Adverse Impact & AI Fairness Calculator
Run the four-fifths (80%) rule on selection rates, test whether the difference is statistically significant, and compare fairness metrics side by side.
Example scenarios
Group A
Any group you are comparing — the rule is undirected, so it does not matter which one you call A.
Group B
The second group. Whichever group has the lower selection rate is the one the ratio is measured on.
General information, not legal advice. Adverse impact analysis is fact-specific, and a finding here is a screening result rather than a determination. Consult qualified employment counsel before acting on it.
Related Calculators
You might also find these calculators useful
Accuracy, Precision & Recall Calculator
Classification metrics from a confusion matrix
AI Governance Maturity Model
Score nine NIST AI RMF topics and read your GOVERN-MAP-MEASURE-MANAGE profile
Model Drift Calculator
Test whether a model's accuracy drop is real or noise
Confidence Interval Calculator
Confidence interval for a mean or proportion — margin of error and steps
Run the four-fifths rule the way the regulation actually states it
Enter applicants and selections for two groups and this calculator returns the impact ratio, the four-fifths comparison, a two-proportion significance test, and the number of selections the lower-rate group is short. Most four-fifths calculators stop at the ratio. The regulation does not: it says smaller differences may still constitute adverse impact where they are statistically significant, and greater differences may not where they rest on small numbers. Both of those readings need a significance test, so both are computed here.
What is adverse impact and the four-fifths rule?
Adverse impact is a substantially different rate of selection that works to the disadvantage of a race, sex or ethnic group. The four-fifths rule, from the Uniform Guidelines on Employee Selection Procedures (29 CFR Part 1607), is the screening device federal enforcement agencies use: a selection rate less than four-fifths — 80% — of the rate for the group with the highest rate will generally be regarded as evidence of adverse impact. The comparison is against the highest-rate group, which is why the ratio here is undirected. The guidelines then qualify the screen in both directions, which is the part most tools leave out.
Impact ratio
Impact ratio = lower selection rate ÷ highest selection rate ≥ 0.80How to run an adverse impact analysis
Choose Selection data if you have applicant and selection counts, or Model predictions if you have a confusion matrix per group.
Enter the number of applicants considered and the number selected for each group. Count everyone the procedure was applied to, not only those who completed it.
Pick a significance level. 0.05 is the usual default; 0.01 is stricter and 0.10 more permissive.
Read the impact ratio against 80%, then read the p-value against your alpha. The highlighted row of the reference table tells you which of the four readings applies.
If the ratio is below 80%, check the shortfall figure to see how many selections separate the groups in practice.
Repeat the analysis for each component of a multi-part selection procedure, not only for the procedure as a whole.
Where adverse impact analysis is used
Hiring and promotion screens
Tests, interviews, education requirements and physical standards are all selection procedures under the Uniform Guidelines, and each component can be examined separately.
Reductions in force
Layoff selection is a selection procedure. Running the ratio and the significance test before a reduction is announced is standard practice.
Credit and lending decisions
Disparate impact analysis on approval rates is a routine part of fair-lending review, using the same ratio and the same two-proportion test.
Algorithmic screening tools
Where a model does the screening, the same test applies to its outputs — and the model-predictions mode adds the fairness metrics that describe how the errors fall.
Vendor and tool evaluation
Comparing selection rates across groups is the first check to run on any scoring tool before it goes into a decision process.
Monitoring over time
The guidelines suggest looking at impact over a longer period when counts are small. Running the same test each cycle is how a pattern becomes visible.
Why this calculator is built this way
The ratio alone is not the test
29 CFR 1607.4(D) qualifies the 80% screen twice — smaller differences can still be adverse impact when statistically significant, and larger ones may not be when based on small numbers. A calculator that returns only a ratio cannot express either sentence, so this one runs a two-proportion test alongside it.
Practical significance, in people
The guidelines ask for significance in practical as well as statistical terms. The shortfall figure says how many additional selections the lower-rate group would need to reach four-fifths parity — the gap expressed as people rather than as a percentage.
No combined fairness score
Demographic parity, equalized odds and predictive parity cannot all hold when two groups have different base rates. Averaging them into one number scores a model against a target no model can reach, so the metrics here are reported side by side and left for you to weigh.
Named for what it computes
The accuracy gap is called accuracy parity, not calibration. Calibration compares outcome rates at a given predicted score and needs the model's scores, which a confusion matrix does not carry.
Frequently asked questions
From 29 CFR 1607.4(D): a selection rate for any race, sex or ethnic group which is less than four-fifths (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 comparison is always against the highest-rate group, so the ratio does not depend on which group you enter first.
Because the rule itself requires it in both directions. The same paragraph says smaller differences may nevertheless constitute adverse impact where they are significant in both statistical and practical terms, and that greater differences may not constitute adverse impact where they are based on small numbers and are not statistically significant. A ratio on its own cannot express either statement.
Yes. With a large enough sample a gap of a few percentage points is measurable rather than accidental, and the guidelines explicitly contemplate that case. This calculator reports it as a distinct outcome rather than passing it silently.
It means the counts are too small to tell the difference from chance. The guidelines address this directly and suggest looking at the procedure's impact over a longer period, or where it has been used in the same way elsewhere, before drawing a conclusion.
The number of additional selections the lower-rate group would need for its rate to reach four-fifths of the higher rate, holding the other group fixed. It expresses the gap in people, which is what practical significance means.
0.05 is the common default. Use 0.01 when you want stronger evidence before calling a difference real, and 0.10 when missing a real difference is the costlier mistake. Choose it before you look at the result, not after.
Because the metrics are mutually incompatible. When two groups have different base rates, a model cannot have equal precision and equal false positive rates at the same time — this is the impossibility result of Chouldechova (2017) and Kleinberg, Mullainathan and Raghavan (2016). Averaging incompatible criteria produces a number that penalises every model for a trade-off none can escape. The metrics are shown side by side instead.
Because it is not calibration. Calibration asks whether, among everyone given the same predicted score, the outcome rate is the same across groups — that needs the model's scores. A confusion matrix only carries counts, so what can be computed from it is accuracy parity, and it is labelled accordingly.
No. The rule is a screening device used by enforcement agencies, not a safe harbour. It is applied to a selection procedure as a whole and, in the circumstances the guidelines describe, to its individual components. This tool gives general information, not legal advice.