Here we present a large-scale, prospective social contact patterns study, the first of its kind to be conducted in Cambodia. We estimated that (at least prior to the Covid-19 pandemic) the average Cambodian person had a mean of 31.2 social contacts per day, although this figure falls to 16.3 when counting only the contacts that were individually reported in the contact diary. The average number of self-reported diary contacts are similar to the Mossong et al., (2008) European study (13.4 contacts per person per day) and a study in a rural coastal population of Kenya (17.7 contacts per person per day 16, but higher than other studies in Hong Kong 18, UK 17 and Zimbabwe 23 which reported an average of 7 to 8.1 contacts per person per day. Similar to our study, the study in Zimbabwe also asked participants to list ‘casual’ additional contacts with 64% of people reporting fewer than additional 10 contacts per person per day 23. Such heterogeneities highlight the need for the collection of context-specific mixing patterns data, although the definition of contact used in the various studies and method of recording contacts will also play a role 18.
Consistent with studies in other settings, contacts in Cambodia were highly age-assortative; this was particularly evident in school-aged children but was also observed in other age-groups, such as young adults 13,16,23. Tertiary diagonals, representing inter-generational mixing, were also evident in our contact matrices, and indeed appeared more prominent compared with data from European countries (Mossong et al., 2008). This likely reflects the higher prevalence of multi-generational households in Cambodia, further supported by our data showing that these contacts tended to be reported at home with household members, particularly in rural areas. This has important implications for respiratory infections such as influenza and SARS-CoV-2, where younger individuals may have higher exposure rates due to their relatively high contact rates, and elderly individuals may be more susceptible severe disease 7.
Home and school are well documented as high-intensity contact settings important for infectious disease transmission 24,25. Interventions aiming to reduce contact rates of vulnerable individuals (such as elderly shielding) may be more challenging to implement in LMICs given the household structures. Model simulations indicate that reducing contact rates among younger individuals in LMICs through school closures can have secondary protective effects on older household members 26. Conversely, concerns have also been raised surrounding potential adverse consequences of such measures, for example if children continue to mix with peers outside of school whilst increasing contact rates with their grandparents, or if there is health care worker absenteeism due to childcare responsibilities 27,28.
Few, if any, previously published contact surveys have differentiated between household and non-household contacts when describing contacts taking place at home (at least in their reported results). Notably, over a quarter of all diary contacts (27.7%) reported in our study occurred at home with non-household members. Given that reduction in rates of inter-household mixing has been a key component in the SARS-CoV-2 pandemic response, this may highlight a need for future contact surveys and modelling studies to consider heterogeneities in contact rates between household vs. non-household members at home. Contacts made in settings outside of home, work or school (defined as “other” settings), also accounted for a substantial proportion of contacts, particularly in rural areas, lending support for the role of NPIs such as public space closures during periods of elevated transmission risk.
A previous study by Prem et al., (2017) combined contact data from the European POLYMOD survey with socio-demographic data with to project social contact matrices for 152 countries, including Cambodia. The overall patterns in our empirical contact matrices were broadly comparable to those in the synthetic, country-level matrices generated by Prem et al., (2017), although we observed a lower relative contribution of contacts in work settings than predicted. Moreover, we observed marked differences between rural and urban areas in terms of the number, intensity and settings of contacts. The higher number of contacts observed in rural areas, particularly in “other” settings, may initially seem counter-intuitive, given higher population densities in urban areas. However, this might be explained by socio-cultural factors such as close-knit rural communities, particularly when considering our survey’s definition of a contact as a two-way conversation. Such factors may also be reflected in higher observed rates of contact made across age groups with visitors at home. Meanwhile, the contacts of urban participants tended to be of higher intensity in terms of both duration and the proportion of contacts which were physical.
Relatively few previous studies have investigated differences in contact patterns between urban and rural populations. However, our findings are consistent with a study in Kenya, which reported higher contact rates in rural vs peri-urban participants 16, while studies in Zimbabwe 29 and China 20 did not observe any significant differences. Recognizing that early outbreaks of SARS-CoV-2 were mostly observed in cities with subsequent spread to rural areas, Prem et al., (2021) recently updated their synthetic contact matrices to include custom contact matrices for rural and urban settings.
The data analysed in our study were collected over 12 days in May 2012, therefore temporal changes are not captured and changes in social mixing patterns may have occurred since that time, particularly following the emergence of the SARS-CoV-2 pandemic. However, given the lack of data from this region, and LMICs in general 2, such data are still highly valuable. The POLYMOD data collected in 2005/2006, and “synthetic” contact matrices based on these data, continue to be widely used in infectious disease transmission models, including those developed for SARS-CoV-2 6,9,30,31. Statistical approaches can be used to adjust for changing population demographics, and our results can also be used to inform and compare with future surveys to identify temporal trends.
Selection bias may have occurred as we observed lower employment rates in our survey participants than has been reported in country census data as a whole. This is likely due to the recruitment via household visits, and future studies recruiting in a similar way may seek to avoid this during collection by having sampling quotas for this variable. We sought to correct for this in our analyses by using sampling weights for employment status (along with other factors such as weekday vs weekend). Such weighting is useful, but care should be taken to avoid raking with too many variables with many levels, and ensuring sufficient data points with each cross-tabulation of these selected variables. This should help to avoid extreme sampling weights and minimise design effects 32.
Our study used paper-based contact diaries for self-reporting of contacts; other studies have used online surveys, retrospective interviews, or made inferences regarding contacts through wearable devices 33. Leung et al., (2017) found that those with more years of education and higher income levels were more likely to choose online questionnaires, as opposed to paper based, when both were offered in a study in Hong Kong. They also found that participants using paper questionnaires reported a significantly greater number and duration of contact than those using online questionnaires 18. Cambodia is classified as an LMIC and, according to the 2010 Demographic Health Survey (DHS), only 19% of rural residents had electricity (compared to 90% of urban residents), necessitating the use of household visits and paper-based questionnaires to reduce selection bias.
Prospective surveys are preferred as they allow participants to remember more details of the contacts 2,34. The use of an interviewer to retrospectively review and/or complete the contact diaries with participants through prompts enhanced completeness, and also inclusion, for example of illiterate participants (estimated at 30% of females in rural areas in 2012) 35. However, while participants were asked to report all contacts within the diary, it is worth noting that less than half of our study participants believed their diaries to be complete when asked about this at the end of the interview. (Participants were not given prior notice that this question would be asked.) On the other hand, it seems likely that many of the ‘supplementary’ (non-diary) contacts reported would have been more casual in nature, and thus of lower epidemiological relevance. Indeed, a larger proportion of these supplementary contacts were non-physical contacts compared to contact-diary reported contacts. Furthermore, the age distributions of supplementary contacts were reasonably similar to those of the diary contacts (to the extent that these could be compared), suggesting that truncation of the latter would not have biased the age-contact matrices to any substantial degree.
The contact matrices (Table S2) and other data presented here can be used to inform mathematical models of disease transmission in Cambodia. Indeed, the results from this survey were recently used to parameterise an age-structured Susceptible-Exposed-Infectious-Removed (SEIR) model of SARS-CoV-2, to conduct a scenario analysis of NPI strategies on transmission and healthcare resource capacities at country and province-level 36, under work conducted for the WHO Western-Pacific Regional Office Covid-19 modelling consortium. More broadly, the study also contributes towards addressing knowledge gaps on social-mixing patterns and population mobility in LMIC and Southeast Asian contexts, and provides empirical evidence on within-country variation between urban and rural areas.