[Paper Review] Towards Algorithmic Fidelity: Mental Health Representation across Demographics in Synthetic vs. Human-generated Data

Shinka Mori, Oana Ignat+2 more|arXiv (Cornell University)|Mar 25, 2024

Mental Health Research TopicsPsychology3 citations

TL;DR

This paper introduces HeadRoom, a synthetic dataset of 3,120 depression-triggering stressor posts generated via GPT-3, controlling for race, gender, and time (pre- and post-COVID-19). It evaluates algorithmic fidelity by comparing synthetic data to human-generated mental health data, finding that GPT-3 captures some real-world stressor distributions across demographics, though with notable biases and limitations in representation and sensitivity.

ABSTRACT

Synthetic data generation has the potential to impact applications and domains with scarce data. However, before such data is used for sensitive tasks such as mental health, we need an understanding of how different demographics are represented in it. In our paper, we analyze the potential of producing synthetic data using GPT-3 by exploring the various stressors it attributes to different race and gender combinations, to provide insight for future researchers looking into using LLMs for data generation. Using GPT-3, we develop HEADROOM, a synthetic dataset of 3,120 posts about depression-triggering stressors, by controlling for race, gender, and time frame (before and after COVID-19). Using this dataset, we conduct semantic and lexical analyses to (1) identify the predominant stressors for each demographic group; and (2) compare our synthetic data to a human-generated dataset. We present the procedures to generate queries to develop depression data using GPT-3, and conduct analyzes to uncover the types of stressors it assigns to demographic groups, which could be used to test the limitations of LLMs for synthetic data generation for depression data. Our findings show that synthetic data mimics some of the human-generated data distribution for the predominant depression stressors across diverse demographics.

Motivation & Objective

• To investigate how Large Language Models like GPT-3 represent depression stressors across diverse racial and gender groups.
• To evaluate the algorithmic fidelity of synthetic data in mirroring real-world mental health data distributions.
• To identify biases and limitations in LLM-generated mental health data, especially concerning underrepresented demographics.
• To provide a reproducible framework for generating and analyzing synthetic depression data using controlled prompts.
• To caution researchers against using synthetic data uncritically in mental health applications due to potential bias amplification.

Proposed method

• Controlled prompt engineering using GPT-3 to generate 3,120 synthetic blog-style posts about depression stressors, with explicit control over race, gender, and time frame (pre- and post-COVID-19).
• Development of a structured prompt template: 'I want you to act like a {race} {gender} who is feeling depressed. Write a blog post to describe the main source of stress in your life.'
• Semantic and lexical analysis of generated data to identify predominant stressors per demographic group.
• Comparison of synthetic data with human-generated data from the UMD-ODH dataset using topic modeling and keyword-based analysis.
• Use of existing topic models and keyword lists from Aguirre et al. (2022) to enable quantitative comparison between synthetic and real data.
• Public release of the HeadRoom dataset and code via GitHub to support reproducibility and further research.

Figure 1: Topic Modeling: topic proportions between race and gender intersectionality – African American women vs. African American men. The bars represent confidence intervals. The closer to the graph extremities, the more prevalent the topics are for the corresponding demographics

Experimental results

Research questions

• RQ1 RQ1: What are the depression stressors identified by GPT-3 for different demographic groups, and does it capture known demographic biases in mental health stressors?
• RQ2 RQ2: How does synthetic data about depression stressors compare to human-generated data across race and gender demographics?
• RQ3 RQ3: To what extent does GPT-3 exhibit 'algorithmic fidelity' in representing real-world distributions of depression stressors across diverse groups?

Key findings

• GPT-3-generated synthetic data captures some of the most prevalent depression stressors found in human-generated data, such as financial strain and family-related concerns, across multiple demographic groups.
• The model exhibits demographic bias in stressor attribution, with certain stressors being over- or under-represented depending on race and gender combinations.
• Stressor patterns in synthetic data show partial alignment with real-world distributions from the UMD-ODH dataset, indicating moderate algorithmic fidelity.
• The model does not generate mentions of suicide or self-harm, which are common in real human depression texts, suggesting possible safety or training data limitations.
• The synthetic dataset is limited in size (3,120 samples) and may not generalize to longer text sequences or more complex linguistic patterns.
• Model limitations include lack of explainability, potential changes in future GPT-3 versions, and a training cutoff in June 2021, restricting post-2021 relevance.

Figure 2: Topic Modeling: topic proportion between different demographics, as detected in GPT-generated data and in real-life data. Colors represent different races and genders: Men – purple, Women – orange, Asian – magenta, African American – green, Hispanic – blue, and White – red. The bars repres