时间：2021年10月27日13:30

地点：腾讯会议室（ID：586219144）

主题：A Computational Account for Biases in Correlation Judgments

主讲：何黎胜

主办：学术期刊编辑部

主讲人简介：

上海大学悉尼工商学院副教授，英国University of Warwick管理学博士、美国University of Pennsylvania博士后。主要研究方向为决策心理学与行为经济学，长期运用实验方法和计算建模相结合，探究人的管理与经济决策行为。主要讲授课程有决策行为与心理、人类学习行为，以及风险决策、跨期决策、社会决策的心理与认知神经机制等。

在Management Science、Psychological Review、Cognitive Psychology、Decision、《应用心理学》等国内外权威学术期刊发表论文多篇，兼任Psychological Review、Decision、Journal of Experimental Psychology:General等国际学术期刊审稿人。

讲座简介：Correlation judgments are at the core of belief formation and economic decisions. Existing laboratory and field studies have repeatedly suggested that decision makers show behavioral biases in correlation judgments made using scatterplots (2D graphs that display the co-occurrence of two signals). Prominently, people tend to underestimate the correlations between two signals in scatterplots and this underestimation bias is stronger when the scatterplot is displayed in a landscape view than in a portrait view. Yet, little is known about how the biases arise. Here, we propose that decision makers are Bayesian learners who perform “mental regressions” that take as input the attended points (i.e. pairs of x and y coordinates) and update the slope of the regression line as output. Accordingly, cognitive biases can arise from suboptimal visual information search in the scatterplot. We tested the model predictions with an eye-tracking experiment (N=103). The experiment involved two within-subject treatments. The Landscape treatment displayed the scatterplot in a 1200px*800px view while the Portrait treatment displayed the scatterplot in an 800px*1200px view. Each treatment involved 20 randomly generated scatterplots. We found that the Bayesian learning model that only took the attended points as inputs successfully replicated the classic behavioral finding that decision makers underestimate the correlations and that the underestimation bias is stronger in the Landscape treatment than in the Portrait treatment. The model also predicted trial-level estimation biases at a high accuracy rate (R²=0.34). We further show that it is the under-sampling of peripheral regions that leads to the underestimation bias, as well to systematic differences between the Landscape and Portrait treatments. Our study showcases how eye-tracking data, when combined with computational models, can inform cognitive mechanisms in behavioral and experimental research.