Explainer: What is Disparate Treatment?

Disparate impact is accidental discrimination. Disparate treatment is deliberate - and the law treats them very differently.

The One-Sentence Definition

Disparate treatment is intentional discrimination: a decision-making process that treats an individual differently because of a protected characteristic (race, gender, age, religion, national origin, etc.) - regardless of whether that characteristic is named explicitly or encoded through a feature that serves as a stand-in for it.

Why This Matters

Most algorithmic fairness work focuses on disparate impact - unequal outcomes produced by a model that never explicitly uses a protected attribute. Disparate treatment is the other half of discrimination law, and it is the half that is harder to escape in court.

Where disparate impact asks "did outcomes differ across groups?", disparate treatment asks "was the protected attribute a factor in the decision?" A model that directly includes Gender, Race, or Age as input features is, by definition, engaging in disparate treatment - even if the resulting outcome gap is small. The intent is embedded in the design.

This matters in practice because:

Including a protected attribute as a model feature is per se disparate treatment under Title VII, the Equal Credit Opportunity Act (ECOA), and the Fair Housing Act - regardless of whether you can measure a downstream gap
Courts distinguish between the two doctrines sharply: disparate impact claims require statistical evidence of outcome disparity; disparate treatment claims require evidence that the protected characteristic was used, not just that outcomes differed
Every biased model in this repo was first trained with protected attributes included - that is, every unfair.py script is a disparate treatment example before it becomes a disparate impact one

Disparate Treatment vs Disparate Impact - The Distinction That Matters

These two doctrines are often confused. They are different legal claims with different burdens of proof and different fixes.

	Disparate Treatment	Disparate Impact
Definition	Protected attribute was used as a factor in the decision	Neutral process produces unequal outcomes across groups
Intent required?	Yes - the attribute must have been used, deliberately or by design	No - no intent required; the outcome gap is the evidence
Legal standard	Title VII §703(a) - treating someone differently because of race, sex, etc.	Title VII §703(k) - neutral practice with unjustified disparate effect
Proof	Show the attribute was an input or a determinative factor	Show the statistical outcome gap (four-fifths rule, etc.)
Defence	Bona Fide Occupational Qualification (BFOQ) - narrow, rarely succeeds	Business necessity + job-relatedness
Algorithmic form	Protected attribute or known proxy included as a model feature	Protected attribute excluded but proxies remain; outcome gap persists
Fix	Remove the protected attribute from the feature set	Remove the protected attribute and its proxy variables
Metric	Feature presence audit - is it in the model at all?	Disparate impact ratio, demographic parity gap

The critical asymmetry: you can have disparate treatment without disparate impact (the attribute is used but outcomes happen to be equal), and disparate impact without disparate treatment (no protected attribute was used but proxies recreated the gap). In practice, disparate treatment usually produces disparate impact - but the legal exposure exists independently.

Real-World Proof: Every `unfair.py` in This Repo

Every biased model in this repo is a disparate treatment example. The protected attribute is a direct model input. The outcome gap is the downstream disparate impact. They are the same model - looked at through two different legal lenses.

COMPAS - `unfair.py`

# DISPARATE TREATMENT: Race is a direct model feature
X = pd.get_dummies(df[[
    'race',            # ← protected attribute, fed directly to the classifier
    'Sex_Code_Text',
    'CustodyStatus',
    'MaritalStatus'
]])

Race is an explicit input. The model is permitted to use race as a predictive signal. That is disparate treatment. The resulting 86.77pp gap between Black and White defendants is the disparate impact it produces.

AI Fair Recruitment - `unfair.py`

# DISPARATE TREATMENT: Gender and Age are direct model features
features = ['Gender', 'Age', 'Experience_Years', 'Technical_Test_Score',
            'Education_Level', 'Previous_Companies', 'Distance_from_Company']

Gender is a direct input. Age is both an input and a proxy for gender (women in the dataset more often have career gaps, so age encodes gender signal twice over - once directly, once through correlation). The model was designed to see these attributes. The 20.9pp hire rate gap is the disparate impact.

The Fix - What Removing Disparate Treatment Looks Like

# fair.py: protected attribute and its proxy removed
features = ['Experience_Years', 'Technical_Test_Score']
# Gender removed ✓  - eliminates disparate treatment
# Age removed ✓     - eliminates proxy-based disparate treatment

Removing the protected attribute ends the disparate treatment. Whether it also ends the disparate impact depends on whether proxies remain - which is why proxy variable analysis is a required step in every audit. See proxy-variables.md.

Detection Code

Audit whether protected attributes are in the feature set

import pandas as pd

# Standard protected attributes under US federal law
PROTECTED_ATTRIBUTES = {
    'race', 'ethnicity', 'color', 'national_origin', 'origin',
    'sex', 'gender', 'pregnancy',
    'age',
    'religion',
    'disability',
    'marital_status',       # ECOA
    'family_status',        # Fair Housing Act
}

def disparate_treatment_audit(feature_columns, protected=PROTECTED_ATTRIBUTES):
    """
    Flag any feature that is a protected attribute or contains one as a substring.
    Returns a list of flagged column names.
    """
    flagged = []
    cols_lower = {col: col.lower().replace(' ', '_').replace('-', '_')
                  for col in feature_columns}

    for original, normalised in cols_lower.items():
        for attr in protected:
            if attr in normalised:
                flagged.append(original)
                break

    if flagged:
        print(f"⚠ Disparate treatment risk - protected attributes in feature set:")
        for f in flagged:
            print(f"  · {f}")
    else:
        print("✓ No protected attributes detected in feature set.")
        print("  Run a proxy variable audit next - see proxy-variables.md")

    return flagged


# Example - AI Fair Recruitment biased model
features_unfair = ['Gender', 'Age', 'Experience_Years', 'Technical_Test_Score',
                   'Education_Level', 'Previous_Companies', 'Distance_from_Company']

disparate_treatment_audit(features_unfair)
# ⚠ Disparate treatment risk - protected attributes in feature set:
#   · Gender
#   · Age

features_fair = ['Experience_Years', 'Technical_Test_Score']
disparate_treatment_audit(features_fair)
# ✓ No protected attributes detected in feature set.
#   Run a proxy variable audit next - see proxy-variables.md

Check a trained sklearn model's feature list

from sklearn.pipeline import Pipeline

def audit_sklearn_model(model, feature_names, protected=PROTECTED_ATTRIBUTES):
    """
    Given a fitted sklearn model (or Pipeline) and its input feature names,
    flag any protected attributes present in the training features.
    """
    print("=== Disparate Treatment Audit ===")
    print(f"Features used: {list(feature_names)}\n")
    flagged = disparate_treatment_audit(feature_names, protected)

    if flagged:
        print(f"\nThis model engaged in disparate treatment.")
        print(f"Remove flagged features and recheck for disparate impact via")
        print(f"proxy variable audit before redeployment.")
    return flagged


# Example
import pandas as pd
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split

df = pd.read_csv('AI_Fair_Recruitment_Dataset.csv')
feature_cols = ['Gender', 'Age', 'Experience_Years', 'Technical_Test_Score']
X = pd.get_dummies(df[feature_cols])
y = (df['HiringDecision'] == 1).astype(int)

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
model = RandomForestClassifier(n_estimators=100, random_state=42)
model.fit(X_train, y_train)

audit_sklearn_model(model, feature_cols)

The full two-stage check: treatment → impact

def full_discrimination_audit(feature_columns, y_true, y_pred, group_series):
    """
    Stage 1: Disparate treatment - are protected attributes in the feature set?
    Stage 2: Disparate impact   - does the outcome gap exceed the four-fifths threshold?

    Run both regardless: treatment can exist without impact, and impact without treatment.
    """
    print("=== Stage 1: Disparate Treatment ===")
    flagged = disparate_treatment_audit(feature_columns)

    print("\n=== Stage 2: Disparate Impact ===")
    results = pd.DataFrame({'y_pred': y_pred, 'group': group_series})
    rates = results.groupby('group')['y_pred'].mean()
    ratio = rates.min() / rates.max()
    passes = ratio >= 0.80

    print("Positive prediction rates by group:")
    print(rates.round(4).to_string())
    print(f"\nDisparate impact ratio: {ratio:.3f}  "
          f"{'✓ passes' if passes else '✗ fails'} four-fifths rule")

    return {'disparate_treatment': flagged, 'di_ratio': round(ratio, 3), 'passes_eeoc': passes}

The Proxy Problem: Indirect Disparate Treatment

Removing the protected attribute stops direct disparate treatment. It does not stop indirect disparate treatment - where a correlated feature serves as a stand-in for the protected attribute the model is no longer permitted to see.

Courts have recognised this. In Watson v. Fort Worth Bank & Trust (1988), the Supreme Court held that facially neutral criteria that serve as proxies for protected characteristics can constitute disparate treatment - not just disparate impact - when the intent to use them as substitutes can be shown.

In algorithmic systems, proving intent is difficult. But the structure of indirect disparate treatment is mechanical and detectable:

from scipy.stats import chi2_contingency

def proxy_treatment_check(df, feature_col, protected_col, threshold=0.05):
    """
    Test whether a feature is statistically associated with a protected attribute.
    A significant result (p < threshold) indicates potential indirect disparate treatment
    if the feature is included despite known correlation.
    """
    ct = pd.crosstab(df[feature_col], df[protected_col])
    chi2, p, dof, expected = chi2_contingency(ct)
    flag = p < threshold

    print(f"{feature_col} ~ {protected_col}:  χ²={chi2:.2f}, p={p:.4f}  "
          f"{'⚠ correlated - proxy risk' if flag else '✓ not significantly correlated'}")
    return flag

# Example - German Credit Lending: employment tenure as age proxy
proxy_treatment_check(df, feature_col='employment', protected_col='age_class')

If a feature is significantly correlated with the protected attribute and you knowingly include it in the model, the case for indirect disparate treatment strengthens. The safest position is to document the decision either way. See proxy-variables.md for the full proxy detection methodology.

Limitations

Intent is legally required but technically invisible. Disparate treatment doctrine requires that the protected characteristic was used as a factor - which in ML terms means it was in the feature set, or a proxy was included with knowledge of its correlation. A model that produces equal outcomes using race as a feature technically engaged in disparate treatment even though the outcome was fair. The law cares about the process, not just the result.

The BFOQ defence is narrower than most employers assume. The Bona Fide Occupational Qualification exception - which allows using a protected attribute if it is genuinely necessary for the job - is extremely narrow under Title VII. Sex is permissible as a criterion for a role in a women's shelter. Age is permissible for an airline pilot under FAA safety rules. Race is almost never a permissible BFOQ. Do not rely on BFOQ as a general defence for including protected attributes in a model.

Removing protected attributes does not eliminate all legal exposure. A model with no protected attributes in its feature set can still produce a disparate impact ratio that fails the four-fifths rule - because proxy variables remain. Clearing the disparate treatment audit is the first step, not the last. Always follow it with a disparate impact audit and a proxy variable check.

Intersectionality is not captured by single-attribute audits. Auditing for Gender alone and Race alone can both pass while a model systematically discriminates against Black women specifically. Run the feature audit and the impact audit across all relevant intersecting combinations where sample sizes allow.

Disparate Impact

Disparate impact is the outcome-based counterpart - a facially neutral process that produces unequal results across groups. You can have one without the other, but in practice biased models produce both. The fix for disparate treatment (remove the attribute) is necessary but not sufficient to fix disparate impact. See disparate-impact.md.

Proxy Variables

The mechanism by which disparate treatment persists after the protected attribute is removed. If you drop Gender but keep Age - which correlates with gender in the dataset - the model reconstructs gender signal from age. That is proxy-based disparate treatment. See proxy-variables.md.

Demographic Parity

The fairness metric that directly measures the downstream outcome of disparate treatment: equal positive prediction rates across groups. A model free of disparate treatment may still fail demographic parity if proxies remain. See demographic-parity.md.

AI Fair Recruitment/ - unfair.py includes Gender and Age directly; fair.py removes both. The cleanest before/after disparate treatment example in the repo.
COMPAS/ - unfair.py includes race and CustodyStatus (a race proxy) directly. Removing both in fair.py ends both the disparate treatment and reduces the disparate impact gap from 86.77pp to 15.69pp.
German Credit Lending/ - unfair.py includes age and employment (an age proxy). The proxy relationship is documented in the notebook.
explainers/disparate-impact.md - the outcome-based counterpart; the four-fifths rule and how it is measured
explainers/proxy-variables.md - how indirect disparate treatment survives attribute removal and how to detect it
explainers/demographic-parity.md - the downstream metric that disparate treatment violations typically produce

Disparate Treatment