Intersectional Categories in a Bias Audit: Sex by Race/Ethnicity

The Table That Doesn't Match the Headline Number

Say your company renews an AEDT contract every January, and this year you actually open the vendor's posted bias audit summary before signing off. The top-line numbers look fine — overall selection rates for men and women are close, and the race/ethnicity breakdown looks reasonable too. Then you scroll down and hit a second table: rows like "Female, Hispanic or Latino," "Male, Asian," "Female, Black or African American." Some of these intersectional rows show impact ratios well below the ones you just saw. A couple look alarming. A couple have footnotes about small sample size. You don't know which numbers to trust, or what you're even supposed to do with a table like this before you renew.

This is what an intersectional bias audit for sex and race/ethnicity is built to surface — patterns that a sex-only or race-only breakdown can hide entirely. By the end of this piece, you'll know why the cross-tab exists, how to read a selection rate and impact ratio inside it, and how to tell a real gap from a small-sample artifact before you take anything to your auditor or your leadership.

What an Intersectional Bias Audit for Sex and Race/Ethnicity Actually Measures

Local Law 144 requires an annual independent bias audit of any automated employment decision tool used on NYC candidates or employees, plus a public summary posted on the employer's website naming the AEDT's distribution date (Crowell & Moring LLP, 2023; Epstein Becker Green, 2023). An AEDT, under the law, is any computational process built on machine learning, statistical modeling, or similar techniques that produces a simplified output — a score, classification, or recommendation — used to substantially assist or replace human discretion in an employment decision (Perkins Coie, 2023).

The bias-audit portion of that summary typically doesn't stop at "selection rate by sex" and "selection rate by race/ethnicity" reported separately. It also breaks the population down by the combination of the two — sex crossed with race/ethnicity — producing a longer table of intersectional groups. This is the intersectional layer people mean when they talk about an intersectional bias audit for sex, race, and ethnicity. The exact category labels and grouping conventions a given auditor uses can vary by methodology, so if you're comparing two vendors' summaries side by side, confirm the specific taxonomy each one used rather than assuming they match.

Why the Cross-Tab Exists at All

The reason intersectional reporting matters is straightforward once you see it worked through: an AEDT can clear a sex-only comparison and a race-only comparison and still disadvantage one specific combination of the two. A tool might select men and women at roughly even rates overall, and select White and Black candidates at roughly even rates overall, while still selecting Black women at a meaningfully lower rate than every other group. Neither of the two single-category breakdowns would catch that pattern on its own — only the intersection does.

That's the whole logic of running an intersectional bias audit for sex and race/ethnicity instead of two separate single-axis audits: the two axes can look clean independently while the cell where they overlap does not.

Reading the Selection Rate and Impact Ratio for Each Group

Once you're looking at the intersectional table, you're reading the same two numbers you'd read anywhere else in a bias audit summary — just recalculated for each combined group.

Selection rate is the share of a group's applicants (or candidates, or employees, depending on the AEDT's use case) who advanced or were selected.

Impact ratio compares each group's selection rate to the highest-selected group's rate. The long-standing four-fifths rule, drawn from the EEOC's Uniform Guidelines, treats a selection rate below 80% of the top group's rate as a potential signal of adverse impact worth investigating (via Assessment Systems, 2024).

Here's a worked example, using round numbers to show the method rather than a real vendor's figures. Say "White, male" is the highest-selected intersectional group at a 30% selection rate. "Black, female" comes in at 22%. The impact ratio is 22 ÷ 30 = 73.3% — below the 80% line, which flags that cell for closer review under the four-fifths framework. Compare that against a group at 27%: 27 ÷ 30 = 90%, comfortably above the threshold. The math doesn't change between the single-axis and intersectional tables — only which groups you're running it on. If you want the full mechanics of that calculation, including how selection rate and impact ratio are computed from raw counts, that's covered separately.

The Small-Sample Problem Built Into Intersectional Reporting

Splitting an applicant pool by sex and then again by race/ethnicity multiplies the number of groups fast — and at a 50-to-250-employee company, some of those intersectional cells can end up with a handful of applicants. A single hiring decision can swing a small cell's selection rate by ten or twenty points. That's not evidence of bias; it's evidence of a small denominator.

This is a real limitation, not a hypothetical one. Research examining published LL144 audits has found that many may under-report disparities because of missing demographic data, opaque aggregation choices, and metrics that don't reflect how the tool is actually deployed day to day (ACM FAccT, "Auditing the Audits," 2025). A separate study of 391 employers found that only 18 had posted an audit report at all, and only 13 had posted the required transparency notice (ACM FAccT, Wright & Muenster et al., 2024). If the underlying reporting is thin to begin with, an intersectional cell built on top of it deserves extra scrutiny before anyone treats it as conclusive — in either direction.

A wide gap in a large intersectional cell is a signal worth raising. A wide gap in a cell with a handful of candidates is a reason to ask for the underlying counts before you draw a conclusion.

Where This Fits Your Compliance Operation, Not a Legal Determination

Nothing in this article, and nothing in a workbook or reading guide built around it, is legal advice, and none of it performs, certifies, or signs a bias audit. That work belongs to the independent auditor engaged under Local Law 144, with no financial or employment relationship to your company or your AEDT vendor. What a documented reading method gives you is a defensible, repeatable way to review what that auditor published — so that when you take a question about an intersectional cell to counsel or back to the vendor, you're asking a specific, well-formed question instead of a vague one. If you're new to reading a published summary end to end, or want the fuller compliance picture LL144 sits inside, those are worth reading alongside this piece.

Enforcement of all this has been uneven so far. A December 2025 Comptroller's audit covering July 2023 through June 2025 found the city's own enforcement "ineffective" — DCWP found only one of 32 reviewed companies non-compliant, where the Comptroller's auditors found 17 (OSC, 2025). That gap is exactly why reading the underlying table yourself, instead of trusting a headline pass/fail, matters.

Your First Move: Read the Cross-Tab With a Method, Not a Guess

Before you sign off on a renewal, flag a vendor to your auditor, or brief leadership on what a published summary actually shows, you need a repeatable way to turn a row of percentages into a defensible read — not a gut call on which numbers "look bad." The Four-Fifths Rule & Impact-Ratio Reading Guide walks the calculation step by step, including how to apply it to intersectional cells, with a companion calculator so you're not doing the division by hand under deadline pressure. Download it, run it against the last summary your vendor posted, and you'll know exactly which cells — if any — are worth a real conversation.