search

samueldnj / hierProdAnalysis

A simulation-estimation model for that compares the performance of hierarchical multispecies and single species surplus production stock assessment models.
0 stars 0 forks source link

Models that converge and find good MLEs can't find parameter values for positive Bt in MCMC #10

Closed samueldnj closed 8 years ago

samueldnj commented 8 years ago

Species 1 and 2 won't return all MCMC trials, as the Bt values are too low (pospen >0). Checking over the first row of par values (or turning off the pospen switch), q ~ 2300 when it should be q ~ 1

This is weird, as the model converges to a local min and provides good estimates of all leading pars, including q. Perhaps the Hessian is bad and the proposal distribution crazy.

Replicate the issue using seed value 1001 for zero parameter space points in species 1, or seed = 1 for 268 successful parameter space points (out of 20000 requested at runtime)

Maybe the large number of random effects is doing this. I am thinking about changing to an ADMB-RE model, where the REs are integrated over as a matter of course. ssProdRE and msProdRE are early attempts at this that work on current pin and dat structure - no success fitting yet.

samueldnj commented 8 years ago

May have been fixed when I diagnosed another problem: the pospen multiplier was way to big, causing the likelihood value to blow up too much and warping gradients. I reduced it, and the mcmc output seems to have settled, as well as seeming to have reduced the number of problem seed values.

samueldnj commented 8 years ago

Nope, just found it: -mceval calls require you to nominate a pin and dat if different to the default. I was using whatever was laid down as the default in the directory, which as giving the crazy bad runs.

This doesn't remove the possibility of non-positive values (some SS models still return as few as 40% of the MCMC trials if rejecting the pospen > 0 trials) but fixes it enough for me to close this issue.

Huzzah!