An integrated model of seasonal changes in stock composition and abundance with an application to Chinook salmon

Study system

Chinook salmon are harvested by American and Canadian commercial, recreational, and First Nations fisheries, prior to and during return migrations to freshwater spawning habitats. In Canada, management decisions are often applied at the scale of DFO’s Pacific Fishery Management Areas (PFMAs) (DFO, 2018). A PFMA may contain multiple subareas, with a PFMA denoted by a numeric and a subarea by an alphabetical. Our analysis focused on PFMAs throughout southern British Columbia (i.e., the west coast of Vancouver Island (WCVI) and the Canadian portions of the Salish Sea), which we aggregated into six catch regions based on proximity and shared oceanographic features (Fig. 1). Note that we moved two PFMA subareas (13M and 13N), located in the northern Strait of Georgia, to that catch region from the Queen Charlotte/Johnstone Strait region. The timing and location of fisheries restrictions within a given PFMA may change to avoid stocks of concern, as well as changes in quota as determined under the Pacific Salmon Treaty. For example, in recent years, management actions have resulted in the commercial troll fishery (largely restricted to outside portions of WCVI) shifting from harvesting in the late fall through early spring to harvesting only in late summer. The WCVI and inside (i.e., Canadian Salish Sea plus Queen Charlotte Strait) sport fisheries operate year-round with area-specific retention regulations, but the majority of effort occurs from early summer to early fall.

Data collection

Tissue samples for genetic stock identification (GSI) were collected from commercial and recreational fisheries by two independent sampling programs. Genetic stock assignments were performed using cBAYES and a reference baseline derived from microsatellite markers that consisted of 268-populations with more than 50,000 individuals (Beacham et al., 2006).

Genetic samples were collected from the commercial troll fishery, predominantly dockside and at processing facilities, from 2007–2015. Sampling was performed by observers contracted by Fishery and Ocean Canada’s (DFO) Mark Recovery Program to recover CWTs. The Mark Recovery Program aims to sample 20% of the landed WCVI commercial catch and these samples were further sub-sampled to provide GSI samples with a target of 4% of the monthly catch. We note that GSI samples from the commercial fishery could be attributed to a catch region and landing day, but not to a specific spatial location or harvest date because trollers may fish multiple PFMAs and remain at sea for several days before landing their catch. Additional GSI samples originated from two other WCVI fisheries. The first was a contracted test fishing troller with an at-sea observer, which gathered samples in May, June, and September of 2008–2011 from locations in the NWVI catch region immediately before and after the standard commercial opening. The second was the T’aaq-wiihak fishery, a First Nations economic-opportunity fishery, that operates on the west coast of Vancouver Island. T’aaq-wiihak fishery openings may occur at different times than standard commercial fisheries and we only retained samples in this analysis that overlapped with commercial openings. Commercial catch (individual fish) and effort (boat days) data were retrieved from DFO’s Fisheries Operations System database for all PFMAs and years in which GSI samples were available. These data consist of daily individual catch values from mandatory vessel logbooks, as well as the number of licensed vessels operating on a given day in each PFMA. Data from the commercial fishery were restricted to the northwest Vancouver Island (NWVI) and southwest Vancouver Island (SWVI), otherwise referred to as “outside”, catch regions (Fig. 1). We included commercial catch and effort data from 2007–2015 with catch and composition data available for all months except July in SWVI (18,470 individual samples; Figs. S1; S2).

Genetic samples were collected from the recreational fishery by dockside creel survey observers, as well as via a citizen science program, Avid Anglers, which is a collaboration between DFO, recreational harvesters, sport fishing guides, and the Pacific Salmon Foundation. The program encourages individuals who are consistent fishers (e.g., guides) to provide size and location data for all salmon that they encounter (kept and released), as well as sampling at least one fish per day for GSI. We aggregated the recreational GSI data to catch region and month to ensure adequate sample sizes and facilitate comparisons with commercial data. Recreational catch data were obtained from DFO’s creel database, which records estimates of monthly catch (individual fish) and effort (boat days) at the subarea level based on dockside fisher interviews and regular fly-overs (English, Searing & Nagtegaal, 2002). Sufficient GSI samples and catch data from recreational fisheries were only available from PFMAs within “inside” regions (i.e., Queen Charlotte and Johnstone Strait, the Strait of Georgia, and Juan de Fuca Strait; Fig. 1) and, with the exception of southern Strait of Georgia GSI data, were only available for a subset of months. Therefore, we fit the recreational models to data from these catch regions for months between January and December, with specific ranges differing among regions (Fig. S2). Recreational catch/effort data spanned 2009–2019 and stock composition data (11,729 individual samples) spanned 2014–2019 (Figs. S1; S2). All biological sampling was covered by a blanket Section 52 license relevant to Fisheries and Oceans Canada field activities for management purposes.

We pooled populations with similar freshwater distributions and run timing to generate two sets of stock aggregates. The first grouping (Regional, Table 1) contained aggregates based on those defined in Pacific Salmon Treaty documents (CTC, 2019), with the modification that several regional aggregates that are less common in southern BC fisheries were pooled and that Columbia River stocks were grouped based on common juvenile patterns of marine dispersal (Fisher et al., 2014). The second, Canadian-centric grouping retained fine-scale Canadian-origin aggregates most relevant to southern BC management decisions and pooled all other aggregates (Table 1). Although the PST Chinook technical committee recently included higher resolution stock groupings for certain Canadian stocks (CTC, 2020), we use the older regional groupings to improve readability and because the relevant stocks are shown in the Canadian-centric results. Our groupings largely pool stocks with evidence of similar marine distributions based on CWT recoveries (Weitkamp, 2010); however, aggregating stocks will necessarily obscure population-specific distributions. We recommend using down-scaled model versions when a small number of stocks are of interest. Additionally, we note that most stock aggregates contain a mix of hatchery- and wild-origin populations; however, the relative proportion of each varies among aggregates, as well as years, and we did not attempt to distinguish between patterns in hatchery and wild abundance.

Stock-specific distribution model

We assumed catch per unit effort (CPUE) could be used as a proxy for relative abundance in a given spatio-temporal strata (here catch region and month), though we note that hyperstable catch rates can violate this assumption (Harley, Myers & Dunn, 2001). To predict standardized CPUE (i.e., catch of all stocks given a fixed mean effort), we modeled catch (individual fish) C as a negative binomial process, with mean μ and inverse dispersion ϕ (Var[C] = μ + μ²/ϕ), via a log link and a generalized additive model (GAM). We included log effort as an offset (i.e., fixed the effort coefficient at one), and allowed changes in seasonal abundance to be described by splines. The GAM followed the general form:

(1a) $$C_i\sim \mathrm{N}\mathrm{e}\mathrm{g}\mathrm{B}\mathrm{i}\mathrm{n}\left({\mu }_i,\varphi \right),$$

(1b) $$\log ({\mu }_i)={\alpha }_y+{\alpha }_p+\sum _{m=1}^M{f}_{m_p}(m_i)+\log (b_i),$$

(1c) $${\alpha }_y\sim (\begin{array}{cc}\mathrm{N}\mathrm{o}\mathrm{r}\mathrm{m}\mathrm{a}\mathrm{l}\left({\mu }_{{\alpha }_{\mathrm{y}}},{\sigma }_{{\alpha }_{\mathrm{y}}}^2\right),& ify=1\\ \mathrm{N}\mathrm{o}\mathrm{r}\mathrm{m}\mathrm{a}\mathrm{l}\left({\alpha }_{\mathrm{y}-1},{\sigma }_{{\alpha }_{\mathrm{y}}}^2\right),& ify>1\end{array}$$

where i is an observation, α_y is a random intercept for year y modeled as a random walk, α_p is a fixed intercept for PFMA p, f_m is a PFMA-specific smooth function for month m, and b is the number of boat days (representing effort). Each smoother is represented by a sum of k basis functions, multiplied by corresponding coefficients (Wood, 2011). When data were available for the full annual cycle (e.g., commercial and southern Strait of Georgia composition data) we fixed k at four and fit the model with cubic cyclic splines, which constrained estimates for the first and last months of the year to converge on one another (Wood, 2006). When data were available for a fraction of the annual cycle, we fixed k at three and fit the model using thin plate splines. Values of k were chosen after preliminary model runs indicated higher values failed to converge or resulted in unrealistic seasonal patterns.

To facilitate comparison with the stock composition data, which could not be reliably assigned to PFMAs in the commercial fishery, we estimated μ_mr (monthly catch within each catch region r) as the sum of its component PFMAs:

(2) $${\mu }_{m_r}=\sum _{p=1}^P{\mu }_{m_p}.$$

We modeled stock composition as a Dirichlet-multinomial process—a compound distribution that accounts for variability in observed proportion data, similar to the approaches used by Thorson et al. (2017) and Douma & Weedon (2019). We assumed predicted stock proportions q are related to a vector of observed stock proportions θ and sample size n:

(3) $$\mathbf{q}\sim \mathrm{M}\mathrm{u}\mathrm{l}\mathrm{t}\mathrm{i}\mathrm{n}\mathrm{o}\mathrm{m}\mathrm{i}\mathrm{a}\mathrm{l}\left(\mathbf{\theta },n\right).$$

Variability in stock proportions, associated with uncertainty in the assignment probabilities for individual fish, is approximated using a Dirichlet distribution described by a vector of positive parameters λ representing its mean and variance.

A benefit of using a Dirichlet-multinomial, rather than multinomial, distribution is its ability to incorporate uncertainty in individual stock assignments (i.e., non-whole number observations of stock-specific counts in a sample), while accounting for variation among spatio-temporal strata in sampling effort. This contrasts with many GSI analyses where threshold probabilities are used to assign individuals to a population (e.g., 75%), effectively assuming perfect identification and excluding data associated with ambiguous stock assignments.

In practice, we implemented the model by computing the integrated Dirichlet-multinomial distribution (Thorson et al., 2017). Since individual assignment probabilities were identified at the scale of spawning populations (i.e., sub-stock units), we first aggregated GSI data into the stocks described in Table 1 by summing, within an individual, all probabilities associated with populations belonging to stock s. We next defined a sampling event j as all the individuals collected on a given day and within a given catch region and created vector ν_j by summing the assignment probabilities of all sampled individuals. Thus ν_j has length equal to the total number of stocks S and sums to n_j (i.e., ν_j is equivalent to n_j θ_j). Using the gamma function within the likelihood function of the Dirichlet-multinomial allows it to be defined for non-negative (rather than whole number only) sample sizes:

(4) $$L(\mathbf{q},\rho |\nu ,n)={\displaystyle \frac{\mathrm{\Gamma }(n+1)}{{\prod }_{s=1}^S\mathrm{\Gamma }({\nu }_s+1)}{\displaystyle \frac{\mathrm{\Gamma }(\rho )}{\mathrm{\Gamma }(n+\rho )}\prod _{s=1}^S{\displaystyle \frac{\mathrm{\Gamma }({\nu }_s+\rho q_s)}{\rho q_s},}}}$$

where γ represents the gamma function and ρ is a parameter representing overdispersion resulting from the Dirichlet distribution (Thorson et al., 2017).

We modelled seasonal patterns in stock composition similarly to Eq. (1b) using a logit link and the GAM:

(5a) $$logit({\mathbf{q}}_j)={\beta }_y+{\beta }_r+\sum _{m=1}^M{g}_{m_r}(m_j),$$

(5b) $${\beta }_y\sim (\begin{array}{cc}\mathrm{N}\mathrm{o}\mathrm{r}\mathrm{m}\mathrm{a}\mathrm{l}\left({\mu }_{{\beta }_y},{\sigma }_{{\beta }_y}^2\right),& ify=1\\ \mathrm{N}\mathrm{o}\mathrm{r}\mathrm{m}\mathrm{a}\mathrm{l}\left({\beta }_{y-1},{\sigma }_{{\beta }_y}^2\right),& ify>1\end{array}$$

where β_y is a random intercept for year y modeled as a random walk, β_r is a fixed intercept for catch region r, and g_m is a catch region-specific smooth function for month m. We used the same number of basis functions and the same spline types for the stock composition component of the model as the aggregate abundance component. To ensure model convergence we replaced zero stock assignment probablities in θ_j with very small values (0.00001). Similarly, parameters associated with stock-region combinations that were never observed (CA/OR-coast and Fraser Spring 4.2 in the northern Strait of Georgia) were fixed at zero.

We used the negative binomial and Dirichlet-multinomial model components to predict seasonal changes in standardized CPUE and stock composition, respectively, in a given catch region. We first generated predictions of standardized CPUE by assuming effort was fixed at a fishery-specific (i.e., commercial or recreational) mean value. We then used the product of predicted standardized CPUE and the probability of encountering a given stock to predict standardized stock-specific CPUE, a proxy for stock-specific abundance. By integrating across random effects, these predictions can be interpreted as representing an average year. Thus the model allows inferences to be made about the seasonal distribution of stocks in various catch regions, while accounting for interannual differences in sampling effort. Note that we also present estimates of predicted stock composition fit separately from abundance data because recreational GSI data were available for several months that lacked complete catch and effort data.

We fit the model to four datasets: regional stock aggregates captured in the commercial fishery (outside regions), regional stock aggregates captured in the sport fishery (inside regions), Canadian-origin stock aggregates captured in the commercial fishery, and Canadian-origin stock aggregates captured in the sport fishery (Table 1). We focus on regional aggregates in the main text, but include Canadian-centric results as a supplement.

We also completed two supplemental analyses to account for data collection methods that could impact our conclusions. First, we evaluated the effect of including the Avid Angler (i.e., voluntarily submitted) samples, which may be a less representative sample of stock composition, by repeating the composition analysis for the recreational fisheries, either including or excluding the Avid Angler samples. We then compared the predictions of models fit to the full and restricted datasets. Excluding the Avid Angler samples reduced the total number of samples by ∼50% and this analysis was limited only to early summer to fall months. These results are presented in the main text and as supplemental figures. Second, we evaluated the effect of fisheries restrictions in Juan de Fuca Strait and the southern Strait of Georgia on abundance and stock composition estimates. Management of Chinook salmon fisheries in southern BC includes spatio-temporal closures, as well as mark- and size-selective fisheries, to reduce impacts on stocks of concern. We included GSI samples collected from released individuals in all our analyses and, in a supplemental methods section, evaluated whether these samples adequately represent the activity of the recreational fishery.

We emphasize that estimates of commercial and recreational CPUE are independent, relative proxies for abundance and their scales should not be considered equivalent for several reasons. First, each dataset was collected over different time periods. Second, estimates of recreational catch and effort are substantially less precise than commercial equivalents due to differences in reporting requirements. Third, catchability differs between the fisheries because of differences in gear type and regulations, resulting in distinct relationships between CPUE and abundance. Fourth, the parameters in each fishery’s model were estimated independently and predictions were generated with effort standardized to a fishery-specific mean. Thus the “effort effect” is different and not directly comparable between commercial and recreational models.

We identified parameter values that maximized the marginal likelihood with respect to fixed effects while integrating across random effects via the the non-linear minimizer nlminb in R 4.0.2 (R Core Team, 2020). We first used the mgcv R package to define model matrices containing the appropriate splines to estimate seasonal and effort effects (Wood, 2011). We then fit the models with TMB, which implements the Laplace approximation to integrate across random effects and the generalized delta method to compute standard errors of all fixed and random effects, as well as derived quantities (Kristensen et al., 2016). We computed the standard errors on predictions in link space and calculated confidence intervals as Wald confidence intervals. The TMB framework allows the model to be fit quickly and the model’s structure is sufficiently flexible to accommodate a wide range of spatially stratified composition and catch data. Code to reproduce the analysis is available at https://github.com/CamFreshwater/chinDist and DOI 10.5281/zenodo.4672524, while the functions that allow fitting the integrated model to any similarly formatted dataset are available as an R package stockseasonr (DOI 10.5281/zenodo.4672540).

Regional variation

Although Chinook salmon standardized CPUE, a proxy for relative abundance, generally peaked in mid-summer, seasonal trends differed among regions. Broadly, this variation suggests Chinook salmon use habitats on the west coast of Vancouver Island differently to regions within the Salish Sea, though both fall within Healey’s category of a continental shelf distribution (Healey, 1983). On the west coast of Vancouver Island, seasonal peaks in relative abundance were broad due to a mix of resident and migratory life-history types, as well as a particularly diverse stock composition. Conversely, the seasonal peak in relative abundance in Juan de Fuca Strait was compressed, consistent with the region being used predominantly as a migratory corridor for a smaller number of stocks returning to systems within the Salish Sea. The other inside regions (Queen Charlotte and Johnstone straits, as well as the northern and southern Strait of Georgia), exhibited much weaker seasonal cycles. The lack of an obvious seasonal peak in relative abundance in the southern Strait of Georgia, which had robust estimates of catch and effort throughout the year, suggests that substantial numbers of fish remain resident year-round or return to inside waters before recruiting to the fishery (45 or 62 cm fork length depending on PFMA). Chinook salmon residence has been relatively well documented in Puget Sound (O’Neill & West, 2009; Chamberlin et al., 2011; Chamberlin & Quinn, 2014; Arostegui et al., 2017); however, in the Strait of Georgia, this life-history strategy has not received extensive attention in the primary literature (but see Healey & Groot, 1987).

Coast-wide analyses indicate the stock composition of adult Chinook salmon varies spatially, with marine distributions correlated with freshwater life history (Healey, 1983; Fisher et al., 2014) and ocean entry location (Weitkamp, 2010). We found that each region’s catch was dominated by three or four stock aggreagates with resident populations eventually replaced by migratory stocks in summer and fall. Yet seasonal patterns in composition were highly variable among regions. Some of these patterns are intuitive (e.g., stocks that migrate directly from freshwater to the California Current are rarely observed within the Salish Sea). Others, however, are less obvious. For example, Puget Sound stocks were common in southern outside (SWVI) and inside (southern Strait of Georgia and Juan de Fuca Strait) regions during winter and spring, but in regions slightly further north they were replaced by similar stocks (i.e., predominantly fall run, subyearling) from the Columbia River basin or the Strait of Georgia.

Stock-specific patterns

Patterns of stock-specific relative abundance (i.e., stock-specific standardized CPUE) within southern BC emphasize subtle differences in distribution and migratory behaviour among Chinook salmon populations that are lost when broadly categorizing stocks as continental shelf residents or offshore migrants. For example, the standardized CPUE of many subyearling, fall run populations was greatest in NWVI (e.g., Strait of Georgia, upper Columbia River summer/fall, CA/OR coastal), but lower Columbia River and Fraser River late run populations were more abundant in SWVI. The relative difference in standardized CPUE between NWVI and SWVI also ranged from ambiguous (e.g., Puget Sound) to dramatic (e.g., Washington coastal). While ocean-type stocks that are observed in southeast Alaskan troll fisheries are often referred to as far-north migrating (CTC, 2019), our seasonal estimates of stock-specific abundance suggest differences in distribution may also occur further south. Such stock-specific patterns are likely a result of differences in marine maturation grounds, but may also be influenced by migratory behaviours, such as travel speed or distance from shore, that moderate a stock’s exposure to fisheries.

Chinook salmon residence in southern continental shelf, and even nearshore, habitats during non-migratory periods is well known by fisheries managers and described qualitatively in Pacific salmon life history texts (e.g., Riddell et al., 2018). However, by quantifying seasonal trends in relative abundance, we are able to better resolve stock-specific distributions. First, Puget Sound fish were abundant throughout the year in both SWVI and NWVI, peaked in abundance in late winter or early spring in the southern Strait of Georgia, and migrated through Juan de Fuca Strait in late summer. This diversity may represent multiple allopatric components of the same cohort or perhaps movements between basins associated with ontogeny. Substantial numbers of Puget Sound Chinook salmon are resident within the sound during winter and early spring (O’Neill & West, 2009; Chamberlin et al., 2011; Chamberlin & Quinn, 2014), and individuals may exhibit restricted localized distributions (Arostegui et al., 2017).

Second, Columbia River (except lower spring run populations) and Strait of Georgia stocks were also present in NWVI during the winter and spring. Yet seasonal patterns in abundance were much more pronounced than in Puget Sound fish, suggesting the majority of each stock migrates north or offshore. Individuals caught in winter and spring may represent an early arrival by northern/offshore fish that will mature the following year or the southern portion of those populations’ marine distribution.

Third, Fraser River late run (fall run, subyearling fish) and Strait of Georgia (predominantly fall run, subyearling east coast Vancouver Island populations) stocks were abundant early in the year in the Strait of Georgia. Yet there was also a late summer peak in the abundance of these stocks on the west coast of Vancouver Island, followed by an increase in Juan de Fuca Strait. These patterns suggest that Fraser River late run and Strait of Georgia stocks may exhibit multiple, distinct migratory behaviors. Some proportion appear to either return early to inside waters via Johnstone Strait (a migratory pulse that has not been well-resolved with available data) or else remain resident within the Salish Sea until they recruit into the fishery. Another component of these stocks appears to migrate from the continental shelf through Juan de Fuca Strait immediately prior to spawning migrations. If, as we suspect, some fish remain resident within the Salish Sea, there may be substantial implications for management. For example, previous estimates of poor overwinter survival from acoustically tagged east coast Vancouver Island Chinook salmon may be biased low (Neville, Beamish & Chittenden, 2014).

Conservation implications

Near year-round occupancy of the west coast of Vancouver Island and the Strait of Georgia by Chinook salmon emphasizes these locations should be considered foraging and maturation habitats for a relatively large number of stocks. Importantly, each region has oceanographic characteristics that differ from each other, as well as from offshore and northern areas (Ware & McFarlane, 1989; Mackas & Coyle, 2005). The west coast of Vancouver Island is the approximate location of the bifurcation of the North Pacific Current and straddles downwelling-dominated regions to the north and upwelling-dominated regions to the south (Ware & McFarlane, 1989). Interannual variation in the latitude of the bifurcation point and basin-scale climate forcing (e.g., the Pacific Decadal Oscillation) influences primary productivity and zooplankton community composition, with subsequent effects on higher trophic levels, including salmon (Peterson, 2009; Sydeman et al., 2011; Malick et al., 2017). Conversely, the Strait of Georgia is a protected coastal sea that is strongly influenced by estuarine circulation and freshwater inputs, rather than vertical transport (LeBlond, 1983). These traits have resulted in zooplankton communities (Mackas et al., 2013) and fish population dynamics (e.g., herring; Cleary et al., 2020) that are distinct from those of the continental shelf.

Distinct spatio-temporal distributions may contribute to variation among salmon stocks in growth and productivity. While declines in Chinook salmon body size and age-at-maturity are widespread, their extent varies among regions (Ohlberger et al., 2018; Oke et al., 2020). Given that Pacific salmon size- and age-at-maturity are influenced by growth late in marine residence (Quinn, 2018), stock-specific adult marine distributions may be essential to identifying mechanistic drivers of these declines. Already CWT recovery data suggest Chinook salmon may exhibit stock-specific shifts in marine distribution as climate change progresses, which could magnify divergent responses to what is commonly viewed as a shared environment (Shelton et al., in press). From a metapopulation perspective, variation in marine distributions may contribute to salmon portfolio effects and stabilize aggregate abundance (Freshwater et al., 2018). Indeed Chinook salmon populations entering the Salish Sea, a region that we have illustrated to contain greater variability in marine behaviour, exhibit weaker synchrony in productivity than stocks that enter the California Current or Gulf of Alaska (Dorner, Catalano & Peterman, 2017; Ruff et al., 2017).

Interspecific interactions, including predation, are likely impacted by differences in spatio-temporal distribution among Chinook salmon populations. Resident killer whales preferentially target Chinook salmon and early in the summer depend heavily on spring and summer run yearling Fraser River stocks (predominantly from the upper and middle portions of the watershed) (Hanson et al., 2010). Although the absolute abundance of yearling Fraser River Chinook salmon (all early run timing populations) is less than summer and fall run subyearling stocks (CTC, 2019), our estimates of relative abundance suggest that yearling stocks may be particularly available to southern resident killer whales due to their compressed migration through Juan de Fuca Strait. Conversely, Fraser River fall run stocks have a more dispersed spatial and temporal distribution within southern BC, which may reduce their availability as prey.

Data and model structure assumptions

Our predictions of stock-specific abundance depend on imperfect sampling of fisheries-dependent data, resulting in several assumptions. First, we accounted for annual variation via random effects. Such an approach was necessary due to imbalanced sampling across the annual cycle among years and relatively few samples for many strata; however, without year-specific fixed effects, the model will underestimate anomalous boom or bust years in stock-specific abundance. Second, estimates of catch and effort for the recreational fisheries data, derived from creel surveys and overflights, are particularly uncertain. Since this observation error was not incorporated into our predictions, the uncertainty associated with standardized CPUE estimates in inside regions is underestimated.

Third, it is unclear how precisely stock-specific harvest reflects stock-specific relative abundance because fisheries-independent estimates of mature Chinook salmon abundance in marine areas are not available. For example, catchability may vary seasonally due to changes in the depth distribution or behaviour of fish, as well as changes in fleet composition (e.g., highly motivated and skilled fishers may make up a greater proportion of the fleet during winter). Management actions, such as area closures and non-retention periods, will impact fisher behaviour and may also decouple catch from abundance. As a result, we can provide only minimum estimates of relative abundance of Fraser River early run stocks in Juan de Fuca Strait and the southern Strait of Georgia, as well as WCVI stocks in NWVI and SWVI, because spatio-temporal closures are used to minimize incidental harvest of those stocks (DFO, 2012; Dobson, Holt & Davis, 2020). Estimates of composition may be particularly biased in Juan de Fuca Strait and the southern Strait of Georgia, where recreational fisheries release relatively large numbers of individual that are difficult to genetically sample (Dobson, Holt & Davis, 2020; see Supporting Information for additional information). These and other difficulties associated with fisheries-dependent data emphasize the importance of developing robust catch sampling programs and of leveraging multiple data sources to inform fisheries management decisions.

Age structure is not incorporated into many salmon distribution models derived from GSI data (including the one presented here), but is a logical addition to increase ecological realism. CWT recoveries suggest older juvenile (Trudel et al., 2009) and adult (Weitkamp, 2010) Chinook salmon age classes, within a stock, are distributed further from their natal streams. GSI analyses incorporating age data from scale samples or associated CWT indicators could be used to resolve how robust this pattern is and increase the spatio-temporal resolution of movement models. Disentangling age- and stock-specific marine distributions may be necessary to identify the mechanisms driving declines in older age classes (Ohlberger et al., 2018; Oke et al., 2020).

Our model estimates stock-specific abundance as a function of total abundance and stock composition. Yet stock-specific abundance is commonly inferred directly from CWT recoveries, which are expanded for tagging and sampling effort, then used to calculate total abundance and stock composition as necessary. Such an approach approximates reality, where total abundance is the sum of stock-specific abundance, and is necessary for CWT recoveries because the entire catch is not tagged (thus composition estimates cannot be applied to total catch). We believe, however, that this method is less appropriate for GSI data because the expansion factors that are necessary to account for variable sampling effort when analyzing CWT recoveries are not readily available for GSI samples. We account for variable sampling effort by explicitly incorporating sample size into the stock composition component of the model, which has the additional benefit of accounting for uncertain stock assignments. Nevertheless, comparing inferences derived from CWT and GSI data using a common model, perhaps by estimating stock-specific abundance as latent variables constrained by total abundance, would be a valuable addition.