West Coast Guidelines

The SSC has approved use of the geostatistical delta-GLMM for use when estimating abundance indices using data from the West Coast Groundfish Bottom Trawl Survey, and has requested guidelines regarding its use. The following is a living document and I (JimT) encourage edits.

When using the SpatialDeltaGLMM package for estimating abundance indices for this purpose, I recommend:

That the analyst use the latest official release of the software, rather than the development version, i.e., by clicking on the release tab, and then installing the latest release:

# Install and load devtools package
devtools::install_github("nwfsc-assess/geostatistical_delta-GLMM", ref="3.2.0")

The analyst should then reference this release number in any documentation, and should properly cite the appropriate references from the main page (depending upon the usage of ECE options, vessel effects, etc.).

That analysts use as a default the following configuration:

• 1000 polygons in the piecewise constant approximation, i.e., n_x=1000
• Vessel-year effects turned on, and vessel effects turned off, i.e., VesselConfig = c("Vessel"=0, "VesselYear"=1)
• Spatial and spatiotemporal variation turned on for both encounter probability and positive catch rate components, i.e., FieldConfig = c("Omega1"=1, "Epsilon1"=1, "Omega2"=1, "Epsilon2"=1)
• Pass included as a covariate with an estimated effect on both encounter probability and positive catch rate components, i.e., Q_Config = c("Pass"=1)
• Using geometric anisotropy, i.e., Aniso = 1
• A prolonged search for optimal spacing of knots in the piecewise constant approximation, i.e., Kmeans_Config = list( "Locs"=c("Samples","Domain")[1], "nstart"=100, "iter.max"=1e3)

For each model, the analyst should check for model convergence:

• They should check that the final gradient of the marginal likelihood with respect to fixed effects is <0.01 for all fixed effects.
• They should check that the generalized delta-method (calculated using sdreport()) generates positive (not NA) estimates for all fixed, random, and derived values.
• For their own understanding of the model, the analyst should check the output of sdreport() to see whether any random effect has been driven to zero. This is evidenced by the marginal standard deviation hitting the lower bound for spatiotemporal variation (SigmaE1, SigmaE2), spatial variation (SigmaO1, SigmaO2), or vessel-year effects (SigmaVT1, SigmaVT2). It is fine if these hit their lower bound, given a lower bound of <0.01, but the author should be aware that this model fit is essentially "turning off" these effects.

For each model, the analyst should also check for evidence of goodness of fit. It is up to the assessment author to interpret how this affects model selection via AIC (although I, JimT, do not recommend including any model with poor visual fit in a QQ-plot in the set of models considered for model selection). Authors can do this by:

• Looking at the plot entitled "Q-Q plot", where a model that explains the data will generally have data fall on the one-to-one line.
• Looking at the plot entitled "Posterior predictive" for outliers.

That the analyst run the geostatistical delta-GLMM with different settings for the distribution of residual error. Any of these models that have converged (and possibly that pass some goodness-of-fit inspection) could then be selected among using AIC.

• Specifically, I recommend gamma and lognormal models, plus the gamma-ECE and lognormal-ECE models for species where there is either biological evidence of schooling, or poor fit to the QQ plot, i.e., ObsModel = {1,2,11,12}