Environmental and individual drivers of animal movement patterns across a wide geographical gradient



  1. Within the rapidly developing field of movement ecology, much attention has been given to studying the movement of individuals within a subset of their population's occupied range. Our understanding of the effects of landscape heterogeneity on animal movement is still fairly limited as it requires studying the movement of multiple individuals across a variety of environmental conditions. Gaining deeper understanding of the environmental drivers of movement is a crucial component of predictive models of population spread and habitat selection and may help inform management and conservation.
  2. In Ontario, woodland caribou (Rangifer tarandus caribou) occur along a wide geographical gradient ranging from the boreal forest to the Hudson Bay floodplains. We used high-resolution GPS data, collected from 114 individuals across a 450 000 km2 area in northern Ontario, to link movement behaviour to underlying local environmental variables associated with habitat permeability, predation risk and forage availability.
  3. We show that a great deal of observed variability in movement patterns across space and time can be attributed to local environmental conditions, with residual individual differences that may reflect spatial population structure.
  4. We discuss our results in the context of current knowledge of movement and caribou ecology and highlight potential applications of our approach to the study of wide-ranging animals.


Movement ecology is a rapidly growing field of research aimed at understanding the ‘causes, mechanisms and spatiotemporal patterns of movement and their role in various ecological and evolutionary processes’ (Nathan et al. 2008). One potential way to test putative causal mechanisms is to link observed movement patterns with the spatial and temporal variability in their underlying environmental conditions. As experimental manipulation is often unpractical in many ecological systems, a promising observational approach is to use ‘space as a surrogate’ (McIntire & Fajardo 2009) – comparing the movement behaviour of multiple individuals across wide geographical and seasonal gradients, thus encompassing a variety of environmental conditions. Despite the benefits of such large-scale comparative studies, the high cost and technical challenges involved in tracking animals in the wild have heretofore resulted largely in studies focused on local populations within a limited portion of their occupied range.

The expected squared displacement of an animal's path, E(R2), has long been a central concept in the study of animal movement (Kareiva & Shigesada 1983). It is directly proportional to the population rate of spread (Turchin 1998) and may serve to quantify migration and dispersal patterns (Bunnefeld et al. 2010; Börger & Fryxell 2012). Furthermore, E(R2) has been nominated recently as a single summary statistic that best captures key statistical properties of animal movements (Nouvellet, Bacon & Waxman 2009). Here we demonstrate the utility of this summary statistic in linking local environmental conditions to animal movement rates. While the observed displacement of an individual is but a single realization of a stochastic movement process, E(R2) reflects the mean displacement expected under a set of assumptions regarding the statistical properties of the process. If these assumptions hold true, E(R2) allows one to better assess the general patterns and accordingly identify the ecological drivers of animal movement. This approach holds great potential for dealing successfully with a central challenge in movement ecology – scaling up animal movement across heterogeneous landscapes (Morales & Ellner 2002).

Movement patterns result from interactions between animals and their environments (Johnson et al. 1992; Schick et al. 2008). Theory suggests that movement rates should decrease when animals travel through resource-rich areas (Pyke, Pulliam & Charnov 1977; Avgar, Kuefler & Fryxell 2011), supported by numerous empirical studies (e.g. Klaassen, Nolet & Bankert 2006; Fryxell et al. 2008; Kuefler, Avgar & Fryxell 2012). Moreover, physical impediments, such as thick undergrowth or deep snow, may slow movement rates (Morales & Ellner 2002; Schooley & Wiens 2004). Conversely, unprofitable local conditions, such as sites with low forage availability or high predation risk, may stimulate increased movement rates (e.g. Englund & Olsson 1996; Gilliam & Fraser 2001; Frair et al. 2005), as would physical properties that increase habitat permeability, such as roads or waterways (e.g. James 1999). Analysing movement attributes in the light of local environmental conditions may thus help identify habitat preferences of free-ranging animals (Fauchald & Tveraa 2003; Barraquand & Benhamou 2008; Bastille-Rousseau, Fortin & Dussault 2010) while controlling for the mechanical effects of habitat permeability. Such an approach presents an attractive alternative to resource selection functions that have come into popular use in recent years (Boyce & McDonald 1999), because local trajectory responses may be less sensitive to the arbitrary definition of available resources (Beyer et al. 2010).

Movement patterns, as well as their relationship with underlying environmental conditions, are often scale dependent (Fryxell et al. 2008; Rivrud, Loe & Mysterud 2010). Animals may respond to their environment at specific spatial scales and the resulting movement patterns may be only discernible at a narrow range of temporal resolutions. Consequently, linking movement patterns to their underlying environmental drivers is often a complex task. Here we deal with this task by re-sampling animal movement paths at three different temporal resolutions and matching the resulting movement data with environmental variables sampled at the appropriate spatial resolution (as determined by the spatial scale of movement at each temporal resolution).

We explore the connection between landscape attributes and animal movement patterns for a population of woodland caribou (Rangifer tarandus caribou) that inhabit a vast geographical range, which encompasses a variety of environmental conditions. Woodland caribou occupy the southern range of caribou in North America and are considered a threatened subspecies over most of this range (COSEWIC 2002). Whereas several previous studies have focused on comparative analysis of caribou movement characteristics (e.g. Bergman, Schaefer & Luttich 2000), and their temporal dynamics (e.g. Ferguson & Elkie 2004a; Rettie & Messier 2001), few studies have investigated possible environmental drivers of these characteristics. Noteworthy exceptions are Johnson et al.'s studies (2002a,b) that estimate scale-specific effects of cover type on caribou movement in north-central British Columbia. The extensive geographic range of caribou in Ontario, from the boreal forest in the south-west to the Hudson Bay floodplains at the north-eastern extent, provides a unique opportunity to investigate the attributes and environmental correlates of movement in a wide-ranging herbivore across a broad range of ecological conditions.

We evaluate the relative importance of spatiotemporal variables associated with three main functional aspects of the landscape (forage availability, predation risk and habitat permeability) as potential drivers of caribou movement patterns. Interpretation of the effects of different habitat variables on biological processes is often hampered by the covariation characterizing spatial and temporal data (Rossi et al. 1992; Bini et al. 2009). For example, in our system, areas regenerating after natural or anthropogenic disturbance often support high population densities of moose (Alces alces; Rempel et al. 1997), the preferred prey species for wolves (Canis lupus; Boertje, Valkenburg & McNay 1996), which in turn are the major predator of woodland caribou (Bergerud & Elliot 1986; Seip 1992; McLoughlin et al. 2003). It is accordingly plausible that caribou might avoid or travel faster through regenerating forest stands, as supported by some field evidence (Cumming, Beange & Lavoie 1996; James et al. 2004; Vors et al. 2007; Leblond et al. 2011). On the other hand, during the growing season, regenerating forest stands often have abundant annual vesicular plants and may thus be attractive as foraging sites (Schaefer & Pruitt 1991; Leblond et al. 2011) in which caribou tend to linger. Here we deal with such discrepancies by classifying environmental variables into functional groups and making a priori predictions of their effects, based on our current knowledge of caribou ecology. This approach increases our confidence in the biological interpretation of our results and allows us to identify fundamental relationships between local environmental conditions and animal movement patterns. Consequently we are able to account for a great deal of the variability observed in caribou movement patterns across space and time and gain insights into habitat preference and population structure. We discuss implications of our results for caribou conservation and major applications of our approach in the context of movement ecology.

Materials and methods

Data Collection

The telemetry data employed in this study were collected by the Ontario Ministry of Natural Resources (OMNR). During February–April of 2009 and 2010, 165 adult female caribou across Northern Ontario were net-gunned from a helicopter and fitted with a GPS-Argos telemetry collars (Telonics Inc., 932 E. Impala Avenue, Mesa, AZ, 85204-6699, USA). Capturing and handling were performed according to Ontario animal care regulations. GPS locations were recorded every 5 h during the first 13 months after collaring, and every 25 h thereafter, and were transmitted to OMNR via an Argos satellite uplink every 5 days. Telemetry points collected during the first 24 h after capturing, the last 24 h before a mortality event, and those with obvious errors (e.g. a roundtrip to a remote location performed over a short time; <0·1% of all fixes) were excluded from our analysis. Overall, our telemetry data set was composed of more than 200 000 GPS points, taken over a period of 3 years, across a 450 000 km2 area.

Each telemetry point was affiliated with eight environmental variables characterizing that locality at that time. Environmental variables were collected from three main sources. MODIS-based normalized difference vegetation index (NDVI) values (temporal resolution: 16 days, spatial resolution: 250-m pixels) were obtained via the Land Processes Distributed Active Archive Center at the U.S. Geological Survey Earth Resources Observation and Science Center. Values of ΔNDVI (indicative of rate of vegetative growth rather than standing biomass; Pettorelli et al. 2005) were calculated as the difference between the current NDVI values and those 16 days previous. Snow depth data (temporal resolution: 3 h, spatial resolution: 40-km pixels) were obtained via the North America Regional Reanalysis data (NARR data set DSI-6175), from NOAA Operational Model Archive and Distribution System. Land cover types and distances to different landscape features (spatial resolution: 25-m pixels) were obtained from the Ontario Provincial Land Cover Database. This data base is based on the Thematic Mapper sensor on Landsat Earth-Resource satellites using data frames recorded between 1999 and 2002 (Spectranalysis Inc., unpublished data). Annual forest harvest and natural disturbance spatial layers available from the OMNR were used to update the land cover classification to ensure it accurately reflected habitat conditions during the years of animal monitoring.

Environmental variables were recorded for each GPS point at three spatial scales, within radii of 250, 950 and 9300 m. These radii correspond to half of the median displacements (i.e. half the length of a typical step) of caribou over the entire study period at three temporal resolutions: 5, 25 h and 30 days, respectively. Temporally dynamic variables were averaged over the relevant temporal steps (e.g. a 25-h step is affiliated with the average of eight snow depth values occurring at that locality during that time). All spatial processing was conducted using ArcGIS desktop 9.2 and 9.3 (ESRI Inc., 380 New York Street Redlands, CA 92373).


Explanatory variables were classified according to their expected impact on three functional aspects of the landscape. To account for potential interactions between environmental variables and time of year (e.g. lakes may enhance movement when frozen during winter yet impede it during the summer), all explanatory variables were separated into summer (May–October) and winter (November–April) effects.

The first functional aspect of the landscape is habitat permeability. Permeability should be positively correlated with water cover during the winter months (when lakes are frozen and may serve as movement corridors), and with open habitat (a combination of all treeless land cover classes) year round. Hence, these variables are expected to have a positive effect on movement rates. Note, however, that treeless areas may still have dense vegetation cover impeding movement and hence weakening this positive effect. Conversely, permeability should be negatively correlated with winter snow depth, which is thus expected to have a negative effect on movement rates. Snow has been shown to inhibit caribou movement in Alberta (Stuart-Smith et al. 1997) and as a factor affecting feeding habitats in British Colombia (Johnson, Parker & Heard 2001).

The second functional aspect of the landscape is forage availability. During winter months, when caribou feed mainly on lichen (usually associated with coniferous habitats), conifer forest cover (a combination of all land cover classes dominated by coniferous species) should be positively correlated with forage availability (Brown, Rettie & Mallory 2006). During summer months, ΔNDVI, reflecting vegetative growth or forage quality, should be positively correlated with forage availability. Finally, NDVI, reflecting vegetative cover or forage quantity (Pettorelli et al. 2005), should be positively correlated with forage availability year round. Note that, whereas winter NDVI values do not reflect caribou winter forage (i.e. lichen) per se, it is likely to reflect coniferous habitats (associated with lichen) and may contain information that is missing or not updated, in the conifer forest cover layer. Several studies have used NDVI to account for habitat selection (but not movement) by caribou (Gustine et al. 2006; Thomas, Johnson & Griffith 2006) and reindeer (Hansen et al. 2009a,b). All three variables are expected to have negative effects on movement rates owing to the tendency to linger in resource-rich habitats.

The last functional aspect we consider here is predation risk. The main predators of caribou in our system are wolves, which in turn rely on moose as their major prey. Both deciduous forest cover (a combination of all land cover classes dominated by deciduous species) and regenerating forest cover (cuts and burns that are <20 years old) should be positively correlated with moose occurrence, and thus with high predation risk (Vors et al. 2007), and are expected to have positive effects on movement rates year round. During the summer months, variables associated with refuge habitats (James et al. 2004), namely conifer forest cover and water cover (caribou are excellent swimmers and are known to use islands during spring and summer), should be negatively correlated with predation risk and are thus expected to have negative effects on movement rates. These predictions are summarized in Table 2.

Data Processing

Fine-scale movement patterns were characterized by three basic attributes of a biased correlated random walk (Benhamou 2006; Codling, Plank & Benhamou 2008) at two temporal resolutions. A single step was defined as two consecutive GPS locations, taken either 5 or 25 h apart. Each step, i, at each temporal resolution, is characterized by its length, li, (in metres), and its heading, θi – its orientation relative to the true north (in radians). The cosine of the angular difference between the headings of two consecutive steps, cos(θiθi−1), is the step's directional persistence, c. The preferred direction, γ, is defined as the overall heading of the individual's movement path over an entire calendar month (from the first to the last GPS points). The cosine of the angular difference between the preferred direction and the step's heading, cos(γθi), is the step's directional bias, q. Note that steps and turns defined this way are not ‘natural’ steps performed by the animal, but rather derived variables based on a sample of animal locations recorded at regular intervals over time (Turchin 1998).

To reduce the strong autocorrelation stemming from working with sequential observations in space and time, data were aggregated to obtain monthly summaries of movement attributes and environmental variables per individual per month. A monthly minimum of 20 valid 5-h steps and five valid 25-h steps was imposed. To avoid potential biases, individuals that were represented by <3 summer months, or <3 winter months, were excluded from our analysis (data for 38 excluded individuals were used to test our approach's predictive abilities – see Results). The remaining monthly data set included 114 individuals with an average of 11 months per individual (range: 6–13) and was based on, on average, 980 steps of 5 h per individual (range: 327–1838), 9308 steps of 5 h per month (range: 7342–10 877) and 90 steps of 5 h per individual per month (range: 25–149).

Movement attributes were summarized for each individual, at each month, for each of the two finer temporal resolutions. The mean directional persistence, E(c), is a measure of the animal's tendency to maintain its previous movement heading. The mean directional bias, E(q), is a measure of the animal's tendency to maintain its average monthly movement heading. To account for potential biased estimations resulting from ql or cl cross-correlations (e.g. a tendency to move faster when moving towards a preferred direction), both means were taken as weighted averages based on step lengths. Note that E(c) and E(q) are not independent, as the directional persistence increases with the directional bias (Benhamou 2006). To enable straightforward calculation of the expected monthly squared displacements, E(R2), we classified the path of each animal at each month as either a biased random walk (BRW) or a correlated random walk (CRW), based on the correlation between directional persistence and directional bias (Benhamou 2006). If the correlation between c and q, cor[q,c], was positive and significant, the path was classified as a BRW and its E(R2) was calculated as (Codling, Plank & Benhamou 2008 and refs therein):

display math(eqn 1)

Where E(l) and E(l2) are the mean step length and squared step length, and n is the expected number of steps in a month (= 144 for 5-h steps and = 29 for 25-h steps). Otherwise, the path was classified as a CRW and its E(R2) was calculated as (Benhamou 2006 and refs therein):

display math(eqn 2)

We recorded E(R2), E(l), E(l2), E(c), E(q) and cor[q,c], for each animal, at each month, and for each of the two finer temporal resolutions. At the monthly temporal resolution, we recorded the observed monthly displacement, the straight-line distance travelled by each animal during each calendar month, to be coupled with the environmental variables measured at the appropriate spatial resolution (i.e. within a 9300-m radius around the first GPS point of the month) and compared with E(R2) values obtained at the two finer temporal resolutions.

For each individual, at each month, and at each spatiotemporal resolution, mean values of all environmental variables were calculated by averaging all available values for a specific variable at a specific resolution. Individual monthly range centroids (northing and easting) were calculated as the harmonic mean of all northing and easting values for each individual in each month.

Statistical Analysis

At each spatiotemporal resolution, the processed data set consisted of 6–13 data points (i.e. months) for each individual. Linear mixed effects models were used to account for the nested data structure and lack of independence among repeated measures. Mixed effect models enable treating individual animal identity as a random effect, thereby assigning a different intercept to each individual in the population, while treating environmental effects as fixed. This allows one to estimate regression coefficients for the entire population while accounting for individual differences. Function lme within package nlme in program r (version 2.14.0; http://cran.r-project.org/) was used, with an exponential spatial and/or temporal autocorrelation structure for the residuals.

The response variables in our analysis are the individual monthly movement attributes: √E(R2), E(l), √E(l2), E(c), E(q), cor[q,c] and observed monthly displacement. All displacement variables (√E(R2), E(l), √E(l2) and observed monthly displacement) were log-transformed to reduce heteroscedasticity. All correlation variables (E(c), E(q) and cor[q,c]) were logit-transformed. The explanatory variables are the mean environmental variables experienced by each study animal during each month, at the appropriate spatial resolution. All variables were scaled (by subtracting the within-season mean and dividing by the within-season standard deviation) so as to standardize their coefficients (Schielzeth 2010). The same candidate set of explanatory variables (Table 2) was used for all dependent variables. The appropriate residual exponential autocorrelation structure (i.e. temporal, spatial, both or none) was selected for each response variable based on AIC competition using the full set of explanatory variables (semivariograms are provided in Appendix S1, Supporting information).

Model selection was performed according to the general guidelines in Zuur et al. (2009; chapter 5). We started by fitting the full model (including 14 explanatory variables; Table 2) for each response variable, and then gradually reducing complexity, based on AIC competition, to obtain the best combination of explanatory variables (interactions were not considered). Finally, as an indicator of goodness-of-fit, the portion of variance in each response variable, explained by the selected set of environmental variables (i.e. fixed effects), was calculated as the difference between the variance of the response variable and the variance of the population-level residuals of the best model (i.e. considering only fixed effects), divided by the variance of the response variable (coded in r as: (var(response)−var(resid(best.model, level = 0)))/var(response)).


Expected monthly displacement values, √E(R2), calculated based on either BRW or CRW, matched well with observed monthly displacement values and explained 68% (at the 5-h resolution) and 65% (at the 25-h resolution) of observed variability (Fig. 1). However, our BRW/CRW-based expectations tended to overestimate observed monthly displacement values at the higher end of the range (Fig. 1). Overestimations were positively associated with high NDVI values (explaining c. 10% of the variation in √E(R2)-observed monthly displacement deviations) and had an additional individual (i.e. random) component (explaining an additional c. 10%). The frequency of monthly paths classified as BRW was 13% at the 5-h resolution and 7% at the 25-h resolution. All movement attributes varied substantially over time and space (Table 1 and Fig. 2). Overall, movement rates were highest in the north and during the winter (Fig. 2).

Figure 1.

Observed vs. expected monthly displacements [observed monthly displacement vs. √E(R2)]. Circles represent expectations based on a correlated random walk (CRW) while triangles represent expectations based on a biased random walk (BRW). Small markers represent expectations based on 5-h steps while large markers represent expectation based on 25-h steps. The black line represents a perfect match (for reference).

Figure 2.

Seasonal means of observed (blue) and expected (at 5-h resolution; red) monthly displacements. Points represent the centroids of the seasonal range of each individual. Colour gradient represents interpolated (krigged) values of monthly displacements. The grey area in the north-west is Hudson Bay. Expected monthly displacements at 25-h resolution show very similar trends.

Table 1. Caribou movement attributes
Temporal resolutionMovement attributeaLower 95% CIMedianUpper 95% CI
  1. a

    By order of appearance, these attributes are as follows: the mean step length, the mean squared step length, the mean directional bias, the mean directional persistence and the mean squared displacement.

5 hE(l) (m h−1)46·14196·95615·90
E(l2) (m h−1)87·30341·65948·57
E(R2) (km per month)5·1723·00183·05
25 hE(l) (m h−1)23·93124·18511·02
E(l2) (m h−1)33·85184·71661·66
E(R2) (km per month)4·0826·93192·89
30 daysObserved monthly displacement (km per month)0·8620·30197·67

Most environmental variables considered here were selected as predictors of monthly movement rates during at least one of the seasons, and the direction of their effects matched our predictions (Table 2). The only notable exception is the positive effect of conifer forest cover on observed monthly displacement during winter. Both regenerating forest cover and deciduous forest cover, variables affiliated with predation risk, had negligible effects on monthly movement rates at all resolutions. The same was true for open habitat (affiliated with permeability) during the summer.

Table 2. Monthly displacements – predictions and results
 Functional aspectEnvironmental variablePredicted effectObserved monthly displacement (30 days)E(R2) (25 h)E(R2) (5 h)
  1. NDVI, normalized difference vegetation index. A priori predictions and results for the effects of different landscape variables on woodland caribou movement across space and time at three different spatiotemporal resolutions. As all variables were scaled, the regression coefficients presented here are standardized and their magnitudes indicate their effect size.

SummerHabitat permeabilityOpen
Forage availabilityNDVI−0·71−0·64−0·52
Predation riskConifer−0·17−0·21
WinterHabitat permeabilityOpen0·250·390·36
Snow depth−0·29−0·24−0·31
Forage availabilityNDVI−0·29−0·25−0·26
Predation riskDeciduous
Random effects standard deviation0·400·360·38
Variance explained by fixed effects (%)304342
Total explained variance (%)505960

Our models accounted for most of the observed variability in caribou monthly displacement across space and time, a considerable portion of which was explained by the environmental variables (Table 2). Regression coefficients magnitude indicate that NDVI had the largest effect on monthly displacement during the summer whereas variables affiliated with habitat permeability were most influential during the winter. The observed monthly displacement model included fewer fixed effects and had a lower goodness-of-fit, than the two √E(R2) models.

BRW/CRW components (E(l), E(l2), E(c), E(q) and cor[q,c]) varied in their response to environmental variables but overall reflected the trends observed in √E(R2) (while this was true for both the 5- and 25-h resolutions, we report here the results for the former; Table 3). Environmental conditions had the greatest influence on E(c) whereas cor[q,c] was the least sensitive component to the variables considered here. Note that, contradictory to our prediction, habitat variables thought to be affiliated with winter predation risk had negative effects on both E(l) and E(l2) whereas summer ΔNDVI had a positive effect on E(q). Overall, predicted √E(R2) values, calculated based on the predicted values for each of the biased CRW components, (CRW was assumed if cor[q,c] ≤ 0 and BRW was assumed otherwise) matched observed monthly displacement values well (explaining c. 40% of observed variability) and were unbiased (linear regression intercept and slope did not significantly differ from 0 and 1, respectively).

Table 3. Biased correlated random walk (CRW) components (5 h)
 Functional aspectEnvironmental variableE(l)E(l2)E(q)E(c)cor[q,c]
  1. NDVI, normalized difference vegetation index. The effects of different landscape variables on the spatiotemporal variability in the components of caribou biased CRW at 5-h resolution. As all variables were scaled, the regression coefficients presented here are standardized and their magnitudes indicate their effect size.

SummerHabitat permeabilityOpen0·33
Forage availabilityNDVI−0·27−0·44−0·39−0·72−0·30
Predation riskConifer−0·37
WinterHabitat permeabilityOpen0·28
Snow depth−0·44−0·41−0·24−0·21
Forage availabilityNDVI−0·32−0·28−0·43−0·19−0·34
Predation riskDeciduous−0·11
Residuals autocorrelation range1·2 months0·6 months1·5 km0·6 months0·39 km
Variance explained by fixed effects (%)424317398
Total explained variance (%)5456255616

Random effects in our statistical models reflect individual differences that are not explained by the fixed (environmental) variables (Tables 2 and 3). Approximately 20% of the variability in monthly displacement rates was captured by individual identities. Individual effects did not vary with respect to the animal's estimated age at the time of capture. Individual effects did show a clear spatial trend, however, with a general increase from south to north (Fig. 3a). This trend was best modelled as a piecewise liner model of the individual effects as function of latitude (we found no significant effect of longitude). The model, fitted using function piecewise.linear within package SiZer, outcompeted any continuous model (based on AIC), indicating a breakpoint in the spatial trend at latitude 13 028 647 m (Lambert conformal conic projection; Fig. 3b).

Figure 3.

Individual effects on √E(R2) at 5-h resolution. (a) A map illustration of spatial trends in estimated individual movement rates (i.e. random effects), possibly reflecting spatial population structure (see Discussion). Dots represent the centroids of the annual range of each individual. Colour gradient represents interpolated (krigged) values of individual √E(R2) in km. (b) Individual movement rates as function of range latitude. Lines represent a fitted piecewise linear model.

We evaluated the predictive power of our approach using movement data previously omitted from our analysis. We used the best models selected for each of the biased CRW components (at 5-h resolution; Table 3), together with latitude-dependent piecewise liner models of their random effects, to predict √E(R2) for 38 individuals over 160 months, that had been excluded from our statistical models. These predictions were compared against the observed monthly displacement values (Fig. 4). Approximately 22% of the variance in observed values was explained by our predictions. The linear regression intercept was, however, significantly smaller than 0, whereas the regression slope was significantly larger than 1, indicating a tendency of our models to over-predict at low displacement rates and under-predict at high displacement rates.

Figure 4.

Observed vs. predicted (at 5-h resolution) monthly displacements for animal months omitted from our analysis. The black line represents a perfect match (for reference).


Several techniques have been developed to help identify profitable foraging areas based on attributes of the animal's movement path (reviewed in Barraquand & Benhamou 2008). These techniques vary from the intuitive ‘first passage time’ analysis (Fauchald & Tveraa 2003) applied, for example, to bottlenose dolphins in the Inverness Firth (Bailey & Thompson 2006), through to sophisticated state-space models (Morales et al. 2004) applied, for example, to elk in the Rocky Mountains (Frair et al. 2005), to residence time analysis applied to Arctic fulmars in the Faroe Islands (Barraquand & Benhamou 2008). Recently, Boettiger et al. (2011) quantified African elephant movement by calculating the smallest ellipse that contained all GPS points within a 24-h window and then linking the attributes of these ellipses to habitat variables using linear filtering. Our focal response variable, E(R2), is inversely proportional to the residence time proposed by Barraquand & Benhamou (2008) and may be further used to link spatiotemporal habitat distribution to population distribution and redistribution patterns (Turchin 1991).

Regardless of their methodological approach, the vast majority of current studies in movement ecology focus on a few individuals, occupying a small subset of the population range. This calls into question the degree to which estimated parameters truly reflect movement properties at the population level. The spatial and temporal extent of our study yielded strong statistical leverage for the identification and relative ranking of multiple forage-related landscape attributes, giving us a firmer basis for characterizing the study population.

Foragers, from ants (Avgar, Giladi & Nathan 2008) to antelopes (Fryxell, Wilmshurst & Sinclair 2004), change their movement behaviour in response to local forage abundance. For ungulates, time spent feeding is inversely, and disproportionally, related to overall movement rate (Owen-Smith, Fryxell & Merrill 2010). Our results clearly support the hypothesis that forage quality and availability (measured by NDVI and ΔNDVI during the summer and conifer forest cover and NDVI during the winter) suppresses movement rates by woodland caribou. This could explain why previous resource selection studies have found that caribou are most commonly found in conifer stands as well as regenerating forest stands (Ferguson & Elkie 2004b; Brown et al. 2007). Other studies of large herbivores have similarly found strong links between food resource availability and probability of habitat use, including gazelles, elk, reindeer and elephants (Fryxell, Wilmshurst & Sinclair 2004; Frair et al. 2005; Mueller et al. 2007; Fryxell et al. 2008; Hansen et al. 2009b; Boettiger et al. 2011).

Support for the predation risk hypothesis was more equivocal. Presumed indicators of winter predation risk (deciduous and regenerating forest cover) had little effect on caribou movement. Indicators of summer refuge habitat (conifer forest and water cover) were more strongly supported, which might indicate greater responsiveness by female woodland caribou to predation pressure during the summer months, when caribou calves are most vulnerable. Previous studies of other large ungulates have similarly found ambiguous results. Plains zebras in Kenya travel faster, yet with sharper turns, in areas of high predation risk (Fischhoff et al. 2007). Elk travel faster through areas of high predation risk in the Rocky Mountains of Canada (Frair et al. 2005). A second study in Yellowstone National Park, however, found high variability in the response of individual elk to local wolf densities, with some individuals showing reduced movement rates through areas of high predation risk and no obvious trend at the population level (Forester et al. 2007).

Such ambiguity might well be expected from a theoretical standpoint, as animals might reduce their movement rate to reduce the probability of encountering a predator or alternatively increase their movement rate to avoid high predation areas all together. Indeed, it has been suggested that woodland caribou avoid predators at the home-range level rather than through more fine-grained patterns of habitat use (Rettie & Messier 2000). Coarse scale range selection would quite likely result in low explanatory power of predation-related habitat variables at a finer spatial scale. Moreover, the dynamic nature of predation risk across space and time likely weakens its correlation with remote-sensed habitat characteristics. Hence, without direct assessment of local habitat utilization by predators at the appropriate spatiotemporal scale, evaluation of movement responses to predation risk based on habitat attributes alone is necessarily tentative.

Physical or structural attributes of the landscape may also affect permeability to mobile herbivores. Owing to the inherent spatiotemporal covariation of landscape attributes, accounting for habitat permeability is therefore crucial for unbiased inference regarding the effects of other functional characteristics of the landscape on movement patterns (Gaillard et al. 2010). Previous studies of other large herbivores have demonstrated that habitat permeability can have direct fitness consequences owing to its impact on energy balance (e.g. deep snow, thick vegetation or elevation gradients; Parker, Robbins & Hanley 1984; Dailey & Hobbs 1989). While the magnitude of the response of woodland caribou movement rate to habitat permeability does not necessarily translate into absolute energy expenditure, it does allow a comparison of the relative effects of different structural attributes, perhaps providing insights into the energetic consequence of future environmental change (Mysterud et al. 2008).

Temporally dynamic environmental variables are rarely used in studies of animal movement and habitat preference. Movement patterns reflect interactions between animals and their environment (Schick et al. 2008) and understanding them often requires consideration of the temporally dynamic nature of these environments (Mueller et al. 2011). For example, spatial distributions of Thomson's gazelles in Africa and Mongolian gazelles in Asia continually change from month to month, like a shifting mosaic (Fryxell, Wilmshurst & Sinclair 2004; Mueller et al. 2007). Similarly, home-range size of red deer at multiple temporal scales is driven by temporally dynamic variable such as temperature, precipitation, day length and snow cover (Rivrud, Loe & Mysterud 2010). Hence, studies of animal movement can greatly benefit from the incorporation of temporally dynamic environmental variables, as demonstrated by the relative large effect size of NDVI, ΔNDVI and snow depth in our study. Indeed, comparison between the movement characteristics of different ungulates species suggests that major classes of movement, such as migration and nomadism, might reflect an adaptation to the spatiotemporal dynamics of resource distribution across the landscape (Mueller & Fagan 2008; Mueller et al. 2011).

Our analysis was conducted at three spatiotemporal scales to account for possible scale-specific relationship between movement patterns and environmental variables. Additional analysis (not reported here) indicates that the portion of variance explained by the fixed effects declines whenever there is a mismatch between the temporal resolution of the response variable and the spatial resolution of the explanatory variables. While our results do not indicate substantial difference between the two finer temporal resolutions (i.e. 5 and 25 h), inference of environmental drivers of movement at a monthly resolution was slightly weaker, possibly due to the coarse spatial resolution of the environmental variables.

Analysis of telemetry data is often confounded by the nested nature of such data, where each individual is represented by multiple data points. Here we have dealt with the resulting pseudoreplication by using mixed effects models, allowing us to infer environmental effects (i.e. the fixed effect) on caribou movement while taking into account the hierarchical nature of the data. Additional useful information may lay in the estimated individual intercepts (i.e. the random effects). Spatial trends evident in the random effects (Fig. 3) might indicate environmental gradients (other than the ones accounted for here) affecting movement attributes. Alternatively, such trends could be attributed to innate behavioural differences among individuals, either in the propensity to move or in the response to local conditions, reflecting underlying spatial population structure.

Woodland caribou in Ontario is commonly delineated into a sedentary forest-dwelling ecotype and a migratory forest-tundra ecotype (Thomas 1995; Harris 1999; COSEWIC 2002; OWCRT 2008; Nagy et al. 2011). Previous distinctions between these woodland caribou ecotypes have been based on qualitative characterizations of migratory and calving behaviours. The spatial patterns of individual effects reflected in Fig. 3 provide a more rigorous quantitative corroboration of ecotype delineation based on fine-scale movement behaviour. The latitudinal breakpoint detected in this pattern may indicate the transition zone between these two ecotypes and thus inform management policy.

We have demonstrated the use of an extensive data set of telemetry and remote-sensed environmental variables to link patterns of movement to local conditions experienced by individual animals across a vast landscape. The approach presented here has several potential applications. First, as a behaviour-based indicator for habitat preference, our approach might complement traditional analysis based on resource selection functions, providing the basis for predictive models of animal occurrence (Moorcroft & Barnett 2008; see Bastille-Rousseau, Fortin & Dussault 2010 for a detailed discussion on residency time analysis). Unlike most resource selection analyses (relying on a comparison of habitat use vs. availability), our approach does not depend on a definition of availability, enabling a straightforward consideration of temporally dynamic variables and allowing the incorporation of random effects (but see Gillies et al. 2006; Fieberg et al. 2010). Second, habitat-dependent movement rates could help with development of spatially explicit models of gene flow or range expansion across novel or altered landscapes. Third, variables affecting habitat permeability are often crucial components in spatially explicit energetic models and may serve to infer the effects of future climate change on energy expenditure by animals in the wild (Parker, Robbins & Hanley 1984; Dailey & Hobbs 1989; Mysterud et al. 2008). Lastly, residual random effects may be useful in identifying sub-population structure. Fuelled by rapid advancement in both telemetry and remote sensing technologies, these applications will advance our understanding and conservation of wide-ranging wildlife species.


We thank D. Kuefler, B. van Moorter and an anonymous reviewer for helpful comments on an earlier draft of the manuscript, and staff of the Ontario Ministry of Natural Resources who supported delivery of the telemetry monitoring field work. This work was supported by Natural Sciences and Engineering Research Council of Canada (NSERC) Discovery, Strategic, and Collaborative Research and Development Grants (to J.M.F.), NSERC Vanier Fellowship (to T.A.) and the Ontario Ministry of Natural Resources.