep values if we need != 0
for idx in const_idx:
value = self.exog[:, idx].mean()
if value == 1:
self.k_constant = 1
self.const_idx = idx
break
values.append(value)
else:
# we didn't break, no column of ones
pos = (np.array(values) != 0)
if pos.any():
# take the first nonzero column
self.k_constant = 1
self.const_idx = const_idx[pos.argmax()]
else:
# only zero columns
check_implicit = True
elif self.k_constant == 0:
check_implicit = True
else:
# shouldn't be here
pass
if check_implicit:
# look for implicit constant
# Compute rank of augmented matrix
augmented_exog = np.column_stack(
(np.ones(self.exog.shape[0]), self.exog))
rank_augm = np_matrix_rank(augmented_exog)
rank_orig = np_matrix_rank(self.exog)
self.k_constant = int(rank_orig == rank_augm)
self.const_idx = None
def test_rank(self):
import warnings
with warnings.catch_warnings():
warnings.simplefilter("ignore")
X = standard_normal((40, 10))
self.assertEquals(tools.rank(X), np_matrix_rank(X))
X[:, 0] = X[:, 1] + X[:, 2]
self.assertEquals(tools.rank(X), np_matrix_rank(X))
def test_rank(self):
import warnings
with warnings.catch_warnings():
warnings.simplefilter("ignore")
X = standard_normal((40,10))
self.assertEquals(tools.rank(X), np_matrix_rank(X))
X[:,0] = X[:,1] + X[:,2]
self.assertEquals(tools.rank(X), np_matrix_rank(X))
def __init__(self, sys, sigma=None, dfk=None):
if len(sys) % 2 != 0:
raise ValueError("sys must be a list of pairs of endogenous and \
exogenous variables. Got length %s" % len(sys))
if dfk:
if not dfk.lower() in ['dfk1', 'dfk2']:
raise ValueError("dfk option %s not understood" % (dfk))
self._dfk = dfk
M = len(sys[1::2])
self._M = M
# exog = np.zeros((M,M), dtype=object)
# for i,eq in enumerate(sys[1::2]):
# exog[i,i] = np.asarray(eq) # not sure this exog is needed
# used to compute resids for now
exog = np.column_stack(np.asarray(sys[1::2][i]) for i in range(M))
# exog = np.vstack(np.asarray(sys[1::2][i]) for i in range(M))
self.exog = exog # 2d ndarray exog is better
# Endog, might just go ahead and reshape this?
endog = np.asarray(sys[::2])
self.endog = endog
self.nobs = float(
self.endog[0].shape[0]) # assumes all the same length
# Degrees of Freedom
df_resid = []
df_model = []
[df_resid.append(self.nobs - np_matrix_rank(_)) for _ in sys[1::2]]
[df_model.append(np_matrix_rank(_) - 1) for _ in sys[1::2]]
self.df_resid = np.asarray(df_resid)
self.df_model = np.asarray(df_model)
# "Block-diagonal" sparse matrix of exog
sp_exog = sparse.lil_matrix(
(int(self.nobs * M),
int(np.sum(self.df_model + 1)))) # linked lists to build
self._cols = np.cumsum(np.hstack((0, self.df_model + 1)))
for i in range(M):
sp_exog[i * self.nobs:(i + 1) * self.nobs,
self._cols[i]:self._cols[i + 1]] = sys[1::2][i]
self.sp_exog = sp_exog.tocsr() # cast to compressed for efficiency
# Deal with sigma, check shape earlier if given
if np.any(sigma):
sigma = np.asarray(sigma) # check shape
elif sigma == None:
resids = []
for i in range(M):
resids.append(
GLS(endog[i],
exog[:, self._cols[i]:self._cols[i + 1]]).fit().resid)
resids = np.asarray(resids).reshape(M, -1)
sigma = self._compute_sigma(resids)
self.sigma = sigma
self.cholsigmainv = np.linalg.cholesky(np.linalg.pinv(\
self.sigma)).T
self.initialize()
def _initialize(self):
"""
Initializes the model for the IRLS fit.
Resets the history and number of iterations.
"""
self.pinv_wexog = np.linalg.pinv(self.exog)
self.normalized_cov_params = np.dot(self.pinv_wexog, np.transpose(self.pinv_wexog))
self.df_resid = np.float(self.exog.shape[0] - np_matrix_rank(self.exog))
self.df_model = np.float(np_matrix_rank(self.exog) - 1)
self.nobs = float(self.endog.shape[0])
def initialize(self):
"""
Initialize a generalized linear model.
"""
# TODO: intended for public use?
self.history = {"fittedvalues": [], "params": [np.inf], "deviance": [np.inf]}
self.pinv_wexog = np.linalg.pinv(self.exog)
self.normalized_cov_params = np.dot(self.pinv_wexog, np.transpose(self.pinv_wexog))
self.df_model = np_matrix_rank(self.exog) - 1
self.df_resid = self.exog.shape[0] - np_matrix_rank(self.exog)
def __init__(self, sys, sigma=None, dfk=None):
if len(sys) % 2 != 0:
raise ValueError("sys must be a list of pairs of endogenous and \
exogenous variables. Got length %s" % len(sys))
if dfk:
if not dfk.lower() in ['dfk1','dfk2']:
raise ValueError("dfk option %s not understood" % (dfk))
self._dfk = dfk
M = len(sys[1::2])
self._M = M
# exog = np.zeros((M,M), dtype=object)
# for i,eq in enumerate(sys[1::2]):
# exog[i,i] = np.asarray(eq) # not sure this exog is needed
# used to compute resids for now
exog = np.column_stack(np.asarray(sys[1::2][i]) for i in range(M))
# exog = np.vstack(np.asarray(sys[1::2][i]) for i in range(M))
self.exog = exog # 2d ndarray exog is better
# Endog, might just go ahead and reshape this?
endog = np.asarray(sys[::2])
self.endog = endog
self.nobs = float(self.endog[0].shape[0]) # assumes all the same length
# Degrees of Freedom
df_resid = []
df_model = []
[df_resid.append(self.nobs - np_matrix_rank(_)) for _ in sys[1::2]]
[df_model.append(np_matrix_rank(_) - 1) for _ in sys[1::2]]
self.df_resid = np.asarray(df_resid)
self.df_model = np.asarray(df_model)
# "Block-diagonal" sparse matrix of exog
sp_exog = sparse.lil_matrix((int(self.nobs*M),
int(np.sum(self.df_model+1)))) # linked lists to build
self._cols = np.cumsum(np.hstack((0, self.df_model+1)))
for i in range(M):
sp_exog[i*self.nobs:(i+1)*self.nobs,
self._cols[i]:self._cols[i+1]] = sys[1::2][i]
self.sp_exog = sp_exog.tocsr() # cast to compressed for efficiency
# Deal with sigma, check shape earlier if given
if np.any(sigma):
sigma = np.asarray(sigma) # check shape
elif sigma == None:
resids = []
for i in range(M):
resids.append(GLS(endog[i],exog[:,
self._cols[i]:self._cols[i+1]]).fit().resid)
resids = np.asarray(resids).reshape(M,-1)
sigma = self._compute_sigma(resids)
self.sigma = sigma
self.cholsigmainv = np.linalg.cholesky(np.linalg.pinv(\
self.sigma)).T
self.initialize()
def _initialize(self):
"""
Initializes the model for the IRLS fit.
Resets the history and number of iterations.
"""
self.pinv_wexog = np.linalg.pinv(self.exog)
self.normalized_cov_params = np.dot(self.pinv_wexog,
np.transpose(self.pinv_wexog))
self.df_resid = (np.float(self.exog.shape[0] -
np_matrix_rank(self.exog)))
self.df_model = np.float(np_matrix_rank(self.exog) - 1)
self.nobs = float(self.endog.shape[0])
def initialize(self):
"""
Initialize a generalized linear model.
"""
#TODO: intended for public use?
self.history = {'fittedvalues' : [],
'params' : [np.inf],
'deviance' : [np.inf]}
self.pinv_wexog = np.linalg.pinv(self.exog)
self.normalized_cov_params = np.dot(self.pinv_wexog,
np.transpose(self.pinv_wexog))
self.df_model = np_matrix_rank(self.exog)-1
self.df_resid = self.exog.shape[0] - np_matrix_rank(self.exog)
def contrastfromcols(L, D, pseudo=None):
"""
From an n x p design matrix D and a matrix L, tries
to determine a p x q contrast matrix C which
determines a contrast of full rank, i.e. the
n x q matrix
dot(transpose(C), pinv(D))
is full rank.
L must satisfy either L.shape[0] == n or L.shape[1] == p.
If L.shape[0] == n, then L is thought of as representing
columns in the column space of D.
If L.shape[1] == p, then L is thought of as what is known
as a contrast matrix. In this case, this function returns an estimable
contrast corresponding to the dot(D, L.T)
Note that this always produces a meaningful contrast, not always
with the intended properties because q is always non-zero unless
L is identically 0. That is, it produces a contrast that spans
the column space of L (after projection onto the column space of D).
Parameters
----------
L : array-like
D : array-like
"""
L = np.asarray(L)
D = np.asarray(D)
n, p = D.shape
if L.shape[0] != n and L.shape[1] != p:
raise ValueError("shape of L and D mismatched")
if pseudo is None:
pseudo = np.linalg.pinv(D) # D^+ \approx= ((dot(D.T,D))^(-1),D.T)
if L.shape[0] == n:
C = np.dot(pseudo, L).T
else:
C = L
C = np.dot(pseudo, np.dot(D, C.T)).T
Lp = np.dot(D, C.T)
if len(Lp.shape) == 1:
Lp.shape = (n, 1)
if np_matrix_rank(Lp) != Lp.shape[1]:
Lp = fullrank(Lp)
C = np.dot(pseudo, Lp).T
return np.squeeze(C)
def contrastfromcols(L, D, pseudo=None):
"""
From an n x p design matrix D and a matrix L, tries
to determine a p x q contrast matrix C which
determines a contrast of full rank, i.e. the
n x q matrix
dot(transpose(C), pinv(D))
is full rank.
L must satisfy either L.shape[0] == n or L.shape[1] == p.
If L.shape[0] == n, then L is thought of as representing
columns in the column space of D.
If L.shape[1] == p, then L is thought of as what is known
as a contrast matrix. In this case, this function returns an estimable
contrast corresponding to the dot(D, L.T)
Note that this always produces a meaningful contrast, not always
with the intended properties because q is always non-zero unless
L is identically 0. That is, it produces a contrast that spans
the column space of L (after projection onto the column space of D).
Parameters
----------
L : array-like
D : array-like
"""
L = np.asarray(L)
D = np.asarray(D)
n, p = D.shape
if L.shape[0] != n and L.shape[1] != p:
raise ValueError("shape of L and D mismatched")
if pseudo is None:
pseudo = np.linalg.pinv(D) # D^+ \approx= ((dot(D.T,D))^(-1),D.T)
if L.shape[0] == n:
C = np.dot(pseudo, L).T
else:
C = L
C = np.dot(pseudo, np.dot(D, C.T)).T
Lp = np.dot(D, C.T)
if len(Lp.shape) == 1:
Lp.shape = (n, 1)
if np_matrix_rank(Lp) != Lp.shape[1]:
Lp = fullrank(Lp)
C = np.dot(pseudo, Lp).T
return np.squeeze(C)
def isestimable(C, D):
""" True if (Q, P) contrast `C` is estimable for (N, P) design `D`
From an Q x P contrast matrix `C` and an N x P design matrix `D`, checks if
the contrast `C` is estimable by looking at the rank of ``vstack([C,D])``
and verifying it is the same as the rank of `D`.
Parameters
----------
C : (Q, P) array-like
contrast matrix. If `C` has is 1 dimensional assume shape (1, P)
D: (N, P) array-like
design matrix
Returns
-------
tf : bool
True if the contrast `C` is estimable on design `D`
Examples
--------
>>> D = np.array([[1, 1, 1, 0, 0, 0],
... [0, 0, 0, 1, 1, 1],
... [1, 1, 1, 1, 1, 1]]).T
>>> isestimable([1, 0, 0], D)
False
>>> isestimable([1, -1, 0], D)
True
"""
C = np.asarray(C)
D = np.asarray(D)
if C.ndim == 1:
C = C[None, :]
if C.shape[1] != D.shape[1]:
raise ValueError('Contrast should have %d columns' % D.shape[1])
new = np.vstack([C, D])
if np_matrix_rank(new) != np_matrix_rank(D):
return False
return True
def isestimable(C, D):
""" True if (Q, P) contrast `C` is estimable for (N, P) design `D`
From an Q x P contrast matrix `C` and an N x P design matrix `D`, checks if
the contrast `C` is estimable by looking at the rank of ``vstack([C,D])``
and verifying it is the same as the rank of `D`.
Parameters
----------
C : (Q, P) array-like
contrast matrix. If `C` has is 1 dimensional assume shape (1, P)
D: (N, P) array-like
design matrix
Returns
-------
tf : bool
True if the contrast `C` is estimable on design `D`
Examples
--------
>>> D = np.array([[1, 1, 1, 0, 0, 0],
... [0, 0, 0, 1, 1, 1],
... [1, 1, 1, 1, 1, 1]]).T
>>> isestimable([1, 0, 0], D)
False
>>> isestimable([1, -1, 0], D)
True
"""
C = np.asarray(C)
D = np.asarray(D)
if C.ndim == 1:
C = C[None, :]
if C.shape[1] != D.shape[1]:
raise ValueError('Contrast should have %d columns' % D.shape[1])
new = np.vstack([C, D])
if np_matrix_rank(new) != np_matrix_rank(D):
return False
return True
def __init__(self, sys, indep_endog=None, instruments=None):
if len(sys) % 2 != 0:
raise ValueError("sys must be a list of pairs of endogenous and \
exogenous variables. Got length %s" % len(sys))
M = len(sys[1::2])
self._M = M
# The lists are probably a bad idea
self.endog = sys[::2] # these are just list containers
self.exog = sys[1::2]
self._K = [np_matrix_rank(_) for _ in sys[1::2]]
# fullexog = np.column_stack((_ for _ in self.exog))
self.instruments = instruments
# Keep the Y_j's in a container to get IVs
instr_endog = {}
[instr_endog.setdefault(_, []) for _ in iterkeys(indep_endog)]
for eq_key in indep_endog:
for varcol in indep_endog[eq_key]:
instr_endog[eq_key].append(self.exog[eq_key][:, varcol])
# ^ copy needed?
# self._instr_endog = instr_endog
self._indep_endog = indep_endog
_col_map = np.cumsum(np.hstack((0, self._K))) # starting col no.s
# move this check to whiten since we're not going to build a full exog?
for eq_key in indep_endog:
try:
iter(indep_endog[eq_key])
except:
# eq_key = [eq_key]
raise TypeError("The values of the indep_exog dict must be\
iterable. Got type %s for converter %s" % (type(del_col)))
# for del_col in indep_endog[eq_key]:
# fullexog = np.delete(fullexog, _col_map[eq_key]+del_col, 1)
# _col_map[eq_key+1:] -= 1
# Josef's example for deleting reoccuring "rows"
# fullexog = np.unique(fullexog.T.view([('',fullexog.dtype)]*\
# fullexog.shape[0])).view(fullexog.dtype).reshape(\
# fullexog.shape[0],-1)
# From http://article.gmane.org/gmane.comp.python.numeric.general/32276/
# Or Jouni' suggetsion of taking a hash:
# http://www.mail-archive.com/[email protected]/msg04209.html
# not clear to me how this would work though, only if they are the *same*
# elements?
# self.fullexog = fullexog
self.wexog = self.whiten(instr_endog)
def __init__(self, sys, indep_endog=None, instruments=None):
if len(sys) % 2 != 0:
raise ValueError("sys must be a list of pairs of endogenous and \
exogenous variables. Got length %s" % len(sys))
M = len(sys[1::2])
self._M = M
# The lists are probably a bad idea
self.endog = sys[::2] # these are just list containers
self.exog = sys[1::2]
self._K = [np_matrix_rank(_) for _ in sys[1::2]]
# fullexog = np.column_stack((_ for _ in self.exog))
self.instruments = instruments
# Keep the Y_j's in a container to get IVs
instr_endog = {}
[instr_endog.setdefault(_,[]) for _ in iterkeys(indep_endog)]
for eq_key in indep_endog:
for varcol in indep_endog[eq_key]:
instr_endog[eq_key].append(self.exog[eq_key][:,varcol])
# ^ copy needed?
# self._instr_endog = instr_endog
self._indep_endog = indep_endog
_col_map = np.cumsum(np.hstack((0,self._K))) # starting col no.s
# move this check to whiten since we're not going to build a full exog?
for eq_key in indep_endog:
try:
iter(indep_endog[eq_key])
except:
# eq_key = [eq_key]
raise TypeError("The values of the indep_exog dict must be\
iterable. Got type %s for converter %s" % (type(del_col)))
# for del_col in indep_endog[eq_key]:
# fullexog = np.delete(fullexog, _col_map[eq_key]+del_col, 1)
# _col_map[eq_key+1:] -= 1
# Josef's example for deleting reoccuring "rows"
# fullexog = np.unique(fullexog.T.view([('',fullexog.dtype)]*\
# fullexog.shape[0])).view(fullexog.dtype).reshape(\
# fullexog.shape[0],-1)
# From http://article.gmane.org/gmane.comp.python.numeric.general/32276/
# Or Jouni' suggetsion of taking a hash:
# http://www.mail-archive.com/[email protected]/msg04209.html
# not clear to me how this would work though, only if they are the *same*
# elements?
# self.fullexog = fullexog
self.wexog = self.whiten(instr_endog)
def fullrank(X, r=None):
"""
Return a matrix whose column span is the same as X.
If the rank of X is known it can be specified as r -- no check
is made to ensure that this really is the rank of X.
"""
if r is None:
r = np_matrix_rank(X)
V, D, U = L.svd(X, full_matrices=0)
order = np.argsort(D)
order = order[::-1]
value = []
for i in range(r):
value.append(V[:, order[i]])
return np.asarray(np.transpose(value)).astype(np.float64)
def fullrank(X, r=None):
"""
Return a matrix whose column span is the same as X.
If the rank of X is known it can be specified as r -- no check
is made to ensure that this really is the rank of X.
"""
if r is None:
r = np_matrix_rank(X)
V, D, U = L.svd(X, full_matrices=0)
order = np.argsort(D)
order = order[::-1]
value = []
for i in range(r):
value.append(V[:, order[i]])
return np.asarray(np.transpose(value)).astype(np.float64)
def setupClass(cls):
from .results.results_regression import Longley
data = longley.load()
data.exog = add_constant(data.exog, prepend=False)
res1 = OLS(data.endog, data.exog).fit()
res2 = Longley()
res2.wresid = res1.wresid # workaround hack
cls.res1 = res1
cls.res2 = res2
res_qr = OLS(data.endog, data.exog).fit(method="qr")
model_qr = OLS(data.endog, data.exog)
Q, R = np.linalg.qr(data.exog)
model_qr.exog_Q, model_qr.exog_R = Q, R
model_qr.normalized_cov_params = np.linalg.inv(np.dot(R.T, R))
model_qr.rank = np_matrix_rank(R)
res_qr2 = model_qr.fit(method="qr")
cls.res_qr = res_qr
cls.res_qr_manual = res_qr2
def setupClass(cls):
from .results.results_regression import Longley
data = longley.load()
data.exog = add_constant(data.exog, prepend=False)
res1 = OLS(data.endog, data.exog).fit()
res2 = Longley()
res2.wresid = res1.wresid # workaround hack
cls.res1 = res1
cls.res2 = res2
res_qr = OLS(data.endog, data.exog).fit(method="qr")
model_qr = OLS(data.endog, data.exog)
Q, R = np.linalg.qr(data.exog)
model_qr.exog_Q, model_qr.exog_R = Q, R
model_qr.normalized_cov_params = np.linalg.inv(np.dot(R.T, R))
model_qr.rank = np_matrix_rank(R)
res_qr2 = model_qr.fit(method="qr")
cls.res_qr = res_qr
cls.res_qr_manual = res_qr2
def add_indep(x, varnames, dtype=None):
'''
construct array with independent columns
x is either iterable (list, tuple) or instance of ndarray or a subclass of it.
If x is an ndarray, then each column is assumed to represent a variable with
observations in rows.
'''
#TODO: this needs tests for subclasses
if isinstance(x, np.ndarray) and x.ndim == 2:
x = x.T
nvars_orig = len(x)
nobs = len(x[0])
#print('nobs, nvars_orig', nobs, nvars_orig)
if not dtype:
dtype = np.asarray(x[0]).dtype
xout = np.zeros((nobs, nvars_orig), dtype=dtype)
count = 0
rank_old = 0
varnames_new = []
varnames_dropped = []
keepindx = []
for (xi, ni) in zip(x, varnames):
#print(xi.shape, xout.shape)
xout[:, count] = xi
rank_new = np_matrix_rank(xout)
#print(rank_new)
if rank_new > rank_old:
varnames_new.append(ni)
rank_old = rank_new
count += 1
else:
varnames_dropped.append(ni)
return xout[:, :count], varnames_new
def add_indep(x, varnames, dtype=None):
"""
construct array with independent columns
x is either iterable (list, tuple) or instance of ndarray or a subclass of it.
If x is an ndarray, then each column is assumed to represent a variable with
observations in rows.
"""
# TODO: this needs tests for subclasses
if isinstance(x, np.ndarray) and x.ndim == 2:
x = x.T
nvars_orig = len(x)
nobs = len(x[0])
# print('nobs, nvars_orig', nobs, nvars_orig)
if not dtype:
dtype = np.asarray(x[0]).dtype
xout = np.zeros((nobs, nvars_orig), dtype=dtype)
count = 0
rank_old = 0
varnames_new = []
varnames_dropped = []
keepindx = []
for (xi, ni) in zip(x, varnames):
# print(xi.shape, xout.shape)
xout[:, count] = xi
rank_new = np_matrix_rank(xout)
# print(rank_new)
if rank_new > rank_old:
varnames_new.append(ni)
rank_old = rank_new
count += 1
else:
varnames_dropped.append(ni)
return xout[:, :count], varnames_new
def fit(self, q=.5, vcov='robust', kernel='epa', bandwidth='hsheather',
max_iter=1000, p_tol=1e-6, **kwargs):
'''Solve by Iterative Weighted Least Squares
Parameters
----------
q : float
Quantile must be between 0 and 1
vcov : string, method used to calculate the variance-covariance matrix
of the parameters. Default is ``robust``:
- robust : heteroskedasticity robust standard errors (as suggested
in Greene 6th edition)
- iid : iid errors (as in Stata 12)
kernel : string, kernel to use in the kernel density estimation for the
asymptotic covariance matrix:
- epa: Epanechnikov
- cos: Cosine
- gau: Gaussian
- par: Parzene
bandwidth: string, Bandwidth selection method in kernel density
estimation for asymptotic covariance estimate (full
references in QuantReg docstring):
- hsheather: Hall-Sheather (1988)
- bofinger: Bofinger (1975)
- chamberlain: Chamberlain (1994)
'''
if q < 0 or q > 1:
raise Exception('p must be between 0 and 1')
kern_names = ['biw', 'cos', 'epa', 'gau', 'par']
if kernel not in kern_names:
raise Exception("kernel must be one of " + ', '.join(kern_names))
else:
kernel = kernels[kernel]
if bandwidth == 'hsheather':
bandwidth = hall_sheather
elif bandwidth == 'bofinger':
bandwidth = bofinger
elif bandwidth == 'chamberlain':
bandwidth = chamberlain
else:
raise Exception("bandwidth must be in 'hsheather', 'bofinger', 'chamberlain'")
endog = self.endog
exog = self.exog
nobs = self.nobs
exog_rank = np_matrix_rank(self.exog)
self.rank = exog_rank
self.df_model = float(self.rank - self.k_constant)
self.df_resid = self.nobs - self.rank
n_iter = 0
xstar = exog
beta = np.ones(exog_rank)
# TODO: better start, initial beta is used only for convergence check
# Note the following doesn't work yet,
# the iteration loop always starts with OLS as initial beta
# if start_params is not None:
# if len(start_params) != rank:
# raise ValueError('start_params has wrong length')
# beta = start_params
# else:
# # start with OLS
# beta = np.dot(np.linalg.pinv(exog), endog)
diff = 10
cycle = False
history = dict(params = [], mse=[])
while n_iter < max_iter and diff > p_tol and not cycle:
n_iter += 1
beta0 = beta
xtx = np.dot(xstar.T, exog)
xty = np.dot(xstar.T, endog)
beta = np.dot(pinv(xtx), xty)
resid = endog - np.dot(exog, beta)
mask = np.abs(resid) < .000001
resid[mask] = ((resid[mask] >= 0) * 2 - 1) * .000001
resid = np.where(resid < 0, q * resid, (1-q) * resid)
resid = np.abs(resid)
xstar = exog / resid[:, np.newaxis]
diff = np.max(np.abs(beta - beta0))
history['params'].append(beta)
history['mse'].append(np.mean(resid*resid))
if (n_iter >= 300) and (n_iter % 100 == 0):
# check for convergence circle, shouldn't happen
for ii in range(2, 10):
if np.all(beta == history['params'][-ii]):
cycle = True
break
warnings.warn("Convergence cycle detected", ConvergenceWarning)
if n_iter == max_iter:
warnings.warn("Maximum number of iterations (1000) reached.",
IterationLimitWarning)
e = endog - np.dot(exog, beta)
# Greene (2008, p.407) writes that Stata 6 uses this bandwidth:
# h = 0.9 * np.std(e) / (nobs**0.2)
# Instead, we calculate bandwidth as in Stata 12
iqre = stats.scoreatpercentile(e, 75) - stats.scoreatpercentile(e, 25)
h = bandwidth(nobs, q)
h = min(np.std(endog),
iqre / 1.34) * (norm.ppf(q + h) - norm.ppf(q - h))
fhat0 = 1. / (nobs * h) * np.sum(kernel(e / h))
if vcov == 'robust':
d = np.where(e > 0, (q/fhat0)**2, ((1-q)/fhat0)**2)
xtxi = pinv(np.dot(exog.T, exog))
xtdx = np.dot(exog.T * d[np.newaxis, :], exog)
vcov = chain_dot(xtxi, xtdx, xtxi)
elif vcov == 'iid':
vcov = (1. / fhat0)**2 * q * (1 - q) * pinv(np.dot(exog.T, exog))
else:
raise Exception("vcov must be 'robust' or 'iid'")
lfit = QuantRegResults(self, beta, normalized_cov_params=vcov)
lfit.q = q
lfit.iterations = n_iter
lfit.sparsity = 1. / fhat0
lfit.bandwidth = h
lfit.history = history
return RegressionResultsWrapper(lfit)
def fit(self, q=.5, vcov='robust', kernel='epa', bandwidth='hsheather',
max_iter=1000, p_tol=1e-6, **kwargs):
'''Solve by Iterative Weighted Least Squares
Parameters
----------
q : float
Quantile must be between 0 and 1
vcov : string, method used to calculate the variance-covariance matrix
of the parameters. Default is ``robust``:
- robust : heteroskedasticity robust standard errors (as suggested
in Greene 6th edition)
- iid : iid errors (as in Stata 12)
kernel : string, kernel to use in the kernel density estimation for the
asymptotic covariance matrix:
- epa: Epanechnikov
- cos: Cosine
- gau: Gaussian
- par: Parzene
bandwidth: string, Bandwidth selection method in kernel density
estimation for asymptotic covariance estimate (full
references in QuantReg docstring):
- hsheather: Hall-Sheather (1988)
- bofinger: Bofinger (1975)
- chamberlain: Chamberlain (1994)
'''
if q < 0 or q > 1:
raise Exception('p must be between 0 and 1')
kern_names = ['biw', 'cos', 'epa', 'gau', 'par']
if kernel not in kern_names:
raise Exception("kernel must be one of " + ', '.join(kern_names))
else:
kernel = kernels[kernel]
if bandwidth == 'hsheather':
bandwidth = hall_sheather
elif bandwidth == 'bofinger':
bandwidth = bofinger
elif bandwidth == 'chamberlain':
bandwidth = chamberlain
else:
raise Exception("bandwidth must be in 'hsheather', 'bofinger', 'chamberlain'")
endog = self.endog
exog = self.exog
nobs = self.nobs
exog_rank = np_matrix_rank(self.exog)
self.rank = exog_rank
self.df_model = float(self.rank - self.k_constant)
self.df_resid = self.nobs - self.rank
n_iter = 0
xstar = exog
beta = np.ones(exog_rank)
# TODO: better start, initial beta is used only for convergence check
# Note the following doesn't work yet,
# the iteration loop always starts with OLS as initial beta
# if start_params is not None:
# if len(start_params) != rank:
# raise ValueError('start_params has wrong length')
# beta = start_params
# else:
# # start with OLS
# beta = np.dot(np.linalg.pinv(exog), endog)
diff = 10
cycle = False
history = dict(params = [], mse=[])
while n_iter < max_iter and diff > p_tol and not cycle:
n_iter += 1
beta0 = beta
xtx = np.dot(xstar.T, exog)
xty = np.dot(xstar.T, endog)
beta = np.dot(pinv(xtx), xty)
resid = endog - np.dot(exog, beta)
mask = np.abs(resid) < .000001
resid[mask] = ((resid[mask] >= 0) * 2 - 1) * .000001
resid = np.where(resid < 0, q * resid, (1-q) * resid)
resid = np.abs(resid)
xstar = exog / resid[:, np.newaxis]
diff = np.max(np.abs(beta - beta0))
history['params'].append(beta)
history['mse'].append(np.mean(resid*resid))
if (n_iter >= 300) and (n_iter % 100 == 0):
# check for convergence circle, shouldn't happen
for ii in range(2, 10):
if np.all(beta == history['params'][-ii]):
cycle = True
warnings.warn("Convergence cycle detected", ConvergenceWarning)
break
if n_iter == max_iter:
warnings.warn("Maximum number of iterations (" + str(max_iter) +
") reached.", IterationLimitWarning)
e = endog - np.dot(exog, beta)
# Greene (2008, p.407) writes that Stata 6 uses this bandwidth:
# h = 0.9 * np.std(e) / (nobs**0.2)
# Instead, we calculate bandwidth as in Stata 12
iqre = stats.scoreatpercentile(e, 75) - stats.scoreatpercentile(e, 25)
h = bandwidth(nobs, q)
h = min(np.std(endog),
iqre / 1.34) * (norm.ppf(q + h) - norm.ppf(q - h))
fhat0 = 1. / (nobs * h) * np.sum(kernel(e / h))
if vcov == 'robust':
d = np.where(e > 0, (q/fhat0)**2, ((1-q)/fhat0)**2)
xtxi = pinv(np.dot(exog.T, exog))
xtdx = np.dot(exog.T * d[np.newaxis, :], exog)
vcov = chain_dot(xtxi, xtdx, xtxi)
elif vcov == 'iid':
vcov = (1. / fhat0)**2 * q * (1 - q) * pinv(np.dot(exog.T, exog))
else:
raise Exception("vcov must be 'robust' or 'iid'")
lfit = QuantRegResults(self, beta, normalized_cov_params=vcov)
lfit.q = q
lfit.iterations = n_iter
lfit.sparsity = 1. / fhat0
lfit.bandwidth = h
lfit.history = history
return RegressionResultsWrapper(lfit)
def check_rank(self, J):
rank = np_matrix_rank(J)
if rank < np.size(J, axis=1):
raise ValueError("Rank condition not met: "
"solution may not be unique.")