Abstract for: The challenge of model complexity: improving the interpretation of large causal models through variety filters

While large causal models provide detailed insights for those who develop them, they often suffer from being inaccessible to outsiders of the modeling process. This is particularly regrettable in cases where large models are carefully crafted by experts and hold potential lessons to learn for academics and practitioners interested in the particular research field. To address this problem, we propose a set of variety filters, i.e. tools to reduce model complexity, to turn the interpretation of large causal models more efficient. A primary variety filter is the recently published algorithmic detection of archetypal structures (ADAS) method. ADAS is a method for the identification of influential structures within causal models, facilitating both model diagnosis and policy design. However, when applied to complex models, ADAS might come to its limits because the number of identified archetypal structures might be too high for a meaningful interpretation. Hence, we propose two additional filters, as precursors to ADAS: structural model partitioning and interpretive model partitioning. We demonstrate the proposed variety filters on the basis of the Obesity System Map—a model containing 108 variables and 297 interdependencies. Through the filters, we are able to attenuate model complexity drastically, while enhancing the comprehension of the model.