Mathematical models are increasingly used to evaluate and inform infectious disease prevention and control policy. At their heart all models must make assumptions about how individuals contact each other and transmit the infectious agent. Until now, modellers have relied on proxy measures of contacts and calibration to epidemiological data. For instance, household size, class size, transport statistics, and workplace size distribution have been used in recent models to define the contact structure [
2,
3,
33,
34]. Our study complements those relying on proxy measures by using direct estimates of the number, age, intimacy levels, and distribution of actual contacts within various settings. The analysis of population-based contact patterns can help inform the structure and parameterisation of mathematical models of close-contact infectious diseases.
One of the most important findings of our study is that the age and intensity patterns of contact are remarkably similar across different European countries even though the average number of contacts recorded differed. This similarity implies that the results may well be applicable to other European countries, and that the initial phase of spread of newly emerging infections in susceptible populations, such as SARS was in 2003, is likely to be very similar across Europe and in countries with similar social structures.
Another major insight gained from our study comes from the observation that the contacts made by children and adolescents are more assortative than contacts made by other age groups. That is, most of the individuals contacted by children and teenagers are of very similar age, and these contacts tend to be of long duration. This pattern is likely to be the main reason why children and teenagers are and have been an important conduit for the initial spread of close-contact infections in general and for influenza in particular [
11,
14] and our preliminary modelling work confirms this.
Our study allows us to assess and quantify the risk of transmission in different settings. We took a number of different measures of “closeness of contact,” including duration and frequency of contact and whether skin-to-skin contact occurred. These measures correlated highly with each other, such that the longer-duration contacts tended to be frequent and to involve physical contact (and vice versa). More-intimate contacts are likely to carry a greater risk of transmission. Furthermore, these types of contact tend to occur in distinct social settings: the most-intimate contacts occur at home or in leisure settings, whereas the least-intimate tend to occur while travelling. Thus, the risk of infection in these settings can be inferred to vary. This variation has important implications for contact tracing during outbreaks of a new infection. Our results suggest that if efforts concentrate on locating contacts in the home, school, workplace, and leisure settings, on average more than 80% of all contacts would be found.
We have used simulations to expand on two particular types of contacts (physical and nonphysical) and to sketch the consequences of the observed contact patterns on the age distribution of incidence in the initial phase of an epidemic, when a new infectious disease is introduced into a completely susceptible population. As shown clearly by our simulations, the highest incidence of infection will occur among the younger age classes (5–19 y) for all countries. It is tempting to link such contact patterns to the observation during the 1957 Asian influenza A H2N2 pandemic that the first few generations of infection primarily affected those aged 11–18 y [
35]. However, we note that our survey did not address the clustering of contacts; such clustering of contacts might result in less-pronounced differences in age-specific incidence than suggested by our calculations. Addressing the frequency of clustered contacts, duration and type of contact, differential impact of pathogen on different age groups, time correlation of contacts, and assortative mixing by demographic factors other than age should be key priorities for future research.
One of the major assumptions behind our approach is that talking with or touching another person constitutes the main at-risk events for transmitting infectious diseases. There may be other at-risk events that our methodology does not capture, such as being in a confined space or in close physical proximity with other individuals and not talking to them [
23]. Such events are difficult to record or to measure without using intrusive and expensive surveillance methods, and are probably of lower risk than the communication events captured by our approach. Similarly, our framework does not apply to pathogens that, in addition to the respiratory route, can be also spread by other means, for example, the sewage contamination events for SARS [
8]. Although we believe that it is plausible that the contact patterns observed in our study are predictive of disease transmission, further work is clearly needed to establish the types of contacts that represent transmission risks for different diseases and to determine the circumstances under which lower-intensity contacts could be epidemiologically relevant. The data reported in this study should not be considered a substitute for epidemiological studies that quantify, for instance, the intensity of transmission of influenza in households, schools, or other settings. However, this study does provide invaluable data on the relative importance of “leisure” and “other” contacts, which are very difficult to assess in other ways, and it highlights the relatively small contribution of personal contacts during travel based on our approach of defining a contact.
Using contact diaries in the general population was a feasible method for our specific study objectives, but as with all self-reported data, future research should validate our findings with different approaches, including interviews or direct observation. The latter might be particularly useful in assessing contacts of young children who spend time in day-care centres and kindergartens, because parental proxy reporting for young children is likely to be problematic. Despite the limitations of self-reported egocentric data [
36], contact diaries can provide extensive details regarding contact structures and have been used successfully for social network analysis [
37]. Our contact diaries yielded detailed information about intimacy, frequency, and epidemiological relevance of contacts with an acceptable burden on respondents. In five countries, participants were given the opportunity to report whether they had any problems filling in the diary. The low proportion reporting problems (4% in adults, 4.9% in older children self-reporting contacts, and 4.9% in parents as proxy for children) suggest that the contact diary was readily accepted and understood by responding participants.
A further limitation of our study is that the comparison of contact patterns between countries is complicated by the variations of diary design (see
Table S1), recruitment, and follow-up methodology (see
Table S1). Our surveys were conducted in each country by different commercial companies with different recruitment and follow-up methods. Conducting surveys on contact behaviour and networks that entail a certain burden on participants and follow identical methodology in different countries is a challenging task, given that cultural factors in response also play a role. Further research is definitely warranted to determine optimal survey methodologies in different international settings, including developing countries, to improve comparability of contact data. Diaries used in BE, DE, FI, and NL instructed respondents not to record all of their professional contacts, but to provide an estimate if they had a lot of them. The reason for this instruction was to try to capture information from those people who make very large numbers of contacts (shop assistants and bus drivers, for instance), given that it might be very difficult or impossible for such people to fill out the full contact diary. This instruction may have lead to some underreporting of contact frequencies and thus have affected the distribution of age and circumstance of contacts for these four countries, although we have taken account of this possibility to some extent using a censored model. Additional analyses for these countries that combine and compare the estimated frequency of professional contacts with the diary data will provide additional insights about the number of contacts for all countries. The differences between diaries do not, however, affect the age-specific pattern, nor the similarity in age-specific patterns found across countries.
Our survey is, to our knowledge, the first population-based prospective survey of mixing patterns pertinent to the spread of airborne and close-contact infectious diseases performed in several European countries using a similar diary methodology. The quantification of these mixing patterns shows a remarkable similarity in degree of assortativeness, which likely results in similar patterns of spread in different populations. This finding represents a significant advance in our understanding of the spread of these infectious diseases and should help to improve the parameterisation of mathematical models used to design control strategies.
See "
is an indicator variable for censoring, yi is the number of observed contacts, Xi is the vector of explanatory variables, and P is the probability function of the negative binomial distribution:
c} and {Y,Yc}. Using both constraints, the set of rules can further be filtered by the lift, which is defined as the ratio of the relative frequency of X ∩ Y and the product of relative frequencies of X and Y. The lift can be interpreted as the ratio of the rule's observed support to the support expected under independence. Greater lift values indicate stronger associations. Additionally, a Chi-square test for the rule-corresponding two-by-two table consisting of cells X ∩ Y, X
