SUPPLEMENTARY MATERIALS
Dominance and social information use in a lizard
FontiKar, Martin J. Whiting, Daniel W.A. Noble
Stability of dominance relationships
We assessed the stability of the dominance relationship within pairs by staging another round of contests (n = 26, two pairs did not re-fight). Previous studies in lizards have shown that winners of the first fight tend to consistently win in the second fight with the same individual (Podarcis hispanica, López and Martín 2001). However, aggression towards a previous rival significantly decreases in subsequent encounters and so we predicted these contests would result in lower levels of overt aggression (Fox and Baird 1992; López and Martín 2001; Whiting 1999). As a consequence, in interactions where the opponents did not interact aggressively (7/28 contests), we assigned the dominant and subordinate individual using fleeing and tail-waving behaviors, as these are considered subordinate displaysin this species(Kar et al. 2016). In 91% (21/23) of pairs the dominance relationship was stable. Five of 28 pairs (18%) did not interact during the second round of contests. Given that the dominance relationship was stable in 91% of pairs, we assumed that dominance relationship within the five pairs that did not interact were also stable.
Randomization analysis with uncertain dominance relationships
There were seven pairs of lizards where their dominance relationship were uncertain as the pairs either did not fight in the second round of contests (5/26) or their dominance statuses had changed (2/26). As a consequence, we randomized the dominance status of these lizards by creating 10,000 permutated datasets to assess whether changes in the dominance status of these lizards influenced our results.We found that our original model coefficients and standard errors presented in the paperencompass all the possible coefficient values generated by the permutated datasets (Fig. S2). This suggests that our estimates are relatively conservative and potential misclassifications of observer dominance status have little effect on our overall conclusions for either task.
Figure S1 Histograms of coefficient values from generalized linear models exploring the effects of a lizard’s dominance status and treatment group on the mean trials taken to learn the a) association task b) reversal task. The coefficient values were generated from 10,000 permutated datasets where the dominance status of seven lizards were randomly allocated. The black point represents the coefficient value from the original model presented in the paper. The error bars represent the standard error from the original model presented in the paper.
Training and social demonstration
All lizards were trained to displace a lid from a dish to access a food reward. We used a similar training protocol as Noble et al. (2014),Clark et al. (2014) and Leal and Powell (2011), however we made some small changes to expedite the training process. This approach quickly trained many of the lizards that were having difficulty learning how to flip a fully covered lid and was particularly suitable for our needs given that we were not interested in instrumental learning abilities and it allowed us to alleviate time constraints during the training period. For a schematic overview of the training phases, see Fig. S2.
Briefly, lizards were first trained to eat from an open dish containing a mealworm. Lizards had to eat from the open dish a minimum of 5/6 times before graduating to the next task (phase 1). Once lizards achieved the criterion, a yellow lid was placed over the food dish so that it covered 75% of the dish (phase 2). Again, lizards had to reach the criterion before graduating to a yellow lid that fully covered the dish and that required lizards to use their snout to open it (phase 3). Some lizards had difficulty learning this task (n = 29/56) and to expedite learning we provided them with two dishes (phase 4). One of these dishes contained a lid that was 75% covered while the second dish completely covered the food well. The dish that had a fully or 75% closed lid varied between trials (either the right or left). Lizards had to eat from both dishes a minimum of 5/6 times before graduating to the next phase where both lids completely covered the food well (phase 5). Lizards that did poorly on this task (i.e. ones that did not eat from both dishes in at least 2/3 of their last trials; n = 16) were then given two dishes that were 98% covered (phase 6). The purpose of this was to train the lizards to associate each dish with food when visual cues were absent, but to not make access to this food particularly difficult (i.e. lizards could simply slide the lid off). In contrast, lizards that ate from both dishes in 2/3 of their last trials (6 trials total) were allowed to continue with one lid fully closed and the second 75% (phase 4) closed until they achieved criterion. In 13 instances, lizards that achieved the 5/6 criterion for phase 4 (two dishes - 75% and fully covered) but then were unsuccessful in the first two trials of phase 5 (two dishes fully covered) were returned to phase 4.
Figure S2Schematic overview of phases used to train lizards to displace a lid to access a food reward. Lizards must eat from the dishes 5 out of 6 consective times to advance to the next phase. Lizards that did not eat from both dishes in 2/3 consecutive trials in Phase 4 were given the an easier task (Phase 6). As lizards improved in Phase 6 and ate from both dishes in 2/3 conseuctive trials, they were returned to Phase 4 task. Lizards that struggled to eat from both fully covered dishes in the first two trials of Phase 5 were returned to Phase 4. Lizards that were assigned to be the demonstrator continued training after Phase 5.
Once all lizards were proficient at displacing lids, training for observer lizards was terminated to prevent satiation. Demonstrators continued training on a new apparatus consisting of two covered dishes; one was covered by a white lid while the other was covered by a blue lid (Fig. 1). However, the white lid was fixed to the dish and contained no mealworm, so that the demonstrator lizard could unambiguously open the blue lid during social demonstration for the association task. Demonstrators were required to eat from the blue dish 5/6 consecutive times before social demonstration commenced.
FigureS3Flow chart depicting the timeline of the social demonstrations for demonstrators and observers in the social treatment group. LC represents learning criterion. Please note that control demonstrators and observers viewed each other for the same amount of time, see Methods for more details
Extended statistical analyses
We first assessed the robustness of our learning criterion by tallying the number of correct choices after a lizard reached the learning criterion lizards that had five or more trials beyond the trial they learnt. We tested whether the number of correct choices was significant according to an exact binomial choice test (association task: n = 24, reversal task: n =23). In the association task, 18/24 (75%) lizards that had five or more trials beyond the trial they learnt chose the correct dish significantly more than expected by chance. While in the reversal task, 20/23 (88%) lizards that had five or more trials beyond the trial they learnt chose the correct dish significantly more than expected by chance. These results suggest that our learning criterion was mostly sufficient in categorizing lizards that learnt from those that did not.
We assessed whether body condition (residuals from a linear model between log body mass and log SVL) and the number of attempts differed between dominant and subordinate lizards using a generalized linear model (GLM), as individuals in poorer condition may be more motivated to do the task and learn faster. We found that body condition did not differ between dominant and subordinate individuals (Estimate = 0.014, SE = 0.008, t = 1.588, P = 0.113)
We also tested whether the dominance status of demonstrator lizards (subordinate or dominant) affected their motivation to execute the tasks in response to the observer lizards by comparing the number of trials where a demonstrator did not do the task using a GLM with negative binomial errors. All observer lizards, irrespective of their dominance status attempted 80% of their trials, suggesting that motivation did not affect their learning. Additionally, the number of times lizards did not demonstrate in the association task and reversal task was not significantly different between dominant or subordinate demonstrators (Association task - GLM: estimate = -0.383, SE = 0.3672, z = -1.043, P= 0.297; Reversal task - estimate = -1, SE = 0.6872, z = -1.455, P= 0.166).
Focusing on rates of learning, we modeled the ‘probability of choosing the correct dish first’ and the probability of ‘choosing only the correct dish’ across trials using generalized linear mixed models (GLMM) with binomial errors (‘logit’ link). We also modeled the log transformed latency to displace the correct lid across trials using a GLMM with Gaussian errors. We tested whether there was social information use by comparing these three variables between control and social lizards (i.e. we pooled lizards in the social and control groups regardless of dominance status). We also tested whether dominant and subordinate lizards used social information difference by fitting an interaction between dominance status and treatment group.
For our GLMMs, we used the MCMCglmm package in R (Hadfield 2010; R Development Core Team 2010). We included a random intercept and slope for trial number within each individual to account for non-independence of observations on the same lizards. In all models, trial number was also included as a fixed effect .We ran all our GLMMS using a prior specification of V = and nu = 0.002 for the random effect variance-covariance matrix. We used default uniform priors for our fixed effects.We ran three separate MCMC chains for each model and compared the convergence of chains using the Gelman-Rubin test in the R package ‘coda’ (Plummer et al. 2006). Each chain was run for 2 000 000 iterations and sampled at every 5000 iterations (thinning interval) with a burn-in of 10,000 iterations. We visually inspected the trace plots of our samples to ensure chains were mixing well and performed Geweke and Heidelberg auto-correlation diagnostics to check our samples were not strongly correlated. Parameter estimates we report are pooled posterior modes across the three MCMC chains and 95% credible intervals. Parameter estimates were considered significant when the credible intervals did not include zero.
Finally, we calculated Bayes factors for all our GLMMS to test how likely our data occurred under H1comparedH0. In all cases, H1 represented our full model and H0was a reduced model with the relevant parameter excluded.
GLMM results
In the association task, there were no differences between treatment groups or dominance statuses in the probability of making a correct choice or latency to displace the correct dish. However, lizards in the control treatment had ahigher probability of choosing the correct dish only, compared to social lizards (Table S1a).There was a trend for dominant lizards to have a higher probability of choosing the correct dish and choosing the correct dish only. However, Table S2a reveals thatthese probabilities, as well as the latency to displace the blue lid, did not depend on treatment groups, a lizard’s dominance status, or their interaction.
In the reversal task, there were no differences between control and social lizards in the probability of choosing the correct dish and only the correct dish, as well as, the latency to displace the white dish (Table S1b). Additionally, these variablesdid not depend on a lizard’s treatment group, dominance status, or their interaction (Table S2b).
In the association task, there was very weak evidence for H0 in our models testing whether for differences between social and control treatment groups (i.e. we pooled lizards in the social and control groups regardless of dominance status) (Table S2a). Whereas in the reversal task, there was very weak to no evidence for H1in models testing whether for differences between social and control treatment groups (Table S2b).
In our models testing for differences between dominant and subordinate lizards, we found very weak to no evidence for H0 in both the association task and the reversal (Table S4)
Table S1Pooled posterior modes and 95% credible intervals from a Bayesian Markov chain Monte Carlo generalized linear mixed effects model (MCMC-GLMM) examining the effects of a lizard’s treatment group (social or control, i.e. we pooled lizards in the social and control groups regardless of dominance status), standardized SVL and trial number on the log odds of making a correct choice, the log odds of choosing the correct dish only and the log latency to displace a lid from the correct dish in the: a) association task (Nobs = 498, Nlizards = 28) and b) the reversal task ( Nobs = 710, Nlizards= 27). Bolded estimates are significant.a) Association Task
Log odds of making correct choice / Log odds of choosing correct dish only / Latency to displace lid from correct dish
Estimate / Lower / Upper / Estimate / Lower / Upper / Estimate / Lower / Upper
Intercept / 0.59 / -0.34 / 1.21 / -3.63 / -4.78 / -2.35 / 5.90 / 5.56 / 6.25
Treatment CONTROL / 0.69 / -0.27 / 1.88 / 1.41 / 0.23 / 3.22 / 0.17 / -0.51 / 0.67
Scaled SVL / 0 / -0.53 / 0.52 / -0.71 / -1.47 / 0.09 / -0.23 / -0.44 / 0.08
Trial number / 0.16 / 0.07 / 0.23 / 0.13 / 0.05 / 0.21 / -0.05 / -0.06 / -0.02
b) Reversal Task
Intercept / 0.94 / -0.21 / 1.85 / -2.72 / -4.56 / -1.5 / 5.34 / 4.82 / 5.62
Treatment CONTROL / 0.56 / -0.85 / 2 / 0.74 / -1.20 / 2.32 / 0.11 / -0.47 / 0.75
Scaled SVL / -0.21 / -0.91 / 0.47 / -0.70 / -1.57 / 0.22 / -0.16 / -0.44 / 0.17
Trial number / 0.05 / 0.01 / 0.1 / 0.10 / 0.04 / 0.15 / -0.01 / -0.03 / -0.01
Table S2Bayes factor (BF)for H1 over H0given our data, in the a) association task and b) reversal task. H1 = Response ~ treatment + standardized SVL+ trial number + (1+ trial number | ID) number. H0= Response ~ standardized SVL+ trial number + (1+ trial number | ID)
a) Association Task / BF10
Log odds of making correct choice / 0.51
Log odds of choosing correct dish only / 0.06
Latency to displace lid from correct dish / 0.03
b) Reversal Task
Log odds of making correct choice / 1.35
Log odds of choosing correct dish only / 4.27
Latency to displace lid from correct dish / 1
Table S3Pooled posterior modes and 95% credible intervals from a Bayesian Markov chain Monte Carlo generalized linear mixed effects model (MCMC-GLMM) examining the effects of a lizard’s dominance status (DOM or SUB), treatment group (social or control), standardized SVLand trial number on the log odds of making a correct choice, the log odds of choosing the correct dish only and the log latency to displace a lid from the correct dish in the: a) association task (Nobs = 498, Nlizards = 28) and b) the reversal task ( Nobs = 710, Nlizards= 27). Bolded estimates are significant. Main effects are presented from a model without the interaction
a) Association Task
Log odds of making correct choice / Log odds of choosing correct dish only / Latency to displace lid from correct dish
Estimate / Lower / Upper / Estimate / Lower / Upper / Estimate / Lower / Upper
Intercept / 0.18 / -0.80 / 1.02 / -4.20 / -5.18 / -2.39 / 5.96 / 5.55 / 6.34
Status DOM / 0.65 / -0.60 / 1.38 / 0.97 / -0.53 / 2.48 / -0.08 / -0.62 / 0.41
Treatment CONTROL / 0.64 / -0.40 / 1.85 / 1.46 / -0.21 / 2.91 / 0.02 / -0.55 / 0.68
Scaled SVL / -0.06 / -0.65 / 0.48 / -0.70 / -1.80 / 0 / -0.17 / -0.46 / 0.14
Trial number / 0.15 / 0.08 / 0.24 / 0.15 / 0.04 / 0.20 / -0.05 / -0.07 / -0.02
StatusTreatment / 0.36 / -2.14 / 1.99 / 0.13 / -2.73 / 3.23 / 0.16 / -1.02 / 0.94
b) Reversal Task
Intercept / 1.15 / -0.12 / 2.32 / -2.63 / -4.25 / -0.98 / 5.32 / 4.82 / 5.71
Status DOM / -0.82 / -2.09 / 0.76 / -0.63 / -2.46 / 0.80 / -0.10 / -0.64 / 0.50
Treatment CONTROL / 0.38 / -0.80 / 2.17 / 0.74 / -1.05 / 2.79 / 0.18 / -0.51 / 0.79
Scaled SVL / -0.20 / -0.84 / 0.59 / -0.64 / -1.51 / 0.50 / -0.04 / -0.42 / 0.23
Trial number / 0.05 / 0.01 / 0.09 / 0.10 / 0.04 / 0.15 / -0.01 / -0.03 / -0.01
StatusTreatment / 0.10 / -2.77 / 2.59 / -0.72 / -4.22 / 2.53 / 0.36 / -0.83 / 1.45
Table S4Bayes factor (BF)for H1 over H0given our data, in the a) association task and b) reversal task. H1 = Response ~ status + treatment + standardized SVL+ trial number + (1+ trial number | ID) number. H0= Response ~ treatment + standardized SVL+ trial number + (1+ trial number | ID)
a) Association Task / BF10
Log odds of making correct choice / 0.91
Log odds of choosing correct dish only / 0.05
Latency to displace lid from correct dish / 0
b) Reversal Task
Log odds of making correct choice / 0
Log odds of choosing correct dish only / 0.38
Latency to displace lid from correct dish / 1
REFERENCES
Clark BF, Amiel JJ, Shine R, Noble DW, Whiting MJ (2014) Colour discrimination and associative learning in hatchling lizards incubated at ‘hot’and ‘cold’temperatures Behav Ecol Sociobiol 68:239-247
Fox SF, Baird TA (1992) The dear enemy phenomenon in the collared lizard, Crotaphytus collaris, with a cautionary note on experimental methodology Anim Behav 44:780-782
Hadfield JD (2010) MCMC methods for multi-response generalized linear mixed models: the MCMCglmm R package J Stat Softw 33:1-22
Kar F, Whiting MJ, Noble DWA (2016) Influence of prior contest experience and level of escalation on contest outcome Behavioral Ecology and Sociobiology 70:1679-1687 doi:10.1007/s00265-016-2173-4
Leal M, Powell BJ (2011) Behavioural flexibility and problem-solving in a tropical lizard Biology letters doi:10.1098/rsbl.2011.0480
López P, Martín J (2001) Fighting rules and rival recognition reduce costs of aggression in male lizards, Podarcis hispanica Behav Ecol Sociobiol 49:111-116
Noble DW, Byrne RW, Whiting MJ (2014) Age-dependent social learning in a lizard Biology letters 10 doi:10.1098/rsbl.2014.0430
Plummer M, Best N, Cowles K, Vines K (2006) CODA: Convergence diagnosis and output analysis for MCMC R news 6:7-11
R Development Core Team (2010) R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria
Whiting MJ (1999) When to be neighbourly: differential agonistic responses in the lizard Platysaurus broadleyi Behav Ecol Sociobiol 46:210-214
APPENDIX – R FUNCTION USED FOR DATA PERMUTATION
Function to permutate only the lizards that have uncertain dominance statuses. Data is the data frame you wish to permutate; varib = the variable in the data frame you wish to permutate. The name of the variable must be provided as a character; ids = the character string of individual identifiers that you need to change or permutate data for; n = the number of permutated datasets. Note that this function requires the ‘plyr’ package
permutate <- function(data, varib, ids, n){
datlist <- list() # create an empty list for each permutated dataset
nPer <- length(ids) # number of levels
var <- unique(data[,varib]) # Extract unique levels from the variable you want permutated
replicates <-plyr::ddply(data[data$id %in% ids,], .(id), summarise, n = length(status))$n
for(i in 1:n){
# randomly sample the levels to be permutated
rand <- rep(sample(var, size = nPer, replace = TRUE), replicates)
# Change relevant ids with new random category
permute <- data[data$id %in% ids,][,varib] <- rand
#Assign new data to list
datlist[[i]] <- data
}
return(datlist)
}