Climate and population density drive changes in cod body size throughout a century on the Norwegian coast

Abstract

Understanding how populations respond to changes in climate requires long-term, high-quality datasets, which are rare for marine systems. We estimated the effects of climate warming on cod lengths and length variability using a unique 91-y time series of more than 100,000 individual juvenile cod lengths from surveys that began in 1919 along the Norwegian Skagerrak coast. Using linear mixed-effects models, we accounted for spatial population structure and the nested structure of the survey data to reveal opposite effects of spring and summer warming on juvenile cod lengths. Warm summer temperatures in the coastal Skagerrak have limited juvenile growth. In contrast, warmer springs have resulted in larger juvenile cod, with less variation in lengths within a cohort, possibly because of a temperature-driven contraction in the spring spawning period. A density-dependent reduction in length was evident only at the highest population densities in the time series, which have rarely been observed in the last decade. If temperatures rise because of global warming, nonlinearities in the opposing temperature effects suggest that negative effects of warmer summers will increasingly outweigh positive effects of warmer springs, and the coastal Skagerrak will become ill-suited for Atlantic cod.

Climate change poses one of the most serious challenges to sustainable marine fisheries (1). For many temperate species, warming is predicted to improve conditions for growth and survival at higher latitudes but be detrimental to lower-latitude populations, where organisms may already be experiencing conditions close to physiological limits (2). Indeed, for the commercially important Atlantic cod (Gadus morhua), past warming has negatively affected recruitment for some southerly stocks (3), leading to projections of stock collapse with future climate warming (4). In addition to affecting recruitment, climate can affect phenotypic traits, such as body size, through plastic responses to environmental change and evolutionary processes (5, 6). However, long time series of phenotypic trait values are rare, making it difficult to distinguish climatic effects from other intrinsic and extrinsic factors, such as density dependence or harvesting (7, 8). Here, we present a unique 91-y time series of individual juvenile lengths from surveys conducted along the Norwegian Skagerrak coast to assess how climate and density dependence affect length distributions in Atlantic cod.

At the population level, length distributions reflect individual variation in hatching date, growth rate, and size-selective mortality, all of which may vary with temperature in complex ways (7, 9–13). For instance, warmer temperatures increase the metabolic scope for growth but only up to a thermal optimum, which depends on food abundance (14), life stage (15), and population (9). Above this optimum, warming is expected to reduce growth rates (9). Also, the temperature regime experienced by young cod will vary seasonally, and warming at different times of the year may have different effects on cod growth, spawn timing, and size-specific survival (1). Juvenile growth and body size are key phenotypic traits that are likely to influence population dynamics through effects on survival, maturation, fecundity, and offspring quality (12, 13, 16, 17). Hence, it is important to understand how body size is influenced by climate change.

Our data are 91 y of length measurements from beach seine surveys for juvenile cod conducted for scientific purposes by the Norwegian Institute of Marine Research, giving us the advantage of high-quality, long-term fishery-independent data on a prerecruit life stage. Sampling is conducted every autumn at an average of 110 stations spread across 250 km of the Norwegian Skagerrak coast (Fig. 1) (18) and targets cod hatched the previous spring (referred to as age 0). The Flødevigen surveys are especially remarkable for their consistency: only three scientific personnel have led the surveys since 1919, each of whom overlapped with another for at least 5 y to ensure consistency in methods, station locations have been marked exactly, and the wooden boat used today is a replica of the one used in 1919. The near-century time frame of this study includes extended warm and cool periods in Northern European climate (1).

Fig. 1.

Beginning in 1919, juvenile cod were measured annually at an average of 110 stations along the Skagerrak coast of southeast Norway. (A) Black dots show beach seine stations included in this study grouped into 15 fjord groups. Sea temperatures at 1 m were recorded near Flødevigen Research Station in region 5. (B) Time series of mean age-0 cod lengths (±1 SD) are shown for 3 of 15 fjord groups (Fig. S1 shows time series from all regions).

Previous work on Skagerrak cod has shown inter- and intracohort density dependence in population dynamics (18), but studies of the effects of climate and density on length distributions have been less conclusive (19, 20), possibly because they have not accounted for spatial population structure. Since the early days of this survey, the spatial structuring of populations within and among fjords has been evident (21), and recently supported by genetics (22, 23) and life-history studies (17, 24). This structure is maintained through retention of early life-history stages (eggs and larvae) (25) and limited juvenile and adult dispersal (26, 27). Spatial population structure can provide stabilizing benefits to fisheries and ecosystems (28, 29) but is not always considered in management of fish stocks or incorporated into ecological studies. By using mixed-effects models, we are able to simultaneously model changes in the mean and variance of age-0 length distributions while accounting for potential spatial diversity. We first quantify how the mean and variance in age-0 length have responded to changes in climate and cod density over the past nine decades, and second, we ask whether there is evidence for spatial structuring in age-0 length distributions along the Skagerrak coast.

Results and Discussion

Survey stations were initially grouped into 15 fjord groups along the Southeast Norwegian coast based on geographic proximity and shoreline features (Fig. 1 and Fig. S1). Between 1919 and 2009, the lengths of 127,540 individual cod were measured. Mixture models (30) (Methods) applied to data grouped by fjord group and year allowed us to classify 102,626 of these individuals as likely age-0 cod based on length distributions. The temporal extent of the time series varied between fjord groups, with new regions added to the survey in 1936, 1953, and 1962 (Fig. 1 and Fig. S1); however, balanced data are not required for mixed-effects models (31).

Seasonal Effects of Temperature.

We detected clear effects of temperature on length distributions of cod throughout the Skagerrak. Comparison of linear mixed-effects models with different covariates resulted in a model that included highly significant but opposite effects of spring and summer sea surface temperatures on age-0 cod length [natural logarithm (ln) transformed] (Table 1). Length in the fall was positively associated with the average temperature during the spawning season in the spring (February to April) but negatively associated with summer temperatures (July to September), when juveniles have settled in near-shore demersal habitat. The estimated effects translate to a 3.1% increase in length with every 2 °C increase in spring temperatures and a 3.1% decrease in length for every 2 °C increase in summer temperatures. Spring and summer temperatures were only moderately correlated (r = 0.33, n = 91 y).

Table 1.

Restricted maximum likelihood (REML) parameter estimates for the best model of effects on age-0 length [ln(cm)]

Model component	Estimate	SE	DF	t value	P value	95% CI
Fixed effects
Intercept	2.276	0.0160	96,622	141.7	<0.0001
ExpL (0, 1)	−0.0375	0.0068	4,886		<0.0001*
ExpH (0, 1)	0.0070	0.0139	4,886
SpringSST (°C)	0.0157	0.0028	1,096	5.6	<0.0001
SummerSST (°C)	−0.0154	0.0034	1,096	−4.5	<0.0001
SDens [ln(N)]	−0.0051	0.0017	4,886	−2.9	0.0033
Day	0.0021	0.0005	4,886	4.1	<0.0001
Year	−0.0005	0.0004	1,096	−1.1	0.2583†
Random effects (SD units)
σ0 (year slope by f)‡	0.0014					0.0009–0.0023
σ1 (f)	0.0562					0.0372–0.0847
σ2 (y in f)	0.0876					0.0819–0.0937
σ3 (s in y in f)	0.1053					0.1020–0.1087
σ (ε)§	0.2241					0.2215–0.2268

Covariates were centered by subtracting the mean before inclusion in the model. DF, degrees of freedom.

*The significance of exposure level was tested as F_2,4886 = 17.377 and P < 0.0001.

^†A fixed effect of year is included, although not significantly different from zero, to allow for the highly significant random year response at the fjord-group level.

^‡Nested groups (fjord group, year, and station) for the random effects are indicated by f, y, and s, respectively.

^§Error variance was an exponential function of SpringSST (Eq. 2). Estimated δ = −0.0410 (95% CI = −0.0444 to −0.0375). See Table S1 for a comparison with other variance covariates.

We also found strong evidence for changes in length variation tied to spring temperatures (Table 1 and Table S1), with warmer springs resulting in narrower age-0 length distributions the next autumn. This was determined by modeling the within-station residual variance in lengths as a function of potential covariates (Methods) (31, 32). Other single-covariate variance models were also supported over the null model, indicating some evidence for decreased trait variation with increased cod density, increased summer temperatures, and through time (Table S1). However, the model with SpringSST far outperformed the others in terms of Akaike's information criterion (AIC) (Table S1). The estimated exponential variance parameter (δ) corresponds to a 15.1% decrease in variance with a 2 °C warming in SpringSST.

Combining these results, we document that warmer springs resulted in larger juvenile cod the next autumn, with less variation in lengths within a cohort (Fig. 2). Temperature-driven changes in the onset and duration of spawning could affect the mean and variance in juvenile lengths by determining the length of time juveniles have to grow before they are sampled. In some populations of Atlantic cod, spawning occurs earlier in warm years (10, 33), presumably because of accelerated maturation of gonads (11). The pattern observed in Skagerrak cod is consistent with these observations. Likewise, the variance in juvenile cod lengths could reflect variation in spawn timing within and among individuals, suggesting that cod spawning may occur over a shorter period in warm years (34). Alternatively, the smallest or latest hatched larvae may have poor survival in warm springs because of a mismatch with the timing of zooplankton production (35) or size-specific mortality caused by other factors such as cannibalism or competition with larger age-0 cod (36). An apparent narrowing of length distributions with warmer springs could also occur if juvenile cod exhibit a size-based shift in habitat preferences and larger fish migrate out of the survey area. Unfortunately, few data exist on movement at this life stage, but evidence suggests that individuals up to 20 cm defend near-shore territories for several months after settling (36).

Fig. 2.

Observed age-0 lengths during the coldest (gray dashed bars) and warmest (open bars) quartiles of (A) SpringSST and (B) SummerSST.

Compared with spring temperatures, summer temperatures had the opposite effect on mean age-0 length in the autumn. The negative relationship between summer temperatures and cod length is counter to previous empirical studies, which have found that warmer temperatures tend to increase growth rates in Atlantic cod (8, 37, 38). The optimum temperature for growth (T_opt.G) of juvenile Atlantic cod is likely between 12 °C and 15 °C depending on size (table 2 in ref. 15), and as temperatures warm above T_opt.G, growth rates decline rapidly (9, 15). Mean summer temperatures in the coastal Skagerrak are often well above optimum for juvenile growth [mean July to September temperatures at the surface and 19 m were 15.9 °C (n = 91 y) and 14.6 °C (n = 49 y), respectively] (Fig. 3); thus, it is likely that coastal Skagerrak cod are experiencing decreased growth rates because of high metabolic costs in warm summers. Further analysis with generalized additive models (GAMs) supports this hypothesis, showing an increasingly negative effect of temperatures beginning above ∼16 °C (Fig. 4). Temperature may also affect growth indirectly through effects on prey communities. In the North Sea, warming temperatures have decreased the quantity and quality of plankton prey for cod (39); therefore, food limitation may be more likely in warm summers, exacerbating the physiological effect of warm temperatures on growth.

Fig. 3.

Average spring (A) and summer (B) sea temperatures recorded near Flødevigen Research Station at 1-m (black circles and lines) and 19-m (gray circles) depths. Temperatures for 1919–1923 (black crosses) are reconstructed based on a regression with Torungen SST (Spring R² = 0.94, n = 78; Summer R² = 0.92, n = 80). Temperatures are only shown if data exist at that depth for at least 80% of the days during the season.

Fig. 4.

Partial plots of GAM responses, with 95% confidence intervals (CIs) obtained by a modified wild bootstrap approach. Linear effects were included as follows (estimate with bootstrapped 95% CI): day [1.7 × 10⁻³ (−2.7 × 10⁻⁴ to 3.6 × 10⁻³)] and year [−9.5 × 10⁻⁴ (−6.7 × 10⁻⁴ to 5.0 × 10⁻⁴)]. Exposure and fjord group were included as factors.

An alternative explanation for the observed decrease in mean length associated with warm summers could be a behavioral response. If larger juvenile cod moved from the shallow near-shore zone to adjacent deeper parts of the fjord basins to avoid high temperatures, beach seine catches would be biased toward smaller individuals, producing an apparent trend in size with temperature. If this were the case, then autumn lengths should be more strongly associated with temperatures during the beach seine survey (SurveySST) than SummerSST, which was not the case (ΔAIC = 10, including all other final model covariates). Thus, a short-term size-biased behavioral shift seems unlikely as an explanation for the negative relationship between length and summer temperatures.

Density-Dependent Growth.

Climatic effects often occur simultaneously with intrinsic population processes, such as density-dependent growth. Cod density, measured as the station-specific catch of cod ≤20 cm to represent the local density of potential competitors, had a significant negative effect on length (Table 1). Although the causal basis for this association merits further research, this result matches predictions for density-dependent growth, which has previously gone undetected when analyzing this system as a whole and not accounting for local spatial structure (19, 20). Density can affect growth of juveniles through competition for high-quality nursery habitat (36). At low densities, before nursery habitat is saturated, we would not expect competition to reduce growth rates, and indeed, further analysis using GAMs indicated a threshold-type response, with no apparent effect of density up to ∼50 cod per station, after which there was a strong negative effect on length (Fig. 4). On average, only 5% of beach seine tows have contained >50 cod. This has dropped to less than 1% in the last decade, indicating that growth is only rarely constrained by density. We also modeled the effects of average regional cod density (average catch by fjord group) but found that regional density estimates had a weaker relationship with length than station-specific estimates (ΔAIC = 6, including all other covariates). This underlines the importance of considering density–length relationships at the appropriate spatial resolution, because the relationship may become harder to detect when data are aggregated over large spatial scales, as is common with fisheries data.

Spatial Structure.

Olsen et al. (20) found no significant effect of summer temperature or population density on the mean length of age-0 Skagerrak cod. Likewise, Lekve et al. (19) failed to detect an effect of density or summer temperature on length, although they did report a positive effect of spring temperatures. Our ability to detect these relationships was greater for three likely reasons. First, we included data from a broader set of stations along the coast. Second, we considered density effects on local scales, rather than a coast-wide scale, which likely better matches the scale at which individual juvenile cod experience their environment. Third, we allowed for spatial heterogeneity in length dynamics, which is more realistic for coastal cod populations with proven genetic structure (23) and spatial differences in life-history traits (17, 24).

Spatial variation in lengths partly corresponded to differences in the exposure of survey stations to the open Skagerrak. A fixed effect of station exposure (Exp) in the final model indicated that cod in protected inner fjords were smaller, on average, than cod in more exposed areas at the mouths of fjords (Table 1). This matches previous findings (19, 40) and could be because of differences in the timing of juvenile settlement, feeding conditions, or influence of larval cod advected from the North Sea to the outer Skagerrak coast (41). Evidence of genetic differentiation between larval cod in exposed and sheltered regions (42) makes it unlikely that this pattern is because of a seaward movement of larger individuals.

The random effects in the mixed-effects models indicated further unexplained spatial structure in juvenile cod lengths. AIC-based model selection (Methods) strongly favored a nesting structure for the random effects with station nested within sampling year nested within fjord group (Table S2). Fjord group was supported as the outermost grouping factor, rather than year, indicating persistent spatial structure in the data. Estimated random intercepts for fjord groups show that cod in regions 8 and 10 were relatively small, whereas cod in nearby region 9 were substantially larger than in any other region in the Norwegian Skagerrak (Fig. 5A). There was also spatial heterogeneity in long-term trends, as indicated by a highly significant random effect of year as a continuous variable at the fjord-group level (Table 1 and Fig. 5B). These roughly fell along a latitudinal gradient, with regions in the north generally decreasing in length through time and regions in the south generally increasing in length.

Fig. 5.

Estimated coefficients showing regional differences in mean length and long-term time trends. (A) Fjord group-specific intercepts (β₀ + b_1f) are shown, transformed from the natural logarithmic scale into centimeters. The transformed intercepts represent the predicted length given the mean values of all covariates and intermediate level of exposure. (B) Linear time trends are transformed to show the percentage change in length over 90 y ((exp((β₈ + b_0f) × 90) − 1) × 100). The causal basis for the regional differences in cod length remains a topic for future research.

The spatial variation in mean lengths and length trends could have both a genetic and an environmental component. Previous studies have found genetic substructure among adult cod from different fjords (22, 23) matched to the scale of spatial variation in life-history traits (17, 24). Thus, genetic differences in growth rates may underlie the spatial variation in length distributions. It is also likely that conditions for growth and survival vary along the coast. For instance, fjord group 11 has the largest estimated decreasing trend in length (Fig. 5B) and is among the most industrialized and polluted regions (43). Salinity and oxygen concentrations may also affect cod growth; however, we lack the appropriate data to include these in our study. Furthermore, spatial variation in length and length trends could be caused by persistent local temperature deviations from the sea surface temperature (SST) measured at Flødevigen. Temperature records throughout the Skagerrak and North Sea are highly correlated in their interannual variability ( among five Skagerrak SST records) (Table S3) (44), but individual fjords may be warmer or colder on average, subsequently affecting local cod growth rates. Spatial variation and trends in the abundance of larger cannibalistic cod, because of differences in fishing pressure or otherwise, could affect age-0 size distributions through size-selective mortality. Being unable to distinguish between the possible mechanisms underlying spatial variation in cod lengths, we offer these as hypotheses for future studies.

Conclusions

We have shown that Atlantic cod along the Norwegian Skagerrak coast are clearly affected by climate-associated changes in their thermal environment as expressed by changes in length distributions, complementing work on other stocks that shows climatic effects on recruitment (3, 39, 45). Although the demographic consequences of these changes in length distributions are unknown, many ecological processes depend on size or growth rate. For instance, overwinter survival in temperate fishes is often poorer for smaller individuals (12). Thus, survival rates to age 1 may depend on the size of age-0 cod before winter, which is likely to be smallest after a cold spring and a warm summer. Reproductive output is size-dependent, with larger females producing more viable offspring (16). Finally, fish with slower growth rates may mature later (17), increasing their exposure to natural and fishing mortality before spawning. Therefore, the observed climate-driven changes in size distributions could ultimately affect the dynamics of Skagerrak cod populations.

Downscaled climate models forecast mean SSTs in the Skagerrak to increase by 2–3 °C in the next century (46). The effects of this temperature increase on size distributions of cod will depend on the seasonality of warming and how close populations are to their thermal limits (1). Our study indicated that cod in the coastal Skagerrak are already experiencing summer temperatures above the optimum for growth. A similar rate of warming in the spring could partly offset the effects on mean length, but nonlinearities in the temperature response (Fig. 4) suggest that negative effects of warmer summers will increasingly dominate over positive effects of warmer springs at higher temperatures. Furthermore, exposure to high temperatures takes a physiological toll, and mortality rates increase after temperatures surpass optimum (9). Although age-1 and older cod may behaviorally thermoregulate by moving to colder waters (47), age-0 juveniles are typically strongly site-attached with small (10–1,000 m²) shallow-water home ranges (36). Thus, unless they are able to adapt to warmer temperatures, cod may no longer find suitable habitat along the Norwegian Skagerrak coast, with consequences for the ecology and economy of this coastal region. We hypothesize that other temperate species at the lower-latitude edges of their ranges may show similar season-dependent temperature responses, which may be revealed through analyses at the appropriate temporal and spatial scales.

Methods

Separating Age-0 from Age-1 Cod.

At each beach seine station, the total lengths of all individual cod were measured to the nearest centimeter, or in the case of very large catches, a random subset of ∼100 cod was measured and total catch was recorded (see ref. 18 for a detailed description of the Flødevigen monitoring program). Age-0 cod were separated from age-1 cod based on a mixture analysis (30) applied to length-frequency distributions. Because this was not possible for each station in each year because of small sample sizes, we grouped the stations into fjord groups based on fjord or geographic region (Fig. 1). Assuming the observed lengths represent a mixture of age-0 and age-1 cod, we fit two log-normal distributions to the data for each fjord group in each year to determine the probability (p(age 0)_yfl) that a fish of a given length (l) in year (y) and fjord group (f) was age 0. If there was not a clear age-1 mode, we assigned p(age 0) as equal to one for all fish ≤18 cm (the mean length at which an estimated 50% of fish in all samples pooled by year were age 0). We drew a random number from a U(0, 1) distribution for each individual fish, and if the random number was less than p(age 0)_yfl, we retained that fish for our analysis. Comparison of final model parameter estimates for datasets constructed using alternate random draws indicated that the final results were invariant to the stochasticity introduced here.

Length Covariates.

Sea surface temperatures for February to April (SpringSST) and July to September (SummerSST) were averaged based on daily observations at 1-m depth in the bay near the Flødevigen Research Station (Fig. 1), where temperatures have been measured consistently since 1924. Temperatures were also measured at 19-m depth, which may better represent conditions experienced by spawning adults, but the record had significant temporal gaps. Interannual fluctuations in mean February to April temperatures at 1 m and 19 m were highly correlated (r = 0.95, n = 62 y); thus, we used the 1-m record to retain the greatest temporal coverage. Summer temperatures at 1 m and 19 m were less correlated because of thermal stratification (r = 0.65, n = 49 y). Sea temperatures measured at Flødevigen were highly correlated (r ≥ 0.94) with records from elsewhere in the Skagerrak (Table S3), and we reconstructed seasonal SSTs for 1919–1923 based on regressions with historical temperatures from nearby Torungen Lighthouse (Fig. 3). Because cod recruitment and size have previously been linked to the North Atlantic Oscillation (NAO) (19, 45), we also considered the winter (December to March) principal component-based NAO index (48) (http://www.cgd.ucar.edu/cas/jhurrell/indices.data.html#naostatdjfm), corresponding to the months just before and during spawning. SpringSST was correlated with the NAO index (r = 0.76, n = 91 y), and NAO was not included in the final model, indicating that local temperature measurements were a better predictor of length dynamics in the Skagerrak. SurveySST was the average 1-m temperature at Flødevigen from September 8 to October 11, when the beach seine survey occurs.

Cod density was calculated for each year at two spatial resolutions: total catch per station (SDens) and average catch among stations within fjord groups (GDens). To describe the density of potential competitors of age-0 cod, we used the catch of individuals ≤20 cm, which, in some years, included individuals estimated to be age 1. Density was log-transformed before inclusion in models.

Spatial variation in length has previously been reported in this system (19, 49), particularly related to the degree of exposure of the site to the open Skagerrak. Stations were assigned to one of three exposure levels: low (fjords or sheltered fjord-like areas), medium (partly sheltered by islands), or high (exposed open coast).

We accounted for variation in the date of sampling by including day of the year (January 1 = 1) as a covariate (Day). In the last two decades, most stations were surveyed within the same 20-d period, but sampling day was more variable in earlier years. Finally, Year was included to capture any unexplained long-term trends in length.

Modeling.

We used linear mixed-effects models (31) to model changes in the mean length as well as variance in lengths using the nlme library (50) in R (51). Nested random effects accounted for the lack of independence among individual fish lengths within a station, within a fjord, and within the same year. Our response variable was natural log-transformed length, and our candidate models included both fixed and random effects, such as the following (Eq. 1)

where ln(l_yfsi) is natural log-transformed body length of individual i within station s, fjord group f, and year y, βs are fixed-effects coefficients, and ExpH and ExpL are levels of exposure (coded 0 or 1) such that β₁ and β₂ represent the differences in ln(l) between medium and high exposures and medium and low exposures, respectively. Year represents years from 1919 to 2009, allowing for a linear trend through time. The bs are normally distributed random fjord group, year, and station effects, respectively. The random-effects structure shown above is one example of the multiple structures considered, including alternative orders of nesting and structures with random slopes for each covariate at the fjord-group level.

The variance in residual lengths (ε_fysi) was modeled as a function of possible covariates within the mixed-effects modeling framework; for example (Eq. 2),

where δ describes the estimated change in variance with SpringSST. We compared this with models of the residual variance varying exponentially with SummerSST, SDens, and Year, and by exposure and fjord group (SI Text).

Model selection proceeded in an iterative manner. First, the full suite of fixed effects was included in the model as above, and different random-effects structures were compared using Akaike's information criterion (AIC) (31) with restricted maximum likelihood (REML) parameter estimation. With the selected random-effects structure in place, we then selected which fixed effects to retain using AIC with maximum likelihood (ML) parameter estimation. Using the most parsimonious fixed-effects structure, we then reassessed the random-effects structure and iterated until settling on a final model. Inspection of standardized residuals (raw residuals divided by the corresponding SEs) and random-effects estimates from the final model did not reveal any evident outlier observations or strong departures from normality.

Finally, to detect possible nonlinearities in how cod length is influenced by climate and density, we fit a GAM (52) with the final selected fixed effects from the linear mixed-effects model, here modeled as smooth functions (splines) to allow for more complex responses. It was not possible to fit a generalized additive mixed model because of the large size of our dataset. Day and Year were retained as linear effects, and fjord group was added as a factor to account for spatial patterns in the mean length. We limited the number of knots to three (SpringSST and SummerSST) and four (SDens) to reflect plausible biological responses. GAMs were fit using the mcgv library of R (51, 52), and confidence intervals were calculated using a modified wild bootstrap approach that accounted for within-year correlation and heteroscedasticity in the residuals (45).