How to Use Image-Based Vignettes in Survey Experiments: An Image-Editing Approach

• In Event: 20858 - Reimagining How We Measure and Model the Social World: Causation, Flows, and Artificial Intelligence

Abstract

Survey experiments have played a crucial role in identifying causal effects in social science research. While most studies utilize text-based vignettes, emerging research increasingly presents respondents with image-based vignettes. However, existing studies often select images that either exhibit confounding biases, thus lacking internal validity, are highly stylized, thus lacking external validity, or are too costly to create. We propose a simple yet effective solution based on image editing using generative AI models. Researchers can edit existing images to produce new ones that differ from the original control image only by the intended treatment. We discuss the assumptions and practical details of image editing and extend two previous studies to demonstrate that our methods can accurately identify the effects of visual elements hypothesized to induce treatment, in both simple and conjoint experiments. This study contributes to the literature on survey experiments, causal inference, and the use of images as data.

Author

• Han Zhang, Brown University