Attribution Bias: 

The Source of Overconfidence

Economists and Psychologists have documented the prevalence of overconfidence in various settings: entrepreneurs believe their ideas are overconfident in the probability with which their businesses will succeed, and we observe excessive entry into markets (Camerer and Lovallo, 1999); people who are at high risk of having Huntington's disease are overconfident about their health and have sub-optimal healthcare savings (Oster et al., 2013); employees overestimate the quality of their match to an employer and do not look for alternative opportunities often enough (Hoffman and Burks, 2020). These examples show that overconfidence is a behavioral bias that is prevalent and widespread, and it leads people to engage in costly behaviors. I designed and implemented a laboratory experiment where I collected data that allowed me to identify the mechanism through which these incorrect beliefs prevail. 

In the experiment, participants answered trivia questions on six different topics and then participated in a survey. In the survey, they were asked how many questions they thought they answered correctly. In the plots below, we see that 43% of participants guessed their score incorrectly, with 23% underestimating their performance and 20% overestimating it. They were also asked how sure they were about their guess. As the right panel shows, subjects who were initially overconfident expressed significantly more certainty than those who correctly estimated or underestimated their performance.

Initial Specification

The share of subjects who initially underestimated their ability, correctly estimated their ability and who overestimated their ability across all topics.

Initial Certainty

The average level of certainty expressed by participants for each of the topics that were tested and by initial belief about their performance. A certainty level of 100 would correspond to being completely sure, while a certainty of 33 is a random guess.

After the initial trivia and survey, the participants had to make choices to receive noisy information. The task was designed so that the choices would differ depending on what belief updating mechanism was being used. The three mechanisms that I considered were: 

These have been proposed by theoretical research and provide hypotheses that I can test with my data.  The models make predictions regarding two things: the evolution of beliefs and the response in behavior after observing either a success (good news) or a failure (bad news).

Changes in beliefs about own performance in the trivia questions from initial guess to guess after 11 rounds of information

The plot above illustrates the main patterns in beliefs and behavior that are present in the data. In the first plot, we see how beliefs about our performance in the trivia evolved after receiving information. The main takeaways are that initial beliefs are sticky, and overconfidence is the most sticky belief. This means that whatever belief people have at the beginning of the task is the most likely belief that they will have at the end. The fact that 71% of those who started overestimating themselves continue to do so after multiple rounds of information while only 52% remain underconfident indicates that there is a key difference in how people react to information depending on what the initial belief is. This can be related to different things:

We can discard the first theory by looking at the plot on the left. This plot separates participants of the experiment into three groups:

The plot further classifies participants into two types of situations: participants who were at risk of falling into a learning trap and participants who could not have fallen into a learning trap. 

We see that most of the participants who managed to learn correctly were prone to falling into traps but did not. Furthermore, the share of participants whose behavior is consistent with being in a learning trap is only 17%. 

These results strongly indicate that the prevalence of overconfidence is not directly linked to the presence of traps.

On the other hand, when I look at the way in which subjects react to good and bad news, I find a very asymmetric response. The plot on the right shows how the choices that participants made responded to each type of news. In this plot, good news corresponds to information that is surprisingly good, given what the person has seen in the past. This corresponds to a positive news difference. The larger the news difference, the better the news. On the other hand, a negative news difference corresponds to surprisingly bad news.

In this setting, if people were incorporating information in a fully Bayesian manner, we would expect to see a positive slope for both types of news. Surprisingly, good news should lead participants to a higher choice, whereas surprisingly, bad news would have the opposite effect (refer to column 4 in the regression table below)

The observed behavior corresponds to people incorporating the information in a direction consistent with Bayes' rule when they observe bad news but in the opposite direction when they observe good news. In addition, the reaction to bad news is significantly exaggerated relative to what a fully Bayesian agent would have done. This type of behavior is consistent with a belief updating procedure in which bad news is overly blamed on exogenous parameters (circumstances) while good news is attributed to the person's own performance and not the exogenous circumstances. This is what the psychology literature refers to as self-attribution bias and has been modeled in economics through motivated beliefs. 

From the theories that have been proposed by the literature to study the prevalence of overconfidence, I find that self-attribution bias is the only one that can explain the patterns that I observe in the data. 

Attribution bias has been modeled in economics by a modified Bayes rule (Benjamin, 2019). The model is a generalized Bayes rule where the likelihood of observing each type of information is distorted in a self-serving manner. If the outcome is positive (a success), then the rule inflates the likelihood that such an outcome was generated by a high ability and deflates the likelihood of that outcome having been generated by a low ability but good exogenous factors. On the contrary, if the outcome is negative, the rule will do the opposite: inflate the likelihood of a failure due to bad exogenous factors and deflate the probability of the outcome coming from a low ability. 

In the academic paper, I go over the exact model and estimate the structural parameters to gain some insight into potential heterogeneity in the mechanisms being used by different participants. I find that biased updating is overwhelmingly the most used mechanism, but there is some evidence of dogmatism and Bayesian hypothesis testing as well. 

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

(raw data available upon request)