Measurement Checkpoint: What Does a Wage Gap Mean?

This checkpoint concerns the H-1B employer–worker relationship after its petition and survey records have been turned into a wage comparison. The task is to decide what a negative gap can support—and what it cannot—before choosing among competing estimates. The useful judgment is not “which percentage feels right?” but “what kind of claim is this percentage making, and what evidence would be required to make a stronger claim?”

A wage gap is a chain of comparisons

The H-1B program does not produce one natural wage number. It produces a petition attached to a particular worker and employer, inside a system where new for-profit visas are scarce and selected through a lottery. An analyst then places that petition beside a population of U.S.-born workers drawn from a different data system. The resulting estimate inherits every important choice in that chain: which workers enter, which earnings count, which job and place are treated as comparable, and which differences are held constant.

That is why the broad and adjusted results can both be meaningful. A broad comparison can show that H-1B workers are concentrated in highly paid technical jobs and major metropolitan areas. A controlled comparison can show that, after selected observable differences are accounted for, the petition wages in the revised NBER analysis are lower than the wages predicted for a constructed U.S.-born baseline. The second statement is narrower, not automatically more “real” in every context. It answers a different question.

The data architecture makes the chain visible. The LCA describes a position and its occupation; the I-129 identifies a beneficiary, employer, worksite, and wage offer after a lottery win; the ACS supplies annual earnings for a selected population of U.S.-born workers. The regression connects those observations through common categories such as age, education, occupation, industry, and geography. It does not observe the H-1B worker standing beside the exact U.S.-born applicant whom the employer would otherwise have hired.

The conditional estimate is therefore a comparison generated by a model. In the revised paper, the raw specification reports an H-1B advantage of about 4.4 percent, while the fully controlled specification reports an adjusted disadvantage of about 16.1 percent. The movement between them is informative because it shows how much the composition of the two samples matters. It is also a warning: a coefficient is not independent of the variables and categories used to produce it.

A useful way to read the result is as a claim ladder. Each step can be relevant, but each step requires more evidence than the one before it.

flowchart TD A["Observed petition and survey wages"] --> B["Average comparison"] B --> C["Conditional statistical association"] C --> D["Legal required-wage comparison"] D --> E["Individual underpayment or causal claim"] E --> F["Policy welfare conclusion"] G["Additional evidence required at each step"] -.-> D G -.-> E G -.-> F

The arrows should not be read as automatic escalation. A conditional estimate does not become a legal finding merely because it follows an average comparison. A policy conclusion does not become established merely because it follows a regression.