[ADMB Users] phases in state-space model

[Mollie Brooks]

Hi,
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
PARAMETER_SECTION
init_bounded_number logitrho(0,9); //autocorrelation coefficient on the logic scale
init_vector ucoeffs(1,ncoeffs); //coefficients of linear part of model
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

[H. Skaug]

Hi Mollie,

I always estimate all fixed effects paramaters first (using one or more phases).

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.

hans

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.

Thanks!
Mollie

Mollie Brooks, PhD
Postdoctoral Researcher, Population Ecology Research Group http://www.popecol.org
Institute of Evolutionary Biology & Environmental Studies, University of Zürich




_______________________________________________
Users mailing list
Us...@admb-project.org
http://lists.admb-project.org/mailman/listinfo/users

[dave fournier]

The other rule of thumb is that you estimate first the parameters you
know least about.


The rationale for this is more or less that if you already knew the
value of a prameter you
would not estimate it at all -- just fix it at its known value.

Dave