Subtle biases introduced in equity studies through data anonymization

Area of Science:

• Educational research
• Data privacy
• Sociology of education

Background:

• Educational equity research relies on accurate data analysis.
• Data anonymization techniques like differential privacy are used to protect sensitive information.
• Assessing the impact of anonymization on equity-related variables is critical.

Purpose of the Study:

• To investigate the trade-off between data anonymization and utility for educational equity research.
• To evaluate the impact of (ε, δ)-Differential Privacy on microdata from the Brazilian National Student Performance Exam (ENADE).
• To assess how anonymization affects the representation of sociodemographic groups in educational equity analysis.

Main Methods:

• Application of the (ε, δ)-Differential Privacy model to microdata.
• Clustering of both original and anonymized datasets.
• Evaluation of anonymization effects on student sociodemographic variables (gender, race, income, parental education).

Main Results:

• Anonymization preserves overall data structure but can suppress or misrepresent minority groups.
• Differential privacy may introduce biases, potentially hindering the promotion of educational equity.
• The utility of anonymized data for socio-educational equity analysis is significantly impacted.

Conclusions:

• Domain experts are essential for interpreting anonymized data in equity studies.
• Anonymization efforts must carefully consider their impact on key group categories to avoid distorting findings.
• Preventing bias in anonymized data is vital for the validity of data-driven policies promoting educational equity.