Machine learning models help make decisions in lending, hiring, admissions, and fraud review. Accuracy and AUC matter, but they do not reveal whether outcomes differ across demographic groups. Fairness metrics provide formal definitions that teams can compute, report, and monitor. In workshops and case studies from a data scientist course in Delhi, these metrics are often introduced alongside confusion matrices because both are needed for responsible evaluation.
Statistical Parity (Demographic Parity)
Statistical parity—often called demographic parity—is an outcome-based definition. It requires that the rate of positive predictions be the same across groups. Let A be a protected attribute (for example, gender), Ŷ be the predicted label, and Ŷ = 1 represent a positive decision (approved, eligible, shortlisted). The formal condition is:
P(Ŷ = 1 | A = a) = P(Ŷ = 1 | A = b)
for any two groups a and b.
Common summaries are:
- Parity difference: P(Ŷ=1|A=a) − P(Ŷ=1|A=b)
- Parity ratio: P(Ŷ=1|A=a) / P(Ŷ=1|A=b)
Demographic parity is easy to interpret because it speaks directly about who receives positive outcomes. Its limitation is that it ignores the true label Y. If the base rate of Y = 1 differs between groups, equalising selection rates can raise errors in at least one group. In practice, teams sometimes treat demographic parity as a policy constraint (for example, “keep parity difference within a bound”) rather than an absolute requirement.
Equal Opportunity
Equal opportunity focuses on how the model treats truly qualified cases. It requires equal true positive rates (TPR) across groups. Using Y as the ground-truth label:
P(Ŷ = 1 | Y = 1, A = a) = P(Ŷ = 1 | Y = 1, A = b)
This means: among individuals who actually qualify (Y = 1), each group should have the same chance of receiving a positive prediction. Equal opportunity is often chosen when false negatives are the main harm, such as denying credit to borrowers who would repay.
Operationally, compute a confusion matrix per group and compare:
TPR = TP / (TP + FN)
If one group has a lower TPR, the model misses more qualified cases in that group. Unlike demographic parity, equal opportunity does not require equal overall selection rates; it allows selection rates to differ if base rates differ.
Disparate Impact and the 80% Rule
Disparate impact is commonly used as a compliance-style check based on selection rates. It compares an advantaged and disadvantaged group using a ratio:
DI ratio = P(Ŷ = 1 | A = disadvantaged) / P(Ŷ = 1 | A = advantaged)
A widely used screening threshold is the “four-fifths rule” (80% rule): DI ratio ≥ 0.8. If the ratio is below 0.8, the disadvantaged group’s selection rate is less than 80% of the advantaged group’s rate, signalling the process should be reviewed.
Passing the 80% rule does not guarantee fairness in other ways. A model can meet this ratio while still having unequal error rates. Treat disparate impact as an early warning indicator, not a complete fairness verdict.
Using Fairness Metrics in Practice
A practical fairness review can be embedded into standard validation:
- Define what Ŷ = 1 means and which mistakes matter most (false positives vs false negatives).
- Choose a fairness target that matches policy: demographic parity for equal access, equal opportunity for equal treatment of qualified cases, and disparate impact for quick screening.
- Compute metrics per group on a held-out test set. Report selection rate, TPR, and sample sizes so gaps are interpretable and not driven by tiny groups.
- Inspect threshold effects. If the model outputs scores, changing the decision threshold can shift both accuracy and fairness gaps. If gaps remain unacceptable, consider mitigation such as reweighting data, improving features, or fairness-aware training and post-processing (with appropriate governance).
Teams often practise these steps in applied exercises during a data scientist course in Delhi to learn how metric choice changes conclusions.
Conclusion
Statistical parity (demographic parity) targets equal positive outcome rates across groups. Equal opportunity targets equal true positive rates among those who truly qualify. Disparate impact compares selection rates using a ratio, often checked against the 80% rule. Because these metrics represent different fairness goals, teams should pick the one that matches the decision context and monitor it over time. The habit you build—whether in a data scientist course in Delhi or in production reviews—makes fairness measurable rather than aspirational.