CMStatistics 2021: Start Registration
View Submission - CMStatistics
Title: Design of experiments for autoregressive networks Authors:  Ben Parker - Brunel University (United Kingdom) [presenting]
Steven Gilmour - KCL (United Kingdom)
Vasiliki Koutra - King's College London (United Kingdom)
Abstract: In much traditional experimental design methodology, it is assumed that experimental units are unaffected by other experimental units, or occasionally that there is some simple structure that defines their common behaviour (for example blocking). We expand recent research on designing experiments on networks. Here we develop a methodology for designing experiments on a network where experimental units are related by an autoregressive model, such that the response of each experimental unit depends on neighbouring units specified by a general adjacency matrix. For example, these may be experiments where we measure: i) the popularity of an advert that is spread on an online social network; ii) the effectiveness of an agricultural treatment where responses from one plot are correlated with their neighbours; iii) a spatial network where responses are linked based on geographical closeness. We demonstrate some simple (pseudo-Bayesian) designs on these networks. We show the importance of accounting for this autoregressive effect, and that neglecting it in the experimental design can produce very inefficient experiments.