B0940
Title: Optimal design of two-stratum experiments in the presence of autocorrelation
Authors: Daniel Palhazi Cuervo - University of Antwerp (Belgium) [presenting]
Peter Goos - Universiteit Antwerpen (Belgium)
Kenneth Sorensen - University of Antwerp (Belgium)
Abstract: There are several experimental scenarios in which a complete randomization is not possible. This usually happens when not all observations can be made under homogeneous conditions, or when there are factors whose levels are hard or expensive to change. For these cases, it is possible to generate experimental designs that explicitly take into account these limitations. This kind of designs, called two-stratum designs, groups the observations that are made under similar experimental conditions. The data from these experiments is usually analyzed assuming that every pair of observations within a given group has the same correlation. We study experimental scenarios in which not all pairs of observations within each group are correlated to the same extent. Instead, we consider observations that are closer to each other (either in time or space) to be more correlated than observations that are further apart. This might be expected when the observations are carried out sequentially, or when the experimental conditions are slightly different for each observation within the group. We generate optimal designs for two-stratum experiments under these conditions, and we compare their statistical efficiency to that of traditional designs (constructed assuming that all observations within a group are correlated to the same extent).
Journal