Social Contacts and Mixing Patterns Relevant to the Spread of Infectious Diseases — TEI-tagged view

Joël MossongpersName, Niel HenspersName, Mark JitpersName, Philippe BeutelspersName, Kari AuranenpersName, Rafael MikolajczykpersName, Marco MassaripersName, Stefania SalmasopersName, Gianpaolo Scalia TombapersName, Jacco WallingapersName, Janneke HeijnepersName, Malgorzata Sadkowska-TodyspersName, Magdalena RosinskapersName, W. John EdmundspersName (2008date). “Social Contacts and Mixing Patterns Relevant to the Spread of Infectious Diseases.” PLoS Medicinetitle·j.

Object of study — age-, intensity-, duration-, frequency-, and setting-specific social-contact patterns relevant to respiratory and close-contact infection spread

Research question — what empirically observed contact patterns characterize European populations and what age-specific epidemic incidence follows if infections spread through those contacts

Contribution — the first large-scale, multicountry, population-based prospective diary surveyterm·mth of epidemiologically relevant social contactsterm, with contact matrices and initial-epidemic simulations for model parameterisation

Full text — tagged

Social Contactsterm and Mixing Patternsterm Relevant to the Spread of Infectious Diseases Abstract Mathematical modelling of infectious diseases transmitted by the respiratory or close-contact routeterm (e.g., pandemic influenza) is increasingly being used to determine the impact of possible interventions. Although mixing patternsterm are known to be crucial determinants for model outcome, researchers often rely on a priori contact assumptions with little or no empirical basis. We conducted a population-based prospective surveyterm·mth of mixing patternsterm in eight European countries using a common paper-diary methodology.

7,290 participants recorded characteristics of 97,904 contacts with different individuals during one day, including ageterm, sexterm, locationterm, durationterm, frequencyterm, and occurrence of physical contactterm. We found that mixing patternsterm and contact characteristics were remarkably similar across different European countries. Contact patterns were highly assortative with ageterm: schoolchildrenterm and young adults in particular tended to mix with people of the same ageterm. Contacts lasting at least one hour or occurring on a daily basis mostly involved physical contactterm, while short durationterm and infrequent contacts tended to be nonphysical. Contacts at home, school, or leisure were more likely to be physical than contacts at the workplace or while travelling. Preliminary modelling indicates that 5- to 19-year-olds are expected to suffer the highest incidence during the initial epidemic phase of an emerging infection transmitted through social contactsterm measured here when the population is completely susceptible.

To our knowledge, our study provides the first large-scale quantitative approach to contact patterns relevant for infections transmitted by the respiratory or close-contact routeterm, and the results should lead to improved parameterisation of mathematical modelsterm used to design control strategiesterm.

Introduction Preparing for outbreaks of directly transmitted pathogens such as pandemic influenza [ 1 – 3 ] and SARS [ 4 – 9 ], and controlling endemic diseases such as tuberculosis and meningococcal diseases, are major public health priorities. Both can be achieved by nonpharmaceutical interventions such as school closure, travel restrictions, and contact tracingterm, or by health-care interventions such as vaccination and use of antiviral or antibiotic agents [ 2 , 10 – 13 ]. Mathematical modelsterm of infectious disease transmission within and between population groups can help to predict the impact of such interventions and inform planning and decision making. Contact ratesterm between individuals are often critical determinants of model outcomes [ 14 ]. However, few empirical studies have been conducted to determine the patterns of contact between and within groups and in different social settings.

In comparison to HIV and sexually transmitted diseases [ 15 – 17 ] and drug/needle sharing networks [ 18 ], where a number of large-scale empirical studies have been conducted on contact patterns, relatively little effort has been devoted to infections spread by respiratory droplets or close contact. Instead, the contact structureterm for these infections has been assumed to follow a predetermined pattern governed by a small number of parameters that are then estimated using seroepidemiological data [ 19 , 20 ]. A small number of studies have attempted to directly quantify such contact patterns, but they were conducted in small or nonrepresentative populations [ 14 , 21 – 25 ]. Hence, it is unclear to what extent the results can be generalized to an overall population and across different geographical areas. To address this lack of empirical knowledge, we present here results from, to our knowledge the first, large-scale, prospectively collected, population-based surveyterm of epidemiologically relevant social contactterm patterns. The study was conducted in eight different European countries using a common paper diary approach and covering all ageterm groups. We use these data to assess how an emerging infection could spread in a wholly susceptible population if it were transmitted by the social contactsterm measured here.

Methods Surveyterm·mth Methodology Information on social contactsterm was obtained using cross-sectional surveysterm·mth conducted by different commercial companies or public health institutes in Belgium (BE), Germany (DE), Finland (FI), Great Britain (GB), Italy (IT), Luxembourg (LU), The Netherlands (NL), and Poland (PL). The recruitment and data collection were organised at the countryterm·var level according to a common agreed quota sampling methodologyterm and diary design. The surveysterm·mth were conducted between May 2005 and September 2006 with the oral informed consent of participants and approval of national institutional review boards following a small pilot study to test feasibility of the diary design and recruitment [ 26 ].

Surveyterm·mth participants were recruited in such a way as to be broadly representative of the whole population in terms of geographical spread, ageterm, and sexterm. In BE, IT, and LU, surveyterm·mth participants were recruited by random digit dialling using land line telephones; in GB, DE, and PL surveyterm·mth participants were recruited through a face-to-face interview; surveyterm·mth participants in NL and FI were recruited via population registers and linked to a larger national sero-epidemiology surveyterm·mth in NL. Children and adolescents were deliberately oversampled, because of their important role in the spread of infectious agents. For more details on the surveyterm·mth methodology in the various countries, see Table S1 .

Briefly, only one person in each household was asked to participate in the study. Paper diariesterm were either sent by mail or given face to face to participants. Participants were coached by telephone or in person on how to fill in the diary.

Diaries recorded basic sociodemographic information about the participant, including employment status, level of completed education, household composition, ageterm, and sexterm. Participants were assigned a random day of the weekterm to record every person they had contact with between 5 a.m. and 5 a.m. the following morning. Participants were instructed to record contacted individuals only once in the diary. A contact was defined as either skin-to-skin contact such as a kiss or handshake (a physical contactterm), or a two-way conversation with three or more words in the physical presence of another person but no skin-to-skin contact (a nonphysical contactterm). Participants were also asked to provide information about the ageterm and sexterm of each contact person. If the ageterm of a contact person was not known precisely, participants were asked to provide an estimate of the ageterm range (the midpoint was used for data analysis). For each contact, participants were asked to record locationterm (home, work, school, leisure, transport, or other), the total durationterm of time spent together (less than 5 min, 5–15 min, 15 min to 1 h, 1–4 h, or 4 h or more) as well as the frequencyterm of usual contacts with this individual (daily or almost daily, about once or twice a week, about once or twice a month, less than once a month, or for the first time).

Diaries were translated into local languages (see Text S1 for the diary used in GB) and are available on request in the following languages: Dutch, English, French, Finnish, German, Italian, Polish, Portuguese, and Swedish. Diaries for young children were filled in by a parent or guardian on their behalf. Older children who obtained parental consent were given diaries with simplified language to fill in on their own (see Table S1 for more details).

Data Analysis Main effects of covariates (ageterm, sexterm, household sizeterm, and countryterm·var) on numbers of contacts were assessed using multiple censored negative binomial regressionterm [ 27 ]. The data were right censored at 29 contacts for all countries because of a limited number of possible diary entries in some countries. Additionally, a sensitivity analysis was performed to assess the effects of different handling of professional contacts between the countries.

The log-likelihood function ll for the censored negative binomial was where w i is the weight of observation i, is an indicator variable for censoring, y i is the number of observed contacts, X i is the vector of explanatory variables, and P is the probability function of the negative binomial distribution: where μ = exp( X i β); β is the vector of coefficients and α is the overdispersion parameter.

Sampling weightsterm—the inverse of the probability that an observation is included because of the sampling design—were calculated for each countryterm·var separately, based on official ageterm and household sizeterm data of the year 2000 census round data published by Eurostat ( http://epp.eurostat.ec.europa.eu/ ) (see Table S2 ) and used to correctly estimate population-related quantities. Overall statistics should be considered indicative of general trends and levels, but specific statistical representativity for the whole of Europe is not claimed, since participating countries, although geographically and socially diverse, are not a representative or random selection at the European level.

Association Ruleterm Analysis Mining association rulesterm is a tool for discovering patterns between variables in large databases [ 28 ]. Let X , Y denote disjoint nonempty items in the contact surveyterm·mth, such as daily frequencyterm, durationterm of more than 4 h, and physical contactterm. Association rulesterm are rules of the form X → Y that measure how likely the event Y is, given X . In this context X is called antecedent while Y is called consequent. Rules are typically extended to include more items in the antecedent but are restricted to include only one item in the consequent. The length of the rule is defined as the total number of items in both antecedent and consequent.

Selecting interesting rules from the set of all possible rules is based on various measures of significance and interest. The best-known are supportterm, confidenceterm, and liftterm. The supportterm of an association ruleterm X → Y is defined as the relative frequencyterm of X ∩ Y . Finding rules with high supportterm can be seen as a simplification of the learning problem called “mode finding” or “bump hunting.” The confidenceterm of a rule is the conditional probability P ( Y | X ) indicating what percentage of times the rule holds and thus measuring the association between { X , X c } and { Y , Y c }. Using both constraints, the set of rules can further be filtered by the liftterm, which is defined as the ratio of the relative frequencyterm of X ∩ Y and the product of relative frequencies of X and Y . The liftterm can be interpreted as the ratio of the rule's observed supportterm to the supportterm expected under independence. Greater liftterm values indicate stronger associations. Additionally, a Chi-square testterm for the rule-corresponding two-by-two table consisting of cells X ∩ Y, X c ∩ Y, X ∩ Y c , X c ∩ Y c , where c refers to the complementing set of items, can be used to test statistical significance of the association. Whenever the Chi-square distribution seemed inappropriate due to small sample size, a Fisher exact testterm was used. For a more extensive overview of applying association rulesterm on contact data see [ 29 ].

Contact Surface Smoothingterm Contact surface smoothingterm was performed by applying a negative binomial model on the aggregated number of contactsterm (both physical and nonphysical) over 5 y age bandsterm for both responders and contacts using a tensor product splineterm as a smooth interaction term [ 30 , 31 ].

Epidemiological Modelling: Simulating the Initial Phase of an Epidemicterm We explore the age-specific incidenceterm of infection during the initial phase of an epidemicterm of an emerging infectious disease agent that spreads in a completely susceptible populationterm. We focus on the generic features of epidemic spread along the transmission route that is specified by physical and nonphysical contactsterm as defined here. We partition the population into 5 y age bandsterm, and we group all individuals aged 70 y and older together. This process results in 15 ageterm classes. We denote the number of at-risk contacts of an individual in ageterm class j with individuals in ageterm class i by k ij . We take k ij as proportional to the observed number of contactsterm (both physical and nonphysical) that a respondent in age bandterm j makes with other individuals in age bandterm i . The matrix with elements k ij is known in infectious disease epidemiology as the next generation matrixterm K [ 32 ]. The next generation matrixterm can be used to calculate the distribution of numbers of new cases in each generation of infection from any arbitrary initial number of introduced infections. For example, when infection is introduced by one single 65-y-old infected individual into a completely susceptible populationterm, we can denote the number of initial cases in generation 0 by the vector x 0 = (0,0,0,0,0,0,0,0,0,0,0,0,0,1,0) T . The expected numbers of new cases in the i th generation are denoted by the vector x i , and this vector is calculated by applying the next generation matrixterm K i times to the initial numbers of individuals x 0 , that is, x i = K i x 0 . For large i , the vector x i will be proportional to the leading eigenvectorterm of K . We find that, in practice, the distribution of new cases is stable after five generations; that is, the distribution no longer depends on the precise ageterm of the initial case. The incidence of new infections per age bandterm is obtained by dividing the expected number of new cases per ageterm class by the number of individuals in each ageterm class. To facilitate comparison among countries, we normalized the distribution of incidence over ageterm classes such that for each countryterm·var the age-specific incidencesterm sum to one.

Results Description of Sample A total of 7,290 diaries covering all contacts made by respondents during a full day were collected in eight countries ranging from 267 in NL to 1,328 in DE (see Table 1 ). 37.6% of participants in our surveyterm·mth were under 20 y of ageterm, 12.4% of participants were over 60 y of ageterm, and the medians were 28 y in BE (the lowest) to 33 y in DE (the highest). Returns of diaries by female participants showed a slight excess in all countries (ranging from 50.8% in FI to 55.7% in DE). In all countries except DE, single-person households were underrepresented in our sample ( Table S2 ). This can be partially explained by the fact that children and adolescents were deliberately oversampled, and they are more likely to live in larger households.

Number of Recorded Contacts per Participant per Day by Different Characteristics and Relative Number of Contactsterm from the Weighted Multiple Censored Negative Binomial Regressionterm Model Overall, 35.3% of the participants were in full-time education, 32.6% employed, 11% retired, 6.1% home-makers, 3.6% unemployed or seeking employment, whereas 8.6% recorded “other,” and 2.8% failed to record their occupation. The proportion employed or in full-time education was fairly consistent across the eight countries; the other categories differed somewhat between countries.

Number of Contactsterm A total of 97,904 contacts with different persons were recorded (mean = 13.4 per participant per day) in the diaries. On average, German participants reported the fewest daily number of contactsterm (mean = 7.95, standard deviation [SD] = 6.26) and Italians the highest number (mean = 19.77, SD = 12.27). The contact distributions in all countries are slightly skewed, the skewness statistics ranging from 0.62 in IT to 2.96 in DE ( Figure S1 ). Analysis of the total number of reported contacts with a multiple regression model shows a consistent pattern of contact frequencyterm by ageterm, with a gradual rise in the number of contactsterm in children, a peak among 10- to 19-y-olds, followed by a fall to a lower plateau in adults until the ageterm of 50 and a sharp decrease after that ageterm ( Table 1 ). Living in a larger household sizeterm was associated with higher number of reported contacts. Weekdays were associated with 30%–40% more contacts than Sundays. The influence of the countryterm·var in which the surveyterm·mth was performed was also apparent ( Table 1 ), even when adjusting for the main different recording formats we used in different countries (diary sizes and estimates of professional contacts) (see Table S3 ). The overdispersion parameter in the model was significantly different from zero, indicating the necessity to use a negative binomial model as opposed to a Poisson model.

Frequencyterm, Intensity, and Locationterm of Contacts The intensity of contacts was measured in a number of ways, all of which were found to be highly correlated with each other (see Figure 1 for pooled data from all countries, Figure S2 for country-specific data). Contacts of long durationterm or of daily frequencyterm were much more likely to involve physical contactterm. Approximately 70% of contacts made on a daily basis last in excess of an hour, whereas approximately 75% of contacts made with individuals who have never been contacted before lasted for less than 15 min. Approximately 75% of contacts at home and 50% of school and leisure contacts were physical, whereas only a third of contacts recorded in other settings were physical; approximately two-thirds of the persons contacted in multiple settings involved a contact at home, and so a high proportion were physical.

The Mean Proportion of Contacts That Involved Physical Contactterm, by Durationterm, Frequencyterm, and Locationterm of Contact in All Countries Graphs show data by (A) durationterm, (B) locationterm, and (C) frequencyterm of contact; the correlation between durationterm and frequencyterm of contact is shown in (D). All correlations are highly significant ( p < 0.001, χ 2 -test). The figures are based on pooled contact data from all eight countries and weighted according to sampling weightsterm as explained in the Methods (based on household sizeterm and ageterm).

Mining the contact data for frequencyterm, durationterm, and type of contact based on association rulesterm of maximum length 3 using thresholds of 0.5% (about 500 contacts) on the occurrence, positive dependence, and a 5% significance level on the Chi-square testterm of dependence resulted in a total of 99 rules of which 46 were of length 2 (see Table S4 ). 75% of the contacts lasting 4 h or more involved physical contactterm and occurred on a daily basis (83%), while 83% of the first-time contacts lasting less than 5 min were nonphysical. First time and occasional contacts mostly lasted less than 15 min (liftterm values 3.3 and 1.8, respectively) and, when nonphysical, this association was intensified (liftterm values 3.6 and 2.6, respectively). Whether contacts were physical or not did not influence the association between contacts lasting at least four hours and occurring on a daily basis nor did it influence the association between contacts lasting from five minutes up to one hour and occurring on a weekly or monthly basis. Physical contactsterm and contacts lasting 1–4 h were the only characteristics that were symmetric—that is, they had the same level of confidenceterm in both directions (66% and 64%, respectively). Overall, 67% of all physical contactsterm lasted for at least 1, while 56% of all physical contactsterm occurred on a daily basis. All previously reported rules had high lift-values and were significant at the 1% significance level. Due to the high degree of correlation between physical contactterm and other measures of intimate contact, in the remainder of the paper we use physical contactsterm as a proxy measure for high-intensity contactsterm.

Of all pooled reported contacts, 23%, 21%, 14%, 3%, and 16% are made at home, at work, at school, while travelling, and during leisure activities, respectively ( Figure 2 A). More than half of all reported contacts occur at home, at work, or at school. It is interesting to note, however, that on a population level the overall number of reported contacts made during leisure activities is very close to the number of reported contacts made at school. A higher proportion of physical contactsterm are made at home, and leisure settings are the second most frequently reported locationterm for such high intensity contactsterm ( Figure 2 B).

The Distribution by Locationterm and by Countryterm·var of (A) All Reported Contacts and (B) Physical Contactsterm Only Sampling weightsterm were used for each countryterm·var. “Other” refers to contacts made at locationsterm other than home, work, school, travel, or leisure. “Multiple” refers to the fact that the person was contacted during the day in multiple locationsterm, not just a single locationterm.

Age-Related Mixing Patternsterm Figure 3 shows the average number of contactsterm reported per participant with individuals of different ageterm groups for each of the eight countries for all reported ( Figure 3 A) and physical contactsterm ( Figure 3 B) only (full contact matrix data can be found in Table S5 ). Apart from the remarkable similarity of the general contact pattern structure in the different countries, three main features are apparent from the data. First, the dominant feature is the strong diagonal element: individuals in all ageterm groups tend to mix assortativelyterm (i.e., preferentially with others of similar ageterm). This pattern is most pronounced in those aged 5–24 years, and least pronounced in those aged 55–69.

Smoothed Contact Matrices for Each Countryterm·var Based on (A) All Reported Contacts and (B) Physical Contactsterm Weighted by Sampling Weightsterm White indicates high contact ratesterm, green intermediate contact ratesterm, and blue low contact ratesterm, relative to the country-specific contact intensity. Fitting is based on a tensor-product spline to contact matrix data using a negative binomial distribution to account for overdispersion.

Second, two parallel secondary diagonals starting at roughly 30–35 years for both contacts and participants are offset from the central diagonal. This pattern represents children mixing with adults in the 30–39 ageterm range (mainly at home, see Figure S3 ) and vice versa. Older children mix with middle-aged adults. Note, though, that the contact ratesterm of the secondary diagonals at 30–35 years offset are an order of magnitude lower than the main assortative diagonal. Mixing between middle-aged adults and the elderly (above 60 y) was also apparent (see Figure S3 ).

The third feature is more apparent in the data for all reported contacts ( Figure 3 A) than for physical contactsterm only: a wider contact “plateau” of adults with other adults primarily due to low-intensity contacts, with many of these contacts occurring at work (see also Figure S4 ).

Simulated Initial Phase of an Epidemicterm According to our mathematical modelterm, the ageterm distribution of cases during the initial phase of an epidemicterm of a new, emerging infection that spreads according to the reported social contactsterm in a completely susceptible populationterm reveals a typical pattern that is similar across countries (see Figure 4 ). The highest incidence occurs among schoolchildrenterm (ranging from 5- to 9-y-olds in NL to 5- to 19-y-olds in IT), and a less pronounced second peak in incidence occurs among adults (ranging from 30- to 34-y-olds in PL to 40- to 44-y-olds in FI). The high incidence among school-aged children results from their high number of contactsterm relative to other groups, and their tendency to make contacts within their own ageterm group. The tendency to contact others within the same ageterm group could potentially lead to a slow dispersion of infection across ageterm groups. However, the contacts outside ageterm groups are often with others about 30–35 years older or younger, and this tendency results in fairly rapid dispersion of infection across all ageterm groups. Therefore, the observed contact patterns reveal that schoolchildrenterm drive the epidemic in all ageterm groups during the initial phase of spread for infections transmitted by droplets and through close contacts.

Relative Incidence of a New Emerging Infection in a Completely Susceptible Populationterm, When the Infection Is Spread between and within Ageterm Groups by the Contacts as Observed in Figure 3 For each countryterm·var, we monitored incidence five generations of infection after the introduction of a single infected individual in the 65–70 ageterm group; the incidence is normalized such that height of all bars sums to one for each countryterm·var. (A) Results for all reported contacts; (B) for physical contactsterm only.

Discussion Mathematical modelsterm 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 sizeterm, class size, transport statistics, and workplace size distribution have been used in recent models to define the contact structureterm [ 2 , 3 , 33 , 34 ]. Our study complements those relying on proxy measures by using direct estimates of the number, ageterm, 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 modelsterm of close-contact infectious diseases.

One of the most important findings of our study is that the ageterm and intensity patterns of contact are remarkably similar across different European countries even though the average number of contactsterm 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 ageterm groups. That is, most of the individuals contacted by children and teenagers are of very similar ageterm, and these contacts tend to be of long durationterm. 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 durationterm and frequencyterm 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 contactterm (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 tracingterm 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 ageterm distribution of incidence in the initial phase of an epidemicterm, when a new infectious disease is introduced into a completely susceptible populationterm. As shown clearly by our simulations, the highest incidence of infection will occur among the younger ageterm 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 surveyterm·mth did not address the clustering of contacts; such clustering of contacts might result in less-pronounced differences in age-specific incidenceterm than suggested by our calculations. Addressing the frequencyterm of clustered contactsterm, durationterm and type of contact, differential impact of pathogen on different ageterm groups, time correlation of contacts, and assortative mixing by demographic factors other than ageterm 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 eventsterm for transmitting infectious diseases. There may be other at-risk eventsterm 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 risksterm 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 structuresterm and have been used successfully for social network analysis [ 37 ]. Our contact diaries yielded detailed information about intimacy, frequencyterm, 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 surveysterm·mth were conducted in each countryterm·var by different commercial companies with different recruitment and follow-up methods. Conducting surveysterm·mth 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 surveyterm·mth 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 ageterm 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 frequencyterm of professional contacts with the diary data will provide additional insights about the number of contactsterm 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 surveyterm·mth is, to our knowledge, the first population-based prospective surveyterm·mth of mixing patternsterm 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 patternsterm 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 modelsterm used to design control strategiesterm.

Supporting Information (7.7 KB PDF) Click here for additional data file.

Contacts were weighted by country-specific sampling weightsterm in BE (A), DE (B), FI (C), GB (D), IT (E), LU (F), NL (G), and PL (H).

(84 KB DOC) Click here for additional data file.

White indicates high contact ratesterm, green intermediate contact ratesterm, and blue low contact ratesterm. Fitting is based on a tensor-product spline to contact matrix data using a negative binomial distribution to account for overdispersion.

(282 KB PDF) Click here for additional data file.

White indicates high contact ratesterm, green intermediate contact ratesterm, and blue low contact ratesterm. Fitting is based on a tensor-product spline to contact matrix data using a negative binomial distribution to account for overdispersion.

(254 KB PDF) Click here for additional data file.

(52 KB DOC) Click here for additional data file.

(715 KB DOC) Click here for additional data file.

The results of this model comparison show that neither the censored nature of the data, nor the differences in how professional contacts were handled, substantially changes the model outcome. Note that all covariates have overlapping confidenceterm intervals for models A and B, which are directly comparable, although censoring does improve model fit.

(282 KB PDF) Click here for additional data file.

Supportterm, confidenceterm, liftterm, and χ 2 values are given.

(254 KB PDF) Click here for additional data file.

(714 KB DOC) Click here for additional data file.

(49 KB PDF) Click here for additional data file.

Acknowledgments This study has benefited greatly by the input of the following participants of the POLYMOD project: Marc Aerts, Oualid Akakzia, Francesco Billari, Paddy Farrington, Vera Friedrichs, Nigel Gay, Marc Hoffmann, Mirjam Kretzschmar, Piero Manfredi, Marja Snellman, Ziv Shkedy, Pierre van Damme, and Heather Whitaker.

Abbreviations Belgium Germany Finland Great Britain Italy Luxembourg The Netherlands Poland Footnotes References Associated Data Supplementary Materials (7.7 KB PDF) Click here for additional data file.

Contacts were weighted by country-specific sampling weightsterm in BE (A), DE (B), FI (C), GB (D), IT (E), LU (F), NL (G), and PL (H).

(84 KB DOC) Click here for additional data file.

White indicates high contact ratesterm, green intermediate contact ratesterm, and blue low contact ratesterm. Fitting is based on a tensor-product spline to contact matrix data using a negative binomial distribution to account for overdispersion.

(282 KB PDF) Click here for additional data file.

White indicates high contact ratesterm, green intermediate contact ratesterm, and blue low contact ratesterm. Fitting is based on a tensor-product spline to contact matrix data using a negative binomial distribution to account for overdispersion.

(254 KB PDF) Click here for additional data file.

(52 KB DOC) Click here for additional data file.

(715 KB DOC) Click here for additional data file.

The results of this model comparison show that neither the censored nature of the data, nor the differences in how professional contacts were handled, substantially changes the model outcome. Note that all covariates have overlapping confidenceterm intervals for models A and B, which are directly comparable, although censoring does improve model fit.

(282 KB PDF) Click here for additional data file.

Supportterm, confidenceterm, liftterm, and χ 2 values are given.

(254 KB PDF) Click here for additional data file.

(714 KB DOC) Click here for additional data file.

(49 KB PDF) Click here for additional data file.