Experimental Design
Subjects who are randomly matched in pairs play a one-shot Trust Game. Each of them receives an endowment of 10 points.
The task is as follows. Participant A can decide how many of these 10 points they want to share with Participant B. Participant B then receives triple the number of points that Participant A shared. Participant B can then decide to share their total amount of points (their endowment plus the share from Participant A) equally with Participant A, or if they want to keep all points to themselves. Importantly, sharing equally means that Participant B splits points such that both participants receive the same amount of points.
We use the strategy method to elicit subjects’ decisions. First, we ask subjects to pick the amount they would share, given they are Participant A. Second, they decide for each amount they could have received from Participant A if they shared the points equally or kept all the points to themselves, given they are Participant B. Finally, a random mechanism decides whether they are Participant A or B with a 50% chance.
In our experimental treatments, we vary two aspects: subjects' ability to communicate with each other within pairs and group assignments.
In total, we have six treatments:
BASE/Comm – no groups, communication;
BASE/NComm – no groups, no communication;
NM/Comm – groups without meaning, communication;
NM/NComm – groups without meaning, no communication;
M/Comm – groups with meaning, communication;
M/NComm – groups with meaning, no communication.
In treatments with communication (marked Comm), participants enter a chat where they can write to their matched partner for 180 seconds before they make their decisions. In treatments without communication (marked NComm), subjects cannot communicate before making a decision.
We introduce three types of group assignments described below: BASE, NM, and M. Subjects can be members of groups C or L, and they observe their own group membership and the group membership of their Trust Game partner. In a baseline treatment (BASE), subjects there are no groups. In treatments NM, subjects choose to join group C or L. Importantly, in NM, we explicitly tell subjects that groups C and L have no further meaning. In treatments M, we introduce meaning to the group assignment: We explain that C stands for Pro-Choice and L for Pro-Life. The screen says: “Please indicate that you would like to belong to group C if you identify as Pro-Choice. According to the Cambridge Dictionary, pro-choice (adj.) means supporting the belief that a pregnant woman should have the freedom to choose an abortion if she does not want to have a baby”. Analogously, the screen says: “Please indicate that you would like to belong to group L if you identify as Pro-Life. According to the Cambridge Dictionary, pro-life (adj.) means opposed to the belief that a pregnant woman should have the freedom to choose an abortion if she does not want to have a baby”.
The experiment starts off with general instructions and explanations of the Trust Game. This identification number will be later displayed to their matched partner as well. In treatments M, we explain what the group membership in C and L stands for. Then, we elicit the binary choice for the group membership and the intensity of the preference for the group membership in treatments M and NM. Notably, subjects in M make an informed choice, while those in NM simply choose between two letters with no further meaning.
In the next step, we match subjects in pairs. Our study focuses on the out-group effect. Hence, in treatments with groups, we always match an individual from group L and an individual from group C. We add the disclaimer to the description of our study: “Please note that, in this study, you interact with people who may have different opinions than you. The chance that this happens is higher than in reality”. Subjects observe their partner's group membership. Matched subjects always belong to the same treatment. It means that there is no information asymmetry between them. For example, either both subjects in the pair know the meaning behind the group membership, or neither knows.
In treatments M/Comm, NM/Comm, and BASE/Comm, subjects communicate for 180 seconds. To them, we provide additional instructions about the communication stage. Their communication is open, i.e., they can type their messages and send them to their matched partner. There is no predetermined structure of their communication. Subjects cannot skip the communication stage. They also cannot shorten or extend the communication time. There are no restrictions on topics subjects can discuss. However, participants are instructed not to share personal information such as names, phone numbers, etc.
After subjects have made their decisions and before they receive feedback, we elicit their beliefs about their partner's decisions. We incentivize belief elicitation using the binarized scoring rule. In addition, we ask subjects whether and why they associate the reported perception of their partner's decision with their group membership. This part of the belief elicitation is not incentivized.
In the post-experimental survey, we collect data on subjects' socioeconomic characteristics, including age, gender, education level, and fluency in English. Moreover, we use validated survey questions to elicit positive and negative reciprocity and risk aversion, loss aversion, and social image concerns. We elicit the Pro-Life/Pro-Choice views for subjects in treatments BASE and NM. After the questionnaire is complete, subjects observe the outcome of the Trust Game and the respective payment, as well as the payments for belief elicitations and the incentivized items from the post-experimental survey.
HYPOTHESES.
For Hypotheses below, we refer to the “out-group effects”. We assume that out-group effects are absent in treatments BASE (by definition, i.e., no groups) and present in treatments NM and M, with out-group effects being stronger in M than in NM due to the polarizing nature of group membership in M. In the paper, in addition to providing the intuition, we formalize the hypotheses theoretically.
Hypothesis 1 (Shared and shared back amounts). The amounts shared and shared back in the Trust Game decrease if the out-group effects become stronger.
We hypothesize that the amount sent as Participant A is, on average, larger in BASE than in NM, and larger in NM than in M. With respect to the amount shared back, for a strategy method setup, it implies that the minimal amount received to share back as Participant B in smaller in BASE than in NM, and smaller in NM than in M.
Hypothesis 2 (Communication quality). Communication becomes *unfriendlier* if the out-group effects become stronger.
We hypothesize that individuals are friendlier to each other in BASE than in NM, and friendlier to each other in NM than in M.
To analyze the content of communication, we use natural language processing. Natural language processing is a common approach to quantifying text data. We call communication friendlier if it scores higher on the emotional scale. Before we analyze the content of the communication, we process the communication data or corpus to be expressive: correcting spelling mistakes, eliminating words that have no meaningful content, i.e. stopwords, and reducing words to their linguistic root. The corpus equals a matrix where each row represents a chat or document, and each column represents a token. Tokens can be words in the documents, or, e.g., a number or a letter. We then analyze the friendliness of a chat. A chat is quantified as friendlier the higher the tokens score on the emotional scale.
Hypothesis 3 (Communication impact). The positive effect of communication on sharing and sharing back in the Trust Game becomes smaller if the out-group effects become stronger.
We hypothesize that the effect of communication on shared amounts in the Trist Game is larger in BASE than in NM, and larger in NM than in M. Likewise, we hypothesize that the effect of communication on shared amounts in the Trust Game is larger in BASE than in NM, and larger in NM than in M. Our paper focuses on increasing differences between subjects and will potentially lead to higher degrees of hostility in certain environments, for example, if subjects belong to different groups in treatment M/Comm.
Due to the directional hypotheses, we will rely on one-sided Mann-Whitne U tests and regression analyses to test the hypotheses described above. For the non-parametric tests, we will use individual subjects as the unit of observation. We plan to use the regression analysis to account for the interdependence of individual observations after communication. Furthermore, we will analyze communication content to shed further light on communication effects. We plan to use machine learning methods to analyze chat data in an attempt to discover not only whether but *how* communication affects these outcomes (see Hypothesis 2). The specific computation method depends on the characteristics of actual data (e.g., length of the messages, number of topics within the messages, variance across the messages, etc.). We plan to cluster data based on the communication content in treatment M/Comm and to analyze heterogeneous treatment effects. Our pre-registered sample size accounts for clustering.