A pseudo-likelihood approach for estimating diagnostic accuracy of multiple binary medical tests

Authors

• Wei Liu, Department of Mathematics, Harbin Institute of Technology, Harbin
• Bo Zhang, Division of Biostatistics, Office of Surveillance and Biometrics, Center for Devices and Radiological Health, Food and Drug Administration, Silver Spring, MDFollow
• Zhiwei Zhang, Division of Biostatistics, Office of Surveillance and Biometrics, Center for Devices and Radiological Health, Food and Drug Administration, Silver Spring, MD
• Baojiang Chen, University of Nebraska Medical Center, Omaha, NE
• Xiao-Hua Zhou, University of Washington

Date of this Version

2015

Document Type

Article

Citation

Computational Statistics and Data Analysis 84 (2015) 85–98

Comments

U.S. Government Work

Abstract

Latent class models with crossed subject-specific and test(rater)-specific random effects have been proposed to estimate the diagnostic accuracy (sensitivity and specificity) of a group of binary tests or binary ratings. However, the computation of these models are hindered by their complicated Monte Carlo Expectation–Maximization (MCEM) algorithm. In this article, a class of pseudo-likelihood functions is developed for conducting statistical inference with crossed random-effects latent class models in diagnostic medicine. Theoretically, the maximum pseudo-likelihood estimation is still consistent and has asymptotic normality. Numerically, our results show that not only the pseudo-likelihood approach significantly reduces the computational time, but it has comparable efficiency relative to the MCEM algorithm. In addition, dimension-wise likelihood, one of the proposed pseudolikelihoods, demonstrates its superior performance in estimating sensitivity and specificity.