Effects of the social environment on movement-integrated habitat selection

Abstract

Background

Movement links the distribution of habitats with the social environment of animals using those habitats. Despite the links between movement, habitat selection, and socioecology, their integration remains a challenge due to lack of shared vocabulary across fields, methodological gaps, and the implicit (rather than explicit) historical development of theory in the fields of social and spatial ecology. Given these challenges can be addressed, opportunity for further study will provide insight about the links between social, spatial, and movement ecology. Here, our objective was to disentangle the roles of habitat selection and social association as drivers of movement in caribou (Rangifer tarandus).

Methods

To accomplish our objective, we modelled the relationship between collective movement and selection of foraging habitats using socially informed integrated step selection function (iSSF). Using iSSF, we modelled the effect of social processes, i.e., nearest neighbour distance and social preference, and movement behaviour on patterns of habitat selection.

Results

By unifying social network analysis with iSSF, we identified movement-dependent social association, where individuals took shorter steps in lichen habitat and foraged in close proximity to more familiar individuals.

Conclusions

Our study demonstrates that social preference is context-dependent based on habitat selection and foraging behaviour. We therefore surmise that habitat selection and social association are drivers of collective movement, such that movement is the glue between habitat selection and social association. Here, we put these concepts into practice to demonstrate that movement is the glue connecting individual habitat selection to the social environment.

Background

Movement is defined by a change in spatial location and is the behavioural link between the physical space an animal occupies and the conditions, risks, and resources that exist in their environment [1, 2]. In the context of the social environment, which is defined as the nature, quality, and patterning of relationships among members of a group or population [3], movement represents the connection between the distribution of resources and the social structure of animals that consume those resources [4]. Disentangling the social and spatial drivers of movement is a formidable challenge within behavioural ecology. In many cases, research omits the social contexts within which animals move to, from, and within the areas that contain foraging resources [5, 6]. Spatially-explicit models of sociality highlight that some gregarious species aggregate at areas associated with profitable foraging resources [7], whereas some territorial species only interact at territory edges [8]. Sharing space, either at foraging sites, territory edges, or elsewhere within an animal’s range is required to form the social environment. For example, animals are predicted to select habitat as a function of the profitability and availability of the habitat [9]. A logical extension can be made to conspecifics; individuals form groups based on their familiarity with conspecifics and the profitability of associating with familiar conspecifics. We aim to quantify the relative importance of habitat and conspecifics by developing a socially informed integrated step selection analysis, a movement-based method that accounts for the relative intensity of selection for habitats and social neighbours.

For social animals, individual movement shapes social encounters and subsequent interactions with conspecifics that can affect consensus decision making associated with collective movement [10, 11]. Further complicating our understanding of collective movement is the idea that the type, quality, and distribution of habitats on the landscape can constrain or promote collective movement [12]. For example, dense vegetation impedes visibility, which could influence movement of individuals and reduce the probability a group remains together [13, 14]. In addition, individual movement and habitat selection are affected by the distribution of resources. For example, patchily distributed foraging resources could facilitate large aggregations, whereas homogenously distributed foraging resources could result in a reduction in social associations [5]. While the physical space an individual, or group, occupies is an important driver of animal movement [4], the familiarity of animals within a social group can also affect how animals move through the landscape [11].

Not all social groups are equal; some groups contain unfamiliar individuals (i.e., anonymous groups) [15], while others contain familiar individuals [16]. For anonymous and familiar groups, social foraging occurs when the costs (e.g. competition) and benefits (e.g. reduced predation risk) of an individual’s foraging behaviour are linked with the foraging behaviour of conspecifics [17]. Social foraging can be most beneficial when social information about foraging resources comes from familiar individuals [18]. For example, when foraging resources are unpredictable, familiar individuals obtain reliable information from conspecifics to increase foraging efficiency [19, 20], such that time searching for forage is reduced in favour of more time spent foraging. In the context of movement and habitat selection, theory on social foraging and the benefits of social familiarity provides a framework through which the costs and benefits of collective movement can be explored [17, 21].

Apparent social familiarity or preference is the long-term repeated social association due to shared space at the same time. Although individuals often associate with many conspecifics, non-random repeated social associations with certain individuals form the basis for social preference [22]. Proximately, long-term social relationships can influence collective movement via the reliability of information transfer about foraging resources or predator risk [23, 24], while ultimately they can enhance fitness [25]. The social environment can be influenced by the availability of foraging resources, but social groups can also be composed of individuals with similar physiological or nutritional requirements that occupy the same locations. Apparent social preference may therefore arise as a function of spatial constraints [26], including physical barriers, such as rivers or mountains. Disentangling social preference from spatial constraint could inform our understanding of collective movement and habitat selection [27, 28].

Here, we use contemporary tools from movement ecology and social network analysis and integrate aspects of socioecology, spatial ecology, and movement ecology. We test the effects of the social environment on patterns of habitat selection behaviour by parameterizing socially informed integrated step selection models (Fig. 1). Using caribou (Rangifer tarandus) as a model system, our over-arching goal was to disentangle the roles of habitat selection and social association as drivers of collective movement, when the availability and distribution of foraging resources are variable. To accomplish our goals, we had an over-arching objective to test how social information influences patterns of selection for foraging resources. Social information was measured using proximity to conspecifics (i.e., nearest neighbour distance) and time-lagged social association as proxies for information flow. Objective 1 was to quantify patterns of habitat selection in the presence of conspecifics. We incorporated proximity to conspecifics (i.e., nearest neighbour distance) as a proxy for social information transfer, such that individuals near one another are more likely to exchange information than individuals that are further away from one another (O1). We therefore expect that individuals will be more likely to select foraging habitat (lichen) when conspecifics are near rather than far. Objective 2 was to quantify patterns of habitat selection in the presence of familiar conspecifics. Here, we incorporated time-lagged social association as a proxy for reliable information flow by decomposing social relationships into familiar and unfamiliar contexts. Specifically, we expect that individuals with stronger short-term social preference should collectively select foraging habitat, which tends to be more open and therefore facilitate grouping (O2). The corollary is that individuals should also take short steps in the presence of conspecifics. For O2 we assessed short-term time-lagged social preference using weekly measures of association and tested whether recency bias of association affected patterns of foraging habitat selection. O2 relies on the assumption that from a movement ecology perspective, shorter steps typically represent foraging behaviour and longer steps represent searching behaviour [29].

Fig. 1

Summary of the data pipeline used to generate integrated step selection function (iSSF) models. Primary data sources were landcover data and caribou GPS relocation data, which were combined to determine the physical locations of animals on the landscape. The pairing of animal locations and landcover data was used to generate the comparison of used to available points (panel a), which is the response variable in iSSF models, as well as the habitat type in which a given relocation occurred: lichen (defined in text as open-forage) and forest (panel b). Caribou relocation data were also used to generate two movement parameters (panel c) and aspects of the social environment (panels d and e). Our movement parameter included step length, which is the linear distance between consecutive relocations. The social environment included nearest neighbour distance (panel d) and weekly social networks and the dyadic SRI generated based on a moving-window as a proxy for short-term social preference (panel e). The bottom row represents a graphical formulation of our iSSF models, where habitat selection (1:20 ratio of used to available relocations) was regressed against habitat type (lichen and forest), movement parameters (step length and turn angle), nearest neighbour distance, and weekly dyadic SRI

Methods

Caribou as a model system

We investigated patterns of movement, space use, and social behaviour for caribou on Fogo Island, Newfoundland, Canada. Fogo Island is a small (~ 237km²) island situated ~ 12 km off the northeastern coast of Newfoundland with a humid continental climate. The dominant habitat types on Fogo Island consist of coniferous and mixed forests of balsam fir (Abies balsamea), black spruce (Picea mariana) and white birch (Betula papyrifera), as well as bogs, lakes, and barren rock. Between 1964 and 1967, 26 caribou were introduced to Fogo Island from the Island of Newfoundland [30]. Currently, Fogo Island has a population of approximately 300 caribou and unlike many of the other herds in Newfoundland [31], caribou on Fogo Island have a relatively stable population that has not declined in recent years. In addition, caribou on Fogo Island are sedentary and do not display any migratory or long-distance movements.

Caribou live in fission–fusion societies [32], and throughout much of their range, caribou forage primarily on lichen, grasses, sedges, and other deciduous browse with access to these resources changing between the seasons [33]. During winter (January to March), the landscape is covered by snow, and caribou forage primarily on lichen [34]. Lichen is heterogeneously distributed, and access is impeded by snow and ice cover. Caribou dig holes in the snow, termed craters, to access lichen in the winter, often where snow depth is relatively shallow (~ 30–60 cm deep). Consequently, caribou have limited access to lichen buried under the snow and tend to re-use established craters. To cope with this limitation, caribou use conspecific attraction and social information transfer to gain access to foraging opportunities [35]. In winter, caribou activity budgets suggest that caribou spend ~ 50% of their time foraging, while ~ 40% of their time is spent lying down or ruminating, 7% of their time is spent walking or trotting, and 3% of their time is spent standing [36, 37]. In addition, caribou typically avoid forested habitats due to deep snow in forests and lack of access to forage opportunities [38], whereas most open habitats on Fogo Island are windswept in the winter, facilitating foraging and movement [33].

We used GPS location data collected from Fogo Island caribou (2017–2019) to assess the relationship between social behaviour, habitat selection, and movement (see supplementary information for details on collaring procedures). For all analyses, we restricted locations to only include relocations from the first 75 days of each year (1 January–16 March). Each relocation was assigned to a given habitat classification that was extracted from Landsat images with 30m × 30m pixels [39]. We pooled habitats into three categories. Open habitats that caribou move through consisted of wetland (21.2% of habitats on Fogo Island), rocky outcrops (6.7%), and water/ice (9.2%) habitats. Forest habitats consisted of conifer scrub (27.0%), conifer forest (10.4%), mixed-wood (9.2%), and broadleaf forest (0.2% of habitats) habitats. Finally, we considered lichen barrens (11.7%) as caribou foraging habitat. Overall, the proportion of GPS relocations (see below) that fell within each habitat were: 54.2% of GPS relocations were in open habitats, 30.4% of GPS relocations were in lichen habitat, 10.8% of GPS relocations were in forested habitats, and 4.4% of locations were in undesignated habitats. We then calculated the proportion of each habitat type (i.e., open foraging, open moving, or forest) within 200 m around each used and available point location (see below).

Adult female caribou (n = 26 individual caribou, n = 76 caribou-years) were immobilized and fitted with global positioning system (GPS) collars (Lotek Wireless Inc., Newmarket, ON, Canada, GPS4400M collars, 1250 g). Prior to analyses, we removed all erroneous and outlier GPS locations following Bjørneraas et al. [40]. Collars collected data throughout the year and were programmed to collect locations every two hours. Specifically, we removed implausible GPS fixes, including those that were recorded in the ocean and those that exceeded the maximum diameter of Fogo Island (30 km). We removed individuals with collar failure during the study period (n = 16 caribou-years), individuals that swam to nearby adjacent islands (n = 3 caribou-years), mortality (n = 3), or individuals without data in the winter (n = 16). Given that habitat selection and kernel analyses can be biased by fix success rates [41, 42], we removed individuals with high rates of collar failure (i.e., where > 20% of fixes were missing), resulting in a total of 21 individual caribou (n = 38 caribou-years). Overall, average relocation success rate for the 21 individuals included in subsequent analyses was 98.8%. We did not collar all female caribou in the herd, however, and collared individuals were randomly selected from the population. We therefore assume that our sample of collared animals was randomly distributed. Although associations between collared and uncollared animals were unrecorded, we assumed that our networks (see below) were unbiased representations of the relative degree of social association among all caribou.

Formulating integrated step selection models

To accomplish Objective 2, we formulated integrated step selection functions (iSSF); a type of movement model that simultaneously incorporates movement and habitat selection within a conditional logistic regression framework (Fig. 1) [43,44,45]. As in other resource and step selection analyses [46], iSSF models habitat selection as a binomial response variable where ‘use’ represents the location an animal was observed and ‘availability’ represents the geographical area an animal could potentially use but was not necessarily observed (Fig. 2). iSSF defines availability based on probability distributions of step lengths and turn angles [43], where a step is the linear connection between consecutive relocations, and turn angle is the angular deviation between the headings of two consecutive steps [47]. We generated available steps and turn angles based on the distributions informed by observed population-level movement behaviour using the amt package in R [48]. First, we sampled step lengths from a gamma distribution of observed step lengths for the study population; values were log-transformed for analysis. The statistical coefficient of log-transformed step length is a modifier of the shape parameter from the gamma distribution originally used to generate available steps [43]. Second, we sampled turn angles (measured in radians) for available steps from observed values between $-\pi$ and $\pi$ following a Von Mises distribution. Each observed relocation was paired through a shared start point with 20 available steps generated from step-length and turn-angle distributions and compared in a conditional logistic regression framework. In addition to generating available movement parameters, we also generated an available social environment (see below). To evaluate the predictive performance of our model, we used k-fold (k = 5) cross validation [49] following the methods of Fortin et al. [50]. For details on k-fold cross validation see Appendix 2.

Fig. 2

Schematic of integrated step selection functions in the context of social network analysis. Available (random) steps are generated based on the distribution of used step lengths (thin dashed orange and blue lines) and turn angles. We compared used (observed) to available (random) nearest neighbour distance based on fine-scale movement decisions of individuals. Blue lines represent used (dark thick lines) and available (light dashed lines) steps of individual i and orange lines represent used (dark thick lines) and available (light dashed lines) steps of individual j. The dashed grey line $({NN}_{i,j}$) represents the observed nearest neighbour distance between i and j at t₃. For each set of available steps, we re-calculated nearest neighbour distance, denoted by a dashed grey line (${NN}_{i,j}{\prime}$) between available steps for i and j, which represents the available nearest neighbour distance at a given iteration. Step length is the distance between the used start (e.g. t_j,2) point and the step end point (t_j,3) and turn angle is calculated as the angular deviation between the previous step heading (grey line and ${\theta }_{i}$ and ${\theta }_{j}$) and the subsequent used step

Social network analysis

To accomplish Objective 1, we generated proximity-based social association networks. Nodes in the networks represented individual caribou and edges represented the frequency of association based on proximity between individuals. We generated social networks based on proximity of locations between individual caribou during each week throughout the winter to assess the role of short-term social preference on patterns of habitat selection (O1). For weekly networks (see below for details), we assumed association between two individuals when simultaneous locations (i.e. GPS relocations that occurred within 5 min of each other) were within 50 m of one another [32, 35]. We selected the 50 m threshold based on the standard distance applied to assign individuals to groups in studies of ungulate group size and social behaviour [51] and based on sensitivity analyses in our system. We deem 50 m as an appropriate threshold for caribou as lower thresholds yield relatively sparse networks and higher thresholds yield saturated networks [52]. We applied the ‘chain rule’ to define groups, where each discrete GPS fix was buffered by 50 m and we considered individuals in the same group if 50 m buffers for two or more individuals were contiguous. In some cases, the distance between discrete fixes of two individuals exceeded 50 m because contiguous buffers overlapped with at least one other central individual to form a ‘chain’ of buffers [51]. Groups were defined based on the gambit-of-the-group assumption, which was used to construct social networks [53]. We weighted edges of social networks by the strength of association between dyads of caribou using a dataset specific simplification of the simple ratio index [54], SRI:

$$SRI=\frac{x}{x+{y}_{AB}}$$

where x is the number of times individuals A and B were within 50 m of each other and y_AB is the number of simultaneous fixes from individuals A and B that were separated by > 50 m [55]. To generate networks, we used GPS location data using the R [56] packages spatsoc [57] and igraph [58].

Weekly networks

To accomplish Objective 1, we iteratively generated weekly social networks using a moving window approach and calculated the observed SRI to be included as a covariate in our iSSF model (see formulating iSSF section). The first network was calculated for 1 January to 7 January, the second was 2 January to 8 January, and so on. Weekly networks contained 84 relocations per individual (12 relocations per day). For each of these networks, we used dyadic values of SRI as a proxy for short-term social preference. We used a three-step process. First, to incorporate SRI within the iSSF framework, we determined the identity and distance (m) of each individual’s nearest neighbour at each relocation. Second, for each focal individual and their nearest neighbour at each relocation, we matched the time-lagged dyadic SRI value for the prior week. For example, for individual A at 12:00 on 8 January, we determined the nearest neighbour was individual B and we extracted the dyadic SRI value for these individuals for the previous week. Third, we repeated steps one (i.e., calculate the identity and distance to nearest neighbours) and two (i.e., match the time-lagged dyadic SRI to nearest neighbours) for each observed relocation and the accompanying set of ‘available’ relocations defined by random steps generated in the iSSF. Therefore, each individual at each relocation had one observed weekly dyadic SRI value and 20 ‘available’ weekly dyadic SRI values (see dashed lines on Fig. 2 for example of available steps and how these steps are related to nearest neighbour distance).

Modelling collective movement and habitat selection

To accomplish Objective 2, we fit three iSSF models with a series of fixed and random effects following Muff et al. [62] (see Fig. 1 for iSSF breakdown). When fitting iSSFs, the model is formulated by incorporating a selection-free movement kernel, which considers step length and turn angle distributions, and a selection kernel, which considers habitat types [59]. Parameters in both the movement and selection kernels can be simultaneously estimated and interactions between movement and habitat types can be included [59]. Importantly, an iSSF produces a mechanistic movement model that is empirically parameterized [43, 48].

We took advantage of the fact that the conditional logistic regression model is a likelihood-equivalent to a Poisson model with stratum‐specific fixed intercepts. The approach outlined by Muff et al. [62] uses a mixed modelling approach which allows intercepts and/or slopes to vary by individual, while also incorporating shared information that is present in the data from different individuals [60]. For social species that may move collectively, and therefore have correlated movement trajectories, varying intercepts by individual is recommended to account for correlation within nested groupings of locations [61]. However, we are unable to nest individual within groups in our models because caribou live in fission–fusion societies [32], where group membership changes through space and time.

Following Muff et al. [62], all variables included in the fixed effect structure were also included in the random effect structure. Leveraging the method described by Muff and colleagues allows us to quantify population level effects of covariates (fixed effects) while controlling for individual variation in each covariate (random effects), which ultimately ensures fixed effect coefficients are more robust than previously described two step approaches [62]. For models to yield unbiased results, each fixed effect must also be parameterized as a random slope at the individual animal level. With GPS data, the quantity of data is higher (i.e., ~ 900 observations per animal) than most observational studies over a similar time-period. The quantity of data therefore allows the development of more complex models. Each fixed effect and interaction term has biological relevance for our objectives, indicating that a single large model is more effective for the comparison of effect sizes than many smaller models. From a single model, we test the importance of covariates relative to one another, therefore enabling us to test our objectives related to social foraging.

Our initial model included the proportion of lichen, forest, and open habitat within 200 m of the point location, the natural log-transformed step length, natural log-transformed nearest neighbour distance, and log-transformed time-lagged weekly dyadic SRI (O1, see Fig. 1 for model visualization). Given high collinearity among habitat types, we removed open habitat from our model for all subsequent analyses (Table S2). We natural log-transformed step length, nearest neighbour distance, and weekly dyadic SRI to account for potential decay function in each variable, to scale variables, and to ensure model convergence. We then generated three follow up models. First, we fit a null model that only included log-transformed step length, lichen habitat, and forest habitat within 200 m of the point location as well as interactions between step length and both habitat types (Table S3). Second, we fit a model that include the same terms as model 1, but also included turn angle in an interaction with both habitat types (Table S3). Third, we fit a model that included social variables: nearest neighbour distance (m), which was measured as the distance between a focal individual and the nearest collared conspecific and was calculated for all used and available steps and time-lagged weekly dyadic SRI. In our third model, we also included interactions between: 1) step length and each of the proportion of lichen and forest habitats within 200 m of the point location, 2) nearest neighbour distance and step length, 3) nearest neighbour distance and each of the proportion of lichen and forest habitats within 200 m of the point location as separate interactions, 4) log-transformed SRI and each of the proportion of lichen and forest habitats within 200 m of the point location. For interactions that included nearest neighbour distance, we used either distance at the start of a step or at the end of the step, depending on the other variable in the interaction (Fig. 2). Specifically, for the interaction between step length and nearest neighbour distance (O2), we used distance at the start of the step because the likelihood of taking a shorter or longer step is predicted to vary based on the distance to conspecifics before the step is taken. By contrast, for interactions between habitat variables and nearest neighbour distance, we used distance at the end of the step because the likelihood of selecting a given habitat is predicted to vary based on the distance to conspecifics when that habitat is being selected, i.e., at the end of the step. To assess model fit, we compared the null model to the model with social variables using AIC [63]. We fit our iSSF model using the glmmTMB package in R [64].

Calculating effect sizes

We calculated individual-level relative selection strength (RSS) to demonstrate how habitat features influenced selection [65]. Our model was therefore interpreted using RSS, which is the biological effect size extracted from conditional logistic models [65]. Interpretation of RSS provides information about how a covariate is selected or avoided across an environmental gradient of what is available to the individual. We calculated the strength for selecting one step over another that differed in the habitat value where those steps ended. RSS was calculated for each habitat type (i.e., forest or lichen habitats) as a function of nearest neighbour distance and the shared dyadic SRI between nearest neighbours.

Results

Social networks were generally well connected such that most individuals associated with most other individuals at least once during the winter (Fig.’s S1, S2, and S3). On average, graph density (i.e. the proportion of realized pairwise connections in each network) was 0.83 (2017 = 1.00; 2018 = 0.78; 2019 = 0.71). Moreover, individuals tended to cluster geographically, suggesting a possible spatial driver of social network structure (Fig.’s ,S1, S2, and S3yadic SRI values for weekly networks were skewed to zero with a long tail (median across all networks [± SD]: 0.03 ± 0.05), indicating that even though most individuals were connected at least once throughout the winter, it is clear some dyads consistently demonstrated social preference for one another at shorter timeframes.

We found evidence for an effect of the social environment on patterns of habitat selection (Table 1; Fig. 3). Specifically, our iSSF model which included social variables was considered a better fit than our null models (ΔAIC = 19,000, Table S3). Given the social nature of caribou, the effect of the social environment on selection was nuanced, and we found partial support for our expectation of social foraging. Overall, individuals took longer steps than expected (0.34, 95% CI: 0.18, 0.50; Table 1) and tended to moved more slowly when selecting lichen (– 0.38, 95% CI: – 0.45, − 0.31) and forest habitat (− 0.36, 95% CI: − 0.43, − 0.31) (Fig. 4; Table 1). Meanwhile, individuals tend to be closer to their nearest neighbours when selecting lichen habitat (0.55, 95% CI: 0.46, 0.65; Table 1) and relative selection strength for lichen decayed as nearest neighbours were further away (Fig. 5). Though the interaction between forest habitat and nearest neighbour distance was non-significant (− 0.01, 95% CI: − 0.08, 0.11; Table 1), relative selection strength for forest habitat decayed as nearest neighbours were further away (Fig. 5). Finally, individuals tended to share a higher dyadic SRI value with their nearest neighbours when selecting lichen habitat (− 4.03, 95% CI: − 4.64, 3.43), relative to the availability of lichen habitat (Table 1), while there was no relationship between forest habitat and SRI (− 0.21, 95% CI: − 0.78, 0.36; Table 1). Taken together, relative selection for lichen habitat was stronger than forest habitats as a function of nearest neighbour distance and SRI (Fig. 5). Our k-fold cross-validation had high scores (rho = 0.81 SE ± 0.03), demonstrating our model was better than random at predicting where caribou moved (see Fig. S4 to view coefficients and error for variables in each fold).

Table 1 Summary of integrated step selection function. We considered covariates significant if the 95% confidence intervals did not overlap zero

Fig. 3

Habitat selection coefficients and 95% confidence intervals for our integrated step selection model. Coefficients that do not overlap zero are deemed significant (see Table S3). Colours denote which habitat each fixed effect term was associated with and in cases where a term was not associated with habitat, points were coloured grey. Step length, nearest neighbour distances (both start and end), and the simple ratio index were log-transformed. Measures of distance at the end of a step are included when the distance term is a single term in a model or included in an interaction with habitat type, but measures of distance at the start of the step are included when the distance term is in an interaction with step length. These terms model the effect of nearest neighbour distance at the end of a step on the likelihood of selecting a given habitat type because in the case of habitat selection, proximity to conspecifics is more relevant for where an animal is going. By contrast, the effect of nearest neighbour distance at the start of a step on the likelihood of taking a shorter or longer step because in the case of movement, proximity to conspecifics is more relevant for where an animal is starting. Note, the lower panel is a zoom of the upper panel for enhanced visualization of coefficients with relatively small confidence intervals

Fig. 4

Relationship between expected speed (m/hr) as a function of changes in the proportion of lichen and forest habitats within 200 m of a given point location for individual caribou

Fig. 5

Relative selection strength of forest and lichen habitats as a function of nearest neighbour distance (m) (top panels) and shared dyadic simple ratio index between nearest neighbours (bottom panels). The dotted horizontal line represents no response, while values above the line indicate the population is selecting to be closer to that habitat than expected or to have higher shared dyadic SRI than expected, and below the dotted line the population is selecting to be farther from that habitat than expected or to have a lower shared dyadic SRI than expected. Interpretation for RSS values are that individuals generally tend to select to be near to conspecifics when selecting lichen habitat relative to its availability, whereas the response to nearest neighbours in forest habitat relative to its availability is limited. Meanwhile, individuals select for familiar nearest neighbours in all habitats

Discussion

Our study examined apparent social preference in the context of shared space use using socially informed integrated step selection functions. We integrate contemporary social network analysis within a traditional movement ecology and habitat selection framework. Although social networks were well connected at the population level (Fig.’s S1, S2, and S3), we highlight two forms of social preference, including long-term temporal stability of associations among individuals, and an effect of short-term social familiarity on habitat selection. Further, we found that individuals tended to select foraging habitat near familiar individuals, which could be one way that individuals reduce competition for forage at cratering sites. Based on our unification of social network analysis with integrated step selection functions, we highlight the influence of collective movement and preferred associations on habitat selection and foraging.

Our findings reveal that caribou are social in nearly all circumstances, although we observed a social hierarchy of movement-dependent social associations. Specifically, individuals tended to select to be closer to nearest neighbours in lichen habitat relative to forest habitats regardless of the familiarity of nearest neighbours. Within the movement ecology literature for ungulates, there is an assumption that slower movement in a given habitat represents foraging behaviour and faster movement represents searching behaviour [29]. Our results support this assumption; individuals moved slowly in lichen habitat. In conjunction with our observations in the field and past work in our system, we consider lichen habitat as key foraging habitat and dietary item for caribou in winter [33, 34, 52]. Individuals are more likely to trust social information about food sources and predation risk from familiar individuals, but the potential costs are an increase in competition at foraging patches. Despite the potential for competition, individual selection for lichen habitat was strongest in close proximity to conspecifics, suggesting a potential trade-off between the costs (i.e. competition) and benefits (i.e. access to information or anti-predator vigilance) of social foraging. Lichen habitat is typically open, suggesting the possibility that individuals may remain in visual and vocal contact, thereby facilitating and maintaining social cohesion during foraging [66]. Our results are also corroborated by other ungulate systems. In bison (Bison bison), the social environment in combination with recent knowledge of local foraging options dictated whether individuals followed, or left, a group [67]. Moreover, in the bison system, the costs and benefits of foraging in a group are moderated by collective decision making [68] and collective movement [69], both of which are likely involved in the foraging decisions made by caribou. Here, we elucidate potential behavioural mechanisms (i.e., foraging) that influence the frequency and magnitude of social associations.

The emergent geometry of collective movement and spatial arrangement of individuals in a group appears to change as individuals adjust their behaviour based on the availability of resources and the presence of familiar conspecifics [70]. Assamese macaques (Macaca assamensis) distance from one another during foraging, but move collectively between foraging sites [71], while individual giraffes (Giraffa camelopardalis) show social preference for conspecifics during foraging, but not during movement [24]. Interestingly, macaques foraged in closer proximity to individuals of similar dominance rank, but for giraffes it was unclear whether observed social preference was the result of passive or active assortment. For caribou, dominance hierarchies are linear and typically driven by body size [72], suggesting that social preference in caribou could also be related to dominance. Our ability to delineate aspects of the social environment between collective movement and habitat selection is useful for disentangling passive or active assortment, for example dominance rank, conspecific attraction, or the transfer of information about foraging resources.

We assumed that moving with familiar conspecifics is the result of information transfer about the location or quality of cratering sites. Despite clear patterns of social foraging, the question remains how individuals balance potential competition at craters in the winter, which can be substantial [72]. Selection for lichen habitat relative to its availability in groups could reflect the use of social information about the location of foraging sites [32] or predation [73], both of which could counter-act potential for direct competition. Moreover, it is possible that direct competition among familiar individuals is low; an assumption which requires further testing. Craters can vary in size and distribution [33]; however, craters may only be large enough for a single individual to forage at a time [74]. We propose that while caribou generally have larger group sizes in winter [75], groups vary in size based on movement and habitat selection behaviour presumably to balance the trade-off between competition and information acquisition. Furthermore, female caribou often have antlers, which unlike males, persist into winter. Females are hypothesized to use their antlers to defend craters and exert dominance over both males and females without antlers [72, 76]. Our interpretation is corroborated by theory used to explain fission–fusion dynamics, where individuals are expected to split and merge through space and time to reduce conflict and competition during foraging [77, 78].

We wish to highlight an important caveat related to sample size in our study. Despite the small number of individuals each year (n = 21 total individuals; N = 38 caribou-years), the number of GPS relocations per individual was high (n ~ 900 relocations). Street et al. [79] highlight that as few as two individuals (but ideally > 20) with 100 relocations per individual may be used to infer population-level habitat selection using selection function type analyses. Our sample size of individuals and relocations per individual far exceed minimum thresholds, and we are confident our findings are stable for individuals and populations [79].

Within a socially informed integrated step selection framework, we bridge the theoretical and methodological gap between social network analysis, movement ecology, and habitat selection. We demonstrate that although caribou are social in nearly all situations, individuals tend to be the most social when selecting lichen relative to its availability. Our synthesis of integrated step selection functions with social networks to describe the relationships between sociality, movement, and habitat selection is an important step towards identifying the roles of physical space and animal space use as factors influencing the social environment [12]. Moreover, individual variation in phenotypes attributable to movement or habitat selection may affect how individuals experience the social environment [80, 81]. Movement, habitat selection, and social behaviour are clearly linked; as van Moorter et al. [2] described movement as the ‘glue’ connecting habitat selection to the physical location of a given set of habitats, we posit that movement is the glue connecting collective habitat selection to the social environment.