In ecosystems impacted by anthropogenic disturbance, honey bees may help fill the pollination void left by declines in non-honey bee pollinators . Lastly, where honey bees reach high densities, as reported in some areas of their introduced range , they may exploit enough food resources to compete with other pollinators . These phenomena are of broad ecological, evolutionary, and conservation interest, but to our knowledge, there currently exists no quantitative synthesis on the numerical importance of honey bees as floral visitors in natural ecosystems worldwide, either in their native or introduced range.Here, we use a meta-analysis to address the question of honey bee importance by taking advantage of a recent trend in pollination research—the documentation of community-level plant-pollinator interaction networks . Pollination network studies match the identity and frequency of each type of pollinator visiting each plant species within a locality . While these studies are performed to investigate a variety of questions , data from pollination networks provide an excellent opportunity to investigate the importance of honey bees in natural habitats, not the least because the role of honey bees has rarely been their focus . Here, we compile a database of 80 pollination networks from natural and semi-natural habitats from all continents except Antarctica, as well as several oceanic islands, including regions where honey bees are native and places where they have been introduced. These networks allow us to address four questions regarding the ecological importance of honey bees in natural habitats. What proportion of floral visits in natural habitats are due to honey bees? Do honey bees reach higher numerical dominance in their non-native range? How are honey bee visits distributed among plant species?

For instance, plant plastic pot what proportion of plant species is not visited by honey bees, and for what proportion do honey bees contribute the majority of visits? Finally, network studies often use visitation frequency as a proxy for pollinator importance . To further assess the value of honey bees as pollinators, we compile data on per-visit pollination efficiency of honey bees relative to other floral visitors from studies on 35 plant species. Using these data, we address a fifth research question: How does the per-visit pollination efficiency of honey bees compare to the average non-honey bee pollinator?We used two approaches to compile our data set of pollination networks. First, we performed a literature search using the ISI Web of Science database with the search terms [pollinat* network], [pollinat* web], and [pollinat* visit* community] from October 2014 to August 2016. Second, we downloaded pollination network data from the Interaction Web DataBase of the National Center for Ecological Analysis and Synthesis website and the Web of Life Ecological Networks Database . From the latter two databases, we downloaded all plant-pollinator interaction network datasets available as of December 2014 that reported visitation frequency in addition to the presence / absence of interaction between plant and pollinator taxa. Each data point in our study consists of a weighted pollination network in which the set of interactions between each plant and pollinator pair is weighted by a measure of visitation frequency . We defined a network as the sum of recorded plant-pollinator interactions in all study sites from a single study that fell within a 50-km diameter circle, regardless of the number of study plots that constitute the network. Sites within the same study that are separated by more than 50 km were treated as separate networks. When we encountered networks from different studies that were less than 50km apart, we excluded those studies that sampled a smaller number of plants or pollinators, or documented fewer interactions.

All networks retained for analyses met the following criteria. The data were collected in natural or semi-natural habitats; agricultural, urban, experimental, or otherwise managed habitats were excluded. Each included network consisted of observations on five or more plant species when pooled across study sites; networks that focused only on select plant taxa with specialist pollination syndromes were excluded from analyses. Included networks documented a broad range of pollinators; studies that had a narrow taxonomic scope or that explicitly excluded honey bees from data recording were excluded. Because we are primarily interested in quantifying the importance of honey bees in natural areas free of human interference, we excluded data from study sites that are known to be heavily influenced by honey bee colonies stocked for adjacent agricultural pollination. Thus, our estimates of honey bee numerical importance may be conservative with respect to mosaic landscapes where natural habitats are intermixed with agriculture, but achieve a closer representation of the role of honey bees in natural areas worldwide, overall. We also did not exclude networks from localities outside of the honey bee’s climatic niche, or where honey bees have never been introduced. In all, we obtained 80 networks from 60 peer-reviewed studies, two graduate theses , and our own study of plant-pollinator interactions in San Diego’s scrub habitats .For each network, we obtained the following data from their associated publications or from study authors when data were not available from publications: latitude, longitude, and final year of data collection. When these data were not available and authors could not be reached, we used the approximate geographical center of the study locality listed in the publication, and the year of publication as the last year of data collection.

We defined the native status of honey bees based on ; in Great Britain , where the native status of honey bees is uncertain, we treated honey bees as native rather than introduced, but classifying honey bees there as introduced in that location did not substantially alter our results. We also extracted the following information from each study, when available: the proportion of total floral visits contributed by honey bees, the proportion of plant species receiving at least one visit by honey bees, and the rank of honey bees with respect to both the total number of interactions and the proportion of plant species visited. Additionally, we used geographic information system analysis to obtain elevation data and bioclimatic variables for each network based on its GPS coordinates. We also assigned each network as being on an island or a mainland; the latter category includes all continents as well as large islands > 200,000 km2 , namely Great Britain , Honshu , and Greenland. For relevant studies for which raw data were not available, we contacted the corresponding authors to request data, or, in cases where data could not be shared, requested summary statistics on plant-pollinator interactions. When raw numeric data were unavailable from the publication or from authors, we used ImageJ to extract data from figures, where possible . Due to the different methodologies and data-reporting requirements of each study, not all of the above mentioned variables were extracted from all networks. Comparison of honey bees and bumble bees in pollination networks: Because studies vary in the level of detail with which individual species of floral visitors other than Apis mellifera are reported, we cannot reasonably compare frequencies of honey bee visitation with those of other single species across all of our networks. However, data are sufficiently detailed in 66 of our 80 networks to enable comparison of honey bees and bumble bees ; the latter are the only other pollinator group with a similar pattern of local numerical abundance and widespread introduction compared to honey bees . We compared the network-level relative visitation frequency of honey bees with that of all bumble bee species combined using a paired t-test. Since our goal was to compare global patterns of numerical importance, this analysis did not exclude networks in which honey bees, bumble bees, or both taxa were absent. It is worth noting that the leaf cutter bee Megachile rotundata , another widely introduced pollinator , was not reported in any of our 80 networks. Drivers of honey bee visitation frequency among pollination networks worldwide: We used multiple linear regression models to examine environmental factors that may contribute to variation in the network-level frequency of floral visits by honey bees. Networks where honey bees were not recorded were excluded from this analysis because of the variety of reasons that could explain their absence, black plastic plant pots ranging from studies that were outside the geographical or altitudinal range of the honey bee , to studies where honey bees were undetected despite being present in the ecosystem .

Inclusion of networks that documented no honey bee visits using a zero-inflated multiple beta regression model in Program R v.3.3.1 did not qualitatively alter our results . The response variable in these regression models was the proportion of all floral visits in each network contributed by honey bees, logit-transformed to improve normality . To identify the environmental model that best explains network-level honey bee visitation frequency, we generated models containing all possible combinations of the following explanatory variables : latitude, longitude, altitude, land category , and bioclimatic variables relating to temperature and precipitation . To incorporate bioclimatic variables, we first performed Principal Components Analysis to avoid constructing models with highly collinear terms. We performed one PCA for the 11 variables measuring temperature , and a separate PCA for the eight bioclimatic variables measuring precipitation ; these analyses enabled us to reduce bioclimatic variables to the first two principal components of the temperature variables and the first two principal components of the precipitation variables . We used R package glmulti to generate the candidate models and to select the best model using corrected Akaike’s Information Criterion scores. We also used the resulting “best” environmental model to address the questions of whether or not the network-level frequency of honey bee visits depends on their native status and the year of data collection, by adding these two variables to the “best” environmental model, both individually and together. Distribution of honey bee visitation frequency across plant species: We examined the distribution of honey bee relative visitation frequency across plant species as measured by the proportion of visits to each plant species contributed by honey bees. In this analysis, we included 46 networks in which at least one visit by a honey bee was recorded, and data on the proportion of total visits contributed by honey bees were available for each studied plant species. We pooled all plant species from all networks, and did not correct for cases in which the same plant species occurs in more than one network. Given the breadth of geographical areas and ecological contexts represented by networks in our study, the same plant species is expected to be served by different pollinator assemblages in distinct networks. Because plant species receiving few visits overall may tend to have extreme values of proportion of visits by honey bees, we also repeated this analysis after restricting the dataset to plant species with 10 visits recorded. Pollination efficiency of honey bees: We used two approaches to compile our data set. Second, we examined the literature cited sections of each of the studies found through the first approach for additional studies that were not captured in the literature search. Data points in this analysis consist of studies of focal plant species that compared honey bees and at least one other pollinator taxon with respect to pollen deposition, seed set, or fruit set resulting from a single visit by an individual floral visitor . In a small number of cases, we used ImageJ to extract data from figures when raw data were not available. In all, we obtained 33 studies reporting single-visit pollination efficiency data for 35 plant species, spanning 23 plant families . Of these, 19 plant species in 16 families were undomesticated, and 16 plant species in 7 families were grown in agricultural settings. Multiple metrics of per-visit efficiency were available from some studies. We used or calculated seed set data whenever available since it is the most closely related to plant reproductive fitness , fruit set when no seed counts were available, and pollen deposition when measures of seed and fruit set were unavailable. For each plant species in each study, we calculated the average single-visit pollination efficiency of non-honey bee pollinators as the numerical mean efficiency metric of all non-honey bee visitors studied. Then, we calculated the relative single-visit pollination efficiency of honey bees by dividing honey bee pollination efficiency by the average efficiency of non-honey bee floral visitors studied.