> Agent view (Markdown) of https://www.nycbiasaudit.com/blog/impact-ratio-calculation
> Publisher: Holistic AI (https://www.holisticai.com). Site index for agents: https://www.nycbiasaudit.com/llms.txt · Everything in one file: https://www.nycbiasaudit.com/llms-full.txt

# Selection rates and impact ratios, explained simply

Published: 2026-10-21 · By Holistic AI

The two numbers at the heart of every NYC Local Law 144 bias audit, explained with one small example: what a selection rate is, how an impact ratio works, and what 0.80 really means.

**In short:** A selection rate tells you how often a tool moves a group forward. An impact ratio compares each group to the group the tool favours most. Local Law 144 requires both, for every group, in every bias audit.

## Start with the selection rate

1. Take one group of candidates.
2. Count how many of them the tool moved forward.
3. Divide by how many there were in total.

That's the selection rate.

> **THE TWO FORMULAS** **Selection rate** = people moved forward ÷ people in the group. **Impact ratio** = a group's selection rate ÷ the highest group's selection rate.

## Then the impact ratio

Find the group with the highest selection rate.

Divide every other group's rate by that one.

The top group always scores 1.00. Everyone else lands somewhere below it.

## One small example

Say a resume screener reviewed two groups.

| Group | Selection rate | Impact ratio |
|---|---|---|
| Group A | 30% | 1.00 |
| Group B | 24% | 0.80 |

Group B's ratio is 24 divided by 30, which gives 0.80.

That's the whole calculation. Simple to do. Easy to do wrongly at scale.

## What does 0.80 mean?

Local Law 144 doesn't set a pass mark.

Most people use 0.80 as a reference point. It comes from the four fifths rule the Equal Employment Opportunity Commission has used for decades. Below it, a difference usually deserves a closer look.

A low ratio isn't automatically illegal. It is the kind of number a regulator or a plaintiff will notice, and your summary is public.

## What if the tool gives scores instead of decisions?

Then you use a scoring rate.

Find the median score for everyone.

For each group, work out the share of people who scored above it.

Then compare groups exactly as before.

## The part people miss

The rules ask for more than sex and race/ethnicity on their own.

They also ask for every combination of the two, such as women in one race/ethnicity category compared with men in another.

This matters because combined groups can show a gap that both simple views hide.

You also have to report how many people fell into an unknown category, even though they aren't part of the ratios.

## What your published summary has to show

- The date of the most recent bias audit
- Where the data came from, and why
- How many people were assessed, including the unknown category
- Selection or scoring rates for every group
- Impact ratios for every group
- The date you started using the tool

## Where audits go wrong

When New York's Comptroller reread a set of published audits, the calculation problems were familiar.

- Candidates grouped into tiers instead of being compared group by group
- Data sources that were never explained
- Testing run by the vendor rather than an independent auditor

None of it is complicated. It just needs care.

## Try it on your own numbers

Our [free bias check](/bias-check) runs these exact calculations in your browser. Nothing is uploaded. It's a quick way to see where you stand before a full independent audit.

*This article is general information, not legal advice. The example figures are illustrative.*

---

Holistic AI carries out independent Local Law 144 bias audits. What the service includes: https://www.nycbiasaudit.com/services.md. To commission a bias audit or ask a question, book a call at https://www.nycbiasaudit.com/demo or email we@holisticai.com.