Latent Class Models for Medical Studies with Replicate Observations: Alternative Latent Structures for Observer Agreement and Diagnostic Accuracy

• Activity Number: 200
• Type: Contributed
• Date/Time: Monday, August 8, 2005 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): Andrew Baughman*+ and Michael Haber and Huiman X. Barnhart
• Companies: Centers for Disease Control and Prevention and Emory University and Duke University
• Keywords: Latent class models ; Conditional dependence ; Observer agreement ; Diagnostic accuracy ; Replicate measurements ; Binary data

Abstract

In this paper, we develop a new latent class model, termed the three-stage model, for observer agreement studies with replicate binary measurements in the absence of a gold standard diagnostic test. The three-stage model integrates two previous latent class models that differ in their underlying latent structure (Baker, Freedman, and Parmar 1991; Dawid and Skene, 1979). The new model allows for simultaneous assessment of the performance of individual observers (i.e., sensitivity and specificity) and interobserver agreement. An alternative parameterization of the three-stage model, termed the intraobserver correlation model, is presented and shown to be mathematically equivalent to the three-stage model. The equivalence between these two models demonstrates that adding latent variables to a latent class model produces the same results as adding latent correlation parameters. The intraobserver correlation model provides a simple approach to modeling conditional dependence among replicate measurements. The three-stage model is applied to a published dataset of radiologists' readings of mammograms.