Fit for the future? Environmental covariates and random effects in stock assessment

class: right, middle, my-title, title-slide

.title[
# Fit for the future? Environmental covariates and random effects in stock assessment
]
.subtitle[
## ICES ASC Session M: Controversial opinions in stock assessment and fisheries management 17 September 2025
]
.author[
### Sarah Gaichas, Micah Dean, Jon Deroba
]

---

class: top, left

background-image: url("https://github.com/sgaichas/HSpresentations/raw/main/docs/images/AtlanticMapGOMandKlaipedacrop.png")
background-size: 1070px
background-position: bottom

# Environmental covariates and random effects

.pull-left-60[
## Controversial opinion:

Random effects often improve model fit and solve model diagnostic problems, but may mask signals that could be explained by and projected with mechanistic processes.

Fitting the assessment model isn't our ultimate goal!  
Catch advice for future years is.

]

.pull-right-40[
## Atlantic herring, *Clupea harengus*

<img src="20250917_ICESASC_ecovRE_Gaichas_files/figure-html/herring-1.png" alt="Atlantic herring illustration, credit NOAA Fisheries, from https://www.fisheries.noaa.gov/s3/styles/original/s3/2022-08/640x427-Herring-Atlantic-NOAAFisheries.png" width="356" />
]

???
Both random effects and environmental covariates can help explain modeled population dynamics, but they may interact in unexpected ways.

RE represent cumulative unknown drivers, which can be difficult to project into the future without understanding mechanisms, while environmental covariates represent possibly predictable but incomplete mechanistic drivers, whose influence may change over time.

RE solve a lot of diagnostic issues in models (e.g. retrospective patterns) and are considerably easier to implement than environmental covariates, which often require investment in research and development.

Considerable effort has been invested in evaluating potential environmental drivers of stock productivity, mortality, and/or availability in recent Northwest Atlantic stock assessments. Additional recent investment in regional ocean modeling is aimed at operational short term prediction of environmental conditions such as bottom temperature. Environmental drivers have been successfully included in stock assessments using the Woods Hole Assessment Model (WHAM), a state-space modeling framework. Here we review lessons learned and controversial opinions generated from attempts to include environmental recruitment covariates within a stock assessment for Atlantic herring. Environmental covariates were tailored to herring life history and implemented based on mechanistic hypotheses. Despite significant correlations with recruitment estimated from a prior (non-state-space) assessment model, none of the covariates were included in the final state-space model. We draw two main conclusions. First, even in a state of the art model there are limits on environmental covariate inclusion. For example, mechanistic linkages available for recruitment covariates dwindle if a stock recruit relationship cannot be estimated. Second, tradeoffs may exist between including mechanistic drivers as covariates and including random effects that account for variation from unidentified mechanisms. Herring recruitment covariates had much stronger explanatory power in the absence of numbers at age (NAA) random effects, but model fit was much better with NAA random effects. Given that some environmental covariates can be forecast, while some random effects cannot or it is not clear how, this presents a dilemma. Do we risk sacrificing short term prediction capability (needed by management) for assessment model fit (needed to get models accepted for use in management)?

---
# What drives recruitment of Atlantic herring?

.pull-left[
![Richardson paper](https://github.com/sgaichas/HSpresentations/raw/main/docs/images/Micah_RichardsonPaper.png)
![:img Haddock head, 50%](https://github.com/sgaichas/HSpresentations/raw/main/docs/images/Micah_HaddockHead.png)![:img Herring eggs, 50%](https://github.com/sgaichas/HSpresentations/raw/main/docs/images/Micah_HerringEggs.png)

]

.pull-right[
![Atlantic herring R/SSB and haddock predation index](https://github.com/sgaichas/HSpresentations/raw/main/docs/images/Micah_HerringRSSBhaddock.png)
Haddock egg predation looked promising based on the previous assessment and extensive data analysis

]

???
![Atlantic herring spawning areas in the Gulf of Maine, Northweast Atlantic](https://github.com/sgaichas/HSpresentations/raw/main/docs/images/Micah_HerringSpawningMap.png)

---
# New improved herring assessment, no covariate impact

.pull-left[
State space assessment has much better diagnostics and stability in the [WHAM assessment framework](https://timjmiller.github.io/wham/) designed to incorporate environmental covariates

Yet the haddock predation index had negligible impact on recruitment

]

.pull-right[
Why?

Limited ways to implement covariates: no stock recruit relationship

Best fit with numbers at age transition random effects (NAA RE): correlations among ages, not years

![NAA RE implemented](https://github.com/sgaichas/whamtest-covariates/raw/main/WHAMfits/mm205-test/plots_png/results/NAA_dev_tile.png)
]

???
New state space model developed with NAA RE, best fit with numbers at age transition correlations among ages but not years

Image of NAA RE pattern

![ecov fit with NAA RE](https://github.com/sgaichas/whamtest-covariates/raw/main/WHAMfits/mm205-test/plots_png/diagnostics/NAA_4panel_stock_1_region_1_age_1.png)

Good model diagnostics

Including the covariate in the model: only one approach possible because stock recruit relationship not estimable

Model estimates recruitment as annual iid RE around an estimated fixed effect recruitment scaling parameter ("mean-ish" recruitment)

Model estimates a fit to covariate data, then estimates parameters of relationship with recruitment scaling parameter (lag-1-linear)

Image of fit to haddock predation index covariate

---
# Strong covariate effect, fit poorer without NAA RE

.pull-left[
An experiment fitting the model without NAA RE on ages 2+ and with the environmental covariate resulted in poorer model fit but a much stronger covariate effect; and a different recruitment signal

]

.pull-right[

<div class="tabwid"><style>.cl-fe9a096a{}.cl-fe970238{font-family:'Helvetica';font-size:11pt;font-weight:normal;font-style:normal;text-decoration:none;color:rgba(0, 0, 0, 1.00);background-color:transparent;}.cl-fe984c2e{margin:0;text-align:left;border-bottom: 0 solid rgba(0, 0, 0, 1.00);border-top: 0 solid rgba(0, 0, 0, 1.00);border-left: 0 solid rgba(0, 0, 0, 1.00);border-right: 0 solid rgba(0, 0, 0, 1.00);padding-bottom:5pt;padding-top:5pt;padding-left:5pt;padding-right:5pt;line-height: 1;background-color:transparent;}.cl-fe984c2f{margin:0;text-align:right;border-bottom: 0 solid rgba(0, 0, 0, 1.00);border-top: 0 solid rgba(0, 0, 0, 1.00);border-left: 0 solid rgba(0, 0, 0, 1.00);border-right: 0 solid rgba(0, 0, 0, 1.00);padding-bottom:5pt;padding-top:5pt;padding-left:5pt;padding-right:5pt;line-height: 1;background-color:transparent;}.cl-fe985958{width:0.75in;background-color:transparent;vertical-align: middle;border-bottom: 1.5pt solid rgba(102, 102, 102, 1.00);border-top: 1.5pt solid rgba(102, 102, 102, 1.00);border-left: 0 solid rgba(0, 0, 0, 1.00);border-right: 0 solid rgba(0, 0, 0, 1.00);margin-bottom:0;margin-top:0;margin-left:0;margin-right:0;}.cl-fe985962{width:0.75in;background-color:transparent;vertical-align: middle;border-bottom: 1.5pt solid rgba(102, 102, 102, 1.00);border-top: 1.5pt solid rgba(102, 102, 102, 1.00);border-left: 0 solid rgba(0, 0, 0, 1.00);border-right: 0 solid rgba(0, 0, 0, 1.00);margin-bottom:0;margin-top:0;margin-left:0;margin-right:0;}.cl-fe985963{width:0.75in;background-color:transparent;vertical-align: middle;border-bottom: 0 solid rgba(0, 0, 0, 1.00);border-top: 0 solid rgba(0, 0, 0, 1.00);border-left: 0 solid rgba(0, 0, 0, 1.00);border-right: 0 solid rgba(0, 0, 0, 1.00);margin-bottom:0;margin-top:0;margin-left:0;margin-right:0;}.cl-fe985964{width:0.75in;background-color:transparent;vertical-align: middle;border-bottom: 0 solid rgba(0, 0, 0, 1.00);border-top: 0 solid rgba(0, 0, 0, 1.00);border-left: 0 solid rgba(0, 0, 0, 1.00);border-right: 0 solid rgba(0, 0, 0, 1.00);margin-bottom:0;margin-top:0;margin-left:0;margin-right:0;}.cl-fe985965{width:0.75in;background-color:transparent;vertical-align: middle;border-bottom: 1.5pt solid rgba(102, 102, 102, 1.00);border-top: 0 solid rgba(0, 0, 0, 1.00);border-left: 0 solid rgba(0, 0, 0, 1.00);border-right: 0 solid rgba(0, 0, 0, 1.00);margin-bottom:0;margin-top:0;margin-left:0;margin-right:0;}.cl-fe98596c{width:0.75in;background-color:transparent;vertical-align: middle;border-bottom: 1.5pt solid rgba(102, 102, 102, 1.00);border-top: 0 solid rgba(0, 0, 0, 1.00);border-left: 0 solid rgba(0, 0, 0, 1.00);border-right: 0 solid rgba(0, 0, 0, 1.00);margin-bottom:0;margin-top:0;margin-left:0;margin-right:0;}</style><table data-quarto-disable-processing='true' class='cl-fe9a096a'><thead><tr style="overflow-wrap:break-word;"><th class="cl-fe985958">Model</th><th class="cl-fe985962">conv</th><th class="cl-fe985962">pdHess</th><th class="cl-fe985962">NLL</th><th class="cl-fe985962">dAIC</th><th class="cl-fe985962">AIC</th></tr></thead><tbody><tr style="overflow-wrap:break-word;"><td class="cl-fe985963">NAAon_ecovoff</td><td class="cl-fe985964">TRUE</td><td class="cl-fe985964">TRUE</td><td class="cl-fe985964">-1,757.763</td><td class="cl-fe985964">0.0</td><td class="cl-fe985964">-3,255.5</td></tr><tr style="overflow-wrap:break-word;"><td class="cl-fe985963">NAAon_ecovon</td><td class="cl-fe985964">TRUE</td><td class="cl-fe985964">TRUE</td><td class="cl-fe985964">-1,758.735</td><td class="cl-fe985964">0.0</td><td class="cl-fe985964">-3,255.5</td></tr><tr style="overflow-wrap:break-word;"><td class="cl-fe985963">NAAoff_ecovon</td><td class="cl-fe985964">TRUE</td><td class="cl-fe985964">TRUE</td><td class="cl-fe985964">-1,688.659</td><td class="cl-fe985964">136.2</td><td class="cl-fe985964">-3,119.3</td></tr><tr style="overflow-wrap:break-word;"><td class="cl-fe985965">NAAoff_ecovoff</td><td class="cl-fe98596c">TRUE</td><td class="cl-fe98596c">TRUE</td><td class="cl-fe98596c">-1,679.479</td><td class="cl-fe98596c">152.5</td><td class="cl-fe98596c">-3,103.0</td></tr></tbody></table></div>

![NAA RE off](https://github.com/sgaichas/whamtest-covariates/raw/main/WHAMfits/mm207-ecovon/plots_png/results/NAA_dev_tile.png)

]

???
![ecov fit without NAA RE](https://github.com/sgaichas/whamtest-covariates/raw/main/WHAMfits/mm207-ecovon/plots_png/diagnostics/NAA_4panel_stock_1_region_1_age_1.png)

And a different recruitment time series? NAA RE on ages soaking up some uncertainty that could be explained mechanistically

Image comparing recruitment deviations with and without NAA RE and impact of haddock predation index

Need methods to address multiple covariates: larval optimal temperature duration similar results to haddock predation

---
#  Our perception of the stock changes

.pull-left[
Biomass estimates are similar between models

But recruitment is different, and we don't know which is right because we lack a recruitment index.

Further, including recruitment covariates alters reference point calculations.
]

.pull-right[

![modcompare all](https://github.com/sgaichas/whamtest-covariates/raw/main/compare_png/compare_rel_status_kobe.png)

Different assumptions affect reference point calculations

]

???
We had several potential drivers so there could be more than one. But which mechanism to include?

We got a similar effect with a different environmental covariate--optimal larval temperature. Try to include both??

![:img with NAA no ecov, 50%](https://github.com/sgaichas/whamtest-covariates/raw/main/WHAMfits/mm204-ecovoff/plots_png/results/SSB_Rec_time_stock_1.png)![:img no NAA no ecov, 50%](https://github.com/sgaichas/whamtest-covariates/raw/main/WHAMfits/mm206-ecovoff/plots_png/results/SSB_Rec_time_stock_1.png)
![:img with NAA and ecov, 50%](https://github.com/sgaichas/whamtest-covariates/raw/main/WHAMfits/mm205-test/plots_png/results/SSB_Rec_time_stock_1.png)![:img no NAA and ecov, 50%](https://github.com/sgaichas/whamtest-covariates/raw/main/WHAMfits/mm207-ecovon/plots_png/results/SSB_Rec_time_stock_1.png)

---
.pull-left[
# Discussion

We need more investigation of NAA RE and interactions with covariates, reference point calculations, and projections.

These NAA RE can't be projected, while some recruitment covariates could. NAA RE were retained in the model but not covariates due to fit criteria combined with parsimony.

Do we risk sacrificing short term prediction capability (needed by management) for assessment model fit (needed to get models accepted for use in management)?

Past performance does not guarantee future returns!
]

.pull-right[
<img src="20250917_ICESASC_ecovRE_Gaichas_files/figure-html/unnamed-chunk-3-1.png" width="504" />

# Thoughts? Thank you!
]

.footnote[
Slides available at https://sgaichas.github.io/HSpresentations 
Contact: <sgaichas@hydrascientificllc.com>
]

???
A lot of effort and review goes into identifying the model that best fits historical data, then we do projections separately with a single best fit model, but projections are what's used to set catch advice in fishery management.