New — The State of AI Hiring 2026. Every number sourced→
Ethics & Compliance

Impact Ratio and the Four-Fifths Rule: Formula and Worked Example

The impact ratio formula under NYC Local Law 144: selection and scoring rates, intersectional tables, the 2% rule and the EEOC four-fifths rule.

Impact ratio formula with an illustrative three-row table: White 1.000, Asian 0.974, Black or African American 0.749.

The formula: Impact ratio = a category's selection rate ÷ the selection rate of the most selected category

Race/ethnicity categoryApplicantsAdvancedSelection rateImpact ratio
White78029337.6%1.000
Asian38013936.6%0.974
Black or African American3209028.1%0.749

Illustrative data, not real. White is the most selected category, so each category's rate is divided by 37.6%. The full worked example is further down the page.

TL;DR: The impact ratio divides a category's selection rate by the selection rate of the most selected category. New York City's Local Law 144 requires the ratio to be calculated and published in three tables but sets no threshold. The 0.80 figure is the EEOC's four-fifths rule of thumb (29 CFR § 1607.4(D)), a screening reference; exactly 0.800 is not "less than" four-fifths.

General information, not legal advice. Have counsel confirm how this applies to you.

Updated 21 October 2026. First publication. This page is reviewed quarterly, and again whenever the Uniform Guidelines, the DCWP rules or the DCWP FAQ change.

Key takeaways

  • The impact ratio divides a category's selection (or scoring) rate by the rate of the most selected (or highest scoring) category, per 6 RCNY § 5-300 (DCWP Notice of Adoption, April 2023).
  • Local Law 144 requires three tables (sex, race/ethnicity, and their intersections) plus a count of individuals in an "unknown" category, and sets no minimum ratio.
  • The four-fifths rule is a federal rule of thumb: a ratio below 0.80 "will generally be regarded" as evidence of adverse impact, and smaller or larger differences can still matter, or not matter, depending on numbers and statistical significance (29 CFR § 1607.4(D)).
  • A sex-only table can hide a larger gap in an intersectional group, as the worked example below shows.
  • A self-check on this page is not a bias audit. Local Law 144 requires an independent auditor.

The impact ratio is simple division. What makes it harder under New York City's Local Law 144 is the form the rule requires: three tables, a count of individuals in an "unknown" category, an optional 2% exclusion, and a separate scoring-rate version for tools that score rather than select. A two-group calculation that ends at 0.80 reaches none of that.

This page gives the formula as the rule states it, the three tables the rule requires, and a worked example in which a sex-only table looks unremarkable while one intersectional group sits far lower. Two points frame everything that follows. The audit requires the ratio, but, as the page on how Local Law 144 treats impact ratios explains, the law sets no threshold. The familiar 0.80 figure comes from a separate federal rule of thumb.

What is the impact ratio, and how do you calculate it?

The impact ratio compares how often each category of people moves forward. In the rule, it is the selection rate for a category divided by the selection rate of the most selected category. For tools that score candidates, it is the scoring rate for a category divided by the scoring rate of the highest scoring category (6 RCNY § 5-300, DCWP Notice of Adoption, April 2023). The most selected category therefore always has a ratio of 1.00.

The same source defines the two rates:

  • Selection rate: the number of individuals in the category moving forward or classified, divided by the number of individuals in the category who applied or were considered.
  • Scoring rate: the rate at which individuals in a category "receive a score above the sample's median score."

To calculate it:

  1. Count the applicants (or individuals considered) in each category, and how many of them advanced.
  2. Divide the advanced count by the applicant count to get each category's selection rate.
  3. Identify the highest selection rate.
  4. Divide each category's selection rate by that highest rate.

A quick two-group example shows the mechanics. Suppose 100 men and 50 women apply, and 40 men and 12 women advance. The selection rates are 40% and 24%, so the women's impact ratio is 0.60. For the women's rate to reach 80% of the men's rate, which is 32%, 16 of the 50 women would have to advance.

That arithmetic is the whole formula. Local Law 144 applies it across many more categories, and the result is one of the numbers an audit reports; what a bias audit measures is covered on its own page. The next section turns the four steps into a tool.

The three tables to produce

A spreadsheet is enough for a self-check. Choose the mode first: selection, for a tool that advances or classifies candidates, or scoring, for a tool that produces a score.

What you need

  • Applicants and selected counts for the seven race/ethnicity categories and the two sex categories. In scoring mode, the entries are applicants and the number scoring above the sample's median.
  • A count of individuals in the "unknown" category.

What to produce

  • Three tables (sex, race/ethnicity, and sex by race/ethnicity), each with the selection rate and impact ratio, computed from unrounded values.
  • A note of categories under the 2% threshold, which "may be excluded." Exclusion is the independent auditor's decision and requires a written justification.
  • The "unknown" count, reported but not calculated.
  • Ratios below 0.80 labeled as below the EEOC four-fifths reference point. That is not a Local Law 144 threshold.

What it is not. A self-check, not a bias audit. Local Law 144 requires an independent auditor. If you need the audit itself, how to conduct an AI bias audit walks through the process and the steps for vetting an auditor. SiftFirst also publishes a CSV-based self-check for readers who prefer a different format.

Test vectors from the rule

These are DCWP's own published examples, recomputed here from unrounded values.

Example (source: Notice of Adoption, § 5-301)PublishedRecomputed from unrounded values
Selection, Male: 1,390 applicants, 667 selectedRate 48%, ratio 1.00667/1,390 = 47.99%
Selection, Female: 1,181 applicants, 555 selectedRate 47%, ratio 0.979555/1,181 = 46.99%; 0.4699 ÷ 0.4799 = 0.979
Race/ethnicity, Black or African American (Not Hispanic or Latino): 390 applicants, 170 selectedRatio 0.84(170/390) ÷ (412/797) = 0.843
Scoring, Male and FemaleScoring rates 54.3% and 44.7%; ratio 1.00 and 0.82Not recomputed here

In the rule's selection example, 250 individuals of unknown sex or race are reported, not calculated.

How Local Law 144 uses the impact ratio

The rule requires three tables: one for sex, one for race/ethnicity, and one for the intersections of the two. The audit also reports the count of individuals the tool assessed in an "unknown" category. That count is reported but not calculated (6 RCNY § 5-301(b)–(c), Notice of Adoption).

The 2% exclusion. The independent auditor "may exclude a category that represents less than 2% of the data." If it does, the summary must give the justification and the excluded category's applicant count and rate (§ 5-301(d)). Whether DCWP applies the 2% test table by table is a question for counsel and the auditor, and the worked example below shows why the answer matters.

No threshold. DCWP's FAQ states: "The Law does not require any specific actions based on the results of a bias audit." (DCWP FAQ II.2, 29 June 2023.) According to DCWP's statement of basis in the Notice of Adoption, the calculations are consistent with Section 1607.4. The arithmetic matches the federal guideline, but the rule itself contains no minimum ratio.

Where the numbers end up. The primary source for the required tables is the Notice of Adoption linked above. For a real published summary, see Pfizer's Bias Audit Summary of the HiredScore Grade System, published May 2026. It sets out three impact-ratio definitions and excludes groups under 2%. It is shown here as an example of format, not as an endorsement.

Worked example: why a sex-only table hides a gap

The data below are illustrative, not real. They are built to show how the three required tables can tell three different stories about the same hiring step. The 1,520 individuals in the calculation exclude 60 additional individuals in the "unknown" category, which are reported but not calculated.

Same illustrative data in three tables; the sex table shows women at 0.860 while the intersectional table shows Black women at 0.618.

SexRace/ethnicityApplicantsAdvancedSelection rateImpact ratio
MaleWhite40016040.0%1.000
MaleAsian2007638.0%0.950
MaleBlack or African American1504832.0%0.800
MaleNative American or Alaska Native20630.0%excluded (under 2%), still reported
FemaleWhite38013335.0%0.875
FemaleAsian1806335.0%0.875
FemaleBlack or African American1704224.7%0.618
FemaleNative American or Alaska Native20525.0%excluded (under 2%), still reported

The page shows the eight cells of a simplified 2 × 4 version. The rule uses the full set of categories.

Collapsing the same data into the other two tables gives the following.

TableCategoryApplicantsAdvancedSelection rateImpact ratio
SexMale77029037.7%1.000
SexFemale75024332.4%0.860
Race/ethnicityWhite78029337.6%1.000
Race/ethnicityAsian38013936.6%0.974
Race/ethnicityBlack or African American3209028.1%0.749
Race/ethnicityNative American or Alaska Native401127.5%0.732

The example makes three points.

  1. The sex table alone hides a larger gap. Women's ratio is 0.860 in the sex table. In the intersectional table, Black women sit at 0.618, because the sex table averages them with White and Asian women whose rate is 35.0%.
  2. Exactly 0.800 is not "less than" four-fifths. Black men advance at 32.0% against 40.0% for White men, a ratio of exactly 0.800. The federal guideline draws its line at "less than four-fifths," so this group sits on the line, not below it. What that means for a given employer is a question for counsel.
  3. The same group can be over 2% in one table and under it in another. Native American or Alaska Native applicants are 2.6% of the 1,520 in the race/ethnicity table (40 individuals), so that category stays in. Its two intersectional cells are 1.3% each (20 individuals), and each falls under the 2% line. If the 2% test applies table by table, the group is included in one table and excludable in another.

AI hiring bias examples sets out the documented cases that make these numbers matter.

What does the four-fifths rule say?

The four-fifths rule appears in the federal Uniform Guidelines on Employee Selection Procedures. The text of 29 CFR § 1607.4(D) (2024 edition) says that a selection rate for a group that is "less than four-fifths (4/5) (or eighty percent) of the rate for the group with the highest rate will generally be regarded by the Federal enforcement agencies as evidence of adverse impact."

The same paragraph adds two qualifications:

  • Smaller differences "may nevertheless constitute adverse impact" where they are significant in both statistical and practical terms.
  • Greater differences "may not constitute adverse impact where the differences are based on small numbers and are not statistically significant."

"Less than" means exactly 0.800 is not below the line

The rule reads "less than four-fifths." A ratio of exactly 0.800 does not meet that wording, as the Black men in the worked example show. A ratio of 0.799 does.

Scale of impact ratios from 0.618 to 1.000 showing that exactly 0.800 sits on the four-fifths line, with the 29 CFR 1607.4D wording quoted.

Is the four-fifths rule the law?

No. It is a rule of thumb in federal guidelines, and the guidelines say the agencies "generally" regard a lower ratio as evidence of adverse impact. The statutes, including Title VII, and state and city laws are the law; the four-fifths rule is a screening reference.

Adverse impact and impact ratio

The two terms are related but not interchangeable. The impact ratio is the number: one category's rate divided by the highest rate. Adverse impact is the conclusion the federal guideline says a low ratio may be evidence of. A ratio can be calculated for any table; whether it amounts to adverse impact depends on the numbers, their significance and the legal question being asked.

The legal theory behind adverse impact claims, disparate impact, is being tested in court. The Workday case covers it; the claims there are allegations, and no court has found Workday liable. As of 28 October 2026, the CourtListener docket shows a Fourth Amended Complaint filed on 10 July 2026 and a motion to certify a class filed on 14 September 2026, with a hearing set for 9 March 2027.

Does the four-fifths rule still apply in 2026?

That is a question for counsel, and the facts below are as of 15 October 2026. Executive Order 14281 (23 April 2025), Section 4, tells agencies to "deprioritize enforcement" of disparate-impact liability (White House). The EEOC's May 2023 technical assistance page on AI and adverse impact returned a not-found error on 15 October 2026; Warden AI reports its removal, and no EEOC announcement was found.

How far state and city regimes and private litigation continue to use the 0.80 figure as a reference point is also a counsel question. This page does not predict enforcement. Whatever the federal posture, selection data is read best alongside the evidence behind each decision, and the page on evidence-based AI interviews covers how that evidence is recorded.

FAQ

How do you calculate impact ratio?

Divide a category's selection rate by the selection rate of the most selected category. In practice, count applicants and advanced individuals for each category, divide to get each rate, find the highest rate, and divide every category's rate by it. For scoring tools, the same steps apply to scoring rates.

What is a good impact ratio?

A ratio of 1.00 means a category advanced at the same rate as the most selected category. Local Law 144 sets no cut-off, and the EEOC's 0.80 is a screening reference, not a target. Counsel interprets any result.

Is there a minimum acceptable impact ratio under Local Law 144?

No. The law requires the number to be calculated and published, not to meet a mark. DCWP's FAQ states that the Law "does not require any specific actions based on the results of a bias audit."

Is the four-fifths rule the law?

It is a rule of thumb in 29 CFR § 1607.4(D), which federal enforcement agencies "generally" treat as evidence of adverse impact when a rate is less than four-fifths of the highest rate. It is a guideline for screening, not a statute.

What is the 2% rule in a bias audit?

The independent auditor may exclude a category that represents less than 2% of the data. The summary must give the written justification and the excluded category's applicant count and rate. Whether the test applies table by table is a question for counsel and the auditor.

How is the impact ratio calculated for scoring tools?

Take the median score of the full sample. A category's scoring rate is the share of its individuals who score above that median, and the impact ratio is a category's scoring rate divided by the highest category's scoring rate.

What does an intersectional impact ratio look like?

It is the same division applied to sex-by-race/ethnicity cells. In the worked example above, Black women advance at 24.7% against 40.0% for White men, a ratio of 0.618, while the sex-only ratio for women is 0.860.

Does the four-fifths rule still apply in 2026?

The rule text appears in the 2024 edition of the CFR, but federal enforcement posture has shifted, as described above. Whether and how it applies to a particular employer is a counsel question.

Need the audit itself?

An independent auditor performs it; if you want to see the candidate-level evidence a hiring tool records, see how an evidence pack records one decision.

ComplianceAI hiring law
Was this useful?

Comments

No comments yet — start the conversation.

Leave a comment

Your email is never published. Comments appear once a moderator approves them.

Replying to
Plain text only. Be kind and stay on topic.

By commenting you agree to our privacy policy.