A graph-theoretic approach for testing causal effects under interference.

David Puelz from UT Austin and the Salem Center presents an approach for randomization tests of causal effects under general forms of interference.  The key idea is to represent a null hypothesis of spillovers as a bipartite graph and condition the test on a biclique in this graph. The approach is completely algorithmic and is applicable under arbitrary structures of interference, including clustered and spatial interference.  We connect the size of the conditioning biclique to statistical power, and we apply off-the-shelf graph clustering methods to find such bicliques efficiently and at scale.  Our working application is a large-scale policing experiment in Medelliìn, Colombia, where interference has a spatial structure.