Background
Infectious diseases have substantial impact on public health and economy and warrant constant monitoring and follow-up. Initially, transmission models relied on untested assumptions about “at risk events”. In recent years, disease transmission models have been informed by social contact surveys as, e.g., those obtained from the large-scale European POLYMOD project [
1], in which participants had to report about their contact behavior by age, gender, frequency, etc. Social contact patterns have been successfully used as proxies for transmission of close-contact diseases, such as influenza and mumps, under the social contact hypothesis [
2]. The use of social contact data helps estimating key epidemiological parameters [
3,
4], behavioral changes [
5‐
7] and demographic change [
8] in the context of disease transmission. The number of social contact surveys to collect empirical data on human contact behavior has increased substantially over recent years [
9].
Next to social contact data, also time use data have proven their value for explaining infectious disease data using the
time use approach in which mixing patterns can be estimated from the time spent at a given location [
10‐
12]. The reported presence over time at different locations during a day enables the estimation of the exposure time among age groups [
12]. The notion of co-presence is complementary to reported social contacts, hence time use data are useful to capture “at risk events” that fall outside the definition of a social contact.
The integration of both the “time use approach” and the “social contact approach” has lead to estimation of “suitable contacts” [
11]. De Cao et al. [
11] showed that the interplay between exposure time and social contacts appeared to be important to study the transmission of Varicella-Zoster virus (VZV) whereas the transmission dynamics of parvovirus-B19 was best captured using only exposure time. This study was based on two independently collected surveys (social contacts and time use).
A systematic review of social contacts surveys [
9] revealed that only a limited number of contact studies investigated the relationship between contacts and distance whereas next to the number of contacts, contact dispersal is of essence to capture disease transmission dynamics. Disease counts have been modeled using power law dispersal kernels [
13‐
15], though questions remain whether this also holds for social mixing behavior. A study in Great Britain [
16] collected information on the distance from home for each contact and observed a decrease in contact duration with increasing distance from home. In addition, age-specific contact patterns and temporal differences with respect to weekdays and weekends have been described [
9], but gender-specific behavior in social mixing is less reported, though could contribute to the parameterisation of mathematical transmission models [
17,
18].
In this work, we use data from one single survey to compare the social contact, time use and suitable contact approach. An important element in this survey [
18] is the recording of the distance from home for every reported location in the time use survey. In particular, we (1) compared mixing patterns based on time use and social contact data, and explored covariates such as age, gender, location, etc.; (2) analysed social mixing patterns and estimated basic reproduction numbers by distance from home; (3) evaluated the value of time use and social contact data sources to explain seroprevalence data for VZV and parvovirus-B19, and influenza-like illness (ILI) incidence data in Belgium.
Discussion
We analysed time use and social contact data and compared their use as proxy for effective contacts governing disease transmission for disease transmitted through the respiratory route. Our dataset is unique since it provides both time use and social contact data from the same participants, avoiding possible differences due to sample biases. In our analysis we identified the main drivers in shaping everyday time use and linked this info to social contact patterns.
The reported time use patterns in Belgium are quite similar to the published patterns for Italy [
37], but different from the results from Zimbabwe [
38]. In Zimbabwe, the working-age participants and children less than 6 years old reported much less time at work and school respectively, compared with participants of the same age group in Belgium and Italy. We found that males spent on average
\(\pm 2\) hours more at work than females, which is in line with previous work [
39,
40]. We also found that both males and females living with children spend more time at home than people living without children, which is consistent with what was found in [
39]. The expected temporal patterns were observed, with more time spent at “other” locations during weekends and holiday periods. The time spent at home seemed not to be affected by the type of day. Our gender-specific analysis of the time use data indicated that participants were prone to spend more time with the same gender when they are young, and more time with the other gender when they are older. This result differs from the observed gender-specific contact rates as reported in [
18], where the assortativity is reported to be higher for same-gender contacts.
Power law dispersal has been useful to model disease counts [
13‐
15], though questions remained whether this also holds for social contact behavior, the driver of transmission dynamics [
2]. Danon et al. [
16] showed a relation between clustering and distance from home, with high clustering within two miles, dominated by home contacts, but the highest value of clustering occurring at a distance 50 miles or more from home. The authors hypothesize that this might be due to differences in the purpose behind contacts made at various distances.
In our study, we observed an increase in the number of contacts by distance during weekdays, until an age-specific distance-threshold. A large-scale study in Taiwan [
41] reported that 52.7% of contacts took place at a distance less than 1 km from home, 29.2% at a distance 1–9 km from home, 14.6% at a distance 10–49 km from home. In our study, this pattern was clearly age-specific. Half of our reported contacts during regular weekdays for children between [0–18) years of age took place less than 1 km from home. During weekends, we observed more contacts at +10 km from home. In general, we observed that the average number of contacts decreased by distance, though if people made the effort to travel, they made it count in terms of social contact behavior. As such, social contact patterns conditional upon presence at each distance provided useful info to inform individual-level behavior. A study in China quantified the distances from home based on the latitude and longitude of each reported contact and observed an increase in assortative mixing when contacts were made further from home [
42]. We observed similar patterns with the construction of conditional social contact matrices by distance. A study in the United Kingdom [
43] requested for infants to report the maximum distance travelled from home, but, to date, did not report results thereof. Other survey designs (e.g., [
38]) included the distance between home and work but did not report results related to this information, yet.
The age-specific social contact and time use pattern followed a similar trend of assortativeness, which stresses once more the tendency for people to have contacts with someone of similar age. In addition to that, strong mixing among generations (parent–child) was present in both the social contact and time use matrices. The inter-generation mixing is mostly observed at home, and was more pronounced in the exposure time matrix than in the contact matrix. With our unified survey, we can confirm the contrasting effect of the relative small number of social contacts at home compared to the large amount of time spent at home, as observed in [
10,
38].
Social contact matrices provide useful data to estimate disease transmission dynamics in terms of the transmission dynamics and relative incidence [
1,
2]. As such, we estimated the R
\(_{0}\) for each distance conditional upon presence, and observed a clear increase by distance. This reflects an increasing transmission potential by distance from the individual-level perspective. The latter is of interest for individual-based and some meta-population models where individuals join other sub-populations at distance. Assuming a constant (or decreased) social mixing behavior by distance conditional upon presence might not be optimal. Some individual-based models handled this by the use of location-specific mixing patterns irrespective of distance from home [
44].
We compared the value of different social contact features (duration, physical/non-physical, etc.) to inform transmission models for parvovirus-B19 and VZV. By scoring the model-based prevalence with Belgian serological data, we found that physical contacts provided the best proxy for both parvovirus-B19 and VZV. In terms of contact duration, the best model fit was obtained with physical contacts of long duration (more than 4 h) for parvovirus-B19 and (more than 1 h) for VZV. Goeyvaerts et al. [
3] reported the best fit to VZV with physical contacts of at least 15 min; this result is also in line with the study of [
4]. However, in the previous studies, not all combinations in terms of contact duration and physical/non-physical contacts were analysed, which explains the new “best estimate” in our study.
We also compared the results of the contact approach, the time use approach and the suitable contact approach in fitting serological data of VZV and parvovirus-B19. In the case of VZV, the suitable contact approach provided the best fit, while for parvovirus-B19, the time use approach gave the best fit. Our results are consistent with the findings of De Cao et al. [
11], although we observed much higher parameter estimates for
\(q_2\) (0.13 vs 0.001 for parvovirus-B19 and 0.94 vs 0.37 for VZV), but the confidence intervals of these parameter estimates were overlapping.
We also tested the value of contact matrices and exposure time matrices in fitting the dynamic transmission model to weekly ILI incidence data in the season 2010–2011 in Belgium. Exposure time matrices provided a better fit to ILI than overall contact matrices and the integration of these two types of matrices did not improve the fit to the data. Within the social contact approach, physical contacts provided a better proxy for the risks of influenza transmission than non-physical contacts. We found that
\(R_{0}\) of the best model is 1.43, this result is consistent with a systematic review of estimates of
\(R_{0}\) for different types of influenza [
45] and a study in UK [
34] that also used contact matrices to fit to ILI incidence data. However, this value is much lower than the seasonal estimates of the reproduction number in [
33], in which physical contact matrices from the Belgian POLYMOD data were used to fit to ILI data over multiple influenza season from 2003 to 2009. Fewer ILI cases were reported in the season 2010–2011 as compared with previous seasons, which partly explained the difference in
\(R_{0}\). In addition, note that small differences in model parametrization entail substantial differences between the estimates of
\(R_{0}\), which was also mentioned in [
33].
In our study, we combined information from time use and social contact data to gather information on human mixing patterns. One of the main advantages with respect to previous work is that both sources of information came from the same survey. To keep participants’ burden as low as possible, time use information was collected with rather large time slots and participants were asked to fill in only one location for each time slot. However, the comparison with more refined time use surveys performed in Flanders [
39,
40] confirms that we were able to well characterize exposure patterns at an aggregated level. Therefore, we expect that this limitation did not substantially affect our results. Under the Proportionate Mixing Assumption, age-specific exposure matrices could be biased for some locations, e.g. public transportation or ‘Other’ location. In this study, we assume that age specific-contact patterns and time use patterns did not change over 10 years, the gap (in years) between the collection of serology and social contact data. This assumption is partially supported by the work of Hoang et al on contact patterns over a time span of 5 years [
18].
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.