Reducing perceptions of discrimination (follow-up to AEARCTR-0009592 and AEARCTR-0011806)

Participants start by taking the same baseline survey as participants in the prior experiments (quizzes related to proofreading skills and demographic questions) and are paired to a prior participant with similar baseline aptitude for the work task, education level, and demographics. The intervention provides information about how that paired worker was evaluated and assigned to the easier of two proofreading tasks in an earlier experiment. There are two key elements to the information that workers learn about how their paired worker was evaluated. First, they learn about the procedure itself (what information was used to make the decision, and who the decision-maker was). Second, they learn about the workers that were previously assigned the harder task by the same decision-maker in an even earlier sample. Unlike earlier trials, I measure comprehension of the provided information at multiple intervals, and reinforce the treatment information via confirmation or correction of participants' responses to the comprehension questions.

The decision-making procedures are:
1. A manager who knows demographics when deciding who to hire
2. A manager who does not know demographics when deciding who to hire
3. An algorithm that does not use demographic information
4. An algorithm that does use demographic information

Among those who learn about decisions made by a demographic-blind manager or algorithm, participants are cross-randomized to learn the demographics of the previously-promoted workers or not. Those who learn about decisions made by a non-blind manager or algorithm always learn the demographics of the previously-promoted workers. These three “previously-promoted” workers were primarily white men due to the demographic composition of that earlier sample; whether participants learn about a decision-maker who historically promoted three white men or two white men and one woman is cross-randomized.

2026-06-25

2026-07-10

Perceived discrimination

The main measure of perceived discrimination is the explicit measure of perceived discrimination used in the original experiment: Mentioning race, gender, bias, or discrimination in a free-response question about what needed to be different about their paired participant's profile to be assigned to do the harder of the two proofreading tasks

Comprehension of the different decision-making procedures, secondary measures of perceived discrimination, beliefs about the likelihood of being assigned to each task in the future if evaluated under the same decision-making procedure, and reservation wages to be evaluated and assigned to tasks in the future

Secondary measures of perceived discrimination are complaints about discrimination or bias, and answering “yes” to questions about being hired if their gender or race was different.

Participants take the same baseline screening quizzes as in the earlier experiments and answer demographic questions. Participants are randomly assigned to learn about one of the decision-making procedures and whether they will learn about an example in which three white men were promoted or two white men and a woman were promoted, and paired to a similar prior participant who not promoted after being evaluated by that type of decision-maker who made that particular past decision. Participants learn about the evaluation of their paired worker and answer the survey questions underlying all main outcomes.

Participants complete the screening quizzes and answer questions about their demographics, then move immediately into the experimental manipulation. They are randomly assigned to learn about decisions made by a blind or non-blind manager or algorithm (four arms) and cross-randomized to either learn about a case in which three white men were previously promoted or a case in which two white men and a woman were previously promoted. They are paired to a previous participant who was evaluated in a way that aligns with the current participant's assigned treatment status and who matches them as closely as possible on several demographic characteristics. All paired participants are white women, and participants are paired to previous participants who scored in the same quintile as them on the screening quizzes and have the same education. (All participants are also white women). If possible, they are also matched on whether they are currently employed outside Prolific and age and income bins. The likelihood of being paired to a past participant is weighted such that the treatment arms are balanced on characteristics like how the previously-promoted workers' education and quiz scores compare to the paired participant. All of the possible paired participants were evaluated by white, male managers with college degrees who were employed outside Prolific and passed attention checks during the evaluation procedure—and this is all that participants learn about the managers.

Workers learn about how their paired worker was evaluated (and not promoted) by a manger or an algorithm and what information that decision-maker had available at the time. Then, I elicit comprehension of that information and reinforce the correct answer after each question. Participants know in advance that there will be incentivized comprehension questions.

Then, they learn about who was promoted by the same manager/algorithm in the past. Among those evaluated by a blind manager or algorithm, participants are further cross-randomized to learn the demographics of those previously-promoted workers or not (as in one prior experiment). Then, participants are asked why they think their paired worker was not promoted. This question underlies the main measure of perceived discrimination, as in the prior experiments. Then, I elicit participants' beliefs that they will be assigned to each type of task if they are evaluated by the same manager/algorithm in the future and the lowest wage at which they would be willing to be evaluated and offered one of the two tasks. Like in the first two experiments, this reservation wage is incentivized via a BDM mechanism which participants know will be implemented for a random subset of participants. Next, I elicit comprehension of the information that decision-makers had available to them for a second time (now, after learning or not learning the demographics of previously-promoted workers), and again correct any misperceptions. Finally, I elicit the secondary measures of perceived discrimination that were also used in the prior experiments.

Participants are randomly assigned to treatment groups using Qualtrics' randomizer feature.

Individual

No

N/A

3600 individuals

Blind manager, don't see previously-promoted workers' demographics: 200 participants
Blind manager, do see previously-promoted workers' demographics: 800 participants
Non-blind manager, do see previously-promoted workers' demographics: 800 participants
Blind algorithm, don't see previously-promoted workers' demographics: 200 participants
Blind algorithm, do see previously-promoted workers' demographics: 800 participants
Non-blind algorithm, do see previously-promoted workers' demographics: 800 participants

In each of the above groups, half of the participants see previously-promoted workers who are all white men and the other half do not.

Research question 1: Does changing hiring procedures change rates of perceived discrimination, holding the outcomes of those procedures fixed? To answer this research question, I am interested in comparing blind- to non-blind managers, blind to non-blind algorithms, blind managers to blind algorithms, or non-blind managers to non-blind algorithms, all when workers learn the demographics of previously-promoted workers. I calculate conservative MDEs by focusing on the comparisons of two groups at a time without control variables, but will estimate the results jointly in one regression with controls as described in the PAP. Without conditioning on whether participants see three white men or not, each of these calculations compare two groups of 800 participants. In the most conservative case (when 50% of the comparison group perceives discrimination), I will be powered to detect a 7.0pp difference between any two groups (based on analytical power calculations). Research question 2: Does minority-group representation change rates of perceived discrimination? By construction, half of each of these four groups will see that three white men were previously promoted. Again in the most conservative case where 50% of the comparison group perceives discrimination), comparing two groups with N=400 is powered to detect differences larger than 9.8pp. Comparisons to “pure control”: In both prior experiments, less than 1% of participants have perceived discrimination when evaluated by a demographic-blind decision-maker without learning the demographics of previously-promoted workers; this group serves as a pure control. Comparing any of the four main treatment arms (each N=800) to the pure control group with the same type of decision-maker (N=200) will be powered to detect differences larger than 4.0pp, assuming that 1% of the pure control group perceives discrimination. Comparing either subsample that sees three white men or not (N=400) to the pure control group (N=200) is powered to detect differences larger than 4.4pp.