|
Field
Experimental Design (Public)
|
Before
People are randomly matched into pairs and assigned one of two roles: Sender and Receiver (Called persons A and B in the experiment). Each player observes two boxes on their screen. One of the boxes is red, the other box is blue. It is common knowledge that each one of these two boxes contains 10 identical cards. Each card gives a certain amount of points to the Sender and a certain amount of points to
the Receiver. Importantly, the Sender can observe what points are written on the cards in the two boxes while the Receiver cannot.
In the instructions, we explain to both players that the two boxes can contain the following cards:
"With a probability of 75%, i.e., on average in 3 of 4 cases,
one of the boxes contains 10 cards that each give
10 points to person A and 10 points to person B,
and the other box contains 10 cards that each give
15 points to person A and 15 points to person B.
With a probability of 25%. i.e., on average in 1 of 4 cases,
one of the boxes contains 10 cards that each give
10 points to person A and 10 points to person B,
and the other box contains 10 cards that each give
20 points to person A and 0 points to person B."
First, person A will be asked to recommend to person B one of the two boxes to pick 10 cards from. They can either recommend the blue or the red box.''
The Sender is asked which one of the two boxes to recommend to the Receiver. The recommendation of the Sender is not binding and is displayed to both players for the remainder of the experiment. The Receiver then has to choose how many cards to pick from which box. They are aware that they will pick 10 cards in total.
TREATMENTS.
We have a between-subject design with two treatments: NOCOMM and COMM.
In both treatments, the Sender recommends one of the boxes to the Receiver. The Receiver observes the recommendation and makes their choice of 10 cards (from all cards from the red box to all cards from the blue box, with any possible integer combination in between).
In COMM, both matched participants additionally enter an open chat *after* the Sender has made the recommendation and *after* the Receiver has made their first (preliminary) choice. The Sender is not informed about the Receiver's choice of cards. The Receiver thus makes their second choice during/after the communication. Both players are informed that the Receiver will have the opportunity to revise their choice. This experiment aims to compare how participants' suspicion changes with and without communication. We include the first choice in COMM to have additional information on what *exact aspects* of communication affect suspicion: the anticipation (i.e., the fact) of communication versus the contents thereof. In COMM, only the second choice is incentivized.
TIMELINE.
1. Participants read general instructions and answer control questions.
2. The Sender recommends one of two boxes.
3. The Receiver observes the Sender's recommendation.
4 The Receiver enters in integers how many cards to pick from which box. This choice is not revealed to the Sender.
5 (COMM). Participants enter an open chat. On the same screen, the Receiver enters in integers how many cards to pick from which box.
6. We elicit beliefs. Belief elicitation for the Sender is "How many cards do you think the other person has drawn from the red (blue) box?" based on whether they recommended a red (blue) box. Belief elicitation for the Receiver is
"Based on your subjective estimation - what points do you think were in the two boxes?" and "Please tell us, how likely you think it is that the other person has recommended the box with cards giving each one of you 10 points per card."
7. Participants fill out the questionnaire.
HYPOTHESES.
We formulate the following hypotheses.
First of all, we need to establish:
Hypothesis 1 (Suspicion). Without communication, Receivers significantly deviate from the Senders' recommendations, i.e., display significant levels of suspicion.
The focus of the paper is the impact of communication on suspicion. Two possible countervailing effects may play a role. On the one hand, communication might be beneficial, i.e., increase Senders' trustworthiness such that suspicion decreases. On the other hand, communication might result in more suspicion by making Senders' potentially misaligned incentives more salient and leading Receivers to think more about Senders' motives. Alternatively, the Sender might communicate in a way that raises suspicion, e.g., by making contradicting statements or by not giving satisfactory answers to questions by the Receiver. We, therefore, formulate a two-sided hypothesis on the difference between suspicion in NOCOMM and suspicion in COMM in the second, final decision of the Receiver.
Hypothesis 2 (Communication effect on suspicion). With communication, Receivers deviate more or less from the Senders' recommendations.
To disentangle these effects, we will use the within-subject differences in the Receivers' level of suspicion before and during/after the chat. Furthermore, we will analyze the communication content. We expect that a high level of suspicion/negativity in the chat will correspond to higher levels of suspicion in terms of the Receivers' choices and vice versa.
PREPARATION.
In preparation for this experiment, we conducted a pretest where in 100% of cases one box contains (10,10) payoffs and the other box contains (12,12) payoffs. The pretest ensures us that the noise generated by subjects' confusion and/or misunderstanding of the instructions is minimal and hence the Receivers' deviations from the Senders' recommendations indeed capture suspicion. We collected 15 independent observations (30 subjects). On average, Receivers deviated by less than one card out of ten. Importantly, 73.3% of Receivers did not deviate at all. We specifically invited newly recruited subjects who never participated in the experiments before to document the upper bound of noise in participants' behavior.
_______________________________________________
Due to the independent data structure, we will rely on two-sided Mann-Whitney U tests and regression analyses to test the hypotheses described above. For the non-parametric tests, we will use a matched pair of subjects as the unit of observation. We will analyze communication content to shed further light on communication effects on suspicion in humans. We plan to use (unsupervised) machine learning methods to analyze chat data. 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.). Furthermore, we plan to use the regression analysis to account for the communication characteristics.
|
After
Update 14.10.2024: Upon completion of the data collection according to the original pre-registration, we add an additional treatment COMM+SIGNAL to deepen the understanding of how the content of communication shapes suspicion. All details regarding the new treatment are integrated below.
We have adjusted the exchange rate for the experiment from previously 16 points = 1 Euro to now 30 points = 1 Euro. This was done to accommodate for the additional earnings in the new stage while ensuring that behaviour would not be influenced by incentive effects.
People are randomly matched into pairs and assigned one of two roles: Sender and Receiver (Called persons A and B in the experiment). Each player observes two boxes on their screen. One of the boxes is red, the other box is blue. It is common knowledge that each one of these two boxes contains 10 identical cards. Each card gives a certain amount of points to the Sender and a certain amount of points to
the Receiver. Importantly, the Sender can observe what points are written on the cards in the two boxes while the Receiver cannot.
In the instructions, we explain to both players that the two boxes can contain the following cards:
"With a probability of 75%, i.e., on average in 3 of 4 cases,
one of the boxes contains 10 cards that each give
10 points to person A and 10 points to person B,
and the other box contains 10 cards that each give
15 points to person A and 15 points to person B.
With a probability of 25%. i.e., on average in 1 of 4 cases,
one of the boxes contains 10 cards that each give
10 points to person A and 10 points to person B,
and the other box contains 10 cards that each give
20 points to person A and 0 points to person B."
First, person A will be asked to recommend to person B one of the two boxes to pick 10 cards from. They can either recommend the blue or the red box.''
The Sender is asked which one of the two boxes to recommend to the Receiver. The recommendation of the Sender is not binding and is displayed to both players for the remainder of the experiment. The Receiver then has to choose how many cards to pick from which box. They are aware that they will pick 10 cards in total.
TREATMENTS.
We have a between-subject design with three treatments: NOCOMM, COMM and COMM+SIGNAL.
In all treatments, the Sender recommends one of the boxes to the Receiver. The Receiver observes the recommendation and makes their choice of 10 cards (from all cards from the red box to all cards from the blue box, with any possible integer combination in between).
In COMM+SIGNAL, there is an additional stage (Dictator stage) before the matched Sender-Receiver pair interacts. In this stage, all Senders enter a version of a Dictator Game that mimics the payoffs of the misaligned case of the Sender-Receiver game with a *different* anonymous Receiver. In the Dictator Game, each Sender chooses between two options: one yielding 100 points for each player and one yielding 200 points for the Sender and 0 points for the Receiver. Importantly, the Sender’s payoff choice in the Dictator stage will be observable to their Receiver (who is not the one the decision was made for) along with their recommendation. Receivers observe their payoffs from the Dictator stage at the end of the experiment.
In COMM and COMM+SIGNAL, both matched participants additionally enter an open chat *after* the Sender has made the recommendation and *after* the Receiver has made their first (preliminary) choice. The Sender is not informed about the Receiver's choice of cards. The Receiver thus makes their second choice during/after the communication. Both players are informed that the Receiver will have the opportunity to revise their choice. This experiment aims to compare how participants' suspicion changes with and without communication. We include the first choice in COMM and COMM+SIGNAL to have additional information on what *exact aspects* of communication affect suspicion: the anticipation (i.e., the fact) of communication versus the contents thereof. In COMM and COMM+SIGNAL, only the second or third choice, respectively, is incentivized.
TIMELINE.
1. Participants read general instructions and answer control questions.
2. (COMM+SIGNAL) The Sender enters the Dictator stage and chooses between (100,100) and (200,0).
3. (COMM+SIGNAL) The pairs are rematched.
4. The Sender recommends one of the boxes.
5. The Receiver observes the Sender's recommendation.
6. The Receiver enters in integers how many cards to pick from which box ("prior deviation"). This choice is not revealed to the Sender.
7. (COMM+SIGNAL) The Receiver observes the Sender’s recommendation and payoff of the dictator stage.
8. (COMM+SIGNAL) The Receiver enters in integers how many cards to pick from which box ("informed deviation"). This choice is not revealed to the Sender.
9. (COMM and COMM+SIGNAL). Participants enter an open chat. On the same screen, the Receiver enters in integers how many cards to pick from which box (“final deviation”).
10. We elicit beliefs. Belief elicitation for the Sender is "How many cards do you think the other person has drawn from the red (blue) box?" based on whether they recommended a red (blue) box. Belief elicitation for the Receiver is "Based on your subjective estimation - what points do you think were in the two boxes?" and "Please tell us, how likely you think it is that the other person has recommended the box with cards giving each one of you 10 points per card."
11. Participants fill out the questionnaire.
HYPOTHESES.
We formulate the following hypotheses.
First of all, we need to establish:
Hypothesis 1 (Suspicion). Without communication, Receivers significantly deviate from the Senders' recommendations, i.e., display significant levels of suspicion.
The focus of the paper is the impact of communication on suspicion. Two possible countervailing effects may play a role. On the one hand, communication might be beneficial, i.e., increase Senders' trustworthiness such that suspicion decreases. On the other hand, communication might result in more suspicion by making Senders' potentially misaligned incentives more salient and leading Receivers to think more about Senders' motives. Alternatively, the Sender might communicate in a way that raises suspicion, e.g., by making contradicting statements or by not giving satisfactory answers to questions by the Receiver. We, therefore, formulate a two-sided hypothesis on the difference between suspicion in NOCOMM and suspicion in COMM in the second, final decision of the Receiver.
Hypothesis 2 (Communication effect on suspicion). With communication, Receivers deviate more or less from the Senders' recommendations.
To disentangle these effects, we will use the within-subject differences in the Receivers' level of suspicion before and during/after the chat. Furthermore, we will analyze the communication content. We expect that a high level of suspicion/negativity in the chat will correspond to higher levels of suspicion in terms of the Receivers' choices and vice versa.
The experimental evidence from the two original treatments (COMM and NOCOMM) reveals that participants deviate significantly less when communication takes place. However, the contents of the chats are quite homogeneous and shed little light on *why* communication is so effective in reducing suspicion. The new treatment COMM+SIGNAL offers an additional signal to Receivers regarding the behavior of their Sender in a similar game where Senders themselves make a decision. We hypothesize that this additional intervention will result in a more heterogeneous communication content that will shed additional light on suspicion channels.
Receivers might interpret their Senders behavior differently. If the Sender chose a selfish outcome in the Dictator stage, they might be more likely to choose a selfish outcome again (i.e., they are selfish in general) or they might be less likely to choose the selfish outcome again because they have already earned a substantial payoff from the Dictator stage. Likewise, if the Sender chose a generous outcome in the Dictator game, they might be more likely to choose a generous outcome again (i.e., they are generous in general) or they might be less likely to choose the generous outcome again because they have not earned a lot in the Dictator stage and used it to “trick” the Receiver to trust them. Hence, we formulate a two-sided hypothesis consisting of two parts:
Hypothesis 3 (Dictator stage effect on suspicion).
(a) With communication and the Dictator stage, Receivers deviate more or less from the Senders' recommendations if their Sender chose a selfish outcome in the Dictator stage.
(b) With communication and the Dictator stage, Receivers deviate more or less from the Senders' recommendations if their Sender chose a generous outcome in the Dictator stage.
PREPARATION.
In preparation for this experiment, we conducted a pretest where in 100% of cases one box contains (10,10) payoffs and the other box contains (12,12) payoffs. The pretest ensures us that the noise generated by subjects' confusion and/or misunderstanding of the instructions is minimal and hence the Receivers' deviations from the Senders' recommendations indeed capture suspicion.
Due to the independent data structure, we will rely on two-sided Mann-Whitney U tests and regression analyses to test the hypotheses described above. For the non-parametric tests, we will use a matched pair of subjects as the unit of observation. We will analyze communication content to shed further light on communication effects on suspicion in humans. We plan to use (unsupervised) machine learning methods to analyze chat data. 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.). Furthermore, we plan to use the regression analysis to account for the communication characteristics.
|