4.1 Variation in vital rates

Changes in vital rates were identified as a key consideration across all five geographical areas discussed during the workshop (Table 1), with changes in fecundity, natural mortality and age-at-maturity potentially leading to low-frequency variation in productivity, a pattern observed in many stocks, including some SPFs.

4.2 Spatial considerations

Spatio-temporal dynamics are an essential component of how SPFs behave and how predators and fisheries exploit them. Spatial issues include hyperstability, migration, stock mixing, spatial differences in predator needs, particularly for central place foragers susceptible to localized depletion, and spatial differences in the implementation of harvest control rules, such as when multiple countries share the same resource. How these are represented in the OM will depend on the research question and the performance objectives for the stock. Stock identification (the identification of biologically distinct stocks for the purposes of management) and the spatial mixing between these management units are an issue for several stocks, including: Pacific herring (Clupea pallasii, Clupeidae) in the Gulf of Alaska whose growth and fecundity vary spatially; Northeast Atlantic sprat (Sprattus sprattus, Clupeidae) (Hunter et al., 2019; McKeown et al., 2020; Quintela et al., 2020); anchovy and sardine in the Mediterranean Sea (Fiorentino et al., 2014); two stocks of Atlantic herring (Clupea harengus, Clupeidae), namely, western Baltic spring-spawning herring and central Baltic herring (ICES, 2018a); and South African sardine (de Moor et al., 2017; Weston et al., 2015).

Many SPFs undergo range contractions when abundance is lower (i.e. the “basin model” of MacCall, 1990), which can affect productivity if it results in depensatory dynamics (Holling, 1959; Liermann & Hilborn, 2001; Swain & Benoît, 2015; Walters & Kitchell, 2011). Conversely, range collapses might be explained using information about biomass, changes in recruitment or spatial vulnerability to disease (e.g. MacCall, 1990). These behaviours can also affect the precision and accuracy of abundance indices when range expansions in high-abundance years result in surveys missing a proportion of the stock (Barange et al., 2009; Hilborn & Walters, 1992; MacCall et al., 2016). For example, in the California Current, where sardine habitat appears to extend northward at high population sizes and under specific oceanographic conditions, the extent of the spring acoustic survey for sardine depends on an environmentally based estimate of spawning habitat extent to ensure that all potential spawning habitat is sampled (Zwolinski et al., 2011). While expansion–contraction effects can be modelled empirically, there is seldom a mechanistic way to model them.

Recruitment can also vary spatially, and the limitations of recruitment by habitat may drive dynamics. For example, optimal spawning habitat for Pacific sardine is associated with specific oceanographic conditions (Nieto et al., 2014; Reiss et al., 2008; Weber & McClatchie, 2010; Zwolinski et al., 2011). Recruitment can be spatially correlated because of sensitivity to local environmental conditions (e.g. in north Atlantic and northeast Pacific herring stocks; Zheng, 1996). Stocks can also be expected to be less productive at their margin of distribution if they follow the basin model (MacCall, 1990; Saraux et al., 2014). Alternatively, the underlying stock–recruit relationship might change with local depletion because density-dependent effects occur at small spatial scales (e.g. Casini et al., 2011). An OM may need to consider local dynamics if they lead to emergent patterns at the stock scale (MacCall et al., 2019; Punt, MacCall et al., 2016). This is possible with some ecosystem models that have small-scale dynamics and have been used for MSE (e.g. Fulton et al., 2011). As the stock–recruit relationship varies spatially, so will carrying capacity, which might be easier to model implicitly using depensatory stock–recruitment relationships instead of explicitly with a spatial model (Liermann & Hilborn, 2011). Spatial models linked to oceanographic models (i.e. “end-to-end” ecosystem models) can generate spatio-temporal dynamics in recruitment, but conditioning these such models is so resource-intensive that applying them at a resolution relevant to management is often not an option.

Interactions between predators and SPFs are also inherently spatial, and predator populations may not have access to the entire region of SPF distribution. Spatial models can capture some of the dynamics and behaviour of predators that forage from a fixed location like a breeding colony (e.g. Barrett et al., 2006). Some spatial structure can be captured using Models of Intermediate Complexity (MICE; Plagányi et al., 2014) with limited spatial structure (e.g. Punt, MacCall et al., 2016; Robinson et al., 2015). Alternatively, spatial structure could be approximated by modelling an (annually varying) proportion of the population to be within the foraging area and thus subject to the predation of a predator. Short time steps (e.g. monthly or quarterly) may be required to effectively model individual prey movement.

If fishing pressure differs throughout a SPF’s distribution range, the MSE may require spatial structuring in the OM to adequately account for these differences. The effect of fishing activities on the dynamics and movement of SPFs should be considered in spatial models (Sydeman, Thompson et al., 2017; Okamoto et al., 2020). Plausible spatial distributions of fishing activity can be generated using random utility models of fisher behaviour and port-level landings (e.g. Rose et al., 2015). While explicit modelling of areas is preferable where data permit, in some cases, selectivity can be used as a proxy for area (the “areas as fleets” approach; Cope & Punt, 2011; though some simulation studies now indicate that this approach is sub-optimal). For example, spatially disaggregated fishing effort may be adequately reflected by age-disaggregated selectivity if the SPFs move through different areas as they grow (e.g. South African anchovy de Moor & Butterworth, 2016). SPFs often occupy an area that spans the jurisdictions of multiple countries. For example, Pacific sardine are fished across Canada, the USA and Mexico (Punt, MacCall et al., 2016). The South Peru–North Chile anchovy stock is managed independently between Peru and Chile, with different assessments (Canales et al., 2020). In cases where the stock is not managed cooperatively across its range, management outcomes for the whole stock will depend on the combined effects of each country's management plan. While the benefits of cooperative management across the stock's range have been documented for some taxa (e.g. Pacific hake [Merluccius productus, Merlucciidae]; Hamel et al., 2015), SPFs are often large stocks, the majority of which are fished by more than one country (139 of 224 stocks, based on FAO wild capture production from 2019). Additionally, the need to work cooperatively in a short timeframe is unique to SPFs, given that acquiring data can sometimes take longer than the time for which the fish is available to the fishery (for example, recruitment fisheries for age-0 individuals). In cases where data are not available for all jurisdictions, simulations can include alternative plausible ranges for historical and future catches for countries for which dynamic management strategies are not available, such that fishery impacts on the entire stock can be modelled. Attempting to model dynamics across jurisdictions can identify data gaps where more information should be gathered (Fulton et al., 2011).

The choice of spatio-temporal dynamics in the OM depends on the research question, the scale of the dynamics, the resolution of the data, and the objectives of the strategic management decision being investigated. For example, the OM could explicitly include space or stock structure if users want to evaluate management strategy performance with respect to benefits to central place foragers such as seabirds (e.g. sardine and anchovy in the California Current, Punt, MacCall et al., 2016; South African sardine, de Moor et al., 2017). Alternatively, if users want to evaluate performance with respect to spatially-dependent harvest rates or implementation error (e.g. overshooting quotas), such as when multiple countries with different management structures harvest the same straddling stock, the OM could (minimally) be structured to represent the stock distributed by country. Spatial differences in recruitment, natural mortality, growth and fleet dynamics may necessitate a spatially explicit model. Choosing the appropriate spatial scale to represent is a central challenge, as misspecification can lead to incorrect representation of recruitment and of the management values that are used to estimate reference points (Kapur et al., n.d.). Other considerations include conceptual decisions about the hypotheses being tested and tactical decisions about how mortality and recruitment occur over space (Punt, 2019).

Not all spatial dynamics can or should be included in an MSE. One example of this is the case of sardine in South Africa, where the harvest control rule implemented after MSE testing provides total sardine quotas (spatially disaggregated by west/south coast in some years). In addition to this harvest control rule, an island closure experiment has restricted purse seine fishing around some key African penguin breeding islands (e.g. Ross-Gillespie & Butterworth, 2020; Sherley, 2020). In this case, the relatively small spatial scale of the 20-km radius around the penguin breeding islands was not explicitly modelled in the MSE but under EBFM, the impact of fishing on penguins is still being evaluated (de Moor, 2018). Spatial issues are likely to become even more urgent in the future, as climate change continues to influence the movement and distribution of SPFs. Management strategy evaluations that allow decision-makers to balance tradeoffs for multiple jurisdictions will be useful.

4.3 Multispecies interactions

SPFs are characterized by their role as prey for a diverse assemblage of predators, including predatory fish, marine mammals and seabirds (Engelhard et al., 2014; Pikitch et al., 2014; Smith et al., 2011). The availability of alternative prey for predators is a key consideration, but there is considerable uncertainty surrounding interactions between prey species and how these impact predators.

Properly characterizing the ecological relationships among multiple SPF stocks and how predators might respond to fluctuations in their prey is an important step in generating realistic dynamics for each species and determining the outcomes for fisheries that harvest them and predators that may depend on more than one species. Negative correlations between species abundance can be evidence of competition, apparent competition or opposite drivers. Thus, if there is little information or conflicting information about intraguild relationships, OMs with multiple forage species may need to include a range of possible options. For example, sardine–anchovy relationships can appear to be positive (McClatchie, 2012) or negative (Chavez et al., 2003), and different biological and non-biological processes can be invoked to explain the same observations (Hosack et al., 2013; Siple et al., 2020). Competitive relationships between SPFs can also change over the life cycle of a species, as in the Norwegian Sea where Atlantic mackerel feed on herring larvae (Allan et al., 2021; Skaret et al., 2015), while the adults of both species compete for planktonic prey on their feeding grounds (Nikolioudakis et al., 2019). There are several possible scenarios for intraguild relationships in fisheries where SPFs are part of a multispecies fishery. Considering different interactions between exploited SPF stocks could help to capture dynamics associated with competition or apparent competition due to food availability.

Predator–prey relationships—particularly functional responses and predation rates—may be one of the most important ecosystem processes driving the dynamics of SPFs. If possible, empirical data on functional response relationships (the shape of the relationships between predator demographic rate and prey abundance) should be used to parameterize predator responses (Koehn et al., 2021). These data may be available from meta-analyses, for example, seabird functional responses to changes in forage fish abundance (Cury et al., 2011) or from case studies of individual populations (e.g. Robinson et al., 2015; Suryan et al., 2006). Alternatively, statistical relationships representing the presence and strength of pairwise species interactions, including competition, can be estimated directly from time-series data using multivariate autoregressive models (Ives et al., 2003; Lindegren et al., 2009). However, even in data-rich cases, the shape of predator functional responses or estimated interaction terms are subject to uncertainty (Sherley et al., 2015; Sydeman, Piatt et al., 2017). Hence, alternative OMs might include different shape parameter values for the functional response.

Realized prey is a function of preferred prey, spatial overlap and the availability of other prey items (Blanchard et al., 2014). If predators eat non-preferred prey when it is abundant but a more preferred prey if available, a single-predator, single-prey model would be unable to sufficiently capture the plausible range of predation rates experienced by the prey (e.g. Pinnegar et al., 2003; Plagányi et al., 2014). Care should also be taken as to the importance of the focal SPF in the ecosystem. While one or two species of SPF are the key forage species in some ecosystems (i.e. “wasp-waisted” ecosystems; Cury et al., 2000) SPFs can be part of a diverse forage category, in which case models with predators must be flexible enough to characterize the dynamics of other forage species and the potential for key predators to prey-switch (e.g. with sardine and anchovy in the California Current; Koehn et al., 2016). Ecosystem models with the appropriate degree of structure can be helpful in distinguishing whether a given species is a “key” component or not (Plagányi & Essington, 2014). Incorrect assumptions about functional responses could lead to over- or under-estimation of the impacts of SPF fisheries on predators (Plagányi & Butterworth, 2005). Therefore, it is important to consider uncertainty in the strength and shape of predator–prey functional responses when assembling the reference set of OMs and robustness tests. The effects of a disturbance or a large increase in predator abundance may not propagate through the food web in a manner that is consistent across ecosystems, and the observed strength of trophic cascades depends on the response of the particular SPF to changes in food availability, as well as density dependence (Heath et al., 2014). These processes can be represented in OMs implicitly in the form of time-varying natural mortality and uncertainty in stock–recruit relationships, respectively.

The number of species, sites or ecosystem components included in the OM is important decisions to consider for SPFs because they are never at the top of the food chain; thus, there is no natural limit on model complexity. Accounting for additional non-predator–prey food web relationships may result in more realistic OMs than single-species models, although sometimes it suffices to model ecosystem components such as predation with a realistic range of natural mortalities (Deroba et al., 2018) or allow the OM to explicitly account for predator needs (Fulton et al., 2014; Pikitch et al., 2012). Theoretically, including more species increases the likelihood that the OM better represents the real world, provided the ecological processes are well understood, but the increased number of parameters required for a more complex OM involves evaluating an ever-larger set of robustness trials, which can quickly become prohibitive when parameters are unknown and cannot be estimated reliably. This may be the case when there are one or more forage species without stock assessments. In some cases, other analyses can be used to quantify the benefits of including additional ecological information to models (e.g. economic analyses; Essington et al., 2018) and to optimize the degree of model complexity (Collie et al., 2016), given the model error, that is, the uncertainty in the input variables propagating to the model output (Saltelli, 2019).

When OMs include predators as drivers of prey abundance, it is important for OMs to generate plausible levels of variability in predator and prey abundances if the objectives of management are based on desired levels of predator reproduction or abundance. The latter has been the primary way that multispecies interactions have been considered in MSEs for SPFs so far. For example, the MSE for Atlantic herring in the USA evaluated control rules based on the needs of seabirds and piscivorous fish (Deroba et al., 2018; Feeney et al., 2018) and represents one case where an MSE approach has been used explicitly to support EBFM (Townsend et al., 2019). The MSE for California Current sardine and anchovy included performance metrics related to the abundance of predators (sea lions and brown pelicans) and the probability that those predators fell below abundance thresholds (Punt, MacCall et al., 2016). Penguin feeding has also been included in the MSE for South African sardine, where performance metrics included the projected impact of alternative sardine management strategies on penguin dynamics based on the estimated relationship with sardine abundance (de Moor, 2018; Robinson et al., 2015).

Generating realistic amounts of variation in natural mortality in single-species models for SPFs may be difficult for some stocks. Proxies such as predator abundance may be used to bracket a realistic range of values and develop plausible scenarios for low-frequency variability in natural mortality. For example, in the Barents Sea, cod abundance is used to simulate capelin mortality, and in the North Sea, a multispecies assessment model is used to calculate natural mortality for several species (ICES, 2020a,b). Hazen et al. (2019) review the ways that predator diet, movement, life history, reproduction and demography can be used to make inferences about unobserved ecosystem components including the abundance of SPFs (fish and invertebrates) and provide guidance for how to employ them to detect changes at different timescales. It is important to ensure that predator abundance proxies for natural mortality are linked to the life-history parameter that is most likely to be actually involved in the relationship. For instance, changing predator abundance may be more appropriate to link to depensatory recruitment than natural mortality if predators select younger prey.

The complexity of the OM should be informed by the research question and management priorities and how ecosystem information is to be used, that is, to generate dynamics or to inform ecosystem reference points. Complex ecosystem models are useful for making strategic management decisions but are generally not recommended for use as operating models in an MSE (Punt, Butterworth et al., 2016; Sainsbury et al., 2000). They can, however, be used to determine reference points or harvest rates that might satisfy predator needs (Pikitch et al., 2012; Smith et al., 2011), before those reference points are evaluated in a single-species MSE. MICE models are more practical for making tactical, practical and quantitative management decisions. For example, Smith et al. (2015) used a small pelagics-specific variant of an Atlantis ecosystem model (Atlantis-SPF) to evaluate reference points for the Commonwealth Small Pelagic Fishery; for northern British Columbia herring. Surma et al. (2018) applied an ecosystem model to ascertain ecological outcomes for harvest control rules, while socio-economic impacts were assessed by industry and community members for their societal desirability (Lam et al., 2019). Pikitch et al. (2012) added variability to Ecosim food web models to test a broad range of harvest control rules. On the other hand, MICE models can address both tactical and strategic questions, because they can be conditioned on data, provide estimates of current abundance and exploitation while retaining important ecological interactions. Predator effects on prey can be modelled explicitly when data are available to condition a MICE model (e.g. the NWACS-MICE model built for Atlantic menhaden; SEDAR, 2020). These simpler models can be beneficial to use as OMs, as they illustrate how a management strategy performs in complex situations without the uncertainty and computational burden associated with more complex ecosystem models.

4.4 Considerations for modelling assessment error and reference points in the estimation model (EM)

The OM should generate pseudo-data about the biological system that match what would be available in reality. In practice, these data would either be used directly in a harvest control rule (empirical management strategies; for example advice for North sea horse mackerel is based on changes in survey indices; ICES, 2019a, b, c), as input to an assessment when the harvest control rule depends on assessment model outputs (model-based Management Strategies), or both, when the control rule depends on both data and results from an assessment. For model-based management strategies, the Estimation Model (EM) should either be the existing or proposed assessment method to be used to manage the stock in question; or generate uncertainty similar to the existing or proposed assessment method.

The inclusion of spatially explicit and/or time-varying components in the OM will assist the evaluation of the performance of a management strategy for spatially structured SPFs relative to its objectives. In particular, performance metrics using reference points may require spatially explicit or time-varying reference points (RPs). RPs are benchmarks to which stock level and fishing intensity are compared for management purposes (e.g. B_MSY). RPs for determining stock status should be chosen appropriately given the “boom and bust” dynamics exhibited by many SPF populations. Some harvest control rules are linked to RPs. In some cases, externally estimated RPs (i.e. not estimated within an assessment) are used in management and can easily be input as time-invariant values in the MSE. In other instances, harvest control rules depend on annually estimated reference points output from the latest assessment (the EM) and time-varying changes to the RPs calculated by the EM should be considered. Where spatially explicit advice is to be provided, the EM would need to estimate spatially explicit reference points or spatially explicit relative stock status to inform the harvest control rule. This would ensure that performance of dynamic or spatially explicit management strategies is evaluated given realistic errors in the estimation of stock status and reference points.

EMs that include other ecosystem components are rare, but in some cases, they may help to meet new EBFM-related challenges. Tommasi et al. (2017) found through an MSE process that harvest guidelines explicitly incorporating sea surface temperature predictions as a driver of future SPF productivity increased biomass and yield and reduced the probability of biomass or yield falling below accepted thresholds when combined with harvest restrictions at low biomass. Environmental conditions are used in some harvest guidelines to ensure catch quotas are responsive to recruitment variation (e.g. in Pacific sardine; Kuriyama et al., 2020) and have been proposed in others as a way to avoid bias in biomass estimates (Peruvian anchoveta; IMARPE, 2019). Predator data collection could improve the characterization of uncertainty in EMs of SPFs and could allow predator abundance (Hollowed et al., 2000) or predator diet composition (Velarde et al., 2015; Zador & Yasumiishi, 2018) to be used as an indicator of prey abundance and distribution within management strategies. Predator diets may also be used as an indicator for SPF community composition (Elliott et al., 2015; Sydeman, Thompson et al., 2017). However, approaches that use ecosystem information in EMs for SPFs are rare and the OM would need to be constructed to generate data that reflects the uncertainty associated with using such information for management purposes (De Oliveira et al., 2005).