Disparate Impact Calculator
Whether a model or process selects one group at a lower rate than another — the disparate impact ratio, the four-fifths rule, the rate difference, and how many more selections would bring the groups into line.
Compare the selection rate for the group of concern with the rate for the group selected most often.
How the disparate impact calculator works
Compare the selection rate for the group of concern with the rate for the group selected most often. The ratio between them is the disparate impact ratio, and the long-standing four-fifths rule treats anything under 0.8 as evidence of adverse impact — not proof, but enough to have to explain.
It is the first fairness test to run on a hiring model, a credit model or any classifier whose decisions land on people, because it needs no ground truth: only who was selected.
Formula: ratio = rate_B / rate_A (A = the group with the higher rate); four-fifths rule: ratio ≥ 0.8
Worked examples
| Inputs | Disparate impact ratio | Note |
|---|---|---|
| 30% against 20% | 0.6667 | ratio 0.67 — fails |
| Within four fifths | 0.8667 | 0.87 — passes |
| Small samples | 0.6667 | fails the rule, but z is small |
FAQFrequently asked questions
What is disparate impact?
A practice that selects one group at a substantially lower rate than another, whatever the intent. The ratio of the two selection rates is the standard measure.
What is the four-fifths rule?
A rule of thumb from US employment guidelines: a selection rate below 80% of the highest group's rate is treated as evidence of adverse impact. Many fairness toolkits adopt the same threshold.
Is failing the rule proof of bias?
No — it is a screening test. Small samples fail it by chance, and a job-related, validated criterion can justify a gap. It means you have to look.
What does the z-score add?
A rough sense of whether the gap could be sampling noise. Below about 2 the difference is not statistically reliable; above it, it is unlikely to be chance.
Which metric should I use for a model?
This one first, because it needs no labels. Then equalised odds — comparing error rates between groups — which needs ground truth and catches unfairness this ratio misses.
Where these figures come from
- Vaswani et al. (2017) — Attention Is All You Need — the transformer architecture the memory arithmetic follows
- Kaplan et al. (2020) — Scaling Laws for Neural Language Models — the compute relationship used for training estimates
- Hoffmann et al. (2022) — Training Compute-Optimal Large Language Models — the tokens-per-parameter guidance ("Chinchilla")
- IEEE 754 — Standard for Floating-Point Arithmetic — the numeric formats behind bytes per parameter
- AI Safety Institute — the UK government AI body
Last checked: September 2026. The relationships here are architectural, not vendor-specific: bytes per parameter follow the numeric format, KV-cache size follows the transformer definition, and token-per-word ratios come from published tokeniser behaviour.