Skip to main content
Dryad

Data set for: Behavioral traits that define social dominance are the same that reduce social influence in a consensus task

Cite this dataset

Jordan, Alex (2020). Data set for: Behavioral traits that define social dominance are the same that reduce social influence in a consensus task [Dataset]. Dryad. https://doi.org/10.5061/dryad.qz612jmbz

Abstract

Dominant individuals are often most influential in their social groups, affecting movement, opinion, and performance across species and contexts. Yet behavioral traits like aggression, intimidation, and coercion, which are associated with and in many cases define dominance, can be socially aversive. The traits that make dominant individuals influential in one context may therefore reduce their influence in other contexts. Here we examine this association between dominance and influence using the cichlid fish Astatotilapia burtoni, comparing the influence of dominant and subordinate males during normal social interactions and in a more complex group-consensus association task. We find that phenotypically dominant males are aggressive, socially central, and that these males have a strong influence over normal group movement, whereas subordinate males are passive, socially peripheral, and have little influence over normal movement. However, subordinate males have the greatest influence in generating group-consensus during the association task. Dominant males are spatially distant and have lower signal-to-noise ratios of informative behavior in the association task, potentially interfering with their ability to generate group-consensus. In contrast, subordinate males are physically close to other group members, have a high signal-to-noise ratio of informative behavior, and equivalent visual connectedness to their group as dominant males. The behavioral traits that define effective social influence are thus highly context-specific and can be dissociated with social dominance. Thus, processes of hierarchical ascension in which the most aggressive, competitive, or coercive individuals rise to positions of dominance may be counter-productive in contexts where group-performance is prioritized.

Methods

Captive Astatotilapia burtoni descended from a wild caught stock population (41) were maintained in stable community tanks (26°C, pH 7.5-8.0, 12:12 light:dark cycle) until transfer to the experimental arenas (205 liters, 108 x 54.6 x 42.7 cm, 17 - 20cm cm water depth, 26°C, pH 7.5-8.0, 12:12 light:dark cycle). Groups consisted of both males and females between 40 and 70mm SL, though only males were used as informants since they have clear phenotypic indicators of social dominance, whereas females, although likely having social dominance hierarchies, have no reliable visual indicators of dominance status. In analyses of routine group interactions, we used six groups of ten fish (n = 60 individuals). In the association tasks, we used eight groups of eight fish during initial training (n = 64 individuals), followed by eight further groups each with either a subordinate informant (n = 56 additional individuals) or a dominant informant (n = 56 additional individuals). These informants were taken from the initial group of eight naïve individuals, which had now successfully responded to the association tasks. One of these group sets was abandoned due to technical failure and was not included in the final data set, leaving seven groups used in analyses of the association task (n = 154 individuals in association trials). All work was conducted in compliance with the Institutional Animal Care and Use Committee at The University of Texas at Austin.

 

Experimental paradigm

The behavioral task measured the number of trials taken to reach group-consensus in a simple association task using a food reward and colored light-emitting diodes. Experiments were conducted in experimental arenas with two automatic fish feeders (EHEIM) mounted on opposite ends of each tank. The motor control pins of the feeders were rewired and externally controlled by a digital I/O switch slaved to an Arduino Uno microcontroller that also controlled one diffuse RGB LED mounted directly under each feeder (code available github.com/jordanlabmpi/). Four times a day (0830, 1130, 1430, 1730) for five consecutive days, the association stimulus would be remotely triggered using the Arduino microcontroller, which randomly assigned both LEDs to simultaneously display one of two colors (RGB 255,60,0 [yellow-orange] or 0,255,255 [cyan]) for three seconds, followed by three seconds of no stimulus; the Arduino would then trigger the autofeeder associated with the LED that displayed the yellow-orange stimulus, providing a portion of Tetramin flake food. Neither of these color stimuli elicits an innate response in the focal animals, due for example to inherent color preferences (42), allowing their use as conditioned stimuli in an association learning paradigm. However, the color of the rewarded stimulus affected the time taken for the association to be achieved, and we therefore kept the rewarded stimulus color consistent throughout all trials and randomized the location of colors to prevent spatial learning. A networked Logitech HD 1080p webcam was mounted above each tank and automatically scheduled to record for one minute before and after each training event using iSpy open source security camera software.

 

Association task

Using the protocol described above, groups of eight A. burtoni (4 males, 4 females) underwent the training four times a day for five days. Animals remained in the experimental tanks described above for the entire five days, and were not disturbed or moved between association trials. Behavior and interactions were recorded for 20 minutes prior to, during, and 10 minutes following the stimulus onset. Group behavioral response to the task during all trials was scored as the proportion of individuals in the group that responded to the light stimulus by swimming towards it in the three seconds of LED stimulus prior to the three second pause and subsequent delivery of food. A successful group response was defined as seven or more of the eight group members swimming directly toward the positive stimulus in less than one second of stimulus onset, in two or more consecutive trials. The second successful trial was used as the value in subsequent analyses (e.g. if a group responded successfully in trials 11 and 12, we recorded a successful response for that group at trial 12).

 

Within the five-day training period, all groups showed a behavioral shift from a lack of coordinated movement to a consensus movement toward the conditioned stimulus. After five days, all initially naïve groups reached consensus movement towards that correct cue, and subsequently one dominant male (“dominant”) and one subordinate male (“subordinate”) were placed into new groups (3 males, 4 females; total group size 8 individuals) that were naïve to the association task (Figure 2a). For groups with dominant males, all three other males were smaller than the dominant, while for groups with subordinate males, at least one male was larger than the subordinate. We did not observe any dominance shifts (i.e. a dominant becoming a subordinate in a new group, or vice-versa) in these group transitions. Seven groups each with either a dominant or subordinate informant were then placed in identical training protocols as previously and the time taken to group-consensus measured.

 

Deep-learning based automated tracking and analysis of behavior

We trained an implementation of a Mask and Region-based Convolution Neural Network, or Mask R-CNN (43, 44), on a small set of 34 manually labeled images. Given the simplicity of this highly specific segmentation task and using standard image augmentations such as rotation, horizontal or vertical flipping, this allowed the accurate detection and segmentation of individual fish in each of the video frames. In terms of raw data, Mask R-CNN predictions resulted in binary pixel masks for each frame and individual respectively. These masks were then skeletonized into 1 px midlines along each mask’s long axis using morphological image transformations. Subsequently, this allowed the estimation of fish spine poses (45) as seven equidistantly spaced points along these midlines. The second spine point represents an individual’s head position and the vector pointing from the second to the first spine point its orientation. This positional data was then used to automatically reconstruct continuous fish trajectories using a simple, distance-based identity assignment approach. Accuracy and high detection frequency were visually verified with a Python-based GUI (45) developed within the lab, that was also used to manually correct false identity assignments and losses. Mask R-CNN predictions resulted in a mean coverage of 96.3% throughout all analyzed videos and automatic trajectory assignment in an average of 14 losses per individual. 1.6% of all detections were false positives or poorly segmented, resulting in a mean coverage of 94.8% in the manually corrected trajectories. See SI Appendix Video 1 for a visualization of the tracking pipeline. All code is available at https://github.com/jordanlabmpi/social-influence.

Behavioral, visual, and spatial connectivity analyses

In order to examine baseline differences in the behavior of dominant and subordinate males in social contexts, we placed six additional groups of 10 individuals in identical tanks as described above and filmed their behavior for 5 minutes in the absence of external stimuli (‘routine social context’). We calculated the behavioral, visual, and spatial interactions between all fish of each group. To estimate the number of behavioral interactions that dominant and subordinate males had with other group members, trajectory data was used to determine events with elevated swimming speed (above the 95th percentile of the speed distribution). The first two individuals passing this threshold in such events were treated as event initiator and responder, and a delay time between the two individuals was calculated (Figure 1c, d). All other group members passing the threshold while either the initiator or the responder were still at elevated speed were considered to be part of the same event, but not counted as direct responders. Each separate event lasted until the speed of all initiators or responders fell below the threshold again.

 

From these dyadic initiator-responder events, we created behavioral interaction networks using the ’networkx’ package (46) in Python. Here, the count of the directed, pairwise events between each pair of network nodes defines the weight of the respective edge (directed from initiator to responder). This allowed the calculation of out-edges Katz centrality (47) as a measure of behavioral influence in standard conditions (Figure 1). Related to this, we also calculated the initiator count for each fish as the number of events in which an individual was the initiator. Additionally, the ratio of the total time spent in these events above the speed threshold to the full duration of the recording was calculated for each fish, constituting individual hypothetical noise frequency in the social training context (fast, directed movement in absence of LED stimulus).

 

Spatial connectivity between group members was calculated as their mean pairwise distances. We then computed the visual connectivity as the mean angular area subtended by each individual on the retinas of all other group members, utilizing the contours of the Mask R-CNN detection results as occluding objects in a ray-casting approach (Figure 3a). Casting rays from both eyes of a focal fish towards these contours (including the focal individual), we modelled the nearly complete field of view known from other freshwater fishes (48). These measures generated three connectivity scores for each dominant and subordinate group member: a behavioral (‘interaction’) connectivity, spatial (‘association’) connectivity, and visual connectivity. Finally, we conducted a principal component analysis (PCA; Figure 3e, See SI Appendix Figure 5a) on the speed threshold event ratio (noise frequency) and connectivity scores. After determining the appropriate number of clusters using silhouette analysis (49), we performed a clustering analysis of the principal components (50) (k-means clustering, k = 2; See SI Appendix Figure 5b) to assess the overall consistency of metrics derived from trajectory and network analyses with the phenotypic indicators of dominance that we used to identify dominant from subordinate individuals.

 

Data analysis and statistics

For time-to-consensus movement analysis, we conducted a Kaplan-Meier Survival Analysis (51) with the ‘survival’ package in R (52), using the first of two consecutive trials in which seven or more individuals responded to the stimulus onset as time to criterion, and right-censoring groups that did not complete the social learning task. We then fit a Cox Proportional-Hazards model with the absence or social status of the informed individual as the single covariate. This allowed the comparison of the survival estimates between the groups that were initially naïve to the stimulus during the training regime and the groups with both dominant and subordinate informants. Further, we validated the proportional hazard assumption for each of the groups using the same package in R.

 

For comparisons of the baseline behavioral traits of dominant and subordinate fish, we either performed network randomization tests in Python or linear models in R. In the case of network centrality, the mean angular area subtended on the retina, and the mean pairwise distance, network randomization tests were necessary because these metrics, by definition, are non-independent for individuals of the same group and social network. Further, the initiator count is also based on dyadic interactions and should be considered non-independent within the networks. Therefore, we preformed network randomization (i.e. node randomization by assigning the dominance status to an individual that was randomly drawn from the group, n = 1000) for each of the six ’routine social context’ groups to construct null models in which social dominance is detached from the respective response variable (53). For each randomization of the six networks and for each response variable, an estimate was calculated as the mean difference of the respective metric between the assigned dominant individual and the mean of the remaining, assigned subordinate individuals. These estimates can be considered as null distributions for the test statistics, and were used to calculate two-tailed p-values for the actually observed differences between dominant and subordinate individuals. Accordingly, the null hypotheses that the observed differences were drawn from the respective null models were rejected when the corresponding p-value was smaller than, or equal to, the significance level α = .05. See SI Appendix Figures 1-3 for visualizations of the network randomization tests.

 

By contrast, the noise frequency (ratio of time spent in above speed threshold events to trial duration) is not dependent between individuals of a group. Here, we fit a linear regression model with social status as predictor and noise frequency as response variable. Further, we tested the model’s assumptions of normality of residuals with the Shapiro-Wilk test and homoscedasticity with the Breusch-Pagan test (54).

Usage notes

Dataset for: Behavioral traits that define social dominance are the same that reduce social influence in a consensus task

- "delay_times.csv", "learning.csv" and "social_parameters.csv" were used in statistical analyses
- "tracks_T[3, 4, 5, 6, 7, 12].csv" provide simplified tracking data (only head position per individual per frame)

All analysis, raw data and further scripts can be found at https://github.com/jordanlabmpi/social-influence

Funding

Deutsche Forschungsgemeinschaft, Award: DFG Cluster of Ex- cellence 2117 “Centre for the Advanced Study of Collective Behavior” Grant 422037984

National Science Foundation, Award: IOS1354942