Machiavellian strategist or cultural learner? Mentalizing and learning over development in a resource-sharing game

Theorists have sought to identify the key selection pressures that drove the evolution of our species' cognitive abilities, life histories and cooperative inclinations. Focusing on two leading theories, each capable of accounting for many of the rapid changes in our lineage, we present a simple experiment designed to assess the explanatory power of both the Machiavellian Intelligence and the Cultural Brain/Intelligence Hypotheses. Children (aged 3-7 years) observed a novel social interaction that provided them with behavioural information that could either be used to outmanoeuvre a partner in subsequent interactions or for cultural learning. The results show that, even after four rounds of repeated interaction and sometimes lower pay-offs, children continued to rely on copying the observed behaviour instead of harnessing the available social information to strategically extract pay-offs (stickers) from their partners. Analyses further reveal that superior mentalizing abilities are associated with more targeted cultural learning - the selective copying of fewer irrelevant actions - while superior generalized cognitive abilities are associated with greater imitation of irrelevant actions. Neither mentalizing capacities nor more general measures of cognition explain children's ability to strategically use social information to maximize pay-offs. These results provide developmental evidence favouring the Cultural Brain/Intelligence Hypothesis over the Machiavellian Intelligence Hypothesis.

Notes: Coe cients are presented as odds ratios, so "1" indicates no e↵ect. Standard errors and confidence intervals are robust and use two-way clustering on both individuals and sites. 95% confidence intervals are reported below each coe cient in parentheses. Round of the game was treated as a continuous variable. Sex was centered on the percentage of males to ease interpretation of the other coe cients for the entire sample. For those interested in significance testing, ***, **, and * indicate p-values below 0.01, 0.05 and 0.  Notes: Coe cients are presented as odds ratios, so "1" indicates no e↵ect. Standard errors and confidence intervals are robust and use two-way clustering on both individuals and sites. 95% confidence intervals are reported below each coe cient in parentheses. Round of the game was treated as a continuous variable. Sex was centered on the percentage of males to ease interpretation of the other coe cients for the entire sample. For those interested in significance testing, ***, **, and * indicate p-values below 0.01, 0.05 and 0.  Notes: Coe cients are presented as odds ratios, so "1" indicates no e↵ect. Standard errors and confidence intervals are robust and use two-way clustering on both individuals and sites. 95% confidence intervals are reported below each coe cient in parentheses. Round of the game was treated as a continuous variable. Sex was centered on the percentage of males to ease interpretation of the other coe cients for the entire sample. For those interested in significance testing, ***, **, and * indicate p-values below 0.01, 0.05 and 0.1.

6
We modeled whether the relationships between cognitive ability and overimitation presented in the main 7 text (see Table 3) could be further qualified by examining the associations between the three subscales of identify synonyms and antonyms of said word. The visual matching test has participants identify (e.g., point 13 to) as many of matching pairs of numbers in a row of six numbers as quickly as they can in a three-minute 14 time period. The subcsale analyses reveal that the already small e↵ect of cognitive ability on overimitation 15 may be driven mostly by verbal and visual matching scores rather than concept formation (see Table S4 for 16 details). Notes: Coe cients are presented as odds ratios, so "1" indicates no e↵ect. Standard errors and confidence intervals are robust and use two-way clustering on both individuals and sites. 95% confidence intervals are reported below each coe cient in parentheses. For those interested in significance testing, ***, **, and * indicate p-values below 0.01, 0.05 and 0.  Notes: Coe cients are presented as odds ratios, so "1" indicates no e↵ect. Standard errors and confidence intervals are robust and use two-way clustering on both individuals and sites. 95% confidence intervals are reported below each coe cient in parentheses. The CONTROL condition (Intercept; controlling for other variables) is the reference category for condition e↵ects. Round of the game was treated as a continuous variable. Sex was centered on the percentage of males to ease interpretation of the other coe cients for the entire sample. For those interested in significance testing, ***, **, and * indicate p-values below 0.01, 0.05 and 0.1. Notes: Coe cients are presented as odds ratios, so "1" indicates no e↵ect. Standard errors and confidence intervals are robust and use two-way clustering on both individuals and sites. 95% confidence intervals are reported below each coe cient in parentheses. The CONTROL condition (Intercept; controlling for other variables) is the reference category for condition e↵ects. Round of the game was treated as a continuous variable. Sex was centered on the percentage of males to ease interpretation of the other coe cients for the entire sample. For those interested in significance testing, ***, **, and * indicate p-values below 0.01, 0.05 and 0.1. Notes: Coe cients are presented as odds ratios, so "1" indicates no e↵ect. Standard errors and confidence intervals are robust and use two-way clustering on both individuals and sites. 95% confidence intervals are reported below each coe cient in parentheses. The CONTROL condition (Intercept; controlling for other variables) is the reference category for condition e↵ects. Round of the game was treated as a continuous variable. Sex was centered on the percentage of males to ease interpretation of the other coe cients for the entire sample. For those interested in significance testing, ***, **, and * indicate p-values below 0.01, 0.05 and 0. is subtly but importantly di↵erent from inferring something normative. As an analogy, young basketball 32 players might watch an experienced player shoot using an underhand technique (e.g. NBA star Rick Barry).

33
They might assume that you must shoot underhanded in basketball (or else it doesn't count and causes a 34 'turnover'); or, they might see this as the usual approach that people take in shooting, but that you can 35 shoot overhand if you prefer (but others may think it is a bit odd). To examine this question, we conducted 36 a small supplemental study in which children played the sticker game in an identical manner as in our main 37 study. Following the game, children were asked a series of questions regarding the interaction to determine 38 how children understood the 'rules' of the game.

46
Participants in this study were randomly assigned to one of two conditions (EVEN: N = 20; SELFISH: N = 19. The CONTROL condition from Study 1 was not replicated here, as there was no cause for 48 concern regarding imitation e↵ects as allocations were occluded from the participants' view. We included 49 the SELFISH condition (but not the NICE condition) because if responses to follow-up questions in the 50 SELFISH condition indicate that children understood that the stickers could be distributed di↵erently than 51 how they had observed, yet continued to imitate the unfavorable uneven distribution that resulted in reduced 52 sticker payo↵s, we could be more confident that these behaviors are the result of a propensity for imitation 53 and not a lack of understanding or strict rule following. All participants played the game with the same two 54 female experimenters who played the same role (proposer or responder) with each participant. Otherwise, 55 the sticker game proceeded exactly as described in the main study. After the game, the experimenter who 56 had played as proposer in the sticker game asked the participant six questions. These questions are described 57 in tandem with the results below.

59
In this section, we first show that we replicated the relevant results from the main text in this supplemental 60 experiment and then explore how our participants understood the game using our interview protocol.

61
S 3.2.1 Replicating relevant results

62
As in the main study, children's allocations were strongly influenced by the allocation strategy they saw in 63 the observation phase (see Figure S1 and Table S8 for model summary details). Note that the regression 64 coe cients here, expressed in odds ratios, are relative to the SELFISH condition, (not a CONTROL condition 65 as is presented in the main text), which is why they are so large. The confidence intervals are large because 66 with 80 total observations in EVEN Condition, we have only 5 uneven observations. Nevertheless, the main 67 results for these conditions in the main text are replicated here. Notes: Coe cients are presented as odds ratios, so "1" indicates no e↵ect. Standard errors and confidence intervals are robust and clustered on individuals. 95% confidence intervals are reported below each coe cient in parentheses. The SELFISH condition (Intercept; controlling for other variables) is the reference category for condition e↵ects. Round of the game was treated as a continuous variable. Sex was centered on the percentage of males to ease interpretation of the other coe cients for the entire sample.

68
For those interested in significance testing, ***, **, and * indicate p-values below 0.01, 0.05 and 0.1. The post-game interviews of these participants unfolded as follows. First, at the completion of game, the 70 experimenter exclaimed that the other research assistant had forgotten the rules of the game, and asked 71 whether or not the child could teach her how to play the game. The child was then asked to indicate 72 whether not the experimenter was allowed to distribute stickers in (1) an even manner (two in each basket),

73
(2) uneven manner (three in one basket and one in the other), and (3) Figure S1: Predicted probability of even distributions in the two conditions of Study 2 across the four rounds (Panel A) and age (Panel B). Predictions were generated from Model 4 in S8. The shaded regions show the 95% confidence intervals based on subject-level clustering. The grey lines reproduce predicted estimates from the SELFISH condition in Study 1 for comparison. Study 1 recruited children of a wider age-range than Study 2.
as did 2 out of 13 in the SELFISH condition. Then, children's memory of their own behaviors in the game 89 was assessed in the same manner. Six children incorrectly remembered their own decisions: 2 in the EVEN 90 condition and 4 in the SELFISH condition.

91
Following these memory checks, the experimenter recounted how the model distributed stickers in the 92 observation phase and what the child did at test and then asked, "Could you have put the stickers in the 93 baskets in any other way?". The results were almost identical with those above. Again, nearly two-third of 94 participants explained that they could have deviated from the demonstrators' allocation. However, in the 95 EVEN condition, 12 out of 20 children again thought that an uneven distribution would not be allowed.

96
In the SELFISH condition, 2 children out of 19 thought only the demonstrators uneven allocation was 97 allowed-that is, 17 children thought they could deviate from what they saw the demonstrator do.

98
Next, participants were asked, "Would you have been allowed to just take the stickers without even 99 putting them into the baskets?" The answer to which is technically 'yes', however we wanted to see if 100 children understood this situation to be a game with a certain set of boundary conditions. And unlike the and 12 in the SELFISH condition); 9 in the EVEN condition hinted at an explicit understanding that an even