Kernel weighted influence measures

Abstract

To asses the sensitivity of conclusions to model choices in the context of selection models for non-random dropout, several methods have been developed. None of them are without limitations. A new method called kernel weighted influence is proposed. While global and local influence approaches look upon the influence of cases, this new method looks at the influence of types of observations. The basic idea is to combine the existing influence approaches with a non-parametric weighting scheme. The kernel weighted global influence offers a possible solution to the problem of masking, while the kernel weighted local influence can be seen as a tool to better understand the source of influence. (C) 2004 Elsevier B.V. All rights reserved.