-
PDF
- Split View
-
Views
-
Cite
Cite
Simon P Ripperger, Sebastian Stockmaier, Gerald G Carter, Tracking sickness effects on social encounters via continuous proximity sensing in wild vampire bats, Behavioral Ecology, Volume 31, Issue 6, November/December 2020, Pages 1296–1302, https://doi.org/10.1093/beheco/araa111
- Share Icon Share
Abstract
Sickness behaviors can slow the spread of pathogens across a social network. We conducted a field experiment to investigate how sickness behavior affects individual connectedness over time using a dynamic social network created from high-resolution proximity data. After capturing adult female vampire bats (Desmodus rotundus) from a roost, we created “sick” bats by injecting a random half of bats with the immune-challenging substance, lipopolysaccharide, while the control group received saline injections. Over the next 3 days, we used proximity sensors to continuously track dyadic associations between 16 “sick” bats and 15 control bats under natural conditions. Compared to control bats, “sick” bats associated with fewer bats, spent less time near others, and were less socially connected to more well-connected individuals (sick bats had on average a lower degree, strength, and eigenvector centrality). High-resolution proximity data allow researchers to flexibly define network connections (association rates) based on how a particular pathogen is transmitted (e.g., contact duration of >1 vs. >60 min, contact proximity of <1 vs. <10 m). Therefore, we inspected how different ways of measuring association rates changed the observed effect of LPS. How researchers define association rates influences the magnitude and detectability of sickness effects on network centrality. When animals are sick, they often encounter fewer individuals. We tracked this unintentional “social distancing” effect hour-by-hour in a wild colony of vampire bats. Using bat-borne proximity sensors, we compared changes in the social network connectedness of immune-challenged “sick” bats versus “control” bats over time. “Sick” bats had fewer encounters with others and spent less time near others. Associations changed dramatically by time of day, and different measures of association influenced the sickness effect estimates.
INTRODUCTION
As a pathogen spreads across a population, changes in social behavior can change the rate of pathogen spread. Transmission rates can increase when parasites manipulate host behavior (Klein 2003; Ezenwa et al. 2016) or decrease when healthy individuals avoid sick individuals (Kiesecker et al. 1999; Behringer et al. 2006; Boillat et al. 2015; Poirotte et al. 2017). In eusocial insects, sick individuals might altruistically self-isolate or be excluded by their colony mates (Heinze and Walter 2010; Baracchi et al. 2012; Bos et al. 2012; Cremer et al. 2018; Stroeymeyt et al. 2018). A simpler mechanism causing reduced transmission is that infected individuals often show “sickness behavior,” which includes increased lethargy and sleep, and reduced movement and sociality (e.g., Van Kerckhove et al. 2013; Lopes et al. 2016; Stroeymeyt et al. 2018). This sickness-induced “social distancing” does not require cooperative traits and might be common across group-living species. It might therefore be important for modeling pathogen transmission as a social network changes over time (i.e., a dynamic social network; Pinter-Wollman et al. 2014; Farine 2017).
Creating a dynamic social network to track the effects of sickness behavior over time requires datasets with high temporal and spatial resolutions that are sufficient to answer the biological question. However, for many questions and study designs, there are severe logistical limitations imposed by how social associations are sampled. Automated tracking of animal associations typically occurs in the lab (e.g., Stroeymeyt et al. 2018) or at specific locations in the field, such as at camera traps (e.g., McCarthy et al. 2019), at feeders (e.g., Aplin et al. 2015), or at nest boxes (e.g., Lopes et al. 2016). In contrast, proximity sensors can continuously measure the times and durations of encounters among free-ranging animals (Ripperger et al. 2020), and are therefore a potentially powerful tool for understanding how changes in individual behavior reshape a social network over time.
Here, we induced sickness behavior in wild-caught vampire bats using injections of lipopolysaccharide (LPS). By inducing sickness behavior without an active pathogen, LPS injections isolate the effects of sickness behavior from the effects of manipulation by a specific pathogen (Stockmaier et al. 2018). Next, we tagged both the “sick” bats (injected with LPS) and control bats (injected with only saline) with proximity sensors (Ripperger et al. 2020), released them back into their wild colony, and tracked changes in their association rates over time. Based on known effects of LPS on the physiology and behavior of captive vampire bats, including lethargy and dramatic reductions in social behaviors such as social grooming and contact calling (Stockmaier et al. 2018; Stockmaier, Bolnick, Page, Carter 2020; Stockmaier, Bolnick, Page, Josic, et al. 2020), we predicted lower association rates for pairs of a “sick” bat and a control bat than for pairs of control bats, and lower overall connectivity for “sick” bats compared to control bats, in a natural setting.
We indeed found evidence that LPS-induced sickness behavior reduced network centrality. In addition, we visualized how association rates and the effect size of LPS changed over time hour by hour, and we show how estimates of the sickness effect change as social network connections (edges) are defined at different thresholds of encounter proximity or duration. Using high-resolution data, studies can define spatial and temporal thresholds to match the exposure time or proximity required for a particular pathogen to transmit between individuals.
METHODS
Sampling wild vampire bats
We captured bats from a colony of common vampire bats (Desmodus rotundus) inside a hollow tree at Lamanai, Belize. Before sunset on 24 April 2018, we blocked all exits from the roost except one. We used a hand net to cover this exit and set up mist nets in front of it to capture any bats that evaded the hand net. Bats that hit either net were removed instantly. Before midnight, we captured male vampire bats only. The first female exited the roost at around 0200 h. In total, we captured about 100 vampire bats (including 41 females) until 0500 h the next morning. We released the adult males and juveniles back into the roost and kept only adult females in individual cotton cloth bags, and measured their mass to ensure it did not differ between the randomly assigned treatment and control groups (difference in mass = 0.17 g [95% confidence interval {CI}: −2.9, 2.4]).
Bats that use multiple roosts, such as common vampire bats, will often switch to another roost when captured at the same roost repeatedly, and we did not know if or how many other roosts were used by these bats. Therefore, to minimize disturbance and the chance of bats abandoning the roost, we chose to capture and treat the test and control bats only once, rather than capture them first to apply proximity sensors and then recapture the same bats to induce sickness behavior. As a control period, we therefore used the subsequent days after the treatment period, when we expected the LPS effects to disappear.
Inducing sickness behavior
To ensure treatments were balanced, we randomly assigned female bats to the test or control treatment by flipping a coin until either the test or control group reached half the sample of bats. We injected the individuals in the test group under the dorsal skin with 70–100 µL of LPS (LPS in phosphate-buffered saline, L2630 Sigma-Aldrich, St Louis, MO) at a dosage of 5 mg/kg, following previous studies with this species (Stockmaier et al. 2018; Stockmaier, Bolnick, Page, Carter 2020). Bats in the control group received an injection of the same volume per body mass of phosphate-buffered saline. One hour after injection, we released 34 females, tagged with proximity sensors, back into their roost (along with untreated bats that were excluded from the study because of advanced pregnancy), but we analyzed data from the 31 females that did not remove their proximity sensors.
Proximity tracking
To track dyadic associations among the bats, we used custom-built proximity sensors (see Duda et al. 2018 and Ripperger et al. 2020 for details). The sensors weighed 1.8 g (including battery and housing) and were glued to the dorsal fur using skin-bonding latex adhesive (Montreal Ostomy Skin-Bond). Tag weights were 4.5–6.9% of each bat’s mass, in accordance with recommendations for short-term tracking of bats (O’Mara et al. 2014). Each proximity sensor broadcasted a signal every 2 s. This signal wakes every proximity sensor within 5–10 m from “sleep mode,” and initiates dyadic encounters between the sender and all receivers. As long as a dyad remains within reception range, the encounter duration and the maximum received signal strength indicator (RSSI) are updated every 2 s. Although RSSI values can change with orientation and environmental factors like humidity, vampire bats roost in a predictable orientation (upside down), and roosts have fairly stable temperatures and humidity. When no signal is received from an encountered partner for 10 s (five times the sampling rate), the encounter is terminated and stored to on-board memory along with the IDs of the partners, a timestamp of the start of the encounter, encounter duration, and the maximum RSSI. The duration and maximum RSSI of each encounter can be used as an estimate for a minimum distance between two tagged bats during the encounter. As a measure of relative proximity, we defined the “proximity index” as the percent quantile of all maximum RSSI values.
To define an association, we used a proximity index of 85% (i.e., the top 15% of all encounters ranked by signal strength, Supplementary Figure S1). We chose this value because previous work showed that this RSSI value (−27 dbm) corresponds to a proximity of less than 50 cm and used this value to link association networks in the wild to grooming networks for the same bats in captivity (Ripperger et al. 2019). Association durations were the sum of encounter durations within each dyad.
We could remotely download proximity sensor data without disturbing the bats inside the tree because we placed the antennas of the base station (for encounter data download) inside the roost at ground level below the bats. Each download event is stored to an SD card with a timestamp of data reception and an RSSI of the received package. We connected to the base station each morning from outside the roost without entering the roost.
We excluded data from three sensors, which apparently dropped off the bat, either inside (n = 2) or outside (n = 1) the roost, evident from the sensor’s constant contact with the base station and no evidence of the sensor exiting or entering the roost. We therefore used association data from 16 “sick” bats and 15 control bats.
Network construction
We created social networks where edges connect pairs of bats that were associated and edge weights were association durations. To measure the overall effect size for LPS injections, we created one network for a “treatment period” that was 3–9 h postinjection (1700–2300 h), when captive bats showed LPS-induced behavioral changes (Stockmaier et al. 2018). Our analysis was also constrained to within this general time window for two other reasons. First, we observed, both during the capture night and in the subsequent proximity data, that bats began to leave the roost to forage after midnight (females departed between 0200 h and 0500 h). Vampire bats depart roosts individually to forage, so if “sick” bats were less likely to forage or spent less time foraging, this would reverse the effect of LPS on social interaction rates during that time window. Second, observations of captive vampire bats suggest that individuals are often asleep from sunrise to sunset (0600–1800 h). If both healthy and “sick” bats are asleep, their rate of associations will not differ during these hours because changes in encounter rates require movement. In summary, we expected to see clear effects during the treatment period, which was also the part of the night when all bats would freely interact within the roost.
Based on previous captive studies in vampire bats using the same dosage of 5 mg LPS per kg body mass (Stockmaier et al. 2018), we expected 48 h to be sufficient time for the bats to return to baseline behavior. In LPS-injected rats, social exploration of a juvenile conspecific mostly returned to normal after 24 h (Deak et al. 2005). For comparison with the social network during the treatment period, we also created two more social networks for the corresponding periods of the next days, 24 and 48 h later. We compared identical hours across days because degree centrality can change greatly within a day (e.g., Figure 2 in Ripperger et al. 2020), due to circadian rhythms and other diel effects. To track associations at higher temporal resolution within each day, we also created social networks for each hour. We used these to plot the mean differences between groups and dyad types over time.
Ethical approval
This work was approved by the Institutional Animal Care and Use Committee at the American Museum of Natural History (Protocol # AMNHIACUC-20180123).
Statistical analysis
To test the effect of LPS on social connectedness, we first fit a general linear mixed-effects model with treatment (LPS, saline) and day (1, 2, and 3) as fixed effects and bat as a random effect. The response variable was one of three network centrality measures: degree (number of sampled bats encountered), strength (time spent with other sampled bats), and eigenvector centrality (a measure of both direct and indirect connectedness that accounts not only for the number of encountered bats but also for the connectedness of those encountered bats). We then extracted the standardized coefficients for the overall treatment effect and the interaction between treatment effect and day. Then, we fit a linear model for the observations within each day separately and extracted those standardized coefficients for the treatment effect. To test if effect sizes were greater than expected by chance, we fit the same model to 10,000 null datasets in which the treatment identity was assigned randomly among bats, then we calculated the proportion of the null expected coefficients that were greater than the actual observed coefficients to get one-sided P values. This node-label permutation procedure accounts for the nonindependent and non-normal structure of the network data, by holding the network edge structure constant in the observed and expected networks. To get two-sided P values, we used the proportion of the absolute value of expected coefficients that were greater than the absolute value of the observed coefficient. To test the robustness of our results, we also did an alternate analysis testing the effects within each hour (rather than over the whole period), by including hour as a random effect in the model and using the same permutation approach (1000 permutations).
To investigate assortativity of “sick” and control bats over time, we calculated for each hour the association probability (proportion of possible pairs that were associated) and the mean association duration (total seconds per period) for three types of dyads: control–control, control–sick, and sick–sick. For all 95% CI, we used basic nonparametric bootstrapping with 5000 iterations.
Measuring effects of network construction on effect size of sickness behavior
Defining social network edges requires deciding what minimum value of proximity or duration is sufficient for an “association.” Dyadic association rates are typically the sum of a series of encounters or co-occurrences in a period of sampled time. With proximity sensor data, this decision of how to define an encounter requires balancing the trade-off between maximizing the sample size of encounters versus filtering for encounters that are more meaningful (e.g., closer proximities or longer durations). After conducting our main analysis above, we inspected how the size of the LPS effect changed with variations in how associations (and hence network edges) might be defined. To do this, we resampled our data using a range of different thresholds for defining encounters, then we plotted changes in the number of observations and the effect size of the treatment (defined as the unstandardized model coefficient within the treatment period). To assess detectability, we plotted the two-tailed permutation-based P values (1000 permutations per P value).
We investigated how three different ways of defining encounter-based association would influence the effect size of sickness behavior on degree centrality. First, to investigate the effect of minimum duration, we defined encounters at one fixed threshold of proximity (85% as in our original analysis), but used 21 increasing thresholds of minimum duration from 0–1200 s. Second, to investigate the effect of minimum proximity, we defined encounters using one fixed threshold of duration (0 s as in our original analysis), but then used 12 increasing thresholds of the proximity index from 76% to 98%. We used this range because proximity thresholds below 76% are not informative given the volume of the roost, and thresholds above 98% lead to small sample sizes and tests that are clearly underpowered. Third, to assess if the LPS treatment effect size was robust across different proximity thresholds while controlling for number of observations, we used 10 proximity index thresholds from 75% to 94%, and we randomly subsampled (without replacement) the same number of observations in each case. We always sampled 5258 encounters, which is 95% of the total observations at the highest proximity index (94%). At each proximity index threshold, we obtained 200 effect size estimates with different random subsamples. These three procedures tell us about the detectability and robustness of our effect sizes across different ways of defining encounters and hence association rates.
RESULTS
LPS-injected bats were less socially connected
Compared to the control group, “sick” (LPS-injected) bats associated with fewer groupmates (lower degree centrality), spent less time with others (lower strength centrality), and were less socially connected to more well-connected groupmates when considering both direct and indirect connections (lower eigenvector centrality), and these effects declined after the treatment period (Figure 1, Table 1). During the 6 h of the treatment period, a “sick” bat associated on average with four fewer associates than a control bat. A control bat had, on average, a 49% chance of associating with each control bat ([95% CI: 44, 54], n = 105 pairs), but only a 35% chance of associating with each “sick” bat ([31, 38], n = 240 pairs). During the treatment period, “sick” bats spent 25 fewer min associating per partner. While the mean association time for all possible pairs of control bats was 15 min/h [13, 18], it was only 10 min/h for a control and “sick” bat ([8, 11], Figure 2).
Model estimates of LPS treatment on network centrality. Standardized coefficients (β) and two-sided P values are reported for degree, strength, or eigenvector centrality. Two-sided P values are based on 10,000 node-label permutations
Fixed effect . | Degree . | Strength . | Eigenvector . | . | ||
---|---|---|---|---|---|---|
. | β . | P . | β . | P . | β . | P . |
Interaction (treatment × day) | 0.60 | 0.001 | 0.31 | 0.015 | 0.49 | 0.003 |
Treatment (with interaction) | −1.47 | 0.0005 | −0.91 | 0.002 | −1.41 | 0.005 |
Treatment (treatment period) | −0.86 | 0.0077 | −0.86 | 0.014 | −0.81 | 0.024 |
Treatment (24 h later) | −0.11 | ns | −0.36 | ns | −0.41 | ns |
Treatment (48 h later) | 0.29 | ns | 0.02 | ns | −0.003 | ns |
Fixed effect . | Degree . | Strength . | Eigenvector . | . | ||
---|---|---|---|---|---|---|
. | β . | P . | β . | P . | β . | P . |
Interaction (treatment × day) | 0.60 | 0.001 | 0.31 | 0.015 | 0.49 | 0.003 |
Treatment (with interaction) | −1.47 | 0.0005 | −0.91 | 0.002 | −1.41 | 0.005 |
Treatment (treatment period) | −0.86 | 0.0077 | −0.86 | 0.014 | −0.81 | 0.024 |
Treatment (24 h later) | −0.11 | ns | −0.36 | ns | −0.41 | ns |
Treatment (48 h later) | 0.29 | ns | 0.02 | ns | −0.003 | ns |
Model estimates of LPS treatment on network centrality. Standardized coefficients (β) and two-sided P values are reported for degree, strength, or eigenvector centrality. Two-sided P values are based on 10,000 node-label permutations
Fixed effect . | Degree . | Strength . | Eigenvector . | . | ||
---|---|---|---|---|---|---|
. | β . | P . | β . | P . | β . | P . |
Interaction (treatment × day) | 0.60 | 0.001 | 0.31 | 0.015 | 0.49 | 0.003 |
Treatment (with interaction) | −1.47 | 0.0005 | −0.91 | 0.002 | −1.41 | 0.005 |
Treatment (treatment period) | −0.86 | 0.0077 | −0.86 | 0.014 | −0.81 | 0.024 |
Treatment (24 h later) | −0.11 | ns | −0.36 | ns | −0.41 | ns |
Treatment (48 h later) | 0.29 | ns | 0.02 | ns | −0.003 | ns |
Fixed effect . | Degree . | Strength . | Eigenvector . | . | ||
---|---|---|---|---|---|---|
. | β . | P . | β . | P . | β . | P . |
Interaction (treatment × day) | 0.60 | 0.001 | 0.31 | 0.015 | 0.49 | 0.003 |
Treatment (with interaction) | −1.47 | 0.0005 | −0.91 | 0.002 | −1.41 | 0.005 |
Treatment (treatment period) | −0.86 | 0.0077 | −0.86 | 0.014 | −0.81 | 0.024 |
Treatment (24 h later) | −0.11 | ns | −0.36 | ns | −0.41 | ns |
Treatment (48 h later) | 0.29 | ns | 0.02 | ns | −0.003 | ns |

Network centrality was lower in LPS-injected bats compared to control bats. “Sick” LPS-injected bats (dark nodes) were less socially connected than control bats (light nodes) and this effect diminished over the 48 h of observation. In network graphs, edge weights are association durations (log10-transformed), and spatial positions are based on the graph embedder (GEM) force-directed layout algorithm. The time-series panel (left side) shows that three measures of mean (with standard error) centrality (degree, strength, and eigenvector) were lower in the “sick” test group (dark triangles) compared to the control group (light circles) during the treatment period. This difference diminished during the corresponding post-treatment periods (solid vertical lines). Effect size panels (right side) show the mean (and bootstrapped 95% CIs) centrality measures of each group during the entire treatment and post-treatment periods. Strength centrality was calculated with association hours rather than seconds.

LPS-injected “sick” bats associated less with each other and with saline-injected control bats. Time series (left side) shows the hourly association duration (top) and mean association probability (bottom) for dyads with two control bats (black circles), two “sick” bats (light green squares), and one “sick” and control bat (red triangle). Effect size panels (right side) show the mean of aforementioned association values (and bootstrapped 95% CIs) for a control bat and another control bat (black circles, c-c) or a control and a “sick” bat (red triangles, s-c) during treatment and both post-treatment periods.
Our results were robust to an alternate analysis of the hourly networks. Effect sizes of LPS on centrality by hour were smaller with larger CIs when compared with the same effects summed over 6 h, but we still detected the same effects of the treatment only within the initial treatment period (n = 31, degree: β = −0.55, P = 0.014; strength: β = −0.59, P = 0.010; eigenvector centrality: β = −0.54, P = 0.019, see Table 1 for comparison with the 6-h networks).
At about 0200 h, when we observed that the female bats began departing the roost to forage, both the number of associates (degree) and total time of associations (strength) declined sharply in both groups (Figure 1). The earlier effects of sickness behavior (the difference between the group means) were no longer clear during the periods when we expected the bats to be foraging (past midnight to 0600 h) and later sleeping (0600–1800 h, Figures 1 and 2).
Effect sizes and detectability of sickness effects depend on network construction
When we increased the minimum encounter duration for defining an association, we observed that the estimate of the effect size for the LPS treatment grew larger but eventually became smaller as the sample size of observations decreased (Figure 3A). The effect also became harder to detect when only using encounters over 15 min. When we increased the minimum proximity index for defining an encounter, we again observed that the treatment effect estimate grew larger, but the P values also increased as the number of observations declined (Figure 3B). When controlling for the number of observations, the treatment effect was relatively stable across different proximity thresholds (Supplementary Figure S2).

Effect of network edge thresholds on treatment effect size. Network edges are based on dyadic association rates, which are the sum of a series of encounters defined by a minimum duration (A) and proximity (B). Redefining association using different thresholds of duration and proximity can alter the magnitude and detectability of treatment effects on degree centrality. Panel A shows how the minimum threshold of encounter duration (x axis) affects the estimate of the treatment coefficient (top), sample size of dyadic encounters (middle), and permutation test P value (cropped at 0.05). Panel B shows how the same measures are affected when encounters are defined using different thresholds of relative spatial proximity. Dashed vertical lines show the duration and proximity index thresholds we used in the original main analysis. Gray triangles indicate estimates with P values >0.05.
DISCUSSION
“Sick” bats, which were injected with the immune-challenge agent LPS, had a lower network centrality than the control bats injected with saline (Figure 1). Pairs of a “sick” and a control bat had on average a lower probability of association and a lower association time than pairs of two control bats (Figure 2). These effects were not caused by spatial assortativity of “sick” bats (e.g., “sick” bats clustering closer together at a certain location away from others) because associations between two “sick” bats were even lower than between a “sick” and a control bat (Figure 2). “Sick” bats were therefore less often in close proximity to other sampled bats in the roost.
Previous captive studies revealed that LPS-injected bats were physiologically immune-challenged, slept more, moved less, engaged in social grooming with fewer partners in a flight cage, spent less time grooming neighboring bats when forced into close association, and produced fewer contact calls that attract affiliated groupmates (Carter and Wilkinson 2016; Stockmaier et al. 2018; Stockmaier, Bolnick, Page, Carter 2020; Stockmaier, Bolnick, Page, Josic, et al. 2020). In those past experimental trials, one focal common vampire bat at a time was injected with the same dose of LPS (relative to body weight), and the interactions with group members were focal sampled in controlled captive environments. In this field experiment, we showed that these effects of sickness behavior extend to proximity-based association durations and social network connectivity in the natural environment. Sickness behavior can therefore slow the spread of a pathogen that is transmitted at higher probability with higher rates of physical contact (e.g., grooming) or closer proximity.
When we visualized the sickness effects over time, we observed that mean association rates were lowest during the first night, probably because catching and treating the bats influenced their social behavior (e.g., effect of the proximity sensor on self-grooming), but these effects existed across both the test and control group. As expected, differences between groups began after LPS effects are known to occur from previous captive studies (Stockmaier et al. 2018) and declined afterward. We also observed evidence for changes in association and the sickness effect size during times when bats were likely to be sleeping (after sunrise at 0600 h) or foraging (the sudden drop in associations seen around 0300 h in Figures 1 and 2). Dynamic networks are useful for revealing such temporal patterns and could also help to account for ecological effects on association rates, for example, when individuals respond to predation risk during certain times of the day, or huddle for warmth during colder periods.
Besides sickness behavior reducing interaction rates, social networks can also be reshaped following an infection through three other processes. Individuals can actively avoid contact with infected conspecifics (e.g., lobsters: Behringer et al. 2006, bullfrog tadpoles: Kiesecker et al. 1999, mice: Boillat et al. 2015, mandrills: Poirotte et al. 2017), some human and eusocial societies can actively and cooperatively restructure their social networks (e.g., Cremer et al. 2018; Stroeymeyt et al. 2018), and parasites can manipulate host behavior to restructure host networks in favor of parasite transmission (Klein 2003; Ezenwa et al. 2016). However, these processes can also interact on either a developmental and an evolutionary timescale, as for example, when infection-induced changes in host behavior induce feedbacks that alter parasite manipulation behavior (Ezenwa et al. 2016).
Depending on the goals of a study, sickness behavior or pathogen transmission can be measured or modeled at varying spatial and temporal scales. Networks are often based on co-occurrence at a resource. For example, studies using passive integrated transponder tags to track free-ranging mice showed a decreased probability of sharing a refuge (Lopes et al. 2016), and social networks in bats are typically defined based on individuals occupying the same roost on the same day (Wilkinson et al. 2019). Proximity sensors allowed us to continuously measure proximity even within a single roost and day. On a larger spatial and temporal scale, pathogen transmission crucially depends on movements between roosts and sites (e.g., rabies in vampire bats; Blackwood et al. 2013). Conceptually, a within-site social network (mapping transmission rates between individuals) can be considered as embedded within a single node of a larger network mapping transmission rates between sites (e.g., Montiglio et al. 2020).
When defining network edges with continuous proximity data, there is a trade-off between having more observations of encounters and selecting those encounters that are most meaningful, for example, closer encounters that are more relevant for a given transmission mode or pathogen. In this study, we used resampling to show the effect of defining encounters (and hence association rates) at various thresholds of minimum duration and distance (Figure 3), and we recommend resampling procedures for testing the robustness of an effect across the range of durations or distances that are biologically meaningful (Supplementary Figure S2). As tracking technology improves the capacity to create dynamic animal social networks from large, high-resolution datasets, we expect researchers to gain new insights into the patterns and processes underlying the spread of pathogens, information, or behavioral states.
Acknowledgements
We thank Michelle Nowak for help with fieldwork and data collection, Brock Fenton, Nancy Simmons, and Dan Becker for help with logistics and permits, Frieder Mayer for funding, and Björn Cassens and Niklas Duda for technological support. We thank the editor, Kelsey Moreno, and two reviewers for suggestions that improved the manuscript.
FUNDING
This work was supported by the Deutsche Forschungsgemeinschaft (research unit FOR-1508 to Frieder Mayer) and by the National Geographic Society (research grant WW-057R-17 to G.C.).
Conflict of interest:
The authors declare that no conflict of interest exists.
Author contributions:
G.G.C. conceived of the study. S.P.R. and G.G.C. participated in fieldwork, data collection, analysis, and writing. S.S. created the LPS treatments and participated in writing. All authors gave final approval for publication and agree to be held accountable for the work performed therein.
Data availability: Analyses reported in this article can be reproduced using the data and R-code provided by Ripperger and Carter (2020).
REFERENCES