[Question]: using MCMC when using Fs as parameters (Fmethod=2 or Fmethod=4 and annual Fs estimated for some fleets)

[shanae-allen]

Describe the bug

Hello SS community, When I specify Fmethod = 2 in SS version 3.30.15 or Fmethod=4 (with annual Fs estimated for some fleets) in SS version 3.30.22.1, MCMC does not align with the base model results (chains that converge have higher LL compared to the base model and posterior distributions do not overlap with base model estimates). However, when I use the hybrid F option (Fmethod=3) the MCMC distributions align with the base model results. Is this possibly an issue with MCMC or could it be implying that a model with estimated Fs is unidentifiable? For context, I’m not using the -nuts command, only -mcmc. Thanks! Shanae

To Reproduce

1. Compare log-likelihoods and posterior distributions when using -mcmc to when solving an SS model normally. SS_input_files.xlsx

Expected behavior

I expected the log-likelihoods and posterior distributions to align with the base model, as is the case when Fmethod = 3 or Fmethod = 4 but all phases = 99. SS_input_files_hybrid_F.xlsx

Screenshots

No response

Which OS are you seeing the problem on?

Windows

Which version of SS3 are you seeing the problem on?

3.30.22.1 and 3.30.15

Additional Context

No response

[iantaylor-NOAA]

Hi Shanae, Thanks for reporting this. I haven't seen this issue before. It's hard to know if it's an issue of identifiability of the F parameters or somehow just bad initial values for the Fs leading to a different range of Fs. Using adnuts could overcome initial conditions by more efficient sampling which would allow longer runs with more chains.

Could you share your posteriors.sso and derived_posteriors.sso files for the two runs by attaching to this issue or by emailing them to nmfs.stock.synthesis@noaa.gov.

Note that in general the MCMC results shouldn't be expected to match too precisely to the base model in SS3 models as discussed in Stewart et al. (2013) https://www.sciencedirect.com/science/article/pii/S0165783612002081

[Rick-Methot-NOAA]

I am not surprised to see this difference, especially if the se on the observed catch is very small. With hybrid F, the F's are internally adjusted within each MCMC draw to maintain a nearly exact match to the catch, hence nil logL value. But with F as parameter, the F's themselves are being controlled by the MCMC algorithm, so will typically not match the catch exactly and with a small catch_se, the catch logL will be substantial.

[shanae-allen]

Thank you Rick and Ian for your quick feedback! I will email the SS input files and posterior files for another model (Yellowtail in SS version 3.30.15) for which we are also seeing this issue. The MCMC for the Mutton model I previously shared with Fmethod = 4 is showing convergence issues.
As for the log standard errors for the catch, in both Yellowtail and Mutton models there are a mixture of commercial fleets with very low se's (< .1) and recreational fleets that have average se's around 0.3 - 0.4. Would the higher se's for the rec fleet increase the catch logL as much?