New strategies for the detection of influential observations

Efficient algorithms for diagnosing influential data points are investigated. Techniques examining potentially influential subsets are considered. Given a list of candidate observations, a new row-dropping algorithm (RDA) computes all possible observation-subset regression models. It employs a Cholesky updating algorithm using Givens rotations. The algorithm is organized via the all-subsets tree. The number of cases needed to be considered by multiple-row methods rapidly exhausts available computing power. The tree's structure is exploited to effect a parallel algorithm. Strategies using statistical information to prune the tree and narrow the search space are investigated.