Social encounter networks: characterizing Great Britain

Abstract

A major goal of infectious disease epidemiology is to understand and predict the spread of infections within human populations, with the intention of better informing decisions regarding control and intervention. However, the development of fully mechanistic models of transmission requires a quantitative understanding of social interactions and collective properties of social networks. We performed a cross-sectional study of the social contacts on given days for more than 5000 respondents in England, Scotland and Wales, through postal and online survey methods. The survey was designed to elicit detailed and previously unreported measures of the immediate social network of participants relevant to infection spread. Here, we describe individual-level contact patterns, focusing on the range of heterogeneity observed and discuss the correlations between contact patterns and other socio-demographic factors. We find that the distribution of the number of contacts approximates a power-law distribution, but postulate that total contact time (which has a shorter-tailed distribution) is more epidemiologically relevant. We observe that children, public-sector and healthcare workers have the highest number of total contact hours and are therefore most likely to catch and transmit infectious disease. Our study also quantifies the transitive connections made between an individual's contacts (or clustering); this is a key structural characteristic of social networks with important implications for disease transmission and control efficacy. Respondents' networks exhibit high levels of clustering, which varies across social settings and increases with duration, frequency of contact and distance from home. Finally, we discuss the implications of these findings for the transmission and control of pathogens spread through close contact.

Keywords: epidemic; infectious disease; network; social contact; survey.

Figure 1.

(a) Spatial distribution of respondents in the GB (4689 individuals provided a valid postcode); dots are colour-codes, so that regions of the highest density are in red, whereas low-density regions are in blue. There is good agreement between the location of respondents and major urban areas. (b) Example of an egocentric network collected by our survey. (c) Distribution of household sizes from the postal (blue) and online (red) surveys, compared with the national average (grey) showing that households of size 1 and 2 are over-represented. (d) Proportion of the respondents of a particular age and gender from the postal (blue) and online (red) surveys; the black lines show the estimated GB population percentages for 2009. These highlight the lack of young children and that males below 60 are under-represented.

Figure 2.

Distributions of the number of contacts and total contact time per individual. (a,b) The correspondence between distributions from our survey distributions and distributions from other estimates of human contacts—in particular, the POLYMOD study [16], the SocioPatterns study [19] and the EpiSims model [20]. In (a), we show the frequency of respondents with relatively low numbers of contacts, whereas in (b), we plot the cumulative frequency on a logarithmic scale to provide a clearer visual representation and highlighting the tail of the distribution. (c) The distribution of the total contact time on a logarithmic scale, with the error bars showing the confidence intervals from 1000 bootstrapped samples, and (d) shows the relationship between the total contact time and the number of contacts, the blue points showing results for each respondent and the red line showing the mean values and confidence intervals from 1000 bootstrap samples.

Figure 3.

Relationship between age of respondent time, contact hours and clustering. (a) Box and whisker plots of the estimated total contact time, reported by age group. The black dots show median values, the grey boxes show the inter-quartile range and the whiskers extend to 1.5 times the size of the inter-quartile range from the quartiles; all points beyond that are considered to be outliers and are omitted for clarity. The blue line depicts the mean values together with the 95% confidence limits of those means derived from 1000 bootstrap samples for age ranges with more than five respondents. (b) The mean contact hours in each age group are partitioned into only those that involved touch (which may be considered more likely to pass infection; red) and those that are conversation only (black). (c) Box and whisker plots for the weighted clustering measured for each respondent.

Figure 4.

Heterogeneity in the number of contact hours by occupation. Box and whisker plot showing the median, quartiles and 95 percentiles of contact hours; occupations are ordered by median number of contact hours. For each occupation category, we show the number of respondents in brackets (work days and non-work days) and distinguish between days at work or school (red), from non-working days (green); for some occupations (pre-school, home, retired and unemployed, shown in grey), such a distinction is not possible. For each category, where the contact hour distributions are significantly greater or less than the total sampled population they are shown with a left-facing triangles or right-facing triangles symbol, respectively, and with an circles denotes when no significant difference is observed.

Figure 5.

The variation in total contact hours and weighted clustering with other covariates measured in the survey. The total contact hours (red and top x-axis) depicts the average total time a respondent spends with contacts of a particular type. Weighted clustering of respondents ego networks (blue and lower x-axis), captures the proportion of transitive links between contacts of a given type. Confidence intervals are calculated by bootstrapping from the respondent sample and duration per contact.