In a laboratory experiment where I collected data from undergraduate students, I studied the forces behind belief formation about others. I find that when participants initially overestimate the ability of others, they tend to ignore or under-weigh negative signals about that person's performance. This allows incorrect beliefs such as "men are good at math" to prevail. Instead, when subjects initially underestimate the ability of others, they over-weigh positive signals. This is consistent with a belief updating procedure as the following: "I believe women are not good at science in general; however, if a woman is good at it, she must very good since she is an outlier." 

In the experiment, I measure people's abilities by their performance in 6 short questionnaires on different topics.  I then show participants the characteristics of someone who participated in a previous session and ask them to guess the scores for that person in each of the topics. The plots below show the beliefs for each group of people. The plot on the left shows how people perceive their own abilities. We see that :

The panel on the right shows how those same groups are perceived by others. Interestingly,  the group that underestimated itself the most (American females) is overestimated by others the most. While underestimation is the most common for non-American females. 

The figure on the right disaggregates the same data by topic and provides further insight into the role of stereotypes in these beliefs. The experiment was designed so that there would be female-typed topics (culture and Verbal Reasoning) as well as male-typed topics (Sports and Video Games, Science and Technology, Math) and neutral topics (US Geography).

In this plot, a positive score means that, on average, people tend to overestimate the ability of that group for that topic. While a negative score means the opposite, that group is underestimated in the corresponding topic. Non-American women are perceived to have the lowest scores in all topics except pop culture. While American men receive the highest score overall. 

These reported beliefs conform to the prevalent stereotypes.

In this project, I study how these initial beliefs evolve after multiple rounds of endogenous feedback. Participants in the experiment need to make choices that affect the information that they receive. This information is informative about the true ability of the other participant and also of an exogenous variable. The participant has to disentangle the two forces to make optimal choices. However, if they are very certain of their initial belief and if such belief is incorrect, the data allows them to interpret it in a way that confirms their belief, thus becoming even more certain of it.  The table below illustrates the changes in beliefs after 10 rounds of information acquisition.

The figure on the left is a transition matrix for the beliefs about other participants. The first row of the matrix contains all of the people who initially underestimated the other; The second row represents all participants who initially correctly estimated the other person's ability, and the third row is everyone who initially overestimated it. The columns correspond to what their belief was at the end of the experiment. Again, each column corresponds to under, correct, and over-estimating the ability.

The data shows that overestimation is incredibly persistent: 70% of those who initially overestimated the other participant continue to do so even after receiving abundant information. 

On the other hand, underestimation is not nearly as persistent. Only 41% of those who initially underestimated the other continued to do so by the end. However, these participants present a different interesting pattern. For those who initially underestimated the other and realize that they have done so, a lot of them tend to over-compensate for the initial bias and, as a consequence, end up overestimating the ability of others. This sort of behavior is consistent with what has been found in field experiments as well (see Bohren, Imas and Rosenberg, 2019).   

Interestingly, a similar over-correction is not present when the initial belief overestimates the ability of others. These two facts seem to suggest that the participants of the study are overly generous in how they interpret data about others. Whether this is a pattern that is present outside of the laboratory setting is not clear. In fact, this particular behavior could be due to experimenter demand effects, where subjects know that this data will be analyzed and over-corrected to give the experimenter a different impression of who they are. In this case, to show that they are not biased against a particular group (mostly non-american females).

Because this was an incentivized experiment, not learning the true score of the other participant implied making sub-optimal choices. Which in turn translated into lost earnings. The plot below shows the share of optimal choices made by three different groups: those who managed to learn the truth, those who found a way to sustain their initial belief (these subjects are said to be in a learning trap), and those who do not belong to either of those categories.

We see that those who managed to learn the truth are making choices that maximize their earnings more than 90% of the time by the end of the experiment. While those who did not learn to make optimal choices less than half the time. This amounts to a loss of around 25%. The incentive in this experiment was quite low, and a 25% loss would be around 2 USD. It would be interesting to study how learning responds to higher incentives. Perhaps when stakes are high, people are less attached to their initial beliefs and evaluate information in a less biased manner. 

Find out more in the paper and find the code to generate these plots in my GitHub repository. This data was collected together with my data on attribution bias and overconfidence, and the repository contains the analysis for both.

(raw data available upon request)