exploring phases with von Bertalanffy growth parameter estimates

[CJ Schlick]

Hello, 

I am working on a SS model (version 3.30.22) that uses multiple K parameters estimates that change based on age (specifically change at age 1 and age 6). I have been exploring the phase estimations around these parameters and came across a scenario that I cannot explain. I was hoping to have some thoughts on what may be occurring here. I can provide some more details if needed, but I am unsure what exactly would be helpful (I'm new to SS). 

I had a model that the only estimated parameters with increased individual gradients (L_at_Amax_Fem_GP_1, VonBert_K_young_Fem_GP_1, and Age_K_mult_Fem_GP_1_a_1) and the maximum gradient was 0.00237. None of the parameters were at bounds, so I thought I would just change the phases. Specifically, I changed the phases so the model estimated K and then L for the von Bertalanffy growth parameters. This did improve the model performance but it did not change the parameter estimates. Only the standard deviation of those estimates. Is this something that others have run into? I appreciate any thoughts on this. 

Here are the parameter estimates for 3 runs for those specific values I was exploring: 

Parameter estimates with L estimated in phase 2 and K estimated in phase 3: 

L_at_Amin_Fem_GP_1 22.4204000 2 6.000 25.00 21.3265000 OK 0.1677110 -0.000391837 No_prior NA NA NA OK
L_at_Amax_Fem_GP_1 108.6030000 2 90.000 130.00 107.4760000 OK 1.1766900 -0.00107904 No_prior NA NA NA OK
VonBert_K_young_Fem_GP_1 0.4312210 3 0.100 0.70 0.4260610 OK 0.0114684 -0.00237066 No_prior NA NA NA OK
Age_K_mult_Fem_GP_1_a_1 0.4969810 3 0.010 1.00 0.5228720 OK 0.0071857 -0.00221227 No_prior NA NA NA OK
Age_K_mult_Fem_GP_1_a_6 0.2260770 3 0.010 1.00 0.2582340 OK 0.0187676 -0.000260195 No_prior NA NA NA OK
CV_young_Fem_GP_1 0.1561940 6 0.050 0.40 0.1862220 OK 0.0023731 0.000278441 No_prior NA NA NA OK
CV_old_Fem_GP_1 0.0241375 6 0.010 0.30 0.0212113 OK 0.0026502 0.000214822 No_prior NA NA NA OK

Parameter estimates with L estimated in phase 3 and K estimated in phase 2: 

L_at_Amin_Fem_GP_1 22.4204000 3 6.000 25.00 21.3265000 OK 0.1677100 -1.6521e-05 No_prior NA NA NA OK
L_at_Amax_Fem_GP_1 108.6030000 3 90.000 130.00 107.4760000 OK 1.1767800 -3.77117e-06 No_prior NA NA NA OK
VonBert_K_young_Fem_GP_1 0.4312210 2 0.100 0.70 0.4260610 OK 0.0114709 -2.95577e-05 No_prior NA NA NA OK
Age_K_mult_Fem_GP_1_a_1 0.4969810 3 0.010 1.00 0.5228720 OK 0.0071868 0.00015103 No_prior NA NA NA OK
Age_K_mult_Fem_GP_1_a_6 0.2260770 3 0.010 1.00 0.2582340 OK 0.0187692 6.29547e-05 No_prior NA NA NA OK
CV_young_Fem_GP_1 0.1561940 6 0.050 0.40 0.1862220 OK 0.0023727 -1.55911e-06 No_prior NA NA NA OK
CV_old_Fem_GP_1 0.0241375 6 0.010 0.30 0.0212113 OK 0.0026498 -0.000220241 No_prior NA NA NA OK

Parameter estimates with L estimated in phase 3, K estimated in phase 2, and age at k estimated in phase 2. This did result with a final maximum gradient of <0.0001. 

L_at_Amin_Fem_GP_1 22.4204000 3 6.000 25.00 21.3265000 OK 0.1676740 1.49883e-05 No_prior NA NA NA OK
L_at_Amax_Fem_GP_1 108.6030000 3 90.000 130.00 107.4760000 OK 1.1766400 2.0769e-05 No_prior NA NA NA OK
VonBert_K_young_Fem_GP_1 0.4312210 2 0.100 0.70 0.4260610 OK 0.0114678 8.71366e-06 No_prior NA NA NA OK
Age_K_mult_Fem_GP_1_a_1 0.4969810 2 0.010 1.00 0.5228720 OK 0.0071856 1.88925e-05 No_prior NA NA NA OK
Age_K_mult_Fem_GP_1_a_6 0.2260770 2 0.010 1.00 0.2582340 OK 0.0187642 -1.62146e-05 No_prior NA NA NA OK
CV_young_Fem_GP_1 0.1561940 6 0.050 0.40 0.1862220 OK 0.0023718 -1.70572e-05 No_prior NA NA NA OK
CV_old_Fem_GP_1 0.0241375 6 0.010 0.30 0.0212113 OK 0.0026482 7.96499e-05 No_prior NA NA NA OK

Thank you for your help. 

[Richard Methot - NOAA Federal]

Hi CJ,

The very small differences you are seeing in the final gradient and parameter variances is well below any level of concern that I would have.  The fact that the parameter values themselves are identical is a good sign.  Phasing shouldn't matter if the starting parameter values are reasonably close to the solution region.  The problem that can occur with phasing is when some parameter has a far off starting value and is not estimated until a later stage.  In this case, the model works on doing as much improvement as it can with the other parameters and perhaps goes into a cul-de-sac that it cannot get out of when the far off parameter starts to get estimated.  This can happen when jitter is used with very wide parameter bounds.

Rick

--
You received this message because you are subscribed to the Google Groups "SS3 - Forum" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ss3-forum+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/ss3-forum/a46285da-efb0-4024-8bd9-64ebda1371d8n%40googlegroups.com.