Journal list menu

Volume 14, Issue 6 e4588
Open Access

Combining stable isotopes, trace elements, and distribution models to assess the geographic origins of migratory bats

Jamin G. Wieringa

Corresponding Author

Jamin G. Wieringa

Department of Evolution, Ecology and Organismal Biology, The Ohio State University, Columbus, Ohio, USA

Ohio Biodiversity Conservation Partnership, Columbus, Ohio, USA


Jamin G. Wieringa

Email: [email protected]

Search for more papers by this author
Juliet Nagel

Juliet Nagel

Appalachian Lab, University of Maryland Center for Environmental Science, Frostburg, Maryland, USA

Search for more papers by this author
C. J. Campbell

C. J. Campbell

Department of Biology, University of Florida, Gainesville, Florida, USA

Search for more papers by this author
David M. Nelson

David M. Nelson

Appalachian Lab, University of Maryland Center for Environmental Science, Frostburg, Maryland, USA

Search for more papers by this author
Bryan C. Carstens

Bryan C. Carstens

Department of Evolution, Ecology and Organismal Biology, The Ohio State University, Columbus, Ohio, USA

Search for more papers by this author
H. Lisle Gibbs

H. Lisle Gibbs

Department of Evolution, Ecology and Organismal Biology, The Ohio State University, Columbus, Ohio, USA

Ohio Biodiversity Conservation Partnership, Columbus, Ohio, USA

Search for more papers by this author
First published: 20 June 2023
Handling Editor: Brooke Maslo


The expansion of industrial-scale wind-energy facilities has not only increased the production of low-carbon emission energy but has also resulted in mortality of wildlife, including migratory bats. Management decisions can be limited by a lack of understanding of the geographic impact of bats killed at wind-energy facilities. Several studies have leveraged stable hydrogen isotope ratios (δ2H) of bat fur to illuminate this issue but are limited in the precision of conclusion because δ2H values vary primarily across latitudinal and elevational bands. One approach to increase the precision of geographic assignment is to combine independent inferences about spatial location from additional biomarkers and other related information. To test this possibility, we assigned known-origin individuals of three bat species (Lasiurus borealis, L. cinereus, and Lasionycteris noctivagans) commonly killed at on-shore wind-energy facilities in North America to a probable origin using δ2H values, trace element concentrations, and species distribution models. We used cross-validated calibrated combined model tuning to determine the degree to which assignment probabilities improved when combining datasets. We found that combining markers typically performed better than single approaches. For Lasiurus borealis and L. cinereus, combining all three data sources outperformed any single or other combined approach. With an accuracy set at 80%, an average of 39.7% and 36.0% of each species' total geographic range was considered a potential origin, respectively; stable hydrogen alone included 51.8% and 50.6% of the total geographic area. In contrast, for Lasionycteris noctivagans, including trace elements did not increase precision and adding distribution data to δ2H values only improved precision by 0.6%. Thus, we found that a combination of multiple biomarkers typically, but not always, outperforms single-marker approaches and optimized combinations of different markers outperform equal weighting of each marker. From a practical perspective, δ2H values performed better than trace elements alone; in cases where cost is a limiting factor, the stable hydrogen should be the single biomarker used in conjunction with species distribution models. Overall, these results highlight the importance of validating methods for each species they are applied to and show that combining information from intrinsic biomarker approaches is a useful tool to document bat movements.


An emerging threat to the populations of several species of migratory bat is the expansion of wind-energy generation (Frick et al., 2020). In North America, three species of migratory bats (Lasiurus borealis, L. cinereus, and Lasionycteris noctivagans) comprise up to 80% of bats killed at all on-shore wind-energy facilities (Arnett & Baerwald, 2013). Without a drastic reduction in these mortality rates, there is a potential for 90% decline in Lasiurus cinereus population numbers within 50 years, with similar population pressures likely affecting L. borealis and Lasionycteris noctivagans (Davy et al., 2021; Frick et al., 2017; Friedenberg & Frick, 2021; Rodhouse et al., 2019). The risk to migratory bats posed by wind-energy facilities highlights the need for an improved understanding of bat migration and, more specifically, the seasonal origins of individuals dying at wind-energy facilities. Understanding whether the impacts of wind turbines might be concentrated locally or be spread over a broad geographic range is essential to predicting population-level impacts and prioritizing the geographic areas in which to target conservation actions (Cryan & Barclay, 2009; Voigt et al., 2012). For example, if most mortalities represent bats that reproduce and live locally, managers might prioritize mitigation in areas surrounding wind-energy facilities, whereas if mortalities happen predominately on non-local bats, then broader scale coordination may be needed. Furthermore, if bats killed at wind-energy facilities are migrants, then information about their geographic origins could similarly allow for more targeted conservation and management decisions (Hayes et al., 2019; Kunz et al., 2007).

At present, biomarkers represent a useful approach for generating information about the geographic origins of migratory bats (Fleming, 2019; Hobson et al., 2019). Biomarkers are stable individual features of a tissue that reflect the location where the tissue was formed. They represent a scalable alternative for studying animal movements relative to other approaches such as GPS tags or bands that are not feasible for many bat species (Brigham, 1988; Ellison, 2008; Popa-Lisseanu & Voigt, 2009). Two common types of biomarkers for tracing animal movements are chemical markers, such as isotopes and trace elements that are incorporated into fur or other tissue, and genetic markers, which are passed from parent to offspring. Genetic markers to determine origin have not been used for Lasiurus borealis, L. cinereus, or Lasionycteris noctivagans, because studies to date have indicated they display little population genetic structure (Pylant et al., 2016; Sovic et al., 2016; but see also Monopoli et al., 2020). Thus, most work has focused on stable isotopes, primarily stable hydrogen isotope ratios or δ2H (Baerwald et al., 2014; Cryan et al., 2004, 2014; Pylant et al., 2016), and more recently, trace elements (Wieringa et al., 2020). In addition, while not a biomarker, information from distributional data, such as species distribution models (SDMs) or occurrence record density, can also be used to understand migration and is sometimes used in conjunction with biomarkers (Cryan et al., 2014). This is because SDMs represent the suitability of a geographic area for an organism, with not all locations considered equally suitable (e.g., Fournier et al., 2017; Ruegg et al., 2017; Zhu et al., 2020).

The use of δ2H as a biomarker has generated strong evidence of the capacity for long-distance migration for many bat species across a latitudinal gradient (e.g., Baerwald et al., 2014). For example, Wright et al. (2020) used δ2H values to demonstrate that Miniopterus schreibersii undertakes long-distance migrations across Europe. δ2H analysis has also demonstrated that L. noctivagans are capable of a latitudinal migration up to 2500 km (Fraser et al., 2017). While some have applied these data to attempt to identify the origin of bats killed at wind-energy facilities (e.g., Pylant et al., 2016), δ2H analysis generally provides coarse spatial resolution information. This coarse resolution reflects a low precision of assignment, reflected in the percent area of a species range that a given biomarker suggests as a likely origin While it is difficult to give specific numbers, probable origins using only δ2H values can be on the order of continent-wide in North America (Fournier et al., 2017). Using this definition, a higher precision reflects a smaller predicted area that is considered a likely origin. Because geographic patterns of δ2H values of precipitation vary primarily along latitudinal or elevational gradients, subsequent models typically lack spatial resolution when predicting the origin of an individual along longitudinal dimensions (Hobson et al., 2019).

Other chemical biomarkers such as trace elements and other stable isotopes typically reflect different geographical patterns and so offer the potential for additional precision when identifying animal origins (Wieringa et al., 2020). Trace element markers were reasonably successful at modeling the origins of bats but also lacked spatial resolution in predictions. Although trace elements represent limited overall ability to assign individuals to small geographic areas of origin, they reflect a different geographic pattern than δ2H values. Specifically, δ2H values have a broadly latitudinal differentiation across North America, whereas trace elements have a more varied pattern with more varied spatial orientation (Wieringa et al., 2020). These differences in geographic orientation of resolution suggest that measurements of δ2H values and trace elements from the same individual could, in conjunction, increase the precision for assigning origins of individual bats.

Combining information from multiple biomarkers may increase the precision of sourcing individuals to specific locations (e.g., Clegg et al., 2003; Kelly et al., 2005). For example, in migratory birds, researchers have demonstrated an increased precision for determining summering grounds by combining genetics with δ2H data (Clegg et al., 2003; Gómez-Díaz & González-Solís, 2007; Kelly et al., 2005; Ruegg et al., 2017; Rundel et al., 2013). Combining multiple stable isotopic markers also improves model assignment (Popa-Lisseanu et al., 2012; Segers & Broders, 2015). In the same vein, researchers have combined δ2H and trace element data (Holder et al., 2014; Szép et al., 2009) or combined δ2H data with species distribution data (Cryan et al., 2014; Fournier et al., 2017). However, while some exceptions exist (see Cryan et al., 2014; Popa-Lisseanu et al., 2012; Segers & Broders, 2015), most of these studies have focused on migratory birds. Despite calls for this type of analysis to be applied to a broader range of organisms (Hobson et al., 2019), to date the application of combining intrinsic markers in migratory bat species has so far been limited, largely due to the statistical and sampling issues in doing so. Specifically, enough samples need to be collected from across each species range of known origin, which can be difficult to accomplish in a migratory species for which permits, justifiably so, can be difficult to obtain, and time and money are limited.

Here, we assessed the combining of multiple biomarkers (δ2H and trace elements) and distribution data for three species of migratory bats (Lasiurus borealis, L. cinereus, and Lasionycteris noctivagans). As combining multiple markers is not guaranteed to improve assignment (Hobson et al., 2019), it is important that researchers understand the potential improvement in the resolution of sourcing individuals provided by each approach for cost/benefit determination. To tackle this challenge, we (1) collected biomarker data from the same known-origin individuals across the species' range for three species of migratory bats, (2) combined information from multiple biomarkers to assess origin of individuals from each species, and (3) determined the improvement to precision while maintaining an 80% accuracy after combining markers. Our goal is to provide researchers with guidelines for biomarker application that can be used to design future studies for sourcing individuals which can be used to generate information important for developing species-specific conservation plans.


Sample collection

Samples were collected from across the geographic range of each species studied (Figure 1). Sources for these samples included carcasses salvaged at wind-energy facilities and museum specimens. All samples used were collected during summer months (June and July), presumably after molting and prior to fall migration (Fraser et al., 2013) and can be assumed to be of known origin. All samples have been used previously in other research studies, and so we combined past data sources with data generated for this study (Table 1). In total, 279 samples were used: 96 for Lasiurus borealis, 134 for L. cinereus, and 49 for Lasionycteris noctivagans. Specifically, Lasiurus borealis samples were used in Pylant et al. (2016), Campbell et al. (2020), and Wieringa et al. (2020) to test the use of δ2H and trace element information for the assignment of origin. Samples for L. cinereus and Lasionycteris noctivagans were used in δ2H analyses only (Campbell et al., 2020; Pylant et al., 2016) and so trace element data were generated for these samples as well, as described below. Based on sampling time, we assume that the location of sampling is close to the summering location for that individual, indicating that the fur composition represents the δ2H and trace element values in that area.

Details are in the caption following the image
Sample locations for each species as compared to their range as given by the IUCN. (A) Lasiurus borealis; (B) L. cinereus; (C) Lasionycteris noctivagans.
TABLE 1. Origin of data used in this study for each data type and each species.
Species Stable hydrogen isotopes Trace elements SDM
Lasiurus borealis Pylant et al. (2016); Campbell et al. (2020) Wieringa et al. (2020) Wieringa et al. (2021)
Lasiurus cinereus Pylant et al. (2016); Campbell et al. (2020) This study Wieringa et al. (2021)
Lasionycteris noctivagans Pylant et al. (2016); Campbell et al. (2020) This study Wieringa et al. (2021)

Isotope assignment

We used data on δ2H values originally described in Pylant et al. (2016) and Campbell et al. (2020). Detailed descriptions of the sample collection and δ2H analysis can be found in those publications.

To assign fur samples to a geographic origin, we relied on an isoscape, or a geographic model of variation in δ2H values of precipitation. We applied an isoscape reflecting the months of tissue growth (Fraser et al., 2013), which for all study species occurs in the summer months (Cryan et al., 2014; Fraser et al., 2017; Pylant et al., 2014). We used an isoscape reflecting the isoscape δ2H values of precipitation for June–August based on elevation, latitude, and latitude2 as spatial predictors at a 2 km2 resolution in North America (IsoMAP job 66098; Bowen et al., 2014). For a given individual, δ2Hhair was transformed to δ2Hprecip using mean regression coefficients (Campbell et al., 2020). The probability of origin modeling was conducted using the “isocat” R package (Campbell, 2020; Campbell et al., 2020). The probability of an individual having originated at a given cell of the isoscape was predicted using the following equation:
f y μ , σ = 1 2 π σ 2 exp 1 2 σ 2 y μ 2 ,
where the probability of that given cell being the origin with a given δ2Hprecip value y is f(y|μ,σ), given an expected mean (μ) and combined error (σ) within the δ2Hprecip isoscape and mean regression coefficients. Surfaces created were normalized to sum to 1 and were fit to an isoscape with an equal-area projection.

Trace element assignment

Trace element data for Lasiurus borealis samples were taken from Wieringa et al. (2020). Data for samples of L. cinereus and Lasionycteris noctivagans were generated for this study following the methods described in Wieringa et al. (2020). For a given individual, as with the δ2H analyses, hair samples were cleaned with Triton X and rinsed with acetone as recommended by Flache, Becker, et al. (2015). Fur was then placed in low-density polyethylene tubes after gold trichlorite was added at 1 ppm as a mercury stabilizer. Fur was then digested using 1 mL of HNO3 and placing the sample on a heating block at 70°C for 1 h. The samples were allowed to cool to room temperature and were diluted to a final volume of 10 mL with a 10 ppb indium control standard.

The samples were then analyzed with a ThermoFinnigan Element 2 high-resolution inductively coupled plasma mass spectrometry (ICP-MS) at the Trace Element Research Lab (TERL) at The Ohio State University. The samples were compared to a narrow-range calibration curve to determine the concentration of elements present in each sample. During a run, every 10–15 samples, the blank and a calibration standard were reanalyzed to ensure the accuracy of our results. We then corrected each sample using the internal standard using the integrated software.

A trace element basemap (similar in principle to an isoscape, using spatially weighted PCA values for trace element concentrations) was created using a leave-three-out approach converting sample trace element concentrations to spatial weighted principal components using the R package “gwpca,” during which all but three samples were used to create a basemap for North America, after which those three samples were assigned to the basemap (Wieringa et al., 2020). Assignment to the basemap was done using the following equation from Tonra et al. (2015):
f y * μ i , σ = 1 σ 2 π e y * μ i 2 2 σ 2 ,
where f(y*|μi,σ) is the likelihood that an individual with a given principal component fur value (y*) originated from cell I, μi is the predicted spatially weighted principal components analysis (sPCA) value of cell I, with a variance within a sampling area (σ). For each individual, then a map of probability across cells in North America was normalized.

Species distribution modeling

We used SDMs generated in Wieringa et al. (2021). For this study, we only used SDMs from summer months as that is when the fur for all other analyses was grown on the sampled bats. To do this, we created ensemble SDM for each species combining June and July distribution models from Wieringa et al. (2021) via averaging (reflecting the mean probability of occurrence). The ensemble models were then normalized to sum to 1.

Combining assignment models

We assessed whether there was any increase in the precision of assignment resulting from combining data from δ2H analysis of fur, trace element analysis of fur, and seasonally explicit SDMs to generate a single joint probability surface for each known-origin individual, as has been suggested for bats (Brewer et al., 2021). For each individual of each species, we used individual probability surfaces generated from SDM, trace element, and δ2H data sets and then combined individual probability layers into a single surface using two approaches. The first approach is a direct multiplication of normalized probability surfaces as under the assumption of independence among the biomarkers, the combined probability is the product of the probability surfaces (Brewer et al., 2021). We also utilized a Bayesian approach introduced in Rundel et al. (2013), cross-validated calibrated combined (CVCC) model tuning, in which each parameter is independently flattened or sharpened to generate and then multiplied to create an overall surface. The sharpening and flattening terms were then varied to produce the best-performing model using this equation (Rundel et al., 2013):
P k S , I ref , T ref P sdm k w P S I k , I ref x P S T k , T ref y π k ,
where Iref is the isoscape basemap samples are assigned to, Tref is the trace element basemap, k is a cell (location) on each raster, S is the sample δ2H or trace element values, so Psdm(k) is the habitat suitability at a given cell (k), P(SI|k,Iref) is the probability of sample (S) given the location and basemap for δ2H, P(ST|k,Tref) is the probability of sample (S) given the location and basemap for trace elements, π(k) is a flat basemap, and w, x, and y are powers each surface is raised to. Each of these contributing models from a different data set is then raised to varying powers ranging from 10−1 to 10 to sharpen (power >1) or flatten (power <1) the probability surfaces. This approach was used to iteratively manipulate the continuous layers to find the best-performing combination. The final combined maps were rescaled to range from probabilities of 0 to 1. This allows for easy implementation of comparing results at different probability thresholds (below).

Model evaluation

To determine the best combination, we evaluated each iteration using several metrics. First, we used the area under the curve (AUC) metric to estimate that the probability that a randomly selected true origin location will have an assignment probability that is greater than the assignment probability of a randomly selected non-origin location. We calculated AUC using randomly generated background points, the known sampling location, and the “sdmvspecies” package (Duan et al., 2015). This approach jointly evaluates the accuracy and precision of different models, and a higher value of AUC is an indication of better performance (Rundel et al., 2013). Second, we also used a distance-based approach by quantifying the shortest distance between the highest probability cell and the known origin (Rundel et al., 2013). This allows us to understand if the highest value cell or averaged 95th quantile cell location are good representations of known origin. However, as noted in Campbell et al. (2020), distance to the highest value cell is subject to large amounts of bias. They showed that the error rates for minimum distances to the 99th and 95th quantile probability thresholds were very high for these species. To account for this, we also calculated the distance to the average 75th quantile cell.

Finally, to further assess accuracy and precision of a given joint model, we quantified both measures at a variety of probability of origin of a given location using a variety of probability thresholds (probability at known origin cell >0.5, >0.66, and >0.75). In this case, accuracy is our ability to correctly predict the known origin (bats sampled prior to migration) above a determined threshold, while precision is the area of the species range above the same threshold. In general, a higher accuracy and lower precision are desirable. With these thresholds, we can quantify the degree to which model predictions are correct and how much combining markers has reduced the possible origins relative to the total range for a given species. Another way to compare accuracy and precision is to set the accuracy threshold and then determine precision at that threshold. If known-origin samples are used, as was done here, we can select a particular accuracy (here, at 80%) to directly compare precision metrics. The best model was assessed based on the 80% precision metrics as this allowed for direct comparisons between model combinations.


We first present results for trace elements for species that have not been previously analyzed for this biomarker and then present SDMs used in this study for all three species. Finally, we compare the joint analyses of involving δ2H, trace element, and SDM information on a species-by-species basis.

Trace elements

Trace element concentrations for Lasiurus cinereus and Lasionycteris noctivagans are reported in Table 2 and Wieringa (2023). The concentrations of trace elements in the fur of these two species are similar to those previously reported in other species of bats (Flache, Becker, et al., 2015; Flache, Czarnecki, et al., 2015; Hickey et al., 2001; Wieringa et al., 2020; Zocche et al., 2010). The high SD values shown suggest there is spatial variation in the values for samples (Wieringa et al., 2020). Trace elements were incorporated in the fur of each species differently, for example, the values of mercury (Hg) in the fur of Lasiurus borealis, L. cinereus, and Lasionycteris noctivagans vary between species at levels of 4.8, 3.4, and 6.7 ppb, respectively (ANOVA; p = 0.04). As a result, each spatially weighted PCA should be determined using samples from each species, as is the case for δ2H data (Wassenaar, 2019).

TABLE 2. Trace element concentration values (parts per billion) for Lasiurus cinereus and Lasionycteris noctivagans.
Element Lasiurus cinereus Lasionycteris noctivagans
Mean SD Mean SD
Al 290.8 1141 168.7 332.5
Ni 12.6 18.2 15.1 23.9
Cu 10.4 3.3 9.9 3.9
Rb 0.2 1.0 0.2 0.3
Y 0.1 0.4 0.1 0.1
Mo 0.2 0.3 0.3 0.5
Sn 1.2 3.3 1.1 2.6
Ba 2.5 6.7 2.2 1.7
Cs 0.1 0.1 0.1 0.1
Ce 0.3 1.6 0.2 0.3
Hg 3.4 10.7 6.7 7.8
Mg 167.8 506.0 335.4 915.6
Al 272.2 991.2 172.5 341.4
Mn 6.4 21.5 6.6 7.9
Fe 208.6 993.3 77.7 99.4

To expand our prediction area from the limits of our sampling to the entire species' range, we followed the approach outlined in Wieringa et al. (2020) linking fur concentrations to soil and using soil values to increase the possible area of prediction. Soil data were taken from the USGS and Natural Resources Canada, available at and, respectively. In brief, 4857 low-density sites (1 site per 1600 km2) were selected and samples were collected from soil layers of 0–5 cm, Soil A (topsoil), and Soil B + C composite (deeper soil up to 1 m). We determined the linear relationship between the soil and fur sPCA values. This relationship had a correlation of 0.34 and 0.39 for Lasiurus cinereus and Lasionycteris noctivagans, respectively, while Lasiurus borealis had a correlation of 0.24 (Wieringa et al., 2020). These results were used to make predictions of likely origin and can be found in each species section.

Species distribution modeling

The ensemble summer maps generated for each species are visualized in Figures 3-5. The predicted distributions resemble those presented in other studies (Hayes et al., 2015; Wieringa et al., 2021).

Species-by-species performance of different approaches

A summary can be found in Figure 2 showing the best performance of each combination while detailed results are found in Wieringa (2023). We discuss these results on a species-by-species basis below.

Details are in the caption following the image
Bar chart showing the best-performing combination of markers as determined by the mean percent of species range predicted (y-axis) at an 80% accuracy (with error bars of 1 SD). Species are LABO (Lasiurus borealis), LACI (L. cinereus), and LANO (Lasionycteris noctivagans). Each combination is represented by All (isotope, trace element and species distribution model [SDM]), ISO (isotope), ISO + SDM (isotope and SDM), ISO + TE (isotope and trace elements), TE (trace elements), and TE + SDM (trace elements and SDM). Lower values of percent species range indicate better performing combinations of markers.

Lasiurus borealis—When using varying probabilities at the known-origin site, we obtained results for accuracy and precision that were similar to previous studies (Campbell et al., 2020; Pylant et al., 2016; Wieringa et al., 2020). Using a probability of origin at a given cell >0.5 δ2H had an accuracy of 73.9% with a prediction area comprising 45.9% (SD: 16.5%) of the species range (Wieringa, 2023). Similarly, for trace elements with the cell probability >0.5, we obtained an accuracy of 73.9% while the precision was 46.1% (SD: 26.6%) of the species range. Using the same cell probability when combining data across multiple sources, the best-performing model was three data sources combined which yielded an accuracy of 44.8% while predicting 14.7% (SD: 7.6%) of the species range (Wieringa, 2023). When setting thresholds to reflect ~80% accuracy for all approaches, CVCC with powers of 0.6, 0.2, and 1 for all markers combined (δ2H, trace elements, and SDMs, respectively) performed most accurately (Figure 3). Using the 80% accuracy threshold (0.265), a mean 39.7% (SD: 12.3%) of the species range was predicted as likely origin using the CVCC combination of powers, outperforming all other single markers and combinations. The combination of all models predicts likely origins containing the smallest mean geographic area (Figure 2), and the combined approach predicts less area than any individual approach (Figure 3). This combination also reflected a relatively short distance (655 km [SD: 401 km]) between the highest value cell and the sampling location compared with the other combinations used (Wieringa, 2023).

Details are in the caption following the image
Probability of origin using different approaches for one individual Lasiurus borealis. The black dot indicates the known origin of the sample. SDM is the species distribution model for that species.

Lasiurus cinereus—When using cell probability >0.5, δ2H analysis had a low mean accuracy of 26.9% while predicting a mean of 26.0% (SD: 3.2%) of the species range as likely origins. For trace elements, a cell probability >0.5 had a mean accuracy of 53.7% while predicting 37.7% (SD: 13.7%) of the species range. The best combination based on a cell probability >0.5 was a CVCC combination of δ2H and trace elements with powers of 1 and 5, respectively, reflecting an accuracy of 13.4% and precision of 2.7% (SD: 2.4%). At the 80% accuracy threshold values, a CVCC combination of δ2H, trace elements, and SDMs with powers of 0.2, 1, and 5, respectively performed best (Figure 4). This combination, a threshold of 0.0000327, reflected a mean accuracy of 79.9% while predicting 36.0% (SD: 15.4%) of the species range on average, while δ2H and trace elements predicted 50.6% and 74.9% of the species range, respectively. Using the CVCC combination, the highest value cell was 1485 km (SD: 88 km) from the known origin, while we obtained an AUC of 0.767. While neither of these were the best metrics out of all comparisons, they are within 1 SD of the best of those metrics.

Details are in the caption following the image
Probability of origin using different approaches for one individual Lasiurus cinereus. The black dot indicates the known origin of the sample. SDM is the species distribution model for that species.

Lasionycteris noctivagans—Finally, for L. noctivagans cell probability >0.5, δ2H had a mean accuracy of 42.8% while predicting an average of 21.8% (SD: 2.0%) of the species range. These values were consistent with previously reported values (Campbell et al., 2020; Pylant et al., 2016). Trace elements, while having a higher mean accuracy of 69.4%, also predicted on average 64.0% (SD: 13.7%) of the species range. The best combination using the cell probability >0.5 was trace elements and SDM having a mean accuracy of 42.3% while predicting an average of 6.8% (SD: 13.5%) of the species range. When comparing the 80% accuracy threshold precisions, we find that δ2H analysis is the best-performing single approach with a mean precision of 46.9% (SD: 4.3%) while trace elements had a precision of 87.0% (SD: 13.6%). Based on combined data, CVCC outperformed unweighted multiplication (all powers equal to 1), with δ2H and SDMs performing the best overall. The best combination using CVCC was δ2H to the power of 0.2 and SDM to the power of 0.6 with a threshold of 0.1713 (Figure 5). This combination gave a mean precision of 46.4% (SD: 4.2%). While not a substantial improvement over δ2H alone (46.4% vs. 46.9%), 0.5% of L. noctivagans range is still ~4700 km2. As in the other species, the AUC and distance metrics (0.798 [SD: 0.146] and 1910 [SD: 958]) were not the best overall but all within 1 SD of the best-performing for those metrics.

Details are in the caption following the image
Probability of origin using different approaches for one individual Lasionycteris noctivagans. The black dot indicates the known origin of the sample. SDM is the species distribution model for that species.

Link between accuracy and precision

When comparing our results, we set our accuracy to 80% to allow for direct comparisons between our approaches. Our results support the notion that there exists a clear tradeoff between accuracy and precision (e.g., Campbell et al., 2020). There was a clear relationship between the two when using any single approach (single biomarkers or a combination) as we vary the threshold values (Appendix S1: Figure S1).


We investigated the impact of combining multiple different biomarkers and distribution data on the accuracy and area of prediction (Hobson et al., 2019) by studying three species of North American migratory tree bats. Our results demonstrate that combining multiple biomarkers and SDMs improved our prediction ability for at least two species (Lasiurus borealis and L. cinereus) when compared to single biomarker approaches (δ2H or trace elements), while for Lasionycteris noctivagans, we found δ2H and SDMs together performed slightly better than δ2H alone.

Previous studies using multiple different biomarkers have shown similar shown results. For example, Crowley et al. (2021) and Kruszynski et al. (2021) measured δ2H and strontium isotopes and found that, in combination, both markers increased the prediction power in three species of raptors and Pipistrellus nathusii, respectively. However, as noted in Hobson et al. (2019), combining data from multiple biomarkers does not always lead to an improvement in assignment precision or accuracy because a variety of factors can negate potential improvements. For example, markers may not show different patterns or one marker may not add additional resolution relative to another marker's performance. For example, Pekarsky et al. (2015) found that combining stable oxygen isotope δ2H data did not improve the accuracy of origin models of Eurasian cranes, largely because geographic variation in both isotopes of precipitation is driven by similar processes.

Our results demonstrate both possibilities by showing an improvement in assignment precision when combining biomarkers for two of three species but no improvement in the third. This is further demonstrated as each species' performance varied, at the level of both on individual biomarkers and when data combining biomarkers was used. Further, the improvement between species was not consistent when the same type of data were combined. This is expected based on previous studies in other species. In birds, one comparison of trace elements and δ2H showed that trace elements performed best but that the two species (Riparia riparia and Hirundo rustica) studied showed different patterns of successful assignment to populations (Szép et al., 2009). In another study, Crowley et al. (2021) found an order of magnitude improvement when determining the origin of bird samples while others, such as Gómez-Díaz and González-Solís (2007), found an improvement in accuracy of ~8.5% when assigning seabirds to colony. This level of improvement, while based on different metrics, is similar to our results. Overall, previous results and those from our study support the claim by Hobson et al. (2019) that increased accuracy and precision is not a given when using data from multiple biomarkers. This supports the idea that the benefits of using multiple makers on assignment accuracy should be assessed on a species-by-species basis with known-origin individuals.

There are many possible explanations as to why these combinations (or single approaches) vary in performance. Much of the variation likely is due to influences of the sources and uptake of each biomarker before incorporation into animal tissue. One study comparing trace elements and δ2H found that there were yearly differences in the feather δ2H values, and, within a year, even differences between feathers of the same colony (Szép et al., 2009). These variations were attributed to differential habitat use within a single species. We would expect the same to be true for different species of bats, as they feed on different insects, behave differently, and have different internal bioregulation (Hickey et al., 1996; Reimer et al., 2010; Smith & Hatch, 2015).

The trace elements detected in these species that deposited in the hair come from one of a few sources, most likely from the intake of water and food. As all three species have overlapping ranges, it seems most likely that any differences observed are due to differences in diet. Here, we found that the two closely related species (Lasiurus borealis and L. cinereus) showed similar trends in their assignment performance. This indicates that there may be some biological features (such as differences in diet or behavior) of Lasionycteris noctivagans that make the use of some biomarkers (e.g., trace elements, which performed poorly), less accurate in this species. In terms of diet, past work indicates that both Lasiurus species eat similar prey. Previous studies have shown that both Lasiurus borealis and L. cinereus feed predominantly on Lepidoptera (Clare et al., 2009; Hayes et al., 2019; Reimer et al., 2010). However, the second most common species varies likely leading to the differences observed. On the other hand, Lasionycteris noctivagans eats a much more varied diet (Reimer et al., 2010) that is still largely Lepidoptera but includes many other types of prey. These results suggest that understanding the diet of target species has a large impact on the performance of trace elements in the overall assignment of individuals. We also note that age class should also be considered when sampling specimens as diet can change from juvenile to adult (Reimer et al., 2010; Rolseth et al., 1994).

In addition to the types of biomarkers described in this study, other methods and data sources exist for determining the origin of migratory individuals. For example, Popa-Lisseanu et al. (2012) combined stable isotopes of three elements (carbon, nitrogen, and hydrogen), finding that they were able to increase the accuracy of the origin assignment for migratory bats from <50% for a single isotope to 93% when using all three. Similarly, Kruszynski et al. (2021) showed that combining strontium isotopes with δ2H provided benefit for determining the origin of migratory bats in Europe. Another promising possibility, especially for L. noctivagans, is genetic data. Lasionycteris noctivagans has been hypothesized to have an east/west split (Cryan, 2003), and previous work using mtDNA has shown some indication of genetic structure (Monopoli et al., 2020). Finally, possible data sources include sulfur isotope analysis (Kabalika et al., 2020), tagging technology (Castle et al., 2015), MOTUS network (Dowling & O'Dell, 2018), and many others (see Brewer et al., 2021). However, while these should be investigated, combined approaches should first be tested on known-origin individuals to validate their usefulness (Hobson et al., 2019).

Using biomarkers can be a valuable approach to understanding the migration of bats as they are difficult to capture, and more traditional approaches that require recapture are likely not feasible due to low recapture rates (Schorr et al., 2014). However, to be helpful, the markers implemented must reflect geographic variation relevant to the scale and direction of movement. For example, δ2H values of precipitation have a clear latitudinal gradient (Hobson et al., 2019). Additional markers subject to a similar latitudinal gradient would, at best, offer small potential gains in model precision. We recommend utilizing those markers reflecting diverse different patterns of geographic variation to best increase model precision. Trace elements reflect diverse geographic patterns that do not reflect latitudinal variation, making them suitable for combining with δ2H (Wieringa et al., 2020). Similarly, SDMs reflect habitat suitability, which is dependent on many different environmental factors independent of latitudinal variation. By combining δ2H and trace elements, we determined the areas where overlap occurs, which allows modelers to maintain accuracy while increasing model precision. As a result, we recommend using a combination of trace elements, δ2H, and SDMs for the best predictions, the current best practice, for tracking migratory tree bats in North America. Given funding limitations, using δ2H analysis in conjunction with SDMs is both affordable and relatively accurate. Our results highlight that combining biomarkers does improve assignment ability of migratory bat species and confirm that varying the powers using CVCC (Rundel et al., 2013) is a worthwhile step using known-origin individuals for validation.

Combining multiple biomarkers is a promising technique to increase the spatial precision of origin modeling and inform many aspects of bat biology and conservation research. Although δ2H values alone have been applied successfully (e.g., Baerwald et al., 2014; Lehnert et al., 2014; Pylant et al., 2016), the range of likely origins lacked resolution with some estimates of likely origins thousands of kilometers apart. The results of this study highlight that combining markers can be useful at better resolving the origin of migratory species, but that this is not always the case and so several important issues need to be considered before taking this step in a new species. First, to accomplish this type of approach, dozens of samples need to be analyzed for each species in the study system, and as seen in L. noctivagans, this may not lead to better assignment. Second, careful thought should be given to which additional markers to utilize. For example, in some species, δ2H may not be ideal, especially if there is no North/South movement (or elevational movement) within the species as δ2H variation may not be able to resolve longitudinal patterns of movement. The same concern applies to trace elements, as in some cases the spatial resolution may be insufficient for improved assignment. In the end, the best practices for this type of approach are to use a variety of biological informed markers from over 100 individuals (for wide ranging species), sampled from across the species range, using known-origin individuals for validation, validated for each species studied, with the possibility of failure acknowledged.

Understanding migration routes and migratory status of these three species is an important step for better placement of wind-energy development by using this information to avoid locating new turbines along bat migration routes. As most bat mortality happens during migration (Arnett et al., 2008), careful siting alone could reduce bat mortality or inform the need for existing facilities to mitigate against increasing fatality. While this study adds to our understanding of bat migration and can aid in our understanding of the migratory status of those killed by wind energy, the single best approach to wind farm mortality is curtailment (e.g., Adams et al., 2021; Măntoiu et al., 2020; Smallwood & Bell, 2020; Voigt et al., 2022) or smart curtailment (Hayes et al., 2019). Overall, there needs to be a reduction in the mortality caused by wind energy. The methods presented here do offer information that could be useful for mitigation but is not a strategy in and of itself. Instead, this approach can aid conservation managers on the geographic impacts wind-energy development is having on populations. As noted in Voigt et al. (2012), better understanding the catchment area of mortality due to wind-energy facilities may allow managers to better mitigate the effects of mortality due to turbines.


We thank the Trace Element Research Lab (TERL) at The Ohio State University for their assistance with the processing of samples. Specifically, we thank Anthony Lutton for help with scheduling lab equipment use and for help with interpreting our results and John Olesik and other members of the TERL for general assistance. We also thank the Carstens and Gibbs labs in the Department of EEOB at Ohio State for their advice and assistance with the editing of this manuscript. Finally, we thank Erin Hazelton and Jonathan Sorg, Ohio Division of Wildlife, for assistance with grant administration. This work was supported by a grant (GRT00046616) from the Competitive State Wildlife Grants Program to Ohio State University and the University of Maryland Center for Environmental Science as jointly administered by the US Fish and Wildlife Service, the Ohio Division of Wildlife, and the Maryland Department of Natural Resources. This study is a contribution from the Ohio Biodiversity Conservation Partnership. Samples for this project were provided by the Smithsonian Institution, National Museum of Natural History, Division of Mammals.


    The authors declare no conflicts of interest.


    Isotope data are available via Pylant et al. (2016) and Campbell et al. (2020), trace element data for hoary and silver-haired bats are available on Dryad (Wieringa, 2023), eastern red bats data are available via Wieringa et al. (2020), and species distribution models are available via Wieringa et al. (2021). All spatial layers generated from the raw data are available from Dryad (Wieringa, 2023,