The contribution of social behaviour to the transmission of influenza A in a human population

Abstract

Variability in the risk of transmission for respiratory pathogens can result from several factors, including the intrinsic properties of the pathogen, the immune state of the host and the host's behaviour. It has been proposed that self-reported social mixing patterns can explain the behavioural component of this variability, with simulated intervention studies based on these data used routinely to inform public health policy. However, in the absence of robust studies with biological endpoints for individuals, it is unclear how age and social behaviour contribute to infection risk. To examine how the structure and nature of social contacts influenced infection risk over the course of a single epidemic, we designed a flexible disease modelling framework: the population was divided into a series of increasingly detailed age and social contact classes, with the transmissibility of each age-contact class determined by the average contacts of that class. Fitting the models to serologically confirmed infection data from the 2009 Hong Kong influenza A/H1N1p pandemic, we found that an individual's risk of infection was influenced strongly by the average reported social mixing behaviour of their age group, rather than by their personal reported contacts. We also identified the resolution of social mixing that shaped transmission: epidemic dynamics were driven by intense contacts between children, a post-childhood drop in risky contacts and a subsequent rise in contacts for individuals aged 35-50. Our results demonstrate that self-reported social contact surveys can account for age-associated heterogeneity in the transmission of a respiratory pathogen in humans, and show robustly how these individual-level behaviours manifest themselves through assortative age groups. Our results suggest it is possible to profile the social structure of different populations and to use these aggregated data to predict their inherent transmission potential.

Figure 1. Schematic of model framework.

By dividing the population into different numbers of age groups and contact classes, it was possible to recreate a number of commonly used model structures. If only one age groups and one contact classes were included, the framework produced a simple mass-action model, in which all individuals had the same risk of infection. When there was only one contact class in each age group, we obtained an age-structured model. Alternatively, when only one age group was used, risk of infection depended only on the contact class an individual was in.

Figure 2. Risk of infection in different models.

(A) Possible model structures. Given the size of the Hong Kong dataset, the maximum possible number of age and/or contact groups in a particular model was limited 60. (B) and (E) Results from model X, which has 20 age groups, each containing one contact class. Each point represents one of the 762 individuals surveyed, with position based on reported age and total number of contacts, and colour showing risk of infection predicted by the model. (C) and (F) Results from model Y (1 age group with 20 contact classes). (D) and (G) Results from model Z (5 age groups, each with 5 contact classes). Models are either based on all reported contacts (B, C and D), or close contacts only (E, F and G). R₀ = 1.5.

Figure 3. Comparison of different models in Figure 1A.

(A) Model based on all contacts with relative susceptibility of over-18s, α, equal to one. (B) Model based on close contacts with α = 1. (C) Model based on all contacts with variable α. (D) Model based on close contacts with variable α. Colour shows model support under the Akaike Information Criterion (AIC). Note that models with numerous contact classes in B and D had some classes consisting solely of individuals – some of whom had been infected – that had no reported close contacts. The likelihood of such people seeing infection given the model assumptions was zero; the difference in AIC was therefore infinite.

Figure 4. Comparison of model fits to data, with classes sorted by empirically observed risk of infection.

Thick blue line, model prediction; light blue bars, data. Error bars give 95% binomial confidence interval. (A) Model based on all contacts with 10 age groups and 1 contact class in each. (B) Model based on close contacts with 10 age groups and 1 contact class in each. (C) Model based on all contacts with 10 age groups and 2 contact classes. (D) Model based on close contacts with 10 age groups and 2 contact classes in each. All models have variable relative susceptibility in the over-18s.

Figure 5. The social resolution of influenza transmission.

(A) Detailed analysis of AIC for models with age structure only and α = 0 (i.e. top rows in Figure 3A–B), with transmission based on: red, total contacts; blue, close contacts. (B) AIC for age-structured models with variable α. (C) Performance of best-supported model in Figure 4B, which has 20 age groups and transmission based on close contacts, against data. Light grey bars show observed proportion of individuals that are seropositive, with 95% binomial confidence interval given by error bars. Blue solid line shows model prediction. (D) Comparison of residuals for model in Figure 4C (blue line) and equivalent model with 35 age groups (green line).