Addressing discretization-induced bias in demographic prediction

Area of Science:

• Social Sciences
• Computer Science
• Political Science

Background:

• Demographic imputation is vital for auditing disparities and political targeting.
• Current methods often discretize continuous predictions, leading to potential bias.

Purpose of the Study:

• To investigate the phenomenon of discretization bias in demographic imputation.
• To introduce and evaluate a novel method for mitigating this bias.

Main Methods:

• Analysis of argmax labeling for race/ethnicity imputation using real-world data.
• Development and testing of a joint optimization approach with a data-driven threshold heuristic.

Main Results:

• Argmax labeling significantly undercounts Black voters (e.g., 28.2% in North Carolina).
• The proposed joint optimization method effectively eliminates discretization bias.
• Negligible individual-level accuracy loss was observed with the new method.

Conclusions:

• Discretization bias in demographic imputation has serious implications for downstream applications.
• Calibrated continuous models alone cannot resolve this bias; specialized methods are necessary.
• Researchers and practitioners must carefully consider the consequences of discretizing demographic predictions.