Four-Fifths Rule

Where the rule comes from

The rule is in the Uniform Guidelines on Employee Selection Procedures, adopted in 1978 by the EEOC, the Department of Labor, the Department of Justice and the Civil Service Commission (now the Office of Personnel Management). Section 1607.4(D) says a selection rate for any race, sex or ethnic group that is less than four-fifths of the rate for the group with the highest rate will generally be regarded by the federal enforcement agencies as evidence of adverse impact, while a greater ratio generally will not.

The same section is careful to call it a practical guide, not a legal test. It notes that smaller differences can still be adverse impact when they are statistically and practically significant, and that greater differences may not be when the numbers are small. Courts treat it the same way: a useful first screen, not the final word on disparate impact.

How to calculate the impact ratio, step by step

1. Pick the step. Run the calculation for each decision point separately: knockout questions, an assessment, a phone screen, interviews, offers.
2. Count by group. For each group, record how many people were considered at that step and how many advanced. You need voluntary self-identification data for this.
3. Compute selection rates. Advanced divided by considered, for each group.
4. Find the highest rate. That group is the reference point.
5. Divide. Each group's rate divided by the highest rate gives its impact ratio.
6. Flag anything under 0.80, then check whether the difference is statistically significant.

A common slip is comparing groups' shares of hires instead of selection rates. If women are 25% of hires and 25% of applicants, that is not adverse impact even though women are a minority of hires. Another is using the wrong denominator, such as counting everyone who started an application rather than everyone who reached the step.

Worked example across a whole funnel

A contact center hiring for a 40-seat class runs this funnel for one quarter. Hispanic applicants: 600 apply, 420 pass knockout questions (70%), 210 pass the call simulation (50% of 420). White applicants: 500 apply, 375 pass knockouts (75%), 225 pass the simulation (60% of 375).

Knockout step: 0.70 ÷ 0.75 = 0.93. No flag. Simulation step: 0.50 ÷ 0.60 = 0.83. Above 0.80, but close.

With samples this size, a 10-point gap in pass rates is very likely statistically significant even though it passes the four-fifths check. A careful analyst would look at the simulation's scoring, check whether accent is being scored when only clarity matters, and keep monitoring. This is exactly the case the Guidelines describe: the ratio passes, the statistics still deserve attention.

Limits of the 80% rule

The ratio is easy to compute and easy to misread.

• Small samples swing wildly. With 10 applicants per group, one person changes the ratio by 10 points or more. Use Fisher's exact test or similar for small numbers, and pool data across time periods where it is reasonable.
• Large samples hide significant gaps. With thousands of applicants, a ratio of 0.88 can be well beyond chance. Run a significance test alongside the ratio, typically a two-standard-deviation threshold.
• It flags, it doesn't decide. A ratio under 0.80 does not make a practice illegal. The employer can defend it by showing it is job-related and consistent with business necessity, supported by validation evidence.
• It covers race, sex and ethnicity in the Guidelines. Age claims under the ADEA use their own analysis, but running the same check for applicants 40 and over is a sensible habit.

Use in AI bias audits

The impact ratio has become the standard metric for auditing automated hiring tools. New York City's Local Law 144 requires a bias audit of automated employment decision tools, and the city's rules require reporting selection rates and impact ratios by sex, race/ethnicity and their intersections. For tools that produce scores rather than pass/fail outcomes, the rules use a scoring rate: the share of each group scoring above the median.

For any tool that screens or ranks candidates at scale, from resume parsers to chatbot knockout logic to AI interview scoring, compute impact ratios before launch on historical or pilot data, then monitor quarterly. High-volume funnels generate enough data to spot problems quickly, so there is little excuse for waiting until an annual audit.

Common mistakes

• Checking only final hires. Adverse impact usually hides at one step.
• Ignoring missing self-ID data. If 40% of applicants decline to identify, note it and check whether non-responders differ.
• Adjusting scores by group to fix the ratio. Race norming of test scores is prohibited by the Civil Rights Act of 1991.
• Running it once. Sourcing mix shifts by season and channel, so ratios drift.

This is general information about US federal guidance. Requirements differ for federal contractors and under state and local AI laws, so involve counsel when a ratio flags.