|
Abstract:
|
In a retrospective analysis of a historical data set, the data may have multiple sustained shifts and/or outliers. Especially with individual observations, subgroups of only one observation, the presence of out-of-control observations may disrupt parameter estimation, in turn masking their presence. The goal is partitioning the data so that all pairs of successive parts have statistically significantly different mean vectors and no finer partition has this property. A part is a set of observation(s) between consecutive boundaries. Robust estimators, such as the minimum volume ellipsoid, are computationally intensive for large amounts of data. Making use of the sequential nature of sustained shifts can reduce the computational burden. This research generalizes earlier univariate methods, presenting alternative approaches and preliminary comparison of their effectiveness.
|
