Accounting for space and uncertainty in real‐time location system‐derived contact networks

Abstract Point data obtained from real‐time location systems (RTLSs) can be processed into animal contact networks, describing instances of interaction between tracked individuals. Proximity‐based definitions of interanimal “contact,” however, may be inadequate for describing epidemiologically and sociologically relevant interactions involving body parts or other physical spaces relatively far from tracking devices. This weakness can be overcome by using polygons, rather than points, to represent tracked individuals and defining “contact” as polygon intersections. We present novel procedures for deriving polygons from RTLS point data while maintaining distances and orientations associated with individuals' relocation events. We demonstrate the versatility of this methodology for network modeling using two contact network creation examples, wherein we use this procedure to create (a) interanimal physical contact networks and (b) a visual contact network. Additionally, in creating our networks, we establish another procedure to adjust definitions of “contact” to account for RTLS positional accuracy, ensuring all true contacts are likely captured and represented in our networks. Using the methods described herein and the associated R package we have developed, called contact, researchers can derive polygons from RTLS points. Furthermore, we show that these polygons are highly versatile for contact network creation and can be used to answer a wide variety of epidemiological, ethological, and sociological research questions. By introducing these methodologies and providing the means to easily apply them through the contact R package, we hope to vastly improve network‐model realism and researchers' ability to draw inferences from RTLS data.


| INTRODUC TI ON
Real-time location systems (RTLSs) allow for spatial positioning and tracking of animate and inanimate objects in real time (Li, Chan, Wong, & Skitmore, 2016). Data sets generated by RTLSs are incredibly versatile and can be used in conjunction with other geographic data (e.g., remotely sensed data) to answer a wide variety of ecological research questions pertaining to individual-and population-level animal behaviors (Kays, Crofoot, Jetz, & Wikelski, 2015).
The most fruitful areas of RTLS data application have been disease ecology and epidemiology. Variation in contact is one of the most important drivers of disease transmission. By quantifying interanimal and environmental (i.e., relating to abiotic components of a given area) contacts, researchers can examine contact variation and the role that social behavior and spatial proximity have in shaping disease transmission in study populations (Chen et al., 2013;Dawson et al., 2019;Harris, Johnson, McDougald, & George, 2007;Leu, Kappeler, & Bull, 2011;Mersch, Crespi, & Keller, 2013;Nagy, Ákos, Biro, & Vicsek, 2010;Spiegel et al., 2016). The integration of contact data with network analysis has led to increased understanding of the drivers of contact and subsequent disease transmission . Early work by Hamede, Bashford, McCallum, and Jones (2009) showed that contact among Tasmanian devils varies between mating and nonmating season, but all members were connected in a single component, making the population highly susceptible to disease spread. Additional studies have further investigated types of social behavior and other interactions underlying transmission (Blyton, Banks, Peakall, Lindenmayer, & Gordon, 2014;Silk, Drewe, Delahay, & Weber, 2018).
Recent advances in RTLS technologies have provided researchers with the tools to more easily, accurately, and consistently identify when and for how long individuals are in contact with one another Mersch et al., 2013;Pfeiffer & Stevens, 2015; Strandburg-Peshkin, Farine, Couzin, & Crofoot, 2015). Real-time location systems based on radio-frequency identification (RFID) and global positioning system (GPS) technologies are becoming increasingly accurate, with positional accuracies often <2 m (Chen et al., 2013;Dawson et al., 2019;King et al., 2012;Schiffner et al., 2018), and able to fix individuals' locations over increasingly small temporal intervals (e.g., 1-10 s) (Dawson et al., 2019;Kays et al., 2015;Schiffner et al., 2018). Increases in RTLS accuracy and fix intervals translate to decreased uncertainty about animals' activities at a given time point Swain, Wark, & Bishop-Hurley, 2008). Accompanying researchers' increased ability to draw inferences about animal behavior from RTLS data, use of RTLS data in animal contact network modeling is becoming increasingly common (Krause et al., 2013;White, Forester, & Craft, 2017). In these network models, nodes (e.g., individuals and specific locations) are connected to one another by edges (i.e., contacts), which often represent instances when ≥2 nodes were observed within a specified distance threshold (SpTh) of one another (e.g., ≤1 m) over a predefined time period (Craft, 2015;Farine & Whitehead, 2015;White et al., 2017).
Contact networks are frequently used to evaluate individuals' behaviors, resource use, and disease transmission risk in wildlife and livestock populations (Croft, Madden, Franks, & James, 2011;, but it is often unclear if proximity-based network edges are truly representative of real-world pathogen transmission opportunities (Craft & Caillard, 2011;Craft, 2015;Davis, Abbasi, Shah, Telfer, & Begon, 2015). In many cases, positional accuracy is a limiting factor when deciding how to define contacts, as researchers cannot identify specific interactions between individuals (e.g., grooming and mating) if spatial accuracy is too coarse (Brookes, VanderWaal, & Ward, 2018;Leu et al., 2011).
As RTLSs only produce point data, even when positional accuracy is ≈100% (i.e., approximately all RTLS-reported coordinates correspond to individuals' true geographic locations) RTLS-derived contact networks may represent an incomplete picture of potential contacts in a given biological system. For example, point location data collected by ear tag-or collar-based tracking devices, which are often deployed in livestock-and wildlife-monitoring studies, respectively (Chen et al., 2013;Dawson et al., 2019;Theurer et al., 2012;Tsalyuk et al., 2019;Strandburg-Peshkin et al., 2015;Swain et al., 2008), are not sufficient for describing the space occupied by individuals' bodies. Therefore, contacts involving areas relatively far from the head cannot be captured without introducing substantial amounts of noise and uncertainty (Dawson et al., 2019;Figure 1).
Uncertainty related to contact precision within a relatively large SpTh leads to epidemiologically (i.e., contacts during which pathogens may be transmitted to susceptible individuals) and sociologically relevant interactions (i.e., contacts representative of specific behaviors known to indicate significant social relationships) involving body parts not equipped with tags being potentially excluded from contact networks or misidentified as noise (Blyton et al., 2014;Dawson et al., 2019). Without this information, network modelers may draw incorrect conclusions regarding the frequency of interanimal interactions (e.g., attraction or avoidance) and pathogen transmission potential in animal populations.
Here, we solve this problem by describing how to incorporate animals' physical space at RTLS fix intervals into RTLS-derived animal contact networks, ensuring that signal capture pertaining to the whole of tracked individuals' physical space is maximized. We present novel procedures for deriving polygons from RTLS point data while maintaining distances and orientations associated with individuals' relocation events (see Section 2.1) and demonstrate the versatility of this methodology for network modeling using three network creation examples (see Section 2.3). Additionally, in creating our networks, we establish a procedure to adjust definitions of "contact" to account for RTLS positional accuracy.
Thus, we ensure that all true contacts in our systems of interest are likely captured and represented in generated networks. By introducing these methodologies and providing the means to easily apply them through the contact R package, we hope to vastly improve network-model realism and researchers' ability to draw inferences from RTLS data.

| Steps for polygon derivation
Accounting for objects' physical space in real time involves interpolating polygon vertices from RTLS data points. By doing so, we create 2-dimensional objects representative of areas covered by tracked individuals' bodies from 1-dimensional objects describing RTLS tags' point locations. Throughout this section, we refer to an example wherein we want to generate polygons covering each individual calf whose point locations are reported by a cattle monitoring RTLS ( Figure 2). All terms described in Section 2.1 are listed in Table 1.
For n tracked individuals, we define a set of planar RTLS data (x, y)-coordinate pairs as {loc i } for individuals i = 1, … ,n, at sequential fix intervals t = 1, … ,T, where T is the total number of fix intervals over the course of the study period. Each polygon vertex, V itl , is derived from a single RTLS-reported point location contained in {loc i } , loc it (i.e., (x loc it , y loc it )) and denotes a specific (x, y)-coordinate pair, (x itl , y itl ). The variable, l = 1, … ,L, identifies unique poly it vertices. Each polygon, poly it , represents the area contained within the vertex set , and L is an integer ≥ 3 describing the number of vertices in {V it }. For example, if poly it is defined using four vertices, unique vertices in Effectively, we want to transform each unique point location in a data set into a unique polygon with L vertices. Before we can derive {V it }, however, we must first consider where each V itl is located relative to a unique loc it on individuals' bodies. In other words, we know where tracking devices are located on animals' bodies (e.g., ear and neck), but before we can transform these point locations into polygon vertices, we must decide where these new points will exist on animals' bodies as well (e.g., nose and tail). In our calf example, tags are located on the left ear of each individual, and we assume animals' sizes and proportions were equivalent and stable over the observation period ( Figure 2a). We decide a priori where {V it } will be located on planar, polygonal representations of space around of animals' bodies, which we refer to as "planar models" (Figure 2b).
We use the star denotation to distinguish variables in planar models from their empirical counterparts (e.g., loc * it and V * itl ). Area described by each poly it is restricted to the shape presented in these planar models, however, this limitation can be overcome to some extent by creating different models for each tracked individual, and/or updating planar models over time (t).
). This is the Euclidian distance between loc * it and each V * itl , and is equivalent to the distance between loc it (i.e., RTLS-reported point location) and V itl (i.e., desired polygon  poly it Area contained within vertices described in {V it }|L. loc i The most-recent previously reported location for individual i with a different (x, y)-coordinate pair than loc it (i.e., ≤ t − 1).
it If gyroscopic data are available: the observed angle of movement reported by a gyroscopic measurement device (e.g., gyroscopic accelerometer) at time t. If no gyroscopic data are available: the absolute angle of line loc i loc it measured from a horizontal axis intersecting loc i . The planar-model counterpart to loc i ; describes an assumed location of loc * it at time , and is used to identify the angular orientation of the modeled individual.

{V *
it } A set containing the (x, y)-coordinate pairs of vertices described in a planar model; indicates where vertices should exist relative to The planar-model counterpart to ; describes the absolute angle of line loc * i loc * it measured from a horizontal axis intersecting loc * i . dist itl The Euclidean distance between loc it and V itl . * l The absolute angle of line V * itl loc * it measured from a horizontal axis intersecting loc * it .
vertex). Once we know the distance between loc it and the vertex of interest, we can (c) identify (x, y)-coordinate pairs that lie dist itl planar units (e.g., meters) from loc it in a 360 • − ( * − ( * l + it )) counter-clockwise direction relative to a horizontal axis intersecting loc it ( Figure 2c). This is the transformation. (d) Repeat steps 2 and 3 for each vertex l.
In the above formula, * l is the absolute angle of line V * itl loc * it measured from a horizontal axis in loc * it . The variable it is the observed angle of movement reported by a gyroscopic measurement device (e.g., gyroscopic accelerometer) at time t and allows us to account for changes in the orientation of animals' bodies attributed to movement while keeping * l fixed. Incorporating this variable into the {V it } derivation formula ensures that {poly i } appropriately represents animals' physical orientation (i.e., what direction they face at time t), which may change between times t and t + 1. In many cases, it may be unknown. For example, if gyroscopic and RTLS data were not collected concurrently (e.g., animals were outfitted with GPS transmitters, but not gyroscopic accelerometers), researchers would not intrinsically know animals' orientations. In these cases, it can be estimated by calculating the absolute angle of line loc i loc it measured from a horizontal axis intersecting loc i , the most-recent previously reported location for individual i with a different (x, y)-coordinate pair than loc it (i.e., ≤ t − 1). The variable * is the planar-model counterpart to it and describes the shape's original orientation.

| Assumptions and limitations of polygon derivation
There are a couple limitations that researchers must take into account when using this procedure. Firstly, when deriving polygon vertices from RTLS points, researchers must justify how polygons relate to real-world physical space by clearly explaining rationales for polygons' shapes, sizes, and behaviors. As previously noted, areas represented by polygons are rigid and restricted to shapes described in planar models. Though these shapes can be updated over time, to elevate the likelihood that polygons truly represent real-world spatial features, our polygon derivation methodology is best used to model space with never-changing or infrequently changing dimensions. For example, because the size and shape of a baboon's body frequently changes based on its activities (e.g., walking baboons are quadrupedal, but they often sit on their haunches when stationary), using our methodology to create polygons representative of baboons' physical bodies may produce inaccurate results. Conversely, as ungulates' body shapes and sizes are generally unchanging over short time periods, when modeling these species, we can be relatively confident that polygons generated using our methodology consistently reflect real-world physical space. This is not to say that our methodology cannot be used to model regularly changing shapes, however. In these cases, researchers must utilize multiple planar models (i.e., one for each spatial form), determine criteria for switching between them (e.g., use one model when animals are observed moving slower than a specified speed, and another when their speed exceeds the stated limit), and accept that the added complexity of the system may increase risk of erroneous inference.
Secondly, in the absence of paired gyroscopic data, when it must be estimated, we must make four assumptions to account for directionality changes associated with animal movement while maintaining positional relationships between V it and loc it . First and foremost, (a) we assume that RTLS fix intervals are sufficiently small and allow RTLSs to capture all changes in animals' movement direction (i.e., animals do not face unknown directions in-between fix intervals). The minimum required temporal resolution will vary based on the system being modeled. For example, if modeling an animal that is largely sessile and slow moving, we may assume that 10min fix intervals are sufficient for capturing movement directions.
When modeling frequently moving animals, however, sub-minute fix intervals are likely required to capture all directional changes.
Additionally, (b) because we rely on observed animal movements to define it , we cannot know which direction animals are facing until the first relocation event occurs. Thus, we cannot create polygons representative of animals' physical orientations at the first time point, or any time points before relocations occur (i.e., in poly it , t ≥ 2).
Furthermore, (c) we assume that individuals only move forward and in a straight line, as is common practice when calculating many pathbased movement metrics (e.g., angle of movement and step length; Miller, 2015). Finally, (d) when creating polygons representative of space occupied by animals' bodies, we assume that when the length of line loc i(t−1) loc it is below a certain threshold (e.g., 0.1 m), individuals' physical locations and orientations remain unchanged. This immobility threshold allows us to discount orientation changes due to observed movements so miniscule that the majority of the modeled physical space is likely unaffected (e.g., head shaking), or movements caused by inaccurate RTLS reporting.

| Data sets
In the following subsections, we generate direct contact (see Section 2.3.3) and visual contact (see Section 2.3.4) networks using two previously published RTLS-generated data sets, which we refer to as calves and baboons. Neither of these data sets include any gyroscopic information about animals' movements. Therefore, as part of the polygon derivation procedure, we estimated it values using the previously described calculation and accepted the associated assumptions and limitations.
In a previous paper (Dawson et al., 2019), we published the calves data set, which contains RTLS data for n = 70 beef cattle (Bos taurus) calves confined in a feedlot pen. Calves were approximately 1.5 years old with estimated 1.5-m nose-to-tail lengths and 0.5-m shoulder widths. Data were obtained using a radio telemetry-based RTLS, where 90% of points fell within ±0.5 m of individuals' true locations, at a temporal resolution of 5-10 s (i.e., fixes for each individual were obtained every 5-10 s) on 2 May 2016. To standardize the temporal resolution of this data set at 10 s, we smoothed individuals' movement paths (i.e., observed consecutive relocations) using the methodology we previously described in Dawson et al. (2019), and by doing so we obtained (x, y) coordinates representative of individuals' average location at each 10-s interval in the study period.
The baboons data set, collected by Strandburg-Peshkin et al. handling-induced influences from the data, we removed the first and last days of data in baboons. Additionally, we removed the first and last hours from each day in the data set (i.e., 03:00:00-03:59:59 UTC and 14:00:00-14:59:59 UTC). We did this because the number of individuals observed during each second of these hours was highly variable, an effect potentially caused by tracking devices powering on/off at different rates during these periods. Finally, we standardized the temporal resolution of our subset at 1-s fix intervals by smoothing individuals' daily movement paths (Dawson et al., 2019). Thus, we were able to create a baboons subset containing 23 animals' geographic locations at 1-s fix intervals between 04:00:00 and 13:59:59 UTC from 2 August to 13 August 2012. We used this subset for polygon derivation and subsequent network creation. As our polygon derivation methodology requires animals' locations to be expressed as planar coordinates, we transformed the data using an azimuthal equidistant projection centered on the data centroid (Barmore, 1991).

| Processing software
To simplify polygon derivation and network creation, we developed the contact package for R (v. 3.6.0, R Foundation for Statistical Computing). This package is available for download on the CRAN repository and was specifically built to process spatiotemporal data into point-or polygon-based contact and social networks (Figure 3).
It contains 20 + functions for cleaning, interpolating, randomizing, and creating networks from spatiotemporal data, and the principal functions are briefly described in Table 2. All RTLS data processing was carried out in R using RStudio (v. 1.1.463, RStudio Team), utilized contact functions, and is described in Appendices S1 and S2.

| Direct contact network creation
We know that in animal populations, social interactions can increase the risk of pathogen transmission within dyads (Drewe, 2010;Blyton et al., 2014). In animal production systems, enteric pathogens (e.g., E. coli and Salmonella spp.) are often present on animals' hides, where they can be directly transmitted to hosts during social interactions or bumping (Keen & Elder, 2002;Nastasijivec, Mitrovic, & Buncic, 2013;Villarreal-Silva et al., 2016). Because social relationships between cattle frequently involve increased physical contacts between dyad members (e.g., grooming, mounting, and butting; Gibbons, Lawrence, & Haskell, 2009;Horvath & Miller-Coushon, 2019;MacKay, Turner, Hyslop, Deag, & Haskell, 2013), there is likely an increased risk for hideto-hide or hide-to-mouth pathogen transfer between socially interacting individuals (Blyton et al., 2014). We aimed to create networks representative of direct contacts between calves (i)  findDistThresh a Sample from a multivariate normal distribution to create "in-contact" point pairs based on RTLS accuracy, and generate a distribution describing average distances between point pairs.
randomizePaths Generate randomized movement paths over defined temporal intervals for each individual according to methods described by Spiegel et al. (2016). repositionReferencePoint a Translates planar point locations to a different location fixed distances away, given a known angular offset, while maintaining angular orientations of movements. This function is the basis for polygon derivation from point locations, as it allows for vertex placement around planar models.
contactTest Compare empirical contact distributions to null models using various testing methods (e.g., χ 2 goodness-of-fit, Mantel, 1967) to evaluate if observed contacts occur more or less frequently than would be expected at random, respectively.
a Indicates functions based on novel procedures described within this manuscript. We then calculated the upper 99% CI for the resulting expected distance distribution to be used as our adjusted SpTh value for contact network creation. In this way, we estimated that a SpTh of 0.56 m likely captures ≥99% of contacts, as previously defined (i.e., polygon intersections).
To demonstrate differences resulting from differing contact definitions, we created two distinct categories of contact network sets. In the "precise" set, contacts were said to occur when polygons intersected (i.e., SpTh = 0; Figure 6), and in the "expected" set contacts occurred when polygon edges were within 0.56 m of one another. The "expected" set can also be considered to have been created using relatively large polygons compared to the "precise" set ( Figure 7). Each network set contained three time-aggregated, undirected contact networks: (a) the "fullBody" contact network describing any instance of polygon intersection (i.e., ∑ n i=1 {Poly Calf it }∩{Poly Calf }) or interpolygon distances ≤ 0.56 m, (b) a "head.head" bipartite contact network describing instances when head polygons intersect (i.e., ∑ n i=1 {Poly H it }∩{Poly H }), or are ≤0.56 m from one another, and (c) a "head.posterior" bipartite contact network describing instances when head polygons intersected (i.e., ∑ n i=1 {Poly H it }∩{Poly P }) or were within 0.56 m of posterior polygons. In each of these networks, edge formation was limited to polygons describing different individuals (i) (e.g., no polygon-based intersection can exist between Poly H it and Poly A it ). Network edges associated with each dyad were weighted by contact frequency over the 24-hr study period. We used Welch's ANOVAs (Welch, 1947) and post hoc Games-Howell tests (Games & Howell, 1976) to evaluate differences in mean node degree, contact duration (i.e., number of consecutive time points edges existed between node pairs), and per-capita sum contacts between all networks. Additionally, we used two-sided Mantel tests (Mantel, 1967) to evaluate correlations between intra-set contact matrices (i.e., "precise" and "expected" sets were evaluated separately). Mantel tests were each based on 10,000 graph permutations, and for all statistical analyses we set an α-level of 0.05. We did not evaluate correlations between "precise" and "expected" matrices because contact definitions are mutually exclusive and would not be concurrently implemented when modeling real-world systems. Code for polygon and network creation can be found in Appendix S1.

F I G U R E 5
Implications for real-time location system accuracy on proximitybased contact determination. Reported point locations (dark blue) may not necessarily represent true locations, but rather, will fall within a certain true location range (beige). As such, in-contact points like those shown in the inset figure may misrepresent interactions within the tracked population, if true locations actually fall outside of contact-threshold distances (light blue) from one another

| Visual contact network creation
Primate social behaviors are often driven by visual cues (Bielert & Van der Walt, 1982;Janson & Di Bitetti, 1997) We initially defined "contact" as occurring when a GPS point, loc it , intersected a polygon, Poly VIS it (i.e., distance between loc it and Poly VIS it equaled 0). We adjusted this SpTh to account for accuracy of the baboons data set (i.e., approximately 100% of points fall within ±0.26 m from true locations) using the methodology described in Section 2.3.3. By sampling 1,000,000 in-contact point location pairs from the multivariate distribution  [[0,0,0,0],(0.26∕3.89) 2 I], we determined that a SpTh value of 0.109 m likely captures ≥ 99% of contacts, as previously described. Edges in this bipartite network, with independent node sets {Poly VIS } and {loc}, were weighted by contact frequency over the study period. We report the mean per-capita number of expected contacts per second, as well as the mean observed duration of contacts and daily node degree (i.e., number of baboons within individuals' visual fields). Code for visual contact network creation and summarization can be found in Appendix S2.

| Calf networks
All ANOVA results indicated that differences in network metrics existed, with p-values < 2.2e −16 , and post hoc Games-Howell test  Table 3. On average, "expected" contact networks consistently had greater contact durations and per-capita sum contacts than their "precise" counterparts, highlighting the effect of relatively large SpTh values on network realization. Our results suggest that these metrics scale with polygon size. That is to say, just as increasing SpTh values lead to inflated contact frequency in pointbased proximity contact networks (Dawson et al., 2019), our work here suggests that the presence of larger polygons translates to increased probability that polygons intersect, and therefore morefrequent and longer-duration contact events. Average node degree generally followed the same trend, but all graphs aside from the precise-set "head.head" one were nearly complete.
We also observed strong correlations between intra-set graphs ( Figure 8). Mantel tests suggested that all intra-set matrices were related, with a p-value < .001. We found that "fullBody" graphs were consistently moderately to highly correlated with others, which is not surprising given that the "head.head" and "head.posterier" graphs were subsets of the former. Furthermore, in the case of the expected-set "head.head" and "head.posterier" graphs, when graphs were not subsets of one another but polygons involved in contacts overlapped (Figure 7), we observed a relatively high correlation value.
The presence of a moderate correlation between the precise-set "head.head" and "head.posterier" graphs, which did not overlap, is especially interesting. Though we did not examine specific dyadic relationships and potential correlations at the dyad level, our findings suggest that animals with more head-to-head contacts will likely report increased head-to-posterior contacts as well. This means that when modeling social relationships in cattle populations, it may be sufficient to use head-to-head interactions alone to identify dyads with high social affinity. On the other hand, this is not necessarily the case for modeling pathogen transmission. Assuming that our "precise" and "expected" networks reflect true interactions at least to some extent and that observed contacts are not solely a function of differences in polygon sizes, our results suggest that head-to-head contacts occur less frequently than head-to-posterior contacts, but the two contact types are inter-related in this system. Presumably then, under the assumption that probability of transmission given contact is stable, RTLS-derived direct pathogen transmission models of similar systems wherein only head-to-head contacts are effectively represented (Chen, Ilany, White, Sanderson, & Lanzas, 2015;Dawson et al., 2019) likely under-represent dyadic interactions where pathogens may be transferred from the posterior of one animal to the mouth of the other, or vice versa.
We must note here that these findings are based on analyses of data collected over a single day and therefore may not be wholly reflective of contact patterns in this population. That said, we have demonstrated that transforming point locations into bodily polygons (e.g., animal heads, and posteriors) allows us to characterize observed contact events based on what polygons intersect (e.g., head-to-head). By doing so, we gain the ability to assess how different modes of contact, which may be indicative of different social behaviors (Figure 6), may affect pathogen transmission. Thus, contacts involving RTLS-derived polygons can provide insight into both physical contact-mediated direct pathogen transmission events, which are difficult, if not impossible, to observe in many field studies (Blyton et al., 2014).

| Baboon network
We found that, on average, baboons observed 5.39 (SD = 1.02) other tagged individuals at any given second and visual contacts lasted an average of 3.67 (SD = 4.95) seconds. The maximum duration of a visual contact was 701 s (i.e., ≈12 min), and the average daily degree was 18.13 (SD = 1.74). It is necessary to note that, though we defined "visual contacts" as instances when baboon points were observed within visual field polygons, in actuality, observers may not have necessarily been actively watching "contacted" individuals during these time points (e.g., observers' eyes may have been closed, they may have been otherwise focused on other objects). Furthermore, we assumed that baboons' views were unobstructed and viewing distances were stable during the study period, which is almost certainly an oversimplification of real-life vision. Future studies may TA B L E 3 Mean network connectivity metrics for contact networks with and without RTLS accuracy adjustment (i.e., "expected" and "precise" network sets, respectively)

| Data processing considerations
Previous work has described in detail how difficult defining animal interactions from RTLS point data can be, as "contact" definitions must be specific to the system researchers are attempting to model (Craft, 2015;Farine & Whitehead, 2015;White et al., 2017). When defining point-based contacts, researchers must clearly describe their rationale for selecting contact definitions, and because each definition inherently makes a number of assumptions (e.g., animals outside a given distance threshold do not pose an infection risk), network modelers must also acknowledge these unique assumptions and associated limitations in their work (Dawson et al., 2019).
Defining polygon-based contacts is less ambiguous, as interactions occur when spatial objects (i.e., points, lines, or polygons) intersect (Mersch et al., 2013 modeled. Increasing the SpTh/polygon area using our procedure will likely introduce noise into the system (Dawson et al., 2019), but without doing so, researchers cannot be confident that a majority of real-world contacts are truly represented in generated contact networks. Luckily, animal tracking technologies (e.g., global positioning system and radio telemetry tags) are advancing rapidly, becoming increasingly lightweight and accurate Thomson et al., 2017). As these technologies advance, and newer devices are deployed, the need to inflate SpTh values will decrease, and resulting contact networks will better reflect real-world interactions.
F I G U R E 8 Calf network comparisons. (a) Correlation plot describing the sign and magnitude of correlations between "precise" networks. (b) Correlation plot describing the sign and magnitude of correlations between "expected" networks Aside from the aforementioned nuanced difference in how contacts are defined, polygon data can be stored and processed in much the same ways as point location data (e.g., network data can be stored as adjacency lists, and edge lists). One process that necessitates additional consideration for polygon-based networks, however, is network randomization. Network randomization procedures traditionally involve randomizing point locations prior to contact network creation, generating null models wherein contacts occur at random, then comparing null and empirical models to test hypotheses about contact occurrence (Farine & Whitehead, 2015;Spiegel et al., 2016;Farine, 2017). Polygons derived from point locations can also be randomized to create null models using the same methodologies, but researchers must decide a priori if randomization procedures will be implemented before or after polygon generation.
If polygons are to be oriented using gyroscopic data rather than RTLS data (i.e., if researchers do not rely on observed animal movements to define it values), there would be no difference in randomization outcomes regardless of the chosen order. If polygon orientations are to be calculated using point location data alone, however, randomizing point locations prior to polygon derivation will also randomize subsequently calculated polygon orientations.
Alternatively, if randomization procedures were to be implemented following polygon creation in this example (i.e., polygon locations themselves are randomized), polygon orientations will reflect those described in the empirical data set. Either randomization protocol described herein can be a useful tool for hypothesis testing and can be easily implemented through the "randomizePaths" function in our contact package.

| CON CLUS ION
Using the methods described herein and the associated contact package for R, researchers can derive polygons from RTLS points.
We have demonstrated these polygons are highly versatile for contact network creation and can be used to answer a wide variety of epidemiological, ethological, and sociological research questions.
We hope that by utilizing our methods and the tools provided, researchers can vastly improve network-model realism and increase their abilities to draw inferences from RTLS data sets. tions. Finally, we want to thank the journal editors and reviewers who helped us to revise and refine this manuscript. Without their insight, the procedures we describe herein would be less accessible and much more difficult for readers to understand.

CO N FLI C T O F I NTE R E S T
None declared.