An opportunity is offered to calculate Bayesian posterior distributions
for specified parameters.
Posterior distributions are created by numerically integrating,
in user chosen increments, across multidimensional parameter space,
the posterior probabilities for values of the specified parameter
of interest.
This procedure can be extremely time consuming, but some mitigation
is achieved by giving the analyst the opportunity to choose the
increment size (or equivalently the grid resolution) for numerical
integration.
Nevertheless, although SmartStats © can integrate up to six
(6) additional parameters into the calculation of the posterior
distribution, only one (1) or two (2) dimensional integrations are
recommended.
Smartstats© will guide you through the procedures to generate
the posterior distributions for a chosen parameter, but note that
this procedure requires carefully preparation of the parameter space
prior to executing the calculation of the posterior distributions.
The covariance matrix for the current
model fit can be an excellent guide for choosing the relevant parameter
space, particularly for unconstrained parameter space, in that an
analyst can use information from the most probable point estimates
and the covariance matrix to propose
reasonable domains for the parameter space.
However, the covariance matrix
is naïve about the parameter space, thus user modification
of the parameter space may be required if, for example, the parameter’s
domain is restricted to between zero (0) and one (1) as in a survival
rate.
The need for such a user intervention in this case follows from
SmartStats © initial naïve presumption of a Gaussian (normal)
form for the posterior distribution such that a large enough standard
error would propose values less than zero (<0) or greater than
one (>1) to be within the parameter space for a survival rate.
One dimensional plots of the posterior distribution of the parameter
of interest will be generated by SmartStats ©, however, if
only one other parameter is integrated into the posterior distribution,
a bivariate contour plot of the joint posterior distribution is
offered as well.
The numerical output of the procedure to calculate Bayesian posterior
distributions can be saved in a text file.
Hopefully, sometime in the not too distant future the use of a
Markov Chain Monte Carlo (MCMC) algorithm to calculate posterior
distributions will be implemented in SmartStats ©.
|