def attest_ind(a, b, dim=None):
""" Return the t-test statistics on arrays a and b over the dim axis.
Returns both the t statistic as well as the p-value
"""
# dim = a.ndim - 1 if dim is None else dim
x1, x2 = ma.mean(a, dim), ma.mean(b, dim)
v1, v2 = ma.var(a, dim), ma.var(b, dim)
n1, n2 = (a.shape[dim], b.shape[dim]) if dim is not None else (a.size, b.size)
df = float(n1+n2-2)
svar = ((n1-1)*v1+(n2-1)*v2) / df
t = (x1-x2)/ma.sqrt(svar*(1.0/n1 + 1.0/n2))
if t.ndim == 0:
return (t, statc.betai(0.5*df,0.5,df/(df+t**2)) if t is not ma.masked and df/(df+t**2) <= 1.0 else ma.masked)
else:
prob = [statc.betai(0.5*df,0.5,df/(df+tsq)) if tsq is not ma.masked and df/(df+tsq) <= 1.0 else ma.masked for tsq in t*t]
return t, prob
def aF_oneway(*args, **kwargs):
dim = kwargs.get("dim", None)
arrays = args
means = [ma.mean(a, dim) for a in arrays]
vars = [ma.var(a, dim) for a in arrays]
lens = [ma.sum(ma.array(ma.ones(a.shape), mask=ma.asarray(a).mask), dim) for a in arrays]
alldata = ma.concatenate(arrays, dim if dim is not None else 0)
bign = ma.sum(ma.array(ma.ones(alldata.shape), mask=alldata.mask), dim)
sstot = ma.sum(alldata ** 2, dim) - (ma.sum(alldata, dim) ** 2) / bign
ssbn = ma.sum([(ma.sum(a, dim) ** 2) / L for a, L in zip(arrays, lens)], dim)
# print ma.sum(alldata, dim) ** 2 / bign, ssbn
ssbn -= ma.sum(alldata, dim) ** 2 / bign
sswn = sstot - ssbn
dfbn = dfnum = float(len(args) - 1.0)
dfwn = bign - len(args) # + 1.0
F = (ssbn / dfbn) / (sswn / dfwn)
if F.ndim == 0 and dfwn.ndim == 0:
return (F,statc.betai(0.5 * dfwn, 0.5 * dfnum, dfwn/float(dfwn+dfnum*F)) if F is not ma.masked and dfwn/float(dfwn+dfnum*F) <= 1.0 \
and dfwn/float(dfwn+dfnum*F) >= 0.0 else ma.masked)
else:
prob = [statc.betai(0.5 * dfden, 0.5 * dfnum, dfden/float(dfden+dfnum*f)) if f is not ma.masked and dfden/float(dfden+dfnum*f) <= 1.0 \
and dfden/float(dfden+dfnum*f) >= 0.0 else ma.masked for dfden, f in zip (dfwn, F)]
return F, prob
def __call__(self, data, weight=None):
if not self.use_attributes is None:
new_domain = orange.Domain(self.use_attributes, data.domain.classVar)
new_domain.addmetas(data.domain.getmetas())
data = orange.ExampleTable(new_domain, data)
if self.stepwise and self.stepwise_before:
use_attributes=stepwise(data,add_sig=self.add_sig,remove_sig=self.remove_sig)
new_domain = orange.Domain(use_attributes, data.domain.classVar)
new_domain.addmetas(data.domain.getmetas())
data = orange.ExampleTable(new_domain, data)
# continuization (replaces discrete with continuous attributes)
continuizer = orange.DomainContinuizer()
continuizer.multinomialTreatment = continuizer.FrequentIsBase
continuizer.zeroBased = True
domain0 = continuizer(data)
data = data.translate(domain0)
if self.stepwise and not self.stepwise_before:
use_attributes=stepwise(data,weight,add_sig=self.add_sig,remove_sig=self.remove_sig)
new_domain = orange.Domain(use_attributes, data.domain.classVar)
new_domain.addmetas(data.domain.getmetas())
data = orange.ExampleTable(new_domain, data)
# missing values handling (impute missing)
imputer = orange.ImputerConstructor_model()
imputer.learnerContinuous = orange.MajorityLearner()
imputer.learnerDiscrete = orange.MajorityLearner()
imputer = imputer(data)
data = imputer(data)
# convertion to numpy
A, y, w = data.toNumpy() # weights ??
if A is None:
n = len(data)
m = 0
else:
n, m = numpy.shape(A)
if self.beta0 == True:
if A is None:
X = numpy.ones([len(data),1])
else:
X = numpy.insert(A,0,1,axis=1) # adds a column of ones
else:
X = A
# set weights
W = numpy.identity(len(data))
if weight:
for di, d in enumerate(data):
W[di,di] = float(d[weight])
D = dot(dot(numpy.linalg.pinv(dot(dot(X.T,W),X)), X.T), W) # adds some robustness by computing the pseudo inverse; normal inverse could fail due to singularity of the X.T*W*X
beta = dot(D,y)
yEstimated = dot(X,beta) # estimation
# some desriptive statistisc
muY, sigmaY = numpy.mean(y), numpy.std(y)
muX, covX = numpy.mean(X, axis = 0), numpy.cov(X, rowvar = 0)
# model statistics
SST, SSR = numpy.sum((y - muY) ** 2), numpy.sum((yEstimated - muY) ** 2)
SSE, RSquare = SST-SSR, SSR/SST
R = numpy.sqrt(RSquare) # coefficient of determination
RAdjusted = 1 - (1 - RSquare) * (n - 1) / (n - m - 1)
F = (SSR / m) / (SST - SSR / (n - m - 1)) # F statistisc
df = m - 1
sigmaSquare = SSE / (n-m-1)
# standard error of estimated coefficients
errCoeff = sqrt(sigmaSquare * inv(dot(X.T,X)).diagonal())
# t statistisc, significance
t = beta / errCoeff
df = n-2
significance = []
for tt in t:
try:
significance.append(statc.betai(df*0.5,0.5,df/(df+tt*tt)))
except:
significance.append(1.0)
# standardized coefficients
if m>0:
stdCoeff = (sqrt(covX.diagonal()) / sigmaY) * beta
else:
stdCoeff = (sqrt(covX) / sigmaY) * beta
model = {'descriptives': { 'meanX' : muX, 'covX' : covX, 'meanY' : muY, 'sigmaY' : sigmaY},
'model' : {'estCoeff' : beta, 'stdErrorEstimation': errCoeff},
'model summary': {'TotalVar' : SST, 'ExplVar' : SSE, 'ResVar' : SSR, 'R' : R, 'RAdjusted' : RAdjusted,
'F' : F, 't' : t, 'sig': significance}}
return LinearRegression(statistics = model, domain = data.domain, name = self.name, beta0 = self.beta0, imputer=imputer)
def __call__(self, data, weight=None):
if not self.use_attributes is None:
new_domain = orange.Domain(self.use_attributes,
data.domain.classVar)
new_domain.addmetas(data.domain.getmetas())
data = orange.ExampleTable(new_domain, data)
if self.stepwise and self.stepwise_before:
use_attributes = stepwise(data,
add_sig=self.add_sig,
remove_sig=self.remove_sig)
new_domain = orange.Domain(use_attributes, data.domain.classVar)
new_domain.addmetas(data.domain.getmetas())
data = orange.ExampleTable(new_domain, data)
# continuization (replaces discrete with continuous attributes)
continuizer = orange.DomainContinuizer()
continuizer.multinomialTreatment = continuizer.FrequentIsBase
continuizer.zeroBased = True
domain0 = continuizer(data)
data = data.translate(domain0)
if self.stepwise and not self.stepwise_before:
use_attributes = stepwise(data,
weight,
add_sig=self.add_sig,
remove_sig=self.remove_sig)
new_domain = orange.Domain(use_attributes, data.domain.classVar)
new_domain.addmetas(data.domain.getmetas())
data = orange.ExampleTable(new_domain, data)
# missing values handling (impute missing)
imputer = orange.ImputerConstructor_model()
imputer.learnerContinuous = orange.MajorityLearner()
imputer.learnerDiscrete = orange.MajorityLearner()
imputer = imputer(data)
data = imputer(data)
# convertion to numpy
A, y, w = data.toNumpy() # weights ??
if A is None:
n = len(data)
m = 0
else:
n, m = numpy.shape(A)
if self.beta0 == True:
if A is None:
X = numpy.ones([len(data), 1])
else:
X = numpy.insert(A, 0, 1, axis=1) # adds a column of ones
else:
X = A
# set weights
W = numpy.identity(len(data))
if weight:
for di, d in enumerate(data):
W[di, di] = float(d[weight])
D = dot(
dot(numpy.linalg.pinv(dot(dot(X.T, W), X)), X.T), W
) # adds some robustness by computing the pseudo inverse; normal inverse could fail due to singularity of the X.T*W*X
beta = dot(D, y)
yEstimated = dot(X, beta) # estimation
# some desriptive statistisc
muY, sigmaY = numpy.mean(y), numpy.std(y)
muX, covX = numpy.mean(X, axis=0), numpy.cov(X, rowvar=0)
# model statistics
SST, SSR = numpy.sum((y - muY)**2), numpy.sum((yEstimated - muY)**2)
SSE, RSquare = SST - SSR, SSR / SST
R = numpy.sqrt(RSquare) # coefficient of determination
RAdjusted = 1 - (1 - RSquare) * (n - 1) / (n - m - 1)
F = (SSR / m) / (SST - SSR / (n - m - 1)) # F statistisc
df = m - 1
sigmaSquare = SSE / (n - m - 1)
# standard error of estimated coefficients
errCoeff = sqrt(sigmaSquare * inv(dot(X.T, X)).diagonal())
# t statistisc, significance
t = beta / errCoeff
df = n - 2
significance = []
for tt in t:
try:
significance.append(
statc.betai(df * 0.5, 0.5, df / (df + tt * tt)))
except:
significance.append(1.0)
# standardized coefficients
if m > 0:
stdCoeff = (sqrt(covX.diagonal()) / sigmaY) * beta
else:
stdCoeff = (sqrt(covX) / sigmaY) * beta
model = {
'descriptives': {
'meanX': muX,
'covX': covX,
'meanY': muY,
'sigmaY': sigmaY
},
'model': {
'estCoeff': beta,
'stdErrorEstimation': errCoeff
},
'model summary': {
'TotalVar': SST,
'ExplVar': SSE,
'ResVar': SSR,
'R': R,
'RAdjusted': RAdjusted,
'F': F,
't': t,
'sig': significance
}
}
return LinearRegression(statistics=model,
domain=data.domain,
name=self.name,
beta0=self.beta0,
imputer=imputer)