Errors in probabilistic reasoning and judgment biases

Abstract

Errors in probabilistic reasoning have been the focus of much psychology research and are among the original topics of modern behavioral economics. This chapter reviews theory and evidence on this topic, with the goal of facilitating more systematic study of belief biases and their integration into economics. The chapter discusses biases in beliefs about random processes, biases in belief updating, the representativeness heuristic as a possible unifying theory, and interactions between biased belief updating and other features of the updating situation. Throughout, I aim to convey how much evidence there is for (and against) each putative bias, and I highlight when and how different biases may be related to each other. The chapter ends by drawing general lessons for when people update too much or too little, reflecting on modeling challenges, pointing to areas of economics to which the biases are relevant, and highlighting some possible directions for future work.

A theory of learning to infer

“…This is analogous to our study in which subjects "attend" more to a source that was more informative in the experienced query distribution, leading to a stronger reaction to that source in future queries. However, studies reported in Benjamin (2018) find greater under-reaction with increasing diagnosticity (Figure 3). We note that these studies predominantly used within-subjects designs, 8 in which the same subject has to make inferences across all levels of diagnosticity.…”

“…While much of the work on under-reaction to the prior discussed above was largely driven by findings in more 'realistic' scenarios, such effects are also found in more laboratory-controlled paradigms like those in C. Miller (1964) andEdwards (1968). In particular, when the parameters of the model in Equation 5 are fit to behavioral data from studies using such laboratory-controlled stimuli, the value of α P is generally between 0 and 1 -indicating that subjects sometimes under-weight the prior in these cases as well, but do not neglect it completely (Benjamin, 2018). This formulation therefore allows for the case where both α P and α L are less than 1, corresponding to a version of the "system neglect" hypothesis proposed by Massey and Wu (2005): both the likelihood and prior are neglected, producing an overall insensitivity to variations in the data-generating process.…”

“…Several hypotheses about the origin of under-reaction were put forth (for a comprehensive review, see Benjamin, 2018). One hypothesis held that subjects compute Bayes' rule correctly, but had an inaccurate understanding of the underlying sampling distributions.…”

Choosing and learning: outcome valence differentially affects learning from free versus forced choices

“…While standard accounts of belief (or value) prescribe that an agent should learn equally well from positive and negative information (e.g., Barto and Sutton, 1998;Benjamin, 2018 for a review), previous studies have consistently shown that people exhibit valence-induced biases (e.g., Lefebvre et al, 2017;Kuzmanovic et al, 2018). These biases are traditionally exemplified in the so-called "good news/bad news" effect whereby people tend to overweight good news relative to bad news when updating self-relevant beliefs (Sharot and Garrett, 2016).…”

Just a Big Misunderstanding? Bias and Bayesian Affective Polarization

“…I model inter-personal feelings by assuming they are based on cognitive beliefs, consistent with recent work from psychology and neuroscience. 4 I seek to complement other theories of hostility, such as "mindless" emotional reaction, social group identity, and motivated reasoning, by showing how hostility can also occur and grow due to Bayesian inference with biased priors ("quasi-Bayesian" inference; Benjamin, 2018). 5 These biases are relevant to a wide range of bilateral relationships, including those that do not involve social groups, and thus provide an explanation for the general tendency toward undue conflict in such relationships.…”