Invariant polar bear habitat selection during a period of sea ice loss

Abstract

Climate change is expected to alter many species' habitat. A species' ability to adjust to these changes is partially determined by their ability to adjust habitat selection preferences to new environmental conditions. Sea ice loss has forced polar bears (Ursus maritimus) to spend longer periods annually over less productive waters, which may be a primary driver of population declines. A negative population response to greater time spent over less productive water implies, however, that prey are not also shifting their space use in response to sea ice loss. We show that polar bear habitat selection in the Chukchi Sea has not changed between periods before and after significant sea ice loss, leading to a 75% reduction of highly selected habitat in summer. Summer was the only period with loss of highly selected habitat, supporting the contention that summer will be a critical period for polar bears as sea ice loss continues. Our results indicate that bears are either unable to shift selection patterns to reflect new prey use patterns or that there has not been a shift towards polar basin waters becoming more productive for prey. Continued sea ice loss is likely to further reduce habitat with population-level consequences for polar bears.

1. Introduction

Global climate change is expected to significantly alter habitat for many species [1,2], leading some to shift their ranges [3] and others to have their ranges significantly restricted [4]. To better anticipate how populations might be affected by climate change, researchers build models to predict how a population's habitat might shift and how much is likely to remain under different climate change scenarios (e.g. [1]). Typically these models are built based on a population's current pattern of habitat selection. If a species is capable of shifting habitat selection to accommodate these changes, however, projections based on current habitat selection patterns might lead to overly pessimistic predictions of future habitat conditions.

The capacity for populations to alter their habitat selection patterns in the face of climate-induced habitat changes is, therefore, an important factor to consider when predicting how climate change might affect a population [5]. Unfortunately, obtaining estimates of habitat selection for a population before and after climate change-related habitat alteration is challenging given that data often do not span the appropriate timeframe, or the predicted changes to habitat have yet to occur. This is not true, however, for polar bears (Ursus maritimus) in the Chukchi Sea (CS) which have exhibited some of the largest reductions of sea ice of any polar bear population [1] and for which space use data exist before and after these large habitat changes [6,7].

Some polar bear subpopulations have already exhibited negative effects of sea ice loss [8,9]. This pattern, however, is not currently ubiquitous across populations. In particular, CS polar bears do not appear to have experienced negative effects of sea ice loss that have affected an adjacent population [9]. It remains unknown whether polar bears in the CS have changed their habitat selection patterns while on sea ice to reflect the observed changes to ice and whether this might play a role in the differences observed between adjacent populations. There has been an increase, however, in the proportion of the population that uses land during summer and the length of time spent there [10], though the majority of bears remain on the ice year-round [10] even with reduced sea ice over productive waters [9]. It remains unclear why bears in the CS subpopulation have yet to experience negative population-level effects experienced by other subpopulations.

Polar bears are generally considered to be specialist predators that require sea ice to access prey and are, therefore, sensitive to sea ice loss [11]. There is some evidence to suggest, however, that polar bears might be capable of adjusting their habitat selection patterns as sea ice conditions change [12]. Habitat selection is a function of what is available to animals [13] but does not necessarily reflect what environmental conditions an animal prefers. Therefore, it is important to also estimate functional responses [14] in selection to address whether polar bear behaviour has exhibited an overall change with sea ice loss or is simply a reflection of changes in what is available.

Shifts in habitat selection with changes in sea ice conditions could imply that polar bear prey are also shifting their space use strategies (e.g. shifting to deeper water as ice retreats). While polar bear prey currently prefer the shallow waters over the continental shelf [15–17], if changing ice and currents in the Arctic Ocean made deeper waters more productive for prey, then we might expect prey to begin using these areas with greater frequency. This would imply that loss of sea ice in summer would have a more limited effect on polar bears than currently thought [18]. Conversely, a lack of change in habitat selection under significantly altered sea ice conditions could imply that prey species have not shifted their space use strategies as sea ice is lost, indicating an overall reduction in habitat for polar bears with the potential for significant population-level effects. Thus, understanding whether polar bear habitat selection patterns have changed from a period before widespread sea ice loss will enhance our understanding of how polar bears deal with sea ice loss and how continued sea ice loss might affect populations.

Our primary objective in this study was to determine how polar bear habitat selection patterns have changed for CS polar bears from a period before (1986–1994) and after (2008–2013) widespread sea ice changes [19]. We then used these results to determine how the distribution and area of highly selected habitat have changed between the two periods. We also were interested in determining whether the level of inter-individual variation in selection has changed through time to indicate if there has been an increase or reduction in the behavioural plasticity of polar bears to respond to change.

2. Material and methods

(a). Study area

The CS subpopulation is one of 19 recognized polar bear subpopulations [20]. Polar bears in the CS subpopulation primarily prey on ringed (Pusa hispida) and bearded seals (Erignathus barbatus) [9]. Other preys have been detected in their diet, but at much lower levels than ringed and bearded seals [9].

The CS is a highly dynamic sea ice system, with sea ice reaching its maximum extent in March, and its minimum extent in September. As a result of climate change sea ice dynamics have changed significantly in the CS. Over the past 30 years, sea ice retreat has occurred an average of 6 days earlier per decade in spring, and sea ice advancement an average of 7 days later per decade in the autumn [19]. These changes have resulted in an average of 44 days each summer (2007–2010) in the CS where there is less than 6250 km² of sea ice (greater than or equal to 50% concentration) over the continental shelf compared with 0 days between 1985 and 1993 [9]. Projections of future ice conditions in the CS indicate continued summer sea ice loss, with the potential for a five month ice-free period by 2040 but still completely ice covered between January and May [21].

(b). Capture and handling

We obtained location data from adult female polar bears captured in the CS population during two periods: 1986–1994 and 2008–2013 (hereafter, historic and contemporary periods, respectively). During both periods, polar bears were captured using a helicopter in March and April and immobilized with a dart containing M99 and an antagonist in 1986 and zolazepam–tiletamine (Telazol or Zoletil) in all other years. Captures during both periods primarily occurred in the southern CS [6,7,22], but during the historic period some captures occurred on Wrangel Island in Russia and in the northern Bering Sea [6,22]. Upon capture, we fitted bears with a satellite telemetry (1986–2010; Telonics, Mesa, AZ, USA) or global positioning system (GPS) collar (2011–2013; Telonics). Transmission intervals for collars varied, ranging from every 2–3 days for satellite telemetry collars [7,22] and 1–4 h for GPS collars [7]. Collars during the historic period had a battery life of approximately 14 months [22]. Battery life was longer for collars deployed during the contemporary period, but we fitted these collars with an automatic drop-off unit set for approximately 14 months post-capture.

(c). Habitat selection

We used methods in Wilson et al. [7] to format location data for analysis of annual step selection patterns between the two periods. Our analysis was restricted to those individuals with location data covering greater than or equal to 1 year, and individuals that did not enter a maternal den during a given year. Because our primary interest was in the annual selection patterns while on sea ice, we also restricted locations used in the analysis to those that occurred on sea ice. Additionally, we omitted locations from the date of a bear's capture in the spring until 1 June of the same calendar year to reduce non-random selection of the capture area. We filtered all locations from satellite telemetry collars with the algorithm developed by Douglas et al. [23] to retain the most accurate location from each duty cycle.

We estimated habitat selection at the scale of weekly steps [24]. To define what was available for polar bears to select, we followed the advice of Forester et al. [24] and sampled available points from empirical distributions of 7-day step lengths and turn angles. We developed separate empirical distributions for each month and each period. For each used location, we drew a random sample of 25 step lengths and turn angles to define the available locations in relation to the used location (hereafter referred to as a cluster). A sample of greater than 20 available locations has been shown to be sufficient for obtaining accurate estimates of habitat selection [25]. Similar to our restrictions on used locations, we only allowed available locations to occur on sea ice.

We used a similar set of variables to model habitat selection as the study by Wilson et al. [7], although we excluded variables that were not available during both study periods or that were not important for predicting polar bear use. This led us to include variables depicting sea ice concentration (resolution = 25 km²; http://nsidc.org/data/nsidc-0051), sea ice concentration squared, ocean depth (resolution approx. 16 km²; http://www.ngdc.noaa.gov/mgg/global/global.html), the standard deviation in sea ice concentration within a 100 km radius of a location (i.e. the approximate mean distance a polar bear moved in a week), the presence or absence of landfast ice (resolution = 25 km²; http://nsidc.org/data/g02172) and step length to reduce bias in selection estimates owing to not adequately accounting for what was available for animals to select [24]. We assessed collinearity between variables by calculating variance inflation factors (VIF), with a variable having VIF > 3 considered collinear with other variables [26]. We found no evidence of collinearity among our variables.

The study by Wilson et al. [7] also found that polar bears in the CS exhibited seasonal patterns of selection for sea ice concentration, ocean depth and sea ice variability. Given the dynamic nature of the CS and clear annual cycles of sea ice formation and melt, we desired to treat habitat selection as a dynamic process. Treating habitat selection as a dynamic process allows us to capture more subtle changes annually than one might obtain from seasonal models. Season-specific models could be especially problematic if seasons were not adequately defined, and the estimate was a mean response during a season that was not actually defined in a biologically meaningful way. While most studies estimate season-specific resource selection functions to capture annual patterns in habitat selection (e.g. [1]), this significantly increases the number of parameters needing to be estimated. Additionally, by estimating selection separately across seasons misses an opportunity to use information from the entire year to inform the overall selection patterns (i.e. the non-time varying mean response). Finally, an important assumption of any habitat selection study is that the probabilities of selection remain unchanged during the period of study [27], a problem that is completely overcome by modelling habitat selection as a continuous variable throughout a year. Thus, we allowed selection to vary across seasons by interacting covariates with trigonomic functions (electronic supplementary material). This method has been used in a variety of contexts to model annual patterns of habitat selection for polar bears [7,28], and diurnal patterns of habitat selection and movement for numerous other species [24,29,30]. Fitting trigonomic functions to model annual variation in selection is a simple form of functional data analysis which has a strong statistical underpinning [31].

It should be noted that even though allowing annual variation in selection imposes a trigonomic wave on these variables, if no annual variation exists in habitat selection for a variable, the coefficient estimates for the trigonomic terms will not differ from zero, indicating no annual variation in selection. Similarly, when one estimates season-specific habitat selection, an assumption of constant selection during a season is imposed on selection estimates, even if selection does vary over the period. As stated above, we feel that estimating annual habitat selection as a continuous process is the most appropriate method for these data given the strong annual sea ice dynamics in the CS and the additional analytical benefits noted.

There is often an implicit assumption in habitat selection studies that preference for a given habitat feature is proportional to its availability [14]. That is, selection will increase (or decrease) linearly with changes in availability. This has been shown to be incorrect for numerous species [14,32]. Given the changes observed to the sea ice in the CS during the study period, it is important that we incorporate estimates of polar bear functional responses into our model. This allowed us to determine whether polar bear preferences for different habitat features have been maintained even after changes in their availability. To incorporate functional responses, we calculated the mean sea ice concentration, ocean depth and variation in sea ice concentration of all available locations within a cluster. We then interacted the observed value for each variable (table 1) at each location with its associated mean value for that variable.

Table 1.

Coefficient estimates (and associated 95% CI) for population-level parameters of a habitat selection model for polar bears in the CS population for two periods: historic (1986–1994) and contemporary (2008–2013). (Covariates included in the model were sea ice concentration (Conc), ice concentration squared (Conc²), ocean depth (Depth), standard deviation in sea ice concentration within a 100 km radius (SDConc), landfast ice (LFI) and step length (Distance). Coefficients for covariates that interacted with time are denoted by ‘.c1’ or ‘.s1’ corresponding to that covariate's interaction with the cosine or sine time-wave functions, respectively. Coefficients representing functional response for a variable are denoted by a ‘Mean.’. Only the ‘Distance’ parameter had estimates with CI that did not overlap between periods.)

parameters	period
	historic		contemporary
	median	95% CI	median	95% CI
Conc	9.56	7.12 to 12.11	10.57	7.34 to 13.96
Conc.c1	3.10	1.07 to 4.98	4.56	1.96 to 7.24
Conc.s1	7.18	4.51 to 9.71	5.66	2.67 to 8.77
Conc2	−4.30	−6.52 to 2.36	−3.15	−5.57 to −0.70
Mean.Conc	−3.93	−6.86 to −0.39	−6.89	−10.57 to −2.70
Depth	5.07	0.75 to 10.57	2.73	−1.80 to 10.06
Depth.c1	2.31	−7.24 to 10.00	5.32	−4.01 to 16.16
Depth.s1	−7.68	−16.35 to 0.56	2.18	−5.24 to 12.16
Mean.Depth	4.73	−14.25 to 13.47	0.05	−10.33 to 11.88
SDConc	3.87	2.35 to 5.27	4.51	2.14 to 5.37
SDConc.c1	1.20	−0.88 to 3.46	3.17	0.44 to 5.80
SDConc.s1	−1.33	−3.55 to 0.95	−1.99	−4.84 to 0.59
Mean.SDConc	−7.06	−10.04 to −3.87	−2.28	−5.20 to 1.37
LFI	−0.29	−0.73 to 0.02	−1.01	−2.00 to −0.30
Distance	−8.46	−9.49 to −7.58	−4.93	−5.91 to −4.05

We scaled each variable so that they ranged between 0 and 1 by subtracting a value from the minimum value of all observations of that variable and dividing by the range of values for that variable [33]. This aided in model convergence and allowed assessment of the relative strength of selection across variables. We obtained separate coefficient estimates for each variable (i.e. mean response, temporal response and functional response) for current and historic periods. To estimate the posterior distribution of habitat selection parameters, we implemented a Bayesian version of conditional logistic regression [34] (electronic supplementary material).

(d). Predicted use

To determine where the resource units with a high probability of selection [27] were distributed across the study area during both periods (hereafter referred to as highly selected polar bear habitat), we derived maps of the predicted probability of selection within the study area (electronic supplementary material). We were also interested in determining whether the distribution of highly selected habitat throughout the year differed between the two periods. To accomplish this, we obtained (from the posterior samples) the mean monthly probability of selection for each pixel and rescaled such that the sum of pixels in the study area equalled 1. We did this for both periods. We then calculated the utilization distribution overlap index [35] which assesses how much overlap exists in the distribution of highly selected in the same month between the two periods. Values of 0 indicate no overlap, and values of more than or equal to 1 indicate a high degree of overlap of two utilization distributions [35]. We also compared the change in area of highly selected habitat with changes in the area of sea ice during September in the study area, obtained from the National Snow and Ice Data Center (https://nsidc.org/data/seaice_index/archives.html) to determine whether changes in habitat were simply the result of changes in sea ice distribution.

3. Results

A total of 89 (n = 57 historic; n = 32 contemporary) polar bears had sufficient location data (i.e. greater than or equal to 1 year) to be included in the study. We used 2412 polar bear locations for habitat selection modelling (n = 1527 historic, n = 885 contemporary) with an average of 33 (s.d. = 17) locations per polar bear for the historic period and 34 (s.d. = 24) locations per polar bear for the contemporary period.

(a). Habitat selection patterns

Our model performed well based on the cross-validation using the out-of-sample collection of points (r_Spearman = 0.86; confidence interval (CI) 0.81–0.90). Polar bears selected areas with higher, but more variable concentrations of sea ice; shallower water and lack of landfast ice (table 1). We detected no significant differences in population-level coefficient estimates between periods (figure 1) except the coefficient for step length, which indicated selection for larger step lengths for bears in the contemporary dataset (table 1). Variation among individuals did not differ between periods for selection of different variables (electronic supplementary material, table S1). Population-level annual patterns of selection for ice concentration, ocean depth and standard deviation in ice concentration all were not significantly different between periods (table 1 and figure 1). Polar bears exhibited positive selection for sea ice concentration year-round but showed the strongest selection during spring and weakest selection during late summer (figure 1).

Figure 1.

Plots of the annual coefficient estimates for ice concentration, ocean depth and the standard deviation in ice concentration. Lines depict median of the posterior distribution for each variable, with the shaded regions depicting the 95% CI of coefficient estimates for the historic (solid grey; 1986–1994) and contemporary (hashed; 2008–2013) periods. Coefficient estimates greater than 0 indicate a positive relationship with a variable, whereas estimates less than 0 indicate a negative relationship. Because ocean depth was measured against sea level, positive coefficient estimates represent selection for shallower areas.

We found no differences in the population-level functional responses between periods for sea ice concentration, ocean depth or variation in sea ice concentration (table 1). As the concentration of sea ice adjacent to a bear increased, selection for ice concentration decreased, eventually leading to avoidance of areas with highly concentrated ice (table 1). This result was similar to selection of areas with variable ice, with bears showing decreasing selection for areas of variable ice when in areas with increasing levels of sea ice variability (table 1). Conversely, polar bears exhibited increasing selection for shallower waters in areas with lower average depth (table 1), indicating polar bears always preferred to be in shallower water.

(b). Predicted use

We detected no differences in the amount of highly selected habitat from December through July between periods (table 2 and figure 2). There was, however, a significant reduction in the amount of highly selected habitat from August through to November during the contemporary period compared with the historic period (table 2 and figure 2). The largest reduction in the area of highly selected habitat occurred in September (table 2), representing a 75% reduction in the amount of highly selected habitat between the two periods. This compares to an average area of sea ice in the study area of 923 864 km² (s.d. = 268 687) during the historic period and 204 478 km² (s.d. = 198 074) during the contemporary period, a 78% reduction between periods. The spatial distribution of highly selected habitat (figure 2) exhibited a high degree of overlap between periods during all months except September, which displayed low correspondence (electronic supplementary material, figure S1).

Table 2.

Estimates of the median monthly area (1000 km²) of highly selected habitat available to polar bears in the CS population during historic (1986–1994) and contemporary (2008–2013) periods. (We defined highly selected habitat as any pixel having a relative probability of use greater than or equal to 0.80. Months with an asterisk (*) indicate non-overlapping 95% CI between the historic and contemporary periods.)

months	habitat area (1000 km2)
	historic		contemporary
	median	95% CI	median	95% CI
January	145	120–339	157	122–475
February	172	108–666	355	145–904
March	211	107–1053	302	130–1011
April	183	108–937	637	120–1127
May	151	106–561	288	110–826
June	129	107–378	179	97–374
July	129	103–253	121	80–209
August*	83	73–132	48	36–62
September*	74	68–89	19	15–24
October*	88	82–101	54	44–67
November	107	99–129	92	87–129
December	159	127–250	159	105–340

Figure 2.

Monthly mean predicted habitat value (i.e. relative probability of use) across the study area for polar bears in the CS population during historic (1986–1994) and contemporary (2008–2013) periods. Predicted maps were developed for the 15th of each month during each year of both periods. These monthly maps were then averaged within periods. Colours range from red (high predicted use) to blue (low predicted use). The solid line depicts the 100% minimum convex polygon of locations used in the analysis.

4. Discussion

Polar bears in the CS did not exhibit significant changes in their habitat selection patterns even after substantial sea ice loss in the region. This lack of change appeared to have no impact on the distribution or area of highly selected habitat during most of the year but led to a substantial decrease in highly selected habitat available to polar bears during summer. Loss of summer sea ice coupled with invariant habitat selection patterns across periods resulted in a 75% decrease in the area of highly selected habitat available to polar bears in September between study periods. Loss of habitat was similar to the overall loss of sea ice in the study area during September and highlights the potential for projected sea ice loss [21] to further decrease summer sea ice habitat for polar bears in the CS.

These results imply that it is unlikely that polar bear prey have shifted their space use to reflect reduced summer sea ice, forcing polar bears to find areas of habitat that have conditions similar to those present during summer before significant sea ice loss (i.e. pack ice over the continental shelf). Thus, it is unlikely that sea ice over the polar basin will provide sufficient access to prey to offset the overall reduction in summer sea ice. This is highlighted in a recent study that found a 50% reduction in the kill rate of seals in the Beaufort Sea by polar bears between the mid-1980s and the mid-1990s [36]. That the reduction of highly selected habitat was directly related to the loss of sea ice in summer indicates that future summer sea ice loss will continue to reduce the availability of highly selected summer sea ice habitat for polar bears, a period that is considered to be a primary driver behind observed declines in other populations [9].

Our observation that loss of highly selected polar bear habitat in the CS is restricted to summer conforms with predictions that CS sea ice loss will mainly occur during summer [1,21]. During this same period, there has been a significant increase in the number of bears summering onshore [10] even though opportunities to forage terrestrially are probably limited [37]. This implies that opportunities to hunt while on sea ice during summer are probably becoming fewer or that fewer polar bears can access productive areas in summer given the reduction in highly selected habitat area. There is active debate on what role terrestrial foods will play to compensate for lost hunting opportunities on sea ice owing to increased time spent on shore [37,38]. Thus, if polar bears are able to compensate for lost hunting opportunities with terrestrial foods [38], then increased land use might not be overly problematic for the CS population. If, however, terrestrial foods will never be able to provide sufficient energy to compensate for lost seal hunting opportunities [37], then polar bears in the CS might begin to experience negative population-level effects of sea ice loss in the near future.

Our observed lack of change in habitat selection patterns is consistent with that found for the Eastern Greenland (EG) polar bear subpopulation before and after sea ice loss [39]. Only during winter did selection for sea ice concentration differ between periods, with EG polar bears exhibiting stronger selection for ice concentration in the 2000s than in the 1990s [39]. Our model did not suggest a similar difference, but might have been masked by the inclusion of functional responses in our model or by not creating season-specific models as in Laidre et al. [39]. Polar bears in the EG and CS subpopulations occur in areas with significantly different ice dynamics [7,39]. That subpopulations residing in areas with different annual sea ice dynamics both showed a general invariance in habitat selection patterns with sea ice loss suggests that polar bears, overall, are unlikely to adjust habitat selection to changes in sea ice.

Polar bears in the CS displayed functional responses in their selection based on what was locally available to a bear. Functional responses have been reported previously for polar bears [12] indicating shifts in preference for different environmental attributes given changes in their availability [14]. Similar to our results, Mauritzen et al. [12] found that the probability of using high sea ice concentrations decreased when bears were located in areas with high sea ice concentrations, but exhibited the opposite pattern when bears were in areas with low concentrations of sea ice. Polar bears live in a highly dynamic environment, requiring them to respond by constantly moving to areas with conditions suitable for finding and capturing prey [16,36]. The lack of change in functional responses between periods implies that polar bears are making similar decisions in response to their local conditions as they did before sea ice changes.

Intraspecific variation in the habitat selection parameters also did not change between the two periods. It is unclear whether the levels of intraspecific variation represent high or low variability, because few studies have examined intraspecific variation in habitat selection. The results imply, however, that variation in polar bear space use strategies remained unchanged between periods. Differences in site fidelity [40] and space use strategies [41] have been documented for several populations which could lead to intraspecific variation in habitat selection. Such behavioural variation, however, has not yet been observed for CS polar bears. Differences in habitat selection among individuals could be a result of different hunting strategies such as waiting at breathing holes for seals, stalking prey or seeking lairs to capture seal pups [42]. Additionally, intraspecific variation could be related to differences in space use patterns between adult females at different life stages (e.g. with and without cubs; [41]).

That polar bears did not alter their habitat selection patterns after significant sea ice loss is not entirely surprising given that polar bear habitat selection is thought to be largely driven by prey availability [43]. Unless seals alter their preferences for sea ice, polar bears are likely to maintain habitat selection preferences. Indeed, there is concern that ringed seals might be unable to adjust to changing sea ice conditions [44]. It is also possible that owing to data limitations, we have missed an important variable (e.g. ice thickness, ice roughness) for which polar bears have shifted selection over time. Additionally, a relatively small sample of polar bears could have reduced our ability to detect differences in selection between time periods if they existed. Even if this were the case, our results imply that the magnitude of change (if it exists) is not large.

Future research should focus on understanding how much prey is available to bears on sea ice while it is at its annual minimum extent and how this availability has changed through time. If seals have never been widely available during summer, then the consequences of reduced summer sea ice habitat might be less severe. Alternatively, the benefits of highly selected habitat during summer might have more to do with earlier access to prey, in which case, a further northward retreat of sea ice could be even more problematic for bears as it not only prolongs the periods of low or no prey access, but also decreases the time available to hunt productive waters. Future research should also focus on relating polar bear space use to ice metrics that are more directly related to seal presence and abundance (e.g. presence/absence of pressure ridges; [16]).

Although we did not find any significant changes in habitat selection between time periods, our analysis highlights the importance of determining whether habitat selection has remained static after climate-related habitat changes have occurred. Our work also provides a framework for the type of analysis that would be required for other species to determine whether current habitat selection models are suitable for extrapolating to future habitat conditions.

Ethics

This research was permitted under the Marine Mammal Protection Act and Endangered Species Act under US Fish and Wildlife Service permit MA046081 and followed protocols approved by Animal Care and Use Committees of the US Fish and Wildlife Service.

Data accessibility

The datasets supporting this article have been uploaded as part of the supplementary material.

Authors' contributions

R.W., E.R., K.R. and M.S.M. aided in the collection and design of field data collection, and conceived of the study. R.W. designed the study, conducted statistical analysis and drafted the manuscript. All authors edited the manuscript and gave final approval for publication.

Competing interests

We have no competing interests.

Funding

Funding for this study was provided by the US Fish and Wildlife Service, the Changing Arctic Ecosystems Initiative of the US Geological Survey Ecosystems Mission Area, a Coastal Impact Assistance Program grant, the National Fish and Wildlife Foundation and the Detroit Zoological Association. The Alaska Department of Fish and Game, North Slope Borough, the Alaska Nanuuq Commission, National Park Service, the communities of Point Hope and Kotzebue, Selawik National Wildlife Refuge, Red Dog Mine, Teck Alaska Inc. and Nana provided significant support for this research. This study benefitted from a course on the use of Bayesian methods in Ecology supported by the National Science Foundation (DEB 1145200).