RESEARCH SYNTHESIS -- LLM-Based Simulations of Survey Responses: A Systematic Literature Review and Synthesis of Methods

Large language models (LLMs) are increasingly used in survey research, including to predict, simulate, or impute responses. Despite a rapidly growing interest in applications to create synthetic survey data, the empirical evidence remains fragmented, with substantial variation in research objectives, model architectures, ``persona'' construction strategies, and validation practices. This paper presents a systematic review of 55 studies published between 2022 and 2025 that use LLMs to generate, predict, or impute responses in survey and related attitudinal or behavioral data. The results suggest that LLM-based survey simulation is best understood as a conditional, use-case-specific tool rather than a general substitute for traditional imputation methods and human respondents. We outline concrete steps toward a more coherent evaluation framework, including use-case-specific metric selection, relational alignment as a standard validity check, and greater transparency in model and design choices.