[ADMB Users] phases in state-space model
hskaug at gmail.com
Tue Jun 18 15:41:07 PDT 2013
I always estimate all fixed effects paramaters first (using one or more
Then I activate both the parameters of the AR-1 state space model in
a single phase. The advantage of this is that plain ADMB (non-ADMB-RE)
is used in the beginning which is really fast.
However, every model has its own features. I recently came across a model
where one of the fixed effects (not a simple regression parameter, though)
was not identifiable without the random effects being active, so
my standard approach failed.
I hope this helps.
On Tue, Jun 18, 2013 at 3:21 PM, Mollie Brooks <mbrooks at ufl.edu> wrote:
> I was wondering if any of you have a rule of thumb for what parameters to
> bring in in what phases of state-space models. I'm thinking about using
> phases to try to speed up and robustify my state-space model.
> My latent state follows an AR(1) process.
> These are the parameters
> init_bounded_number logitrho(0,9); //autocorrelation coefficient
> on the logic scale
> init_vector ucoeffs(1,ncoeffs); //coefficients of linear part of
> init_bounded_number logsigmaSq(-10,2,3);
> objective_function_value jnll;
> random_effects_vector u(1,totobs,1); //latent states
> This is the basic underlying process
> where the Ds are the latent states in random_effects_vector u,
> rho is the autocorrelation coefficient,
> beta is the vector ucoeffs that gets multiplied by the linear predictors
> in X,
> and sigmaSq is the variance of process error.
> Mollie Brooks, PhD
> Postdoctoral Researcher, Population Ecology Research Group
> Institute of Evolutionary Biology & Environmental Studies, University of
> Users mailing list
> Users at admb-project.org
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Users