^{1}

^{*}

^{1}

In this paper, we analyze the quasi-stationary distribution of the stochastic
*SVIR* (Susceptible, Vaccinated, Infected, Recovered) model for the measles. The quasi-stationary distributions, as discussed by Danoch and Seneta, have been used in biology to describe the steady state behaviour of population models which exhibit discernible stationarity before to become extinct. The stochastic
*SVIR* model is a stochastic
*SIR* (Susceptible, Infected, Recovered) model with vaccination and recruitment where the disease-free equilibrium is reached, regardless of the magnitude of the basic reproduction number. But the mean time until the absorption (the disease-free) can be very long. If we assume the effective reproduction number
*R*
* _{p}* < 1 or
, the quasi-stationary distribution can be closely approximated by geometric distribution.

*β*and

*δ*stands respectively, for the disease transmission coefficient and the natural rate.

Measles is a highly contagious viral infection that manifests as a rash associated with signs of respiratory infections. It is caused by a virus of the paramyxovirus family whose reservoir is exclusively human [

The children under 5 years remain the most affected. 90% who die have less than 5 years. In developing countries, like Niger where children under one year old represent 4.32%, those under 5 years old 19.73% and those under 15 years 51.18%. The measles remains one of the main causes of infant mortality [

Our stochastic model is a stochastic SVIR (Susceptible, Vaccinated, Infected, Recovered) model for the measles [

To understand this phenomenon, we study the long time behavior of the process conditioned on non extinction, which leads us to consider the quasi-stationary distribution introduced by Danoch and Seneta in biology. It allows to describe the steady state behaviour of population models which exhibit discernible stationarity before to become extinct [

For the continuous time SVIR model under some conditions on the effective reproduction number [

The main results are stated in theorem 4 and theorem 5. Precisely, let R p be the effective reproduction number. In theorem 4, we prove that, if R p < 1 , the quasi-stationary distribution of the number of infected can be closely approximated by geometric distribution with parameter 1 − R p . However in theorem 5, we note that if R p > 1 + β δ this latter distribution is approximatively geometric with parameter 1 / I e * , where β , δ and I e * stands respectively, for the disease transmission coefficient, the natural death rate and the endemic equilibrium point, for the number of infected.

The rest of the paper is organized as follows: The Section 2 describes the deterministic SVIR model by a system of differential equations. The equilibrium points of the system of differential equations are also given. In Section 3, we use the continuous time Markov chains model to form our stochastic SVIR model [

In what follows, S ( t ) , I ( t ) , R ( t ) denote respectively the number of susceptible, infected and immunized (susceptible vaccinated and recovered patients) at time t.

In this model, the new susceptible (newborns) are introduced at a constant rate n. A fraction, pn, of newborns has acquired immunity by vaccination. The other fraction ( 1 − p ) n remains susceptible. p is the probability that a newborn will acquire immunity after being vaccinated. In addition, we assume that:

● the natural death rate is δ for each compartment.

● infectious patients recover at the rate of γ .

● infectious patients have an additional μ death rate from measles.

● we consider the standard incidence f ( I , S ) = β S I , β is the disease transmission coefficient. β is the average probability of an adequate contact (contact sufficient for transmission) between an infected and a susceptible per unit of time.

The compartment diagram of the transitions in the SVIR model is in

The dynamics of a well-mixed population can be described by the following system of differential equations:

{ d S d t = n ( 1 − p ) − β S I − δ S d I d t = β S I − ( δ + μ + γ ) I d R d t = n p + γ I − δ R (1)

Remark.

1) In the case of equilibrium without disease, the system (1) admits an equilibrium point ( S * 0 , I * 0 , R * 0 ) with

S * 0 = ( 1 − p ) n δ , I * 0 = 0 and R * 0 = n p δ (2)

R 0 = β n δ ( δ + μ + γ ) is the basic reproduction number [

reproduction number is defined by R p = ( 1 − p ) R 0 [

2) If R p > 1 , an endemic equilibrium point appears ( S * e , I * e , R * e ) asymptotically stable [

S * e = δ + μ + γ β , I * e = ( R p − 1 ) δ β et R * e = n p β + γ δ ( R p − 1 ) δ β (3)

Let X t = ( S ( t ) , I ( t ) ) t ≥ 0 be a continuous-time homogeneous Markov chain on the denumerable state space ℕ 2 = { 0,1,2, ⋯ } 2 . First, assume that Δ t can be chosen sufficiently small such that at most one change in state occurs during the time interval Δ t . In particular, there can be either a new infection, a birth, a death, or a recovery. From of state { X t = ( s , i ) } , only the following states are accessible:

( s , i ) ; ( s + 1, i ) ; ( s , i − 1 ) ; ( s − 1, i ) ; ( s − 1, i + 1 ) .

corresponding to the possible transitions starting from the state ( s , i ) (see

Let V ( s , i ) be the set of neighbors of state ( s , i ) :

V ( s , i ) = { ( s + 1, i ) ; ( s − 1, i + 1 ) ; ( s − 1, i ) ; ( s , i − 1 ) }

Setting τ ( s , i ) = n ( 1 − p ) + β i s + δ s + ( μ + δ + γ ) i , the transition rates are defined by:

τ ( s , i ) , ( k , l ) = { n ( 1 − p ) ( k , l ) = ( s + 1 , i ) , s ≥ 0 , i ≥ 0 β i s ( k , l ) = ( s − 1 , i + 1 ) , s ≥ 1 , i ≥ 0 δ s ( k , l ) = ( s − 1 , i ) , s ≥ 1 , i ≥ 0 ( μ + δ + γ ) i ( k , l ) = ( s , i − 1 ) , s ≥ 0 , i ≥ 1 (4)

The transition probabilities of X t = ( S ( t ) , I ( t ) ) are defined by

P ( s , i ) , ( k , l ) ( Δ t ) = ℙ { X t + Δ t = ( k , l ) / X t = ( s , i ) }

We have ∀ s ≥ 0 ,

P ( s , i ) , ( k , l ) ( Δ t ) = { ∀ i > 0 , τ ( s , i ) , ( k , l ) Δ t + o ( Δ t ) if ( k , l ) ∈ V ( s , i ) 1 − τ ( s , i ) Δ t + o ( Δ t ) if ( k , l ) = ( s , i ) ∀ i = 0 , P ( s , 0 ) , ( s , 0 ) ( Δ t ) = 1 (5)

The distribution of X t is P s , i ( t ) = 0 if s < 0 or i < 0 and P s , i ( t ) = ℙ { X t = ( s , i ) } if s ≥ 0 , i ≥ 0 . Therefore, the marginal distributions are given by:

ℙ { I ( t ) = i } = ∑ s ≥ 0 P s , i ( t ) and ℙ { S ( t ) = s } = ∑ i ≥ 0 P s , i ( t )

From the Equations (5), we obtain the Kolmogorov Forward equations, for all s ≥ 0 and i ≥ 0

d P s , i d t = n ( 1 − p ) [ P s − 1, i − P s , i ] + β [ ( s + 1 ) ( i − 1 ) P s + 1, i − 1 − s i P s , i ] + ( μ + γ + δ ) [ ( i + 1 ) P s , i + 1 − i P s , i ] + δ [ ( s + 1 ) P s + 1, i − s P s , i ] (6)

Hence the system of differential equations verified by the mathematical expectations:

{ d S ¯ d t = ( 1 − p ) n − β S ¯ I ¯ − δ S ¯ − β c o v S I d I ¯ d t = β S ¯ I ¯ − ( μ + δ + γ ) I ¯ + β c o v S I d R ¯ d t = n p + γ I ¯ − δ R ¯ (7)

S ¯ ( t ) = ∑ s = 0 + ∞ ∑ i = 0 + ∞ s P s , i ( t ) , I ¯ ( t ) = ∑ s = 0 + ∞ ∑ i = 0 + ∞ i P s , i ( t ) c o v S I ( t ) = ∑ s = 0 + ∞ ∑ i = 0 + ∞ s i P s , i ( t ) − S ¯ ( t ) I ¯ ( t ) and R ¯ ( t ) = ∑ r = 0 + ∞ r ℙ { R ( t ) = r }

In the previous article, see [

Theorem 1. Let T 0 = inf { t ≥ 0 , I ( t ) = 0 } with inf ∅ = + ∞ . Then, for all i ∈ ℕ * , ℙ i [ T 0 < + ∞ ] = 1 and lim t → + ∞ ℙ i [ I ( t ) = 0 ] = 1 .

Theorem 2. Let T 0 = inf { t ≥ 0 , I ( t ) = 0 } with inf ∅ = + ∞ and ( S * 0 = ( 1 − p ) n δ , I * 0 = 0 , R * 0 = n p δ ).

If R p ≤ 1 , then (1) E [ T 0 ] < + ∞ and (2) lim t → + ∞ ( S ¯ ( t ) , I ¯ ( t ) , R ¯ ( t ) ) = ( S * 0 , I * 0 , R * 0 ) .

Theorem 3. Let T 0 = inf { t ≥ 0 , I ( t ) = 0 } , inf ∅ = + ∞ and ( S * e = δ + μ + γ β , I * e = ( R p − 1 ) δ β , R * e = n p β + γ δ ( R p − 1 ) δ β )

If R p > 1 , then (1) E [ T 0 ] = + ∞ and (2) lim t → + ∞ ( S ¯ ( t ) , I ¯ ( t ) , R ¯ ( t ) ) = ( S * e , I * e , R * e )

Unlike the deterministic approach, we note that the epidemic is extinguished independently of the threshold R p with a probability equal to 1. More precisely, if R p ≤ 1 extinction occurs in a time of finite mean, and if R p > 1 the disease eventually disappears in a time of infinite mean. However, before the instant of absorption (which is relatively long) the process passes through a quasi-stationary state. To understand this phenomenon, we study the long time behavior of the process conditioned on non extinction, which leads us to consider the quasi-stationary distribution introduced by Danoch and Seneta in biology.

The term quasi-stationarity refers to the distribution of the Markov chain by conditioning on the event that absorption has not occurred yet [

Definition 4.1. A probability distribution π on the set of transient states E T is called a quasi-stationary distribution for the process ( X t ) t ≥ 0 if for all t ≥ 0 and any measurable set A ⊂ E we have π ( A ) = ℙ π ( X t ∈ A / T 0 > t ) [

Remark

1) Equivalently, π is the unique limiting conditional probability distribution such that ∀ t ≥ 0 , ∀ A ⊂ E T lim t → + ∞ ℙ α { X t ∈ A / T > t } = π ( A ) independently of initial distribution α [

2) If the set of transient states is finite and irreducible, it is well known that the quasi-stationary distribution exist. But if this set is infinite the existence of quasi-stationary distribution is not guaranteed, furthermore even if it exist, it is typically impossible to evaluate it explicitly. One is therefore lead to consider iterative methods [

Theorem 4. Let Q s , i * be the quasi-stationary distribution of the process ( X t ) t ≥ 1 and Q . i * = ∑ s ≥ 0 Q s , i * the marginal distribution of the number I * of infected in a quasi-stationary regime. If R p < 1 , for all i ≥ 1 , Q . i * ≈ ( 1 − R p ) R p i − 1

Proof. For all i , j ≥ 1 , setting

{ P s ( i , j , Δ t ) = ℙ ( I ( t + Δ t ) = j / S ( t ) = s , I ( t ) = i ) P I ( i , j , t , Δ t ) = ℙ ( I ( t + Δ t ) = j / I ( t ) = i ) (8)

we have

P I ( i , j , t , Δ t ) = ∑ s ≥ 0 ℙ ( S ( t ) = s ) P s ( i , j , Δ t ) (9)

and according to the process definition ( X t ) t > 0 ,

P s ( i , j , Δ t ) = { β i s Δ t + o ( Δ t ) if j = i + 1 ( μ + γ + δ ) i Δ t + o ( Δ t ) if j = i − 1 1 − [ β i s + ( μ + γ + δ ) i ] Δ t + o ( Δ t ) if j = i (10)

we deduce that

P I ( i , j , t , Δ t ) = { ( μ + γ + δ ) i Δ t + o ( Δ t ) if j = i + 1 1 − [ β i S ¯ ( t ) + ( μ + γ + δ ) i ] Δ t + o ( Δ t ) if j = i − 1 1 − [ β i S ¯ ( t ) + ( μ + γ + δ ) i ] Δ t + o ( Δ t ) if j = i (11)

As in the case of disease-free equilibrium, lim t → + ∞ S ¯ ( t ) = S * 0 = n ( 1 − p ) δ , thanks to the Equation (11), we have

P I ( i , j , t , Δ t ) = { β i S * 0 Δ t + o ( Δ t ) if j = i + 1 ( μ + γ + δ ) i Δ t + o ( Δ t ) if j = i − 1 1 − [ β i S * 0 + ( μ + γ + δ ) i ] Δ t + o ( Δ t ) if j = i (12)

thus asymptotically the process I ( t ) is a linear birth-death process with infinitesimal generator:

q i j = { λ i if j = i + 1 ν i if j = i − 1 where λ = β n ( 1 − p ) δ and ν = ( μ + γ + δ ) (13)

In this case, under the condition λ < ν , it is well known [

Q . i * ≈ ( 1 − R p ) R p i − 1 or Q . i * = ∑ s ≥ 0 Q s , i * (14)

The proof is completed for the theorem 4. □

Remark. Under the condition R p ≤ 1 , the irreducible Markov chain ( X t ) t > 0 is positive recurrent. Then an unique invariant probability measure π exists and

π ( s , i ) = 1 E ( s , i ) ( τ s , i ) where τ s , i = inf { t > 0 / X t = ( s , i ) } .

Thus, the theorem 4 simply states that for all i ≥ 1 , π i = ∑ s ≥ 0 π ( s , i ) ≈ ( 1 − R p ) R p i − 1 if R p < 1 .

Theorem 5. Let Q s , i * be the quasi-stationary distribution of the process ( X t ) t ≥ 1 and Q . i * = ∑ s ≥ 0 Q s , i * the marginal distribution of the number I * of infected in a quasi-stationary regime. If R p > 1 + β δ , for all i ≥ 1 , Q . i * ≈ 1 I e * ( 1 − 1 I e * ) i − 1

Proof. If R p > 1 , we have lim t → + ∞ S ¯ ( t ) = S * e = μ + γ + δ β = ν β .

The approximation of the process I ( t ) by a birth-death process does not lead to a satisfactory result. In fact we will use the recursive method of Nåssell [

Let Q s , i ( t ) the conditional distribution of X t = ( S ( t ) , I ( t ) ) given that the epidemic has not extinguished:

Q s , i ( t ) = P s , i ( t ) 1 − P .0 ( t ) et Q . i ( t ) = ∑ s ≥ 0 Q s , i ( t ) (15)

Q . i ( t ) denotes the marginal distribution of I ( t ) , conditional on non-extinction. From Kolmogorov forward equations (Equation (6)) we obtain the following system:

Q ′ s , i ( t ) = n ( 1 − p ) Q s − 1 , i ( t ) + β ( s + 1 ) ( i − 1 ) Q s + 1 , i − 1 ( t ) + ( μ + γ + δ ) ( i + 1 ) Q s , i + 1 ( t ) + δ ( s + 1 ) Q s + 1 , i ( t ) − τ ( s , i ) Q s , i ( t ) + ( μ + γ + δ ) Q .1 ( t ) Q s , i ( t ) (16)

The quasi-stationary distribution is the stationary probability distribution { Q s , i * } s ≥ 0, i ≥ 0 satisfying:

0 = n ( 1 − p ) Q s − 1 , i * + β ( s + 1 ) ( i − 1 ) Q s + 1 , i − 1 * + ( μ + γ + δ ) ( i + 1 ) Q s , i + 1 * + δ ( s + 1 ) Q s + 1 , i * − τ ( s , i ) Q s , i * + ( μ + γ + δ ) Q .1 * Q s , i * (17)

the recursive method of Nåsell [

( i + 1 ) Q . i + 1 * = δ R 0 n i Q . i * e S * ( i ) + Q .1 * ( 1 − ∑ k = 1 i Q . k * ) (18)

where e S * ( i ) = ∑ s ≥ 0 s Q s , i * Q . i * is the conditional expectation of S * given that I * = i . Let denote S * ¯ the expectation of S * , by summing the Equation (18) over i ≥ 1 we obtain:

Q .1 * = 1 − δ R 0 n ( S * ¯ + c o v ( S * , I * ) I * ¯ ) (19)

It follows that the marginal distribution Q . i * satisfies the recursion relationship

{ Q . i + 1 * = δ R 0 n i i + 1 Q . i * e S * ( i ) + Q .1 * i + 1 ( 1 − ∑ k = 1 i Q . k * ) Q .1 * = 1 − δ R 0 n ( S * ¯ + c o v ( S * , I * ) I * ¯ ) (20)

Let denote X ^ e = ( S e * , I e * ) the endemic equilibrium point in Equation (3) and α = μ + γ + δ δ . We approximate the process X t − X ^ e by a diffusion process with drift matrix A e and covariance matrix c o v e

A e = δ ( − R p − α R p − 1 0 ) et c o v e = n ( 1 − p ) R p ( 2 R p − ( R p − 1 ) − ( R p − 1 ) 2 ( R p − 1 ) ) (21)

its stationary distribution is approximately bivariate normal with mean 0 and covariance matrix [

Σ = n ( 1 − p ) δ ( R p ) 2 ( α + R p − R p − R p R p − 1 + R p 2 α ) (22)

from Equation (18) we approximate c o v ( S * , I * ) by Σ 12 = Σ 21 = − n ( 1 − p ) δ R p and ( S * ¯ , I * ¯ ) by ( S e * , I e * ) , thus we obtain Q .1 * = 1 I e * .

In other hand since the stochastic means are close to equilibrium points in stationary regime we have set e S * ( i ) = c S e * where c is a constant, and from the Equation (18) necessarily c = ( 1 − Q .1 * ) . It follows that the marginal distribution Q . i * satisfies the recursion relationship

{ Q . i + 1 * = i i + 1 Q . i * ( 1 − Q .1 * ) + Q .1 * i + 1 ( 1 − ∑ k = 1 i Q . k * ) Q .1 * = 1 I e * (23)

The only solution of which is Q . i * = 1 I e * ( 1 − 1 I e * ) i − 1 , pour tout i ≥ 1 . The condition

of the theorem ensures that I e * > 1 so the probability distribution ( Q . i * ) i ≥ 1 is well defined. The proof of theorem 4 is complete. □

Two sample paths of I ( t ) in

with initial distributions the quasi-stationary distribution estimated at Equation (23). In the case R p > 1 , the approximation of the quasi-stationary distribution of I by a diffusion process gives a normal distribution whit mean μ I and variance σ I 2 [

μ I = I e and σ I 2 = n ( 1 − p ) δ ( R p ) 2 ( R p − 1 + ( R p ) 2 ) .

since I > 0 , the approximation of this distribution is

O . i * = ϕ ( i − μ I σ I ) σ I Φ ( μ I − 0.5 σ I ) .

where ϕ and Φ denote the normal density function and the normal cumulative distribution function respectively.

The asymptotic analysis of our model identifies three regions in parameter space with qualitatively different behaviors of the quasi-stationary distribution. R p is significantly greater or less than the deterministic threshold value 1 in the first region or the third region respectively and that R p is in a second region can be the transition region close to the deterministic threshold value 1 namely

1 ≤ R p ≤ 1 + β δ or 1 1 − p ≤ R 0 ≤ δ + β δ ( 1 − p ) . This result is analogous to that of Nåsell [

Note that 1 δ is average life length in the target population. For measles the target group remains the children of 0 to 5 years. The mortality rate of children before five years of age in Niger is 280% [

Another approach for understanding the dynamics of the system before absorption is the ratio of means approach. Given the initial state, the ratio of expectations distribution (RE) is defined as a ration between the time that the process spends at each transient state and the expected time to absorption (provided that the expected time to absorption is finite). Precisely let T i be the time that the process ( X t ) t ≥ 0 , starting from state j, spending in state i before absorption, the

RE is defined by [

q i k = { λ ( i ) if k = i + 1 μ ( i ) if k = i − 1 (24)

we have [

Q ^ j i = 1 μ ( i ) ∑ k = 1 min ( j , i ) ∏ n = 1 i − 1 λ ( n ) μ ( n ) ∑ i = 1 N 1 μ ( i ) ∑ k = 1 min ( j , i ) ∏ n = 1 i − 1 λ ( n ) μ ( n ) , 1 ≤ i , j ≤ N (25)

as N → ∞ and ( X t ) t ≥ 0 is a linear birth-death process with infinitesimal generator given in Equation (13), It is easy to see that

Q ^ j i = ( 1 − λ ν ) ( λ ν ) i − 1 = Q . i *

Indeed, the RE is a another natural measure of the behavior of absorbing Markov chains before absorption, but the approximation is good only if the convergence to quasi-stationarity is relatively fast [

The authors declare no conflicts of interest regarding the publication of this paper.

Seydou, M. and Tessa, O.M. (2021) Approximations of Quasi-Stationary Distributions of the Stochastic SVIR Model for the Measles. Journal of Applied Mathematics and Physics, 9, 2277-2289. https://doi.org/10.4236/jamp.2021.99145