Predicting Self-Initiated Preventive Behavior Against Epidemics with an Agent-Based Relative Agreement Model Journal of Artificial Societies and Social Simulation

Human self-initiated behavior against epidemics is recognized to have significant impacts on disease spread. A few epidemic models have incorporated self-initiated behavior, and most of them are based on a classic population-based approach, which assumes a homogeneous population and a perfect mixing pattern, thus failing to capture heterogeneity among individuals, such as their responsive behavior to diseases. This article proposes an agent-based model that combines epidemic simulation with a relative agreement model for individual self-initiated behavior. This model explicitly represents discrete individuals, their contact structure, and most importantly, their progressive decision making processes, thus characterizing individualized response to disease risks. The model simulation and sensitivity analysis show the existence of critical points (threshold values) in the model parameter space to control influenza epidemic including minimum values for the initially positive population size, the communication rate, and the attitude uncertainty. These threshold effects shed insights on preventive strategy design to deal with the current circumstances that new vaccines are often insufficient to combat emerging communicable diseases.


Introduction 1.1
Recent outbreaks of emerging infectious diseases, such as the new H1N1 flu in 2009 and Ebola in 2014, have attracted substantial interests in understanding human responsive behavior against diseases, from which practical intervention policies could be suggested (Funk et al. 2010;Perisic & Bauch 2009).Since effective vaccines require sufficient time to develop and manufacture, the best human response at an early stage of epidemics is the self-initiated preventive behavior, for example, wearing facemasks, performing hand hygiene, taking antiviral drugs, and avoiding close contact with people with symptoms (CDC 2008).Many studies have argued that these self-initiated behaviors can significantly change the course of epidemics (Poletti et al. 2009;Sahneh et al. 2012;Wu et al. 2012).Knowledge on the drivers and decision processes toward self-initiated behaviors is critical to the full understanding of disease spread and the early control of epidemics (Meloni et al. 2011).
1.2 Although many computational and mathematical models have been developed to understand disease transmission dynamics, few of them have, until recently, incorporated the human self-initiated behavior (Funk et al. 2014;Sahneh, et al. 2012).Previous models of human behavior in the context of epidemics have been concentrated on gauging the effectiveness of various institutionally enforced public health measures such as school closures (Funk et al. 2010).Challenges in modeling self-initiated behavior lie in how to depict individuals' perception of disease risks, their influence on each other, and their psychological process to make decisions, along with the feedback of behavior on disease spread (Fenichel et al. 2011;Funk et al. 2014;Sahneh et al. 2012).To date, the few models that have considered human self-initiated behavior are mostly designed by a classic populationbased approach, or in other words, depicted by mean-field differential equations (Fenichel et al. 2011;Funk et al. 2009;Kiss et al. 2010;Reluga 2010).These population-based models often assume a homogeneous population and a perfect mixing pattern, thus failing to capture the heterogeneity among individuals, such as personal traits, individualized contact structure, and individualized decision making processes (Bian & Liebner 2007;Koopman & Lynch 1999).There is a compelling need to develop a new class of disease-behavior models that explicitly differentiate individuals, their contacts, and decision making processes.The individual-based (also called agent-based) approach has been gradually recognized as a promising solution, but its application so far remains limited in integrating epidemics and human self-initiated behavior (Epstein et al. 2008;Sahneh et al. 2012).

1.3
This article proposes an agent-based model that combines epidemic simulation with a relative agreement model for human selfinitiated behavior.The relative agreement model has been developed in sociology and used to study dynamics of social opinions (Deffuant et al. 2005), but never been used in the context of epidemics.The inclusion of the relative agreement model allows a representation of step-by-step decision making processes for individuals, which has been widely recognized but not fully represented in existing models.With the proposed model, a sensitivity analysis is conducted to explore potential strategies that could promote preventive behaviors among the population and help control epidemics.The remainder of this article describes the model construction and its parameterization, and discusses simulation outcomes and research implications.

Methodology design
Conceptual model and assumptions

2.1
From a modeling perspective, both the transmission of diseases and the self-initiated adoption of preventive behavior can be conceptualized as diffusion processes.The disease spreads from individual to individual through their physical contacts, while the self-initiated preventive behaviors disperse through their social contacts, such as communication and observation (Amblard & Deffuant 2004;Glanz et al. 2002;Hägerstrand 1967).These two diffusion processes spread simultaneously and interact with one another.Once infected by diseases, individuals may adopt preventive behavior to protect others, and meanwhile motivate others to adopt as well.Conversely, the adoption of preventive behavior limits the transmission of disease by reducing the likelihood of infection.In this article, the disease-behavior diffusion model is established based on six assumptions: 1. Individuals in a population are linked together by a network structure.Individuals have contact with one another through the network.2. Every individual has an infection status, which can be susceptible, latent, infectious or recovered (Anderson & May 1992;Kermack & McKendrick 1927).The contact between individuals through the network triggers the change of infection status, which then follows the natural history of the disease.3. Every individual also holds an attitude toward the preventive behavior, which can be either positive, neutral, or negative.
The initial attitude is a mixed product of the individual's knowledge, experiences, perceived barriers and benefits (Glanz et al. 2002).The attitude varies over time due to interpersonal influences, and determines whether to adopt the preventive behavior or not.4.During an epidemic, individuals observe and communicate with each other through the network, which in turn influences their attitudes toward the adoption of preventive behavior.The communication and observation between individuals drives the diffusion of preventive behavior throughout the population.5.The adoption of preventive behaviors can reduce the transmissibility of influenza to different degrees, dependent on infection status of individuals.6.The manifestation of disease symptoms will add positive effects to an individual's attitude toward the adoption of preventive behavior.

2.2
To build a working simulation model, each of these six assumptions are formulated and then programmed as elaborated in following steps.
Step 1: Modeling the social network (Assumption 1) 2.3 A hypothetic network of 5,000 individuals is modeled for simulation.The network structure is generated with following algorithm.First, individuals are grouped into families with an average size of 2.5 members per family, consistent with that of a typical American family.A k-mean clustering method is used to group individuals based on their proximity.Second, same individuals are redistributed and then grouped into workplaces with an average size of 10 members per workplace (similar to a typical American workplace) with the k-mean clustering method based on their proximity.Individuals are linked to all members in their families and workplaces, resulting in approximately 12 contacts per individual on average.This network is highly clustered because individuals are grouped based on proximity and meanwhile has short average path length due to two groups of connections each individual has.This algorithm, thus, produces a network with a small-world structure (Watts & Strogatz 1998).

2.4
For simplicity, the model experiment assumes that both infectious diseases and human self-initiated behaviors diffuse through an identical network composed of people's close contacts, including family members and co-workers.Empirical studies on human behavior and social networks have reported that social distance between individuals have significant effects on their behavior change, such as lifestyle, substance use, and smoking (Bauman & Ennett 1994;Christakis & Fowler 2007, 2008).Individuals' attitude and behavior are more likely to be changed by people socially closer to them, such as their spouses and co-workers, among whom the transmission of diseases would also occur.Those who are socially distant to individuals may inform them, but are less likely to influence their attitude and behavior.Hence, simulating the diffusion of behavior and disease on the same network could be reasonable.
Step 2: Modeling the diffusion of influenza (Assumption 2) http://jasss.soc.surrey.ac.uk/18/4/6.html 2 25/01/2016 2.5 Influenza is taken as a typical example of infectious diseases, and its diffusion is simulated by changing the infection status of every individual over time.As shown in Figure 1, once having contact with an infectious individual, a susceptible individual may be infected by the influenza virus and enter into the latent status.The likelihood of infection is formulated as a probability referred to as the transmission rate (r), based on which the Monte-Carlo method can be applied to determine the transmission.The infection starts a latent period, during which influenza virus develops internally and cannot be transmitted.The end of latent period moves the individual into an infectious status and enables the individual to transmit influenza virus to other susceptible contacts.During the infectious period, the individual may develop symptoms of influenza or remain asymptomatic.At the end of the infectious period, the individual recovers, and develops immunity in the remaining period of an epidemic.The diffusion of influenza was programmed by tracking susceptible contacts of infectious individuals during every simulation day and emulating the transmission events through the Monte-Carlo method.
Figure 1.The infection events and status in the natural history of influenza Step 3: Modeling the diffusion of preventive behaviors (Assumption 3&4)

2.6
In behavioral science, the diffusion of a self-initiated behavior in a population can be modeled as gradual variations of individuals' attitudes toward adoption (Granovetter 1978;Hägerstrand 1967;Valente 1995).Following sub-steps depict the formulation of individuals' decision making process, their attitude toward adoption, and the exchange of individuals' attitudes.
Step 3.1: Decision process toward adoption 2.7 An individual's decision making is formulated as a staged process with a series of adoption status, following the concept of Precaution Adoption Process Model (PAPM) (Deffuant et al. 2005;Weinstein & Sandman 1992).As shown in Figure 2, the individual starts with an Unaware status of the preventive behavior.At some time point, this individual may be Informed about the preventive behavior through communication with network contacts.Once informed, this individual develops an initial attitude (positive, neutral or negative) toward the adoption.If the initial attitude is positive or neutral, this individual is moved to an Engaged status, and is willing to spread this information to others.Otherwise, this individual will disregard the information (staying at Informed).After engaging in discussion for a certain period (the discussion period t c ), this individual will move to a decision status (Deciding about Adoption).The decision making finally leads to one of three results according to the current attitude: Adopting if currently holding a positive attitude, or Not to Adopt if holding a negative attitude, or return to the Engaged stage for further discussion if holding a neutral attitude.At the Adopting stage, if an individual can keep the positive attitude for a certain period (the maintenance period t m ), then the adoption is Maintained; otherwise, the individual will give up the adoption during the maintenance period.

2.8
To formulate individuals' attitude toward adoption, the model introduces a concept of attitude segment from previous work by Deffuant et al. (2002).The attitude segment of individual i is defined as a mathematic range within a i ± u ai , where a i is the attitude value and u ai is the associated uncertainty.The attitude value (a i ) is a negative or positive real number, initially drawn from a normal distribution N(m, σ).The associated uncertainty (u ai ) is a positive real number that represents a mix between individual i's uncertainty, conviction, and openness to the attitudes of others.For simplicity, it is initialized with the same value for every individual.
2.9 Turning back to Figure 2, if a i − u ai > 0, individual i is positive toward adoption.If a i + u ai < 0, individual i has a negative attitude.
Otherwise (a i + u ai ≥ 0 and a i − u ai ≤ 0), individual i is neutral (Deffuant et al. 2002).Individuals' attitude values vary over time by discussions with social contacts, and as a result, the staged decision process is propelled.
Step 3.2: Discussion between individuals 2.10 Corresponding to Assumption 4, once an individual has contact with another, there is a probability γ that they will discuss the adoption issue.This probability is assumed constant in the following experiments for model simplicity, but can be formulated in a more sophisticated manner (discussed in Section 3).During the discussion, an individual's attitude will be influenced by others following rules of "relative agreement" (Deffuant et al. 2002).These rules assume that: 1) the attitude segment with low uncertainty (e.g., strong minded people) are more influential than that with high uncertainty; 2) the more the overlap between attitude segments, the more likely that the individuals may influence one another; 3) when the overlap between two attitude segments is too small, there is no influence.These "relative agreement rules" are formulated as follows: 2.11 Given that individual i has an attitude segment a i ± u ai , the influence of individual i on individual j, when they communicate, is determined by their attitude agreement, i.e., the overlap section (h ij ) of the two individuals' attitude segments minus the nonoverlap section (2u oi − h ij ) (Equation 1 and illustration in Figure 3).
where 2.12 Then, the relative agreement between i and j is calculated by the agreement divided by the total range of i's attitude segment If h ij > u i (the relative agreement > 0), then the modifications of a j and u aj by the interaction with i are updated as Equation 3: 2.13 By varying individuals' attitudes according to their social network and these formulated interaction rules, the model emulates the decision making process (Figure 2) for every individual.As a result, the variation of individuals' adoption status collectively leads to the diffusion of preventive behaviors over the population.
Step 4: Interactions between the two diffusion processes (Assumption 5&6) 2.14 Previous three steps have formulated the two diffusion processes independently, and the last step is to model their interactions.
For model simplicity, the interactions are formulated into two factors corresponding to Assumption 5 and 6, namely, the efficacy of preventive behaviors (p) and the change of attitude by infection (q).Specifically: 1.If Individual i is susceptible, the adoption of preventive behaviors can reduce the probability of being infected by p%; 2. If Individual i is in the latent status, the adoption can reduce the probability of becoming infectious by p%/2, following the same assumption from Longini et al. ( 2005); 3.If Individual i is already infectious, the adoption may reduce the probability of infecting others by p%/2, following the same assumption from Longini et al. (2005); 4. When Individual i starts to manifest symptoms, the attitude value will move toward positivity (Assumption 6).For the sake of simplicity, the attitude value a i is assumed to increase by q% of its uncertainty u ai , i.e., a i + q% ⋅ u ai .Further, if Individual j has discussed with i, the attitude value will increase by q%/2 of associated uncertainty u aj , i.e., a j + q% ⋅ u aj /2; 2.15 Upon the completion of the four steps described above, a disease-behavior (dual) diffusion model is established.This model is then parameterized for simulation and sensitivity analysis.
Model parameterization for simulation 2.16 Based on a pilot study and the literature, Table 1 specifies a default setting for model parameters during the simulation.The transmissibility of influenza r, its latent period t l and infectious period t i are set to be close to a moderate seasonal influenza (Chowell et al. 2007;Heymann 2004).Without much information about individuals' initial attitude, the mean value (m) and its standard deviation (σ) are set to 0 and 0.1, respectively, to generate a population with equal proportions of individuals holding positive, neutral and negative attitude before the simulation, hereinafter called initially positive, neutral and negative individuals, respectively.The uncertainty of individuals' attitude (u a ), the probability of discussion (γ) are adopted from previous settings of Deffuant et al. 2005.The efficacy of preventive behavior on infection (p) follows findings on the use of antiviral drugs (Longini et al. 2004).Change of attitude by infection (q) 50% 2.17 All individuals are initially assigned susceptible and unaware of the adoption.The proportions of individuals with positive, neutral, and negative attitude are equally split in the population.At the first day of simulation, one individual is randomly selected to be infectious to initialize the diffusion of influenza.Meanwhile, another individual is randomly selected to be informed to initialize the diffusion of preventive behavior.The simulation of both diffusion processes takes a daily time step over 120 days.
Scenario and sensitivity analysis 2.18 For model outcomes, the diffusion of influenza is described by an epidemic curve that depicts the percent of infection in the population through the course of the epidemic.The diffusion of preventive behaviors is described by an adoption curve that shows the percent of adoption in the population during the diffusion.In addition, the evolution of individuals' attitudes is also displayed by plotting respective proportions of positive, neutral and negative individuals in the population during the diffusion.
2.19 This research first compares how one of the two processes diffuses with and without the other.In other words, the dual-diffusion scenario (from the default parameter settings) is compared to an adoption-only scenario (no disease epidemic: transmission rate r = 0), and to an influenza-only scenario (no behavior diffusion: the probability of communication γ = 0).For each of the three scenarios, the average value of 100 realizations is used to construct the epidemic curve, adoption curve, and the attitude evolution curves.
2.20 Furthermore, a sensitivity analysis is conducted to explore effective strategies to control influenza epidemics.In particular, this research is interested in the minimum number of initially positive individuals required to control an epidemic by self-initiated preventive behavior.The proportion of initially positive individuals in the population, denoted as β 0 , is varied from 0 to 100% by increasing the mean of initial attitude distribution (m), and represents the effects of previous information campaigns or memory from last outbreaks in the population.The parameter β 0 is paired with three other parameters respectively in the sensitivity analysis, including: 1) the transmission rate of influenza r, varying from 0 to 0.2.When r ranges from 0.08-0.1, it represents a seasonal influenza scenario where the basic reproductive number R0 is between 1.2-1.3(Chowell et al. 2007).When r varies from 0.1-0.2, it represents a pandemic influenza scenario in which R0 increases to 1.6-1.7 (Mills et al. 2004;Yang et al. 2009); 2) the probability of discussion γ (from 0 to 1).When γ increases to very high values, like 90% or 100%, the modeled information diffusion could somewhat match the effects of mass media; and 3) the uncertainty of individuals' attitude (u a ), ranging from 0.01 to 0.2 with an increment of 0.01.Low values represent conservative populations, while high values represent progressive populations.Each pair of parameters is given alternative values while holding all of other parameters as default in Table 1.For any pair of parameter values, 100 model realizations are performed to estimate the percent of infection in the population.The epidemic is assumed to be controlled if the resultant percent of infection is less than 10% of the population.

Simulation results and discussion
Scenario analysis 3.1 Figure 4 compares the diffusion of self-initiated preventive behavior with and without an influenza epidemic.For the adoptiononly scenario (transmission rate r = 0), the adoption curve starts with an abrupt rise during the first 20 days, and then increases slowly till the last day (the red dashed line in Figure 4a).Approximately 50% of the population would adopt the preventive behavior.The early abrupt rise can be attributed to the initially positive individuals, who are informed through network communication and quickly adopt the preventive behavior.The later slow increase can be explained as a number of initially neutral individuals changing their attitudes to positive after discussions (the red curve with circles in Figure 4b), and finally adopting the preventive behaviors.Compared to the adoption-only scenario, adding the diffusion of influenza causes a marked increase in adopters (the blue solid line in Figure 4a).The total number of adopters has the same rise as the adoption-only scenario before Day 20, but remains a fast increase from Day 20 to 60.This later fast increase is because a number of negative individuals change to positive attitudes due to infection (the blue curve with diamonds in Figure 4b).More individuals, therefore, finally adopt the preventive behaviors and the percent of adoption reaches a higher level (about 70% of the population).

3.2
Figure 5 illustrates the difference between the spread of influenza with and without the diffusion of preventive behaviors.For the influenza-only scenario (the probability of discussion γ = 0), influenza can infect nearly entire population within 30 days (the red dash line in Figure 5a).Due to the lack of discussion, individuals are unaware of the preventive behavior, and no one starts the decision making process toward adoption (the flat red curves in Figure 5b).Once the diffusion of preventive behaviors is added ( γ = 0.25), the transmission of influenza can be significantly slowed with only half of the population being infected (the blue solid line in Figure 5a).Through discussion, an increasing number of individuals are gradually aware of the adoption issue, change their attitude to positive, and decide to adopt preventive behaviors, which reduce the transmissibility of influenza (See the blue curves in Figure 5b).

3.4
Another parameter that impacts epidemics is the probability of discussion between individuals (γ). Figure 6b shows that a 10% chance of discussion between individuals (γ = 0.1) is the minimum requirement to control an influenza epidemic.Given that all individuals are initially unaware of the preventive behavior, they need to be gradually informed to start their adoption processes.With a 10% chance of discussion, β 0 must be greater than 90% (almost all individuals) to control the epidemics.If the probability of discussion can be raised to 0.5, the β 0 only needs to reach 50% for a successful control.A higher probability of discussion means that individuals can be informed faster and make earlier decision toward adoption.The increasing number of early adopters could timely limit the transmission of influenza and thus reduces the disease incidence.This result suggests the importance of encouraging individuals to discuss the adoption issue.A mass media bombardment strategy could be useful to attract public attention onto the adoption issue through television, newspapers, and the internet.

3.5
Figure 6c indicates the effect of initial uncertainty of attitude (u a ) on the total infection.The general pattern shows that a conservative population with low attitude uncertainty (e.g., u a = 0.02) requires more initially positive individuals to control the epidemics.When the uncertainty elevated to 0.2, only 5% of the population is needed to initially hold positive attitude for a successfully disease control.There are several two possible reasons.First, the greater uncertainty means wider individual attitude segments (Figure 3), and thus their attitude segments are more likely to overlap, which in turn triggers attitude influence between them.Second, the greater uncertainty associated with individuals' attitude also means their attitude would have a big turn toward adoption, when they were infected or when they were aware of infection risks from surrounding people.More adopters would occur in the early phase of disease transmission and prevent the wide spread of the disease.

3.6
Still in its infancy, the presented model is so far a crude approximation to the reality with following limitations.First, the infectious diseases and human self-initiated behaviors can diffuse through different networks, while this research assumed an identical network for model simplicity.This assumption can be relaxed by linking individuals with two networks: a contact network for disease transmission and a communication network for behavior diffusion.The two diffusion processes can be simulated on the two networks respectively.Second, the information diffusion through the mass media and Internet are not considered, which may play an important role in the diffusion of preventive behaviors.An easy solution is to add into the model a broadcasting rate (as a parameter), i.e., the chance of an individual being informed through the mass-media or internet.With such a rate, a number of individuals can be randomly selected to be informed in every time step.Third, the communication rate between individuals is assumed constant for model simplicity.To be more realistic, this rate can vary with the size of epidemic, i.e., the more infections in the population the more likely individuals will discuss.For example, the discussion rate can be formulated as a function of the total number of infections at a time step: where N I (t) is the number of infections at time step t and β is the coupling parameter that relates discussion to infection.The compensation for these limitations can improve the model accuracy, but also increase the complexity.It is argued that the goal of modeling is not to look for exactly what may happen, but rather for differences in outcomes between different scenarios (McKenzie 2004).In this sense, the proposed model framework could be a valuable prototype to guide the design of more realistic working models.

4.2
Using the proposed model, this research also explored critical points (threshold values) in the model parameter space to control influenza epidemic solely by self-initiated preventive behavior, for example, minimum values for the initially positive population size, the communication rate, and the attitude uncertainty.For disease control, it is not practical to have all individuals holding positive attitude toward adoption at the beginning of an epidemic.Many individuals may be reluctant to adopt preventive behaviors because of their knowledge, prior experience, incomes, occupations, and other factors (Nichol & Zimmerman 2001;Toy et al. 2005).The sensitivity analysis estimates that at least 60% of the population with an initial positive attitude is necessary to successfully control a moderately transmissible influenza.This threshold would elevate as the transmissibility of influenza rises, but can also be reduced by encouraging discussions between individuals or by increasing the mind-openness of individuals' toward preventive behavior.These threshold effects shed insights on preventive strategy design to deal with the current circumstances that new vaccines are often insufficient to combat emerging communicable diseases.

Figure 2 .
Figure 2. The staged process of an individual to adopt the preventive behaviors.The squares and diamonds represent adoption stages.The comments on arrows describe the conditions for transition.

Figure 3 .
Figure 3.The influence of individual i on individual j based on the "relative agreement rules" proposed byDeffuant et al. ( 2002)

Figure 4 .
Figure 4.The comparison of the dual-diffusion scenario (blue) to the adoption-only scenario (red).(a) The adoption curves depict the percent of adoption in the population over time.The solid line represents the dual-diffusion scenario (transmission rate r = 0.1), while the dash line represents the adoption-only scenario with r = 0. (b) The evolution of the proportion of positive (square), neutral (circle) and negative (diamond) individuals in the population over time.

Figure 5 .Figure 6 .
Figure 5.The comparison of the dual-diffusion scenario (blue) to the influenza-only scenario (Red).The epidemic curves depict the percent of infection in the population over time.The solid line represents the dual-diffusion scenario (discussion probability γ = 0.25), while the dash line represents the influenza-only scenario with γ = 0. (b) The evolution of the of positive (square), neutral (circle) and negative (diamond) individuals in the population over time.
proposed a new prototype model to simultaneously simulate the disease transmission and self-initiated preventive behaviors against the disease.Based on the relative-agreement rules, this model explicitly represents discrete individuals, their contact structure, and most importantly, their progressive decision making processes, resulting in individualized responses to disease risks.This model allows disease transmission to alter individuals' decision making toward preventive behavior, which in turn affects the disease transmission.The comparison study shows that any model that fails to account for this interaction mechanism could not accurately portray a realistic epidemic.

Table 1 :
The default disease-behavior scenario for model simulation