Methods/Statistics
Diagnostic accuracy of tests for SARS-CoV-2 acute infection: Distinguishing measurands from target conditions Joanna Merckx* Joanna Merckx Ian Schiller Yap Boum Patrick M Bossuyt Nandini Dendukuri
Test accuracy evaluation for SARS-CoV-2 infection is complicated by the lack of a perfect reference. Additionally, the available tests have different measurands further challenging performance estimation. The current literature reports that antibody test sensitivity varies according to time since symptom onset. Such reporting however conflates two separate parameters: the time-varying prevalence of antibodies (the measurand), and the time-invariant ability of the assay to measure antibodies. We improve the commonly used latent class model (LCM) via the decomposition of the diagnostic accuracy question into its elements: i) the tests under evaluation, ii) their measurands and iii) the target conditions. We use directed acyclic graphs (DAG) to visualize these elements. We use LC analysis (LCA) to model the relationships in the DAG and Bayesian inference to obtain estimates of accuracy of a PCR, antigen, and rapid IgM and IgG antibody test for the diagnosis of the target conditions acute and past SARS-CoV-2 infection and their prevalence. We introduce two random effects to capture the dependence between the measurands due to acute and past infection. We represent the results as posterior distribution medians with 95% credible intervals (CrI) and compare with a measurand naïve LC model. We apply our methods to a Cameroonian cohort of 1,194 adults tested at multiple-time points in the early pandemic. We estimated the prevalence of acute and past infection as 20% (95%CrI 17; 24) and 26% (95%CrI 20; 32), respectively. Sensitivity and specificity estimates are provided for all tests, measurands and their target conditions. We learned that by distinguishing target conditions and measurands and using a DAG to clarify their relations we were able to build a LCM with more insightful results on the performance of a range of COVID-19 tests. Similar models can enrich estimates of test accuracy and prevalence of infectious diseases specific target conditions in future studies.
