Designing experiments under bipartite interference

Develop experimental designs for bipartite interference settings—where interventions are applied to one set of interventional units that influence outcomes on a second set of outcome units via partially overlapping exposure mappings—so that researchers can learn effectively about policy-relevant interventions that allocate treatments across interventional units.

Background

The paper discusses interference in complex systems where the no-interference assumption fails. In the bipartite interference setting, interventions are applied to one population (e.g., power plants) and outcomes are measured on another (e.g., sensor locations), with overlapping influence patterns.

The authors note that while such systems are common and important in practice, there is a gap in how to design experiments that enable effective learning about policies, such as which interventional units to treat given their overlapping effects on outcome units.

References

Designing experiments that allow researchers to learn effectively about policies of interest that have this form remains an open question.

— Challenges in Statistics: A Dozen Challenges in Causality and Causal Inference  (2508.17099 - Cinelli et al., 23 Aug 2025) in Section: Interference and Complex Systems — Example: Bipartite Graph (Figure: General bipartite interference graph)

Several directions remain open. Natural extensions include unit-specific allocation and assignment probabilities, covariate adjustment in the spirit of CA-ERL , weighted experiment graphs, and designs that optimise the allocation split across concurrent experiments jointly with the exposure distribution . On the theoretical side, characterising the efficiency frontier beyond the linear reweighting class, and relaxing the homogeneity assumption underlying variance estimation, both appear fruitful.

— EARL: Exposure- and Allocation-Reweighted Linear Estimator for Bipartite Experiments with Partial Assignment  (2609.31022 - Kurennoy, 25 Sep 2026) in Conclusion, Section 7