model-based-indices

Code used for annual model based index production by the AFSC GAP ModSquad in GOA and the Bering Sea. Model-based-indices are produced using the spatiotemporal delta-glmm implemented in VAST, software developed by Jim Thorson

Authors

Primary contact: @Lewis-Barnett-NOAA

Mission

Provide VAST estimates of abundance and their standard error from GAP survey data for stock assessment authors in conjunction with traditional design-based estimators.

• TOR or Terms Of Reference
• Annual stock assessment request form
• Annual ESR request form
• ESP submission tool

GAP Work Plan

• Species requests & alternate model-based runs/settings in from SSMA, ESP, ESR by 15 January
• Run VAST with previous year survey data included as requested by SA author (March - April, code frozen by May)
• Merge new frozen code (index & data retrival code) in git repo
• Upload all VAST output files for hindcast and data (as an .Rdat) used to produce the indices to the appropriate region hindcast folder in the ModSquad google drive folder
• Notify the assessment lead & ModSquad lead (Lewis) that you’ve completed hindcasts (March - April)
• Run hindcast model structure again after survey with updated data (August - Sept., completed by 30 September) and upload results and data (as an .Rdat) used to produce the indices to the appropriate region production folder on google drive
• Merge final index code & data pull code onto repo
• Notify the assessment lead & ModSquad lead (Lewis) that you’ve completed the production estimates

General Production Timeline Estimate

March - April is the time window for any model exploration/iteration, as requested

• December - ModSquad leadership meets to prioritize species + research plans
• January 15 - Deadline for SAFE authors to submit product requests
• late January - ModSquad planning meeting with program leads
• February 15 - SAFE author's revised requests given ModSquad capacity and discussion from REFM meeting
• March 15 - TOR memo finalized and posted on the web
• April 1 - Crab hindcasts completed
• May 1 - Groundfish hindcasts completed (stop all research, decide on which to not recommend for use in base model)
• August 25 - Model-based estimates for EBS crabs completion target date
• September 25 - Model-based estimates for NBS crabs target date
• September 30 - Model-based estimates for groundfishes completion target date, pending bottom trawl data QAQC timeline
• October 15 - Final deadline for completing groundfish model-based estimates

VAST Resources

• VASTGAP/ModSquad folder here
• ModSquad training materials here
• VAST common troubleshooting here
• Identification of valid hauls for Bering Sea VAST indices, from Jason here

Minimum standard of versions of software and key dependencies (i.e., using these versions or later)

• 2024:
  • Rv4.0.2: VAST v3.9.0, FishStatsUtils v2.10.0, cpp VAST_v13_1_0, TMB v1.7.22, Matrix v1.4-0, DHARMa 0.4.5; or,
  • MRAN v4.0.2: VAST v3.9.0, FishStatsUtils v2.10.0, cpp VAST_v13_1_0, TMB v1.7.16, Matrix v1.2-18, DHARMa v0.3.2
• 2023:
  • Rv4.0.2: VAST v3.9.0, FishStatsUtils v2.10.0, cpp VAST_v13_1_0, TMB v1.7.22, Matrix v1.4-0, DHARMa 0.4.5; or,
  • MRAN v4.0.2: VAST v3.9.0, FishStatsUtils v2.10.0, cpp VAST_v13_1_0, TMB v1.7.16, Matrix v1.2-18, DHARMa v0.3.2
• 2022:
  • Rv4.0.2: VAST v3.8.2, FishStatsUtils v2.10.0, cpp VAST_v13_1_0, TMB v1.7.22, Matrix v1.4-0, DHARMa 0.4.5; or,
  • MRAN v4.0.2: VAST v3.9.0, FishStatsUtils v2.10.0, cpp VAST_v13_1_0, TMB v1.7.16, Matrix v1.2-18, DHARMa v0.3.2
• 2021: Rv4.0.2: VAST v3.6.1, FishStatsUtils v2.8.0, cpp VAST_v12_0_0, TMB v1.7.18, Matrix v1.2.18
• 2020: VAST v3.3.0, FishStatsUtils v2.5.0, cpp VAST_v8_2_0

Initial model settings to generate model-based indices

Initial Model Setting	Suggested alternative setting (if needed)
purpose = "index2” in make_settings()	NA
knots = 750 in make_settings()	knots = 500, 1000; 50 for age comps
Poisson-link delta-gamma observation model1: ObsModel = c(2,1) in make_settings()	option 2: Tweedie ObsModel = c(10,2)2 option 3: delta-lognormal ObsModel = c(1,1)
knot_method = ‘grid’ in fit_model()	knot_method = ‘samples’ if necessary to aid convergence or for comparison to a previous model fit
fine_scale = TRUE in make_settings()	fine_scale = FALSE
max_cells = 2000 in make_settings()	max_cells = n_x * 10
Npool = 100 in make_model()	Npool = 20 to 99
bias.correct = TRUE in make_settings()	NA
refine = TRUE in fit_model()	refine = FALSE
spatiotemporal fields: “IID” default settings for FieldConfig in make_settings()	model spatiotemporal components (epsilon) as a first-order autoregressive process “AR1” (required for extremely unbalanced data) or “0” (if necessary to aid convergence)
anisotropy is on (use_anisotropy = TRUE) in make_settings()	anisotropy off (use_anisotropy = FALSE) if necessary to aid convergence
no vessel effects, catchability or density covariates in fit_model()	may include covariates in cases where their incorporation has been previously demonstrated to improve model fit (e.g., a spatially varying response to cold-pool extent when generating abundance indices combining the EBS and NBS); in these cases, covariates will be centered and scaled prior to fitting

Note: if a request to only include fish west of X degrees longitude, just note that there are two components to this request: (1) excluding data from east of X degrees longitude and (2) specifying this as a "strata" boundary in the model settings in the VAST code

R support

• Installation Guide for R 4.0.2, required to run MRAN here

Footnotes

1. Noting that for species with 100% encounters in any year we will use c(2,4) instead of c(2,1), or the equivalent setting for the lognormal
   ↩

2. Tweedie also involves additional changes to RhoConfig and FieldConfig to ensure that there is only a single linear predictor being estimated, as documented elsewhere ↩