A Measure of Agreement for Interval or Nominal Multivariate Observations

Abstract

This article addresses the problem of accounting overall chance-corrected interobserver agreement among the multivariate ratings of several judges. Modifying an approach by Berry and Mielke, a generalization of Cohen’s kappa coefficient is proposed. The generalized statistic accounts agreement for multivariate interval or nominal observations among several judges. The statistic’s metric is conventional, and in the univariate case it is equivalent to existing extensions of the kappa coefficient to several observers, as well as to a certain intraclass correlation coefficient. Application is exemplified using hypothetical and real-life data sets.