Evaluating Algorithmic Approaches to Uncover Racial, Ethnic, and Gender Disparities in Scientific Authorship

Abstract

To explore the capabilities of race/ethnicity and gender prediction algorithms in uncovering patterns of authorship distribution in scientific paper submissions to a major peer-reviewed scientific journal (AJPH), we analyzed 17 667 manuscript submissions from the United States between 2013 and 2022. We used machine-learning algorithms to predict corresponding authors’ race/ethnicity (Asian, Black, Hispanic, White) and gender categories based on name-derived probabilities to compare the predictive performance of these algorithms and their impact on disparity analysis. Predicted White authors dominated submissions and had the highest acceptance rates (21.1%), while predicted Asian authors faced the lowest (14.9%). Predicted women, despite being the majority, had lower acceptance rates (17.9%) than men (20.5%), a trend consistent across most racial/ethnic groups. Different algorithms revealed similar disparities but were limited by biases and inaccuracies in predicting race and ethnicity. Manuscript acceptance rates revealed disparities by race/ethnicity and gender; predicted White and male authors had the highest rates. While machine-learning algorithms can identify such patterns, their limitations necessitate combining them with self-identified demographic data for greater accuracy.