Assessing Group-level Gender Bias in Professional Evaluations: The Case of Medical Student End-of-Shift Feedback

Emmy Liu, Michael Henry Tessler, Nicole Dubosh, Katherine Mosher Hiller, Roger Levy

Abstract: Although approximately 50% of medical school graduates today are women, female physicians tend to be underrepresented in senior positions, make less money than their male counterparts and receive fewer promotions. There is a growing body of literature demonstrating gender bias in various forms of evaluation in medicine, but this work was mainly conducted by looking for specific words using fixed dictionaries such as LIWC and focused on recommendation letters. We use a dataset of written and quantitative assessments of medical student performance on individual shifts of work, collected across multiple institutions, to investigate the extent to which gender bias exists in a day-to-day context for medical students. We investigate differences in the narrative comments given to male and female students by both male or female faculty assessors, using a fine-tuned BERT model. This allows us to examine whether groups are written about in systematically different ways, without relying on hand-crafted wordlists or topic models. We compare these results to results from the traditional LIWC method and find that, although we find no evidence of group-level gender bias in this dataset, terms related to family and children are used more in feedback given to women.

Introduction

Despite women comprising approximately 50% of medical school graduates, they remain underrepresented in senior leadership roles and continue to face a persistent gender pay gap. A leading hypothesis for this disparity is the accumulation of "allocational harms"—inequities in professional evaluation that begin early in a physician's career. While previous research has identified gender bias in high-stakes documents like residency letters of recommendation, less is known about the "day-to-day" formative feedback medical students receive during clinical rotations. This study investigates whether the language used in routine end-of-shift evaluations differs systematically by gender, even when objective performance scores are identical.

Methodology

The researchers employed a novel Prediction Residual Analysis using Large Language Models (LLMs) to move beyond traditional dictionary-based approaches.

Figure 6: The Prediction Residual Analysis workflow, showing how a fine-tuned language model predicts a competency score from narrative comments to calculate the residual error.

The process involved three main steps:

  1. Anonymization: To prevent the model from relying on explicit gender cues, all names were replaced with "x" and gendered pronouns were neutralized to "they."
  2. Model Training: A bert-base-uncased model was fine-tuned to predict a student’s "global competency score" (ranging from 0 to 3) based solely on the anonymized narrative comments.
  3. Residual Analysis: The core metric was the residual error ($\delta = \text{actual score} - \text{predicted score}$). A consistent under-prediction for one gender would indicate that the text written about that group is systematically less positive than their numeric performance warrants.

Additionally, the team used the Linguistic Inquiry and Word Count (LIWC) method to check for specific thematic categories like "agentic" (e.g., ambitious) or "communal" (e.g., caring) language.

Dataset

The study utilized the NCAT-EM (National Clinical Assessment Tool for Emergency Medicine) dataset, comprising 3,162 individual assessments from 13 U.S. institutions between 2017 and 2019.

Figure 1: Structure of the NCAT-EM tool, including categorical evaluations, free-text comments, and demographic data for both students and assessors.

The final dataset included evaluations for 1,767 male and 1,395 female students. Each entry contains short free-text comments, categorical assessments across various medical skills, a global competency score, and demographic information for both students and faculty assessors.

Key Results

The analysis yielded two primary findings:

  • Absence of Systemic Residual Bias: The BERT model achieved a prediction accuracy of 46%, which is comparable to human performance (50.7%). Crucially, there was no significant difference in the prediction residuals between male and female students ($p = 0.517$). This suggests that, at a group level, the "valence" of the narrative feedback aligns with numeric scores regardless of the student's gender.
  • Linguistic Divergence in Social Themes: While overall valence was balanced, the LIWC analysis revealed a highly significant difference in "Social-communal" themes ($p < 6 \times 10^{-16}$). Comments for female students were significantly more likely to mention families, babies, and children (18.4% for women vs. 8.26% for men), often in the context of patient interactions or perceived suitability for pediatrics.
  • Agentic/Communal Parity: Unlike studies of high-stakes recommendation letters, this formative feedback did not show significant differences in "agentic" or "communal" (kindness/agreeableness) descriptors between genders.

Discussion and Limitations

The findings suggest that the "snowball effect" of gender inequity in medicine may manifest more prominently in high-stakes summative documents than in daily formative feedback. The use of model residuals provides a robust way to detect bias without requiring a pre-defined "gold standard" for unbiased text.

However, the study is limited by the inherent "noise" in short shift comments, which may not always fully reflect the reasoning behind a numeric grade. Furthermore, the data is specific to Emergency Medicine and treats gender as a binary variable. Future research should apply this residual method to larger, more diverse professional datasets to determine if these patterns hold across different sectors and specialties.

Summary updated

Abstract

Although approximately 50% of medical school graduates today are women, female physicians tend to be underrepresented in senior positions, make less money than their male counterparts and receive fewer promotions. There is a growing body of literature demonstrating gender bias in various forms of evaluation in medicine, but this work was mainly conducted by looking for specific words using fixed dictionaries such as LIWC and focused on recommendation letters. We use a dataset of written and quantitative assessments of medical student performance on individual shifts of work, collected across multiple institutions, to investigate the extent to which gender bias exists in a day-to-day context for medical students. We investigate differences in the narrative comments given to male and female students by both male or female faculty assessors, using a fine-tuned BERT model. This allows us to examine whether groups are written about in systematically different ways, without relying on hand-crafted wordlists or topic models. We compare these results to results from the traditional LIWC method and find that, although we find no evidence of group-level gender bias in this dataset, terms related to family and children are used more in feedback given to women.
BibTeX
@inproceedings{liu2022assessing,
  author = {Emmy Liu and Michael Henry Tessler and Nicole Dubosh and Katherine Mosher Hiller and Roger Levy},
  title = {{Assessing Group-level Gender Bias in Professional Evaluations: The Case of Medical Student End-of-Shift Feedback}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2022},
}
Figure