157 Estimating Equation Approaches for Integer-Valued Time Series Models
∗
∂m
t
(
θ
t−1
)
∂a
t
∗
−1
(
θ
t−1
)
K
t
= K
t−1
1 −
a
t−1
(
θ
t−1
) +
m
t
(
θ
t−1
)
∂θ ∂θ
∂M
t
(
θ
t−1
)
∂b
t
∗
−1
(
θ
t−1
)
−1
+ b
∗
t−1
(
θ
t−1
) +
M
t
(
θ
t−1
)
K
t−1
. (7.21)
∂θ ∂θ
∗
Since −E
∂g
C
(θ)
F
t−1
denotes the optimal information matrix based on the rst t observations,
∂θ
∗
it follows that K
t
−1
=−
t
s=1
∂g
C
∂
(
θ
θ
s−1
)
can be interpreted as the observed information matrix
associated with the optimal combined EF g
C
∗
(θ). The proof of this theorem is given in Thavaneswaran
et al. (2015). �
In an interesting recent paper, Fokianos et al. (2009) described estimation for linear and
nonlinear autoregression models for Poisson count time series, and used simulation studies
to compare conditional least squares estimates with maximum likelihood estimates. Sim-
ilar to Fisher (1924), we compare the information associated with the corresponding EFs,
and show that the optimal EF is more informative than the conditional least squares EF.
In the class of estimating functions of the form G ={g
m
(θ) : g
m
(θ) =
n
1
a
t−1
(θ)m
t
(θ)},
t=
the optimal EF is given by g
m
∗
(θ) =
t
n
=1
a
t
∗
−1
(θ)m
t
(θ), where a
t
∗
−1
=
−
∂μ
∂
t
θ
(θ)
m
1
.The
t
optimal EF and the conditional least squares EF belong to the class G, and the optimal value
of a
t−1
is chosen to maximize the information. Hence I
g
∗
m
− I
g
CLS
is nonnegative denite. It
follows from page 919 of Lindsay (1985) that the optimal estimates are more efcient than
the conditional least squares estimates for any class of count time series models.
Note that g
n
∗
= 0 corresponds in general to a set of nonconvex, nonlinear equations.
The formulas for (7.17) through (7.21) may be easily coded as functions in R. For each
data/model combination, use of the EF approach in practice requires soft coding of the rst
four conditional moments, derivatives of the rst two conditional moments with respect
to model parameters, and specication of initial values to start the recursive estimation.
Example 7.6
Consider a zero-inated regression model. Let {y
t
}denote a count time series with excess
zeros, and assume that the mean, variance, skewness, and kurtosis of y
t
conditional on
F
t−1
are given by
μ
t
(θ) = (1 − ω
t
)λ
t
,
σ
2
t
(θ) = (1 − ω
t
)λ
t
(1 + ω
t
λ
t
),
ω
t
(1 + 2ω
t
)λ
2
+ 3ω
t
λ
t
+ 1
γ
t
(θ) =
t
,and
((1 − ω
t
)λ
t
)
1/2
(1 + ω
t
λ
t
)
3/2
ω
t
(6ω
2
− 6ω
t
+ 1)λ
3
+ 6ω
t
(2ω
t
− 1)λ
2
+ 7ω
t
λ
t
+ 1
κ
t
(θ) =
t
t
t
,
(1 − ω
t
)λ
t
(1 + ω
t
λ
t
)
2
therefore corresponding to the moments of a ZIP(λ
t
, ω
t
) distribution with pmf
ω
t
+ (1 − ω
t
) exp(−λ
t
),if y
t
= 0,
p(y
t
|F
t−1
) =
(1 − ω
t
) exp(−λ
t
)λ
y
t
t
/y
t
!,if y
t
> 0.