Accounting for syntax in analyses of countersinging reveals hidden vocal dynamics in a songbird with a large repertoire

Animal Behaviour 131 (2017) 23e32

Contents lists available at ScienceDirect

Animal Behaviour journal homepage: www.elsevier.com/locate/anbehav

Accounting for syntax in analyses of countersinging reveals hidden vocal dynamics in a songbird with a large repertoire Richard W. Hedley a, *, Kaleda K. Denton a, Robert E. Weiss b a b

Department of Ecology and Evolutionary Biology, University of California Los Angeles, Los Angeles, CA, U.S.A. Department of Biostatistics, UCLA Fielding School of Public Health, Los Angeles, CA, U.S.A.

a r t i c l e i n f o Article history: Received 13 November 2016 Initial acceptance 27 December 2016 Final acceptance 8 May 2017 MS. number: A16-00990R2 Keywords: birdsong Cassin's vireo countersinging syntax vocal interaction

Identifying the signalling strategies employed by animals during vocal interactions is a challenge, especially for species with large vocal repertoires. We propose that efforts to study such vocal dynamics can benefit by integrating models of syntax into their analyses. In this study, we conducted playback experiments on Cassin's vireo, Vireo cassinii, to examine the role of syntax, and more specifically, shared syntactic patterns, in countersinging. We presented 11 males with song sequences ordered according to population norms, and with sequences whose order deviated from population norms. We did not find evidence that individuals markedly altered their responses based on the syntax of the playback, either in their physical approach to the speaker or in the quantity of song they delivered in response. We did, however, find evidence that syntax was important in governing their choice of phrase types in response to the playbacks. Subjects did not match the playback phrase types. Instead, they engaged in a vocal behaviour referred to as song advancing, where they responded to a stimulus phrase type by singing the phrase type that most often followed the stimulus in their own normal song sequences. When playback sequences were ordered according to population norms, song advancing resulted in birds pre-empting the upcoming playback phrase type or delivering another of the prior playback phrase types (i.e. delayed matching) at higher rates than when playback sequences deviated from population norms. The detection of song advancing was only possible with the explicit inclusion of syntax in our analysis, suggesting that studies of the vocal interactions of species with repertoires of multiple vocalizations can benefit from consideration not only of a subject's repertoire, but also their syntax. © 2017 The Association for the Study of Animal Behaviour. Published by Elsevier Ltd. All rights reserved.

Countersinging interactions between neighbouring songbirds facilitate the establishment and maintenance of territory boundaries (Yasukawa, 1981). In species that possess repertoires of discrete song types, such interactions can involve complex dynamics and apparently sophisticated exchange of information (Burt, Bard, Campbell, & Beecher, 2002; Burt, Campbell, & Beecher, 2001; Molles, 2006; Searcy & Beecher, 2009). Many species have been shown to employ pattern-specific responses, meaning they adjust their song pattern in response to the song of their rival (Todt & Naguib, 2000). This can include subtly altering the acoustic structure of an individual song type (Vehrencamp, Yantachka, Hall, & de Kort, 2013), or altering their choice of song type altogether (Falls, 1985; Krebs, Ashcroft, & Orsdol, 1981; Stoddard, Beecher, Campbell, & Horning, 1992). Identifying and studying pattern-

* Correspondence: R. W. Hedley, 621 Charles E. Young Drive South, Room 3113, Los Angeles, CA 90095, U.S.A. E-mail address: [email protected] (R. W. Hedley).

specific responses is fundamental to our understanding of the evolution of vocal complexity in birds. This is particularly true of species with large song repertoires, in which the dynamics of vocal interactions can be complex, and where the functional significance of large song repertoires remains a topic of debate (Byers & Kroodsma, 2009). A central challenge in the study of pattern-specific responses is determining what factors lead a bird to deliver a particular song type at a particular time. This aim makes clear that if we seek to understand a singer's vocal behaviour, we should attempt to identify any and all influences on song choices, whether internal or external. It is well established that many, and probably all, songbirds abide by an internal syntax (Berwick, Okanoya, Beckers, & Bolhuis, 2011), meaning they arrange repertoire elements into nonrandom sequences characterized by frequent transitions between some pairs of elements and rare transitions between others. In light of this, the answer to the question may sometimes be as simple as to say that the bird delivered song type B at time t because he delivered song type A at time t
1, and typically

http://dx.doi.org/10.1016/j.anbehav.2017.06.021 0003-3472/© 2017 The Association for the Study of Animal Behaviour. Published by Elsevier Ltd. All rights reserved.

24

R. W. Hedley et al. / Animal Behaviour 131 (2017) 23e32

transitions from A to B. Although this study focuses on a species in which syntax governs the order of song types (or phrase types) within a bout, the same reasoning can be applied to syllable types in species where syntax governs the ordering of syllables within a song. When a bird interacts with a rival, it may abide by its internal syntax, or its song choices may be influenced by external factors such as the rival's song. In black-throated grey warblers, Setophaga nigrescens, for instance, the repertoire is partitioned into type I and II songs, and playback of any song lead males to respond with type I songs (Morrison & Hardy, 1983). In this case, the warbler's song sequence was disrupted by an external stimulus, leading the bird to deliver a type I song regardless of its prior singing behaviour. A similar behaviour is song type matching, where a bird repeats a perceived song type shortly after it was delivered by a rival (Falls, 1985; Krebs et al., 1981; Stoddard et al., 1992). Song type matching can be thought of as an interruption of a bird's normal song sequence. Given that syntax is likely to play such a large role in determining song type choices, and that pattern-specific responses constitute disruptions to a bird's normal song sequences, it is surprising that few studies have assessed the role of syntax in countersinging interactions. Those that have done so have occasionally documented additional behaviours. For example, Verner (1975) noted that western marsh wrens, Cistothorus palustris, deliver their songs by cycling through their large song repertoires in nearly stereotyped orders, and that the sequences employed by neighbouring males are similar. When Verner (1975) broadcast a sequence to a bird, ‘the subject anticipated each song type to be played next by the recorder’ (emphasis his, page 283). The birds, he proposed, cued in on the previous playback song, which led them to jump ahead in a shared sequence. Similar observations have been made in wood thrush, Hylocichla mustelina (Whitney, 1985), and nightingales, Luscinia megarhynchos (Todt, 1971). We henceforth refer to this behaviour as ‘song advancing’, defined as responding to a perceived song with the song type that typically follows it in one's own preferred song sequences. Song advancing would seem to be challenging, if not impossible, to detect and study without prior knowledge of an individual bird's syntax. The integration of syntactic models into analyses of vocal interactions is therefore warranted, and may even be necessary if we wish to identify an exhaustive list of the vocal behaviours of any species, not only because syntax itself may guide a bird's song choices, but also because some pattern-specific responses may be syntactic in nature, as in the case of song advancing. We sought to take such an approach using playback experiments to examine the countersinging dynamics of Cassin's vireo, Vireo cassinii. Males of this species possess repertoires of ~50 phrase types (equivalent to song types in other species) and sing according to a structured syntax, wherein the identity of upcoming phrase types can be predicted with >55% accuracy if the previous phrase type is known (Hedley, 2016a). Song sequences in this species show two additional phenomena. First, phrase types are arranged into clusters that consistently appear together in sequences (Fig. 1a). These clusters have been shown to be stable over time with respect to the phrase types contained therein, and the song sequences can be well described using Markov models (Hedley, 2016a). Second, cluster composition is often shared between individuals (Fig. 1b). That is, not only do neighbours overlap in their song repertoires (Hedley, 2016b), their syntax appears to be shared to an extent as well. Our experiments examined the role of shared song syntax in countersinging interactions. Observations of natural countersinging interactions, such as the interaction depicted in Fig. 1c, suggest that birds interact nonrandomly using phrase types that are

shared between the two participants. To test the role of syntax in these interactions, we presented each bird with a playback trial containing phrase types that normally occurred adjacently in sequences (‘typical’ trials) and one containing phrase types that rarely occurred adjacently in sequences (‘atypical’ trials), and examined the dynamics of each bird's response to the playbacks. Our statistical approach to analyse vocal responses was motivated by the work of Kroodsma (1975), who proposed that the probability of matching in response to a phrase type depends on four factors: (1) the frequency of occurrence of that phrase type overall; (2) the transition probability from the bird's most recent phrase type to the playback phrase type; (3) the amount of time since the bird most recently sang the phrase type in question; and (4) the vocal behaviour of other males within earshot. We employed a model that incorporated properties of syntax from the songs of the subjects, and thereby effectively controlled for (1), (2) and (3). The influence of other males (4) is precisely what we hope to understand with playback experiments, and while it is possible that songs from distant background males may affect the subject, such effects are likely to be minimal relative to the effect of a playback speaker positioned within the territory to simulate a territorial intrusion. Although Kroodsma (1975) referred only to song type matching, our model was flexible enough to be applied to song advancing as well. We used this approach to show that Cassin's vireos engage in song advancing at levels exceeding the rate expected by chance. METHODS Study Site and Species Research was conducted in a 1 km2 valley of mixed coniferedeciduous forest near Volcano, CA, U.S.A. (UTM: 10 S 706584 4262742, datum WGS 84). Experiments were approved by the Animal Research Committee at the University of California Los Angeles (ARC number 2013-041-03A) and conducted under U.S. Fish and Wildlife Service bird banding permit number 23809 and California Scientific Collecting Permit number 12750. Prior to the initiation of this study, 11 male Cassin's Vireos were colour banded for individual identification. These 11 birds were all of the males present at our study site in 2015. Males in our study population possess repertoires composed of an average of 51 phrase types (range 31e60), which they deliver in structured sequences with immediate variety, meaning phrase types are rarely repeated consecutively (Hedley, 2016b). Phrase types are short (<0.7 s long), highly stereotyped, and can be readily identified by a trained observer with >99% accuracy (Hedley, 2016a). Phrase types are widely shared in this population, such that the repertoires of any two males tend to overlap by about 50% (Hedley, 2016b). Details regarding the singing style of this species are examined in more detail elsewhere (Hedley, 2016a, 2016b). Nonexperimental Recording Sets We constructed a set of recordings of each of the 11 individuals made under nonexperimental conditions. In general, these recordings were made when birds were singing individually on their territory and not interacting closely with another bird. The nonexperimental recording sets contained a total of 62 395 phrases (mean ¼ 5672 phrases per bird, range 1498e14 101) and 300 recordings (mean ¼ 27 recordings per bird, range 9e60). The nonexperimental recording sets of five individuals spanned 3 years, those of two individuals spanned 2 years and those of four individuals included only recordings from 2015.

R. W. Hedley et al. / Animal Behaviour 131 (2017) 23e32

25

Phrase type ID

40 30 (a)

30 (b)

25

25

20

20

Shared cluster 1

15

30

Shared cluster 2

5

Unique to bird 2

20

15

10

0

(c)

Shared phrase types

10

10

5 0

20

40

60

80

100 120

0

0

20

40 60 80 100 120 Position in sequence

0

Unique to bird 1

0

50

100

150

Figure 1. Syntactic patterns are shared between individuals. (a, b) Two separate sequences of solo singing by two individuals (read from left to right). Phrase type ID (ordinate) was assigned to each phrase type on the order of appearance in (a). Phrase types unique to (b) were assigned numbers >23 as a continuation of the numbering scheme in (a). Phrase types are often delivered in clusters, appearing as groups of points on the plot. Two groups of phrase types that are clustered similarly in the repertoires of both birds are highlighted with red and blue backgrounds, illustrating that syntactic patterns, in addition to phrase types, are often shared between individuals in this species. We presented individuals with sets of phrase types that were often observed clustered together in song sequences in our population (typical trials), and sets of phrase types that were never clustered together (atypical trials). In this example, a typical trial could be composed of phrase types in either shared cluster 1 or shared cluster 2, while an atypical trial might combine phrase types from shared cluster 1, shared cluster 2 and other phrase types to form an abnormal sequence. (c) A countersinging interaction between two birds across a territorial boundary (bird 1 ¼ open black circles; bird 2 ¼ filled red circles). Shared phrase types are shown in the middle of the plot, with unshared phrase types above and below this region. When one bird delivered a shared phrase type, the other bird often sang that same phrase type shortly thereafter, suggesting that the song choices of the two birds are not independent.

The nonexperimental recording sets served three purposes: to determine the phrase types in each bird's repertoire; to parametrize a Markov model describing each bird's syntax; and to identify transitions between phrase types common in the sequences of multiple individuals, for the purposes of designing playback sequences. Importantly, previous work has not found evidence of marked reorganization of syntax within or between years in a given bird's song sequences (Hedley, 2016a), so a nonexperimental set of recordings is likely to give a good representation of a bird's syntax. Playback Design Each colour-banded bird was subjected to two playback trials containing five phrase types per trial. To assemble the playback sequences, we first selected 55 phrase types from the population. We assembled 11 typical trial sequences and 11 atypical trial sequences such that each phrase type appeared in one typical and one atypical sequence. To determine which phrase types to combine into typical and atypical trials, we overlaid Markov transition matrices for all birds calculated from the nonexperimental recording set to identify transitions that were both common in the nonexperimental recording set as a whole, and delivered by multiple individuals. Typical trials were assembled by identifying sets of five phrase types that frequently occurred adjacently in sequences, and atypical trials were assembled by identifying sets of five phrase types that rarely or never occurred adjacently. As confirmation that the typical and atypical sequences had the desired properties, we again examined the nonexperimental recording set sequences. Each set of five phrase types could give rise to 20 possible unique transitions (excluding repetitions), for a total of 440 unique transitions across the 22 trial sequences. In the nonexperimental recording set of 62 395 phrases, transitions between phrase types assembled into typical trials occurred 19 697 times, whereas transitions between phrase types in atypical trials occurred just 220 times. Similarly, each possible unique transition in the typical sequences occurred in the sequences of 2.37 ± 1.61 (mean ± SD) individuals in the nonexperimental recording set, while each possible transition in the atypical trials occurred in the sequences of only 0.26 ± 0.37 individuals. When assigning a typical and atypical sequence to each individual, we applied two additional constraints. First, birds were

never subjected to the same phrase type in the typical and atypical trials, to avoid habituation that might occur with multiple presentations of the same stimulus. Second, each bird was presented with some phrase types that were present in their repertoires and some that were not, to simulate a real intruder and examine whether presentation of an unshared phrase type may elicit its delivery from a bird. The number of shared and unshared phrase types presented (either three or four shared phrase types and one or two unshared) was balanced across the typical and atypical trials for each bird. The five phrase types chosen for a given trial were arranged into a sequence of 25 phrases, where each type occurred five times. Within this sequence, phrase order was randomized in such a way that consecutive phrases were never of the same type. Phrases were delivered every 2 s, as is typical for this species, so the sequence took 50 s to complete, and was followed by 10 s of silence. This 1 min segment repeated five times, comprising a 5 min playback period. Because of the differences in the phrase types contained in each sequence, and the randomization of the sequence, no two sequences in the experiments were alike. We sourced phrase type exemplars from high-quality recordings of 12 individuals in the same study area during 2013 and 2014, but the 10 phrase types presented to each bird need not have originated from the same individual. Instead, phrase types were combined into artificial sequences to simulate an unfamiliar intruder. Three individuals were unintentionally presented with one to three phrase type exemplars recorded from themselves at an earlier date. Since subjects may respond differently to their own songs and those of a rival (Falls, 1985; Stoddard et al., 1992), we ran the main analyses again without these individuals. The main conclusions did not change upon removal of these three individuals, but we included them in the analysis presented here to avoid unnecessarily reducing the sample size. A summary of the results from analysis without these three individuals is provided in Supplementary material 1. Playback Field Methods We conducted playbacks from 28 April to 31 May 2015 between 0700 and 1100 hours. An attempt was made to conduct both trials during the same phase of nesting, either during territory

26

R. W. Hedley et al. / Animal Behaviour 131 (2017) 23e32

establishment prior to nesting (N ¼ 2 individuals) or during incubation (N ¼ 9 individuals) when the male was off the nest. Ten subjects received the two trials separated by 1e3 days, but one individual's nest failed between trials, so the second trial was postponed until incubation began on a subsequent nest 23 days later. We placed a Jambox speaker (Jawbone, Inc., https://jawbone.com/) about 1.8 m up in a tree within the bird's known territory, with flags 10 m on each side to aid in distance estimation. The speaker was placed in the same location for both trials, and presentation order of the two trials was randomized. Phrase amplitude was normalized and broadcast at ~80 dB measured 1 m from the speaker. A single observer conducted all trials. The observer stood ~20 m from the speaker, recorded the subject's songs with a Sennheiser MKH20-P48 microphone and Telinga parabolic reflector onto a Marantz PMD661 recording device, and dictated into the microphone the bird's estimated distance from the speaker. The experimental period lasted 12 min: the first 2 min were passive observation, the subsequent 5 min coincided with the broadcast of the playback file, and the last 5 min comprised a second period of passive observation. Two observers examined the spectrogram of the resulting .wav file in the program Praat (Boersma & Weenink, 2014). The observers annotated to phrase type the phrases from the speaker and subject by assigning each phrase type a two-letter code (see Hedley, 2016b), and noted the subject's estimated distance from the speaker when each phrase was delivered. When the observers differed in their annotation of a phrase type (<1% of phrases), they discussed the discrepancy to arrive at a consensus. Statistical Analyses Initially, we examined whether seven different pattern-specific responses occurred above chance levels in the vocal responses of the subjects. These seven behaviours were as follows: (1) immediate matching (IM), defined as repeating the most recent playback phrase type; (2) delayed matching (DM), defined as repeating any of the phrase types that had previously emanated from the speaker during that trial; and (3e7) five forms of song advancing (SA1 to SA5). The first form of song advancing (SA1) was defined as responding to the most recent stimulus phrase type with the phrase type that most commonly follows it in the bird's normal syntax, as determined from that bird's nonexperimental recording set. The second form was defined as responding to a stimulus with its second most common successor in the subject's nonexperimental recording set, and so on until the fifth form of song advancing, SA5, defined as responding to a stimulus with its fifth most common successor. For each of the above behaviours, we employed a logistic regression model that took into consideration each bird's known syntax determined from the nonexperimental recording set. The model incorporates Markov transition probabilities to estimate the probability that the vocal behaviour in question would have occurred by chance each time the subject sang a phrase. Chance probabilities varied throughout the experiment as a function of the most recent phrase type delivered by the bird, so the model effectively controlled for the influence of syntax on these probabilities. We considered each of the subjects' phrases during the 5 min playback trials to be an independent Bernoulli trial, where the outcome can be either 1 or 0 with some probability of each outcome. These Bernoulli trials were modelled using a logistic regression model

yi jmi  Bernoulliðmi Þ

(1)

logitðmi Þ ¼ logitðpðyi ¼ 1jxi
1 ÞÞ þ a0 þ bj

(2)

In this model, yi designated whether phrase i from the bird did (yi ¼ 1) or did not (yi ¼ 0) constitute the behaviour in question (i.e. IM, DM or SA1-5, which were each analysed separately). Independent of this, mi was the estimated probability of engaging in the behaviour at phrase i, which itself depended on the values on the right side of equation (2). The first of these values, p(yi ¼ 1jxi-1), was the probability of the behaviour occurring by chance at phrase i given the subject's previous phrase type xi-1. This chance probability p(yi ¼ 1jxi-1) was calculated using a Markov model, smoothed with Backoff smoothing (Jurafsky & Martin, 2000) and Witten-Bell discounting (Witten & Bell, 1991), and parametrized from the subject's nonexperimental recording set (see Supplementary material 1). The value a0 was an intercept, and the values bj were random effects to account for the fact that multiple responses were derived from each subject. In the above model, the value of the intercept is of primary interest: an intercept (a0) with posterior density significantly greater than zero indicates that the behaviour in question occurred more often than expected by chance. The model can be thought of as assessing whether or not syntax alone could account for the number of occurrences of a behaviour, or whether the bird appeared to intentionally deviate from its syntax to increase the rate of the behaviour in question. In the IM model and SA1 to SA5 models, the chance probability p(yi ¼ 1jxi-1) was simply the probability of singing the single phrase type that would constitute a match or advance at phrase i given the bird's previous phrase type. In the DM model, however, the chance probability was the sum of the probabilities of singing each of the phrase types that would constitute a delayed match, as defined above. After assessing which of the seven vocal behaviours occurred above chance levels, we added additional predictors to the models to assess whether the treatment type (typical or atypical) or the subject's distance from the speaker affected the tendency to engage in each behaviour. The addition of a treatment parameter tested whether individuals altered their vocal behaviour in response to the syntax of the playback. The addition of a distance parameter tested whether birds altered their vocal behaviours based on their proximity to the speaker, as might be expected if these vocal behaviours are associated with the conveyance of aggression or if they reflect a bird's motivational state. Treatment (TMT) was coded as 0 for all phrases in atypical trials, and 1 in typical trials. Distance (D) was the bird's estimated distance from the speaker when each phrase was delivered. The treatment and distance variables were not collinear (R2 ¼ 0.039). Upon adding these predictors to the model, the updated model took on the following structure:

yi jmi  Bernoulliðmi Þ

(3)

logitðmi Þ ¼ logitðpðyi ¼ 1jxi
1 ÞÞ þ Zi a þ bj

(4)

Where Zi is a vector containing a 1 followed by the values of the two additional predictor variables described above, and a is a vector of the intercept (a0) followed by two coefficients (aTMT, aD). In this model, the value of the intercept is of little interest, since it represents the tendency to engage in song advancing when all other covariates are zero. Instead, we are interested in the other coefficients. A significantly nonzero value of one of these coefficients represents a significant effect of that predictor on the tendency of the subjects to engage in song advancing. The magnitude of a coefficient represents the influence of a one-unit change in the independent variable on the probability of engaging in song advancing on the logit scale, a nonlinear scale. On the logit scale, a

R. W. Hedley et al. / Animal Behaviour 131 (2017) 23e32

difference of one is equivalent to an increase in probability from 0.005 to 0.0125, from 0.05 to 0.135, or from 0.5 to 0.731. A more extensive description of logistic regression and the logit scale is given by Sokal and Rohlf (1995). To identify the best fit model to explain variation in each behaviour, we also conducted model selection using deviance information criteria (DIC). DIC is an approach commonly employed in conjunction with Markov chain Monte Carlo methods (Spiegelhalter, Best, Carlin, & van der Linde, 2002). It measures the fit of models with different numbers of parameters while penalizing more complex models. In our case, we compared models with and without the treatment and distance coefficients. Lower DIC values indicated models that better fit the data. The playback sequences contained shared and unshared phrase types, but subjects never responded to a playback with an unshared phrase type, indicating that our prior assessment of which phrase types were in each bird's repertoires was accurate. As a result, our analysis only considered the subject able to engage in immediate matching or song advancing if (1) the most recent playback phrase type was shared and (2) the subject had not delivered any intervening phrases since the most recent playback phrase. Delayed matching was less constrained, since the bird could engage in delayed matching at any time as long as at least one shared phrase type had been delivered by the playback. Immediate matching and song advancing models were therefore based on lower sample sizes (N ¼ 1125 phrases) than the delayed matching model (N ¼ 2438 phrases). Analyses were conducted using a Bayesian framework where coefficients were assigned uninformative priors (normal with mean ¼ 0). Fixed effects were assigned a wide prior SD of 10. Random effect SDs were assumed to have unknown variance, and the precision (i.e. 1/(SD2)) was assigned a prior based on a gamma distribution with shape and scale parameters of 1, per the recommendations made by Dey, Ghosh, and Mallick (2000). Analysis was carried out via 25 000 Markov chain Monte Carlo simulations in Jags (Plummer, 2003), version 3.3.0, implemented in R (R Core Team, 2016) with the package R2Jags (Su & Yajima, 2015). All data for this experiment is available on Figshare (http://dx.doi.org/ 10.6084/m9.figshare.3410110).

27

matching or song advancing relative to chance levels. We ran the model for each of the seven behaviours (IM, DM, SA1-5), and results are shown in Table 1 and Fig. 2. Only SA1 occurred significantly more often than was expected by chance (P ¼ 0.0026), indicating that birds often responded to perceived phrase types by singing the phrase type that would most commonly follow it in their own normal song sequences. SA1 remained significant at a Bonferronicorrected significance threshold of a ¼ 0.05/7 ¼ 0.007. Effects of Treatment and Distance on Vocal Behaviours Next we sought to examine whether the treatment type or a subject's distance from the speaker influenced the tendency to engage in song advancing (SA1). To do so, we added treatment type and distance as predictors in the model. Results for the first form of song advancing (SA1) are shown in Table 2. Treatment type had a significant negative effect on the tendency to engage in song advancing, indicating a stronger tendency to engage in song advancing in atypical trials. The effect of distance was not significant. Model selection confirmed these results: the best model for SA1 was a model including treatment, but not distance, as a predictor. Model selection results for SA1 are shown in Table 3. Table 1 Analysis of the behaviours that occurred at levels greater than expected by chance in the vocal responses of Cassin's vireos Vocal behaviour

Coefficient (a0)

SD

P

Immediate matching (IM) Delayed matching (DM) Song advancing (SA1) Song advancing (SA2) Song advancing (SA3) Song advancing (SA4) Song advancing (SA5)

0.04 0.11 1.03 0.45 0.29
0.20
0.43

0.34 0.18 0.33 0.31 0.31 0.33 0.38

0.45 0.27 0.0026 0.072 0.16 0.27 0.12

A significantly positive coefficient estimate represents a behaviour that occurred more often than expected by chance. P values represent the proportion of the posterior density that was smaller than zero (for positive coefficient estimates) or larger than zero (for negative coefficient estimates).

RESULTS

Expected Observed

Strength of Response to Playback Treatments

Matching or Advancing Relative to Chance Levels When we ran our model without additional predictor variables, the intercept represented the tendency for the birds to engage in

***

0.1 Proportion of phrases

Individuals exhibited a strong physical response, approaching towards the speaker in every trial (Supplementary material 2, Fig. S2), and responded vocally with an average of 111 phrases (range 12e145) per 5 min playback interval. There were no clear differences in the physical response to the trials with typical phrase type sequencing patterns compared to those with atypical sequencing. The minimum distance from the subject to the speaker did not differ between the two treatments (Wilcoxon signed-ranks test: T ¼ 7, N ¼ 11, P ¼ 1), nor did the amount of time spent within 10 m of the speaker (T ¼ 35, N ¼ 11, P ¼ 0.16), the latency to approach within 10 m of the speaker (T ¼ 13, N ¼ 7, P ¼ 0.94), or the number of changes in position, either towards or away from the speaker (T ¼ 24.5, N ¼ 11, P ¼ 0.80). In addition, we found no differences in the total number of phrases (T ¼ 40.5, N ¼ 11, P ¼ 0.53) or in the number of different phrase types (T ¼ 33, N ¼ 11, P ¼ 0.61) delivered in response to the two treatments during the 5 min playback periods.

0.05

0

IM

DM

SA1

SA2

SA3

SA4

SA5

Vocal behaviour Figure 2. Summary of song type matching and song advancing in response to playback. Expected and observed rates of immediate matching (IM), delayed matching (DM) and song advancing (SA1-5) are shown. Expected values were calculated using the P formula for the Poisson binomial distribution: EðxÞ ¼ ni¼1 pi (see Supplementary material 1; Johnson, Kemp, & Kotz, 2005). ***P ¼ 0.003.

28

R. W. Hedley et al. / Animal Behaviour 131 (2017) 23e32

Table 2 Factors affecting the tendency to engage in song advancing Parameter

Coefficient

SD

P

Intercept (a0) Treatment (aTMT) Distance (aD)

1.46
0.62
0.0097

0.42 0.30 0.024

0.0012 0.021 0.34

Significantly nonzero coefficient estimates represent factors that affected the tendency for the subjects to engage in song advancing, and the sign (positive or negative) indicates whether this was a positive or negative effect. P values represent the proportion of the posterior density that was smaller than zero (for positive coefficient estimates) or larger than zero (for negative coefficient estimates).

The negative treatment coefficient for SA1 indicates that there was a stronger tendency to engage in song advancing in the atypical trial than in the typical trial. Closer inspection, however, reveals that this was driven not by higher absolute rates of song advancing in the atypical trials, but by lower expected rates of song advancing in the atypical trials; song advancing occurred at approximately equal rates between the two treatments (Fig. 3). It may not be the case that the treatment itself was salient to the birds. An alternative interpretation is that subjects simply tended to engage in song advancing about 10% of the time, regardless of the treatment and their prior singing behaviour. At the individual level, 10 of 11 subjects engaged in song advancing (SA1) above chance levels. We quantified the overall strength of the tendency to engage in song advancing for each individual by dividing the observed rate of song advancing by the rate

Table 3 Results of model selection to examine the best model for the first form of song advancing (SA1) Parameters

DIC

DDIC

Interceptþtreatment Intercept Interceptþtreatmentþdistance Interceptþdistance

615.74 617.91 618.26 620.07

0 2.17 2.52 4.34

Lower deviance information criteria (DIC) values indicate a better model fit, while accounting for the complexity of the model. DDIC reflects the fit of the model relative to the top model.

0.12

Expected Observed

Rate of song advancing

0.1

0.08

0.06

Pre-emption of Upcoming Playback Phrases

0.04

0.02

0

expected by chance. These values ranged from 0.75 to 3.38. Interestingly, the two birds that engaged in song advancing at the highest rates were the two that were exposed to playbacks during territory establishment. These two individuals engaged in song advancing at 3.38 and 2.41 times chance levels, while the remaining nine birds that were tested later in the nesting cycle engaged in song advancing between 0.75 and 2.13 times chance levels. Although suggestive of an influence of breeding stage on the tendency to engage in song advancing, the small sample sizes preclude meaningful statistical analysis of this trend, so we leave this to future studies. We also ran similar analyses for the six other vocal behaviours (IM, DM, SA2-5) by adding treatment and distance as predictors of those behaviours. Results are provided in Supplementary material 1. Neither treatment nor distance showed a significant effect on any of these behaviours, with the exception of the DM model, which showed a significant positive effect of treatment. Delayed matching occurred at a rate of 10.4% in atypical trials and 14.2% in typical trials. Model selection gave broadly similar results: the top model for the other behaviours was the model containing only the intercept, again with the exception of DM, where the top model included the intercept and the treatment parameter. Results of model selection for these six behaviours are provided in Supplementary material 1. Since DM did not occur above chance levels overall, we suspected that the positive treatment coefficient in the DM might be a by-product of song advancing, rather than an active behaviour on the part of the birds. In typical trials, the similarity between the subject's own syntax and the syntax of the playback would mean that advancing in response to a playback phrase would likely result in the delivery of another playback phrase type (i.e. a delayed match). This would not be expected to be the case in atypical trials, because the atypical trials were composed of phrase types that do not normally occur near each other in sequences within this population. To investigate whether song advancing could explain the differences in delayed matching rates between typical and atypical treatments, we calculated rates of delayed matching when birds engaged in song advancing, and when they did not, in typical and atypical trials. Results are illustrated in Fig. 4a. Rates of delayed matching were significantly different between these four contexts (Fisher's exact test: P < 0.001). Post hoc comparisons revealed that song advancing in typical trials led to greater rates of delayed matching than in the other three contexts (P < 0.001 for all comparisons). Song advancing in atypical trials had the opposite effect, by decreasing the rate of delayed matching relative to the other three contexts (P < 0.005 for all comparisons). When birds did not engage in song advancing, delayed matching occurred at rates that did not significantly differ between typical and atypical trials (P ¼ 0.78). All significant comparisons remained significant at a Bonferroni-corrected significance threshold of a ¼ 0.05/6 ¼ 0.008. These results support the idea that the difference in delayed matching rates between typical and atypical trials can be explained as a by-product of song advancing.

Atypical trials

Typical trials

Figure 3. Rates of song advancing in typical and atypical trials.

Given that song advancing involves responding with the phrase type most likely to follow the stimulus phrase in a preferred sequence, a reasonable hypothesis regarding this behaviour is that it might lead to the pre-emption of a rival's upcoming phrase type that has yet to be delivered. Pre-emption of a rival, however, would only be expected to be successful if the two participants deliver their songs in similar orders. To test this, we investigated the rate of these ‘pre-emptive matches’ of the upcoming playback phrase type

R. W. Hedley et al. / Animal Behaviour 131 (2017) 23e32

(a)

(b) a

b

c

c

Rate of delayed matching

37/55 0.6

0.4

0.2 62/488 71/528

0

0/54 Song Song Other Other advancing, advancing, responses, responses, typical atypical typical atypical trials trials trials trials

0.15 Rate of pre-emptive matching

0.8

29

a

b

b

b

7/52

0.1

0.05 16/482

0

14/521

0/54 Song Song Other Other advancing, advancing, responses, responses, typical atypical typical atypical trials trials trials trials

Figure 4. Rates of (a) delayed matching and (b) pre-emptive matching of the upcoming playback phrase when the subjects engaged in song advancing (SA1) versus when they did not (‘other responses’) in typical and atypical trials. Numbers above bars indicate sample sizes of matches (delayed or pre-emptive) divided by the number of instances of the behaviour (song advancing or other responses) in each trial type (typical versus atypical). Letters above bars indicate significant differences, assessed with Fisher's exact tests. Bars with the same letter were not significantly different from one another (all pairwise P > 0.35); bars with different letters were significantly different (all pairwise P  0.005).

when the subjects engaged in song advancing (SA1) versus when they did not in typical and atypical trials. Each phrase was classified as either an instance of song advancing or not. Not mutually exclusively, each phrase was also classified as either a pre-emptive match, if it was the same phrase type as the subsequent playback phrase type, or not. Pre-emptive matches were observed 37 times throughout the experiments, and were observed at least one time in the responses of eight of the 11 subjects. Fig. 4b depicts the rate of pre-emption when birds engaged in song advancing, and when they did not, in typical and atypical trials. Rates of pre-emptive matching were significantly different between these four contexts (Fisher's exact test: P ¼ 0.003). Post hoc comparisons revealed that song advancing in typical trials led to greater rates of pre-emptive matching than in the other three contexts (P  0.005 for all comparisons), but that the other three contexts did not differ significantly from one another (P > 0.35 for all comparisons). All significant comparisons remained significant at a Bonferronicorrected significance threshold of a ¼ 0.05/6 ¼ 0.008. Thus, song advancing sometimes led birds to sing the upcoming playback phrase type before it had emanated from the speaker, but only when the playback phrase types were ordered according to locally typical syntax. When playback phrase types were arranged according to locally atypical sequencing patterns, song advancing still occurred at similar rates, but did not affect the tendency for subjects to pre-empt the playback phrase type. The similar patterns regarding delayed and pre-emptive matching shown in Fig. 4a and b can be explained based on the structure of typical and atypical playback sequences. Since playbacks were composed of repeated clusters of five phrase types, an upcoming phrase type was likely to have already occurred earlier in the playback. Therefore, instances of pre-emptive matching, in our experiments, constituted a subset of delayed matching outcomes. In much the same vein as we proposed for delayed matching, these results suggest the possibility that pre-emptive matching may result as a by-product of the tendency for birds in this population to deliver their songs in similar orders and to engage in song advancing. Alternatively, pre-emptive matching may be an active process, perhaps facilitated through familiarity with another bird's preferred song sequences. In that case, it might be expected to

increase as the birds become more familiar with the ordering of phrase types in the playback sequences, which in our experiments contained five repeated 1 min segments. We used logistic regression to test whether the rate of pre-emptive matching changed as a function of the time (in seconds) of each response phrase from the start of the playback. The relationship was not significant (btime ± SE ¼
0.0032 ± 0.0022, P ¼ 0.141), indicating that the rate of pre-emptive matching did not change markedly as the experiments progressed. DISCUSSION Responses to Typical and Atypical Playback Sequences Cassin's vireos did not alter their response to playbacks on the basis of the syntactic characteristics of the playback sequence. In terms of their physical response, various metrics intended to measure the strength of their approach did not differ in response to the typical and atypical playback sequences. Similarly, the quantity of their song output or the number of phrase types delivered in response did not differ between treatments. Overall, song sequences arranged according to population norms and those with abnormal ordering both appear to be perceived as a strong territorial intrusion and do not appear to evoke markedly different responses in this species. In previous experiments that have examined responses to altered song sequences, results have been mixed. Skylarks, Alauda arvensis, for instance, exhibited heightened levels of aggression in response to altered song sequences (Briefer, Rybak, & Aubin, 2013). California thrashers, Toxostoma redivivum, and Eurasian wrens, Troglodytes troglodytes, showed the opposite effect, responding less strongly to altered sequences than to unaltered ones (Holland, Dabelsteen, & Paris, 2000; Taylor, Brumley, Hedley, & Cody, 2017). In other species, including European robins, Erithacus rubecula (Bremond, 1968) and indigo buntings, Passerina cyanea (Emlen, 1972), alterations to song order did not appear to affect the response of receivers. It appears doubtful, given these disparate results from different species, that syntax has the same function in all species. The precise information encoded by syntax is an active area of ongoing study. It

30

R. W. Hedley et al. / Animal Behaviour 131 (2017) 23e32

is worth noting that in Cassin's vireos, syntax governs the order of phrases (i.e. songs) within an extended bout of song, where successive phrases are typically separated by two or more seconds. From a theoretical standpoint, syntactic patterns spanning such long gaps would seem poorly suited to encode information (e.g. individual identity or species identity), since a listener would have to wait several seconds to hear even a few consecutive phrases. In contrast, in the species mentioned above where differential responses to altered and unaltered song sequences have been observed, syntax governs transitions between elements within a continuous burst of song, where elements are separated by gaps on the order of a fraction of a second in duration. In those species, differences between altered and unaltered song sequences would likely be apparent within a second or two, which may facilitate discrimination based on sequence properties. Whether the rate of delivery of song elements is related in a systematic way to the function of syntax is speculative, at present, and awaits additional exploration in a wider suite of species. Countersinging Dynamics in Cassin's Vireo Although subjects did not give markedly different responses to the two playback treatments, we nevertheless found evidence of a role for the subject's own syntax in guiding vocal responses, through the vocal behaviour called song advancing. Males engaged in song advancing by frequently responding to a playback phrase with the phrase type that would most commonly follow the stimulus in the subject's normal song sequences (Fig. 2). Engaging in song advancing resulted in increased rates of pre-emptive matches, where the subject delivered the next playback phrase type before it was broadcast by the speaker, and delayed matches, where the subject delivered a playback phrase type that had come from the speaker previously. Pre-emptive matching and delayed matching were facilitated by shared syntactic patterns present in the playback, and song advancing did not lead to pre-emptive matches or delayed matches when the playback contained sequences whose syntax was atypical for the population under study (Fig. 4). These results hint towards a central role for song syntax in determining song choices during countersinging in this species and help to shed light on the dynamics depicted in Fig. 1c, dynamics that we have observed frequently in vocal interactions in this species. While the patterns appear to show that one bird delivers a shared phrase type that is soon copied by the other bird, our results now lead us to suspect that the dynamics instead involve one bird engaging in song advancing and pre-empting the song of the other by jumping ahead in a sequence that is common to both participants. A conclusive appraisal of this view of countersinging dynamics in Cassin's vireo will require analyses focusing on natural interactions, to complement our experimental results. Notably, some authors have proposed that singing behaviours comprise a graded signalling system (Burt et al., 2001), where low or moderate levels of threat might be met with song type matching, and higher levels of threat with more aggressive behaviours such as soft song (Ballentine, Searcy, & Nowicki, 2008; Searcy, Anderson, & Nowicki, 2006; Searcy & Beecher, 2009). If this paradigm can be applied to Cassin's vireos, it is possible that our playback experiments, which simulated a strong territorial intrusion, may have elicited only a subset of the full suite of singing strategies employed by this species. Additional behaviours, such as immediate matching, may become apparent as analyses are expanded to a greater variety of contexts. Another question that might be clarified through investigations of natural interactions is whether pre-emptive matching is an active process, rather than a by-product of song advancing. We did

not find evidence for increased rates of pre-emptive matching towards the end of playbacks, as might be expected if familiarity with a sequence facilitated pre-emptive matching. While our results are consistent with the view that pre-emptive matching is a byproduct of song advancing, it is also possible that the 5 min playback period was too short for familiarity with a sequence to develop. This would not be problematic in natural vocal interactions, since neighbours repeatedly interact with each other over the course of days and weeks during the breeding season. If pre-emptive matching is an active behaviour, birds might alter their choice of phrase type based on prior familiarity with their rival's syntax, and their choices may vary as a function of the identity of the bird with which they are interacting. If it is instead a byproduct, birds would only be expected to select their response with reference to their own syntax, rather than their rival's syntax, through the process of song advancing. Research is currently ongoing to examine whether our experimental results and interpretations generalize to natural countersinging interactions. Descriptions of song advancing in at least three bird species in the laboratory (nightingale: Todt, 1971; marsh wren: Kroodsma, 1979; wood thrush: Whitney, 1985) and one species in the wild (Cassin's vireo, this study) suggest that this behaviour is more widespread than currently appreciated. Despite evidence of this behaviour dating back over 40 years, few studies have focused on understanding its role in communication or documenting it in additional species. In stark contrast, vocal matching behaviours have been documented in dozens of species as distantly related as songbirds, parrots and cetaceans over the same time frame (reviewed in King & Mcgregor, 2016). The critical difference between these two behaviours is that matching involves a stimulus and response that share obvious acoustic structure, while song advancing involves a stimulus and response that can differ arbitrarily in their acoustic structure. This difference leads to a vast divergence in the detectability of these two behaviours. For example, vocal matching is often apparent to the human ear (Kroodsma, 1971) and readily discernible on a spectrogram (Stoddard et al., 1992), while song advancing would seemingly be impossible to detect by ear, by spectrogram, or by any method that does not consider a subject's vocal responses in light of their syntax. We propose, on the basis of our findings, that animal communication research can benefit greatly by incorporating and studying syntax, and that a full understanding of the complexities and nuances of the vocal interactions of animals may require such an approach. Potential Functions of Song Advancing A significant question not addressed by our results is the question of the role of song advancing in communication. One possibility is that song advancing may convey aggression from sender to receiver. Searcy and Beecher (2009) outlined criteria for identifying aggressive signals in avian communication. The first of these is to demonstrate that the behaviour occurs more often than expected by chance, a criterion that is supported in the case of song advancing in Cassin's vireos. Another criterion is that the behaviour should be more prevalent during aggressive contexts. Previous authors have argued that aggressive encounters are expected to be more frequent and intense prior to the initiation of nesting. Higher rates of song type matching early in the season has been put forth as evidence that matching acts as a signal of aggression (Beecher, Campbell, Burt, Hill, & Nordby, 2000). We noted that the birds tested during the territory establishment phase in our experiment showed higher rates of song advancing than those tested during incubation. Although this is suggestive of a trend, and worthy of further investigation, our sample of just two individuals in the pre-

R. W. Hedley et al. / Animal Behaviour 131 (2017) 23e32

nesting phase is not sufficient to draw a conclusion. We also did not observe a clear relationship between a bird's distance from the speaker and the tendency to engage in song advancing, but again our modest sample of just 11 individuals makes interpretation of negative results challenging. Perhaps the most critical consideration if a signal is to be of biological importance is that it be salient to other birds, the third criterion of Searcy and Beecher (2009). Our results shed little light on this point. As mentioned above, the behaviour is difficult to detect, so salience to a receiver cannot be assumed. It seems plausible, however, that neighbouring birds may be familiar with each other's syntax, either because their neighbour's syntax is similar to their own, or because they have interacted with that neighbour previously. The tendency for song advancing to result in pre-emption of a rival's song may be another means by which it can be detected by a receiver. Assessing receiver responses to song advancing would require interactive playbacks and would be another productive avenue of inquiry into this behaviour. On the whole, our results neither fully satisfy nor contradict the criteria for aggressive signals outlined by Searcy and Beecher (2009), but opportunities for follow-up experiments are plentiful. Another possible function of song advancing is that it may encode dominance relationships. Following up on the field studies of Kroodsma (1979), Verner (1975) demonstrated that, among two captive marsh wrens, the dominant bird sought to maintain a leading position in the delivery of shared song sequences, while the subordinate bird trailed. Song advancing might achieve a similar end, by allowing a bird to jump ahead of his rival in a shared sequence, thereby establishing a leading role in a countersinging interaction. The trailing bird may ‘match’ the leader simply by continuing to sing its normal song sequence. Whether this behaviour is actually related to dominance, as proposed by Kroodsma (1979), remains an open question. An alternative is that such vocal interactions are performances whose evolution is mediated by the responses of eavesdropping conspecifics (Logue & Forstmeier, 2008). Results of two-speaker playback experiments on nightingales showed that leaders in matched countersinging interactions elicit more attention than followers from male and female eavesdroppers alike (Bartsch, Wenchel, Kaiser, & Kipper, 2014). If these results extend to Cassin's vireos, asymmetrical costs and benefits of song advancing may be imposed, not by the participants themselves, but by eavesdroppers. The extent to which leaderefollower dynamics apply to the specific case of song advancing is one of many questions that remains to be addressed in future studies. AUTHOR CONTRIBUTIONS R.W.H. designed and conducted experiments, processed and analysed the data and wrote the manuscript. K.K.D. assisted with postprocessing and annotation of data and helped draft the manuscript. R.E.W. assisted with statistical analysis. COMPETING INTERESTS We have no competing interests. Acknowledgments This work was funded by National Science Foundation Award Number 1125423 and the Ralph W. Schreiber Ornithology Research Award from the Los Angeles Audubon Society. We thank Charles

31

Taylor and David Logue for feedback on the manuscript and Martin Cody for input on experimental design. Supplementary material Supplementary material associated with this article is available, in the online version, at http://dx.doi.org/10.1016/j.anbehav.2017. 06.021. References Ballentine, B., Searcy, W. A., & Nowicki, S. (2008). Reliable aggressive signalling in swamp sparrows. Animal Behaviour, 75, 693e703. Bartsch, C., Wenchel, R., Kaiser, A., & Kipper, S. (2014). Singing onstage: Female and male common nightingales eavesdrop on song type matching. Behavioral Ecology and Sociobiology, 68, 1163e1171. Beecher, M. D., Campbell, S. E., Burt, J. M., Hill, C. E., & Nordby, J. C. (2000). Songtype matching between neighbouring song sparrows. Animal Behaviour, 59, 21e27. Berwick, R. C., Okanoya, K., Beckers, G. J. L., & Bolhuis, J. J. (2011). Songs to syntax: The linguistics of birdsong. Trends in Cognitive Sciences, 15, 113e121. Boersma, P., & Weenink, D. (2014). Praat: Doing phonetics by computer (version 6.0.29). http://www.praat.org/. mantique et les e le ments vecteurs Bremond, J.-C. (1968). Recherches sur la se d'information dans les signaux acoustiques du rouge-gorge (Erithacus rubecula). Terre Vie, 2, 109e220. Briefer, E. F., Rybak, F., & Aubin, T. (2013). Does true syntax or simple auditory object support the role of skylark song dialect? Animal Behaviour, 86, 1131e1137. Burt, J. M., Bard, S. C., Campbell, S. E., & Beecher, M. D. (2002). Alternative forms of song matching in song sparrows. Animal Behaviour, 63, 1143e1151. Burt, J. M., Campbell, S. E., & Beecher, M. D. (2001). Song type matching as threat: A test using interactive playback. Animal Behaviour, 62, 1163e1170. Byers, B. E., & Kroodsma, D. E. (2009). Female mate choice and songbird song repertoires. Animal Behaviour, 77, 13e22. Dey, D. K., Ghosh, S. K., & Mallick, B. K. (2000). Generalized linear models: A Bayesian perspective. New York, NY: M. Dekker. Emlen, S. T. (1972). An experimental analysis of the parameters of song eliciting species recognition. Behaviour, 41, 130e171. Falls, J. B. (1985). Song matching in western meadowlarks. Canadian Journal of Zoology, 63, 2520e2524. Hedley, R. W. (2016a). Complexity, predictability and time homogeneity of syntax in the songs of Cassin's vireo (Vireo cassinii). PLoS One, 11, e0150822. Hedley, R. W. (2016b). Composition and sequential organization of song repertoires in Cassin's vireo (Vireo cassinii). Journal of Ornithology, 157, 13e22. Holland, J., Dabelsteen, T., & Paris, A. (2000). Coding in the song of the wren: Importance of rhythmicity, syntax and element structure. Animal Behaviour, 60, 463e470. Johnson, N. L., Kemp, A. W., & Kotz, S. (2005). Univariate discrete distributions. Hoboken, NJ: J. Wiley. Jurafsky, D., & Martin, J. H. (2000). Speech and language processing: An introduction to natural language processing, computational linguistics, and speech recognition. Englewood Cliffs, NJ: Prentice-Hall. King, S. L., & Mcgregor, P. K. (2016). Vocal matching: The what, the why and the how. Biology Letters, 12, 20160666. Krebs, J., Ashcroft, R., & Orsdol, K. (1981). Song matching in the great tit Parus major L. Animal Behaviour, 29, 918e923. Kroodsma, D. E. (1971). Song variations and singing behavior in the rufous-sided towhee, Pipilo erythrophthalmus oregonus. Condor, 73, 303e308. Kroodsma, D. E. (1975). Song patterning in the rock wren. Condor, 77, 294e303. Kroodsma, D. E. (1979). Vocal dueling among male marsh wrens: Evidence for ritualized expressions of dominance/subordinance. Auk, 96, 506e515. Logue, D. M., & Forstmeier, W. (2008). Constrained performance in a communication network: Implications for the function of song-type matching and for the evolution of multiple ornaments. American Naturalist, 172, 34e41. Molles, L. (2006). Singing complexity of the banded wren (Thryothorus pleurostictus): Do switching rate and song-type diversity send different messages? Auk, 123, 991e1003. Morrison, M. L., & Hardy, J. W. (1983). Vocalizations of the black-throated gray warbler. Wilson Bulletin, 95, 640e643. Plummer, M. (2003). JAGS: A program for analysis of Bayesian graphical models using Gibbs sampling. In Proceedings of the 3rd International Workshop on Distributed Statistical Computing, 2003 (pp. 1e10). https://www.r-project.org/ conferences/DSC-2003/Proceedings/. R Core Team. (2016). R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing. Searcy, W. A., Anderson, R., & Nowicki, S. (2006). Bird song as a signal of aggressive intent. Behavioral Ecology and Sociobiology, 60, 234e241. Searcy, W. A., & Beecher, M. D. (2009). Song as an aggressive signal in songbirds. Animal Behaviour, 78, 1281e1292. Sokal, R. R., & Rohlf, F. J. (1995). Biometry: The principles and practice of statistics in biological research. New York, NY: W. H. Freeman.

32

R. W. Hedley et al. / Animal Behaviour 131 (2017) 23e32

Spiegelhalter, D. J., Best, N. G., Carlin, B. P., & van der Linde, A. (2002). Bayesian measures of model complexity and fit. Journal of the Royal Statistical Society B (Statistical Methodology), 64, 583e639. Stoddard, P. K., Beecher, M. D., Campbell, S. E., & Horning, C. L. (1992). Songtype matching in the song sparrow. Canadian Journal of Zoology, 70, 1440e1444. Su, Y.-S., & Yajima, M. (2015). R2jags: using R to run ‘JAGS’. https://cran.r-project.org/ package¼R2jags. Taylor, C. E., Brumley, J. T., Hedley, R. W., & Cody, M. L. (2017). Sensitivity of California thrashers (Toxostoma redivivum) to song syntax. Bioacoustics, 4622, 1e12. Todt, D. (1971). Aquivalente und konvalente gesangliche Reaktionen einer extrem regelmabig singenden Nachtigall (Luscinia megarhynchos L.). Zeitschrift für Vergleichende Physiologie, 71, 262e285.

Todt, D., & Naguib, M. (2000). Vocal interactions in birds: The use of song as a model in communication. Advances in the Study of Behavior, 29, 247e296. Vehrencamp, S. L., Yantachka, J., Hall, M. L., & de Kort, S. R. (2013). Trill performance components vary with age, season, and motivation in the banded wren. Behavioral Ecology and Sociobiology, 67, 409e419. Verner, J. (1975). Complex song repertoire of male long-billed marsh wrens in eastern Washington. Living Bird, 14, 263e300. Whitney, C. L. (1985). Serial order in wood thrush song. Animal Behaviour, 33, 1250e1265. Witten, I. H., & Bell, T. C. (1991). The zero-frequency problem: Estimating the probabilities of novel events in adaptive text compression. IEEE Transactions on Information Theory, 37, 1085e1094. Yasukawa, K. (1981). Song and territory defense in the red-winged blackbird. Auk, 98, 185e187.