The H-1B Wage Gap Is a Measurement Problem Before It Is a Career Verdict
The subject is the capped, for-profit H-1B program and the employer–worker relationship it creates; the final question is what its wage evidence can responsibly tell us. The answer is not that measurement makes every concern disappear. It is that the meaning of a gap comes before the verdict attached to it. Once the comparison, clock, map, and institutional constraint are made explicit, the evidence supports a narrower and more useful judgment about workers, employers, compliance, and policy.
What the competing numbers actually say
Several figures in the supplied material can sound like direct contradictions. H-1B workers are reported to earn more than the average native-born worker. The revised NBER analysis reports a raw H-1B advantage of about 4.4 percent, then an adjusted disadvantage of about 16.1 percent after adding controls. The Economic Innovation Group reports a preferred corrected gap of 5.1 percent and a worksite-geography result of 5.3 percent. These are not interchangeable measurements of one settled quantity.
The broad comparison answers a selection question. H-1B workers are concentrated in technical occupations, high-wage metropolitan areas, and industries that employ relatively few native workers. EIG reports median H-1B pay of $120,000 in 2024 against $67,000 for the average native-born worker. That comparison shows that the program selects people into a high-earning part of the labor market. It does not tell us whether the H-1B worker earns less than a native worker doing similar work.
The adjusted NBER comparison answers a conditional question. It places petition wages beside wages predicted for U.S.-born workers with specified education, age, gender, occupation, industry, and location characteristics. In the revised paper, adding those controls reverses the raw result. The coefficient says that H-1B petition wages are lower than the constructed native baseline within that sample and specification. It does not say that every H-1B worker is paid less than every comparable native worker, or that visa status alone caused the difference.
EIG’s estimate answers a methodological challenge. EIG says it reconstructed the analysis without direct access to Borjas’s code and data, then changed the alignment of wage years, geographic controls, and sample filters. Its preferred specification reports a much smaller gap. That is important because it identifies plausible ways the original comparison could be biased. It is not the same as observing a definitive corrected wage distribution. The NBER manuscript disputes EIG’s timing treatment and argues that payment periods, rather than simple cohort-year matching, should anchor the deflation.
The March 2026 IZA discussion paper and the revised August 2026 NBER paper also belong to a version history rather than a pair of independent replications. The IZA abstract reports a 15 percent average disadvantage and a $97,000–$154,000 revenue-maximizing fee range under its framing. The revised NBER abstract reports 16 percent and $95,000–$152,000 for moderate excess demand. The nearby values show continuity in the author’s analysis, while the changed figures and revised text show why the version of a working paper matters. They should not be combined as if two unrelated research teams had confirmed the same estimate.
A fair reconciliation therefore has three layers. First, the program attracts and places high-skill workers in high-paying jobs. Second, under the revised NBER construction, new H-1B petition wages sit below a predicted native baseline after selected observable differences are accounted for. Third, EIG’s reconstruction argues that choices about time, geography, and sample symmetry reduce the gap substantially. The evidence supports all three statements at their respective levels. It does not supply a single percentage that can replace them.
This is why averaging 16.1 percent and 5.1 percent would be analytically meaningless. An average of disputed estimates does not resolve the disagreement; it hides which wage period, geographic unit, native baseline, and model assumptions generated each number. The right comparison is between the estimands—the questions each construction is trying to answer—not between the rhetorical force of the headlines.
What the evidence does—and does not—justify
The evidence is strongest when it describes the selected population and the sensitivity of the estimate. The revised NBER sample contains new, approved, full-time, cap-subject H-1B petitions filed by for-profit employers for workers aged 21 to 50. The native baseline is a constructed group of U.S.-born private-sector workers with a college degree and full-time, year-round employment. The observations come from different systems: an individual petition wage on one side and survey earnings on the other. This design can reveal a conditional pattern, but its boundaries are part of the result.
The employer evidence makes the boundaries humanly consequential.