Skip to main content
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2026 Oct 1.
Published in final edited form as: Epidemiology. 2026 Jul 6;37(6):756–764. doi: 10.1097/EDE.0000000000002019

Temporal Connectivity of Social Contact Networks in Urban and Rural India

Dehao Chen 1, Prasanna Samuel 2, Rajan Srinivasan 2, Adrien Le Guillou 1, Venkata Raghava 2, Saad B Omer 3, Benjamin A Lopman 1, Samuel M Jenness 1
PMCID: PMC13624582  NIHMSID: NIHMS2210434  PMID: 42406882

Abstract

Background

The epidemic potential of infectious diseases depends on how contacts connect individuals over time—a form of temporal connectivity that has rarely been quantified in resource-poor settings. We used the forward-reachable path (FRP)—the proportion of population reachable from an index person via direct or indirect connections—to quantify temporal connectivity in contact networks relevant for acute respiratory transmission.

Methods

From empirical social-contact data collected in rural and urban Tamil Nadu, India, we derived contact-location–specific network statistics. These statistics were used to parameterize dynamic network models, simulate daily networks over one year, and compute FRPs.

Results

In both rural and urban networks, mean FRPs rose sharply on day 1, then either increased steadily at school and work or plateaued at home and at locations included in the other layer (that is, locations other than home, school, and work). By day 365, mean FRPs followed the order: home (0.06% [rural] and 0.03% [urban]) < school (11.96% [rural] and 9.14% [urban]) < work (12.55% [rural] and 26.72% [urban]) < other (40.54% [rural] and 67.99% [urban]). The mean FRP peaked at home among those aged ≥60 years, at school among those aged 10–19 years, and at work among those aged 40–59 years.

Conclusions

Although FRP at home was bounded by household size, reachability expanded substantially through school, work, and other contacts. These findings indicate high temporal connectivity and substantial epidemic potential for acute respiratory transmission in these settings.

Keywords: Social networks, Epidemic potential, Acute respiratory transmission, India

INTRODUCTION

As highlighted by the recent COVID-19 pandemic, transmission of many infectious diseases is mediated through close human-to-human contacts. These social contacts can be represented as a network, in which individuals are nodes and their contacts are edges. Since the early 2000s, the field of network epidemiology has developed markedly, but much of the work relied on static networks (1), whose contacts (edges) are presumed fixed over the duration of an epidemic. This representation simplifies social contact patterns that are inherently transient: the formation of new connections can be accompanied by the dissolution of old ones, creating pathways along which infection can move through a population. Dynamic network models, which allow edges to form and dissolve, better capture these features. While static models characterize lower-order statistics such as mean degree (the average number of current or extant contacts across individuals) and assortativity (the proportion of within-group contacts), dynamic models also preserve higher-order temporal properties like the forward-reachable path (FRP) (2,3) that are more analogous to dynamic representations of epidemic potentials, such as R0.

The FRP length is the total number of nodes reachable from an index node via temporally ordered edges within a time interval (2,3). In a population where contacts represent possible disease transmission events, the FRP approximates the upper bound of epidemic potential inherent in a dynamic network. The FRP does not predict epidemic size, which also depends on the pathogen’s natural history (e.g., transmissibility, duration of infection), but it provides a baseline estimate of network temporal connectivity that can be compared across individuals and populations. Because the FRP is disease-agnostic, it may serve as a foundation for transmission models of any pathogen spread via the same contact networks. Although the FRP has been used to evaluate the epidemic potential of sexually transmitted infections (4), epidemiologic applications to social contact networks relevant for acute immunizing disease transmission across both high- and low-income settings remain uncommon. We therefore use the FRP in this study to quantify temporal connectivity among individuals in dynamic social contact networks.

Dynamic networks can be modeled using the temporal exponential random graph model (TERGM), which provides a statistically principled framework for representing contact dynamics in the context of infectious disease transmission. This framework allows disease transmission models to be applied on top of the dynamic network simulations, Such network-based transmission models have been used to study context-specific COVID-19 transmission and the effect of control measures on cruise ships (5) and in US jails (6). However, these studies primarily focused on transmission outcomes; the dynamic network properties (e.g., temporal connectivity) themselves remain relatively understudied.

This lack of information about dynamic network properties is especially pronounced in low- and middle-income countries, where studies of infectious disease network dynamics remain scarce (7). The burden of infectious diseases varies between rural and urban areas in India (8), which may be driven in part by differences in epidemic potential across these settings. The GlobalMix study, which collected empirical contact data across multiple resource-poor settings including India (9), provides an opportunity to quantify temporal connectivity in dynamic networks. In this study, we leverage social contact data from rural and urban Tamil Nadu, India to parameterize dynamic network models and assess temporal connectivity using the FRP, with implications for the epidemic potential of acute respiratory disease transmission.

METHODS

Study Design and Field Procedures.

This analysis used empirical data from the GlobalMix study within a TERGM statistical approach for estimation and simulation. The study design of the GlobalMix study has been described in detail elsewhere (9). Briefly, between July 2022 and February 2023, we enrolled 63 households each from a rural site (Kilarasampet village, population 117,808) and an urban site (an urban ward in Vellore district, population 257,977) in Tamil Nadu, India (126 households total). At each site, we enrolled 624 participants, each of whom completed a two-day social contact diary about the location, contact age, and duration of all social contacts in each day. A total of 9,785 and 9,761 contacts were reported in the rural and urban sites, respectively. Hereafter, we refer to each site’s total population as target population and the enrolled participant sample as the study population.

Estimation of Forward-Reachable Path.

We estimated the FRP from the GlobalMix contact network data using the four-step method illustrated in Figure 1. We adapt the FRP estimation framework introduced by Armbruster et al (2), previously applied to sexual contact networks by Anderson et al (4), for use with social contact networks. Analyses were performed in R using the EpiModel package v2.5.0 (10), in which we designed a streamlined FRP estimation algorithm based on the general FRP estimator developed in the tsna package (11). The analysis scripts for the following procedures are available at: https://github.com/EpiModel/COVID-GlobalMix/.

Figure 1.

Figure 1.

Workflow for estimating forward-reachable paths (FRPs)

Step 1: Network Statistics Estimation.

Step 1 has two parts. First, Step (1a) involves estimation of per-capita egocentric network statistics, representing the contact characteristics among the study participants. Second, Step (1b) involves estimation of population-level target statistics by adjusting the per-capita egocentric network statistics from Step (1a) by the target population’s demographic (i.e., age) distribution and then scaling them up to represent the contact characteristics of the target population. These target statistics will serve as the fitting targets for TERGM estimation in Step 2.

In Step (1a), we derived per-capita egocentric network statistics from the contact diary for four locations — home, school, work, and other (any locations other than home, school, and work) — in each study site. These statistics included: 1) the egocentric mean degree of participants (i.e., the average number of contacts per participant), 2) the age-specific mixing proportions between participants and contacts, and 3) the duration of contact. Hereafter, we refer to each study site (rural, urban) as a network and each contact location (home, school, work, other) as a layer. Considering a strong negative correlation in school and work degrees in both urban and rural sites, we additionally calculated a degree-based extension of statistics 1): their cross-layer egocentric mean degrees (the mean degrees at one layer conditional on having contacts at the other layer). For statistic 3), we set the layer-specific mean acquaintance duration (i.e., the duration of knowing a contact) as the duration of contact. The mean acquaintance duration was calculated by converting acquaintance duration categories from the contact diary to numeric values and taking a weighted average using the proportion of daily contact as the weight. Because our goal was to estimate FRP relevant for acute respiratory transmission, we excluded contacts with duration less than 15 minutes (i.e., non-close contacts) in the above calculations (12). We further detailed the mathematical formulations and calculation steps for these per-capita network statistics in eAppendix 1.

In Step (1b), we used the estimates from Step (1a) to calculate two layer-specific population-level target statistics: the total number of edges between individuals (nodes) in an age group a and those in age group b, and the number of nodes having contacts at a layer, conditioning on the contact status of the other layer, for school and work layers only. For computational efficiency, we defined the modeled population (a simulated set of nodes representing individuals in each site’s contact network, comprised 11,781 and 25,798 nodes in the rural and urban networks, respectively) as 10% of each site’s target population (the full population a simulated network was intended to represent) size. eFigure 1 shows the age distribution of the study and modeled populations; the age distribution of the modeled population in each network reflected that in the target population, which skewed to the older groups, whereas the study population (the surveyed individuals from whom contact data were collected) was skewed to the younger groups. We used the modeled population’s age distribution to derive age-specific population sizes and calculated the ratio of each age group’s proportion in the modeled population to its proportion in the study population as a post-stratification weight. To calculate the first population-level target statistic (the total number of edges between individuals in different age groups), for each directed pair of age groups (e.g., group a contacted b, a −>b), we multiplied the egocentric mean degree of group a (estimated in Step 1a) by its modeled population size and by the normalized contact frequency with group b. We repeated this process for b −> a and summed the edge numbers of both directions as the total edge count between a and b (a—b). To derive the second population-level target statistic (the number of nodes with contacts in a layer conditioning on contact status in the other layer), we first applied post-stratification weights for the cross-layer egocentric mean degree to adjust for the differences in age distribution between the study and target populations (eFigure 1). Then, we multiplied the resulting conditioned mean degree (e.g., mean school degree among nodes with work contacts) by the number of nodes having contact at the conditioning layer. We calculated the layer-specific mean degree from the total edges counts and reported them as empirical individual-level network statistics. We further detailed the mathematical formulations for these population-level target statistics in eAppendix 2.

Step 2: Network Model Estimation.

This step fitted separate temporal exponential random graph models (TERGMs) to the specified target statistics of the school, work, and other layers of each site (estimated in Step 1) in order to estimate formation coefficients, the log odds of a partnership between any two nodes in a given layer. In each TERGM, we used a transformed empirical mean acquaintance duration as the estimated edge duration coefficient, which determined edge dissolution (i.e., the termination of an existing edge over time) under an exponential distribution. —The dissolution probability is thus inversely proportional to the edge duration. We initialized nodal age attributes using the age distributions of the target population and the nodal contact attribute using the empirical age-stratified contact proportion of the study population. Because of dependency between the models pertaining to the school and work layers, we initialized the school layer contact attribute using the work layer degree (i.e., the number of edges a node had in the work layer) and vice versa. At the other layer, an TERGM with edges-only formation coefficients implies a Poisson degree distribution (13). The observed degree distribution at this layer exhibited an inflated zero-degree count (i.e., a higher proportion of isolates [individuals without contact] than a Poisson distribution would predict). To capture this non-Poisson feature, we initially added a non-parametric isolate term (degree(0)) to the TERGM. However, the joint edge–isolate model did not converge, instead exhibiting a bimodal stationary distribution—a known phase transition in this model family (13). As an alternative, we treated isolate status as a fixed nodal attribute in network estimation, that is, we estimated edge formation only among nodes with initial contacts at the other layer. Both approaches are statistically principled and practically achieve the same result—fitting the non-parametric degree distribution with its inflated zero count—but the latter avoids the convergence problem. We assumed that household members contacted all other household members and that these ties were permanent (i.e., no edge dissolution), effectively making the home layer static and saturated at the household level. Therefore, we did not estimate a TERGM for the home layer. We assumed no travel between the rural and urban sites. We further detailed the probabilistic model specifications for generating these nodal attributes in eAppendix 3.

We conducted standard dynamic network model diagnostics to assess goodness-of-fit to the target statistics (10). We simulated 30 networks from each fitted TERGM and compared the simulated mean of each target statistic with its corresponding value from Step (1b). Adequate fit was defined as a <5% relative difference between the simulated and target values. This threshold is consistent with the general principle that Markov Chain Monte Carlo (MCMC)-based network simulations should reproduce their target statistics with minimal bias, analogous to convergence diagnostics for Markov Chain Monte Carlo estimation. This criterion has been applied in prior TERGM-based studies, including analyses of forward-reachable paths in sexual contact networks (4) and network-based models of SARS-CoV-2 transmission (14).

Step 3: Network Simulation.

This step applied the estimated TERGM coefficients from Step 2 to simulate edge formation and dissolution for the dynamic (school, work, and other) layers in daily time steps over a one-year period (from day 0 to day 365). This period aligns with the typical time scale over which respiratory epidemics unfold. We assumed the modeled population was closed, so the nodes were fixed. On day 0, all nodes existed without edges (no contacts). We simulated each layer independently using its fitted TERGM to generate the baseline network on day 1. Starting from day 1 and continuing through day 365, we assigned the daily simulated network degree of the school to the work layer and vice versa; we simulated the other layer using its own degree. For the static home layer, adapting the approach from Wallrafen-Sam et al (14), we used the empirical household composition and the layer-specific mean degree to simulate household identifiers and edges fully saturated in each household. We replicated the entire simulation 100 times to quantify uncertainty across simulations. These simulated daily networks were used as inputs for FRP computation in Step 4.

Step 4: Estimation of Forward-Reachable Path.

The FRP is computed at the individual level (2,3). Because our modeled population represented a subpopulation of the target population, we operationally defined FRP as the proportion of the modeled population reachable from an index node via direct or indirect connections—that is, the ratio of a node’s FRP length to the total network size. Since all nodes were without edges at network initiation (day 0), the FRP length of each node was 1 (i.e., the node itself)—corresponding to an FRP of 1/11,781 for the rural network (N=11,781). If 5,000 nodes formed edges and were directly or indirectly connected to each other at day 1, and an index node formed an edge with any of these 5,000 nodes, its FRP would increase sharply from 1/11,781 to 5,001/11,781. Such strong interconnectivity is characteristic of our social contact networks and is explained further in the Discussion.

As time progressed and daily edges formed and dissolved, a node’s FRP increased cumulatively as new connections occurred. For each node, we used the layer-specific daily network edge list (all contacts each day) from the simulated network to compute FRP over days 1 to 365. To facilitate interpretation, we report FRP as a percentage in the Results. Across the 100 simulations, we first calculated, at each time step, the mean FRP across nodes within each simulation. We then reported the median and 95% confidence interval (2.5th and 97.5th percentiles) of these 100 means. To illustrate node-level variability within a single simulation, we also characterized the FRP distribution across nodes by its mean, standard deviation, median, and interquartile range. We performed these calculations separately for each network layer, stratified by nodal age group. In addition to presenting the full one-year time series, we also summarized FRP at selected time points (e.g., day 180) to provide snapshots during the simulation.

RESULTS

Empirical Network Statistics.

Table 1 presents individual-level network statistics. The home layer had the highest overall mean degree (5.44 and 6.19 in the rural and urban networks), followed by the other, school, and work layers. Across all layers except school, the rural mean degree exceeded that of the urban network. At the other layer, 64.2% and 82.6% of the rural and urban participants reported at least one contact. The conditional mean degrees at the school and work layers, among those with ≥1 contact, were much lower than their corresponding overall mean degrees, and these conditional values were lower in the rural network than in the urban network for both layers.

Table 1.

Individual-level network statistics for contact networks in rural and urban populations in Tamil Nadu, India

Home   School   Work   Other
Mean Degree Rural Urban Rural Urban Rural Urban Rural Urban
 Overall 5.44 6.19 1.21 (0.02) 1.02
(0.13)
0.92
(0.02)
1.31 (0.08) 3.67 4.62
Age group a$
0-9 6.54 7.38 1.74 1.22 0 0 3.95 4.45
10-19 7.31 6.83 6.17 4.70 0 0 4.91 4.60
20-29 5.29 6.14 0.48 0.16 1.08 1.33 3.44 5.43
30-39 5.85 6.67 0.50 0.16 1.24 2.02 3.55 4.55
40-59 4.91 5.49 0.28 0.16 1.78 2.84 3.81 4.61
60+ 3.87 4.67 0 0 0.41 0.38 2.67 3.82
Percentage of non-isolates$ - - - - - - 64.17% 82.64%
Assortative mixingb
0-9 16.16% 22.86% 82.54% 82.47% 0% 0% 27.78% 23.86%
10-19 22.73% 26.18% 79.49% 90.95% 0% 40.74% 46.20% 51.85%
20-29 13.97% 24.87% 13.04% 0% 30.85% 8.82% 29.27% 27.30%
30-39 15.78% 16.01% 4.76% 54.55% 32.26% 26.98% 26.58% 28.80%
40-59 19.71% 18.13% 0% 0% 68.55% 88.11% 44.77% 47.99%
60+ 14.85% 17.51% 0% 0% 11.11% 5.13% 22.94% 18.06%
Mean contact duration# - - 5 3 18 12 12 10
a

Because the model overestimated this percentage at the other layer, we reported its empirical value from study participants of the GlobalMix study for adjustment in network estimation.

b

The mean contact durations were presented in years but fitted by our models in days; the age groups were in years.

At school, the highest mean degree occurred among nodes aged 10-19 years — 6.17 and 4.70 in the rural and urban networks, respectively — whereas those aged ≥ 20 years had a mean degree < 1. At work, mean degrees were < 1 for those aged ≤ 19 or ≥ 60 years, while the middle age groups had mean degree ranging from 1.08 to 2.84 for both networks. At both the home and other layers, mean degrees were >2 across all age groups. At home, the degree peaked at ages 10-19 years and then declined with age, whereas the other layer had a smaller variation between age groups, and those aged ≥ 60 years had the lowest values. The distributions of the age-specific assortative mixing proportion across layers generally mirrored those of the age-specific mean degrees. eFigure 2 shows the empirical age-mixing proportion matrices—high proportions occurred not only in assortative mixing (e.g., 0.83 for 0–9 with 0–9 at rural school and 0.88 for 40–59 with 40–59 at urban work) but also in non-assortative mixing (e.g., 0.78 for ≥ 60 with 40–59 at rural work and 1.00 for 20–29 with 10–19 at urban school). Mean acquaintance duration was the longest at the work layer (18 and 12 years in rural and urban networks, respectively) and shortest at the school layer (5 and 3 years in rural and urban networks, respectively).

eTables 1 and 2 http://links.lww.com/EDE/C366 present the models’ fitted target statistics. All differences between simulated values and their corresponding targets were <5%, indicating an adequate goodness-of-fit.

Forward-Reachable Path.

Figure 2 shows the mean FRPs from 100 simulations over the full 1-year period; eTables 3–6 http://links.lww.com/EDE/C366 present mean FRPs at days 30, 180, and 365 as snapshots of these trends (see Figure 2). Across both rural and urban networks, mean FRPs followed the order of home < school < work < other. In all layers, the means rose sharply from time 0 to time 1, except in the school layer for nodes aged ≥ 60 years and in the work layer for nodes < 19 years, whose mean degree = 0. After day 1, mean FRPs plateaued across age groups at home (0.06% in the rural network and 0.03% in the urban network) and at the other layer (40.54% in the rural network and 67.99% in the urban network by day 365). In comparison, mean FRPs at school and work increased steadily after day 1, reaching 11.96% (rural) and 9.14% (urban) at the school layer, and 12.55% (rural) and 26.72% (urban) at the work layer by day 365 (eTables 3–6 http://links.lww.com/EDE/C366).

Figure 2.

Figure 2.

Mean forward-reachable paths (FRP, percentage of population reached) in rural and urban networks in India over a 1-year period. In each panel, the age-specific 50th (solid line), and 2.5% and 97.5% percentiles (colored error band) of mean FRP estimates across 100 network simulations were shown. The y-axes were different across settings.

Figure 3 shows the mean FRPs at day 180 to illustrate age-specific patterns. At home, nodes aged ≥ 60 years had the highest mean FRP, followed by those aged 10-19, 30-59, 0-9, and 20-29 years. At school, those aged 10-19 had the highest FRP (33.45% and 27.60% in the rural and in urban networks, respectively), followed by those aged 0-9, 20-59, and ≥60 years (mean degree = 0). At work, the highest FRP occurred among those aged 40-59 (22.93% in rural and 47.95% in urban), followed by those aged 20-39, ≥60, and 0-19 years (mean degree = 0). The other layer showed less distinct age patterns, with FRPs ranging from 30.82% to 47.48% in the rural and from 60.02% to 74.89% in the urban network (eTables 3–6 http://links.lww.com/EDE/C366).

Figure 3.

Figure 3.

Boxplots of mean forward-reachable paths (FRP, percentage of population reached) across layers in the rural and urban networks at day 180 from 100 simulations. The mean FRP of each simulation was calculated as the average across nodes. The y-axis scale for the home layers was much lower, as reachability was limited at the household level. The y-axes were different across settings.

Figure 4 shows the time series of nodal FRPs over 1 year from a single representative simulation. eTables 7 and 8 http://links.lww.com/EDE/C366 present summary statistics of nodal FRPs at days 30, 180, and 365 (see Figure 4). The ranking of layer-specific FRPs in this run mirrored the mean FRP ranking from 100 simulations — except in the rural network where, by day 365, the school layer’s mean nodal FRP slightly exceeded that of the work layer. Unlike the mean FRP, which jumped sharply only at day 1 (Figure 2), nodal FRPs at all layers except home increased at various points throughout the year. After the initial rise, nodal FRPs at school and work grew gradually, reaching mean values of 11.84% (rural) and 9.23% (urban) at school, and 11.49% (rural) and 26.43% (urban) at work by day 365. At the other layer, nodal FRPs plateaued at the overall contact proportions, with medians of 63.65% (rural) and 82.48% (urban) and mean of 40.52% (rural) and 68.03% (urban) by day 365. At home, mean nodal FRPs reached 0.06% (rural) and 0.03% (urban) at day 1 and remained unchanged thereafter. Notably, except at the urban work layer, more than half of nodes at the school and work layers of both networks maintained an FRP of 0% (nodes without edges) throughout the entire time series (eTables 7 and 8 http://links.lww.com/EDE/C366). Figure 5 shows the frequency distributions of nodal FRPs at day 180 by layer and age group. At the home layer, FRP distributions were unimodal across all age groups. In the non-home layers, distributions were bimodal in some settings — one mode at FRP = 0% and the second among nodes with edges. The age distributions of nodes in the second mode mirrored the ranking of age-specific mean degrees in those layers — higher FRPs corresponded to higher mean degrees.

Figure 4.

Figure 4.

Nodal forward-reachable paths (percentage of population reached) in the rural and urban networks over a 1-year period. The y-axis scale for the home layers was much lower, as reachability was limited at the household level. The y-axes were different across settings.

Figure 5.

Figure 5.

Frequency distributions of nodal forward-reachable paths (FRP, percentage of population reached by a node) in the rural and urban networks at day 180. The x-axis showed the nodal FRPs at a specific layer and age group, and the y-axis showed the frequency of nodes with these nodal FRPs by network. The x-axis scale for the home layers was much lower, as reachability was limited at the household level. The x-axes were different across settings.

DISCUSSION

Using the FRP estimator of temporal connectivity, we quantified epidemic potential in social contact networks representing rural and urban settings in South India. In the rural and urban networks, the population-level mean and nodal FRPs rose sharply to a threshold and then increased steadily (at school and work layers) or plateaued (at home and the other layers) (Figure 2, Figure 4). The mean FRP followed the order home < school < work < other, indicating the relative overall epidemic potential across contact locations, with the other layer having the highest epidemic potential and the home layer the lowest (Figure 2). Age-specific peaks also shifted by layer: at home the oldest age group (≥60 years) had the highest FRP, at school it was adolescents (10–19 years), and at work middle-aged adults (40–59 years) (Figure 3). The other layer exhibited less clear age patterns, likely because it contained all other locations (e.g., market, public transportation) visited by people from across all age groups. Nodal FRPs at the home location corresponded to household size, ranging from single-person to multigenerational households. At the other layer, nodal FRPs were bounded by the proportion of individuals with any contact in those locations, whereas at the school and work layers, ≥ 50% of nodes maintained an FRP of 0% throughout the simulation in both networks (except in the urban work layer) (Figure 4). Overall, this analysis provides statistically principled estimates of epidemic potential across these population settings in South India, as well as across contact locations and demographics. These estimates are useful in themselves to understand the upper bound of disease transmission, and they will be useful as the foundation of future network-based epidemic models.

The FRP dynamics described above are shaped by the formation of giant network components, i.e., large clusters of interconnected nodes. Network theory has established that a mean degree of 1 is a critical threshold for giant component formation (15). In our estimated networks, the overall mean degrees exceeded this threshold in all layers except the urban work layer; even there, the age-specific mean degree for working-age nodes (20-59 years) remained above the threshold. This, combined with the long edge durations (and correspondingly low dissolution probabilities), suggests that a giant component formed in each dynamic layer (school, work, and other) and persists over time. However, layer-specific constraints limited the size of these giant components: at the other layer, the proportion of non-isolated nodes capped both the FRP and the component size; and at the school and work layers, edge formation capacities were constricted by cross-layer effects. These traits made the component magnitude more realistic and prevented the giant components from encompassing the full population, yielding more realistic estimates of epidemic potential. In other words, the network layers allowed large interconnected clusters to form, but these clusters did not include the entire modeled population.

The giant components set the upper threshold for each node’s FRP. Whenever an isolated node formed an edge with the giant component, its FRP length jumped from 1 to the component size—producing the sharp day-1 increase observed across the dynamic layers. At the school and work layers, subsequent gradual increases in nodal FRP resulted from that node forming additional edges with other isolated nodes. At the school and work layers (except the urban work layer), this upper threshold declined over time, reflecting shrinkage of the giant component, likely because the conditioned mean degree was much lower than the overall mean degree. At the other layer, isolate status was fixed, so the FRP jump was driven entirely by edge formation among the predetermined non-isolated nodes and the giant component. At the home layer, where no giant component formed and edges were fixed, each component corresponded to a simulated household, so the nodal FRP distribution mirrored household size distribution. Whereas the mean FRP captures a layer’s overall epidemic potential, the nodal FRP captures individual-level reachability. Because giant components existed in each non-home layer, nodes that visited multiple layers contributed to cross-layer connectivity, potentially linking these separate giant components and increasing reachability toward the full network. In the limit, this cross-layer connectivity would produce network saturation (FRP = 100% for all nodes), providing a structural basis for population-wide pathogen transmission.

These FRP patterns for social contacts contrast with patterns observed in sexual partnership networks among men who have sex with men (MSM) in the US, where mean degrees were all under 1 (4). As expected, reachability in social contact networks was substantially greater. In sexual networks, where mean degrees were below 1 and giant components were likely absent, temporal FRPs increased gradually with sigmoid and curvilinear patterns, reflecting the slow accumulation of connections over time. Although the FRP measures epidemic potential, infectious disease transmission also depends on pathogen-specific natural history (e.g., transmission probability per contact), and the FRP better predicts the transmission path when that probability approaches 100% (4). Hence, the FRP may be more useful for approximating epidemic size for pathogens of acute infections (e.g., influenza virus, SARS-CoV-2, respiratory syncytial virus) than for chronic infections such as HIV.

Our study has several limitations stemming from the empirical contact data used to parameterize the network target statistics (9). First, GlobalMix participants might have been more likely to recall contacts with people they knew well (e.g., family members, schoolmates or workmates) than with strangers, potentially underestimating contact proportions at the other layer. Second, for children under ten years old, guardians reported contacts on their behalf, which may introduce inaccuracies for this age group. Third, because India contact diaries were surveyed between mid-2022 and early-2023 amid the COVID-19 pandemic, the contact patterns might not generalize to non-pandemic contexts without social distancing measures. Fourth, the time span of the survey was short (under one year), precluding parameterization of seasonal fluctuation in contact patterns (e.g., holidays, start and end of school year). Fifth, given we did not have empirical data on travel between the rural and urban sites to support the parameterization, we did not model interconnectivity between the two networks. Additionally, despite leveraging computational parallelization and high-performance computing, estimating the FRP for the full target population exceeded the capacity of our computation resources, so we modeled only a subset of that population. Despite these limitations, to our knowledge, this is the first study to characterize FRP in social contact networks using multilayer network models parameterized from comprehensive network statistics representing diverse mixing traits and network structures.

The FRP—depicting temporal reachability across key social-contact settings—offers a quantitative measure for epidemic potential without requiring a full-scale transmission model. Giant network components are present in all contact locations of our social contact networks except home, driving the sharp rise in an individual’s FRP and accelerating reachability. Visits to multiple layers likely connect these separate giant components, extending reachability toward the entire population and maximizing epidemic potential. This may be curtailed by interventions that limit visits across multiple locations, as observed during the COVID-19 pandemic. Future work should validate these FRP estimates against cumulative-incidence projections from full transmission models built upon the same network framework.

Supplementary Material

Appendix

Acknowledgements:

This study would not have been possible without the generous time and support of the individuals who participated. We are grateful for the work of field teams and data managers at the Christian Medical College. We also acknowledge the entire Global Mix collaboration that includes additional investigators from Mozambique, Guatemala and Pakistan. We would like to thank Maria Garcia Quesada for her input in deriving the network statistics.

Funding:

This work was supported by grants R01 HD097175 and R01 AI161399 from the National Institutes of Health.

Footnotes

Conflict of Interest: We declare no conflicts of interest.

Data and Computing Code:

The data are available upon request. The computing code for this study can be accessed at https://github.com/EpiModel/COVID-GlobalMix.

REFERENCES

  • 1.Bansal S, Read J, Pourbohloul B, Meyers LA. The dynamic nature of contact networks in infectious disease epidemiology. J Biol Dyn. 2010. Sep;4(5):478–89. doi: 10.1080/17513758.2010.503376 [DOI] [PubMed] [Google Scholar]
  • 2.Armbruster B, Wang L, Morris M. Forward reachable sets: Analytically derived properties of connected components for dynamic networks. Netw Sci (Camb Univ Press). 2017. Sep 1;5(3):328. doi: 10.1017/NWS.2017.10. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3.Moody J. The Importance of Relationship Timing for Diffusion. Social Forces. 2002. Sep 1;81(1):25–56. doi: 10.1353/SOF.2002.0056 [DOI] [Google Scholar]
  • 4.Anderson EJ, Weiss KM, Morris MM, Sanchez TH, Prasad P, Jenness SM. HIV and Sexually Transmitted Infection Epidemic Potential of Networks of Men Who Have Sex With Men in Two Cities. Epidemiology. 2021. Sep 1;32(5):681–9. doi: 10.1097/EDE.0000000000001390 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Jenness SM, Willebrand KS, Malik AA, Lopman BA, Omer SB. Dynamic network strategies for SARS-CoV-2 control on a cruise ship. Epidemics. 2021. Dec 1;37:100488. doi: 10.1016/J.EPIDEM.2021.100488 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6.Schneider I, Wallrafen-Sam K, Kennedy S, Akiyama MJ, Spaulding AC, Jenness SM. Interventions for SARS-CoV-2 prevention among Jailed adults: A network-based modeling analysis. Infect Dis Model. 2025. Jun 1;10(2):628–38. doi: 10.1016/J.IDM.2025.02.001 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.Tripathy JP, Lakshmi P, Anand T, Deshmukh PR. Designing, Implementing and Optimising a Capacity-Building Model for Infectious Disease Modelling in India. Ann Glob Health. 2024;90(1):84. doi: 10.5334/aogh.4606 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Sangar S, Thakur R. Infectious diseases in India: assessing the role of household amenities and socio-demographic determinants. J Public Health (Bangkok). 2021. Apr 20;30(12):2809–17. doi: 10.1007/s10389-021-01549-3 [DOI] [Google Scholar]
  • 9.Aguolu OG, Kiti MC, Nelson K, Liu CY, Sundaram M, Gramacho S, et al. Comprehensive profiling of social mixing patterns in resource poor countries: A mixed methods research protocol. PLoS One. 2024. Jun 1;19(6):e0301638. doi: 10.1371/JOURNAL.PONE.0301638 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Jenness SM, Goodreau SM, Morris M. EpiModel: An R Package for Mathematical Modeling of Infectious Disease over Networks. J Stat Softw. 2018. Apr 20;84:1–47. doi: 10.18637/JSS.V084.I08 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11.Bender-deMoll S, Morris M. Tools for Temporal Social Network Analysis [R package tsna version 0.3.6]. CRAN: Contributed Packages. 2025. May 5. doi: 10.32614/CRAN.PACKAGE.TSNA [DOI] [Google Scholar]
  • 12.Infection Control Guidance: SARS-CoV-2 | COVID-19 | CDC [Internet]. [cited 2025 Jul 20]. Available from: https://www.cdc.gov/covid/hcp/infection-control/index.html
  • 13.Butts CT. On the equivalence of the edge/isolate and edge/concurrent tie ERGM families, and their extensions. Journal of Mathematical Sociology. 2016;40(1):1–6. doi: 10.1080/0022250X.2015.1112386 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.Wallrafen-Sam K, Quesada MG, Lopman BA, Jenness SM. Modelling the interplay between responsive individual vaccination decisions and the spread of SARS-CoV-2. Epidemics. 2025. Jun 1;51:100831. doi: 10.1016/J.EPIDEM.2025.100831 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.Solomonoff R, Rapoport A. Connectivity of random nets. Bull Math Biophys. 1951. Jun;13(2):107–17. doi: 10.1007/BF02478357 [DOI] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Appendix

Data Availability Statement

The data are available upon request. The computing code for this study can be accessed at https://github.com/EpiModel/COVID-GlobalMix.

RESOURCES