Using hierarchical models to estimate effects of ocean anomalies on north-west Pacific Chinook salmon Oncorhynchus tshawytscha recruitment


  • R. Sharma,

    Corresponding author
    1. Columbia River Inter-Tribal Fish Commission, 729 NE Oregon St, Suite 200, Portland, OR 97232, U.S.A.
      Tel.: +1 503 736 3590; fax: +1 503 235 4228; email:
    Search for more papers by this author
  • M. Liermann

    1. Fish Ecology Division, Northwest Fisheries Science Center, National Marine Fisheries Service, National Oceanic and Atmospheric Administration, 2725 Montlake Blvd. E, Seattle, WA 98112, U.S.A.
    Search for more papers by this author

Tel.: +1 503 736 3590; fax: +1 503 235 4228; email:


The high variability in survival over the past three decades of north-west Pacific Chinook salmon Oncorhynchus tshawytscha is summarized for 24 stocks and analysed using hierarchical Bayesian models. Results from a simple model indicate that recruitment anomalies appear to be correlated in time and space. A simple model with a covariate based on basin-scale effects (Pacific Decadal Oscillation and El Niño Southern Oscillation) and local-scale effects (sea surface temperature, SST anomaly) was introduced to explain this variability. The model still exhibited residual patterns that were removed when a random-walk component was added to the model. The analysis indicates that recruitment is negatively related to SST anomaly for all stocks and the effect of basin-scale variables is negligible. The effect of climate over the next century is expected to result in estimated recruitment declining by an average of 13% for O. tshawytscha stocks coastwide.


Chinook salmon Oncorhynchus tshawytscha (Walbaum) is a high valued species for recreational and commercial fisheries and is of high cultural significance for Native American tribes in the north-west Pacific region of North America (Hunn, 1990). Highly variable adult returns by age, a complex life history, and a complicated multi-jurisdictional fishery, however, have created substantial challenges for understanding their dynamics and management. When this complexity is combined with a management process that is slow to adapt, it is not surprising that a number of populations have been overfished during periods of low productivity (Peterman & Pyper, 2000). A better understanding of how productivity responds to ocean conditions could lead to better models of recruitment and therefore improved management. While scientists have tried to understand how Oncorhynchus spp. recruitment and survival varies with respect to marine environmental conditions (Hare & Francis, 1995; Hare et al., 1999; Magnusson, 2002, Mueter et al., 2002; Peterman et al., 2003; Beamish et al., 2004), the work done on O. tshawytscha has been focussed on particular populations, e.g. Skagit River, U.S.A. (Greene et al., 2005) or Snake River, U.S.A. (Hinrichsen & Fisher, 2009), has been dominated by hatchery populations (Magnusson, 2002) or has been limited to catch data (Hare & Francis, 1995). While results from these studies specific to O. tshawytscha have helped identify marine environmental conditions that are probably important drivers of their population dynamics, as far as is known, there are no geographically extensive models that link the marine environment with recruitment. The intent of this paper is to assess the degree to which marine environmental variables can explain O. tshawytscha recruitment variation for north-west Pacific populations and use them in a predictive sense for management.

Oncorhynchus spp. typically leaves the freshwater environment to enter an ocean that is in a state of flux. Depending on the timing of the spring transition, out-migrants could encounter a strong upwelling event, the downwelling conditions that dominate the west coast of North America in the winter months or the transition between the two (Bakun, 1996). Immediately after ocean entry, juvenile O. tshawytscha spend a large amount of time on the coastal shelf either migrating northwards towards Alaska or staying close to their natal rivers (Trudel et al., 2009). This early marine life-history phase has been shown to be extremely important with many studies linking interannual patterns of these conditions with indicators of Oncorhynchus spp. ocean survival (Logerwell et al., 2003; Peterson & Schwing, 2003; Lawson et al., 2004; Peterson et al., 2006).

Large-scale ocean-atmospheric processes like the El Niño Southern Oscillation (ENSO, Wolter & Timlin, 1998) and the Pacific Decadal Oscillation (PDO, Mantua et al., 1997) influence basin-scale winds and currents that drive the oceanic systems in the north-east Pacific Ocean (i.e. the north-west Pacific coast region of North America). Thus, coastal waters in the north-east Pacific Ocean are influenced by atmospheric conditions not only over the adjacent areas (as indexed by the PDO, Mantua et al., 1997) but also over equatorial waters, especially during El Niño events (Peterson et al., 2006). While the multivariate ENSO index is a composite of many variables [sea surface temperature (SST) zonal surface wind, sea-level pressure, surface air temperature and cloud fraction] at latitudes 5° north and south of the equator (Wolter & Timlin, 1998), the PDO is derived entirely from SST anomalies in the North Pacific Ocean (N of 20° N). Local-scale conditions are described by SST at a particular station rather than the entire North Pacific Ocean as the PDO does. Both PDO and ENSO have been found to relate to survival and abundance either directly through juvenile Oncorhynchus spp. abundance (Peterson et al., 2006) or through catch levels (Francis & Hare, 1994; Mantua et al., 1997; Hare et al., 1999). In addition, a number of studies have shown a negative relationship between SST anomalies and recruitment of Oncorhynchus spp. (Coronado & Hilborn, 1998; Ryding & Skalski, 1999), while Magnusson (2002) found a dome-shaped relationship (implying an optimal range of temperatures) between O. tshawytscha survival (estimated as the number of adults that survive from juvenile-known releases from a hatchery) and SST, and another between coho salmon Oncorhynchus kisutch (Walbaum) survival and SST. The relationships are assumed to be driven by the duration and the strength of upwelling events during and prior to smolt out-migration where strong upwelling (as indicated by lower SSTs) tends to lead to higher productivity and food availability (Peterson et al., 2006).

Sharma (2009) related patterns in the metrics SST, ENSO and PDO to survival trends for multiple O. tshawytscha stocks coastwide. The analysis focused on tagged stocks that were either natural populations or the best representation of natural populations (i.e. hatchery tagging programmes that are managed to mimic wild fish behaviour). Low-to-moderate negative correlations were detected between estimated survivals and ocean covariates SST, ENSO and PDO along the west coast of North America (Sharma, 2009). The strongest relationships found were with SST. Similar although weaker results were detected with ENSO and PDO for the stocks analysed.

The models presented in this paper build on those developed in Parken et al. (2006) and Liermann et al. (2010), where catchment size was related to density-dependent parameters. Hierarchical models presented in Liermann et al. (2010) simultaneously predict recruitment for multiple populations while assuming that demographic parameters controlling density-dependent and density-independent processes come from common coastwide distributions. That is, while parameters such as the coefficient relating productivity to SST are allowed to vary between populations, they are assumed to be independent values from a common distribution. In Liermann et al. (2010), strong residual trends that were observed in many of the populations were accounted for with random walks which allowed for time-varying productivity. The primary hypothesis in this study was that adding the ocean covariates would account for some of the variability modelled by the random walks. Specifically, there would be a negative relationship between the recruitment residuals and SST (as found in Sharma, 2009) and also a negative relationship between basin-scale indicators, namely PDO and ENSO, and recruitment residuals for most stocks from the north-west Pacific coast of North America in this analysis.

Materials and methods

Ocean sea surface temperature data

Sea surface temperature was calculated as the mean monthly SST during the primary months of ocean entry for O. tshawytscha smolts (April to September). The SST stations (Fig. 1) chosen as explanatory variables varied by population but were all located along the coastal shelf from 42 to 60° N. Specific station–population pairings were based on Sharma (2009) where survival estimates derived using coded wire tags were compared to SST from different stations at locations along the coastal shelf. The raw SST data were obtained from the Comprehensive Ocean Atmospheric Data Set (COADS, 2008), the National Data Buoy Center (NDBC, 2008) for US waters, and from British Columbia lighthouse data for Canadian waters (BCLD, 2008). All SST time series were standardized by subtracting the mean and dividing by the s.d. as in Smith & Reynolds (2004).

Figure 1.

Oncorhynchus tshawytscha stocks and associated sea surface temperature (SST) stations used in the analysis. Irregular polygons associate stocks with SST station. The names of the rivers are as follows: Klukshu (1), Situk (2), Taku (3), King Salmon (4), Andrew Creek (5), Stikine (6), Unuk (7), Chickamin (8), Blossom (9), Keta (10), Kitsumkalum (11), Harrison (12), Cowichan (13), Skagit (14), Quillayute (15), Queets (16), Humptulips (18), Chehalis (19), Lewis, Nehalem (20), Upper Columbia (21), Siltez (22), Siuslaw (23) and Deschutes (24).

The ENSO (ESRL, 2008) and PDO (JISAO, 2009) metrics are large basin-scale indices that represent conditions in the tropical Pacific and northern Pacific Ocean, respectively. Based on the results of Sharma (2009), an average value for the months of April, May and June was used for ENSO while PDO values were averaged for May and June. PDO and ENSO effects do not vary by populations, whereas SSTs do vary by population as local-scale data are used.

Spawner – recruit and catchment size data

Spawner–recruit data and catchment size were compiled for 24 populations of O. tshawytscha distributed along the north-west coast of North America from Alaska to Oregon including British Columbia (Table I). Recruitment was defined as the number of adults from the same age-class that would have survived to maturity without harvest, and spawners was calculated as the number of 2 year ocean age and older fish (stream-type O. tshawytscha implies an age 3 year fish and ocean type an age 2 year fish). Catchment size was defined as the total drainage area (km2) minus the area upstream of man-made barriers and natural barriers on fourth order or fifth order stream segments (Strahler, 1957) for respective stream-type or ocean-type O. tshawytscha populations (Healey, 1991). A complete description of the data can be found in Parken et al. (2006).

Table I.  Oncorhynchus tshawytscha populations used in the analysis with life-history type (O, ocean type; S, stream type), catchment size, latitude (the river mouth location with reference to the ocean), the first year of data, number of years of spawner–recruit data, sea surface temperature covariate location and time and per cent of residual variance explained by the ocean covariate as opposed to the random walk with ocean covariate (full) model. The spawner–recruit, latitude and catchment size data are from Parken et al. (2006) (with some minor updates)
River nameLife historyCatchment size (km2)Latitude (° N)First yearLength of time series (years)Sea surface temperature covariate location and monthPer cent change from base model residual variance explained by
Ocean modelFull model
KlukshuS26060·119761654·15° N 133·03° W September2655
SitukO17659·419761854·15° N 133·03° W September459
TakuS1553958·419731954·15° N 133·03° W September−25
King SalmonS9358·019712154·15° N 133·03° W September622
Andrew CreekS12656·719752454·15° N 133·03° W September−325
StikineS1533756·619772254·15° N 133·03° W September−108
UnukS221356·119772254·15° N 133·03° W September2340
ChickaminS169655·819772254·15° N 133·03° W September1188
BlossomS17655·419772254·15° N 133·03° W September1490
KetaS19255·319772254·15° N 133·03° W September2682
KitsumkalumS225554·519841454·15° N 133·03° W September−316
HarrisonO761149·219841548·18° N 123·32° W June615
CowichanO122748·819811948·18° N 123·32° W June478
SkagitO419848·419712848·18° N 123·32° W June543
QuillayuteO131347·919811148·00° N 124·00° W July2383
QueetsO116447·519771848·00° N 124·00° W July1756
HumptulipsO63547·019771848·00° N 124·00° W July246
ChehalisO439047·019762048·00° N 124·00° W July1419
Lewis River FallsO81645·919642848·00° N 124·00° W July2125
NehalemO172845·719672544·00° N 124·00° W July123
Upper Columbia SpringsS11443445·619393148·00° N 124·00° W JulyNANA
SiletzO52344·919672544·00° N 124·00° W July141
SiuslawO201044·019652744·00° N 124·00° W July−850
DeschutesO110145·619772248·00° N 124·00° W July−332

Model definition

The models of Liermann et al. (2010) were adapted to include three metrics of ocean condition (SST, ENSO and PDO). To provide a baseline for comparison, the base model of Liermann et al. (2010) was implemented, with no attempt made to address temporal patterns in residuals. The next series of models added the ocean covariates and the random walk. Models were evaluated based on their ability to explain temporal residual patterns, overall fit and parsimony. Because the portion of the model linking catchment size to the density-dependent parameter was not of primary interest here, that section of the model was described in a separate section (Liermann et al., 2010).

The spawner–recruit relationship

The spawner–recruit data for each population were modelled using the Ricker (1954) spawner–recruit function:


where Ri,j and Si,j are the ith year's recruits and spawners for population j, and pj and Bj are the density-independent (productivity) and density-dependent parameters, respectively. Specifically, inline image is the slope of the relationship when S = 0, and Bj is the spawner level at which maximum recruitment occurs. The wi,j are normally distributed independent random variables with mean zero and population-specific s.d. σj. The population-specific precision values (inline image were assigned diffuse gamma prior distributions with shape and rate parameters both set to 0·0001 (Gelman et al., 1995). The Ricker (1954) spawner–recruit function was chosen because it tended to give more realistic parameter estimates than other spawner–recruit functions (Myers et al., 1999) and has previously been used extensively with O. tshawytscha data (Pacific Salmon Commission, 1999).

Productivity is defined as:


where X = PDOi,j, Y = SSTi,j and Z = ENSOi,j (the ocean covariates described above), the α are the population-specific slopes relating productivity to the ocean covariates, rj is the population-specific intercept and inline image is the population-specific random walk. The different models that are evaluated include just rj (the base model), rj and the ocean covariates, and rj, the ocean covariates, and the random walk.

The random walk

The random walk is defined as:


The zi,j follow a normal distribution with mean zero and population-specific s.d. inline image. A diffuse gamma prior is assigned to the random-walk precisions, inline image, where the shape and rate parameters are 0·001.

Upper level parameters governing productivity

In the second level of the hierarchical model, the population-specific productivity parameters are assumed to come from distributions described by a second set of parameters (hyperparameters). The population-specific rj and αn,j parameters are assumed to follow normal distributions with mean rM + rDLj (where Lj = life-history type), and αM,n + λnXj + αD,nLj(whereX = latiude), and and αs.d.,n. Here, n takes the values 1 to 3 depending on the ocean covariate. The two means are life-history specific, where rD and aD,n are constants added to the mean for ocean-type populations (Lj = 0 for stream type and Lj = 1 for ocean type). To allow the relationships between the ocean covariates and the productivity to change with latitude, the λn parameters are added. The life-history-specific productivity parameter rD is included in all models based on model assessment in Liermann et al. (2010). The latitude and life-history components for the α are included if assessment of simpler models without these parameters suggests they will improve fit. The s.d. parameters rs.d. and αs.d. are assigned uniform priors on the interval from zero to 100 (Gelman et al., 1995). All other hyperparameters were assigned normal priors with mean zero and large s.d. (31·6). Notice that a s.d. of 31·6 corresponds to a precision (inverse of the variance) of 0·001.

Upper level parameters governing density dependence

The density-dependent parameters log Bj are assumed to follow a normal distribution with mean a + aDLj + (b + bDLj)lnWj and s.d. Bs.d.. Here, Lj is the life-history type (Lj = 0 for stream type and Lj = 1 for ocean type). The intercept and slope for the relationship between log Bj and log catchment size (ln Wj) are therefore a and b for stream-type fish and (a + aD) and (b + bD) for ocean-type fish. The ln Wj values are centred (the mean is subtracted from all values) to aid in computation. The Wj are equal to these centred values exponentiated. All further references to ln Wj and Wj will indicate the centred values. The parameters a, b, aD and bD were assigned normal priors with mean zero and large s.d. (31·6).

Model evaluation

In Liermann et al. (2010), the component of the model relating catchment size to population capacity was shown to fit the data well. The primary focus in this paper was on the density-independent productivity term (equation 2). The residuals from the spawner–recruit relationship (the Ricker model) were plotted against year and spawners, and the productivity parameters were plotted against latitude and life-history type. These graphical assessments at the two levels were formalized using posterior-predictive P-values (Gelman et al., 1996) defined as Pr⌊T(yrep,θ) ≥T(y,θ)|H,y⌋, where y are the data, yrep are data simulated from the posterior distribution, θ are the parameters, H is the model and T is a statistic used to assess fit (the number of runs, for example). Values close to zero or 1 indicate problems with the model fit. Posterior-predictive P-values were calculated for autocorrelation in the residuals.

The relative predictive performance of the different models was evaluated using the deviance (Spiegelhalter et al., 2002) and the number of upper level (population) parameters. More emphasis was placed on model fit and parameter interpretation than predictive performance. As such, this analysis should be viewed as a model-based exploratory analysis.

Monte-carlo markov-chain analysis

Monte-Carlo Markov-Chain (MCMC) integration was used to estimate posterior distributions for the parameters of interest (Gelman et al., 1995). The WinBUGS (Spiegelhalter et al., 1999; and R2WinBUGS (Sturtz et al., 2005) software were used. The two-step regression approach of Parken et al. (2006) was used to assign initial values, and the chain was run for 11 million iterations with a 1 million iteration burn in and thinning to every 1000th draw. The resulting chains were analysed for convergence by inspecting the parameter traces, autocorrelation plots and plots of the parameters against each other to assess cross-correlation. The convergence diagnostics of Geweke (1992) and Heidelberger & Welch (1981), as implemented in the R package CODA (Convergence Diagnostics and Output Analysis; Plummer et al., 2006), were also applied.


The spawner–recruit residuals from Liermann et al. (2010) showed some common temporal trends, with populations closer to each other tending to agree more (Fig. 2). Adding the three ocean covariates to the base model reduced the overall deviance, improving the fit for a number of populations (Tables I and II). Adding the ocean variables, however, did not eliminate the strong temporal patterns in the residuals seen in many of the populations (Liermann et al., 2010), suggesting a lack of model fit (Table II). Adding the random walks to the ocean model (to create the full model) improved the fits substantially with posterior-predictive P-values for autocorrelation falling between 0·9 and 0·1 for all populations except one, and the deviance was reduced almost by half. The distribution of ocean coefficients across populations varied substantially by covariate, with only the SST coefficients showing consistent differences from zero (Fig. 3). Plotting the mean coefficient values of the full model for the three indices against latitude and life-history type (Fig. 4) suggested potential interactions between latitude and the effect of PDO. This may also be confounded by life-history-type as stream-type fish are primarily found in far northern latitudes, whereas ocean-type fish are found further south in the north-west Pacific coast of North America. When this component was added to the full model, the interaction with latitude had posterior probability concentrated above zero for PDO (95% high posterior density interval = 0·002 and 0·026; median = 0·014). The importance of the random walk and ocean effects depended on the population (Table I). In some cases, adding the ocean covariates improved the model fits marginally (<6% change in residual variance, e.g. Situk, King Salmon, Cowichan, Humptulips, Skagit and Nehalem; Table I), whereas in other cases adding the ocean covariate resulted in a large improvement (>17% change in residual variance, e.g. Queets, Quillayute, Unuk, Klukshu and Keta). In most cases (other than Taku and Stikine), adding the ocean covariates and random walk reduced the residual variance by at least 10%. The SST data for the upper Columbia Springs stock were unavailable to analyse and was thus dropped from the final analysis.

Figure 2.

Standardized residuals based on maximum likelihood Ricker model fits plotted against Oncorhynchus tshawytscha brood year. Populations ordered from bottom to top by latitude with latitude (decimal degrees north) following each of the population names (inline image, positive residuals; inline image, negative residuals). The areas of the circles are proportional to the size of the residuals with inline image and inline image corresponding to +1 and +2 s.d., respectively.

Table II.  Model comparison using deviance and model fit using posterior-predictive P-values. The third and fourth columns are the number of Oncorhynchus tshawytscha populations for which the posterior-predictive P-values were outside the range 0·1–0·9 (indicating a lack of fit). The posterior-predictive P-values were for autocorrelation with respect to years and spawners
ModelDeviance0·9 < P-values < 0·1 for years0·9 < P-values < 0·1 for spawners
Base (model 1)853152
Ocean only (model 2)488144
Full (model 3)45711
Figure 3.

Posterior distributions for the ocean coefficients (a) Pacific Decadal Oscillation, (b) sea surface temperature and (c) El Niño Southern Oscillation from the full model (random walk + ocean).

Figure 4.

Mean coefficients of spawner–recruit model fits by life-history-type (inline image, ocean-type populations; inline image, stream-type populations) and (a) Pacific Decadal Oscillation, (b) sea surface temperature and (c) El Niño Southern Oscillation and latitude.

The joint effect of the random walk and ocean variables resulted in substantial ranges in productivity for many populations (Fig. 5). This information could be used on various spatial scales to understand how productivity changes would affect management targets such as spawners at maximum sustainable yield [SMSY; Fig. 5(g)–(i)]. Hence, as productivity of a stock declines, the derived management target also drops substantially as shown for all populations and is unsustainable if productivity drops below 1.

Figure 5.

Temporal trends in Oncorhynchus tshawytscha spawners at maximums sustainable yield, recruitment and productivity with 90% credible intervals (c.i.) in three rivers, (a), (d), (g) Unuk, (b), (e), (h) Cowichan and (c), (f), (i) Skagit. (a), (b), (c) Loge recruits with the base model fit (inline image), the ocean only fit (inline image) and the full model fit (inline image). (d), (e), (f) The temporal trend in productivity (er) along with ±90% c.i. (g), (h), (i) the temporal trends in spawners at maximum sustainable yield (MSY) with 90% c.i.

If both SST and PDO changed two s.d., the full model predicted a change of >30% in productivity averaged across all populations. For most populations, however, the random walk accounted for most of the variability in productivity. Global climate models project a 1° C warming over the north-west coast SSTs of North America over the next century. A 1° C change in SST resulted anywhere from 0·1 to 2 s.d. in the SST anomaly depending on the population, with a median of 1·3. Increasing the SST anomaly in the model by 1·3 resulted in a decrease in productivity of c. 13% averaged across all populations. This was quite small relative to the unexplained variability in productivity modelled by the random walk (Fig. 5).


Numerous papers have discussed the relationship of the negative phases of the ENSO and PDO metrics to positive survival of Oncorhynchus spp. (Hare & Francis, 1995; Hare et al., 1999; Magnusson, 2002). For O. tshawytscha, however, these studies have been limited in geographic extent (Greene et al., 2005; Hinrichsen & Fisher, 2009), relied on hatchery populations (Magnusson, 2002), or used catch data (Hare & Francis, 1995). In the present study, temporal variation in recruitment data from natural O. tshawytscha populations spanning a large proportion of the species' geographic range was modelled using spawning stock size and three metrics previously linked to O. tshawytscha survival. As such, these results are in the currency most valuable to managers of these populations, i.e. recruitment, and are generally applicable to populations throughout the species' range. Linking the populations through a hierarchical model provides a mechanism for moving towards managing at a larger scale. By thinking about the distribution of productivities across populations, and how productivities respond in aggregate and individually to environmental variables, management will probably be more robust. By allowing for productivity to change in predictive models of recruitment, managers are more likely to avoid over-and-under fishing as populations respond to environmental forcing, and a pathway is provided towards a better understanding of how climate change will affect these populations.

As a whole, the three environmental metrics applied in this work explained a relatively small proportion of the previously unexplained recruitment residuals. The model results, however, generally agree with the literature, and the predicted productivity changed by >30% in response to moving from low to high values of the covariates. The coefficients relating SST to productivity were all predicted to be negative, while for the PDO covariate the coefficients were predicted to be negative for lower latitudes, moving towards zero or slightly positive for northern populations. This maybe confounded with life-history type and should be investigated further with new data. The latitudinal response, however, is consistent with other studies (Hare & Francis, 1995; Hare et al., 1999) and the single northern latitude ocean-type population in the present study has PDO coefficient close to zero, thereby providing additional support for this hypothesis. The substantial temporal trends in the recruitment residuals that remained after including the ocean covariates could be effectively accounted for with the random-walk models (i.e. temporal trends in residuals are no longer evident). While the random walk does not provide an environmental mechanism, it can be useful for prediction since changes in the random walk from one year to the next are relatively small.

Potential methods of improving model prediction include the construction of metrics that more closely match the location and scale at which smolts from specific populations experience the marine environment (Levin, 1992) and building a better understanding of the physical and biological interactions (Mackas et al., 1985) that affect O. tshawytscha smolts during early ocean residence. In a study of O. tshawytscha stocks in the Skagit River, Greene et al. (2005) related interannual changes in recruitment to a set of carefully selected predictors based on a detailed understanding of freshwater life history and when smolts in this population were in transition between the riverine, estuarine, nearshore and ocean environments. Metrics such as frequency and magnitude of flood events, estuary condition and nearshore ocean conditions during fish residence together explained >70% of the annual variability in ln recruits per spawners. Peterson et al. (2006) examined the interrelationships between several physical and biological indices including the PDO, an ENSO metric, water temperature, copepod species composition and Oncorhynchus spp. survival. They found, for example, that Oregon (U.S.A.) hatchery O. kisutch survival tended to be higher in years when the relative abundance of coldwater copepods was higher, possibly because such zooplankton includes species with high fat contents. Copepod species composition was in turn shown to be related to water temperature and the PDO index. While data better tailored to specific populations and an improved understanding of the ecology driving O. tshawytscha survival would be valuable contributions, managers can only use data that are available. Moving from theoretical models and case studies will require difficult empirical work that includes the meticulous preparation of disparate sources of data, as was the case for the spawner–recruit data sets and environmental covariates used in this analysis, and the data in the studies cited above.

While more spatially and biologically appropriate ocean metrics will undoubtedly improve the predictive models, some of the variability in recruitment residuals is due to noise introduced from substantial uncertainty in population-specific estimates of escapement, harvest rates, age composition and mortality, and through model mis-specification. For some populations, regional harvest rates are used. This could lead to spurious correlation in temporal patterns that could be misinterpreted as a response to common environmental effects. A better understanding of the aggregate effect of these errors on population estimates would provide an understanding of how much predictive improvement is possible, and a more realistic context in which to interpret model results.

The coastwide management of O. tshawytscha is dependent on models of recruitment. In particular, spawners at maximum sustainable yield (SMSY) is used as an index for setting management targets. Current models of recruitment largely assume constant productivity leading to the potential for overfishing populations during periods of low productivity. It is shown here that the spawner–recruit data for 23 studied populations tend to support the literature as described above, suggesting that productivity varies with time. The model-predicted patterns in productivity in turn lead to substantial patterns in SMSY. To the degree that models such as those presented in this study can predict these changes in productivity, managers can adapt to different regimes to avoid overfishing and to take advantage of years of high productivity. In addition, projected changes in oceanic conditions due to global climate change can be used to assess the potential effects of such environmental changes on populations of O. tshawytscha.

This study attempted to relate phenomena across different temporal and spatial scales to recruitment for O. tshawytscha in the north-west Pacific coast of North America, building on the work of Sharma (2009) and Liermann et al. (2010) to understand variation in recruitment anomalies over large temporal and spatial scales. This information could be of vital importance to the overall viability of the species as it can be used to formulate early predictions of O. tshawytscha survival coastwide and incorporated into an adaptive management framework based on these projections. Principles of precautionary management (FAO, 1996; Richards & Maguire, 1998) could be used as guidelines and the risk to the resource could be minimized by setting lower fishing targets when conditions are poor ensuring that spawning biomass would not be threatened across all areas being affected by the ocean fisheries. Finally, climate change patterns that would warm the ocean in the future could pose a threat to the overall viability of O. tshawytscha in the west coast of the U.S.A. and Canada. If managers of such populations are to prepare adequately for this eventuality, predictive models such as those presented in this study will be essential.

The authors would like to thank D. Graves [Columbia River Inter-Tribal Fish Commision (CRITFC), Portland, OR] for generating the map in Fig. 1. In addition, the authors would like to acknowledge C. Parken (Department of Fisheries and Oceans Canada) for sharing the spawner–recruitment data with us, P. Roni (National Oceanic and Atmospheric Administration) for reviewing this and two anonymous reviewers for very helpful reviews in improving the manuscript, as well as the Guest Editor, I. Winfield for making substantial effort in improving the clarity of this manuscript. Finally, the authors would like to thank Bonneville Power Administration (BPA) for providing partial funding for this work.