What is Impact Ratio?

Why It Matters

The impact ratio is the closest thing employment AI has to a shared unit of account. It reduces a hiring process to a comparison: for each group, what share advanced, against the share that advanced most. A number near 1.00 means the groups advanced at similar rates; a number well below it means one group was filtered out more often, and that is a fact a regulator, a plaintiff, or a works council can act on.

Its virtue is that it can be computed from data an organisation already holds. Its danger is that it looks like a verdict when it is an observation.

How It Works

Take the selection rate for each group: the number who advanced divided by the number who applied. Then divide each group’s rate by the highest rate among the groups. The result is the impact ratio for that group.

The 0.80 reference point comes from the four-fifths rule in the EEOC Uniform Guidelines on employee selection procedures, where a selection rate below four-fifths of the highest rate is treated as evidence of adverse impact. New York City’s rules align the calculation with those guidelines and require the race, ethnicity and sex categories employers already report on, including intersectional combinations.

The ratio is evidence rather than a finding. Falling below the reference point invites scrutiny; clearing it does not close the question.

Where It Breaks

The first failure is sample size. A requisition that advanced twelve people produces ratios that swing on a single decision, and reporting them without the underlying counts invites conclusions the data cannot support.

The second is cadence. A ratio computed once a year describes a system that has since changed its criteria, its model version, or the roles it screens. Disparity is a property of a running process, and annual measurement is a lagging indicator of a problem that was live in March.

The third is the choice of denominator. The reference group is the most-selected category, so the ratio moves when that group’s rate moves, which means a system can appear to improve while the group at the bottom is unchanged.

The fourth is the proxy. A model can clear the reference point while a feature correlated with a protected characteristic does the deciding, which is why the ratio belongs alongside a disparate impact analysis of the features rather than standing in for it.

How Flytebit Handles It

We compute selection rates and impact ratios from the decisions the system actually influenced, with the criteria version and the reviewer attached, and we keep the underlying counts next to every ratio so the number can be read honestly. A ratio crossing the reference point triggers an examination of the criteria and the feature set rather than an explanation. The industry application is on our HR & Workforce Technology page, and the measurement approach is our evaluation work.