Self-Isolation and Testing Behaviour During the COVID-19 Pandemic: An Agent-Based Model

Following its appearance in late 2019, the SARS-CoV-2 virus or COVID-19 has become the most significant global pandemic since the 1918 influenza pandemic. By 15 March 2022 the World Health Organization recorded over 456 million confirmed COVID-19 cases worldwide, and over 6.04 million deaths (https://covid19.who.int/). Excess mortality estimates place the global death toll between January 2020 and December 2021 at 18.2 million (Wang et al., 2022). The scale and severity of this global crisis, and the subsequent need for severe public health restrictions to be implemented worldwide, has generated significant interest in using agent-based models (ABMs) to simulate the spread of the pandemic and its effects.

ABMs enable modellers to examine the impact of individual behaviour on population-level outcomes, and in the context of the global pandemic, this approach has been used to examine how individual behaviour influences the capacity for various interventions to reduce the spread of COVID-19. For example, Rajabi et al. (2021) have investigated the impact of social distancing restrictions and travel restrictions on COVID-19 spread; Koehler et al. (2021) have developed an ABM designed to assist policy-makers in making decisions on lifting pandemic restrictions; and Almagor and Picascia (2020) have investigated the efficacy of contact-tracing apps in containing the spread of the virus. While the use of ABMs has proliferated during the COVID-19 crisis, the question of how ABMs can be used most effectively remains a focus of significant debate: See, for example, Dignum (2021), the subsequent review (Chattoe-Brown, 2021), and numerous comments at https://rofasss.org/tag/jasss-covid19-thread/.

Building simulation models of the pandemic and its effects requires an understanding of human behavioural responses to the spread of the virus in their community. Born et al.’s (2021) empirical study indicates that people do respond to the risk imposed by a pandemic by reducing their mobility. Using data from Google COVID-19 Community Mobility Reports (https://www.google.com/covid19/mobility/) for Sweden (a country where compulsory lockdowns have not been imposed), this work shows that the mobility of the population decreased considerably during the first wave of the pandemic, although significantly less than in countries where lockdown policies were implemented. Other empirical studies shed some light on the factors affecting self-isolation and adherence to lockdown policies. Smith et al. (2020), in their study of factors associated with self-isolation in the United Kingdom, found evidence that self-isolation increases with increased worry about COVID-19 and increased perceived likelihood of catching the virus. Moreover, they found that self-isolation is affected by the perceived social norm and by the help received from someone outside the household. Pullano et al. (2020) found that, in France the reduced mobility after the lockdown was positively associated with the number of hospitalizations (which can be thought of as a proxy for the risk posed by the virus) and the socioeconomic conditions. Similarly, A. L. Wright et al. (2020), investigating the level of compliance with shelter-in-place policies in the United States, found that compliance with the policy increases with local income.

However, despite the widespread use of agent-based modelling, relatively few COVID-19 models have examined the impact of behavioural factors on how individuals respond to the risks posed by the pandemic. Instead, much of the extant modelling work is highly abstracted, and uses very simplified behavioural models. For example, Wilder et al. (2020) proposed an agent-based model in which out-of-household contacts are based simply on a country-specific contact matrix containing the mean number of daily contacts agents of an age group have with agents from each of the other age groups. Similarly, Silva et al. (2020) presented a model that allows agents to have differential exposure to risk according to their economic status, but does not include a facility for agents to modify their behaviour in accordance with their own perception of risk.

One of the first works accounting for the adaption of the behaviour of individuals during a pandemic was Epstein et al. (2008). Here, agents could be independently infected by the virus and by the fear of the virus. If infected by fear, agents self-isolate at a certain rate, until they recover from fear, ending their self-isolation. Shastry et al. (2022) have recently made progress in this area of COVID-19 modelling research by including a model of risk tolerance in an agent-based simulation of COVID-19 spread. In this model, agents have individual levels of risk tolerance, which is a decreasing function of age, and is adjusted towards the mean risk tolerance of groups with which they interact. The model focused on a U.S. policy context in which shelter-in-place orders were issued but were not legally enforceable; this meant that agents were able to freely defy the restrictions, if their risk tolerance allowed.

In this article, we present an agent-based model of COVID-19 spread that builds upon these foundations and includes a nuanced model of self-isolation behaviour. This behavioural module enables a feedback process, in which pandemic dynamics influence agent behaviour, which then influences pandemic dynamics in turn. The central contribution of this model is a behavioural module that allows households to reduce their social interaction as a response to the perceived risks posed by the virus; these risk perceptions are influenced by public information provided about the virus, the prevalence of COVID-19 infections within an agent’s network of neighbours, and the tendency for their neighbours to self-isolate.

While the current model is a proof-of-concept, we propose that simulations including these behavioural elements may help policy-makers to design more effective interventions during future global health crises, including future waves of COVID-19 variants or other novel pathogens. Non-pharmaceutical interventions require compliance from the population in order to be maximally effective, and a deeper understanding of how public information and social context influence compliance may enable the development of more effective messaging.

The simulation unfolds in two stages: In the first stage a demographic process proceeds in one-year steps, from year 1860 to year 2020, to create a population whose demographic and socioeconomic characteristics roughly replicate those of the U.K. population. In the second stage, the COVID-19 spread is simulated through one-day steps, starting at the beginning of the year 2020 for 360 days. In this article, we provide a brief summary of the demographic module, and refer the reader to Gostoli and Silverman (2019, 2020), Noble et al. (2012), and Silverman et al. (2013) for more details about this module (while the complete Python 2.7 source code for the simulation is available in our GitHub repository at https://github.com/UmbertoGostoli/Pandemic-Behaviour-Model/tree/Pandemic-Only-Sim).

We will describe in detail the novel modules introduced in this article, which are:

• the agent’s social network;

• the virus exposure settings and processes;

• the behavioural module (defining isolation and testing behaviour); and

• the course of the virus.

This stage begins with the generation of an initial population of couples, which are randomly distributed on a 8 × 12-cell grid approximating the geography of the United Kingdom. Each cell of the grid represents a town, and within each town a number of houses is created proportional to the U.K. population density. Each year, a series of demographic events drive the population’s dynamics: births, marriages, divorces, and deaths. Empirical population data (in the form of U.K. Census data) is integrated into the model’s demographic processes in 1951.

With a certain age-specific probability the couple’s female will give birth. Agents enter adulthood at the working age of 16: At this point they can either start working or continue in education, a choice that is repeated at two-year intervals, until the age of 24.¹ After education, agents become employed, taking a salary that is a function of the education level they have reached (which is a stochastic function of their parents’ socioeconomic status; see section 2.1.4). When agents reach retirement age (set at 65 in these simulations), they retire from employment. Mortality rates in the model follow Noble et al. (2012) and use a Gompertz-Makeham mortality model until 1951. From that point we use mortality rates drawn from the Human Mortality Database (https://www.mortality.org/). Lee-Carter projections generate agent mortality rates from 2009.

Once they reach working age, agents can form partnerships. Agents are paired randomly with probabilities that depend inversely on their age, geographical distance from one another, and socioeconomic differences. Age-specific annual divorce probabilities determine whether a couple dissolves their partnership in each year.

Relocation happens most frequently due to agents finding a partner in a different town. Male agents relocate to new houses once a partnership dissolves, and any children produced by that partnership stay with the mother. Retired agents with care needs may move in with one of their adult children, with a probability determined by their care need level and the amount of care supply in their child’s household. Orphaned children are adopted by a household in their kinship network, or by a random family if there are no available households in their kinship network. Apart from these specific, event-driven cases, households also relocate to another town with a certain probability that is inversely proportional to the relative cost of relocation, defined as the ratio between the total cost of relocation and the households’ per capita income.

Agents are placed in one of five socioeconomic status (SES) groups, based on the Approximated Social Grade from the Office for National Statistics (https://www.ukgeographics.co.uk/blog/social-grade-a-b-c1-c2-d-e). The model contains a social mobility process: An agent is assigned the SES group associated with the education level the agent has reached, with a probability of moving further up the education ladder depending on the household income and parental level of education. The introduction of SES groups has a number of effects on the various stages of the agent life course: A higher SES is associated with lower mortality and fertility rates, higher hourly salaries, and higher salary growth rate.

Every employed agent receives an hourly salary that is a function of the agent’s SES and cumulative work experience. On the basis of the total income of the household’s members, each household is assigned an income quintile, which affects a range of processes during the pandemic stage, such as the size of social networks, the probabilities to visit certain venues, the isolation propensities, and the probabilities to develop different conditions (see the next section for details).

The process is repeated until, for each agent, a social network of the desired size has been created. These personal networks are associated with the networks connecting the households. Each household H is associated with:

• - 

  the households of the relatives of household H’s members; and

• - 

  the households of the friends of household H’s members.

The two networks have different roles in the information processing module through which agent behaviour is determined. While the personal friends network affects agent self-isolation propensity, after this has been formed at a first, individual, level, through agent tendency to adjust behaviour towards the mean of the network, the households network affects the formation of individual self-isolation propensity itself, by allowing agents to observe the occurrence of pandemic events and, therefore, providing them with the information they use to develop their subjective probabilities of these events.

In the second stage, the COVID-19 spread is simulated from year 2020 for a period 360 days, through one-day time steps. We assume that, initially, the virus is brought into the U.K. from abroad by international travellers, a process that we will call exogenous infection. Then, the virus spreads within the U.K. by means of two main spreading mechanisms: social interaction and within-household interaction. We assume that the social interaction process depends on a household’s decisions to isolate. After the incubation period, the infected individuals develop various conditions and, at the end of the infection period, either recover or die, as shown in Figure 2.

In our model, social interaction takes place in a series of venues the agents attend. At the beginning of the pandemic stage, a number of venues is created in each town proportional to the town’s population. In each day, we allocate agents to the venues in their town in a way that is consistent with the age-specific interaction matrix for the U.K., as estimated by Prem et al. (2017). Moreover, we assume that a certain percentage of agents visits other towns in each period, to account for daily intra-urban commuting.

During the pandemic, the number of people attending venues decreases as people become aware that interaction with other people carries a risk of them getting the virus and, as a consequence, they self-isolate. We assume that agent self-isolation rate is the result of an individual assessment of the risks and of social processes happening through their social network and their household membership.

As for the first, the individual element, the behavioral framework we introduce in this work is based on two cognitive processes: The first is the process through which agents discount probabilistic gains (and losses); the second is the process through which agents estimate probabilities of negative events. The two cognitive processes interact to determine the individually-determined agent self-isolation rate: First, agents estimate the probabilities of being infected and, if infected, of developing various conditions (i.e., being hospitalized, being intubated, or dying); then, with these probabilities agents compute the expected cost of unrestricted movement (which can be also considered the expected gain of self-isolation). We will consider the two modules in turn.

Agent subjective probabilities are determined by taking into account, sequentially:

• the publicly available information and

• the direct observation of COVID-19 events (within the agent’s social network)

Then, with these probabilities, the expected cost of infection is calculated, and from this, the individually-determined isolation rate. The final isolation rate is the result of an incremental adjustment of this individually-determined isolation rate through the effect of the social pressure acting upon each agent, i.e., the individually-determined rates of the agents’ social network and household members.

The flowchart of the behavioural determination of the agent isolation rate is shown in Figure 3.

In order to compute the expected cost of infection V, the relevant probabilities are the probability of infection, p^f; and, if infected, the probability of being hospitalized p^h, of being intubated p^v, and dying p^d.

From a normative point of view, decision theory dictates that when given alternative choices, individuals should select the alternative with the greatest expected benefit, given by the sum of the products between the probability of the choice’s outcome and the value individuals attach to that outcome (Pratt et al., 1995). Harris et al. (2009), however, found experimental evidence that probabilities and outcomes are not independent: Severe events are usually associated with a higher occurrence probability than “neutral” events. Moreover, they showed that this relationship is mediated by the extent to which the occurrence of the event is controllable by the agents. W. F. Wright and Bower’s (1992) study, on the other hand, suggests a possible causal link between the severity of outcomes and the estimation of their probability, by showing that the subjective probability of events is affected by an individual’s mood: Happy subjects overestimate the likelihood of positive events and underestimate that of negative events; sad subjects display the opposite tendencies, overestimating bad and underestimating good events. On the basis of these studies, we can argue that the higher the severity of a negative event, the stronger the effect of its occurrence will be on an individual’s negative mood and, therefore, the stronger an agent’s overestimation of the probability of its occurrence.

This individually determined cost, V_i, determines both the individual self-isolation propensity, r_i and the agent’s propensity to get tested for COVID-19.

According to this equation, the more distant is the kinship relationship between two households, the smaller weight the help that could potentially be transferred between the two households will have on their self-isolation propensity, with the parameter ϕ determining how fast the influence of the potential help provision decreases as the kinship distance increases.

The agent isolation rate is then used in the social interaction process to determine venue attendants. For each venue, after the pool of people who would normally attend is determined, each agent in the pool is removed with a probability equal to the agent’s self-isolation rate. The parameters of the process determining the self-isolation rate are shown in Figure 4.

For agents taking a test, the result can be positive if they have been infected or in the case of a false positive, or the result can be negative if they have not been infected or in the case of a false negative. In any case, if the test result is positive, they self-isolate completely, and take a new test after a fixed number of days that is a parameter of the model.

The flowchart for the course of the virus is shown in Figure 6. Susceptible agents can become exposed through an exogenous process (for international travelers), or through interaction with other members of their household, or interaction with other agents they meet in their social activities.

Once an agent has become exposed, it is assigned one infection course over four possible courses: asymptomatic; symptomatic not hospitalized; hospitalized not in intensive care; in intensive care. In accordance with empirical studies, we assume that the probability of developing more serious conditions grows with age, decreases with social status, and is higher for males than for females (Abate et al., 2020; Brazeau et al., 2020; Ferguson et al., 2020; Guilmoto, 2020; Public Health of England, 2020; Verity et al., 2020; Yanez et al., 2020).

Upon exposure, the agent is also assigned an incubation period, which, in line with empirical observations (Lauer et al., 2020; McAloon et al., 2020), is drawn from a log-normal distribution with a mean of about 5 days, and a recovery period, whose length depends on the severity level assigned to the agent, in order to reproduce a log-normal distribution of the recovery period with a mean of about 12 days at the population level. The exposed agent is also assigned a viral load, which is drawn from a standard uniform distribution. The agent’s viral load determines its contagiousness.

After the incubation period, the agent starts to develop symptoms (if not asymptomatic) and, in line with empirical observations (He et al., 2020), we assume that the exposed agent becomes infectious 2 days before the emergence of symptoms. The severity of symptoms of not-hospitalized symptomatic agents is differentiated by assigning them a symptoms severity index, between 0 and 1 exclusive, with the probability of the agent being assigned a higher value increasing with its viral load and its age, decreasing with its income quintile, and being higher for males than for females. The closer to 1 the symptoms severity index, the more severe the symptoms are, and the greater the reduction of the agent’s mobility. Therefore, the agent’s mobility is given by the minimum between the illness-affected mobility and the mobility resulting from the behaviourally determined isolation rate of its household. Finally, the agent’s symptoms severity index determines the probability that the agent will take a test.

After the recovery period, a share of agents die, with a probability that also depends on age, social class, and gender. All other agents recover and we assume that they are immune to COVID-19 thereafter.

In this section, we present the results of a benchmark simulation repeated 20 times (the black lines representing the average, and the green bands around it, the standard deviations).² In Figure 7 we show the share of susceptible agents in the population. Contrarily to the usual logistic curve of the traditional susceptible-exposed-infectious-recovered (SEIR) model, where the number of susceptible agents decreases quite fast in the central period of the pandemic, we can see here that the number of susceptible agents decreases quite slowly, with almost 65% of agents still susceptible after 360 days.

Correspondingly, the number of infections, shown in Figure 8, grows relatively quickly until its peak, after around 60 days, to slowly decrease thereafter.

A similar dynamics follows the number of new cases and hospitalizations, shown respectively in Figure 9 and Figure 10. The number of cases increases quite quickly to its peak after around 60 days, to decrease quite slowly from that point, through a series of irregular fluctuations.

Figure 11 shows the total attendances (i.e., the sum of all agents attending venues in a given day), whose dynamics is the result of the self-isolation generated by the behavioural model proposed in this article. We can see that agent self-isolation propensity increases up to the peak of the pandemic, where attendances are about 20% lower than the pre-pandemic figure. After the peak, attendances tend to recover but at a relatively slow pace, in line with the decrease of cases.

In the next three figures, we show how the pandemic has different outcomes depending on the agent’s income quintile. Empirical studies have found conflicting evidence about the relationship between socioeconomic status and outcome severity. While some studies have found a positive relationship (Hawkins et al., 2020), other studies failed to find a clear relationship (Ingraham et al., 2021; Khan et al., 2021; Little et al., 2021). Our simulations confirm the complexity of the relationship between socioeconomic status and COVID-19 infections and outcomes. While the number of infections increases with the income quintiles (Figure 12), as we look at outcomes of increasing severity, the income quintiles that are most affected tend to move towards the lowest one: The central quintiles are those with the highest number of hospitalizations (Figure 13) and the second quintile is that with the highest number of intubations (Figure 14). Note that in our model we do not consider the effect of comorbidity, which may account for a large part of the relationship between socioeconomic status and COVID-19 outcomes: These differences among income quintiles, therefore, are only the effect of the interaction among the different demographic structures, social interaction, and behavioural differences in the interaction patterns of agents belonging to different income levels.

Finally, Figure 15 shows the out-degree distribution of the infection network, where the degree represents the number of infected individuals by each contagious agent. We can see that the relationship between the frequency of degrees and the degrees on the log–log scale is approximately linear, (the R² being −0.97), with a scaling parameter of around 2.5, a typical value of scale-free networks. The fact that the model reproduces a scale-free infections network is rather surprising if we consider that from a behavioural point of view all the agents have a similar behaviour (apart from “linear” differences related to their age and socioeconomic status). Note that in the real world, the so called super-spreaders, are likely to infect more than 12–13 agents. However, we have to consider that this result is highly dependent on the size of the population, which, because of computational constraints, we scaled down by a factor of 1 to 5,000.

Finally, we conduct a sensitivity analysis to estimate the effect of a set of six parameters on the simulation’s output. As the aim of this article is to introduce a behavioural model of self-isolation, the outcome we focus on is the total number of attendances during the pandemic: the lower this number, the higher the degree of self-isolation, on average. As for the choice of the parameters, we focused on six crucial processes of the complex behavioural model proposed, as described below:

• - 

  β, in Equation 27, determining the degree to which agent self-isolation propensity is sensitive to agent income level I and the value of the estimated cost V;

• - 

  σ, in Equation 24, determining the sensitivity of the agent’s subjective estimate of the probability of infection to the relative number of new infections in the agent’s town;

• - 

  ξ, in Equation 19, determining the effect of the relative number of new hospitalizations, intubations, and deaths on the agent’s subjective estimate of the related probabilities;

• - 

  λ, in Equation 29, determining the effect of social norms on the agent self-isolation propensity;

• - 

  τ, in Equation 31, determining the effect of the availability of the kinship network’s help on the household’s mean self-isolation propensity; and

• - 

  ρ, in Equation 23, which is the fear of extinction rate.

From the sensitivity analysis we can see that within our theoretical framework, the parameter that most affects the degree of self-isolation is σ, the parameter determining the effect of the relative number of new infections in the agent’s town on the agent’s estimate of the probability of infection. (See Table 1.) The second most important parameter is β, determining the effect of the product between the income level and the value of the expected cost of infection V. We can notice that V depends on the probability of infection, whose value depends on σ. However, because of Equation 3, the leverage of σ happens to be greater than that of β. The third most important parameter is the fear of extinction rate.

On the other hand, the other three parameters have a much smaller effect. The most important parameter among these three is ν, which regulates the effect of a kinship network’s help on a household’s mean self-isolation, whereas the effect of the social norms and, especially, of the relative number of the various conditions (i.e., new hospitalizations, intubations, or deaths) is quite negligible. This last result is quite surprising if we consider that the relative number of new infections has the largest effect on the outcome. The reason is that public information is only one of the determinants of the subjective probabilities of events: These probabilities are also affected by the events that agents observe (or, do not observe) in each period within their household’s network. Since the three events of hospitalization, intubation, and death are relatively rare for a large part of the population, the fear of extinction process is predominant over the effect of public information.

In this article we demonstrate a proof-of-concept model that simulates interactions among the behavioural adaptations of agents to the COVID-19 pandemic and the course of the pandemic itself. Our behavioural model allows households to reduce their social interactions due to their perceived risks of infection, which provides a first step towards the development of more policy-relevant agent-based models of COVID-19 and future pandemics. The results show that the propensity of agents to self-isolate in the presence of risk has a pronounced effect on the course of the pandemic, which is reflective of real-world outcomes.

We propose that this model can provide a basis for further exploration of behavioural responses to pandemics and their impact on disease transmission. The simulation also allows for the more detailed investigation of inequities in COVID-19 outcomes between different socioeconomic status groups. Future iterations of this model can examine additional health behaviours adopted in response to COVID-19, such as mask-wearing, social distancing, and vaccines, in order to further explore the behavioural responses to the wide variety of pharmaceutical and non-pharmaceutical interventions used during the pandemic.

Umberto Gostoli and Eric Silverman are part of the Complexity in Health Improvement Programme supported by the Medical Research Council (MC_UU_00022/1) and the Chief Scientist Office (SPHSU16). This work was also supported by U.K. Prevention Research Partnership MR/S037594/1, which is funded by the U.K. Research Councils, leading health charities, devolved administrations, and the Department of Health and Social Care.