Skip to main content
Journal of Applied Statistics logoLink to Journal of Applied Statistics
. 2020 Jun 23;48(11):1975–1997. doi: 10.1080/02664763.2020.1783521

A dependent counting INAR model with serially dependent innovation

Masoumeh Shirozhan 1, Mehrnaz Mohammadpour 1,CONTACT
PMCID: PMC9041821  PMID: 35706430

ABSTRACT

To provide a more flexible model of count data, we extend the first-order integer-valued autoregressive model with serially dependent innovations based on the dependent thinning operator. This model is appropriate for modelling the number of dependent random events affecting each other when the number of new cases depend on the previous count through a linear functional relationship. Several statistical properties of the model are determined, parameters are estimated by some methods and their properties are studied via simulations. This study was carried out to investigate the efficiency of the new model by two real count data sets, the number of contagious diseases and robbery.

Keywords: Alternative dependent thinning operator, INAR model, modified conditional least square, overdispersion

2010 Mathematics Subject Classifications: 62M10, 62M05, 62M20, 62G05, 62P25

1. Introduction

In recent years, many developments have been made in modelling and analysis of count time series and various characteristics of the data have been taken into account in these developments. The most of the studies on modelling count time series are based on the thinning operators with independent counting variables, [1–4,8,12]. Ristić et al. [13] introduced a new model based on the dependency on current population via the thinning operator. Proposing such models is appropriate to model the number of dependent random events affecting each other, which may only produce more new random events or vanish after a period similar to the spread of parasitic diseases, the commission of vagrancy offences and survival or collapse of some companies in economy. Some weaknesses and tedious calculus in describing the properties of the model, motivated the authors to provide an alternative operator. Nastić et al. [9] modified the definition of dependent counting series in order to make it more simple and flexible and offered an alternative dependent thinning operator.

The common feature of the above models is that the innovations which represent the new cases are independent and identically distributed. Weiß [18] proposed a new model by motivating a certain type of infinite-patch metapopulation model. He introduced integer-valued autoregressive (INAR) model with state-dependent innovations to model overdispersed count time series.

The purpose of this paper is to provide a development in the analysis of count time series. We introduce a new INAR(1) model on the basis of the alternative dependent thinning operator containing dependent Bernoulli counting series, where the innovations would be allowed to be serially dependent. As Weiß [18], we assume the distribution of the innovation at time t depends on the previous count Xt−1 through a linear functional relationship that make the model more flexible for applications in practice. The interpretation of the model is the same as the dependent counting INAR(1) model in such a way that the count of new cases depends on the previous count Xt−1. This model may be fitted to the highly contagious disease data and describe the mechanism of the spread of a disease.

The rest of the paper is outlined as follows. In Section 2, we construct a new INAR(1) model and investigate some properties of the process. In Section 3, conditional maximum likelihood, modified conditional least square, modified Yule-Walker and modified maximum empirical likelihood methods are applied to estimate the parameters of the model and the performance of the estimators is assessed by a simulation study. Finally in Section 4, we apply the proposed model and the maximum likelihood estimation procedure to two real data sets and illustrate the best performance of the proposed process among some competitive INAR(1) models. In order to confirm the adequacy of the new model, the residual analysis of the two data series are discussed. Consequently, forecasting of the data based on classical, Bayesian and bootstrap approaches are considered for the selected model.

2. The model and its properties

Nastić et al. [9] defined alternative generalized binomial thinning operator α⋄θ as

α⋄θX=U1+U2+⋯+UX,

where X is a non-negative integer-valued random variable, 0≤α≤θ≤1 and {Ui}i∈N is a sequence of random variables defined as Ui=ViZ, where {Vi}i∈N is a sequence of i.i.d random variables with Bernoulli (θ) distribution and Z is Bernoulli (α/θ) distributed random variable. They showed that {Ui}i∈N is a sequence of dependent Bernoulli distributed random variables. It is shown that random variable U1+…+Un,n≥1 is a mixture of a zero distributed random variable with proportion 1−(α/θ) and a random variable with binomial (n,θ) distribution with proportion α/θ.

Now we introduce an INAR(1) model {Xt} based on the alternative generalized binomial thinning operator as

Xt=α⋄θXt−1+εt, (1)

where εt|Xt−1∼Po(aXt−1+b);b>0, a,α,θ∈[0,1),α<θ and given Xt−1 the random variables α⋄θXt−1 and εt are independent of each other. We shall refer to model (1) as an alternative dependent counting INAR process of the first-order with serially dependent innovation (ADSDINAR(1)).

The practical motivation for the process is the need to model the count of dependent random events, which may produce more new random events or vanish after a period in such a way that the count of new cases, εt, depends on the previous count Xt−1. The INAR(1) model has an interpretation in terms of the spread of the viral disease. In the case of the INAR(1) model, if Xt−1 is the number of new patients during the time interval (t−2,t−1], α⋄θXt−1 will be the number of surviving patients from the previous month, which may lead to more new patients or probably treat and the virus debilitated after a period and εt becomes the number of new patient infected from others in the current month.

The ADSDINAR(1) model is equivalent to the usual INAR(1) models with the random coefficient thinning operator as

Xt=dA∘Xt−1+εt, (2)

where ‘○’ is the binomial thinning operator and A is a discrete random variable with probability mass function given by P(A=θ)=1−P(A=0)=αθ. The ADSDINAR(1) model is related to several other models. Let us look at the boundary case α=0, the random coefficient thinning operator reduces to a binomial thinning operator. But from (1), we recognize that the boundary cases α=0 and a = 0 lead to the INARCH(1) model [17] and ADINAR(1) model [9], respectively. So the ADSDINAR(1) model forms a bridge between the ADINAR(1) and INARCH(1) models.

Proposition 2.1

The ADSDINAR(1) process {Xt} defined in (1) to be stationary satisfies a condition as α+a<1.

The one step conditional probability mass function (pmf) and probability generating function (pgf) of the process {Xt} are obtained as

PXt=i|Xt−1=j=1−αθe−aj+baj+bii!+αθ∑k=0min(i,j)jkθk1−θj−ke−aj+baj+bi−k(i−k)!,

and

φXts=eb(s−1)1−αθφXt−1ea(s−1)+αθφXt−1(1−θ+θs)ea(s−1).

The one step conditional expectation and variance are obtained as

EXt|Xt−1=α+aXt−1+b, (3)

and

VarXt|Xt−1=α(θ−α)Xt−12+α(1−θ)+aXt−1+b. (4)

From (3) and (4), the unconditional expectation and variance of the stationary process {Xt} are

μX=EXt=b1−α+a,

and

σX2=VarXt=α(θ−α)E2(Xt)+1−αθE(Xt)1−a2−2aα−αθ,

which show that the ADSDINAR(1) process is overdispersed.

The following proposition formalizes some probabilistic properties of the ADSDINAR(1) model.

Proposition 2.2

Suppose {Xt} following the ADSDINAR(1) model is stationary. Then we obtain the following properties

  1. The conditional expectation is
    EXt|Xt−k=(α+a)kXt−k+(1−(α+a)k)μX, (5)
    and when k→∞, implies limk→∞E(Xt∣Xt−k)=μX, which is the unconditional mean.
  2. The conditional variance is
    VarXt|Xt−k=Mk−(a+α)2kXt−k2+μX2−2μXXt−kα(θ−α)(a+α)2k−Mkα2−αθ−(a+α)k−Mka+α−M+α(1−θ)+aXt−k(a+α)k−Mka+α−M+μXα(1−θ)+a1−Mk1−M−(a+α)k−Mka+α−M+b1−Mk1−M, (6)
    where M=a2+2aα+αθ. When k→∞, implies limk→∞Var(Xt∣Xt−k)=σX2, which is the unconditional variance.
  3. The autocovariance function of the process and innovation are
    γXk:=CovXt,Xt−k=a+αkγX0,
    and
    γεj:=Covεt,εt−j=aa+αj−1aασX2+σε2.

Proof.

(i) The conditional expectation can be computed easily. (ii) For conditional variance, we have

Var(Xt|Xt−k)=EVar(Xt|Xt−k)|Xt−k+VarE(Xt|Xt−1)|Xt−k=α(θ−α)EXt−12|Xt−k+α(1−θ)+aEXt−1|Xt−k+b+(a+α)2VarXt−1|Xt−k=α(θ−α)E2Xt−1|Xt−k+α(1−θ)+aEXt−1|Xt−k+b+MVarXt−1|Xt−k,

where

Var(Xt−1|Xt−k)=α(θ−α)E2Xt−2|Xt−k+α(1−θ)+aEXt−2|Xt−k+b+MVarXt−2|Xt−k.

By repeating this procedure k times, the k-step conditional variance can easily be obtained.

(iii) Since

Covεt,Xt−1=ECovεt,Xt−1|Xt−1+CovEεt|Xt−1,EXt−1|Xt−1=CovaXt−1+b,Xt−1=aVarXt,

and

CovXt,Xt−1=Covα⋄θXt−1+εt,Xt−1=(α+a)VarXt,

the autocovariance function at lag k is obtained as

CovXt,Xt−k=ECovXt,Xt−k|Xt−k+CovEXt|Xt−k,EXt−k|Xt−k=Covα+akXt−k+(1−(α+a)k)μX,Xt−k=α+akVar(Xt).

For computing the autocovariance function of the innovations, first, we need the following covariance functions

CovXt,εt=Covα⋄θXt−1+εt,εt=aασX2+σε2,

and

CovXt,εt−j=CovEXt|Xt−1,εt−j,Eεt−j|Xt−1,εt−j+ECovXt,εt−j|Xt−1,εt−j=Covα+aXt−1+b,εt−j=α+aCovXt−1,εt−j=α+ajCovXt−j,εt−j=α+ajaασX2+σε2.

Hence

Covεt,εt−j=CovEεt|Xt−1,εt−j,Eεt−j|Xt−1,εt−j+ECovεt,εt−j|Xt−1,εt−j=CovaXt−1+b,εt−j=aα+aj−1aασX2+σε2.

3. Estimation

In this section, we will investigate several methods for parameter estimation of the stationary ADSDINAR(1) model based on a realization X1,…,Xn of this process.

3.1. Conditional maximum likelihood estimation

We describe parameter estimation of the model by the conditional maximum likelihood (CML) method. The log-likelihood function for sample observation X1,…,Xn can be written as

ℓα,θ,a,b=lnLα,θ,a,b=∑t=2nln⁡PXt=i|Xt−1=j=∑t=2nln1−αθe−aj+baj+bii!+αθ∑k=0min(i,j)jkθk1−θj−ke−aj+baj+bi−k(i−k)!.

The CML estimates of the unknown parameters α,θ,a,b are obtained numerically by maximizing the log-likelihood function. Numerical maximization of the log-likelihood function is accomplished by using the functions nlm or optim from R software, taking the conditional least squares estimates as initial values of the function.

3.2. Modified conditional least square estimation

The modified conditional least squares estimators based on a combination of conditional least squares estimators and maximum likelihood estimators have been introduced by Miletić Ilić et al. [7]. The estimator of the parameters a,b,α and θ are obtained by two steps. In the first step, the modified conditional least squares estimator of the parameters a, b and α are obtained by minimizing the squares of the conditional errors

Q(α,a,b)=∑t=2nXt−(α+a)Xt−1−b2. (7)

After some calculations, we obtain the conditional least square estimators of the parameters a,α and b as follows

aˆcls=(n−1)∑t=2nXtXt−1−∑t=2nXt∑t=2nXt−1(n−1)∑t=2nXt−12−∑t=2nXt−12−αˆcml,

where αˆcml is the CML estimator of the parameters α used as the initial estimate of the parameter and using aˆcls, we deduce

αˆcls=(n−1)∑t=2nXtXt−1−∑t=2nXt∑t=2nXt−1(n−1)∑t=2nXt−12−∑t=2nXt−12−aˆcls,

and

bˆcls=∑t=2nXt−(aˆcls+αˆcls)∑t=2nXt−1(n−1).

In the second step, the parameter θ can be estimated by minimizing the function

S(θ)=∑t=2nVt−Var(Xt|Xt−1)2, (8)

where Vt=(Xt−E(Xt|Xt−1))2=(Xt−(α+a)Xt−1−b)2 and Var(Xt|Xt−1) is defined in (6). Hence

θˆcls=∑t=2nXt−12−Xt−1Vˆt+αˆcls2Xt−12−(aˆcls+αˆcls)Xt−1−bˆclsαˆcls∑t=2n(Xt−12−Xt−1)2,

where Vˆt=(Xt−(αˆcls+aˆcls)Xt−1−bˆcls)2.

3.3. Modified Yule-Walker estimation

The modified Yule-Walker method, proposed by Miletić Ilić et al. [7], considers a combination of the Yule-Walker estimators and maximum likelihood estimators. The modified Yule-Walker estimators of the unknown parameters a,b,α and θ are obtained as follows. Since

E(Xt)=b1−(a+α),Corr(Xt,Xt−1)=a+αandVar(Xt)=α(θ−α)E2(X)+(1−αθ)E(X)1−a2−2aα−αθ,

using the sample mean (X¯), sample autocorrelation function of lag 1 (C) and sample variance (SX2) we have

bˆyw=X¯(1−C),aˆyw=C−αˆcml,

where αˆcml is the CML estimator of the parameter α used as the initial estimate of the parameter. It is obvious that

αˆyw=C−aˆyw.

Also

θˆyw=(1−aˆyw2−2aˆywαˆyw)SX2+αˆyw2X¯2−X¯αˆywX¯2−X¯+SX2.

3.4. Modified maximum empirical likelihood estimation

In this section, we consider the modified empirical likelihood (MEL) method for the ADSDINAR(1) model, which has the following three steps. In the first step, we obtain the maximum MEL estimators of the parameters a and b as follows. By taking the derivative of Q(⋅) defined in (7) with respect to β=(a,b), the estimating equation obtained as

−12∂Q(β)∂β=∑t=2nmt(β)=0,

where mt(β)=(m1,t(β),m2,t(β))′ with m1,t(β)=Xt−1(Xt −(αˆcml+a)Xt−1−b), m2,t(β)=Xt−(αˆcml+a)Xt−1−b. Note that in this step, the CML estimator of the parameter α is considered as the initial estimate of the parameter. Following Qin and Lawless [11], we can define the log MEL function as

LME(β)=∑t=1nlog⁡(1+d′(β)mt(β)), (9)

where d(β) satisfies

1n∑t=1nmt(β)1+d′(β)mt(β)=0.

The maximum MEL estimator (MMELE) for the parameter β is defined by minimizing the function LME(β) in (9), i.e.

βˆmmel=arg⁡minβLME(β).

In the second step, by substituting aˆmmel and bˆmmel in the CLS estimating Equation (9), CLS estimating equation takes the form of

−12∂Q(α)∂α=∑t=2nmt(α),

where mt(α)=Xt−1(Xt−(α+aˆmmel)Xt−1−bˆmmel) and the MMELE for the parameter α will be obtained by minimizing the function log MEL of α that is written similar to (9).

In the third step, we derive the maximum MEL estimator of the parameter θ. By considering the function S defined in (8), we have

−12∂S(θ)∂θ=∑t=2nmt(θ),

where

mt(θ)=αˆmmelXt−1(Xt−1−1)[Vt−αˆmmel(θ−αˆmmel)Xt−12−(αˆmmel(1−θ)+aˆmmel)Xt−1−bˆmmel],Vt=Xt−(αˆmmel+aˆmmele)Xt−1−bˆmmel2.

Similarly, the MMELE of the parameter θ will be obtained by minimizing the function log MEL (9) for the parameter θ.

3.5. Simulation

The performance of the methods of estimation is assessed by a small Monte Carlo simulation using different sample sizes over 1000 replications. Tables 1 and 2 provide the mean and root mean square errors (RMSE) of the estimates besides the computation time of the CML and MMEL estimates. Based on the tables, we find that the estimates are convergent to their values. Also, increasing the sample size yields smaller RMSE. As it was expected, the CMLEs have the smallest RMSE for all parameters. To show the performance of the MMEL method, we compared the simulation results of MMEL method with CML method via RMSE and computational running time under sample size n=100,500,1000, 5000. Moreover, we can conclude that similar to CML method, the MMEL method works well with lower computational running time. The simulation results are summarized in Table 2.

Table 1. Simulation results for the CLS and YW estimates of some true values of the parameters for the ADSDINAR(1) model with RMSEs.

  (a,b,α,θ)=(0.4,2,0.3,0.5)
n aˆcls bˆcls αˆcls θˆcls aˆyw bˆyw αˆyw θˆyw
100 0.375768 2.183633 0.307481 0.482424 0.368277 2.171793 0.289401 0.511055
St. err. 0.098932 0.473303 0.085799 0.083423 0.105332 0.466128 0.092748 0.094463
500 0.388978 2.112854 0.297715 0.501588 0.39333 2.120684 0.290969 0.508769
St. err. 0.080831 0.288514 0.053678 0.068124 0.081992 0.297057 0.054264 0.067812
1000 0.395183 2.089864 0.299434 0.501726 0.396187 2.096887 0.295491 0.505972
St. err. 0.059033 0.222947 0.050469 0.046833 0.063597 0.239674 0.051947 0.033246
5000 0.398669 2.029853 0.299638 0.498448 0.397709 2.026601 0.298455 0.504221
St. err. 0.038091 0.181278 0.045609 0.040611 0.039754 0.177829 0.050466 0.032878
  (a,b,α,θ)=(0.3,5,0.4,0.8)
100 0.280194 5.163146 0.395394 0.810327 0.288713 5.18775 0.388315 0.789709
St. err. 0.050056 0.822221 0.081136 0.068007 0.047275 0.670502 0.083791 0.064693
500 0.285119 5.11669 0.397124 0.805145 0.283734 5.104828 0.395204 0.791496
St. err. 0.048341 0.685117 0.056589 0.055234 0.040585 0.577114 0.065722 0.053915
1000 0.289663 5.092519 0.398367 0.801154 0.288413 5.098834 0.404227 0.798004
St. err. 0.040062 0.437553 0.040999 0.033526 0.039032 0.465406 0.047224 0.031946
5000 0.294253 5.079339 0.399203 0.799074 0.298224 5.086172 0.399902 0.800307
St. err. 0.037284 0.299316 0.031125 0.030224 0.036946 0.333505 0.025175 0.029088

Table 2. Simulation results for the MMEL and CML estimates of some true values of the parameters for the ADSDINAR(1) model with RMSEs and computational running times.

  (a,b,α,θ)=(0.4,2,0.3,0.5)  
n aˆmmel bˆmmel αˆmmel θˆmmel CPU time aˆcml bˆcml αˆcml θˆcml CPU time
100 0.383467 2.14611 0.296074 0.488409 239.74 0.3867159 2.097073 0.296483 0.511085 389.74
St. err. 0.071163 0.399008 0.040033 0.079363   0.082424 0.322815 0.076802 0.078582  
500 0.398139 2.070159 0.292011 0.492191 922.03 0.393879 2.042568 0.299131 0.503701 1817.67
St. err. 0.047987 0.252115 0.025973 0.054868   0.049827 0.19593 0.046671 0.041789  
1000 0.396296 2.032147 0.299234 0.497765 2090.92 0.397381 2.014196 0.300328 0.501323 3969.34
St. err. 0.03304 0.190941 0.017606 0.039676   0.034628 0.138715 0.031667 0.030222  
5000 0.399321 2.008217 0.2992396 0.499934 14644.75 0.399513 2.005066 0.299571 0.500817 59197.95
St. err. 0.015609 0.090791 0.008256 0.019783   0.017615 0.060441 0.016674 0.018505  
  (a,b,α,θ)=(0.3,5,0.4,0.8)  
100 0.291237 5.133475 0.397088 0.791612 455.95 0.289992 5.151365 0.397785 0.803061 834.14
St. err. 0.042209 0.648061 0.067613 0.053087   0.043512 0.738561 0.052138 0.042161  
500 0.293849 5.048327 0.399629 0.791936 2567.54 0.296752 5.059745 0.398869 0.800128 4373.97
St. err. 0.011706 0.20324 0.040087 0.037449   0.018274 0.321541 0.023252 0.018307  
1000 0.299301 5.011436 0.399706 0.796051 4113.87 0.299161 5.013221 0.399984 0.799651 9870.34
St. err. 0.009918 0.155858 0.028459 0.014226   0.013443 0.228749 0.016671 0.012199  
5000 0.299977 5.010116 0.400525 0.799523 35059.66 0.299893 5.011854 0.399992 0.800029 127783.91
St. err. 0.006204 0.084056 0.013713 0.009929   0.008548 0.140225 0.010074 0.008631  

4. Applications

In this section, we investigate application of the ADSDINAR(1) process by using two real count data sets obtained from the Forecasting Principles site (http://www.forecastingprinciples.com) and Robert Koch Institute: SurvStat@RKI 2.0, (https://survstat.rki.de).

4.1. Robbery data

The first data set represents the monthly counts of robbery, reported in the 27th police car beat in Pittsburgh. The data consist of 144 observations, from January 1990 to December 2001. The sample autocorrelation function (ACF) and partial autocorrelation function (PACF) of the data are displayed in Figure 1. The sample PACF plot suggests that the first-order autoregressive model is appropriate for the data. Moreover, the stationarity of the data series is justified by Augmented Dickey-Fuller test. The null hypothesis of non-stationarity was rejected by the test with a p-value of 0.03105 (at 0.05 significance level). The data is empirically overdispersed with dispersion index IˆX=1.484447. The ADSDINAR(1) model is compared to some competitive INAR(1) models: PINAR(1) [2], GINAR(1) [3], NBIINAR(1) [1], NBRCINAR(1) [16], NGINAR(1) [12], PLINAR(1) [8], EPINAR(1) [18], DCGINAR [13] and ADCINAR(1) [9]. For each INAR model, we obtained the CML estimates together with approximate standard errors (in parentheses, computed from the Hessian of the log-likelihood in R software). The following criteria were used for comparison: Akaike information criterion (AIC), Bayesian information criterion (BIC), Hannan-Quinn information criterion (HQIC) and consistent Akaike information criterion (CAIC). As we know, the four information criteria are based on the log-likelihood function and directly proportional to the number of the parameters in a model. The root mean squares of differences of observations and predicted values (RMS) are also obtained. The smaller values of statistics are better to fit. The results are given in Table 3. As it can be seen from this table, the values of the goodness-of-fit statistics and RMS are the smallest for the ADSDINAR(1) model. Therefore, we can conclude that the ADSDINAR(1) model works well for the data.

Figure 1.

Figure 1.

The ACF and PACF of robbery data.

Table 3. CML estimates with approximate standard errors and some goodness-of-fit statistics for robbery data.

Model CMLE   AIC BIC HQIC CAIC RMS
PINAR(1) λˆ=1.819397(0.12359)   545.99 551.93 548.41 546.07 1.7924
  αˆ=0.245856(0.04287)            
GINAR(1) pˆ=0.659631(0.02784)   586.29 592.23 588.71 586.37 1.8134
  αˆ=0.325251(0.06358)            
NGINAR(1) pˆ=2.322233(0.41239)   569.21 575.15 571.63 569.31 1.8868
  αˆ=0.659536(0.09536)            
NBRCINAR(1) nˆ=7.609481(0.60522)   541.14 550.05 544.76 541.31 1.78817
  pˆ=0.759703(0.01457)            
  ρˆ=0.274231(0.07613)            
NBIINAR(1) nˆ=6.790903(0.56165)   541.33 550.24 544.95 541.50 1.7856
  pˆ=4.027299(0.33371)            
  ρˆ=0.300681(0.05153)            
PLINAR(1) θˆ=0.768386(0.07808)   561.37 567.31 563.79 561.46 1.8043
  αˆ=0.307616(0.06703)            
EPINAR(1) ρˆ=1.621e−7(0.04498)   544.59 553.5 548.21 544.76 1.7868
  aˆ=0.371996(0.04748)            
  bˆ=1.454575(0.11633)            
DCGINAR(1) μˆ=2.288857(0.31023)   575.83 584.74 579.45 576.01 1.7978
  αˆ=0.445493(0.06168)            
  θˆ=0.716618(0.09749)            
ADCINAR(1) μˆ=2.578704(0.31008)   577.3 586.21 580.92 577.47 1.7884
  αˆ=0.367281(0.06714)            
  θˆ=0.908267(0.08592)            
ADSDINAR(1) aˆ=0.192057(0.04898)   539.52 551.41 544.35 539.81 1.6944
  bˆ=1.718345(0.12941)            
  αˆ=0.092476(0.04659)            
  θˆ=0.877033(0.23186)            

Robbery as a criminal act is often committed by groups of people working in agreement with each other. So they cannot decide independently. The ADSDINAR(1) model contains dependent Bernoulli random variables as a counting series can provide good results the robbery data. This model describes the counts of robbery at the time t−1 which may repeat with the probability θˆ=0.1054 at time t or at the same time; it may activate immediate police action and the police arrest all the groups with the probability of 1−(αˆ/θˆ)=0.8946. The number of new cases at time t has linear mean rate in the sense that the baseline mean bˆ=1.718345 is increased by aˆXt−1=0.192057Xt−1.

Eventually, all the measures lead to the superiority of the new INAR model in the data set and the estimates are well defined in the model.

4.2. Giardiasis disease data

The second data set represents the weekly counts of Giardiasis disease, reported from Hamburg. It consists of 469 observations, from January 2010 to December 2018. First, we briefly demonstrate the basic information of the data set through some plots. The sample ACF and PACF of the data are displayed in Figure 2. The sample PACF plot indicate that the first-order autoregressive model is appropriate for data series.

Figure 2.

Figure 2.

The ACF and PACF of Giardiasis data.

The p-value 0.01 of Augmented Dickey-Fuller test indicates the stationarity of the data set (at 0.05 significance level). The data is empirically overdispersed with a dispersion index of IˆX=1.255933. The ADSDINAR(1) model is compared to some competitive INAR(1) models. For each INAR model, the obtained CML estimates together with approximate standard errors are shown in Table 4, as well as AIC, BIC, HQIC, CAIC and RMS. As it can be seen from this table, the values of the goodness-of-fit statistics and RMS are the smallest for the ADSDINAR(1) model. Therefore, we can conclude that the ADSDINAR(1) model works well for the Giardiasis data.

Table 4. CML estimates with approximate standard errors and some goodness-of-fit statistics for Giardiasis data.

Model CMLE   AIC BIC HQIC CAIC RMS
PINAR(1) λˆ=2.413827(0.07537)   1833.81 1842.11 1837.08 1833.84 1.8251
  αˆ=0.109615(0.02323)            
GINAR(1) pˆ=0.687419(0.01442)   2003.13 2011.43 2006.39 2003.16 1.8901
  αˆ=0.322517(0.03427)            
NGINAR(1) pˆ=2.412727(0.20637)   1977.91 1986.20 1981.16 1977.92 1.9774
  αˆ=0.541251(0.05193)            
NBRCINAR(1) nˆ=11.60225(0.29358)   1824.45 1836.91 1829.35 1824.51 1.8245
  pˆ=0.810662(0.00541)            
  ρˆ=0.128534(0.04587)            
NBIINAR(1) nˆ=11.60916(0.40827)   1824.54 1836.99 1829.44 1824.59 1.8245
  pˆ=4.913591(0.17344)            
  ρˆ=0.128414(0.02658)            
PLINAR(1) θˆ=0.290257(0.04199)   1944.68 1952.98 1947.95 1944.71 1.8583
  αˆ=0.659872(0.03544)            
EPINAR(1) ρˆ=4.3175e−7(0.0241)   1869.36 1881.81 1874.25 1869.41 1.8759
  aˆ=0.341973(0.02517)            
  bˆ=1.578462(0.06585)            
DCGINAR(1) μˆ=2.442642(0.17806)   1979.25 1991.71 1984.15 1979.31 1.8936
  αˆ=0.394369(0.03476)            
  θˆ=0.558566(0.05925)            
ADCINAR(1) μˆ=2.826236(0.21637)   1980.32 1992.77 1985.22 1980.37 1.8651
  αˆ=0.341249(0.03426)            
  θˆ=0.615021(0.05216)            
ADSDINAR(1) aˆ=0.074864(0.02573)   1818.96 1835.56 1825.49 1819.05 1.7811
  bˆ=2.311641(0.07831)            
  αˆ=0.073899(0.02383)            
  θˆ=0.960087(0.25181)            
               

Now we describe the mechanism of the spread of the contagious Giardiasis disease by ADSDINAR(1) model. If Xt−1 is the number of new cases of disease that develops during the time interval (t−2,t−1], θˆ=0.960087 will be the probability that each of these cases transmits infection until time t and 1−(αˆ/θˆ)=0.9231 will be the probability that all the cases have been treated. The number of new cases arriving in the interval (t−1,t] infected from other has linear mean rate in the sense that the baseline mean bˆ=2.311641 is increased by aˆXt−1=0.074864Xt−1 with the probability of αˆ/θˆ=0.0769.

Eventually, all the measures lead to the superiority of the new INAR model in the data set and the estimates are well defined in the model.

4.3. The residual analysis and model adequacy

In the following section, we assess the adequacy of the model proposed and fitted to the data.

A hypothesis test for detecting a serial dependence of the innovations is equivalent to test H0:a=0 against H1:a≠0. For this purpose, we use the likelihood ratio test (LRT) statistic 2{log⁡L0(φ,b,λ))−log⁡L1(φ,a,b,λ)}, where L0(⋅) is the likelihood function of the INAR(1) process with alternative generalized binomial thinning operator and Poisson innovation and L1(⋅) is the likelihood function of the ADSDINAR(1) model. Note that the LRT statistic follows the chi-square distribution (asymptotically) with 1 degree of freedom, [15]. The p-values of LRT for both data sets are 0.009 and 0.0235. Hence, a significant serial dependence of the innovations would be desirable.

Tools in the diagnostic checking of dynamic structure can be based on the standardized Pearson residuals given by

et=Xt−E(Xt|Xt−1)Var(Xt|Xt−1),

where the population quantities are replaced by their estimated counterparts in ADSDINAR(1) model. The residual analysis is shown in Figures 3 and 4. We depict the sample autocorrelation function and cumulative periodogram plots of the Pearson residuals for both data series. Figures 3(a) and 4(a), plot the ACF of the Pearson residuals. There is no evidence for any correlation between the residuals and the findings are confirmed by p-values 0.332946 and 0.466017 of Ljung-Box test. Based on cumulative periodogram plots in Figures 3(b) and 4(b), it is clearly shown that the residuals are randomly distributed and do not have specified trend. Figures 5 and 6 are the non-randomized probability integral transformation (PIT) histogram (Czado et al. [5]), which should be uniform if a model is adequate.

Figure 3.

Figure 3.

ACF and cumulative periodogram plot of the Pearson residuals for robbery data.

Figure 4.

Figure 4.

ACF and cumulative periodogram plot of the Pearson residuals for Giardiasis data.

Figure 5.

Figure 5.

Non-randomized PIT histogram and marginal calibration plot (robbery data).

Figure 6.

Figure 6.

Non-randomized PIT histogram and marginal calibration plot (Giardiasis data).

The relative performance of a model within a given group of competing models can be assessed using scoring rules. Scoring rules are regularly used in decision analysis to measure the quality of probabilistic predictions by assigning a numerical score based on the predictive distribution and the observed data. We consider the logarithmic, quadratic and spherical scores which are defined as logs(P,xt)=−log⁡P(xt), qs(P,x)=−2P(xt)+∥P∥2 and sphs(P,x)=−(P(xt)/∥P∥), respectively, where P(xt) is the probability mass of the predictive distribution at the observed count and ∥P∥2=∑i=0∞P(i)2. Table 5 shows mean scores for all competing models of both data series. The scores prefer the ADSDINAR(1) model among all of the competitive INAR models for both data series. In conclusion, our diagnostic tools confirm the superiority of the ADSDINAR(1) model.

Table 5. Mean scores for the both data series.

  Robbery data Giardiasis data
Model logs qs sphs logs qs sphs
PINAR(1) 1.89 −0.3 −0.83 1.95 −0.24 −0.55
GINAR(1) 2.02 −0.16 −0.49 2.12 −0.16 −0.48
NGINAR(1) 1.96 −0.24 −0.63 2.07 −0.21 −0.39
NBRCINAR(1) 1.86 −0.29 −0.81 1.89 −0.24 −0.55
NBIINAR(1) 1.86 −0.29 −0.86 1.89 −0.23 −0.53
PLINAR(1) 1.98 −0.21 −0.54 2.4 −0.1 −0.35
EPINAR(1) 1.87 −0.19 −0.43 1.99 −0.16 −0.4
DCGINAR(1) 1.98 −0.26 −0.72 2.11 −0.23 −0.61
ADCINAR(1) 2.01 −0.21 −0.51 2.11 −0.2 −0.52
ADSDINAR(1) 1.82 −0.37 −0.91 1.78 −0.32 −0.69

4.4. Forecasting

Forecasting is an integral part of time series analysis to check the adequacy and predictability of the selected model. Here, we discuss the classical, Bayesian and bootstrap methods for forecasting the model.

4.4.1. Classical prediction

One of the most common procedures for classical prediction in time series models is to use conditional expectation. Based on (5), the h-step ahead predictor is

Xˆt+h=EXt+h|Xt=α+ahXt+(1−(α+a)h)μX.

In practice, the parameters a, b and α are replaced by their corresponding CML estimates.

4.4.2. Bayesian prediction

The Bayesian predictive probability function assumes that both Xt+h and the vector of parameters are random. Since Beta and Gamma distributions are conjugate of binomial and Poisson distributions, respectively, we use them for prior distributions of the parameters of the model as

a∼Betaa1,b1,α∼Betaa2,b2,θ∼Betaa3,b3,b∼Gammac,d.

Also consider that a,b,α and θ are independent. The posterior distribution of (a,b,α,θ) can be written as

Pa,b,α,θ|Xt∝LXt,a,b,α,θ|X1Pa,b,α,θ=LXt,a,b,α,θ|X1aa1−11−ab1−1αa2−11−αb2−1×θa3−11−θb3−1bc−1e−db,

where L(Xt,a,b,α,θ|X1) is the conditional likelihood function.

The Bayesian predictive function of Xt+h given Xt is given by

fxt+h|xt∝∫α∫θ∫φ∫λfxt+h|xt,α,θ,φ,λPα,θ,φ,λ|xtdλdφdθdα.

So, for Bayesian approach, we need to compute the h-step conditional pmf, f(xt+h|xt,α,θ,φ,λ). Since a closed form expression for the h-step conditional pmf is only available for h=1, we apply the simulation scheme given in [6] to approximate f(xt+h|xt,α,θ,φ,λ). Also, due to the complicated form of Bayesian predictive function f(xt+h|xt), the algorithm given in [14] is used to estimate Xt+h after some adaptation with simulation scheme as the following steps.

  1. From the sample X1,…,Xt, calculate an initial estimate for γ=(a,b,α,θ);

  2. Using the adaptive rejection Metropolis sampling (ARMS) within Gibbs methodology, calculate from the full conditional distribution of the parameters γ, a sample (γ1,…,γm);

  3. For i=1,…,m, j=1,…,B and by (2) simulate
    Xt+h(i,j)=Ai∘Xt+h−1(i,j−1)+εt+h(i,j),
    where the first term on the right-hand side is generated from a binomial distribution with parameters Ai,Xt+h−1(i,j−1) and the last term is generated from the Poisson distribution with parameter aiXt+h−1(i,j−1)+bi;
  4. Compute an approximation of f(Xt+h|Xt,ai,bi,αi,θi) as fˆ(Xt+h|Xt=x,ai,bi,αi,θi)=nxh/B, where nxh denotes the number of draws Xt+h(i,j),j=1,…,B, equal to x;

  5. For each i (i=1,…,m) sample Xt+h,i from fˆ(Xt+h|Xt,ai,bi,αi,θi), using the inverse transform method adapted to integer variables, that is,

    1. sample u from uniform (0,1),
    2. calculate the least integer-valued v:∑j=0vfˆ(Xt+h=j|Xt,ai,bi,αi,θi)≥u,
    3. consider Xt+h,i=v.

Note that the steps 3 and 4 have been applied for the approximation of f(xt+h|xt,α,θ,φ,λ), where the other steps are used to estimate Xt+h from f(xt+h|xt).

Based on sampling Xt+h,1,…,Xt+h,m, the h-step ahead predictor of Xt+h can be calculated from the sample mean Xˆt+h=(1/m)∑i=1mXt+h,i.

The time series plots of the Gibbs sampling chains of parameters and the running mean plots which are the plots of the iterations against the mean of the draws up to each iteration are given in Figures 7–10. These plots indicate that the Gibbs sampling chains are well mixed. Figures 11 and 12 show the posterior densities of the parameters for each real data set.

Figure 7.

Figure 7.

Time series plots of the parameters chain in MCMC (robbery data).

Figure 8.

Figure 8.

Time series plots of the parameters chain in MCMC (Giardiasis data).

Figure 9.

Figure 9.

Mean plots of the parameters (robbery data).

Figure 10.

Figure 10.

Mean plots of the parameters (Giardiasis data).

Figure 11.

Figure 11.

Posterior densities of the parameters (robbery data).

Figure 12.

Figure 12.

Posterior densities of the parameters (Giardiasis data).

4.4.3. Sieve bootstrap approach

The classical and Bayesian predictions do not preserve the integer-valued nature of the data in generating forecasts when the time series is integer-valued. To preserve the integer-valued nature of data, we use a bootstrap approach as a distribution free alternative. Some bootstrap approaches have been proposed. Among others, we employ the bootstrap method proposed by Pascual et al. [10] after some modifications to ADSDINAR(1) model as the following steps.

  1. Estimate the generalized binomial thinning parameters (α,θ) with the YW estimators;

  2. Compute residuals εˆt=Xt−αˆXt−1, for t=2,…,n;

  3. Construct the empirical distribution for modified residuals ε~t defined by ε~t=[εˆt], where [.] represents the value rounded to the nearest integer;

  4. For b=1,…,B, define the bootstrap series Xtb by
    Xtb=αˆ⋄θXt−1b+εtb,
    where εtb for t=1,2,…,n is an i.i.d. sample from the residuals computed previously with X0b=0;
  5. Based on Xtb, compute the YW estimates of the parameters (αˆb,θˆb), then replace the theoretical moments by the sample counterparts: the sample mean and variance and solve the following equations
    EXt(1−α)=EεtVar(Xt)=α(θ−α)E2(Xt)+α(1−θ)E(Xt)+αθVar(Xt)+Var(εt)
  6. Estimate (α,θ) by considering the sample means αˆ=(1/B)∑i=1Bαˆib and θˆ=(1/B)∑i=1Bθˆib;

  7. Compute future bootstrap observations by recursion
    Xˆt+hb=∏i=0h−1At+i∗∘Xtb+∑i=1h−1∏l=1iAt+h−l∗∘εt+h−ib+εt+hb,
    where A∗ is a discrete random variable with probability mass function given by P(A∗=θˆ)=1−P(A=0)=αˆ/θˆ.

In order to evaluate and compare the different prediction methodologies, we compute h-step ahead forecasts ( h=1,2,…,10) and use the symmetric mean absolute percentage error (SMAPE) which is well defined if the observed value is zero, given by

SMAPE=1H∑h=1H2∣Xn+h−Xˆn+h∣Xn+h+Xˆn+h,

where H represents the number of predictions realized. The classical, Bayesian and SB forecasts of the data series are presented in Tables 6 and 7. In general, it can be noted that SMAPE values of Bayesian point predictors are smaller than classical and bootstrap predictors, but bootstrap predictors are integer, same as the nature of the real data.

Table 6. h-step ahead predictions of robbery data.
    Point prediction  
h Real data Bootstrap Classical Bayesian
1 0 1 2.0028 1.5454
2 0 1 2.2882 1.6363
3 1 1 2.3694 1.5051
4 2 3 2.3925 1.5956
5 2 2 2.3991 1.5657
6 2 3 2.4009 1.3737
7 0 1 2.4015 1.4747
8 1 3 2.4016 1.4646
9 0 2 2.4016 1.4141
10 1 0 2.4017 1.5151
  SMAPE 1.18 1.1003 1.0029
Table 7. h-step ahead predictions of Giardiasis data.
    Point prediction  
h Real data Bootstrap Classical Bayesian
1 4 2 2.7569 2.7373
2 2 2 2.7208 2.3535
3 2 4 2.7154 2.2323
4 1 6 2.7146 2.0505
5 2 4 2.7145 2.2222
6 10 4 2.7145 2.8383
7 5 2 2.7144 2.6262
8 4 1 2.7143 2.8828
9 2 3 2.7143 2.202
10 1 5 2.7143 2.2727
  SMAPE 0.8076 0.5551 0.4399

To check the adequacy and predictability of the selected model, ADSDINAR(1), the two data sets and their predicted values are plotted and displayed in Figures 13 and 14, where the CML estimates of the model parameters are given in Tables 3 and 4. It can be seen that the fitted values are closed to the original data series, which indicates that the ADSDINAR(1) model can provide a good forecasting for the data series. As it can be seen, there are some outliers in Giardiasis time series and the fitted data does not capture all the variability of Giardiasis time series.

Figure 13.

Figure 13.

Real data (–) with its predicted values (–) for robbery data.

Figure 14.

Figure 14.

Real data (–) with its predicted values (–) for Giardiasis data.

5. Conclusion

In this paper, we have developed dependent counting INAR model with serially dependent innovations for modelling overdispersed counts. The practical prominence of the model was also confirmed by two real data sets. As the classical and Bayesian predictions do not coherently preserve the integer nature of the data, the bootstrap approach is also used as a distribution free alternative.

Disclosure statement

No potential conflict of interest was reported by the author(s).

References

  • 1.Al-Osh M.A. and Aly E-E.A.A., First order autoregressive time series with negative binomial and geometric marginals, Commun. Stat. - Theor. M. 21 (1992), pp. 2483–2492. doi: 10.1080/03610929208830925 [DOI] [Google Scholar]
  • 2.Al-Osh M.A. and Alzaid A.A., First-order integer-valued autoregressive (INAR(1)) process, J. Time. Ser. Anal. 8 (1987), pp. 261–275. doi: 10.1111/j.1467-9892.1987.tb00438.x [DOI] [Google Scholar]
  • 3.Alzaid A.A. and Al-Osh M.A., First-order integer-valued autoregressive (INAR(1)) process: distributional and regression properties, Stat. Neerl. 42 (1988), pp. 53–61. doi: 10.1111/j.1467-9574.1988.tb01521.x [DOI] [Google Scholar]
  • 4.Bakouch H.S., Mohammadpour M., and Shirozhan M., A zero-inflated geometric INAR(1) process with random coefficient, Appl. Math. 63 (2018), pp. 79–105. doi: 10.21136/AM.2018.0082-17 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Czado C., Gneiting T., and Held L., Predictive model assessment for count data, Biom. J. 65 (2009), pp. 1254–61. [DOI] [PubMed] [Google Scholar]
  • 6.Gorgi P., Integer-valued autoregressive models with survival probability driven by a stochastic recurrence equation, J. Time. Ser. Anal. 39 (22018), pp. 150–171. doi: 10.1111/jtsa.12272 [DOI] [Google Scholar]
  • 7.Miletić Ilić A.V., Ristić M.M., Nastić A.S., and Bakouch H.S., An INAR(1) model based on a mixed dependent and independent counting series, J. Stat. Comput. Simul. 88 (2018), pp. 290–304. doi: 10.1080/00949655.2017.1388380 [DOI] [Google Scholar]
  • 8.Mohammadpour M., Bakouch H.S., and Shirozhan M., Poisson-Lindley INAR(1) model with applications, Braz. J. Probab. Stat. 32 (2016), pp. 262–280. doi: 10.1214/16-BJPS341 [DOI] [Google Scholar]
  • 9.Nastić A.S., Ristić M.M., and Miletić Ili ć A.V., A geometric time series model with an alternative dependent Bernoulli counting series, Commun. Stat. - Theor. M. 46 (2017), pp. 770–785. doi: 10.1080/03610926.2015.1005100 [DOI] [Google Scholar]
  • 10.Pascual L., Romo J., and Ruiz E., Bootstrap predictive inference for ARIMA processes, J. Time. Ser. Anal. 25 (2004), pp. 449–465. doi: 10.1111/j.1467-9892.2004.01713.x [DOI] [Google Scholar]
  • 11.Qin J. and Lawless J., Empirical likelihood and general estimating equations, Ann. Stat. 22 (1994), pp. 300–325. doi: 10.1214/aos/1176325370 [DOI] [Google Scholar]
  • 12.Ristić M.M., Bakouch H.S., and Nastić A.S., A new geometric first-order integer-valued autoregressive (NGINAR(1)) process, J. Stat. Plan. Inference 139 (2009), pp. 2218–2226. doi: 10.1016/j.jspi.2008.10.007 [DOI] [Google Scholar]
  • 13.Ristić M.M., Nastić A.S., and Miletić Ili ć V.A., A geometric time series model with dependent Bernoulli counting series, J. Time. Ser. Anal. 34 (2013), pp. 423–446. doi: 10.1111/jtsa.12023 [DOI] [Google Scholar]
  • 14.Shirozhan M. and Mohammadpour M., A new class of INAR(1) model for count time series, J. Stat. Comput. Simul. 88 (2018a), pp. 1348–1368. doi: 10.1080/00949655.2018.1431237 [DOI] [Google Scholar]
  • 15.Shirozhan M. and Mohammadpour M., An INAR(1) model based on the Pegram and thinning operators with serially dependent innovation, Commun. Stat. - Simul. Comput. (2018b). Available at 10.1080/03610918.2018.1521975. [DOI] [Google Scholar]
  • 16.Weiß C.H., Thinning operations for modeling time series of count-a survey, ASTA-Adv. Stat. Anal. 92 (2008), pp. 319–341. doi: 10.1007/s10182-008-0072-3 [DOI] [Google Scholar]
  • 17.Weiß C.H., The INARCH(1) model for overdispersed time series of counts, Commun. Stat. - Simul. Comput. 39 (2010), pp. 1269–1291. doi: 10.1080/03610918.2010.490317 [DOI] [Google Scholar]
  • 18.Weiß C.H., A poisson INAR(1) model with serially dependent innovations, Metrika, 78 (2015), pp. 829–851. doi: 10.1007/s00184-015-0529-9 [DOI] [Google Scholar]

Articles from Journal of Applied Statistics are provided here courtesy of Taylor & Francis

RESOURCES