Detecting departures from the conditional independence assumption in diagnostic latent class models: a simulation...

Area of Science:

• Biostatistics
• Statistical modeling
• Diagnostic accuracy studies

Background:

• Latent class models estimate diagnostic accuracy without a gold standard.
• Assuming test independence can bias results when dependencies exist.
• Residual correlation plots and chi-squared statistics are used to check for test dependencies.

Purpose of the Study:

• To evaluate the performance of residual correlation plots and chi-squared statistics in identifying dependent diagnostic tests within latent class models.
• To assess the reliability of these tools across various simulation scenarios.

Main Methods:

• Simulated 504 data set combinations varying sample size, prevalence, covariance, sensitivity, and specificity.
• Generated 1000 datasets per combination from a model with four tests and dependence between tests 1 and 2.
• Fitted a conditional independence model in a Bayesian framework and analyzed bias, coverage, and lack-of-fit detection.

Main Results:

• Residual correlation plots and chi-squared statistics correctly identified the dependent test pair only ~10-12% of the time.
• These tools incorrectly flagged tests 3 and 4 as dependent in over 50% of simulations.
• Lack of overall fit was detected in 64-74% of models, and the conditional independence model showed biased estimates for correlated tests.

Conclusions:

• Residual correlation plots and chi-squared statistics are unreliable for identifying conditional dependence between diagnostic tests.
• These methods have low power for detecting overall model fit issues.
• Failure to account for test dependence leads to significant parameter estimation bias.