Longitudinal surveys often need to compare responses across waves or datasets, but questions that look alike may not measure the same thing. A new arXiv paper highlights that surface-level textual similarity is not enough for this task, noting that two questions can be lexically similar yet assess entirely different concepts.
The authors describe a "conditional information retrieval" approach for survey harmonisation, presumably using additional context to determine conceptual equivalence. The work addresses a core challenge in secondary data analysis: linking questions that are conceptually equivalent despite different phrasing. The paper is available on arXiv under identifier 2610.26173.