key: cord-0907413-0rxofyzq
authors: McGail, Alec M.; Feld, Scott L.; Schneider, John A.
title: You are only as safe as your riskiest contact: Effective COVID-19 vaccine distribution using local network information
date: 2022-04-05
journal: Prev Med Rep
DOI: 10.1016/j.pmedr.2022.101787
sha: 81e09c8b77989c74b6dc11a6e989a1250a5acb91
doc_id: 907413
cord_uid: 0rxofyzq

When vaccines are limited, prior research has suggested it is most protective to distribute vaccines to the most central individuals – those who are most likely to spread the disease. But surveying the population’s social network is a costly and time-consuming endeavour, often not completed before vaccination must begin. This paper validates a local targeting method for distributing vaccines. That is, ask randomly chosen individuals to nominate for vaccination the person they are in contact with who has the most disease-spreading contacts. Even better, ask this person to nominate the next person for vaccination, and so on, what we call chained popularity nomination. To validate this approach, we simulate the spread of COVID-19 along empirical contact networks collected in two high schools, in the United States and France, pre-COVID. These weighted networks are built by recording whenever students are in close spatial proximity and facing one another. We show here that popularity nomination performs significantly better than random vaccination, and on par with strategies which assume a full survey of the population. These results are robust over a range of realistic disease-spread parameters, as well as a larger synthetic contact network of 3000 individuals. 192 words.

 Using simulation to evaluate nomination of most popular contacts for vaccination.

 Simulating spread of COVID-19 across two contact networks among high-schoolers.

 Targeting in this way can reduce spread to the suscptible population by 25% or more.

 Results are robust in a synthetic network replicating spread in a small town.

 Results are robust across a wide range of infectiousness, and mistaken nomination.

Although viable vaccines for COVID-19 are now widely in use, accessibility to the vaccine is progressing slowly through the world. As of this writing (April 20, 2021), 40% of the U.S.

population have received at least one dose of the vaccine, almost 50% of the U.K. and 62% of Israelis. However, only 6.5% of the world's population has had at least one dose and less than 1% in Africa. In light of this stark reality, and in response to the possible need for distributing new vaccines to fight new strains of this virus or others, we should find ways to improve the effectiveness of the limited vaccines a community, school, or nation has available.

The current dominant method for vaccine prioritisation is to first vaccinate those most vulnerable, then front-line workers most likely to be exposed to the disease, working eventually towards herd immunity at around 70% of the population vaccinated. This overall strategy for targeting has been recently shown through simulation to be optimal in avoiding hospitalization and death (Jahn et al. 2021) . To reduce total infections and deaths we can employ a more nuanced targeting strategy which aims at those who are most likely to spread the disease. This methodology does not necessarily supplant the prioritisation just mentioned but may be used to complement it. For example, within a nursing home, local nomination strategies could be used to choose who to vaccinate first and may yield important transmitters as opposed to an age or comorbidity approach. In addition, after these highest risk groups are vaccinated a targeted approach could be used to vaccinate the population at large.

Prior work has shown that individuals most central in the disease-spread network are the most important targets for vaccines (Jia, Shi, Yang, & Fu 2020; Pastor-Satorras & Vespignani 2002; Dezső & Barabási 2002) , and even in the specific context of COVID-19 (Jadidi 2020) . Nunner, van de Rijt, & Buskens (2022) recently demonstrated that targeting occupational categories as a proxy for connectedness in a contact network is significantly effective. Some work has pointed to Vaccination Nomination 4 the importance of decentralised methods for nominating those who should be vaccinated (Holme 2004; Ke & Yi 2006; Cohen, Havlin & ben-Avraham 2003; Lee, Rocha, Liljeros & Holme 2012; Hébert-Dufresne, Allard, Young & Dubé 2013; Taghavian, Salehi & Teimouri 2017) . One compelling method chooses a random individual and nominates for vaccination a random of their contacts. This "random nomination strategy" relies on the Friendship Paradox, the fact that these random contacts will be more connected than random individuals are (Feld 1991) . This strategy has been shown to be more effective than random vaccination (Wang et al. 2016; Manzo & van de Rijt 2020 for COVID in particular). A related strategy asks individuals to recall who they interacted with most recently, relying on the recurring nature of interactions. Lee, Rocha, Liljeros & Holme (2012) have shown that this method outperforms random nomination.

In this paper we suggest the nomination of most popular contacts (NP) as a practical and effective method. In this method administrators choose an individual at random and ask them to nominate a contact of theirs who has disease-spreading contact with the most people. Although similar strategies have been proposed and evaluated in the physics literature (Holme 2004; Ke & Yi 2006; Wang et al. 2016) , they have been ignored by epidemiologists and policymakers. One possible reason, which leads to the central contribution of this paper, is that the models they use, and the networks on which they evaluate these strategies, are simplifications at best. Holme (2004) finds that chained popularity nomination is the most effective local targeting strategy of those he analysed, but he does not test this using realistic contact networks, nor perturbing the model of disease spread or (of course) calibrating this model to COVID-19 in particular. This paper evaluates the strategy using more realistic simulations, which simulate the spread of COVID-19 on contact networks measured from physical interactions in a real-world setting, and presents results in a digestible form, in the hopes of spurring renewed policy interest in decentralised targeting strategies for vaccines. Our analyses show a marked robustness of the effectiveness of nomination of most popular contacts over a wide range of disease spread models over three contact networks, and with some loosening of the assumption that individuals can accurately nominate their most contacted contact. The DATA and METHODS sections describe the contact networks, targeting strategies, and simulation methodologies. We conclude with 

Epidemiological models often assume homogeneous mixing, where an infected individual has equal probability of infecting anyone else in the population. And when epidemiologists employ a networked approach, they often use an unweighted network, where contact either exists or does not between each pair of interactants, with no variation. Both these assumptions are patently false (Bioglio et al. 2016) , and variation in contact proves instrumental to accurate modelling of disease spread (Manzo & van de Rijt 2020; Stehlé et al. 2011) . In this paper we include the more practical and differentiated structure of the contact network of two high schools as they operated pre-COVID. We focus on high-school students as they are the most likely to not follow public health or other authority recommendations around social distancing and mask wearing.

Furthermore, in the United States the rate of infection is twice as high in those aged 12-17 compared to 5-11-year-olds (Leeb et al. 2020) .

In these two studies (HS-1, Mastrandrea, Fournet & Barrat 2015, and HS-2, Salathé et al. 2010 ), students wore battery-powered Bluetooth transmitters / receivers which exchange packets of information whenever students are in close physical proximity with each other. The signals do not travel as far through solid objects, including students' bodies. They most reliably communicate when students are face-to-face and within a distance of approximately 6 feet (in HS-1) and approximately 3 feet (in HS-2). The sensors have been designed to reliably determine colocation in each 20 second interval, resulting in an extremely high-resolution contact network.

These temporal contact networks have been extensively and independently evaluated for the purpose of studying respiratory diseases whose main vector of transmission is across such short distances. As such they are an ideal source for plausible simulations of the spread of COVID-19.

The physical and social structures of these two schools yield social networks that are similar in some ways, and different in others. The French lycée (HS-1) is split into three grades, each of which are split into three classes. Students mostly interact within their own class, but in the hallways, during lunch, and before and after school, we see many more cross-class contacts, especially within grade. Some students act as a bridge between classes and there is strong age homophily. The American high school (HS-2) has many of the same characteristics, but split across four grade levels, and with much more between-class interactions. This high school is also more than double the size. These structural elements, amongst others which I have not noted, are embedded as features of the HS-1 and HS-2 networks upon which we simulate the spread of COVID-19 in this paper. In all, the differences between these two schools and the two sensor methodologies offer strong robustness checks to the results we present.

The third contact network we use in this paper represents a small town of 3000 residents and was generated procedurally using SEIRS+ (McGee et al. 2021) . The algorithm (an adaptation of FARZ) reproduces the clustering and degree distribution observed on average in the United

States. This network also reproduces the age distribution of United States citizens, and the differential probability of contact between those of different age groups. Within each age group a community structure is generated, representing primary schools, secondary schools, workplaces, and elderly community structures. Average degree by age group was matched to an empirical measurement of contact networks in the United States (Mossong et al. 2008) . Individuals from different age groups are then grouped into households, matching the distribution of household sizes and the household age demographics of the United States. Vaccination Nomination 7 

The contact network and a realistic model of disease spread together constitute a complete understanding of the spread of a disease. But both are heterogeneous across local contexts, and to some extent unmeasurable. How exactly COVID-19 spreads depends on many factors, and our understanding of these factors is still incomplete. For example, the probability that an infected individual spreads the disease to another person in one day is very hard to measure and depends on a variety of factors. Masking, ventilation, the physical arrangement of a space, these all contribute to reducing the probability of contagion. These propensities are also affected by individual attributes such as age, and a social context's relation to the outside environment. Different vaccination strategies have differential relative effectiveness depending on the number of vaccines administered, and who is available for vaccination at all, which in turn depends on institutional, logistic, and individual psychological factors. Individuals differ in how COVID-19 affects them and their subsequent infectiousness, and the scope of this variability and its relationship with network position are not entirely known. In response to the existing knowledge and extant uncertainty of how COVID-19 spreads, we test a range of modifications to the central nomination strategy, in addition to perturbing average infectiousness, the variation in infectiousness across the population, the percent vaccinated, the number of initial infections, all in addition to testing in the context of the three contact networks described above. The following subsections detail the vaccine targeting strategies and simulation methodology we consider.

R -Random Each person is equally likely to be vaccinated. D -Degree -First determine the number of contacts each person has (their so-called degree).

Choose the N people who have the highest degree.

NR -Random nomination -1) Choose a person at random. 2) From their unvaccinated contacts choose a person to be vaccinated at random. 3) Repeat steps 1 and 2 until N individuals are vaccinated. As shown by Feld (1991) , randomly nominated individuals are on average more Vaccination Nomination 11 central than randomly selected individuals (R), and this method is a common benchmark for decentralized strategies. 

We use a stochastic SEIR (susceptible, exposed, infected, removed) model for disease spread, following the model contributed and compiled by McGee et al. (2021) . In this model, a susceptible person is exposed (E) to COVID-19 by one of their infected (I) contacts. The exposure occurs at a randomly chosen time, drawn from an exponential distribution with mean depending on the infectiousness of the infected person, the susceptibility of the exposed person, and the amount of in-person contact they share. Once exposed, an individual will after some time move into the infected (I) state where they can expose others, and some time later will move into the recovered (R) state, no longer infectious or susceptible to infection. The latent period between the exposed and infected states, and the recovery time, are also drawn from exponential distributions, with expected means which are somewhat different for each person. In the case of the synthetic network, these parameters are tuned to match what we know of their agedependence in the case of COVID-19. For full details on the distributions we used for these parameters, see the online supplement.

The daily probability of spread for COVID-19 has been measured to be anywhere from less than 0.05 to 0.2 on average for those who come into contact in that day (Mwalili et al. 2020; Jiang et al. 2020; Carcione et al. 2020 ), corresponding to a wide range of R 0 anywhere from near 1.0 to upwards of 5. This does not only reflect an uncertainty of the "true" transmissibility of COVID.

In addition, it reflects the heterogeneity of this average transmissibility across different contexts.

A typical estimate in the literature is 2.5 (e.g. as used in Manzo and van der Rijt), but in this paper we vary R 0 along this entire range, assessing to what extent differences in average transmissibility may change the overall results. We generate R 0 for each individual based on a gamma distribution. For the central models in this paper we assume a relatively low coefficient of variation CV[R 0 ] = 0.2, which describes the variation in personal R 0 across the population, although the supplement checks robustness with respect to this parameter. Note also that we assume that those aged 0-19 are half as susceptible to infection as those aged 20+.

To begin the simulation, according to one of the targeting methods detailed in the previous section, we assume that some group of individuals had been vaccinated at the start of the Vaccination Nomination 13 simulation, and are not at all susceptible to the infection. They are fully removed from the disease-spread network. Then we randomly infect some number of unvaccinated individuals. We then run the simulation for 100 days. The measure we use for the effectiveness of any given strategy is the total number of individuals who entered the exposed (E) state at any time in the 100 days. For each set of parameter values, we run 500 independent simulations, choosing again who to vaccinate and who to infect, in order to estimate accurately the properties of the distribution of total infections under these scenarios. We report uncertainty in our estimate of the true mean infected by the standard error. For uncertainty in the percent improvement over not vaccinating or random vaccination, we bootstrap from the sampled simulation results. We collect 10,000 samples with replacement, with sample size 500, and calculate the quantiles of the relevant ratios corresponding to a 95% CI.

Because HS-1 and HS-2 are empirically gathered contact datasets, we assume that the measured contact is the only contact on which disease may spread. However, the synthetic network generates strong contact ties according to what we know of institutional and family ties. This would leave out spread which occurs in public and interstitial spaces. And so for this network only we assume a propensity of random spread to any other node in the network, in addition to the propensity of spread along network ties. This is constant throughout, set at 20% of an individual's total disease-spread contact.

Average individuals' infectiousness 1, 2.5, 4

The average of individual infectiousness, which is the expected number of additional infections in a completely connected population, given a seed infection.

Coefficient 

Network on which disease spreads.

HS-1 and HS-2 are described in Data. The Synthetic network is described in Robustness Checks.

First we present the relative effectiveness of these vaccination strategies under one scenario. That show that moderate error in nominating highly connected contacts will for the most part not derail the effectiveness of this strategy. We might then ask how much error would it take.

Once individuals are nominating from their top 10 or more friends, we found little difference between NP(N) and NR in all circumstances. Likewise for random Gaussian error, once the standard deviation of this error is greater than 30 contacts, NP(ε) is indistinguishable from NR (see Figure S6 ). We would expect these strategies to monotonically approach NR (as distinct from R), and attribute the minor deviations from this pattern to the stochasticity of simulations.

We also varied the number of individuals initially infected, as well as the variation in individuals' R0, and found no substantive differences ( Figures S4 and S5 ). In addition, the number of individuals initially infected, a proxy for the extent of outside infections introduced, does not affect the order of strategies, but shows that the more dire the threat from outside, the more effective are these targeted strategies relative to random vaccination, at least in the parameter ranges we consider here ( Figure S5 ). In addition, we varied CV[R0], as a proxy for wider variability in infectiousness independent of network position. With dramatic increase in this variability, there was not an appreciable increase in effectiveness of the targeting strategies relative to random vaccination ( Figure S7 ).

For a deeper look at all the realised runs of the simulation we use in this paper as well as the remainder of these robustness checks, or to extend these results to new empirical settings or differently specified models of disease spread or vaccination nomination, see the accompanying repository at https://www.github.com/amcgail/episim/. We also verified here that the ordering of daily contact in HS-1 and HS-2 were irrelevant for the outcomes reported here through explicitly simulating these dynamics.

Vaccination Nomination 18 

This paper tested the relative effectiveness of a strategy for choosing how to allocate limited vaccines to maximise their effectiveness. That is, to choose a random person and have them nominate from their contacts the individual with the most contacts. This method is aimed at logistical feasibility, allowing the administrator to survey individuals as they receive vaccinations, needing just one survey response to administer the first vaccine. This strategy performs significantly better than randomly distributing vaccines, but also performs better than choosing random contacts of individuals, a classic decentralized targeting strategy, and often even better than simply targeting those with the highest degree, assuming we have a full survey of the population of interest. Concretely, the majority of reasonable parameter combinations showed better than a 20% reduction in infections amongst the susceptible compared to randomly vaccinating, on average across 500 simulations. One may object that individuals' reports of the interaction profiles of their contacts may prove more erroneous than self-reports. Yet, through the inclusion of the NP(ε) and NP(N) strategies, we were able to show that moderate error in nomination does not cripple the effectiveness of the method.

This simulation is necessarily limited, not including all features of realistic COVID-19 spread.

Future work can explore the inclusion of various competing strains of COVID-19, variation in the initial composition of individuals in terms of having or having had COVID-19, heterogeneity in the effectiveness of vaccination which is not total and wanes over time, the temporality of contact networks, and many other complexities. In addition, a separate paper could address the variation of effectiveness under different measures, as compared to total number infected with COVID-19 after 100 days as analysed in this paper. It is also possible that most central individuals may not be optimal targets in practice. The most central may differ from the general population in various other ways, which may correlate with their unwillingness to be vaccinated (as in the model of Wells, Klein, & Bauch 2013) , or their probability of already being immune to the disease. We do not in this paper address the concrete issues of implementation such as the right survey strategy to approximate this theoretical model, and leave this to future work. Such work could additionally investigate how the targeting strategy proposed here could make use of the personal relationships between interviewee and target, to mobilize interpersonal trust and Vaccination Nomination 22 communication to convince an individual to get vaccinated, along the same lines of respondentdriven sampling (Heckathorn 1997) .

N/A. No students were infected in this simulation study.

In order to replicate what we present here, and to encourage extension of our methodology and results, all code and data used in this paper are available at https://www.github.com/amcgail/episim/.

Recalibrating disease parameters for increasing realism in modeling epidemics in closed settings

A Simulation of a COVID-19 Epidemic Based on a Deterministic SEIR Model. Frontiers in Public Health

Efficient immunization strategies for computer networks and populations

Halting viruses in scale-free networks

Why Your Friends Have More Friends Than You Do

Global efficiency of local immunization on complex networks

Respondent-Driven Sampling: A New Approach to the Study of Hidden Populations

Efficient local strategies for vaccination and network attack

A Conjectural Experiment to Observe the Effect of Conditional locked-down in an Epidemic

Targeted covid-19 vaccination (Tav-covid) considering limited vaccination capacities-an agent-based modeling evaluation

Immunization strategies in directed networks

Mathematical models for devising the optimal SARS-CoV-2 strategy for eradication in China, South Korea, and Italy

Immunization for scale-free networks by random walker

Exploiting temporal network structures of human interaction to effectively immunize populations

COVID-19 Trends Among School-Aged Children -United States

Halting SARS-CoV-2 by Targeting High-Contact Individuals

Contact patterns in a high school: A comparison between data collected using wearable sensors, contact diaries and friendship surveys

Predicting and controlling infectious disease epidemics using temporal networks

Model-driven mitigation measures for reopening schools during the COVID-19 pandemic

Targeted Vaccination for COVID-19 Using MobileCommunication Networks

Social Contacts and Mixing Patterns Relevant to the Spread of Infectious Diseases

SEIR model for COVID-19 dynamics incorporating the environment and social distancing

Vaccination Nomination 25

Prioritizing high-contact occupations raises effectiveness of vaccination campaigns

Immunization of complex networks

Immunization strategies in networks with missing data. ArXiv

A high-resolution human contact network for infectious disease transmission

Simulation of an SEIR infectious disease model on the dynamic contact network of conference attendees

A local immunization strategy for networks with overlapping community structure

Statistical physics of vaccination

Policy Resistance Undermines Superspreader Vaccination Strategies for Influenza