Conservatism and “copy-if-better” in chimpanzees (Pan troglodytes)

Social learning is predicted to evolve in socially living animals provided the learning process is not random but biased by certain socio-ecological factors. One bias of particular interest for the emergence of (cumulative) culture is the tendency to forgo personal behaviour in favour of relatively better variants observed in others, also known as the “copy-if-better” strategy. We investigated whether chimpanzees employ copy-if-better in a simple token-exchange paradigm controlling for individual and random social learning. After being trained on one token-type, subjects were confronted with a conspecific demonstrator who either received the same food reward as the subject (control condition) or a higher value food reward than the subject (test condition) for exchanging another token-type. In general, the chimpanzees persisted in exchanging the token-type they were trained on individually, indicating a form of conservatism consistent with previous studies. However, the chimpanzees were more inclined to copy the demonstrator in the test compared to the control condition, indicating a tendency to employ a copy-if-better strategy. We discuss the validity of our results by considering alternative explanations and relate our findings to the emergence of cumulative culture. Electronic supplementary material The online version of this article (doi:10.1007/s10071-016-1061-7) contains supplementary material, which is available to authorized users.

chimpanzee, while ignoring possible exchange attempts by the chimpanzee who just exchanged a token with the experimenter (to prevent the building up of frustration, the experimenter attuned to the speed of the subject, which worked adequately due to the continued and predictable exchange behaviour of the demonstrator).
During testing, the subject and the demonstrator had all three token-types available ad libitum. The demonstrator, however, could only use one particular tokentype due to access to the other token-types being blocked with transparent barricades.
The subject was unable to see these barricades, resulting in the desired situation where the subject plausibly perceived the demonstrator making a free choice between the available token-types.
Throughout testing, the experimenter maximized the number of demonstratortrials that the subject would observe by i) having all commodities concentrated in one small space (token containers, exchange windows, food rewards), ii) placing the exchanged tokens on a table connecting the exchange-windows of the subject and the demonstrator, iii) calling out the name of the subject during the demonstrator-exchange, and iv) timing the demonstrator-exchange such that the subject would likely observe it (i.e. most subjects remained close to the exchange window throughout the testing phases, yet some walked around in a somewhat predictable pattern). Nevertheless, not all demonstrator exchanges were observed by the subjects, so we included "observed demonstrator-trials" (ad-hoc assessed during the experiment by the experimenter) as an offset term in our models.
Notably, by counterbalancing the token-types that the subjects were trained on in relation to the token-types used by the demonstrator, we were able to control for the effect of potential token preferences (these factors were additionally controlled for in the models). Similarly, by using three different token-types, we were able to control for individual learning: in case subjects abandoned their trained token and started exploring the other tokens, they may not have necessarily selected the demonstrator's token, which would have been the case if we had only used two different token-types.

Analysis
We opted for initial permutation tests because the token choices were non-independent, both within and across subjects, and the binomial Generalized Linear Mixed Models (GLMMs) could not ascertain whether certain token-types were chosen more than specific others due to the fact that we used three instead of two token-type alternatives (i.e. the GLMM intercept estimate would indicate whether token-type A was used more often than token-types B and C combined).
In the first GLMM, the response variable was "yes/no used trained token", in keeping with predictions based on previous research [2,3]. In the second GLMM the response variable was "yes/no copied the demonstrator", in keeping with our 'copy-ifbetter' hypothesis (based on [4,5]).
The GLMMs were fitted in R [6] using the function lmer of the R-package lme4 [7]. The significance of the full model as compared to the null model (comprising only the control variables, including the random effects) was established using a likelihood ratio test (R function anova with argument test set to "Chisq"; [8]). To allow for a likelihood ratio test we fitted the models using Maximum Likelihood (rather than Restricted Maximum Likelihood; [9]). P-values for the individual effects were based on likelihood ratio tests comparing the deviances of the full and respective reduced models.

Do chimpanzees copy if better?
Here, we present additional information regarding chimpanzees' proclivity to copy the demonstrator. First, chimpanzees' individual choices on their first trials after having observed the demonstrator are depicted in Table S1. Table S1. Proportion of first trials (2 per condition, given that each condition comprised 2 sessions of 10 trials each) in which the demonstrator was copied. In the control condition, the demonstrator received a similar-value reward for exchanging as the subject (carrot), in the test condition the demonstrator received a higher-value reward for exchanging than the subject (banana). When the demonstrator was not copied (e.g. when the proportion is "0"), the subject either chose its trained token or the random one. Note that individual variation exists such that only some chimpanzees (n=4) copied-ifbetter on their first trials.  Figure S1. This figure adds information regarding chimpanzees' responses over time. Figure S1. Proportion of demonstrator copying per trial for each condition (control condition = "carrot"; test condition = "banana"). The proportions are relatively low due to the fact that instead of copying the demonstrator, chimpanzees were mostly conservative and chose their trained token. Yet, when they copied the demonstrator, they did so more frequently when the demonstrator received banana (test condition) compared to when the demonstrator received carrot (control condition). The dotted line indicates the model estimation, the grey area the 95% confidence intervals.

Chimpanzee Control condition Test condition
Third, we present results from a derived GLMM in which subjects' choices for their trained token-type are omitted, such that their choice comprises a dichotomous one between copying the demonstrator or using the third, random option. This model is structured in the exact same way as the reported GLMM with response variable "yes/no copied the demonstrator", yet adds insight on the subjects' relative choosing between copying the demonstrator and the third, random option across conditions in one model (n=143). The effect of "condition" also in this model is highly significant (χ 2 = 36.13, Δdf = 1, p < 0.001; Estimate ± SE = -10.84 ± 3.72), corroborating the findings reported in the main text.
Finally, we present the GLMM results for all estimated parameters in the copyif-better model (response variable is "yes/no copied the demonstrator"). We present both the model summaries and the single term analyses based on model comparisons (comparing the deviance of the models with and without the concerning parameter).
Note that in order to determine whether a factor significantly contributes to the prediction of subjects' responses, the single term analyses (based on Maximum Likelihood Ratio tests) results are conclusive [9].

Interpretation of control variables
In our experimental setup and in our GLMM analyses, we controlled for the effects of the specific token-type the subject was trained on (trained.token), the specific tokentype used by the demonstrator (demonstrator.token) and the order by which subjects participated in the test and control conditions by counterbalancing all relevant factors and incorporating them in our GLMM models as fixed effects, respectively. We found no significant effects of order and trained.token, and a significant effect of demonstrator.token (see Table S2b). Inspecting the model summaries, we find that whenever the demonstrator was using the green or grey token, she was more likely to be copied compared to when she was using the blue token (see Table S2a). Additionally, we report that chimpanzees seemed decreasingly inclined to copy the demonstrator over time across conditions (main effect for trial; see Table S2a). Lastly, in line with the results depicted in Table S1, we found significant differences between the chimpanzees in their proclivity to apply a copy-if-better strategy across the entire set of trials, for which we controlled in the random effect structure of our GLMM (subject: χ 2 = 40.2, Δdf = 1, p < 0.001). We chose not to report the results of these control variables in the main text, however, because the results relevant to our hypotheses (the effect of "condition") exist despite their respective influences (also see L129-132 in main text).

DISCUSSION
Note that our paradigm simultaneously represents an inequity aversion design with the additional feature of offering the disadvantaged individual a way out of his predicament.
By rewarding the demonstrator with a higher value reward than the subject, while both individuals engage in a similar token-exchange procedure (i.e. there is no difference in the amount of required effort between the individuals), we created an inequity between the two individuals that is typically used in inequity aversion studies (e.g. Bräuer et al. 2009; Brosnan et al. 2010). In contrast to these studies, however, our paradigm incorporated an option for the disadvantaged individual to undo the inequity and therefore level the playing field (also see Hopper et al. 2013). We envisage that offering individuals alternatives to the inequity inflicted upon them will yield interesting insights into their (lack of) flexible application of (social) strategies.