Relationship between household attributes and contact patterns in urban and rural South Africa

TL;DR

The study investigates how within-household contact patterns in South Africa depend on household attributes (site, head gender, and household type) using high-resolution proximity-sensor data from rural and urban settings. By constructing and analyzing contact matrices, it shows that interactions are predominantly intergenerational and within households, with seasonality and household composition shaping the total contact time. A multilevel epidemiological analysis reveals that age-based heterogeneity dominates over gender alone, but combining age and gender yields the strongest influence on potential transmission (R0) in extended households. The findings highlight the importance of incorporating rich household-structure metadata into epidemic models, while noting limitations due to sample size and scope (only within-household contacts) that call for larger, richer data collection in sub-Saharan Africa.

Abstract

Households play a crucial role in the propagation of infectious diseases due to the frequent and prolonged interactions that typically occur between their members. Recent studies have emphasized the need to include socioeconomic variables in epidemic models to account for the heterogeneity induced by human behavior. While sub-Saharan Africa suffers the highest burden of infectious disease diffusion, few studies have investigated the mixing patterns in the countries and their relation with social indicators. This work analyzes household contact matrices measured with wearable proximity sensors in a rural and an urban village in South Africa. Leveraging a rich data collection describing additional individual and household attributes, we investigate how the household contact matrix varies according to the household type (whether it is composed only of a familiar nucleus or by a larger group), the gender of its head (the primary decision-maker), the rural or urban context, and the season in which it was measured. We show the household type and the gender of its head induce differences in the interaction patterns between household members, particularly regarding child caregiving, suggesting they are relevant attributes to include in epidemic modeling.

Figures (6)

• Figure 1: Summary of household and population compositions disaggregated by age and gender and according to the household classification.Left: fraction of households disaggregated by site, household type, and gender of the household head. The inner level indicates the fraction of households in the rural and the urban sites. The second layer categorizes households by type, while the third layer categorizes households by the gender of the household head. Right: number of individuals per age-gender for varying household grouping strategies. The column "Classification" indicates the different ways of characterizing households considered in the main text. The rows that are not separated by a horizontal line indicate non-overlapping groups of households (e.g., Rural and Urban), while overlaps exist across lines separated by horizontal lines (e.g., Rural and Nuclear). We report, in parentheses, the number of households in each group. The subsequent columns indicate the number of individuals in each age group: "M" stands for male, "F" for female, "Ch" for children, "Ado" for adolescents, and "Adu" for adults. The column "Total" reports the total population in each household type (the sum over the row), while the row "Total" is the population per age-gender group. The rightmost bottom cell reports the total population.
• Figure 2: Contact matrices. Each frame represents one of the aggregations described in Section \ref{['sec:res']}, plus the one obtained by grouping measurement waves. The matrices' entries denote the daily average contact duration per household by gender-age group, expressed in minutes and normalized by the population sizes, and are obtained following the procedure detailed in Section \ref{['sec:boot']}. The number and color-code report the average value of the bootstrap, while the numbers in the bracket are the $10\%$ confidence interval. The population sizes for each aggregation level are reported in Table \ref{['tab:hhdistr']}. The matrices are indexed by gender ("M for male, and "F for female) and age ("Ch" for children, "Ado" for adolescents, and "Adu" for adults).
• Figure 3: Household interaction time per individual.Top row: daily household interaction times normalized by the population, disaggregated by wave, site, household type, and gender of the household head. The $y$-axis indicates the sum of all entries on the upper diagonal of the normalized household contact matrices shown in Figure \ref{['fig:HCM']}. The distributions are obtained from a bootstrap sample, as described in Section \ref{['sec:boot']}. Bottom row: ridge regression. Box plot of the ridge regression coefficients obtained from several bootstrap randomizations. The whiskers show the $5-95$ confidence interval. The face color is gray if the confidence interval crosses zero, and it is orange otherwise.
• Figure 4: Interaction rates disaggregated by age and gender.Top row: daily interaction times by age and gender, normalized by the population and disaggregated by wave, site, household type, and gender of the household head. The $y$-axis indicates the sum of the normalized household contact matrices rows. The distributions are obtained from a bootstrap sample, as described in Section \ref{['sec:boot']}, and aggregated according to the measurement wave (first frame), the measurement site (second frame), the household type (third frame), and the gender of the household head (fourth frame). We highlight in orange the interactions involving male children M-Ch and in light blue those involving adolescents M/F-Adu. Bottom row: ridge regression. Box plot of the ridge regression coefficients obtained from several bootstrap randomizations. The whiskers show the $5-95$ confidence interval. The face color is gray if the confidence interval crosses zero, and it is orange otherwise.
• Figure 5: Interactions with children.Top row: daily interaction rates by age and gender normalized by the population, disaggregated by wave, site, household type, and gender of the household head. The plot shows the histograms of the interaction rates with children (male and female) per individual, for adults and adolescents of both genders. The histograms are obtained from the bootstrap sample described in Section \ref{['sec:boot']} section and aggregated according to the measurement wave (top left frame), the measurement site (top right frame), the household type (bottom left frame), and the gender of the household head (bottom right frame). Bottom row: ridge regression. Box plot of the ridge regression coefficients obtained from several bootstrap randomizations. The whiskers show the $5-95$ confidence interval. The face color is gray if the confidence interval crosses zero, and it is orange otherwise.