CMStatistics 2021: Start Registration
View Submission - CMStatistics
Title: Randomized graph cluster randomization Authors:  Johan Ugander - Stanford University (United States) [presenting]
Hao Yin - Stanford University (United States)
Abstract: Causal inference under network interference provides a formal framework for measuring network effects using randomized experiments. Experimental designs based on graph cluster randomization (GCR), randomizing units at the level of network clusters, have been shown to greatly reduce variance when measuring network treatment effects, compared to unit-level random assignment. But even so the variance is very often prohibitively large. A randomized version of the GCR design is proposed which is descriptively named randomized graph cluster randomization (RGCR), and which uses a random clustering rather than a single fixed clustering. By considering an ensemble of many different cluster assignments, this design avoids a key problem with GCR where a given unit is sometimes ``lucky'' or ``unlucky'' in a given clustering, thereby greatly reducing the variance of network treatment effect estimators in both theory and across extensive simulations.