Subtle biases introduced in equity studies through data anonymization

This work investigates the trade-off between data anonymization and utility, particularly focusing on the implications for equity-related research in education. Using microdata from the 2019 Brazilian National Student Performance Exam (ENADE), the study applies the (ε, δ)-Differential Privacy model to explore the impact of anonymization on the dataset’s utility for socio-educational equity analysis. By clustering both the original and anonymized datasets, the research evaluates how group categories related to students’ sociodemographic variables, such as gender, race, income, and parental education, are affected by the anonymization process. The results reveal that while anonymization techniques can preserve overall data structure, they can also lead to the suppression or misrepresentation of minority groups, introducing biases that may jeopardise the promotion of educational equity. This finding highlights the importance of involving domain experts in the interpretation of anonymized data, particularly in studies aimed at reducing socio-economic inequalities. The study concludes that careful attention is needed to prevent anonymization efforts from distorting key group categories, which could undermine the validity of data-driven policies aimed at promoting equity.