A Longitudinal Approach to the Analysis of Social Media Engagement: The Case of Anger-Driven Climate-Skeptic Message Propagation During the 2021 German Elections

Social media have become an important space for political agenda formation and mobilization, with user engagement playing a key role in spreading messages. Accordingly, prior research has extensively examined social media users’ engagement and sharing behaviors. In this study, we examine the advantages of longitudinal modeling for analyzing social media engagement compared with cross-sectional approaches, focusing on the relationship between emotional reactions—particularly anger—and content propagation. Although cross-sectional approaches are commonly used to analyze social media data, engagement patterns are inherently temporal and therefore naturally call for a longitudinal, time-sensitive approach. We argue that longitudinal methods can be more effective than cross-sectional ones for analyzing time-evolving engagement dynamics. These methods indeed represent a middle ground between familiar cross-sectional approaches and sophisticated time-series techniques, avoiding some pitfalls of the former while relying on simpler assumptions than the latter. Longitudinal methods, in fact, reduce omitted variable bias, help account for time-varying factors—including algorithmic amplification and network propagation confounders—and accommodate irregular or sparse social media data. We empirically explore the differences between longitudinal and cross-sectional analysis by comparing estimates of the effect of anger on sharing using 1,137 environmentally themed Facebook posts by German political parties during the 2021 federal election, focusing on Alternative für Deutschland (AfD) and Die Grünen (The Greens), which represent opposite ends of the environmental policy spectrum. Cross-sectional estimates were derived using three sampling strategies: the last observation per post, the first post-election observation, and a randomly selected observation per post. Bayesian multilevel regression with a negative binomial specification was applied across both longitudinal and cross-sectional models. Our results indicate that longitudinal modeling yields more conservative and precise estimates, whereas cross-sectional methods tend to exaggerate effect sizes and interparty differences. All models suggest that anger is positively associated with sharing for AfD and negatively for Die Grünen, but longitudinal analysis provides greater inferential stability by controlling for time-invariant confounders and algorithmic amplification. Overall, the findings underscore the value of incorporating temporal dynamics into social media research, while also highlighting the challenges of applying longitudinal approaches to digital trace data, particularly with regard to data access.