Learning from Convenience Samples: Fine-Tuning LLMs to Predict Voting Choices from Biased Survey Samples
The study tested whether fine-tuning Large Language Models (LLMs) can accurately predict German voting choices when trained on biased convenience samples, addressing issues like non-response and missing data. Results showed that fine-tuned LLMs significantly outperform zero-shot models and sometimes traditional classifiers under biased sampling conditions, suggesting they are a robust tool for imputing survey data.