You are on page 1of 12

Journal of Statistical Planning and Inference 191 (2017) 1–12

Contents lists available at ScienceDirect

Journal of Statistical Planning and Inference


journal homepage: www.elsevier.com/locate/jspi

Estimation of growth rate in second order branching process


Akanksha S. Kashikar
Department of Statistics, Savitribai Phule Pune University, Ganeshkhind, Pune, 411007, India

article info a b s t r a c t
Article history: The classical BGW process assumes first order dependence, whereas many real life datasets
Received 21 March 2016 exhibit a second or higher order dependence. Further, in some situations, there is a need for
Received in revised form 30 December 2016 a model which allows for simultaneous reproduction by a parent and its offspring. Kashikar
Accepted 14 June 2017
and Deshmukh (2015) have proposed a second order Bienaymé–Galton–Watson (SOBGW)
Available online 27 June 2017
model to deal with such situations. This paper discusses estimation of growth rate for
SOBGW model when only the generation sizes are observed. This model is further used
Keywords:
Growth rate to model the swine flu data for Pune, India and La-Gloria, Mexico.
Higher order branching process © 2017 Elsevier B.V. All rights reserved.

1. Introduction

Branching process is a special type of Markov chain and is widely used in a variety of fields like biological, social and
engineering sciences. Epidemiological studies also use branching processes extensively (cf. Bailey, 1975 chapters 6 and 8
and the references therein). This modelling proves to be useful for estimating the growth rate of an epidemic. The severity of
the epidemic can be decided using this growth rate. Furthermore, estimation of extinction probability enables us to decide,
whether the disease will be completely wiped out from the given closed community. Keeping in mind these efficacies of the
branching process approach, we attempted to model the daily count of swine flu (H1N1) cases in Pune city when there was
an outbreak of swine flu in July–August 2009. However, though the dataset contains zeros in between the first and the last
observation, the epidemic does not die out at that point. We continue to get nonzero generation sizes at some of the later
time points. This is not possible in the classical BGW setup, as 0 is the absorbing state, i.e., the generation size Zn is zero for
some n implies that the process has become extinct at or before nth generation. This is because the BGW model assumes
that an individual reproduces only once in its lifetime and dies immediately after producing the offspring, i.e., the lifetime is
assumed to be one unit. Thus, Zn corresponds to the offspring of individuals at generation (n − 1). This is not the case for the
data on number of swine flu cases in Pune city. For this disease, the period of infectiousness is more than one day. In many
cases, suspecting an ordinary flu, a patient is not isolated on the first day so that on the second day also (s)he continues to
spread the disease. Then, most likely (s)he is removed (now suspecting swine flu). Hence, the dependence is at most at lag
2. Moreover, data are available only on the daily basis. This is one of the reasons why age dependent branching process is
not necessary and SOBGW (easier to deal with and having discrete time structure) model is better. When one tries to model
this dataset using time series approach, AR(2) model gives a better fit than AR(1) model, as discussed in Kanade and Rajarshi
(2010).
The study of first order branching processes started with the problem of extinction of family names. Bienaymé (1845) was
the first to explain this phenomenon mathematically. This problem was further studied by Galton and Watson (1875). Hence
such a simple branching process is known as Bienaymé–Galton–Watson (BGW) process. This process assumes that all the
individuals are statistically identical and reproduce according to the same probability law known as an offspring distribution.
The book by Harris (1963) gives a complete exposure to the developments in the field of branching processes, up to that

E-mail address: akanksha.kashikar@gmail.com.

http://dx.doi.org/10.1016/j.jspi.2017.06.003
0378-3758/© 2017 Elsevier B.V. All rights reserved.
2 A.S. Kashikar / Journal of Statistical Planning and Inference 191 (2017) 1–12

time, whereas Athreya and Ney (1972) present a unified treatment of the theory of branching processes and various related
limit theorems. Jagers (1975) discusses BGW process and its many variations. Inference in branching processes is discussed
in detail by Guttorp (1991) and Basawa and Prakasa Rao (1980, Chapter 2). Inferential aspects for various parameters of a
branching process have also been discussed by (cf. Dion, 1974, 1975) and (cf. Heyde, 1974, 1975).
An extension of the classical BGW process to allow for random lifetimes was formulated in the form of age-dependent
processes proposed by Bellman and Harris (1952) and Sevastyanov (1964). In these processes, the lifetime of an individual
is a continuous random variable. Hence, every individual lives for a random amount of time and at the time of death gives
birth to a random number of offspring. A general process, which is an extension of age-dependent branching processes
(cf. Jagers, 1975) extends this further to address the issue of multiple reproductions during the life of the individual thereby
allowing the simultaneous reproduction by a parent and its offspring. Estimation of parameters in age-dependent processes
by various methods has been discussed by Hoel and Crump (1974) and Johnson et al. (1979) among others and more recently
by Cao (2015). However, not much work has been done on estimation of parameters of general process, especially when the
entire family tree is not observed.
These observations led to SOBGW model as discussed in Kashikar and Deshmukh (2015). Such a model accommodates the
possibility of an individual reproducing more than once in its lifetime. It can also be a suitable model in the case of network
marketing. In the case of network marketing, some schemes allow a person and his/her offspring (i.e. a person recruited
by him/her) to sell the products at the same time. To incorporate such situation, the model allows for the simultaneous
reproduction by an individual and its offspring.
In Section 2, we state the SOBGW model and its basic probabilistic properties. Section 3 discusses estimation of growth
rate when only the generation sizes are observed. Consistency and asymptotic normality of the estimator of growth rate are
also established in the same section. Lastly, the model is applied to two real life datasets in Section 4.

2. SOBGW model

In SOBGW, we assume that an individual reproduces at age 1 and also at age 2 and then dies. Hence, if Zn is the number
of offspring born at nth generation and Z0 = 1, i.e. the process starts with one ancestor, we define,
Zn−1 Zn−2
∑ ∑
Z1 = ξ1 (0), Zn = ξi (n − 1) + ψj (n − 2); ∀ n ≥ 2, (1)
i=1 j=1

where, ξi (n − 1) is a random variable indicating the number of offspring produced at time n by ith individual in (n − 1)st
generation and ψj (n − 2) is a random variable indicating the number of offspring produced at time n by jth individual in
(n − 2)nd generation. We assume that {ξi (n − 1)} and {ψj (n − 2)} are two independent sequences, each being a sequence
of nonnegative integer valued independent and identically distributed (i.i.d.) random variables. The process {Zn , n ≥ 0}
defined as above is SOBGW process. By definition of Zn in Eq. (1), it follows that the process is a second order Markov
chain with state space {0, 1, 2, . . .}. Kashikar and Deshmukh (2015) study the limiting behaviour of the generation sizes
for SOBGW process and it is established that the process can be categorized into three categories, viz. subcritical, critical and
supercritical depending
√ on whether µ1 + µ2 is less than 1, equal to 1 or greater than 1 respectively. The growth parameter
ρ = (µ1 + µ21 + 4µ2 )/2 acts as a criticality parameter, since ρ = 1 implies µ1 + µ2 = 1 and ρ < 1 (ρ > 1) iff
µ1 + µ2 < 1 (µ1 + µ2 > 1). For SOBGW process, if mn denotes E(Zn ), then mean and variance of the generation sizes are
given by,
n
an+1 − bn+1 ∑ an−j bj
mn = E [Zn ] = = , (2)
2n (a − b) 2n
j=0
√ √
where, a = µ1 + µ21 + 4µ2 and b = µ1 − µ21 + 4µ2 and
[ n−2
] [ n−2 ]
∑ ∑
V (Zn ) = σ1 mn−1 +
2
mn−2−i mi+1 + σ2
2 2
mn−2−i mi .
2
(3)
i=0 i=0

Further, we have,

V (Zn ) = σ12 mn−1 + σ22 mn−2 + µ21 V (Zn−1 ) + µ22 V (Zn−2 ) + 2µ1 µ2 Cov (Zn−1 , Zn−2 ). (4)

V (Zn+1 )
Theorem 1. If ρ > 1, i.e., in the supercritical case, V (Zn )
→ ρ 2 as n → ∞.

Proof. From Eq. (3), we have,


n−2
[ ]
V (Zn ) mn−1 ∑ m2i+1 m2i
=σ 2
1 + mn−2−i σ 2
1 +σ 2
2 . (5)
(E(Zn ))2 (E(Zn ))2 (E(Zn ))2 (E(Zn ))2
i=0
A.S. Kashikar / Journal of Statistical Planning and Inference 191 (2017) 1–12 3

From Eq. (2), it is easily seen that the first term on the right hand side of the above equation converges to 0, as n → ∞. Let
Ki = 1 − (b/a)i . Then, using the value of mi from Eq. (2), we get,
n−2
[ ] n−2
∑ m2i+1 m2i 1 ∑
mn−2−i σ 2
+σ 2
ρ i−n−2 Kn−i−1 σ12 ρ 2 Ki2+2 + σ22 Ki2+1 = σ12 Pn + σ22 Qn ,
[ ]
1 2 = (6)
(E(Zn ))2 (E(Zn ))2 K1 Kn2+1
i=0 i=0
[ ]
ρ −(n+2) ∑n−2 i+2 ρ −(n+2) ∑n−2 i
i=0 ρ Kn−i−1 Ki2+2 i=0 ρ Kn−i−1 Ki+1
2
where, Pn = and Qn = . Consider,
K1 K 2
n +1
K1 K 2n+1

n−2
ρ −(n+2) ∑
Pn = ρ i+2 Kn−i−1 Ki2+2
K1 Kn2+1
i=0
n−2
[ ( )i+2 ( )2i+4 ( )n−i−1 ( )n+1 ( )n+i+3 ]
ρ −(n+2) ρ 2 ∑ b b b b b
= ρ i
1−2 + − +2 −
K1 Kn2+1 a a a a a
[(i=0 ( )2 { −1
ρ −1 − ρ −n ρ (b/a)n−1 − ρ −n
) }
1 b
= − 2
K1 Kn2+1 ρ−1 a (ρ b/a) − 1
( )4 { −1 2 2 n−1 } ( )n−1 { −1
ρ (b /a ) −ρ −n
ρ (a/b)n−1 − ρ −n
}
b b
+ −
a (ρ b2 /a2 ) − 1 a (ρ a/b) − 1
( )n+1 { −1 n+3 { −1 }]
ρ −ρ −n
ρ (b/a) n−1
− ρ −n
} ( )
b b
−2 − .
a ρ−1 a (ρ b/a) − 1

Therefore, as n → ∞,
ρ −1 ρ −1 b
[ ]
1 a−1
Pn → − = .
K1 ρ − 1 ρa − b K1 (ρ − 1)(ρ a − b)
Similar rearrangements of terms yield
[ n−2
]
ρ −(n+2) ∑ a−b
lim Qn = lim ρ i
Kn−i−1 Ki2+1 = .
n→∞ n→∞ K1 Kn2+1 K1 ρ 2 (ρ − 1)(ρ a − b)
i=0

Substituting these limits and using Eq. (2), we get,


V (Zn+1 ) V (Zn+1 ) (E(Zn ))2 (E(Zn+1 ))2
= → ρ2. □
V (Zn ) (E(Zn+1 ))2 V (Zn ) (E(Zn ))2

Theorem 2. If ρ > 1, Corr(Zn , Zn−1 ) → 1, as n → ∞

Proof. Using Eq. (4) and Theorem 1, we have,


ρ 3 − µ21 ρ − µ22 ρ −1
lim Corr(Zn , Zn−1 ) = . (7)
n→∞ 2µ1 µ2
Further, from the definition of ρ , we have
⎛ √ ⎞2 √
µ1 + µ21 + 4µ2 2µ21 + 2µ1 µ21 + 4µ2 + 4µ2
ρ2 = ⎝ ⎠ = = µ1 ρ + µ2 . (8)
2 4

Substituting this in the numerator of RHS of (7), we get,


µ22
ρ 3 − µ21 ρ − µ22 ρ −1 = ρ (µ1 ρ + µ2 ) − µ21 ρ −
ρ
µ2
= µ1 (µ1 ρ + µ2 ) + µ2 ρ − µ21 ρ − 2
ρ
µ2
[ ]
= µ2 µ1 + ρ −
ρ
= 2µ1 µ2 .
Hence, the theorem is proved. □
4 A.S. Kashikar / Journal of Statistical Planning and Inference 191 (2017) 1–12

Thus, in the supercritical case, for large n, Zn and Zn−1 are almost perfectly linearly related. Hence, we can write,
Zn ≈ an + bn Zn−1 . Therefore, V (Zn ) ≈ b2n V (Zn−1 ). Using Theorem 1, we get b2n → ρ 2 . Further, E(Zn ) ≈ an + bn E(Zn−1 ).
For large n, E(Zn ) ≈ ρ n and bn ≈ ρ . This gives an ≈ 0, for large n. Thus, we get, Zn ≈ ρ Zn−1 , for large n. Thus, it is of interest
a.s.
to estimate ρ as it is related to growth of the process. Kashikar and Deshmukh (2015) have proved that ρZnn → W ′ . Thus, ρ
plays the role of growth rate for the second order branching process. Therefore, we proceed with the estimation of ρ .

3. Growth rate

In this section, we propose an estimator for growth rate when only the generation sizes are observed and study the
limiting behaviour of this estimator. We propose the following estimator for ρ .
∑n
Zt
ρ̂ = ∑n t =2 .
t =2 Zt −1

Theorem 3. In the supercritical case, on the non-extinction path, ρ̂ is consistent for ρ .

Zn a.s.
Proof. Using the fact that ρn
→ W ′ and Toeplitz lemma, consistency of ρ̂ can be established easily. □
To study the asymptotic distribution of ρ̂ , consider ρ̂ − ρ . Denoting it by Bn , we have,
∑n
i=2 (Zi − ρ Zi−1 )
Bn = ∑ n . (9)
i=1 Zi−1
Most of the asymptotic theory related to branching processes has been developed via martingales. However, in the present
setup, E(Zi − ρ Zi−1 |Fi−1 ) is not zero and hence the numerator of Bn is not a martingale. Therefore, we rewrite the numerator
of Bn as follows.
n n n
∑ ∑ ∑
(Zi − ρ Zi−1 ) = (Zi − µ1 Zi−1 − µ2 Zi−2 ) + (µ1 Zi−1 + µ2 Zi−2 − ρ Zi−1 ). (10)
i=2 i=2 i=2

The first term on the RHS is a martingale. Thus, to establish the normality of the first term with appropriate norming, we
Zn
use the result that E(Z → W almost surely and in quadratic mean (Kashikar and Deshmukh, 2015). The second term is
n)
handled subsequently in Theorem 5.
Let us define,
n n
∑ ∑
Sn = (Zt − µ1 Zt −1 − µ2 Zt −2 ) = Ut (say). (11)
t =2 t =2

This gives, E(Ut |Ft −1 ) = 0 and hence, {Sn } forms a martingale. We use the martingale central limit theorem by Scott (1978).
To state the theorem, we need to construct a sigma field D on RN , the space of all real sequences. We first define a metric
d(x, y) as,
⎛ ⎞
m m
∑ ∑
d(x, y) = sup χ ⎝ xj , yj ⎠ ,
m≥0
j=0 j=0

where, χ (a, b) = |a − b|/(1 + |a − b|) for real numbers a, b. The sigma-field D is the Borel field generated by the metric d.
We recall that probability measure defined on the usual Borel field can be extended to one on D (cf. Guttorp, 1991 p. 55).
Then, we have the following.
∑n
Scott’s Theorem. Let (Sn , Fn ) be a martingale on a probability space (Ω , F, P). Write Sn = j=1 Uj , sn = E(Sn ) < ∞ and
2 2
∑n
j=1 E(Uj |Fj−1 ), where F0 = {Φ , Ω }. For ρ > 1 and δ > 0, define random sequences in R by
2 2
Vn = N

ηn = (ηn0 , ηn1 , . . .),


where,

⎨ Un−j , j = 0, 1, . . . , n − 3,
ηnj = δ sn
0, j ≥ n − 2.

Let

η = (WY0 , WY1 , . . .)
A.S. Kashikar / Journal of Statistical Planning and Inference 191 (2017) 1–12 5

where W ≥ 0 is independent of Y0 , Y1 , . . . , which in turn are independent with Yj ∼ N(0, ρ −j ). Also let

ζn = (ζn0 , ζn1 , . . .) = sn Vn−1 ηn


and
ζ = (Y0 , Y1 , . . .).
The sequence {ζn } has a random norming instead of a fixed one. Suppose that,
P
n Vn → W , with E(W ) < ∞,
2 2
s− (12)

n sn−r → ρ , r ≥ 0,
2 2 −r
s− (13)
P
and E(eisUn /(δ Vn ) |Fn−1 ) → exp(−s2 /2) on {W > 0}. (14)
Then for every E in D such that P(η ∈ ∂ E) = 0 and every F in F
P(ηn ∈ E , F |W > 0) → P(η ∈ E)P(F |W > 0),
and for every E in D such that P(ζ ∈ ∂ E) = 0 and every F in F
P(ζn ∈ E , F |W > 0) → P(ζ ∈ E)P(F |W > 0).
Following lemmas establish the three conditions, viz. (12)–(14) of the Scott’s martingale central limit theorem.

Lemma 1. With the above notations, in the supercritical case,


Vn2 p
→ W,
s2n
where W is a non-negative random variable having finite expectation.

Proof. Consider,
E(Ut2 ) = E (Zt − µ1 Zt −1 − µ2 Zt −2 )2
[ ]

= E E (Zt − µ1 Zt −1 − µ2 Zt −2 )2 |Ft −1
[ { }]

= E σ1 Zt −1 + σ22 Zt −2 .
[ 2 ]

Thus, E(Ut2 ) is finite for each fixed t. Also, for t < s, E(Ut Us ) = 0. Therefore, we get,
n
∑ ∑
s2n = E(Ut2 ) = E σ12 Zt −1 + σ22 Zt −2 < ∞,
[ ]
∀n ∈ N
t =2
n n
∑ ∑
Vn2 = E(Ut2 |Ft −1 ) = (σ12 Zt −1 + σ22 Zt −2 ).
t =2 t =2

To study the limiting behaviour of Vn2 /s2n , consider,


∑n
t =2 (σ1 Zt −1 + σ2 Zt −2 )
2 2
Vn2
= n
].
t =2 E σ1 Zt −1 + σ2 Zt −2
2
∑ [ 2 2
sn
Let

Xt = σ12 Zt −1 + σ22 Zt −2 . (15)


Therefore, we get
∑n
Vn2 t =2 Xt
2
= ∑ n .
sn t =2 E [Xt ]

Using Toeplitz lemma, we get that the limit of the above sequence is the same as that of Xn /E(Xn ). Thus, from the limiting
behaviour of the process (Kashikar and Deshmukh, 2015 Theorem 3) and properties of almost sure convergence, it can
be proved that Vn2 /s2n converges to W almost surely. As almost sure convergence implies convergence in probability, the
result follows. Further, the quadratic mean convergence of Zn /E(Zn ) (Kashikar and Deshmukh, 2015 Theorem 3) implies
that E(W ) = 1. □

s2n−r
Lemma 2. Continuing with the same notations as above, in the supercritical case, the limit of equals ρ −r , ∀r ≥ 0.
s2n
6 A.S. Kashikar / Journal of Statistical Planning and Inference 191 (2017) 1–12

Proof. Using the notations from the proof of the previous lemma, we have,
∑n−r
s2n−r E(Xt )
= ∑tn=2 .
s2n t =2 E(Xt )
We have,
a
E(Xt ) = ρ t −2 et ,
a−b
[ ( )t ] [ ( )t −1 ]
where, et = σ12 ρ 1 − ba + σ22 1 − ba . Further,

lim en = σ12 ρ + σ22 < ∞.


n→∞

Therefore,
∑n−r t −2
s2n−r ρ et
lim = lim ∑tn=2 t −2
n→∞ s2
n
n→∞ ρ et
[(t ∑
=2
)( ∑ ) ( ∑n−r t −2 )]
n−r t −2 n
t =2 ρ t =2 ρ ρ
t −2
et
= lim ∑n ∑tn=2 t −2 .
t =2 ρ t =2 ρ
∑n−r t −2
t =2 ρ
n→∞ t − 2 et

ρ , is a divergent series. Therefore, the first factor converges to σ12 ρ + σ22 , whereas the second
∑ t −2
In the supercritical case,
factor converges to its reciprocal. Thus, we get
∑n−r t −2
s2n−r ρ
lim = lim ∑tn=2 t −2
n→∞ s2n n→∞
t =2 ρ
ρ n−r −1 − 1
= lim
n→∞ ρ n−1 − 1

= ρ −r .
Hence the lemma follows. □
[ ( )⏐ ]
Lemma 3. With En = E exp isUn
⏐ Fn−1 , in the supercritical case, on the non-extinction path (i.e., on the set W > 0),

δ Vn

p
En → exp(−s2 /2).

Proof. Consider,
[ )⏐ ( ]
isUn ⏐
En = E exp
⏐ Fn−1
δV ⏐
⎡ ⎛ n ⎞⏐ ⎤
is [(Zn − µ1 Zn−1 − µ2 Zn−2 )] ⏐

= E ⎣ exp ⎝ √∑ ⎠⏐ Fn−1 ⎦ .
δ (σ1 Zt −1 + σ2 Zt −2 )
2 2



σ 2 Zn−1 +σ22 Zn−2
Let δn = √∑
1
. Thus, En becomes,
2 (σ1 Zt −1 +σ22 Zt −2 )
⎡ ⎛ ⎞⏐ ⎤
isδn (Zn − µ1 Zn−1 − µ2 Zn−2 )

⎠⏐ Fn−1 ⎦ .

En = E ⎣ exp ⎝ √
δ σ12 Zn−1 + σ22 Zn−2

Using the definition of Xn in (15), δn2 simplifies to,


Xn
δn2 = ∑
X
( t ) (∑ )( )
Xn E(X ) E(X )
= ∑ t ∑ n
E(Xn ) X E(Xt )
( )t
a.s. E(Xn )
→ lim ∑
n→∞ E(Xt )
ρ n−2
= lim ∑ t −2
n→∞ ρ
A.S. Kashikar / Journal of Statistical Planning and Inference 191 (2017) 1–12 7

(ρ − 1) ρ n−1
[ ]
= lim
n→∞ ρ ρ −1 − 1
n

ρ−1
= = δ 2 > 0 as ρ > 1.
ρ
Using definition of Zn , we have,
⎡ ⎛ ⎞⏐ ⎤
is(Zn − µ1 Zn−1 − µ2 Zn−2 )


E ⎣ exp ⎝ √ ⎠⏐ Fn−1 ⎦
σ12 Zn−1 + σ22 Zn−2


⎡ ⎛ (∑ ) ⎞⏐⏐ ⎤
ξi − µ1 Zn−1 + ψj − µ2 Zn−2 ⏐

is
= E ⎣ exp ⎝ √ ⎠⏐ Fn−1 ⎦
σ1 Zn−1 + σ2 Zn−2
2 2


⎡ ⎛ ⎞⏐ ⎤ ⎡ ⎛ ⎞⏐ ⎤
ξ i − µ1 Zn−1 ψ µ
∑ ⏐ ∑ ⏐
j − 2 Zn−2
⎠⏐ Fn−1 ⎦ ,
⏐ ⏐
= E ⎣ exp ⎝is √ ⎠⏐ Fn−1 ⎦ E ⎣ exp ⎝is √ (16)
σ12 Zn−1 + σ22 Zn−2 ⏐ σ12 Zn−1 + σ22 Zn−2 ⏐
⏐ ⏐

since, {ξi } and {ψj } are independent sequences. The first term on the right hand side of Eq. (16), can be written as,
⎡ ⎛ ⎞⏐ ⎤ ⎡ ⎛ √ ⎞⏐ ⎤

ξi − µ1 Zn−1
⏐ ∑
ξ − µ Z σ 2
1 Zn−1

⎠⏐ Fn−1 ⎦ = E ⎣ exp ⎝is √i 1 n−1
⎠⏐ Fn−1 ⎦ .
⏐ ⏐
E ⎣ exp ⎝is √ √
σ1 Zn−1 + σ2 Zn−2 σ1 Zn−1 σ1 Zn−1 + σ2 Zn−2 ⏐
2 2
⏐ 2 2 2

Examining the behaviour of the square of the second term in the exponent in the above equation, we get,
]−1
σ12 Zn−1 σ22 Zn−2
[
= 1+ 2
σ12 Zn−1 + σ22 Zn−2 σ1 Zn−1
]−1
σ22 E(Zn−2 ) (Zn−2 /E(Zn−2 ))
[
= 1+ 2
σ1 E(Zn−1 ) Zn−1 /E(Zn−1 )
aσ12
→ . (17)
aσ12 + 2σ22
Therefore, using Eq. (17) and applying central limit theorem to {ξi } (a sequence of i.i.d. random variables), we get
⎡ ⎛ ⎞⏐ ⎤
ξi − µ1 Zn−1 aσ12
∑ ⏐ ( 2( ))
p −s
.

E ⎣ exp ⎝is √ ⎠⏐ Fn−1 ⎦ → exp (18)
σ 2 Zn−1 + σ 2 Zn−2 ⏐
⏐ 2 aσ12 + 2σ22
1 2

Similarly, we get
⎡ ⎛ ⎞⏐ ⎤
ψ j − µ2 Zn−2 2σ22
∑ ⏐ ( 2( ))
p −s
.

E exp is √
⎣ ⎝ ⎠
⏐ Fn−1 → exp 2
⏐ ⎦ (19)
σ12 Zn−1 + σ22 Zn−2 ⏐ aσ12 + 2σ22

Thus, using Eqs. (16), (18) and (19), we get


⎡ ⎛ ⎞⏐ ⎤
is(Zn − µ1 Zn−1 − µ2 Zn−2 )
⏐ ( 2)
p −s
.

E ⎣ exp ⎝ √ ⎠⏐ Fn−1 ⎦ → exp (20)
2
σ Zn−1 + σ Zn−2
2 2


1 2
( )
−s2
Using δn → δ and Eq. (20), we get, En → exp 2
. □

Thus, using Lemmas 1–3, all the three conditions of the Scott’s theorem are verified. This can now be used to establish
the asymptotic normality of the first term in RHS of Eq. (10).

Theorem 4. In the supercritical case, on the non-extinction path,


(Zt − (µ1 Zt −1 + µ2 Zt −2 )) D

√∑ → Z ∼ N(0, 1).
(σ12 Zt −1 + σ22 Zt −2 )
8 A.S. Kashikar / Journal of Statistical Planning and Inference 191 (2017) 1–12

Proof. Applying Scott’s martingale central limit theorem to Sn defined in Eq. (11), we get,
n−2

Vn−1 Sn = Vn−1 Uj+2
j=0
n−2

= Vn−1 Un−j
j=0
n−2

= Vn−1 δ sn−2 η(n−2)j
j=0


=δ ζ(n−2)j since ζ(n−2)j = 0 for j ≥ (n − 2)
j=0
n−2

= δ lim sup ζ(n−2)j
j=0

= δ lim sup h(ζn ),


∑n−2
where h(ζn ) = j=0 ζ(n−2)j is a continuous function. From Scott’s martingale central limit theorem stated above, it follows
that, P [ζn ∈ E |W > 0] converges to P [ζ ∈ E |W > 0] for every E such that P [ζ ∈ ∂ E ] = 0, where ζ = (Y0 , Y1 , . . .) such that
{Yi , i = 0, 1, . . .} is a sequence of independent random variables with Yi ∼ N(0,∑ ρ −i ). Hence, ∑
by the continuous mapping
n−2 n−2
theorem (Billingsley, 1968 Theorem 5.1), h(ζn ) converges in law to h(ζ ) = lim sup i=0 Yi . But, i=0 Yi is a sequence of sum
of independent random variables with zero mean. Thus, it forms a martingale. Furthermore,
(⏐ n−2 ⏐) n−2
⏐∑ ⏐ ∑
E ⏐ Yi ⏐ ≤ E(|Yi |) < ∞.
⏐ ⏐
⏐ ⏐
i=0 i=0

Thus, by martingale convergence theorem (Breiman, 1992 Theorem 5.14), there exists a random variable Y such that
∑n−2 a.s. ∑n−2 ∑n
i=0 Yi → Y . To identify the random variable Y , we note that i=0 Yi → Y in distribution. But i=1 Yi has normal
distribution with mean zero and variance βn2 given by,
( n−2 ) n−2
∑ ∑ 1 − ρ −(n−1) ρ
β = var
2
Yi = ρ −i = → = δ −2 .
n
1 − ρ −1 ρ−1
i=0 i=0
[∑ ]
n−2
Thus, for any real x, P i=0 Yi ≤ x = Φ (xβn−1 ), where Φ (x) is the cdf of standard normal random variable and is continuous.
Therefore, Φ (xβn−1 ) → Φ (xδ ), and hence,

Yi converges in law to a normal random variable with mean zero and variance
δ −2 .

δ Yi → Z ∼ N(0, 1).

This completes the proof. □


The next theorem establishes the asymptotic normality of ρ̂ using the results derived so far.

Theorem 5. In a supercritical SOBGW process, on the non-extinction path,


( n
)1/2
ρ 2 (ρσ12 + σ22 )
( )
D

Zi (ρ̂ − ρ ) → N 0, .
(ρ 3/2 + µ2 )2
i=1

Proof. Using (9) and (10), we have,


∑n ∑n ∑n
i=1 (Zi − ρ Zi−1 ) i=1 (Zi − µ1 Zi−1 − µ2 Zi−2 ) i=1 ( µ1 Zi−1 + µ2 Zi−2 − ρ Zi−1 )
Bn = ∑ n = ∑n + ∑n . (21)
i=1 Zi−1 i=1 Zi−1 i=1 Zi−1
Using (8), we have µ1 − ρ = −µ2 /ρ . Substituting this in (21), we get,
∑n
− µ1 Zi−1 − µ2 Zi−2 ) µ2 ni=1 (Zi−1 − ρ Zi−2 )

i=1 (Zi
Bn = −
ρ
∑n ∑n
i=1 Zi−1 Zi−1
∑n ∑ni=1
(Z i − µ1 Zi −1 − µ 2 Z i−2 ) µ 2 Z i−2
= i=1 − Bn−1 ∑in=1 .
ρ
∑n
Z
i=1 i−1 Z
i=1 i−1
A.S. Kashikar / Journal of Statistical Planning and Inference 191 (2017) 1–12 9

This gives,
( n
)1/2 [
µ2
∑ ( ) ]
Z
∑ i−2 Bn−1

Zi−1 Bn +
ρ Zi−1
i=1
⎛ ⎞ ⎛ √∑ ⎞
(σ12 Zi−1 + σ22 Zi−2 )
∑n
(Z i − µ 1 Zi −1 − µ2 Z i−2 )
= ⎝ i√=1 ⎠⎝ ∑n ⎠. (22)
( i=1 Zi−1 )1/2
(σ1 Zi−1 + σ2 Zi−2 )
∑ 2 2

Using Theorem 4, we get that the first factor on the right hand side of Eq. √
(22) converges in distribution to a standard
σ22
normal variable. Further, the second factor converges a.s. to a constant σ12 + ρ
. Thus, if we define Dn as Dn =
σ22
[ ( ) ]
µ2 ∑ Zi−2 D

Zi−1 )1/2 Bn + we have Dn → D, where D ∼ N(0, σ12 +

( ρ Zi−1
Bn−1 , ρ
).
( n
)1/2 [
µ2
∑ ) ](
∑ Zi−2
Dn = Zi−1 Bn + Bn−1 = Gn + Fn Gn−1 , (23)
ρ

Zi−1
i=1
( ) (∑ )1/2
µ µ
where, Gn = ( i=1 Zi−1 )1/2 Bn and Fn = ρ2 ∑ Zi−2
∑n
Zi−1
. The sequence {Fn } converges a.s. to a positive constant F = ρ 3/22 .
P
consistent for ρ , Bn → 0. Further, on the non-extinction path, ( Zi−1 )−1/2 also converges to 0 in probability. As
Since, ρ̂ is∑

n 1/2
a result, ( i=1 Zi−1 ) Bn is bounded in probability and hence, we get
Gn−1 P
→ 1.
Gn
Dn a.s. D D
Let Mn = 1+Fn
. Since Fn → F , Mn → 1+F
. Using Eq. (23), we have
Dn − Gn − Gn Fn
Mn − Gn = .
1 + Fn
Further, using the same equation, we also have
Dn − Gn − Fn Gn−1 = 0 = Dn − Gn − Gn Fn + Gn Fn − Gn−1 Fn
= Dn − Gn − Gn Fn + Fn (Gn − Gn−1 ).
Therefore,
⏐ ⏐
⏐ Fn (Gn − Gn−1 ) ⏐
|Mn − Gn | = ⏐⏐ ⏐
1 + Fn ⏐
⏐ ⏐
Fn ⏐ Gn − Gn−1 ⏐
= ⏐ ⏐ |Gn |
1 + Fn ⏐ Gn ⏐
⏐ ⏐
⏐ Gn − Gn−1 ⏐
≤ Fn ⏐⏐ ⏐ |Gn |
⏐ Gn
= Fn |Ln | |Gn | (say).
P P
Since Gn /Gn−1 → 1, we have |Ln | → 0. Thus, we can use the following result from Billingsley (1968, Problem 2, Page 28).

D P D
Result. For random vectors Xn , Yn , X and random variables Zn that if Xn → X , |Xn − Yn | ≤ Zn |Yn | and Zn → 0, then Yn → X .
Hence, we have
D D
Gn → .
1+F
Thus, Gn converges in distribution to a normal random variable, with mean 0 and variance given by,
σ22 ρ 2 (ρσ12 + σ22 )
( )
(1 + F )−2 σ12 + = .
ρ (ρ 3/2 + µ2 )2
(∑n )1/2 ρ 2 (ρσ12 +σ22 )
D
( )
Therefore, from (21), we observe that i=1 Zi (ρ̂ − ρ ) → N 0, (ρ 3/2 +µ2 )2
. □

In classical BGW process, the offspring mean itself acts as the growth rate. SOBGW process reduces to classical BGW
process, if the offspring distribution at lag 2 is taken to be degenerate at 0. Therefore, if we put µ2 = 0 and σ22 = 0 in
Theorem 5, we get the well-known result for the normality of the Harris estimator of the offspring mean in BGW process.
10 A.S. Kashikar / Journal of Statistical Planning and Inference 191 (2017) 1–12

Fig. 1. QQ plot for ρ̂ .

Table 1
ACF—PACF for square-root transformed Pune data.
Lag ACF PACF
1 0.320 0.320
2 0.539 0.486

Fig. 1 gives a QQ plot of standardized values of ρ̂ for 5000 samples, each of size 40 with both the offspring distributions
as Poisson with µ1 = 0.8 and µ2 = 0.5.

4. Data analysis

In this section, we apply the SOBGW model to two datasets related to spread of swine flu. The first dataset consists of
number of cases tested positive on each day in Pune, India and the second dataset records the same for La-Gloria, Mexico.
For Pune, the data are recorded from July, 15 to August, 4, 2009. In the case of La-Gloria, the data are recorded from March,
9 to April, 13, 2009. Thus, we have 21 observations in Pune data and 47 observations in La-Gloria data. The Pune dataset is
from Kanade and Rajarshi (2010), whereas La-Gloria dataset is taken from Fraser et al. (2009).
In both the datasets, there are zeros in between the first and the last observations. However, there are no consecutive
zeros. Further, Kanade and Rajarshi (2010) tried to model Pune data using time series approach and found that AR(2)
gives a better fit than AR(1) model. They have considered square-root transformed data for eliminating non-stationarity.
The values of partial autocorrelation function (PACF) and autocorrelation function (ACF) for this square-root transformed
dataset suggest strong second order dependence. In fact, for Pune data, dependence at lag 2 is higher than the dependence
at lag 1 (see Table 1). This suggests that the SOBGW model might be a good fit to these datasets. Therefore, we model both
the datasets using SOBGW and obtain the estimates of ρ . For both the datasets, the estimate of ρ turns out to be greater
than 1, indicating that the process may not become extinct on its own (see Table 2). In Pune, the government took some
measures such as closing down the educational institutes, theatres, malls etc. which are justified by our findings. These
control measures were implemented after 5th August. This external interference lowered the growth rate of the process
and that finally resulted in elimination of the disease. Hence, in the present analysis, we have considered the data only till
4th August. Due to similarity of the symptoms of swine flu with those of seasonal flu, in the initial days of the epidemic,
many people from Pune did not seek/receive appropriate medical attention. As a result, they were not quarantined. Further,
since a city like Pune consists of many susceptibles and many channels through which people can come in contact with
the infected population, in the absence of any external measures such as those mentioned above, the branching process
has practically infinite resources. Hence, the SOBGW model can be appropriate for the initial period of the epidemic. As the
awareness about the epidemic increased, the timely quarantining of the patients, shutting down public places definitely
caused a decline in the resources available for the growth of the epidemic. However, this constraint is not applicable to the
data points considered in the current analysis as it is restricted to the first three weeks of the epidemic.
A.S. Kashikar / Journal of Statistical Planning and Inference 191 (2017) 1–12 11

Fig. 2. Graph of observed vs. expected values for LaGloria data.

Fig. 3. Graph of approximate standardized residuals for Pune data.

To assess the goodness of fit of the model, we apply two methods. First, we compute the fitted values of Zn as ρ̂ Zn−1 . For
both the datasets, the graph of actual versus fitted values of Zn turns out to be approximately linear. The graph for La-Gloria
data is shown in Fig. 2. The second criterion for checking goodness of fit involves computing approximate standardized
Z −ρ̂ Zt −1
residuals as √ t . Ideally, the denominator should contain the offspring variances. However, since it is difficult to
Zt −1 +Zt −2
estimate them, we replace the term σ12 Zt −1 + σ22 Zt −2 by another term of similar order. For both the datasets, the values of
the residuals are satisfactory. The graph of these residuals with respect to time for Pune data is shown in Fig. 3.
One of the most obvious responses to an epidemic is carrying out mass vaccination. However, due to limited financial
and medical resources, only a section of a population can be vaccinated. Further, vaccines may have their own side-effects.
Hence, it may not be advisable to vaccinate people belonging to subgroups of the population vulnerable to such side effects,
such as say children, pregnant women etc. Hence, we are interested in estimating the minimum proportion of population
that needs to be vaccinated in order to bring the growth of the epidemic under control. A subcritical process dies out on its
own. Hence, to bring the process under control, we need to reduce the size of the pool of the susceptibles. Becker (1976)
has suggested a simple formula to estimate the proportion of vaccination using the fact that reduction in the percentage
12 A.S. Kashikar / Journal of Statistical Planning and Inference 191 (2017) 1–12

Table 2
Estimates for H1N1 datasets.
Data ρ̂ ν̂
Pune 1.0563 5%
La-Gloria 1.0222 1%

of susceptibles will result in an equivalent reduction in the offspring mean. In SOBGW setup, the criticality of the process
is determined by growth rate. Hence, we need to reduce the proportion of susceptibles to such an extent that the value
of ρ becomes less than 1. Therefore, if ν is the minimum value of the proportion of vaccination required to guarantee the
extinction of the process, it can be estimated using the formula ν ≥ 1 − 1/ρ̂ . Table 2 shows that since the estimates of the
growth rates in both the cases are just slightly above 1, the proportion of population required to be vaccinated is also quite
low.

Acknowledgements

The author would like to thank Dr. S. R. Deshmukh and Dr. M. B. Rajarshi for their valuable suggestions. The author is
also thankful to the two referees whose suggestions have helped improve this paper considerably. The author would like
to acknowledge the financial support received from Board of College and University Development, Savitribai Phule Pune
University for part of this work.

References

Athreya, K.B., Ney, P.E., 1972. Branching Processes. Springer-Verlag, Berlin.


Bailey, N.T.J., 1975. The Mathematical Theory of Infectious Diseases and Its Applications. Griffin, London.
Basawa, I.V., Prakasa Rao, B.L.S., 1980. Statistical Inference for Stochastic Processes. Academic Press, London.
Becker, N., 1976. Estimation for an epidemic model. Biometrics 32 (4), 769–777.
Bellman, R., Harris, T.E., 1952. On age-dependent binary branching processes. Ann. of Math. 55, 280–295.
Bienaymé, I.J., 1845. De la loi de multiplication et de la durée des familles. Soc. Philomath. Paris 5, 37–39.
Billingsley, P., 1968. Convergence of Probability Measures. John Wiley and Sons, New York.
Breiman, L., 1992. Probability, SIAM ed.
Cao, X., 2015. Inference for Age-Dependent Branching Process and their Applications, A Ph.D. Dissertation Submitted to George Mason University.
Dion, J., 1974. Estimation of the mean and the initial probabilities of a branching process. J. Appl. Probab. 11 (4), 687–694.
Dion, J., 1975. Estimation of the variance of a branching process. Ann. Statist. 3 (5), 1183–1187.
Fraser, C., Donnelly, C.A., Cauchemez, S., Hanage, W.P., Van Kerkhove, M.D., Hollingsworth, T.D., Griffin, J., Baggaley, R.F., Jenkins, H.E., Lyons, E.J., Jombart, T.,
Hinsley, W.R., Grassly, N.C., Balloux, F., Ghani, A.C., Ferguson, N.M., Rambaut, A., Pybus, O.G., Lopez-Gatell, H., Alpuche-Aranda, C.M., Chapela, I.B., Zavala,
E.P., Guevara, D.M.E., Checchi, F., Garcia, E., Hugonnet, S., Roth, C., The WHO Rapid Pandemic Assessment Collaboration, 2009. Pandemic potential of a
strain of influenza a (H1N1): Early findings. Science 324, 1557–1561.
Galton, F., Watson, H.W., 1875. On the probability of extinction of the families. J. Anthropol. Soc. London 4, 138–144.
Guttorp, P., 1991. Statistical Inference for Branching Processes. John Wiley & Sons, Inc. New York.
Harris, T.E., 1963. Branching Processes. Springer, New York.
Heyde, C.C., 1974. On estimating the variance of the offspring distribution in a simple branching process. Adv. Appl. Probab. 3, 421–433.
Heyde, C.C., 1975. Remarks on efficiency in estimation for branching process. Biometrika 62, 49–55.
Hoel, D.G., Crump, K.S., 1974. Estimating the generation-time distribution of an age-dependent branching process. Biometrics 30 (1), 125–135.
Jagers, P., 1975. Branching Processes with Biological Applications. John Wiley and Sons, London.
Johnson, R.A., Susarla, V., Ryzin, J., 1979. Bayesian non-parametric estimation for age-dependent branching processes. Stochastic Process. Appl. 9 (3),
307–318.
Kanade, R.A., Rajarshi, M.B., 2010. Outbreak of strain of influenza a (H1N1) in pune city during July - October 2009. Calcutta Statist. Assoc. Bull. 62 (247–248),
299–308.
Kashikar, A.S., Deshmukh, S.R., 2015. Probabilistic properties of second order branching process. Ann. Inst. Statist. Math. 67 (3), 557–572.
Scott, D.J., 1978. A central limit theorem for martingales and applications to branching processes. Stochastic Process. Appl. 6, 241–252.
Sevastyanov, B.A., 1964. Age-dependent branching processes. Theory Probab. Appl. 9, 521–537.

You might also like