INTRODUCTION
Observed and projected Arctic warming is approximately twice the global mean (Overland et al., 2019), and record low Arctic sea ice extent has occurred for more than a decade (Stroeve & Notz, 2018), increasing the risk of disturbance to marine wildlife from vessel traffic (Post et al., 2013). The Chukchi Sea is one of eight marginal Arctic seas, and its vessel traffic is currently modest. However, most growth in maritime transport operations will likely occur in locations such as the Chukchi Sea that are undergoing rapid lengthening of the ice-free season and lie along routes leading to the northeast and northwest passages (Silber & Adams, 2019).
The Pacific walrus (Odobenus rosmarus divergens) is an Arctic/subarctic species of conservation concern that uses sea ice for important life history events and activities, including calving and resting. The population winters (November–April) and breeds (January–March) within the pack ice in the Bering Sea (Fay, 1982). When the pack ice has retreated north from the Bering Sea, females and young, as well as some mature males, spend May–October in the Chukchi Sea (Fay, 1982). They inhabit the marginal ice zone (areas of water with broken, floating ice) over the continental shelf for most of this season, and they have hauled out in large numbers on land in northwestern Alaska between early August and late October in recent years (2007, 2009–2011, and 2013–2022) when ice has retreated entirely off the continental shelf (Fischbach et al., 2016; Jay et al., 2012).
Vessel disturbance could displace walruses from preferred habitats, interfere with indigenous subsistence harvests, and affect walrus activity budgets (proportions of time spent foraging, swimming, or resting). Changes in activity budgets can lead to changes in body condition and possibly changes in reproductive and survival rates (Udevitz et al., 2017). Lactating females and dependent young summer in the Chukchi Sea and have higher energetic needs, making them most likely to be affected by altered activity budgets (Jay et al., 2012). Autumn body condition of adult female walruses may be reduced by mid-century due to sea-ice-loss-induced changes in activity budgets alone (Udevitz et al., 2017), and by that time, if not already, most of the Arctic Ocean basin will likely be accessible to vessels capable of limited ice-breaking for eight months of the year (Sokolova et al., 2020; Stephenson et al., 2011) and accessible to open water vessels for at least two to four months of the year (Melia et al., 2016). Thus, increased vessel disturbance has potential to compound the energetic effects of sea ice loss, particularly if it decreases the amount of time spent foraging or hauling out (resting).
Wildlife managers in countries of the circumpolar north need information on how vessels and other anthropogenic ocean activities may affect walrus activity and energy budgets. Importantly, this includes information on how walruses respond to vessels when they are in-water foraging, in addition to when they are hauled out resting. Vessel effects on walrus activity budgets are a pressing concern, and due to lack of empirical data, the U.S. Fish and Wildlife Service has been limited to evaluating these effects on walruses using expert opinion (Harwood et al., 2019).
Notably, the most common method of assessing pinniped behavioral responses to vessels (shipboard line of sight) is insufficient to quantify activity and energy budgets. Shipboard line-of-sight studies detect pinnipeds flushing off ice or land into water; however, line-of-sight observations alone cannot determine sustained in-water behavioral responses, and therefore cannot be coupled with bioenergetic models to quantify changes as a result of disturbance in energy required or acquired by the animal (Mikkelsen et al., 2019). Furthermore, distances at which disturbances have been observed to cause pinnipeds to stop resting and flush into the water are probably smaller than distances at which pinnipeds respond to vessels if they are already in-water foraging because of differences in the properties of sound, and its perception by pinnipeds, in air versus in water (Kastak & Schusterman, 1998).
Distances at which walruses flush off ice into water have not been well quantified, but these flushing distances for other ice-associated pinnipeds can be short, often <1 km (Jansen et al., 2010; Lomac-Macnair et al., 2019; Wilson et al., 2017). However, distances at which walruses in water might respond to vessels are unknown. This information is similarly lacking for other pinnipeds (Erbe, Dähne, et al., 2019; Erbe, Marley, et al., 2019), and we are not aware of any more recent work that addresses this information gap. Furthermore, sound propagates much more effectively in water than in air, and preliminary behavior disturbance thresholds set by the National Oceanic and Atmospheric Administration (NOAA) and applied to walruses are half as loud in water than in air (NMFS, 2021; Scholik-Schlomer, 2015). This combination of better sound propagation in water and perhaps lower levels of sound being needed to cause in-water disturbance means that longer disturbance distances should be considered for in-water walruses than for hauled out walruses.
Determining a potential response distance of in-water walruses to vessels is further complicated because vessel type and activity, and physical conditions (such as water depth and ice cover) all contribute to the distance that sound propagates in water (Erbe, Marley, et al., 2019; Halliday et al., 2017). For example, broadband noise encompassing the sensitive hearing range of walruses (Reichmuth et al., 2020) from a 98-m-long icebreaker in the Beaufort Sea (not necessarily breaking ice) only exceeded the 120 dB re 1 μPa preliminary in-water behavior disturbance threshold set by NOAA (Scholik-Schlomer, 2015) at distances up to 2 km. In contrast, a smaller, 80-m-long tanker in the Beaufort Sea produced sound that exceeded the 120 dB re 1 μPa threshold at distances up to 10 km under certain physical conditions, and up to 52 km under other physical conditions.
The available walrus habitat-specific information comes from an area of the northeastern Chukchi Sea close to the Hanna Shoal, a walrus foraging hotspot (Jay et al., 2012). There, a drilling operation produced broadband sound levels (also encompassing the sensitive hearing range of walruses) of ~120 dB re 1 μPa at 16 km away, compared with a peak level of ~110 dB re 1 μPa of wind-driven ambient noise (Quijano et al., 2019). Some of the loudest sounds associated with these drilling operations were from support vessels, and the spatial distribution of support vessels strongly influenced the acoustic footprint of drilling-related activities (Quijano et al., 2019).
Our goal was to assess sustained walrus behavior responses to vessels when the walruses were in water (either foraging or not) at distances that might be relevant to in-water walruses, and if sample sizes permitted, also at shorter distances that would likely be relevant to hauled out, resting walruses. Technological advances with sound and movement recording tags (e.g., DTAGS) are beginning to allow measurement of sustained pinniped reactions to vessel noise for a few individuals (Mikkelsen et al., 2019), but such methods are currently infeasible for walruses due to capture and handling constraints. By contrast, walrus telemetry data are plentiful (as they are for many other pinnipeds) and automatic identification system (AIS) vessel tracking data are readily obtained; thus, these data currently offer the only opportunity to explore the effects of vessel disturbance on walrus activity.
We explored the potential to leverage existing walrus Argos satellite telemetry data (i.e., hourly behavior and location) from 2012 to 2015 (Citta et al., 2018; Jay et al., 2022, 2017; Udevitz et al., 2017) and AIS vessel tracking data to quantify possible effects of vessel exposure on walrus activity budgets and to determine the feasibility of combining telemetry data and AIS data for marine mammals, in general. Because our telemetry data were not collected specifically to study walrus responses to vessels, they are not ideally suited to our questions. However, logistical and financial costs associated with assessing walrus movements and behavior, combined with the pressing need for information on walrus response to vessel noise (Harwood et al., 2019), warranted a thorough evaluation of currently available data.
We had three objectives: (1) quantify measurement error and develop a filter to prevent misclassification of vessel presence versus vessel absence that would otherwise result from the substantial location error inherent to Argos data; (2) test for effects of vessel exposure on walrus behavior using both the misclassification filter and particularly sensitive statistical techniques that account for confounding covariates and thus help adjust for the small sample size of walruses exposed to vessels; and (3) conduct power analyses to determine sample sizes needed to evaluate effects of vessel exposure on walrus behavior at successively closer distances, and either conduct those analyses if power was sufficient or determine if a different study design might be required for further evaluation of walrus behavior responses to vessels.
METHODS
Walrus telemetry data
We obtained walrus location and behavior data by attaching 218 satellite-linked tags to walruses resting on sea ice or land (Appendix S1: Section S1) and studied their locations and behavior in the Chukchi Sea during the months of June through December 2012–2015 (Figure 1). Tag transmissions were received by satellite-borne Argos sensors (Collecte Localisation Satellites, 2011) and processed to estimate tag locations and location accuracy. Tags also collected conductivity and pressure readings that were processed onboard (Fischbach & Jay, 2016; Appendix S1: Section S1) and transmitted to the satellite as hourly behavior states relevant to bioenergetics models: hauled out (resting), foraging (always in water and mostly underwater), or in-water not foraging (mainly swimming at or near the surface). Because Argos location estimates are prone to substantial error and received irregularly in time, we filtered them with the Douglas Argos-filter algorithm (Douglas et al., 2012) and used a continuous-time correlated random walk model (Johnson et al., 2008) to estimate an hourly location (and location error) to associate with the hourly walrus behavior (Appendix S1: Section S2).
[IMAGE OMITTED. SEE PDF]
Vessel tracking data
Vessel locations were obtained through a satellite-relayed marine AIS that monitors marine traffic via transmissions from participating vessels. We obtained satellite-collected AIS data processed by Exact Earth (Cambridge, Ontario) for 2012–2014 through the Marine Exchange of Alaska (), and for 2015 through Harris Corporation Maritime Geospatial Solutions (Melbourne, Florida).
Frequent AIS transmissions were received (typically many per minute), and the overwhelming majority provide precise locations. Nevertheless, we applied a speed filter (Mcconnell et al., 1992) to remove implausible locations, and in the unusual situations with substantial time gaps, we ended a sequence if the gap was >30 min long. We then regularized vessel locations within each sequence to predict a vessel location for every minute. We later combined these by-minute vessel locations with predicted hourly walrus locations to determine if a walrus in any given hour was vessel-exposed.
Statistical analyses
Objective 1: Quantify measurement error and develop a misclassification filter for vessel presence versus vessel absence
Argos-based location estimates are imprecise (even after applying an Argos-filter algorithm and movement model); thus, we expected imprecision would make it impossible to distinguish walrus to vessel distances that were only a few kilometers apart. In addition, measured distances between objects are biased high when locations for one object (vessels) are measured with much greater precision than for the other object (walruses); thus, we also expected bias to be a substantial problem at close distances but not at larger distances. We quantified this imprecision and bias by bootstrapping walrus location errors (Appendix S1: Section S3), confirming that we could not use distance from walrus to vessel as a continuous predictor. We, instead, needed to define vessel presence near a walrus as a categorical covariate that was specifically developed to avoid misclassification due to either imprecision or measurement-error-induced bias.
To accomplish this, we defined vessel presence versus vessel absence based on two different measured radii from an hourly walrus location, where a measured radius included all distances up to that radius, as they were observed with measurement error. We defined vessel presence as ≥10 min of vessel locations within a 15-km measured radius of an hourly walrus location: the 15-km measured radius was a smaller, but similar number to the 16-km distance to which drilling and vessel noise near the Hanna Shoals walrus feeding hotspot met or exceeded the 120 dB re 1 μPa in-water behavioral disturbance threshold set by NOAA (Quijano et al., 2019). We defined vessel absence as no vessel locations within a 20-km measured radius of an hourly walrus location. To avoid misclassification, we removed any hourly walrus locations that did not meet either set of criteria. We chose the 10-min criterion to ensure that vessel presence was more than transitory.
By employing one measured radius to define vessel presence and a second, 5-km larger measured radius to define vessel absence, we removed observations where vessels appeared to be within a 5-km-wide “doughnut” around the hourly walrus location. This removed almost all points where vessel presence or absence would have been misclassified due to location error. In other words, the 5-km-wide doughnut between 15 and 20 km from a walrus served as our misclassification filter. We also used bootstrap simulations (Appendix S1: Section S3) to estimate the true vessel presence–absence boundary: the true radius around a walrus within which a vessel would be considered present, and beyond which it would be considered absent, had there been no walrus location error.
Objective 2: Test for effects of vessel exposure on walrus behavior
We followed a five-step process. In Step 1, we defined the dependent variable, response-hour walrus behavior, to be conditional on behavior in the previous hour, and in Step 2, we defined the independent variable of interest (vessel exposure) to capture the first hour in which a walrus encountered a vessel. Information obtained from Objective 1 (Quantify measurement error and develop a vessel presence–absence misclassification filter) was essential to the definition of vessel exposure. In Step 3, we defined covariates previously shown to predict walrus behavior, in Step 4, we matched covariates, and in Step 5, we tested for an effect of vessel exposure on walrus behavior using multinomial logistic regression. These five steps are explained below under their subheadings.
Step 1: Define dependent variable as conditional walrus behavior
Our dependent variable was conditional walrus behavior. In other words, we wanted to know the probability a walrus would be foraging, in-water not foraging, or hauled out in the response hour, given (i.e., conditioned on) its known behavior the previous hour. Conditional behavior is an appropriate dependent variable because properties of sound and its perception by pinnipeds differ depending on whether the sound is heard through the air (e.g., previously hauled out) or through the water (e.g., previously foraging) (Kastak & Schusterman, 1998).
To implement this conditional approach, we generated one data set for each of the three categories of previous behavior (foraging, in-water not foraging, and hauled out), where each data set contained response-hour behavior as the dependent variable, vessel exposure as the independent variable of interest, and covariates (air temperature, wind speed, haul-out substrate, benthic biomass, and bout hours of previous behavior). The following analytical details are identical for the analysis of each of the three data sets.
Step 2: Define vessel exposure, the independent variable of interest, to capture the first hour a walrus encountered a vessel
The walruses we tracked encountered vessels infrequently and at low densities, thus any sufficiently disturbed walrus could likely swim away to an area without vessels. This suggested that any walrus with a vessel present nearby for multiple consecutive hours was likely undisturbed during those hours. As a result, including multiple consecutive hours of vessel presence in our analysis would have biased our results toward observations where walruses were undisturbed. To control for this, we analyzed only the first hour in which a walrus had a vessel present nearby, and we refer to these first hours of vessel presence as vessel-exposed. To implement this, we defined vessel exposure using consecutive pairs of hours: a response hour was vessel-exposed if a vessel was present near the hourly walrus location in the response hour (≥10 min of vessel locations within a 15-km measured radius), but not in the previous hour (no vessel locations within a 20-km measured radius). A response hour was unexposed if no vessel was present near the hourly walrus locations in both the response hour and the previous hour. Any hours that were not classified as either exposed or unexposed were removed from the data set.
Step 3: Define covariates
We included four environmental covariates and one behavioral covariate previously shown to predict walrus behaviors (Jay et al., 2017; Udevitz et al., 2009). The two covariates describing weather conditions at response-hour walrus locations (air temperature and wind speed) were obtained from the National Oceanic and Atmospheric Administration (). They were based on the North American Regional Reanalysis (NARR), which provides weather metric estimates at 3-h intervals on an approximately 32.5 × 32.5 km grid (Mesinger et al., 2006). Each metric was interpolated by inverse distance weighting of the points closest to the response-hour walrus location, first in space, and then in time.
The remaining two environmental covariates (haul-out substrate and benthic biomass) described resource availability at response-hour walrus locations. Haul-out substrate was a 4-category variable that identified whether and what kind of haul-out substrate(s) were available to a walrus within a 5-km measured radius. The four categories were ice only, land only, ice and land, and no ice or land (Appendix S1: Section S4).
We indexed biomass of walrus prey at each response-hour walrus location from a predictive surface of benthic macrofaunal biomass (ln[grams of carbon per square meter]). The predictive surface was generated with universal kriging based on van Veen grab samples collected at benthic sampling stations during 2008–2012 (Beatty et al., 2016; Jay et al., 2017). Because benthic sampling was restricted to a subregion of the Chukchi Sea (Figure 1), and we lacked a defensible way to extrapolate beyond that subregion, we indicated that benthic biomass values were missing (NA) for response-hour walrus locations outside this region.
We included a single behavioral covariate: bout hours of previous behavior. Walrus behaviors occur in bouts (consecutive hours of the same behavior), and the number of consecutive hours for which a walrus has already engaged in the same behavior influences the probability it will switch to a new behavior. To account for this, for each response hour, we not only identified the behavior associated with the previous hour (previous behavior) but also the number of consecutive hours a walrus had already engaged in that behavior, which is the bout hours of previous behavior (Appendix S1: Section S5).
Step 4: Match on covariates
We devised an analysis based on matching, which accounts for covariate imbalance (in this case, covariates with markedly different values in the vessel-exposed group vs. the unexposed group) and is particularly sensitive for the following reason. Correct matching on confounders obviates a linear model's need to obtain correct causal regression relationships of the effect of interest and all confounders on the response variable, across the entire range of confounder values that occur in the data. In addition, our use of matching allowed us to include covariates, such as benthic biomass, with a substantial proportion of missing values that were not possible to impute and therefore could not be included in a linear model. We used all vessel-exposed response hours, and we matched 10 unexposed response hours to each vessel-exposed response hour based on covariate values. We used exact matching for haul-out substrate due to its importance for walrus behaviors (Jay et al., 2017; Udevitz et al., 2009), and we used propensity score-based matching to obtain the best possible balance for the remaining covariates (Pan & Bai, 2015; Ramsey et al., 2019; Rosenbaum & Rubin, 1983; see Appendix S1: Section S6 and Table S1).
We evaluated balance of each matched data set by comparing standardized effect sizes (SES) of all main effects and two-way interactions. The SES for any covariate (x) is a function of its sample means under vessel-exposed () and unexposed () conditions, and the sample standard deviation under vessel-exposed conditions (, such that . The most commonly used cutoffs to indicate that data are well balanced are |SES| < 0.1 or |SES| < 0.25 (Stuart et al., 2013). We ensured that all main effects and two-way interactions had |SES| < 0.25 and, when possible, we also ran models that included as regression covariates any effects where 0.10 < |SES| < 0.25 (see Step 5 below).
Step 5: Test for an effect of vessel exposure on walrus behavior using multinomial logistic regression
We expected walruses disturbed by vessels to forgo foraging or hauling out in favor of in-water not foraging behavior. We tested whether there was any difference in proportions of time spent in the three behavior categories (foraging, in-water not foraging, and hauled out) for vessel-exposed versus unexposed response hours using an unconditional multinomial logistic regression (Appendix S1: Section S7) for each of the three matched data sets (one data set for each category of previous behavior). Primary models contained the single effect of vessel exposure, and we tested its significance using likelihood ratio tests. We confirmed results from primary models by fitting doubly robust models (Pan & Bai, 2015) that also included as regression covariates any main effects or interactions that may still have been unbalanced after matching (0.10 < |SES| < 0.25) and had no missing values.
Objective 3: Power analyses
We conducted power analyses (Appendix S1: Section S8) to determine the proportion of times a predetermined effect size could be detected based on prespecified sample sizes of vessel-exposed response hours (25–200, by increments of 25), the standard specifications of a p value of 0.05 and a power of 80% (Sims et al., 2007), and observed variation from the primary analysis (obtained from Objective 2: Test for effects of vessel exposure on walrus behavior). We examined two types of changes from the vessel-unexposed response-hour behavior probabilities: (1) vessel exposure that reduced the proportion of response-hour time spent foraging in favor of in-water not foraging and (2) vessel exposure that reduced the amount of response-hour time spent hauled out in favor of in-water not foraging. We examined 20% and 40% reductions in both the probability of foraging and the probability of being hauled out based on expert opinion that reductions in time spent foraging could be ~20% due to disturbance from continuous sound exposure such as from vessels and ~40% due to disturbance from impulsive sound exposure such as from seismic surveys (Harwood et al., 2019). To make these reductions meaningful, we limited our analyses to cases where the vessel-unexposed response-hour probability for foraging or hauling out was at least 0.05.
We used the results of these power analyses to determine whether sample sizes in our data were large enough for us to detect an effect based on three different definitions of vessel exposure: (1) vessel presence defined as ≥10 min of vessel locations within a 15-km measured radius and vessel absence defined as no vessel locations within a 20-km measured radius (definition used in primary analysis); (2) vessel presence defined as ≥10 min of vessel locations within a 10-km measured radius and vessel absence defined as no vessel locations within a 15-km measured radius; and (3) vessel presence defined as ≥10 min of vessel locations within a 5-km measured radius and vessel absence defined as no vessel locations within a 10-km measured radius.
RESULTS
Objective 1: Quantify measurement error and develop a misclassification filter for vessel presence versus vessel absence
Our bootstrap-based evaluation of measurement error used location error from 126,546 h of walrus behavior from 218 walruses. Measured distances between simulated walrus locations and vessels were positively biased (i.e., the measured distances were, on average, larger than the true distances; Figure 2). Bias was largest when true distances were small. On average, a vessel and walrus with a true distance of 0, 1, 5, or 10 km apart had measured distances of 1.8, 2.1, 5.3, or 10.2 km apart. Imprecision was large and relatively constant (Figure 2). Standard errors ranged from 1.4 to 1.6 km, and 95% confidence interval lengths (upper confidence limit minus lower confidence limit) ranged from 5.5 to 7.0 km. Bias in distance measurements resulted from large walrus location error combined with negligible vessel location error. If vessel location error was as large as walrus location error, distance measurements would be unbiased but substantially less precise.
[IMAGE OMITTED. SEE PDF]
Our misclassification filter (the 5-km doughnut ring between the 15-km measured radius we used to define vessel presence and the 20-km measured radius we used to define vessel absence) markedly reduced misclassification rates due to the measurement error quantified above. Without the misclassification filter (i.e., when vessel presence and vessel absence were both defined within a 15-km measured radius), the misclassification rate at 15-km true distance was ~50% and dropped to ~10% at ±2 km from that distance (Appendix S1: Figure S1). When the misclassification filter was used, the true vessel presence–absence boundary was ~17 km, roughly halfway between the 15-km measured radius used for defining vessel presence and the 20-km measured radius used for defining vessel absence. Most importantly, the misclassification rate never exceeded 5%, and it was only 1% at ±2 km from the true vessel presence–absence boundary (Appendix S1: Figure S1).
Objective 2: Test for effects of vessel exposure on walrus behavior
Ninety-five of the 218 walruses (44%) were exposed to vessels during June through November, providing 206 vessel-exposed response hours (103 for previously foraging, 71 for previously in-water not foraging, and 32 for previously hauled out). Our analyses used these 206 vessel-exposed response hours and matched to them 2060 of the >120,000 unexposed response hours to obtain the best covariate balance (Figure 1).
Sixty-four vessels contributed to walrus vessel exposure. Almost all of the 206 vessel-exposed response hours involved a single vessel (200, 97%) and the rest involved two vessels (6, 3%), giving a total of 212 vessel hours, once the second vessels were counted. Vessel type (Table 1) and size varied. Almost all vessel hours (198, 94%) resulted from moderate-sized vessels (range of length × width = 207–3681 m2), although all vessel sizes ranged from 10 to 11,505 m2.
TABLE 1 Number of vessel hours (proportions) that contributed to vessel-exposed response hours by vessel type for each of the three data sets (previous behaviors).
| Data set | Vessel type | Vessel hours (proportion) |
| PF | Research vessel | 33 (0.31) |
| PF | Icebreaker | 24 (0.22) |
| PF | General cargo ship | 16 (0.15) |
| PF | Tug | 10 (0.09) |
| PF | Oil products tanker | 8 (0.07) |
| PF | Oil tanker | 3 (0.03) |
| PF | Other | 13 (0.12) |
| PF | Total | 107 (1.00) |
| PIWNF | General cargo ship | 23 (0.32) |
| PIWNF | Oil products tanker | 15 (0.21) |
| PIWNF | Research vessel | 9 (0.12) |
| PIWNF | Oil tanker | 6 (0.08) |
| PIWNF | Tug | 6 (0.08) |
| PIWNF | Icebreaker | 4 (0.05) |
| PIWNF | Other | 10 (0.14) |
| PIWNF | Total | 73 (1.00) |
| PHO | Oil products tanker | 6 (0.19) |
| PHO | Research vessel | 6 (0.19) |
| PHO | General cargo ship | 5 (0.16) |
| PHO | Icebreaker | 5 (0.16) |
| PHO | Oil tanker | 4 (0.13) |
| PHO | Refrigerated cargo ship | 2 (0.06) |
| PHO | Other | 4 (0.13) |
| PHO | Total | 32 (1.00) |
Covariates in the raw data (before matching) were unbalanced, with the absolute value of the standardized effect size (|SES|) as large as 1.87, and over half the covariate main effects with |SES| ≥ 0.25 (Figure 3). Matched data were well balanced: no variable had |SES| > 0.24 and each haul-out substrate category had an |SES| of exactly 0.
[IMAGE OMITTED. SEE PDF]
Covariate imbalance varied among the three raw data sets, but clearly unbalanced covariates (|SES| ≥ 0.25) did not trend in opposite directions (Table 2). Based on |SES| ≥ 0.25 in one or more data sets, walruses were more likely to be exposed to vessels in areas with land only or no ice or land nearby, and less likely to be exposed to vessels in areas with ice only nearby. The proportion of vessel-exposed response hours occurring in a land-only area could be up to 0.35 higher than the proportion of unexposed response hours occurring in a land-only area (Table 2, previously hauled out), the proportion with no ice or land could be up to 0.22 higher (Table 2, previously foraging), and the proportion with ice only could be up to 0.34 lower (Table 2, previously hauled out). In addition, walruses were more likely to be exposed to vessels in conditions that were windier, colder, and either outside the region where benthic biomass data were collected, or if in that region, in areas with higher biomass.
TABLE 2 Means and standard deviations (SD) for main covariate effects for vessel-exposed data, matched unexposed data, and raw unexposed data for each of the three data sets (previous behaviors).
| Data set | Covariate | Vessel exposed | Matched unexposed | Raw unexposed | |||
| Mean | SD | Mean | SD | Mean | SD | ||
| PF | Ice onlya | 0.46 | 0.50 | 0.46 | 0.50 | 0.66 | 0.47 |
| PF | Land only | 0.01 | 0.10 | 0.01 | 0.10 | 0.02 | 0.14 |
| PF | No ice or landa | 0.53 | 0.50 | 0.53 | 0.50 | 0.31 | 0.46 |
| PF | Ice and land | 0 | 0 | 0 | 0 | 0 | 0.06 |
| PF | Bout hours of previous behavior (h) | 9.46 (0.92) | 8.29 (0.27) | 9.53 (0.92) | 8.51 (0.27) | 10.9 (0.92) | 9.26 (0.28) |
| PF | Wind speed (m/s)a | 5.96 | 3.13 | 5.88 | 3.06 | 5.17 | 2.66 |
| PF | Temperature (°C) | 1.04 | 2.05 | 0.88 | 2.04 | 1.33 | 1.92 |
| PF | Biomass (ln[g C/m2])a | 2.58 (0.91) | 0.41 (0.28) | 2.63 (0.91) | 0.35 (0.29) | 2.47 (0.94) | 0.41 (0.24) |
| PIWNF | Ice onlya | 0.10 | 0.30 | 0.10 | 0.30 | 0.35 | 0.48 |
| PIWNF | Land onlya | 0.31 | 0.47 | 0.31 | 0.46 | 0.19 | 0.39 |
| PIWNF | No ice or landa | 0.59 | 0.50 | 0.59 | 0.49 | 0.44 | 0.50 |
| PIWNF | Ice and land | 0 | 0 | 0 | 0 | 0.02 | 0.14 |
| PIWNF | Bout hours of previous behavior (h) | 5.17 (1) | 5.56 (0) | 4.85 (1) | 3.68 (0) | 5.28 (1.00) | 6.58 (0.03) |
| PIWNF | Wind speed (m/s)a | 7.29 | 3.28 | 7.54 | 3.17 | 5.56 | 2.82 |
| PIWNF | Temperature (°C) | 1.10 | 1.68 | 0.92 | 1.47 | 1.40 | 2.38 |
| PIWNF | Biomass (ln[g C/m2])a | 2.49 (0.61) | 0.40 (0.49) | 2.59 (0.61) | 0.30 (0.49) | 2.27 (0.78) | 0.57 (0.41) |
| PHO | Ice onlya | 0.47 | 0.51 | 0.47 | 0.50 | 0.81 | 0.39 |
| PHO | Land onlya | 0.50 | 0.51 | 0.50 | 0.50 | 0.15 | 0.36 |
| PHO | No ice or land | 0 | 0 | 0 | 0 | 0.03 | 0.16 |
| PHO | Ice and land | 0.03 | 0.18 | 0.03 | 0.17 | 0.01 | 0.12 |
| PHO | Bout hours of previous behavior (h) | 10.44 (1) | 9.51 (0) | 9.66 (1) | 8.46 (0) | 9.98 (1.00) | 8.10 (0.07) |
| PHO | Wind speed (m/s)a | 5.93 | 2.96 | 5.42 | 2.58 | 3.72 | 1.90 |
| PHO | Temperature (°C)a | 0.38 | 1.78 | 0.48 | 1.70 | 1.24 | 1.92 |
| PHO | Biomass (ln[g C/m2])a | 2.69 (0.47) | 0.09 (0.51) | 2.7 (0.47) | 0.08 (0.50) | 2.52 (0.83) | 0.32 (0.37) |
Primary models did not show a significant effect of vessel exposure on walrus behaviors during the response hour (p values for each of the three analyses: previously foraging = 0.50, previously in-water not foraging = 0.82, and previously hauled out = 0.69). Regardless of vessel exposure, the probability a previously foraging walrus remained foraging was ~0.9, and the probability it transitioned to in-water not foraging was ~0.1 (Table 3 probability scale; Appendix S1: Table S2 logit scale with variance–covariance matrix). Regardless of vessel exposure, a previously in-water not foraging walrus transitioned to foraging with a probability of ~0.15, remained in-water not foraging with a probability of ~0.8, and transitioned to hauled out with a probability of ~0.05. Regardless of vessel exposure, a previously hauled out walrus transitioned to foraging with a probability of ~0.03, transitioned to in-water not foraging with a probability of ~0.10, and remained hauled out with a probability of ~0.87.
TABLE 3 Estimated probabilities from primary models that walruses will be foraging, in-water not foraging, or hauled out during the response hour, as predicted by vessel exposure category for each of the three data sets (previous behaviors).
| Data set | Vessel exposure category | Foraging | In-water not foraging | Hauled out |
| PF | Vessel exposed | 0.91 | 0.08 | 0.001 |
| PF | Unexposed | 0.88 | 0.12 | 0.005 |
| PIWNF | Vessel exposed | 0.15 | 0.80 | 0.04 |
| PIWNF | Unexposed | 0.15 | 0.79 | 0.06 |
| PHO | Vessel exposed | 0.03 | 0.13 | 0.84 |
| PHO | Unexposed | 0.03 | 0.08 | 0.89 |
We fit doubly robust models for the previously in-water not foraging and previously hauled out data sets because they contained covariates that may not have been ideally balanced (0.10 ≤ |SES| < 0.25). The model for the previously in-water not foraging data set had the additional covariates of temperature, bout hours of previous behavior, and their interaction, and it showed no effect of vessel exposure on response-hour walrus behavior (p = 0.86). The model for the previously hauled out data set had the additional covariates of wind, temperature, haul-out substrate, bout hours of previous behavior, and the two-way interactions between wind and the other covariates. It also showed no effect of vessel exposure on response-hour walrus behavior (p = 0.61).
Objective 3: Power analyses
Power analyses revealed that a sample size of 50 vessel-exposed response hours reliably detected a 20% reduction in the probability that a previously hauled out walrus remained hauled out in the response hour (Table 4), but a sample size of 75 vessel-exposed response hours was required to reliably detect a 20% reduction in the probability that a previously foraging walrus remained foraging in the response hour. Smaller sample sizes could reliably detect a 40% reduction in these same haul-out or foraging probabilities, but expert opinion suggests a 20% reduction is appropriate for disturbance from continuous noises such as vessels, and a 40% reduction is appropriate for disturbance from impulsive noises such as seismic surveys (Harwood et al., 2019). Power to detect a 20%–40% reduction in the probability a previously in-water not foraging walrus transitioned to foraging or hauled out was low across all sample sizes examined (25–200). This is largely because the probability a walrus spent the response hour foraging or hauled out was already low if it spent the previous hour in-water not foraging. As a result, the proportional reduction in the logits (9%–35%) was much lower in the previously in-water not foraging analysis than in the other two analyses (57%–94%).
TABLE 4 Proportions of parameter values achieving ≥80% power and mean power (in parentheses) over all parameter values when response-hour foraging or hauling out time was reduced in favor of time in-water not foraging.
| Data set | Sample size | Power for 20% reduction | Power for 40% reduction |
| PF | 25 | 0.00 (0.45) | 0.99 (0.93) |
| PF | 50 | 0.33 (0.75) | 1.00 (1.00) |
| PF | 75 | 0.93 (0.90) | 1.00 (1.00) |
| PF | 100 | 1.00 (0.96) | 1.00 (1.00) |
| PF | 125 | 1.00 (0.99) | 1.00 (1.00) |
| PF | 150 | 1.00 (0.99) | 1.00 (1.00) |
| PF | 175 | 1.00 (1.00) | 1.00 (1.00) |
| PF | 200 | 1.00 (1.00) | 1.00 (1.00) |
| PHO | 25 | 0.00 (0.60) | 1.00 (0.97) |
| PHO | 50 | 0.95 (0.87) | 1.00 (1.00) |
| PHO | 75 | 1.00 (0.96) | 1.00 (1.00) |
| PHO | 100 | 1.00 (0.99) | 1.00 (1.00) |
| PHO | 125 | 1.00 (1.00) | 1.00 (1.00) |
| PHO | 150 | 1.00 (1.00) | 1.00 (1.00) |
| PHO | 175 | 1.00 (1.00) | 1.00 (1.00) |
| PHO | 200 | 1.00 (1.00) | 1.00 (1.00) |
The sample size of vessel-exposed response hours from our primary analysis gave excellent power to detect a relevant (20%) reduction in the probability a previously foraging walrus would remain foraging, but sample sizes were not useful for the other previous behaviors, and analyses using smaller radii to define vessel presence and absence would not provide useful power for any previous behaviors (Table 4). In particular, in the analysis we conducted for previously foraging walruses (vessel presence defined within a 15-km measured radius and vessel absence defined within a 20-km measured radius), the sample size of 103 vessel-exposed response hours gave almost complete power to detect a 20% reduction in the probability a previously foraging walrus would remain foraging. Had we reduced each measured radius in the previously foraging analysis by 5 km (vessel presence defined within a 10-km measured radius and vessel absence defined within a 15-km measured radius), the resulting sample size of 57 vessel-exposed response hours would have given complete power to detect a 40% reduction in the probability a previously foraging walrus would remain foraging, but very low power for a 20% reduction. Had we reduced each measured radius in the previously foraging analysis by another 5 km (vessel presence defined within a 5-km measured radius and vessel absence defined within a 10-km measured radius), the resulting sample size of 15 vessel-exposed response hours would not have been useful to detect a 20% reduction and would have been questionable for a 40% reduction.
Sample size of vessel-exposed response hours for previously hauled out walruses in the primary analysis (32) was lower than for previously foraging walruses and too small to detect a relevant (20%) reduction in the probability a walrus remained hauled out (Table 4). Because hauled out walruses likely react to vessels at a much shorter distance than do in-water walruses, we focus on the power to detect a response of previously hauled out walruses using vessel presence defined within a 5-km measured radius and vessel absence defined within a 10-km measured radius. At these radii, the resulting sample size of vessel-exposed response hours (4) would have been far too low to detect a difference. Even if we had taken the measured radius at face value (i.e., used the 5-km radius to define both vessel presence and vessel absence), we would have still had too low a sample size (12) to detect a difference.
DISCUSSION
We found no evidence of foraging walruses altering their behavior in response to vessel exposure within 17 km, although this reflected a variety of vessels and environmental conditions. The vessels walruses encountered varied in size, type, and activity, and we had no way to measure sound levels reaching walruses. Thus, lack of response by foraging walruses to vessel presence up the 17-km true vessel presence–absence boundary reflected their responses to this mix of vessels and environmental conditions (such as bathymetry and presence of sea ice) that affect sound propagation through water.
The resolution of this work was coarse due to data constraints. Sample sizes of vessel-exposed walrus response hours were small even when measured distances between vessels and walruses were taken at face value. In addition, measured distances were highly biased when vessels and walruses were close, and imprecision was large at all distances. These factors contributed to our need to conduct initial analyses by defining vessel presence within a 15-km measured distance from an hourly walrus location and vessel absence within a 20-km measured distance from an hourly walrus location, resulting in a large, 17-km distance defining the true boundary between vessel presence and vessel absence. Nevertheless, this boundary is similar to the 16-km distance from Quijano et al. (2019), which we considered a general target for the maximum distance at which in-water walruses might respond to vessels. Furthermore, our analysis had almost complete power to detect a relevant (20%) reduction in the probability a previously foraging walrus would remain foraging after vessel exposure. Thus, although we could not determine at what distance vessel exposure affects walrus behavior, our findings put a 17-km upper bound on the distance at which the vessels encountered may disturb foraging walruses. When more situation-specific information is lacking, this distance could be used as a conservative buffer for routing vessel traffic relative to areas of high use by foraging walruses.
Power in our initial analyses was insufficient to detect a 20% reduction in the probability a previously hauled out walrus would remain hauled out after vessel exposure or a 20% reduction in the probability that a previously in-water not foraging walrus would switch to foraging or hauling out after vessel exposure. In addition, the 17-km true vessel presence–absence boundary is unrealistically large to apply to hauled out walruses; thus, we limit any above inferences to previously foraging walruses. We explored the potential to analyze our data using successively smaller radii to define vessel presence and absence, but none provided sample sizes large enough for acceptable power. Sample sizes for previously hauled out walruses were particularly low. Within our smallest (and therefore potentially most relevant) set of radii (vessel presence defined within a 5-km measured radius and vessel absence defined within a 10-km measured radius), the number of vessel-exposed walrus response hours was 4, too low to detect a difference. Even if we had ignored measurement error (i.e., used the 5-km measured radius to define both vessel presence and vessel absence), the number of vessel-exposed response hours would have been 12, still too low to detect a difference.
We went to great lengths to assure our analyses were as sensitive as possible to take full advantage of telemetry and vessel AIS data, types of data commonly available to marine mammal scientists. This involved a simulation analysis to quantify and account for Argos-based walrus location error. Although we cannot provide general guidelines based on our simulations, we suggest that marine mammal researchers who are considering combining vessel AIS data with Argos-based location data consider the following challenges we faced. First, a walrus that appeared to be 2 km away from a vessel was, on average, somewhere between 0 and 1 km from a vessel. Second, imprecision in distances between vessels and walruses was large and relatively constant no matter how far the vessel and walrus were from each other. Our analyses produced a 1.5-km SE or ~6-km 95% confidence interval length (upper confidence limit − lower confidence limit) in distances between a walrus and a vessel. Thus, we defined a 5-km doughnut-shaped zone of uncertainty around a radius defining vessel presence in which we discarded observations to avoid high misclassification rates. These biases and uncertainties would also apply to other forms of disturbance besides vessels if the disturbance source location was measured with negligible error, and the animal location data exhibited error with a high magnitude similar to ours.
We also applied very sensitive statistical matching techniques that used all vessel-exposed observations and matched the unexposed observations to them based on confounding covariates (confounders). Matching has only recently been introduced into the ecological literature (Ramsey et al., 2019), but it is often used in the human social, health, and behavioral sciences for making causal inference based on observational data, which are inherently confounded (Pan & Bai, 2015). An example of confounding in this study is the presence of sea ice. Walruses are less likely to be exposed to vessels in the marginal ice zone than in open water because most vessels avoid ice. However, walruses need a substrate on which to haul out. As a result, not correctly accounting for the confounding effect of ice presence during the response hour could make it appear that walruses haul out less when vessels are nearby even if it is only due to lack of sea ice as a resting platform.
Correct matching on confounders obviates the need to obtain correct causal regression relationships of the effect of interest and all confounders on the response variable, across the entire range of confounder values that occur in the data. This is particularly important because correct causal relationships are not reliably identified by the most parsimonious model. Rather, by using a matching technique that retained all vessel-exposed response hours and matched them with the unexposed response hours, we obtained an effect of exposure standardized to the exposed (Sjölander & Greenland, 2013). In other words, we predicted the effect of vessel exposure on response-hour walrus behaviors under the conditions in which walruses are exposed to vessels. Thus, the goal of matching was to limit the scope of inference to conditions observed with vessel exposure, with the intent of obtaining the most sensitive, unbiased possible inference under those conditions. When our matching procedures may not have been fully successful (as indicated by 0.10 ≤ |SES| < 0.25), we combined them with regression modeling to form doubly robust models to ensure our primary model tests were valid. In addition, because our matching techniques could handle missing data (whereas a regression could not), we were able to include benthic biomass as a covariate and account for its values when they existed, but still use the 54 of 206 vessel-exposed response hours where benthic biomass information was not available and not possible to impute. Finally, our matching method still allowed us to account for temporal autocorrelation within the multinomial response framework by using the “bout hours of previous behavior” variable.
We could not account for two sources of disturbance that may have limited our ability to detect an effect of vessel exposure on behavior of foraging walruses, although neither source would make it more likely to detect an effect that did not exist. The first disturbance source, seismic surveys typically performed for oil and gas exploration, has an extensive sound footprint that may mask less intense sounds, such as vessels in transit, or may disturb animals not exposed to vessel traffic. Keen et al. (2018) estimated that seismic surveys conducted in the Chukchi Sea from August 14 to October 31, 2014, produced a sound pressure level of 120 dB re 1 μPa at ~80 km. While we know that 15% of the response hours used in the matched analysis for previously foraging walruses occurred during this time frame, these response hours occurred over a >600,000-km2 area, and we do not know where in the Chukchi Sea or on which days the seismic surveys occurred. As such, it is impossible to evaluate the potential effect of these seismic surveys on our results.
The second disturbance source we could not account for was vessels that did not use AIS during our study, meaning some hourly walrus locations may have had vessels present, even though vessels appeared absent based on AIS (ghost vessels). Retrospectively ground-truthing AIS data is challenging at best, and information (including satellite imagery) that could be used for this purpose for our study area and period was insufficient. However, for two reasons, any effect of ghost vessels was probably small. First, we think AIS data captured the vast majority of moderate- to large-sized vessel activity in our study area. Cargo vessels over 500 gross tons (GT), vessels over 300 GT on international voyages, and passenger vessels are required to transmit AIS data (International Maritime Organization, 2009). While AIS data transmission was voluntary for other vessels, such as industrial support vessels and skiffs operated by researchers or coastal residents, such transmissions are routine for industrial support vessels. Second, given the extremely low apparent rate of vessel encounters (1 out of every 665 response hours for already foraging walruses), the number of ghost vessels would have to have been extremely large compared with the number of vessels using AIS to have a substantial effect.
More detailed analyses of walrus behavioral responses to vessels will not be possible with existing data due to the caveats we have described. However, potential vessel disturbances for walruses are expected to increase as vessel traffic in the Arctic increases; thus, it remains a critical topic for continued research (Halliday et al., 2017). Further study of sustained responses may be feasible using focal following methods combined with sound and movement recording tags; however, this would require controlled approaches to walruses shortly after tagging. In such a study, a walrus would be tagged, and after disturbance due to tagging ameliorated, researchers would conduct controlled approaches to the walrus with a research vessel or with a small research skiff fitted with underwater loudspeakers capable of simulating the approach of large vessels. Telemetry with real-time reporting to the vessel would be essential because walruses typically occur in large aggregations, even when at sea (Speckman et al., 2011), making visual focal follows impossible. Given the limited spatial resolution of walrus locations with Argos data from this study, any future study of vessel effects on walrus behavior would also require modifying walrus tags to incorporate quick fix pseudoranging and inertial navigation technology (e.g., quick fix GPS) to enable more frequent relocation of walruses with precision on the order of 100 m (Costa et al., 2010; Tomkiewicz et al., 2010).
AUTHOR CONTRIBUTIONS
Conceptualization: Rebecca Taylor, Chadwick Jay, and Anthony Fischbach. Data curation: Rebecca Taylor, Anthony Fischbach, William Beatty, and Justin Crawford. Formal analysis: Rebecca Taylor, William Beatty, and Anthony Fischbach. Funding acquisition: Chadwick Jay and Lori Quakenbush. Investigation: Chadwick Jay, Lori Quakenbush, Anthony Fischbach, and Justin Crawford. Methodology: Rebecca Taylor. Project administration: Chadwick Jay and Lori Quakenbush. Software: Rebecca Taylor, William Beatty, and Anthony Fischbach. Supervision: Chadwick Jay and Lori Quakenbush. Writing—original draft: Rebecca Taylor. Writing—review and editing: Rebecca Taylor, Chadwick Jay, William Beatty, Anthony Fischbach, Lori Quakenbush, and Justin Crawford.
ACKNOWLEDGMENTS
USGS walrus tagging was funded by the USGS Changing Arctic Ecosystem initiative through the Wildlife Program of the Ecosystem Mission Area, the USGS-BOEM Outer Continental Shelf program, and the USGS-USFWS Science Support program. USGS thanks crews of the R/V Norseman II for supporting offshore tagging, crews of NOAA's Aerial Surveys of Arctic Marine Mammals program and Clearwater Air for aerial reconnaissance, the community of Point Lay, Alaska, especially Marie Tracey (North Slope Borough liaison), Leo Ferreira III, Warren Harding-Lampe, and the Point Lay Fire Department for assistance with onshore tagging, and Anatoly Kochnev of ChukotTINRO for onshore walrus tagging in Chukotka. ADFG walrus tagging was funded by BOEM (Contract M09PC00027). ADFG thanks walrus hunters Clarence Irrigoo, Perry Pungowiyi, and Edwin Noongwook for sharing their knowledge of walrus behavior and ecology and greatly enhancing tagging success by their expertise, and crews of the R/V Norseman II and Professor Multanovskiy for support during tag deployment. We appreciate the interest and support of the Eskimo Walrus Commission. We appreciate extremely helpful conversations on statistical approaches with Mark Udevitz, Kristen Hunter, and Luke Miratrix and his lab group. Ioannis Kosmidis and Martin Elff graciously answered questions regarding theory underlying their R packages (“brglm2” and “mclogit”), and Devin Johnson and Josh London helped extensively with implementation of their R package “crawl.” Brian Taras, Dan Esler, Morten Tange Olsen, and four anonymous reviewers provided helpful comments on previous versions of this manuscript. The findings and conclusions in this article are those of the authors and do not necessarily represent the views of the U.S. Fish and Wildlife Service. Any use of trade, firm, or product names is for descriptive purposes only and does not imply endorsement by the U.S. Government.
CONFLICT OF INTEREST STATEMENT
The authors declare no conflict of interest.
DATA AVAILABILITY STATEMENT
Data (Fischbach et al., 2023) are available from USGS: .
Beatty, W. S., C. V. Jay, A. S. Fischbach, J. M. Grebmeier, R. L. Taylor, A. L. Blanchard, and S. C. Jewett. 2016. “Space Use of a Dominant Arctic Vertebrate: Effects of Prey, Sea Ice, and Land on Pacific Walrus Resource Selection.” Biological Conservation 203: 25–32.
Citta, J. J., L. F. Lowry, L. T. Quakenbush, B. P. Kelly, A. S. Fischbach, J. M. London, C. V. Jay, et al. 2018. “Overlap of Marine Mammal Distributions and Core Use Areas.” Deep‐Sea Research Part II: Topical Studies in Oceanography 152: 132–53.
Collecte Localisation Satellites. 2011. Argos User's Manual. Ramonville‐Saint‐Agne: Collecte Localisation Satellites.
Costa, D. P., P. W. Robinson, J. P. Y. Arnould, A. L. Harrison, S. E. Simmons, J. L. Hassrick, A. J. Hoskins, et al. 2010. “Accuracy of ARGOS Locations of Pinnipeds at‐Sea Estimated Using Fastloc GPS.” PLoS One 5: [eLocator: e8677].
Douglas, D. C., R. C. Weinzierl, S. Davidson, R. Kays, M. Wikelski, and G. Bohrer. 2012. “Moderating Argos Location Errors in Animal Tracking Data.” Methods in Ecology and Evolution 3: 999–1007.
Erbe, C., M. Dähne, J. Gordon, H. Herata, D. S. Houser, S. Koschinski, R. McCauley, et al. 2019. “Managing the Effects of Noise from Ship Traffic, Seismic Surveying and Construction on Marine Mammals in Antarctica.” Frontiers in Marine Science 6: 647.
Erbe, C., S. A. Marley, R. P. Schoeman, J. N. Smith, L. E. Trigg, and C. B. Embling. 2019. “The Effects of Ship Noise on Marine Mammals—A Review.” Frontiers in Marine Science 6: 606.
Fay, F. H. 1982. “Ecology and Biology of the Pacific Walrus, Odobenus rosmarus divergens Illiger.” North American Fauna 74: 1–279.
Fischbach, A. S., and C. V. Jay. 2016. “A Strategy for Recovering Continuous Behavioral Telemetry Data from Pacific Walruses.” Wildlife Society Bulletin 40: 599–604.
Fischbach, A. S., C. V. Jay, W. S. Beatty, R. L. Taylor, J. A. Crawford, and L. T. Quakenbush. 2023. “Walrus Haulout and In‐Water Activity Levels Relative to Vessel Interactions in the Chukchi Sea: 2012–2015.” U.S. Geological Survey Data Release. [DOI: https://dx.doi.org/10.5066/P9IO8AZJ].
Jay, C. V., R. L. Taylor, A. S. Fischbach, M. S. Udevitz, and W. S. Beatty. 2017. “Walrus Haul‐Out and In Water Activity Levels Relative to Sea Ice Availability in the Chukchi Sea.” Journal of Mammalogy 98: 386–96.
Johnson, D. S., J. M. London, M. Lea, and J. W. Durban. 2008. “Continuous‐Time Correlated Random Walk Model for Animal Telemetry Data.” Ecology 89: 1208–15.
Kastak, D., and R. J. Schusterman. 1998. “Low‐Frequency Amphibious Hearing in Pinnipeds: Methods, Measurements, Noise, and Ecology.” The Journal of the Acoustical Society of America 103: 2216–28.
Keen, K. A., B. J. Thayre, J. A. Hildebrand, and S. M. Wiggins. 2018. “Seismic Airgun Sound Propagation in Arctic Ocean Waveguides.” Deep Sea Research Part I: Oceanographic Research Papers 141: 24–32.
Lomac‐Macnair, K., J. P. Andrade, and E. Esteves. 2019. “Seal and Polar Bear Behavioral Response to an Icebreaker Vessel in Northwest Greenland.” Human–Wildlife Interactions 13: 277–89.
Mcconnell, B. J., C. Chambers, and M. A. Fedak. 1992. “Foraging Ecology of Southern Elephant Seals in Relation to the Bathymetry and Productivity of the Southern Ocean.” Antarctic Science 4: 393–8.
Melia, N., K. Haines, and E. Hawkins. 2016. “Sea Ice Decline and 21st Century Trans‐Arctic Shipping Routes.” Geophysical Research Letters 43: 9720–8.
Mesinger, F., G. Dimego, E. Kalnay, K. Mitchell, P. C. Shafran, W. Ebisuzaki, D. Jović, et al. 2006. “North American Regional Reanalysis.” Bulletin of the American Meteorological Society 87: 343–60.
Mikkelsen, L., M. Johnson, D. M. Wisniewska, A. van Neer, U. Siebert, P. T. Madsen, and J. Teilmann. 2019. “Long‐Term Sound and Movement Recording Tags to Study Natural Behavior and Reaction to Ship Noise of Seals.” Ecology and Evolution 9: 2588–601.
NMFS (National Marine Fisheries Service). 2021. “Takes of Marine Mammals Incidental to Specified Activities; Taking Marine Mammals Incidental to Lighthouse Repair and Tour Operations at Northwest Seal Rock, California.” Federal Register 86(171): 50304–20.
Overland, J., E. Dunlea, J. E. Box, R. Corell, M. Forsius, V. Kattsov, M. S. Olsen, J. Pawlak, L. Reiersen, and M. Wang. 2019. “The Urgency of Arctic Change.” Polar Science 21: 6–13.
Pan, W., and H. Bai. 2015. Propensity Score Analysis: Fundamentals and Developments. New York: The Guilford Press.
Post, E., U. S. Bhatt, C. M. Bitz, J. F. Brodie, T. L. Fulton, M. Hebblewhite, J. Kerby, S. J. Kutz, I. Stirling, and D. A. Walker. 2013. “Ecological Consequences of Sea‐Ice Decline.” Science 341: 519–24.
Quijano, J. E., D. E. Hannay, and M. E. Austin. 2019. “Composite Underwater Noise Footprint of a Shallow Arctic Exploration Drilling Project.” IEEE Journal of Oceanic Engineering 44: 1228–39.
Ramsey, D. S. L., D. M. Forsyth, E. Wright, M. McKay, and I. Westbrooke. 2019. “Using Propensity Scores for Causal Inference in Ecology: Options, Considerations, and a Case Study.” Methods in Ecology and Evolution 10: 320–31.
Reichmuth, C., J. M. Sills, A. Brewer, L. Triggs, R. Ferguson, E. Ashe, and R. Williams. 2020. “Behavioral Assessment of In‐Air Hearing Range for the Pacific Walrus (Odobenus rosmarus divergens).” Polar Biology 43: 767–72.
Rosenbaum, P. R., and D. B. Rubin. 1983. “The Central Role of Propensity Score in Observational Studies for Causal Effects.” Biometrika 70: 41–55.
Scholik‐Schlomer, A. R. 2015. “Where the Decibels Hit the Water: Perspectives on the Application of Science to Real‐World Underwater Noise and Marine Protected Species Issues.” Acoustics Today 11: 36–44.
Silber, G. K., and J. D. Adams. 2019. “Vessel Operations in the Arctic, 2015–2017.” Frontiers in Marine Science 6: 573.
Sims, M., D. A. Elston, M. P. Harris, and S. Wanless. 2007. “Incorporating Variance Uncertainty into a Power Analysis of Monitoring Designs.” Journal of Agricultural, Biological, and Environmental Statistics 12: 236–49.
Sjölander, A., and S. Greenland. 2013. “Ignoring the Matching Variables in Cohort Studies—When Is it Valid and Why?” Statistics in Medicine 32: 4696–708.
Sokolova, Y. V., O. S. Devyataev, E. V. Afanasyeva, and Y. M. Titova. 2020. “Comparison of Independent Navigation of LNG Carriers of Type Yamalmax and their Transition with an Icebreaker Escort.” Russian Arctic 11: 39–58.
Speckman, S. G., V. I. Chernook, D. M. Burn, M. S. Udevitz, A. A. Kochnev, A. Vasilev, C. V. Jay, A. Lisovsky, A. S. Fischbach, and R. B. Benter. 2011. “Results and Evaluation of a Survey to Estimate Pacific Walrus Population Size, 2006.” Marine Mammal Science 27: 514–53.
Stephenson, S. R., L. C. Smith, and J. A. Agnew. 2011. “Divergent Long‐Term Trajectories of Human Access to the Arctic.” Nature Climate Change 1: 156–60.
Stroeve, J., and D. Notz. 2018. “Changing State of Arctic Sea Ice across all Seasons.” Environmental Research Letters 13: [eLocator: 103001].
Stuart, E. A., B. K. Lee, and F. P. Leacy. 2013. “Prognostic Score‐Based Balance Measures Can be a Useful Diagnostic for Propensity Score Methods in Comparative Effectiveness Research.” Journal of Clinical Epidemiology 66: S84–90.
Tomkiewicz, S. M., M. R. Fuller, J. G. Kie, and K. K. Bates. 2010. “Global Positioning System and Associated Technologies in Animal Behaviour and Ecological Research.” Philosophical Transactions of the Royal Society B, Biological Sciences 365: 2163–76.
Udevitz, M. S., C. V. Jay, A. S. Fischbach, and J. L. Garlich‐Miller. 2009. “Modeling Haul‐Out Behavior of Walruses in Bering Sea Ice.” Canadian Journal of Zoology 87: 1111–28.
Udevitz, M. S., C. V. Jay, R. L. Taylor, A. S. Fischbach, W. S. Beatty, and S. R. Noren. 2017. “Forecasting Consequences of Changing Sea Ice Availability for Pacific Walruses.” Ecosphere 8: [eLocator: e02014].
Wilson, S. C., I. Trukhanova, L. Dmitrieva, E. Dolgova, I. Crawford, M. Baimukanov, T. Baimukanov, et al. 2017. “Assessment of Impacts and Potential Mitigation for Icebreaking Vessels Transiting Pupping Areas of an Ice‐Breeding Seal.” Biological Conservation 214: 213–22.
You have requested "on-the-fly" machine translation of selected content from our databases. This functionality is provided solely for your convenience and is in no way intended to replace human translation. Show full disclaimer
Neither ProQuest nor its licensors make any representations or warranties with respect to the translations. The translations are automatically generated "AS IS" and "AS AVAILABLE" and are not retained in our systems. PROQUEST AND ITS LICENSORS SPECIFICALLY DISCLAIM ANY AND ALL EXPRESS OR IMPLIED WARRANTIES, INCLUDING WITHOUT LIMITATION, ANY WARRANTIES FOR AVAILABILITY, ACCURACY, TIMELINESS, COMPLETENESS, NON-INFRINGMENT, MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE. Your use of the translations is subject to all use restrictions contained in your Electronic Products License Agreement and by using the translation functionality you agree to forgo any and all claims against ProQuest or its licensors for your use of the translation functionality and any output derived there from. Hide full disclaimer
© 2023. This work is published under http://creativecommons.org/licenses/by/4.0/ (the "License"). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.
Abstract
Arctic marine mammals have had little exposure to vessel traffic and potential associated disturbance, but sea ice loss has increased accessibility of Arctic waters to vessels. Vessel disturbance could influence marine mammal population dynamics by altering behavioral activity budgets that affect energy balance, which in turn can affect birth and death rates. As an initial step in studying these linkages, we conducted the first comprehensive analysis to evaluate the effects of vessel exposure on Pacific walrus (
You have requested "on-the-fly" machine translation of selected content from our databases. This functionality is provided solely for your convenience and is in no way intended to replace human translation. Show full disclaimer
Neither ProQuest nor its licensors make any representations or warranties with respect to the translations. The translations are automatically generated "AS IS" and "AS AVAILABLE" and are not retained in our systems. PROQUEST AND ITS LICENSORS SPECIFICALLY DISCLAIM ANY AND ALL EXPRESS OR IMPLIED WARRANTIES, INCLUDING WITHOUT LIMITATION, ANY WARRANTIES FOR AVAILABILITY, ACCURACY, TIMELINESS, COMPLETENESS, NON-INFRINGMENT, MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE. Your use of the translations is subject to all use restrictions contained in your Electronic Products License Agreement and by using the translation functionality you agree to forgo any and all claims against ProQuest or its licensors for your use of the translation functionality and any output derived there from. Hide full disclaimer
Details
1 U.S. Geological Survey, Alaska Science Center, Anchorage, Alaska, USA
2 U.S. Fish and Wildlife Service, Marine Mammals Management, Anchorage, Alaska, USA
3 Alaska Department of Fish and Game, Arctic Marine Mammal Program, Fairbanks, Alaska, USA