Gender Discrimination in Teammate Choice

Students take part in a two-session, lab-in-field experiment in an undergraduate economics course. In session 1, they complete incentivized tasks that measure agentic traits (assertiveness, diligence, self-reliance, confidence, and dominance) and cognitive ability, plus a short background survey.

In session 2, they rank anonymized classmate profiles. The intervention is the information shown on those profiles. Across rounds, cards vary by gender and by information on agentic traits, dominance, cognitive ability, and research-topic overlap. Some rounds hold other characteristics fixed so that profiles differ only by gender. Other rounds vary trait or ability information, including a clearer versus noisier cognitive-ability signal. A smaller set of rounds varies domestic versus international student status without showing gender. Rankings from these rounds are used to help assign real course project teams.

2026-02-01

2026-12-14

1.Test of gender discrimination against a 50% benchmark: whether the probability of ranking the man above the otherwise identical woman differs from 0.5, overall and separately for men and women evaluators.
2.Pairwise gender choice by precision of the cognitive-ability signal
3.Pairwise preference for higher versus lower agentic-trait profiles, holding gender fixed
4.Pairwise gender choice when agentic-trait or dominance information is shown

All primary outcomes are built from session 2 ranking ranks. Students rank profiles from 1 (most preferred) to 4 (least preferred). Ties are allowed. The analysis does not use the full four-way ranking as the outcome. It uses pairwise comparisons implied by those ranks.

For any two profiles on the same screen, the constructed outcome is 1 if the focal profile receives a strictly better (numerically lower) rank, 0 if the other profile receives a strictly better rank, and 0.5 if the two profiles tie. Analysis is at the evaluator-by-comparison level.
Under no gender preference, the pairwise gender-choice mean should be 0.5 (ties pull toward 0.5; strict choices should split evenly). Discrimination is defined as a statistically significant departure from 0.5. This is estimated overall and separately by evaluator gender.
Outcome 1 uses same-type comparisons only: the man and woman cards show the same other information (same interest match, same trait standing, and the same cognitive descriptor when one is shown). The focal profile is the man.

We do this analysis split by evaluator gender as well. The object of interest is same-gender preference: whether women are more likely than men to rank the woman above the otherwise identical man, and whether men are more likely than women to rank the man above the otherwise identical woman.

Outcome 2 uses the same constructed gender-choice variable as outcome 1, compared across rounds with no cognitive score, a noisy score (4 or more out of 8), and a precise high score (7 out of 8). When a cognitive score is shown, all cards in that round carry the same score.

Outcome 3 holds gender fixed and compares the higher-trait profile with the lower-trait profile. The focal profile is the higher-trait card. This is constructed separately for the average of four traits (assertiveness, diligence, self-reliance, and confidence), the average of five traits (those four plus dominance), and dominance only.

Outcome 4 uses the same constructed gender-choice variable as outcome 1, compared across rounds with no trait scores, rounds that show the four-trait average, rounds that show the five-trait average, and rounds that show dominance only.

Pairwise preference for higher versus lower research-topic interest match, holding gender fixed
First-order beliefs about otherwise identical Man versus Woman classmates (contribution, competence, communication and coordination, comfort, and conflict)
Second-order beliefs about what own-gender and opposite-gender classmates believe on those same gender-only items
First-order beliefs about high agentic-trait and high-dominance Man and Woman profiles (contribution, competence, communication and coordination, and conflict)
Willingness to pay for team gender composition
Own ranking of profile-card elements
Beliefs about the class ranking of profile-card elements
End-of-semester peer-evaluation ratings of teammates
Pairwise preference for domestic versus international student status

Outcome 1 is constructed from session 2 ranks in the same way as the primary pairwise outcomes: 1 if the focal profile receives a strictly better rank, 0 if the other profile does, and 0.5 if they tie. Gender is held fixed. The comparison is higher versus lower research-topic interest match. The focal profile is the higher-match card. This is collected only in sections that complete the research-topic survey.

Outcomes 2 through 5 come from the session 2 mechanism module, administered after all ranking rounds.

Outcome 2 is the student’s own belief about a classmate whose profile differs only by gender. For the Man card and the Woman card, students rate expected contribution relative to a fair share, competence, ease of communication and coordination, comfort working closely, and likelihood of creating costly conflict. The constructed objects are Man–Woman gaps on each item, and whether those gaps line up with same-gender ranking choices.

Outcome 3 asks what the student thinks other students believe, not what the student personally believes. One block asks what the average student of the respondent’s own gender believes about the Man and Woman cards. A second block asks what the average student of the other gender believes about those cards. The same contribution, competence, communication and coordination, comfort, and conflict items are used. The constructed objects are own-gender versus other-gender belief gaps, and whether anticipated others’ beliefs line up with ranking choices. Second-order answers are incentivized by closeness to the leave-one-out mean of first-order answers in the named gender group.

Outcome 4 uses six first-order trait screens: Man and Woman, each shown as high on the four-trait average, high on the five-trait average (including dominance), or high on dominance only. Items are contribution, competence, communication and coordination, and conflict. Comfort is not asked on trait screens. The constructed objects are belief gaps between high-trait or high-dominance profiles and the gender-only profiles, and whether those inferences differ by profile gender and by evaluator gender.

Outcome 5 is the maximum the student is willing to pay, from 0 to 100 dollars, for three team compositions: all teammates of their own gender, majority own gender, and majority other gender. One student from the semester may be selected for payment and implementation.

Outcome 6 is collected after the mechanism module. Students rank six profile elements from most to least important for their own teammate choice: gender, cognitive ability, research-topic interest match, average of four traits, average of five traits, and dominance only. Ties are allowed.

Outcome 7 uses the same six elements. Students rank them by what they believe the average student in the class prefers.

Outcome 8 comes from the fielded end-of-semester Qualtrics peer-evaluation survey. Each student rates each teammate on contribution relative to a fair share, competence, and related project-performance items.

Outcome 9 compares a domestic-student profile with an international-student profile. These cards do not show gender. This is a small part of the design. The focal profile is specified in analysis (for example, the domestic-student card, or the card that matches the evaluator’s own status).

The mechanism module also includes two unpaid harassment-exposure items and an optional open-text question about why the student ranked otherwise identical Man and Woman profiles as they did. Those items are exploratory moderators and descriptive text. They are not confirmatory secondary endpoints.

This is a lab-in-field experiment in undergraduate economics courses. Enrolled students are the participants. The study has two sessions, held on different days, plus a later peer-evaluation survey after the group project.

In session 1, all students complete the same incentivized trait and cognitive-ability tasks and a background survey. Some tasks are completed in pairs, so session 1 is run with an even number of students present.

In session 2, all students rank anonymized classmate profiles. The main variation is within student: each person sees several rounds in which the information on the cards changes. Some rounds differ only by gender. Other rounds add information on agentic traits, dominance, cognitive ability, or research-topic overlap. Round order and the left-to-right order of cards are randomized. Students may give the same rank to more than one profile.

Most students are in the standard version, which includes research-topic overlap on some cards. One course section does not complete the topic survey and does not see that information. After the ranking rounds, students answer questions about beliefs, how much they would pay for different team gender compositions, and which profile elements matter to them.

Rankings are used to help form real course project teams of four or five. The later survey asks students to rate their teammates after the project.

Not available

Randomization is done by computer in the oTree session, using Python’s random number generator at the participant level. For each student, round order within each ranking block and the left-to-right order of cards on each screen are randomly shuffled. Team-assignment processing order is also shuffled by computer after session 2.

Individual student. Within a section, every student sees the same ranking blocks for that section. What is randomized at the student level is round order within each block and the left-to-right order of cards on each screen. After session 2, the order in which students are processed for team assignment is also randomized at the individual level.

No

5 classes -> 200 students. A pilot(N=37) for this experiment took place in Feb 2026. The planned number of observations includes the pilot data.

200 students. A pilot(N=37) for this experiment took place in Feb 2026. The planned number of observations includes the pilot data.

Within subject experiment -> 200 students. A pilot(N=37) for this experiment took place in Feb 2026. The planned number of observations includes the pilot data.

With about 200 student respondents, alpha = 0.05 (two-sided), 80% power, and standard errors clustered by student, the minimum detectable effect sizes for the main gender-preference outcomes are as follows. The outcome is the pairwise ranking measure coded 1 / 0 / 0.5 (man strictly preferred / woman strictly preferred / tie). This outcome has an approximate standard deviation of 0.40. Discrimination against a 50% benchmark: whether P(rank man above otherwise identical woman) differs from 0.5. Overall MDE ≈ 0.06 (6 percentage points, about 0.15 SD). Within each evaluator-gender group, MDE ≈ 0.08 to 0.10 (8 to 10 percentage points, about 0.20 to 0.25 SD). Evaluator-gender gap in that same probability (same-gender homophily): MDE ≈ 0.10 (10 percentage points, about 0.25 SD). Change in that probability across information packets (agentic traits, dominance, or cognitive-signal precision): planned MDE ≈ 0.08 (8 percentage points, about 0.20 SD).