Rethinking statistical approaches for serological data analysis for viral surveillance

Abstract

A robust serological surveillance system for zoonotic pathogens is imperative for both early detection and advancing knowledge of emerging diseases. A statistical analysis plan that is aligned to research and epidemiological goals requires a purposeful choice among alternative methods for differentiating seronegative from seropositive samples, estimating seroprevalence, and estimating risk factors associated with seropositivity. The common standard deviation-based cutoff (e.g., 3sd) approach is simple to implement and understand, but fails to appropriately propagate uncertainty in serostatus assignments to any risk factor analysis. Methods such as Gaussian mixture models, which jointly estimate serostatus, risk factors, and their uncertainty, can alleviate the dichotomy created by the cutoff approach. Yet, because of a lack of empirical guidance of method performance, it remains difficult to choose a robust analysis method for a given serological dataset. Here we examine the performance of both cutoff and clustering approaches using simulated datasets that represent the epidemiological, biological, and immunological data generation process. We focus on understudied pathogens for which validated serological assays do not exist, as is common in emerging viruses in wildlife. We quantify coverage (the proportion of time 95 % confidence intervals contain the true value) and bias (systematic differences between true values and model point estimates) of model estimates for individual serostatus assignments, population seroprevalence, and regression coefficients for serostatus risk factors. In nearly all scenarios, Bayesian mixture models provide the highest coverage and lowest bias. Only with very low seroprevalence (∼ < 3 %) and large differences in signal between seronegative and seropositive individuals will a cutoff provide low bias and near-nominal coverage. Given poor coverage of risk factor regression coefficients, we advise against using a cutoff approach for quantifying determinants of seropositivity.