This repository contains the data, experimental code, model outputs, and analysis materials accompanying Gender Associations in LLM-Mediated ADHD Self-Diagnosis.
The study evaluates gender associations in LLM-mediated ADHD self-diagnosis using an adapted Context Association Test (CAT) under explicit and implicit gender cueing.
The repository contains the experimental inputs, model outputs, analysis code, and visualization materials used in the study.
Experimental inputs and reference data are provided under data/, model responses and derived evaluation results under results/, and the corresponding experimental, analysis, and visualization code under src/. Generated figures and tables are available under out/.
API credentials are not included. Scripts requiring access to the OpenAI API expect an OPENAI_API_KEY to be supplied through a local environment file or environment variable.
self_diagnostic_bias_llms/
├── data/ # Experimental inputs and reference data
│ ├── base_symptom_clauses/ # Base symptom clauses used to construct stimuli
│ ├── context_lists/ # Final context sentences used in experiments
│ ├── mapping/ # Data used in the Reddit-to-DIVA mapping procedure
│ ├── names/ # SSA name data used for implicit gender cueing
│ ├── proxy_data/ # Clinical reference data
│ └── README.md # Data documentation
│
├── src/ # Experimental, analysis, and visualization code
│ ├── utils/ # Shared configuration and helper functions
│ ├── prompts/ # Prompt templates used in the LLM pipeline
│ ├── main.py # LLM-based experimental pipeline
│ ├── mapping_pipe.py # Reddit-to-DIVA mapping procedure
│ ├── figures_study1.py # Study 1 figures
│ ├── figures_study2.py # Study 2 figures
│ ├── plots_LLM_rstudio.Rmd # R Markdown tables and visualizations
│ ├── simple_stats.py # Descriptive statistics and summaries
│ └── README.md
│
├── results/ # Model outputs and derived evaluation results
│ ├── responses/ # Raw model responses
│ │ ├── s1_ex_responses.csv
│ │ ├── s1_im_responses.csv
│ │ ├── s2_ex_responses.csv
│ │ └── s2_im_responses.csv
│ │
│ ├── metrics/ # CAT-derived evaluation metrics
│ │ ├── s1_ex_metrics.csv
│ │ ├── s1_im_metrics.csv
│ │ ├── s2_ex_metrics.csv
│ │ └── s2_im_metrics.csv
│ │
│ ├── summaries/ # Analysis summaries used for tables and figures
│ └── README.md
│
├── out/ # Generated outputs
│ ├── figures/ # Figures and visualizations
│ ├── tables/ # Generated tables
│ └── README.md
│
├── requirements.txt # Python package dependencies
├── LICENSE # Repository license
└── README.md # Project documentation
Abbreviations: s1/s2 = Study 1/Study 2; ex = explicit cueing; im = implicit cueing.
The analyses were conducted using Python and R. Required Python dependencies are listed in requirements.txt.
| Tool | Installation |
|---|---|
| Python 3.12.2 | Install Python and the dependencies listed in requirements.txt. |
| R 4.4.2 + R Markdown | Install R via CRAN. R Markdown files can be run using RStudio or another compatible environment. |
For scripts that query the OpenAI API, set the OPENAI_API_KEY environment variable before running the experimental pipeline.
Gender Associations in LLM-Mediated ADHD Self-Diagnosis © 2026 Matilde Elene Hansen and Ross Deans Kristensen-McLachlan.
This work is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0).
If you use this work, please cite:
Citation forthcoming.
The paper has been accepted to EVAL4SD (First Workshop on Evaluating LLMs for Specialized Domains), co-located with KONVENS 2026. The final published version in the ACL Anthology is forthcoming.
This work was partially supported by the Danish National Research Foundation (Grant No. DNRF193) through TEXT: Center for Contemporary Cultures of Text, Aarhus University.