Our long-term dolphin research has been run on a seasonal basis (typically austral winter–spring) since 1982 off Monkey Mia (in the eastern gulf of Shark Bay) and 2007 off Useless Loop (in the western gulf of Shark Bay). Data for this study were collected during 25 focal follows (Altmann 1974) of first-order allied male dolphins and their female consort from a small (5.4 m) research vessel in May–June 2016 in Shark Bay’s western gulf, and August–September 2016 and June–September 2017 in the eastern gulf.
We analysed focal follows of first-order male alliances herding a female. Herding is defined as an aggressively maintained association, where two to three males use vocal and physical threats to coerce a female to accompany them. Males engage in normal daily activities, such as foraging, travelling and resting, while herding a female, as well as in social and sexual behaviours directed at the female (e.g. synchronous displays, Connor and Krützen 2015). Individual dolphins in this population are well marked and thus were identified by trained observers on the research vessel via their unique dorsal fin shape and scars. Individual identification was corroborated with photo-identification data collected using a Canon 50D camera and 100–400 mm IS lens. During each focal follow, we verified the following variables at every 5-min interval: group membership and size, as defined by the 10 m ‘chain rule’ (Smolker et al. 1992); predominant group behavioural state (rest, travel, forage and socialise—see ESM for definitions); and predominant group spread, which was visually estimated and classified as tight (inter-animal distance < 2 m or < 1 body length distance (BLD)), moderate (inter-animal distance 2–5 m or 1–2 BLD), spread (inter-animal distance 5–10 m or 2–5 BLD) or widespread (inter-animal distance > 10 m or > 5 BLD). All changes in group membership (e.g. arrivals and departures of individuals) were recorded when they occurred during focal follows. In 2017, additional data were collected on which male was closest to the female during each 5 min observation period. Distance measures used were the same as used for group spread, e.g. if a trio of males are widespread (> 10 m apart) but one of them is tight (< 2 m) with the female, then the predominant group spread is widespread but the closest male is tight. The closest male to the female is considered to be the guard. We systematically recorded predominant closest male to the female every 5-min and also recorded all cases of guard switches when they occurred during focal follows.
Our hydrophone array consisted of four HTI-96 MIN series (flat frequency response: 0.002–30 kHz ± 1 dB) towed at 1 m depth around our research vessel in a rectangular formation (ca. 2 × 3.5 m), as per that outlined in King et al. (2018). Recordings were made onto a TASCAM DR-680 MKII multi-track recorder at a sampling rate of 96 kHz. A spoken track was used to note the bearing (compass bearing, where the boat’s bow was 0°), distance (m) and identification of the focal individuals at each surfacing. Focal follows were synchronised with the acoustic recording at the start of each follow. All recordings used in the analysis were made when the engine was switched off.
Acoustic recordings were analysed by inspecting the spectrograms (FFT length 1024, Hamming window) in Adobe Audition CC 2017. All occurrences of whistles and pops were identified and visually graded based on their signal-to-noise ratio (1: signal is faint but visible on the spectrogram, 2: signal is clear and unambiguous, 3: signal is prominent and dominates; Kriesell et al. 2014). Frequency-modulated whistles were identified as either continuous in their frequency contour pattern or multi-looped whistles. Multi-loop whistles were defined as a repeated modulation pattern that could be separated by periods of stereotyped silence up to 250 ms in length (Esch et al. 2009).
All vocalisations graded 2 and 3 were included in the analysis. In order to demonstrate that the vocalisations used in this analysis reliably came from our focal group (and, where possible, to identify which group member was vocalising), we localised a subset of whistles and pops (Table S1).
Only calls with good signal-to-noise ratios were used for the localisation analysis. Localisation accuracy of the array was calculated using custom-written MATLAB routines to calculate 2D averaged MINNA (minimum number of receiver array) localisations using the methods described in Wahlberg et al. (2001) and Schulz et al. (2006). The array was calibrated using two different frequency-modulated dolphin whistles, each approximately 1.5 s in duration with a frequency range of 4–20 kHz, as well as two different pop trains previously recorded from this population. Acoustic localisation accuracy for whistle directions (n = 75) were calculated as 76% within ± 15° of the true location, and 99% within ± 30°. Localisation accuracy for pop directions (n = 50) were calculated as 68% within ± 5°, 94% within ± 10°, and 100% within ± 15° of the true location. However, variation in estimated direction within a train was low, with < 2° difference between sequential pops in the same train.
All statistical procedures were conducted in R (R Core Team 2018). We summed the number of each vocalisation type (whistles and pop trains) in each 5-min observation period and modelled them against behavioural state and group spread (as factors). Any 5-min periods in which group membership was unknown were removed from the analysis (n = 5). Our data were both highly zero-inflated, due to the fact that dolphins can be silent for extended periods of time (Table S1), and temporally correlated, due to the nature of focal follows. Thus, to account for temporal correlation in call production (e.g. bouts), we used Generalised Estimating Equations (GEEs) using the geepack package in R (R Core Team 2018). We built models with an autoregressive correlation structure, where the focal follow number was the blocking unit, so that calls were correlated within each focal follow but were independent between follows. Zero-inflation can be addressed with zero-altered or hurdle models that are partitioned into two parts (a binary process that models zero and positive counts; and a zero-truncated Poisson process that models only positive counts; Zuur et al. 2009), but these models do not account for temporal correlation. The hurdle model, however, can be carried out manually using binomial and Poisson generalised linear models, which provide the same results in terms of estimated parameters and standard errors (Zuur et al. 2009). We therefore built two types of GEE for each vocalisation type: (1) a call occurrence model to identify how behavioural state and group spread influence the occurrence of a call type, considering binomial distribution to evaluate the presence or absence of calls per observation period; and (2) a call frequency model to identify how behavioural state and group spread affect the frequency of calls when they occur, considering Poisson distribution to model the positive counts of each vocalisation type per observation period. For the frequency model, the logarithm of the number of animals per observation period was included as an offset to account for differences in group sizes when additional individuals joined or left the focal group. We selected the most parsimonious model with the Quasi-likelihood Information Criterion (QIC; Pan 2001) using the MuMIn package in R (Bartoń 2009) and sequential Wald tests (anova function in R). Where ΔQIC < 4 between the best models, we used model averaging on the top set of models (Grueber et al. 2011). All models are presented in Table S2.
To explore the relationship between communication and coordinated behaviour in more detail, we built two Generalised Linear Mixed Models with binomial family. In the first model, the response variable was ‘arrival’ (0 = no, and 1 = yes), defined as a new individual(s) joining the focal group during that 5-min period (n = 620 across 16 first-order alliances). In the second, we used a subset of data from 2017 where information was available on the closest consorting male to the female (n = 194 across 11 first-order alliances), where the response variable was ‘change in closest male to female’ (0 = no change, and 1 = change). For both models, to control for repeated measures of individuals, we set first-order alliance as a random effect. Predictor variables for both models were pop train rate (number of pop trains/group size) and whistle rate (number of whistles/group size). We selected models using Akaike Information Criterion (AIC) and sequential Wald tests (anova function in R) and, where ΔAIC < 4 between the best models, then model averaging was carried out on the top set of models (Grueber et al. 2011). All models are presented in Table S3 and Table S4.