Transportable Causal Effect Estimation across Networks under Interference

Estimating causal effects from networked data is central to social science, public health, and policy evaluation. In these settings, a unit's outcome depends on its own treatment and, through interference, on the treatments of its neighbors. Consequently, an intervention tested on one network often fails to yield the same effect when deployed on a different population. In practice, the training network and the deployment network differ in topology, node-covariate composition, and spillover pathways. This paper investigates whether causal effects—including direct, spillover, and total effects—can be validly estimated for a deployment network using experimental data from a different source network combined only with passive observations from the target network.

The authors model the discrepancies between populations using a selection diagram extended to the network setting. Specifically, the framework separates covariate shift (changes in node attributes and assignment mechanisms) from structural network shift (changes in graph topology, degree distribution, and neighbor covariate composition). Under graphical conditions and a stratified-interference assumption, the authors derive transport formulas for direct, spillover, and total causal effects. These formulas explicitly separate which interventional mechanisms from the source domain remain invariant and which observational distributions from the target domain must be reweighted.

To operationalize the theoretical findings, the authors develop TranCE (Transported Causal Effects), a doubly-robust algorithm. TranCE combines an interventional outcome model fitted on the source network, a domain density-ratio correction estimated from both networks, and known neighbor treatment propensities. By employing cross-fitted inference, the estimator achieves asymptotic normality and remains consistent if either the outcome model or the propensity score model is correctly specified. Extensive experiments on semi-synthetic social network benchmarks and a real-world weather-insurance field experiment confirm the method's effectiveness over existing single-network or independent transport baselines.