Best practices for modeling coastal pelagics in SS3?

[Elizabeth Gugliotti - NOAA Affiliate]

Below is an email thread from a user to the SS3 team and response from Ian Taylor that might be useful to other users. If you have thoughts/experience/etc. with coastal pelagics, please feel free to comment on this topic!

Initial email from Joel Rice

Is there a 'best practices' for coastal pelagics? In the Pacific we have changes (decreases) over the last few years, along with changes in maturity at age. We have some survey data, and some commercial CPUE as well, mostly trending down in recent years. Initial analysis with SAM shows that the trajectories of spawners per recruit without fishing has dropped to the lowest level in 50 years. 

A couple of questions, 

I remember reading/hearing that the use of length data and length-selectivity in combination with use of empirical weight-at-age file is not advised. Is that still the general thinking?

Is the inclusion of time varying growth a recommended step when we have wtatage.ss? 

We have a recruitment index, is it better to set this to type 33, or just use type 3, and set the age comp selectivity to 0?

There is also an egg survey, that the group has argued that is reflective of SSB. Should that be option 30? What does that actually do? Does it track the maturity at age * numbers at age?

Response from Ian Taylor

Regarding weight-at-age: The empirical weight-at-age option in SS3 was created to bypass the results of parametric growth. In hindsight, it might have been better to remove the growth parameter controls when this option is chosen. If you use length composition data in combination with the wtatage you would just end up with an internally inconsistent model where the length comps would inform the growth parameters, but they would not match the weight-at-age matrix which is used for much of the dynamics. You would be able to estimate length-base selectivity, but the average weight of the selected fish would not match the weight that's being applied from the wtatage file.

I've been thinking of creating a flowchart to describe the decision points around weight-at-age vs parametric growth. But here's a simplistic version of my thinking:

1. If you have complex time-varying growth AND lots of age data relative to the number of ages in the population, then empirical weight-at-age is a good choice.
2. If you have fewer ages, OR your species is longer lived so there are more gaps in the age data, OR you don't have evidence (or enough data to estimate) time-varying growth, parametric growth that is static over time is better.
3. If you have a moderate amount of age data and see evidence of time-varying growth, but have gaps in time or age dimensions of the weight-at-age information, you can consider modeling time-varying parametric growth, which could vary by year or cohort, but you're unlikely to be able to estimate both types of variation.

Models with empirical-weight-at-age are simpler, run faster, and more amenable to using MCMC. One drawbacks of the wtatage option is that you are throwing away length data which often spans a longer period of time than the age comps (or applying an age-length-key to use the length data in a less ideal way than integrating it into the model). Another is that you have to make more assumptions about what weight-at-age is applied prior to your first years with data and what weight-at-age is applied in the future. That's also a problem for models with parametric time-varying growth, where in one case you're typically averaging over years of your wtatage matrix and in the other projecting or averaging over time-varying parameter values.

Regarding indices of SSB or recruitment: the choice for a recruitment index depends on the assumptions around stock-recruit relationships. If you think the index is of the absolute number of recruits, then I think using survey type 33 is probably equivalent to a fleet that selects only age-0 fish (you could check to confirm). If you think the index is about positive or negative recruitment events, independent of spawning output, then one of the survey types that is focused on recdevs (31, 32, or 36). For an egg survey, option 30 is indeed the best choice. It's basically just using the SSB * catchability to calculate the expected index value compare to each observation, which differs from the calculation of vulnerable biomass in that it is applying maturity curve instead of selectivity AND only based on females (in a 2-sex model).