Irrational Statistical Discrimination

Last registered on September 10, 2019


Trial Information

General Information

Irrational Statistical Discrimination
Initial registration date
September 10, 2019

Initial registration date is when the trial was registered.

It corresponds to when the registration was submitted to the Registry to be reviewed for publication.

First published
September 10, 2019, 9:01 AM EDT

First published corresponds to when the trial was first made public on the Registry after being reviewed.


Primary Investigator

Lund University

Other Primary Investigator(s)

PI Affiliation
University of Essex

Additional Trial Information

On going
Start date
End date
Secondary IDs
In the absence of complete information about individual characteristics, rational employers might systematically discriminate against workers from given identity groups. We develop a model in which employers are either bayesian updaters or conservatives. We theoretically show that, if the share of conservative employers is sufficiently high, minority workers might be more discriminated against than with bayesian updaters. The pre-registered experiment tests this hypothesis.
External Link(s)

Registration Citation

Campos-Mercade, Pol and Friederike Mengel. 2019. "Irrational Statistical Discrimination." AEA RCT Registry. September 10.
Experimental Details


Intervention Start Date
Intervention End Date

Primary Outcomes

Primary Outcomes (end points)
Whether a worker educates, Whether an employer chooses to hire the yellow worker
Primary Outcomes (explanation)

Secondary Outcomes

Secondary Outcomes (end points)
Secondary Outcomes (explanation)

Experimental Design

Experimental Design
Experimental Design Details
We will test our hypotheses using a lab experiment:
- Each session consists of 32 subjects, which are randomly divided into two metagroups of 16 subjects each.
- In each pool, we randomly assign the roles of 4 employers, 2 workers of high ability and yellow type, 3 workers of medium ability and yellow type, 1 worker of low ability and yellow type, 2 workers of high ability and orange type, 2 workers of high ability and orange type, and 2 workers of low ability and orange type.
- There are 60 rounds. In each round, 4 of the 6 workers of each colour are selected and form a pool. For example, in a given round, the pool of the yellow workers might consist of 1 high ability, 2 medium abilities, and 1 low ability, and the pool of the orange workers might consist of 1 high abilities, 1 medium abilities, and 2 low abilities. The pools are formed such that the expected ability of the workers in the yellow pool is higher than the expected ability of the workers in the orange pool.
- Each employer is matched with one yellow and one orange worker. The employer knows the pool composition from which the workers have been drawn, be she does not know their ability. The workers know their own ability, and see the pool from which they have been drawn, but do not know the ability of the other worker.
- Each round consists of two stages. In the first stage, each worker decides whether to pay a cost to pursue education. If the worker pursues education, the probability that he graduates is 0% if he is low ability, 80% if he is medium ability, and 100% if he is high ability. In the second stage, the employer sees whether each worker graduated. She then decides whether to hire one worker or none of them. The payoffs are such that employers are not interested in hiring non-graduated workers. But if they graduated, they prefer that the worker has as high ability as possible.
- The matching is done such that, if both workers graduated, bayesian employers will pick yellow 50% of the times and orange 50% of the times. However, naive employers will only pick yellow.

The main treatment variation takes place in round 31:
- In round 31, the computer counts the number of times that each of the 8 employers chose to hire the yellow worker. Using this as a proxy for naiveté, it re-matches the employers in the two metagroups: it assigns the 4 more naive employers to the first metagroup, and the 4 more bayesian employers to the second metagroup. Hence, on expected terms workers in each metagroup are as likely to face a naive employer during rounds 1-30. However, in rounds 31-60, workers in the first metagroup are more likely to be matched with naive employers, while workers in the second metagroup are more likely to be matched with bayesian employers.
Randomization Method
The matching randomization is done by the program z-tree during the experiment.
Randomization Unit
Was the treatment clustered?

Experiment Characteristics

Sample size: planned number of clusters
Sample size: planned number of observations
320 (10 sessions of 32 subjects each)
Sample size (or number of clusters) by treatment arms
240 workers, 120 in metagroup one and 120 in metagroup two
Minimum detectable effect size for main outcomes (accounting for sample design and clustering)
Main test: The main hypothesis that we will test is whether, after round 31, orange workers in metagroup one are less likely to educate than orange workers in metagroup two. We will restrict this analysis to medium and high ability workers, since low ability should in theory never educate. We will use a regression analysis where the outcome variable is whether a worker educates in each period. We will explain it with a dummy indicating whether the round is after round 31, this dummy interacted with a dummy indicating whether the megagroup is the first or the second one, and subject fixed effects. We performed a power analysis after running the first two sessions. We could not perform it before since our priors regarding the underlying data generating process were too noisy. To perform the power analysis, we bootstrap the sample from the first two sessions to reach 320 subjects. We then randomly assign them to one of the two metagroups and perform the regression described above. We run 10000 simulations and store the (on average null) treatment effects for each regression. We use these data to estimate that with 320 subjects we have 80% power to detect an effect of 0.112 percentage points at the 5% level. In other words, if orange workers in the naive pool are 0.112 pp less likely to educate than orange workers in the bayesian pool, we have an 80% chance to capture a significant effect. Secondary tests: - Employers: Before runnuing the main test, the first hypothesis that we will test is whether employers who can choose between two graduated workers are indeed more likely to hire the yellow worker than the orange worker. This test is essential for the rest of the analysis. We expect to easily reject the test that on average employers hire yellow workers half of the time. - Employers: We will study whether employers become more bayesian or naive over time. - Emplyoers: We will explore whether employers' naiveté is correlated with their answers in a post-experimental survey. - Workers: We will also study whether yellow workers choose differently depending on the megagroup that they are placed in, as predicted by the theory. - Workers: We will also explore whether workers' decisions is correlated with their answers in a post-experimental survey.

Institutional Review Boards (IRBs)

IRB Name
IRB Approval Date
IRB Approval Number


Post Trial Information

Study Withdrawal

There are documents in this trial unavailable to the public. Use the button below to request access to this information.

Request Information


Is the intervention completed?
Data Collection Complete
Data Publication

Data Publication

Is public data available?

Program Files

Program Files
Reports, Papers & Other Materials

Relevant Paper(s)

Reports & Other Materials