A latent variable model for the analysis of variability in the classification of radiographs of pneumoconioses

The interpretation of chest radiographs for pneumoconioses is subjective and susceptible to inter- and intra-observer variability. The Istatistic has traditionally been used to assess interobserver agreement. In this paper a different paradigm for the analysis of radiographic readings for pneumoconiosis is introduced. This framework involves log-linear modelling and considers the radiograph and the reader together. The results of the analysis are estimates of the relative 'abnormal appearance' of each radiograph and the relative tendency of each reader to read 'high' or 'low'. Log-linear analyses could be useful to ensure quality and to place a reader's codings in the context of those of his peers.