ESRA 2027 sessions by theme

Back to Overview of Sessions

Session Details

As survey researchers have been exploring the potentials and pitfalls of large language models (LLMs) in simulating survey responses, less attention has been paid to longitudinal panels. This session focuses on data quality in hybrid panels – designs that combine human-generated responses with LLM-generated responses based on respondents’ characteristics and previous responses. Such designs are increasingly discussed in computational and commercial contexts as a possible remedy for declining response rates, increasing nonresponse bias, and rising costs of longitudinal survey research. Not collecting all data from all panel members could allow researchers to concentrate resources on a subset of the panelists only, while also reducing respondent burden. However, the value of hybrid panels remains far from established: designs, implementations, and evaluation frameworks spanning both survey-methodological and substantive quality aspects are largely missing. What, if anything, can silicon responses add to longitudinal panel research beyond established imputation, weighting, or planned missingness approaches – and, conversely, how might longitudinal data improve the quality of silicon samples?
The session invites contributions addressing: (1) Hybrid designs versus established alternatives like imputation, weighting, and planned missingness; (2) quality-informed decision rules for allocating human versus silicon data across respondents and items; (3) content and methods for persona and digital-twin construction, including the role of prior-wave and auxiliary data (e.g., digital trace or interview data) and in-context learning versus fine-tuning; (4) calibration (e.g., MRP, PPI) and validation techniques for representation, measurement, and substantive quality in hybrid data; (5) trade-offs in cost/resources, burden, and data quality; and (6) fitness-for-purpose, including populations and topics where silicon responses are likely to fail.
We welcome methodological and applied contributions from academic, official-statistics, and commercial settings, with particular attention to transparent validation of when hybrid panels offer genuine methodological gains.