An empirical Bayes method for serotype case-carrier ratios, with an application to Group B streptococcus

Abstract

ABSTRACTBackgroundCase-carrier ratios quantifying the relative pathogenicity of serotypes can inform vaccine formulations for antigenically-diverse pathogens. However, sparse serotype-specific counts in epidemiologic datasets may undermine such analyses, most notably for rare serotypes that pose emergence risks in vaccinated populations. This challenge is well-illustrated in Group B streptococcus (GBS), where serotype III dominates in both carriage and disease.MethodsWe develop an empirical Bayes random-effects model based on conjugate Dirichlet-multinomial distributions of serotype frequencies in carriage and disease states. We validate the model using simulated datasets, and apply it to data from 15 paired sets of GBS isolates from intrapartum rectovaginal colonization (n=3403) and neonatal invasive disease (NID; n=1088), 16 from blood (n=2352) and cerebrospinal fluid (n=780) neonatal specimens, and 3 from fatal (n=173) and non-fatal (n=1684) neonatal invasive infections.ResultsOur method accurately recovers parameters in simulated datasets. Using this approach, we confirm that GBS serotype III exhibits the greatest invasiveness, followed by serotype Ia with a 75.3% (95%CrI: 43.7-93.8%) lower estimate. Enhanced invasiveness of serotypes III and Ia is most evident in late-onset disease. Non–hexavalent-vaccine serotypes, which are rare in carriage and disease, generally show lower invasiveness; serotype IX/non-typeable GBS, the most prevalent cause of non–vaccine-preventable disease, is 98.7% (81.7-99.9%) and 94.2% (13.9-99.6%) less invasive than serotypes III and Ia, respectively.ConclusionsWe present a strategy for measuring associations of serotype with carrier and disease states in the presence of sparse counts, avoiding biases that exist in common ad-hoc approaches.