Survival of Svalbard pink-footed geese Anser brachyrhynchus in relation to winter climate, density and land-use



    1. CEFE, UMR 5175, CNRS, 1919 Route de Mende, 34293 Montpellier Cedex 05, France; National Environmental Research Institute (NERI), Frederiksborgvej 399, 4000 Roskilde, Denmark
    Search for more papers by this author

    1. CEFE, UMR 5175, CNRS, 1919 Route de Mende, 34293 Montpellier Cedex 05, France; National Environmental Research Institute (NERI), Frederiksborgvej 399, 4000 Roskilde, Denmark
    Search for more papers by this author

    1. CEFE, UMR 5175, CNRS, 1919 Route de Mende, 34293 Montpellier Cedex 05, France; National Environmental Research Institute (NERI), Frederiksborgvej 399, 4000 Roskilde, Denmark
    Search for more papers by this author

and present address: Marc Kéry, Swiss Ornithological Institute, 6204 Sempach, Switzerland. E-mail:


  • 1Global change may strongly affect population dynamics, but mechanisms remain elusive. Several Arctic goose species have increased considerably during the last decades. Climate, and land-use changes outside the breeding area have been invoked as causes but have not been tested. We analysed the relationships between conditions on wintering and migration staging areas, and survival in Svalbard pink-footed geese Anser brachyrhynchus. Using mark–recapture data from 14 winters (1989–2002) we estimated survival rates and tested for time trends, and effects of climate, goose density and land-use.
  • 2Resighting rates differed for males and females, were higher for birds recorded during the previous winter and changed smoothly over time. Survival rates did not differ between sexes, varied over time with a nonsignificant negative trend, and were higher for the first interval after marking (0·88–0·97) than afterwards (0·74–0·93). Average survival estimates were 0·967 (SE 0·026) for the first and 0·861 (SE 0·023) for all later survival intervals.
  • 3We combined 16 winter and spring climate covariates into two principal components axes. F1 was related to warm/wet winters and an early spring on the Norwegian staging areas and F2 to dry/cold winters. We expected that F1 would be positively related to survival and F2 negatively. F1 explained 23% of survival variation (F1,10 = 3·24; one-sided P = 0·051) when alone in a model and 28% (F1,9 = 4·50; one-sided P = 0·031) in a model that assumed a trend for survival. In contrast, neither F2 nor density, land-use, or scaring practices on important Norwegian spring staging areas had discernible effects on survival.
  • 4Climate change may thus affect goose population dynamics, with warmer winters and earlier springs enhancing survival and fecundity. A possible mechanism is increased food availability on Danish wintering and Norwegian staging areas. As geese are among the main herbivores in Arctic ecosystems, climate change, by increasing goose populations, may have important indirect effects on Arctic vegetation. Our study also highlights the importance of events outside the breeding area for the population dynamics of migrant species.


Global change, including change in climate, land-use and land-use relevant policy, is predicted to produce severe effects on animal and plant communities (Hughes 2000). As a response to climate change, the timing of annual events such as breeding and migration may be altered (Crick et al. 1997; Dunn & Winkler 1999; Butler 2003; Jenni & Kéry 2003), and the abundance of some species may change tremendously (Fogarty & Murawski 1998; Frederiksen et al. 2004). Ranges of individual species may move polewards, expand or decline in extent, and in mountain areas, move towards higher elevations (Parmesan et al. 1999; Thomas & Lennon 1999). Finally, even whole biomes, i.e. characteristic range assemblages, may move or change in size. Ultimately, such broad patterns must have an explanation in terms of changed survival, fecundity or dispersal rates of individual species or groups of species. However, the population dynamic mechanisms responsible for large-scale patterns have only rarely been elucidated (but see Crozier 2004).

Such mechanisms will likely be complex and may be direct or indirect. For instance, climate change may directly affect the demographic rates of a herbivore. Alternatively, it may have indirect effects, e.g. via an effect on vegetation, which (via food supply) has repercussions on the population of the herbivore. Or there may first be an effect on the population of a herbivore, which (via herbivore pressure) also has effects on the vegetation. Hence, climate change is likely to change the interactions between vegetation and its interacting animal species.

Arctic environments are fragile and slight changes in climate and/or interacting communities or populations of animals may have severe effects on them (Phoenix & Lee 2004; Ims & Fuglei 2005). Several species of geese are among the chief herbivores in Arctic environments (Kerbes, Kotanen & Jefferies 1990). Their interaction with plant communities may thus be particularly sensitive to climate change. World-wide, populations of many Arctic goose species have greatly increased during recent decades (see examples in Madsen, Cracknell & Fox 1999; Menu, Gauthier & Reed 2002; Frederiksen et al. 2004). To identify the causes underlying these population changes and to test for possible climate effects, detailed population dynamics studies are needed that relate components of population change such as survivorship to climate covariates. For migratory birds, reasons for population change need not only be sought on the breeding but also on wintering and migration staging areas (Newton 2004).

We analysed the survival of pink-footed geese Anser brachyrhynchus breeding on Svalbard. This population has almost tripled between 1965 and 2003 (Fox et al. 2005). Madsen, Frederiksen & Ganter (2002) suggested that the main cause for this may be changed conditions in terms of relaxation of hunting pressure and land-use on the wintering and migration staging areas rather than changes on the breeding grounds, whereas effects of climate changes remain to be evaluated. Svalbard pink-footed geese provide an attractive system for studying goose population dynamics: the population is essentially closed geographically and since the winter 1989/90, an extensive mark–resight scheme is underway in a large part of the area where the entire population concentrates at different times of the winter. We estimate sex- and age-specific annual survival rates and test for time trends and relationships with local and regional winter climate, goose density, land-use in the Danish wintering, and goose scaring practices in the northern Norwegian spring migration staging areas.


Study species pink-footed geese are medium-sized grey geese that breed in the Arctic and winter in temperate coastal zones (Fox et al. 1997; Madsen et al. 1999). The Svalbard population spends October–April along the North Sea coast between Belgium and Denmark. Data on the breeding cycle, migration, wintering, and hunting regulations can be found in Madsen et al. (1999). Hatching date on Svalbard is about 5 July, and goslings fledge after about 50 days, about 25 August (J. Madsen, unpublished data).

markrecapture data

The study period covered 14 winters from 1989/90 to 2002/03 inclusive. Subsets of these data formed the basis for earlier analyses (Madsen & Noer 1996; Madsen et al. 2002). Our larger data set comprised 1796 birds, 1790 of which were ringed in Denmark in the nine springs (15 March−4 May) of 1990–92, 1994–95, 1998, and 2000–02. Ringing took place in a small area where most of the population is concentrated at that time. Mean ringing date was 28 March. Birds were sexed and aged (first year vs. adult) and obtained a metal leg ring and a plastic neck-band with a unique three-digit code. Resightings were recorded during intensive campaigns over a very large part of the winter quarters in the Netherlands, Belgium and Denmark, as well as on Norwegian stopover sites (Madsen et al. 1999) during early October until mid-May (See for details.) Mean resighting date was 1 February, hence, the first survival interval after marking lasted 10 months (28 March−1 February), while all subsequent survival rates pertain to a full 12 month interval (1 February−1 February). No data are available on juvenile survival for the first 7 months of life after fledgling.

Individually marked birds were resighted up to 75 (mean 18) times per winter. During the study period, a total of 115 218 resightings of marked birds were made in later winters. A total of 71 birds were found dead (mostly shot). We did not analyse these recovery data because of its small sample size and because these data would mostly reflect mortality due to hunting rather than natural mortality. However, among birds recovered dead, some were reported to be seen alive repeatedly even after their date of death. This indicates that rings were sometimes misread. To account for misreads, we required a minimum of three observations of an individual during a given winter to record that individual as resighted. Obviously, ‘resightings’ of birds earlier recovered dead were not counted.

analysis of recapture and survival

We used capture–recapture models of the Cormack–Jolly–Seber (CJS) family (Lebreton et al. 1992) to estimate survival rates, and the slopes of their relationships with covariates. Under this model, mortality is confounded with permanent emigration, but we believe we are in fact estimating true rather than apparent survival. Resightings took place in a densely populated part of Europe in virtually the entire known wintering range, outside of which there are very few resightings.

Goodness of fit

The first step in such an analysis is to make sure that the data conform sufficiently well to the assumptions made by the model. A CJS model by the four groups (sex and age at marking crossed) did not fit the data (inline image = 558·50, P < 0·001). Most lack of fit was due to contributions of test 3.SR in birds ringed as adults and of test 2.CT in all four groups (see Lebreton et al. 1992 for an explanation). These six components added 94% to the total Chi-square goodness of fit statistic. The signed test statistics (see Choquet et al. 2004a) indicated that newly marked adults were seen again more frequently than adults marked earlier. The direction of the effect diagnosed by test 2.CT indicated strong trap-happiness, a form of recapture heterogeneity, for all four groups (Pradel 1993).

We accommodated these effects by introducing additional parameters into the model. Thus, for adults during the first period after marking we estimated a separate survival rate, which is denoted in the models by a2. For all four groups, we fitted a different resighting rate for winters after birds had been sighted and for winters they had not and thus accounting for trap dependence. This effect is denoted by m in our models (Pradel 1993). Our most general model was then φ,, where age denotes true age with two age classes (first year and older). It contains separate parameters for the survival rate of adults in the time interval immediately after marking, and trap dependence in all juveniles and adults as an effect of age (denoted by a2 for survival and by m for recapture rates). The goodness of fit statistic for this model is obtained by summing over the remaining cells (i.e. excluding cells for test 3.SR in birds ringed as adults and for test 2.CT in all four groups) as inline image = 32·44, P = 0·99. Hence, a model with a separate survival rate for adults during the first interval after marking (i.e. a transient parameterization for adult survival, Pradel et al. 1997), and a separate recapture rate for all individuals in those winters after they had been captured (i.e. a trap-dependence parameterization for recapture, Pradel 1993) fitted the data very well and was a valid starting point for further model optimization. Reasons for departures from the CJS assumptions are examined later.

Modelling recapture and survival rates

The goodness of fit analysis identified an adequate global model for recapture and survival rates. To obtain more parsimonious models and to identify covariates affecting survival rate, we reduced the complexity of this model in four steps. Our interest focused on survival rates so we started reducing the recapture part of the model to achieve maximum power to distinguish patterns in survival rates. First, we modelled patterns in recapture rate that could be explained by the factors sex, age and time. In our search for parsimony we reduced the global model by constraining parameters to be equal across groups or time periods. Second, we checked if a time trend in recapture rate could explain temporal variation in a more parsimonious way. Third, we retained the recapture part of the model as identified in steps 1–2 and applied the procedure from step 1 to the survival part of the model, i.e. we identified patterns in survival rates attributable to sex, age or time. Finally and fourth, we tested for relationships between survival rate and climate and other factors (see below) by introducing these factors as covariates to constrain the unstructured, annual variation of survival rates.

Parameter identifiability

CJS models with trap response can suffer from nonidentifiable parameters. We checked identifiability of all parameters in our most parsimonious model using the numerical version of a formal approach to detection of parameter redundancy (Catchpole & Morgan 1997) implemented in M-SURGE 1·7. [See Choquet et al. (2004b) and Gimenez et al. (2004) for further details.]

Decreasing survival estimates as an artefact of resighting heterogeneity

Resighting rates that differ among individuals (resighting heterogeneity) may induce an underestimation of survival rates and lead to the spurious detection of negative trends for survival estimates (Lebreton 1995). Carothers (1979) and Buckland (1982) found that this bias was most noticeable during the last few years of a study. To test if a negative trend in survival estimates may merely be the result of resighting heterogeneity, we examined the slope estimates of a trend in survival estimated over increasing lengths of time. If a negative trend estimate is constant and genuine, and not just an artefact due to heterogeneity, the slope estimate should not become less negative as data from more years are used to estimate it.

covariate analysis for climate, density, and land-use

We considered 20 covariates for local and regional climatic and other factors acting outside the breeding quarters (Appendix 1) and that were likely to have either a direct effect on survival or an indirect effect via resource availability. As a measure of local climate, we used precipitation sum and mean monthly temperature. Up to 1999/2000, these data were extracted from a GIS at the Wildfowl and Wetlands Trust (Mark O’Connell and Larry Griffin, pers. comm.) for the locations of the main Svalbard pink-footed goose wintering concentrations in Belgium, the Netherlands and Denmark and for the two staging areas in northern Norway (Trondheimsfjord and Vesterålen). Data for the remaining two winters were obtained from the respective national meteorological services for the following locations: Brugge (Be), Den Helder (NL), as a weighted-average over the stations in Jutland (DK), and for Trondheim (Trondheimsfjord/No) and Sortland (Vesterålen/No).

As measure of regional climate, we used start of spring in Denmark and in the Norwegian staging areas Trondheimsfjord and Vesterålen as determined on the basis of the normalized difference vegetation index (NDVI). Specifically, we used the Julian day when an NDVI threshold was reached that corresponds closely with ground data on the start of spring such as onset of leafing or start of plant pollen production (Høgda, Karlsen & Tømmervik 2006). As another large-scale weather index, we used the North Atlantic Oscillation (NAO), specifically the average winter NAO for December and January, and the extreme and the mean NAO during pink-footed goose migration (May, October). Extreme NAO during the migration months may be indicative of a higher frequency of storms inducing greater mortality. All climatic covariates were quantified during a time-window matching our survival analysis, i.e. the time-window during which climate covariates were measured did not extend over the limit of each survival interval defined by 1 February. This was the reason why we did not use the standard winter NAO index (e.g. Hurrell 1995), which, as a December–March average, sits squarely on these limits.

Land-use variables were selected on the basis of a recent review of major changes in land-use that has affected goose habitat use in the winter range (Fox et al. 2005). The major changes were attributed to increased use of winter cereals and a slight decrease of pastures in Denmark. In Vesterålen in northern Norway, farmers have organized a campaign to scare pink-footed geese away from sheep grazed pastures. Local campaigns were carried out in the mid-1990s, but from spring 1999, the campaign has been intensified to cover the most important part of the staging area, with consequent increased short-stoppings by geese and decreased accumulation of body stores prior to the final migration to the breeding grounds (Madsen 1995; J. Madsen, unpubl. data). We categorized years as no scaring (0), extensive scaring (1) and intensive scaring (2) years.

We used principal components analysis (PCA) of the 16 climatic covariates to combine their variation in two synthetic variables (F1, F2). To test for climate effects on survival, we tested F1 and F2 using an analysis of deviance (anodev) within the most parsimonious model for survival (Skalski 1996; V. Grosbois et al., unpublished). Let D0, DX and Dt be the deviance of models with constant annual survival, with annual survival constrained to be a linear function of covariate X, and with fully time-specific (i.e. annual) survival rates, respectively. In our anodev, the total temporal variation in survival (D0 −Dt) is partitioned into a part explained by a logistic-linear relationship with F1 or F2 (D0 − DX) and an unexplained residual part (DX − Dt). We constructed the following mean deviance ratio fX:

fX = [(D0 − DX)/(dfX − df0)]/[(DX − Dt)/(dft − dfX)],

where df0, dfX and dft are the associated degrees of freedom. To test the significance of covariate X we compared fX with an F-distribution on df1 = dfX − df0 and df2 = dft − dfX.

Based on our knowledge of pink-footed goose biology and the relationship between synthetic PCA axes and weather covariates, we predicted the expected direction of effect a priori (see later). This enabled us to test for an effect of climate in a one-sided test with increased power. We also used anodev to test the effects of four nonclimate covariates: population density, extent of winter wheat cultures and of pasture in western Jutland/Denmark and of the scaring index in northern Norway. We used u-care (Choquet et al. 2004a) for goodness of fit analyses, m-surge (Choquet et al. 2004b) to fit capture–recapture models and test parameter identifiability, and GenStat (Payne et al. 1993) for all other analyses.


patterns in resighting and survival rates

Resighting rate did not depend on age or sex but on whether a bird was already resighted during the previous winter (trap-dependence effect m) and on time (t) in an interactive fashion. The interaction m.t was present in all five most parsimonious models (Table 1), and the most parsimonious model contained just the m.t effects (Table 1). However, the unstructured time variation in the trap-dependence effect m could be parsimoniously described by a logistic-linear trend over time denoted as T (Table 2). The resulting, overall most parsimonious structure for resighting rate was Thus, there were different time trends in resighting rate for each combination of sex and two groups formed by birds resighted during the previous winter and those that were not.

Table 1.  Analysis of the resighting rate of Svalbard pink-footed geese in relation to factors age (with two age classes), sex and time. The survival part is φ for all models, with a first-year age-effect to account for a possibly different juvenile survival rate as well as for a separate adult survival for the interval following marking in adults (see text). Model notation follows Choquet et al. (2004b). Most parsimonious structure in bold
Resighting modelKDeviAICΔAICMeaning of model
  1. K – number of estimable parameters, Devi– model deviance, AIC– Akaike's information criterion, ΔAIC– difference of a model's AIC and the AIC of the model with the lowest AIC.·575764·5716·33Fully interactive effects of trap dependence, age, sex, and time + t 845589·465757·46 9·22As model 1, but time variation parallel for all combinations of trap dependence, age and sex
m + age + sex + t 825594·075758·07 9·83Additive effects of trap dependence, age, sex, and time 725616·285760·2812·04Interactive effects of trap dependence, age, and sex. No time effects
m + age + sex 705616·355756·35 8·11Additive effects of trap dependence, age, and sex. No time effects
m.age.t 965561·595753·59 5·35Interactive effects of trap dependence, age, and time. No sex effects
m.age + t 815594·525756·52 8·28As model 6, but time variation parallel for all combinations of trap dependence and age
m.t + age 905572·685752·68 4·44As model 6, but with a constant difference between juvenile and adult birds for all combinations of trap dependence and time
m + age + t 815596·905758·9010·66Additive effects of trap dependence, age, and time. No sex effects
m.age 695618·115756·11 7·87Interactive effects of trap dependence and age. No sex or time effects
m + age 695618·795756·79 8·55Additive effects of trap dependence and age·865752·86 4·62Interactive effects of trap dependence, sex, and time. No age effects + t 825590·015754·01 5·77As model 13, but time variation parallel for all combinations of trap dependence and sex
m.t + sex 925565·655749·65 1·41As model 13, but with a constant difference between males and females for all combinations of trap dependence and time
m + sex + t 815592·895754·89 6·65Additive effects of trap dependence, sex, and time. No age effects 705614·605754·60 6·36Interactive effects of trap dependence and sex. No age or time effects
m + sex 695616·405754·40 6·16Additive effects of trap dependence and sex
m.t 895570·245748·24 0·00Interactive effects of trap dependence and time
m + t 805594·525754·52 6·28Additive effects of trap dependence and time
m 685618·205754·20 5·96Trap dependence constant over sex, age, and time
Table 2.  Analysis of time trends in the resighting rate of Svalbard pink-footed geese. Those models in Table 1 that contain a time effect and had at least some support by AIC (ΔAIC ≤ 8, Burnham & Anderson 1998) were constrained by a logit-linear time trend. The survival part of the model and notation are as in Table 1
Resighting modelKDeviΔAIC ≤ 8AICMeaning of model
Previous most parsimonious model (from Table 1)
m.t895570·2405748·240 6·12Model 18 of Table 1 shown for comparison
Models with time trends:
m.age.T725606·2945750·294 8·17Separate time trends for juv and for ad that were seen during the previous winter and for those that were not (trap-dependence effect)
m.T + age715607·2595749·259 7·14Separate time trends for trap-dependence effect plus constant difference between juveniles and adults·1215742·1210·00Separate time trends for each combination of sex and trap-dependence effect + T715611·4675753·46711·35Interactive effects of trap dependence and sex plus an additive time trend
m.T + sex715602·1845744·184 2·06Separate time trends for trap-dependence effect plus a constant difference between sexes
m + sex + T705612·4255752·42510·30Additive effects of trap dependence and sex plus an additive time trend
m.T705606·7085746·708 4·59Different time trends for trap dependence
m + T695616·0185754·01811·90Additive effect of trap dependence and time trend

Keeping the resighting part of the model as identified previously (i.e., we then modelled survival during the first survival interval for adults (Table 3a). The most parsimonious structure assumed a constant survival rate for the first interval after marking (model a(1)) and the model with time dependence (model a(1).t) was almost tied (ΔAIC = 0·20). Keeping the a(1) structure for adult survival during the first interval after marking, we examined the remaining structure in survival (Table 3b). The best representation of survival rate was model a(1) + juv + adult.t that assumed time dependence in adult survival. As there was strong evidence for time variation in survival also during the first interval after marking (model a(1).t, Table 3a), we tried a model with constant difference of survival for the first interval after marking and afterwards and parallel time-variation superimposed (model (a(1) = juv) + adult + t, Table 3b). This model was the most parsimonious overall and was used to obtain estimates of resighting and survival rates and to test for effects of time-dependent covariates. According to the tests conducted in m-surge 1·7, all its parameters were identifiable.

Table 3.  Analysis of the survival rate of Svalbard pink-footed geese in relation to factors age, sex and time. (a) Model selection for adult survival during the first interval after marking, denoted by a(1). In addition, these models all contain a(2) effects for survival after the first interval after marking, denoted by a(2) for adults. (b) Final survival models. The resighting part of the model is the most parsimonious structure, ΔAIC, as identified in Table 2. Model 6 of 3a is repeated as model 1 in 3b. Notation is as in Table 1. calculated separately for each subtable
Survival modelKDeviΔAICAICMeaning of model
(a) Effects on adult survival only during first interval after marking
a(1).sex.t745594·125742·12 6·94Different male and female survival for the first survival interval at each time
a(1).sex.T545632·675740·67 5·49Different time trends for males and females
a(1).sex525633·155737·15 1·98Constant difference in male and female transient survival
a(1).t635609·385735·38 0·20Common, time-varying survival for males and females
a(1).T525633·015737·01 1·83Common time trend in transient survival common for males and females
a(1)515633·185735·18 0·00Common, constant transient survival for males and females
(b) Effects on survival of birds marked as juveniles, and for adults after the first interval after marking
a(1) + a(2)·185735·1819·39Different survival rates for each combination of age, sex and time
a(1) + a(2).age.sex135707·665733·6617·87Different survival rates for each combination of age and sex
a(1) + age.t305673·585733·5817·79Survival rate time-dependent for both ages independently
a(1) + juv + adult.t 225676·615720·61 4·82Survival rate constant for juveniles but time-dependent for adults
(a(1) = juv) + adult + t225671·785715·78 0·00Constant survival difference between the first interval after marking and later intervals with added, parallel time-variation
(a(1) = juv) + adult + T115703·755725·75 9·96Constant survival difference between the first interval after marking and later intervals with an added common trend
a(1) + juv.t + adult195705·955743·9528·16Survival rate time-dependent for juveniles but constant for adults
a(1) + age115709·175731·1715·38Different survival rate for juveniles and adults
a(1) + sex.t355662·505732·5016·71Survival rate time-dependent for both sexes independently
a(1) + sex115726·365748·3632·57Different survival rate for males and females
a(1) + t225691·635735·6319·84Survival rate different for each interval
a(1) + a(2)105726·835746·8331·04Survival rate constant over time

estimates of resighting and survival rate

Under this most parsimonious model, resighting rates were very high for birds seen during the previous winter and low for those not seen during the previous winter and they changed over the study period (Fig. 1). Survival rate during the first interval after marking ranged 0·88–0·97 and 0·74–0·93 afterwards (Fig. 2). Dropping time variation from this model yielded an estimated average survival of 0·967 (SE 0·026) for the first and of 0·861 (SE 0·023) for later survival intervals, regardless of sex and age. These survival rates are not directly comparable, as the former pertain to the approximate 10-month interval from 28 March to 1 February while the latter to a full 12-month interval. Exponentiating 0·967 by 12/10 resulted in an extrapolated annual survival rate during the first interval of 0·961.

Figure 1.

Resighting rates (with 1 SE) (a) for male and (b) for female Svalbard pink-footed geese during 14 winters (1989/90–2002/03) under the most parsimonious model (model 4a in Table 3) for those birds sighted during the previous winter (filled dots) and for those that were missed (open dots). Modelling this ‘trap-dependence’ effect takes account of resighting heterogeneity.

Figure 2.

Survival rates (with 1 SE) during 14 winters (1989/90–2002/03) for Svalbard pink-footed geese for the first survival interval after marking (10 months; filled dots) and for later intervals (12 months; open dots) under the most parsimonious model (model 4a in Table 3). For both groups, estimates under a logit-linear trend model (4b in Table 3) are also shown; however, the trend was not supported by the data (ΔAIC = 9·97) and is suspicious on grounds of the presence of resighting heterogeneity (see Figs 1 and 3).

A linear trend model showing a decline in survival rate over the study period was not well supported by the data (ΔAIC = 9·96; model (a(1) = juv) + adult + T, Table 2b). Furthermore, it was likely that the negative slope estimate in that model was an artefact owing to individual heterogeneity in resighting rates, as the estimated slope became less negative, when the number of years to estimate it was increased (Fig. 3).

Figure 3.

Relationship between the slope estimate of a time trend in survival and the number N of winters over which a trend is estimated, starting from the first winter (1989/90). Error bars are SEs, and the dashed line marks a zero slope, i.e. no trend. The slope estimates become less negative with an increasing number of years, which indicates that a negative slope estimate is artefactual and due to resighting heterogeneity (see text for explanation).

covariate relationships with survival rate

We considered 16 climate covariates that might affect survival (Appendix 1). There were substantial correlations among them (mean absolute correlation coefficient r: 0·28) with 17 among 120 pairs r > 0·5. Two synthetic variables (F1 and F2) from a PCA between them captured 56% of the total variation contained in the covariates (F1 37%, F2 19%). F1 was correlated mostly with warm and wet winters and an early spring in the Norwegian spring staging areas and F2 with dry and cold winters (Appendix 2). Hence, based on our knowledge of pink-footed goose biology, we clearly expected F1 would be positively and F2 negatively related to survival. The model with fully time-dependent survival rates (Table 3b, model (a(1) = juv) + adult + t) had a deviance of 5671·785 on 22 d.f. while the model with constant survival rates (not in Table 3b) had 5709·210 on 10 d.f. The total deviance due to time-dependent variation in survival was thus 37·425 on 12 d.f. Of this, F1 explained 23·4% and was on the verge of statistical significance (Table 4a; one-sided test for a negative effect of F1: F1,10 = 3·24; P = 0·051). F2 explained only 4·2% and was not significant (F1,10 = 0·59; P = 0·231).

Table 4.  Analysis of deviance (anodev) of the relationship between survival rate and climate quantified by the first (F1) and second (F2) PCA axis based on 16 climate covariates (see Appendix 1). The last three columns show parameter estimates for the slope of a linear trend, F1 and F2. Tests for F1 and F2 are one-sided as we predicted their direction a priori
  1. Notation: d.f. – degrees of freedom, Devi– deviance, %Devi–% deviance explained, MD– mean deviance, F and p – F-statistic and associated P-value.

(a) Analysis without time trend: overall, long-term variation in survival rates
F1 1 8·7623·4 8·763·240·0510·070 (0·024)
F2 1 1·59 4·2 1·590·590·2310·065 (0·025)0·052 (0·044)
Residual1027·0872·4 2·71     
(b) Detrended analysis: short-term, annual fluctuations in survival rates
Trend 1 5·4514·6 5·452·360·368−0·657 (0·240)
F1 110·4127·810·414·500·031−0·873 (0·291)0·070 (0·025)
F2 1 0·75 2·0 0·750·330·291−0·809 (0·295)0·070 (0·025)0·030 (0·036)
Residual 920·8155·6 2·31     

Climate as expressed by F1 and F2 did not change linearly over the time of our study (F1: slope estimate = −0·03, F1,11 = 0·14; P = 0·720; F2: slope estimate = −0·07, F1,11 = 0·92; P = 0·358). However, we conservatively tested for effects on survival of F1 and F2 also in a detrended analysis (Table 4b). Results remained virtually unchanged; F1 explained 27·8% and hence a significant part of the total temporal variation in survival rates (F1,9 = 4·50; one-sided test for negative effect of F1: P = 0·031) while F2 still only explained 2·0% and hence no significant part of the variation in survival. Hence, warmer winters and years with advanced springs on the Norwegian spring staging areas were associated with significantly higher survival rates of pink-footed geese; and these winter climate conditions explained a significant amount of variation in survival rates.

In contrast, there were no significant effects on survival of goose population density (F1,11 = 0·004; P = 0·950), land-use in western Denmark (extent of winter wheat: F1,11 = 0·255; P = 0·623, extent of pasture: F1,11 = 0·004; P = 0·954) or of the degree of scaring in north Norway (F1,11 = 0·252; P = 0·625).


We estimated the average annual survival rate of Svalbard pink-footed geese at 0·861 (SE 0·023). This estimate is fairly similar to those obtained by nine earlier studies that are summarized in Frederiksen et al. (2004; 0·79–0·86, unweighted average 0·83). We found a clear signal of winter and spring climate in winter-to-winter survival rates of Svalbard pink-footed geese during 1989–2002. The first PCA axis extracted from 16 winter climate covariates explained about one-quarter of the variation in annual survival. Warmer and wetter winters in an area between the Netherlands and Denmark, and advanced springs in Denmark and northern Norway, were associated with higher survival. We think that it is unlikely that winter weather had a direct effect on mortality rates, and that a causal chain ‘winter weather–vegetation (food base)–food intake gain rates–survival’ is more probable. Conditions towards the end of the wintering period in Denmark and in the staging areas in northern Norway may be particularly limiting. In late winter and spring, plenty of food is needed for Arctic-nesting geese to accumulate the resources required for both the home-bound migration as well as for the subsequent breeding (Ebbinge 1992).

We did not find density-dependent mortality, which is somewhat surprising given the large change of population size observed in this species (Fox et al. 2005). Further, there were no discernible effects of land-use in western Denmark as quantified by the extent of two crops important for the geese: pasture and winter wheat. This is also surprising, in that changes in land-use, along with hunting regulations, are generally assumed to be among the chief factors that have caused the population increases observed in recent decades (Madsen et al. 2002), although the major changes in hunting regulations causing increased survival took place already in the 1970s (Ebbinge 1991), i.e. before the start of our study. It is possible, though, that a more detailed analysis, where survival is related to more specific land-use categories in a finer temporal and spatial stratification, might still uncover such relationships. Alternatively, a mismatch between the duration of our mark–recapture study (started in the 1989/90 winter) and the full period of the population increase of the Svalbard pink-footed geese (starting in the 1960s) may prevent a full appreciation of the effects of land-use factors on survival. This is even more so, as some of the main changes in land-use may have taken place before 1989. Also, land use changes may have effects on population size that are mediated by reproduction rather than survival.

We could not detect any negative effects on survival of scaring practices on the spring staging areas in northern Norway. Several explanations are possible for this finding: (1) the effect of scaring has truly been small or nil; (2) a too small portion of the total population has been affected for an effect to be measurable at the population level; (3) the period of intensive scaring was too short to yield significant results (however, may be the cause of the tendency for decline in survival); (4) scaring affected demographic parameters other than survival; and (5) owing to a possible reproduction/survival trade-off scaring may actually have a positive effect on survival. For instance, reproduction may have been depressed in the subsequent breeding season (Madsen 1995; J. Madsen, unpubl. data) and this may lead to increased survival. An individual-based dynamic programming model predicts that intensive scaring will cause declines in food intake rates, which will ultimately have negative fitness consequences in terms of reproduction and survival (Klaassen et al. 2006). As the spring staging areas in northern Norway are considered important for the population dynamics of Svalbard pink-footed geese, a deeper look into the population dynamics consequences of scaring would be valuable.

The evidence for a trend in survival rates of Svalbard pink-footed geese is doubtful. A plot of survival rate estimates against time (Fig. 2) visually suggested a decline and this was also found in an earlier analysis of a subset of our data (Madsen et al. 2002). However, AIC much favoured a model with unstructured annual variation in survival rates over that with a trend (ΔAIC = 9·96), and a formal test of trend using anodev (Table 4b) did not show a trend to be significant either.

An important consideration when detecting a negative trend in survival is resighting heterogeneity. It is not widely enough appreciated that individual differences in resighting rates may bias low survival rates (Prévot-Julliard, Lebreton & Pradel 1998), especially towards the end of a study period (Carothers 1979; Buckland 1982), and that this may induce an artefactual negative trend in survival estimates over time. This is worrisome as some resighting heterogeneity is bound to be present in almost any animal population. Here, we introduce an informal test for a spurious decline of survival estimates. With increasing length of the time over which a trend is estimated, the trend estimates become less negative. This strongly suggests that a tendency for decline in survival is artefactual. It cannot even be ruled out that survival was in fact increasing.

There is a danger in studies such as ours. Climate, as quantified for instance by NAO (Hurrell 1995), is changing. Hence, everything else that changes simultaneously and for whatever reason may be declared as being affected by climate. We tried to control for this by relating climate, as quantified by our first PCA axis F1, both directly and indirectly to variation in survival. We did this by fitting F1 not only alone but also in a model containing a trend in survival to test for a relationship between year-to-year fluctuations in survival around a long-term trend and the putative climate driver of such a change. This ensures robust conclusions about a relationship between climate and survival (V. Grosbois et al., unpublished).

In addition, our analysis uncovered other interesting effects. Survival rates did not differ by sex, but by ‘trap-age’: geese had a higher survival rate during the first survival interval after marking than thereafter, even after accounting for unequal time of exposure (10 vs. 12 months), which may perhaps indicate some collar loss. Further, resighting rates differed by sex and were much higher for birds recorded during the previous winter than for those not recorded. Such individual heterogeneity in recapture rates has often been found (Grosbois & Thompson 2005; Sandvik et al. 2005). As geese are most often resighted rather than recaptured, such ‘trap-happiness’ may be due to the behaviour of the observers rather than the birds themselves. If observers repeatedly visit the same sites that harbour the same individuals, a positive autocorrelation of sightings will be expected. This may for example be the case in Denmark during spring, when geese are attracted to sites with bait grain. In particular, in one site, which is also the location of capture of geese in most years, resightings are made on a daily basis which will contribute to the trap-happiness effect. However, heterogeneous use of space by the geese may also contribute to resighting heterogeneity. Resighting rates also changed smoothly over time. This may be due to steady changes in observation pressure or to slowly moving site preferences of marked geese. The latter is the case in spring, where geese have increasingly used sites in northern Jutland in Denmark and mid Norway with lower observation pressure.


Global, including climate, change is predicted to drastically affect the population dynamics of many plant and animal species. Here, we show that warm and wet winters and springs were associated with higher survival rates in an Arctic goose species. This is consistent with the observed increase of that population over the past 35 years and the simultaneous trend towards warmer and wetter winters over much of northern Europe (Hurrell 1995). The increase of the Svalbard pink-footed goose population may thus partly be attributable to a progressive change of winter weather that provides more secure winter feeding conditions.

Geese are among the main herbivores in terrestrial Arctic ecosystems and their pre-nesting congregations and large breeding colonies may exert a considerable grazing pressure on the tundra vegetation (Kerbes et al. 1990). Our study thus suggests an interesting indirect effect of climate change on Arctic vegetation, one mediated by the population dynamics of a herbivore, similar to what Sedinger et al. (2006) found for reproduction instead of survival. It appears that climate change leads to changed survival rates and increasing populations of the geese and consequently to greater herbivore pressure in the Arctic tundra. This demonstrates that climate change may have both direct and indirect effects on biomes and populations of animals and plants. Our results also point out the importance of events outside the breeding season for the population dynamics of migratory bird species (Newton 2004; Schaub et al. 2005). This is a crucial insight for the management or conservation of such species, in that their population dynamics should not be analysed in terms of events in one of the visited areas only (e.g. only in the breeding area).


The analyses reported here would have been impossible without the help of Rémi Choquet. The long-term mark–resight project was partly funded by the Danish Research Councils, the Danish Ministry of Environment, the Norwegian Directorate for Nature Management, the Norwegian Research Councils (LANDRING Programme). The analysis was carried under the EU 5th Framework project FRAGILE (EVK2-2001-00235). We thank M. O’Connell and L. Griffin from WWT for providing data for analysis. For discussions, we thank them and V. Grosbois, R. Pradel, E. Rees and M. Trinder. For climate data, we thank K.A. Høgda (NORUT), M. O’Connell (WWT), I.M. Nordin (Met No) and I. Tombre (NINA). For comments on drafts of this paper, we thank M. Frederiksen, M. Lindberg, M. Klaassen, M. Schaub and an anonymous referee.