Measurement errors in caries diagnosis: some further latent class models

Area of Science:

• Statistics
• Dental Diagnostics
• Psychometrics

Background:

• Latent class models are used for analyzing relative errors in discrete measurements, such as caries diagnosis.
• Previous models fitting caries data required interaction terms between raters, challenging the assumption of local stochastic independence.

Purpose of the Study:

• To investigate an alternative approach to latent class models for caries diagnosis data.
• To explore if increasing the number of latent classes can achieve similar or better model fit without rater interaction terms.

Main Methods:

• Analysis of empirical data from caries diagnosis (N = 3,869 teeth, 5 dentists).
• Comparison of an unrestricted three-class model (generalizing Carlos-Senning assumptions) with a restricted four-class model (resembling latent distance models).
• Evaluation of model fit using statistical criteria.

Main Results:

• Both the three-class and four-class models provided good to excellent fit to the caries data.
• The restricted four-class model demonstrated an excellent fit, comparable to latent distance models.
• Increasing the number of classes, with parameter restrictions, achieved the desired model fit without rater interactions.

Conclusions:

• Increasing latent classes is a viable alternative to including rater interaction terms for improving model fit in discrete measurement analysis.
• The findings support the utility of latent class models and latent distance models in dental diagnostics.
• The existence of multiple solutions for the four-class model highlights the need for caution regarding model identifiability.