ABSTRACT
Latent-class analysis can be used to analyze data obtained from multiple raters who independently assign the same set of targets to nominal categories. In this situation, the assessment of the degree to which raters agree is of considerable interest because it reflects the reliability of the classifications. Two different latent-class analysis approaches have been suggested for analyzing nominal rater agreement data. The first approach assumes that disagreement among raters occurs because of response error. The second approach distinguishes two different types of targets. I refer to the two approaches as the response-error approach and the targettype approach, respectively. I show that the response-error approach yields estimates of sensitivities as well as rater-specific and category-specific error rates. The response–error approach does not, however, yield a simple overall summary of the rater reliability. Although the target–type approach does not yield rater–specific or category–specific conclusions that ignore the target type, it yields an intuitively appealing reliability statistic that can be interpreted as the proportion of systematic agreement.
