Implementing Causal Perception: Competing SCMs and Situated Fairness

Research Question and Approach

Traditional approaches to algorithmic fairness often assume a single, objective ground-truth model of the world. This paper challenges that premise by implementing the 'causal perception' framework, which models how different agents—each holding their own Structural Causal Model (SCM)—interpret the same information. The author operationalizes both structural perception (disagreement on the causal graph) and parametrical perception (disagreement on causal weights). By developing algorithms for interventional and counterfactual reasoning, the study allows researchers to quantify the divergence between competing worldviews using various distance metrics.

Implementation and Findings

The author demonstrates the framework using the German Credit dataset, comparing two idealized agents with competing SCMs. The implementation covers both the second rung (interventional) and third rung (counterfactual) of Pearl’s causal ladder. The results reveal that even minor disagreements in causal structure propagate into significant differences in fairness metrics, such as demographic parity. For instance, two agents differing on a single causal edge reached substantially different parity gaps and disagreed on a significant portion of individual loan decisions. The study highlights that the 'perception verdict'—whether two agents are considered to be in disagreement—is highly sensitive to the chosen distance metric (e.g., Wasserstein-2, Kullback-Leibler, or Total Variation) and the threshold used to define significant divergence.

This work shifts the focus of fairness research from finding a single 'correct' model to acknowledging that bias is situated. In multi-expert decision systems, such as credit lending or medical diagnosis, different stakeholders may legitimately hold different causal beliefs. By providing the tools to quantify these differences, the framework enables a more nuanced understanding of why experts disagree and how their underlying worldviews shape their fairness assessments. It suggests that instead of trying to average away disagreement as noise, practitioners should treat competing causal interpretations as a core component of the decision-making landscape.