Male and female bees show large differences in floral preference

Background Intraspecific variation in foraging niche can drive food web dynamics and ecosystem processes. In particular, male and female animals can exhibit different, often cascading, impacts on their interaction partners. Despite this, studies of plant-pollinator interaction networks have focused on the partitioning of the floral community between pollinator species, with little attention paid to intraspecific variation in plant preference between male and female bees. We designed a field study to evaluate the strength and prevalence of sexually dimorphic foraging, and particularly resource preferences, in bees. Study design We observed bees visiting flowers in semi-natural meadows in New Jersey, USA. To detect differences in flower use against a shared background of resource (flower) availability, we maximized the number of interactions observed within narrow spatio-temporal windows. To distinguish observed differences in bee use of flower species, which can reflect abundance patterns and sampling effects, from underlying differences in bee preferences, we analyzed our data with both a permutation-based null model and random effects models. Findings We found that the diets of male and female bees of the same species were often dissimilar as the diets of different species of bees. Furthermore, we demonstrate differences in preference between male and female bees. We show that intraspecific differences in preference can be robustly identified among hundreds of unique species-species interactions, without precisely quantifying resource availability, and despite high phenological turnover of both bees and plant bloom. Given the large differences in both flower use and preferences between male and female bees, ecological sex differences should be integrated into studies of bee demography, plant pollination, and coevolutionary relationships between flowers and insects.


Findings
We found that the diets of male and female bees of the same species were often dissimilar as the diets of different species of bees. Furthermore, we demonstrate differences in preference between male and female bees. We show that intraspecific differences in preference can be robustly identified among hundreds of unique species-species interactions, without precisely quantifying resource availability, and despite high phenological turnover of both bees and plant bloom. Given the large differences in both flower use and preferences between male and female bees, ecological sex differences should be integrated into studies of bee demography, plant pollination, and coevolutionary relationships between flowers and insects. PLOS

Introduction
Intraspecific variation in traits and behavior, including foraging niche, has important consequences for species interactions and conservation [1,2]. Sexual dimorphism is a large source of individual niche variation, and an important factor in plant-animal interactions, such as seed dispersal [3]. Sexual dimorphism underlies adaptation, speciation, and the way in which animals exploit their ecological niche [4,5]. Morphological, behavioral, and life-history dimorphisms can also drive the form and function of ecosystems, for example when predator sex ratio drives the community composition of lower trophic levels, affecting the physical and chemical properties of the environment [6,7]. Though ecological dimorphisms were first studied in vertebrates [8], they are common across taxa, including insects [9]. Surprisingly, in bees (Hymenoptera, Apoidea) for which both foraging [10] and sexual dimorphism [11] have been well studied, sexually dimorphic foraging has rarely been documented. Intraspecific variation in floral preference is known for social [12] and to a lesser extent, solitary bees [13,14], yet most community-level studies focus on species-level interactions, and specifically on how female bees forage.
Male bees differ from their better-studied female counterparts in their life history and ecology. Female bees construct, maintain, provision, and defend nests, whereas male bees primarily seek mates [15]. Both sexes drink floral nectar for their own caloric needs, but only females collect pollen to provision young, and thus forage at greater rates. Pollens from different flower species (the term we use throughout for the flowers from a species of plant) tend to be distinct not only in morphology but also in terms of nutritional content, and both of these (factors drive plant-specific foraging by bees [16]. While recent work shows variation in nectar is more important than previously acknowledged [17], even species of bees that specialize narrowly on pollen hosts (oligolects) tend to nectar from many flower species. It is unclear what factors drive male bee preferences, though the criteria males use to select floral partners probably differ from those used by females.
Although female bees are more prolific pollinators due to the greater time they spend foraging at flowers, when male bees have been studied, they prove to be important pollinators as well. This is true not only in specialized oil-or scent-collecting pollination systems, where males would be predicted to be important [18][19][20], but also for males simply foraging for nectar [21][22][23]. Male bees may also be particularly relevant for bee conservation. Males may be limiting in declining populations, either because genetic diversity is necessary for the development of female offspring as a result of complementary sex determination, or because mate or sperm limitation results from poor male condition [24][25][26]. As the dispersing sex in most bee species, males may be crucial for gene flow and metapopulation persistence even when they are not locally limiting [27,28].
Observed differences in resource use, which reflects the overlap of consumers and resources (availability) as well as consumer preference, may fail to reveal more essential differences in foraging niche. Preferences may be more important than use alone in the context of species conservation, and may mediate the strength of selection imposed by interaction partners. Preference-the use of a resource in excess of its relative availability-is challenging to measure, because both resource use and availability must be known. Floral resource availability for pollinators is particularly hard to quantify outside an experimental context because the appropriate scale and units of floral resource availability are unclear. The composition, amount, and supply rate of pollen and nectar per flower, the number of flowers per inflorescence, of inflorescences per individual, and the number and distribution of individual plants over the square kilometers of a bee's foraging range are all important components of availability [29]. Furthermore, floral availability can change rapidly over time. However, differences in flower use between bees foraging at the same place and time indicate differences in preference, which may occur between species, or between individuals of the same species.
In this study, we assess differences between floral preferences of male and female bees in the field. We collected bees foraging on flowers in meadows in New Jersey, USA. In order to observe preference differences, we collected as many individuals as possible during replicated, short (3-day) windows, during which we assumed floral availability and bee abundance were constant at each site. We compare the species composition of flowers visited by males and females of the most common bee species across the entire study as a naïve measure of differences in preference between the sexes. Then, using random effects models, we assess when differential flower species use by male and female bees likely arises from sex-specific floral preference, as opposed to shifting overlap between foragers and floral resources (i.e. changes in availability without differences in preference). Specifically, we ask 1) How much do male and female bee diets overlap?
2) To what degree are particular flower species disproportionately visited by bees of one sex?
3) To what extent are differences in floral use driven by preference, rather than phenological differences between male and female bees?

Study design and data collection
Because absolute preference is nearly impossible to observe outside of an experiment, we designed our study to reveal differences in preference between groups of bees. In order to collect a large number of males and females from many native bee species, we selected six meadows (sites) in New Jersey, USA with a high abundance and diversity of flowers. These seminatural meadows were managed for pollinator-attractive, summer-blooming forbs through seed addition, and a combination of mowing, burning, and weed removal. Most flower species present in the meadows are native to the eastern United States. We collected our data during peak bloom and maximum day length (6 June to 20 August, 2016), and during good weather (sunny enough for observers to see their own shadow, no precipitation). We visited each site for three consecutive good weather days over five evenly spaced sampling rounds in the 11-week period of our study. In all analyses, we assume that bees and flowers detected at a site within one 3-day sampling round co-occurred. In contrast, we assume that turnover of both plant species in bloom and bee species activity can occur in the~10 days between sampling rounds. During each 3-day sampling round, an observer walked parallel transects through the meadow (which ranged in size from 0.8-2.2 ha; mean = 1.4 ha), observing every open flower within a moving 1-m semicircle, and net-collecting any bee seen actively foraging, which we defined as contacting anthers or collecting nectar from a flower ( Figure A in S1 File). We collected all bee species except Apis mellifera L., the domesticated western honey bee, because Apis males do not forage. Observations began as soon as pollinator activity picked up in the morning (7-9 am) and continued into the late afternoon or evening until pollinator activity slowed substantially. Observers sampled nearly continuously, in 30-minute timed collection bouts with short breaks in between. If inclement weather precluded a minimum of six 30-minute sampling bouts in a day, we added an additional day to the sampling round as soon as weather permitted.
Flower species were identified in the field by the data collector. Bee species were identified using a dissecting microscope and published keys; Jason Gibbs (University of Manitoba), Joel Gardner (University of Manitoba), and Sam Droege (USGS) assisted with identification for bees in the genera Andrena, Anthophora, Coelioxys, Halictus, Heriades, Hoplitis, Hylaeus, Lasioglossum, Megachile, Melissodes, Nomada, Osmia, Pseudoanthidium, Ptilothrix, Sphecodes, Stelis, and Triepeolus, and at least one of them confirmed voucher specimens for every species. We determined every specimen to species except for the following four complexes: Most bees in the genus Nomada with bidentate mandibles (ruficornis group) were treated as one species. All specimens from the Hylaeus species complex that includes Hylaeus affinis, H. modestus, and at least one additional species, informally dubbed "species A," were treated as a single species, denoted Hylaeus affinis-modestus, because females cannot be reliably distinguished. There is a cryptic species in the genus Halictus unlikely to occur in our area, Halictus poeyi, which is not morphologically distinct from H. ligatus; we treat all specimens in this complex as Halictus ligatus. We could not confidently separate all specimens of the two closely related Lasioglossum species Lasioglossum hitchensi and L. weemsi. Thus, we treat all specimens from either species as one, denoted Lasioglossum hitchensi-weemsi.
All bee specimens are curated in the Winfree lab collection at Rutgers University, and the data used in this paper, along with R scripts used in data analysis and figure preparation, are available from the Dryad Digital Repository (doi:10.5061/dryad.c3rr6q1). No specific permits were required to collect these data, however, we obtained permission to access meadow habitats and sample insects from Mercer County, the Institute for Advanced Study, Somerset County, and the Raritan Headwaters Association

Analytical methods
We performed all statistical analyses and simulations using R 3.5.1 [30].

1) How much do male and female bee diets overlap?.
To compare the diets of male and female bees, we used the Morisita-Horn index of resource overlap [31,32]. This dissimilarity index compares the proportion of all female bees found on each flower species to the proportion of all male bees found on each flower species. In other words, it compares the contribution of each flower species to female diets (where this term includes the food that females collect for themselves and also to feed to young) to the contribution of the same flower species to male diets. The Morisita-Horn index ranges from zero (completely similar) to one (maximally dissimilar), and has several good properties for our purposes. First, it uses proportions, placing visits from male and female bees on the same scale, even though most visits come from females. Second, it is much more sensitive to large proportions than to small ones, thereby down-weighting the contribution of flower species for which we have little information. Third, the Morisita-Horn estimates are resilient to undersampling and uneven sample size between groups [33].
To determine whether the male-female differences we observed exceeded those expected by chance, we compared the observed compositional dissimilarity between flower visits from male and from female bees to dissimilarity measures from a null model that randomly permuted the bee sex associated with each flower-visit record. This permutation holds constant the total number of male and of female visits, and the total number of visits to each flower species from both sexes combined ( Figure B in S1 File). The range of dissimilarity values from this simulation is the difference we would observe in our sample, if there were no true difference in flower species use between males and females of the same bee species. We evaluated the hypothesis that male and female diets overlap less than would be expected by chance; thus we use a one-sided alpha of p<0.05. We iterated this null simulation 9999 times, which was sufficient to stabilize p-values near our chosen alpha (North, Curtis & Sham 2002). When the observed dissimilarity was greater than 9500 of the 9999 simulated dissimilarities, we concluded that we had detected a difference in the pattern of floral visitation between conspecific male and female bees, given the observed diet breadth and abundance of each sex. We also computed the mean null model value, and a 95% confidence interval for this mean using the 0.025 and 0.975 quantiles of the dissimilarity values generated for each null model.
To compare the diet overlap we observed between sexes to a meaningful benchmark, interspecific diet overlap, we repeated the same null model analysis, this time comparing females of the focal species to females of other species. To compare the results, we present the difference between the observed dissimilarities and the null dissimilarities for each female-male and species-species comparison. We performed one analysis for each bee species for which we collected at least 20 visitation records for each sex (19 species). This sample size threshold is arbitrary, but null model variance shrinks with sample size, such that apparent patterns for species with smaller sample sizes are rarely interpretable (Figure C in S1 File).
Because we analyze 19 bee species, females of each species are compared to 18 others. We then compared the female-male difference (observed minus mean null dissimilarity in flower communities visited) to the analogous species-species difference (observed minus null dissimilarity).
For this analysis, which evaluates holistic differences between male and female bees of the same species, we combined observations across the full season and all sites. This allows us to observe foraging niche differences that are driven by flower and/ or bee phenology, in addition to any sex-specific floral preference.
2) To what degree are particular flower species disproportionately visited by bees of one sex?. To answer this question, we fit a random intercepts model to our entire data set of 153 bee species to determine whether particular flower species are disproportionately visited by male or female bees, and whether the answer varies by bee species. In our model, bee sex is the response, and flower species, bee species, site, and their interactions are all random effects; we included no fixed effects. The random effects provide partial pooling, which is especially useful when there are many levels, few data associated with some or all levels, and/ or inconsistent amounts of data across levels [34]. We can infer disproportionate visitation by male vs. female bees for a flower species when predicted odds of visitors to that flower species being male are especially high or low.
We statistically control for variation in the overall sex ratio across bee species through a random intercept of bee species, and variation in sex ratios across sites, through random intercepts for site, and the site-bee species and site-flower species interactions. Although we deemed it unlikely that, within bee species, sex ratios at birth vary greatly across space, any variability attributed to site terms could result from differences in bee sex ratios, or from differential overlap of bee foraging activity and flower bloom across space. We call this model the "summed model" because we sum interactions observed across the entire season (all five sampling rounds) at each site. In the summed model, the relationship between phenological overlap and the odds of flower-visiting bees being male would be incorporated into the species effects. This perspective is helpful for considering flower species' contributions to the overall diets of male versus female bees. We fit the model with the R package lme4 [35] with the following call: Summed model lme4::glmer(bee_sex~(1|site)+ (1|flower_species)+ (1|bee_species)+ (1|flower_species: bee_species)+ (1|site:bee_species)+ (1|site:flower_species), family = "binomial", data = data) We included bee species and site as random, rather than fixed, effects to directly compare the variability in bee sex associated with each of these predictors to the variability associated with flower species (preference). Comparing the overall variability across these groups was more important to us than assessing predictions on a per-site or per-bee-species basis. We fit flower species, the primary covariate of interest, as a random effect to facilitate model fitting (fewer degrees of freedom) as well as interpretation. In our summed model, we included all two-way interactions, but omitted the three-way interaction, bee species by flower species by site. Although the sort of context-dependent preference this term could represent (e.g. males from bee species 1 prefer flower A at one site (relative to females), but shun it at another) may exist in nature, it is unlikely we would estimate it accurately in our model. We confirmed model convergence by comparing several fitting methods using the allFit function in lme4 [35], which all showed similar parameter estimates (Table A in S1 File). We tested whether residuals from our model were overdispersed using Bolker's function "overdisp" [36], and visually assessed our additivity assumptions with binned residual plots [34] ( Figure D in S1 File).
3) To what extent are differences in floral use driven by preference, rather than phenological differences between male and female bees?. Over the 11 weeks of our study, we observed turnover in bee species, in flower bloom, and within-bee species changes in sex ratio. Therefore, phenological overlap between male versus female bees and the bloom period of particular flower species, rather than preference of those bees for those flowers, may explain much of the variation in sex ratio we observed across visitors. In question 3, we are explicitly interested in distinguishing sex-specific diet preferences from variable use resulting from seasonal resource availability and male vs. female abundance. We do this in the "seasonal model" by incorporating sampling round (our measure of phenology) as an additional random intercept effect, along with random intercepts for the interactions between sampling round and the other covariates. We chose to include sampling round as a random effect because this enables direct comparison to all other terms in both models. We ignored the three-and four-way interactions between bee species, flower species, and other covariates. We fit this model with the following call in the R package lme4, with new terms in bold: Seasonal model glmer(bee_sex~(1|site)+ (1|flower_species)+ (1|bee_species)+ (1|flower_species:bee_species)+ (1|site:bee_species)+ (1|site:flower_species)+ (1|sampling_round)+ (1|site:samplin-g_round)+ (1|flower_species:sampling_round)+ (1| bee_species:sampling_round)+ (1|site:bee_species:sampling_round)+ (1|site:flow-er_species:sampling_round), family = "binomial", data = data) Our index of preference for both the summed model and the seasonal model is the change in predicted odds that a bee is male when the flower species it visits is given. To describe the importance of model terms, we calculated a bootstrapped median odds ratio using code from Seth [37], which gives the expected difference in odds that a flower-visiting bee is male between levels of a predictor [38]. For example, a median odds ratio of five for the main effect of sampling round would indicate that the odds of a flower-visiting bee being male differ by about a factor of five between sampling rounds, while a median odds ratio of one would indicate that the odds of a flower-visiting bee being male do not change across rounds. If the median odds ratio is large for flower species in both models (and the random effects predictions for each species are consistent across both models), we could say that there are intrinsic (i.e. not simply phenological) properties of flower species identity that male or female bees prefer. If flower species is a strong predictor of bee sex in the summed model but not in the seasonal one, we would still conclude that flower species often contribute more strongly to the diet of one sex than the other, though these differences may not arise due to differing preferences. If the sampling round terms have large median odds ratios, then accounting for phenology is critical for identifying differences in preference in addition to differences in use.
Because male bees require less pollen for their own diets and do not collect pollen to provision offspring, we predict that they would be less likely to visit flower species that do not produce nectar than females would. Post-hoc, we examined whether the predicted odds that visitors to nectarless flower species would be male were lower than for flower species known to produce both pollen and nectar in the seasonal model (Appendix A in S1 File).

Results
In total we collected 18,698 bee specimens belonging to 152 bee species (Table B in S1 File) from a total of 109 flower species (Table C in S1 File), which together comprised 1417 unique species-species interactions. Although the ratio of male to female bees was highly variable across bee species (Table B in S1 File), roughly 18% of specimens were male (n = 3372). Thus, the overall ratio of male to female bees we collected was 0.22, although this ratio varied markedly between flower species (Fig 1).

How much do male and female bee diets overlap?
We found that male and female bee diets overlap significantly less than would be expected given random sampling of the flowers visited by both sexes (Fig 2), and that the differences in diet composition between male and female bees of several species were of similar magnitude to the differences in diet between species of bee (Fig 3). The patterns we observed did not result from a single tendency across all bee species, such as males always visiting a nested subset of flower species visited by females ( Figure E in S1 File). To what degree are particular flower species disproportionately visited by bees of one sex?
The sex ratio of flower-visiting bees varied across species of flower (Fig 2). After controlling for bee species identity (the strongest predictor of sex in our models, Fig 4), and site, we still found that some flower species received a disproportionate number of male bee visitors (Figs 4  and 5). The median odds ratio for the main effect of flower species was 3.6 (bootstrapped CI 3.0-4.2) in our summed model, indicating that, typically, the visitor sex ratio differs between two flower species by more than a factor of 3. Furthermore, we observed sex-based differences in flower use specific to particular bee species: the median odds ratio for the flower species by bee species interaction in our summed model was nearly as large (median = 3.1, bootstrapped CI 3.0-3.3) as the main effect of flower species. By contrast, sex ratios did not differ between sites (median odds ratio for main effect of site = 1).
To what extent are differences in floral use driven by preference, rather than phenological differences between male and female bees?
The flower species blooming in our system turned over throughout our 11-week sampling period, with several highly visited species blooming for only one of the three months during the observed statistic minus, for each pairwise comparison, the mean dissimilarity in the null model. Each panel focuses on a bee species (panel name) and shows: above the label "sex", observed diet dissimilarity between male and female bees of the focal species, minus the average null dissimilarity resulting from randomly permuting the sex identity of each visit record; above the label "sp", observed diet dissimilarity between female bees of the focal species and each other bee species, minus the average null dissimilarity resulting from randomly permuting the species identity of each visit record.
https://doi.org/10.1371/journal.pone.0214909.g003 which we sampled. This turnover, along with potential sex-specific bee flight seasons (e.g. males emerge first in many solitary species, but are not produced until the end of the colony cycle for many social species), means that differences in diet between male and female bees could reflect seasonal availability and use, without also indicating preference differences between the sexes. Indeed, phenology predicts bee sex somewhat, with the odds of a flower-visiting bee being male expected to change by a factor of 1.5 (bootstrapped CI 1.1-1.9) between sampling rounds (Fig  4). Phenological patterns of male vs. female flight seasons vary across bee species; the median odds ratio for the bee species by sampling round interaction is 2.2 (bootstrapped CI 2.1-2.3) (Fig 4). Even after accounting for these effects, however, there remains a strong association between the species of flower a bee visits and its sex (Figs 4 and 5).
The relative effects of each flower species on the sex of its visitors were changed very little by accounting for phenology; Pearson and Spearman correlations between the random effect of flower species in the seasonal model and the same random effect in the simpler summed model were both 0.98. Each flower species recorded in our study appears in Table C in S1 File, along with the number of male and female bees collected from it and the conditional mode of the random effect prediction from the seasonal model. In addition to finding overall preference difference between males and females, we found evidence for bee-species-specific difference in floral preferences between the sexes (median odds ratios in both models for the bee Flower species is an important predictor of bee sex even after accounting for phenology (seasonal model). For each term ("bee" = bee species, "flower" = flower species, "round" = sampling round) in each model, the median odds ratio (+/-95% bootstrapped credible interval) indicates the expected difference in odds that a flower-visiting bee is male between two levels. For example, a median odds ratio of 3.7 for the flower species term means the odds of a visitor being male are expected to differ by a factor of 3.7 between two randomly selected species of flower. https://doi.org/10.1371/journal.pone.0214909.g004 Male and female bee foraging species by flower species interaction > 2.8). A post-hoc examination of random effects predictions was consistent with our prediction that male bees avoid flower species that do not produce nectar, with the odds of visiting bees being male nearly twice as high, on average, for flower species that do produce nectar (Figure E in S1 File). Each point is the conditional mode of the random effects prediction (the random-effects analog to an estimate), for a flower species that received at least 20 visits, on the logit scale. Zero represents the odds of a visitor being male on a random flower, and -2 or 2 indicates a~7 fold decrease or increase in those odds, given flower species identity. Error bars are the square root of the conditional variances on the conditional mode � 1.96, and can be interpreted as the expected range in which the random effect for a particular flower truly lies, analogous to 1.96 times the standard error of the mean for a fixed effect. https://doi.org/10.1371/journal.pone.0214909.g005

Discussion
We found strong evidence for sexually dimorphic foraging preferences in bees. At the level of individual flower species, we found that commonly, a disproportionate number of bee visitors were male, and that sexually dimorphic preferences drove these patterns. We found that the difference between the flower species visited by male and female bees of the same species was similar in magnitude to differences between females of different species. The partitioning of the floral community among bee species is a primary focus of pollination ecology and ecological network analysis [39], but male bees are typically disregarded or lumped together with their female counterparts. Our study suggests this may represent an important oversight. Further, our study provides strong confirmation of the few studies that investigate the foraging behavior of male bees, which found that males play a unique role in plant pollination [21][22][23]40]. Lastly, our result implies that male bees contribute substantially to the complexity of plant-pollinator networks in nature, and that network analyses might benefit from separating males and females into different nodes [2,3].
Phenology, a previously reported mechanism for distinct use of floral resources by male and female bees [21,41], explained some variation in the sex ratio of flower-visiting bees, but was less important than flower species identity over the period of our study. We expected to find an effect of phenology because both the identity of the flower species blooming within sites, and also the sex ratio of foragers within bee species, vary across the season. Males emerge first in most solitary bees. In contrast, for social species, initial broods usually consist primarily of female workers, then males and reproductive females emerge at the end of the colony cycle [15]. To account for the possibility that phenology explained the disproportionate use of many flower species by one sex of bee, we extended our summed model by adding phenology terms. Surprisingly, phenology only weakly predicted the sex of flower-visiting bees. This is despite the fact that, as predicted by natural history, the sampling round(s) in which males were relatively more prevalent depended on bee species (the bee species by sampling round interaction was much bigger than the sampling round main effect; Fig 4). This indicates that our evidence for floral preference differences between male and female bees was robust to accounting for seasonal turnover in flower species bloom, bee species flight seasons, and the sex ratios within bee species.
Patterns in bee-flower interaction data can arise from the sampling process itself [42,43]. Our analyses control for these patterns. To evaluate diet overlap, we used a dissimilarity index that downweights rarely used resources, and implemented a null model that accounts for differences that could arise from sampling effects or the fact that females outnumber males in our dataset by nearly a factor of 5. To evaluate preference, we used random effects models that incorporated all (nearly 19,000) observations, and shrank extreme values for rarely observed species-species interactions towards the global mean for each effect. Thus, our estimates for sex-specific preferences should be robust to the inevitable under-sampling of rarer taxa. Establishing differences in preference between categories of bees such as males and females, even when resource availability is seasonal and difficult to quantify, is possible using methods such as these, though absolute preference remains elusive.
Some studies show, and conservation practice assumes that floral diversity is associated with more bee individuals and diversity, although this pattern could arise from many processes [44][45][46][47][48], Complementary flower species use between the sexes implies one mechanism by which a bee species could benefit from a diversity of flower choices: a resource used in small proportions at the species level may be crucial for fitness in one sex. Such a dependency would likely be overlooked when individuals of both sexes are pooled before analysis.
Within pollinating insects, sexually dimorphic preferences and contribution to plant reproduction have been reported before [49][50][51]. Though studies examining foraging differences between male and female bees [52][53][54][55] and pollination by male bees [22,23,40,[56][57][58] have been few outside sexual mimicry and scent collection pollination systems, they found that male and female bees visit different flowers, and that male bees could be important pollinators. Male bees may be especially implicated in long-distance pollen transfer [18,[58][59][60], although outside the tropics there is little direct evidence this is true.
Mating behaviors of male bees [11,15,[61][62][63][64][65][66] are better known than foraging behaviors. The two, however, are likely closely linked. Mating-related selection may drive differences in the sensory systems of male and female bees [67][68][69], or even their approaches to learning [70,71]. The mate seeking behaviors of male bees, such as patrolling routes [65] or seeking flowers visited by conspecific females [72] could generate differences from females via complementarity (males visiting flower species not visited by females), or nestedness (one sex primarily visiting a subset of species visited by the other). We found evidence for both ( Figure E in S1 File). Divergent floral preferences between sexes may reflect nutritional needs or mating behavior, but could also reflect visual or olfactory sensitivities that differ between the sexes [67,68].
Lastly, nutritional rewards likely drive differences in flower species preference between male and female bees. Whereas most female bees collect both nectar and pollen, male bees forage primarily for nectar to fuel flight [15]. Thus, we predicted that male bees would avoid flowers that produce no nectar. Indeed, in both our models, the predicted odds of a bee visiting a nectar-less flower species being male were approximately half that of a bee visiting flower species that produce nectar (Figure E in S1 File). Within flower species that produce pollen and nectar, we found large variation in the relative preferences of male and female bees. Further investigation could reveal which floral traits mediate these sex-specific flower preferences and visitation rates in bees.
Scaling up, it is currently unknown how the distinct foraging niches of male bees mediate either the robustness of pollinator communities to species loss and environmental perturbations [14,81,82], or the effectiveness of different habitat ameliorations [49,55,83]. This study suggests that both questions warrant further investigation.
Supporting information S1 File. Sampling scheme. (a) The six study sites in central New Jersey, USA. (b) Schematic sampling diagram (not to scale). One observer walked parallel 2m transects covering the entire sampling area. Each 30-minute sampling bout resumed where the previous one left off; observers typically covered the entire meadow once over a 3-day sampling round. (c) The southwestern-most site in peak bloom ( Figure A). Schematic cartoon of our simulation for the dissimilarity values associated with our null hypothesis that diets of male and female bees do not differ. (a) Each collection record for each bee species associates the sex of an individual bee to the flower species from which it was collected. (b) To compute the dissimilarity between males and females, we compare all visits to each flower species from males (purple vector) to all visits to each flower species from females (green vector). (c) The Morisita-Horn index summarizes the differences between the two vectors as a value between 0 (identical) and 1 (maximally dissimilar). (d) For our null model, we shuffle the sex column from our observation table. (e) This produces two null vectors. The row and column sums for the matrices in (b) and (c) are identical, but the elements can differ. (f) For our null model, we compute the dissimilarity between the null vectors. We repeated steps d-f 9999 times to generate confidence intervals for the null hypothesis that the sex of a visiting bee is unrelated to the flower species it is collected from. When comparing the flower species visited by different species of bee, we conducted an analysis identical except that rather than comparing two sexes of the same species, we compared two species of the same sex (i.e. exchanging "sex" and "species" throughout Figure A in S1 File) ( Figure B). Effect size for diet dissimilarity is independent of sample size, while standardized effect is strongly driven by the number of individuals of the sex with the fewest records. a) Observed Morisita-Horn dissimilarity in flower communities visited by male and female bees of a single species, minus average null dissimilarity vs. the number of records for the less frequently observed sex. b) Observed minus null dissimilarity in composition of flowers visited by male and female bees of a single species, scaled by the variation in the null model, versus the number of records for the less frequently observed sex ( Figure C). Binned residual plots for each model show minor violation of the additivity assumption. Residuals and predicted values on the probability scale ( Figure D). Seasonal model predictions are consistent with the hypothesis that male bees avoid flower species that do not produce nectar, relative to females. Each point is the random effect prediction (change in odds that a bee visiting that flower is male) for a flower species. Boxplots show the 25th, 50th, and 75th percentiles, with whiskers extending to more extreme values within 1.5x the interquartile range ( Figure E). Methods for post-hoc analysis of male avoidance of nectar-free flowers (Appendix A). Model convergence confirmed based on similar parameter estimates across fitting routines. For each model, the estimate for each term is given for each of 6 fitting algorithms in the R package lme4. Subsequent analyses used parameter estimates in yellow, in both cases tied for the highest estimated likelihood with other very similar fits (Table A). Bee species with number of female and male specimens collected (Table B). Number of male and female visitors to each plant species, and bias towards attracting male bee visitors. This bias is the random effect prediction from the seasonal model, which indicates the change in log(odds) that a visiting bee is male when the species of flower it visits is given; greater values indicate male bias (Table C)