A0726
Title: Score augmented Frobenius distance with applications in causal inference
Authors: Siyun He - University of Michigan (United States) [presenting]
Abstract: A novel method for causal inference is proposed in panel settings with network-valued outcomes by introducing the score augmented Frobenius distance, a metric that compares networks after sorting their adjacency matrices by node-level structural scores. These scores, which incorporate both observed covariates and structural features (e.g., centrality), serve to align nodes across networks under the assumption that structural roles, rather than identities, drive treatment effects. This sorting induces an equivalence class over node permutations, allowing valid comparisons between networks with unobserved heterogeneity. It is shown how the causal Frobenius distance can be used to extend standard difference-in-differences and synthetic control methods to settings where the outcome is a network. The framework is applicable to a variety of empirical settings, including social networks, trade networks, and institutional relationships, where interventions affect structural properties rather than specific node labels. Formal identification results are provided, asymptotic behavior is discussed under network sampling, and performance on simulated and real-world policy interventions is demonstrated.
