Closed samueldnj closed 7 years ago
ssProd fits fine with an auto-correlation factor fixed, need to do some testing to discover if it works well as a free parameter
Tough for msProd, this extension involves adding a cholesky factor and an autocorrelation factor. The model won't converge, nor will it estimate the cholesky factor entries. There are now (nS+1)*nT random effects, so I'm thinking of moving this to ADMB-RE
ssProd can fit a free AR(1) factor rho.
msProd has too many REs, I think, and has trouble fitting with fixed rho and fixed cholesky factor. I think I need to migrate this to ADMB-RE or TMB, or reduce the number of REs by compressing the AR(1) and correlations: produce correlated REs across the species, then assume the proc error is AR(1) developed from those correlated REs
found the problem for msProd: by splitting the proc error into 2 components, and re-purposing the original prior on proc error variance to be for the single-species component (zetat), I had removed the prior on the multispecies component epst. A quick check by using the first species' prior on the epst variance converged the model.
When going forward and estimating priors instead of keeping them fixed, this is going to take some care. Try defining hyper-priors to keep the model on track (Issue #13 and similar)
Okay, I think I can close this issue. Covariance structures are being estimated for both components of proc error in msProd, and the AR(1) is being estimated in ssProd.
Needed to add a new prior for epst variance, but once this was added the model fit without going crazy.
Add the ability to estimate covariance in both estimators
ssProd ADMB
Use the Cholesky decomposition to model autocorrelation in the single species process error deviations.
msProd ADMB
Same as ssProd, but also estimate multispecies covariance in process error (and in osb error?? flatfish appear in the same survey)
Big question: do I combine both covariance structures into one (somehow, Hadamard product??) and just estimate a single process error deviation for each time step, or do I keep them split?