Information processing plays an important role in many life decisions, but the same information may not always be interpreted the same way. For example, despite scientific consensus on climate change, individuals may have di↵erent interpretations of the likely impact that persist for long periods of time (Kahan et al., 2011, 2012;
Fryer Jr et al., 2019). Receiving a bad grade in math may lead some to pursue non-STEM degrees, while others may see it as a challenge to persist. Some may even believe the earth is flat despite apparent evidence. Indeed, personal experience and motivation may play a role in individual interpretation.
When information could be potentially irrelevant, individuals may be overly inclined to discount it entirely when it contradicts their prior beliefs. For example, if one believed there was a chance new research was driven by political or commercial interests, it may be overly easy to discount this as untrustworthy or irrelevant. In contrast, when this information conforms to our pre-existing beliefs, it may be much harder to account for this possibility. Since Bayes’ Rule implies we should still update our posteriors according to the new information in both cases, ignoring such information because it may be irrelevant results in incorrect beliefs persisting for much longer.
In this paper, we examine a laboratory experiment investigating people’s ability to process new information, and study the di↵erence in belief updating when infor-mation aligns with or is against the prior belief. Specifically, we consider a two step procedure, in which subjects first independently draw information from one of two urns, and use this information to update beliefs about the state of that urn. Then,
each subject learns the stated belief of another randomly chosen subject for the urn that he or she drew from. However, the second subject’s urn may or may not be the same as the first subject.
In the face of conflicting information, a subject should correctly infer that the other subject is more likely to have drawn from a di↵erent urn. However, this does not mean that the other subject could not have drawn from the same urn, just that it is less likely. However, even in a neutral context, we document that subjects appear to overestimate the probability that conflicting information comes from a di↵erent (and hence, irrelevant) urn. Thus, in our neutral setting subjects underestimate the chance that conflicting information implies that their own initial information was misleading. Conversely, in the face of confirming information, subjects overestimate the probability that this information comes from the same urn, and underestimate the chance that seemingly confirming information may be irrelevant (coming from a di↵erent urn).
Fryer Jr et al. (2019) introduce a model to depict why polarization in people’s beliefs would occur in many settings where information is open to interpretation.
An important theoretical prediction from this paper is that polarization increases when people interpret an ambiguous signal as a certain signal for a particular state based on their current beliefs. Their online Amazon Mechanical Turk experiments show that when subjects observe a sequence of information, they indeed form biased interpretation of evidence in the face of ambiguous ones and results in polarization in issues like climate change and death penalty.
In contrast, we provide three main contributions. First, the polarizing beliefs in
Fryer Jr et al.(2019) stem from ambiguous information that is incorrectly inferred to be informative. In our paper, we explore how individuals incorporate information that is contradictory to their current beliefs, rather than how they misinterpret non-informative signals. Thus, even in purely informative spaces, we show improper Bayesian updating. Secondly, in our experiment, we explore a politically neutral context with objective outcomes, as opposed to the politically charged context with subjective outcomes (as the scale of interpretation in Fryer Jr et al. (2019) may be itself di↵erently interpreted based on prior beliefs). Lastly, we provide evidence that contradictory information is not misinterpreted as consistent with beliefs, but is instead tends to be misattributed as irrelevant. Collectively, these results can potentially explain why, despite the general scientific consensus on climate change, individuals may form beliefs that cause them to ignore this information. In other words, given the relative paucity of ambiguous information in climate change, it may be that individuals instead infer that unambiguous contradictory information is instead from an untrustworthy or irrelevant source.
Failure of Bayesian updating is documented in several papers, includingTversky and Kahneman(1973) andGrether(1980). They find that subjects ignore base-rate information contrary to Bayes rule, resulting in representativeness heuristic. Holt and Smith(2009) show that subjects tend to over/underweight low/high prior prob-abilities. Compare to cognitive incompetence to perform Bayes rule, recent studies focus on asymmetrically processing information. Eil and Rao (2011) investigate how subjects update di↵erently between neutral and ego-relevant information like beauty and intelligence. They find that subjects respond less when the information
is bad (suggesting one’s beauty or intelligence is inferior) and this e↵ect only occurs in non-neutral settings. However, Coutts (2019) do not find the “good news-bad news e↵ect” in their experiments. How people process ego relevant information is still debating.
Besides, our paper is an extension of rich literature of confirmation bias, which was documented in economics (Babcock et al., 1995) and psychology (Lord et al., 1979). Confirmation bias describes people’s tendency to interpret the information in a fashion that is biased toward confirming one’s prior belief. Glaeser and Sunstein (2013) introduces a model to show how balanced information leads to polarization.
They suggest that the same information have diametrically opposite e↵ect for those who have confirming and conflicting priors. Our experiment provides experimental evidence and illustrates a possible mechanism of this phenomenon.