The Detection of Influential Observations for Allocation, Separation, and the Determination of Probabilities in a Bayesian Framework

Abstract

Normal theory separation and allocation problems are discussed from a predictive point of view. Influence statistics are defined and employed to ascertain the impact that particular observations will have on the inferential goals—allocation of future observations, separation between populations, and the determination of probabilities for future cases. Methods are illustrated on a collection of financial data taken from Johnson and Wichern (1982).

KEY WORDS:

• Influence measures
• Predictive distribution
• Discriminant analysis
• Posterior odds
• Outlier