Some Uses of Order Statistics in Bayesian Analysis

by Seymour Geisser
Technical Report No. 604
School of Statistics
University of Minnesota
February, 1995

Research supported in part by grant NIGMS 25271.


Introduction

Order statistics, although playing a prominent role in frequentist methodology, especially in nonparametric inference, are not often featured in Bayesian analysis. One area, however, where order statistics can be of interest to Bayesians is in the detection of outliers or discordant observations. These situations are such that there is an observation that appears to be somewhat removed from the remaining ones but no discernible alternative can readily be specified for the potential discordancy. Alternatives, although sometimes considered, Dixit (1994), require additional distributional assumptions and prior probabilities over and above the original model assumptions that initially an investigator may not be prepared to contemplate. In these situations a simple Bayesian test of significance may be appropriate in determining whether an observation or several of them are discordant or if it is necessary to contemplate alternative model assumptions for the entire data set. Sections 2 through 6 will consider Bayesian discordancy testing. Another area involves situations where the calculation of the probability that the minimum (maximum) of a set of future observables is greater (smaller) than some critical threshold. More generally we shall be interested in the chance that R out of M future values are in some interval or some set. This essentially involves the Rth future order statistic. This will be the subject of sections 7 and 8.


Click here to download the complete PostScript technical report.