Building social cohesion through virtual intergroup contact: Mobile app experiment in Bangladesh

We will develop an original mobile game application (app) and conduct a randomized experiment in Bangladesh. In the game, players will form a group and collaboratively manage a fruit farm.

In the game, three players will form a group and collaboratively manage a fruit farm. Each group member’s information is limited to only their nationality (Bangladeshi or Rohingya), while all other details remain anonymous. In addition, the game does not include any communication features, preventing players from interacting with other group members.
The game is simple, where the only task is to open the app and harvest the fruits that appear on the tree by tapping the screen. The game interface displays both the total harvest of the group and the individual harvests of each of the three members.
Players engage in this game with their group for five days, which constitutes one round. In total, there are six rounds to be played, with group members being randomly shuffled at the beginning of each round. At the end of each round, players receive rewards based on the total fruits harvested by their group. The maximum group reward per round is set at 900 Bangladesh Taka (BDT), with a total of 5,000 fruits generated in every round. Consequently, the value of each fruit is set at 0.18 BDT. Regardless of individual contributions, the total reward is divided equally among the three members. Actual payments are made in a lump sum after the completion of all six rounds and once all rewards are finalized. The total rewards will be transferred through bKash, an electronic payment system commonly used in Bangladesh.
Before the game begins, participants will be informed that Rohingya players may be included in their team. They will also be told that computer players (bots) may be in the team. In reality, all Rohingya players are bots because of the practical difficulties in invloving Rohingya refugees. However, this information will not be disclosed to the participants. In our pre-test, the majority of participants believed that the Rohingya players were real human players.

2024-10-30

2024-12-31

The primary outcome variables are general attitudes towards refugees, hostility through an incentivized joy-of-destruction experiment, and implicit bias using the list experiment technique.

We plan to recruit up to 2,500 smartphone users in Bangladesh. Participants will be randomly assigned to one of the four treatment arms or the control group prior to the beginning of the game. Each group will consist of 500 participants.

Participants will be randomly assigned to one of the four treatment arms or the control group prior to the beginning of the game. Each group will consist of 500 participants. The randomization is at the player level without any stratification. In the control group, each team is formed with two Bangladeshi human participants and one Bangladeshi bot. The bot's performance level will dynamically reflect the real-time average performance of the Bangladeshi players in this group.
In the first treatment group, participants will continuously engage with a Rohingya bot from the first round through the fifth round. During this period, the performance of the bots will be adjusted to reflect the real-time average performance of Bangladeshi participants in the control group. In the final sixth round, participants will form teams identical to those in the control group, consisting of two Bangladeshi human participants and one Bangladeshi bot.
In the second and third treatment groups, teams will also always include one Rohingya player represented by a bot until the fifth round, but the performance level of the bot will vary. For the second group, the bot’s capability will be enhanced to 150%, simulating virtual contact with a high-performing Rohingya, whereas in the third group, the bot’s performance will be decreased to 50%, presenting an interaction with a lower-performing Rohingya. Consistent with the first treatment group, participants in these groups will team up with a Bangladeshi bot in the final sixth round.
Finally, in the fourth treatment group, we reduce the frequency of virtual contact with Rohingya to one-third compared to the first treatment.

The randomization will be done in the app using random numbers.

The randomization is at the player level without any stratification.

No

No cluster

2,500 smartphone users

Each group will consist of 500 participants.

This section details the power analysis conducted for our game experiment, where the unit of analysis is individual level. In this experiment, we anticipate a 100% take-up rate for all treatment arms, as participants are automatically assigned to one of the treatment groups. It is important to note that failure to access the game during its duration does not count as attrition in this study, as it potentially reflects both attrition and free-riding behavior. Instead, attrition is defined strictly as the failure to respond to the post-game endline survey. In a pre-test conducted in February 2024, the observed attrition rate was 13%. Using this figure as a reference, we conservatively estimated the attrition rate to be 15% for this power analysis. We performed power calculations for selected primary outcomes: the burning amount in the Joy-of-Destruction (JoD) experiment, the social cohesion index, and the index of beliefs about refugees’ abilities. These calculations are based on a hypothesis test with a significance level of 5% and 80% power, comparing two trial arms with 500 participants each in balanced treatment assignment. For the burning amount in the JoD experiment, our analysis builds upon the study by Higuchi et al. (2024), which implemented the JoD experiment against Rohingya refugees in 2020, targeting 1,679 households in Bangladesh. They reported a mean payment of 23.9 BDT and a standard deviation of 28.1 BDT to participants to decrease donations. For the indexes measuring social cohesion and beliefs about refugees’ abilities, we use standardized scales (z-scores) with a mean of 0 and a standard deviation of 1. In our power analysis, the anticipated effect size with a 15% attrition rate is 0.19, which Cohen (1988) categorizes as a small effect. We are powered to detect a MDE of 5.41 BDT for the burning amount in the JoD experiment, which is approximately 23% of the mean reported in previous studies. For the social cohesion index and the index of beliefs about refugees’ abilities, the MDEs are set at 0.19 z-scores. These MDEs are considered reasonable for our experimental design, especially in light of findings from prior studies, such as one indicating that each standard deviation increase in distance from the refugee camp correlated with an increase of up to 7.5 BDT in the amount paid in the JoD experiment (Higuchi et al., 2024).