Disparate Impact Calculator for AI Fairness Audits

Calculator Inputs

Audit Name

Use a clear label for your fairness review.

Privileged Group Label

Example: Male, Group A, Existing Segment.

Unprivileged Group Label

Example: Female, Group B, Target Segment.

Privileged Group Total

Total records, applicants, or predictions reviewed.

Privileged Group Selected

Approved, hired, passed, or positively predicted.

Unprivileged Group Total

Total records for the comparison group.

Unprivileged Group Selected

Positive outcomes for the comparison group.

Impact Ratio Threshold

Use 0.80 for the four-fifths rule benchmark.

Example Data Table

Audit Name	Privileged Group	Privileged Total	Privileged Selected	Unprivileged Group	Unprivileged Total	Unprivileged Selected	Impact Ratio
Resume Screening	Reference Group	200	120	Protected Group	180	72	0.667
Interview Shortlist	Reference Group	150	78	Protected Group	140	67	0.920
Loan Approval Model	Segment A	320	176	Segment B	290	130	0.814

This sample illustrates how the calculator can compare positive outcomes across groups in hiring, lending, admissions, and risk scoring workflows.

Formula Used

Selection Rate
Selection Rate = Selected Cases / Total Cases

Disparate Impact Ratio
Disparate Impact = Unprivileged Selection Rate / Privileged Selection Rate

Selection Rate Gap
Rate Gap = Privileged Selection Rate - Unprivileged Selection Rate

Expected Unprivileged Selections at Parity
Expected Selections = Privileged Selection Rate × Unprivileged Total

Selection Shortfall
Shortfall = Expected Unprivileged Selections - Actual Unprivileged Selections

Four-Fifths Rule Interpretation
If the impact ratio is below 0.80, many audits flag potential adverse impact for closer review.

Two-Proportion Z Test
z = (Unprivileged Rate - Privileged Rate) / Standard Error
This helps estimate whether the observed difference may be statistically meaningful.

How to Use This Calculator

Enter a descriptive audit name and label both comparison groups clearly.
Provide total cases reviewed for each group.
Enter selected or positive outcomes for both groups.
Set the impact threshold, usually 0.80 for the four-fifths rule.
Click the calculate button to generate ratios, gaps, significance checks, and the Plotly chart.
Download the result summary as CSV or PDF for documentation, governance logs, or stakeholder review.

Why This Metric Matters in AI and Machine Learning

Disparate impact analysis helps teams evaluate whether a model, workflow, or screening rule creates materially different positive outcome rates across groups. It is widely used in fairness reviews for hiring systems, credit decisioning, education screening, fraud controls, and automated ranking. The metric does not diagnose root cause by itself, but it quickly highlights where deeper analysis is needed.

For a stronger governance process, combine this calculator with confusion matrix checks, calibration review, subgroup performance analysis, data quality inspection, and policy review. Fairness is multidimensional, so impact ratio should be interpreted with context, sample size, legal requirements, and domain expertise.

Frequently Asked Questions

1) What is disparate impact?

It compares positive outcome rates between groups. A lower ratio can indicate that one group receives approvals, selections, or passes less often than another.

2) What does the 0.80 threshold mean?

It reflects the four-fifths rule. If the unprivileged group rate is below 80% of the privileged group rate, many reviews flag potential adverse impact.

3) Does a low ratio prove discrimination?

No. It signals potential risk and the need for further investigation. Root causes may involve policy design, data imbalance, sampling, or process errors.

4) Can I use this for model predictions?

Yes. Treat positive predictions, approvals, or recommended actions as selected outcomes. The calculator works for screening, ranking, approval, and classification workflows.

5) Why include a z score and p value?

They add a statistical view of rate differences. They help judge whether an observed gap may be more than random variation, especially with larger samples.

6) What is selection shortfall?

It estimates how many additional unprivileged selections would be needed to match the privileged group’s selection rate, given the same unprivileged sample size.

7) Should I rely on one fairness metric alone?

No. Combine impact ratio with accuracy, false positive rates, false negative rates, calibration, and subgroup error analysis for a stronger fairness assessment.

8) When is this calculator most useful?

Use it during pre-deployment review, model monitoring, policy updates, vendor assessment, or when auditing automated decisions for hiring, lending, admissions, or access control.