Breeding Cooperation: Cultural Evolution in an Intergenerational Public Goods Experiment

This paper investigates the evolution of cooperation across multiple laboratory generations in an experimental public goods game. Theories of cultural evolution show how cooperative equlibria can be supported by the transmission of behavioral norms across generations. These types of cultural evolutionary processes are important for political science topics ranging from public policy to political participation. One of the best-established findings in the massive literature on experimental social dilemmas is that within-game communication increases cooperation. We find that it is possible to breed cooperation by selectively exposing later generations of subjects to cooperative messages from previous generations. We propose a number of potential reasons for the fact that between-generation communication (or advice) is at least as strong as within-game communication. This paper investigates the evolution of cooperation in an intergenerational public goods experiment. Theories of cultural evolution show how cooperative equlibria can be supported by the transmission of behavioral norms across generations. These types of cultural evolutionary processes are important for political science topics ranging from public policy to political participation because they are an important process by which behaviors and attitudes are transmitted among interacting individuals. One of the best-established findings in the massive literature on experimental social dilemmas is that within-game communication increases cooperation. We find that it is possible to breed cooperation by selectively exposing later generations of subjects to cooperative messages from previous generations. The effects of cultural transmission are at least as strong as within-game communication.

that norms of reciprocity and punishment of non-cooperators are passed across generations, reducing the costs of learning how to cooperate in each subsequent generation.
Such patterns of cultural transmission are crucial to the stability of cooperation in real-world societies and political systems, although the concepts of cultural evolutionary models have received little attention in political science. Cultural evolution is implicated in broad-scale processes of political socialization and citizen behavior, including cooperation problems of voting and tax compliance. Within formal political institutions such as Congress, norms of reciprocity are passed down from senior legislators to junior legislators in order to support various forms of bargaining, committee structures, and coalition-building. One of the current hot topics in public policy is the role of collaborative institutions for building cooperation in complex conflicts involving multiple issues and organizations (O'Leary et al. 2006;Sabatier et al. 2005;Lubell et al. 2002). A major challenge of collaborative institutions is maintaining agreement and cooperation in the face of high levels of turnover among participants; overcoming this challenge requires cultural transmission of cooperative norms. Ongoing cooperation games in the real world nearly always involve multiple generations of participants, and thus it is critical to understand how social learning and cultural transmission supports different behavioral norms. Our experimental design creates a situation that mirrors this realworld situation and examines how norms of cooperation are passed among individuals in a simpler setting.
It is difficult to stake out new territory in the mature (with early experiments dating to the 1950s) and vast literature on experimental social dilemmas (Komorita et al. 1992;Kollock 1998); researchers who claim to be covering new ground will often find similar examples among the thousands of experimental papers. Intergenerational experiments are no exception-although relatively new in this literature, some initial studies have been conducted using a variety of types of social dilemmas and methods of cultural transmission. One set of experiments finds that advice from previous generations has the strongest effect when it is announced publicly to all experimental subjects and generally recommends cooperative behavior (Ananish et al. 2009;Chaudhuri et al. 2006). Another series of experiments suggest that while cultural transmission has an extremely strong influence on behavior in later generations, there are some interesting asymmetric effects such as perpetuating unequal divisions of gains from cooperation (Schotter and Sopher 2003), decreasing levels of trust among senders in a trust game but increasing levels of trustworthiness among receivers (Schotter and Sopher 2006), and considerations of fairness being more important for receivers than senders in an ultimatum game (Schotter and Sopher 2007). A common thread in these experiments is that many of the findings would not emerge without the use of an intergenerational experiment.
Our experiment contributes to this literature by thinking about breeding cooperation in a manner similar to Darwin's observations on breeding pigeons. In particular, we aim to direct the cultural evolution of cooperation by selectively exposing future generations to the most positive cooperative advice. Breeding cooperation in this manner has direct implications for real-world games such as collaborative policy, where program designers make choices about who to include in future games, and also select among past examples as models for current collaborations. We next turn to the design and results of our initial experiments.

Experimental Design and Procedures
Undergraduate student subjects in group sizes ranging from 3 to 9 played an N-Person public goods game. Individuals could contribute between zero and 20 tokens to a common fund. Contributions to the common fund were doubled and returned to all members of the group regardless of whether or not they contributed. Subjects began the game with an endowment of between 50 and 150 tokens, although the vast majority of subjects began with an endowment of 100 tokens. (After an initial period of testing to calibrate final payments to average around $15, we settled on an initial endowment of 100 tokens.) Each token was worth $.05 in US dollars. Subjects accumulated or lost tokens over the course of the experiment, and were paid in cash at the end of the game in proportion to the amount of tokens they had accumulated, plus a $5 show-up payment. Subjects earned on average $15 and between $10-20. In the baseline progenitor generation (F0 generation in lab breeding jargon), some groups were allowed to communicate in each round (via computer messages) while other groups had nocommunication. The communication versus no-communication conditions were designed to reproduce the best-established finding in the social dilemma literature that communication increases cooperation (Dawes et al. 1977;Orbell et al. 1988;Sally 1995;Balliet 2009). Each game lasted 20 rounds, although subjects were not told this at the outset of the experiment. We did have some unpaid subjects in these pilot experiments, which we describe in more detail later and control for in the analysis. All experiments were run via a computer network, using the Gameweb browser-enabled software developed by Richard McElreath and colleagues (http://sourceforge.net/projects/gameweb/). All subjects initially read a series of instructions informing them of the payoff structure and details of the game operation. Subjects were informed that interactions in the game and payment amounts would be anonymous. Subjects were provided with an introductory screen in which they could test how different levels of contribution amounts would affect the payoff distributions. Subjects could experiment with this screen for as long as they wanted. We did not explicitly test subject understanding of the game by requiring them to correctly answer a series of questions because we wanted to increase the potential strength of social learning and reduce the strength of individual learning.
In each round, subjects contributed an amount to the common fund, and then were able to view (anonymously) the contributions, per-round, and cumulative payoffs for each individual in the game. After viewing these, subjects were able to send a message to other participants (if they were in a session that included within-game communication, and then read (anonymously) the comments of other players.
To make the game intergenerational, each subject was asked to provide written advice (between generation communication) to the next group of subjects at the end of the game. This advice was then provided to the next generation (successor generation; or F1) of subjects that came into the lab. In particular, we selected the most cooperative advice from the set of F0 groups to provide to the F1 generation. We provide the details for this process below. Unlike the Chaudhuri and Schotter et. al. experiments, we did not pay past generations for the performance of future generations, which provides an incentive for previous generations to leave good advice. Our subjects have no incentive to improve the performance of future generations, although we do select the most cooperative advice as discussed later. This pilot experiment only includes one F1 generation; later experiments will extend the number of generations although the results show that intergenerational effects are strong even in one generation. The F1 generations are also run under two conditions, one with communication and one without communication. Table 1  Each message sent could have multiple phrases and thus might be coded in more than one category. The coding scheme is best conceptualized as quantifying different "bits" of each message. We double-coded a random sample of the advice, and there is an acceptable level of intercoder reliability.
Thus we can analyze the "tone" of the messages in the progenitor and successor generations where communication is occurring. This provides some insight into whether or not the advice from the previous generation shifted the dialog among subjects in ways that go beyond just exhorting higher levels of cooperation.
The content coding also allows us to demonstrate the tone of the advice that we selected to provide to future generations. For example, of the 5 pieces of advice provided to the successor generation in this experiment, 4/5 specifically advocated full cooperation, 2/5 contained positive moral statements, 1/5 contained negative moral sanctions against cooperation, and 3/5 contained group building statements the explicitly mentioned words like "we" and other team-oriented comments (remember that each piece of advice was coded for multiple elements). This was by far the most cooperative set of advice from all the progenitor groups. Note however that the most cooperative advice did not come from the most cooperative progenitor group in terms of cooperation. While cooperative advice and behavior are positively correlated (see below), there is enough variation that they are not the same rank ordering. Also, a group that experiences non-cooperation may provide more cooperative advice in an attempt to help future generations obtain better results. Figure 1 reports the mean contribution by round for each of our conditions and clearly shows the effect of cultural transmission. As expected, communication substantially increases cooperation within the progenitor generation. However, cooperation in the successor generation (even without communication) is higher than cooperation in the progenitor generation with communication. In other words, the advice from previous generations alone transmits cooperative norms in a strong enough way to maintain cooperation when subjects cannot talk to one another within the game. In fact, the effect of advice alone produces nearly as much cooperation as seen in the successor generation with communication. Communication does enhance cooperation a small amount relative to the no-communication successor generation a small amount, but does not have nearly the same magnitude of effect as in the progenitor generation.

Results
The benefits of within-game communication are reduced in the successor generation because subjects are relying on the advice from their predecessors. This is a very similar effect to Schotter and Sopher (2003), where advice locks subjects into one of two possible cooperative equilibriums in a Chicken game. One caveat for this pilot study is the F1-no communication condition consists of entirely unpaid subjects; we control for this in the regression below although the effect of advice alone may be less dramatic if all subjects were paid.
[ Figure 1 about here] the successor generation appears to follow the cooperative advice given by the progenitor generation, with higher levels of exhortations for "full" cooperation (e.g. "Contribute all 10 tokens people"; Figure 3) and higher levels of praise (Figure 4). The increases in cooperative messages are concentrated in early rounds of the games, suggesting that cultural transmission provides later generations a "head start" in the evolution of cooperation. Circumstantial evidence for this conjecture is in Figure 1, where cooperation is actually fairly low in the initial rounds of the F0/communication condition, suggesting that it took some time for subjects to coordinate. Cultural transmission essentially reduces the costs of this early learning period.
Reflecting the disappearance of chronic defectors, the level of disapproval is consistently lower throughout the rounds of the successor generation ( Figure 4).  Table 1 provides a statistical analysis using a linear regression model of contributions and also an ordered logit model where the dependent variable is coded as (0=No contribution; 1=Greater than zero, less than full; 2=Full contribution). The ordered logit is appropriate for the non-normal distribution of contribution behaviors.
The independent variables are 1/Round (the inverse of game length, to capture the gradual decline in cooperation), group size, a dummy variable for unpaid subjects, a dummy variable for communication condition, a dummy variable for successor generation, and an interaction term between successor generation and communication.
The combination of dummy variables means the intercept captures the baseline progenitor condition without communication for paid subjects. All results are statistically significant, and the model explains about 28 percent of the variance in contribution behavior.

Conclusion
Even in this pilot study with some unpaid subjects, the experiments reported here clearly show that it is possible to breed cooperation in the short-term by selectively exposing later generations to advice that recommends cooperation. The effect of intergenerational advice is strong enough to effectively substitute for in-game communication as a promoter of cooperation. Cultural evolution models suggest that vertical and oblique social learning reduces the cost of individual learning, such as that seen when subjects are attempting to coordinate in the progenitor games with communication. Put simply, people perform better when you tell them how to succeed.
Cultural transmission may also be particularly strong because the previous generation is seen as experts with previous experience; they have prestige that should be appreciated regardless of the informational content of their advice. In addition, advice from the previous generation may be viewed as more trustworthy than current generation advice because the previous generation is not competing for the same resources, and has no incentive to send deceptive messages of cooperation and then It would also be interesting to examine the effect of intergenerational advice in the absence of selecting for a particular type (e.g. cooperative) of advice. There is a difference between the content of the advice (e.g.; cooperative versus non-cooperative) and the mode of transmission (e.g.; within-generation versus intergenerational), and there are some reasons to expect that intergenerational advice may have a stronger effect independent of message content.
It is also interesting to think about the possibility of breeding different lineages of behavior, for example it should be possible to breed defection by exposing later generations to the most non-cooperative advice. A particularly interesting possibility in this vein is there may be an asymmetry such that it is easier (or harder) to breed cooperation versus defection. Finally, we need to understand how such intergenerational processes interact with institutional arrangements such as the potential to punish defectors and reward cooperators. Are institutions complements or substitutes for institutional mechanisms like punishment, that are known to increase cooperation (Fehr and Gachter 2000;Fehr and Gachter 2002) (65) individuals Note: First number is paid groups, individuals. In parentheses is total groups, individuals. Number of decisions (unit of analysis in the statistical analysis) is #individuals multiplied by 20 rounds.