[R-sig-dyn-mod] FME Package -help
Hi Adriele, I think, it is not possible to send attachments over the help list. Perhaps you can copy paste the code. I am not an expert on Bayesian statistics, but to my knowledge parameter estimation in a Bayesian and a frequentist setting differ especially for parameters in combination with uninformative data, i.e. parameters that cannot be determined from the data. In the Bayesian setting you basically get the parameter distribution back that you set as prior. In the frequentist setting where you maximize the log-likelihood, you find flat directions in parameter space, i.e. no curvature of the log-likelihood -> large variance/covariance. Might this be the reason? Best regards, Daniel
On Mo, 2015-09-28 at 13:32 -0300, Adriele Giaretta Biase wrote:
I am working on my thesis, at the University of S?o Paulo ? Brazil,
I am using the delayed rejection and adaptative Metropolis (included in the
FME Package) and Bootstrapping (using the Nelder-Mead algorithm) to
estimate the best set and parameter distributions for an EDO model and a
very small dataset (n = 27, divided in four groups of n = [4,5,9,9].
Besides, I have estimated the parameters for each individual animal
(minimizing quadratic deviation).
Although the estimate of parameter values are similar between those
methods, the covariance was extremely different. The parameters have
Gaussian distribution. In the larger groups, the Bootstrap and individual
based covariances estimates seems to converge. However, it is not the case
for the Bayesian.
Please, could you help me making sure that I am using the FME package
correctly to estimate the Bayesian covariances. Would those differences be
expected? I include the code for your information.
Thanks in advance,
Adriele.