Bumble bee species distributions and habitat associations in the Midwestern USA, a region of declining diversity

We based our sampling design on Ward et al. (2014) and the US Fish and Wildlife B. affinis survey protocol (2017). Bee surveys were performed between the hours of 0900 and 1800 on days without precipitation, with little or no wind, and with air temperature above 21 °C. Upon each site visit, a team of one to four trained observers recorded bee visits to flowers for a total of 90 person minutes, not including netting and handling time. If there was one observer, that person netted for the full 90 min. If two observers were present, each netted independently for 45 min. For three or four observers, the time per person was reduced accordingly. The observers spread out and walked slowly through the best available foraging habitat, recording all bee visits to flowers and the flower species on which they were observed. We recorded species identity and social caste of each bumble bee. Bumble bees were either identified to species on the wing, or were netted, identified using a field guide (Colla et al. 2011), photographed if needed for identification confirmation, and re-released on site. To minimize our impact and because of the large number of bumble bees observed at some sites (as many as 452 in a single 90-min survey), the majority of bees were not collected. We searched systematically, moving between flower patches often, to minimize double-counting of individual bees.

We estimated the available flower resources for bees at each site on every visit by surveying flowers along four 25 m by 1 m transects (100 m² total area). The four transects were placed within the site in a way that represented the flower community in the area where bees were surveyed. For instance, if we had surveyed flowers along the forest edge, within mowed paths, and in open field, we would haphazardly place transects within each of these areas to capture the primary floral resources used by bumble bees, choosing areas roughly based on the amount of time we spent surveying each. In each transect, we recorded the number of flower units of each species in bloom. Species that exceeded 1000 flower units per 25 m transect were recorded as “1000.” Flower units were defined in “bee walkable clusters,” with one unit as the number of flowers a bee could visit by walking before it would have to fly to the next cluster (Saville 1993). What constituted a flower unit was defined separately for each plant species (but consistently across all observers), based on floral morphology and on observations of bee foraging behavior. Plant species were identified using a field guide (Newcomb 1989) or were photographed for later identification.

To evaluate the landscape around each site, we used remotely-sensed data to calculate the proportion of forest (deciduous and conifer), shrub, fields (alfalfa, clover, other hay, idle cropland, grassland, pasture, and developed open space), developed areas (low, medium, and high intensity), wetlands (woody and emergent herbaceous), pollinator-supporting crops (fruits and vegetables that flower before harvest), and non-pollinator-supporting crops (grains, tree crops, sod-grass, and other crops that are harvested before flowering) using ArcGIS software (ESRI 2016). This simplified land cover classification scheme was modified from the USDA Cropland Data Layer (CDL 2018 version; USDA 2019). We considered landscape composition at three spatial scales—a 1.0-km, 1.5-km, and 2.0-km buffer radius areas from the geographic site centroid—based on bumble bee foraging distances reported in the literature (Dramstad 1996; Osborne et al. 1999; Walther-Hellwig and Frankl 2000a, b; Wood et al. 2015). We used generalized linear models and AICc values to determine that the 2.0-km landscape scale better reflected variation in Bombus abundance and richness.

We restricted our flower richness and abundance measures to only include flower species that were visited by at least one bee during our surveys. For most analyses, we grouped two morphologically similar bumble bee species—B. auricomus and B. pensylvanicus—because not all were distinguished in the field, and they accounted for a low proportion of observations in any given survey. Two rare species in our dataset—B. citrinus (3 individuals observed) and B. borealis (a single individual)—were retained for abundance and richness analyses but omitted from community ordination to avoid their disproportionate influence on pairwise distances among sites and ordination configuration (McCune et al. 2002). Abundance data were natural log transformed as needed to improve their fit to a Gaussian distribution. To assess differences in bee diversity between sites, independently of abundance, data were rarefied down to 19 bumble bees per survey and the average species richness of 1000 random bootstrap replicates was calculated (rrarefy function, ‘vegan’ package). Fifty surveys in which fewer than 20 bumble bees were found were excluded from rarefaction analysis. We were able to visit a minority of sites twice, but they were visited in different years and different times, so due to high background variation in bee survey yield they were treated as independent observations. Most analyses were conducted using R, version 3.5.3 (R Development Core team 2019). JMP software version 14 (2018) was used to construct generalized linear models with logit link functions on presence/absence data for rarer species.

We used multiple regression models to evaluate how habitat, flower resources, landscape, and time of season affected overall bumble bee abundance and diversity in our surveys. To investigate whether common species responded differently to environmental features, models were also constructed separately for each of the three most abundant species. Models of bumble bee abundance and raw and rarefied species richness were fitted with the following potential main effects: growing degree day (GDD; determined using the online calculator at https://​www.​oardc.​ohio-state.​edu/​gdd/​default.​asp), latitude, flower abundance, flowering species richness, habitat type, and the proportion of forest (hereafter “forest”) and urban developed area (hereafter “developed”) in the landscape. We also included the two-way interaction terms between habitat and all local and landscape level variables that have been shown to modify the quality of a specific habitat for bumble bees (e.g., Ahrné et al. 2009, Cariveau and Winfree 2015; Carvell et al. 2011; Heard et al. 2007; Hines 2005; Lanterman et al 2019; Reeher et al. 2020; Schochet et al. 2016), as well as GDD, which helps account for phenological effects on bumble abundance and richness (Lanterman et al 2019). Finally, we included a flower abundance x flower richness interaction to help account for the overall availability of flower resources. A gaussian distribution was selected for each response variable based on the model adjusted R² values under different distributions and Shapiro tests for normality of model residuals. Prior to model construction, we checked for collinearity of continuous predictor variables by calculating a Pearson correlation matrix. There was a moderate negative correlation between developed land and forest (Pearson r = -0.53); all others were less than 0.5. Continuous predictor variables were standardized and centered by subtracting the mean from each observation then dividing by the standard deviation. For each model, we calculated Variance Inflation Factors (VIFs) for each independent variable and found no evidence of multicollinearity. To compare the amount of variance explained by each factor or interaction in our models, we calculated their partial eta-squared values (Richardson 2011). We used the Tukey procedure for pairwise contrasts of mean abundance and richness between all habitat types within the model framework (package ‘lsmeans;’ Lenth 2016). Complete model summaries are given in the Electronic Supplementary Material, Tables S2–S7.

We investigated the factors influencing the probability of occurrence of the less common species (B. fervidus, B. vagans, B. perplexus, B. auricomus, and B. pensylvanicus) using generalized linear models with a logit link function on presence versus absence data. We focused these analyses on broad-scale factors because we suspected that the presence/absence data would be less sensitive than abundance data to snapshot measures of local environmental quality, such as flower abundance and richness. Therefore, for each of these five species we analyzed its presence versus absence as a function of habitat type, landscape context, latitude, and their first-order interactions. We used the Firth adjusted maximum likelihood method to estimate parameter values. Non-significant terms were dropped from models sequentially, starting with the lowest likelihood ratio chi square value, and each reduced model was compared to the previous using AICc. For each species we retained the model that had the most independent variables and an AICc at least 2 lower than the next best fit model. There were not enough sightings of B. citrinus or B. borealis to analyze.

To visualize bumble bee community composition, we used non-metric multidimensional scaling (NMDS) ordination (metaMDS function, ‘vegan’ package; Oksanen et al. 2016). We tested for differences in community composition by habitat type and month using permutational analysis of variance adonis tests (adonis function, 500 permutations, ‘vegan’ package) based on Bray–Curtis dissimilarity between sites. To visualize the influence of landscape and other continuous variables on bee community composition, we fitted significant predictor vectors to the ordination plot using the envfit function in the ‘vegan’ package.

Indicator species analysis was used to determine if bumble bee species were associated with particular habitat types (indval function, ‘labdsv’ package, 500 iterations; Roberts 2016). An indicator value was calculated for each species—habitat combination as the product of the relative frequency and relative average abundance of a species at sites of a particular habitat type, taking into account the number of sites in that habitat group.

To visualize differences in habitat use by each bumble bee species, we standardized abundance values for the June-July samples by dividing the mean abundance per survey by the maximum value for that species across habitats. These values were plotted using radar charts. Bombus auricomus and pensylvanicus were plotted separately for this analysis using the subset of individuals from the B. auricomus/pensylvaniucs morphotype that we had identified to species with confidence.

Sites intentionally planted with flowers (seeded hayfields, restored meadows, restored roadsides, and planted urban patches) consistently yielded high numbers of bumble bees (planted 85 ± 68, n = 142; unplanted 48 ± 42, n = 108) and more than four species (planted 4.3 ± 1.2, unplanted 3.7 ± 1.2). Bumble bee abundance per survey in June and July was highest in hayfields (115 ± 74, n = 7), and lowest in early successional natural fields (40 ± 31, n = 52). Planted urban patches also attracted large numbers of bumble bees (89 ± 102, n = 13) and had relatively high species richness (4.5 ± 1.0).

Total bumble bee abundance varied widely (from 2 to 377 in June/July surveys) and responded strongly to habitat in multiple regression models. Landscape factors and habitat by latitude interactions explained small, but significant amounts of the variation in abundance (Table 2). Bumble bee abundance generally increased with the proportion of developed area and forest in the landscape (Electronic Supplementary Material Table S2), although species varied in their relationship with forest (see individual species analyses below). Habitat effects interacted with latitude such that bumble bee abundance increased with latitude in planted meadows, roadsides and urban patches (Table S2). Bumble bee abundance trended higher in sites with abundant flowers, but not significantly (Table 2). Note, there were no consistent differences between habitat types in flower abundance (F = 0.99, df = 6, 246, p = 0.43) or in flowering species richness (F = 0.78, df = 6, 246, p = 0.59).

We found between one and seven bumble bee species per survey (mean 4.0 ± 1.2). Raw bumble bee species richness varied with habitat and the interaction between habitat and the amount of forest in the landscape (Table 2, Electronic Supplementary Material Table S3). In particular, Bombus species richness was greater in planted meadows than in early successional natural fields (according to Tukey’s pairwise contrasts of abundance by habitat within the model framework, t = 3.20, df = 198, p = 0.03). Rarified bumble bee richness (standardized to 19 individuals per site) declined with flower richness but did not differ among habitat types (Table 2).

Bumble bee community composition varied with landscape setting (Fig. 2), and certain landscape factors emerged as important determinants of species’ abundances (see individual species analyses below). In general, species assemblage was influenced by nearby forest (envfit vector correlation with NMDS axis 1 and 2 scores, r² = 0.07, p < 0.01), developed area (r² = 0.03, p = 0.04), and non-pollinator-supporting crops in the landscape (r² = 0.03, p = 0.02), in addition to GDD (r² = 0.13, p < 0.01). Although significant, the r² values indicating correlation between landscape components and bumble bee community composition were very low.

Community composition was also significantly, but weakly influenced by habitat type (adonis test F = 2.51, R² = 0.06, df = 6, 246, p < 0.01; Fig. 2). We used radar plots to visualize more clearly the differences in habitat affiliations between species. Almost all species reached their maximum abundance in one of the planted habitats (Fig. 3), while only one reached maximum abundance in an unplanted habitat type (B. pensylvanicus on roadsides). Planted hayfields harbored the greatest abundances of three of the most common species, while several of the less common species were frequently encountered in planted roadsides (Fig. 3). However, those were some of the least abundant habitat types (7 planted hayfields and 11 planted roadsides sampled in June/July), so the results should be interpreted with caution. According to indicator species analysis, Bombus bimaculatus (iv = 0.61, freq = 216, p < 0.01) and B. griseocollis (iv = 0.65, freq = 240, p < 0.01) were associated with habitats planted with wildflowers, rather than unplanted fields and roadsides (Fig. 3a). Bombus perplexus was associated with urban planted patches (iv = 0.22, freq = 45, p = 0.03; Fig. 3b), and Bombus auricomus with planted roadsides (iv = 0.23, freq = 34, p = 0.02; Fig. 3b). There were no significant habitat associations for B. pensylvanicus, likely due to low sample size (n = 28 individuals identified with certainty). Shrubby and early successional natural field habitats were not favored by any bumble bee species.

Species-specific analyses of abundance of the three most common bumble bee species (B. impatiens, B. griseocollis, and B. bimaculatus) reveal similarities and differences in their responses to local, landscape, and geographic variables (Table 3). In multiple regression models there was a significant effect of habitat type on the abundances of all three species, primarily due to their high numbers in planted meadows and roadsides habitats (Table 3, Electronic Supplementary Materials Tables S5–S7). In models of Bombus impatiens abundance the effect of habitat interacted with latitude, increasing with latitude in planted meadows and urban patches. However, there were more urban patches available to sample in the northern part of the state than in the southern. Bombus griseocollis and B. bimaculatus were both significantly more abundant in planted meadows than in early successional unplanted fields (according to Tukey’s pairwise contrasts of abundance by habitat within the model framework: B. griseocollis t = − 3.98, df = 198, p < 0.01; B. bimaculatus t = − 5.02, df = 198, p < 0.01). As for landscape scale and geographic factors, Bombus impatiens tended to increase with the proportion of developed land and forest in the landscape (Table 3, Electronic Supplementary Material, Table S5). Bombus griseocollis abundance also increased with percent developed land, but not with forest (Table 3, Table S6). Bombus bimaculatus abundance generally declined with GDD and latitude (Table 3, Table S7). However, this species did increase in abundance with latitude in certain managed habitats (significant latitude x habitat interaction for planted meadows, roadsides, and urban patches; Electronic Supplementary Materials, Table S7).

Bombus bimaculatus and B. griseocollis abundances were weakly positively correlated with one another on a site by site basis (June and July surveys only, Pearson correlation r = 0.12, t = 1.90, df = 228, p = 0.06). Bombus impatiens abundance was weakly positively correlated with that of B. griseocollis (r = 0.16, t = 2.53, df = 239, p = 0.01), but not with B. bimaculatus (r = 0.06, t = 0.88, df = 231, p = 0.38), likely because of the earlier and shorter period of activity of the latter species.

The probability of site occupancy for the less common species (based on presence/absence data) showed that three species, B. fervidus, B. perplexus, and B. vagans, were more likely to be found at sites located in more forested landscapes (Table 4, Fig. 4) and at higher latitudes (Table 4, Fig. 5a–c). The positive effect of forest was especially strong for B. vagans (Fig. 4). The response to forest by B. fervidus varied with latitude; in northern latitudes the probability of occurrence increased strongly with forest, but in southern latitudes it declined slightly with the amount of forest in the landscape, despite similar ranges of percent forest across latitudes (Fig. 4 inset). In contrast, the predicted probability of B. pensylvanicus occurrence was negatively associated with latitude (Table 4, Fig. 5d) and forest (Table 4, Fig. 4), indicating that it was more common in open landscapes in the southern part of the state. The occurrence of B. auricomus, a species often reported to have a similar niche to the rarer B. pensylvanicus (Hobbs 1965; Liczner and Colla 2020), was not predicted by the amount of forest in the landscape. Occurrence of only one species was significantly predicted by habitat type (planted roadside for B. auricomus). The three most common species, Bombus impatiens, B. griseocollis, and B. bimaculatus, were widely distributed and did not show strong geographic patterns in occurrence (see Electronic Supplementary Material Fig. S1 for distribution maps).

The predominance of Bombus impatiens was consistent across all habitats and increased over the season. Other studies of bumble bee communities in eastern North America have found similar dominance of B. impatiens and suggest that this species is increasing in relative abundance compared to other species (Colla and Packer 2008; Cameron et al. 2011; Jacobson et al. 2018; Richardson et al. 2019). No clear explanation for the success of B. impatiens has surfaced, though our results suggest that its highly general habitat use and long season of activity help it maintain large populations in many landscape contexts. In turn, large populations likely buffer this species from loss of genetic diversity and demographic stochasticity that can lead to population decline (Herrmann et al. 2007; Dreier et al. 2014). High diversity in worker size within B. impatiens nests may also allow colonies to respond more efficiently than some other species to environmental changes (Austin and Dunlap 2019), like fluctuations in food availability. Bombus griseocollis and B. bimaculatus were also consistently present and abundant across our surveys. All three of these potential competitors were abundant in the same habitats, raising the question of what subtle differences in their biology allow them to coexist. Our findings demonstrate that these three common Bombus species respond differently to habitat management and landscape. Bombus impatiens showed fewer distinct responses than the other two species (Fig. 2), seeming to flourish in almost all habitats. In contrast, Bombus griseocollis responded positively to forested landscapes in planted meadow or shrubby habitats, and showed GDD-specific increases in planted meadow habitats (perhaps where its favored milkweed food plants (Villalona et al. 2020) were abundant). Bombus bimaculatus responded differently, with an increase at high latitudes, varied responses to GDD-habitat combinations, and an earlier peak in abundance. Although there are certainly broad similarities in these three dominant species, recognizing their differences will be necessary for effective conservation of bumble bee communities. When generalizing about the response of a taxonomic group to habitat quality, teasing apart species-specific effects helps to avoid attributing to all species patterns that are driven primarily by the dominant species. Analysis of species-specific foraging preferences may reveal further differences in their niches.

Several species of bumble bee that were found frequently in our survey of Ohio are thought to be declining or rare in other areas. Two of these, B. vagans and B. fervidus were each found in > 37% of our surveys (for reports of decline elsewhere see Grixti et al. 2009; Williams et al. 2009; Colla and Packer 2008; Colla et al. 2012; Bartomeus et al. 2013; Bushmann and Drummond 2015; Jacobson et al. 2018). On the other hand, our results agree with those of others that B. perplexus is relatively uncommon (Colla and Packer 2008; Williams et al. 2009), found in only 21% of surveys. Due to the lack of rigorous historical datasets on bee abundance and diversity in Ohio, we were not able to investigate temporal trends in abundance as has been done in other recent literature (Colla and Packer 2008; Grixti et al. 2009; Cameron et al. 2011; Colla et al. 2012; Jacobson et al. 2018; Richardson et al. 2019; Wood et al. 2019). Though the stability of various species’ populations is outside of the scope of this paper our results provide species-specific abundance and distribution data that could be useful for future population stability analyses. In places that lack historical data for comparison, such as Ohio, what we can do is focus on how current bee distributions reflect current environments and use these data to inform conservation strategies.

Although our survey was extensive, we did not have enough observations of the rarer species to evaluate confidently possible influences on their abundance. This result is sobering considering the scope of our study; we had > 450 h of observation over two years, which provided > 23,000 bumble bee sightings, a level of sampling that is expensive and uncommon. It is not clear whether rarity of these species is consistent with historical patterns, reflects recent declines in their abundance, or is especially apparent because of increased abundances of the three most common species. Nonetheless, we were able to detect marked patterns regarding occurrence (presence/absence) and biogeography of the rarer bumble bees. Bombus perplexus, which favored forested landscapes and the northern part of the state, was associated with urban wildflower plantings. Bombus vagans and B. fervidus were also more common in forested landscapes and at northern latitudes, but not associated with urban patches. A species of conservation concern, B. pensylvanicus (MacPhail et al. 2020; Richardson et al. 2019), that is affiliated with grasslands in our and other studies (Colla and Dumesh 2010; Williams et al. 2014), was more frequently found in the southern part of the state. Therefore, conserving or enhancing grassland habitat with flowering forbs may be enough in the southern part of the state to support diversity, but in the north habitat improvements near forests may be particularly effective.

One of the motivations for this survey was a desire to document the current distribution and abundance of rare species, especially Bombus affinis and B. terricola. The former species was once common in Ohio and was recently listed as federally endangered in the USA, as a result of precipitous declines after about the year 1999 (Colla and Packer 2008; Grixti et al. 2009; Williams et al. 2009; Cameron et al. 2011; Colla et al. 2012; Jacobson et al. 2018). Our results do not provide evidence that either of these species is now surviving in Ohio. However, B. affinis has been observed in low abundance in the past several years in the nearby states of Iowa, Illinois, Minnesota, and Wisconsin and in isolated mountain meadows in Virginia and West Virginia (S. Droege, personal communication; P. Reeher, personal observations in Virginia). Therefore, further monitoring may yet reveal its continued presence or recolonization in Ohio.