Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Overcondence Marie-Pierre Dargnies January 7, 2015
Marie-Pierre Dargnies
Overcondence
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
What is overcondence? In the standard framework, the economic agent is supposed to be fully rationnal and aware of his performance level and of the accuracy of his knowledge Overcondence can mean dierent things
Self-assessment are often overly optimistic: condence measured by dierence between self-assessment of performance and actual performance People are wrong when they are 100% sure they are right (overcondence about accuracy= calibration): measured with condence intervals
Marie-Pierre Dargnies
Overcondence
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Are people really overcondent? (Benoit and Dubra 2011)
Statements like "90% of people think they drive better than the median" Of course, only 50% of people can be better than the median But, does it necessarily imply overcondence? Is it true that only fty percent of people can rationally believe themselves better than the median?
Marie-Pierre Dargnies
Overcondence
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Are people really overcondent? (Benoit and Dubra 2011) (2) Let's consider drivers can be of three skill levels (good, medium, bad) with equal probability Bad drivers have an 80% chance of causing an accident, medium drivers a 40% chance and good drivers never cause accidents Initially, no driver knows his or her own skill level, and so each person (rationally) evaluates himself as no better or worse than average In period 1, everyone drives and learns something about his skill, based upon whether or not he has caused an accident Each person is then asked how his driving skill compares to the rest of the population. How does a driver who has not caused an accident reply? Marie-Pierre Dargnies
Overcondence
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Are people really overcondent? (Benoit and Dubra 2011) (3) Bayes' rule 1/3∗1/5 1 1/3∗1/5+1/3∗3/5+1/3 = 9 (Medium skill / No accident) = 1/3∗1/51+/31∗/33/∗53/5+1/3 = 31 (High skill / No accident) = 1/3∗1/5+11//33∗3/5+1/3 = 95 Such a driver thinks there is a 59 chance that his skill level
P (Low skill / No accident) = P P
the top third of all drivers Since 53 of drivers have not had an accident, themselves as better than average Marie-Pierre Dargnies
Overcondence
3 5
is in
rationally rank
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Are people really overcondent? (Benoit and Dubra 2011) (4)
Does not mean overcondence does not exist But one should be cautious when nding "evidence" of it... There are nevertheless proper ways to measure overcondence
Marie-Pierre Dargnies
Overcondence
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Calibration Ability to assign subjective probabilities to outcomes that correspond to the objective frequency of those outcomes Applies to (large) sets of judgments, a single judgment e.g "There is a 90% chance that Spain will win the next world cup", can't be assessed for calibration Calibration: Average condence-% Correct Zero=> perfect calibration Positive numbers=> overcondence (you know less than you think) Negative numbers=> undercondence Good calibration means that beliefs match objective frequencies Marie-Pierre Dargnies
Overcondence
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Calibration For 100 predictions made with 90% of condence, (about) 90 should be correct Usually: for all questions the proportion of correct answers is lower than the assigned probability People are often asked to state a 90 percent condence interval for several uncertain quantities. On average, 90% subjective condence intervals only contain the correct answer between, say, 30% and 50% of the time (Glaser and Weber, 2004) Glaser et al nd an even higher level of miscalibration for professionnal traders more than 10% of correct answers belong to the 10% intervals provided, showing that subjects exhibit undercondence in their knowledge at the 10% level (hard-easy eect?) Marie-Pierre Dargnies
Overcondence
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Discrimination
Also applies to (large) sets of judgments, a single judgment e.g "There is a 90% chance that Spain will win the next world cup", can't be assessed for discrimination Discrimination: Average condence with correct itemsAverage condence with incorrect items Zero=> no discrimination (you have no idea whether your answers are correct or incorrect) The higher the number, the better the discrimination
Marie-Pierre Dargnies
Overcondence
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Why is miscalibration relevant to nance? In a nancial market context, Benos (1998) and Odean (1998) show theoretically that miscalibration leads to poor performance Biais, Hilton, Mazurier, Pouget (2005) and Bonnefon et al : Miscalibration is found to reduce trading performance on experimental markets Barber and Odean (2001) nd that men (who are known to be more overcondent than women) trade approximately 45 percent more than women, and earn lower portfolio returns as a result of more frequent trades
Marie-Pierre Dargnies
Overcondence
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Overcondence in performance Measured as the dierence between one beliefs about his performance and one's actual performance Example: Guess your rank for a task previously solved by yourself and a group of people The guessed rank is then compared to the actual rank Men and women are typically found to be overcondent, men more so than women (Niederle and Vesterlund 2007,...) Important consequences to the gender dierences in condence: Partly explains why women shy away from competitive environment (hence, glass ceiling eect, gender wage gap) Marie-Pierre Dargnies
Overcondence
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Application to nance: Malmendier and Tate (2005a, 2005b) Overcondence measured in two ways:
CEO's beliefs about the future performance of their company infered from their personal portfolio transactions (does the CEO bet his own money on company's performance) CEOs classied as overcondent based on their portrayal in the press
Overcondent managers overestimate the returns to their investment projects and view external funds as unduly costly overinvest when they have abundant internal funds curtail investment when they require external nancing Marie-Pierre Dargnies
Overcondence
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Gender dierences in beliefs
Women & men react similarly to information, but are fed dierent signals (stereotypes...) started with dierent priors
Women & men receive the same information but process it dierently
Marie-Pierre Dargnies
Overcondence
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Managing self-condence (Mobius, Niederle, Niehaus and Rosenblat (2012)) How do subjects process information? Does information processing dier across gender? Task subjects care about: IQ test Track subjects' beliefs about belonging to the top 50% performance Repeatedly provide noisy feedback as a simple binary signal whether subject's performance is among the top 50 percent (correct with 75% chance) Allows to calculate a clean bayesian benchmark and to compare how dierent subjects with the same prior react to the same sequence of signals Marie-Pierre Dargnies
Overcondence
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Managing self-condence (2)
After a 3 min IQ test : Chance that your performance is in the top 50%? Solicit belief (x) between 0 and 100, incentives are given through a Becker-DeGroot-Marschak-like procedure: The computer randomly picks a number y between 0 and 100 If x > y , the subject earns 5$ if her score is in the top 50% If x ≤ y , the subject earns 5$ with y% probability
Marie-Pierre Dargnies
Overcondence
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Managing self-condence: Feedback
After eliciting beliefs to be in top 50% Feedback: Yes, No: correct with 75% chance (50% chance of correct feedback, 50% chance random feedback) Re-elicit beliefs Feedback 2: Re-elicit beliefs... We can compare updating to Bayesian benchmark, due to the binary state: Top 50% or not
Marie-Pierre Dargnies
Overcondence
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Managing self-condence: Results
Conservatism: Subjects update less than Bayesians in response to both negative and positive information Asymmetry: Subjects adjust more to positive than negative information Gender dierences: Men are less conservative than women, no signicant dierence in asymmetry These biases are substantially less pronounced in a placebo experiment where ego is not at stake
Marie-Pierre Dargnies
Overcondence
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Gender and competition
Gender dierences in economic success Glass ceiling Gender wage gap...
Theories so far:
Discrimination Dierences in preferences, in ability
Marie-Pierre Dargnies
Overcondence
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Gender and competition
Women and men may dier in their propensity to compete, select into competitions, hard tasks Lab experiments to study questions of gender dierences in performance and preference for performing in tournament Control self-selection issue Measure performance No issue of discrimination, or believed discrimination No issue of "career concerns" or "time commitment"
Marie-Pierre Dargnies
Overcondence
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Gender and competition: Niederle and Vesterlund (2007)
Basic idea: have people choose between two remuneration scheme, a piece rate and a tournament Dierence between piece rate and performance:
Payment depends on the performance of the other participants Payment is uncertain
Possible gender dierence in risk-aversion (big debate) Byrnes, Miller and Schafer (1999), Eckel and Grossman (2008): If anything, women are more risk-averse
Marie-Pierre Dargnies
Overcondence
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Niederle and Vesterlund (2007) (2) What can explain dierences in selection into competitive environments? Dierences in performance Beliefs about relative performance Risk attitudes Feedback aversion Taste for Competition
Marie-Pierre Dargnies
Overcondence
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Niederle and Vesterlund (2007) (3) Experimental design: 80 subjects Groups of 2 women and 2 men perform a real task multiple times under dierent compensations (given as experiment progresses) Performance of others not known until end of experiment No mention of gender
Add up 5 two-digit numbers for 5 mn: Performance is the number of correct answers (little gender dierence in performance)
Marie-Pierre Dargnies
Overcondence
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Niederle and Vesterlund (2007) (4)
Task 1- Piece Rate: 50 cents per correctly solved problem Task 2-Tournament: Groups of 2 men and 2 women (gender not mentioned) The participant who solves the most (correct) problems in the group receives 2 $ per correct problem Task 3- Choose compensation scheme for the next 5- minute addition task: Piece rate or Tournament (Performance is compared to task-2 tournament performance of the other participants)
Marie-Pierre Dargnies
Overcondence
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Niederle and Vesterlund (2007) (5) Given task 2 performance: 30% (40%) of women and 30% (45%) of men could gain from entering the tournament Who enters: 35% of Women and 73% of Men Performance does not predict entry for Women, weakly for Men Men are more overcondent than women For both, women and men, better beliefs predict more entry into the tournament Conditional on beliefs, women enter the tournament signicantly less often than men Marie-Pierre Dargnies
Overcondence
Introduction Are people really overcondent? Calibration and discrimination Overcondence in performance and gender dierences
Niederle and Vesterlund (2007) (6) Task-4: Submitting Piece-Rate Performance: Choose payment for task-1 piece-rate performance (PR or tournament) Beliefs, risk and feedback aversion account for 57% of original gender gap Women decide not to enter tournaments because of: Lack of condence in one's ability Psychic costs of performing once more or in tournaments Only somewhat: Aversion to feedback about tournament performance and Risk aversion
Marie-Pierre Dargnies
Overcondence