Background
The role of social contacts in the spread of respiratory infections has been discussed extensively in the year 2020 due to the global outbreak of severe acute respiratory syndrome coronavirus type 2 (SARS-CoV-2) [
1,
2]. As of March 2021, over 100 million confirmed cases and over 2.5 million deaths have been recorded worldwide [
2]. SARS-CoV-2 is primarily transmitted via droplets and aerosols, so person-to-person contacts are a strong determinant of transmission dynamics [
2‐
4]. Non-pharmaceutical interventions (NPIs) focusing on the reduction of person-to-person contacts are one of the cornerstones of the pandemic response. In the middle of March 2020, Germany mandated school and kindergarten closures, postponed academic semesters, prohibited visiting of nursing homes and restricted the number of people allowed at public and private gatherings in an attempt to protect the vulnerable groups [
5]. In the following weeks, contact reduction measures were implemented on a population level by regulating the maximum number of close social contacts outside one’s household and by closing non-essential shops as well as places for leisure activities [
5]. After a considerable reduction in reported case numbers, federal governments decided to ease these restrictions gradually starting at the beginning of May 2020.
Social contact patterns are known to be a critical factor for the transmission dynamics of respiratory infections [
4,
6‐
10]. However, empirical social contact data have been scarce before the emergence of SARS-CoV-2 [
11‐
13]. One exception is the POLYMOD study, a large-scale survey that described social mixing patterns in eight European countries [
12]. In 2005/2006, POLYMOD measured contacts of more than 7000 participants across eight European countries [
12]. Contact patterns observed in POLYMOD have been widely used to parametrize various mathematical models of infectious disease dynamics [
3,
4,
12,
14].
During the SARS-CoV-2 pandemic, contact surveys were initiated in several countries to understand the effect of contact precaution measures on social contact patterns [
3,
4,
10,
15‐
19]. While contact surveys offer a direct approach to social contact patterns, they are time- and cost-intensive and need to be initiated actively. Mobile phone-based mobility data offer a complementary approach to infer changes in contact patterns in a population. Google and Apple granted free access to anonymized mobility data in a global attempt to provide insights into the change of mobility during the pandemic given different physical distancing policies [
20,
21]. Several SARS-CoV-2 modelling studies assumed that aggregated mobility data can be used as a proxy for the actual number and intensity of contacts of individuals in a defined population, although mobility data measure only certain dimensions of contact behaviour. In this article, we present survey-based social contact data for the first wave of the pandemic in Germany and assess their ability to reflect transmission dynamics 10 days later (measured by reported reproduction number (
R estimates)) when compared to open source population mobility data from Google and Apple [
20‐
22].
Discussion
In this study, we quantified the relative reduction in contacts based on contact survey data and publicly available mobility data. We found that both data sources represent different dimensions of transmission dynamics; changes in contact patterns measured in survey data represented transmission dynamics (measured as R) better than the changes measured in aggregated mobility data independently of the introduction of contact- and mobility type-specific weights and the use of scaling factors. Non-pharmaceutical interventions introduced in Germany during the first wave of the SARS-CoV-2 pandemic were, however, associated with both a considerable reduction in social contacts reported in contact surveys as well as with reductions in mobility patterns. The results of our study indicate that deriving contact behaviour from mobility data alone, as it was often the case in political decision-making during the first and second wave, is not suitable for making real-time inferences on the effects of public health measures on the transmission dynamics in a population. Mobility data used in this study suggested that contact behaviour went back to normal almost instantly after the contact reduction measures were relaxed, which did not reflect the observed R values. A reason for that might be that people still tried to minimise close contacts outside their own households and maximised distance to the contacts they had, although their mobility, e.g. back to work, already reached almost pre-pandemic levels. Therefore, a complementary approach including both aspects, i.e. social contact behaviour as well as mobility behaviour, is necessary to fully reflect transmissions dynamics. Although repeated contact surveys need considerable investment in terms of time and costs, the potential benefits and financial savings if used as a near real-time proxy for transmission dynamics on a population level are likely to outweigh the efforts needed. Benefits include a better preparedness towards expected case numbers as well as earlier information on the effect of newly introduced contact reduction measures, which allows timely adaptation if needed.
In our study, we found a 73% mean reduction in contacts across the first four waves of COVIMOD (i.e. from April to June 2020) which is consistent with studies from other European countries [
3,
4]. Even though the reported number of daily contacts increased over the survey waves, it was still considerably lower than in POLYMOD, indicating sustainable behaviour change even after the end of the strictest contact reduction measures. We found an increased variance in the reported daily number of contacts as the COVIMOD waves progressed, with the maximum number of contacts increasing from 16 in survey wave 1 to 674 in wave 4, while median contact numbers were not affected similarly. Since SARS-CoV-2 has been shown to be associated with a high variance in the number of transmissions arising from one infectious individual [
33], this sharp increase in the maximum number of contacts has huge implications for the risk of superspreading events as the direct aftermath of the end of public health interventions. Participants aged 60 and above reported fewer contacts in all COVIMOD waves as well as a larger reduction to pre-pandemic values when compared to children and middle-aged persons. This should be taken into account when assessing the effects of vaccination prioritisation strategies in combination with NPIs, as people in this age group are known to be more vulnerable to SARS-CoV-2 infections [
15].
We further observed a smaller and more stable reduction in home contacts than in work, educational and leisure time contacts, which confirms that reduction in contacts is location-specific [
3]. This is reasonable as most of the social distancing measures implemented at that time had their main impact outside the household. We confirmed that the majority of remaining contacts under strict contact reduction measures happens between life partners and parents and children, which mirrors the huge role of this transmission setting under contact reduction measures [
7,
34].
When introducing contact- and mobility type-specific weights representing different transmission probabilities for home and non-home contacts/mobility, we were able to considerably reduce the differences in estimates for transmission dynamics when compared to the reported R values 10 days later, even if scaling factors had been fitted to the different data source models before. However, the remaining differences were in all analyses much smaller for estimates based on contact survey data than for mobility data. These results show that the presented approach might be suitable for a near real-time estimation of transmission dynamics based on contact survey data alone or in combination with mobility data. A data-driven estimation (based on contact survey data) of the relative transmission risk at home compared to non-home transmission resulted in estimates very similar to those derived from setting-specific secondary attack rates reported in the literature. Our results indicate that the differentiation in home and non-home contacts based on contact survey data supports the representation of the true role of different types of contacts for transmission dynamics.
Our analyses suggest that the use of contact survey data, especially after weighing for home and non-home contacts together with an additional scaling factor, can indeed be used as an early marker of current transmission dynamics, especially if they are mainly determined by contact reduction measures. We show that aggregated mobility data offer a different behavioural perspective but can also contribute to a better understanding of how transmission dynamics might develop in near real-time. The analyses performed in this study were rather simplistic by nature, as they aimed to provide an overall estimate of transmission dynamics without differentiating by too many different factors and without a formal dynamic mathematical model. In reality, the information provided about changes over time in contact settings, intensities and frequencies with contact partners offers especially for contact survey data but also for mobility data much wider perspectives. Since these analyses require a dynamic modelling approach taking into account various other assumptions not necessarily available in the early phases of an epidemic, they might not be as suitable for near real-time communication with decision-makers as the simpler approaches presented here. However, future analyses should focus on using the available contact and mobility data to construct and validate multi-layer mathematical models which take into account mobility data for large scale movements and contact survey data for small scale effect contacts, and this combines the strengths of the different data sources.
Our study has several limitations. COVIMOD data are not fully representative of the German population since some adult participants with under-aged children living in their households were invited to provide information as a proxy for their children. Moreover, the elderly (> 70 years) and the very young (< 10 years of age) are underrepresented in COVIMOD. We tried to correct for that by introducing weights for sex, age and household size; however, there were no relevant differences in the results of the unweighted and weighted analyses. Participants in COVIMOD were asked to record their contacts retrospectively so that different forms of information bias could have been introduced. For example, it might be challenging to remember a higher number of contacts, or the participants’ willingness to report high numbers of contacts individually might be lower as this is quite tedious and time-consuming. We tried to minimise this by allowing the participants to record group contacts. We also cannot rule out that COVIMOD attracted specifically participants who adhered to social distancing rules as these individuals might be more likely to respond to health surveys. This could have led to an overestimation of the relative reduction of contacts and could explain the gap between relative reductions in social contacts and reported
R values. We tried to minimise this bias by using an established online panel not focusing on healthcare questions as the platform for COVIMOD. Even though contact-related questions were similarly phrased between POLYMOD and COVIMOD, POLYMOD was paper-based, and COVIMOD surveys were web-based. Previous research suggested that participants might report more contacts in paper-based surveys than in web-based surveys [
11,
35]. Future research will be conducted on the differences between web- and paper-based contacts during the pandemic. However, our findings are consistent with other studies that examined social contact patterns under strict contact reduction measures [
3,
4,
15,
36]. We used aggregated mobility data in our study that were freely available and have been discussed as a potential real-time proxy for SARS-CoV-2 transmission dynamics. Although we took advantage of two different data sources representing complementary ways to define mobility, our results cannot be automatically generalised to other ways of measuring mobility (e.g. based on individual movement patterns). The
R values derived from RKI represent the changes in transmission dynamics based on contact reduction measures as well as population immunity, while contact survey data and mobility data can only assess the former. Since population immunity was below 1% in the study period, this is unlikely to have played a major role in this analysis but needs to be taken into account for future studies. Application of scaling factors, which include information on developing population immunity, might be a useful tool for later phases of an epidemic.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.