The zones of tension between Africanized honey bee and European honey bees based on + 10% of the threshold from the binary maps of both taxa. 

Context in source publication

Context 1

... correlated (r ≥ 0.70 or ≤ -0.70), variables with lower biologic relevance were dropped (following Jarnevich et al. [24]). Each of the four statistical models began with the same set of environmental variables (Table 1); however, each algorithm except random forest incorporated variable selection, therefore the final set of variables used in each model was slightly different. The GLM employed standard stepwise regression using Akaike’s information criterion. We used 10 fold cross-validation to assess model performance. After all four statistical models were run for European and Africanized honey bees, we combined the SAHM [30] continuous probability surface outputs in ArcGIS v. 10.1 to produce ensemble maps of the predictions. The sensitivity equals specificity threshold for each model was used to classify each grid cell as suitable or unsuitable [40]. This threshold rule ensures presence locations are equally likely to be false positives as absence locations. The resulting binary maps for each model were then added, resulting in a map that displayed areas where one, two, three, or all four models agree [24]. Finally, to evaluate the “tension zone”, where habitat suitability for European and Africanized honey bees currently overlap, we reclassified the probability surface for each model into a binary map using (threshold value +/- 0.10), combined the four models into an ensemble for each lineage, and then further combined the two ensemble maps into a new ensemble where values of zero, one, or two (number of models in agreement, respectively) equal no tension and values of three or four equal tension where at least three of the statistical models predict European and Africanized honey bees overlap in habitat suitability. Second, mitochondrial DNA analysis on 89 Africanized bees allowed us to model the potential distribution of 72 occurrences of the A1 mitotype (common in Africa, Brazil, and Mexico), and 17 occurrences of the A26 mitotype (common in SW Africa, but rare in the Americas), using the same Maxent approach that we used for Tamarix modeling above, and, to avoid over-fitting the model, we again selected only two variables (minimum temperature of coldest quarter and season length) as predictors. We used other AHB locations as background following the target background approach [41]. These environmental variables have been shown to be important drivers of Africanized bee distribution [24]. Similar to models described above, we used the maximum of sensitivity plus specificity as the threshold for binary classifications. In all our models, we recognized that our results may have been affected by small sample size, but Maxent has worked well with small sample sizes [42, 43]. Our results also may be affected by sample location bias (a point we return to in the discussion). Regardless of the species, we used the area under the receiver-operating characteristic (ROC) curve (AUC) to assess model accuracy with each validation dataset. AUC, a measure of model fit, has been widely used in several model comparison studies. Usually AUC values of >0.9 indicate high accuracy, values of 0.7–0.9 indicate good accuracy, and values 0.5 (random) to 0.7, indicate low accuracy. We minimized the potential problems associated with the reliance on AUC values [44] by: (1) selecting models appropriate for the data (e.g., presence or presence-absence models); (2) selecting only models that performed well in previous model comparison studies [45]; and (3) eliminating one of each pair of highly cross-correlated predictor variables prior to modeling (i.e., in the honey bee models). We also used percent correctly classified (threshold dependent metric where sensitivity equals specificity) as an indication of model performance. Preliminary Maxent models were reasonably strong (test AUC > 0.83) for Tamarx ramosissima , T. chinensis , and the hybrid T. ramosissima x T. chinensis , despite the modest sample size of about 30 observations per taxa (Fig. 1). We were surprised that the extent of the potential habitat suitability for the hybrid T. ramosissima x T. chinensis hybrid was slightly greater than the parent taxa ( T. ramosissima or T. chinensis; Fig. 1). Selected response curves from the Maxent models for minimum temperature of the coldest month and annual precipitation showed that the hybrid T. ramosissima x T. chinensis response curves were intermediate between T. ramosissima and T. chinensis (Fig. 2), describing a more general response to the environmental variables. All statistical models produced test AUC values greater than 0.85 and correctly classified suitable habitat greater than 74%, indicating strong performance (Table 1). The models show that Africanized honey bee and European honey bees have diametrically opposite areas of habitat suitability and occurrence (Fig. 3). We overlaid binary probability maps of the Africanized honey bee and European honey bee distributions to examine the zones of tension in the two taxa (Fig. 4). Each map was calculated by reclassifying the continuous model prediction to be ±10% around the binary threshold value for each model, which identifies the areas of moderate suitability. The western Great Basin, western Great Plains, and southern Appalachian mountains are regions highlighted by these models, where the European honey bees may be struggling to maintain dominance. Maxent models of the two dominant mitotypes of the Africanized honey bees were the strongest models of this study (AUC ~0.90; Fig. 5). Selected response curves for minimum temperature of coldest month and annual precipitation suggest that the two mitotypes may respond differently to environmental drivers. The A26 mitotype is associated with slightly cooler and drier sites compared to the more common A1 mitotype (Fig. 5) We realize that we have small datasets for the Tamarix genotypes, and for the mitotypes of the Africanized honey bees, due largely to the high costs of obtaining genetic data. Small sample size and sample location bias may have affected our model results, but sample sizes as low as n=10 with Maxent performed well [42, 43]. Our purpose here is to demonstrate the utility of modeling hybrid swarms and wildly introgressing populations to provide preliminary risk maps as “mapped hypotheses” to be improved with additional field work and modeling [22]. Since the various genotypes of Tamarix and mitotypes of bees were modeled in similar ways, we think they serve our purpose.. Our preliminary investigation of Tamarix hybrids and Africanized honey bees yielded a similar, frightening result: at least some hybrids may be highly invasive by: (1) responding differently to environmental drivers than the parent taxa (Figs. 2 and 5); or (2) occupying slightly different habitats than the parent taxa (Figs. 1 and 5). This complementarity of taxa distributions may contribute to the spread, extent, and persistence of hybrid taxa. Other mechanisms may also be important for hybrid persistence such as avoiding pathogens and other adaptations [46]. We provided two recent examples of hybrid swarms that are not without conservation concerns. Friedman et al. [47] and Gaskin and Kazmer [29] found that the F1, F2, and backcrosses to two parent species, T. ramosissima and T. chinensis , may represent 83–87% of the tamarisk invasion in the western United States. Recent studies have found that this complicates the biological control of the genus, because increased levels of T. ramosissima introgression resulted in higher investment in roots and tolerance to defoliation, and less resistance to biological control agents often used by managers to decrease the range and spread of this genus [48]. Plant hybridization, particularly in this harmful invasive species, may hinder containment efforts in natural areas. Likewise, the hybridized Africanized honeybees may have many evolutionary advantages over European honeybees and native bees. The Africanized honeybees have faster growth rates than European honeybees, resulting in more swarms to dominate an area [49]. And, since European honeybee queen mate disproportionately with African drones, this is likely to cause rapid displacement of European honeybee genes in a colony. For example, when a queen is inseminated with an equal portion of African drone semen and European honeybee semen, “the queens preferentially use the African semen first to produce the next generation of workers and drones, sometimes at a ratio as high as 90 to 10” [49]. According to bee expert, David Roubik [50], Africanized honeybees “groom themselves more often than Italian bees, making them less likely to get sick from mites and other parasites” that plague European honeybees. In South America, the Africanized honeybees withstand broader environmental gradients, from desert to rainforest. In the more extreme environments “the European bee will just starve to death, the Africanized bee is foraging, bringing things in, and getting by,” said Roubik. Combined with their aggressive behavior, these traits and characteristics may explain why Africanized honeybees are displacing native bees in Mexico and elsewhere [51]. We provided two recent examples of hybrid swarm incidents from a growing literature of examples. Hybrid vigor has been demonstrated for the invasive alien Brazilian peppertree ( Schinus terebinthifolius Raddi., Anacardiaceae) in Florida [52]. Vilà and D’antonio [53] demonstrated hybrid vigor for clonal growth in an invasive alien Carpobrotus (Aizoaceae) in coastal California. The native California Tiger Salamanders ( Ambystoma californiense ) hybridized with the introduced Barred Tiger Salamanders ( Ambystoma tigrinum mavortium ) resulting in a more robust hybrid [54]. As global trade increases, we expect many more closely related species to come in contact [55], setting the stage for swarms of hybrids and rapid evolution. We echo the concerns of Lee [56] more than a decade ago that conservationists should ...