Latent class bivariate model for the meta-analysis of diagnostic test accuracy studies

Tags: Oral
Eusebi P1, Reitsma J2, Vermunt J3
1Department of Epidemiology, Regional Health Authority of Umbria, Italy, 2Department of Clinical Epidemiology, Biostatistics and Bioinformatics, Academic Medical Center Amsterdam, Netherlands, 3Faculty of Social and Behavioural Sciences, Department Methodology and Statistics, University of Tilburg, Netherlands

Background: Several statistical methods for meta-analysis of data from diagnostic test accuracy studies have been proposed. The Bivariate Model is a rigorous method for this purpose by directly analyzing estimates of sensitivity and specificity.

Objectives: Our research is motivated by a re-analysis of data of a published systematic review (1). In this meta-analysis three imaging techniques were compared for the diagnosis of lymph node metastasis in women with cervical cancer. Observing the heterogeneity amount on the data we aim at implementing new methods for helping the understanding on the relationship between sensitivity and specificity.

Methods: We propose the Latent Class Bivariate Model, an extension of the Bivariate Model by means of a discrete latent variable for finding clusters of studies.

Results: Several type of models were fitted with Latent GOLD software (2). The best model is the Latent Class Bivariate Model with the type of test as a nominal covariate. This model detected two latent classes of studies. Studies belonging to the first latent class show lower sensitivity but higher specificity and almost all the studies are X-ray computed tomography (CT) and magnetic resonance imaging (MRI); in that class sensitivity and specificity appear to be negatively correlated. Studies belonging to the second latent class show lower specificity but higher sensitivity and almost all the studies are lymphangiography (LAG); in that class sensitivity and specificity appear not to be correlated.

Conclusions: What is added by the latent approach is that it provides an explanatory and confirmatory tool for investigating and testing different patterns of heterogeneity across studies. We tested the performance equivalence of CT and MRI and the different correlation between sensitivity and specificity in LAG and CT/MRI studies. Additional insight and data-driven hypothesis can be generated for future subgroup meta-analysis.

References

1. Scheidler J, Hricak H, Yu KK, Subak L, Segal MR. Radiological evaluation of lymph node metastases in patients with cervical cancer. A meta-analysis. JAMA. 1997;278(13):1096-101.

2. Vermunt JK, Magidson J. (2008). LG-Syntax User’s Guide: Manual for Latent GOLD 4.5 Syntax Module, Belmont, MA: Statistical Innovations Inc.