Avoiding bias when inferring race using name-based approaches.

Area of Science:

• Sociology of Science
• Computational Social Science
• Bibliometrics

Background:

• Racial disparities in academia are a significant issue.
• Quantitative analysis of these inequalities is hindered by a lack of author race data.
• Algorithmic methods, like name-based racial inference, offer a potential solution but may introduce bias.

Purpose of the Study:

• To assess algorithmic bias in name-based racial inference for academic authors.
• To evaluate different methods of inferring race from author names.
• To inform more equitable research on racial disparities in science.

Main Methods:

• Utilized U.S. Census and mortgage application data to infer race for U.S. authors in Web of Science.
• Examined the impact of using given names, family names, and different inferential approaches (thresholds vs. continuous distributions, imputation).

Main Results:

• The accuracy of name-based racial inference differs significantly across racial and ethnic groups.
• Threshold-based methods were found to underestimate Black authors and overestimate White authors.
• Algorithmic bias can be introduced depending on the specific inference approach and data used.

Conclusions:

• Name-based racial inference requires careful consideration to mitigate bias.
• Recommendations are provided to improve the accuracy and fairness of these methods.
• This work establishes a foundation for less-biased investigations into racial disparities in scientific research.