By continuing to browse this site you agree to us using cookies as described in About Cookies
Notice: Due to essential maintenance the subscribe/renew pages will be unavailable on Wednesday 26 October between 02:00- 08:00 BST/ 09:00 – 15:00 SGT/ 21:00- 03:00 EDT. Apologies for the inconvenience.
Dr Morten Frederiksen, CEFE/CNRS, 1919 Route de Mende, F-34293 Montpellier Cedex 5, France, Fax: + 33 4 67 41 21 38, E-mail: firstname.lastname@example.org
1. The increasing population of cormorants (Phalacrocorax carbo sinensis) in Europe since 1970 has led to conflicts with fishery interests. Control of cormorant populations is a management issue in many countries and a predictive population model is needed. However, reliable estimates of survival are lacking as input for such a model
2. Capture–recapture estimates of survival of dispersive species like cormorants suffer from an unknown bias due to permanent emigration from the study area. However, a combined analysis of resightings and recovery of dead birds allows unbiased estimates of survival and emigration.
3. We use data on 11 000 cormorants colour-ringed as chicks in the Danish colony Vorsø 1977–97 to estimate adult survival and colony fidelity. Recent statistical models allowing simultaneous use of recovery and resighting data are employed. We compensate for variation in colour-ring quality, and study the effect of population size and winter severity on survival, as well as of breeding success on fidelity by including these factors as covariates in statistical models.
4. Annual adult survival fluctuated from year to year (0·74–0·95), with a mean of 0·88. A combination of population size in Europe and winter temperatures explained 52–64% of the year-to-year variation in survival. Differences in survival between sexes was less than 1%. Cormorants older than ≈ 12 years experienced lower survival, whereas second-year birds had survival similar to adults. Colony fidelity declined after 1990 from nearly 1 to ≈ 0·90, implying 10% permanent emigration per year. This change coincided with a decline in food availability.
5. Apparently, survival was more severely affected by winter severity when population size was high. This could be caused by saturation of high-quality wintering habitat, forcing some birds to winter in less good habitat where they would be more vulnerable to cold winters. There was thus evidence for density dependence in adult survival, at least in cold winters.
6. The high population growth rate sustained by European Ph. c. sinensis in the 1970s and 1980s can partly be accounted for by unusually high survival of immature and adult birds, probably caused by absence of hunting, low population density and high food availability.
The cormorant [Phalacrocorax carbo (L.)] is a large fish-eating, colonial waterbird with an almost cosmopolitan distribution (Orta 1992). Two subspecies breed in Europe: Ph. c. carbo (L.), which mainly nests along rocky coasts in the North Atlantic, and Ph. c. sinensis (Blumenbach), which nests in freshwater and sheltered marine habitats. Trends in the European breeding populations of the two subspecies have differed markedly since at least 1970. Phalacrocorax carbo carbo has increased slowly (at a few per cent per year) in most breeding areas (Debout, Røv & Sellers 1995), whereas populations of Ph. c. sinensis have increased dramatically from a near-threatened status. The increase started in the Netherlands, continued in Denmark and later reached Germany, Sweden and Poland (van Eerden & Gregersen 1995). Before the rate of increase levelled off in the mid-1990s, the breeding population in north-west Europe had increased more than 20-fold to ≈100 000 pairs (Bregnballe 1996).
Because cormorants eat exclusively fish, the increase in Ph. c. sinensis has led to conflicts with human interests in both breeding, staging and wintering areas over most of Europe (Kirby, Holmes & Sellers 1996; Adamek, Klinger & Staub 1997; van Dam & Asbirk 1997). Both fish-farms and open-water fisheries complain about cormorant depredation, and management agencies in several countries have been met with demands for reductions in cormorant damage and/or populations. This situation has led to attempts to construct models to project population growth and evaluate the effects of management initiatives (Lebreton & Gerdeaux 1996; Bregnballe, Goss-Custard & Durell 1997). However, these attempts have been hampered by the lack of reliable estimates of most population parameters. Since cormorants are long-lived birds, adult survival is very important for population growth rate (Lebreton & Clobert 1991); almost no information is available on this parameter and the factors affecting it for Ph. c. sinensis.
Two types of data collected through bird ringing schemes are often used to estimate survival of wild birds: recoveries of dead individuals, and recaptures or resightings of live individuals, usually within a restricted study area. A large body of theory has been developed for the estimation of survival from either of these two types of data (recaptures, e.g. Lebreton et al. 1992; recoveries, e.g. Brownie et al. 1985). However, each of these sources of information has its shortcomings for estimation purposes (Clobert & Lebreton 1991).
Estimates of survival from pure capture–recapture (or resighting) studies are biased downwards if permanent emigration from the study area occurs, since mortality and emigration are confounded. Permanent emigration may be considered negligible in highly philopatric species [e.g. shag Phalacrocorax aristotelis (L.), Catchpole et al. 1998], particularly for breeding adults, but when estimating prebreeding survival it can rarely be ignored without risk of drawing misleading conclusions. Inter-colony dispersal is known to occur in cormorants, both among immatures and breeding adults (Bregnballe & Gregersen 1995). Survival estimates from pure resighting studies may therefore be substantially negatively biased by permanent emigration in this species. When only recovery data are analysed, it is problematic to identify age effects on survival if birds are only ringed as chicks (Catchpole, Freeman & Morgan 1995). Furthermore, recovery data are often quite sparse and confidence limits of the estimates obtained are correspondingly wide.
In some studies, both recovery and recapture/resighting data are collected. Several authors have compared survival estimates from both types of data (Francis & Cooke 1993) or attempted to model them simultaneously (Buckland 1980; Mardekian & McDonald 1981; Lebreton et al. 1995). However, none of these approaches made use of the opportunity to estimate emigration (or fidelity) probabilities explicitly from the ratio of apparent survival (capture–recapture) to ‘true’ survival (recoveries; but see Ganter & Cooke 1998).
The theory for a combined analysis of recovery and recapture (or resighting) data was developed by Burnham (1993), and this type of model was first applied by Szymczak & Rexstad (1991) in an analysis of gadwall (Anas strepera L.) data. Catchpole et al. (1998) developed the model further by including age effects. The strength of this type of analysis is that, like recovery analysis, it allows an estimate of survival that is unbiased by emigration from the study area, while precision is increased by the inclusion of recapture or resighting information. At the same time, it allows permanent emigration to be estimated directly, and thus the study of which factors affect emigration is facilitated in a way analogous to what has recently become standard practice in capture–recapture studies of survival (Lebreton et al. 1992; Cooch, Pradel & Nur 1997; Cooch & White 1999).
Since 1977, recoveries and resightings of colour-ringed birds have been collected in a Danish cormorant population during phases of rapid population growth and subsequent stagnation. Here, we use a combined analysis to study survival and colony fidelity in this population.
In this paper, we (i) provide a robust estimate of adult survival of cormorants; (ii) investigate factors affecting adult survival, both individual (age, sex) and external (population size and winter severity); and (iii) provide data on variation over time in emigration of breeders. We use these results to answer two questions: (a) what is gained by the combined analysis, and (b) can survival estimates contribute to an understanding of population trends? Results concerning first-year survival and philopatry are presented in a companion paper (Frederiksen & Bregnballe 2000).
Study area and data collection
The study took place at Vorsø, a 60-ha island in Horsens Fjord, Denmark (55°52′N, 10°01′E) (Fig. 1). The colony at Vorsø is the oldest and one of the biggest cormorant colonies in Denmark. Cormorants settled here in 1944, and the population was kept at a low level (below 500 pairs) by shooting of adults and young until 1970. The number of breeding pairs then increased until 1991 (5000 pairs); since then colony size has fluctuated and then declined to 3100 pairs in 1998. The cormorants nest in trees, which are eventually killed by the birds' excrements. Until ≈1990, the western part of the colony was centred around a small pond. Gradually, all trees around the pond died, and the western subcolony has been spreading to adjacent scrub and woodlands since ≈1987.
In the autumn of 1983, the present 10-m tall observation tower was built in the then central part of the colony (Fig. 1). Access to the tower is through a covered walkway, so that birds are not disturbed by observers entering and leaving. The tower stands next to the pond, which is used for roosting and preening by birds breeding in the western subcolony. Thus, while the number of nests in the vicinity of the tower has decreased markedly since 1987 because of the death of nesting trees, it has remained possible to follow a large proportion of the birds present in the colony from the tower, including breeders, non-breeders and visitors.
We used data on 11 169 cormorants ringed as chicks during 1977–97 in the Vorsø colony (Fig. 2a). The birds received both a standard metal ring on one leg and a coloured plastic ring engraved with a three-character alphanumeric code on the other. Colours used were blue with white engraving, and white or yellow with black engraving.
The types of plastic and engraving techniques used have not been constant over the study period, leading to variation in colour ring quality and, consequently, in risk of wear and loss of colour rings (Bregnballe et al. 1997b). Based on known instances of ring loss and inspection of recovered colour rings, we have divided the rings used into three quality categories (Fig. 2a). In order to account for heterogeneity in colour ring loss, we split the data into three subsets based on these quality categories. Birds that lost their colour rings effectively disappeared from the colony, since they could no longer be resighted. However, they could still be recovered. Colour ring loss should thus affect resighting probability much more than reporting probability and should mimic permanent emigration from the study colony.
Beginning in 1978, resightings of colour-ringed cormorants have taken place annually in the Vorsø colony. Resighting effort increased gradually until 1983 and peaked in 1984 following the construction of the observation tower. Since 1985, the effort has been stable. One to several observers have searched for colour-ringed cormorants 1–3 hours per day, 6–7 days per week during the entire breeding season, with special effort being expended during periods of massive arrival of birds in early spring, as well as during the period of establishment of territories and nest-building. Since 1984, most resightings have been made from the main observation tower (1–3 observation bouts/day). However, supplementary observations have been made from the ground in other parts of the colony. In total, ≈ 300 000 resightings have been made in the colony during the study period.
The colour rings used were readable with a ×20–60 telescope at up to 400 m. However, the majority of resightings in the colony occurred at much shorter distances, where variation in readability of colour rings is unlikely to have had a measurable effect on the resighting probability. In most years, an experienced observer continuously checked the list of resighted birds as data were transcribed to card files. Obvious mistakes (non-existent codes, mistaken colours, birds already dead, etc.) were eliminated at an early stage in this way.
Resightings took place throughout the period when birds were present in the colony, i.e. from early February (in mild winters) until the beginning of October. However, the majority of individuals were observed in spring. For example, in 1994, when 949 birds ringed on Vorsø in previous years were observed, 760 (80%) of these were seen at least once before 1 May and 873 (92%) at least once in the period 1 March – 31 May.
There was a large variation in the number of times each bird was observed in a given year, which was mainly explained by individual variation in breeding activity and nesting site within the colony. Thus, in 1994 a total of 1406 individuals (including newly fledged juveniles and birds ringed in other colonies) were observed between one and 132 times. Because only summary data have been computerized prior to 1994, we disregard this variation and treat all birds ringed on Vorsø in previous years and seen at least once in a given year equally. Observations of newly fledged juveniles were not used and birds ringed in other (Danish or foreign) colonies were excluded from the analysis.
Figure 2b shows the number of individuals that were resighted annually in the colony 1978–97. With each bird counted only once in a given year, we use 18 238 resightings in the combined analysis.
We used 1687 recoveries reported before 1 September 1998 of cormorants known to be dead before 1 May 1998. Recoveries in the colony were also used. However, cases when only the (metal or colour) ring was found were excluded.
For maximum correspondence with resighting data, we defined the recovery year as starting on 1 May, which corresponds well to the end of spring migration. The number of birds recovered in each year is shown in Fig. 2b.
In 24 cases, a bird was resighted in the colony early in spring and subsequently recovered before the end of April in the same year. Because of the above definition of the recovery year, these birds appeared to have been observed after they were dead. We addressed this problem by moving the recovery date of such birds to the next recovery year, as if they had died after 1 May.
Combined analysis of recovery and resighting data
First, we constructed capture histories for each individual. These included information on the year of ringing, years when the bird was observed in the colony and, when appropriate, year of recovery, all coded as a series of 1s and 0s. These capture histories constituted the input data for the software packages used.
In a combined analysis, four types of parameters are estimated (Burnham 1993; White & Burnham 1999): S, the probability that a marked animal survives from one occasion to the next (survival probability); F, the probability that a surviving marked animal remains in the study area, i.e. does not emigrate permanently (fidelity probability); p, the probability that a surviving marked animal that has not emigrated permanently is observed in the study area (recapture/resighting probability); and, r, the probability that a dead marked animal is found and the ring number reported (reporting probability).
The combined analysis was carried out in mark (Cooch & White 1999; White & Burnham 1999), the only software package directly including this facility. mark estimates model parameters iteratively by maximum likelihood methods and provides facilities for goodness-of-fit testing and model selection. For model selection and estimation of precision, we used QAIC (Quasi-likelihood Akaike's Information Criterion), an information-theoretic measure that ensures a good balance between under- and over-fitting (bias and precision) when the most general model does not fit the data (Anderson, Burnham & White 1994; Burnham & Anderson 1998). QAIC is defined for each model as (− 2 log likelihood/ĉ) + 2 np, where ĉ is a variance inflation factor estimated from goodness-of-fit tests and np is the number of parameters estimated in the model. If the general model fits the data, the expected value of ĉ is 1. Since the absolute value of QAIC is uninformative, we present here ΔQAIC, the difference in QAIC between the model in question and the model with the lowest QAIC (Burnham & Anderson 1998).
Survival analysis of known breeders
One potential problem of the combined approach is that it is not possible to take ‘trap-dependence’ into account. Trap-dependence is a general phenomenon, which includes all situations when the probability of observing a bird in a given year depends on whether or not it was observed in the previous year. We suspected the presence of trap-dependence in this study because birds changed breeding sites within the colony, between sites where they were easily observed and others where they were only observed if they showed up in the pond close to the observation tower. A consequence of these moves is ‘trap-happiness’: birds seen in the previous year are much more likely to be seen than those not seen in the previous year (Sandland & Kirkwood 1981). If trap-dependence occurs and is not accounted for in the model, a serious negative bias in survival estimates may be the outcome (Pradel 1993). Given the structure of the input data (capture histories), it is not possible to model age- and trap-dependence simultaneously. However, in a pure resighting data set it is possible to split the capture histories at each resighting, in such a way that ‘age’ becomes a surrogate for number of years since the last resighting (Pradel 1993). Real age effects can no longer be estimated, but trap effects can be included.
Another reason for carrying out a survival analysis based on resightings of known breeders was that we wanted to investigate possible sex effects on survival. Cormorant chicks could not be sexed from morphology at ringing and we therefore needed a sample of known-sex birds to study sex effects. A total of 1428 birds colour-ringed as chicks on Vorsø were known with certainty to have bred there (produced eggs in at least one season) up to 1997. A majority of these birds were sexed through observations of behaviour, but 63 (4·4%) remained unsexed and were excluded from the analysis. This produced a slight positive bias in estimated survival, since birds that died soon were more likely not to have been sexed (Buckland 1982).
For this analysis, we used only resightings in the colony of known breeders. Capture histories were started at the first known breeding attempt of each bird, and all subsequent resightings were included. The period analysed ran from 1981, when breeding colour-ringed birds were first recorded, up to and including 1998. Thus defined, the data set included 3317 resightings of 695 males and 3128 resightings of 670 females—as above, only one resighting per bird per year was used.
In this analysis, the parameters estimated were φ, the local or apparent survival probability (biased by permanent emigration; in principle the product of S and F) and p, the resighting probability defined as above.
The analysis was carried out in surge 5·1 (Cooch et al. 1997) and initial goodness-of-fit tests in release (Burnham et al. 1987), as modified by Pradel (1993) to include tests for trap-dependence. Here, too, we used QAIC for model selection and estimation of precision.
The model notation we use here is based on Lebreton et al. (1992), with further refinements. This notation symbolizes model structure in a way that illustrates the conceptual link with general linear modelling. For each parameter type, model structure is denoted by subscripts. When more than one factor is involved, an asterisk means that the model includes interaction terms, while a plus sign indicates a model with only main effects (additive model). As an example, the model (φt+g, pt) has additive effects of group and time (parallelism among groups over time) on apparent survival, as well as variation over time in resighting probability.
Subscripts used here are listed in Table 1. When model structure differs between two age groups for a given parameter type (typically survival), we use an extended notation; for example, Sc,t means that the model has a constant first-year survival and an adult survival varying over time.
Table 1. . Model subscripts used in the two analyses
Two age classes (first-year and adult)
Quadratic trend over age beyond 1 year
Time-dependence (year-to-year variation)
S, F, r, p, φ
Time-dependence, but constant 1985–97
Linear trend over time
Group effect (colour ring quality)
‘Trap-dependence’ (p dependent on whether the bird was observed on the previous occasion)
p (for breeders only)
Mean winter temperature (Dec-Feb) in Denmark
Mean amount of cold (see text) in Denmark
North-west European breeding population of Ph. c. sinensis
Breeding success (chicks fledged/clutch) in the Vorsø colony
For brevity, we use ‘survival in year i’ to mean ‘survival from 1 May of year i to 30 April of year i + 1’. For example, survival in 1995 means survival over the winter 1995–96. Similarly, survival of 1-year-old (or second-year) birds means survival from age 1 year to age 2 years.
Estimation of survival probabilities relies heavily on selection of an appropriate statistical model. The recommended procedure when selecting a model to describe a capture–recapture or recovery data set is to fit a predefined, limited set of biologically realistic models, and then to select the model having the lowest QAIC as a basis for inference (Burnham & Anderson 1998). It is important to consider carefully which models should be included; this ensures that inferences from the selected model (or a set of models with ΔQAIC values less than ≈ 6–10) are valid (Burnham & Anderson 1998). We used models with full time-dependence (year-to-year variation) and some degree of age-dependence (usually two age-classes), as well as models in which parameters were constrained to be functions of one or more external covariates. Models in which ‘biological’ parameters (S, F, φ) were held constant over the study period were not considered, since we found them obviously unrealistic, particularly since the large sample size should allow us to demonstrate even modest variation over time.
We tried to relate survival and fidelity probabilities to various external covariates. Biological considerations led us to suspect that survival might be influenced by population density and winter severity. Therefore, we used the number of breeding pairs of Ph. c. sinensis in north-west Europe (Denmark, Sweden, The Netherlands, Germany, Poland; no data were available on the size of winter populations) and two different measures of winter severity, namely mean winter temperature (Dec–Feb) and amount of cold (absolute sum of mean temperature of all days when mean temperature was below 0°C). Both measures of winter severity were extracted from Danish sources (Anonymous 1998; Cappelen & Jørgensen 1998). Mean winter anomalies (deviations from long-term mean temperature) in Denmark and Central Europe were correlated (r = 0·81 and 0·71 between Denmark, and west-central and southern Germany, respectively), indicating that using Danish data was justified. The two measures were highly correlated (1977–97: r = − 0·91), but amount of cold emphasized severe ice winters with periods of intense cold, whereas mean winter temperature was better able to discriminate among mild winters.
For fidelity probability we used colony size (testing for direct density dependence) and breeding success in the colony (measured as number of chicks fledged per clutch) as external covariates, since dispersal in colonial birds has been shown to be correlated with conspecific reproductive success (Danchin, Boulinier & Massot 1998). We also modelled reporting probability as following a linear (presumed negative) trend over time, since such a trend has been found in several other studies (Wernham & Peach 1999), often interpreted as an effect of reduced shooting and/or reduced interest in reporting recoveries.
In both the combined analysis and the analysis of breeders, we used the logit link function to provide the link between external covariates and estimated parameters, and to constrain all estimates to be within the interval [0; 1].
When dealing with age effects on survival beyond age 1 years, we chose to use a quadratic constraint, i.e. modelling survival as a quadratic function of age. This is a parsimonious way to allow for senescence that has been used by several authors (Pugesek et al. 1995; Newton & Rothery 1997). Sex effects were treated by using separate data sets for males and females in the analysis of breeders.
A general model
Because ring loss should mimic permanent emigration, we assumed that variation in colour ring quality would only affect F, the fidelity probability, and therefore only included a group effect for this parameter. We also assumed that parameters differed between first-year and older birds, and thus our starting model was (Sa2*t, pa2*t, rt, Fa2*g*t), incorporating variation over time for two age-classes for all types of parameters. An important assumption of models of recovery data from birds ringed only as young is that the reporting probability r is independent of age (Catchpole et al. 1995); therefore, we did not include a two age-class effect for r.
The bootstrap goodness-of-fit procedure in mark showed that the fit of the general model was reasonable (ĉ = 1·46). The lack of fit may have resulted from both over-dispersion and structural inadequacies of the general model. Goodness-of-fit tests of the resighting data set on breeders (see below) indicated that trap-dependence was important and could have caused part of the lack of fit. Because it was not possible to model trap-dependence in the combined analysis, we used the estimated variance inflation factor in model selection and estimation of precision.
In a combined analysis the number of models quickly becomes very high. Thus, if all combinations of five different parameterizations for each parameter type are used, the total number of models is 54 = 625. Therefore, we do not list here all the models we have fitted. Table 1 shows which parameterizations were tried for the four parameter types; not all combinations were fitted, since some parameterizations were quickly seen to be unrealistic.
Table 2 shows the general model and the final model selected, along with some of the best contenders for each parameter type, thus illustrating the model selection process. The final model (St,A2+t, pt,t(c), rt, Ft+g) had 109 parameters and included time-dependence for all parameter types, plus quadratic age-dependence for adult survival and a group (colour ring quality) effect on fidelity. Adult resighting probability could be modelled as constant in the period following construction of the observation tower (1985–97). Because all parameter types were modelled as time-dependent, parameters were generally not estimable for the last time interval.
Table 2. . Model selection for the combined analysis
Sa2*t, pa2*t, rt, Fa2*g*t
St,A2+t, pt,t(c), rt, Ft+g
St,A2+eur+w1, pt,t(c), rt, Ft+g
Sa2*t, pt,t(c), rt, Ft+g
St,A2+t, pa2*t, rt, Ft+g
St,A2+t, pt,t(c), rT, Ft+g
St,A2+t, pt,t(c), rt, Fbs+g
St,A2+t, pt,t(c), rt, Fa2+t+g
Recovery and resighting probabilities
The estimated recovery probability declined from ≈ 0·30 to ≈ 0·10 over the study period, with a peak in 1982 (Fig. 3). A model with a linear trend in r explained 39% of the year-to-year variation, but had a ΔQAIC of 32·41. Under the selected model, resighting probability for adults increased gradually from 1978 to 1984, and could be considered constant 1985–97 at a level of 0·75 (Fig. 4a).
There was substantial year-to-year variation in adult survival, ranging from 0·88 to 0·92 in good years to ≈ 0·70 in the worst years (Fig. 5). The annual mean for all age groups ranged from 0·74 to 0·95. A combination of the number of breeding pairs in northern Europe and mean winter temperature in Denmark explained 52% of the year-to-year variation in adult survival, although this model was not preferred (Table 2, Fig. 6). Adult survival was particularly low in 1995, the first cold winter in 9 years in northern Europe. At that time, the breeding population had more than tripled to 96 000 pairs since the last cold winter 1986–87. Adult survival was also low in 1978 and 1979, although confidence limits were wide because of the small sample size.
The selected model also included a quadratic effect of age. Figure 7 shows the age effect in 2 years 1993 and 1995; the curves were constrained to be parallel and the distance between them was determined by the year effect. This model indicated that survival was highest for 6-year-olds and declined markedly after c. age 12 years.
We calculated weighted mean adult survival from a model with no age effect on adult survival (Sa2*t, pt,t(c), rt, Ft+g). For the whole study period, the mean survival (weighted by the inverse of SE) was 0·879; when split into two periods, the means were: 1978–89: 0·889 1990–96: 0·858.
Fidelity to the colony
There was good support for using the group effect on fidelity probabilities as a measure of variation in ring loss (ΔQAIC for the model without group effect = 35·34). Estimated fidelity probabilities for the three quality groups followed the expected pattern (high > intermediate > low) and differences were substantial (e.g. 1991: Fhigh= 0·866, Fintermediate= 0·804, Flow= 0·792). In the following, we use only the estimates of fidelity from the group with the highest colour ring quality, since these were least biased by ring loss.
Fidelity, i.e. the probability of returning to the colony given survival, was low in the first years of the study, remained high (around 95%) during the 1980s and then declined after 1990 (Fig. 8). There was some indication of a difference in fidelity between first-year and older birds; ΔQAIC for a model including this effect was 1·26 (Table 2). Under this model, adults showed higher fidelity than first-year birds, although the difference was not very marked. The change in fidelity coincided with a decline in food availability and breeding success at Vorsø (see also Frederiksen & Bregnballe 2000).
Analysis of known breeders
A general model
The RELEASE goodness-of-fit test of the resighting data sets on breeders for model (φt,pt) (the Cormack–Jolly Seber model) was highly significant for both males and females (males: = 314, P < 0·001; females: = 385, P < 0·001). This indicated a severe lack of fit of this model with time-dependent parameters; however, almost all the lack of fit was concentrated in release test component 2.Ct, implying immediate trap-dependence (Pradel 1993). Goodness-of-fit of model (φt,pt*m), including trap-dependence, was almost acceptable for males ( = 65·5, P = 0·0194) and fully acceptable for females ( = 52·7, P = 0·263). We calculated a common variance inflation factor ĉ for this analysis as the summed chi-square statistic divided by its degrees of freedom: ĉ = 118·2/91 = 1·30, and proceeded using this value in model selection and estimation of precision.
We fitted a total of 54 models, that is all combinations of nine models for apparent survival and six for resighting. Table 3 shows the best eight models fitted (all models for which ΔQAIC was less than 10) and their ΔQAIC values. Several models including a sex effect on either apparent survival or resighting probability were highly ranked; the best model (φeur+w1,ps+t+m) contained additive effects of population size, and winter temperature on apparent survival and of year, sex and trap-dependence on resighting probability. However, this model was not convincingly better than several of the others (Table 3) and because of this model selection uncertainty we base our inferences on a set of highly ranked models (Burnham & Anderson 1998). No model including full time-dependence in apparent survival had a ΔQAIC less than 14.
Table 3. . Model selection for the analysis of resightings of breeders
According to the selected model, resighting probability of breeders seen in the previous year declined over the study period, but remained high (above 0·85) until 1995, dropping markedly in 1996 and recovering almost to the previous level in 1997 (Fig. 4b). The ‘trap effect’ was very strong; the probability of resighting a breeder not seen in the previous year was ≈ 0·50 over most of the study period (Fig. 4b). Males seemed to be slightly more likely to be resighted than females. However, the model ranking (Table 3) showed that it was difficult to conclude whether there was a sex effect on apparent survival and/or resighting probability.
Estimates of apparent survival from the analysis of breeders were similar to survival estimates from the combined analysis (Fig. 9). Confidence limits were much narrower, reflecting the fact that most of the lack of fit had been isolated in the trap-dependence of resighting probabilities. In this case, the combination of population size and winter temperatures explained 64% of the year-to-year variation in apparent survival. Models with a sex effect on apparent survival (Table 3) indicated that males may have had a slightly higher apparent survival than females, the difference being less than 1%.
Combined analysis vs. analysis of known breeders
The major advantage of the combined analysis was that it provided an estimate of adult survival that was unbiased by permanent emigration. This allowed us to conclude that part of the post-1990 decline in apparent survival was caused by an increase in emigration, although there has also been a decrease in ‘true’ survival (see below) – a conclusion that could not have been reached if only within-colony data had been used. For the same reason, the estimates of S will also be more useful than φ in future attempts to model population development on a European scale. Furthermore, this new method enabled us to study first-year survival in detail (Frederiksen & Bregnballe 2000) and to take the first steps in an attempt to understand the causes of variation in colony fidelity. This approach should be broadly useful for studies in which good data of both types have been collected, both as a means of getting unbiased estimates of survival, and as a way of quantifying emigration from a study area.
The only major weakness of the combined analysis was that it was not possible to include trap-dependence in the model. We suspect that S was biased slightly downwards because of this, and that precision would have been better if trap-dependence could have been included. The analysis of known breeders was justified by the opportunity to confirm that trap-dependence occurred and by the option of investigating sex effects on apparent survival.
Theoretically, the apparent survival probability φ is the product of ‘true’ survival S and fidelity F, and S should thus always be ≥φ. In this study, the theoretical relationship among the parameters almost held (Fig. 10); it should be noted that the constrained model for φ eliminated some of the year-to-year variation. However, φ was consistently higher than SF, and in 5 years out of 16 φ was higher than S. There are several plausible explanations for this:
1. breeders may have had a higher survival than other adults;
2. breeders may have had a higher fidelity than other adults;
3.φ may have been biased upwards because non-sexed birds were excluded from the breeder data sets; and
4.S may have been biased downwards because it was not possible to take trap-dependence into account in the combined analysis.
We find it likely that explanation (4) was the most important, because the ‘trap’ effect was quite strong, and it has been shown that neglecting trap-dependence can cause an important negative bias in estimated survival probabilities (Pradel 1993).
Apparent survival showed a consistent decline after 1988 (Fig. 9). Through the combined analysis, we were able to attribute this decline to a combination of declining survival after 1991 (Fig. 5) and an increase in emigration, starting in 1986 (Fig. 8). Even emigrants confirmed to be breeding in other colonies have been observed in the main study colony in the same season (J. Gregersen, unpublished data). Therefore, emigration has almost certainly been underestimated and apparent survival overestimated in this study. However, this should not affect survival estimates from the combined analysis.
Causes of variation in adult survival
In this study, a combination of population size in Europe and winter severity explained 52–64% of the year-to-year variation in adult survival of Danish cormorants. It was particularly clear that survival was unusually low in 1995, the only occasion during the study period when a severe winter coincided with a high population size, while a series of severe winters in the mid-1980s were not as clearly associated with increased mortality (Fig. 6). Apparently, severe winters cause extra mortality when density is high, or conversely, density-dependent mortality occurs in severe winters. Durell et al. (2000) also found that the highest mortality of oystercatchers (Haematopus ostralegus L.) occurred when a cold winter coincided with high population density.
A possible mechanism is that when population levels are high, the most attractive wintering areas reach carrying capacity, and the remaining birds are forced to winter in less optimal habitat where they are more vulnerable to the effects of severe winters. There are indications that some wintering areas have, indeed, reached carrying capacity during the explosive growth of the European cormorant population: Suter (1995) found that wintering numbers on at least some Swiss lakes stabilized, while numbers passing through on migration continued to increase. Similarly, the lack of a strong effect of severe winters in the early 1980s on adult survival may have resulted from most birds being able to find high-quality wintering habitat, where they were not severely affected by winter temperatures.
The sudden drop in resighting probability of breeders in 1996 may have been caused by some birds skipping breeding following the cold winter 1995–96. Variations in p over time or age may be an indication of underlying variations in the probability of breeding (Viallefont, Cooch & Cooke 1995; Pradel et al. 1997). Thus, in addition to affecting survival, severe winters may deprive individuals of energy reserves and make it more difficult for them to reach a body condition that allows them to breed.
We have not here identified all factors that may have affected adult survival, since more year-to-year variation was present than the above-mentioned factors accounted for. Particularly, the apparently lower survival in 1978 and 1979 was not explained. Variations in hunting mortality may have been involved; the cormorant was a legal quarry species in some European countries until 1980. Shooting mortality may also have contributed to the lower mean survival in the 1990s, because large numbers of cormorants have been shot as pests in several European countries (Italy: Baccetti 1996; Germany: Keller, von Lindeiner & Lanz 1998; France: J-D. Lebreton, personal communication).
Based on life-history theory, it is generally assumed that long-lived organisms should show senescence in survival (Stearns 1992). However, this has proved difficult to demonstrate in practice, mainly because sample sizes of old birds are inevitably low. A parsimonious way to investigate age-dependence in survival, both initial increases and declines in old age, is to model survival as a quadratic function of age. This approach has been used with success in sparrowhawks (Accipiter nisus (L.), Newton & Rothery 1997) and California gulls (Larus californicus Lawrence, Pugesek et al. 1995). In this study, we also found such an age effect (Fig. 7). Initial increases in survival beyond age 1 year were small (we did not include first-year survival in the constraint, since we assumed that it was discontinously lower than adult survival), but declines in old age were substantial. We cannot exclude the possibility that this decline was partly caused by ring loss; however, the loss of both colour and metal rings is unlikely to have been common.
Van Eerden & Munsterman (1995), when sexing birds at winter roosts by observations of morphology and behaviour, estimated a skewed sex ratio among adult cormorants in western Europe (3 : 2 in favour of males). They speculated that this might be due to differential mortality. Our results show little evidence of a large difference in survival between breeding males and females. Models including a sex effect on apparent survival estimated the difference as ≈1%; this may just as well have been caused by higher colony fidelity in males, which is expected in a species where males defend nesting territories (Greenwood 1980), as by differential survival. A skewed sex ratio may also have been caused by sex differences in prebreeding survival. However, we do not have the data to investigate this.
Implications for population growth and modelling
Motivated by the rapid population expansion and growing conflicts with human interests, two attempts have been made at modelling and predicting population growth of cormorants in Europe (Lebreton & Gerdeaux 1996; Bregnballe et al. 1997a). In both cases, input values for adult survival were mainly based on Kortlandt (1942), who estimated survival based on 2 years of resightings of Dutch cormorants, combined with demographic calculations. He found that survival was 0·78 for 1-year-olds, 0·84 for 2-year-olds, and 0·90 and 0·88 for adult males and females, respectively (later corrected to 0·88 and 0·86, A. Kortlandt, in litt.). Our results show that during the 1980s, adult survival was generally higher (up to 0·95, mean 0·89), whereas survival in some recent years has been lower than Kortlandt's estimates (down to 0·74). Likewise, they provide new information on the extent of year-to-year variation and on the frequency of severe mortality events, as well as on the strength of density-dependent relationships – all input that is needed for a predictive population model.
The Atlantic subspecies Ph. c. carbo has exhibited a lower population growth rate than Ph. c. sinensis (Debout et al., 1995); this is reflected in the lower survival estimates available for this subspecies. Wernham & Peach (1999) found that adult survival declined from ≈ 0·85 to 0·75 over the period 1965–94 in Britain, and Fiske & Røv (1997) estimated survival of 1-year-olds as 0·72 and adult survival as 0·80. Both studies of the Atlantic subspecies were based on recoveries. There is thus circumstantial evidence that the rapid growth phase of Ph. c. sinensis was caused partly by a period of unusually high adult survival, which would have a large effect on population growth rate.
Because of the extent of our study, with many years of data on a large sample of marked birds, and because the study period included periods of varying population sizes and winter conditions, we are confident that our results are more reliable than earlier estimates of cormorant survival. However, some features of our results may be unique to this data set. For instance, we did not find a marked increase in survival from 1-year-olds to adults (Fig. 7), in contrast to all previous studies and also to studies on the closely related shag (Aebischer 1986; Harris et al. 1994; Catchpole et al. 1998; Harris, Wanless & Elston 1998). Even an a posteriori fitting of a model with three age classes failed to show consistently lower survival of 1-year-olds than of adults in our study (data not shown). The absence of such a difference indicates that conditions for Danish immature cormorants were unusually good during the study period, perhaps related to the absence of hunting and to increased food availability due to eutrophication (de Nie 1995).
The most important results of our study can be summarized as follows:
1. The application of a combined analysis of recoveries and resightings produced many new insights in cormorant population dynamics. Unbiased estimates of adult survival showed that survival had been higher and more variable than previously believed, and that a combination of high population density and a severe winter presumably caused unusually high mortality in 1995/96. Emigration from the study colony increased since the mid-1980s.
2. There is strong circumstantial evidence that the rapid increase in the breeding population of Ph. c. sinensis 1970–95 was at least partly driven by unusually high survival of adult and immature birds, perhaps related to the virtual absence of hunting during this period and to the low population density at the beginning of the period.
We thank all the ringers and observers for their conscientious effort through the years, particularly Jens Gregersen. The Ringing Department at the Zoological Museum, University of Copenhagen, kindly supplied most of the ringing and recovery data. Tony Fox, Barbara Ganter, John Goss-Custard, Jean-Dominique Lebreton, Peter Rothery, Solveig Schjørring, Werner Suter and an anonymous referee made valuable comments on previous versions of this manuscript. Financial support for this study was received from the Danish National Forest and Nature Agency.
Received 20 July 1999;revisionreceived 23 October 1999