Ethnic Discrimination in U.S. Healthcare Hiring: Identification from H-1B Cap-Exempt Hospitals

How to cite

Lahey, Joanna; Melo, Vitor; Cejka, Tim. 2026. "Ethnic Discrimination in U.S. Healthcare Hiring: Identification from H-1B Cap-Exempt Hospitals." AEA RCT Registry. AEARCTR-0019358. https://doi.org/10.1257/rct.19358-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Joanna Lahey* , Vitor Melo , Tim Cejka

* Primary investigator

Registry ID
AEARCTR-0019358
DOI
https://doi.org/10.1257/rct.19358-1.0
First registered
Last updated
Status
In development
Trial dates
2026-08-24 – 2027-12-20
Countries
Private
JEL codes
J23 J15 J58 J61
Registry record
http://www.socialscienceregistry.org/trials/19358

Abstract

Ethnic discrimination is a persistent driver of labor market exclusion for workers from historically underrepresented racial and ethnic groups in the United States. For Asian and Hispanic applicants, however, name-based audit estimates conflate two distinct channels: direct ethnic bias and employer beliefs about visa cost. The channels imply different policy remedies, and existing designs cannot tell them apart. We propose a pilot study for a 15,000 resume correspondence audit in US healthcare hiring that recovers ethnic discrimination net of immigration-status effects. We exploit a sharp institutional discontinuity: nonprofit hospitals affiliated with universities are exempt from the H-1B cap and lottery, while otherwise identical nonprofit hospitals are bound by it. We randomly assign immigration status (citizen, lawful permanent resident, F-1 student) and name-signaled ethnicity (White, Hispanic, Indian, Filipino) within cap-subject and cap- exempt hospital strata. We identify ethnic discrimination cleanly within the cap-exempt stratum and decompose its mediation by visa friction in the cap-subject stratum. This pilot will 1. Validate resume signals including names and 2. Explore the feasibility of design features including number of positions available, power analysis calculations given realized response rates, and whether our hypothetical candidates are appropriate matches for our targeted positions.

Intervention

We will submit fictitious applications across distinct healthcare-sector vacancies posted on Indeed.com and individual hospital websites over a 17-month field period, for entry- to early-career positions in registered nursing, physical therapy, and clinical data science. Each vacancy receives exactly one application. Vacancies will be divided equally between two strata: hospitals subject to the H-1B cap and lottery (cap-subject), and hospitals exempt from the cap by virtue of university affiliation or qualifying nonprofit/government research status (cap-exempt). Within each stratum, we will stratify further on bed count, metropolitan-area characteristics, and posting recency to ensure the two arms are comparable on observable demand for labor.

Experimental design

Treatments
At each vacancy, we orthogonally randomize two applicant attributes:
Immigration status: U.S. citizen, or F-1 international-student visa holder will be directly signaled on each resume for all applicants. Each vacancy is assigned to one immigration cell with equal probability.
Name-signaled ethnicity: White, Hispanic, Indian, or Filipino. Each vacancy is assigned to one ethnicity cell with equal probability. The four arms reflect the largest ethnic groups whose names plausibly co-signal immigration-cost concerns to U.S. healthcare employers, and the inclusion of Filipino applicants is motivated by their disproportionate representation in the U.S. nursing workforce.
Education, work history, GPA, and skills are drawn from a curated repository of real healthcare resumes. Education will be recent or upcoming graduates of undergraduate programs from US institutions. White-signaled names list English fluency and some will list Spanish; minority-signaled names list English plus one native language, to neutralize a common confound in audit studies. We will use only female names for nursing positions, but for the pilot will include both male and female names for physical therapy and clinical data science applications. Because we will be submitting multiple resumes to each hospital (though only one resume per open position), we will randomize resume inputs such that the same hospital does not get the same verbiage for different candidates.
Identification strategy
Two contrasts are identified by random assignment within hospital strata:
Citizen vs F-1 visa: visa-related friction, holding ethnicity constant.
White-signaled vs. minority-signaled within each visa cell: ethnic discrimination, holding visa status constant.
For each focal ethnic minority group e ∈ {Hispanic, Indian, Filipino}, we estimate the following linear probability model on a subsample restricted to White-signaled (the reference category) and ethnicity-e-signaled applications:
Yi​ = α + β1Ethnicityi + β2Foreigni + β3Exempti + β4(Ethnicityi x Foreigni) + β5(Ethnicityi x Exempti)
+β6(Foreigni x Exempti) + β7(Ethnicityi x Foreigni x Exempti) + Zi'γ + εi,
In this specification, Ethnicityi indicates an applicant assigned a name signaling minority status e; Foreigni denotes the assignment of an F-1 international-student signal (relative to the U.S. Citizen reference group); and Exempti identifies vacancies within the cap-exempt hospital stratum. The vector Zi incorporates hospital-level controls such as bed count, rural-urban designation, and system affiliation. In expanded specifications, we include resume-level randomization attributes, local labor-market indicators, and state and posting-month fixed effects. We cluster standard errors at the hospital level to account for within-institution correlation.
Direct ethnic discrimination is isolated within the cap-exempt stratum, where the principal sponsorship friction (the H-1B lottery) is absent and the hiring decision is administratively close to that for a citizen. For Citizen applicants in these hospitals, the ethnic-name penalty is recovered by β1 + β5; for F-1 applicants, it is β1 + β4 + β5 + β7. A significant negative value for either sum constitutes evidence of direct bias that cannot be explained by visa-related beliefs. The visa-cost channel is further identified by the stratum contrast in the F-1 penalty: this difference equals β6 for White applicants and β6 + β7 for minority applicants. Thus, the triple-interaction coefficient β7 provides a direct test of whether immigration-policy burdens fall disproportionately on minority applicants relative to White peers. Finally, as a placebo test of our identifying assumption, we expect β5 (which captures the ethnic-penalty differential between cap-exempt and cap-subject hospitals among Citizens) to be statistically indistinguishable from zero, since U.S. Citizens require no sponsorship in either stratum; any divergence would suggest unobserved hospital-level heterogeneity unrelated to the H-1B regime, assuming no crowd-out.
Intuitively, the cap-subject vs. cap-exempt hospital contrast is what isolates ethnic discrimination from visa cost. In cap-subject hospitals, hiring a non-citizen entails a lottery, attorney costs, and timeline risk; in cap-exempt hospitals, an H-1B sponsorship can be filed at any time without lottery, and the hiring decision is administratively closer to that for a citizen. The difference-in-differences of the ethnic-name penalty across hospital strata, within visa cells, is our primary causal estimand for ethnic discrimination net of visa cost. Relative to Oreopoulos (2011), this strata contrast is the source of identification absent in his Canadian design.
One primary concern is that cap-subject and cap-exempt hospitals may differ along other characteristics that may also affect differences in ethnic and racial hiring. We address this concern in several ways. First, we will match cap-subject hospitals along observables such as size and non-profit status. Second, we collect additional observable information on hospitals and their areas to use as controls in some specifications. Third, we are using the difference-in-differences strategy described above to compare outcomes within visa cells.
A secondary concern is that our setting focuses on a specific population (working-age applicants in the US with college degrees) within a specific sector (healthcare). We view this focus as a feature of the design, not a limitation. Foreign-born workers are substantially overrepresented in US healthcare, and immigrant labor has been shown to materially raise staffing levels and improve patient outcomes in long-term care (Grabowski, Gruber, & McGarry, 2025), so hiring decisions in this sector have direct downstream consequences for the low-income and majority-minority communities that depend on safety-net hospitals. The ethnic groups we target are themselves marginalized: US-born Hispanic Americans (predominantly of Mexican descent) and Filipinos face among the highest poverty risks in the United States. Our design captures both the ethnic discrimination they experience regardless of citizenship and the additional discrimination from being perceived as immigrant, since employers may read Hispanic-signaled and Filipino-signaled names as foreign regardless of actual visa status. Finally, US healthcare is the rare setting in which this natural experiment is available: cap-exempt and cap-subject hospitals are comparable on most other dimensions relevant to hiring, which is generally not true of cap-exempt and cap-subject firms in other H-1B-exposed sectors such as technology or finance.

Design details

Randomization method
Randomization will be done in office by a computer using a resume randomizer.
Randomization unit
Results will be clustered at the firm (hospital) level to account for within-institution correlation.
Clusters
up to 491 treatment clusters (the full universe of Non-profit AND major affiliation; not all will be hiring) up to 4,861 control clusters (will be matched a-priori or ex-post to treatment clusters depending on job availability)
Observations
For the full study, we will use 15,000 observations. For the pilot, we will determine the full pilot sample size after an initial pilot of 400 applications.
Arms
We do not know-- the pilot will help determine reasonable numbers.
Primary outcomes
Call-backs and positive call-backs (positive call-backs is a subset of call-backs)
Secondary outcomes
Response quality (substantive interview-request vs. boilerplate auto-reply, coded by LLM-assisted classification with human validation on a 10% audit sample); response latency; explicit mention of visa-sponsorship questions in the response text

Cited by

No citation data available for this registration yet.

Keywords