Adaptive Political Surveys and GPT-4

This repository contains the code and data to reproduce the analysis in "Adaptive political surveys and GPT-4: Tackling the cold start problem with simulated user interactions" (published in PLoS One; an earlier pre-print is available here).

Repository Structure

├── data/               # Data files
│   ├── candidates.csv
│   ├── candidates_reactions.csv 
│   ├── gpt_data.csv
│   ├── gpt_voters.csv
│   ├── questions.csv
│   └── results_zh.csv
├── figures/           # Generated figures
├── notebooks/         # Jupyter notebooks for analysis
│   ├── 1. Smartvote Preprocession.ipynb
│   ├── 2. GPT-4-API.ipynb
│   ├── 3. Cold-Start Dataset.ipynb
│   ├── 4. GPT Voters.ipynb
│   ├── 5. Statistical Model.ipynb
│   ├── 6. Running Simulation.ipynb
│   ├── 7. Results: Data Generation.ipynb
│   ├── 8. Results: Simulation Results.ipynb
│   └── 9. Results: Bias Investigation.ipynb
├── results/          # Generated results
└── src/             # Source code
    └── utils/       # Utility functions

Requirements

• Python 3.9+
• Required packages:
  • pandas
  • numpy
  • matplotlib
  • scikit-learn

Data Description

• candidates.csv: Information about political candidates
• candidates_reactions.csv: Candidate responses to survey questions
• gpt_data.csv: GPT-4 generated responses
• questions.csv: Survey questions in multiple languages
• results_zh.csv: Election results from Zurich

Usage

Run notebooks in order:

• Start with data preprocessing (notebook 1)
• Generate GPT responses (notebook 2)
• Build cold-start dataset (notebook 3)
• Create synthetic voters (notebook 4)
• Train statistical model (notebook 5)
• Run simulations (notebook 6)
• Analyze results (notebooks 7-9)

Citation

If you use this code in your research, please cite:

Bachmann, F., Van Der Weijden, D., Heitz, L., Sarasua, C., & Bernstein, A. (2025). Adaptive political surveys and GPT-4: Tackling the cold start problem with simulated user interactions. PLOS One, 20(5), e0322690. https://doi.org/10.1371/journal.pone.0322690