Revised Version of H-1B Wage Gap Paper Available
Changes and Updates
Since the public release of my NBER Working Paper on the measurement and implications of the H-1B wage gap in mid-February, I have received many comments, and (at the latest count) two written critiques (He and Ozimek, 2026; and Clemens, 2026). I have now revised the paper to incorporate the suggestions and address the questions raised. The key result is unchanged. There is a large H-1B wage gap (-15.5 percent), so that even fees exceeding $100,000 may not change the demand for H-1B workers all that much. The revised paper is here and the code is here. These are the key changes:
1. Emphasize the importance of properly deflating earnings
The revision uses all ACS cross-sections between 2021 and 2024 to correspond with the FY2021-FY2024 H-1B data.[1] Because the sample period covers years of high inflation and because the timing of wage receipts differs between the sample of native workers and the H-1B data, the method used to convert nominal into real dollars can influence the results. Any study of this issue should satisfy the following condition:
Wage data are expressed in constant dollars at the time the earnings were received.
Let me illustrate using the 2024 data. The FY2024 H-1B lottery was held in March 2023. Winning firms filed I-129s for the beneficiaries shortly thereafter and those forms reported the wage offered once employment starts. The data do not report the actual employment start date. But it is unlawful for employment to begin before the later of the date in which the beneficiary’s H-1B status is approved and the beginning of the federal fiscal year (October 1, 2023).
I use the beneficiary’s status approval date to convert earnings into real dollars. Some beneficiaries received status approval on or before October 1, 2023. I assume these workers encountered no further delays and started working on October 1, so that they received three months of salary in calendar year 2023 and another nine months in calendar year 2024. The deflator is the average monthly CPI between October 2023 and September 2024. Similar calculations are done for those who received status approval after October 1 but on or before December 1. Finally, the deflator will be the mean CPI for calendar year 2024 for beneficiaries who received status approval after December 1, 2023.
The ACS native wage data also has a timing issue. The ACS samples a new group of households each month, and earnings refer to the previous 12 months. This means that about half of the earnings reported in the 2024 ACS were earned in 2023. The deflator used for wages in the 2024 ACS is the mean of the annual CPI in 2023 and 2024.
Alternative adjustment exercises that ignore the condition that earnings should be expressed in constant dollars at the time the earnings were received will produce different results, sometimes cutting the estimated wage gap by half. However, deviations from this condition are not conceptually justified.
2. Use alternative native samples, CPS and SIPP
The size of the H-1B wage gap is essentially the same if I use the CPS Annual Social and Economic Supplement (ASEC) to construct the native sample, rather than the ACS (see Table 5, row 2). Apart from helping establish the robustness of the evidence, one advantage of the CPS is that annual earnings are reported for the calendar year prior to the survey (avoiding the overlapping-years deflator issue in the ACS).
The size of the wage gap in the SIPP data is totally different than in the ACS and the CPS: It is essentially equal to zero (Table 4, rows 5 and 6). There are well-documented measurement issues in the SIPP that predict this outcome.
Meyer, Mok, and Sullivan (2015, p. 203) report that “wage and salary income are close to administrative totals in the CPS but substantially lower in the SIPP.” Czakja and Denmead (2012, p. 12) concludes that “wage and salary earnings in the ACS were slightly lower than in the CPS ASEC, but SIPP was 15 percent lower.” Finally, a National Academy report (2009, pp. 127-128): “SIPP captures nearly as much transfer income and substantially more self-employment income but less wage and salary income…As a result, SIPP underestimates total CPS income by 11 percent.” The bias, therefore, results in average native incomes far below the level reported in the CPS. Ironically, the absence of a wage gap in the SIPP suggests that the actual wage gap may be between 11 and 15 percent.
3. Main job earnings versus total earnings
The H-1B wage offer reflects earnings in a single job, while native earnings in the ACS or CPS may be exaggerated because they reflect total earnings in multiple jobs. Another advantage of using the ASEC earnings data is that it can be merged with the CPS Basic monthly files. The monthly files contain a variable indicating if the respondent had multiple jobs in the reference week. I estimated the ASEC regression using the sample of native workers who reported working in only one job in all monthly interviews. The estimated wage gap barely changes (see Table 5, rows 3 and 4).
4. Better geography controls
Several readers noted that the geographic controls should be something larger than a state-PUMA combination. The paper now uses the metropolitan area (of employment in the ACS, and of residence in the CPS) to define the geography fixed effects.
5. Private sector employment and industrial composition
Several readers also pointed out that the correct native baseline for the publicly available H-1B data should be natives who work in private sector firms (as this closely corresponds with the I-129 petitions filed by firms who won the lottery). The native baseline throughout the revised paper consists of natives who work in private sector firms.
Similarly, my initial draft ignored that H-1Bs and natives work in very different industries, and there are well-documented and sizable interindustry wage differences. In fact, the concentration of H-1Bs into a small number of industries is greater than their concentration in particular occupations. The largest industry employing H-1Bs is “Computer Systems Design and Related Services,” and that industry alone employs nearly half of the workforce. The fully specified regression model in the revised paper includes industry fixed effects.
6. Treatment of missing H-1B education data
About 17 percent of the I-129 filings do not report the education of the H-1B beneficiary. When I first noticed the missing data, I hesitated to drop the observations because it seemed that a lot of information would be lost. Instead, I classified those workers as “college graduates” because that is the typical minimum education required for the visa. It is also an assumption that, if anything, would bias the wage gap towards zero. Many of the H-1Bs erroneously classified as college graduates have higher degrees and presumably higher earnings, making it look as if H-1Bs with a college education earn more than the natives. The revised paper (footnote 23) notes that the regression coefficients would be very similar if all the H-1B observations with missing education information were simply excluded from the analysis.
7. The (un)importance of job seniority bias
The H-1B wage offer is for the first year on the job and it is being compared to the average wage of natives, many of whom have substantial job seniority. This comparison may exaggerate the wage disadvantage of H-1B workers. Unfortunately, few data sources can be used to avoid the bias. Although the SIPP contains the relevant information on job tenure, the survey does not correctly measure native earnings (see point 2).
There are two ways to get a sense of how large this bias could be. One is to defer to what the literature concludes about the returns to job seniority, and the second is to derive a simple model that roughly estimates the magnitude.
The literature concludes that the wage growth attributable to job seniority is very small. Altonji and Shakotko(1987) report that the return to a year of seniority is only about 0.6 percent (see also Altonji and Williams, 2005; Williams, 2009). The most recent study concludes: “returns to firm tenure…are close to zero” (Dustmann and Adda, 2023, p. 486). In short, the job seniority bias, if it exists, is small.
It is easy to derive algebraically what the bias would be and use data from the CPS Job Tenure Supplement to get a rough estimate (see pp. 16-18 of the revised paper). The CPS data reveals that the wage gap between natives who have been on the job less than a year (i.e., the “brand-new workers” equivalent to H-1Bs in terms of job seniority) and natives who have been on the job more than a year is only 3.7 percent. This premium implies that the job seniority bias is about +0.029 log points. In short, the bias-free wage gap would still be quite large, about 12.6 percent (or -.155 + .029).
8. A totally different approach: The “likely H-1B” wage gap
Many studies measure labor market outcomes of undocumented immigrants by imputing a “likely undocumented” identifier in survey data like the ACS or CPS (based on an algorithm developed by the Pew Research Center). The likely undocumented imputation is based on observed demographic characteristics (e.g., foreign-born persons who are undocumented cannot be citizens, cannot be married to a citizen or a naturalized citizen, cannot qualify for refugee status, etc.).
The identification of “likely H-1B” status in survey data is even easier (see the discussion in pp. 18-20): a noncitizen who migrated after age 18, has been in the country fewer than 6 years, and works in one of the top 5 industries and/or occupations where over two-thirds of actual H-1Bs work. This imputation yields a sample of likely H-1Bs that is disproportionately composed of Indian or Chinese nationals, and who concentrate in the same 5 cities where actual H-1Bs concentrate.
The wage gap between likely H-1Bs who have been in the country 1 or 2 years (i.e., long enough to have a full year’s worth of earnings and not long enough to have experienced wage growth) and natives exceeds -12 percent (see Table 5). This alternative approach is useful because it provides an independent ballpark figure of the wage gap and can serve as an intuitive prior for the magnitude of the H-1B disadvantage.
[1] The 2024 ACS cross-section was not available when I first began to examine the data, so I used the most recent cross-section available at the time. I knew that adding the 2021 and 2022 cross-sections had little effect on the results, so I did not revisit this decision after the 2024 data was released in December 2025. Some readers have claimed that adding the other cross-sections changes the results of the study, but this claim is incorrect. The estimated gap reported in the revised draft goes from -0.155 (0.003) to -0.176 (0.005) when the native baseline is restricted to only the 2023 data.