Abstract:

Making a decision as a group of individuals is at the core of any society. Group decision making (GDM) is thus a topic across many research fields. In particular, two questions are crucial to evaluate group decisions: (1) Whether the group performance is better or worse than that of the individuals, and (2) how the individuals’ decisions lead to the group decision. Previous studies have found controversial answers for the first question, indicating that group performance actually depends on the situation. Therefore, to better understand GDM, researchers have looked for the key factors that influence the formation of a group decision. Recently, confidence has been shown to play a pivotal role in this process. Bahrami et al. (2010) proposed a “weighted confidence sharing” (WCS) model to describe the information integration process in GDM. Koriat (2012) investigated the situation when “the more confident member dominates” (MCD) the decision of groups with two members. While explaining the performance of GDM, these studies ignored the dynamic information communication process. How the dynamical interaction between members of the group affects GDM is thus unclear. To explore this question, we designed and carried out a dyadic motion direction discrimination task with a varying communication process. In our three experiments, participants first decide individually in what direction random dots is moving and also report their confidence in a scale from 1 to 6 after making the decision. To study the dynamical process of reaching a consensual decision, we designed the experiments as follows. If the decisions of the two participants in a group are consensual, feedback information on the screen will tell them whether their answers are right or wrong; otherwise, they need to repeat the decision after seeing the identical stimulus again and incorporating information about the behavior of the other participant. In Experiment 1 and 2, each participant is informed about the other’s choice, while in Experiment 3 the other’s confidence is additionally reported. This process is repeated round-by-round until they reach a consensus. The task’s difficulty can be adjusted by varying the coherence level of the dot pattern (the fraction of dots moving towards the same direction) and by varying the number of choice alternatives (two directions for Experiment 1, and four directions for Experiment 2 and 3). By fitting the experimental data using a cumulative Gaussian function, we compared the psychometric sensitivities between individuals, dyad and the WCS and MCD models. Furthermore, we built a model based on Markov process to consider the dynamic change of choice probability due to interaction. We found that in all three experiments, the accuracy of the first-round choice, which was done individually without influence of the other, strongly positively correlates with confidence (Pearson's correlation coefficients approaching 0.99). However, in the following rounds, where the individual decision could be influenced by the other’s choices, the correlation of the accuracy with confidence decreases. This decrease is particularly evident in Experiment 3, where participants can gauge the confidence of each other. We further compared in Experiment 2 and 3 the relationship between the probability of changing one’s choice in the next round and the difference of the individual confidences in the current round. Our results show that the probability of changing the choice positively correlates with confidence difference, and the trend is more prominent for Experiment 3, where the participants can see each other’s confidence. This finding implies that confidence does affect each other’s choice during GDM. Further, the psychometric sensitivities hold the relationship for all three experiments, implicating that neither the WCS nor the MCD model can describe the experimental data integrally. Moreover, SMCD is slightly smaller than SB in Experiment 1 and 2, which is reversed in Experiment 3, indicating confidence’s effect on GDM again. In conclusion, our results show that (1) the decision accuracy is positively correlated with individuals’ confidence; (2) the communication of confidence of the other can influence the tendency to change one’s decision, leading to higher probability to follow the other’s choice given that she/he is more confident; (3) the dyad performance is better than both individuals’ performance and both models’ predictions, indicating that the more confident member does not dominate the group decision and Bayesian integration of shared confidence cannot account for the whole group performance; (4) a Markov model considering the change of choice probability due to dynamic interactions described the experimental data well. However, to better understand the dynamics of GDM, we need to refine the experimental design to extend the interaction rounds in the future.

Key words: dyadic decision making, dynamic interaction, confidence, confidence sharing, virtual decision making