Social decision making

Reasoning beyond available data is a ubiquitous feature of human cognition. But while the availability of first-hand data typically diminishes with increasing complexity of reasoning topics, people’s ability to draw inferences seems not to. Reasoners may offset the sparsity of direct evidence with evidence that is inferred by observing the statements and actions of others. But this kind of social meta-inference comes with challenges of its own. In evaluating a claim about an unfamiliar topic, a reasoner might sensibly assume that a person who makes an argument in its favour is in possession of some evidence - but how much? How should the evaluation vary with the number of people arguing on each side? Should repeated arguments carry more weight than distinct ones? How people reason in this situation is likely to depend on their assumptions about the generative process behind communication. Here we present preliminary work towards a computational model of the kinds of inferences required when reasoning from indirect evidence, and we examine candidate model predictions via an experiment investigating the evidentiary strength of consensus in the context of social media posts. By systematically varying the degree of consensus along with the diversity of people and arguments involved we are able to assess the contribution of each factor to evidentiary weight. Across a range of topics where reasoning from first-hand data is more or less difficult we find that while people were influenced by the number of people on each side of an argument, the number of posts was the dominant factor in determining how people updated their beliefs. However, in contrast to well established premise diversity effects, our findings suggest that repeated arguments may carry more weight.

We model large sets of interacting mobile agents whose movement choices are determined in a discrete-choice random-utility framework spanning simple multinomial logit models to crossed-nested logit models that account for velocity-related correlations. The agents are predictive, so their choice utility is in part based on projecting the future positions of other agents they observe. They can have diverse characteristics and individual movement plans consisting of goals about visiting sets of locations. When a plan is disrupted through interactions with other agents in crowded scenarios, they can dynamically create sub-goals to enable them to return to complete their mission. Additive combinations of choice utilities provide a method to combine, weight, and resolve sets competing demands from goals (e.g., heading to the next location), individual preferences (e.g., for speed and inter-personal distance), rules (e.g., passing on the right) and social factors (e.g., following a leader and grouping). We report simulations showing that these agents can competently navigate and achieve their goals in difficult environments and results on Bayesian estimation of agent parameters from movement data. We discuss the potential for this framework to build, parametrize, explore, and predict systems of agents guided by complex and flexibly specified cognitive states.

Polarization is often described as the product of biased information search, motivated reasoning, or other psychological biases. However, polarization and extremism can still occur in the absence of any bias or irrational thinking. In this talk, we show that polarization occurs among groups of decision makers who are implementing rational choice strategies (specifically, random walk / relative evidence choice strategies) that maximize decision efficiency. This occurs because extreme information enables decision makers to make up their minds and stop considering new information, whereas moderate information is unlikely to trigger a decision and is thus under-represented in the sampled information. Furthermore, groups of decision makers will generate extremists – individuals who stop sampling after examining extreme information. In re-analyses of seven empirical studies spanning perceptual and preferential choice, a series of simulations manipulating threshold, bias, and drift rates, and a new study examining politically and affectively charged decisions, we show that both polarization and extremism manifest when decision makers gather information to make a choice (choice task). Polarization did not occur, however, when participants made an inference about the difference between two quantities (estimation task). Estimation therefore offers a theoretically-motivated intervention that can increase the amount of information people consider and reduce the degree of polarization and extremism among groups of individuals.

Improving the accuracy of medical image interpretation is critical to improving the diagnosis of many diseases. Research in human decision-making has shown that a Wisdom of the Inner Crowd approach can improve the accuracy of individual decision-makers. In this approach, repeated judgments from the same decision maker on the same stimuli are aggregated. Since repeated decisions in medical contexts are time intensive and potentially costly, we test whether it is possible to aggregate decisions on not necessarily the same but similar images. In a series of experiments, we use the classification decisions (cancerous vs non-cancerous) collected with novice and expert participants on a set of white blood cell images. To determine the similarity between cell images, we use the latent representations of the images from neural network models. We investigate two different representations, distinguished by how the neural networks were trained. The first representation was obtained by training a neural network on the cancer classification task. The second representation was obtained by training the neural network on an unrelated task (i.e., categorizing natural images, but not cell images). We observe that these methods work better for novices than experts. This suggests that novices and experts have different decision mechanisms, where the novices make random errors while experts are systematically biased. Finally, using a better representation not only allows for larger improvements in accuracy but also allows for aggregation over more images.

Although most previous studies indicated that aggregating group-level decisions is superior to individual decisions, some studies argued that collaboration does not always result in better performance. It is still unclear how task context may influence the group decision advantage. To examine the effect of task rule and task difficulty on the collective decision-making process, we applied Systems Factorial Technology to measure group decision-making efficiency in three visual search experiments (i.e., a T/L conjunction search task): In both Experiments 1 and 3, participants had to report the number of the targets (i.e., AND search rule), and trials including uncertain target numbers were used as catch trials in Experiment 3 to prevent early search termination; In Experiment 2, participants had to detect the presence of any target (i.e., OR search rule). The results revealed supercapacity processing under both task rules by comparing the group to individual subject’s performance, suggesting a collective benefit. Most interestingly, the degree of how the collective benefit is affected by the task difficulty depends on the task rule. With an OR rule, collective benefit was unaffected by the number of distractors whereas with an AND rule, collective benefit increased as the task difficulty increased. To conclude, our research suggested that group decisions can outperform individual decisions by showing more efficient processing; and the efficiency effect is prominent with difficult tasks and exhaustive searching rule conditions, respectively.