Trophic interactions and dynamic herbivore responses to snowpack

Herbivore responses to changing abiotic conditions can have strong impacts on community structure. Changes in herbivore abundance or distribution may influence predator dynamics [1] and persistence [2]. Altered intensity or spatial distribution of herbivory can also affect plant abundance and diversity [3]. A more general understanding of how herbivore distribution, abundance, and foraging are affected by shifting abiotic conditions will greatly improve our knowledge of climate change impacts on natural ecosystems.

While we know generally that climate can strongly affect herbivore responses, we lack knowledge about which factors mediate herbivore distribution in relation to environmental conditions. Species interactions change with abiotic conditions [4], and not only directly alter herbivore abundance [5] and distribution [6], but also to change the way in which herbivores respond to environmental conditions [7]. Indeed, novel outcomes of species interactions are among the most important factors to understand in attempting to predict community responses to climate change [8],[9].

Trophic interactions can strongly influence herbivore distribution and behavior and mediate herbivore impacts on plants. For instance, plant performance can be enhanced in areas that herbivores avoid due to high predation risk [10]. By affecting the spatial distribution of herbivores [11], predation risk also has the potential to shape how plant-herbivore interactions respond to changing abiotic conditions. For example, climate-induced habitat deterioration in habitat patches with high predation risk could potentially have no impact on herbivore distribution if the herbivores already avoided such locales.

Interactions between trophic relationships and abiotic conditions on herbivore distribution have seldom been demonstrated, however. One reason for this is that numerous factors influence herbivore distribution, so assessing interactions between trophic dynamics and climate requires untangling multiple concurrent impacts. Here we assess whether top-down and bottom-up interactions shape responses of a dominant ungulate herbivore along a gradient of abiotic conditions. We focus on elk because of their strong roles in affecting the demography and persistence of woody shrubs [12] that provide habitat for other species [13]. Previously, we demonstrated that the impact of elk herbivory on aspen (Populus tremuloides) in the Yellowstone Ecosystem is strongly affected by snow conditions [14]. Elk browse on aspen during the winter, yet avoid areas with deep or dense snow [14]. But determining whether elk responses to snowpack are mediated by trophic interactions requires a detailed analysis incorporating risk of wolf predation, changing elk body condition, and elk population structure. Wolf predation risk has been suggested to alter elk movements [15] and behavior [16] in ways that could potentially interact with the influence of snow conditions. Like other ungulates [17],[18], elk may face extended periods of food shortage during winter [19],[20], but their tolerance of deep snow may decrease over the winter as the enhanced energetic costs of movement become less affordable. Finally, there are likely important sex- and life stage-based differences in elk behavior [21]. Mature males are generally in worse body condition than females or juveniles at the start of winter from having expended bodily energetic reserves during the rut [cf. [22]; hence they may respond differently to predation risk, deep snow, or interactions among these factors.

We surveyed elk over two winters in Yellowstone along gradients in snow conditions and predation risk, assessing the effects of spatial as well as temporal variance in both snow conditions and predation risk. We hypothesized that the level of elk avoidance of deep snow areas will increase in response to several factors. First (prediction i), avoidance of snow would increase as winter progressed, because deteriorating body condition would increase energetic costs of movement. Second (prediction ii), avoidance would be higher for mature males, who started the winter in poorer body condition than females and juveniles because of the autumn rut. Third (prediction iii) avoidance would increase when wolves were in the immediate vicinity [16], because the higher energetic costs of predator-avoidance movements [23] and extra vigilance [cf. [24] would be unaffordable in combination with deep snow; this prediction is generated by the “landscape of fear” hypothesis, but we note that while elk are avoiding wolves, wolves are seeking out elk –these countervailing forces can potentially balance each other out. Finally (prediction iv) avoidance should increase in areas particularly frequented by wolves, since even temporary behavioral avoidance of wolves by elk [23] could potentially translate into permanent spatial differences in predation risk [25]. Predictions (i) and (ii) relate to bottom-up influences and population structure, respectively. Predictions (iii) and (iv) relate to temporal and spatial variation in predation risk, respectively.

Our survey covered 2,011 site-days in 2007–2008 and 2,839 site-days in 2008–2009. We recorded elk on 549 site-days across both sampling periods; mature males on 210 days, and females and juveniles on 376 days. Across sites, the average proportion of days on which elk were observed was 0.12 (SD = 0.10, range = 0.02 : 0.43). We obtained 212 wolf photographs on 92 site-days. Elk were detected on 18.5% of the site-days on which wolves were also detected. Elk were detected on “wolf” days more often than expected by chance; for days when wolves were detected, the proportion of site-days when elk were also detected (0.19) was significantly higher than the proportion of total site-days on which elk were detected (0.12; two-sample proportion test, Z = −2.15, p = 0.03).

We use snow water equivalent, the depth of the equivalent amount of liquid water contained in the snowpack, as a measure of snow conditions that integrates the depth and density of the pack. The maximum snow water equivalent averaged 15.0 cm (range = 1.68 : 28.33) across sites in 2007–2008 and 11.83 cm (range = 1.07 : 23.10) in 2008–2009. Pairwise correlations between predictor variables were R = 0.33 for winterday and swe, |R| < 0.01 for all other variable pairs.

Cross correlation analysis revealed no significant time lags in the effects of swe on elk presence, or of wolves on the presence of female and juvenile elk. Mature male elk presence was correlated with wolf relative abundance six - eight days previously (Figure 1d). We used six-day lag effects for wolves on mature male elk in the temporally-dynamic models below (but also present zero-lag model results in the SM).

Of the 253 models for temporal effects of snow, predation risk, and body condition, four received model support (i.e., Akaike weight >0.00; Tables 1 and 2). Using model averaging, six variables (including interaction terms) had unambiguous effects on elk occurrence (95% confidence intervals (CIs) did not include zero): winterday, sex, winterday × swe, winterday × sex, swe × sex), and winterday × swe × sex (Figure 2; Table 3). The model-averaged wolf coefficient was statistically negligible (β < 0.01; CI = −0.01: 0.02). Occurrence by females and juveniles was substantially lower at high-SWE sites throughout the winter, while male occurrence was initially somewhat neutral, becoming slightly SWE-averse as winter progressed. The difference in response to SWE among sex and age classes was reduced over the course of the winter (Figure 2). The slope of the relationship with snow water equivalent stayed approximately constant for females and juveniles over the course of the winter (−0.09 in January, −0.10 in April), while the intercept declined slightly more (0.14 in January to 0.12 in April). For mature males, the intercept stayed constant (0.04); the slope declined but remained close to zero throughout the winter (0.02 in January and −0.03 in April).

Number of parameters	∆AIC	AIC wt	Coefficients (SE)
			Winterday	swe	Wolf	Sex	day×swe	day×sex	swe×sex	day×swe×sex
8	0.00	0.62	−0.91 (0.11)	−0.13 (0.15)	np	0.36 (0.12)	−0.50 (0.12)	np	−0.70 (0.14)	−0.42 (0.09)
9	1.72	0.26	−0.91 (0.11)	−0.13 (0.15)	0.02 (0.03)	0.36 (0.12)	−0.49 (0.12)	np	−0.70 (0.14)	−0.42 (0.09)
8	4.19	0.08	−1.09 (0.13)	−0.13 (0.15)	np	0.42 (0.13)	−0.70 (0.11)	0.38 (0.09)	−0.65 (0.14)	np
9	5.91	0.03	−1.09 (0.13)	−0.13 (0.15)	0.02 (0.03)	0.42 (0.13)	−0.70 (0.11)	0.38 (0.09)	−0.65 (0.14)	np

Interaction terms are only shown if present in at least one of these models. Sex is coded as 1 for females and juveniles and 0 for mature males; np = not present in the model.

Number of parameters	∆AIC	AIC wt	Coefficients (SE)
			swe	Wolf	Sex	swe×wolf	swe×sex	wolf×sex	swe×wolf×sex
5	0.00	0.52	−0.02 (2.16)	np	11.30 (3.00)	np	−9.36 (3.04)	np	np
6	2.56	0.15	−0.01 (2.17)	−0.53 (1.54)	11.30 (3.04)	np	−9.36 (3.07)	np	np
7	2.99	0.12	0.01 (2.14)	−0.99 (1.55)	11.21 (2.99)	np	−9.62 (3.03)	np	3.19 (2.18)
7	3.81	0.08	−0.15 (2.16)	−1.07 (1.59)	11.30 (3.02)	1.88 (1.59)	−9.36 (3.06)	np	np
7	5.30	0.04	−0.02 (2.20)	−0.08 (2.20)	11.30 (3.07)	np	−9.33 (3.11)	−0.91 (3.11)	np
8	5.55	0.03	−0.02 (2.16)	−0,08 (2.16)	11.20 (3.02)	np	−9.59 (3.06)	−1.90 (3.12)	3.47 (2.25)
8	5.92	0.03	−0.04 (2.17)	−1.07 (1.60)	11.22 (3.03)	0.55 (2.21)	−9.58 (3.08)	np	2.67 (3.06)
4	6.82	0.02	−4.70 (1.70)	np	11.30 (3.30)	np	np	np	np
5	9.26	0.01	−4.68 (1.69)	−0.53 (1.69)	11.30 (3.34)	np	np	np	np

Sex is coded as 1 for females and juveniles and 0 for mature males; np = not present in the model.

We assessed potential bias in our metric of predation risk by examining whether wolf conditional detectability (i.e., the probability of detection on days when the species is known to occur at a given site) varied with swe or winterday. The probabilities that models of wolf detectability containing swe or winterday were the most parsimonious were <0.01 and 0.01, respectively.

Of the 24 models that we analyzed for spatial (i.e., temporally invariant) effects of snow and predation risk, 10 received model support, with one model receiving 52% of the support (Tables 2 and 4). Using model-averaging, only sex and the swe × sex interaction significantly related to elk site usage (i.e., CIs did not include zero; Figure 3). Elk site usage declined with increasing snowpack for females and juveniles but not mature males; site usage by neither females nor males responded to variation in wolf site usage.

	Temporal model (6-day lag for mature male elk)			Temporal model (no time lags)			Spatial model
Parameter	Mean	SE	95% CI	Mean	SE	95% CI	Mean	SE	95% CI
Winterday	−0.93	0.12	−1.16 : −0.71	−0.92	0.11	−1.15 : −0.70	np
SWE	−0.13	0.12	−0.36 : 0.10	−0.13	0.11	−0.35 : 0.09	−0.16	2.15	−4.37 : 4.04
Wolf	0.00	0.01	−0.01 : 0.02	0.01	0.01	−0.01 : 0.02	−0.32	0.75	−1.80 : 1.15
Sex	0.33	0.13	0.08 : 0.58	0.36	0.12	0.13 : 0.60	11.28	3.02	5.36 : 17.20
Winterday × SWE	−0.49	0.11	−0.71 : −0.28	−0.51	0.11	−0.74 : −0.29	np
Winterday × wolf	0.00	0.00	0.00 : 0.00	0.00	0.00	0.00 : 0.00	np
Winterday × sex	0.07	0.02	0.03 : 0.10	0.04	0.01	0.02 : 0.06	np
SWE × wolf	0.00	0.00	0.00 : 0.00	0.00	0.00	0.00 : 0.00	0.17	0.19	−0.21 : 0.54
SWE × sex	−0.73	0.14	−1.01 : −0.45	−0.69	0.14	−0.96 : −0.41	−9.13	2.96	−14.92 : −3.33
Wolf × sex	0.00	0.00	0.00 : 0.00	0.00	0.00	0.00 : 0.00	−0.09	0.22	−0.52 : 0.33
Winterday × SWE × wolf	0.00	0.00	0.00 : 0.00	0.00	0.00	0.00 : 0.00	np
Winterday × SWE × sex	−0.33	0.08	−0.47 : −0.18	−0.37	0.08	−0.54 : −0.21	np
Winterday × wolf × sex	0.00	0.00	0.00 : 0.00	0.00	0.00	0.00 : 0.00	np
SWE × wolf × sex	0.00	0.00	0.00 : 0.00	0.00	0.00	0.00 : 0.00	0.56	0.42	−0.26 : 1.38

Temporal changes in trophic interactions, population structure, and herbivore responses to abiotic conditions help determine community responses to climate change. We presented four predictions of how responses to snowpack conditions by elk, the dominant ungulate herbivore in the Yellowstone Ecosystem, would be mediated by predation hazard, time (a proxy for body condition), and sex. Our first prediction, that elk would increasingly avoid snow as the winter progressed, which is associated with declining body condition, was supported. The probability of male elk visiting a given site declined as snow accumulated over the course of the winter, particularly at sites with deep snowpack. This sex-based variance in responses to snowpack suggests that sex structure could importantly affect the mean response of a population to changing abiotic conditions.

Our second prediction, that avoidance of deep snow would be particularly strong in mature males due to their poor nutritional state following the autumn rut, was not supported. The probability of occurrence by males did respond to variation in snowpack, particularly from mid-March onwards, but the response of females and juveniles was consistently stronger. Females and juveniles are much more abundant in Yellowstone than mature males [26], leading to the higher model intercept for females and juveniles than males. But their occurrence drops much more rapidly in response to increasing snowpack depth, suggesting either that they die off rapidly or else move to lower elevations, such as outside the national park, as snow accumulation increases. It is unlikely that females are dying off faster; elk survival studies show higher survival for females than males [27],[28], though juvenile survival rates are lower than those of adults [29],[30]. Because over-winter mortality of 1st-year juveniles is higher than that of adult females [30], the individual abundance of adult females + juveniles is likely to decline over the course of many winters. Our study, however, assessed occurrence rates of groups rather than abundance of individuals. Because 1st-year juveniles are not found away from adults during the winter, over-winter juvenile mortality probably does not explain the decline in detection rates of adult female + juvenile groups that we observed. Prior studies have shown that elk movement to lower elevations as the snow increases at high elevations [31], but we suggest that this movement is stronger for females than males. This could be due to higher tolerance of deep snow by males, perhaps because their larger body size facilitates digging for sub-nivean forage, they lack the energetic reserves, or they have a higher probability of avoiding encounters with wolves. Males may also opt to remain at higher elevations inside the park rather than leave the park and be exposed to trophy hunting by humans, though we note that adult female harvest rates adjacent to Yellowstone National Park have also been high [32]. It is also possible that elk daily mobility decreases in deep snow, which would reduce detection rate. We did not, however, detect an increase in the time spent at any given site in response to snow, as would be predicted if mobility were reduced.

Our last two predictions, that the level of avoidance of deep snow by elk would increase on days when wolves were present (temporal effects; prediction iii) or in sites where average wolf usage was higher (spatial effects; prediction iv), were not strongly supported. Though a wolf term was included in two of the top four temporally-dynamic models, local wolf relative abundance had at best slight effects on the probability of elk occurrence at a given site on a given day. The model-averaged wolf coefficient was positive (in contrast to our predictions), very small (0.01), and its confidence intervals included zero. It is possible that wolves and elk avoid deep snowpack equally. We had a reasonably large wolf sample size with which to have detected a wolf effect had it existed, and our use of temporally-lagged models for mature male elk should have further increased our ability to detect wolf effects. Though our sample size still might not have been big enough to statistically detect a wolf effect, we maintain that this in itself is biologically meaningful -wolf abundance in Yellowstone peaked at or just before the time of this study [33]. If we were unable to detect a wolf effect then, we probably would not have detected one before or since. We note that our metric of predation risk almost certainly underestimates true wolf usage of each site because camera traps are imperfect at detecting any species. However, we found that wolf conditional detection probability (using temporal sub-sampling) did not vary with snow water equivalent or time, suggesting that relative site usage is an unbiased index of true site usage by wolves.

Effects of predation risk on herbivore behavior are difficult to assess non-experimentally. On the one hand, the “landscape of fear” hypothesis predicts that herbivores avoid sites where predators are most active, theoretically leading to a negative relationship between predator and prey abundance. Nevertheless, predators track their prey and are likely to be more concentrated where prey base is highest [34], theoretically generating positive predator–prey relationships. These opposing forces potentially cancel each other out [35], possibly explaining the flat relationship between wolf and elk site usage that we observed here. We noted a potential signal of wolves tracking elk; the two-sample proportion test suggested that elk were detected on “wolf” days more often than expected by chance. However, this positive relationship disappeared in the logistic models that include other fixed effects as well as a random effect. Thus wolves and elk do not appear to conform to the simple dynamics proposed by the “landscape of fear” hypothesis, whereby prey avoidance of predators generates spatially- and temporally- consistent negative relationships between predation risk and prey abundance. Indeed it remains unclear how more nuanced predator–prey dynamics would influence herbivory intensity or plant performance. We note, however, that the growth of aspen in Yellowstone is not detectably reduced by wolf-induced predation risk [12],[14] [also see [23] for similar results for other shrub species].

Our results have implications for understanding climate change effects on the Yellowstone Ecosystem. As a dominant herbivore, elk can strongly influence plant distribution and persistence [14]. This herbivory can be regulated by snowpack conditions, which are changing as western North America warms and dries [36]. As we show, elk responses to snow are mediated by sex and (weakly) time, the latter likely a proxy for body condition. This suggests that managers can influence winter spatial variation in elk abundance -and therefore the intensity of elk herbivory- by altering the sex ratio of the herd via sex-specific hunting quotas. If the relative proportion of mature males declines, the average response to snowpack across the population may grow stronger, meaning that plants at high-elevation sites are subject to lower cumulative herbivory. Moreover, assuming that the increased avoidance of deep snow in later winter is in fact due to deteriorating elk body condition, summer plant productivity could influence winter herbivory. Multiple interacting factors, all of which are nuanced and imperfectly understood, affect patterns of herbivory. Increasing our still-limited understanding of these factors and their interactions will greatly enhance our knowledge about climate change impacts on communities.