def covar(self):
"""
Returns
-------
covar : np.ndarray
The covariance matrix for the fitting system
"""
_pvals = np.array(self.varying_parameters())
hess = approx_hess2(_pvals, self.nll)
covar = np.linalg.inv(hess)
self.setp(_pvals)
return covar
def covar(self, target="nll"):
"""
Estimates a covariance matrix based on numerical differentiation
of either the negative log-likelihood or negative log-posterior
probabilities.
Parameters
----------
target : str, {"nll", "nlpost"}
Returns
-------
covar : np.ndarray
The covariance matrix for the fitting system
Notes
-----
Estimation of a covariance matrix can be susceptible to numeric
instabilities. Critically evaluate the matrix before further use.
"""
_pvals = np.array(self.varying_parameters())
if target == "nll":
fn = self.nll
elif target == "nlpost":
fn = self.nlpost
try:
# from statsmodels
# the output from this for the test in test_objective.covar
# is very similar to numdifftools.Hessian, or a chained version
# of approx_derivative
hess = approx_hess2(_pvals, fn)
covar = np.linalg.inv(hess)
except LinAlgError:
sz = np.size(_pvals, 0)
covar = np.full((sz, sz), np.inf)
finally:
self.setp(_pvals)
return covar