Variation in diet of frugivorous bats in fragments of Brazil’s Atlantic Forest associated with vegetation density

,

Palavras chave: Chiroptera, efeito de fragmentação, frugivoria, isótopos estáveis, nicho trófico, Phyllostomidae Habitat fragmentation occurs when a once contiguous environment is divided into remnant fragments, or patches, of various sizes and composition in a disturbed matrix (Franklin et al. 2002). The impact of fragmentation on local animals may be wide-ranging, and either positive or negative (Fahrig 2017). Remnant patch size, number, edge density, composition, and degree of isolation from contiguous habitat have all been shown to significantly affect species biology and persistence potential (Fahrig 2017). While habitat loss can result in lower population sizes and/or decreasing diversity of the community, in many instances, species can persist in fragmented landscapes, and some may even benefit from some degree of natural fragmentation Fenton 2003, 2007).
The question of how landscape fragmentation impacts species biology poses additional challenges for the study of species whose natural behaviors are not readily observable, and dietary niche is not easily characterized. To understand cryptic trophic niche dynamics, the niche concept must first be simplified and defined within quantifiable dimensions rather than the conceptual "role" as proposed by Hutchinson (1957). When trophic niche breadth is calculated, it allows for comparison of how these realized trophic niches may change over time, or among populations in response to disturbance (Bearhop et al. 2004).
Habitat fragmentation in Brazil's Atlantic Forest is the result of extensive human settlement and significant land use conversion for agriculture (Ribeiro et al. 2009;Tabarelli et al. 2010). Despite this, the Atlantic Forest remains one of the largest and most diverse rainforests in the world. The once contiguous forest covering approximately 150 million hectares, is now fragmented into over 200,000 remnant patches representing 11-16% of original forest cover, much of which are small or young successional fragments (Ribeiro et al. 2009). More recent estimates have found approximately 28% forest cover and have (perhaps optimistically) predicted that this can be increased to 35% through continued and expanded conservation and sustainable development efforts (Rezende et al. 2018). The Atlantic Forest is home to the richest mammal fauna in Brazil with 318 species (Graipel et al. 2016;Bogoni et al. 2017), of which approximately one-third are bats Muylaert et al. 2017). This biodiversity "hotspot" (Myers et al. 2000) until recently has been under continued threat of deforestation and further habitat loss could lead to a staggering loss of global biodiversity (Ribeiro et al. 2009). Recent developments have led researchers to more hopeful outlooks (Rezende et al. 2018); however, both old and new emerging challenges to conserving this important region remain. As deforestation and reforestation have gone through several phases since approximately 1500 CE (Dean 1996), Brazil's Atlantic Forest represents a unique study region where habitat patches of various ages, size, isolation, and successional states can be compared, and the long-term effects of fragmentation quantified. Through studying the effects of deforestation, we can further our understanding of how natural processes and animals can support renewal of the degraded areas in this biodiversity hotspot.
In tropical forests, a main driver of natural reforestation is seed dispersal by frugivorous bats (Muscarella and Fleming 2007). The feeding behavior of bats plays a key role, as many fruit-eating bats carry food away from the fruiting tree distributing seeds and facilitating forest regrowth (de Carvalho-Ricardo et al. 2014). Movement patterns and foraging behavior can have significant impacts as bats may spread seeds from fruiting trees across either a wide or a narrow area. In degraded or destroyed habitats it has been shown that this ecosystem service is altered, though not always reduced (in particular for abundant frugivorous species and their commonly associated plants); however, habitat fragmentation does reduce the resources that bats have access to (Laurindo et al. 2019). If fragmentation significantly effects the presence or ecology of frugivorous bats, a cascading effect on seed dispersal and habitat fragmentation may occur; when local biodiversity is negatively impacted by local disturbances forest regrowth is also predicted to be impacted (Muscarella and Fleming 2007).
Bats in the family Phyllostomidae are among the most common in the Atlantic Forest (Pedroso et al. 2020) and are among the most diverse families of bats in the Neotropics (160+ species) including all Neotropical fruit-eating bats (Simmons 2005). Many phyllostomid bats exhibit specialized feeding strategies and are commonly grouped into broad trophic guilds (e.g., sanguivory or nectarivory). However, these guilds do not capture the full breadth of the dietary niche of these animals as many species will seek out food items from outside their assigned guilds. For example, frugivorous and nectarivorous bats often will also often eat insects (Fleming et al. 1972;Clare et al. 2014). Many of the findings of assemblage-level responses of bats to habitat fragmentation are contradictory, with some studies pointing to a significant change in abundance of certain species (Gorresen and Willig 2004;Klingbeil and Willig 2009) and decreases in species richness and diversity (Meyer et al. 2016), while others found little or no change in species composition (Bernard and Fenton 2007). This may have to do with the causes of fragmentation (i.e., natural or anthropogenic) as well as the ages and composition of fragments or the surrounding landscape.
Studies that have examined how animal diets change in response to habitat fragmentation are spread across taxa and systems with conflicting results. Many species that persist in fragmented areas switch from being specialist consumers to generalists such that their realized trophic niche breadth increases (Dunn et al. 2010;Chaves et al. 2012;Araújo et al. 2014). Others might specialize when previously diverse and abundant resources are limited (Layman et al. 2007;Bommarco et al. 2010;Boyle et al. 2012). In many cases specialist and generalist consumers entirely switch to abundant resources such that trophic niche breadth remains the same though trophic position might change; while proportion of resources taken are the same the types of resources differ. Diet switching may also occur under natural seasonal or temporal conditions due to fruiting phenology or other trophic flexibility (Nowak and Lee 2013;Clare et al. 2014). Typically, animals-particularly frugivores as a result of fruiting and flowering phenology-will switch from previously abundant resources to resources which may have been present before but were passed over in contiguous forests (Dunn et al. 2010;Boyle et al. 2012;Chaves et al. 2012). The purpose of this study is to assess how the diets of three species of Neotropical fruit bats are affected by landscape composition.
While metrics such as abundance and diversity are informative and important, changes in diet and behavior of species that remain in fragments may have significant fitness consequences. Therefore, using stable isotope analysis (SIA), we aimed to characterize the relationship between habitat fragmentation and diet of three species of Neotropical frugivorous bats. While traditional measures of realized trophic niche breadth are invasive or rely on observational data, SIA allows for a quantitative characterization of the trophic niche of populations (Bearhop et al. 2004). Based on experimental studies by Epstein (1978, 1981) naturally occurring carbon and nitrogen isotopic ratios (δ 13 C and δ 15 N) have become the most widely used measures of realized dietary niche breadth in terrestrial systems (Kelly 2000). The nitrogen isotopic ratio increases with trophic level; therefore, species that consume insects should have a higher δ 15 N than species which exclusively consume fruit (DeNiro and Epstein 1981). The carbon isotopic ratio is conserved from the environment in tissue, and is dependant on the original source of organic carbon and a calculable rate of fractionation (DeNiro and Epstein 1978). For terrestrial systems, the ultimate source of organic carbon is from plants where the photosynthetic pathway largely determines the δ 13 C ratio.
We hypothesized that the diet of populations would vary with local landscape structure and that because larger-bodied species are capable of long-distance movements, they will be less impacted by small-scale landscape changes. We note that while each study species is widely distributed throughout the Neotropics, bat movement potential was defined by a combination of body size and wing morphology (Morrison 1980;Purvis et al. 2000), records of individual movements (Arnone et al. 2016;Esbérard et al. 2017), and telemetry and recapture studies at the local scale (Heithaus and Fleming 1978;Bianconi et al. 2006;Mello et al. 2008a;Trevelin et al. 2013).
Artibeus lituratus is a canopy forager and are among the largest bats in the Neotropics (mass ≈ 72.5 g, forearm length ≈ 70.1 mm, body length ≈ 85.0 mm); these bats forage over a larger area, and are more likely to fly through open areas than smaller heterospecific frugivores (Morrison 1980;Trevelin et al. 2013). Therefore, according to our hypothesis, small-scale landscape changes should have less of an effect on the diet of A. lituratus than bats with less movement potential (Heithaus and Fleming 1978;Mello et al. 2008a).
Carollia perspicillata and Sturnira lilium are understory foragers and are similar in size (Carollia; mass ≈ 14.6 g, forearm length ≈ 40.0 mm, body length ≈ 51.6 mm; Sturnira; mass ≈ 19.0 g, forearm length ≈ 42.1 mm, body length ≈ 53.5 mm), but have different wing morphology (Tavares 2013;Marinello and Bernard 2014). Examining aspect ratio, S. lilium has more similar wing morphology to A. lituratus than to C. perspicillata; S. lilium also has significantly higher relative wing load than either A. lituratus or C. perspicillata (Marinello and Bernard 2014). Carollia perspicillata is frequently recaptured within the same areas (Bianconi et al. 2006) and tend to avoid flying through open areas (Heithaus and Fleming 1978) and are therefore narrow-ranging. Sturnira lilium is (at least superficially) similar in morphology to C. perspicillata, and has slightly more intermediate movement potential (Bianconi et al. 2006;Mello et al. 2008a). Similar to A. lituratus, S. lilium has been noted to make novel long-distance movements (Esbérard et al. 2017); however, the significance and frequency of these movements are not presently known. For the purposes of this study, we have grouped S. lilium and C. perspicillata together as narrowranging fruit bats due to their similar morphology, more limited movement potential when compared to A. lituratus, and the likelihood that small-scale habitat changes will have a larger effect on their biology than for wide-ranging species; however, we note that S. lilium is likely to be more mobile than C. perspicillata and have intermediate responses to landscape changes.
Using SIA as a proxy for diet, we predicted that (1) the wide-ranging A. lituratus, which can move through open areas (Morrison 1980;Meyer and Kalko 2008), would be less affected by landscape and patch composition than the other two focal species (no significant shifts in δ 13 C and δ 15 N between populations); and (2) narrow-ranging species, C. perspicillata (Heithaus and Fleming 1978;Trevelin et al. 2013) and S. lilium (Mello et al. 2008a), would have intraspecific variation in diet (significant differences in δ 13 C and δ 15 N) explained by the nature of the landscape they reside in such that those in smaller patches would have fewer preferred resources. More broadly, we predicted that bats with restricted nightly ranges that are resident to fragments alter their feeding behavior in response to environmental perturbation more-so than bats that make use of a larger geographic area.

Methods
Sample and data collection.-Sampling occurred between 18 December 2015 and 18 January 2017 in Brazil's Atlantic Forest in Reserva Ecologica de Guapiacu ("REGUA") and in fragments adjacent to this area. There was no sampling during the super-humid season (Delciellos et al. 2018) between late December 2015 and 19 May 2016. Thirteen areas were sampled and allocated as REGUA, REGUA2, REGUA3 for those sampled in the reserve (considered repeated efforts sampling in the same fragment), and 10 fragments designated as F1 through F10. To calculate the landscape characteristics for each sampling location we used the software packages ArcGIS 10.1 and Fragstats 3.1. We used the ESRI base maps and combined these with existing maps available through Instituto Brasileiro de Geografia (IBGE). We then extracted all the areas of forest cover and combined these data with a map of forest remnants obtained from SOS Mata Atlântica (2009). The resulting map ( Fig. 1) containing the remnants of Atlantic Forest in the study area was imported into Fragstats 3.1 where we calculated landscape metrics: Area (ha), Isolation (nearest neighbor distance, m), Perimeter (m), PARA (perimeter area ratio), Forest Cover (percentage of forested area within 500 and 1,000 m buffers), PROX (proximity index of like-fragments within 500 and 1,000 m buffers), and Distance from Source (distance from contiguous forest, m). Additionally two principal component (PC) variables calculated by Delciellos et al. (2016) were recorded for each fragment; PC1 corresponds to the abundance of palms (0.599), Cecropia sp. (0.819), and water courses (0.781) and negatively correlated with the presence of liana vines (−0.831), and PC2 corresponds with overstory (0.750) and understory vegetation density (0.717) and the presence of fallen logs (0.813), where negative PC values indicate less of these features present in the fragment than positive values.
Bats were captured using mist nets over 6-day periods in each fragment. Netting was conducted (on average) two nights consecutively, with 6, 9, and 12 m nets moved between different locations in each fragment each night. Between 7 and 12 nets were set each night on available trails and near places of interest including water courses, fruiting trees, rock crevices, and other potential roost structures for a combined average netting effort of 275.95 m 2 /h for each fragment (Teixeira 2019). While we recognize these animals may use multiple day roosts and are capable of traveling large distances, for the purposes of this study, we assume that bats captured in fragments represent a single population (i.e., A. lituratus captured in the humid season in fragment F8 represent the F8-Artibeus-humid population grouping of this species). Recapture rate was extremely low for all species, and no bats were recaptured outside of the fragment they were originally captured in to the best of our knowledge; this supports our use of population groupings and fragments as independent and distinct study sites (Teixeira 2019). Two or three small discs of flight membrane (patagia) were taken using a biopsy punch of the dactylopatagium major or medius of wings and uropatagium of captured bats. Tissue was subsequently stored in sample vials with silica gel beads for desiccation. Patagium samples can be taken in the field, are minimally invasive, heal rapidly (Pollock et al. 2016), and the tissue has a known isotopic turnover rate of approximately 100-130 days making it an appropriate tissue for SIA (Voigt et al. 2003). However, as this turnover rate is slow, it should be noted that the feeding events that would comprise the isotopic record may include food items consumed from other fragments and it is not possible to determine the precise location where this feeding took place (Mirón et al. 2006). As accuracy in analysis depends on sample size, only tissues from species with n ≥ 5 in each of several fragments were included (Jackson et al. 2011). All research followed American Society of Mammalogists guidelines and ethical guidelines of Queen Mary University of London (Sikes et al. 2016).
Tissue processing and SIA.-Samples were weighed in tin capsules and submitted to the Environmental Isotope Laboratory (EIL) at the University of Waterloo for analysis. As samples were below an ideal target weight of 0.350 mg (patagium ranged in mass from 0.050 to 0.150 mg), a nondiluted CO 2 protocol was used for low-mass samples. The samples were combusted to gas at 1,030°C and put through a 4010 Elemental Analyzer (Costech Instruments, Valencia, California) coupled to a Delta Plus XL (Thermo Fisher, Waltham, MA) continuous flow isotope ratio mass spectrometer (CFIRMS). The output was reported in δ-notation in parts per thousand (‰) anchored against standardized scales (Vienna Pee Dee Belemnite for δ 13 C and AIR for δ 15 N).
Statistical analysis.-Data were processed using R-Studio version 1.1.453 using packages Stable Isotope Analysis in R (SIAR), Stable Isotope Bayesian Ellipses in R (SIBER; Jackson et al. 2011), caret (Kuhn et al. 2019), ggplot2 (Wickham 2018), and cowplot (Wilke 2020). To control for natural variation in diet as caused by seasonal changes, we selected for samples taken within 100 days of the end of the super-humid season (i.e., samples collected between 19 May 2016 and 27 August 2016) and analyzed these separately (seasonal groupings "super-humid" and "humid"). Though bats in fragments and REGUA were captured on consecutive days with few breaks between 19 May 2016 and 18 January 2017, we separated these into seasonal treatment groups due to underlying isotopic baselines being unknown and often correlated with season (Popa-Lisseanu et al. 2015), and other evidence that season may impact body condition in other mammals in the same region (Delciellos et al. 2018).
For each population grouping (e.g., F1-Artibeus-humid) we calculated the median Bayesian-corrected standard ellipse area (SEA.b) in squared parts per thousand (‰ 2 ) to estimate trophic niche breadth and examined the ranges of carbon and nitrogen values. We preformed an omnibus ANOVA (single factor) with a post hoc Tukey-Kramer test to compare if mean δ 13 C and δ 15 N differed significantly between fragments and two-tailed Kolmogorov-Smirnov (KS) tests to compare the range of carbon and nitrogen values. For SEA.b we used SIBER to test the probability that the niche breadth of Population 1 is less than the niche of Population 2 bootstrapped to n = 10,000.
To evaluate the impact of landscape variables on diet, we constructed nine a priori selected linear regression models (Table  1) in the caret package and used Akaike's Information Criterion (AIC) for model ranking and selection (Burnham and Anderson 2002). Due to high correlation between various landscape metrics and small sample size, we limited models to univariate and bivariate models to avoid overfitting with a correlation cutoff of 40% (Burnham and Anderson 2002). We calculated the AICc (correlated for small sample size; Burnham and Anderson 2002) value for each population grouping with samples from ≥3 fragments for each of the three response variables (median SEA.b, mean δ 13 C, mean δ 15 N). For REGUA replicates (REGUA, REGUA2, REGUA3) when comparing between fragments, we selected the replicate with the largest sample size. Though these samples were taken from different parts of the reserve in different habitats, by selecting the largest sample, we feel it would best represent the overall forested block, though we did still statistically test all REGUA replicates for significant differences in median SEA.b, mean δ 13 C, and mean δ 15 N. We then calculated the difference between the best model in each set (lowest AICc value) and all other models, recorded as ∆i values. We then calculated Akaike weights (W i ) which are the probability that the ith model is the best model of the proposed set. For models with ∆i < 2, we normalized the Akaike weights (NW i ) to better represent the top set of candidate models (Burnham and Anderson 2002). For each landscape variable within the normalized model set we present the weighted means of the coefficient (β ± SE) to determine the strength and direction of each relationship.

Results
Sampled fragments were heterogenous in forest composition and degree of isolation and ranged in area from 21 to >60,000 Table 1.-A priori selected candidate models for linear regression analysis of the effect of landscape-scale metrics on the isotopic niche breadth of fruit bat populations in Brazil's Atlantic Forest. Each model was run for species-season pairs with n ≥ 5 in three or more fragments and for three response variables (median SEA.b, mean δ 13 C, and mean δ 15 N). Parameters: AREA-total area of habitat fragment in hectares (ha); ISOLATION-nearest neighbor distance between fragments in meters (m); PC1-principal component corresponding to abundance of palms, Cecropia sp., and water courses; PC2-principal component corresponding to abundance of liana vines and vertical and horizontal vegetation density; PROX1000-proximity index with 1,000 m set buffer. In several cases, there were significant differences between δ 13 C and δ 15 N in population groupings. Across all speciesseason groups, there was more variation in δ 15 N both between and within fragments than δ 13 C. Additionally, there was more variability among fragments for narrow-ranging species than for wide-ranging A. lituratus (Fig. 2). This supports our predictions as there was significantly less variability among population group mean δ 13 C and δ 15 N and ranges between fragments in A. lituratus than in C. perspicillata (both seasons) or S. lilium (ANOVA and KS tests). There were no significant differences in trophic niche breadths among populations of the same species.
The univariate model PC2 was the best model in 7 of 12 possible cases (Tables 3-6). PC2 had the largest weight indicating the overall importance of horizontal and vertical vegetation density in explaining variation in the diet of all three species of fruit bats. For wide-ranging A. lituratus PC2 was the top model for all three response variables (SEA.b, mean δ 15 N, and mean δ 13 C). For SEA.b, the univariate model with PC2 had only a 40% chance of being the best model; for mean δ 15 N, PC2 had a 72% chance of being the best model; for mean δ 13 C, PC2 was the only model remaining after normalization and therefore had a 100% chance of being the best model (Table 3). In all these cases the relationship with the response variables was positive indicating niche breadth, mean δ 13 C, and mean δ 15 N increased with increasing vegetation density (Table 7). For the narrowranging species the univariate model with PC2 was similarly the most frequently selected model (four of nine model sets), and PC1 second most (two of nine; Tables 4-6). There were only two cases where bivariate models were predicted to be the best model (AREA + ISOLATION for Carollia SEA.b in the super-humid season; ISOLATION + PC2 for Sturnira δ 13 C in the super-humid season) and three cases where a landscapescale metric is included in the best models (both aforementioned and PROX1000 for Carollia δ 13 C in the humid season; Tables 4-6).
The variable with the greatest relative importance weight (7 of 12 cases) was PC2 which is most associated with fragment-scale variation in vegetation density. Landscape-level metrics (e.g., AREA, ISOLATION, PROX1000) were rarely (3 of 12 cases) the most significant except for species captured in the super-humid season (Table 7). There are also substantial supporting data indicating the importance of fragment-level PC1, which corresponds to the presence of water courses and Cecropia spp. plants (Delciellos et al. 2016). There is little evidence that other landscape variables impact diet in frugivorous bats (Table 7).

Discussion
Fragment composition, primarily related to vegetation density, rather than landscape-scale metrics (i.e., fragment area, isolation) best explained diet variation among fruit bats. For nearly every species-season pairing and across all response variables, PC2 (positively associated with the presence and abundance of fallen logs, overstory vertical vegetation density, and understory horizontal vegetation density; Delciellos et al. 2016) was the best variable in 8 of our 12 sets of models. After normalization, only 3 of the 12 top models contained some metric related directly to habitat fragmentation (Fahrig 2017). Generally, landscape metrics such as fragment area and isolation explained variation better for narrow-ranging species than wide-ranging species, supporting our predictions. These results suggest that Neotropical frugivorous bat diets are more influenced by habitat composition (i.e., forest quality, successional state, etc.) than by the overall landscape. However, it should be noted that we did not consider the age of fragments or cause of fragmentation (i.e., natural or anthropogenic) in our models as these factors were unknown; species responses may vary significantly between natural and anthropogenically fragmented landscapes (Ripperger et al. 2013;Meyer et al. 2016).
Additionally, there was more variation in δ 15 N than in δ 13 C between resident groups within the same species-season pair. Higher δ 15 N values can imply a higher trophic level diet, suggesting that populations with high δ 15 N are likely consuming more insects. Alternative explanations of shifts in δ 15 N are also possible as we were unable to sample baseline environmental isotopic values; changes in plant resources (Craine et al. 2015), animal-plant interaction (e.g., ant mutualists with  Delciellos et al. (2016). Proximity index (PROX) is described in Carrié et al. (2017).

Fragment
Area (  Cecropia; Sagers et al. 2000), or soil composition and chemistry (Bustamante et al. 2004;Crowley et al. 2011) may also contribute to rises in δ 15 N. At present, insects in the diet of frugivorous bats are often underestimated or disregarded; however, recent studies have shown insects often make up a significant portion of diet (especially as a source of protein) for nectarivorous species (Clare et al. 2014;Orr et al. 2016). Insectivory may be a resource-driven behavior, occurring more in periods or regions with low fruit availability, and in fragments with less preferred fruits there may be an increase in insect consumption for narrow-ranging fruit bats (Clare et al. 2014).
Wide-ranging fruit bats.-For wide-ranging A. lituratus, PC2 (within-fragment vertical and horizontal vegetation density) was a metric in the top three models for every response variable. In all cases the relationship with PC2 was positive-median niche breadth, mean carbon and nitrogen ratios all increased with increasing vegetation density. These suggest that, in denser fragments A. lituratus have a larger insect component in their diet than in more open forests, irrespective of fragment size or distance between fragments. A similar study examining differences in diet of A. lituratus found in fragments and in contiguous forest found an increase of insect consumption in fragments and in seasons when fruit resources were limited (Muñoz-Lazo et al. 2019). Notably, in contrast to our study, Muñoz-Lazo et al. (2019) found that this increased the niche breadth of A. lituratus, while we found no significant differences in niche breadth between different sampling Table 5.-Difference between the top-ranked model and the ith model (∆i) with AICc weight (W i ) and normalized weights (NW i ) for candidate models of Carollia perspicillata in the super-humid season. Models with ∆i value < 2 were used for inference and are presented here.  Table 6.-Difference between the top-ranked model and the ith model (∆i) with AICc weight (W i ) and normalized weights (NW i ) for candidate models of Sturnira lilium in the super-humid season. Models with ∆i value < 2 were used for inference and are presented here.    (2019) did not characterize the locations they sampled beyond "fragment" and "continuous" forest, we found no significant effect of landscape metrics that are typically characterized as habitat fragmentation (i.e., AREA + ISOLATION, AREA, or PROX1000). We did however find that localized fragment-scale forest composition explains variation in diet for wide-ranging fruit bats. Differences in δ 13 C between fragments correlated with PC2 notably followed a pattern consistent with the "canopy effect" whereby plants closer to the forest floor have lower δ 13 C than plants which grow in the canopy (Voigt 2010). Of the populations from fragments that significantly differed from REGUA in δ 13 C, most are sparse, likely primary forests while F6 is dense with liana vines and a few larger trees. With a denser understory, A. lituratus are likely feeding on canopy fruits, contributing to the higher δ 13 C values (Voigt 2010). This may also suggest that during the period prior to capture, these bats were resident to these fragments, or sought a preferred forest type.
Most notable is the model prediction that increased δ 15 N was associated with denser forests, suggesting that in fragments with denser foliage, A. lituratus consume more insects. However, it is unlikely that wide-ranging bats are resident in these fragments; rather, it is more likely they are only foraging in certain fragments for short periods (Bianconi et al. 2006;Bernard and Fenton 2007). It is possible that A. lituratus seek out denser fragments and spend more time commuting between fragments, therefore expending more energy. Powered flight and large body size necessitate high energy expenditure (Peters 1983) and as A. lituratus are typically noted to have a preference for Ficus spp. or Cecropia spp. fruit, they may require more time to seek out preferred food sources, requiring an insect "snack" while searching for, or in transit to known ideal foraging locations (Herrera et al. 2001;Esbérard and Bergallo 2008). With these behaviors accounted for, it is unlikely this pattern is directly related to fragment vegetation density, and seemingly supports our prediction that longranging A. lituratus population diets are less affected by local landscape-scale metrics and their immediate environment than their narrow-ranging counterparts.
Narrow-ranging fruit bats.-As C. perspicillata and S. lilium are (at least superficially) ecologically similar, our models predicted that landscape metrics affect these species diets and realized dietary niche breadths in similar ways. Similar to A. lituratus, PC2 was the most significant variable in the majority of cases; however, PC1, isolation, and area are also included as key contributors in some of the top weighted models. Through examining the coefficients, there is only one case where the standard error overlaps zero (C. perspicillata-humid, δ 15 N), which indicates that this result is likely not significant. As we predicted, landscape-scale metrics appear to be more influential for these species than for A. lituratus as they occur more frequently in the top sets of models.
There were no significant differences in realized trophic niche breadth between populations of the same species; however, our models suggest that under enough environmental stress, trophic niche breadth would be affected by landscape. For C. perspicillata, PC2 was included in the top two models in humid season. Given that C. perspicillata prefer early successional "pioneer" fruit species like Piper spp., this trend likely reflects the availability of this type of fruit, as sparse fragments are more likely to have more of this preferred resource (Mello et al. 2004;Thies and Kalko 2004). Similar to studies of primates in a fragmented landscape (Boyle et al. 2012;Nowak and Lee 2013), when preferred resources are abundant, narrowranging fruit bats exploit them, and when they are rare exhibit a more flexible diet.
In the super-humid season, AREA + ISOLATION was the best model for predicting trophic niche breadth. The relationships are negative, though the parameter estimates are small and may not be biologically relevant. At constant isolation, our model predicts little variation in realized trophic niche breadth between a 10-ha fragment and 10,000-ha fragment. Notably at constant area, between 1 and 100 m in isolation, there is little variation in SEA.b (3.6324-3.1394‰ 2 ); however, at approximately 500 m dietary niche breadth is considerably smaller (1.1474‰ 2 ). Bernard and Fenton (2003) found that C. perspicillata used a network of habitat fragments at a local scale and would cross open areas, though given their limited foraging range and high energy costs associated with flight, it is likely that if a fragment is more isolated, commuting between neighboring fragments would be unlikely. For S. lilium, variation in realized dietary niche breadth among resident groups was best predicted by PC2 and differs from C. perspicillata in the humid season, as the relationship is negative. This reflects that populations in fragments with dense vegetation have a smaller dietary niche breadth than those in sparse fragments. Across all fragments Sturnira have on average a broader trophic niche than Carollia (2.802‰ 2 versus 1.885‰ 2 ) and generally have been noted to have a more variable diet (Mello et al. 2008b). Though biologically similar species in body size, foraging behavior, and range, Carollia and Sturnira respond to landscape composition in different ways. This adds to several records demonstrating species-specific responses to landscape alteration and shows there is no single rule when grouping species that are similar; vulnerable species must be considered as individual species rather than with like-species for effective assessments and conservation action (Arroyo-Rodríguez et al. 2016;Willcox et al. 2017). Sturnira lilium additionally have been noted to make vertical migrations in Brazil (McGuire and Boyle 2013;de Carvalho et al. 2019) and generally have greater movement potential than Carollia. It is likely that their ability to travel larger distances due to differences in their wing morphology (Tavares 2013;Marinello and Bernard 2014), and subsequently their behavior has significant bearing on these findings.
For predicting δ 13 C in Carollia, the reverse pattern of Artibeus was observed in the super-humid season and is also likely a result of the canopy effect as C. perspicillata prefer to feed on low-growing fruits. While fragments with greater vertical vegetation density would likely have a denser canopy, and therefore fewer shade-intolerant Piper spp., there are some species of Piper that are shade-tolerant and they are likely to be abundant particularly on or near the fragment edge (Thies and Kalko 2004). Additionally shadetolerant variants of the same species may have different carbon isotopic composition (Krishnaprasad et al. 2017). Changes in mean δ 13 C between fragments might therefore be related to consuming different species of Piper or having a different quality of diet which might change the fractionation of carbon between the environment and the animal's tissue. In the humid season, the best model for δ 13 C (PROX1000, 28.9% probability of being the best model) was considered spurious as the standard error overlapped zero when examining the parameter estimates (Table 4).
In the super-humid season, mean δ 13 C in Sturnira was best explained by PC1, where higher δ 13 C values were in areas with more abundant Cecropia spp., more water courses, and fewer liana vines (Delciellos et al. 2016). As in the super-humid season it is unlikely that water is a limiting resource, fruit from Cecropia spp. are likely the influencing factor as it has been recorded in the diet of S. lilium (Lobova et al. 2003). The significance of Cecropia fruits in their diets has not been explored and it is possible that an abundance of these fruits in times when preferred food items (i.e., Solanaceae) are less abundant would drive this difference in δ 13 C.
Examining differences in δ 15 N between populations for Carollia in the humid season, PC2 was once again the best (and only) model after normalization. The relationship is negative, suggesting that in sparser fragments Carollia take more insects. Significant differences in δ 15 N range (KS tests) and mean (ANOVA) between fragments and REGUA indicate that different populations likely vary in insect consumption, and perhaps insect abundance is a major contributing factor.
In the super-humid season, the relationship between landscape and δ 15 N was more apparent in both Carollia and Sturnira; however, differences in biology between these species make these trends difficult to interpret. The best model for mean δ 15 N in C. perspicillata populations was PC1 with fragments with more Cecropia sp. seeming to have higher insect consumption. In S. lilium, the best model (ISOLATION + PC2) would predict populations in denser and more isolated fragments to have higher mean δ 15 N. Seasonal fluctuations in insect abundance as relative to fruit availability may explain these different responses. Both Carollia and Sturnira are known to switch resources at different points in the year to avoid resource shortages and perhaps to avoid competition with sympatric species (Heithaus and Fleming 1978;Mello et al. 2008b). Previously we have noted that Carollia seem to have a more specialized diet than Sturnira and therefore may be more sensitive to periods of preferred resource scarcity (Oelbaum et al. 2019). When Piper fruits are less available as they would likely be in the super-humid season in denser fragments (Thies and Kalko 2004;Valentin-Silva et al. 2018), we would predict an increase in δ 15 N (as in Sturnira) as a result of increased insect consumption; however, this is not what our model predicts. Foraging behavior is also noted to change between the super-humid and humid season; however, changes in insect consumption have not been noted (Fleming and Heithaus 1986).
While habitat composition (i.e., vegetation density; PC2) explained variation in δ 13 C and δ 15 N in most cases, landscapescale metrics were only important in one-third of cases. There were more significant effects of landscape metrics in narrowranging species however and more testing is required in more fragments with other, less correlated landscape variables to determine the true drivers of dietary differences. It is unlikely that insect consumption by frugivorous bats is driven by large-scale landscape composition, but it may be linked to local habitat type. Similarly, realized dietary niche breadth is not significantly affected by landscape; however, differences between populations are best explained by habitat type in most cases. Nearest neighbor distance between fragments was the most important landscape-scale variable for narrow-ranging species and had more weight in determining diet than fragment area. The effects of seasonality in diet of these animals should be further explored.