Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Ross-Macdonald Models: Which one should we use?

Ross-Macdonald Models: Which one should we use? Ross-Macdonald models are the building blocks of most vector-borne disease models. Even for the same disease, di erent authors use di erent model formula- tions, but a study of the dynamical consequences of assuming di erent hypothe- ses is missing. In this work we present di erent formulations of the basic Ross- Macdonald model together with a careful discussion of the assumptions behind each model. The most general model presented is an agent based model for which arbitrary distributions for latency and infectious periods for both, host and vectors, is considered. At population level we also developed a deterministic Volterra inte- gral equations model for which also arbitrary distributions in the waiting times are included. We compare the model solutions using di erent distributions for the in- fectious and latency periods using statistics, like the epidemic peak, or epidemic final size, to characterize the epidemic curves. The basic reproduction number (R ) for each formulation is computed and compared with empirical estimations obtained with the agent based models. The importance of considering realistic distributions for the latent and infectious periods is highlighted and discussed. We also show that seasonality is a key driver of vector-borne disease dynamics shaping the epidemic curve and its duration. Keywords: Ross-Macdonald Model, Epidemiology, Delayed Model, Agent Based Model 1. Introduction Preprint submitted to Acta Tropica arXiv:2002.11267v1 [q-bio.PE] 26 Feb 2020 Vector-borne diseases are caused by di erent types of parasites, including viruses and bacteria, which are transmitted by vectors as mosquitoes, sandflies, ticks, and kissing-bugs, among others. According to the World Health Organi- zation, every year there are more than 700 thousand deaths as a consequence of vector-borne diseases [47]. Mosquito-borne diseases of humans include malaria, dengue, zika, chykun- gunya, yellow fever (see for example [41, 11, 20]). Di erent triatomine species transmit Trypanosoma cruzi the causal agent of Chagas disease (see for example [37] and references there in) while Leishmaniasis is transmitted by several species of sandflies [9]. Vector-borne diseases are also common zoonotic diseases. Some forms of Leishmaniasis cannot be transmitted from humans to sandflies and the parasite population survives in a wild cycle including small rodents, dogs, cows and sev- eral species of birds as hosts [3, 9]. West Nile Virus may be transmitted to humans but it is maintained in a cycle which includes several species of birds [46]. Try- panosoma cruzi, the causal agent of Chagas disease is also transmitted to di erent animals including dogs, marsupials, rodents, and others hosts [23]. Ross model was published in 1911 [35] and remains as the basis of countless models for vector-borne diseases. Ross considered a simple model for malaria, with births and deaths but with constant populations and infectious periods expo- nentially distributed. Humans and mosquitoes may be in only two classes: Af- fected and Una ected (what here we will denoted by H ; H ; V ; V ). Then, Ross i s i s model in continuous time reads dH V i i = m (H H ) r H h i h i dt V dV H i i = (V V i)  V v v i dt H where H and V are the numbers of humans and mosquitoes, m is the number of mosquitoes per human (V=H), r is the recovery rate for humans,  is the mor- h v tality rate for mosquitoes, and are the transmission parameters which may be decomposed as = b f p with b the mosquitoes biting rate, f the proportion of j j bites in humans, and p the probability of transmission per bite. Ross formulation is still used but it is not advisable. The parameter m is in fact a dynamical variable. For the original Ross model this was not a problem as he considered constant pop- ulations. However, both vector and host populations may vary in time, therefore, 2 an equivalent formulation, more frequently used, and preferable is dH H i s = V r H (1) h i h i dt H dV H i i = V  V (2) v s v i dt H Models with these rates of infection are broadly known as Ross-Macdonald models, albeit Macdonald’s contribution to the Ross model is not reflected in this model formulation. Macdonald modified the original Ross model, integrating bi- ological information about the mosquito latency period, and introduced the ex- posed class for vectors [26]. Later he considered also the case of super-infection in Malaria disease dynamics (see for an extensive discussion [40]). For the Ross model, the basic reproduction number (R ), defined as the number of secondary host cases produced by a typical infectious host in a completely susceptible population is h v R = r  H h v This celebrated result from Ross [35] shows that the basic reproduction num- ber is proportional to the number of vectors per host (V=H), and therefore, disease transmission may be interrupted if the number of vectors per host is reduced below some threshold. Since the pioneering work of Ross, several extensions of his basic model (Eqs. 1 - 2) were developed including the addition of exposed classes, superinfection, spatiality, time-varying populations, age structure and more (see for example [16, 8, 6, 26, 40, 29, 36]), and applied to the study of di erent infectious diseases such as malaria, dengue, yellow fever, cutaneous Leishmaniasis, Chagas disease, West Nile virus, among others (see for example [44, 39, 26, 2, 7, 43, 46] and references therein). In this work we present a detailed analysis of some general Ross-Macdonald models. We show that the inclusion of exposed classes as well as the distribution of the latent and infectious periods, have significant dynamical consequences. We also show that seasonality is a major factor shapening the epidemics curves. This paper is arranged as follow. In the next section we discuss the general assumptions common of all models presented. In Section 3 several deterministic Ross-Macdonald models are developed considering exposed classes and di erent distributions for the waiting periods. The basic reproduction number is computed in each case. A stochastic agent based model (ABM) is developed in Section 3 4. Numerical results, such as epidemic curves, epidemic final sizes, the basic reproduction number are computed for each model and compared between them in Section 5. The key role of seasonality is also discussed. Finally present the discussion of the results and conclusions. 2. General assumptions and parameters In a Ross-Macdonald model it is assumed that populations are homogeneously mixed. Vector’s bites are evenly divided among hosts, that is, every time a vector bites, chooses a host at random. This hypothesis leads to a frequency dependent transmissions terms proportional to V H =H and V H =H. This central hypothesis i s s i is perhaps what define what a Ross-Macdonald model is. However we want to stress that this assumption is only realistic for small populations like a household. The use of Ross-Macdonald type of models for larger populations will be analyzed elsewhere. Demography. Immigration and emigration are not considered as we are in- terested in the simplest cases. Births are assumed to take place at a (density- independent) rate . Deaths may be described by the mortality or by the survival function. Mortality () is the number of deaths per individual and per unit of time. In general it is an age-dependent rate. The survival function, F(a), is the proportion of individuals still alive at age a, and it is related with the mortality by (s)d s F(a) = 1 e . Epidemiology. Populations are divided in some of the following epidemiolog- ical classes: Susceptible, Latent, Infectious, and Recovered. Latent (or Exposed) individuals are infected but not infectious (and therefore are unable to transmit the disease). Recovered individuals are immune, and therefore do not participate of the transmission process. Duration of the latent period may be described for a survival function of the age of infection, F (s) which gives the proportion of latent individuals who remain latent at age of infection s (age of infection is the time elapsed since first infection). Analogously, F (s) is the proportion of infec- tious individuals who remain infectious after a time s after the end of latency. Alternatively we can use the, (in general) age-of-infection dependent, progression rates (from latency to infectiousness) or recovery rates (from infectiousness to recovery). All the periods considered (lifespan, latency period, infectious period) are ran- dom variables which may be characterized by a probability distribution. The sim- ple, and commonly used case of exponentially distributed periods correspond to 4 constant, age independent, rates. For example using a constant mortality rate imply the assumption of an exponentially distributed lifespan. Parameters defining the di erent periods distributions are: T : Host life expectancy (mean lifespan) T =T : Vector life expectancy, mean infectious period for vectors v vi T : Mean latency period for exposed hosts he T : Mean infectious period for hosts hi T : Mean latency period for vectors ve For the limiting cases of exponentially distributed or fixed periods these param- eters values completely define the probability distributions. In the general case other parameters like the variance of the distribution should be provided. In all cases we considered that vectors are infectious for life. Entomological parameters. Biting rate on hosts (number of bites per vec- tor, per unit of time, on hosts) is denoted by b. Probabilities of transmission per bite are p and p (from vectors to hosts and from hosts to vectors respectively). h v Finally we define = p b, and = p b. h h v v Basic reproduction numbers. For a general Ross-Macdonald model the ba- sic reproduction number may be obtained by simple bookkeeping [15]. One in- fectious host will produce an average of V infected vectors per unit of time. If the mean infectious period for hosts is T , then the total number of infected hi vectors is V T . Only a fraction f will survive the latency period, and there- v hi v fore, the total number of infectious vectors produced by the initial infectious host is V T f . Each infectious vector would produce T host infections (T is v hi v h vi vi the mean infectious period for vectors) and only a fraction f will survive the host latency period. Finally the basic reproduction number is given by R = T T f f (3) 0 h v hi vi h v 3. Deterministic Ross-Macdonald models In a Ross-Macdonald model there are host and vector populations (of size H and V respectively) homogeneously mixed. Each population is subdivided in epidemiological classes. For example, susceptible and infectious host and vector populations (H , H , V V ). Vectors bite at the rate b (daily number of bites per s i s i 5 vector, for example). If p is the probability of infection transmission to hosts per bite, p the probability of vector infection per bite on infectious hosts, then, the rate of infection of susceptible hosts is given by p bV while the rate of infection h i of susceptible vectors by p bV . These functional forms for the infection rates v s are characteristic of all the Ross-Macdonald type models. In the following we will present, discuss and compare the more common deterministic models (without age structure). 3.1. Basic Model One of the most simple, general, and used model is the S IR model for hosts and a S I model for vectors. Mortalities are denoted by  while recovery rates by r. ’s are the recruitment rates. We will assume that all the periods are exponentially distributed and therefore we obtain the following Basic model: dH H s s =  V  H (4) h h i h s dt H dH H i s = V (r +  )H (5) h i h h i dt H dH = r H  H (6) h i h r dt dV H s i =  V  V (7) v v s v s dt H dV H i i = V  V (8) v s v i dt H where  = 1=T and  = 1=T . Mean infectious period for host includes recov- h h v v ery and mortality, and therefore in this case is given by T = 1=(r +  ), from hi h h where recovery rate r can be estimated. In this work we will consider only the case  = 0, but for many species of hosts,   r and therefore we may approxi- h h mate the recovery rate by 1=T . Vectors are assumed to be infectious for life and then  = 1=T = 1=T . v vi v Because in this model there are not latency periods, f = f = 1 and the basic h v reproduction number (Eq. 3) becomes h v (1) R = (9) (r +  ) H h h v 6 The assumption of constant mortality for vectors is plausible as for insects we expect an approximately constant daily probability of death. For hosts like birds, constant mortality is also usually observed. However hosts like humans present a survival of type I: low mortality for ages below the mean followed by a steep decrease in survival. In this case an age structured model for the host population should be used (see for example [36]). However in those cases we have that and therefore we may disregard birth and deaths in the host population h v when studying the short-term dynamics like in a single outbreak, the case we are considering in this work. Infectious period is also assumed exponentially distributed, a not realistic as- sumption. Hosts may lose immunity becoming susceptible again, a case we do not consider in this work. 3.2. Basic Model with exposed classes For both, hosts and vectors, there are latent periods and therefore a more re- alistic model is a S E IR model for hosts and a S E I model for vectors (as in most cases vectors are infectious for life). The basic model with latent classes (SEIR- SEI model) is: dH H s s =  V  H (10) h h i h s dt H dH H e s = V (k +  )H (11) h i h h e dt H dH = k H (r +  )H (12) h e h h i dt dH = r H  H (13) h i h r dt dV H s i =  V  V (14) v v s v s dt H dV H e i = V (k +  )V (15) v s v v e dt H dV = k V  V (16) v e v i dt Here, k and k are the progression rates from latency to infectiousness, and in h v this context are given by k = 1=T with T the mean latency periods ( j = h for j je je hosts, and j = v for vectors). 7 In this case the basic reproduction number is ! ! k k V h v v h (2) R = (17) (r +  ) k +  k +  H h h v v v h h where f = k =(k +  ) are the fractions of exposed individuals who survive the j j j j latency period. The assumptions in this model are the same discussed above but here it is also assumed that latent periods are exponentially distributed a not realistic assumption neither. Once again k   and then  1. h h k + h h 3.3. Models with arbitrary distributions for the waiting periods The assumption of exponentially distributed periods is appealing because the corresponding ODE models have constant parameters. However latency or infec- tious periods are, in general, random variables with non-exponential distributions. In our case, where we are considering that vectors remain infectious for life, the infectious period is the vector lifespan. In this case a constant mortality is a realistic choice and therefore the infectious period is exponentially distributed. However this is not the case of vector’s latent period or the latent and infectious host’s periods. As an example we will first consider the simple case of a S IR S I model. For vectors we have the equations 7-8. For the host population we will consider that the infectious period (T ) is a random variable with probability distribution hi function f (s). As usual, the cumulative distribution is denoted by F(s). The complementary cumulative distribution, F(s) = 1 F(s), is known as the survival function and gives the probability that an individual infected in s = 0 remains infected at time s. Because only the fraction F(t s) of the infections produced at time s survives until time t we obtain the integral Volterra equations H (t) = H (0) V (s)H (s)d s s s i s ¯ ¯ H (t) = H (0)F(t) + V (s)H (s)F(t s)d s i i i s H (t) = H H (t) H (t) r s i Di erentiation of Volterra equation gives the following system of integro- di erential equations, 8 dH H s s = V h i dt H ¯ ¯ dH dF H dF i s h = H (0) + V F(0) + V (s)H (s) (t s)d s i h i i s dt dt H H dt s h = H (0) f (t) + V V (s)H (s) f (t s)d s i h i i s H H Realistic distributions for infectious or latent periods are bell shaped and there- fore survival function is of type I. Then, a simple but realistic distribution is ob- tained for the limiting case of fixed infectious period T . In this case the survival hi function is a step function, the probability density distribution is (t T ), and hi therefore we obtain the delayed equation dH H H (t T ) i s s hi = H (0)(t T ) + V V (t T ) (18) i hi h i h i hi dt H H In the general case of arbitrary distributions in latency and infectious periods for hosts and vectors we have the following Volterra integral equations model, H (s) H (t) = H (0) V (s) d s (19) s s h i H (s) ¯ ¯ H (t) = H (0)F (t) + V (s) F (t s)d s (20) e e he h i he Z Z " # H (s) dF s he ¯ ¯ H (t) = H (0)F (t) + V (s) ( s) F (t ) d sd (21) i i hi h i hi H dt 0 0 H (t) = H H (t) H (t) H (t) (22) r s i e Z Z t t H (s) t (ts) V (t) = V (0)e +  e d s V (s) d s (23) s s v v s 0 0 H (s) t (ts) V (t) = V (0)e + V (s) F (t s)e d s (24) e e v s ve " # Z Z H (s) dF i ve t (ts) V (t) = V (0)e + V (s) ( s) e d sd (25) i i v s H dt 0 0 ¯ ¯ where F , F are the survival functions for the exposed and infectious populations je ji ( j = h for host and j = v for vectors). 9 3.3.1. Gamma distributed periods Realistic probability distribution functions for infectious or latent periods are bell shaped and therefore the survival functions are of type I. While accurate nu- merical solutions of a system of ordinary di erential equations like model 10 - 16 are easily obtained using a Runge-Kutta scheme, for example, integral systems like 19-25 are not that amenable. Gamma distributions are flexible functions with two parameters, the shape parameter k and the scale parameter . Some features of this distribution are par- ticularly appealing. The exponential distribution is a special case of the Gamma distribution when k = 1, while for k ! 1 the Gamma distribution converges to the Dirac delta function. Most importantly, for integer values of k the system (19- 25) is equivalent to a system of ordinary di erential equations with constant rates (see for example [38] and Appendix B). This result allows to obtain numerical solutions of the system of integral equations using a simple numerical scheme like Runge-Kutta. 3.3.2. Delayed Model A simple but realistic distribution for the latent or infectious periods is ob- tained in the limiting case of fixed periods when the survival functions are step functions, and therefore the probability density distributions are Dirac delta distri- butions, (sT ), (sT ), (sT ). In this limiting case, the integro-di erential he hi ve system obtained by di erentiation of the integral equations system (see model A.8-A.14 in the Appendix A) reduces to a system of di erential delayed equa- tions, 10 dH H s s = V (26) h i dt H dH H H (t T ) e s s he = V V (t T ) (27) h i h i he dt H H dH H (t T ) i s he = H (0)(t T ) + V (t T ) i h i he dt H H (t T T ) s he hi V (t T T ) (28) h i he hi dH H (t T T ) r s he hi = V (t T T ) (29) h i he hi dt H dV H s i =  V  V (30) v v s v s dt H dV H H (t T ) e i i ve v ve = V e V (t T )  V (31) v s v s ve v e dt H H dV H (t T ) i i ve v ve = e V (t T )  V (32) v s ve v i dt H where T , T , and T are the (fixed) latency and infectious periods of vectors and ve he hi hosts. As discussed above, vector’s infectious period is assumed exponentially distributed as we considered a constant vector mortality rate. Host mortality is disregarded and then all latent host become infectious (and then f = 1). However v ve only a fraction f = e of infected vectors survive the latency period becoming infectious. Therefore the basic reproduction number is given by 1 V (3) v ve R = T e (33) h v hi 3.4. Relationship between the Basic reproduction numbers, and the basic modi- fied model Suppose that we are studying a host-vector system for which there are esti- mations of the parameters as the mean latency and infectious periods. If latency periods are disregarded and we assume that all the periods are exponentially dis- tributed, we may use the basic model (4-8) for which the basic reproduction num- ber is 11 V h v (1) R = : (r +  ) H h h v However, a more realistic model should include the latency periods. Under the most common, but unrealistic, assumption of exponentially distributed periods the corresponding model is 10-16 and the basic reproduction number is (2) (1) v (1) R = R < R : 0 0 0 k + v v The delayed model (27 32) is a more realistic choice for which the basic re- production number is 1 V (3) v ve R = T e : h v hi v  T v ve Because  e , the basic reproduction numbers for the di erent mod- k + v v els satisfy (1) (2) (3) R > R > R : (34) 0 0 0 Therefore we expect larger and faster epidemics for the simple S IRS I model (Eqs. 4 - 8) than the obtained with the more realistic models. However, it is possible to implicitly include the e ect of latency in the vector population in the basic model (4-8) modifying the equation (8) as dV H i i v ve = e V  V (35) v s v i dt H (3) The basic reproduction number for the Basic modified model (Eq. 35) is R , the same as the most realistic Delayed model. 4. A stochastic agent based Ross-Macdonald model A stochastic version of an ordinary di erential equations model like (10-16) is straightforward. Consider, for example, the simple Ross model dH H i s = V r H h i h i dt H dV H i i = V  V v s v i dt H 12 In this case there are only four events: host infection, vector infection, human recovery, and vector death. The rates of the deterministic model define the prob- abilities of occurrence of the events per unit of time or transition rates. Thus, for example, probability of human infection in an interval t is given by P(H ! H ; t) = V t + o(t) s i h i o(t) where o(t) are higher order terms for which lim = 0. t!0 The interval between consecutive events is exponentially distributed with pa- rameter equal to the sum of the all transition rates. This kind of stochastic mod- els are markovian, the probability of occurrence of any event depend only of the present values of the variables. A stochastic version of the integral Volterra equations model like (19-25) is not that easy (see for example [27]) as the corresponding stochastic model is non- markovian, the dynamics depends of the history of the system. One alternative is to consider Gamma distributed periods with integer shape parameter values, for which the equivalent systems becomes markovian and therefore it is possible to use the stochastic simulation scheme outlined above. We preferred to develop an agent based model (ABM) for which the simula- tion of periods with arbitrary distributions is straightforward. Agent based models are a computational tool which allows to simulate populations dynamics consid- ering the features of each individual in the population and the interaction between them [22]. Agent based models are considered the most realistic models where some features, like the mobility of each individual, can be easily incorporated [12, 31]. Our model considers a SEIR model for the host population and a SEI model for the vector population. 4.1. Modeling disease transmission, progression and recovery. For each host and vector, the followings attributes were considered: The epidemiological status (State) which may take the values SUSCEPTI- BLE, EXPOSED, INFECTIOUS, RECOVERED. The age of infection (the time elapsed from first infection). The age of infectiousness (the time elapsed since progression to the infec- tious status). 13 We considered a fixed time step t. For each vector we generated a (pseudo)random number u with uniform distri- bution in the interval (0,1). If u < 1 ex p(bt) the vector bites a host selected at random which may be infected (if the vector is infectious and the host susceptible) or transmit the infection (if the vector is susceptible and the host infectious) with probabilities p and p respectively. h v Latent or infectious periods are random variables with some probability dis- tribution. We considered the general and flexible case where periods are Gamma distributed. Special cases of the Gamma distribution include exponential distribu- tion (for the shape parameter k = 1) and the limiting case of fixed period (k ! 1). Waiting periods in the di erent epidemiological classes were simulated in the following way. For each newly infected individual we drawn a pseudo random number from the corresponding Gamma distribution. This simulated value of the latency period ( ), plus the current time t, was stored in the variable T change. In all cases T change is a future time at which the individual will change the epidemiological status. When the age of infection becomes greater or equal to  we changed the agent’s state from EXPOSED to INFECTIOUS (in our implementation this is equivalent to the condition t  T change). In a similar way, for each newly infectious individual we drawn a value for the infectious period ( ) and stored it in T change (as before T change is the simulated infectious period plus the actual time t). When the age of infectiousness reach (or surpass) this value the state of the individual was changed from INFECTIOUS to RECOVERED. For the cases of fixed waiting times the transitions are deterministic and are determined by the values selected for the di erent (fixed) periods. Other distributions for the latent and infectious periods may be easily incor- porated as long a generator of random numbers for the corresponding distribution is available. 4.2. Modeling births and deaths. We disregarded host births and deaths. For the vector population we consid- ered a constant mortality ( = 1=T ) and a constant birth rate  . Thus, the v v v v t probability of a vector dying in a time interval t is equal to 1 e . On the other hand, the number of newborns vectors in a time step t was modelled by a Poisson random variable with parameter  t. 4.3. Simulation procedure The simulation procedure used is described in the following pseudo-code: 14 1. Initialization of variables and parameters (a) Set the host (H(0)) and vector (V (0)) population sizes, and the initial conditions H (0), H (0), H (0), H (0), V (0), V (0), V (0). s e i r s e i (b) Set the time step t, the simulation duration t and the current time t sim equal to 0. (c) Set the values of parameters  , p , p , k , k , , b. v v h v h h 2. While t  t and 0  H (t) + H (t) + V (t) + V (t) /* this last sentence sim e i e i interrupts the program when infections cannot takes place anymore */ (a) A random number of susceptible vector are added to the population according to a Poisson distribution with parameter  t (b) For each vector in the population i. A uniform random number in the interval (0; 1) is generated. ii. If the number is less than or equal to bt, the vector bites. The host bitten is chosen at random. If the vector is susceptible and the host bitten is infected A uniform random number in the interval (0; 1) is gener- ated. If the number is less than or equal to p , the mosquito be- comes exposed vector:S tate = EXPOSED, V (t) = V (t) 1, V (t) = V (t) + 1 s s e e Generate a latency period  according to the correspond- ing Gamma distribution. Set an exposed time vector:T change = t +  . If the vector is infected and the host bitten is susceptible A uniform random number in the interval (0; 1) is gener- ated. If the number is less than or equal to p , the host becomes exposed host:S tate = EXPOSED, H (t) = H (t)1, H (t) = H (t) + 1 s s e e Generate a latency period  according to the correspond- ing Gamma distribution. Set an exposed time host:T change = t +  . iii. A uniform random number in the interval (0; 1) is generated. v t iv. If the number is less than or equal to 1 e The vector dies and it is removed from vector population. v. Else 15 If the vector is exposed and vector:T change is less than or equal to the current time t vector:S tate = INFECTED, V (t) = V (t)1, V (t) = V (t) + 1 e e i i (c) For each host i. If the host is exposed and host:T change is less than or equal to the current time t host:S tate = INFECTED, H (t) = H (t) 1, H (t) = H (t) + 1 e e i i Generate a infectious period  according to the corresponding Gamma distribution. Set an infectious time vector:T change = t +  . ii. If the host is infected and host:T change is less than or equal to the current time t host:S tate = RECOVERED, H (t) = H (t)1, H (t) = H (t) + 1 i i r r 5. Some numerical results The simulations start with one host infectious, and all the other individuals susceptible. We used the day as the unit of time. 5.1. Vector to host ratio Host population was constant along the simulations. For the vector population we considered two cases, constant populations and seasonal varying populations. In the first case, the population may fluctuate stochastically about the deterministic equilibrium which was the initial population in our simulations. In the second the population presents seasonal oscillations around this value. The ratio vector per host was set in 1. This value vary from system to system and along the year in seasonal environments. For Aedes aegypty for example values of 0.5 and 1.1 per human were estimated [34, 19]. Studies in malaria estimated the number of anophelines per person in the range 3 to 4 depending on the location [25]. A typical household in high risk areas of T. cruzi transmission has an average of 5 human host while triatomine population may vary from some few individuals to several hundreds [24, 42]. 5.2. Parameter values Vectors life expectancy was set in 10 days, which is of the order of values obtained for most species of mosquito [14] and sandflies [10]. Kissing bugs has a longer life expectancy of about 5 months [32]. 16 Probabilities of transmission ( p , p ) are in general asymmetrical and there is h v a wide range of variation. For malaria those probabilities were estimated in the order of 0.48 and 0.022 [14], while for dengue are close to 1. In this work we used p = p = 0:75. hv vh Human mean latency period is of the order of one week for dengue and about 10 days for malaria [28], while in the case of Chagas this period is between 4 and 15 days [18]. In our simulations we considered a host mean latency period of 6 days. Host mean infectious period is about 3 days for dengue [30], 2 or 4 months for malaria [28]. In the case of Chagas, the acute phase is between 0 and 90 days, period in which the risk to transmit the infection is higher [13]. We used a value of 5 days in our simulations. Latent period for vectors was set in 7 days, a value about the observed in dengue infected mosquitoes. The biting rate b is a parameter harder to estimate. We considered two values, b = 0:3=day and b = 0:5=day, in order to have two cases, low and high basic reproduction numbers. This values are of the order observed for mosquitoes and were used for Ross and Macdonald in their seminal works [25]. 5.3. Epidemic curves In figure 1 we compare numerical solutions of the models for low and high values of the basic reproduction number (b = 0:3 and b = 0:5, respectively) considering the parameter values in the table 1. Parameter Value p 0.75 p 0.75 T 6 [days] he T 5 [days] hi T 7 [days] ve T 10 [days] Table 1: Parameter values used in the simulations. In all cases we set host mortality equal to zero ( =0) while  = 1=T =0.1 days . h v v In table 2 we show the corresponding R and some statistics of the epidemic curves (number of infected host at the epidemic peak, time at which the peak is reached and the final epidemic size as proportion of the total host population size) 2,500 2,250 2,000 600 1,750 1,500 1,250 1,000 0 500 1,000 0 100 200 300 Time Time Figure 1: Solutions of the deterministic models (Host infectious population). Left panel: low R (b = 0:3), right panel, high R (b = 0:5). From left to right: basic model (Eqs. 4 - 8), basic model modified (Eq. 35), SEIR-SEI model (Eqs. 10 - 16), delayed model (Eqs. 26 - 32). Time units in days. for each of the simulations presented in figure 1 corresponding to the di erent deterministic models. Similar results obtained with the agent based model are presented in table 3, where we considered only the case of exponentially distributed periods (corre- sponding to the SEIR-SEI model 10 - 16), and fixed periods (corresponding to the Delayed model 27 - 32). Because for the same parameter values the basic reproduction number of the basic model is greater than the basic reproduction numbers of the other models (see inequality 34), the basic model produces faster epidemics with a higher epi- demic final size. For b = 0:3 (low R ’s), the epidemic final size for the basic model is about two times the epidemic final size obtained with the other models, while the epidemic peak is almost 20 times higher than the obtained with the delayed model (see table 2). The modified basic model (35) has the same basic reproduction number than the delayed model and both models predict almost the same final epidemic sizes. However, the first one produces higher peaks in shorter times, resulting in a epi- demic that spreads through the population faster and runs out earlier. It is impor- tant to note that considering latency periods in hosts and vectors always produces lower epidemics, in comparison with the basic model. To compare the solution of the deterministic SEIR-SEI (Eqs. 10 - 16) and delayed models (Eqs. 26 - 32) with the ABM results, we realized simulations Hi i Model R Epidemic Peak Time Epidemic Final Size Basic model 2.53 759 113 0.86 Basic model modified 1.26 69.7 415 0.365 SEIR-SEI model 1.49 114.3 461.25 0.559 Delayed model 1.26 42.5 733.4 0.37 Model R Epidemic Peak Time Epidemic Final Size Basic model 7.03 2250 45 0.99 Basic model modified 3.50 1126 80 0.925 SEIR-SEI model 4.14 809 140 0.965 Delayed model 3.50 682 175 0.94 Table 2: Basic reproduction number, peak of the epidemic, duration from source case introduc- tion to peak and epidemic final size for the solutions of the di erent Ross-Macdonald models considered in figure 1 left panel (top) and right panel (bottom). following the procedure explained above considering the same parameter values (table 1), population sizes and initial conditions used with the deterministic mod- els. For b = 0:3 (Fig. 2 - left panel and Fig. 3 - left panel) stochasticity dominates the dynamics and realizations of the ABM are qualitatively and quantitatively di erent of the deterministic solutions (see table 3). Epidemic peak is lower in the deterministic case than in the mean value obtained with the ABM simulations. Also, the deterministic values of the epidemic peak and the time at which it is reached are not within the corresponding 95% confidence intervals. Not only individual realizations are qualitatively di erent from the deterministic solutions but the mean of those realizations do not converge to the deterministic results. This behaviour of the stochastic realizations is due to the fact that for low values of the basic reproduction number the stochastic e ects dominate the disease dynamics. In this case stochastic dynamics is not a deterministic drift with noise [4], and therefore we cannot expect that stochastic fluctuations average out. For higher values of R (b = 0:5) stochastic dynamics is quasi deterministic and the realizations of the ABM are similar in shape and size to the deterministic solutions (Fig. 2 - right panel and 3 - right panel). The stochasticity may produce a shift of the epidemic curve (to the left or to the right of the deterministic result), but it does not greatly a ect the height of the peak neither the amplitude of the epidemic curve. 19 20 Epidemic Peak Time Epidemic Final Size Type of model b Mean (95% CI) Mean (95% CI) Mean (95% CI) Exponential 0.3 144.64 (141.36, 147.92) 362.99 (352.1, 373.89) 0.5622 (0.5577, 0.5666) periods 0.5 842.88 (838.60, 847.16) 131.94 (129.80, 134.08) 0.9664 (0.9660, 0.9668) Fixed 0.3 69.48 (67.07, 71.89) 517.91 (497.24, 538.58) 0.3580 (0.3496, 0.3664) periods 0.5 714.5 (710.89, 718.22) 166.30 (163.59, 169.02) 0.9418 (0.9412, 0.9423) Table 3: Peak of the epidemic, duration from source case introduction to peak and epidemic final size for the solutions of the di erent ABM Ross-Macdonald models considering 200 simulations. 1,000 Ross det H_I H_I 800 H_I H_I 100 H_I H_I 0 200 400 600 800 1,000 0 100 200 300 Time Time Figure 2: Disease dynamics considering periods exponentially distributed and the parameters in table 1. In black the deterministic result, and in red the ABM simulation. Left panel: low R (b = 0.3); right panel, high R (b = 0.5). HI H_I H_I H_I 500 H_I H_I H_I 0 200 400 600 800 1,000 1,200 1,400 0 100 200 300 400 Time Time Figure 3: Disease dynamics considering fixed periods and the parameters in table 1. In black the deterministic result, and in red the ABM simulation. Left panel: low R (b = 0.3); right panel, high R (b = 0.5). 5.3.1. Fixed periods vs bell shaped distributed periods By far the most used distribution for the waiting times is the exponential distri- bution, which in the deterministic case leads to model 10-16. However, as already discussed, this is an unrealistic assumption for most cases. Latency and infec- tious periods are expected to have a bell shaped distribution. In the general case the model 19-25 should be used but numerical solutions are harder to obtain in this case. Bell shaped distributions may be modelled by the Gamma distribution, which with to independent parameters (the shape parameter k and the scale param- eter ) may control mean and variance (see 3.3.1). When a Gamma distribution is used for the probability density distribution of the waiting times, a very useful property of system 19-25 is that for integer values of the shape parameter, the Hi Hi Hi Hi system of integral equations is equivalent to a larger system of ordinary di eren- tial equations (see Appendix B). Thus, for k = 1, system B.1-B.13 reduces to the SEIR-SEI model 10-16, while for k ! 1 it converges to the delayed model 26-32. In figure 4 we show di erent Gamma distributions for di erent values of the shape parameter k. For k = 10 there is a high variability in the waiting periods but solutions are close to the solutions obtained with fixed periods (k = 1, see fig. 5). For k = 50 both cases are almost identical. 0.4 0.3 0.2 0.1 0 5 10 15 20 Time Figure 4: Probability density function of Gamma distribution considering  = 7 and di erent values of k: 1, 2, 5, 10 and 50. 5.4. E ects of seasonality Seasonality is a key driver of disease dynamics in most vector-borne diseases. Seasonality a ects vector population dynamics because, for example, vector’s ac- tivity is temperature dependent. In the present case we only consider that vector recruitment is a ected by seasonality (mostly by variations in rainfall). As an simple example we considered harmonic variations of the form =  [1 +  sin(!t)]: v 0 For  = 0 we recover the case of constant recruitment used in fig. 1. For 0 <   1 the vector population oscillates with frequency !. As we show in fig. 1, duration of epidemics may last almost two years. Sea- sonal variation of vector populations significantly reduces epidemic duration. In 0 50 100 150 200 250 300 Time [days] Figure 5: Solutions of deterministic models (Host infectious population) considering the param- eters in Table 1. Left panel: low R (b = 0:3), right panel, high R (b = 0:5). From left to right: 0 0 SEIR-SEI , model with gamma distribution for waiting periods (Eqs. B.1 - B.13) considering k = 1, 2, 5, 10 and 50, and delayed model (Eqs. 26 - 32). For k = 1 model (Eqs. B.1 - B.13 reduces to model (Eqs. 10 - 16). fig. 6 we compare numerical solutions of the delayed model (Eqs. 26 - 32) for dif- ferent values of . Duration of epidemics range between approximately 10 months to a couple of months for  = 0:5 and  = 1. 1 4 0 0 6 0 1 2 0 0 5 0 1 0 0 0 4 0 8 0 0 3 0 6 0 0 2 0 4 0 0 1 0 2 0 0 0 0 0 2 0 0 4 0 0 6 0 0 8 0 0 1 0 0 0 1 2 0 0 1 4 0 0 1 6 0 0 1 8 0 0 0 3 0 6 0 9 0 1 2 0 1 5 0 1 8 0 2 1 0 2 4 0 2 7 0 3 0 0 T i m e T i m e Figure 6: Numerical solutions of the delayed model (Eqs. 26 - 32) for  = 0 (no seasonality, black line),  = 0:5 (red) and  = 1 (blue), for b = 0:3 (left panel) and b = 0:5 (right panel). Fig. 7 shows numerical solutions of the ABM considering fixed periods and seasonality. We can see that in the case in which R is higher (b = 0:5) the ABM results are similar to the deterministic model. Conversely, considering b = 0:3, the curves obtained from the AMB are qualitatively similar to the deterministic case, Hi since the same amount of peaks can be clearly observed in both cases. However, the values reached in these peaks in the case of ABM are higher, in general, in the first and second one and lower in the third one. 120 1,200 1,000 80 800 60 600 40 400 0 0 0 500 1,000 1,500 0 50 100 150 200 250 300 350 Time Time Figure 7: Numerical solutions of the delayed model (Eqs. 26 - 32) with  = 0:5 (red) and the corresponding ABM considering fixed periods (other colours curves), for b = 0:3 (left panel) and b = 0:5 (right panel). In black line the solution of the delayed model (Eqs. 26 - 32) without seasonality ( = 0). 5.5. Computing R from the agent based model To compute R in the case of the agent based model, we have to follow the infectious generation of hosts and vectors. So, the procedure realized is as follow. The first infected host is the only host of first infected generation. The vectors infected by a host of first generation, are vectors of first infected generation. When a vector of first infected generation, infects a susceptible hosts, these host are second infected generation. In general, when a host of infected generation m infects a vector, the infected generation of the vector is m. Then, when a vector of infected generation m infects a host, then infected generation of the host is m + 1. Let H be the number of infected-host generation m. Then, R can be esti- m 0 mated as R  H =H [5]. Due to the stochasticity of the ABM simulations, it is 0 3 2 important to realize a considerable number of simulations and then calculate the mean of the R value estimated for each simulation. An R estimation considering 0 0 200 realizations of the simulations analyzed en the previous section is presented in the table 4. In all the cases the deterministic value of R is within the 95% confidence interval. As we can see in Eqs. 17 and 3.4, given the parameters of the host and vector populations, R is a linear function of the relation V=H. So, varying the relation V=H, we can obtain di erent values of R . Hi Hi Deterministic Estimation R  H =H 0 3 2 Type of model b R Mean (95% CI) Exponential 0.3 1.25 1.28 (1.01, 1.54) periods 0.5 3.49 3.60 (3.16, 4.05) Fixed 0.3 1.49 1.45 (1.21, 1.69) periods 0.5 4.14 4.03 (3.60, 4.46) Table 4: Estimation of R from the ABM model considering 200 simulations. Considering the parameters in the table 1 and a biting rate equal to 0.3 (b = 0:3), we estimated the value of R from the agent based model for di erent values of V=H. The results considering exponentially distributed periods and fixed period are shown in the Fig. 8, respectively. In all the cases, an initial population of 10000 hosts was considered, with only one infected host. Each estimation of the basic reproduction number was realized with 200 simulations. 0 . 6 0 . 8 1 . 0 1 . 2 1 . 4 1 . 6 1 . 8 0 . 8 1 . 0 1 . 2 1 . 4 1 . 6 1 . 8 2 . 0 V / H V / H Figure 8: Empirical estimates of the Basic reproduction numbers (squares, bars are 95% confi- dence interval) obtained with the ABM for the cases of exponentially distributed periods (left) and fixed period (right). Continuous line are the corresponding theoretical values given by expressions (17) and (3.4). As can be seen in the figure 8 numerical estimations of the basic reproductive number are in the 95% confidence interval estimated from the 200 ABM simula- tions. 6. Discussion and Conclusions The use of mathematical models in epidemiology has a long and fruitful tra- dition. However di erent hypotheses about the systems under study may lead to, 25 in some cases, significant di erent results (see for example [1, 45, 46]). In this work we present di erent formulations of the Ross-Macdonald model using ordinary di erential equations, Volterra integral equations and agent based modelling. In the most general case we included latency periods in both vectors and hosts. We also considered general distributions for latency and infectious periods including two simple cases: exponentially distributed periods and fixed periods. As we show in this work, disregarding latency periods has a dramatic e ect in the dynamics. This is quite apparent for low basic reproduction numbers (see Fig. 1). As in most vector-borne diseases vector’s latency periods and life ex- pectancy are of the same order of magnitude, disregarding latency overestimate the basic reproduction number, and therefore we observed faster epidemics with significantly higher peaks. A substantial improvement is achieved with the simple modification (35) which produces the same values of R as the delayed model but still the epidemic curves are significantly di erent. Not only the inclusion of latency periods is important but also its distributions. Using exponentially distributed periods leads to slightly smaller basic reproduc- tion numbers and still a noticeable di erences in the epidemic curves. For the most realistic case of bell-shaped distributed waiting periods we show that numerical solutions of the Volterra integral system 19-25 are close to the so- lutions of the simpler delayed model 26-32, for which numerical simulations may be easily obtained using a Runge-Kutta scheme, for example. For us, the delayed model is therefore the model of choice as it combines realism and simplicity. A central assumption of the Ross-Macdonald models is homogeneous random mixing: probability of biting in a susceptible host is proportional to the frac- tion of susceptible host in the entire population. This hypothesis may hold for some local, relatively small, populations. Larger populations may be modelled using a meta-population approach, for example. If local populations have some degree of synchronization, the total population disease dynamics could be quasi- deterministic (see for example [21]), and perhaps a Ross-Macdonald model may describe the global dynamics of the system. In this work we considered popula- tions of 10 individuals, a large enough population for which it is not obvious that the assumption of homogeneous mixing holds. For both, high and low values of the basic reproduction number, solutions of the deterministic models and the estimation of the R of the agent based model are statistical similar (see fig. 8 and table 4), although the epidemic curve may be significantly di erent from the deterministic solution, especially for low R value (see fig. 2 and 3). 26 Deterministic models, like the SEIR-SEI model 10-16, are simple ordinary di erential equations systems with constant parameters, more amenable for anal- ysis. Numerical integration is straightforward using Runge-Kutta of fourth order, for example. The more realistic choice of fixed periods is modelled by delayed di erential equations. Analysis is more complex for these type of models but numerical integration is easily implemented too. For the agent based model there are not di erences, neither in the diculty of the coding or in the computational cost for both cases, and therefore non- exponentially distributed periods (like fixed periods) is the recommended choice. In our simulations we considered parameter values compatible with some vector-borne diseases in humans like dengue. In all cases the number of vec- tors per host was set equal to one at demographic equilibrium. For low values of the basic reproduction number epidemics obtained with the (most realistic) fixed period models have a duration of more than two years (see Fig. 3, left panel), which is never observed in real epidemics. This results highlights the importance of including seasonality when modelling some vector-borne diseases. Vector pop- ulations usually have seasonal fluctuations, driven by rainfall, for example, which shape the duration of the epidemics (see fig. 6). However, outbreaks sizes are gen- erally a function not only of the abundance of vectors, but also of other variables such as climate, community immunity, host mobility, among others. Vector activ- ity is strongly a ected by temperature and therefore not only seasonal variations are significant but also habitat conditions as the use of air conditioning. The two values of  used correspond to moderate seasonal variations in vector abundance ( = 1=2) as expected in endemic settings, and to marked variations in vector abundances as observed in non-endemic populations. As the homogeneous mixing assumption is expected to hold only for relatively small populations, stochasticity should be considered when modelling such cases. Larger populations may be modelled using a metapopulation approach, something we will explore in forthcoming works. Acknowledgements This work was partially supported by grants CIUNSA 2018-2467 and PICT 2014-2476. JPA is a member of the CONICET. MIS is a postdoctoral fellow of CONICET. References 27 [1] Amaku, M., Azevedo, F., Burattini, M. N., Coutinho, F. A. B., Lopez, L. F., & Massad, E. (2015). Interpretations and pitfalls in modelling vector- transmitted infections. Epidemiology & Infection, 143(9), 1803-1815. [2] Amaku, M., Azevedo, F., Burattini, M. N., Coelho, G. E., Coutinho, F. A. B., Greenhalgh, D., Lopez, L. F., Motitsuki, R. S., Wilder-Smith, A., & Mas- sad, E. (2016). Magnitude and frequency variations of vector-borne infection outbreaks using the Ross-Macdonald model: explaining and predicting out- breaks of dengue fever. Epidemiology & Infection, 144(16), 3435-3450. [3] Anaguano, D. F., Ponce, P., Baldeon, ´ M. E., Santander, S., & Cevallos, V. (2015). Blood-meal identification in phlebotomine sand flies (Diptera: Psy- chodidae) from Valle Hermoso, a high prevalence zone for cutaneous leish- maniasis in Ecuador. Acta tropica, 152, 116-120. [4] Aparicio, J. P., & Solari, H. G. (2001). Population dynamics: Poisson ap- proximation and its relation to the langevin process. Physical Review Let- ters, 86(18), 4183. [5] Aparicio, J. P., & Pascual, M. (2007). Building epidemiological models from R 0: an implicit treatment of transmission in networks. Proceedings of the Royal Society B: Biological Sciences, 274(1609), 505-512. [6] Auger, P., Kouokam, E., Sallet, G., Tchuente, M., & Tsanou, B. (2008). The Ross-Macdonald model in a patchy environment. Mathematical biosciences, 216(2), 123-131. [7] Bacaer ¨ , N., & Guernaoui, S. (2006). The epidemic threshold of vector-borne diseases with seasonality. Journal of mathematical biology, 53(3), 421-436. [8] Bacaer ¨ , N. (2007). Approximation of the basic reproduction number R0 for vector-borne diseases with a periodic vector population. Bulletin of mathe- matical biology, 69(3), 1067-1091. [9] Bates, P. A. (2007). Transmission of Leishmania metacyclic promastigotes by phlebotomine sand flies. International journal for parasitology, 37(10), 1097-1106. [10] Belen, A., & Alten, B. (2006). Variation in life table characteristics among populations of Phlebotomus papatasi at di erent altitudes. Journal of Vector Ecology, 31(1), 35-44. 28 [11] Benelli, G., & Mehlhorn, H. (2016). Declining malaria, rising of dengue and Zika virus: insights for mosquito vector control. Parasitology research, 115(5), 1747-1754. [12] Bian, L. (2004). A conceptual framework for an individual-based spatially explicit epidemiological model. Environment and Planning B: Planning and Design, 31(3), 381-385. [13] Castanera, ˜ M. B., Aparicio, J. P., & Gurtler ¨ , R. E. (2003). A stage-structured stochastic model of the population dynamics of Triatoma infestans, the main vector of Chagas disease. Ecological modelling, 162(1-2), 33-53. [14] Chitnis, N., Hyman, J. M., & Cushing, J. M. (2008). Determining impor- tant parameters in the spread of malaria through the sensitivity analysis of a mathematical model. Bulletin of mathematical biology, 70(5), 1272. [15] Diekmann, O., & Heesterbeek, J. A. P. (2000). Mathematical epidemiology of infectious diseases: model building, analysis and interpretation (Vol. 5). John Wiley & Sons. [16] Dietz, K., Molineaux, L., & Thomas, A. (1974). A malaria model tested in the African savannah. Bulletin of the World Health Organization, 50(3-4), [17] Feng, Z., Xu, D., & Zhao, H. (2007). Epidemiological models with non- exponentially distributed disease stages and applications to disease control. Bulletin of mathematical biology, 69(5), 1511-1536. [18] Filigheddu, M. T., Gor ´ golas, M., & Ramos, J. M. (2017). Enfermedad de Chagas de transmision ´ oral. Medicina cl´ ınica, 148(3), 125-131. [19] Focks, D. A., Brenner, R. J., Hayes, J., & Daniels, E. (2000). Transmission thresholds for dengue in terms of Aedes aegypti pupae per person with dis- cussion of their utility in source reduction e orts. The American journal of tropical medicine and hygiene, 62(1), 11-18. [20] Gardner, L., Chen, N., & Sarkar, S. (2017). Vector status of Aedes species determines geographical risk of autochthonous Zika virus establishment. PLoS neglected tropical diseases, 11(3), e0005487. 29 [21] Gutierrez, J. A., & Aparicio, J. P. (2015). Quasi-deterministic popula- tion dynamics in stochastic coupled maps. Journal of Biological Systems, 23(supp01), S151-S162. [22] Nepomuceno, E.G., Takahashi, R. H. C., & Aguirre, L. A. (2016). Individual-based model (IBM): an alternative framework for epidemiolog- ical compartment models. Revista Brasileira de Biometria, 34(1), 133-162. [23] Noireau, F., Diosque, P., & Jansen, A. M. (2009). Trypanosoma cruzi: adap- tation to its vectors and its hosts. Veterinary research, 40(2), 1-23. [24] Nouvellet, P., Dumonteil, E., & GourbiA¨re, S. (2013). The improbable transmission of Trypanosoma cruzi to human: the missing link in the dynam- ics and control of Chagas disease. PLoS neglected tropical diseases, 7(11), e2505. [25] Macdonald, G. (1955). The measurement of malaria transmission. Proceed- ings of the Royal Society of Medicine, 295-302. [26] Mandal, S., Sarkar, R. R., & Sinha, S. (2011). Mathematical models of malaria-a review. Malaria journal, 10(1), 202. [27] Mohammadi, F. (2016). A computational wavelet method for numerical so- lution of stochastic Volterra-Fredholm integral equations. Wavelet and Lin- ear Algebra, 3(1), 13-25. [28] Molineaux, L., & Gramiccia, G. (1980). The Garki project: research on the epidemiology and control of malaria in the Sudan savanna of West Africa. World Health Organization. [29] O’Regan, S. M., Lillie, J. W., & Drake, J. M. (2016). Leading indicators of mosquito-borne disease elimination. Theoretical ecology, 9(3), 269-286. [30] Otero, M., & Solari, H. G. (2010). Stochastic eco-epidemiological model of dengue disease transmission by Aedes aegypti mosquito. Mathematical biosciences, 223(1), 32-46. [31] Otero, M., Barmak, D. H., Dorso, C. O., Solari, H. G., & Natiello, M. A. (2011). Modeling dengue outbreaks. Mathematical biosciences, 232(2), 87- 30 [32] Rabinovich, J. E. (1972). Vital statistics of Triatominae (Hemiptera: Redu- viidae) under laboratory conditions. I. Triatoma infestans Klug. Journal of Medical Entomology, 9(4), 351-370. [33] Reiner Jr, R.C., Perkins, T.A., Barker, C.M., Niu, T., Chaves, L.F., Ellis, A.M., George, D.B., Le Menach, A., Pulliam, J.R., Bisanzio, D. and Buckee, C., 2013. A systematic review of mathematical models of mosquito-borne pathogen transmission: 1970 - 2010. Journal of The Royal Society Interface, 10(81), p.20120921. [34] Romero-Vivas, C. M., & Falconar, A. K. (2005). Investigation of relation- ships between Aedes aegypti egg, larvae, pupae, and adult density indices where their main breeding sites were located indoors. Journal of the Ameri- can Mosquito Control Association, 21(1), 15-22. [35] Ross, R. (1911). Some Quantitative Studies in Epidemiology. 87, 466 - 467. [36] Sanchez , F., & Calvo, J. G. (2020). Dengue model with early-life stage of ˜ ˜ vectors and age-structure within host. Revista de MatemA¡tica: TeorAa y Aplicaciones, 27(1), 157-177. [37] Schofield, C. J., & Dujardin, J. P. (1997). Chagas disease vector control in Central America. Parasitology Today, 13(4), 141-144. [38] Smith, H. (2011). Distributed delay equations and the linear chain trick. In An Introduction to Delay Di erential Equations with Applications to the Life Sciences (pp. 119-130). Springer, New York, NY. [39] Smith, D. L., & McKenzie, F. E. (2004). Statics and dynamics of malaria infection in Anopheles mosquitoes. Malaria journal, 3(1), 13. [40] Smith, D. L., Battle, K. E., Hay, S. I., Barker, C. M., Scott, T. W., & McKen- zie, F. E. (2012). Ross, Macdonald, and a theory for the dynamics and control of mosquito-transmitted pathogens. PLoS pathogens, 8(4), e1002588. [41] Snow, R. W. (2015). Global malaria eradication and the importance of Plas- modium falciparum epidemiology in Africa. BMC medicine, 13(1), 23. [42] Tomasini, N., Ragone, P. G., Gourbiere, ` S., Aparicio, J. P., & Diosque, P. (2017). Epidemiological modeling of Trypanosoma cruzi: Low stercorarian transmission and failure of host adaptive immunity explain the frequency of mixed infections in humans. PLoS computational biology, 13(5), e1005532. 31 [43] Velasco-Hernandez, ´ J. X. (1991). An epidemiological model for the dynam- ics of Chagas’ disease. Biosystems, 26(2), 127-134. [44] Wilder-Smith, A., & Massad, E. (2018). Estimating the number of unvac- cinated Chinese workers against yellow fever in Angola. BMC infectious diseases, 18(1), 185. [45] Wonham, M. J., Lewis, M. A., Renclawowicz, J., & Van den Driessche, P. (2006). Transmission assumptions generate conflicting predictions in host- vector disease models: a case study in West Nile virus. Ecology letters, 9(6), 706-725. [46] Wonham, M. J., & Lewis, M. A. (2008). A comparative analysis of models for West Nile virus. In Mathematical epidemiology (pp. 365-390). Springer, Berlin, Heidelberg. [47] World Health Organization. (2017). Vector-borne diseases. Retrieved from: https://www.who.int/news-room/fact-sheets/detail/vector-borne-diseases (Last access: September 25, 2019). 32 Appendix A. General SEIR-SEI model with arbitrary distributions for the waiting time We will consider the general case of a SEIR-SEI model. For the host and vector populations we assume that the latency period (T for hosts and T for he ve vectors) and the infectious period for hosts (T ) are random variables with prob- hi ability density distributions f (s), f and f (s), respectively. The cumulative he ve hi distributions are denoted by F (s), F and F (s), respectively. The complemen- he ve hi tary cumulative distribution, F (s) = 1 F (s), is known as the survival function and gives the probability that an individual infected in t = 0 remains infected (or exposed, depending of the case) at time s. Therefore the evolution of the populations can be described by the integral Volterra equations [17]: H (s) H (t) = H (0) V (s) d s (A.1) s s h i H (s) ¯ ¯ H (t) = H (0)F (t) + V (s) F (t s)d s (A.2) e e he h i he Z Z " # H (s) dF s he ¯ ¯ H (t) = H (0)F (t) + V (s) ( s) F (t ) d sd (A.3) i i hi h i hi H dt 0 0 H (t) = H H (t) H (t) H (t) (A.4) r s i e Z Z t t H (s) t (ts) V (t) = V (0)e +  e d s V (s) d s (A.5) s s v v s 0 0 H (s) t (ts) V (t) = V (0)e + V (s) F (t s)e d s (A.6) e e v s ve " # Z Z H (s) dF i ve t (ts) V (t) = V (0)e + V (s) ( s) e d sd: (A.7) i i v s H dt 0 0 Di erentiation of these equations leads to the following system of integro-di erential equations, 33 dH H (t) s s = V (t) (A.8) h i dt H dH H (t) H (s) e s s = H (0) f (t) + V (t) V (s) f (t s)d s (A.9) e he h i h i he dt H H dH H (s) i s = H (0) f (t) + V (s) f (t s)d s i hi h i he dt H Z Z ! H (s) V (s) f ( s)d s f (t )d (A.10) h i he hi 0 0 dH = H (0) f (t) + H (0) f (t) e he i hi dt Z Z H (s) + V (s) f ( s)d s f (t )d (A.11) h i he hi 0 0 dV H (t) s i =  V (t) V (t) (A.12) v v s s dt H dV H (t) e i = e V (0) + V (t) e v s dt H H (s) (ts) V (s) f (t s)e d s V (t) (A.13) v s ve s dV H (s) i i t (ts) = e V (0) + V (s) f (t s)e d s  V (t) (A.14) i v s ve v i dt H Appendix B. Model with Gamma distributions for the waiting periods If the waiting periods are Gamma distributed with mean  and integer shape parameter k, we can apply the linear trick [38] to solve the system of integro- di erential di erential equations (A.8-A.14). Suppose that the latency and infectious periods for hosts and the latency period for vectors are all Gamma distributed with means  and integer shape parameters k (i = 1; 2; 3 respectively). Then it is possible to divide the exposed human population in k compartments such that H = H . In a similar fashion, the 1 e e; j j=1 H and V populations may be divided in k and k clases respectively. i e 2 3 Then, the system of integro-di erential equations is equivalent to the follow- ing system of ordinary di erential equations, 34 dH H s s =  V  H (B.1) h h i h s dt H dH H k e;1 s 1 = V H  H (B.2) h i e;1 h e;1 dt H dH k e;2 = (H H )  H (B.3) e;1 e;2 h e;2 dt dH k e;k = (H H )  H (B.4) e;k 1 e;k h e;k 1 1 1 dt dH k k i;1 1 2 = H H  H (B.5) e;k i;1 h i;1 dt 1 2 dH k i;2 2 = (H H )  H (B.6) i;1 i;2 h i;2 dt dH k i;k = (H H )  H (B.7) i;k 1 i;k h i;k 2 2 2 dt dH k r 2 = H  H (B.8) i;k h r dt dV H s i =  V  V (B.9) v v s v s dt H dV H k e;1 i 3 = V V  V (B.10) h s e;1 v e;1 dt H dV k e;2 = (V V )  V (B.11) e;1 e;2 v e;2 dt dV k e;k = (V V )  V (B.12) e;k 1 e;k v e;k 3 3 3 dt dV k i 3 = V 35 V (B.13) e;k v i dt 3 k k k 1 2 3 X X X where H = H ; H = H , and V = V e e; j i i; j e e; j j=1 j=1 j=1 The Gamma distribution with shape parameter k = 1 is an exponential distri- bution, while for k ! 1, the Gamma distribution tends to a Dirac delta distribu- tion. Thus, as k increases from 1 to 1, the model (Eqs. B.1 - B.13) moves from the SEIR-SEI model (Eqs. 10 - 16) to the delayed model (Eqs. 26 - 32). http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Mathematics arXiv (Cornell University)

Ross-Macdonald Models: Which one should we use?

Mathematics , Volume 2020 (2002) – Feb 26, 2020

Loading next page...
 
/lp/arxiv-cornell-university/ross-macdonald-models-which-one-should-we-use-DvFvf4qjTQ
ISSN
0001-706X
eISSN
ARCH-3343
DOI
10.1016/j.actatropica.2020.105452
Publisher site
See Article on Publisher Site

Abstract

Ross-Macdonald models are the building blocks of most vector-borne disease models. Even for the same disease, di erent authors use di erent model formula- tions, but a study of the dynamical consequences of assuming di erent hypothe- ses is missing. In this work we present di erent formulations of the basic Ross- Macdonald model together with a careful discussion of the assumptions behind each model. The most general model presented is an agent based model for which arbitrary distributions for latency and infectious periods for both, host and vectors, is considered. At population level we also developed a deterministic Volterra inte- gral equations model for which also arbitrary distributions in the waiting times are included. We compare the model solutions using di erent distributions for the in- fectious and latency periods using statistics, like the epidemic peak, or epidemic final size, to characterize the epidemic curves. The basic reproduction number (R ) for each formulation is computed and compared with empirical estimations obtained with the agent based models. The importance of considering realistic distributions for the latent and infectious periods is highlighted and discussed. We also show that seasonality is a key driver of vector-borne disease dynamics shaping the epidemic curve and its duration. Keywords: Ross-Macdonald Model, Epidemiology, Delayed Model, Agent Based Model 1. Introduction Preprint submitted to Acta Tropica arXiv:2002.11267v1 [q-bio.PE] 26 Feb 2020 Vector-borne diseases are caused by di erent types of parasites, including viruses and bacteria, which are transmitted by vectors as mosquitoes, sandflies, ticks, and kissing-bugs, among others. According to the World Health Organi- zation, every year there are more than 700 thousand deaths as a consequence of vector-borne diseases [47]. Mosquito-borne diseases of humans include malaria, dengue, zika, chykun- gunya, yellow fever (see for example [41, 11, 20]). Di erent triatomine species transmit Trypanosoma cruzi the causal agent of Chagas disease (see for example [37] and references there in) while Leishmaniasis is transmitted by several species of sandflies [9]. Vector-borne diseases are also common zoonotic diseases. Some forms of Leishmaniasis cannot be transmitted from humans to sandflies and the parasite population survives in a wild cycle including small rodents, dogs, cows and sev- eral species of birds as hosts [3, 9]. West Nile Virus may be transmitted to humans but it is maintained in a cycle which includes several species of birds [46]. Try- panosoma cruzi, the causal agent of Chagas disease is also transmitted to di erent animals including dogs, marsupials, rodents, and others hosts [23]. Ross model was published in 1911 [35] and remains as the basis of countless models for vector-borne diseases. Ross considered a simple model for malaria, with births and deaths but with constant populations and infectious periods expo- nentially distributed. Humans and mosquitoes may be in only two classes: Af- fected and Una ected (what here we will denoted by H ; H ; V ; V ). Then, Ross i s i s model in continuous time reads dH V i i = m (H H ) r H h i h i dt V dV H i i = (V V i)  V v v i dt H where H and V are the numbers of humans and mosquitoes, m is the number of mosquitoes per human (V=H), r is the recovery rate for humans,  is the mor- h v tality rate for mosquitoes, and are the transmission parameters which may be decomposed as = b f p with b the mosquitoes biting rate, f the proportion of j j bites in humans, and p the probability of transmission per bite. Ross formulation is still used but it is not advisable. The parameter m is in fact a dynamical variable. For the original Ross model this was not a problem as he considered constant pop- ulations. However, both vector and host populations may vary in time, therefore, 2 an equivalent formulation, more frequently used, and preferable is dH H i s = V r H (1) h i h i dt H dV H i i = V  V (2) v s v i dt H Models with these rates of infection are broadly known as Ross-Macdonald models, albeit Macdonald’s contribution to the Ross model is not reflected in this model formulation. Macdonald modified the original Ross model, integrating bi- ological information about the mosquito latency period, and introduced the ex- posed class for vectors [26]. Later he considered also the case of super-infection in Malaria disease dynamics (see for an extensive discussion [40]). For the Ross model, the basic reproduction number (R ), defined as the number of secondary host cases produced by a typical infectious host in a completely susceptible population is h v R = r  H h v This celebrated result from Ross [35] shows that the basic reproduction num- ber is proportional to the number of vectors per host (V=H), and therefore, disease transmission may be interrupted if the number of vectors per host is reduced below some threshold. Since the pioneering work of Ross, several extensions of his basic model (Eqs. 1 - 2) were developed including the addition of exposed classes, superinfection, spatiality, time-varying populations, age structure and more (see for example [16, 8, 6, 26, 40, 29, 36]), and applied to the study of di erent infectious diseases such as malaria, dengue, yellow fever, cutaneous Leishmaniasis, Chagas disease, West Nile virus, among others (see for example [44, 39, 26, 2, 7, 43, 46] and references therein). In this work we present a detailed analysis of some general Ross-Macdonald models. We show that the inclusion of exposed classes as well as the distribution of the latent and infectious periods, have significant dynamical consequences. We also show that seasonality is a major factor shapening the epidemics curves. This paper is arranged as follow. In the next section we discuss the general assumptions common of all models presented. In Section 3 several deterministic Ross-Macdonald models are developed considering exposed classes and di erent distributions for the waiting periods. The basic reproduction number is computed in each case. A stochastic agent based model (ABM) is developed in Section 3 4. Numerical results, such as epidemic curves, epidemic final sizes, the basic reproduction number are computed for each model and compared between them in Section 5. The key role of seasonality is also discussed. Finally present the discussion of the results and conclusions. 2. General assumptions and parameters In a Ross-Macdonald model it is assumed that populations are homogeneously mixed. Vector’s bites are evenly divided among hosts, that is, every time a vector bites, chooses a host at random. This hypothesis leads to a frequency dependent transmissions terms proportional to V H =H and V H =H. This central hypothesis i s s i is perhaps what define what a Ross-Macdonald model is. However we want to stress that this assumption is only realistic for small populations like a household. The use of Ross-Macdonald type of models for larger populations will be analyzed elsewhere. Demography. Immigration and emigration are not considered as we are in- terested in the simplest cases. Births are assumed to take place at a (density- independent) rate . Deaths may be described by the mortality or by the survival function. Mortality () is the number of deaths per individual and per unit of time. In general it is an age-dependent rate. The survival function, F(a), is the proportion of individuals still alive at age a, and it is related with the mortality by (s)d s F(a) = 1 e . Epidemiology. Populations are divided in some of the following epidemiolog- ical classes: Susceptible, Latent, Infectious, and Recovered. Latent (or Exposed) individuals are infected but not infectious (and therefore are unable to transmit the disease). Recovered individuals are immune, and therefore do not participate of the transmission process. Duration of the latent period may be described for a survival function of the age of infection, F (s) which gives the proportion of latent individuals who remain latent at age of infection s (age of infection is the time elapsed since first infection). Analogously, F (s) is the proportion of infec- tious individuals who remain infectious after a time s after the end of latency. Alternatively we can use the, (in general) age-of-infection dependent, progression rates (from latency to infectiousness) or recovery rates (from infectiousness to recovery). All the periods considered (lifespan, latency period, infectious period) are ran- dom variables which may be characterized by a probability distribution. The sim- ple, and commonly used case of exponentially distributed periods correspond to 4 constant, age independent, rates. For example using a constant mortality rate imply the assumption of an exponentially distributed lifespan. Parameters defining the di erent periods distributions are: T : Host life expectancy (mean lifespan) T =T : Vector life expectancy, mean infectious period for vectors v vi T : Mean latency period for exposed hosts he T : Mean infectious period for hosts hi T : Mean latency period for vectors ve For the limiting cases of exponentially distributed or fixed periods these param- eters values completely define the probability distributions. In the general case other parameters like the variance of the distribution should be provided. In all cases we considered that vectors are infectious for life. Entomological parameters. Biting rate on hosts (number of bites per vec- tor, per unit of time, on hosts) is denoted by b. Probabilities of transmission per bite are p and p (from vectors to hosts and from hosts to vectors respectively). h v Finally we define = p b, and = p b. h h v v Basic reproduction numbers. For a general Ross-Macdonald model the ba- sic reproduction number may be obtained by simple bookkeeping [15]. One in- fectious host will produce an average of V infected vectors per unit of time. If the mean infectious period for hosts is T , then the total number of infected hi vectors is V T . Only a fraction f will survive the latency period, and there- v hi v fore, the total number of infectious vectors produced by the initial infectious host is V T f . Each infectious vector would produce T host infections (T is v hi v h vi vi the mean infectious period for vectors) and only a fraction f will survive the host latency period. Finally the basic reproduction number is given by R = T T f f (3) 0 h v hi vi h v 3. Deterministic Ross-Macdonald models In a Ross-Macdonald model there are host and vector populations (of size H and V respectively) homogeneously mixed. Each population is subdivided in epidemiological classes. For example, susceptible and infectious host and vector populations (H , H , V V ). Vectors bite at the rate b (daily number of bites per s i s i 5 vector, for example). If p is the probability of infection transmission to hosts per bite, p the probability of vector infection per bite on infectious hosts, then, the rate of infection of susceptible hosts is given by p bV while the rate of infection h i of susceptible vectors by p bV . These functional forms for the infection rates v s are characteristic of all the Ross-Macdonald type models. In the following we will present, discuss and compare the more common deterministic models (without age structure). 3.1. Basic Model One of the most simple, general, and used model is the S IR model for hosts and a S I model for vectors. Mortalities are denoted by  while recovery rates by r. ’s are the recruitment rates. We will assume that all the periods are exponentially distributed and therefore we obtain the following Basic model: dH H s s =  V  H (4) h h i h s dt H dH H i s = V (r +  )H (5) h i h h i dt H dH = r H  H (6) h i h r dt dV H s i =  V  V (7) v v s v s dt H dV H i i = V  V (8) v s v i dt H where  = 1=T and  = 1=T . Mean infectious period for host includes recov- h h v v ery and mortality, and therefore in this case is given by T = 1=(r +  ), from hi h h where recovery rate r can be estimated. In this work we will consider only the case  = 0, but for many species of hosts,   r and therefore we may approxi- h h mate the recovery rate by 1=T . Vectors are assumed to be infectious for life and then  = 1=T = 1=T . v vi v Because in this model there are not latency periods, f = f = 1 and the basic h v reproduction number (Eq. 3) becomes h v (1) R = (9) (r +  ) H h h v 6 The assumption of constant mortality for vectors is plausible as for insects we expect an approximately constant daily probability of death. For hosts like birds, constant mortality is also usually observed. However hosts like humans present a survival of type I: low mortality for ages below the mean followed by a steep decrease in survival. In this case an age structured model for the host population should be used (see for example [36]). However in those cases we have that and therefore we may disregard birth and deaths in the host population h v when studying the short-term dynamics like in a single outbreak, the case we are considering in this work. Infectious period is also assumed exponentially distributed, a not realistic as- sumption. Hosts may lose immunity becoming susceptible again, a case we do not consider in this work. 3.2. Basic Model with exposed classes For both, hosts and vectors, there are latent periods and therefore a more re- alistic model is a S E IR model for hosts and a S E I model for vectors (as in most cases vectors are infectious for life). The basic model with latent classes (SEIR- SEI model) is: dH H s s =  V  H (10) h h i h s dt H dH H e s = V (k +  )H (11) h i h h e dt H dH = k H (r +  )H (12) h e h h i dt dH = r H  H (13) h i h r dt dV H s i =  V  V (14) v v s v s dt H dV H e i = V (k +  )V (15) v s v v e dt H dV = k V  V (16) v e v i dt Here, k and k are the progression rates from latency to infectiousness, and in h v this context are given by k = 1=T with T the mean latency periods ( j = h for j je je hosts, and j = v for vectors). 7 In this case the basic reproduction number is ! ! k k V h v v h (2) R = (17) (r +  ) k +  k +  H h h v v v h h where f = k =(k +  ) are the fractions of exposed individuals who survive the j j j j latency period. The assumptions in this model are the same discussed above but here it is also assumed that latent periods are exponentially distributed a not realistic assumption neither. Once again k   and then  1. h h k + h h 3.3. Models with arbitrary distributions for the waiting periods The assumption of exponentially distributed periods is appealing because the corresponding ODE models have constant parameters. However latency or infec- tious periods are, in general, random variables with non-exponential distributions. In our case, where we are considering that vectors remain infectious for life, the infectious period is the vector lifespan. In this case a constant mortality is a realistic choice and therefore the infectious period is exponentially distributed. However this is not the case of vector’s latent period or the latent and infectious host’s periods. As an example we will first consider the simple case of a S IR S I model. For vectors we have the equations 7-8. For the host population we will consider that the infectious period (T ) is a random variable with probability distribution hi function f (s). As usual, the cumulative distribution is denoted by F(s). The complementary cumulative distribution, F(s) = 1 F(s), is known as the survival function and gives the probability that an individual infected in s = 0 remains infected at time s. Because only the fraction F(t s) of the infections produced at time s survives until time t we obtain the integral Volterra equations H (t) = H (0) V (s)H (s)d s s s i s ¯ ¯ H (t) = H (0)F(t) + V (s)H (s)F(t s)d s i i i s H (t) = H H (t) H (t) r s i Di erentiation of Volterra equation gives the following system of integro- di erential equations, 8 dH H s s = V h i dt H ¯ ¯ dH dF H dF i s h = H (0) + V F(0) + V (s)H (s) (t s)d s i h i i s dt dt H H dt s h = H (0) f (t) + V V (s)H (s) f (t s)d s i h i i s H H Realistic distributions for infectious or latent periods are bell shaped and there- fore survival function is of type I. Then, a simple but realistic distribution is ob- tained for the limiting case of fixed infectious period T . In this case the survival hi function is a step function, the probability density distribution is (t T ), and hi therefore we obtain the delayed equation dH H H (t T ) i s s hi = H (0)(t T ) + V V (t T ) (18) i hi h i h i hi dt H H In the general case of arbitrary distributions in latency and infectious periods for hosts and vectors we have the following Volterra integral equations model, H (s) H (t) = H (0) V (s) d s (19) s s h i H (s) ¯ ¯ H (t) = H (0)F (t) + V (s) F (t s)d s (20) e e he h i he Z Z " # H (s) dF s he ¯ ¯ H (t) = H (0)F (t) + V (s) ( s) F (t ) d sd (21) i i hi h i hi H dt 0 0 H (t) = H H (t) H (t) H (t) (22) r s i e Z Z t t H (s) t (ts) V (t) = V (0)e +  e d s V (s) d s (23) s s v v s 0 0 H (s) t (ts) V (t) = V (0)e + V (s) F (t s)e d s (24) e e v s ve " # Z Z H (s) dF i ve t (ts) V (t) = V (0)e + V (s) ( s) e d sd (25) i i v s H dt 0 0 ¯ ¯ where F , F are the survival functions for the exposed and infectious populations je ji ( j = h for host and j = v for vectors). 9 3.3.1. Gamma distributed periods Realistic probability distribution functions for infectious or latent periods are bell shaped and therefore the survival functions are of type I. While accurate nu- merical solutions of a system of ordinary di erential equations like model 10 - 16 are easily obtained using a Runge-Kutta scheme, for example, integral systems like 19-25 are not that amenable. Gamma distributions are flexible functions with two parameters, the shape parameter k and the scale parameter . Some features of this distribution are par- ticularly appealing. The exponential distribution is a special case of the Gamma distribution when k = 1, while for k ! 1 the Gamma distribution converges to the Dirac delta function. Most importantly, for integer values of k the system (19- 25) is equivalent to a system of ordinary di erential equations with constant rates (see for example [38] and Appendix B). This result allows to obtain numerical solutions of the system of integral equations using a simple numerical scheme like Runge-Kutta. 3.3.2. Delayed Model A simple but realistic distribution for the latent or infectious periods is ob- tained in the limiting case of fixed periods when the survival functions are step functions, and therefore the probability density distributions are Dirac delta distri- butions, (sT ), (sT ), (sT ). In this limiting case, the integro-di erential he hi ve system obtained by di erentiation of the integral equations system (see model A.8-A.14 in the Appendix A) reduces to a system of di erential delayed equa- tions, 10 dH H s s = V (26) h i dt H dH H H (t T ) e s s he = V V (t T ) (27) h i h i he dt H H dH H (t T ) i s he = H (0)(t T ) + V (t T ) i h i he dt H H (t T T ) s he hi V (t T T ) (28) h i he hi dH H (t T T ) r s he hi = V (t T T ) (29) h i he hi dt H dV H s i =  V  V (30) v v s v s dt H dV H H (t T ) e i i ve v ve = V e V (t T )  V (31) v s v s ve v e dt H H dV H (t T ) i i ve v ve = e V (t T )  V (32) v s ve v i dt H where T , T , and T are the (fixed) latency and infectious periods of vectors and ve he hi hosts. As discussed above, vector’s infectious period is assumed exponentially distributed as we considered a constant vector mortality rate. Host mortality is disregarded and then all latent host become infectious (and then f = 1). However v ve only a fraction f = e of infected vectors survive the latency period becoming infectious. Therefore the basic reproduction number is given by 1 V (3) v ve R = T e (33) h v hi 3.4. Relationship between the Basic reproduction numbers, and the basic modi- fied model Suppose that we are studying a host-vector system for which there are esti- mations of the parameters as the mean latency and infectious periods. If latency periods are disregarded and we assume that all the periods are exponentially dis- tributed, we may use the basic model (4-8) for which the basic reproduction num- ber is 11 V h v (1) R = : (r +  ) H h h v However, a more realistic model should include the latency periods. Under the most common, but unrealistic, assumption of exponentially distributed periods the corresponding model is 10-16 and the basic reproduction number is (2) (1) v (1) R = R < R : 0 0 0 k + v v The delayed model (27 32) is a more realistic choice for which the basic re- production number is 1 V (3) v ve R = T e : h v hi v  T v ve Because  e , the basic reproduction numbers for the di erent mod- k + v v els satisfy (1) (2) (3) R > R > R : (34) 0 0 0 Therefore we expect larger and faster epidemics for the simple S IRS I model (Eqs. 4 - 8) than the obtained with the more realistic models. However, it is possible to implicitly include the e ect of latency in the vector population in the basic model (4-8) modifying the equation (8) as dV H i i v ve = e V  V (35) v s v i dt H (3) The basic reproduction number for the Basic modified model (Eq. 35) is R , the same as the most realistic Delayed model. 4. A stochastic agent based Ross-Macdonald model A stochastic version of an ordinary di erential equations model like (10-16) is straightforward. Consider, for example, the simple Ross model dH H i s = V r H h i h i dt H dV H i i = V  V v s v i dt H 12 In this case there are only four events: host infection, vector infection, human recovery, and vector death. The rates of the deterministic model define the prob- abilities of occurrence of the events per unit of time or transition rates. Thus, for example, probability of human infection in an interval t is given by P(H ! H ; t) = V t + o(t) s i h i o(t) where o(t) are higher order terms for which lim = 0. t!0 The interval between consecutive events is exponentially distributed with pa- rameter equal to the sum of the all transition rates. This kind of stochastic mod- els are markovian, the probability of occurrence of any event depend only of the present values of the variables. A stochastic version of the integral Volterra equations model like (19-25) is not that easy (see for example [27]) as the corresponding stochastic model is non- markovian, the dynamics depends of the history of the system. One alternative is to consider Gamma distributed periods with integer shape parameter values, for which the equivalent systems becomes markovian and therefore it is possible to use the stochastic simulation scheme outlined above. We preferred to develop an agent based model (ABM) for which the simula- tion of periods with arbitrary distributions is straightforward. Agent based models are a computational tool which allows to simulate populations dynamics consid- ering the features of each individual in the population and the interaction between them [22]. Agent based models are considered the most realistic models where some features, like the mobility of each individual, can be easily incorporated [12, 31]. Our model considers a SEIR model for the host population and a SEI model for the vector population. 4.1. Modeling disease transmission, progression and recovery. For each host and vector, the followings attributes were considered: The epidemiological status (State) which may take the values SUSCEPTI- BLE, EXPOSED, INFECTIOUS, RECOVERED. The age of infection (the time elapsed from first infection). The age of infectiousness (the time elapsed since progression to the infec- tious status). 13 We considered a fixed time step t. For each vector we generated a (pseudo)random number u with uniform distri- bution in the interval (0,1). If u < 1 ex p(bt) the vector bites a host selected at random which may be infected (if the vector is infectious and the host susceptible) or transmit the infection (if the vector is susceptible and the host infectious) with probabilities p and p respectively. h v Latent or infectious periods are random variables with some probability dis- tribution. We considered the general and flexible case where periods are Gamma distributed. Special cases of the Gamma distribution include exponential distribu- tion (for the shape parameter k = 1) and the limiting case of fixed period (k ! 1). Waiting periods in the di erent epidemiological classes were simulated in the following way. For each newly infected individual we drawn a pseudo random number from the corresponding Gamma distribution. This simulated value of the latency period ( ), plus the current time t, was stored in the variable T change. In all cases T change is a future time at which the individual will change the epidemiological status. When the age of infection becomes greater or equal to  we changed the agent’s state from EXPOSED to INFECTIOUS (in our implementation this is equivalent to the condition t  T change). In a similar way, for each newly infectious individual we drawn a value for the infectious period ( ) and stored it in T change (as before T change is the simulated infectious period plus the actual time t). When the age of infectiousness reach (or surpass) this value the state of the individual was changed from INFECTIOUS to RECOVERED. For the cases of fixed waiting times the transitions are deterministic and are determined by the values selected for the di erent (fixed) periods. Other distributions for the latent and infectious periods may be easily incor- porated as long a generator of random numbers for the corresponding distribution is available. 4.2. Modeling births and deaths. We disregarded host births and deaths. For the vector population we consid- ered a constant mortality ( = 1=T ) and a constant birth rate  . Thus, the v v v v t probability of a vector dying in a time interval t is equal to 1 e . On the other hand, the number of newborns vectors in a time step t was modelled by a Poisson random variable with parameter  t. 4.3. Simulation procedure The simulation procedure used is described in the following pseudo-code: 14 1. Initialization of variables and parameters (a) Set the host (H(0)) and vector (V (0)) population sizes, and the initial conditions H (0), H (0), H (0), H (0), V (0), V (0), V (0). s e i r s e i (b) Set the time step t, the simulation duration t and the current time t sim equal to 0. (c) Set the values of parameters  , p , p , k , k , , b. v v h v h h 2. While t  t and 0  H (t) + H (t) + V (t) + V (t) /* this last sentence sim e i e i interrupts the program when infections cannot takes place anymore */ (a) A random number of susceptible vector are added to the population according to a Poisson distribution with parameter  t (b) For each vector in the population i. A uniform random number in the interval (0; 1) is generated. ii. If the number is less than or equal to bt, the vector bites. The host bitten is chosen at random. If the vector is susceptible and the host bitten is infected A uniform random number in the interval (0; 1) is gener- ated. If the number is less than or equal to p , the mosquito be- comes exposed vector:S tate = EXPOSED, V (t) = V (t) 1, V (t) = V (t) + 1 s s e e Generate a latency period  according to the correspond- ing Gamma distribution. Set an exposed time vector:T change = t +  . If the vector is infected and the host bitten is susceptible A uniform random number in the interval (0; 1) is gener- ated. If the number is less than or equal to p , the host becomes exposed host:S tate = EXPOSED, H (t) = H (t)1, H (t) = H (t) + 1 s s e e Generate a latency period  according to the correspond- ing Gamma distribution. Set an exposed time host:T change = t +  . iii. A uniform random number in the interval (0; 1) is generated. v t iv. If the number is less than or equal to 1 e The vector dies and it is removed from vector population. v. Else 15 If the vector is exposed and vector:T change is less than or equal to the current time t vector:S tate = INFECTED, V (t) = V (t)1, V (t) = V (t) + 1 e e i i (c) For each host i. If the host is exposed and host:T change is less than or equal to the current time t host:S tate = INFECTED, H (t) = H (t) 1, H (t) = H (t) + 1 e e i i Generate a infectious period  according to the corresponding Gamma distribution. Set an infectious time vector:T change = t +  . ii. If the host is infected and host:T change is less than or equal to the current time t host:S tate = RECOVERED, H (t) = H (t)1, H (t) = H (t) + 1 i i r r 5. Some numerical results The simulations start with one host infectious, and all the other individuals susceptible. We used the day as the unit of time. 5.1. Vector to host ratio Host population was constant along the simulations. For the vector population we considered two cases, constant populations and seasonal varying populations. In the first case, the population may fluctuate stochastically about the deterministic equilibrium which was the initial population in our simulations. In the second the population presents seasonal oscillations around this value. The ratio vector per host was set in 1. This value vary from system to system and along the year in seasonal environments. For Aedes aegypty for example values of 0.5 and 1.1 per human were estimated [34, 19]. Studies in malaria estimated the number of anophelines per person in the range 3 to 4 depending on the location [25]. A typical household in high risk areas of T. cruzi transmission has an average of 5 human host while triatomine population may vary from some few individuals to several hundreds [24, 42]. 5.2. Parameter values Vectors life expectancy was set in 10 days, which is of the order of values obtained for most species of mosquito [14] and sandflies [10]. Kissing bugs has a longer life expectancy of about 5 months [32]. 16 Probabilities of transmission ( p , p ) are in general asymmetrical and there is h v a wide range of variation. For malaria those probabilities were estimated in the order of 0.48 and 0.022 [14], while for dengue are close to 1. In this work we used p = p = 0:75. hv vh Human mean latency period is of the order of one week for dengue and about 10 days for malaria [28], while in the case of Chagas this period is between 4 and 15 days [18]. In our simulations we considered a host mean latency period of 6 days. Host mean infectious period is about 3 days for dengue [30], 2 or 4 months for malaria [28]. In the case of Chagas, the acute phase is between 0 and 90 days, period in which the risk to transmit the infection is higher [13]. We used a value of 5 days in our simulations. Latent period for vectors was set in 7 days, a value about the observed in dengue infected mosquitoes. The biting rate b is a parameter harder to estimate. We considered two values, b = 0:3=day and b = 0:5=day, in order to have two cases, low and high basic reproduction numbers. This values are of the order observed for mosquitoes and were used for Ross and Macdonald in their seminal works [25]. 5.3. Epidemic curves In figure 1 we compare numerical solutions of the models for low and high values of the basic reproduction number (b = 0:3 and b = 0:5, respectively) considering the parameter values in the table 1. Parameter Value p 0.75 p 0.75 T 6 [days] he T 5 [days] hi T 7 [days] ve T 10 [days] Table 1: Parameter values used in the simulations. In all cases we set host mortality equal to zero ( =0) while  = 1=T =0.1 days . h v v In table 2 we show the corresponding R and some statistics of the epidemic curves (number of infected host at the epidemic peak, time at which the peak is reached and the final epidemic size as proportion of the total host population size) 2,500 2,250 2,000 600 1,750 1,500 1,250 1,000 0 500 1,000 0 100 200 300 Time Time Figure 1: Solutions of the deterministic models (Host infectious population). Left panel: low R (b = 0:3), right panel, high R (b = 0:5). From left to right: basic model (Eqs. 4 - 8), basic model modified (Eq. 35), SEIR-SEI model (Eqs. 10 - 16), delayed model (Eqs. 26 - 32). Time units in days. for each of the simulations presented in figure 1 corresponding to the di erent deterministic models. Similar results obtained with the agent based model are presented in table 3, where we considered only the case of exponentially distributed periods (corre- sponding to the SEIR-SEI model 10 - 16), and fixed periods (corresponding to the Delayed model 27 - 32). Because for the same parameter values the basic reproduction number of the basic model is greater than the basic reproduction numbers of the other models (see inequality 34), the basic model produces faster epidemics with a higher epi- demic final size. For b = 0:3 (low R ’s), the epidemic final size for the basic model is about two times the epidemic final size obtained with the other models, while the epidemic peak is almost 20 times higher than the obtained with the delayed model (see table 2). The modified basic model (35) has the same basic reproduction number than the delayed model and both models predict almost the same final epidemic sizes. However, the first one produces higher peaks in shorter times, resulting in a epi- demic that spreads through the population faster and runs out earlier. It is impor- tant to note that considering latency periods in hosts and vectors always produces lower epidemics, in comparison with the basic model. To compare the solution of the deterministic SEIR-SEI (Eqs. 10 - 16) and delayed models (Eqs. 26 - 32) with the ABM results, we realized simulations Hi i Model R Epidemic Peak Time Epidemic Final Size Basic model 2.53 759 113 0.86 Basic model modified 1.26 69.7 415 0.365 SEIR-SEI model 1.49 114.3 461.25 0.559 Delayed model 1.26 42.5 733.4 0.37 Model R Epidemic Peak Time Epidemic Final Size Basic model 7.03 2250 45 0.99 Basic model modified 3.50 1126 80 0.925 SEIR-SEI model 4.14 809 140 0.965 Delayed model 3.50 682 175 0.94 Table 2: Basic reproduction number, peak of the epidemic, duration from source case introduc- tion to peak and epidemic final size for the solutions of the di erent Ross-Macdonald models considered in figure 1 left panel (top) and right panel (bottom). following the procedure explained above considering the same parameter values (table 1), population sizes and initial conditions used with the deterministic mod- els. For b = 0:3 (Fig. 2 - left panel and Fig. 3 - left panel) stochasticity dominates the dynamics and realizations of the ABM are qualitatively and quantitatively di erent of the deterministic solutions (see table 3). Epidemic peak is lower in the deterministic case than in the mean value obtained with the ABM simulations. Also, the deterministic values of the epidemic peak and the time at which it is reached are not within the corresponding 95% confidence intervals. Not only individual realizations are qualitatively di erent from the deterministic solutions but the mean of those realizations do not converge to the deterministic results. This behaviour of the stochastic realizations is due to the fact that for low values of the basic reproduction number the stochastic e ects dominate the disease dynamics. In this case stochastic dynamics is not a deterministic drift with noise [4], and therefore we cannot expect that stochastic fluctuations average out. For higher values of R (b = 0:5) stochastic dynamics is quasi deterministic and the realizations of the ABM are similar in shape and size to the deterministic solutions (Fig. 2 - right panel and 3 - right panel). The stochasticity may produce a shift of the epidemic curve (to the left or to the right of the deterministic result), but it does not greatly a ect the height of the peak neither the amplitude of the epidemic curve. 19 20 Epidemic Peak Time Epidemic Final Size Type of model b Mean (95% CI) Mean (95% CI) Mean (95% CI) Exponential 0.3 144.64 (141.36, 147.92) 362.99 (352.1, 373.89) 0.5622 (0.5577, 0.5666) periods 0.5 842.88 (838.60, 847.16) 131.94 (129.80, 134.08) 0.9664 (0.9660, 0.9668) Fixed 0.3 69.48 (67.07, 71.89) 517.91 (497.24, 538.58) 0.3580 (0.3496, 0.3664) periods 0.5 714.5 (710.89, 718.22) 166.30 (163.59, 169.02) 0.9418 (0.9412, 0.9423) Table 3: Peak of the epidemic, duration from source case introduction to peak and epidemic final size for the solutions of the di erent ABM Ross-Macdonald models considering 200 simulations. 1,000 Ross det H_I H_I 800 H_I H_I 100 H_I H_I 0 200 400 600 800 1,000 0 100 200 300 Time Time Figure 2: Disease dynamics considering periods exponentially distributed and the parameters in table 1. In black the deterministic result, and in red the ABM simulation. Left panel: low R (b = 0.3); right panel, high R (b = 0.5). HI H_I H_I H_I 500 H_I H_I H_I 0 200 400 600 800 1,000 1,200 1,400 0 100 200 300 400 Time Time Figure 3: Disease dynamics considering fixed periods and the parameters in table 1. In black the deterministic result, and in red the ABM simulation. Left panel: low R (b = 0.3); right panel, high R (b = 0.5). 5.3.1. Fixed periods vs bell shaped distributed periods By far the most used distribution for the waiting times is the exponential distri- bution, which in the deterministic case leads to model 10-16. However, as already discussed, this is an unrealistic assumption for most cases. Latency and infec- tious periods are expected to have a bell shaped distribution. In the general case the model 19-25 should be used but numerical solutions are harder to obtain in this case. Bell shaped distributions may be modelled by the Gamma distribution, which with to independent parameters (the shape parameter k and the scale param- eter ) may control mean and variance (see 3.3.1). When a Gamma distribution is used for the probability density distribution of the waiting times, a very useful property of system 19-25 is that for integer values of the shape parameter, the Hi Hi Hi Hi system of integral equations is equivalent to a larger system of ordinary di eren- tial equations (see Appendix B). Thus, for k = 1, system B.1-B.13 reduces to the SEIR-SEI model 10-16, while for k ! 1 it converges to the delayed model 26-32. In figure 4 we show di erent Gamma distributions for di erent values of the shape parameter k. For k = 10 there is a high variability in the waiting periods but solutions are close to the solutions obtained with fixed periods (k = 1, see fig. 5). For k = 50 both cases are almost identical. 0.4 0.3 0.2 0.1 0 5 10 15 20 Time Figure 4: Probability density function of Gamma distribution considering  = 7 and di erent values of k: 1, 2, 5, 10 and 50. 5.4. E ects of seasonality Seasonality is a key driver of disease dynamics in most vector-borne diseases. Seasonality a ects vector population dynamics because, for example, vector’s ac- tivity is temperature dependent. In the present case we only consider that vector recruitment is a ected by seasonality (mostly by variations in rainfall). As an simple example we considered harmonic variations of the form =  [1 +  sin(!t)]: v 0 For  = 0 we recover the case of constant recruitment used in fig. 1. For 0 <   1 the vector population oscillates with frequency !. As we show in fig. 1, duration of epidemics may last almost two years. Sea- sonal variation of vector populations significantly reduces epidemic duration. In 0 50 100 150 200 250 300 Time [days] Figure 5: Solutions of deterministic models (Host infectious population) considering the param- eters in Table 1. Left panel: low R (b = 0:3), right panel, high R (b = 0:5). From left to right: 0 0 SEIR-SEI , model with gamma distribution for waiting periods (Eqs. B.1 - B.13) considering k = 1, 2, 5, 10 and 50, and delayed model (Eqs. 26 - 32). For k = 1 model (Eqs. B.1 - B.13 reduces to model (Eqs. 10 - 16). fig. 6 we compare numerical solutions of the delayed model (Eqs. 26 - 32) for dif- ferent values of . Duration of epidemics range between approximately 10 months to a couple of months for  = 0:5 and  = 1. 1 4 0 0 6 0 1 2 0 0 5 0 1 0 0 0 4 0 8 0 0 3 0 6 0 0 2 0 4 0 0 1 0 2 0 0 0 0 0 2 0 0 4 0 0 6 0 0 8 0 0 1 0 0 0 1 2 0 0 1 4 0 0 1 6 0 0 1 8 0 0 0 3 0 6 0 9 0 1 2 0 1 5 0 1 8 0 2 1 0 2 4 0 2 7 0 3 0 0 T i m e T i m e Figure 6: Numerical solutions of the delayed model (Eqs. 26 - 32) for  = 0 (no seasonality, black line),  = 0:5 (red) and  = 1 (blue), for b = 0:3 (left panel) and b = 0:5 (right panel). Fig. 7 shows numerical solutions of the ABM considering fixed periods and seasonality. We can see that in the case in which R is higher (b = 0:5) the ABM results are similar to the deterministic model. Conversely, considering b = 0:3, the curves obtained from the AMB are qualitatively similar to the deterministic case, Hi since the same amount of peaks can be clearly observed in both cases. However, the values reached in these peaks in the case of ABM are higher, in general, in the first and second one and lower in the third one. 120 1,200 1,000 80 800 60 600 40 400 0 0 0 500 1,000 1,500 0 50 100 150 200 250 300 350 Time Time Figure 7: Numerical solutions of the delayed model (Eqs. 26 - 32) with  = 0:5 (red) and the corresponding ABM considering fixed periods (other colours curves), for b = 0:3 (left panel) and b = 0:5 (right panel). In black line the solution of the delayed model (Eqs. 26 - 32) without seasonality ( = 0). 5.5. Computing R from the agent based model To compute R in the case of the agent based model, we have to follow the infectious generation of hosts and vectors. So, the procedure realized is as follow. The first infected host is the only host of first infected generation. The vectors infected by a host of first generation, are vectors of first infected generation. When a vector of first infected generation, infects a susceptible hosts, these host are second infected generation. In general, when a host of infected generation m infects a vector, the infected generation of the vector is m. Then, when a vector of infected generation m infects a host, then infected generation of the host is m + 1. Let H be the number of infected-host generation m. Then, R can be esti- m 0 mated as R  H =H [5]. Due to the stochasticity of the ABM simulations, it is 0 3 2 important to realize a considerable number of simulations and then calculate the mean of the R value estimated for each simulation. An R estimation considering 0 0 200 realizations of the simulations analyzed en the previous section is presented in the table 4. In all the cases the deterministic value of R is within the 95% confidence interval. As we can see in Eqs. 17 and 3.4, given the parameters of the host and vector populations, R is a linear function of the relation V=H. So, varying the relation V=H, we can obtain di erent values of R . Hi Hi Deterministic Estimation R  H =H 0 3 2 Type of model b R Mean (95% CI) Exponential 0.3 1.25 1.28 (1.01, 1.54) periods 0.5 3.49 3.60 (3.16, 4.05) Fixed 0.3 1.49 1.45 (1.21, 1.69) periods 0.5 4.14 4.03 (3.60, 4.46) Table 4: Estimation of R from the ABM model considering 200 simulations. Considering the parameters in the table 1 and a biting rate equal to 0.3 (b = 0:3), we estimated the value of R from the agent based model for di erent values of V=H. The results considering exponentially distributed periods and fixed period are shown in the Fig. 8, respectively. In all the cases, an initial population of 10000 hosts was considered, with only one infected host. Each estimation of the basic reproduction number was realized with 200 simulations. 0 . 6 0 . 8 1 . 0 1 . 2 1 . 4 1 . 6 1 . 8 0 . 8 1 . 0 1 . 2 1 . 4 1 . 6 1 . 8 2 . 0 V / H V / H Figure 8: Empirical estimates of the Basic reproduction numbers (squares, bars are 95% confi- dence interval) obtained with the ABM for the cases of exponentially distributed periods (left) and fixed period (right). Continuous line are the corresponding theoretical values given by expressions (17) and (3.4). As can be seen in the figure 8 numerical estimations of the basic reproductive number are in the 95% confidence interval estimated from the 200 ABM simula- tions. 6. Discussion and Conclusions The use of mathematical models in epidemiology has a long and fruitful tra- dition. However di erent hypotheses about the systems under study may lead to, 25 in some cases, significant di erent results (see for example [1, 45, 46]). In this work we present di erent formulations of the Ross-Macdonald model using ordinary di erential equations, Volterra integral equations and agent based modelling. In the most general case we included latency periods in both vectors and hosts. We also considered general distributions for latency and infectious periods including two simple cases: exponentially distributed periods and fixed periods. As we show in this work, disregarding latency periods has a dramatic e ect in the dynamics. This is quite apparent for low basic reproduction numbers (see Fig. 1). As in most vector-borne diseases vector’s latency periods and life ex- pectancy are of the same order of magnitude, disregarding latency overestimate the basic reproduction number, and therefore we observed faster epidemics with significantly higher peaks. A substantial improvement is achieved with the simple modification (35) which produces the same values of R as the delayed model but still the epidemic curves are significantly di erent. Not only the inclusion of latency periods is important but also its distributions. Using exponentially distributed periods leads to slightly smaller basic reproduc- tion numbers and still a noticeable di erences in the epidemic curves. For the most realistic case of bell-shaped distributed waiting periods we show that numerical solutions of the Volterra integral system 19-25 are close to the so- lutions of the simpler delayed model 26-32, for which numerical simulations may be easily obtained using a Runge-Kutta scheme, for example. For us, the delayed model is therefore the model of choice as it combines realism and simplicity. A central assumption of the Ross-Macdonald models is homogeneous random mixing: probability of biting in a susceptible host is proportional to the frac- tion of susceptible host in the entire population. This hypothesis may hold for some local, relatively small, populations. Larger populations may be modelled using a meta-population approach, for example. If local populations have some degree of synchronization, the total population disease dynamics could be quasi- deterministic (see for example [21]), and perhaps a Ross-Macdonald model may describe the global dynamics of the system. In this work we considered popula- tions of 10 individuals, a large enough population for which it is not obvious that the assumption of homogeneous mixing holds. For both, high and low values of the basic reproduction number, solutions of the deterministic models and the estimation of the R of the agent based model are statistical similar (see fig. 8 and table 4), although the epidemic curve may be significantly di erent from the deterministic solution, especially for low R value (see fig. 2 and 3). 26 Deterministic models, like the SEIR-SEI model 10-16, are simple ordinary di erential equations systems with constant parameters, more amenable for anal- ysis. Numerical integration is straightforward using Runge-Kutta of fourth order, for example. The more realistic choice of fixed periods is modelled by delayed di erential equations. Analysis is more complex for these type of models but numerical integration is easily implemented too. For the agent based model there are not di erences, neither in the diculty of the coding or in the computational cost for both cases, and therefore non- exponentially distributed periods (like fixed periods) is the recommended choice. In our simulations we considered parameter values compatible with some vector-borne diseases in humans like dengue. In all cases the number of vec- tors per host was set equal to one at demographic equilibrium. For low values of the basic reproduction number epidemics obtained with the (most realistic) fixed period models have a duration of more than two years (see Fig. 3, left panel), which is never observed in real epidemics. This results highlights the importance of including seasonality when modelling some vector-borne diseases. Vector pop- ulations usually have seasonal fluctuations, driven by rainfall, for example, which shape the duration of the epidemics (see fig. 6). However, outbreaks sizes are gen- erally a function not only of the abundance of vectors, but also of other variables such as climate, community immunity, host mobility, among others. Vector activ- ity is strongly a ected by temperature and therefore not only seasonal variations are significant but also habitat conditions as the use of air conditioning. The two values of  used correspond to moderate seasonal variations in vector abundance ( = 1=2) as expected in endemic settings, and to marked variations in vector abundances as observed in non-endemic populations. As the homogeneous mixing assumption is expected to hold only for relatively small populations, stochasticity should be considered when modelling such cases. Larger populations may be modelled using a metapopulation approach, something we will explore in forthcoming works. Acknowledgements This work was partially supported by grants CIUNSA 2018-2467 and PICT 2014-2476. JPA is a member of the CONICET. MIS is a postdoctoral fellow of CONICET. References 27 [1] Amaku, M., Azevedo, F., Burattini, M. N., Coutinho, F. A. B., Lopez, L. F., & Massad, E. (2015). Interpretations and pitfalls in modelling vector- transmitted infections. Epidemiology & Infection, 143(9), 1803-1815. [2] Amaku, M., Azevedo, F., Burattini, M. N., Coelho, G. E., Coutinho, F. A. B., Greenhalgh, D., Lopez, L. F., Motitsuki, R. S., Wilder-Smith, A., & Mas- sad, E. (2016). Magnitude and frequency variations of vector-borne infection outbreaks using the Ross-Macdonald model: explaining and predicting out- breaks of dengue fever. Epidemiology & Infection, 144(16), 3435-3450. [3] Anaguano, D. F., Ponce, P., Baldeon, ´ M. E., Santander, S., & Cevallos, V. (2015). Blood-meal identification in phlebotomine sand flies (Diptera: Psy- chodidae) from Valle Hermoso, a high prevalence zone for cutaneous leish- maniasis in Ecuador. Acta tropica, 152, 116-120. [4] Aparicio, J. P., & Solari, H. G. (2001). Population dynamics: Poisson ap- proximation and its relation to the langevin process. Physical Review Let- ters, 86(18), 4183. [5] Aparicio, J. P., & Pascual, M. (2007). Building epidemiological models from R 0: an implicit treatment of transmission in networks. Proceedings of the Royal Society B: Biological Sciences, 274(1609), 505-512. [6] Auger, P., Kouokam, E., Sallet, G., Tchuente, M., & Tsanou, B. (2008). The Ross-Macdonald model in a patchy environment. Mathematical biosciences, 216(2), 123-131. [7] Bacaer ¨ , N., & Guernaoui, S. (2006). The epidemic threshold of vector-borne diseases with seasonality. Journal of mathematical biology, 53(3), 421-436. [8] Bacaer ¨ , N. (2007). Approximation of the basic reproduction number R0 for vector-borne diseases with a periodic vector population. Bulletin of mathe- matical biology, 69(3), 1067-1091. [9] Bates, P. A. (2007). Transmission of Leishmania metacyclic promastigotes by phlebotomine sand flies. International journal for parasitology, 37(10), 1097-1106. [10] Belen, A., & Alten, B. (2006). Variation in life table characteristics among populations of Phlebotomus papatasi at di erent altitudes. Journal of Vector Ecology, 31(1), 35-44. 28 [11] Benelli, G., & Mehlhorn, H. (2016). Declining malaria, rising of dengue and Zika virus: insights for mosquito vector control. Parasitology research, 115(5), 1747-1754. [12] Bian, L. (2004). A conceptual framework for an individual-based spatially explicit epidemiological model. Environment and Planning B: Planning and Design, 31(3), 381-385. [13] Castanera, ˜ M. B., Aparicio, J. P., & Gurtler ¨ , R. E. (2003). A stage-structured stochastic model of the population dynamics of Triatoma infestans, the main vector of Chagas disease. Ecological modelling, 162(1-2), 33-53. [14] Chitnis, N., Hyman, J. M., & Cushing, J. M. (2008). Determining impor- tant parameters in the spread of malaria through the sensitivity analysis of a mathematical model. Bulletin of mathematical biology, 70(5), 1272. [15] Diekmann, O., & Heesterbeek, J. A. P. (2000). Mathematical epidemiology of infectious diseases: model building, analysis and interpretation (Vol. 5). John Wiley & Sons. [16] Dietz, K., Molineaux, L., & Thomas, A. (1974). A malaria model tested in the African savannah. Bulletin of the World Health Organization, 50(3-4), [17] Feng, Z., Xu, D., & Zhao, H. (2007). Epidemiological models with non- exponentially distributed disease stages and applications to disease control. Bulletin of mathematical biology, 69(5), 1511-1536. [18] Filigheddu, M. T., Gor ´ golas, M., & Ramos, J. M. (2017). Enfermedad de Chagas de transmision ´ oral. Medicina cl´ ınica, 148(3), 125-131. [19] Focks, D. A., Brenner, R. J., Hayes, J., & Daniels, E. (2000). Transmission thresholds for dengue in terms of Aedes aegypti pupae per person with dis- cussion of their utility in source reduction e orts. The American journal of tropical medicine and hygiene, 62(1), 11-18. [20] Gardner, L., Chen, N., & Sarkar, S. (2017). Vector status of Aedes species determines geographical risk of autochthonous Zika virus establishment. PLoS neglected tropical diseases, 11(3), e0005487. 29 [21] Gutierrez, J. A., & Aparicio, J. P. (2015). Quasi-deterministic popula- tion dynamics in stochastic coupled maps. Journal of Biological Systems, 23(supp01), S151-S162. [22] Nepomuceno, E.G., Takahashi, R. H. C., & Aguirre, L. A. (2016). Individual-based model (IBM): an alternative framework for epidemiolog- ical compartment models. Revista Brasileira de Biometria, 34(1), 133-162. [23] Noireau, F., Diosque, P., & Jansen, A. M. (2009). Trypanosoma cruzi: adap- tation to its vectors and its hosts. Veterinary research, 40(2), 1-23. [24] Nouvellet, P., Dumonteil, E., & GourbiA¨re, S. (2013). The improbable transmission of Trypanosoma cruzi to human: the missing link in the dynam- ics and control of Chagas disease. PLoS neglected tropical diseases, 7(11), e2505. [25] Macdonald, G. (1955). The measurement of malaria transmission. Proceed- ings of the Royal Society of Medicine, 295-302. [26] Mandal, S., Sarkar, R. R., & Sinha, S. (2011). Mathematical models of malaria-a review. Malaria journal, 10(1), 202. [27] Mohammadi, F. (2016). A computational wavelet method for numerical so- lution of stochastic Volterra-Fredholm integral equations. Wavelet and Lin- ear Algebra, 3(1), 13-25. [28] Molineaux, L., & Gramiccia, G. (1980). The Garki project: research on the epidemiology and control of malaria in the Sudan savanna of West Africa. World Health Organization. [29] O’Regan, S. M., Lillie, J. W., & Drake, J. M. (2016). Leading indicators of mosquito-borne disease elimination. Theoretical ecology, 9(3), 269-286. [30] Otero, M., & Solari, H. G. (2010). Stochastic eco-epidemiological model of dengue disease transmission by Aedes aegypti mosquito. Mathematical biosciences, 223(1), 32-46. [31] Otero, M., Barmak, D. H., Dorso, C. O., Solari, H. G., & Natiello, M. A. (2011). Modeling dengue outbreaks. Mathematical biosciences, 232(2), 87- 30 [32] Rabinovich, J. E. (1972). Vital statistics of Triatominae (Hemiptera: Redu- viidae) under laboratory conditions. I. Triatoma infestans Klug. Journal of Medical Entomology, 9(4), 351-370. [33] Reiner Jr, R.C., Perkins, T.A., Barker, C.M., Niu, T., Chaves, L.F., Ellis, A.M., George, D.B., Le Menach, A., Pulliam, J.R., Bisanzio, D. and Buckee, C., 2013. A systematic review of mathematical models of mosquito-borne pathogen transmission: 1970 - 2010. Journal of The Royal Society Interface, 10(81), p.20120921. [34] Romero-Vivas, C. M., & Falconar, A. K. (2005). Investigation of relation- ships between Aedes aegypti egg, larvae, pupae, and adult density indices where their main breeding sites were located indoors. Journal of the Ameri- can Mosquito Control Association, 21(1), 15-22. [35] Ross, R. (1911). Some Quantitative Studies in Epidemiology. 87, 466 - 467. [36] Sanchez , F., & Calvo, J. G. (2020). Dengue model with early-life stage of ˜ ˜ vectors and age-structure within host. Revista de MatemA¡tica: TeorAa y Aplicaciones, 27(1), 157-177. [37] Schofield, C. J., & Dujardin, J. P. (1997). Chagas disease vector control in Central America. Parasitology Today, 13(4), 141-144. [38] Smith, H. (2011). Distributed delay equations and the linear chain trick. In An Introduction to Delay Di erential Equations with Applications to the Life Sciences (pp. 119-130). Springer, New York, NY. [39] Smith, D. L., & McKenzie, F. E. (2004). Statics and dynamics of malaria infection in Anopheles mosquitoes. Malaria journal, 3(1), 13. [40] Smith, D. L., Battle, K. E., Hay, S. I., Barker, C. M., Scott, T. W., & McKen- zie, F. E. (2012). Ross, Macdonald, and a theory for the dynamics and control of mosquito-transmitted pathogens. PLoS pathogens, 8(4), e1002588. [41] Snow, R. W. (2015). Global malaria eradication and the importance of Plas- modium falciparum epidemiology in Africa. BMC medicine, 13(1), 23. [42] Tomasini, N., Ragone, P. G., Gourbiere, ` S., Aparicio, J. P., & Diosque, P. (2017). Epidemiological modeling of Trypanosoma cruzi: Low stercorarian transmission and failure of host adaptive immunity explain the frequency of mixed infections in humans. PLoS computational biology, 13(5), e1005532. 31 [43] Velasco-Hernandez, ´ J. X. (1991). An epidemiological model for the dynam- ics of Chagas’ disease. Biosystems, 26(2), 127-134. [44] Wilder-Smith, A., & Massad, E. (2018). Estimating the number of unvac- cinated Chinese workers against yellow fever in Angola. BMC infectious diseases, 18(1), 185. [45] Wonham, M. J., Lewis, M. A., Renclawowicz, J., & Van den Driessche, P. (2006). Transmission assumptions generate conflicting predictions in host- vector disease models: a case study in West Nile virus. Ecology letters, 9(6), 706-725. [46] Wonham, M. J., & Lewis, M. A. (2008). A comparative analysis of models for West Nile virus. In Mathematical epidemiology (pp. 365-390). Springer, Berlin, Heidelberg. [47] World Health Organization. (2017). Vector-borne diseases. Retrieved from: https://www.who.int/news-room/fact-sheets/detail/vector-borne-diseases (Last access: September 25, 2019). 32 Appendix A. General SEIR-SEI model with arbitrary distributions for the waiting time We will consider the general case of a SEIR-SEI model. For the host and vector populations we assume that the latency period (T for hosts and T for he ve vectors) and the infectious period for hosts (T ) are random variables with prob- hi ability density distributions f (s), f and f (s), respectively. The cumulative he ve hi distributions are denoted by F (s), F and F (s), respectively. The complemen- he ve hi tary cumulative distribution, F (s) = 1 F (s), is known as the survival function and gives the probability that an individual infected in t = 0 remains infected (or exposed, depending of the case) at time s. Therefore the evolution of the populations can be described by the integral Volterra equations [17]: H (s) H (t) = H (0) V (s) d s (A.1) s s h i H (s) ¯ ¯ H (t) = H (0)F (t) + V (s) F (t s)d s (A.2) e e he h i he Z Z " # H (s) dF s he ¯ ¯ H (t) = H (0)F (t) + V (s) ( s) F (t ) d sd (A.3) i i hi h i hi H dt 0 0 H (t) = H H (t) H (t) H (t) (A.4) r s i e Z Z t t H (s) t (ts) V (t) = V (0)e +  e d s V (s) d s (A.5) s s v v s 0 0 H (s) t (ts) V (t) = V (0)e + V (s) F (t s)e d s (A.6) e e v s ve " # Z Z H (s) dF i ve t (ts) V (t) = V (0)e + V (s) ( s) e d sd: (A.7) i i v s H dt 0 0 Di erentiation of these equations leads to the following system of integro-di erential equations, 33 dH H (t) s s = V (t) (A.8) h i dt H dH H (t) H (s) e s s = H (0) f (t) + V (t) V (s) f (t s)d s (A.9) e he h i h i he dt H H dH H (s) i s = H (0) f (t) + V (s) f (t s)d s i hi h i he dt H Z Z ! H (s) V (s) f ( s)d s f (t )d (A.10) h i he hi 0 0 dH = H (0) f (t) + H (0) f (t) e he i hi dt Z Z H (s) + V (s) f ( s)d s f (t )d (A.11) h i he hi 0 0 dV H (t) s i =  V (t) V (t) (A.12) v v s s dt H dV H (t) e i = e V (0) + V (t) e v s dt H H (s) (ts) V (s) f (t s)e d s V (t) (A.13) v s ve s dV H (s) i i t (ts) = e V (0) + V (s) f (t s)e d s  V (t) (A.14) i v s ve v i dt H Appendix B. Model with Gamma distributions for the waiting periods If the waiting periods are Gamma distributed with mean  and integer shape parameter k, we can apply the linear trick [38] to solve the system of integro- di erential di erential equations (A.8-A.14). Suppose that the latency and infectious periods for hosts and the latency period for vectors are all Gamma distributed with means  and integer shape parameters k (i = 1; 2; 3 respectively). Then it is possible to divide the exposed human population in k compartments such that H = H . In a similar fashion, the 1 e e; j j=1 H and V populations may be divided in k and k clases respectively. i e 2 3 Then, the system of integro-di erential equations is equivalent to the follow- ing system of ordinary di erential equations, 34 dH H s s =  V  H (B.1) h h i h s dt H dH H k e;1 s 1 = V H  H (B.2) h i e;1 h e;1 dt H dH k e;2 = (H H )  H (B.3) e;1 e;2 h e;2 dt dH k e;k = (H H )  H (B.4) e;k 1 e;k h e;k 1 1 1 dt dH k k i;1 1 2 = H H  H (B.5) e;k i;1 h i;1 dt 1 2 dH k i;2 2 = (H H )  H (B.6) i;1 i;2 h i;2 dt dH k i;k = (H H )  H (B.7) i;k 1 i;k h i;k 2 2 2 dt dH k r 2 = H  H (B.8) i;k h r dt dV H s i =  V  V (B.9) v v s v s dt H dV H k e;1 i 3 = V V  V (B.10) h s e;1 v e;1 dt H dV k e;2 = (V V )  V (B.11) e;1 e;2 v e;2 dt dV k e;k = (V V )  V (B.12) e;k 1 e;k v e;k 3 3 3 dt dV k i 3 = V 35 V (B.13) e;k v i dt 3 k k k 1 2 3 X X X where H = H ; H = H , and V = V e e; j i i; j e e; j j=1 j=1 j=1 The Gamma distribution with shape parameter k = 1 is an exponential distri- bution, while for k ! 1, the Gamma distribution tends to a Dirac delta distribu- tion. Thus, as k increases from 1 to 1, the model (Eqs. B.1 - B.13) moves from the SEIR-SEI model (Eqs. 10 - 16) to the delayed model (Eqs. 26 - 32).

Journal

MathematicsarXiv (Cornell University)

Published: Feb 26, 2020

References