Upward Discrimination in Labor Markets

Intervention(s)

This study implements two jointly randomized interventions embedded in a real job-application flow on JobPaw, Haiti's leading online employment platform. Applicants apply for real field enumerator positions.

Intervention 1 — Randomized Manager Identity (between-subjects)

Upon clicking "Apply Now," each applicant is randomly assigned with equal probability (1/3 each) to one of three binary pairwise manager comparisons:


Arm 1: Black Haitian (BH) vs. White non-Haitian (WNH). Varies race and nationality simultaneously; recovers the total local–foreign authority gap.
Arm 2: Black Haitian (BH) vs. Black non-Haitian (BNH). Holds race constant, varies nationality; recovers the local-nationality component of the authority price.
Arm 3: Black non-Haitian (BNH) vs. White non-Haitian (WNH). Holds foreign status constant, varies race; recovers the race component among foreign managers.


Both manager profiles are displayed simultaneously in randomized left–right order. All profiles are identical in job title, tasks, required qualifications, location, language requirements, and application procedures. Manager identity is signaled via name, stated country of origin, and a short standardized video introduction (filmed from a back angle showing head, neck, and hands) in which types a visible verification code. Applicants must enter the code to proceed, ensuring exposure to the identity signal before making any salary-related judgments. Manager gender is held constant across all profiles.


Intervention 2 — Randomized Wage Gap (within-comparison)

Within the assigned comparison, each applicant is also randomly assigned to one of 15 wage-gap levels with equal probability (1/15 each):

∆w = wA − wB ∈ {−140, −120, −100, −80, −60, −40, −20, 0, 20, 40, 60, 80, 100, 120, 140} USD/month

Individual salaries are symmetric around a midpoint of $570/month (consistent with the local market range for field enumerator positions). A positive ∆w means the Manager A–linked opening pays more than the Manager B–linked opening; a negative ∆w means Manager A pays less. Both salaries are revealed simultaneously after pre-reveal belief measures are collected. The wage gap is the identifying variation used to recover compensating wage differentials.

2026-06-22

2026-08-31

Binary job choice (Yi): A binary indicator equal to 1 if applicant i chooses the Manager A–linked opening and 0 if the applicant chooses the Manager B–linked opening. Recorded after randomized salaries are revealed simultaneously to the applicant.
Compensating wage differential / Willingness to Pay (WTP_p): Recovered from the binary choice model for each pairwise comparison p ∈ {BH/WNH, BH/BNH, BNH/WNH} using the formula:
WTP_p = (0.5 − α̂_p) / β̂_p
where α̂_p is the estimated equal-wage choice share for Manager A and β̂_p is the slope of the choice probability with respect to the wage gap (from the linear probability model). A positive WTP_p means Manager A must offer a wage premium to be chosen by half of applicants; a negative value means applicants accept a wage cut to work under Manager A.

Manager-specific stated expected salary (pre-reveal): The monthly salary applicant i expects to receive if offered the job under Manager A (W^r_iA) and under Manager B (W^r_iB). Both elicited before any salary offer is revealed. Within-person difference: d_ip = W^r_iA − W^r_iB.
Expected-pay premium (S_p): For each comparison p, the mean expected-pay advantage for Manager B relative to Manager A: S_p = E[W^r_iB − W^r_iA]. Estimated as −ϕ̂_0p from the regression: d_ip = ϕ_0p + γ′Xi + u_ip.
Belief–revealed preference wedge (G_p): G_p = WTP_p − S_p. The gap between the revealed compensating differential and the expected-pay premium. Positive values indicate that revealed willingness to work under Manager B exceeds what expected-pay beliefs alone would predict.
Pre-reveal job-quality belief ratings (JobPaw service-improvement items): Within-person pairwise ratings for each manager-linked opening collected before salary revelation, on six dimensions:

Career advancement opportunities (which opening is more likely to lead to promotion within 5 years)
Measure nature of discrimination from belief elicitation

Expected-salary and job-belief measures are secondary mechanism analyses. They are collected before salary revelation and thus capture beliefs about manager-linked openings before the randomized wage gap is introduced. They are not required for identification of the main compensating differentials but help interpret whether revealed choices reflect expected-pay beliefs, perceived career opportunity, residual preferences over authority identity, or a combination.

This is a between-subjects randomized field experiment embedded in a real job-application flow on JobPaw, one of Haiti's leading online employment platforms. Applicants apply for real field enumerator positions — common entry-level roles in NGOs, research organizations, and development projects in Haiti.

Design Motivation: Existing work has studied discrimination by employers, managers, customers, and coworkers against workers. This paper studies the reverse direction: vertical upward discrimination, where discrimination is directed from potential subordinates toward the supervisors under whose authority they would work. The central object is not an attitude or survey response, but a compensating differential — the salary premium required to make workers indifferent between otherwise identical jobs that differ only in supervisor identity. The paper introduces supervisor identity as a priced job attribute.

Randomization: Upon clicking "Apply Now," applicants are randomly assigned to one of three binary manager comparisons (BH/WNH, BH/BNH, or BNH/WNH) with equal probability, and to one of 15 wage-gap levels with equal probability. The two randomizations are joint and occur automatically at the platform level, invisibly to the applicant.

Experiment Flow:


Applicant views real field enumerator job posting on JobPaw → clicks "Apply Now"
Random assignment to comparison arm and wage-gap level (invisible to applicant)
Basic contact and demographic information collected
Both manager profiles displayed simultaneously (no salary shown) ; left–right order randomized
Manager videos watched; 6-character video codes entered (attention/exposure check)
Manager-specific expected salaries elicited for both openings (pre-reveal)
JobPaw service-improvement pop-up: pre-reveal job-quality belief ratings for both openings
Randomized salaries revealed simultaneously for both manager-linked openings
Binary job choice recorded (applicant selects preferred manager-linked opening)
Research consent confirmation and application submission

Not available

Randomization is performed automatically upon the applicant clicking "Apply Now." Both the comparison arm assignment (1 of 3, equal probability) and the wage-gap level assignment (1 of 15, equal probability) are generated jointly by the platform's randomization system and are not visible to the applicant or the research team during data collection.

Individual applicant. Each applicant is independently randomized to a comparison arm and a wage-gap level. The unit of analysis for the primary estimating equation is the individual applicant-comparison observation.

No

N/A

Target sample size is 1,449 eligible completed applications: 483 in the Black Haitian vs. White non-Haitian comparison, 483 in the Black Haitian vs. Black non-Haitian comparison, and 483 in the Black non-Haitian vs. White non-Haitian comparison. Within each comparison, applicants are assigned to one of 15 salary-gap levels, implying a target of approximately 32 eligible completed applications per comparison-by-salary-gap cell. Because the experiment is embedded in a real online job posting, realized sample size depends on applicant flow before the pre-specified application deadline. All eligible completed applications submitted before the deadline will be included, and the research team will not inspect outcome data to decide whether to stop, extend, or reopen recruitment. If realized sample size differs from the target, we will report realized observations by arm and salary-gap level and report realized minimum detectable effects.

N/A

The power analysis follows Hedegaard and Tyran (2018). Their estimated marginal effect of a price increase on the probability of discriminatory choice is β̂_HT = −0.036 per euro (SE = 0.016). Converted to a monthly-USD slope using the ratio of their average earnings (≈$61.25) to our salary midpoint ($570): β̂_$ ≈ (0.036 × 61.25) / 570 ≈ 0.00387 A substantively meaningful authority price is defined as m = 5% of the monthly salary midpoint = $28.50/month. The required sample per comparison for each test: Single authority-price test (H1a or H1b): N ≈ 161 per comparison (two-sided, α = 0.05, power = 0.80) Relative component test (H2a, comparing two WTPs): N ≈ 322 per comparison Additive separability test (H2b, combining three WTPs): N ≈ 483 per comparison (binding constraint) The target of 483 per comparison (1,449 total) is driven by the most demanding test, H2b. If the realized sample falls below target, realized minimum detectable effects will be reported and null results interpreted accordingly.