A0437
Title: Design of experiments on sampled networks
Authors: Luke Ayres - King\'s College London (United Kingdom) [presenting]
Vasiliki Koutra - King's College London (United Kingdom)
Abstract: Experiments on social networks are increasingly important, for example, marketing experiments where the effectiveness of different advertisements given to different users needs to be assessed. Development of methods for optimizing the design of experiments on networks is an active area of research. A significant complication is that in networked experiments, the response of a given unit depends not only on the direct treatment applied to that unit, but also on the indirect effect of treatments applied to connected units. Previous research has focused on the problem of optimal design (treatment allocation) to assess direct treatment effects, indirect network effects, or a combination of both. The focus is on how different network properties, such as edge density, impact different optimality criteria. Studying such properties is particularly important for experiments on large networks where it is likely that not all available units will be used due to cost and/or computation time. Hence, a sub-network will need to be chosen for experimentation, with different choices giving different network properties. The use of different network sampling algorithms is assessed to evaluate the effectiveness of the resulting optimal designs and to demonstrate the important role of network structure in determining design efficiency.
