For employers
Four-fifths rule: how is adverse impact calculated for a hiring test, and when does the 80% line mislead?
The four-fifths rule compares the selection rates of groups of candidates. Under 29 CFR 1607.4(D), federal enforcement agencies generally regard a selection rate for any race, sex or ethnic group that is less than four-fifths (80%) of the rate for the group with the highest rate as evidence of adverse impact. The EEOC questions and answers of March 2, 1979 call it a rule of thumb, not a legal definition.

What is the four-fifths rule?
An enforcement rule of thumb. If one race, sex or ethnic group is selected at less than 80% of the rate of the group selected most often, federal agencies generally treat the difference as evidence of adverse impact.
The rule is paragraph (D) of 29 CFR 1607.4, a section of the Uniform Guidelines on Employee Selection Procedures (1978). A selection rate below four-fifths of the highest rate is generally treated by the federal enforcement agencies as evidence of adverse impact. A rate above four-fifths generally is not.
On March 2, 1979 the agencies published questions and answers on the Guidelines. Question 10 defines adverse impact as "a substantially different rate of selection in hiring, promotion or other employment decision" to the disadvantage of members of a race, sex or ethnic group.
Question 11 explains the threshold. The agencies adopted the 80% figure as a practical way to keep attention on serious discrepancies in selection rates, and the answer says: "This "4/5ths" or "80%" rule of thumb is not intended as a legal definition."
The rule covers more than paper tests. In questions and answers on Internet hiring published on March 4, 2004, the EEOC, OFCCP, the Justice Department and OPM wrote in Question 98 that "Online tests, including tests of specific or general skills, are selection procedures" under the Guidelines, and that employers using them should keep records of their impact.
How do you calculate adverse impact step by step?
For each group, divide the number selected by the number of applicants. Find the highest rate, divide every other rate by it, and check whether any ratio falls below 80%.
Question 12 of the EEOC answers sets out four steps. First, the selection rate of each group: the number of people selected divided by the number who applied. Second, the group with the highest rate. Third, the impact ratio, which is each group's rate divided by the highest rate. Fourth, a comparison of each ratio with 80%.
The worked example in Question 12 has 80 white applicants, of whom 48 were hired, a rate of 60%, and 40 Black applicants, of whom 12 were hired, a rate of 30%. The impact ratio is 30 divided by 60, one-half. Because 50% is below four-fifths, adverse impact is usually indicated.
Question 11 runs the test with four groups. Whites other than Hispanics were hired at 60%, American Indians at 45%, Hispanics at 48% and Blacks at 51%. Against 60%, the ratios are 75%, 80% and 85%. Only the American Indian ratio is below four-fifths, so adverse impact is indicated for American Indians and not for Hispanics or Blacks. The example assumes that each group makes up more than 2% of the labor force.
Under Question 13, the first calculation covers the total selection process for each job. If that process shows adverse impact, the rates are then calculated for each component, such as a written test or an interview. Question 14 counts as the total process every step that leads to the decision, plus the combined results when several routes lead to the same job.
Who counts as an applicant, and which groups are compared?
An applicant is a person who has indicated an interest in being considered for the job. Rates are compared by sex and by race or ethnic group, for each group that is at least 2% of the relevant labor force.
Question 15 says: "The concept of an applicant is that of a person who has indicated an interest in being considered for hiring". The interest may be shown on a form or orally. The term candidate covers current employees who are considered for promotion or training without filing an application. A person who withdraws voluntarily is no longer counted as an applicant, but the records on that person are kept.
For online recruiting, the 2004 notice in the Federal Register (69 FR 10152) lists three conditions in Question 96. The employer has acted to fill a particular position, the individual has followed the employer's standard procedures for submitting applications, and the individual has indicated an interest in that particular position. A resume posted to a general database does not by itself make a person an applicant. Under Question 97, all search criteria used to screen such resumes are subject to disparate impact analysis.
Paragraph (B) of 1607.4 requires records by sex and by these groups: Blacks, American Indians (including Alaskan Natives), Asians (including Pacific Islanders), Hispanics, and whites other than Hispanic. The categories match the EEO-1 report. Question 17 compares men with women and each race or ethnic group with the others. It says there is "no obligation to make comparisons for subgroups" such as white women or Black men, although records on subgroups are still kept.
Question 16 calls for annual adverse impact determinations for each group that makes up 2% or more of the total labor force in the relevant labor area or of the applicable workforce. For hiring, groups above 2% of the applicants are added. For promotions, groups of at least 2% of the employer's workforce are added. Records are kept for all groups, including those below 2%.
What if only a few people took the test?
A ratio below 80% may then carry no weight. When the numbers are small and a difference could be due to chance, the agencies do not assume adverse impact without other evidence and may look at a longer period.
Question 20 gives the case of an employer that selected 3 of 20 male applicants and 1 of 10 female applicants. The rates are 15% and 10%, and the ratio is 66 2/3%, below four-fifths. The answer states that the number of selections is too small to warrant a finding of adverse impact.
Under Question 21, when a difference could easily be the result of chance, or when one different selection would reverse it, "the Federal agencies will not assume the existence of adverse impact, in the absence of other evidence". A pattern that repeats over several years can change that conclusion.
Paragraph (D) of 1607.4 gives the agencies two more sources when the data are too small to be reliable: results of the same procedure over a longer period, and results of the same procedure used in similar circumstances elsewhere. If the employer did not keep the required data, the agencies may infer adverse impact when a group is underutilized compared with its share of the relevant labor market or, for promotions, of the workforce.
Manish Raghavan, assistant professor at the MIT Sloan School of Management, made the same point to the EEOC on January 31, 2023: "it would be hard to draw conclusions if an employer has only hired 3 people". The same gap in rates carries more weight in 300 hires.
Can a ratio above 80% still be adverse impact?
Yes. Under 1607.4(D) a smaller difference counts when it is significant in both statistical and practical terms, or when the employer's actions discouraged applicants of one group disproportionately.
Question 20 has an example with a rule that rejects applicants with arrest records. The rule would disqualify 10% of Hispanic applicants and 4% of other applicants, so the selection rates are 90% and 96%. Ninety is about 94% of 96, above four-fifths. Hispanic applicants are still 2.5 times as likely to be disqualified. Where nationwide statistics show the difference to be significant in both statistical and practical terms, the agencies would consider it adverse impact.
Question 19 says that the 4/5ths rule of thumb "speaks only to the question of adverse impact" and does not permit a 20% level of discrimination. Unlawful discrimination may be proved by other evidence even when the ratio is above 80%.
Question 22 deals with large samples. When many people are selected, "relatively small differences in selection rates may nevertheless constitute adverse impact", and a test of statistical significance is appropriate when the ratio is close to 80%. Question 23 adds that a ratio below 80% usually means adverse impact, except when both the number selected and the difference in rates are very small.
Do courts apply the four-fifths rule or a test of statistical significance?
Courts weigh the numbers case by case. In Watson v. Fort Worth Bank & Trust (1988), Justice O'Connor's plurality opinion said the rule had not provided more than a rule of thumb for the courts.
Watson v. Fort Worth Bank & Trust was decided on June 29, 1988. Footnote 3 of Justice O'Connor's plurality opinion describes the four-fifths rule as an enforcement rule under which adverse impact will not ordinarily be inferred above the 80% line, and says "it has not provided more than a rule of thumb for the courts".
The same footnote turns to statistics: "we have not suggested that any particular number of 'standard deviations' can determine" whether a disparity is significant. Courts "appear generally to have judged the 'significance' or 'substantiality' of numerical disparities" one case at a time, and the plurality left it at that.
The Guidelines chose the ratio over significance tests on purpose. Question 24 explains that "the Guidelines do not rely primarily upon a test of statistical significance" because in small samples even large real differences may not reach conventional significance levels. Under Question 18, significance tests are used mainly when many people are selected.
Raghavan put the difference in his 2023 testimony this way: "The four-fifths rule of thumb considers only effect size, not statistical significance." According to him, employers in practice use "a suite of formal statistical tests" and do not rely on the rule alone.
If the whole process passes, does each step pass too?
For agency enforcement, usually yes under 1607.4(C). In a Title VII lawsuit, no: in Connecticut v. Teal (1982) the Supreme Court held that a nondiscriminatory bottom line does not answer a claim against one test.
Paragraph (C) of 1607.4 is the bottom line rule. If the total selection process has no adverse impact, the agencies generally will not expect the employer to evaluate each component and will not take enforcement action based on a component. There are two exceptions: a procedure that significantly contributes to continuing patterns of job assignments caused by past discrimination, and a procedure, such as a height or weight requirement or a no-arrest rule, that courts or administrative interpretations have held not job related in similar circumstances. Under Question 26, the bottom line concept applies only to enforcement actions defined in Section 16 of the Guidelines and not to the EEOC's processing of individual charges.
In Connecticut v. Teal, 48 Black candidates and 259 white candidates took a written examination for promotion. Twenty-six Black candidates passed, 54.17%, and 206 white candidates passed, 79.54%. The Black pass rate was about 68% of the white rate. At the end of the process the state promoted 11 of the 48 Black candidates (22.9%) and 35 of the 259 white candidates (13.5%).
On June 21, 1982 the Supreme Court ruled 5 to 4 for the employees who failed the examination. Justice Brennan wrote that the bottom line "does not preclude respondent employees from establishing a prima facie case" and that a "nondiscriminatory 'bottom line' is no answer, under the terms of Title VII". Justice Powell dissented, joined by Chief Justice Burger and Justices Rehnquist and O'Connor.
Impact on one job can also count when the whole process has none. In an informal discussion letter of February 3, 2005 on a physical ability test, EEOC staff wrote that "the need for validation can be triggered by the existence of an adverse impact for a particular job", citing Question 27. The letter states that it does not constitute an official opinion of the Commission.
What does an employer do after finding adverse impact?
It validates the procedure for the job or replaces it with a lawful alternative that removes the impact. Between procedures of substantially equal validity, the one with less adverse impact should be used.
Under Question 9, evidence of validity is required only when a procedure adversely affects a race, sex or ethnic group. Question 41 says the Guidelines "generally call for a validity study or the elimination of adverse impact". Under Question 31, an employer may switch to lawful alternative procedures that eliminate the adverse impact instead of validating the original one, and some employers do so while they work on a validation study.
Question 50 states that an alternative procedure with less adverse impact should be used when its validity is substantially the same or greater. A test bought from a vendor does not move the duty elsewhere: under Question 35, a vendor's validity report may be used, but "it is the user's responsibility" to confirm that the evidence meets the Guidelines. Question 44 lets an employer use a procedure developed elsewhere if its use for the particular job is consistent with the Guidelines.
The burden of proof under Title VII, the types of validation and the Ricci v. DeStefano decision on setting aside test results are covered in the article Is pre-employment testing legal.
Does the four-fifths rule work for AI screening tools?
At an EEOC meeting on January 31, 2023, Manish Raghavan of MIT testified that the rule measures only effect size, that a model picking candidates at random can pass it, and that results change with the score cutoff.
Raghavan holds a PhD in computer science from Cornell University and studies machine learning in employment. In his written testimony he noted that passing the four-fifths test says nothing about whether a model predicts what it claims to predict, because a model that selects candidates at random can pass it.
His second point concerns cutoffs. A model gives scores on a scale of 0 to 1. Selection rates for two groups may be close at a cutoff of 0.5 and far apart at 0.75, so "the choice of threshold can affect whether a model passes or fails a statistical test on a given dataset." He recommended that developers "test a model across a range of possible thresholds" when the cutoff used in practice is not known in advance.
His third point is differential validity: the four-fifths rule of thumb "does nothing to prevent differential validity." A model may be more accurate for one group than for another and still pass. He proposed that developers search for models with similar accuracy and smaller gaps in selection rates, a search that, in his words, "should become standard practice for model developers."
He also warned about prospective audits. Their result "depend[s] heavily on the data collected to perform that test": an employer could add applicants of one group to the test data, or collect a small dataset in which a gap is hard to detect.
What does the IAE exam give an employer for its own adverse impact analysis?
For each person it invites, the company sees the status, the total score out of 100 and the rubric criterion scores. The pass mark is 70. Selection rates by group are computed by the employer from its own records.
The company registers on the employer page and receives a link to the company cabinet within one business day. In the cabinet it adds a person with a name, an email and an exam field. The person takes the exam on any of 14 days, in one sitting of up to 2 hours: 5 questions of 10 minutes, 2 cases of 25 minutes and a 10-minute question about their own project. The result, "passed" or "not passed" with scores by section, comes within 5 business days and is approved by an examiner with no ties to the employee or the company.
All candidates in a field take the exam under one program, and the program, sample tasks and scoring rubric are published before the exam. Each candidate gets their own figures and circumstances in the cases. Pasting is disabled, a typing log records how each answer was typed, and a doubtful answer brings one more written question with 5 minutes to answer.
The company sees the status, the total score, the scores on each rubric criterion (correctness, reasoning, risks, clarity) in each task and the certificate PDF if the exam is passed. It does not see the questionnaire, the CV, the answers or the reason for refusal, which appears as Not admitted. The records by sex and by race and ethnic group described in 1607.4(B) stay in the employer's own applicant files, and the employer decides on the interview and on hiring.
A candidate with a disability receives up to 50% extra time for each task, or breaks, on a written request with a supporting document submitted before booking. A result may be appealed within 30 days.
Sources
- 29 CFR 1607.4, Information on impact (LII)
- EEOC, Questions and Answers to Clarify and Provide a Common Interpretation of the Uniform Guidelines on Employee Selection Procedures
- Federal Register, March 4, 2004, Additional Questions and Answers on the Uniform Guidelines as they relate to the Internet (69 FR 10152)
- Watson v. Fort Worth Bank & Trust, 487 U.S. 977 (1988) (LII)
- Connecticut v. Teal, 457 U.S. 440 (1982), U.S. Reports (govinfo)
- EEOC informal discussion letter, Title VII: UGESP, February 3, 2005
- EEOC meeting of January 31, 2023, written testimony of Manish Raghavan
Read also
All articles
Does an employer have to pay for required training, certification exams and study time?
Under the Fair Labor Standards Act the hours an employee spends in training are working time unless four conditions are met together: the training
NYC Local Law 144: when does an AI hiring tool need a bias audit, and what must the employer publish and tell candidates?
Since January 1, 2023, New York City has barred employers and employment agencies from using an automated employment decision tool to screen
Are personality tests legal in hiring, and when does the ADA treat a personality test as a medical examination?
The ADA bars medical examinations of job applicants before a conditional offer, and a psychological test that can identify a mental disorder isPilot until December 31, 2026
Register your company
Free for up to 5 people. A link to the company cabinet arrives by email within one business day.

