Study animals
Three-spined sticklebacks (G. aculeatus, 27 ± 2.4 mm, mean ± SD, standard body length at the time of testing for n = 96 individuals), were collected from the river Cary, Somerset, UK (grid ref: ST 469 303) in September 2016 and were transported to the environmentally controlled fish laboratory at the University of Bristol, UK. The fish were housed in three glass tanks (70 cm (L) × 45 cm (W) × 37.5 cm (H)) of ~50 individuals for 10 months before testing and were fed daily with brine shrimp or defrosted frozen bloodworms (Chironomid sp. larvae). Photoperiod was on 11 : 13 h light : dark cycle and ambient temperature was maintained at 16 °C to prevent the fish from entering reproductive condition.
Experimental set-up
Trials were conducted in an oval-shaped experimental arena (Supplementary Fig. 1; 133.5 (L) × 72 (W) with 62 cm high walls). The arena wall was constructed from white opaque foamed polyvinyl chloride (PVC, 3 mm thickness). The food stimulus could appear through one of four holes (3.5 mm ø) in the arena wall, which were located 5 cm high from the base of the arena. Fish were filmed from above with a Panasonic HC-VX980 video camera in 4 K (3840 × 2178 pixels) and a temporal resolution of 25 frames per second. The camera was fixed centrally above the arena with the lens 161 cm above, and perpendicular to, the water surface. Water was maintained at the same temperature as the holding tanks and was 10 cm deep. An opaque plastic curtain was hung from above camera height to below the height of the arena wall to minimize external disturbances and diffuse the overhead lighting to avoid reflections on the water surface. A second video camera (Panasonic HC-X920) was positioned above the arena and connected to two external monitors behind the curtain. The two monitors were positioned at opposite ends of the arena, allowing the experimenter to view the activity of the fish before presenting the food stimulus. Half of each monitor was covered with card to obscure the view of the half of the arena furthest from the experimenter. This ensured that the location and behaviour of the fish was blinded when the stimulus was presented (see below).
Experimental protocol
One week prior to the experiment, 96 fish were randomly assigned to twelve groups of eight individuals. The assignment was carried out using a complete random block design: 12 fish were caught from 1 of the 3 holding tanks and each was randomly assigned to 1 of the 12 groups. This was repeated 8 times to form the 12 groups of 8. Each group included a minimum of two fish caught from each of the three tanks. This method of group assignment was designed to minimize variation between groups that could, e.g., be generated from bolder fish being caught earlier and all being assigned to the same group30. The fish were given 6 days to habituate in their groups in smaller glass holding tanks (70 (L) × 25 (W) × 37.5 (H) cm). Each holding tank was enriched with a horizontal piece of PVC tubing and an artificial plant.
Trials took place from Monday to Friday over 4 weeks between 31 July and 25 August 2017. Each week, the groups were randomly assigned to one of two sets of six groups. The two sets of groups were tested on alternate days. Hence, a group was never tested more than once every 2 days. Trials occurred between 09:45 and 16:00 each day and the order of testing among groups was randomized within each day.
At the start of a trial, all eight fish were netted into the centre of the arena and allowed 2 min to acclimatize. The experimenter then inserted the tip of a plastic pipette holding a single bloodworm into one of the four holes in the wall of the arena (randomly selected per presentation, Supplementary Fig. 1). The end of the pipette was wrapped in red PVC tape to provide a standardized (17 (L) × 3 (D) mm) visual stimulus that mimicked a bloodworm31. The visual stimulus was presented when all eight fish were on the opposite half of the arena to the selected hole, so that the experimenter was blind to the location and behaviour of the fish due to the obscured monitor display. The minimum possible distance between any individual and the stimulus when it was presented was 43 cm (Supplementary Fig. 1). When a fish swam within two body lengths of the stimulus as viewed on the monitor, the bloodworm was pipetted into the arena as a reward. The stimulus was removed once the bloodworm had been consumed or when all eight fish swam away from the stimulus towards the other side of the arena. If the reward was not immediately consumed, a new stimulus was not presented until the bloodworm had been eaten. During each trial, the stimulus was presented at the first opportunity after a minimum of a 3 min interval between successive presentations had elapsed and until six presentations had been performed (4.2 ± 0.9 min, mean ± SD, time gap between stimulus presentations within a trial). This time interval allowed the group to resume normal swimming behaviour between presentations, but varied depending on when the entire group was on the opposite side of the arena to the selected stimulus position. Limiting the number of presentations to six per trial meant that satiation effects were unlikely to influence the responses of the fish, as this species can consume over 50 bloodworms during a single feeding period32.
In 17 presentations, a fish appeared on the same side of the arena as the selected stimulus position during the presentation of the stimulus. In these cases, the stimulus presentation was repeated after 3 min. To maintain the minimum distance between the stimulus and the nearest fish, after reviewing the video footage, eight presentations were subsequently removed from our analyses, because the fish were quantitatively determined to be on the same side of the arena as the stimulus.
Following a trial, the group were returned to their holding tank. During the experimental period, all 12 groups were fed bloodworm in their holding tanks following the final trial of the day (Monday to Friday) and ad libitum during the weekend. This ensured that all fish had access to food and could maintain their health regardless of their performance in the trials. During the first week of trials, an individual in each of three groups was replaced with individuals naive to the experiment (one due to injury and two deaths of unknown cause) and given 24 h to habituate within their groups in the holding tanks prior to testing. All trials conducted prior to the replacements were excluded from the analyses. Thus, group membership was constant in the final dataset.
Video processing and data extraction
Video files were converted from MP4 to M4V format in Handbrake (version 1.0.7, https://handbrake.fr/) and file resolution was reduced to 1920 × 1080 pixels to increase the speed of automated tracking. We used the automated two-dimensional tracking software idTracker21 to obtain the cartesian coordinate positions (xi, yi) for the centre-of-mass of each tracked individual (i) at each time step (t). To keep track of individual identities within the same group across the trials of the experiment, we re-used the fingerprint (individual specific) references generated by idTracker from the first trial of each group when processing video files from the subsequent trials. To reduce tracking noise, the trajectories of each fish were smoothed using a Savitzky–Golay filter with a span of 0.5 s (~13 frames) and a polynomial of 3 degrees in R package Trajr. Trajectory information for three stimulus presentations were not obtained due to corrupted video files. This resulted in tracked video footage for 468 presentations from 77 trials across the twelve groups.
Statistical analyses
All statistical analyses were conducted in R (Version 3.4.3, R Development Core Team). The likelihood of being first to respond to the stimulus in a group, the response latency of the first responding individual, the arrival latency per individual, and the heading difference of individuals to their nearest neighbours were each analysed separately as response variables in mixed models (see Supplementary Methods for details on the identification of first responding individuals and the calculation of response variables). The statistical significance of fixed effects were tested using F-tests with Kenward–Roger approximated degrees of freedom in R package pbkrtest for LMMs and with LRTs in R package lme4 for GLMMs. Estimates of inter-individual repeatability were generated in R package rptR.
To quantify the position and movement of individual fish and the group when the stimuli were presented, we calculated parameters from the trajectory data based on the position and change in position of the fish in the 0.5 s prior to presentation. Most positional and movement parameters were quantified by their median value over these 13 frames to reduce the weight of outliers and best capture the central tendency of the skewed distributions of the parameters.
To analyse the effects of the six positional and movement parameters on the likelihood that an individual is first in their group to respond to the stimulus, we used GLMMs with a binomial error distribution and logit link function. Response to the stimulus was considered a binary variable where 1 indicated the individual that was first in the group to respond and all other individuals in the group were given a score of zero. Fish identity was included as a random intercept to account for individual-level differences. As there could only be one first responder per group per presentation and the likelihood of being the first individual to respond was relative to other individuals in the group, we generated relative rather than aboslute values of each parameter. To generate relative measures, we divided the median value for each fish by the mean median value of all eight individuals in its group for each of the six parameters, except for relative proportion of time spent on the convex hull edge of the group, which we divided by the group mean proportion. We used these relative measures of movement and position as explanatory variables in our models. In addition, we included relative body length in our analysis due to reported correlations between body size and several factors that could influence responsiveness in fish (e.g., metabolic rate and feeding motivation33 and visual acuity34). We assessed the effect of each parameter by calculating model averaged coefficient estimates and 95% CIs from a set of candidate models that initially included all possible combinations of main effect terms. To ease interpretation of model averaged parameter estimates, all seven parameters were standardized (mean = 0, SD = 1) prior to analysis. We used corrected Akaike information criterion (AICc) values to evaluate our candidate models, owing to a low sample size-to-parameter ratio and used only those models with a cumulative weight of 95% to perform the final model averaging35 (Supplementary Table 2). As seven candidate models were <2 ∆AICc from the top-supported model and because we were interested in which factors have the strongest effect on likelihood of first response, we adopted a full model averaging approach36, implemented in R package MuMIn. In addition, we calculated the RI of the parameters by summing the Akaike weights for models in which they appear using the complete candidate set of models35. An RI tending towards 1 indicates that the parameter appears in the best supported models and an RI tending towards 0 indicates that the parameter appears in the least supported models35. To test whether individual identity accounted for important variation in the likelihood of first response to the stimulus, we compared the goodness-of-fit (deviance) between a binomial GLMM without any explanatory variables and fitted with a random intercept of fish identity to a general linear model with individual identity removed, using a LRT37. We also compared the goodness-of-fit between these models but including the four explanatory variables retained in the top-supported model based on AICc comparisons (Supplementary Table 2, model 1).
We took a similar approach to assess whether the five group-level positional and movement parameters explained variance in the latency of first responders to respond to the stimulus. Negative binomial GLMMs were used with the first responder’s identity nested within-group identity as a random effect. We performed the same AICc model averaging procedure outlined above from an initial set of candidate models including all possible combinations of the five group-level parameters and cumulative days of testing as main effects.
For fish that arrived at the stimulus within 20 s of presentation, we tested for an interaction between arrival order at the stimulus and group polarization on the latency to arrive at the stimulus using a LMM to examine the consequences that collective order had on access to resources for different group members. Group polarization was quantified in the 0.5 s prior to the stimulus presentation. Bearing of the group heading to the stimulus and cumulative days of testing (which had the strongest effects on the response latency of first responders), an interaction between bearing of the group heading to the stimulus and polarization and distance of the group centroid to the stimulus (which we expected to correlate positively with arrival latency) were also included in the model as main effects, and fish identity nested within-group identity was included as a random effect. Arrival latencies were log10 transformed prior to analysis. Post-transformation, the model residuals remained moderately right-skewed (skewness = 0.66 where values from −0.50 to 0.50 approximate symmetry) and further inspection revealed that this was owing to observations of arrival latencies close to the maximum of 20 s and underpredicted by the model. These observations also tended to be multivariate influential outliers in the model with greater than six times the mean Cook’s Distance38. To improve the fit to model assumptions, we re-ran the model of latency to arrive at the stimulus excluding observations with greater than six times the mean Cook’s Distance (n = 48) and we report the results from this analysis (n = 2043 observations) in the main text. However, qualitatively similar results were obtained when all observations (n = 2091) were included in the analysis (Supplementary Table 8). To test whether individual identity accounted for significant variation in arrival latencies, we compared the goodness-of-fit of the LMM to the same model with individual identity removed, using a LRT. We examined whether individuals were consistent in their order of arrival at the stimulus by LRT comparison of a LMM with arrival order as the response variable (sqrt transformed) and individual identity nested within-group identity as a random effect to the same model with individual identity removed. After confirming that individuals were consistent, we tested for an effect of learning that was dependent on arrival order by including an interaction between arrival order and cumulative days of testing in the model of latency to arrive at the stimulus.
Individual identity intercept values for the model of latency to arrive at the stimulus represent whether individuals arrived consistently sooner or later to the stimulus and hence should favour group disorder or group order. To test whether inter-individual variation in arrival latencies predicted the alignment behaviour of individuals to their nearest neighbours during periods outside of the stimulus presentations, we ran a LMM with the mean heading difference to the nearest neighbour (averaged over every frame in the three minutes prior to the stimulus presentations) as the response variable and the individual identity intercept values from a model of latency to arrive at the stimulus controlling for bearing of the group heading to the stimulus, distance of the group centroid to the stimulus, cumulative days of testing, polarization and an interaction between polarization and bearing of the group heading to the stimulus as a main effects (arrival order accounted for some of the inter-individual variation in the random intercepts and was excluded from the model). Mean speed (averaged over every frame in the three minutes prior to the stimulus presentations) was included as an additional main effect to control for the expected positive relationship between alignment and speed2 and group identity was included as a random effect.
All statistical models were inspected for under- and over-dispersion, and to ensure that they complied with assumptions of orthogonality, homoskedasticity, and normality of residuals. Model averaging approaches are generally robust to the effects of collinearity, nevertheless, we checked for issues related to collinearity among predictors by calculating Spearman’s rank correlations (rs, Supplementary Tables 5 and 9) and variance inflation factors (VIFs) in all our models. Only polarization and centroid speed showed evidence of collinearity (Supplementary Table 5); however, there was no evidence for strong (VIF > 3) multicollinearity in our models. For GLMMs, we used the R package DHARMa to interpret the model residuals. Statistical analyses were conducted on distance, time and related measurements in their raw units (i.e., frames and pixels—note these units remained consistent across all trials). For the reporting of data and model predictions in figures and tables the units of distance and time were converted to millimetres (where 1 mm was equal to 2.7 pixels) and seconds (where 1 s was equal to 25 frames), respectively.
Ethics statement
All procedures regarding the use of animals in research followed United Kingdom guidelines and were approved by the University of Bristol Ethical Review Group (UIN UB/17/060).
Reporting summary
Further information on research design is available in the Nature Research Reporting Summary linked to this article.
Source: Ecology - nature.com
