Stick or twist: roe deer adjust their flight behaviour to the perceived

0 downloads 0 Views 1MB Size Report
porting the 'flush early and avoid the risk' (FEAR) hypothesis. Despite this, flight .... Study Site and Species. We recorded the flight behaviour of collared free-ranging roe ..... Jr., & Blumstein, D. T. (2014). Novel effects of monitoring predators on.
Animal Behaviour 124 (2017) 35e46

Contents lists available at ScienceDirect

Animal Behaviour journal homepage: www.elsevier.com/locate/anbehav

Stick or twist: roe deer adjust their flight behaviour to the perceived trade-off between risk and reward ge C. Bonnot a, b, *, A. J. Mark Hewison c, Nicolas Morellet c, Jean-Michel Gaillard b, Nade lie Couriot c, Bruno Cargnelutti c, Yannick Chaval c, Bruno Lourtet c, Lucie Debeffe d, Ophe a cile Vanpe b Petter Kjellander , Ce € Wildlife Research Station, Department of Ecology, Swedish University of Agricultural Sciences, Riddarhyttan, Sweden Grimso CNRS, UMR5558, Laboratoire de Biom etrie et Biologie Evolutive, Universit e Lyon 1, Villeurbanne, France CEFS, Universit e de Toulouse, INRA, Castanet-Tolosan, France d Department of Biology, University of Saskatchewan, Saskatoon, Canada a

b c

a r t i c l e i n f o Article history: Received 4 July 2016 Initial acceptance 5 August 2016 Final acceptance 11 November 2016 MS. number: 16-00591R Keywords: assessment interval deer detection delay flight initiation distance flush early and avoid the rush hypothesis optimal escape theory risk assessment riskeresource trade-off starting distance threat monitoring

Because avoiding predation is crucial for fitness, foraging animals must trade acquisition of high-quality resources against risk avoidance when the best resources occur in locations with high predation risk. Although optimality models predict the distance at which an animal should initiate vigilance and flight, many studies have shown that animals generally flee soon after detecting an approaching threat, supporting the ‘flush early and avoid the risk’ (FEAR) hypothesis. Despite this, flight behaviour varies markedly depending on context, suggesting some behavioural plasticity in the response of prey to a given threat. We evaluated the degree of plasticity in the flight responses of roe deer, Capreolus capreolus, a highly flexible species which thrives in human-dominated landscapes. Based on individually identifiable animals and a standardized flight initiation protocol, we measured the distance at which a deer detected an approaching threat, and the distance at which it subsequently initiated flight. Our results provide strong support for the FEAR hypothesis, suggesting that alert and flight responses are strongly coupled in roe deer. However, the perceived level of both risk (in terms of landscape openness and proximity to human infrastructure) and reward (in terms of habitat quality) influenced the time it took for a deer to detect an approaching threat, and the subsequent time for which the threat was tolerated prior to flight. Overall, our findings indicate that although roe deer minimize monitoring costs when assessing risk by fleeing early, they also adjust their monitoring and flight responses to the local risk eresource trade-off. © 2016 Published by Elsevier Ltd on behalf of The Association for the Study of Animal Behaviour.

Optimal foraging theory states that foraging animals should strike a balance between maximization of energetic benefits and minimization of time spent to acquire a fixed amount of energy (MacArthur & Pianka, 1966; Pyke, Pulliam, & Charnov, 1977; Schoener, 1971), while simultaneously accounting for other constraints that potentially affect fitness. Prey species, such as ungulates, must trade acquisition of high-quality resources against predation or disturbance risk because the highest quality resources are often associated with high risk (Fraser & Huntingford, 1986; Sih, 1980). In a heterogeneous landscape of fear, prey are expected to minimize exposure to risk by adjusting their antipredator

trie et * Correspondence: N. C. Bonnot, CNRS, UMR5558, Laboratoire de Biome  Lyon 1, Villeurbanne, France. Biologie Evolutive, Universite E-mail address: [email protected] (N. C. Bonnot).

behaviour to short-term changes in perceived predation risk , Herna ndez, & Altendorf, 2001; Lima & Bednekoff, 1999; (Laundre Lima & Dill, 1990). For example, vigilance is an antipredator tactic that allows individuals to exploit rich feeding patches, while concomitantly minimizing the probability of predation (Brown, 1999). When encountering a predator, decisions made by prey are crucial for immediate individual fitness (Caro, 2005). For example, survival may be conditioned by the distance at which prey initiate flight from an approaching predator, i.e. the flight initiation distance (Fig. 1). Because human-induced stimuli are often analogous to predation risk for wildlife (Frid & Dill, 2002), measuring flight behaviour by experimentally approaching animals on foot offers a simple and reliable way to measure individual tolerance to perceived predation risk (Miller, Garner, & Mench, 2006; Tarlow & Blumstein, 2007). While flight behaviour has been shown to be a

http://dx.doi.org/10.1016/j.anbehav.2016.11.031 0003-3472/© 2016 Published by Elsevier Ltd on behalf of The Association for the Study of Animal Behaviour.

36

N. C. Bonnot et al. / Animal Behaviour 124 (2017) 35e46

FID

AI DD

AD SD

Start

Vigilance

Flight

Animal’s location

Figure 1. Schematic representation of a flight initiation distance approach depicting the distances measured as a function of the behavioural response of the focal animal: flight initiation distance (FID), assessment interval (AI), alert distance (AD), detection delay (DD) and starting distance (SD).

consistent and repeatable personality trait shaping how individuals distribute themselves in human-dominated landscapes (e.g. Carrete & Tella, 2010 on burrowing owls, Athene cunicularia), many studies on a variety of taxa have shown that animals can adjust their flight behaviour to the perceived level of predation risk (see review by Stankowich & Blumstein, 2005). For example, individuals may initiate flight earlier when approached in a faster and more direct manner (e.g. Cooper, 2009 on striped plateau lizards, Sceloporus virgatus; Stankowich & Coss, 2006 on Columbian black-tailed deer, Odocoileus hemionus columbianus) or when further from refuge habitat (e.g. Dill & Houtman, 1989 on grey squirrels, Sciurus carolinensis). According to economic models of escape behaviour, flight initiation should occur when the costs of staying in terms of risk of death or injury equal the costs of fleeing in terms of loss of foraging opportunities (Cooper & Blumstein, 2014; Cooper & Frederick, 2007; Ydenberg & Dill, 1986). Both types of cost vary in space and time in response to factors that directly or indirectly affect the perceived level of predation risk, the prey's state and/or patch quality (Liley & Creel, 2008; Stankowich & Blumstein, 2005). Thus, a given threat may be perceived differently depending on environmental features (such as distance to refuge or landscape openness) and on characteristics of both prey (such as experience, age or reproductive status) and predator (such as the speed and directness of the approach). The observed flight response is thus shaped by how an individual perceives the risk of predation and how it trades foraging opportunities against risk avoidance. Blumstein (2003) and Stankowich and Coss (2006) refined this model to show, from both theoretical and empirical viewpoints, that animals assess costs and optimize their flight decision only when a predator is close enough to be detected and identified as a threat, but not close enough to provoke immediate flight. The distances at which a predator starts its approach (starting distance) and when it is first detected (alert distance; see Fig. 1) are likely to be crucial in the flight decision of prey. In particular, flight behaviour depends strongly on the starting distance in many bird species (Blumstein, 2003). To explain why this might be so, Blumstein (2010) and then Cooper and Blumstein (2014) proposed the ‘flush early and avoid the rush’ (FEAR) hypothesis, which states that animals flee soon after they have detected a threat, to minimize the costs of monitoring an approaching predator in terms of lost foraging opportunities. According to this hypothesis, the distance at which a potential threat is detected (alert distance) and the distance at which that threat is no longer tolerated (flight distance) are strongly correlated because both monitoring costs and perceived risk increase as the assessment interval increases. The assessment interval describes the period following detection during which prey assess risk and decide when to flee depending on the costebenefit balance of flight (see Fig. 1; Cooper & Blumstein, ndez-Juricic, Jimenez, & Lucas, 2002; Stankowich & 2014; Ferna Coss, 2006). An individual's behavioural response to predation threat thus depends on the degree to which it tolerates a threat

once detected, but also on its ability to detect that threat in the first place. Probably because of the difficulty of reliably distinguishing vigilance in some species (Blumstein, 2010; Cooper, 2005), variation in detection delay and assessment interval have rarely been studied in the wild. However, the ability of prey to detect and monitor threats may have a marked impact on individual fitness because (1) detecting the predator too late may limit the antipredator responses of prey and increase the risk of injury or death and (2) assessing the predator for an overly long time increases both the costs of lost foraging opportunities and predation risk, whereas fleeing immediately without assessing the risk can lead to an inappropriate antipredator response and energy loss (Cooper & Blumstein, 2014; Dugatkin, 1992; Quinn & Cresswell, 2005). In this study, we investigated how variation in perceived risk and reward influence the detection, monitoring and flight behaviour of individually identifiable free-ranging roe deer, Capreolus capreolus, living in a heterogeneous human-dominated landscape where hunting is frequent. Assuming that prey should adjust their behaviour with respect to optimal escape theory, we investigated variation in the distance covered by an approaching observer prior to detection (denoted detection delay hereafter) and the subsequent distance covered before the focal individual fled (denoted assessment interval hereafter) in relation to both the quality of the habitat patch where the animal was foraging and the associated perceived predation risk. In terms of habitat quality, cultivated fields offer rich and concentrated food resources for roe deer. Hence, their exploitation should provide greater energetic rewards than natural meadows where preferred foods are less abundant and more dispersed (Hewison et al., 2009; Morellet et al., 2011). In terms of perceived risk, based on the available literature, we supposed that deer would perceive risk to be higher when foraging in more open landscapes, when far from woodland refuge habitat and when close to human infrastructure (e.g. see Benhaiem et al., 2008; , Morellet, Hewison et al., 2015 for anaBonnot et al., 2013; Padie lyses on the same study site). Finally, although still a matter of debate (Stankowich & Blumstein, 2005; Stankowich, 2008), group size is generally expected to correlate negatively with risk perception in line with the many-eyes hypothesis (Lima, 1995; Pulliam, 1973) or because larger groups are more likely to contain at least one or more responsive individual(s) (Bonnot et al., 2015; Stankowich, 2008). We thus hypothesized that both the detection delay and the assessment interval should be shorter (1) in natural meadows, where perceived reward was assumed to be lower because of lower patch quality, than in crops, (2) in relatively open landscapes, far from woodland refuge and close to human infrastructure, when perceived risk was assumed to be higher, and (3) when group size was larger due to increased detection ability. METHODS Ethical Note In our study site, roe deer have been caught during annual winter captures since 1996 in the context of a long-term ongoing project on roe deer ecology. During this study period (from 2010 to 2015), we captured and marked 237 roe deer (of which 16% were recaptured two to four times). Most caught animals were equipped with VHF or GPS collars (N ¼ 215). All capture and marking procedures were done in accordance with French and European laws for animal welfare (prefectural order from the Toulouse Administrative Authority to capture and monitor wild roe deer and agreement no. A31113001 approved by the Departmental Authority of Population Protection). We used large-scale drives with 30e100 beaters and up to 4 km of long-nets. Roe deer were driven for a variable period lasting generally less than 10 min. Once a deer was

N. C. Bonnot et al. / Animal Behaviour 124 (2017) 35e46

captured, we removed it from the net and transferred it to a wooden retention box providing darkness and ventilation until the marking procedure (see Morellet et al., 2009 for more details). On the rare occasions that we unintentionally captured nontargeted animals (mostly wild boar, Sus scrofa), we removed them from the net and released them immediately on site. Most captured roe deer were tranquilized with an intramuscular injection of acepromazine (calmivet 3cc). This is a short-acting neuroleptic that rapidly reduces the stress response and prevents  et al., 2003), but does not adverse reactions in roe deer (Montane require an antagonist to reverse its rather short-term effects. Marking procedures lasted for approximately 10 min. and were performed by trained and experienced handlers. We recorded body mass, hind foot length, sex, age and body temperature and we collected faeces, ticks, blood (15 ml taken from the jugular vein) and skin (1 mm3 of cartilage taken with an ear punch) samples; all materials were disinfected with 70% ethanol. Needles were discarded after a single use. We equipped deer with ear tags and radiocollars (weighing up to 385 g, representing less than 2.5% of a juvenile's individual body mass; subsequent recapture 1 year later of a subsample of juveniles indicated no adverse effects of the collar on normal growth). All animals were then released on site. Besides some short-term loss of hair on the neck of a few individuals ( 0 and SD > DD > 0; Dumont, Pasquaretta, Re Bogliani, & Hardenberg, 2012). As linear models assume statistical independence among variables, SD could not be included as an explanatory variable in models of the nested distances AI or DD. Hence, we indirectly controlled for varying SD by including the relevant independent nested distance as a fixed effect in all models -Jammes & Blumstein, 2012; Dumont et al., 2012 for (see Chamaille alternative methods). That is, to standardize the analysis for variation in SD, AD was included in the models analysing variation in DD, while FID was included in the models analysing variation in AI (Fig. 1). Finally, we also included identity of the focal individual as a random factor on the intercept to control for repeated observations of individuals in all models. Although we might expect the flight responses of deer to vary in relation to the time elapsed since the end of the hunting season (February), a preliminary analysis indicated no temporal variation in behaviour within the period considered (MarcheApril); hence, we did not include observation date in the model (see Appendix 3 for more details). Therefore, the most complex model to explain DD (or AI) included the fixed effects of AD (or FID), group size, both the two-way interactions mentioned above (PC1)habitat and PC2)habitat) and the random effect of individual identity. We then compared this baseline model with all nested models (i.e. 26 different models for each response variable) using the Akaike's information criterion corrected for small sample size (AICc) and Akaike weights (u) (Burnham, Anderson, & Huyvaert, 2011). The model with the lowest AICc value and the highest weight reflects the best compromise between precision and complexity (Burnham & Anderson, 2002). All analyses were conducted in R version 3.2.2 (R Development Core Team., 2015). PCA was performed using the library ‘ade4’ (Dray & Dufour, 2007) and linear mixed models were fitted using the ‘lmer’ function in the library ‘lme4’ using the ML method (Bates, Maechler, Bolker, & Walker, 2015). We used the ‘dredge’ function  , 2015) to generate and compare in the ‘MuMIn’ library (Barton nested models. RESULTS Variation in Risk Detection: Detection Delay Detection delay (DD) varied from 5 to 309 m (mean ± SD ¼ 68 ± 55 m). The best supported model (AICc ¼ 597.3, u ¼ 0.58) was the most complex model including the two-way interactions between habitat type and PC1 (landscape openness) and between habitat type and PC2 (proximity to human infrastructure) together with the additive effect of group size, while controlling for AD (Table 1). This model indicated that DD decreased with increasing landscape openness when roe deer were feeding in meadows (Fig. 2a) such that they detected the approaching observer three times as fast when the surrounding landscape was relatively open as when it was relatively closed (DD varied from 29 to 87 m over the range of values of PC1). However, DD did not vary markedly in relation to landscape openness when roe deer were feeding in crops (from 31 to 47 m over the range of values of PC1). In addition, DD decreased with increasing proximity to human infrastructure when roe deer were in crops, but this was not the case when they were in meadows (Fig. 2b). This indicates that the time required to detect an approaching observer was more than three times shorter when deer were feeding in crops that were close to human infrastructure than those that were far from such structures (DD varied from 85 to 24 m over the range of values of PC2). Finally, we also found that the time required to detect an approaching observer was shorter for larger groups, as DD decreased slightly with increasing group size (from 56 to 37 m over the range of group sizes observed, Fig. 2c).

N. C. Bonnot et al. / Animal Behaviour 124 (2017) 35e46

39

Table 1 Summaries of the top-ranked candidate models for each distance index (AI and DD) Explanatory variables

log(DD)

Model selection criteria

Group size

Hab

PC1

PC2

Hab)PC1

Hab)PC2

AD

þ

þ þ þ

þ þ þ þ þ þ

þ þ

þ þ þ

þ þ

þ þ þ

þ

þ

þ log(AI þ 1)

þ þ

FID

K

DAICc

u

þ þ þ þ þ

10 9 8 5 6 6 4 6

0.00 1.35 3.08 0.00 0.97 1.66 1.91 2.06

0.58 0.30 0.12 0.26 0.16 0.11 0.10 0.09

Models were ranked following AICc selection criteria using the differences in the values for AICc (DAICc), the number of estimated parameters (K) and the Akaike weight (u) of each model. Of the 26 models tested, we only report the top-ranked models with a DAICc that differed by < 2 from the most supported model in the table (in bold) and the next best model with a DAICc value > 2. Explanatory variables included in each model are indicated with (þ). ‘Hab’ refers to habitat type (crop versus meadow), ‘PC1’ to the first principal component of the PCA indexing a gradient of landscape openness and ‘PC2’ to the second principal component of the PCA indexing a gradient of proximity to human infrastructure.

Variation in Tolerance of Risk: Assessment Interval The assessment interval (AI) varied from 0 to 184 m (mean ± SD ¼ 30 ± 30). The best supported model (AICc ¼ 671.7, u ¼ 0.41) included the additive effect of landscape openness (PC1), while controlling for FID (Table 1). According to this model, AI increased with decreasing landscape openness such that the interval during which deer tolerated the approaching observer prior to flight was twice as long when the surrounding landscape was relatively closed in comparison with relatively open landscapes (AI varied from 14 to 28 m over the range of values of PC1, Fig. 3). Analysis of variation in the F index provided similar results. That is, the best supported model (AICc ¼ 381.5, u ¼ 0.21) included the additive effect of landscape openness (PC1), indicating that tolerance was lower in relatively open landscapes (see Appendix 2 for further details). DISCUSSION Flight behaviour reliably indexes perceived risk in wild populations and has thus been intensively used to assess how prey cope with human-induced disturbance (Frid & Dill, 2002; Tarlow & Blumstein, 2007). Most previous studies have focused on the flight response (see reviews of Stankowich, 2008 and Stankowich & Blumstein, 2005); however, the ability of prey to detect and monitor threats may also have marked impacts on individual fitness (Cooper & Blumstein, 2014; Dugatkin, 1992; Quinn & Cresswell, 2005). Within the context of the landscape of fear , Herna ndez, & Ripple, 2010), we explored how detection, (Laundre monitoring and flight behaviour varied in a hunted roe deer population living in a human-dominated landscape in relation to drivers that are assumed to influence perceived risk. Detection Ability in the Risk-resource Trade-off The detection ability of prey is critical because detecting a predator too late markedly increases the risk of death. For example, in Thomson's gazelles, Eudorcas thomsonii, fawns with more vigilant mothers that detected predators earlier were more likely to survive a cheetah's, Acinonyx jubatus, attack (Fitzgibbon, 1990). In many prey species, individuals adjust their antipredator behaviour, for example their level of vigilance, to changes in the perceived predation risk (Brown, 1999; Lima & Bednekoff, 1999; Pays et al., 2012). Notably, roe deer are more vigilant during the hunting season than when no hunting occurs (Benhaiem et al., 2008). In this study, we showed that the time an individual took to detect an approaching threat is related to both the perceived risk of the threat

and the perceived reward in terms of habitat quality where the deer is feeding. In agreement with our prediction, the time required to detect an approaching threat was lower in risky situations (see , Morellet, Hewison Benhaiem et al., 2008; Bonnot et al., 2013; Padie et al., 2015). That is, deer detected the approach more rapidly, and hence were presumably more vigilant, when foraging in exposed open landscapes, far from refuge woodland habitat (Fig. 2a), or when close to human infrastructure (Fig. 2b), although these effects of the surrounding environment on deer detection ability were modulated by the quality of the habitat that the deer was exploiting. Indeed, while detection varied markedly in relation to landscape openness in meadows, when reward was low, there was no such relationship in crop habitat, when reward was high (see Fig. 2a). This suggests that when feeding in crops in an open and risky landscape, deer are willing to sacrifice vigilance because the moderate additional risk is more than compensated for by the energetic gains available in this rich habitat. However, the higher risk associated with proximity to human infrastructure elicited more rapid detection, and hence potentially higher vigilance levels, only when deer were foraging in rich crop habitat, but not in meadows (Fig. 2b). While this appears to contradict the economic model of escape behaviour, it could be a consequence of higher disturbance in crops within the vicinity of farms due to regular agricultural activities. Overall, the observed patterns of variation in detection ability in relation to the perceived riskereward context are consistent with previous studies on vigilance behaviour in a variety of taxa that reported higher vigilance levels when perceived predation risk was higher (Pays et al., 2012; Stankowich & Blumstein, 2005; Stankowich, 2008). We suggest that roe deer trade off the rewards associated with high-quality food patches against humanrelated risk avoidance by modifying their vigilance behaviour, supporting the hypothesis that the costs of lost foraging opportunities due to antipredator behaviour are higher when exploiting high-quality patches (first prediction). In support of this interpretation, we found that the average level of vigilance expressed by roe deer of the same population increased with both landscape openness and proximity to human infrastructure, while vigilance was also higher when deer were feeding in crops than in meadows (Appendix 4). The observed variation in detection ability in relation to landscape structure could alternatively reflect a nonrandom distribution of individuals in space which differ in their sensitivity to disturbance. Under this hypothesis, less sensitive individuals would be able to use open and exposed habitats, even when the perceived risk of predation is high, by compensating for the higher risk through increased vigilance levels. In contrast, more sensitive

40

N. C. Bonnot et al. / Animal Behaviour 124 (2017) 35e46

Decreasing landscape openness Decreasing landscape openness

Meadow Crop

Late detection

5

4

3 Early detection

2 –4

–2

0 PC1

2

4

Log(assessment interval + 1)

(a)

5 High tolerance

4

3

2 Low tolerance

1

–4

–2

0 PC1

2

4

Figure 3. Assessment interval (log-transformed) of roe deer in relation to PC1 (i.e. describing a gradient of landscape openness). The predicted estimate (dark line) is shown with 95% confidence intervals.

Increasing proximity to human infrastructures

(b)

Log(detection delay)

5

Meadow Crop

Late detection

4

3 Early detection

2 –3

–2

–1

0

1

2

PC2 (c) Late detection

5

4

individuals would prefer safer, relatively closed local landscapes, compensating for increased risk by adjusting their spatial behaviour to remain close to refuge habitat. This kind of betweenindividual variation in willingness to take risks would probably be linked to individual state (Beale & Monaghan, 2004) and/or ale, 2008) which personality traits such as boldness (Martin & Re may strongly influence how individuals perceive predation risk. For example, in previous studies on the same roe deer population, we showed that the use of open habitat varied substantially between individuals in relation to their individual capacity to manage risk (i.e. a risk management syndrome; Bonnot et al., 2015), with potential consequences for individual fitness (Monestier et al., 2015). In many taxa, individual vigilance level is negatively related to group size (Elgar, 1989; Lima, 1995). In our study, we found a weak effect of group size on the rapidity with which roe deer detected an approach, indicating that larger groups detected a threat sooner than solitary individuals (Fig. 2c). More acute flight responses of larger groups, such as longer flight distances or more rapid detection of a threat, support the many-eyes hypothesis, suggesting that large groups are better able to detect threats due to increased collective vigilance (Lima, 1995; Pulliam, 1973). This hypothesis is also supported by the lower vigilance levels observed in larger groups in the same roe deer population (Appendix 4). From his review of ungulate flight behaviour, Stankowich (2008) reported markedly contradictory findings on flight responses in relation to group size, suggesting that the effects of group size on perceived risk are context dependent and require further investigation. Tolerance and FEAR Hypothesis

3 Early detection

2 1

2

3 4 Group size

5

6

Figure 2. Detection delay (log-transformed) of roe deer in relation to (a) habitat type and PC1 (i.e. describing a gradient of landscape openness), (b) habitat type and PC2 (i.e. describing a gradient of proximity to human infrastructure) and (c) group size. Predicted estimates (grey and black lines) are shown with 95% confidence intervals.

Very few studies to date have investigated how prey adjust the time during which they tolerate and monitor the approaching ndez-Juricic et al., 2002; predator prior to flight (but see Ferna Stankowich & Coss, 2006). However, some recent studies have investigated the ratio between alert and flight initiation distances (F ¼ AD/FID) to test the hypothesis that prey species should generally flee as soon as they have detected a threat (i.e. the FEAR hypothesis, Blumstein, 2010; Samia, Nomura, & Blumstein, 2013). Our results support the FEAR hypothesis as the assessment interval was positively correlated with FID (Samia & Blumstein, 2014). To facilitate comparison between studies, we also calculated the F ratio (average F ¼ 0.84) and investigated how this tolerance index varied

N. C. Bonnot et al. / Animal Behaviour 124 (2017) 35e46

in relation to variation in perceived predation risk and reward in the environment (see Appendix 2 for more details). Overall, our results suggest that roe deer minimize monitoring costs when they detect a potential threat which could explain why we failed to find marked effects of the environmental context on the assessment interval. While this was globally the case, we also found evidence that tolerance when monitoring a threat varied in relation to landscape openness (see Fig. 3 and Appendix 2). Indeed, in agreement with our second prediction, roe deer were less tolerant towards an approaching human when perceived risk was higher in open landscapes, far from refuge habitat, suggesting that roe deer adjust their degree of tolerance to the perceived landscape of risk. Conclusion In agreement with optimal escape theory, we showed that a large herbivore thriving in human-dominated landscapes displayed some plasticity in antipredator behaviour, adjusting detection and monitoring to the local risk e reward context. However, the environmental drivers that influenced how quickly deer detected the threat and then how quickly they subsequently fled were not the same, suggesting that prey use different environmental cues for making decisions at each stage of an encounter with a threat. While the flight response of prey generally depends on the distance from which the predator starts its approach, our study provides a simple way to standardize the detection, tolerance and flight metrics for varying starting distances. To better understand the behavioural mechanisms involved in antipredator tactics during predatoreprey interactions, we suggest that future studies should quantify how detection and tolerance vary both within and between individuals in relation to the trade-off between acquisition of high-quality resources and avoidance of predation or disturbance risk. Acknowledgments  de ration We thank the local hunting associations, the Fe partementale des Chasseurs de la Haute-Garonne for allowing us De to work in the Comminges, as well as all coworkers and volunteers  Monestier for for help in collecting data. We are grateful to Chloe her assistance and to four anonymous referees for comments that helped us to improve the manuscript. L.D. and C.V. were funded by the PATCH RPDOC ANR project (ANR-12-PDOC-0017-01) awarded to C.V. from the French National Research Agency. N.C.B. was funded by the PATCH RPDOC ANR project and by the Marie-Claire Cronstedts foundation. References Abbas, F., Morellet, N., Hewison, A. J. M., Merlet, J., Cargnelutti, B., Lourtet, B., et al. (2011). Landscape fragmentation generates spatial variation of diet composition and quality in a generalist herbivore. Oecologia, 167, 401e411. Abbas, F., Picot, D., Merlet, J., Cargnelutti, B., Lourtet, B., Angibault, J.-M., et al. (2013). A typical browser, the roe deer, may consume substantial quantities of grasses in open landscapes. European Journal of Wildlife Research, 59, 69e75. Andersen, R., Duncan, P., & Linnell, J. D. (1998). The European roe deer: the biology of success. Oslo, Norway: Scandinavian University Press.  , K. (2015). MuMIn, multi-model inference. R package version 1.15.1. http:// Barton CRAN.R-project.org/package¼MuMIn. Bates, D., Maechler, M., Bolker, B., & Walker, S. (2015). Fitting linear mixed-effects models using lme4. Journal of Statistical Software, 67, 1e48. Beale, C. M., & Monaghan, P. (2004). Behavioural responses to human disturbance: A matter of choice? Animal Behaviour, 68, 1065e1069. Benhaiem, S., Delon, M., Lourtet, B., Cargnelutti, B., Aulagnier, S., Hewison, A. J. M., et al. (2008). Hunting increases vigilance levels in roe deer and modifies feeding site selection. Animal Behaviour, 76, 611e618. Blumstein, D. T. (2003). Flight-initiation distance in birds is dependent on intruder starting distance. Journal of Wildlife Management, 67, 852e857. Blumstein, D. T. (2010). Flush early and avoid the rush, a general rule of antipredator behavior? Behavioral Ecology, 21, 440e442.

41

Bonnot, N., Morellet, N., Verheyden, H., Cargnelutti, B., Lourtet, B., Klein, F., et al. (2013). Habitat use under predation risk, hunting, roads and human dwellings influence the spatial behaviour of roe deer. European Journal of Wildlife Research, 59, 185e193. Bonnot, N., Verheyden, H., Blanchard, P., Cote, J., Debeffe, L., Cargnelutti, B., et al. (2015). Interindividual variability in habitat use: Evidence for a risk management syndrome in roe deer? Behavioral Ecology, 26, 105e114. Brown, J. S. (1999). Vigilance, patch use and habitat selection: Foraging under predation risk. Evolutionary Ecology Research, 1, 49e71. Burnham, K. P., & Anderson, D. R. (2002). Model selection and multimodel inference: A practical information-theoretic approach. New York, NY: Springer Science & Business Media. Burnham, K. P., Anderson, D. R., & Huyvaert, K. P. (2011). AIC model selection and multimodel inference in behavioral ecology: Some background, observations, and comparisons. Behavioral Ecology and Sociobiology, 65, 23e35. Caro, T. (2005). Antipredator defenses in birds and mammals. Chicago, IL: University of Chicago Press. Carrete, M., & Tella, J. L. (2010). Individual consistency in flight initiation distances in burrowing owls: A new hypothesis on disturbance-induced habitat selection. Biology Letters, 6, 167e170. -Jammes, S., & Blumstein, D. T. (2012). A case for quantile regression in Chamaille behavioral ecology: Getting more out of flight initiation distance data. Behavioral Ecology and Sociobiology, 66, 985e992. Cooper, W. E., Jr. (2005). When and how do predator starting distances affect flight initiation distances? Canadian Journal of Zoology, 83, 1045e1050. Cooper, W. E., Jr. (2009). Optimal escape theory predicts escape behaviors beyond flight initiation distance: Risk assessment and escape by striped plateau lizards Sceloporus virgatus. Current Zoology, 55, 123e131. Cooper, W. E., Jr., & Blumstein, D. T. (2014). Novel effects of monitoring predators on costs of fleeing and not fleeing explain flushing early in economic escape theory. Behavioral Ecology, 25, 44e52. Cooper, W. E., Jr., & Frederick, W. G. (2007). Optimal flight initiation distance. Journal of Theoretical Biology, 244, 56e67. Dill, L. M., & Houtman, R. (1989). The influence of distance to refuge on flight initiation distance in the gray squirrel (Sciurus carolinensis). Canadian Journal of Zoology, 67, 233e235. Dray, S., & Dufour, A. B. (2007). The ade4 package, implementing the duality diagram for ecologists. Journal of Statistical Software, 22, 1e20. Dugatkin, L. A. (1992). Tendency to inspect predators predicts mortality risk in the guppy (Poecilia reticulata). Behavioral Ecology, 3, 124e127. ale, D., Bogliani, G., & Hardenberg, A. (2012). Flight Dumont, F., Pasquaretta, C., Re initiation distance and starting distance: Biological effect or mathematical artefact? Ethology, 118, 1051e1062. Elgar, M. A. (1989). Predator vigilance and group size in mammals and birds: A critical review of the empirical evidence. Biological Reviews, 64, 13e33. ndez-Juricic, E., Jimenez, M. D., & Lucas, E. (2002). Factors affecting intra-and Ferna inter-specific variations in the difference between alert distances and flight distances for birds in forested habitats. Canadian Journal of Zoology, 80, 1212e1220. Ferrari, S. L. P., & Cribari-Neto, F. (2004). Beta regression for modelling rates and proportions. Journal of Applied Statistics, 31, 799e815. Fitzgibbon, C. D. (1990). Anti-predator strategies of immature Thomson's gazelles: Hiding and the prone response. Animal Behaviour, 40, 846e855. Fournier, D. A., Skaug, H. J., Ancheta, J., Ianelli, J., Magnusson, A., Maunder, M., et al. (2012). AD model builder: Using automatic differentiation for statistical inference of highly parameterized complex nonlinear models. Optimization Methods and Software, 27, 233e249. Fraser, D. F., & Huntingford, F. (1986). Feeding and avoiding predation hazard: The behavioral response of the prey. Ethology, 73, 56e68. Frid, A., & Dill, L. (2002). Human-caused disturbance stimuli as a form of predation risk. Conservation Ecology, 6, 11. Hewison, A. J. M., Angibault, J.-M., Cargnelutti, B., Coulon, A., Rames, J.-L., Serrano, E., et al. (2007). Using radio-tracking and direct observation to estimate roe deer Capreolus capreolus density in a fragmented landscape, a pilot study. Wildlife Biology, 13, 313e320. Hewison, A. J. M., Morellet, N., Verheyden, H., Daufresne, T., Angibault, J.-M., Cargnelutti, B., et al. (2009). Landscape fragmentation influences winter body mass of roe deer. Ecography, 32, 1062e1070. Hewison, A. J. M., Vincent, J.-P., Joachim, J., Angibault, J.-M., Cargnelutti, B., & Cibien, C. (2001). The effects of woodland fragmentation and human activity on roe deer distribution in agricultural landscapes. Canadian Journal of Zoology, 79, 679e689. Jeppesen, J. L. (1989). Activity patterns of free-ranging roe deer (Capreolus capreolus) at Kalø. Danish Review of Game Biology, 13, 1e32. , J. W., Herna ndez, L., & Altendorf, K. B. (2001). Wolves, elk, and bison: Laundre Reestablishing the ‘landscape of fear’ in Yellowstone National Park, U.S.A. Canadian Journal of Zoology, 79, 1401e1409. , J. W., Herna ndez, L., & Ripple, W. J. (2010). The landscape of fear: ImpliLaundre cations of being afraid. Open Ecology Journal, 3, 1e7. Liley, S., & Creel, S. (2008). What best explains vigilance in elk: Characteristics of prey, predators, or the environment? Behavioral Ecology, 19, 245e254. Lima, S. L. (1995). Back to the basics of anti-predatory vigilance: The group-size effect. Animal Behaviour, 49, 11e20. Lima, S. L., & Bednekoff, P. A. (1999). Temporal variation in danger drives antipredator behavior: The predation risk allocation hypothesis. American Naturalist, 153, 649e659.

42

N. C. Bonnot et al. / Animal Behaviour 124 (2017) 35e46

Lima, S. L., & Dill, L. M. (1990). Behavioral decisions made under the risk of predation: A review and prospectus. Canadian Journal of Zoology, 68, 619e640. MacArthur, R. H., & Pianka, E. R. (1966). On optimal use of a patchy environment. American Naturalist, 100, 603e609. ale, D. (2008). Animal temperament and human disturbance: Martin, J. G. A., & Re Implications for the response of wildlife to tourism. Behavioural Processes, 77, 66e72. Miller, K. A., Garner, J. P., & Mench, J. A. (2006). Is fearfulness a trait that can be measured with behavioural tests? A validation of four fear tests for Japanese quail. Animal Behaviour, 71, 1323e1334. , C., & Monestier, C., Morellet, N., Gaillard, J.-M., Cargnelutti, B., Vanpe Hewison, A. J. M. (2015). Is a proactive mum a good mum? A mother's coping style influences early fawn survival in roe deer. Behavioral Ecology, 26, 1395e1403. http://dx.doi.org/10.1093/beheco/arv087. , J., Marco, I., Lo pez-Olvera, J., Perpin ~a n, D., Manteca, X., & Lavín, S. (2003). Montane Effects of acepromazine on capture stress in roe deer (Capreolus capreolus). Journal of Wildlife Diseases, 39, 375e386. Morellet, N., Van Moorter, B., Cargnelutti, B., Angibault, J.-M., Lourtet, B., Merlet, J., et al. (2011). Landscape composition influences roe deer habitat selection at both home range and landscape scales. Landscape Ecology, 26, 999e1010. Morellet, N., Verheyden, H., Angibault, J.-M., Cargnelutti, B., Lourtet, B., & Hewison, A. J. M. (2009). The effect of capture on ranging behaviour and activity of the European roe deer Capreolus capreolus. Wildlife Biology, 15, 278e287. , S., Morellet, N., Cargnelutti, B., Hewison, A. J. M., Martin, J.-L., & Chamaille Padie Jammes, S. (2015). Time to leave? Immediate response of roe deer to experimental disturbances using playbacks. European Journal of Wildlife Research, 61, 871e879. , S., Morellet, N., Hewison, A. J. M., Martin, J.-L., Bonnot, N., Cargnelutti, B., et al. Padie (2015). Roe deer at risk, teasing apart habitat selection and landscape constraints in risk exposure at multiple scales. Oikos, 124, 1536e1546. Pagon, N., Grignolio, S., Pipia, A., Bongi, P., Bertolucci, C., & Apollonio, M. (2013). Seasonal variation of activity patterns in roe deer in a temperate forested area. Chronobiology International, 30, 772e785. Pays, O., Benhamou, S., Helder, R., & Gerard, J.-F. (2007). The dynamics of group formation in large mammalian herbivores: An analysis in the European roe deer. Animal Behaviour, 74, 1429e1441. -Jammes, S., Duncan, P., Pe riquet, S., Pays, O., Blanchard, P., Valeix, M., Chamaille et al. (2012). Detecting predators and locating competitors while foraging: An experimental study of a medium-sized herbivore in an African savanna. Oecologia, 169, 419e430. Pulliam, H. R. (1973). On the advantages of flocking. Journal of Theoretical Biology, 38, 419e422. Pyke, G. H., Pulliam, H. R., & Charnov, E. L. (1977). Optimal foraging, a selective review of theory and tests. Quarterly Review of Biology, 52, 137e154. Quinn, J. L., & Cresswell, W. (2005). Escape response delays in wintering redshank, Tringa totanus, flocks, perceptual limits and economic decisions. Animal Behaviour, 69, 1285e1292. R Development Core Team. (2015). R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing. https:// www.R-project.org/. Reimoser, S. (2012). Influence of anthropogenic disturbances on activity, behavior and heart rate of roe deer (Capreolus capreolus) and red deer (Cervus elaphus), in context of their daily and yearly patterns. In A. A. Cahler, & J. P. Marsten (Eds.), Deer, habitat, behavior and conservation (pp. 1e96). Hauppauge, NY: Nova Science Publishers. Samia, D. S., & Blumstein, D. T. (2014). Phi index: A new metric to test the flush early and avoid the rush hypothesis. PloS One, 9, e113134. Samia, D. S., Nomura, F., & Blumstein, D. T. (2013). Do animals generally flush early and avoid the rush? A meta-analysis. Biology letters, 9, 20130016. Schoener, T. W. (1971). Theory of feeding strategies. Annual Review of Ecology and Systematics, 2, 369e403. Sih, A. (1980). Optimal behavior: Can foragers balance two conflicting demands? Science, 210, 1041e1043. Skaug, H., Fournier, D., Nielsen, A., Magnusson, A., & Bolker, B. (2012). Generalized linear mixed models using AD Model Builder. R package version 0.8.0. http://CRAN. R-project.org/package¼glmmADMB. Smithson, M., & Verkuilen, J. (2006). A better lemon squeezer? Maximumlikelihood regression with beta-distributed dependent variables. Psychological Methods, 11, 54. Stankowich, T. (2008). Ungulate flight responses to human disturbance: A review and meta-analysis. Biological Conservation, 141, 2159e2173. Stankowich, T., & Blumstein, D. T. (2005). Fear in animals: A meta-analysis and review of risk assessment. Proceedings of the Royal Society: Biological Sciences, 272, 2627e2634. Stankowich, T., & Coss, R. G. (2006). Effects of predator behavior and proximity on risk assessment by Columbian black-tailed deer. Behavioral Ecology, 17, 246e254. Tarlow, E., & Blumstein, D. T. (2007). Evaluating methods to quantify anthropogenic stressors on animals. Applied Animal Behaviour Science, 102, 429e451. Ydenberg, R. C., & Dill, L. M. (1986). The economics of fleeing from predators. Advances in the study of Behavior, 16, 229e249.

Appendix 1. Results of the principal component analysis

Table A1 Loadings and proportion of the variance explained by the first two components (PC1 and PC2) from the principal component analysis

Proportion of open habitats in surroundings Distance to the nearest wood Distance to the nearest hedgerow Distance to the nearest road Distance to the nearest human infrastructure Total variance explained (%)

PC1

PC2

0.89 0.75 0.65 0.51 0.53 46.4

0.10 0.34 0.47 0.62 0.61 22.1

Appendix 2. Analysis of F index for comparison with AI To test the FEAR hypothesis of a strong positive relationship between alert and flight distances (AD and FID), Samia and Blumstein (2014) proposed the F index, which is the ratio between these two distances, such that, for each observation i:

Pn

F¼1

i¼0

ðADi FIDi Þ ADi

n



(A1)

with n being the sample size. As AD  FID > 0, F ranges within the interval ]0,1]. A high F value supports the hypothesis of a strong association between FID and AD as expected under the FEAR hypothesis, whereas a low F value indicates marked departure from the expected 1:1 relationship. Whereas F represents the mean proportional difference between AD and FID over all observations, we can also calculate a F value per observation, such that:

Fi ¼ 1 

AIi FIDi ¼ ADi ADi

(A2)

High values of Fi indicate a marked difference between AD and FID (i.e. long AI distance), whereas low Fi values indicate a small difference between AD and FID (i.e. short AI). Thus, Fi indexes individual tolerance such that high Fi values should be obtained for less tolerant animals. To test the hypothesis that the tolerance of roe deer is related to spatiotemporal variation in the perceived risk of predation, we fitted beta regression models that are useful for modelling ratios (Ferrari & Cribari-Neto, 2004). To this aim, we transformed Fi to exclude the extreme value 1 (with a classic transformation, see equation (A3); Smithson & Verkuilen, 2006), as beta regression models assume that the dependent variable takes values within the interval ]0,1[ only (Cribari-Neto & Zeileis, 2009).

F; i ¼

Fi $ðn  1Þ þ 0:5 n

(A3)

The environmental factors expected to affect detection and tolerance included the type of habitat patch that the animal was exploiting (crop versus meadow), group size, landscape openness (PC1) and distance to human infrastructures (PC2) (see Fig. A1). We tested the main effect of each factor as well as the two-way interactions between PC1 and habitat and PC2 and habitat. Model selection was based on the AICc value and Akaike weights. The model with the lowest AICc value and the highest Akaike weight reflects the best compromise between precision and complexity (Burnham & Anderson, 2002). All analyses were conducted in R version 3.2.2 (R Development Core Team., 2015). Beta regression models were fitted using the ‘glmmadmb’ function in the library ‘glmmADMB’ (Fournier et al., 2012; Skaug et al., 2014). We used the ‘dredge’ function in the  , 2015) to generate and compare models. ‘MuMIn’ library (Barton

N. C. Bonnot et al. / Animal Behaviour 124 (2017) 35e46

Eigenvalues

43

d.hedgerow

d=2

Popen

d.wood

d.hum.infr

Figure A1. Scatter plot of the principal component analysis (PCA) to summarize landscape descriptors based on the first two principal components (PC1 and PC2). PCA was performed on normalized and mean-centred local environmental variables represented on the graph by Popen (proportion of open habitats in the surroundings), d.hedgerow (distance to the nearest hedgerow), d.wood (distance to the nearest wood), d.hum.infr (distance to the nearest human infrastructure) and d.road (distance to the nearest road; on the graph this coincides with d.hum.infr). Each dot represents an observation. The inset shows the proportion of the total variance explained by each component.

F0 varied from 0.38 to 0.99 (mean ± SD ¼ 0.83 ± 0.13). Explanatory variables included in the top-ranked models were consistent with those retained in models explaining variation in the assessment interval, i.e. the first component of the PCA analysis (PC1) indexing landscape openness (see Tables 1 and A2). Indeed, the model with the most support based on AICc criteria included the additive effect of PC1 (landscape

openness) only. According to this model, F0 decreased with landscape openness (Fig. A2), suggesting that roe deer had higher tolerance levels in relatively closed and safer landscapes than in more open and riskier landscapes (F0 varied from 0.87 to 0.79 from the most open to the most closed landscapes).

Table A2 Summaries of the top-ranked candidate models explaining variation in F0 index Explanatory variables Group size

Model selection criteria Habitat

PC1

PC2

þ þ þ þ

þ þ þ

þ

K

DAICc

u

4 3 5 5 5 4

0.00 1.13 1.19 1.91 1.96 2.05

0.21 0.12 0.11 0.08 0.08 0.07

Models explaining F0 index were ranked following AICc selection criteria using the differences in the values for AICc (DAICc), the number of estimated parameters (K) and the Akaike weight (u) for each model. We only report the top-ranked models with a DAICc that differed by < 2 from the best supported model in the table (in bold) and the next best model with a DAICc value > 2. Explanatory variables included in each model are indicated with (þ).

44

N. C. Bonnot et al. / Animal Behaviour 124 (2017) 35e46

Decreasing landscape openness

(a)

1 Low tolerance

0.9

s(Julian date, 1)

Φ’

0.8

1

0.7 0.6

–1

High tolerance

0.5

0

0.4 –4

–2

0

2

–2

4

PC1

| | | | ||||||| |||||| |||||||||||||||||||||||||||||||||||||||| | | 60 70 80 90 100 110 120 50

Figure A2. Variation in F0 index in relation to PC1 describing a gradient of landscape openness. Predicted estimates from the best model (black line) are shown with 95% confidence intervals (grey shadow).

2

(b)

Appendix 3. Analysis of temporal variation in DD and AI

Table A3 Summaries of the models with and without the Julian date explaining both DD and AI Response variables

Explanatory variables

K DAICc u

log(DD)

Group sizeþHab)PC1þHab)PC2þAD Group sizeþHab)PC1þHab) PC2þADþs(Julian date) PC1þFID PC1þFIDþs(Julian date)

10 0.00 12 4.34

0.90 0.10

5 7

0.88 0.12

log(AIþ1)

0.00 4.01

Comparison of the two models with and without the Julian date to test the effect of time on DD and AI, respectively. Models were ranked following AICc selection criteria using the differences in the values for AICc (DAICc), the number of estimated parameters (K) and the Akaike weight (u) for each model. ‘Hab’ refers to habitat type (crop versus meadow), ‘PC1’ to the first principal component of the PCA indexing a gradient of landscape openness and ‘PC2’ to the second principal component of the PCA indexing a gradient of proximity to human infrastructure.

s(Julian date, 1)

1 Although most observations were performed in March and April, after closure of the official hunting season at the end of February, we might expect that the behavioural responses of roe deer in terms of detection ability and tolerance of threat would vary in relation to the time elapsed since the end of hunting. To test this assumption, we analysed the effect of Julian date on logtransformed DD and AI. That is, we built two generalized additive mixed models (GAMMs) which comprised all factors previously retained for each response variable (see Table A3 and the main text for more details) with and without the Julian date included as a smoothing term. We then compared these models based on the AICc value and Akaike weights (w) (Burnham & Anderson, 2002). For both DD and AI, the best supported model did not include the effect of Julian date (see Table A3, Fig. A3a, b), suggesting that there is no temporal variation in either detection ability or assessment interval in relation to the time elapsed since the official end of hunting.

0

–1

–2

| | | | ||||||| |||| || |||||||||||||||||||||||||||||||||||||||| | | 50 60 70 80 90 100 110 120 Julian date Figure A3. Variation in log-transformed (a) detection delay (DD) and (b) assessment interval (AI) in relation to time (Julian date). Predicted estimates from the best model (black line) are shown with 95% confidence intervals (grey shadow).

Appendix 4. Vigilance levels and perceived risk and reward To better understand the relationship between detection ability and vigilance levels in roe deer, we analysed how vigilance level varied in relation to the perceived risk and reward in the deer's immediate surroundings. From 2010 to 2015, over 8 weeks in spring (MarcheApril), we observed the behaviour of individually identified roe deer outside the hunting season. We recorded vigilance levels (i.e. the proportion of time the animal spent with its head above shoulder level while scanning the surroundings) of undisturbed animals foraging in open habitats. Overall, we collected 397 vigilance observations on 104 different individually identifiable animals. For each observation, we recorded group size, the type of habitat in which the deer was feeding (crop versus meadow), and we described the local environment. Therefore, for each observation, we measured

N. C. Bonnot et al. / Animal Behaviour 124 (2017) 35e46

human infrastructure (indexed by PC2). We tested the main effect of each of these factors as well as the two-way interactions between PC1 and habitat and PC2 and habitat. We also included identity of the focal individual as a random factor on the intercept to control for repeated observations of individuals in all models. We compared this global model to all nested models, i.e. 26 different models. Model selection was based on the AICc value and Akaike weights (u). The model with the lowest AICc value and the highest Akaike weight reflects the best compromise between precision and complexity (Burnham & Anderson, 2002). All analyses were conducted in R version 3.2.2 (R Development Core Team., 2015). PCA was performed using the library ‘ade4’ (Dray & Dufour, 2007) and linear mixed models were fitted using the ‘lmer’ function in the library ‘lme4’ using the ML method (Bates et al., 2015). We used the ‘dredge’ func , 2015) to generate and tion in the ‘MuMIn’ library (Barton compare models. On average, roe deer spent 24 ± 15% of their time in vigilance during a feeding phase (range 0e80%). The best supported model (AICc ¼ 320.3, K ¼ 7, u ¼ 0.33, DAICc ¼ 0.34 to the second most supported model among the four models with a DAICc < 2) included the fixed effects of landscape openness (PC1), proximity to human infrastructure (PC2), group size and habitat type. According to this model, roe deer vigilance levels increased slightly with landscape openness (Fig. A4a), proximity to human

landscape openness around the individual's location (the proportion of open habitat within a radius of 150 m), and the distance from its location to the nearest road, to the nearest human habitation, to the nearest woodland, and to the nearest hedgerow. As for our analysis of DD and AI, we summarized these landscape descriptors by a smaller number of synthetic variables using a PCA (see main text, Table A1 and Fig. A1 for more details). We retained the two first principal components which accounted for 67% of the total variance (respectively, 41% and 26%; see Table A4). Table A4 Loadings and the proportion of variance explained by the first two components (PC1 and PC2) from the principal component analysis performed on the environmental variables

Proportion of open habitats in surroundings Distance to the nearest wood Distance to the nearest hedgerow Distance to the nearest road Distance to the nearest human dwelling Total variance explained (%)

PC1

PC2

0.84 0.76 0.72 0.37 0.34 41.2

0.26 0.35 0.06 0.71 0.77 25.7

We used a square-root transformation on vigilance levels to achieve normality. Then, we fitted linear mixed models to assess whether the level of vigilance varied in relation to habitat type, group size, landscape openness (indexed by PC1) and distance to Decreasing landscape openness

Increasing proximity to human infrastructures

(a)

(b)

0.8

0.8

0.6

0.6

0.4

0.4

0.2 √(Vigilance level)

45

0.2

–4

–2

0 PC1

2

4

–3

0.8 (c)

–2

–1

0 PC2

1

2

3

(d) 0.8

0.6 0.6 0.4 0.4 0.2

0.2

1

2

3 4 Group size

5

6

Crop

Meadow Habitat types

Figure A4. Vigilance level (square root-transformed) of roe deer in relation to (a) PC1 describing a gradient of landscape openness, (b) PC2 describing a gradient of proximity to human infrastructure, (c) group size and (d) habitat type (crop versus meadow). Predicted estimates (blue line) are shown with 95% confidence intervals (grey shadow).

46

N. C. Bonnot et al. / Animal Behaviour 124 (2017) 35e46

infrastructure (Fig. A4b) and group size (Fig. A4c). Overall, these results support our interpretation of the observed patterns of variation in detection ability (see main text), suggesting that roe deer had higher vigilance levels when their perception of risk in their environment was greater. The group size effect also supports

the many-eyes hypothesis. However, we also found that the level of vigilance was higher when roe deer were feeding in crops than in meadows (Fig. A4d) which appears to contradict the economic model of escape behaviour.