An Iterative, Frequentist Approach for Latent Class Analysis to Evaluate Conditionally Dependent Diagnostic Tests

Area of Science:

• Veterinary medicine
• Biostatistics
• Diagnostic test evaluation

Background:

• Latent class analysis (LCA) is crucial for evaluating diagnostic test accuracy without a gold standard.
• A key LCA assumption is conditional independence of tests, often violated when tests share biological principles.
• Ignoring test dependencies can lead to inaccurate accuracy estimates in veterinary diagnostics.

Purpose of the Study:

• To extend the traditional latent class model to incorporate conditional dependencies between diagnostic tests.
• To develop a robust iterative algorithm for estimating test accuracies and prevalence when dependencies exist.
• To compare the performance of the new approach with existing Bayesian methods.

Main Methods:

• Developed an iterative algorithm to estimate conditional dependencies and test accuracies simultaneously.
• The algorithm refines estimates of dependencies and accuracies until model convergence.
• Simulated five scenarios based on veterinary diagnostic tests to validate the approach.

Main Results:

• The proposed iterative method and Bayesian approaches yielded similarly precise results.
• The new method provided more accurate results than the Bayesian approach when initial test accuracies were misjudged.
• Bayesian methods were more precise when incorrect dependencies were assumed in the new approach.

Conclusions:

• The developed iterative algorithm is a valuable addition to existing Bayesian methods for diagnostic test accuracy evaluation.
• This approach offers simpler and more objective estimations, particularly when test dependencies are present.
• The method effectively addresses the limitations of traditional latent class analysis in veterinary medicine.