Visual causality: investigating graph layouts for understanding causal processes

Causal diagrams provide a graphical formalism indicating how statistical models can be used to study causal processes. Despite the extensive research on the efficacy of aesthetic graphic layouts, the causal inference domain has not benefited from the results of this research. In this paper, we investigate the performance of graph visualisations for supporting users’ understanding of causal graphs. Two studies were conducted to compare graph visualisations for understanding causation and identifying confounding variables in a causal graph. The first study results suggest that while adjacency matrix layouts are better for understanding direct causation, node-link diagrams are better for understanding mediated causation along causal paths. The second study revealed that node-link layouts, and in particular layouts created by a radial algorithm, are more effective for identifying confounder and collider variables.