External Validity and Video-based Conjoints: A Study of Candidate Choice
Year: 2026
Type: other
Abstract: In the last decade, political scientists have deployed conjoint survey experiments to understand individuals’ political choices in varied settings, from immigration and relocation decisions to voting. Conjoint designs enable researchers to measure the causal impact of multiple attributes simultaneously and on the same scale (Hainmueller et al. 2014, Abramson et al. 2022, Bansak et al. 2023). With some exceptions (e.g. Abrajano et al. 2018, Vecchiato and Munger 2025, López Ortega and Radojevic 2025), researchers typically provide information about the choices via text in the form of tables, which allow respondents to compare the choices on various randomized attributes. For example, respondents might read text explaining that a specific candidate is a Hispanic man who opposes abortion, with that information conveyed in a table. This reliance on textual cues raises questions about whether individuals process such information in the same ways that they process real-world political information (see esp. Jenke et al. 2021), which they increasingly encounter through videos or other media that convey information using audio and/or visuals. Research on cognition suggests that encountering information in different formats may influence which information is acquired, retained, and subsequently employed. Information that is read may have different impacts than information which is seen or heard, and this may be especially true for visible ascriptive categories such as gender or race. By relying so heavily on tables of text, existing conjoint designs may limit the generalizability of their findings. Here, we develop a pipeline using various Artificial Intelligence (AI) tools, namely, the text-to-image model Nano Banana Pro, Runway ML’s native lip synchronization model, and Cartesia’s text-to-sound model Sonic-3, to embed conjoint treatments in videos and then test the resulting videos experimentally. Specifically, we generate realistic videos of synthetic political candidates using multiple procedures. In one, we blend images of state legislators and then present respondents with videos of hypothetical candidates whose race/ethnicity, gender, age, occupation, and political positions are randomly generated (the “deepfake” module). In another test, we instead use videos of live actors whose occupation and political positions are randomly varied. In a third test, we use the same software procedure as above to animate and give voice to still images of the actors. We then compare the impact of these various videos with standard, table-based conjoints which convey equivalent information.
Source: Open Science Framework
Language: English
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7206170556