The H-1B Wage Gap Really Is That Large

A challenge to my H-1B research doesn’t stand up to scrutiny.

Courtesy U.S. Department of Labor

Jiaxin He and Adam Ozimek of the Economic Innovation Group recently published a critique of my new National Bureau of Economic Research (NBER) paper on the H-1B wage gap—the difference between what H-1B visa recipient immigrants and similar native-born workers make—and its implications for the demand for H-1B workers. They assert that my estimate of a 16 percent gap is too high, for two reasons. First, they claimed that my initial work did not adjust for inflation, even though it did. They assert that the estimated wage gap falls by about half after adjusting for inflation. Second, they argue that the wage gap falls even more if the regressions use an alternative set of geography controls.

In addition to the He-Ozimek critique, I received a review (which the author indicated was not for public distribution) that highlighted potential biases introduced by two modeling issues. This short note addresses both critiques. A forthcoming revision of the NBER working paper will incorporate the various extensions. As I show here, my results are robust to these critiques, particularly when I correct crucial errors in the He-Ozimek calculations.

The table above summarizes the results of both my and the He-Ozimek analyses.¹ The first row reports a 14.1 percent wage gap estimated using H-1B wage data from the FY2021-FY2024 lotteries and the 2023 American Community Surveys (ACS), as I did in my original paper. The regression controls for age, gender, education, geography, and occupation.

These numbers are inflation-adjusted, as the NBER paper notes in reference to a table reporting summary statistics (p. 8: “Both groups of workers have high salaries, slightly over $100,000 [in 2023 dollars]”). The footnote that explained the deflator was inadvertently cut, so it is understandable if a reader missed the fact that the entire analysis was indeed using inflation-adjusted wages.

Row 2 presents the results implied by the He-Ozimek research design, which suggest a smaller wage gap, depending on the exact specification. There is, however, a fundamental error in that design, which can be easily grasped from an email exchange I had with Ozimek (the point is more obscurely made in footnote 4 of the He-Ozimek critique). Ozimek wrote:

“The 2024 fiscal year H-1B data should be merged to 2023 ACS data, and so on for earlier years”

Phrased differently, He-Ozimek calculate the H-1B wage gap by comparing the wage of the FY2024 lottery winners with the wage of natives in the 2023 ACS. This research design decision reflects a misunderstanding of how the process works, contaminating and biasing their analysis. H-1B workers should be compared to native workers in the year when they will be earning the bulk of the salary, not in the year when they were awarded the visa.

To understand why, let’s review the process. The FY2024 H-1B lottery was conducted in March 2023. The actual employment of H-1B recipients begins after their H-1B status is approved and after the federal fiscal year begins on October 1, 2023. In other words, the earliest that FY2024 recipients could actually start earning money was in the last quarter of 2023.

The FY2024 lottery data show that 64.1 percent of the lottery winners had an H-1B status start date (i.e., the date on which H-1B status becomes legally valid) of October 1, 2023; an additional 24.2 percent had a status start date between October 2 and December 31, 2023; and 11.7 percent had a status start date sometime in 2024.

When H-1B status starts is not the same as when workers actually start working. In many cases, both firms and workers encounter scheduling issues that can further delay the employment start date. The available data do not report when the employment (rather than H-1B status) of the lottery winners begins, but many probably started working sometime in 2024. The Department of Homeland Security has documented some long delays. The FY2017-FY2020 data show that 42.8 percent of the “petitions that selected consular processing into the United States and that DHS was able to match with the beneficiary’s arrival data into the United States” arrived with at least a 6-month delay past the requested employment start date or the petition approval date (whichever is later).²

Regardless of when the FY2024 H-1B recipients started working, employers made the wage offers knowing much of their salaries will be paid out in the 2024 calendar year. The H-1B data show whether the petition requested consular processing and the valid status date. These data imply that at most 16.5 percent of the salary offers made to FY2024 lottery winners was paid in calendar year 2023.³

Because of all this, it is obviously incorrect to compare the offer wage of the FY2024 lottery winners with natives in the 2023 ACS. Conceptually, the He-Ozimek research design distorts the usual meaning of wage gaps in labor economics, which typically refer to earnings differences across groups at a point in time. A comparison of H-1B earnings mostly paid in calendar year 2024 with native earnings, none of which were paid in 2024, cannot produce an unbiased estimate of the wage gap, particularly in a period of rapid inflation.

To illustrate the bias, I reanalyze the data using the He-Ozimek methodology. Specifically, I use the annual Consumer Price Index (CPI) to deflate earnings of workers in the H-1B sample. The earnings of FY2024 lottery winners are (erroneously) treated as being in nominal 2023 dollars; the earnings of FY2023 lottery winners are (erroneously) treated as being in nominal 2022 dollars, and are converted into 2023 dollars; etc. The H-1B lottery winners are then pooled and compared to native earnings in the 2023 ACS, which are treated as being in 2023 dollars (more on this shortly). Column 1 of row 2 of the table above shows that the He-Ozimek mismatching indeed cuts the original estimate of the wage gap by about half, to 7.3 percent.

There is, in fact, another inflation-related issue that contaminates the results. The ACS samples a new group of households each month, and respondents report earnings in the past 12 months. About half of the earnings reported in the 2023 ACS were, in fact, earned in 2022. A correct conversion of ACS nominal earnings into real dollars should therefore take account of the fact that ACS earnings almost always span two calendar years. The simplest solution is to use the mean CPI in years t-1 and t to adjust earnings in the year t ACS. Column 2 shows that using the mean deflator raises the estimated wage gap to 9.3 percent even in the He-Ozimek framework.⁴

The combination of the ACS monthly sampling and the incorrect matching of the H-1B and ACS wage data “supersizes” the He-Ozimek bias. Consider again the winners of the FY 2024 lottery. The He-Ozimek methodology assumes that the H-1B wage data is in nominal 2023 dollars, even though the employer’s offer was made knowing that the job cannot start prior to October 1, 2023 (and, in fact, over 83.5 percent of the salary will be paid out in 2024). In addition, He and Ozimek use the 2023 ACS to get native earnings, where half the income was received in 2022. Their calculation treats all income of natives and H-1Bs as being in 2023 dollars. In a period with annual inflation averaging 5 percent, the three-year range in nominal earnings that are treated as being in real 2023 dollars could easily make much of the H-1B wage gap vanish.

To summarize, in the He-Ozimek research design, the earnings of native workers typically come from an earlier time period than the earnings of the H-1B workers they’re matched to. That allows H-1B earnings to benefit from an erroneous inflation adjustment and reduces the H-1B wage gap.

The second critique I received noted that my model ignored that H-1Bs and natives work in very different industries, and there are well-documented and sizable interindustry wage differences. For example, the largest industry employing natives is “Elementary and Secondary Schools” (12.2 percent of natives), while the largest industry employing H-1Bs is “Computer Systems Design and Related Services” (47.5 percent of H-1Bs). To account for this, Columns 3 and 4 in the table above add industry fixed effects to the regression model.⁵ The wage gap increases after controlling for industry. Even using the He-Ozimek specification, the wage gap rises to 12.1 percent.

The second critique also noted that I did not exploit the availability of several ACS cross-sections (a point also raised by He and Ozimek). Row 3 uses the He-Ozimek methodology to illustrate what happens to the estimated wage gap if all the data are pooled and the FY2021 H-1B data is incorrectly compared to the 2020 ACS, the FY2022 H-1B data is incorrectly compared to the 2021 ACS, and so on. The coefficients are essentially the same as those obtained when using only the single 2023 ACS cross-section.⁶

Row 4 reports the estimated wage gaps obtained by using all the ACS cross-sections and the correct matching of native and H-1B workers. This row essentially generalizes the results in the NBER paper. Column 2, in fact, replicates what was done in my original work (as the deflator accounts for the fact that half of ACS earnings are received in the prior calendar year). The estimated wage gap is 15 percent, about the same as in the original paper.

The He-Ozimek critique also notes that the geography controls in my original work are not properly defined, as they combine worksite information for the H-1Bs (at the level of “Public Use Microdata Areas,” or PUMAs, the smallest geography for which the two sources can be matched) with residence information for the natives in the ACS. The ACS, however, also provides information on the state and PUMA of employment.

These variables, however, have a high imputation rate. The quality flag for the variable giving the state of employment indicates that the Census Bureau imputed this variable for 14 percent of natives, with the imputation rate being even higher in states that have many H-1B workers (California, New York, and Texas have rates between 15 and 18 percent). Unfortunately, there are no quality flags for either the PUMA or the metropolitan area of employment. Given the baseline of a 15 percent imputation rate for the state of employment, the state-PUMA combinations that would be used to create the geography fixed effects are likely imputed at much higher rates.

Regardless, row 5 shows that the preferred estimates of the H-1B wage gap in columns 2 and 4 is between 12 and 18 percent, depending on whether industry fixed effects are included in the regression. Finally, row 6 shows that the regressions produce similar estimates of the wage gap when they use the metro area of employment rather than the state-PUMA combinations to define the geography fixed effects.

As noted above, a sizable fraction of the state of employment data is imputed by the Census Bureau. This imputation seems to bias the results. If the regression in row 5 and column 4 (using the state-PUMA of employment and controlling for industry) excluded the subsample of natives whose employment location was imputed, the coefficient would rise to -0.190. This is similar to the coefficient obtained by using the state and PUMA of residence of natives (row 4, column 4).

Finally, it is notable that the size of the H-1B wage gap is essentially the same if the analysis used the Current Population Survey’s Annual Social and Economic Supplement (CPS-ASEC) to construct the native sample, rather than the ACS. One advantage of using the CPS data is that annual earnings are reported for the calendar year prior to the survey (so there is no overlapping-years deflator issue as in the ACS). Some disadvantages are that the CPS sample of native workers is smaller, that the CPS does not report geographic information at any level lower than the metropolitan area, and that the CPS only reports the place of residence (and not of employment).

The year t ASEC reports annual earnings in calendar year t-1. This makes it easier to match H-1B and native earnings. The 2024 H-1B lottery data, which reports earnings mainly paid out in 2024, is then matched with the 2025 ASEC, which also reports earnings paid out in 2024, and so on. As row 7, column 3 of the table shows, the estimated H-1B wage gap is identical to that obtained using the ACS data (after correcting for both the He-Ozimek mismatch error and the deflator of overlapping ACS years). It is 17.2 percent in both cases.

In sum, He and Ozimek made a research design error that contaminates their analysis and biases their conclusions. It is incorrect to claim that “the 2024 fiscal year H-1B data should be merged to 2023 ACS data.” This error likely would not matter much if the study examined wages during a low-inflation period, as the wage data in consecutive cross-sections would be relatively constant. But it matters in the 2021-2024 period, when inflation was high, and produces the false impression that the H-1B wage gap is not large.

Note: The replication code can be downloaded here.

1. While preparing this note, I discovered a coding typo unrelated to the He-Ozimek critique that slightly changes the estimate of the wage gap reported in the NBER draft.

2. A high fraction of H-1B lottery winners (50.2 percent in FY2024) typically request consular processing. A December 2025 report by Bloomberg confirms: “Workers from outside the US, rather than recent international graduates of US colleges already in the country, accounted for more than four out of 10 new H-1B hires approved over the past four years.”

3. To illustrate, 33,492 of the 77,689 FY2024 lottery winners in the sample have a valid status date prior to October 31, 2023, and did not request consular processing. If all started work on October 1, they had 3 months of 2023 salary. However, 24,181 of the winners also have a status date prior to October 31 but requested consular processing. The DHS data suggests that 42.8 percent of this group had their entry delayed until 2024, so only 57.2 percent of this group earned 3 months of 2023 salary. Repeating the calculations for the lottery winners with a valid status date in November 2023 (which potentially provides two months of 2023 salary) or December 2023 (which potentially provides one month of 2023 salary) results in the 16.5 percent statistic.

4. The conceptual issue also arises in the H-1B wage data, but the implied error is very small. Although the distribution of the employment start date is unknown, the fraction of H-1B salary paid in the calendar year prior to the lottery fiscal year is less than 16.5 percent. Using this fraction to weigh the CPI in the two adjacent years reduces the H-1B wage gap by at most 0.008.

5. These fixed effects are constructed using the North American Industry Classification System (NAICS) industry variable in the H-1B data, collapsing it into four digits, and using the IPUMS crosswalk to convert it into codes comparable to the census industry code (i.e., “ind” in the ACS).

6. I initially discovered the similarity when doing exploratory work. It was part of my motivation for keeping things simple and focusing on a single ACS wave in the NBER paper.