Bayesian analysis of influenza sero-epidemiology data

Abstract

Influenza viruses pose significant threats to public health globally. Accurate estimation of the cumulative infection incidence lays the basis for evaluating disease burden and clinical severity of influenza. The traditional approach to infer infection is by using a 4-fold or greater rise in antibodies between paired sera from longitudinal studies. However, this estimation often suffers from the non-bracketing issue – collecting sera after the start or before the end of an influenza epidemic – results in underestimation of influenza cumulative incidence. The estimation could also be biased due to waning of antibodies, but there are limited studies quantifying the antibody dynamics in the long term. The Hong Kong influenza vaccination study provided a valuable opportunity to address these two issues.

Data from the Hong Kong influenza vaccination study, 2009- 2014 were analyzed, where sera were tested for specific antibodies against influenza using Haemagglutination Inhibition Assay. A hierarchical Bayesian model was constructed to estimate cumulative incidence against influenza A(H1N1)pdm09 for children, accounting for the sera collection time. Influenza surveillance data from local health department was obtained to reveal the influenza virus activity in Hong Kong during the study period. Conditioned on the test results by Reverse Transcription Polymerase Chain Reaction, Bayesian models were built for capturing antibody dynamics by quantifying the declining rate and geometric mean fold rise in antibodies (boosting factor) after influenza infection. Waning rates and boosting factors were compared by influenza type/subtype (A(H1N1)pdm09, H3N2 and influenza B (Victoria-lineage)), age and prior TIV vaccination history. All parameters were estimated under a Bayesian framework using Markov chain Monte Carlo method.

The estimated cumulative incidence of A(H1N1)pdm09 infections in children varied considerably over the study period. It decreased considerably from 34.0 in 2009-10 influenza season to 6.7% in 2011-12 influenza season, and remained at 12.8% and 13.7% in the 2012-13 and 2013-14 influenza seasons, respectively, similar to the annual cumulative incidence of seasonal influenza of other types/subtypes. Antibodies against influenza B (Victoria-lineage) viruses showed the most rapid waning over time, compared with A(H1N1)pdm09 and H3N2 over all age groups. Vaccinated groups had a faster decline in antibodies than unvaccinated groups. On the other hand, strongest boosting in antibodies after infection with H3N2 was identified, compared with the other two types/subtypes. The boosting factors in unvaccinated groups doubled that of the vaccinated groups.

The annual cumulative incidence of A(H1N1)pdm09 among children declined to the similar level of seasonal influenza during the period of 2012-14 in Hong Kong. The bias in the estimation of cumulative incidence due to non-bracketing sera collected from cohort study could be corrected for through accounting for sera collection time in the Bayesian models, which provide a method to make full use of non-bracketing data rather than throw the data away. The long-term dynamics of influenza antibodies might be influenced by influenza type/subtype, age and prior TIV vaccination history, but could be characterized by using Bayesian models, which provides a method to do so based on serological data collected on regular intervals.