Abstract:

Multivariate control charts are used for monitoring multiple series simultaneously, for the purpose of detecting shifts in the mean vector in any direction. In the context of disease outbreak detection, interest is in detecting only an increase in the process means. Two practical approaches for deriving directional Hotelling charts are Follmann’s correction and Testik and Runger’s quadratic programming. However, there has not been an extensive comparison of their practical performance. Moreover, in practice many of the underlying method assumptions are often violated and the theoretically-guaranteed performance might not hold. In this work we compare the two directionally-sensitive approaches: a statistically-based approach and an
operations research solution. We evaluate Hotelling charts as well as two extensions to multivariate EWMA charts. We examine practical performance aspects such as robustness to often impractical assumptions, the amount of data required for proper performance, and computational aspects. We perform a large simulation study and examine performance on authentic biosurveillance data.