LLMs struggle to outperform statistical methods in survey data imputation

A new arXiv paper explores the use of large language models (LLMs) for imputing missing values in public opinion survey data. The study, which evaluated 84 in-context learning configurations across six open-weight LLMs, found that while LLMs can improve upon zero-shot persona prompting, they are at best competitive with traditional statistical methods like MICE PMM and TabPFN. The research highlights that LLM imputation is significantly more computationally expensive and sensitive to survey topics, and that common LLM prediction metrics do not reliably indicate imputation quality, underscoring the need for fit-for-purpose evaluation. AI

IMPACT LLMs are not yet a superior alternative to established statistical methods for survey data imputation, requiring careful evaluation for specific tasks.

RANK_REASON Research paper published on arXiv detailing LLM capabilities for data imputation. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• large-language models
• MICE PMM
• TabPFN
• Tobias Holtdirk

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →