Clustering of contacts relevant to the spread of infectious disease

Objective Infectious disease spread depends on contact rates between infectious and susceptible individuals. Transmission models are commonly informed using empirically collected contact data, but the relevance of different contact types to transmission is still not well understood. Some studies select contacts based on a single characteristic such as proximity (physical/non-physical), location, duration or frequency. This study aimed to explore whether clusters of contacts similar to each other across multiple characteristics could better explain disease transmission. Methods Individual contact data from the POLYMOD survey in Poland, Great Britain, Belgium, Finland and Italy were grouped into clusters by the k medoids clustering algorithm with a Manhattan distance metric to stratify contacts using all four characteristics. Contact clusters were then used to fit a transmission model to sero-epidemiological data for varicella-zoster virus (VZV) in each country. Results and discussion Across the five countries, 9–15 clusters were found to optimise both quality of clustering (measured using average silhouette width) and quality of fit (measured using several information criteria). Of these, 2–3 clusters were most relevant to VZV transmission, characterised by (i) 1–2 clusters of age-assortative contacts in schools, (ii) a cluster of less age-assortative contacts in non-school settings. Quality of fit was similar to using contacts stratified by a single characteristic, providing validation that single stratifications are appropriate. However, using clustering to stratify contacts using multiple characteristics provided insight into the structures underlying infection transmission, particularly the role of age-assortative contacts, involving school age children, for VZV transmission between households.