More than the Sum of its Parts? Measuring Image Types in Political Communication

This study presents a path to measuring image types in political communication involving an innovative mixed-methods approach. First, building on visual and political communication literature, we identify image types in an inductive, qualitative coding process. Second, using a large dataset containing EU institutions' visual social media posts, we engage with manual and automated content analysis methods. We employ transformer models and contrast methodological choices in the automated measurement of image types. Informed by manually coded data, our study uses 1) transfer learning to finetune models on the image level, trained on entire pictures, and 2) object detection to link image types with recurring features relevant to each category and make them measurable through individual aspects. This comparative analysis demonstrates the efficacy of the different approaches, providing a hands-on example of, so far, very scarce approaches to deal with visual data in an automated manner, as well as showcasing their validation. The theoretical link and the qualitative work to collect, describe, and connect image types with the literature, moreover, enables the approach to be harnessed for substantial research. Exploring the implications of the methodological choices, we add to the understanding of how visual cues convey political messages and how they might relate to specific goals. Contributing to methodological advancements applicable in various social science disciplines, this study thus provides an innovative view on digital data, combining in-depth reflection of critical constructs with large-scale analysis.