def _altModelMixedEffectNonLinear1Kernel(self, G1, approx, link, penalty):
if self.forcefullrank:
assert False, 'Not working, call Danilo'
assert False, "Not implemented yet."
else:
glmm1 = inference.getGLMM(approx, link, self.Y, G1, None, penalty=penalty)
glmm1.setX(self.X)
glmm1.sety(self.Y)
glmm1.setG(G1)
glmm1.optimize()
assert glmm1.sig02 >= 0.0 and glmm1.sign2 >= 0
if glmm1.sig02 + glmm1.sign2 <= NP.sqrt(NP.finfo(NP.float).eps):
h2 = NP.nan
else:
h2 = glmm1.sig02 / (glmm1.sig02 + glmm1.sign2)
a2 = NP.nan
lik1 = {'nLL':-glmm1.marginal_loglikelihood(),
'h2':h2,
'a2':a2}
lik1['sig02'] = glmm1.sig02
lik1['sig12'] = glmm1.sig12
lik1['sign2'] = glmm1.sign2
for i in range(len(glmm1.beta)):
lik1['beta' + str(i)] = glmm1.beta[i]
stat = 2.0*(self.model0['nLL'] - lik1['nLL'])
return (lik1,stat)
def _nullModelMixedEffectNonLinear2Kernel(self, G0, approx, link, penalty):
if self.forcefullrank:
assert False, "Not implemented yet."
else:
glmm0 = inference.getGLMM(approx,
link,
self.Y,
G0,
None,
penalty=penalty)
glmm0.setX(self.X)
glmm0.setG(G0)
glmm0.sety(self.Y)
glmm0.optimize()
self.model0 = {}
if glmm0.sig02 + glmm0.sign2 <= NP.sqrt(NP.finfo(NP.float).eps):
h2 = NP.nan
else:
h2 = glmm0.sig02 / (glmm0.sig02 + glmm0.sign2)
self.model0['h2'] = h2
self.model0['a2'] = 0.0
self.model0['nLL'] = -glmm0.marginal_loglikelihood()
self.model0['sig02'] = glmm0.sig02
self.model0['sig12'] = glmm0.sig12
self.model0['sign2'] = glmm0.sign2
for i in range(len(glmm0.beta)):
self.model0['beta' + str(i)] = glmm0.beta[i]
def _nullModelMixedEffectNonLinear1Kernel(self, approx, link, penalty):
if self.forcefullrank:
assert False, "Not implemented yet."
else:
glmm0 = inference.getGLMM(approx, link, self.Y, None, None, penalty=penalty)
glmm0.setX(self.X)
glmm0.sety(self.Y)
glmm0.optimize()
self.model0 = {}
self.model0['h2']=0.0
self.model0['a2']=NP.nan
self.model0['nLL']=-glmm0.marginal_loglikelihood()
self.model0['sig02'] = glmm0.sig02
self.model0['sig12'] = glmm0.sig12
self.model0['sign2'] = glmm0.sign2
for i in range(len(glmm0.beta)):
self.model0['beta' + str(i)] = glmm0.beta[i]
def _altModelMixedEffectNonLinear2Kernel(self, G1, approx, link, penalty):
if self.forcefullrank:
assert False, "Not implemented yet."
else:
glmm1 = inference.getGLMM(approx,
link,
self.Y,
self.G0,
G1,
penalty=penalty)
glmm1.setX(self.X)
glmm1.sety(self.Y)
glmm1.setG(self.G0, G1)
glmm1.optimize()
assert glmm1.sig02 >= 0.0 and glmm1.sig12 >= 0.0 and glmm1.sign2 >= 0
if glmm1.sig02 + glmm1.sig12 + glmm1.sign2 <= NP.sqrt(
NP.finfo(NP.float).eps):
# in this case we don't have enough precision to calculate the
# proportion between sig02+sig12 and the total or it does not make sense
# because the covariance of the posterior tends to zero
h2 = NP.nan
else:
h2 = (glmm1.sig02 + glmm1.sig12) / (glmm1.sig02 + glmm1.sig12 +
glmm1.sign2)
if glmm1.sig02 + glmm1.sig12 <= NP.sqrt(NP.finfo(NP.float).eps):
a2 = NP.nan
else:
a2 = glmm1.sig12 / (glmm1.sig02 + glmm1.sig12)
lik1 = {'nLL': -glmm1.marginal_loglikelihood(), 'h2': h2, 'a2': a2}
lik1['sig02'] = glmm1.sig02
lik1['sig12'] = glmm1.sig12
lik1['sign2'] = glmm1.sign2
for i in range(len(glmm1.beta)):
lik1['beta' + str(i)] = glmm1.beta[i]
stat = 2.0 * (self.model0['nLL'] - lik1['nLL'])
return (lik1, stat)
def _altModelMixedEffectNonLinear2Kernel(self, G1, approx, link, penalty):
if self.forcefullrank:
assert False, "Not implemented yet."
else:
glmm1 = inference.getGLMM(approx, link, self.Y, self.G0,
G1, penalty=penalty)
glmm1.setX(self.X)
glmm1.sety(self.Y)
glmm1.setG(self.G0, G1)
glmm1.optimize()
assert glmm1.sig02 >= 0.0 and glmm1.sig12 >= 0.0 and glmm1.sign2 >= 0
if glmm1.sig02 + glmm1.sig12 + glmm1.sign2 <= NP.sqrt(NP.finfo(NP.float).eps):
# in this case we don't have enough precision to calculate the
# proportion between sig02+sig12 and the total or it does not make sense
# because the covariance of the posterior tends to zero
h2 = NP.nan
else:
h2 = (glmm1.sig02+glmm1.sig12) / (glmm1.sig02 + glmm1.sig12 + glmm1.sign2)
if glmm1.sig02 + glmm1.sig12 <= NP.sqrt(NP.finfo(NP.float).eps):
a2 = NP.nan
else:
a2 = glmm1.sig12 / (glmm1.sig02+glmm1.sig12)
lik1 = {'nLL':-glmm1.marginal_loglikelihood(),
'h2':h2,
'a2':a2}
lik1['sig02'] = glmm1.sig02
lik1['sig12'] = glmm1.sig12
lik1['sign2'] = glmm1.sign2
for i in range(len(glmm1.beta)):
lik1['beta' + str(i)] = glmm1.beta[i]
stat = 2.0*(self.model0['nLL'] - lik1['nLL'])
return (lik1,stat)
