Removing the protected attribute from a model's inputs is the single most common first response to a fairness complaint - and on real data, it can leave the gap unchanged, or make it worse.
Fairness through unawareness is the intuition that a model is fair simply because it does not use a protected attribute (race, sex, age) as an input - and it fails whenever other features carry enough correlated information to reconstruct what was removed.
Dropping the protected attribute feels like the obvious fix, and it's usually the first thing anyone tries: no race column, no racial bias. The reasoning has an intuitive appeal that makes it persistent even after it's been shown not to hold - "the model literally cannot see race" sounds like a complete argument.
It isn't, because a protected attribute rarely travels alone. Zip code correlates with race through housing patterns. Name correlates with sex and ethnicity. Employment history and arrest records correlate with race through decades of unequal enforcement and access. A model doesn't need the protected attribute itself if enough of its correlates are still sitting in the feature set - it can reconstruct the same decision boundary from the pieces left behind. This is exactly the mechanism Proxy Variables describes, and it's why "fairness through unawareness" is treated in the fairness literature as a well-documented failure mode, not a live debate.
This repo's own benchmark harness names this exact strategy unawareness (S1 in the mitigation strategies ladder) and runs it on every audit specifically to measure how much it actually helps - which, as the real numbers below show, is not a fixed or guaranteed amount.
The demographic parity gap for the baseline logistic regression model on race (disadvantaged: Black applicants, n=2,943; advantaged: White applicants, n=2,224), using the frozen numbers exactly as faircode/benchmark.py computed them:
| Strategy | Demographic Parity Diff | 95% CI | p-value |
|---|---|---|---|
| S0 baseline (race included) | 0.060 | [0.032, 0.086] | 0.0 |
| S1 unawareness (race dropped) | 0.105 | [0.077, 0.131] | 0.0 |
Dropping race from the model's inputs did not shrink the gap - it grew, from 6.0 points to 10.5 points, and both numbers are tightly estimated and clearly significant (large n, narrow CIs, p essentially 0). This is the opposite of what "fairness through unawareness" predicts.
The core features still available to the model - Supervision_Risk_Score_First, prior-arrest-episode counts across felony/violent/property/drug categories, employment percentage - carry real correlation with race, a well-documented pattern from decades of unequal policing and enforcement. Removing race directly didn't remove that correlation; it just meant the model could no longer see race as a labeled input while still learning from features that track it closely, and the resulting fitted model happened to rely on those correlates more once race itself stopped absorbing part of the signal. faircode/strategies.py's next strategy, unawareness_proxy_removal (S2), goes further and drops the explicitly-listed proxy features too - see Mitigation Strategies for how that step, and the constraint-based strategies after it, actually move this specific gap.
Checks whether dropping a column actually reduces its correlation with the rest of the feature set, instead of assuming it does.
import pandas as pd
from sklearn.linear_model import LogisticRegression
from sklearn.preprocessing import LabelEncoder
def unawareness_gap_check(df, protected_column, feature_columns, group_disadvantaged, group_advantaged):
"""
Fits a simple classifier on `feature_columns` to predict the protected
attribute itself. A high accuracy means the remaining features can
reconstruct the protected attribute almost as well as having it
directly - the exact condition under which dropping it will fail to
remove its effect on downstream predictions.
Parameters:
df: DataFrame containing both the protected column and the
candidate feature columns
protected_column: name of the protected-attribute column
feature_columns: candidate columns to keep after "unawareness"
group_disadvantaged, group_advantaged: the two values of
protected_column to compare
Returns a dict with reconstruction_accuracy (how well the remaining
features predict the dropped attribute) and a plain verdict.
"""
mask = df[protected_column].isin([group_disadvantaged, group_advantaged])
sub = df[mask].dropna(subset=feature_columns + [protected_column])
X = pd.get_dummies(sub[feature_columns])
y = LabelEncoder().fit_transform(sub[protected_column])
model = LogisticRegression(max_iter=1000)
model.fit(X, y)
accuracy = model.score(X, y)
return {
"reconstruction_accuracy": accuracy,
"verdict": (
"Remaining features reconstruct the protected attribute well - "
"dropping it is unlikely to remove its effect."
if accuracy > 0.75 else
"Remaining features poorly predict the protected attribute - "
"dropping it is more likely to actually help."
),
}
# Usage example:
# result = unawareness_gap_check(
# df, protected_column="Race",
# feature_columns=["Supervision_Risk_Score_First", "Prior_Arrest_Episodes_Felony",
# "Prior_Arrest_Episodes_Violent", "Percent_Days_Employed"],
# group_disadvantaged="BLACK", group_advantaged="WHITE",
# )
# print(result)As Tenant Screening shows, dropping the protected attribute can make the measured gap larger, not just fail to shrink it. There is no guarantee of direction, only that the outcome depends on exactly which correlated features remain and how the model reweights them.
This repo's own next strategy after unawareness, unawareness_proxy_removal, only removes a pre-defined, finite list of known proxies - see Proxy Variables for why an unlisted or newly-emerging proxy can slip through even that step.
Some uses of a protected-adjacent attribute are lawful and relevant (age in certain insurance actuarial contexts, for instance); blanket removal without checking what else changes can trade one problem for a different, unexamined one.
"Unawareness" alone gives no signal, on its own, about whether the resulting model is more or less fair - only comparing an actual gap metric before and after, as this repo's benchmark harness does for every audit, can answer that.
- Proxy Variables - the mechanism (correlated features standing in for the removed attribute) that makes fairness through unawareness fail.
- What Are Pre-, In-, and Post-Processing Fairness Mitigations? - where "unawareness" sits as the first, weakest step in a five-strategy ladder, and what the stronger steps do differently.
- What Is Demographic Parity? - the metric used to measure the gap in the Tenant Screening example above.
Tenant Screening/- the audit above, where dropping race increased the measured gap rather than closing it.Insurance Denial/- a second real audit in this repo's benchmark harness where the sameunawarenessstrategy is applied and measured independently.
- Pedreschi, D., Ruggieri, S., Turini, F. (2008): Discrimination-Aware Data Mining - an early formal treatment of why removing a sensitive attribute doesn't remove its statistical effect.
- Kusner, M. et al. (2017): Counterfactual Fairness - proposes a causal alternative that explicitly accounts for what unawareness misses; see also Counterfactual Fairness.
- Barocas, S., Hardt, M., Narayanan, A. (2019): Fairness and Machine Learning - Chapter 1 covers "fairness through unawareness" directly as one of the field's earliest and most persistently re-discovered mistakes.
Part of The Fair Code Project - exposing and fixing algorithmic bias with real data and open code.