What the H-1B Data Actually Observe

To understand the H-1B wage estimate, it helps to examine the evidence chain before debating the percentage. The focus here is what the Labor Condition Application, the individual H-1B petition, and the American Community Survey each observe—and what is lost when their records are joined. That map supports a practical judgment: a petition wage, a survey earning, and a legal required wage are related, but they are not interchangeable measurements of the same thing.

Three records, three views of one hire

The Labor Condition Application, or LCA, begins with a position rather than a person. A firm seeking H-1B workers files an LCA with the Department of Labor and attests that employing workers in the named occupation will not adversely affect the working conditions of similarly employed U.S. workers. One application can cover several positions. It carries a case number, identifies whether the positions are full-time, and records the occupation using a detailed Standard Occupational Classification code.

This makes the LCA a description of the job the employer says it wants to fill. It is useful for locating the occupation and the formal conditions attached to the position, but it does not, by itself, tell us which beneficiary will ultimately take the job or what that person’s full compensation will look like over time.

The individual connection arrives through Form I-129, the petition a firm files with the Department of Homeland Security after winning a place in the lottery. Unlike the LCA, the I-129 identifies a particular foreign national. It records such characteristics as age, gender, country of birth, educational attainment, worksite, and employer identification number. It also reports the salary or wages paid to the beneficiary and includes the case number of the associated LCA.

The distinction matters. The LCA says something about a position that may cover multiple workers; the I-129 says something about a named worker whom a particular firm proposes to employ. The wage in the I-129 is a petition wage offer, and the form instructions require the employer to report the salary or wages paid to the beneficiary. It is much closer to an individual hiring decision than the LCA is. But it remains a record created during sponsorship and hiring, not a complete longitudinal payroll history.

The native comparison comes from a third system: the American Community Survey. The ACS does not identify H-1B status for the foreign-born, so the study constructs a baseline from U.S.-born private-sector wage-and-salary workers aged 21 to 50 who have at least a college degree and work full-time, year-round. In the paper’s revised sample, the comparison contains 581,648 native observations. The H-1B side contains 343,397 approved petitions for new, full-time, cap-subject hires by for-profit employers in the same broad age range.

The three systems therefore divide the work of the estimate. The LCA supplies a position’s occupation and full-time status. The I-129 supplies a specific beneficiary’s proposed wage, characteristics, worksite, and employer. The ACS supplies annual earnings and demographic and job characteristics for a large population of U.S.-born workers. None of the three is a complete picture of the employment relationship. The estimate is created at their interfaces.

Those interfaces also have an order. The LCA describes the employer’s intended position; the firm then enters the lottery; only after winning a slot does it file an I-129 for a named beneficiary. The ACS is sampled independently rather than being linked to the worker whom the H-1B employer considered as an alternative. A successful merge can therefore line up education, age, occupation, industry, and place, but it cannot turn the native observation into the actual rejected applicant or the exact job the H-1B worker filled. It produces a comparison population with shared labels, not a paired experiment.

This is why the word “data” should not obscure the division of responsibility. The employer’s filings are shaped by a legal and administrative process, while the ACS is a household survey designed to measure the population’s work and earnings. The former records an offer associated with a petition; the latter records reported annual earnings over a survey-defined period. The estimate becomes informative when those different observations can be placed on common dimensions, but every common dimension is also a claim about what differences can be safely ignored.

erDiagram LCA_POSITION { string case_number string soc_code string full_time_status string worksite_position } I129_BENEFICIARY { string beneficiary_id string wage_offer string age_gender_education string worksite string employer_id } ACS_NATIVE { string survey_year string annual_earnings string age_education_gender string occupation_industry } CROSSWALK { string geography_map string occupation_map string industry_map } EMPLOYER { string employer_id string firm_name } LEGAL_WAGE_RULE { string prevailing_or_actual string required_wage_basis } LCA_POSITION ||--o{ I129_BENEFICIARY : "case number links" I129_BENEFICIARY }o--|| EMPLOYER : "petitioned by" I129_BENEFICIARY }o--|| CROSSWALK : "worksite and job mapped" ACS_NATIVE }o--|| CROSSWALK : "comparison categories mapped" LEGAL_WAGE_RULE ||--o{ LCA_POSITION : "governs wage conditions"

The diagram’s most important feature is not the number of boxes. It is the direction of the evidence chain. The employer files the LCA and I-129 materials; the petition identifies the beneficiary; the ACS supplies a separate population baseline; and crosswalks are required before the two sides can be compared. A result produced after those steps is necessarily conditional on how the interfaces were designed.

The sample is a boundary, not a census

The H-1B observations are not a random sample of everyone who holds an H-1B visa. They come from new petitions filed by firms subject to the annual cap between fiscal years 2021 and 2024. The study keeps petitions for full-time positions that DHS approved—more than 97 percent of the relevant petitions—and workers aged 21 to 50. It focuses on for-profit firms because those are the employers covered by the 85,000-visa cap that the study is examining.

That choice gives the analysis a coherent institutional population: new lottery-linked hires at capped for-profit employers. It also limits the result. It does not automatically describe workers at universities, school districts, or medical centers whose H-1B visas are cap-exempt. It does not describe an established H-1B workforce in the same way as a new-hire sample. And it does not include every kind of compensation or every stage of a worker’s relationship with an employer.

The native sample is narrowed to resemble the H-1B sample, but the symmetry is not perfect. Native workers are selected from survey responses about annual earnings, full-time year-round work, private-sector employment, education, age, and birthplace. H-1B workers are selected from approved petitions containing a wage offer, a named employer, and a worksite. The two groups are made more comparable by observable characteristics, yet they are not observed under the same reporting system or at the same point in a job’s life.

One example is missing education information. About 17 percent of the I-129 filings do not report the beneficiary’s educational attainment. The paper classifies those workers as college graduates because a college degree is the minimum education requirement for the visa. That is a transparent assumption, and the paper reports sensitivity checks that produce similar gaps when the affected observations are excluded or assigned an “unknown” category. Still, the choice illustrates a general rule: an apparently simple variable can contain an analyst’s judgment about what an absent field probably means.

The same issue appears in the geographic and occupational categories. The I-129 worksite is reported as a ZIP code, while the ACS uses metropolitan areas of employment. A ZIP code can cross more than one metropolitan area, so the paper’s crosswalk assigns it to the area with the largest population; observations that cannot be assigned to an ACS metro receive a state-based fallback. The H-1B data use far more detailed occupation codes than the ACS. The analysis reduces the detailed codes to a common Census occupation classification. Industry codes also need to be translated from the North American Industry Classification System into the ACS’s industry categories.

These translations are necessary to make the comparison possible, but they are not merely clerical. A detailed occupation can be split, combined, or represented differently in a survey classification. A worksite ZIP code can be a rough proxy for the labor market in which a worker competes. A state fallback is not the same geographic information as a metropolitan area. When a regression later says that two workers are in the same occupation or location, that sameness is defined by these crosswalks.

The sample’s restrictions also shape the meaning of “comparable.” The H-1B side is deliberately concentrated in high-skill, new-hire, private-sector employment. The native side is designed to approximate that population using survey variables, not to identify the exact U.S.-born worker whom the employer would otherwise have hired. The estimate can be informative about the selected labor-market comparison while remaining silent about workers outside it.

The new-hire feature is especially consequential. An employer’s petition records the wage attached to entering a job, while many native workers in a survey may have already accumulated experience with their current employer. That does not make either observation unusable; it means the comparison can contain a difference in job stage as well as a difference in visa status. A later analysis can try to estimate that seniority effect, but the data architecture itself cannot make it disappear. The result must be read as a comparison whose population and point in the employment lifecycle were chosen in advance.

There is a similar selection issue on the H-1B side. The data show people whose firms won a scarce slot and whose petitions were approved. They do not show all workers who entered the lottery, all offers that were never pursued, or every H-1B worker already in the country through another route. The observed hires are the survivors of employer demand, lottery selection, petition approval, and the study’s filters. This is precisely the population that is relevant to asking how current lottery-winning employers use the program, but it is not a neutral sample from which every broader claim follows.

A legal wage floor is a different interface

The Department of Labor’s Fact Sheet #62G adds a legal distinction that is easy to miss when all the evidence is described as “wage data.” The fact sheet says that an H-1B employer must pay at least the required wage, defined there as the higher of the prevailing wage or the employer’s actual wage for similarly employed workers. It describes the actual wage in relation to workers with similar experience and qualifications in the specific employment, and the prevailing wage in relation to similarly employed workers in the geographic area.

That rule is not the same as the ACS baseline. The ACS comparison asks how annual earnings differ between a selected group of U.S.-born workers and H-1B petition observations after statistical controls. The DOL rule asks whether a particular employer has met the applicable wage floor for a particular employment. The legal comparison is employer- and place-specific; the statistical comparison is population-based.

The difference is not just a matter of precision. A prevailing wage is a rate associated with an occupation and geographic area, while an actual wage concerns the employer’s similarly employed workers and the beneficiary’s experience and qualifications. A survey baseline can contain people working for many employers and in many establishments even after geographic and occupational controls are added. It is therefore possible for a study to find a population-level difference without answering the legal question of which rate applied to one petition. The two interfaces should inform one another, but they should not be merged into a single test.

The fact sheet also says that employers must pay the guaranteed minimum hours reported on the petition, except when the H-1B worker is unavailable for work for reasons such as a voluntary request for time off, and that the Wage and Hour Division can enforce the wage provisions. Those points matter for compliance, but they do not turn the study’s regression coefficient into a legal finding. A worker could earn less than a constructed native baseline while the employer complied with the applicable required wage. A violation could also be missed by a broad statistical comparison that does not examine the employer’s actual wage records and wage determination.

The fact sheet is a general-information publication revised in 2008. It is therefore most useful for the basic distinction between a legal floor and a market comparison, not for resolving current regulatory details or judging an individual case.

This distinction also clarifies what the three-record system can and cannot do. The LCA and I-129 make the employer’s sponsorship and wage offer visible. The ACS makes a comparison population visible. Neither establishes that the H-1B wage is the exact wage paid over the worker’s career, that the U.S.-born baseline is the worker’s true alternative, or that a gap is unlawful. Those stronger conclusions require evidence beyond the merged sample.

Once the records are understood as different views rather than interchangeable facts, the next analytical problem becomes more precise. The question is not simply why the study reports a gap. It is why adding geographic, skill, occupational, and industry comparisons can reverse the raw result—and whether those controls are comparing the right workers in the right places at the right time.

References