Computational models of contagion can provide insight and foresight into the behavior of particular pathogens for specific populations. The representation of contacts between population members has a critical impact on simulation outcomes [1, 2]. It has long been recognized that models that assume random mixing between aggregate population categories can alter the dynamics and results of the simulation of contagion spread [3], and may inhibit insight into intervention trade-offs [4]. Reflecting such findings, the past decade has witnessed the growing use of agent-based transmission models, which can better capture the impact of individual-level contact patterns on both individual risk and population-wide transmission dynamics. Such models have frequently represented the inter-agent contact patterns as static graphs with edge weights corresponding to threat of infection drawn from uniform, power-law [5], or other [5, 6] distributions.

More recently, researchers have used microelectronic devices such as motes [7–9], RFID tags [10], cell phones [11–14], or custom-built wireless technologies [15] to capture contact distance, duration and location, as well as other parameters, with significantly more accuracy than previous diarying [16, 17] or retrospective self-report [18, 19] methods, and substantially greater accuracy than cellular [20, 21] or WiFi [22] location estimates whose modeling use entails assuming random mixing patterns between individuals in the same or nearby geographic regions. Data collected from such dynamic contact networks has been used to simulate the spread of a pathogen through the recorded network, using both the raw sample data [15, 23, 24], and aggregated versions where infection probability is drawn from an empirical rather than functional distribution [9]. While electronically supported micro-contact data collection offers significant spatial-temporal resolution and compliance advantages over traditional techniques, the relatively high cost and logistical effort involved in a deployment of telemetry systems and their limited on-board energy capacity imposes design tensions between study duration, sampling rate, and participant count. While study design has been shaped by clear cost-economic considerations favoring shorter deployments with larger populations [9] or longer deployments with smaller populations [8, 23], it has been been conducted absent a clear understanding as to how study design or post-study data aggregation affect transmission model accuracy. Reflecting the power hungry character of the data collection devices, studies have frequently favoured larger, high sampling-rate designs of just one [9], two [10, 25] or a handful of days [26] in duration. While some researchers have argued that such short sampling periods capture one or more typical day(s) [9], there remains a dearth of formal evidence for the representative character of short sampling periods, and the effects of limited sampling periods on the quality of subsequent simulation outcomes have not been explored.

Researchers have employed diverse strategies for using micro-contact data in transmission models. In some cases [25], researchers have sought to use traditional static representations, with weights on the link connecting two individuals being proportional to the cumulative contact duration observed between those individuals over the study period. Other studies [9, 23, 24, 27] have employed strategies in which dynamic contact patterns collected over an interval of time are used directly as the contact pattern for that the same time horizon. In initial examinations of the effects of aggregation of microcontact data on simulation outcome, the authors of [28] noted the effects of aggregation on contagion using two sets of empirical data, and the authors of [25] explored the impact of changes in model parameter values and network aggregation measures – including alternate use of both weighted and unweighted static representations – to simulation outcomes. Because the dataset in [28] was collected over a relatively short (2 day) study duration, the researchers replicated the recorded data and variants over longer simulation horizons (60 or 100 days) to ensure sufficient duration to capture the dynamics of the simulated outbreak, and found that application of unweighted homogeneous aggregation schemes led to significant overestimation of infection spread. The study further examined three strategies for extending the time frame, with all three approaches positing that the observed contact recorded over the brief study period are representative of a longer period of time. The researchers found notable epidemic-phase-specific differences between the typical day [9] variants, reflecting whether contacts occur repeatedly with the same set of individuals or different individuals. Most importantly, simulated infection transmission on a duration-weighted aggregated static graph yielded results close to those from a fully disaggregated but replicated dynamic network. However, because of the short study duration, the researchers were unable to evaluate the fidelity of any of the three approaches to simulated infection transmission over a disaggregated dynamic network for the entire simulation time horizon. While this contribution yielded important insights into the effects of aggregation and replication strategies on simulation outcomes, the researchers emphasized the importance of analyzing data from longer-duration studies, to validate the fidelity of the proposed strategies for study data extrapolation.

To help address and expand upon the questions raised in [25] regarding the impact of temporal aggregation and extrapolation, this paper analyzes such effects for the Flunet dataset [8] in the context of an influenza-like illness (ILI). The Flunet dataset is characterized by a two minute-level temporal contact resolution, and a smaller population (N=36) observed over a much longer time period (T=92 days) than other studies [9, 10]. Taking advantage of the longer time duration of this study, we examine agent-based simulation using a previously contributed ILI model [29], and both previously contributed [8, 23, 24] and novel dynamic network structures, and previously used but unevaluated strategies for extending the study timeframe. We report three important findings:

Even the weighted “typical-day” techniques advanced in [9, 25] are fragile because the variance in contact patterns between days can lead to highly variable simulation outcomes;

Confirming with an extended temporal horizon and extending findings of [25], aggregated network representations tend to overestimate the disease burden in the population;

While static synthetic networks with contact distributions can produce results consistent with the full dynamic network, this apparent agreement can be due to an accumulation of counter-balancing inaccuracies leading to a deceptively correct outcome.

Overall, our findings suggest that longitudinal data collected over a prolonged period is required to accurately reconstruct dynamic or static contact networks for the purpose of simulating pathogen spread for at least some subpopulations, and that simulations based on summary networks tend to systematically overestimate the risk of infection. These findings are important because they suggest that traditionally calibrated disease models may contain biased disease parameter estimates to compensate for undetected inaccuracies in the underlying network and mixing models. These biases in disease parameter estimates might then lead to erroneous conclusions about the relative impact of interventions or the rate of disease propagation. Because such inaccuracies may not cancel in the same fashion when investigating intervention effects, additional caution should be employed when using such aggregation and generalization schemes. Unfortunately, while typical day [9, 25] contact data acquisition techniques ease study design, they do not appear to be an adequate solution, as they can yield even more erroneous disease burden estimates than traditional aggregate models. This leads us to finally conclude that larger-scale longer-term longitudinal studies may be required to generate sufficiently accurate descriptions of the contact dynamics in many populations. Such studies are urgently needed to help identify the balance between study size and length required to secure reliable insight into intervention tradeoffs.

Given the H1N1 strain emergence and in anticipation of the significance of the 2009–2010 influenza season, the co-authors launched a previously-described [24] pilot study in the midwestern Canadian city of Saskatoon to electronically collect contact patterns between 36 participants, in addition to their influenza-related health status information. The study was conducted between November 9th, 2009 and February 9th, 2010 – a time period coinciding with the second rise of reported H1N1 cases in the province of Saskatchewan [30]. Each participant was requested to carry a proximity sensor whenever awake during the study period, and to respond to a sequence of weekly health surveys via a web browser. Participants filled out an informed consent form prior to joining the study, as required by the university research ethics board.

To study the impact of network representation on simulation outcomes, an H1N1 infection transmission model [24] was created and parameterized with empirical characteristics of the H1N1 pathogen [31]. The model simulated the spread of infection between agents through their daily interactions. Agents’ daily connectivity – either from empirically collected or aggregate approximations – were imported into the model and defined the potential for infection between discordant agents. Overall, 616 different scenarios – each positing a different connectivity network derived from the data – were simulated, with each such scenario being run for 10,000 realizations by replaying a sequence of the 265,000 thirty-second time slots in the study period. Based on the 100,000 run ensembles with the same model and dynamic network reported in [24] we concluded that 10,000 runs were sufficient, as the maximum depth of infection in [24] was reached 10 times, implying that division by 10 would have a high probability of capturing even the most extreme infection events, while maintaining the probability distribution at significantly reduced computational cost. Because each run could be conducted relatively swiftly (typical run-times for an ensemble were 30 s) having a large number of runs was not computationally prohibitive given the resources available to us.

The underlying disease parameters and distributions were held constant across scenarios. Within a given scenario, stochastics associated with exogenous and endogenous infection transmission and duration of different phases in the natural history of infection induced variability in simulation output.

Transmission model

As the data collection occurred during the H1N1 outbreak, we used an agent based H1N1 SEIR transmission model to simulate the infection dynamics. This section includes a short description of the model; interested readers are referred to a detailed specification in [24]. The simulation model classified each individual in the sample population into one of seven states: Susceptible, Latent, Asymptomatic Infectious, Symptomatic Infectious, Symptomatic Non-Infectious, and Recovered. All the agents in the model started in the Susceptible state, consistent with limited pre-existing population-level immunity to H1N1. A susceptible individual could contract the infection either from exogenous or endogenous sources. Exogenous sources are defined as the population outside the study who were in contact with Flunet participants and could transmit the infection to the monitored individuals, while endogenous sources are other Flunet participants in an infectious state within the simulation. Both endogenous and exogenous infection forces of infection were calculated using data from the 2009 outbreak [30], a prominent H1N1 model [31], and collected contact data [24].

A susceptible agent contracting the infection from either exogenous or endogenous sources transitions to the Latent state. When an individual enters the Latent period, the model computed the duration for each of the subsequent four stages of illness (Figure 1). In determining these durations, we sought to reproduce the observed variability in H1N1 progression by drawing the duration of incubation and duration of symptoms from two log-normal distributions [24]. The duration for other stages were calculated using these two values.

Each infected agent experienced the four illness states sequentially with the passage of time. A person in the Asymptomatic Infectious or Symptomatic Infectious state was considered infective. At each time the infective person triggered potentially transmitting events (e.g. coughs or sneezes) with a specified likelihood. A potentially transmitting event had a given probability of transmission to each susceptible in contact. The simulation model did not consider H1N1 mortality, self-quarantine, antiviral administration, or hospitalization outcomes. While the H1N1 model employed has been previously used the capture the impact of vaccination [24], vaccination was not considered in this study.

Connectivity patterns

This work seeks to analyze and quantify the impact of contact pattern representation on transmission outcomes in agent-based simulation models. Because we had recourse to data offering considerably greater temporal span than most other past contributions in this area, we could more readily examine the impact of varying levels of temporal aggregation on model outcomes. To do so, we looked at three different experimental manipulations, each associated with additional parameter variations focused on a particular type of network representation. A combination of manipulation and parameter variation is termed a scenario.

The first baseline manipulation focused on two reference scenarios, each representing different extremes in the aggregation spectrum. The first employed the Flunet empirical data directly, as it captures the contact patterns between agents with the greatest fidelity. In the dynamic baseline graph, edges have weights of 1 or 0 – that is, they correspond to the existence or absence of a connection during a given timeslot. The dynamic contact network can be visualized as a series of network where the nodes represent participants and each connections represents a pair of participants that were proximate to each other during a time step. This construct is often called a dynamic graph – drawing from the field of Graph Theory – where the participants are nodes, and the connections are the edges. This graph can either be realized in practice as a sparse dynamic graph with edges appearing and disappearing, or a time series of symmetric matrices with binary constituents, with 0 indicating the absence and 1 indicating the presence of an undirected edge.

The second baseline manipulation scenario collapsed the dynamic Flunet graph down to a single static graph. In the static graph scenario, edge weights are replaced by the time averaged contact density between a specific pair of participants over the entire study. This is trivially constructed as the sum of all the dynamic symmetric matrices divided by the number of timesteps. This aggregation approach captures an extreme form of aggregation, in which no changes are made in contact graph structure over time. If contact dynamics had no significant impact on the results, then simulations using this graph should echo the fully dynamic case.

The second manipulation examined typical day approaches. As a method of temporally extrapolating graphs collected over shorter time horizons, both [9] and [15] employed a typical day [9] hypothesis, positing that the day(s) captured in the experiment were generally representative of longer-term contact patterns and could be duplicated through time to extend the dynamic graph until the simulated outbreak dissipated or reached a quasi-static endemic equilibrium. Once an interval was selected, the day-by-day duplication of that interval could be accomplished in one of two ways [9, 25]. In the first approach, the collected contacts are aggregated over that interval – in a manner similar to how the static graph is aggregated over the course of the study – with the aggregated contact network then being applied across the entire simulation time horizon. Another appoach simply replicates the fully dynamic graph for that interval throughout the study period. To evaluate the impact of the typical day hypothesis, the second manipulation consisted of a series of simulations with duplicated days using both the fully dynamic and aggregated typical day scenarios.

Reflecting the fact that many researchers lack recourse to high-fidelity empirical contact data for model integration, the third manipulation focused on fitted synthetic networks. Past contributions have generated random contact networks using small-world, scale free or other network topologies, often with contact probabilities represented as edge weights randomly drawn from other independent distributions. To analyze the impact of such an aggregate network representation on the spread of infection across the simulated population, the dynamic Flunet graph was reduced to distributions for edge weights and node degree and a set of new small-world contact graphs based on these distributions were created. To determine the impact of model fit on simulation results, several edge weight distributions were employed that provided increasingly accurate fit, at the cost of decreasing theoretical rigour.

Unweighted small-world connectivity graphs with 36 nodes were generated using the Watts and Strogatz model [33] in R [34, 35]. Average path length and cluster coefficient were used as measures of similarity with respect to the aggregate empirical Flunet graph. Five hundred possible parameterizations of small-world networks based on different connectivity ranges and rewiring probabilities were each generated with 10,000 realizations. For each such realization, the average of each of the similarity measures were calculated. The connectivity range and rewiring probability parameterization that yielded the highest average similarity measure was selected as the base for unweighted graph generation. The parameters were then used to generate final unweighted small-world networks. Figure 2 shows the resulting distributions of networks selected using this process.

Weights were assigned to the edges of the unweighted small-world network by drawing the value for each edge from a distribution. For each of the scenarios within the synthetic network manipulation, we examined the effects of using three different fitted distributions, as well as drawing from the normalized empirical histogram of pairwise Flunet aggregated contact durations (ACDs).

Fitting of distributions was performed using linear regression in MATLAB. For linear regression to perform properly, data underwent log-log (power law fitting) or log-linear (exponential fitting) transformation prior to performing the piecewise fit. For fitted distributions, the fitted curves were required to achieve a R² value exceeding 99%. Piecewise breakpoints were selected by iteratively changing the breakpoints, performing the regression, and manually selecting the point at which error began to increase sharply, but which still maintained a minimum R² of 99% value for all the piecewise components.

Figure 3 shows one of the curve fits implicitly and three explicitly. The red and blue lines correspond to the two and three piece fits for the data. The data points themselves form a discrete empirical distribution, and the three piece plus the additional three points with the highest aggregated contact duration form a combined functional-discrete distribution, which can be mathematically considered a 6 part piecewise function, where the discrete outliers are described as Dirac delta functions.

Infectious events were generated according to a Poisson process as described in [24]. When such event occurred in the FullD and DayD cases, the state of the contact graph was queried, and connected agents were infected based on the probability of infection for the pathogen. For aggregate representations a joint probability of the edge weight from the static graph and infection probability was used to determine infection probability. This is mathematically equivalent to randomly sampling the contact records from either the DayD or FullD contact records for the DayA and FullA cases.

The results presented here simultaneously underscore the high rate of return in terms of model reliability of investing in the collection of micro-contact data, but suggest that not all such investments confer equivalent value. It is widely recognized that model aggregation (for example, in the imposition of random mixing assumptions) tends to overestimate the spread of contagion. The results presented here suggest the natural extension of this understanding to the network context, where simulations employing purely static networks – such as might be produced by traditional network reconstruction techniques based on contact tracing, diarying or survey instruments – are also demonstrated to be biased towards overestimation of contagion. While this overestimation is pronounced in our experiments, it appears likely that the degree of this overestimation will depend heavily on the characteristics of the contagion process – particularly the speed of transmission relative to the speed of contact formation and dissolution [15]. The distortions associated with imposing a static network for some types of contagion suggest the importance of collecting information regarding network dynamics at a fine temporal granularity for such contexts. For very short-lived pathogens, a typical day approach may be sufficient in that all members of the network will be infected quickly or the infectious transmission will dwindle quickly. For long-lived pathogens such as tuberculosis, an aggregate model may be appropriate, as infection rates are low and latent periods are long. For the flu-like infection we modeled, our results indicate that typical days are too volatile and aggregate representations too permissive.

In addition to highlighting the importance of fine-scale data collection, these results also suggest a minimum efficient scale for such data collection. Such results are important in that many healthcare researchers have sought to defray the cost and logistic burden of electronically mediated micro-data collection by restricting the duration of studies. While such study designs sidestep many challenges associated with power consumption and device failures, they do rely heavily on the assumption that the network dynamics observed during the data collection interval is in some sense representative of the network dynamics over the long term. While such assumptions hold greater plausibility when applied in highly structured settings such as secondary schools [9], the results presented here highlight the elusiveness of identifying a typical day in some populations. Our results suggest that adopting a day as representative can be the source of major variability in simulation results, which can swamp the accuracy benefits conferred by using electronic micro-contact data. Diverse transmission modeling studies in the past several decades have revealed pronounced impacts of population heterogeneity on transmission dynamics, and have recognized the risks to model reliability of positing that the population is composed of homogenous typical persons. The results here suggest that – at least for some settings and sample populations – selecting a typical day can be fully as risky to model reliability as selecting a typical person – even when due diligence has been exercised to filter out manifestly unrepresentative days. While typical day assumptions appear to be hazardous in the context of our limited and highly clustered sample population, future data collection is required to determine whether similar risks extend from choosing typical days for studies with larger and more diverse populations.

Synthetic networks based on randomly generated networks and contact duration distributions are an appealing representation for extending results to larger populations. Our results highlight the need for care in the design of synthetic networks given that even high quality matches to the contact distribution using parametric distribution mixtures can lead to distortions of infection risks, primarily due to the impact of a few important outliers. When assessing the match fidelity between the synthetic results and results for the fully detailed data, it is easy to overlook an underestimation if it cancels out the overestimation due to an aggregated network whose outcomes misleadingly agree with historical or synthetic ground truth estimates. These biases could in turn distort the perceived trade-offs between interventions, or skew other important statistics regarding model dynamics. As is the case for random mixing models [4], even apparently calibrated models which match baseline data well can yield very misleading results when applied to counter-factual scenarios, such as interventions. We have little doubt that it is both important and feasible to build synthetic networks to capture transmission dynamics with high accuracy, but the current study suggests that achieving this goal will require moving beyond even well-calibrated models of network structure and contact duration. We particularly highlight the potential need to scrutinize the convenient assumption (applied here) of independence in the centrality of an agent and that agent’s contact duration distribution. While the close match in mean attack rates observed between the FullA and SWE scenarios suggests that this assumption may impose little distortion, further study is required to assess this assumption’s reliability.

The findings here have implications both for health research and epidemiological methodology. The results suggest that, while it is important to capture micro-contact data in many contexts, attempting to economize by reducing study duration may be penny wise and pound foolish. Use of micro-contact data appears to confer substantial benefits, but those benefits will be compromised – and potentially reversed – unless studies are of sufficient duration to capture day-to-day variability. While these results could be unique to smaller participant pools and flu-like infections, they likely apply to many important pathogens in high-risk populations of note such as institutional populations associated with long-term care facilities which are often not much larger than our participant pool.

We hope that the findings presented here will offer initial guidance in planning micro-contact studies, elevating the prospect that these studies will serve as a cost-effective means of enhancing health insight. Our results regarding the impact of network representation and parametric contact distribution approximations will also inform the data analysis methodologies used to analyze and generalize such data for transmission modeling. We have further described here a general methodology that can be replicated for future studies to assess the impact of aggregation on simulation results.

While our paper has made several important contributions to the field, as with all initial analyses, there are several shortcomings that should be addressed in future work. In fact, a major contribution of the paper is elucidating what the structure of future experiments should be to provide realistic samples of contact and disease dynamics for a given population. In particular, the role of sample size (N), versus study duration (T) is problematic.

Previous contributions [9, 15] have opted for larger N at the expense of T, and attempted to use various techniques to extend the duration of simulations beyond T artificially. We have demonstrated that the error introduced through replicating T is significant and perhaps larger than using a well-tuned aggregate model. However, the limited participant pool of our study prevented us from addressing the issue of sample size. This leads to the possibility that our results are a statistical fluke, and that larger sample size would begin to diminish the heterogeneity in sample days to the point where one day was like any other for a large enough group of people. We believe this outcome to be unlikely due to the small-world nature of dynamic contact networks [36]. Rather than having an averaging effect, adding additional N – particularly for heterogeneous study populations – is likely to capture more partially isolated sub-networks with their own particular contact dynamic patterns and life rhythms. A study employing a typical day strategy must then assume that the sampling period represents a typical day for new subnets. In fact, we conjecture that larger sample sizes will increase the sensitivity to T, as temporal variability is likely to rise with population heterogeneity.

The uncertainty over the apparent tradeoff between study size and duration is a strong motivation for further contact dynamic studies with substantially larger, heterogeneous populations observed over longer time periods. Larger longitudinal studies would permit the analysis of the relative sensitivity of different disease models to simplifications in sample size, population and study duration by permitting knock-out experiments where similar simulations could be performed over statistically significant subsets of the data, and compared against each other and the full data set.