Validation of an automated data collection method for ... - Springer Link

Behav Ecol Sociobiol (2014) 68:1379–1391 DOI 10.1007/s00265-014-1757-0

METHODS

Validation of an automated data collection method for quantifying social networks in collective behaviours Fumiaki Y. Nomano & Lucy E. Browning & Shinichi Nakagawa & Simon C. Griffith & Andrew F. Russell

Received: 10 April 2014 / Revised: 12 June 2014 / Accepted: 12 June 2014 / Published online: 4 July 2014 # Springer-Verlag Berlin Heidelberg 2014

Abstract The social network of preferences among group members can affect the distribution and consequences of collective behaviours. However, the behavioural contexts and taxa in which social network structure has been described are still limited because such studies require extensive data. Here, we highlight the use of an automated passive integrated transponder (PIT)-tag monitoring system for social network analyses and do so in a novel context—nestling provisioning in an avian cooperative breeder, for which direct observation

Communicated by L. Z. Garamszegi Electronic supplementary material The online version of this article (doi:10.1007/s00265-014-1757-0) contains supplementary material, which is available to authorized users. F. Y. Nomano (*) Graduate School of Environmental Sciences, Hokkaido University, Sapporo 0600810, Japan e-mail: [email protected] L. E. Browning Department of Zoology, University of Cambridge, Cambridge CB2 3EJ, UK L. E. Browning Fowlers Gap Arid Zone Research Station, School of Biological, Earth and Environmental Sciences, University of New South Wales, Sydney NSW 2052, Australia S. Nakagawa Department of Zoology, University of Otago, Dunedin 9054, New Zealand S. C. Griffith Department of Biological Sciences, Macquarie University, Sydney NSW 2109, Australia A. F. Russell Centre for Ecology and Conservation, College of Life and Environmental Sciences, University of Exeter, Penryn TR10 9FE, Cornwall, UK

of social behaviours is difficult. First, we used observers and cameras to arrive at a suitable metric of nest visit synchrony in the PIT-tag data. Second, we validated the use of this metric for social network analyses using internal nest video cameras. Third, we used hierarchical regression models with ‘sociality’ parameter to investigate structure of networks collected from multiple groups. Use of PIT tags led to nest visitation duration and frequency being obtained with a high degree of accuracy for all group members, except for the breeding female for whom accurate estimations required the use of a video camera due to her high variability in visitation time. The PIT-tag dataset uncovered significant variability in social network structure. Our results highlight the importance of combining complementary observation methods when conducting social network analyses of wild animals. Our methods can also be generalised to multiple contexts in social systems wherever repeated encounters with other individuals in closed space have ecological implications. Keywords Social network . Collective behaviour . PIT tag . Cooperative breeding . Hierarchical model

Introduction Individuals within populations commonly show non-random social associations with conspecifics. The social preferences underpinning such associations between individuals in a population can be characterised using a social network (Croft et al. 2008; Sih et al. 2009). Over the past decade, significant evidence has accumulated in support of the occurrence of heterogeneous social network structure within animal populations and in a variety of contexts (e.g. group membership, Croft et al. 2005; affiliative behaviour, Blumstein et al. 2009; Voelkl and Kasper 2009; collective behaviour/movement, Nagy et al. 2010). This body of work provides important

1380

new insights into the ecological causes and consequences of social behaviour. For example, descriptions of social network structure are important for the studies of infectious disease (Eubank et al. 2004) and mating systems (e.g. Streatfeild et al. 2011) and suggest that individual behaviour within a group might be less independent than is typically considered (Naug 2009; Voelkl and Kasper 2009; Bode et al. 2011). Despite this, our awareness of the ecological and demographic contexts in which social networks can be quantified in the animal kingdom is still relatively limited. The paucity of empirical work is partly caused by a lack of efficient methods for collecting large amounts of data on social interactions from wild animals over the diversity of behavioural contexts. Data collection for social network analysis is challenging since it requires substantial interaction data for every combination of individuals within a social group or population. Recent developments of automated monitoring systems for wild animals provide one potential solution to this problem (Krause et al. 2011). Automated behavioural recordings are particularly useful in species in which individuals are difficult to observe due to their cryptic nature or fast movement during social interactions, or where they do not habituate well to the presence of observers. Although behavioural monitoring of relatively large animals with GPS (e.g. Leu et al. 2010; Nagy et al. 2010) and proximity (e.g. Ji et al. 2005; Rutz et al. 2012) loggers is growing, the financial costs of attaching such loggers simultaneously to all individuals in multiple groups can be prohibitive, and small animals can seldom carry such loggers for sufficient durations. Passive integrated transponder (PIT) tags or radio frequency identification (RFID) tags provide a low-cost method for automated monitoring of individual behaviour (Gibbons and Andrews 2004; Bridge and Bonter 2011). These tags are typically a 1×6- or 2×12-mm microchip, similar to those used in pet identification, that carry a unique identity code that can be detected by an antenna linked to a reader. The antennae may be set where social interactions are expected and with the readers recording individual identity along with date and time of detection. Depending on the contexts in which the data are collected, the co-occurrence of individuals in antennae proximity can be used as an indicator of general social preference (Kerth et al. 2006; Patriquin et al. 2010) or its temporal patterning relating to collective decision making (Kerth et al. 2006; Gillam et al. 2011). The PIT-tag reader allows monitoring of relatively small animals for days at a time, far longer than conventional methods of human observers and video camera, since they consume little battery power, and the simple data structure of PIT tags (tag ID, date and time) requires little memory. Its application has facilitated the analysis of social network structure in species for which direct observations are difficult (e.g. bats, Kerth et al. 2006; Patriquin et al. 2010; small passerine birds, Aplin et al. 2012). However, the use of PIT-tag data for social network

Behav Ecol Sociobiol (2014) 68:1379–1391

analysis is still limited, and the behavioural contexts that can be studied with PIT-tag approaches have not been fully explored. For example, PIT-tag data have typically been restricted to instances of co-occurrence with relatively long intervening time intervals (e.g. home range overlap, Perkins et al. 2009; day by day changes in roost membership, Kerth et al. 2006, but see Robinson et al. 2009; Aplin et al. 2012, 2013; Farine et al. 2012 for exceptions). Notwithstanding potential shortcomings of using PIT-tagtype techniques, such as the simplicity of data collected, they offer one solution to the primary logistical issue of social network analyses—the cheap and effective collection of substantial quantities of dyadic data. In addition, PIT tags can be used in a significant range of ecological context, including individual visitations to feeders (Mariette et al. 2011; Aplin et al. 2012), roost sites (Kerth et al. 2006; Patriquin et al. 2010; Gillam et al. 2011) and breeding nests (Robinson et al. 2009; Shen et al. 2010; Mariette and Griffith 2012), potentially allowing social networks to be evaluated across contexts within and among species (Aplin et al. 2012; Farine et al. 2012). In this paper, we propose and validate a technique to quantify social networks using PIT-tag monitoring systems in a novel context: synchronous provisioning at breeding nests of a cooperatively breeding passerine, the chestnut-crowned babbler (Pomatostomus ruficeps). Our rationale for using offspring provisioning is to add to the range of contexts in which PIT tags can be used to measure networks and to illustrate that any ecological context can be used as long as it involves animals returning repeatedly to a common location and the method is validated carefully. Furthermore, provisioning of offspring is functionally relevant: not only does it have significant fitness consequences for offspring but have been suggested to represent a form of signalling (Zahavi 1974; Gaston 1978). We first validate the PIT-tag method by calibrating it with nest observations using human observer, external video camera and nest cameras. Second, we compare estimates of nest visit synchronicity and the ingredients of social network analyses, between nest cameras and PIT tags. Third, we confirm that social network structure at both dyad and individual levels shows heterogeneity as expected from non-random interactions among carers, revealing the validity of a PIT-tag system approach in a novel context despite the simplicity of the data generated. For this latter aim, we advocate the use of hierarchical regression models for social network data (Hoff et al. 2002; Cross et al. 2012), which remove the problems inherent in the more commonly employed approach of fitting separate models for each network and attempting to compare between them. Finally, we evaluate and compare the three metrics of network structure using PIT tags and internal nest cameras in order to guide future studies as to the pros and cons of using this method for social network analyses. Our hope is to both encourage and provide directions for future investigations of


social networks in a context for which data can be easily, and is routinely, collected but for which analysis of social networks have hitherto been ignored.

Methods Study species and population The chestnut-crowned babbler is a medium-sized (50 g) passerine bird endemic to semi-arid and arid zones of the southeastern Australia. We have studied a population of this species since 2004 at the University of New South Wales Arid Zone Research Station, Fowlers Gap, NSW, Australia (141.39° E, 31.09° S). The study site, located in the arid zone of far western NSW, is characterised by open chenopod shrubland, with the few trees being largely limited to short linear stands along drainage zones and (dry) creek beds (Portelli et al. 2009; Sorato et al. 2012). Chestnut-crowned babblers breed in units of 2–15 individuals (mean=~6) that include a single breeding female, one or more breeding males and non-breeding helpers (primarily males) (Browning et al. 2012b; Rollins et al. 2012). Nestlings are provisioned from dawn to dusk (ca. 06:00– 18:00 h) for around 23 days before fledging (Russell et al. 2010). Over 90 % of group members have been captured using mist nets, with each individual banded with one uniquely numbered metal band (Australian Bird and Bat Banding Scheme) and three colour bands, as well as being injected subcutaneously in the flank with a 2×12-mm PIT tag (Trovan Ltd., UK). The colour band combinations represent a standard method for identifying birds in the field but are an inappropriate method for quantifying provisioning behaviours in babblers due to their rapid movements and large group sizes. PIT tags can be fitted externally (e.g. Ottosson et al. 2001) or injected subcutaneously (e.g. Nicolaus et al. 2008; Schroeder et al. 2011), with the former being most common in bird studies and the latter being more common in mammals. We advocate the latter where possible because external attachment, such as to leg bands in birds, will result in increased tag loss and the possibility of the tag ensnaring a bird (Jamison et al. 2000; Nicolaus et al. 2008). While care is obviously required with subcutaneous injection, particularly in small animals (e.g. Kurth et al. 2007 for an invertebrate), our experience over the past 7 years in babblers suggests that as long as PIT tags are inserted fully in a downward direction, sealed with standard surgical glue and inserted in an area where the pressure between muscle and skin is minimal, PIT tags can remain in the birds for several years. As such, the use of inserted PIT tags offers a long-term means of gathering large amounts of accurate data for a range of questions (e.g. Browning et al. 2012a, b; Nomano et al. 2013; Young et al. 2013), in a way that causes no obvious deleterious effects to the birds (Nicolaus et al. 2008; Schroeder et al. 2011).

1381

Estimating nest visitation synchrony using standard nest observation techniques Although we advocate the use of PIT-tag systems to generate large amounts of simultaneously collected provisioning data, such data requires validation from standard observation techniques (see below). In the case of this study, we used external nest observations to define a group of synchronous visits, as this forms the basic ingredient of social network analyses and is difficult to determine from PIT-tag data in isolation. We used two approaches to arrive at our definition of synchrony. First, we used 2-h-long direct observations at the nests of 16 breeding units in 2004 when nestlings were 10–15 days old (Russell et al. 2010). As the majority of trees (and hence nests) are located in creek beds, nest observations were conducted by an observer situated at a distance of at least 50 m from the nest tree with a view angle perpendicular to the creek: closer distances affected the behaviour of the birds. This enabled a clear view of both the nest and approximately 200 m of creek in either direction. Given that babblers almost always approach the nest along the creek line (Sorato et al. 2012), this enabled the temporal patterns of nest visitation to be determined. Breeding units varied from 2–13 (mean=6) individuals, and a total of 768 individual nest visits were recorded from 253 ‘group’ visits to the nest area. Second, in 2010, we set up a video camera (Sony Handycam HDR-XR150, Sony Corporation, Japan) framed exclusively on the nest tree from a distance of ~20 m, again positioned with a view perpendicular to the creek. These external video cameras (N=3 nests) monitored the nests for ~14 h in total over 4 days (brood age 10–16 days). The primary purpose was to determine for each nest visit: (a) the proportion of group members arriving at the nest tree together (i.e. synchronously) and (b) the time interval between successive entrances by different group members when they arrived at the tree synchronously versus asynchronously. Together, these observations permitted us to determine how synchronous versus asynchronous nest visits arose and differed in terms of inter-visit intervals; this in turn permitted assignment of synchronous versus asynchronous nest visits in the PIT-tag data. We defined a series of visits in the PIT-tag data and nest video data as a synchronous group when temporal separation between nest entry times of successive visits was less than 1 min apart (‘chain rule’ commonly used to define spatial association; Croft et al. 2008; see results). Nest visit duration and frequency Provisioning data were collected from July to November in 2007 and 2008. Nest visits were automatically recorded using an LID-650 PIT-tag reader (Trovan Ltd., UK) placed at the bottom of the nesting tree and attached via a cable to a coil antenna placed around the entrance of the breeding nest. Babblers breed in domed nests with a small (~8-cm diameter)

1382

entrance hole, positioned near the top of the nest. Fitting a coil of the same diameter ensured that all birds had to pass through the antenna to access the nest. The readers recorded the identity of all nest visitors, as well as the date and time of nest visitation to the nearest second. A 12-V 7.2-A h battery powered the reader and monitored the nest 24 h/day continuously for up to 5 days before the data was downloaded and the battery was replaced. However, the duration of the monitoring period varied because of the need to move readers between concurrently active nests. Overall, PIT-tag monitoring durations ranged from 1 to 19 full days (mean=9.4 days, SD=5.3) across 49 breeding attempts by 32 breeding units). Interpretation of PIT-tag records is not necessarily straightforward where the reader records a pass through the nest entrance without the direction of movement being recorded, as was the case in this study using a single antenna for logistical reasons (see also Mariette et al. 2011; Mariette and Griffith 2012). In babblers, multiple records occur for each nest visit, since individuals commonly remain in the proximity of the antenna for a few seconds before entering and/or exiting the nest. Further, the nest of this species is large (~40 cm in depth), and birds are out of range of the antenna while they are inside the nest chamber feeding young nestlings. To overcome these problems, we used nest cameras in a subset of nests to determine precisely nest visitation durations and frequencies. To this end, MO-S408 pen cameras measuring ~10 mm in diameter (with a 3.1-mm pin hole lens) (Misumi Electronics Corporation, Taiwan) were integrated into the PIT-tag reader system and inserted through the roof of the nest (secured with cable ties) to film the behaviour of birds within (Browning et al. 2012b; Young et al. 2013). Because of the extra battery power required to run the nest camera systems, batteries had to be changed daily, reducing the overall amount of data that could be collected with this system. Overall, nest camera systems were established for 1.1 to 13.6 h/day (mean=4.6, SD=3.5), over 1–6 days (mean=2.6, SD=1.7) in 22 breeding attempts by 19 breeding units. The data obtained from the integrated system allows us to determine, with a high degree of accuracy, nest visit duration and frequency from the PIT-tag system. Characterising the social networks We calculated the number of synchronous visits (as defined above) between all possible dyads of individuals within a breeding group to explore an individual’s social preference during the synchronous visits and used this data to characterise the social network. The observed number of synchronous visits per dyad was used as a measure of the strength of connection for each dyad. The number of encounters at the nest Yi,j for dyad i, j contains both passive encounters principally caused by close visits occurring by chance and synchrony events due to social preference. In other words, individuals frequently visiting a nest can be synchronous more often by


chance than individuals with lower visit rates, even if they have no social motivation to synchronise their visits. We examined three potential indices to capture this effect of individual visit counts and chose Yi,j/min (Ni, Nj) as a synchrony frequency index where Ni (or Nj) is the visit count of individual i (or j) (see Electronic Supplement). This is considered as the strength of synchrony between i and j and was used as a weight of network edge. Example networks for some breeding units were drawn using an R package latentnet (Krivitsky and Handcock 2008). Extracting ‘social’ components from the synchronous nest visits An important aspect of our validation is to investigate whether the social network structure quantified with PIT-tag captures any social component of the study system. We examined whether the synchrony networks obtained from the PIT-tag system are more variable at dyad and individual level than expected when the synchrony is caused solely by passive encounters. For this purpose, we employed generalised linear mixed model (GLMM) incorporating a term called sociality (Hoff et al. 2002; Cross et al. 2012). We assembled all the networks collected from different broods and fitted a single model to the entire data set for each monitoring method at a time. In this way, we can control for variability caused by inter-network differences (i.e. individuals in highly connected networks are expected to be more connected than those in less connected networks). We added, to the model, brood ID as a random intercept to capture the inter-network variability. Synchrony frequency index, Yi,j/min(Ni, Nj) was approximated with binomial distribution. The model was specified as follows, Yi,j,k ~binomial(pi,j,k, xi,j,k), with logit(pi,j,k)=β+φk + δi +δj, where Yi,j,k is the number of synchronous visits, and xi,j,k =min(Ni,k, Nj,k). The term φk is random intercept and k indexes the brood ID, and δi and δj are the sociality random term for variation of connectedness at the individual level (Hoff et al. (2002), see also Whitehead (2008) for similar parameterisation in modelling individual level effect for social interaction rate). The models were fitted using WinBUGS 1.4 (Spiegelhalter et al. 2003) with three chains of 1,020,000 iterations, 20,000 burn-in and 20,000 thinning interval, resulting 150 samples. Convergence was checked using Gelman-Rubin statistics (Gelman et al. 2004). Deviance information criterion (DIC) was used to judge the importance of the three random effects. A model is considered to have higher predictive performance when its DIC value is lower than other models with different sets of terms (Spiegelhalter et al. 2002). If the observed networks have significant heterogeneity at the three levels, DIC value is expected to be lower than that of simpler models without each of the random effects. We used 859 synchrony frequencies by 197 individuals observed with the PIT-tag system in 48 breeding attempts by 31 breeding


units. A WinBUGS code used for the model is presented in the Electronic Supplement. Comparisons of network variables across methods To evaluate the utility of PIT-tag data for descriptive statistics of social network structure, we compared the synchrony frequency index, and three network statistics based on the index, from the PIT-tag data against those from nest video data: node strength (Croft et al. 2008), weighted clustering coefficient (Holme et al. 2007) and sociality (Hoff et al. 2002). Node strength is the sum of all edge weights for a given node (individual) and measures connectedness of individuals. The clustering coefficient measures local clustering of network around a focal node and is known to be associated with spread of information through networks (Newman 2003). Flow of information across carers may be important in coordinated provisioning (Johnstone and Hinde 2006; Johnstone et al. 2014). We calculated the clustering coefficient for weighted networks proposed by Holme et al. (2007). In addition, we evaluated the sociality introduced above. The sociality is a model-based measure of connectedness of individuals similar to node strength. Since sociality is estimated as a parameter in a regression model, it can be estimated conditionally on other parameters in the model, which is an advantage when analyzing multiple networks at a time (see above for model details). However, the model did not include any other parameters except for intercept for the comparisons between PIT-tag and nest video data. Finally, repeatability (intraclass correlation) was calculated to examine the consistency of synchrony frequency and network indices across the three monitoring methods (video with breeding female, without breeding female and PIT tag), using GLMM with an R package rptR (Nakagawa and Schielzeth 2010). Grouping factor in the model was either individual-brood ID or dyad-brood ID. The repeatability of synchrony frequency index was calculated using a model with overdispersed (multiplicative) binomial distribution and logit link function, while that of other network indices was obtained based on a model with Gaussian distribution and identity link function.

Results Validating nest visit synchronicity Nest observations conducted by a human observer indicated that while all individuals within a breeding unit tended to visit the nest area (within 100 m of the nest) as a cohesive unit (86 % of occasions, N=163 group visits), individuals less commonly visited the nest tree together (33 % of occasions). While observing the provisioning ecology of babblers is

1383

difficult, the subset of nest visits successfully recorded suggests that ~49 % (N=468) involved individuals arriving at the nest area with food and flying to the nest tree or surrounding trees. By contrast, 37 % of nest visits involved individuals finding food in the vicinity of the nesting area, the rest involving individuals without food. Individuals arriving in the nesting vicinity with food entered the nest on average 23 s after the preceding bird (95 % confidence interval (CI) 5– 152 s), while those arriving without food and subsequently obtaining it (in the vicinity of the nest) did so after a delay of 136 s (95 % CI 29–305 s). Birds never entered the nest before the previous provisioner had exited. However, on 8 % of occasions, provisioners fed twice within the same groupvisiting bout, indicating that the first birds to arrive at the nest sometimes obtained more food in the vicinity of the nest before the group departed. Overall, those birds that flew directly to the nest tree or surrounding trees with food provisioned within a minute of each other (87 %), whilst those that did not tend to feed at 2-min intervals. The data obtained from external video cameras were in broad accordance with those obtained by direct observation. In particular, a proportion of group members arrived in the nest tree (or surrounding trees) together, each with food, and fed the nestlings in quick succession. Thereafter, and following a short delay, other group members arrived at the nest tree, with such visits tending to be less synchronous than the former ones. For example, the median time interval between successive visits by different individuals (entry-entry interval) was 31 s (inter-quartile range (IQR)=20–46 s) when group members arrived at the nest tree together but was 114 s (IQR=63– 225 s), when individuals arrived separately. In addition, the inter-visit interval between different individuals in synchronous visits (i.e. when birds arrived in the nest tree together) was less than 1 min on 83 % of occasions (Fig. 1a), while for asynchronous visits (i.e. when only a single bird was in the nest tree at a time), inter-visit intervals were under a minute in only 24 % of occasions (Fig. 1b). Thus, the results from both methods led us to define a series of visits in the nest video and PIT-tag data below as synchronous when temporal separation between nest entry times of successive visits was less than 1 min apart (chain rule), as this most effectively optimised the trade-off between acceptance versus rejection of truly synchronous visits (Fig. 1c). Finally, evidence from the internal nest videos showed that, with the exception of the breeding female, who frequently visits the nest without food, the vast majority (~90 %) of nest visits occur with a single prey item that is fed to offspring in over 97 % of occasions (see also Young et al. 2013). Breeding females remain in the nest for a median duration of 107 s, but such durations are highly variable (IQR=20.0 s–14.8 min). This variation precludes the duration and frequency of nest visits to be determined for breeding females using PIT tags in isolation, since our PIT-tag readers do not discriminate entry

1384


(a) 0.4

0.2

0 0

5

10

15

20

25

30

35

40

45

min

(b) Proportion

0.4

0.2

0 0

5

10

15

20

25

30

35

40

45

min 1

(c)

0.8 0.6 0.4 0.2 0 0

1

2

3

4

5

min Fig. 1 The difference of arrival times between different individuals from the video camera scanning nesting trees was concentrated to