Causal inference typically measures treatment effects on scalar or vector outcomes — does the drug reduce blood pressure, does the policy increase income. The mathematics assumes outcomes live in Euclidean space where differences, means, and variances are well-defined.
Luo et al. ask what happens when the outcome is topological. Some treatments change the shape of a distribution rather than its location — creating holes, splitting clusters, connecting components. A drug might not shift a biomarker's mean but might fragment its distribution into distinct subpopulations. A policy might not change average behavior but might create structural voids in the outcome space. These are real effects invisible to conventional estimators.
The framework defines treatment effects through differences in persistence diagrams — summaries of the topological features (components, loops, voids) present across scales. They use power-weighted silhouette functions to convert persistence diagrams into functional summaries, then build doubly robust estimators for the topological treatment effect. The estimator achieves functional weak convergence and supports formal hypothesis testing: is the topological structure of outcomes the same under treatment and control?
The through-claim is about what counts as an effect. Standard causal inference asks whether the treatment moved the outcome. Topological causal inference asks whether the treatment reshaped it. These are independent questions — a treatment can shift the mean without changing the topology, or change the topology without shifting the mean. The shape of the outcome distribution is a degree of freedom that conventional effect estimation treats as noise.