Most Lost Callbacks Come From a Minority of Employers

How concentrated is it?

A fifth of the discriminating firms carried close to half the harm. Kline, Rose and Walters sent more than 83,000 fictitious applications to entry-level vacancies at 108 large US employers: 24 percent drew employer contact within 30 days, distinctively Black names reduced the chance of contact by 2.1 percentage points, and that gap was concentrated among a minority of the firms ¹.

The concentration numbers are worth stating precisely, because both the alarming and the reassuring readings of them are wrong.

• The top quintile of discriminating firms produced nearly half the lost contacts, and the Gini coefficient of employer contact gaps came out around 0.4, which makes discrimination against Black names about as concentrated among these firms as income is among US households ¹.
• 23 of the 108 companies were identified as discriminating while holding the false discovery rate at 5 percent ¹. That is a floor produced by statistical power, not a list of the only offenders, and roughly one of the 23 is expected to be a false positive.
• The other 85 were not cleared. The authors could not reject the hypothesis that all 108 firms weakly favored White names, and they put a lower bound of at least 7 percent of all jobs in the experiment discriminating, rising to at least 20 percent of jobs at the flagged firms ¹.

The design limits travel with the finding. These are very large employers, entry-level vacancies, and portals that were easy to audit. The outcome measured is employer contact within 30 days, which stops well short of an offer. Nothing in it observes an interview.

Why doesn't the industry average describe your company?

Because an average is computed across firms whose behavior points in different directions, and it deletes exactly the variation you are trying to find. The same experiment found no average gender gap in employer contact at all, and underneath it a between-company standard deviation of 2.7 percentage points, roughly symmetric about zero: some firms consistently favored men and others consistently favored women ¹.

Take that as a warning about your own reporting. Any average taken over units that behave differently will cancel opposing gaps and return a number that describes none of them. Averaging across your req families will do to you exactly what averaging across firms did to the gender result there, and the detection problem compounds it: bidirectional gaps lower the probability that any single firm looks significant, so a clean report can mean low power rather than even-handed behavior.

Where the gap sits also moves with the kind of work. A large resume audit of 36,880 applications to 9,220 advertisements for new US college graduates found callbacks 28 to 43 percent lower in management occupations for Black men, Black women, White women and Hispanic men than for otherwise identical White men, with the widest gaps in roles combining high analytical and interpersonal demands with low routine content ². That work is a 2026 preprint under review, and the discretion mechanism is the authors' proposal; the experiment does not manipulate it. Taken carefully, it points where you would expect: the less specified the evaluation, the more room there is for something other than the work to move it. That is the practical case for structure, and it is the same argument that runs through which stage the gap opens.

Cut your own numbers by req family before you cut them by company

Cut the numbers at the level where a decision actually gets made, which sits below the company. A company-wide ratio is an average over req families that hire through different processes, use different screeners and attract different pools, and it will cancel opposing gaps the same way the market average cancelled the gender result. Cut the numbers by job family and by stage first, then look at the company total to see what it hid.

1. Fix the unit before you compute anything. One req family, one stage, one time window. A number without a unit invites a comparison to an industry figure that was computed on a different unit entirely. 2. Report the denominator beside the ratio. Forty applications is not a measurement. Publishing the count alongside the rate is what stops a spreadsheet becoming a claim. 3. Keep what each decision rested on. The criterion applied, and the specific thing in the application that met or missed it. A rate tells you a difference exists; only the record tells you what produced it, and only the record can be disagreed with. 4. Write down what you did with the result. California's amended employment regulations make evidence, or the lack of evidence, of anti-bias testing relevant to a claim that an automated-decision system or a selection criterion discriminated, and to any defense of that claim, including the quality, recency and scope of the effort and the response to the results. The same amendments extend the employment-records retention period from two years to four, naming selection criteria and automated-decision system data expressly ⁴. Testing and then doing nothing is the position that reads worst. That is California regulation, effective October 1, 2025, and how far it reaches your own process is a question for employment counsel.

When the tool doing the screening belongs to a vendor and the data sits on their side, the mechanics change and the questions get sharper: how to run an adverse impact audit when the vendor holds the data is the version of this for bought software.

What if your volume is too small to measure?

Say so, and stop there. A ratio computed on nine hires and four rejections is arithmetic rather than evidence, and publishing it invites a conclusion the data cannot support in either direction. The honest posture is to state the number of decisions behind any figure you release, and to treat a sample that cannot separate signal from noise as a reason to keep better records.

The record and the ratio are different assets. A ratio needs volume before it says anything. A written reason attached to one decision is complete on the day it is written, works at any scale, and is the only thing that lets a specific candidate's outcome be examined at all. Small employers have no access to the first and full access to the second, which reverses the usual assumption that fairness measurement is something only large companies can afford.

The market-wide picture is worth carrying only as context. A meta-analysis of every available US field experiment on hiring discrimination against African Americans or Latinos, 28 studies and 55,842 applications, found Whites receiving on average 36 percent more callbacks than African Americans since 1989, with no detectable change over the following 25 years ³. Carry it as the background your own numbers sit against; it describes no individual employer. The two practical follow-ons are what to do when you have too few hires to measure bias and, for the smallest teams, the smallest hiring process you can defend.

See a sample report >