key: cord-0841096-7bsox19z
authors: Arthur, Ronan F.; Jones, James H.; Bonds, Matthew H.; Ram, Yoav; Feldman, Marcus W.
title: Adaptive social contact rates induce complex dynamics during epidemics
date: 2021-02-10
journal: PLoS Comput Biol
DOI: 10.1371/journal.pcbi.1008639
sha: 803c5837007ca201fefd4d53d121c924490a3eee
doc_id: 841096
cord_uid: 7bsox19z

Epidemics may pose a significant dilemma for governments and individuals. The personal or public health consequences of inaction may be catastrophic; but the economic consequences of drastic response may likewise be catastrophic. In the face of these trade-offs, governments and individuals must therefore strike a balance between the economic and personal health costs of reducing social contacts and the public health costs of neglecting to do so. As risk of infection increases, potentially infectious contact between people is deliberately reduced either individually or by decree. This must be balanced against the social and economic costs of having fewer people in contact, and therefore active in the labor force or enrolled in school. Although the importance of adaptive social contact on epidemic outcomes has become increasingly recognized, the most important properties of coupled human-natural epidemic systems are still not well understood. We develop a theoretical model for adaptive, optimal control of the effective social contact rate using traditional epidemic modeling tools and a utility function with delayed information. This utility function trades off the population-wide contact rate with the expected cost and risk of increasing infections. Our analytical and computational analysis of this simple discrete-time deterministic strategic model reveals the existence of an endemic equilibrium, oscillatory dynamics around this equilibrium under some parametric conditions, and complex dynamic regimes that shift under small parameter perturbations. These results support the supposition that infectious disease dynamics under adaptive behavior change may have an indifference point, may produce oscillatory dynamics without other forcing, and constitute complex adaptive systems with associated dynamics. Implications for any epidemic in which adaptive behavior influences infectious disease dynamics include an expectation of fluctuations, for a considerable time, around a quasi-equilibrium that balances public health and economic priorities, that shows multiple peaks and surges in some scenarios, and that implies a high degree of uncertainty in mathematical projections.

a1111111111 a1111111111 a1111111111 a1111111111 a1111111111

Adapting to a changing landscape of risk during an infectious disease epidemic may pose a significant dilemma for a susceptible individual or for a governing body responsible for the health of susceptible individuals. On the one hand, changing behavior (e.g. through social distancing) can reduce the reproduction number (R 0 ) of an epidemic and save many from death or morbidity [1, 2] . On the other hand, behavior change can reduce an individual's ability to make a living or, for a group of people, can hamper or cause a recession in the economy through decreased production, sales, and investment and increased unemployment, inflation, and debt [3] . This dilemma introduces a behavior change tradeoff for the decision-maker, a balancing act between epidemiological interests and economic interests.

There is growing interest in the role of behavior in infectious disease dynamics (see Funk et al., 2010 [4] for a general review). Behavior relevant to epidemic outcomes is known to change in response to perceived risk during epidemics (e.g. measles-mumps-rubella (MMR) vaccination choices [5] , condom purchases in HIV-affected communities [6] , and social distancing in influenza outbreaks [7] and during the ongoing COVID-19 pandemic [8] ). Although behavior is difficult to measure, quantify, and predict [9] , modelers have adopted a variety of strategies to investigate its role in epidemic outcomes. These strategies include agent-based modeling [10] , network structures that model behavior as a social contagion process [11] or that replace central nodes when sick [12] , and game theoretic descriptions of rational choice under changing incentives, as in the case of vaccination [7, 13, 14] . A common approach to incorporating behavior into epidemic models is to track co-evolving dynamics of behavior and infection [11, [15] [16] [17] .

In epidemic response policy, it is typical to think of behavior change as an exogenouslyinduced intervention without considering associated incentives for the individual or the collective. Due to the interactive relationship between behavior and epidemic dynamics, adaptive behavior should instead be thought of as endogenous to an infectious disease system because it is, in part, a consequence of the prevalence of the disease, which in turn responds to changes in behavior [9, 18 ]. An epidemic system with adaptive behavior responds to the conditions it itself creates, and is thus a complex, adaptive system [19] , subject to the properties and tendencies of such systems.

The interaction between behavioral incentives and epidemic dynamics introduces a negative feedback into the epidemic system. In an important early expansion of Kermack and McKendrick's seminal Susceptible-Infectious-Removed (SIR) model [20] , Capasso and Serio built a self-iterative epidemic model by making the transmission parameter (β) a negative function of the number of infected because "in the presence of a very large number of infectives the population may tend to reduce the number of contacts per unit time." [21] A negative feedback such as this may lead to an endemic equilibrium [22] . This happens because, at low levels of prevalence, the cost of behavior change to avoid disease relative to the risk of infection may not be justified, even though the collective, public benefit in the long-term may be greater. Conversely, as prevalence increases, the probability of infection also increases, thus increasing incentives to adopt protective behavior [13] . If responses are based on outdated information, a negative feedback between prevalence and social contact can produce sustained oscillations in time-series data [23] .

Such periodicity (i.e. multi-peak dynamics) has long been documented empirically in epidemiology [24, 25] . Periodicity can be driven by seasonal contact rate changes (e.g. when children are in school) [26] , seasonality in the climate or ecology [27] , sexual and social behavior change [23, 28] , and host immunity cycling through new births of susceptibles or a decay of immunity over time. Some papers in nonlinear dynamics have studied delay differential equations in the context of epidemic dynamics and found periodic solutions as well [29] . Although it is atypical to include delay in modeling, delay is an important feature of epidemics. Delays of information acquisition, behavioral response, scientific investigation, and those inherent in natural biological processes can affect epidemic outcomes. In the ongoing COVID-19 pandemic, for example, there have been delays in the international recognition of the outbreak [30] , delays in the identification of the virus, delays in the acquisition of reliable information on suspected and confirmed cases [31] , and delays in the development and deployment of competent diagnostics [32] .

Although infectious disease modelers have begun to incorporate adaptive behavior into their models, few studies in the literature capture the competing economic and public health incentives that drive delayed behavioral responses in both individual and group settings during epidemics [33, 34] . Here we develop a theoretical model using both discrete and continuous time and both SIR and SIS compartmental epidemic structures. The model, which is designed to be strategic rather than tactical (sensu Holling [35] ), is adjusted on the principle of endogenous behavior change through an adaptive social-contact rate that can be thought of as either individually motivated or institutionally imposed. We introduce a novel utility function that motivates the population's effective contact rate at a particular time period. This utility function is based on information about the epidemic size that may not be current. This leads to a time delay in the contact function that increases the complexity of the population dynamics of the infection. Results from the discrete-time model show that the system approaches an equilibrium in many cases, although small parameter perturbations can lead the dynamics to enter qualitatively distinct regimes. The analogous continuous-time model retains periodicities for some sets of parameters, but numerical investigation shows that the continuous time version is much better behaved than the discrete-time model. This behavior is similar to that in models of ecological population dynamics, and a useful mathematical parallel can be drawn between these systems.

To represent endogenous behavior change, we start with the classical discrete-time susceptible-infected-susceptible (SIS) model [20] , which, when incidence is relatively small compared to the total population [36, 37] , can be written in terms of the recursions

where at time t, S t represents the number of susceptible individuals, I t the infected individuals, and N t the number of individuals that make up the population, which is assumed fixed in a closed population. We can therefore write N for the constant population size. Here γ, with 0 < γ < 1, is the rate of removal from I to S due to recovery. This model in its simplest form assumes random mixing, where the parameter b represents a composite of the average contact rate and the disease-specific transmissibility given a contact event. In order to introduce human behavior, we substitute for b a time-dependent b t , which is a function of both b 0 , the probability that disease transmission takes place on contact, and a dynamic social contact rate c t whose optimal value, c � t , is the number of contacts per unit time that maximize utility for the individual. c � t is determined at each time t as in economic epidemiological models [34] ,

where c � t represents the optimal contact rate, defined as the number of contacts per unit time that maximize utility for the individual. Here, c � t is a function of the number of infected in the population according to the perceived risks and benefits of social contacts, which we model as a utility function. We assume that there is a constant utility independent of contact, a utility loss associated with infection, and a utility derived from the choice of number of daily contacts with a penalty for deviating from the choice of contacts which would yield the most utility.

This utility function is assumed to take the form

Here U represents utility for an individual at time t given a particular number of contacts per unit time c, α 0 is a constant that represents maximum potential utility achieved at a target contact rateĉ. The second term, À a 1 ðc ÀĉÞ 2 , is a concave function that represents the penalty for deviating fromĉ. The third term,

, is the cost of infection (i.e. morbidity), α 2 , multiplied by the probability of infection over the course of the time unit. The time-delay Δ represents the delay in information acquisition and the speed of response to that information. We note that 1 À I N b 0 À � c can be approximated by

when I N b 0 is small and c I N b 0 � 1: We thus assume I N ðb 0 Þ is small, that is, prevalence is low, and approximate U(c) in Eq 5 using Eq 6. Eq 5 assumes a strictly negative relationship between number of infecteds and contact.

We assume an individual or government will balance the cost of infection, the probability of infection, and the cost of deviating from the target contact rateĉ to select an optimal contact rate c � t , namely the number of contacts, which takes into account the risk of infection and the penalty for deviating from the target contact rate. This captures the idea that individuals trade off how many people they want to interact with versus their risk of becoming infected, or that authorities want to reopen the economy during a pandemic and have to trade off morbidity and mortality from increasing infections with the need to allow additional social contacts to help the economy restart. This optimal contact rate can be calculated by finding the maximum of U with respect to c from Eq 5 with substitution from Eq 6, namely

Differentiating, we have

which vanishes at the optimal contact rate, c � , which we write as c � t to show its dependence on time. Then

which we assume to be positive. Therefore, total utility will decrease as I t increases and c � t also decreases. Utility is maximized at each time step, rather than over the course of lifetime expectations. In addition, Eq 9 assumes a strictly negative relationship between number of infecteds at time t − Δ and c � t . While behavior at high degrees of prevalence has been shown to be nonlinear and fatalistic [38, 39] , in this model, prevalence (i.e., b 0 I t N ) is assumed to be small, consistent with Eq 6.

We introduce the new parameter a ¼ a 2 2a 1 b 0 , so that

We can now rewrite the recursion from Eq 2, using Eq 4 and replacing c t with c � t as defined by Eq 10, as

When Δ = 0 and there is no time delay, f(�) is a cubic polynomial, given by

For the susceptible-infected-removed (SIR) version of the model, we include the removed category and write the (discrete-time) recursion system as [20] . The inclusion of the removed category entails thatĨ ¼ 0 is the only equilibrium of the system Eqs 13-15; unlike the SIS model, there is no equilibrium with infecteds present. In general, since c � t includes the delay Δ, the dynamic approach toĨ ¼ 0 is expected to be quite complex. Numerical analysis of this SIR model shows strong similarity between the SIS and SIR models for several hundred time steps before the SIR model converges toĨ ¼ 0. In the section "Numerical Iteration and Continuous-Time Analog" we compare the numerical iteration of the SIS (Eq 11) and SIR (Eqs 13-15) and integration of the continuous-time (differential equation) versions of the SIS and SIR models.

To determine the dynamic trajectories of (11) without time delay, we first solve for the fixed point(s) of the recursion (11) (i.e., value or values of I such that f(I t+1 ) = I t = I t−Δ ). That is, we solve

From Eq 16, it is clear that I = 0 is an equilibrium as no new infections can occur in the next time-step if none exist in the current one. This is the disease-free equilibrium denoted byĨ. Other equilibria are the solutions of

namely a þĉ � ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ða ÀĉÞ

We label the solution with the + sign I � and the one with the − signÎ. I � > 0 but I � � N if 4αγ/Nb 0 � 0, which is impossible under our assumptions that α and γ are positive. Hence I � is not feasible. Further, under these same conditions,Î � N, andÎ > 0 if

It is important to note that under these conditionsÎ is an equilibrium of the recursion (11) for any Δ � 0. Recall that for the SIR version of this model the only equilibrium isĨ ¼ 0.

Assessing global asymptotic stability in epidemic models is an important task of mathematical epidemiology [40, 41] . The three equilibria of the SIS recursion (11) are qualitatively different. I ¼ 0 corresponds to a disease-free population; I � is greater than N and is therefore not feasible;Î is the only positive feasible equilibrium ifĉb 0 > g=N (this is equivalent to R 0 > 1, where R 0 ¼ Nĉb 0 þ 1 À g) and is, therefore, the most interesting for the asymptotic stability behavior of the epidemic. Mathematical stability analysis of recursion (11) is complicated because of the delay term Δ. However, from (11), if Nĉb 0 > g, the disease-free equilibriumĨ ¼ 0 is locally unstable, and in this caseÎ is indeed feasible.

Local stability ofÎ in (18) is discussed in detail in S1 Appendix. First, in the absence of delay (i.e., Δ = 0),Î is locally stable if j d dI f ðIÞj I¼Î < 1, and the condition for this to hold whenÎ is legitimate is b 0Î ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi

If inequalities (20) and Nĉb 0 > g hold, thenÎ is locally stable. However, even if both of these inequalities hold, the number of infecteds may not converge toÎ. It is well known that iterations of discrete-time recursive relations, of which (12) is an example (i.e., with Δ = 0), may produce cycles or chaos depending on the parameters and the starting frequency I 0 of infecteds.

We begin with numerical analysis of the discrete-time SIS recursion (11) , which includes the delay parameter Δ. Local stability properties of the equilibrium stateÎ, with 0 <Î < N, are shown in the Appendix under the assumption Nĉb 0 > g, which also entails that the diseasefree equilibriumĨ ¼ 0 is locally unstable. In the recursion (11), the number of infecteds at time t will not, in general, be integers, but can be interpreted as the expected number of infected in the population. Further, the dynamics of I t under such a recursion can be very sensitive to the starting condition I 0 , the size of the time delay Δ, and the parameters: N; b 0 ; g;ĉ; and α. The local stability ofÎ, namely whether I t converges toÎ from a starting number of infecteds close toÎ, may tell you little about the actual trajectory of I t from other starting conditions. Table 1 reports an array of dynamic trajectories without delay (Δ = 0) for some choices of parameters. In seven cases, I 0 = 1, and in two cases the numerical iteration of Eq 12 was initiated with I 0 6 ¼ 1. The first three rows show three sets of parameters for which the equilibrium values ofÎ are very similar but the trajectories of I t are different: a two-point cycle, a fourpoint cycle, and apparently chaotic cycling above and belowÎ. In all of these cases, df ðIÞ=dIj I¼Î < À 1. Clearly the dynamics are sensitive to the target contact rateĉ in these cases. The fourth and eighth rows show that I t becomes unbounded (tends to + 1) from I 0 = 1, but a two-point cycle is approached if I 0 is close enough toÎ : df ðIÞ=dIj I¼Î < À 1 in these cases. For the parameters in the ninth row, if I 0 is close enough toÎ there is damped oscillation intoÎ: here À 1 < df ðIÞ=dIj I¼Î < 0. In the case marked � ,Î is locally stable and with a large enough initial number of infecteds, there is damped oscillatory convergence toÎ. In the case marked �� , with I 0 = 1 the number of infecteds becomes unbounded, but in this case,Î is locally unstable (df ðIÞ=dIj I¼Î < À 1), and starting from I 0 close toÎ a stable two-point cycle is approached. (11) with Δ = 0 and the other parameters from the first three lines of Table 1 . Asĉ increases, first there is convergence toÎ, then period doubling to chaos and finally passage to negative infinity.

Stability analysis of the SIS model is more complicated when Δ 6 ¼ 0, and in S1 Appendix we outline the procedure for local analysis of the recursion (11) nearÎ. Local stability is sensitive to the delay time Δ as can be seen from the numerical iteration of (11) for the specific set of parameters shown in Table 2 . Some analytical details related to Table 2 are in S1 Appendix.

The fifth and sixth rows of Table 1 exemplify another interesting dynamic starting from I 0 = 1. I t becomes larger thanÎ (overshoots) and then converges monotonically down toÎ; in each case 0 < df ðIÞ=dtj I¼Î < 1. For the parameters in the seventh row, there is oscillatory convergence toÎ from I 0 = 1 (À 1 < df ðIÞ=dIj I¼Î < 0), while in the last row there is straightforward monotone convergence toÎ. The dependence of the dynamics for recursion (11) on the delay Δ and target contact rateĉ is illustrated for Δ = 0, 1, 2 in S6 Fig. The bifurcation diagram for each Δ shows the shift, summarized in Table 2 , from convergence to period doubling, chaos, and negative infinity, which occurs for smaller values ofĉ as Δ increases.

A continuous-time analog of the discrete-time recursion (11) , in the form of a differential equation, substitutes dI/dt for I t+1 − I t in (11) . We then solve the resulting delay differential equation numerically using the VODE differential equation integrator in SciPy [42, 43] (source code available at https://github.com/yoavram/SanJose). Using the parameters in Table 2 , Figs 1-4 compare the effect of the parameters on the trajectories of the discrete-time and continuous-time SIS model specified in (11) . The number of time steps used in the In Fig 1, with no delay (Δ = 0) and a one-unit delay (Δ = 1), the discrete and continuous dynamics are very similar, both converging toÎ. However, with Δ = 2 the differential equation oscillates intoÎ while the discrete-time recursion enters a regime of inexact cycling aroundÎ, which appears to be a state of chaos. For Δ = 3 and Δ = 4, the discrete recursion "collapses". In other words, I t becomes negative and appears to go off to −1; in Fig 1, this is cut off at I = 0. It is worth noting that if the total population size of N decreases over time, for example, if we take N(t) = Nexp(−zt), with z ¼ 50b 0ĉ g, then the short-term dynamics of the SIS model in (11) begins to closely resemble the SIR version. This is illustrated in S9 Fig, where b 0 ;ĉ; g are, as in S8 Fig, the same as in Fig 2A. With N decreasing to zero, both S and I will approach zero in the SIS model, which explain its apparent similarity to the SIR model.

This simple epidemic model with adaptive social contact produces two possible equilibria, one with zero infecteds, where the disease is eradicated, and one between zero and N, the population size, where the disease is endemic. These equilibria are locally stable under different conditions. Dynamics produced by this model are complex and subject to regime shifts across thresholds in the initial conditions and parameter settings. These dynamics include damped oscillation to the equilibrium, periodic oscillation, chaotic oscillation, and regression to positive or negative infinity. Our stability analysis is carried out in the neighborhood of the equilibria. Although global asymptotic stability analysis of some epidemic models has been possible [29, 40, 41] , the inclusion of the delay Δ seems to make global analysis extremely difficult in general [29] .

Our model makes a number of simplifying assumptions. We assume that all individuals in the population will respond in the same fashion to government policy and that governments or individuals choose a uniform contact rate according to an optimized utility function, which is homogeneous across all individuals in the population. This contact rate will, in practice, vary across the population according to a variety of drivers including, but not limited to, disease state, cultural and religious practices, political affiliation, housing density, occupation, risk tolerance, and age. Finally, we assume that the utility function is symmetric around the optimal number of contacts so that increasing or decreasing contacts above or below the target contact rate, respectively, yield the same reduction in utility. These assumptions allowed us to create the simplest possible model that includes adaptive behavior trade-offs and time delay.

Convergence to an endemic equilibrium when economic and public health trade-offs are included in an epidemic model is consistent with both theory [22] and other models [33] . Our results show certain parameter sets can lead to limit-cycle dynamics, consistent with other behavior change models [23, 44] and negative feedback mechanisms with time delays [45, 46] . This is because the system is reacting to conditions that were true in the past, but not necessarily true in the present. The time scale and the meaning of the delay, Δ, can influence the qualitative dynamics of the epidemic and, under certain conditions, can lead to a stable cyclic epidemic even in the continuous-time version of our model. We note that these distinct dynamical trajectories as seen in our computational experiments come from a purely deterministic recursion. This means that oscillations and even erratic, near-chaotic dynamics and collapse in an epidemic may not necessarily be due to seasonality, complex agent-based interactions, changing or stochastic parameter values, demographic change, host immunity, or socio-cultural idiosyncrasies. In our discrete-time model, there is the added complexity that the non-zero equilibrium may be locally stable but not attained from a wide range of initial conditions, including the most natural one, namely a single infected individual.

This dynamical behavior in number of infecteds can result from mathematical properties of a simple deterministic system with homogeneous endogenous behavior change, similar to complex population dynamics of biological organisms [47] . The mathematical consistency with population dynamics suggests a parallel in ecology, that the indifference point for human behavior functions in a similar way to a carrying capacity in ecology, below which a population will tend to grow and above which a population will tend to shrink. For example, the Ricker Equation [48] , commonly used in population dynamics to describe the growth of fish populations, exhibits similar complex dynamics and qualitative state thresholds. These ecological models are typically structured mathematically in discrete time, while continuous time models are more commonly used in modeling epidemics. There is no a priori reason to prefer the continuous time framework over that in discrete time. It is not clear which strategic approach is more realistic as transmission from an infected to a susceptible individual may happen at anytime, but epidemiologists do tend to frame their thinking in discrete time-steps of days and weeks.

Observed epidemic curves of many transient disease outbreaks typically inflect and go extinct, as opposed to this model that may oscillate perpetually or converge monotonically or cyclically to an endemic disease equilibrium. Including institutional and public efforts that are further incentivized to eradicate, rather than to optimize short-term utility trade-offs, would alter the dynamics to look more like real-world epidemic curves. Beyond infectious diseases that remain endemic to society, outbreaks may also flare up once or multiple times, such as the double-peaked outbreaks of SARS in three countries in 2003 [49] , and surges in fluctuations in COVID-19 cases globally in 2020 [50] . There may be many causes for such double-peaked outbreaks, one of which may be a lapse in behavior change after the epidemic begins to die down due to decreasing incentives [11] , as represented in our simple theoretical model. This is consistent with findings that voluntary vaccination programs suffer from decreasing incentives to participate as prevalence decreases [51, 52] . A recent analysis [53] that incorporated epidemiclike transmission of sentiment opposed to vaccination against an infection found that the transient dynamics of the anti-vaccine sentiment could induce complex dynamics of the disease epidemic. However, this analysis did not incorporate a time delay in the manifestation of the anti-vaccine sentiment. The relation between the spread of the sentiment and of the infection is, therefore, somewhat different from that seen here between an adaptive contact rate and the epidemic dynamics.

One of the responsibilities of infectious disease modelers is to predict and project forward what epidemics will do in the future in order to better assist in the proper and strategic allocation of preventative resources. However, there are limits to the power and precision of such modeling. In our model, allowing for adaptive behavior change leads to a system that is qualitatively sensitive to small differences in values of key parameters. These parameters are very hard to measure precisely; they change depending on the disease system and context and their inference is generally subject to large errors. Further, we don't know how policy-makers weight the economic trade-offs against the public health priorities (i.e., the ratio between α 1 and α 2 in our model) to arrive at new policy recommendations. Geographic and/or cultural variation in our parameter c � t (and concomitant variation in the delay Δ) are likely to affect how epidemic dynamics are affected by such trade-offs.

In our model, complex dynamic regimes occur more often when there is a time delay. If behavior change arises from fear and fear is triggered by high local mortality and high local prevalence, such delays are biologically inherent because death and incubation periods are lagging epidemiological indicators. Lags, whether social, environmental, or biological, mean that people can respond inappropriately to an unfolding epidemic crisis, but they also mean that people can abandon protective behaviors prematurely as conditions improve. Developing approaches to reduce lags or to incentivize protective behavior throughout the duration of any lag introduced by the natural history of the infection (or otherwise) should be a priority in applied research. Policy-makers should also consider the benefit of the long-term utility of early-stage overreaction to outbreaks and consider overriding short-term incentives. In light of the COVID-19 crisis, understanding endogenous delayed behavior change and economic incentives is of crucial importance to outbreak response and epidemic management. We anticipate further developments along these lines that could incorporate long incubation periods and other delays, recognition of asymptomatic transmission, influential heterogeneous drivers, and meta-population dynamics of simultaneous, connected epidemics.

Supporting information S1 Appendix. Local stability of the endemic equilibriumÎ. Conditions are given for various values of the delay time Δ and the parameters in Table 2 Table 1 on the x-axis and its corresponding reproduction number R 0 . The dotted horizontal line delineates the total population size (N = 250). Dynamics exhibit convergence to the endemic equilibrium (including monotonic, overshooting, and damped oscillation) and period doubling to chaos, followed by passage to negative infinity. Table 2 with varying target contact rateĉ. Dynamics progress from convergence to chaos to negative infinity. As Δ increases, transitions between dynamic regimes begin at smaller values ofĉ. 

Social factors in epidemiology

Capturing human behaviour

The macroeconomics of epidemics National Bureau of Economic Research

Modelling the influence of human behaviour on the spread of infectious diseases: a review

Evolving public perceptions and stability in vaccine uptake

The responsiveness of the demand for condoms to the local prevalence of AIDS

Game theory of social distancing in response to an epidemic

Scientific and ethical basis for social-distancing interventions against COVID-19. The Lancet Infectious Diseases

Nine challenges in incorporating the dynamics of behaviour in infectious diseases models

A hybrid epidemic model: combining the advantages of agent-based and equation-based approaches

Coupled contagion dynamics of fear and disease: mathematical and computational explorations

The effect of a prudent adaptive behaviour on disease transmission

Erratic Flu Vaccination Emerges from Short-Sighted Behavior in Contact Networks

A general approach for population games with application to vaccination

The spread of awareness and its impact on epidemic outbreaks

Coupled disease-behavior dynamics on complex networks: A review

The Dynamics of Physiologically Structured Populations

Contact structure, mobility, environmental impact and behaviour: the importance of social forces to infectious disease dynamics and disease ecology

Hidden order: How adaptation builds complexity. Basic Books

A contribution to the mathematical theory of epidemics

A generalization of the Kermack-McKendrick deterministic epidemic model

Economic epidemiology and infectious diseases

Information-related changes in contact patterns may trigger oscillations in the endemic prevalence of infectious diseases

Periodicity in epidemiological models

Gonorrhea transmission dynamics and control

Measles in England and Wales-I: an analysis of factors underlying seasonal patterns

Seasonal and interannual cycles of endemic cholera in Bengal 1891-1940 in relation to climate and geography. In: The Ecology and Etiology of Newly Emerging Marine Diseases

Epidemic cycles driven by host behaviour

Periodic solutions of delay differential equations arising in some models of epidemics

Crippled community governance and suppressed scientific/professional communities: A critical assessment of failed early warning for the COVID-19 outbreak in China

Impact of delays on effectiveness of contact tracing strategies for COVID-19: a modelling study. The Lancet Public Health

Diagnostic testing for the novel coronavirus

Merging economics and epidemiology to improve the prediction and management of infectious disease

Adaptive human behavior in epidemiological models

The strategy of building models of complex ecological systems

Modeling infectious diseases in humans and animals

Time series modelling of childhood diseases: a dynamical systems approach

Choices, beliefs, and infectious disease dynamics

Higher disease prevalence can induce greater sociality: a game theoretic coevolutionary model

Global stability of an SIR epidemic model with information dependent vaccination

Global stability for the SEIR model in epidemiology

SciPy 1.0: fundamental algorithms for scientific computing in Python

Scipy-based delay differential equation (dde) solver

Epidemic models with nonlinear infection forces

Time-delay versus stability in population models with two and three trophic levels

Time delays are not necessarily destabilizing

Simple mathematical models with very complicated dynamics

Stock and recruitment

Different epidemic curves for severe acute respiratory syndrome reveal similar impacts of control measures

Coronavirus disease (2019): situation report

Rational epidemics and their public control

Group interest versus self-interest in smallpox vaccination policy

Modeling anti-vaccine sentiment as a cultural pathogen

The authors thank Kaleda Krebs Denton and W. Brian Arthur for helpful comments on an earlier draft of the manuscript.