The demographic impact of extreme events: stochastic weather drives survival and population dynamics in a long-lived seabird

Authors


*Correspondence author. E-mail: mfr@dmu.dk

Summary

  • 1Most scenarios for future climate change predict increased variability and thus increased frequency of extreme weather events. To predict impacts of climate change on wild populations, we need to understand whether this translates into increased variability in demographic parameters, which would lead to reduced population growth rates even without a change in mean parameter values. This requires robust estimates of temporal process variance, for example in survival, and identification of weather covariates linked to interannual variability.
  • 2The European shag Phalacrocorax aristotelis (L.) shows unusually large variability in population size, and large-scale mortality events have been linked to winter gales. We estimated first-year, second-year and adult survival based on 43 years of ringing and dead recovery data from the Isle of May, Scotland, using recent methods to quantify temporal process variance and identify aspects of winter weather linked to survival.
  • 3Survival was highly variable for all age groups, and for second-year and adult birds process variance declined strongly when the most extreme year was excluded. Survival in these age groups was low in winters with strong onshore winds and high rainfall. Variation in first-year survival was not related to winter weather, and process variance, although high, was less affected by extreme years. A stochastic population model showed that increasing process variance in survival would lead to reduced population growth rate and increasing probability of extinction.
  • 4As in other cormorants, shag plumage is only partially waterproof, presumably an adaptation to highly efficient underwater foraging. We speculate that this adaptation may make individuals vulnerable to rough winter weather, leading to boom-and-bust dynamics, where rapid population growth under favourable conditions allows recovery from periodic large-scale weather-related mortality.
  • 5Given that extreme weather events are predicted to become more frequent, species such as shags that are vulnerable to such events are likely to exhibit stronger reductions in population growth than would be expected from changes in mean climate. Vulnerability to extreme events thus needs to be accounted for when predicting the ecological impacts of climate change.

Introduction

The Earth's climate is changing rapidly, and there is an urgent need to predict the ecological consequences of ongoing and future climate change, including impacts on growth rates and extinction probabilities of wild populations (Clark et al. 2001; Sutherland et al. 2006). To date, most such predictions have focused on the effects of changes in mean values of various climate parameters (Frederiksen et al. 2004; Thomas et al. 2004). However, under most scenarios for future climate change, environmental variability is expected to increase and extreme events are expected to become more common (Solomon et al. 2007). Thus, we need to know how this will affect populations, and whether and how species can adapt. Because population growth is a multiplicative process, increasing between-year variability in demographic parameters and hence annual growth rate will inevitably lead to a reduction in the long-term growth rate, even with no change in mean parameter values (Lewontin & Cohen 1969). Natural selection therefore should favour reduced variability in those fitness components (demographic parameters) that are most tightly linked to asymptotic population growth rate (that have the highest sensitivity/elasticity), a process termed environmental canalization (Gaillard & Yoccoz 2003; Morris & Doak 2004).

To understand patterns, causes and consequences of temporal variability in fitness components, we need to be able to measure it accurately. Reliable estimation of temporal variability requires separation of sampling variance, which in this context is a nuisance parameter, from the underlying ‘process variance’, the true variance at the population level. The best framework for this separation is a mixed (hierarchical) modelling approach, where the variance term for a random annual effect estimates temporal process variance (Gould & Nichols 1998; Loison et al. 2002; Altwegg et al. 2006). Similarly, identification of temporal environmental covariates of demographic parameters is best done in a hierarchical framework (Loison et al. 2002), because other methods suffer from inflated power when temporal variability is pronounced (Link 1999). Robust methods for estimating process variance and identifying temporal covariates, for example using capture–mark–recapture statistics, have been developed only recently and are rarely used, and more detailed analyses of existing long-term demographic data are needed to build up a general understanding of the extent, causes and consequences of temporal variation in demographic parameters.

In long-lived organisms, population growth rate is more sensitive to variation in adult survival than in fecundity-related fitness components (Lebreton & Clobert 1991). Most such organisms are characterized by relatively stable population size, and in accordance with the environmental canalization hypothesis, adult survival varies relatively little between years (Gaillard, Festa-Bianchet & Yoccoz 1998; Sæther & Bakke 2000). Most seabirds fit this pattern, and population change tends to be slow (Croxall & Rothery 1991; Weimerskirch 2002). However, breeding populations of many cormorant species (family Phalacrocoracidae) are prone to periodic crashes, caused by large-scale mortality or non-breeding events (European shag, Phalacrocorax aristotelis (L.): Potts, Coulson & Deans 1980; Aebischer 1986; Harris & Wanless 1996; Brandt's cormorant, Phalacrocorax penicillatus: Boekelheide & Ainley 1989; Nur & Sydeman 1999; Guanay cormorant, Phalacrocorax bougainvillei: Duffy 1983). Cormorants typically also have higher potential fecundity than most other long-lived birds (Weimerskirch 2002), and under favourable environmental conditions, populations can grow by up to 20% per year (Frederiksen, Lebreton & Bregnballe 2001). It is unclear whether this suite of demographic traits, often thought to be adaptations to a highly variable environment (Nur & Sydeman 1999; Weimerskirch 2002), will buffer these species against further increases in environmental variability, or whether a higher frequency of environment-related large-scale mortality events will increase extinction risk. Resolving this uncertainty requires detailed analyses of long-term demographic data covering a wide range of environmental conditions.

Here, we use 43 years of ring-recovery data to examine temporal variability in juvenile, immature and adult survival of European shags (hereafter shags) on the Isle of May in eastern Scotland. The main aims of this paper are to (i) quantify and compare temporal process variance in survival for different age classes; (ii) identify environmental factors driving temporal variation in survival; and (iii) evaluate the impact of extreme mortality events on population dynamics.

Methods

study site and field methods

The Isle of May (56°11′ N, 2°33′ W) is situated in the outer Firth of Forth, eastern Scotland, ≈8 km from the mainland. The population dynamics and demography of shags at this colony have been studied in detail over many years (Aebischer 1986; Aebischer & Wanless 1992; Harris et al. 1994b, 1994a; Harris & Wanless 1996; Harris, Wanless & Elston 1998). Since 1961, the number of occupied shag nests has fluctuated between 259 and 1916, with pronounced crashes in 1975–76, 1993–94 and 2004–05 (Fig. 1). Some of these fluctuations were due to non-breeding events, for example 1975–76 (Aebischer 1986), 1993 (Harris & Wanless 1996) and 1999 (unpublished data), while others were linked to major mortality events, for example 1994 (Harris & Wanless 1996). Shags are inshore foragers and always spend the night on land (Daunt et al. 2006). During the breeding season, Isle of May shags forage both around the island and along the adjacent mainland coast (Wanless, Harris & Morris 1991). At other times of the year their distribution is still centred on the colony with similar numbers dispersing both north and south along the UK east coast, juveniles and immatures dispersing greater distances, on average, than older birds (Harris & Swann 2002).

Figure 1.

Counts of European shag nests on the Isle of May, 1961–2006. No counts were made in 1967, 1968 and 1970–72.

Adult shags (≥2 years old), as well as unfledged chicks, have been ringed on the Isle of May with hard-metal British Trust for Ornithology (BTO) rings since 1963. Unique colour rings were first introduced in 1981 for adults, and in 1997 for chicks. Many birds colour-ringed as adults were originally metal-ringed as chicks, but here these birds are treated as released in the year they were colour-ringed. The extensive ringing effort has resulted in large numbers of live recaptures and resightings at the colony, as well as dead recoveries. In this study, we focused on quantifying temporal variability and identifying environmental covariates of survival; we therefore used only dead recoveries in order to obtain the longest time series possible without unduly complicating the analysis. Ringed chicks recovered as dead before fledging were not included in the data set. A total of 28 221 individuals were released: 19 168 metal-ringed chicks, 2590 metal-ringed adults, 5564 colour-ringed chicks and 899 colour-ringed adults (436 of these were originally ringed as chicks). We used 2938 dead recoveries from the period 1963–2005 reported by members of the public, excluding cases where only the ring was found. The recovery year was defined as 1 July to 30 June, except for the first year after ringing, which extended from the actual ringing date of each bird to 30 June in the following year. The overall mean ringing date was 5 July for chicks and 27 June for adults.

statistical methods

We analysed the ring-recovery data in mark (White & Burnham 1999) using the Seber parameterization (Williams, Nichols & Conroy 2002), in which the estimated parameters are S, the annual survival probability; and r, the probability that a dead ringed bird is found and the ring number reported. All model parameters are probabilities, and a logit-link function ensures that estimates and confidence limits remain within the interval [0;1]. A number of statistical models, each representing a biological hypothesis, were fitted to the data. Subscripts indicate the structure of the model, following the principles of Lebreton et al. (1992): a2 or a3 indicates a model with two or three age classes, t indicates a fully time-specific structure (year as a factor), T indicates a logit-linear trend over time (year as a covariate), CR indicates the presence/absence of a colour ring, and environmental covariates are indicated as listed below. An asterisk between two terms indicates that the model includes an interaction term, and a plus that the model is additive (without interaction).

Goodness of fit was evaluated with the median c-hat procedure in mark, and the estimated variance inflation factor ĉ was used to adjust standard errors and calculate QAICc, the bias- and small sample-adjusted Akaike's information criterion (Burnham & Anderson 2002). Models were then ranked according to QAICc, with lower values indicating better approximating models with a proper balance between under- and overfitting.

Previous studies of shags have shown that survival is lower during the first 2 years of life than among adults (e.g. Aebischer 1986; Catchpole et al. 1998), and there is also evidence for a senescent decline in survival among older birds (Harris et al. 1994b; Harris et al. 1998). We restricted age-dependence in survival to three age classes (first-year, second-year and adult), as we were primarily interested in the temporal variability of survival for each of these age classes, and the addition of further age classes would lead to a loss of power to detect meaningful temporal patterns. For recovery probabilities, we used a two-age-class structure, as first-year juveniles often have separate wintering areas and a different pattern of vulnerability to various sources of mortality, both factors that could lead to different probability of dead ringed birds being found and reported (e.g. Frederiksen & Bregnballe 2000). We also allowed, in the most general model, for dead birds with colour rings potentially being more likely to be reported. Our general model thus had the structure Sa3*t ra2*CR*t, with all parameters time-dependent.

estimating and comparing process variances

When evaluating the impact of demographic variability on population dynamics based on empirical data, it is important to separate variance associated with the sampling process, which is irrelevant in this context, from the underlying process variance (Gould & Nichols 1998). For capture–mark–recapture data, this can be done using the random effects module in mark, which uses the method of moments to provide an estimate of the process variance of a given set of estimated parameters, typically annual estimates of survival (Franklin et al. 2000; Burnham & White 2002). However, in order to compare process variances among different sets of parameters, here age classes, mathematical restrictions on the variance of probabilities must also be taken into account. Briefly, the maximum possible variance associated with a probability is a function of the mean [p(1 – p), where p is the mean]; it is highest at a mean of 0·5 and declines to zero at means of 0 or 1 (Morris & Doak 2004). We followed previous authors (Gaillard & Yoccoz 2003; Morris & Doak 2004; Altwegg et al. 2006) in scaling process variance by the maximum possible variance for the given mean, and used the term ‘relative process variance’. We used the Markov chain Monte Carlo (MCMC) module in mark to estimate process correlations between survival probabilities of the three age classes.

identifying covariates of survival

We adopted a confirmatory rather than an exploratory approach when identifying covariates of survival (Anderson et al. 2001). A small number of candidate covariates were thus chosen, based on theoretical considerations and the results of previous studies. Like other cormorants, shags have a partially wettable plumage (Grémillet, Tuschy & Kierspel 1998; Grémillet et al. 2005) and therefore potentially suffer extensive heat loss in cold and wet conditions. Low temperatures, strong winds and high rainfall could therefore lead to increased mortality, particularly in winter when feeding conditions deteriorate and the time available for foraging is limited (Daunt et al. 2006, 2007). In addition, shags probably forage less efficiently when water turbidity is high (cf. Strod et al. 2005), such as during strong onshore winds or following heavy rainfall. Onshore winds may thus have several potentially interacting negative effects on shags: drenching by waves and spray (shags roost on rocks close to the tide line), increased evaporative cooling and reduced foraging efficiency. Large-scale mortality of Isle of May shags in late winter 1994 was related to an extended period of strong onshore (easterly) winds (Harris & Wanless 1996), and similar patterns had been shown previously in another colony (Potts 1969). In general, shag mortality peaks in late winter. Among 851 Isle of May shags recovered as freshly dead, 31% of first-year birds and 41% of older birds were recovered in February and March. Late winter conditions have also been shown to affect the extent of non-breeding and timing of breeding in Isle of May shags (Aebischer & Wanless 1992; Daunt et al. 2006). Taking into account a likely 2–3-week mean lag between death and recovery (cf. Daunt et al. 2007), we concentrated on February weather in our search for relevant covariates of survival. Daily weather data from Leuchars (28 km north-west of the study site) were extracted and the following synthetic weather variables were calculated: mean daily minimum air temperature (AT), total precipitation (R), and summed onshore wind component (OC). The onshore (easterly) wind component was calculated for each day as mean daily wind speed (in knots) × sin(mean daily wind direction), and set to 0 if wind direction was between 180 and 360° (i.e. westerly). The resulting variable was then summed over all days in February The three environmental covariates were not highly correlated (AT/R: r = −0·09; AT/OC: r = −0·46; OC/R: r = 0·34). Onshore winds are also likely to increase the chance that dead birds are recovered, so we first tested for an effect of OC on recovery probabilities before modelling survival.

Traditionally, important temporal covariates of survival have been identified by fitting ultrastructural models, where annual survival is constrained to be a function (usually on the logit scale) of one or more covariates, and comparing these models with constant and fully time-dependent models using information-theoretic criteria such as AICc (Lebreton et al. 1992). However, in large data sets with strong temporal variation in survival, this approach has inflated power (Link 1999): fully time-dependent models are almost invariably preferred over covariate models, and ultrastructural covariate models are generally preferred over constant models, even when the covariate is a series of random numbers. In the present data set, temporal variation in survival was very strong and fully time-dependent models always had the lowest QAICc (see Results). We fitted models including 10 different series of random numbers (uniformly distributed between 0 and 1) as covariates of adult survival; eight of these were preferred over the constant model by QAICc by a margin of up to 63, indicating very strong support for some of these non-informative models. The best framework for covariate selection is probably hierarchical mixed models with random year effects, but guidelines for this have not yet been established. We therefore selected covariates using an alternative method. We used analysis of deviance (anodev, Skalski, Hoffmann & Smith 1993) F-tests in a combined step-up–step-down approach to identify the best combination of covariates for each age class (Grosbois et al. 2006), starting from the model with all main effects and two-way interactions. This approach does not distinguish between process and sampling variance.

The amount of between-year variation explained by covariates was assessed using anodev. We calculated the proportion of the total between-year variation (deviance) in survival or recovery probabilities explained by a given covariate as (DEVc– DEVx)/(DEVc– DEVt), where c, x and t indicate, respectively, models with no temporal variation, with the covariate, and with full time-dependence.

stochastic population model

We used a stochastic matrix population model to explore how the observed level of temporal variation in survival affected long-term population growth rate. A three age-class model of the population at the start of the breeding season (prebreeding census sensu Caswell 2001) was constructed in ulm (Legendre & Clobert 1995). Observed means and variances of survival were taken from random-effect models on the logit scale; annual values were drawn from these distributions and back-transformed to the real scale, and thus included only process variance. We estimated mean annual fecundity (0·9 chicks per pair, SD 0·38) from our long-term records and drew annual values from this distribution. Some birds start breeding at age 2, but the majority commence at age 3, and some later (Potts et al. 1980; Aebischer 1986); in the model we assumed that all birds start at age 3. The model did not include correlations between fitness components; age of first breeding was assumed to be constant rather than stochastic; and non-breeding of established breeders was not accounted for. The starting population was 1000 females distributed according to the stable age distribution of the equivalent deterministic model. 1000 realizations i were run for 500 years T. We recorded mean stochastic population growth rate:

image

as well as the proportion of extinct trajectories at the end of the simulation (with an extinction threshold of 1). To explore the implications of changes in environmental variability, we re-ran this model with values of process variance for fecundity as well as survival of all three age classes between 50 and 150% of the observed values.

Results

modelling recovery probabilities

We first attempted to identify a parsimonious model for the recovery probability r. The most general model (Sa3*t ra2*CR*t, with year-specific recovery probabilities separately for first-year and older birds, and for birds with and without colour rings) showed some lack of fit (ĉ = 1·21), and we therefore used QAICc in model selection. This model (model 12 in Table 1) could be simplified by eliminating the age and colour-ring effects (models 5–7,9,10), but year-to-year variation in r was strong (model 5 vs. 11). A substantial part of this variation could be explained by a linear trend over time (model 4) or by onshore winds in February (model 8), and the model with both effects explained 56% of the interannual variation according to anodev (model 2). At this stage, we again tested whether additive age or colour-ring effects were important. The two age-class effect was not needed (model 3), whereas colour rings seemed to have an effect on r (model 1). The model selected at this stage for r was retained for survival modelling, and all ΔQAICc values given are relative to this model. Recovery probabilities declined strongly over the study period (β = −0·020 ± 0·0032 SE), from ≈17 to ≈7% for non-colour-ringed birds. Onshore winds in February had a positive effect on r (β = 0·0041 ± 0·0015 SE), corresponding to an increase in recovery probability of up to 4–5% in the windiest winters. Colour-ringed birds were more likely to be recovered (β = 0·22 ± 0·10 SE), although the effect was small (≈2% higher recovery probability).

Table 1.  Model selection for recovery probabilities of ringed European shags on the Isle of May
 ModelQDevianceKΔQAICcVariation explained (%)
  1. QDeviance is the deviance of the model adjusted for lack of fit; K is the number of estimable parameters; ΔQAICc is the difference in QAICc between the model in question and the best model; the amount of total between-year variation explained by one or more covariate(s) is calculated with anodev.

1Sa3*t rCR+OC+T1029·34126 0 
2Sa3*t rOC+T1034·23125 2·8755·9
3Sa3*t ra2+OC+T1033·87126 4·54 
4Sa3*t rT1040·96124 7·5949·6
5Sa3*t rt 987·0616230·46 
6Sa3*t rCR+t 985·8716331·30 
7Sa3*t ra2+t 985·8916331·32 
8Sa3*t rOC1071·8912438·5120·7
9Sa3*t rCR*t 951·5518541·53 
10Sa3*t ra2*t 957·1618955·25 
11Sa3*t r.1094·0212358·63 
12Sa3*t ra2*CR*t 917·4222384·52 

modelling survival

Between-year variation in survival was very strong for all three age classes (Fig. 2); ΔQAICc for models with constant survival was 301 for first-year birds, 24·4 for second-year birds and 649 for adults. As shown previously (Harris & Wanless 1996), adult survival was extremely low (0·27) in 1993/94, and very low adult survival (<0·6) was also observed in 1965/66 and 2004/05. First-year survival was particularly low during 1976/77–1978/79 (cf. Aebischer 1986). Mean survival, as estimated using a random-effects model on the real scale, was 0·513 (± 0·038 SE) for first-year birds, 0·737 (± 0·028 SE) for second-year birds, and 0·858 (± 0·030 SE) for adults. Estimated relative process variance was highest for first-year birds (20·4% of maximum possible), and lower for second-year birds (9·6%) and adults (14·1%). Relative process variance declined more rapidly for adults and second-year birds than for first-year birds when extreme years were dropped (Fig. 3), indicating that extreme events were more important for these older age classes. The MCMC analysis indicated substantial process correlations in survival between the three age classes (Table 2); in particular, second-year and adult survival probabilities were highly correlated. We fitted a set of additive models, where survival was constrained to vary in parallel over time between two or three age groups. The model with parallel variation between second-year and adult survival was slightly better than the fully interactive model (ΔQAICc = −5·32), whereas all other additive models performed poorly (ΔQAICc > 33). We therefore explored relationships between survival and environmental covariates separately for each age class, and also using additive models for second-year and adult survival.

Figure 2.

Estimated survival of first-year, second-year and adult European shags from the Isle of May, 1963–2005. Estimates are derived from a fully time-specific random-effects model on the logit scale (see text for details). Error bars indicate 95% confidence limits.

Figure 3.

Process variance as a proportion of the maximum possible (see text for details) in first-year, second-year and adult survival of European shags on the Isle of May, as a function of the number of extreme years dropped from the estimation.

Table 2.  Process correlations (corrected for sampling covariance) between annual time series of estimated survival probabilities of first-year, second-year and adult European shags
CorrelationMedianSE95% CI
  1. Medians, standard errors and 95% credible intervals are shown.

First vs. second year0·4010·187–0·004–0·731
First year vs. adult0·4660·156 0·145–0·736
Second year vs. adult0·8240·133 0·469–0·972

environmental covariates of survival

As expected with the large sample size and very pronounced year-to-year variation in survival, covariate models were never preferred over fully time-dependent models by QAICc. Furthermore, for first-year and adult survival where year-to-year variation was particularly pronounced, covariate models invariably had a lower QAICc than the constant model, and the most complex model (with all main effects and two-way interactions) had the lowest QAICc of all covariate models (Table S1 in Supplementary Material). QAICc was thus not a useful tool for covariate selection. Stepwise anodev resulted in the following covariates being selected: OC (marginal) for first-year survival, and OC*R for second-year and adult survival (Table S2). The selected covariate models for first-year, second-year and adult survival explained, respectively, 7·3, 15·1 and 42·9% of the annual variation. Because the same covariates were selected for both second-year and adult survival, we used a model including an additive constraint on these two age classes to derive coefficients for the relationship between survival, onshore winds and precipitation (Table 3). Predicted adult survival was high (0·85–0·95) under most conditions, but fell dramatically when both OC and R were high, to 0·15 at the highest observed values (Fig. 4). Similarly, predicted second-year survival was 0·7–0·9 under most conditions, but fell to 0·07 at the highest observed values of OC and R. While the relationship with OC and R accurately predicted the low survival in 1965/66 and 1976/77, it underestimated the magnitude of the mass mortality in 1993/94, and the low observed survival in 2004/05 was unexpected based on these weather variables (Fig. 5). We tested whether variation in population size might explain some of the lack of fit by including the number of breeding pairs and interactions with the selected weather variables as additional covariates of adult survival. According to anodev, this model was far from being preferred (F4,32 = 0·23, P = 0·92), and population size thus could not explain the lack of fit of the best weather-related model.

Table 3.  Coefficients of the preferred logit-scale random-effect model for second-year and adult survival
CoefficientEstimateSE
  1. Coefficients are given on the logit scale and thus are not immediately interpretable; Fig. 4 plots the relationship for adult survival.

Intercept 1·7280·121
Additive age effect–0·7730·081
OC 0·0082680·002219
R 0·0119190·001936
OC.R–0·0003500·000027
Figure 4.

Predicted adult survival as a function of summed onshore component and total precipitation, both in February. Coefficients are from a model with an additive constraint on second-year and adult survival (Table 3).

Figure 5.

Predicted and observed adult survival, 1965–2005. Predicted values are from a constrained model (Fig. 4); observed values from an unconstrained random-effect model (Fig. 2).

stochastic population model

With the observed means and process variances for survival of each age class, mean λs was 0·9838, and 745 of the 1000 trajectories were extinct after 500 years. We recalculated process variances (but not means) of second-year and adult survival after removing the most extreme year, 1993/94. Mean λs was 1·0003, and none of the 1000 trajectories was extinct after 500 years. To simulate the effect of potential changes in environmental stochasticity, we re-ran the model with process variances for survival of all three age classes ranging from 50 to 150% of the observed values. Increasing process variance by 20% led to near-certain extinction after 500 years, whereas reducing it by 30% led to a positive growth rate and certain persistence of the population (Fig. 6).

Figure 6.

Stochastic population growth rate (closed symbols) and probability of extinction (open symbols) as functions of the process variance in survival of all three age classes, relative to observed values.

Discussion

shag survival and weather

Survival of second-year and adult shags was substantially reduced in years when high precipitation (mostly rainfall) and strong onshore (easterly) winds coincided in February (Fig. 4); the interactive model including these covariates explained 43% of the temporal variation in adult survival. To avoid data mining and to obtain a parsimonious model, we decided a priori to focus on February weather rather than search for the time window where the weather–survival correlation was highest. February was chosen mainly because mortality (measured as the number of birds recovered) peaks, on average, at this time, and because major shag ‘wrecks’ in the North Sea (mass occurrences of beached dead birds) usually occur from February onwards (Potts 1969; Harris & Wanless 1996). In addition, detailed studies of overwinter time budgets showed that foraging effort in February was linked to timing of breeding in the following season, with females spending more time foraging in February laying later (Daunt et al. 2006). This suggests that late winter is a stressful period for shags, and it is likely that individuals in poor body condition may struggle to survive if weather is poor. Nevertheless, the timing of peak mortality varied substantially between years (data not shown), and it is likely that a more detailed search would allow us to explain a larger proportion of the between-year variation in survival. First-year survival was highly variable between years (Fig. 2), but not strongly correlated with February weather (Table S1; Table S2). Consistent with this, Potts (1969) showed that the timing of peak mortality of first-year shags varied between both years and colonies. The less clear relationship between late winter weather and survival for first-year birds relative to adults may reflect higher vulnerability of juveniles to environmental conditions in autumn or early winter and/or greater importance of food abundance relative to weather in this age class, due to less developed foraging skills. First-year survival thus may not be highly correlated with any individual weather covariate, because the period of greatest vulnerability varies between years. In addition, shags disperse further from the colony during their first winter than later, on average, which would tend to make the weather covariates we have used here less appropriate for this age class.

The combination of strong easterly winds and heavy rainfall is likely to have had both direct and indirect impacts on shags. Gales and associated heavy rainfall during the breeding season can cause mass mortality among unfledged shag chicks, presumably through hypothermia, and this mortality is most pronounced in nests exposed to the prevailing wind (Aebischer 1993; unpublished data). Because shag plumage is not completely waterproof (Grémillet et al. 1998; Grémillet et al. 2005), adults may succumb to the same factors, particularly in winter when ambient temperatures are relatively low. For the double-crested cormorant Phalacrocorax auritus, Hennemann (1983) showed that birds with wet plumage suffered increased heat loss at low temperatures. However, it is also likely that foraging is inhibited during onshore gales, perhaps because of increased turbidity. Daunt et al. (2006) showed that while the foraging effort of Isle of May shags during winter generally increased when onshore winds dominated, birds stopped foraging completely during the strongest wind episodes. This could reflect increasing energy demands during onshore winds, combined with decreased foraging efficiency when turbidity was very high. Prolonged episodes of onshore winds may thus lead to both increased energy demand and decreased intake rates, a potentially lethal combination for shags, which carry very small fat reserves (D.N. Carss, pers. comm.). Interestingly, while major mortality events (wrecks) of young shags are a regular occurrence on the relatively linear east coast of Britain, which has very few islands and thus little shelter from onshore winds (Potts 1969; Harris & Wanless 1996), they seem to be absent on the west coast of Scotland, which, with its convoluted coastline and many small islands, offers shelter from any wind direction (Swann & Ramsay 1979). Indeed, over the period 1985–2005, variability in shag breeding population size was higher on the Isle of May (CV = 0·55) than on Canna in western Scotland (CV = 0·39; R.L. Swann, pers. comm.) or two colonies in Shetland, where shelter is also available from all wind directions (CV = 0·42 and 0·15; M. Heubeck, pers. comm.).

Survival of adult shags showed an unusually high degree of temporal variation, particularly for a generally long-lived organism. In our study, adult survival varied from 0·27 to 0·98, similar to the range of mean values observed across birds and other annually reproducing organisms (Sæther & Bakke 2000). Like other cormorants (Duffy 1983; Nur & Sydeman 1999), shags also show large between-year variation in fecundity (Aebischer & Wanless 1992; unpublished data) and are capable of rapid population growth. Unusually for seabirds (Weimerskirch 2002), shags occasionally can raise a brood of four chicks successfully (Harris et al. 1994a), and exceptionally can rear two broods in a season (Wanless & Harris 1997). Taken together, these life-history traits constitute what could be termed a demographic ‘boom-or-bust’ syndrome among cormorants, where individuals and populations are able to take advantage of favourable conditions through high fecundity and survival, while suffering high mortality and breeding failures under unfavourable conditions. This life-history syndrome is linked to a set of presumed morphological adaptations to efficient underwater foraging (Grémillet et al. 1999; Grémillet et al. 2001): large, fully webbed (totipalmate) feet, partially wettable plumage, and small fat stores to reduce buoyancy. These morphological traits, particularly plumage and fat stores, are probably linked to the high vulnerability to mortality due to inclement weather demonstrated in this study. In contrast to the standard image of seabirds as highly conservative, ‘prudent’ breeders, shags and other cormorants thus have a ‘risky’ life style and can thrive only at times and locations where prey availability is high (Grémillet, Wanless & Linton 2003).

the impact of temporal variation in survival on population dynamics

We found strong temporal variability in survival for all three age classes (Figs 2 and 3). Consistent with the environmental canalization hypothesis (Gaillard & Yoccoz 2003), relative process variance was highest for first-year survival, which in long-lived organisms has a lower sensitivity/elasticity in terms of mean population growth rate than adult survival (Lebreton & Clobert 1991). The relatively low estimated process variance for second-year survival (Fig. 3), which is also a prebreeding parameter and therefore has the same elasticity as first-year survival, may be related to the higher sampling variance for this age class (mean annual sampling variance: first-year 0·069, second-year 0·079, adult 0·032, cf. confidence limits in Fig. 2). Few studies have quantified relative process variance. Our estimates of relative process variance in shag survival are higher than that found in barn owls, Tyto alba (<0·1 for all age classes) by Altwegg et al. (2006, 2007), but comparable with that of European dippers, Cinclus cinclus (0·176 for adults, Loison et al. 2002). A wider range of species need to be studied before general conclusions can be drawn.

Process variance in second-year and adult survival dropped by ≈50% when the most extreme year, 1993/94, was excluded from the estimation (Fig. 3), whereas process variance in first-year survival, while very high, was much less driven by extreme events. A similar pattern was found for barn owls (Altwegg et al. 2006). This has important implications for the impact of extreme weather events on population growth rate, as illustrated by the results of the stochastic population model. Removing the most extreme year from the estimation of process variance in second-year and adult survival increased predicted growth rate to 1 and essentially eliminated the risk of extinction. On the other hand, a 20–50% increase in process variance for all age classes had a strong negative effect on predicted growth rate and caused extinction of all model trajectories. For adults, one additional year with survival as low as in 1994 would lead to ≈50% increase in process variance. In other words, the frequency and severity of extreme weather-driven mortality events has strong implications for population growth in this species. In fact, our model probably underestimates the impact of increased variance in survival, because it does not account for the high and positive process correlation between second-year and adult survival (Table 2), nor for potential correlations between fecundity and survival. Positive correlations between demographic parameters imply that poor years for survival and reproduction, for example, tend to coincide and thus exacerbate the negative effect of temporal variability on population growth rate (Fieberg & Ellner 2001).

With the exception of sudden cold periods, extreme weather events are predicted to become more frequent under most scenarios for future climate change (Solomon et al. 2007), although to our knowledge no specific predictions for the frequency and duration of easterly gales in winter in the North Sea area are currently available. Increased variability and higher frequency of extreme events are likely to affect most ecosystems in the coming decades. It is likely that species for which one or more demographic parameters are directly affected by high temperatures, wind or precipitation extremes, such as European shags, Manx shearwaters, Puffinus puffinus (Thompson & Furness 1991), bearded tits, Panurus biarmicus (Wilson & Peach 2006), and mouflon, Ovis gmelini musimon× Ovis sp. (Garel et al. 2004), will be disproportionately negatively affected (for review see Parmesan, Root & Willig 2000). Predictions of the ecological effects of climate change thus need to account not only for changes in mean climate, but also for the expected increase in the frequency of extreme events and the associated effects on vulnerable species.

Acknowledgements

Thanks to the Natural Environment Research Council and the Joint Nature Conservation Committee for supporting the long-term studies on the Isle of May, to Scottish Natural Heritage for access to the island and for recent nest counts, and to the Isle of May Bird Observatory for some nest counts, for ringing most of the shags during the first few decades of the study, and for supplying BTO rings throughout the study. Weather data were supplied by the UK Meteorological Office through the British Atmospheric Data Centre.

Ancillary