In pbs-assess/gfsynopsis: Data Synopsis Reports for PBS Groundfish

Comparing the MSSM survey indices to related indices {#sec:comparing-indices}

The Multispecies Small-Mesh Survey is the longest running continuous groundfish survey on the BC coast, therefore it is important to understand how well the data capture trends in species abundances. The change in survey locations over time (e.g., Figure \@ref(fig:spatial-shift)) means that a modelled index, which accounts for latent spatial effects, may be a better approach than a design-based index. We compared the modelled MSSM survey indices to the MSSM design-based indices, modelled SYN WCVI indices, and the modelled commercial CPUE 3CD indices.

The design-based indices are calculated following a standard depth-stratified mean estimator. For each year, an average catch is calculated for each of two depth strata. These averages are then multiplied by each stratum's area. Confidence intervals are estimated using a bootstrap procedure with 1000 replicates, drawing stratified samples with replacement, from all tows [see Appendix F in @anderson2019synopsis].

For the modelled MSSM indices, SYN WCVI indices, and commercial CPUE 3CD indices, we used the model outputs from @anderson2024synopsis. Briefly, spatiotemporal models were fit to the data from the MSSM and the SYN WCVI using the R [@r2023] package sdmTMB [@anderson2022sdmTMB]. Models were only fit to species that had a proportion of positive sets $\ge$ 5%. In the MSSM data, for six species that did not have catch weights before 2003, because of the family-level identification, the proportion of positive sets was calculated for 2003 to present. However, of these species, only Blackbelly Eelpout met the 5% coverage threshold to be modelled. Additionally, in 2022, 64% of the MSSM trawls (57 fishing events) did not have doorspread values and so for these fishing events we calculated the mean doorspread of 2022 and used this for the missing values.

The spatiotemporal models for both the MSSM and SYN WCVI surveys were fit with constant spatial Gaussian Markov random fields (GMRFs) and spatiotemporal GMRFs. If models did not converge [criteria in @anderson2024synopsis], then the spatial random field was turned off (i.e., set to zero) and only the spatiotemporal random field was included. The spatiotemporal component, however, differed in structure between the surveys: for the MSSM indices, the GMRFs were structured as a random walk to account for the changes in the sampling domain over time. For the SYN WCVI, which has a consistent sampling domain over time, the spatiotemporal component was structured using independent GRMFs with a common variance along with independent annual means. For each survey and species, five possible families were considered: Tweedie, delta-gamma, delta-lognormal, Poisson-link delta-gamma, and Poisson-link delta-lognormal [@thorson2018three; @anderson2024synopsis]. From these models, the model with the lowest marginal AIC (Akaike information criterion) was selected and used in this report. We omitted poorly fitted models: those which did not meet convergence criteria [@anderson2024synopsis], those where the maximum estimated CIs were greater than 10 times the maximum estimated biomass, and those where the mean coefficient of variation was < 1.

To derive area-weighted biomass indices, the models were then used to estimate abundance across standardized sampling grids. For the MSSM, the sampling domain was defined by creating a 3 $\times$ 3 km grid covering the extent of all survey locations in WCVI Shrimp Survey Areas 124 and 125. Any grid cell that intersected with more than one survey location was retained. This time period covered the most consistently sampled locations in recent years. We used a subset of the grid cells that were last sampled between 2009 to 2021 (Figure \@ref(fig:historical-grid)). In 2022, additional stations that were infrequently or not sampled since 1975 were sampled and so these grid cells were not used to estimate a biomass index. For the SYN WCVI, we compared indices predicted on the 2 $\times$ 2 km grid covering the extent of the SYN WCVI sampling domain, but also on the 3 $\times$ 3 km MSSM grid.

We derived standardized commercial catch per unit effort (CPUE) indices from [@anderson2024synopsis] outputs using methods described in [@anderson2019synopsis Appendix D]. Briefly, the standardization uses a generalized linear mixed model (GLMM) with a Tweedie observation likelihood and a log link fitted with the 'glmmTMB' R package [@glmmTMB]. Binned (factor) covariates are included for depth, latitude, and month. An additional factor predictor is included for year. Locality (spatial polygon), a locality-year interaction, and fishing vessel are treated as random intercepts. The coefficients on the factor year effects are considered the standardized index.

Comparison of modelled and design-based indices of abundance

Trends in each species index were similar between the model-based and design-based MSSM indices. However, for most species, the design-based index had more variable mean annual estimates compared with the modelled index (Figure \@ref(fig:mssm-model-design)) and larger confidence intervals around these estimates. Design-based indices for Yellowtail Rockfish, Bocaccio, Canary Rockfish, and Redstripe Rockfish had the greatest deviations from their respective modelled indices. For species with insufficient data for modelling (<5% positive sets) or where models did not meet convergence checks, we have shown the design-based index for reference (Figure \@ref(fig:design-indices)).

(ref:model-design-indices) Comparison of modelled (green line) and design-based (pink open circles) indices from the MSSM survey. Time-series with a solid trend line and shaded ribbon for 95% confidence intervals represent an index that has been standardized with a spatiotemporal model. Pink open circles represent design-based mean estimates of relative biomass and vertical lines around the dots represent 95% bootstrap confidence intervals. Vertical axes are scaled between zero and the maximum upper confidence interval value of the modelled index. Then the two indices have been scaled to have the same geometric mean. The axes are set to encompass the model-based confidence intervals. Years before 2003 are shaded grey to indicate that catch observations in the MSSM survey are considered less reliable than modern data. 'Mean CV' is the mean of the annual coefficients of variation (CVs) for the modelled (M) and design-based (D) indices. 'Mean $+$ve sets' indicates the ratio of the mean number (across the years) of sets that captured the species of interest to the mean number of sets. The modelled index was predicted on 3 $\times$ 3 km grid (Figure \@ref(fig:historical-grid)).

knitr::include_graphics(here::here("report/mssm-tech-report/figure/index-model-design_pjs-mode.png"), dpi = NA)

knitr::include_graphics(here::here("report/mssm-tech-report/figure/index-mssm-design.png"), dpi = NA)

Comparison of MSSM to other survey indices in the same region

We visually compared the modelled MSSM indices with the modelled commerical CPUE 3CD and SYN WCVI indices. This included a comparison of SYN WCVI indices predicted on the same grid used for the MSSM index. We compared a 10-year rolling window correlation between the modelled MSSM vs. CPUE 3CD, the MSSM vs. SYN WCVI, and the MSSM vs. SYN WCVI predicted on the MSSM grid to examine if lower correlations in earlier years reflected improvements in sampling procedures.

Overall, the modelled MSSM indices were more variable, with larger fluctuations over time, relative to the other indices (Figures \@ref(fig:mssm-cpue), \@ref(fig:mssm-syn-wcvi), \@ref(fig:mssm-syn-wcvi-mssm)). The trends in the MSSM index and SYN WCVI were similar across more species when we estimated both indices using the same MSSM grid (Figures \@ref(fig:mssm-syn-wcvi-mssm), \@ref(fig:rolling-window-correlation)) than compared to SYN WCVI indices predicted over the full SYN WCVI survey grid (Figure \@ref(fig:mssm-syn-wcvi)). This is expected given that this better captures matching spatial patterns and covariates, such as depth, that can affect species distributions (Figure \@ref(fig:depth-dist)). Twenty out of the twenty-seven species had a correlation $\ge$ 0.5 between the full time-series.

The average correlation between the MSSM and the SYN WCVI predicted on the MSSM grid increased early in the time-series (Figure \@ref(fig:rolling-window-correlation)). However, the comparison of the MSSM with the SYN WCVI indices coincides with the most important change in 2003 of full species identification and quantification (Table \@ref(tab:change-table)). The low correlations in early years appear to be driven by Dover Sole, Slender Sole, English Sole, Pacific Sanddab, and Darkblotched Rockfish, but it is unclear how this could be driven by sorting changes.

The commercial CPUE 3CD data precedes the 2003 change in sorting procedure on the MSSM. However, the rolling window correlations were more variable and generally lower between the MSSM and the CPUE 3CD indices (Figure \@ref(fig:rolling-window-correlation)). The average correlation between the MSSM and the CPUE 3CD indices was lower in the years prior to 2003 than after, which is driven by low correlations in Bocaccio and Redstripe Rockfish in early years. When the full time-series were compared, only 10/21 of the species had correlations $\ge$ 0.5.

knitr::include_graphics(here::here("report/mssm-tech-report/figure/index-mssm-model-cpue3CD.png"), dpi = NA)

knitr::include_graphics(here::here("report/mssm-tech-report/figure/index-mssm-model-syn-wcvi-model.png"), dpi = NA)

knitr::include_graphics(here::here("report/mssm-tech-report/figure/index-mssm-model-syn-wcvi-model-mssm-grid.png"), dpi = NA)

knitr::include_graphics(here::here("report/mssm-tech-report/figure/index-mssm-syn-wcvi-mssm-grid-zoom-in.png"), dpi = NA)

knitr::include_graphics(here::here("report/mssm-tech-report/figure/depth-ranges-mssm-syn-wcvi.png"), dpi = NA)

knitr::include_graphics(here::here("report/mssm-tech-report/figure/index-correlation.png"), dpi = NA)

\clearpage

pbs-assess/gfsynopsis documentation built on March 26, 2024, 7:30 p.m.