Four Guiding Principles for Modeling Causal Domain Knowledge: A Case Study on Brainstorming Approaches for Urban Blight Analysis

TL;DR

The paper tackles the challenge of translating causal domain knowledge gathered through brainstorming (e.g., cognitive mapping) into rigorous causal diagrams for urban blight analysis. It proposes four concrete rules to extract and structure causal variables, treat interaction terms as artificial nodes, constrain edges to direct causal relations, and enforce transitivity-based encoding checks. Applying these rules to digitized cognitive maps and existing urban blight diagrams reveals encoding loops, overlapping clusters, and interaction effects that benefit from modularity-focused restructuring and artificial-nodes representation. The work emphasizes collaboration between domain experts and causal-data-science researchers and outlines future extensions to incorporate temporal and spatial explicit-variable encoding, with potential applicability to other policy-relevant domains like industrial risk assessment.

Abstract

Urban blight is a problem of high interest for planning and policy making. Researchers frequently propose theories about the relationships between urban blight indicators, focusing on relationships reflecting causality. In this paper, we improve on the integration of domain knowledge in the analysis of urban blight by introducing four rules for effective modeling of causal domain knowledge. The findings of this study reveal significant deviation from causal modeling guidelines by investigating cognitive maps developed for urban blight analysis. These findings provide valuable insights that will inform future work on urban blight, ultimately enhancing our understanding of urban blight complex interactions.

Figures (4)

• Figure 1: Grouping related text entities under their corresponding causal variable. The text entities $x_1$ to $x_n$ are articulated by subject matter experts. These text entities describe different values of the causal variable $X$. The relationships between these text entities can be identified as either mutually exclusive, as $X$ cannot take multiple values simultaneously, or as describing the same value of the causal variable.
• Figure 2: Representing interaction entities as artificial nodes. Interaction entities are text entities which describe an interaction between two or more causal variables. These text entities should be represented as artificial nodes in the causal diagram. The artificial nodes encompass the corresponding observed values of the causal variables it represents. The artificial node can be linked to causal variables; however the type of this relation should differ from other causal relations in diagrams.
• Figure 3: Identifying causal relations in causal diagrams. In a causal diagram, two variables represented as nodes are causally connected if there is at least one text entity/value belonging to the first variable that is the direct cause of an text entity/value belonging to the second variable. Causal relations should only be included between two variables if there is no other variable that fully mediates their relationship. Causal relations that include an interaction entity as a cause or an effect should be highlighted as a different relation type, as causal diagrams are intended to link two causal variables rather than a causal variable to a interaction entity represented as an artificial node.
• Figure 4: Devising transitivity principles of causal relations to verify causal variables encoding. Violations to the transitivity principles of causal relations indicates impairments in the causal variables identifications. One potential cause of this issue is that the identified mediator variable could be too general. Alternatively, at least a value of one of the mediator variable is actually an interaction even.