 Research
 Open Access
 Published:
Test for parameter changes in generalized random coefficient autoregressive model
Journal of Inequalities and Applications volume 2014, Article number: 309 (2014)
Abstract
In this paper, we study the problem of testing for parameter changes in generalized random coefficient autoregressive model (GRCA). The testing method is based on the monitoring scheme proposed by Na et al. (Stat. Methods Appl. 20:171199, 2011), and the test statistic relies on the conditional leastsquares estimator of an unknown parameter. Furthermore, under mild conditions, we obtain the asymptotic property of the test statistic. Some simulation studies are also conducted to investigate the finite sample performances of the proposed test.
MSC: 62M10, 91B62.
1 Introduction
Consider the following oneorder generalized random coefficient autoregressive model (GRCA(1)):
where {({\mathrm{\Phi}}_{t},{\epsilon}_{t})}^{\tau} is a random vector with E\left(\begin{array}{c}{\mathrm{\Phi}}_{t}\\ {\epsilon}_{t}\end{array}\right)=\left(\begin{array}{c}{\varphi}_{t}\\ 0\end{array}\right) and Var\left(\begin{array}{c}{\mathrm{\Phi}}_{t}\\ {\epsilon}_{t}\end{array}\right)=\left(\begin{array}{cc}{V}_{\varphi ,t}& {\sigma}_{\mathrm{\Phi}\epsilon ,t}\\ {\sigma}_{\mathrm{\Phi}\epsilon ,t}& {\sigma}_{\epsilon ,t}^{2}\end{array}\right). In addition, {({\mathrm{\Phi}}_{t},{\epsilon}_{t})}^{\tau} is assumed to be independent of {\mathcal{F}}_{t1}=\sigma ({Y}_{t1},{Y}_{t2},\dots ).
The model (1.1) was first introduced by Hwang and Basawa [1]. When {\mathrm{\Phi}}_{t} and {\epsilon}_{t} are mutually independent, the model (1.1) becomes the random coefficient autoregressive model (RCAR), and when {V}_{\varphi}=0, the model (1.1) becomes the usual autoregressive model. Furthermore, the model (1.1) also includes the Markovian bilinear model (see, e.g., [2, 3]), the generalized Markovian bilinear model, and the random coefficient exponential autoregressive model (see [4] for more information) as special cases.
GRCA is designed for investigating the result of random perturbations of a dynamical system in engineering and economic data, and it has become one of the important models in the nonlinear time series context. In several recent years, GRCA has been studied by many authors. For instance, Hwang and Basawa [5] established the local asymptotic normality of a class of generalized random coefficient autoregressive processes. Lee [6] studied the problem of testing the constancy of the coefficient. Moreover, Carrasco and Chen [7] provided tractable sufficient conditions that simultaneously imply strict stationarity, finiteness of higherorder moments, and βmixing with geometric decay rates. In this paper, we consider the problem of testing for parameter changes in GRCA.
The changepoint problem has a long history and began with i.i.d. samples (see, e.g., [8–13]). Observing that time series often suffer from structural changes, statisticians started the study of the changepoint problem for economic time series models (see, e.g., [14, 15]). Recently, the changepoint problem has become very popular in economic time series. Lee and Park [16] considered the monitoring process in time series regression models with nonstationary regressors. Gombay and Serban [17] proposed sequential tests to detect an abrupt change in any parameter, or in any collection of parameters of an autoregressive time series model. By using the cumulative sum test, Kang and Lee [18] studied the problem of testing for a parameter change in a firstorder random coefficient integervalued autoregressive model. Moreover, Na et al. [19] also designed the monitoring procedure in general time series models and applied it to the changes of the autocovariances of linear processes, GARCH parameters, and underlying distributions.
In order to monitor the parameter changes in generalized random coefficient autoregressive model, we employ the monitoring scheme proposed by Na et al. [19]. The test statistic relies on the conditional leastsquares estimator of an unknown parameter, and under mild conditions we also obtain the asymptotic property of the test statistic.
The rest of this paper is organized as follows: In Section 2, we introduce the methodology and the main results. Simulation results are reported in Section 3. A real data analysis is given in Section 4. Section 5 provides the proofs of the main results.
Throughout this paper, we denote pdimensional standard Brown motion by \{{W}_{p}(s),s\ge 0\}. The symbols ‘\stackrel{d}{\to}’ and ‘\stackrel{p}{\to}’ denote convergence in distribution and convergence in probability, respectively, and convergence ‘almost surely’ is written as ‘a.s.’.
2 Methodology and main results
Na et al. [19] proposed a monitoring scheme of detecting parameter changes for general time series models. In what follows, we will first introduce this monitoring procedure briefly, and then we give our main results.
Let \{{Y}_{t}\} be a time series and Θ be unknown parameter which will be examined for the parameter constancy. We assume that Θ is a constant for the historical data {Y}_{1},{Y}_{2},\dots ,{Y}_{T}. Here, we wish to test the following hypotheses based on the estimator {\stackrel{\u02c6}{\mathrm{\Theta}}}_{T} of Θ:
{\mathrm{H}}_{0}: Θ does not change over time t>T versus {\mathrm{H}}_{1}: Θ changes at some time t>T.
We assume that, under {\mathrm{H}}_{0}, {\stackrel{\u02c6}{\mathrm{\Theta}}}_{T} is a consistent estimator of Θ based on {Y}_{1},{Y}_{2},\dots ,{Y}_{T}, its asymptotic variancecovariance matrix Σ is nonsingular and {\stackrel{\u02c6}{\mathrm{\Sigma}}}_{T} is the consistent estimator of Σ based on {Y}_{1},{Y}_{2},\dots ,{Y}_{T}. We can then define
where {\stackrel{\u02c6}{\mathrm{\Theta}}}_{k} is the estimator of Θ at time lag k>T based on {Y}_{1},{Y}_{2},\dots ,{Y}_{k}, \parallel \cdot \parallel denotes a norm on {R}^{p}, and g(s) (s\in (1,\mathrm{\infty})) is a given boundary function.
If \tau (T) is finite, we reject {\mathrm{H}}_{0}. In actual practice, if there exists k\in (T,T+q), with q=2T,3T,5T,10T etc., such that \parallel {\stackrel{\u02c6}{\mathrm{\Sigma}}}_{T}^{\frac{1}{2}}({\stackrel{\u02c6}{\mathrm{\Theta}}}_{k}{\stackrel{\u02c6}{\mathrm{\Theta}}}_{T})\parallel \ge \frac{1}{\sqrt{T}}g(\frac{k}{T}), then we reject {\mathrm{H}}_{0}. Meanwhile, the boundary function g is chosen to satisfy
for a given \alpha \in (0,1) and
Further, under {\mathrm{H}}_{0}, suppose that the estimator {\stackrel{\u02c6}{\mathrm{\Theta}}}_{k} (k=T,T+1,\dots) satisfies the following conditions:
(A_{1}) {\stackrel{\u02c6}{\mathrm{\Theta}}}_{k} can be decomposed as
where {\mathrm{\Theta}}_{0} is a true value of Θ under {\mathrm{H}}_{0}, \{{l}_{t}={({l}_{1t},\dots ,{l}_{pt})}^{\tau},t\ge 1\} is a sequence of pdimensional random vectors and {\mathrm{\u25b3}}_{k}={({\mathrm{\u25b3}}_{1k},\dots ,{\mathrm{\u25b3}}_{pk})}^{\tau} are negligible terms.
(A_{2}) There exists a pdimensional standard Brownian motion \{{W}_{p}(s),s\ge 0\} such that, for some 0<\lambda <\frac{1}{2},
(A_{3}) For each 1\le i\le p, \sqrt{T}{sup}_{k\ge T}{\mathrm{\u25b3}}_{i,k}={o}_{p}(1).
Remark 1 Condition (A_{2}) holds for zero mean stationary martingale difference sequence \{{l}_{t},t\ge 1\} and at this point \mathrm{\Sigma}=E({l}_{1}{l}_{1}^{\tau}) (see [20, 21]).
Based on the above conditions, we have the following lemmas.
Lemma 2.1 Suppose that (A_{1})(A_{3}) hold. If g(s)=c{g}_{1}(s), s\in (1,\mathrm{\infty}), c is a positive constant and {g}_{1} is a given continuous realvalued function with {inf}_{s\in (1,\mathrm{\infty})}{g}_{1}(s)>0, then
Particularly, if g(s)=c, then
Lemma 2.1 is due to Na et al. [19]. Moreover, (A_{1})(A_{3}) and the consistency of {\stackrel{\u02c6}{\mathrm{\Sigma}}}_{T} indicate that
as T\to \mathrm{\infty}. Thus, by using Theorem 3.4 of Chu et al. [22], we have the following lemma.
Lemma 2.2 Suppose that (A_{1})(A_{3}) hold. If g(s)=\sqrt{\frac{s1}{s}({e}^{2}+ln(\frac{s}{s1}))}, s\in (1,\mathrm{\infty}) and {\parallel \cdot \parallel}_{\mathrm{\infty}} is the maximum norm, then
where e is a constant, ϕ and Φ denote the standard normal density and distribution functions, respectively.
Below, we use the above method to detect parameter changes in the model (1.1). Suppose that {\sigma}_{\epsilon ,t}^{2}={\sigma}_{\epsilon ,0}^{2}, {\sigma}_{\mathrm{\Phi}\epsilon ,t}={\sigma}_{\mathrm{\Phi}\epsilon ,0}, {V}_{\varphi ,t}={V}_{\varphi ,0}, t=1,2,\dots . Moreover, we assume that {\varphi}_{t}={\varphi}_{0} for t=1,\dots ,T. Consider the following hypothesis test:
{\mathrm{H}}_{0}: {\varphi}_{t}={\varphi}_{0}, t>T versus {\mathrm{H}}_{1}: {\varphi}_{t} changes at some t>T.
Before we state our main results, we list some regular conditions used in this paper.
(C_{1}) The distributions of {\mathrm{\Phi}}_{t} and {\epsilon}_{t} are absolutely continuous with respect to the Lebesgue measure on {R}^{1} and their densities are strictly positive on some neighborhood of 0.
(C_{2}) {\theta}_{0}={\varphi}_{0}^{2}+{V}_{\varphi ,0}<1.
(C_{3}) E({\mathrm{\Phi}}_{t}^{6})<1 and E({\epsilon}_{t}^{6})<\mathrm{\infty}.
Remark 2 It is shown by Theorem 2.1 of Hwang and Basawa [1] that, under (C_{2}), \{{Y}_{t},t\ge 1\} is stationary and ergodic.
In order to establish the test statistic we need to obtain the consistent estimator of parameter {\varphi}_{0} under {\mathrm{H}}_{0}. Below we further assume that conditions (C_{1})(C_{3}) hold. Based on the recorded data \{{Y}_{0},\dots ,{Y}_{k}\}, the conditional leastsquares estimator {\stackrel{\u02c6}{\varphi}}_{k} of {\varphi}_{0} is obtained by minimizing
with respect to ϕ. Substituting E({Y}_{t}{Y}_{t1})=\varphi {Y}_{t1} in S and solving dS/d\varphi =0 for ϕ, we obtain
Under the conditions (C_{2}) and (C_{3}), the estimator {\stackrel{\u02c6}{\varphi}}_{k} is consistent and asymptotically normal, and its asymptotic variance J={\sigma}_{\epsilon ,0}^{2}{(1{\theta}_{0})}^{2}({\sigma}_{\epsilon ,0}^{2}E{Y}_{0}^{2}+2{\sigma}_{\mathrm{\Phi}\epsilon ,0}E{Y}_{0}^{3}+({\theta}_{0}{\varphi}_{0}^{2})E{Y}_{0}^{4}) (see Hwang and Basawa [1]).
In what follows, we consider an estimate of J. By the ergodic theorem, it is easy to find that, for any integer l>1, \frac{1}{n}{\sum}_{t=1}^{n}{Y}_{t1}^{l} is the consistent estimator of E{Y}_{0}^{l}. Therefore, in order to obtain the consistent estimator {\stackrel{\u02c6}{J}}_{k} of J, we only need to obtain the consistent estimator of \mathrm{\Upsilon}={({\theta}_{0},{\sigma}_{\epsilon ,0}^{2},{\sigma}_{\mathrm{\Phi}\epsilon ,0})}^{\tau}.
We now consider an estimate of ϒ. A conditional leastsquares estimator {\stackrel{\u02c6}{\mathrm{\Upsilon}}}_{k} of ϒ can be obtained by minimizing
where {N}_{t}={({Y}_{t1}^{2},1,2{Y}_{t1})}^{\tau}. Solving dQ/d\mathrm{\Upsilon}=0 for ϒ, we obtain
The following lemma indicates that {\stackrel{\u02c6}{\mathrm{\Upsilon}}}_{k} is the consistent estimator of ϒ.
Lemma 2.3 Suppose that (C_{2}) and (C_{3}) hold. Then, under {\mathrm{H}}_{0}, we have
After obtaining the consistent estimator of {\varphi}_{0} and J, we can establish the following test statistics:
For the test statistics {\tau}_{1}(T), we have the following results.
Theorem 2.1 Suppose that (C_{1})(C_{3}) hold.

(i)
If g(s)=c{g}_{1}(s), s\in (1,\mathrm{\infty}), c is a positive constant, and {g}_{1} is a given continuous realvalued function with {inf}_{s\in (1,\mathrm{\infty})}{g}_{1}(s)>0, then
\underset{T\to \mathrm{\infty}}{lim}{P}_{{\mathrm{H}}_{0}}\{{\tau}_{1}(T)<\mathrm{\infty}\}=P\{\underset{s\in (0,1)}{sup}\frac{\parallel {W}_{1}(s)\parallel}{{g}_{1}(1/(1s))}\ge c\}.Particularly, if g(s)=c and \parallel \cdot \parallel ={\parallel \cdot \parallel}_{2}, then
\underset{T\to \mathrm{\infty}}{lim}{P}_{{\mathrm{H}}_{0}}\{{\tau}_{1}(T)<\mathrm{\infty}\}=P\left\{\underset{s\in (0,1)}{sup}\right{W}_{1}(s)\ge c\}. 
(ii)
If g(s)=\sqrt{\frac{s1}{s}({e}^{2}+ln(\frac{s}{s1}))}, s\in (1,\mathrm{\infty}), and \parallel \cdot \parallel ={\parallel \cdot \parallel}_{\mathrm{\infty}}, then
\underset{T\to \mathrm{\infty}}{lim}{P}_{{\mathrm{H}}_{0}}\{{\tau}_{1}(T)<\mathrm{\infty}\}=22\mathrm{\Phi}(e)+2e\varphi (e).
Remark 3 If g(s)=c and \parallel \cdot \parallel ={\parallel \cdot \parallel}_{2}, we have
By (i) of Theorem 2.1, we can determine the constant c for any significance level \alpha \in (0,1). In fact, since {lim}_{T\to \mathrm{\infty}}{P}_{{\mathrm{H}}_{0}}\{{\tau}_{1}(T)<\mathrm{\infty}\}=P\{{sup}_{s\in (0,1)}{W}_{1}(s)\ge c(\alpha )\}=\alpha, where c(\alpha ) is the 1\alpha quantile point of {sup}_{s\in (0,1)}{W}_{1}(s), we thus have c=c(\alpha ). When p=1, we have c=7.78 for \alpha =0.1. Moreover, if g(s)=\sqrt{\frac{s1}{s}({e}^{2}+ln(\frac{s}{s1}))}, s\in (1,\mathrm{\infty}), and \parallel \cdot \parallel ={\parallel \cdot \parallel}_{\mathrm{\infty}}, then when p=1, we have {e}^{2}=1.488 for the nominal level \alpha =0.1.
3 Simulation results
In this section, we evaluate the performance of the monitoring test through a simulation study.
Consider the following model:
where \{{\epsilon}_{t}\} is i.i.d. normally distributed with mean 0 and variance 1.
We compare the performance of testing methods (i) and (ii) in Theorem 2.1. For the testing method (i), we take g(s)=c and \parallel \cdot \parallel ={\parallel \cdot \parallel}_{2}. In the actual simulation, we reject {\mathrm{H}}_{0} if there exists k\in (T,T+q) such that \parallel {\stackrel{\u02c6}{J}}_{T}^{\frac{1}{2}}({\stackrel{\u02c6}{\varphi}}_{k}{\stackrel{\u02c6}{\varphi}}_{T})\parallel \ge \frac{1}{\sqrt{T}}g(\frac{k}{T}), where the horizon q=2T,3T,5T\text{and}10T, and T=100,200,300,500\text{and}1\text{,}000. In each simulation, 1\text{,}000 observations are discarded to remove initialization effects and a repetition number of 1\text{,}000 is utilized.
In the first simulation, we calculate the probability of rejecting the null hypothesis when it is true at the nominal level \alpha =0.1. The results of the simulations are presented in Table 1, and the figures in parentheses are those for a constant function test.
From Table 1, we see that the test with the boundary function in (ii) of Theorem 2.1 has superiority over that with the constant boundary function. We can also see that the empirical sizes of these two tests tend to decrease as the historical sample size T increases and increase as q increases. But even if T is small or q is large, the empirical sizes of these two testing methods are still very close to the nominal level.
The second simulation study is designed to examine the power. We calculate the probability of rejecting the null hypothesis when the alternative hypothesis is true at the nominal level \alpha =0.1. To do this, we consider the alternative hypotheses as follows:
{\mathrm{H}}_{1}: A change occurs from {\varphi}_{t}=0.1, \alpha =0.3 to {\varphi}_{t}=0.7, \alpha =0.3;
{\tilde{\mathrm{H}}}_{1}: A change occurs from {\varphi}_{t}=0.7, \alpha =0.3 to {\varphi}_{t}=0.1, \alpha =0.3.
In all cases, the changes are assumed to occur at {k}^{\ast}=0.3T and {k}^{\ast}=0.8T. The results of the simulations are presented in Table 2, and the figures in parentheses are those for the constant function test.
From Table 2, we see that the probability of rejecting the null hypothesis when the alternative hypothesis is true tends to increase as T increases or q increases, and the test with the constant boundary function tends to detect changes more efficiently than the test in (ii) of Theorem 2.1. It can also be seen from Table 2 that if {\varphi}_{t} changes from a smaller number to a larger number, then these two test approaches perform very well. Although these two test approaches tend to mistakenly detect changes when {\varphi}_{t} changes from a larger number to a smaller number when T and q are small, both these two tests still perform well when T and q are large. Therefore, when we use the above methods to detect the parameter changes, we suggest that one estimate the parameter firstly. If the parameter is small, we can use a smaller q or T.
4 Real data analysis
In this section, we illustrate how our method can be applied to a practical example. Specifically, we apply the above test methods to model offence data, which are the best indicator of crimes with victims such as robbery, sex offence, homicide, drugs, burglary, arson, etc. We mainly pay attention to drugs data series. The data are available online at the forecasting principles site (http://www.forecastingprinciples.com/index.php?option=com_content&view=article&id=47&Itemid=250). An observation of the time series represents a count of drugs reported in the police car beat in Pittsburgh, during one month. The data consist of 120 observations, starting in January 1992 and ending in December 2001. The data are denoted {y}_{1},{y}_{2},\dots ,{y}_{120}. Figure 1 is the sample path plot for the real data {y}_{t}, t=1,2,\dots ,120.
The sample path plot reveals nonstationarity. Therefore, let {x}_{t}={y}_{t}{y}_{t1}. The sample path plot, the autocorrelation function (ACF), and the partial autocorrelation function (PACF) for the differenced series {x}_{t} is given in Figures 2, 3, and 4, respectively. From Figure 2, we can see that {x}_{t} is from a stationary series. From Figures 3 and 4, we can see that {x}_{t} may come from an GRCA(1) process. Therefore, we consider a model of the data series {x}_{t} by using the following:
where \{{\epsilon}_{t}\} is for i.i.d. random variables.
We assume that ϕ is a constant for the historical data {x}_{1},{x}_{2},\dots ,{x}_{50}. Then we test the following hypotheses:
{\mathrm{H}}_{0}: ϕ does not change over time t>50 versus {\mathrm{H}}_{1}: ϕ changes at some time t>50.
Based on the historical data {x}_{1},{x}_{2},\dots ,{x}_{50}, testing methods (i) and (ii) in Theorem 2.1 both accept the null hypothesis. That is to say, ϕ does not change over time t>50. Further, we use the data {x}_{1},{x}_{2},\dots ,{x}_{120} to estimate the unknown parameters ϕ and α, respectively. Therefore, we can model the data {x}_{1},{x}_{2},\dots ,{x}_{120} by using the model {x}_{t}=(0.3807+0.6794{\epsilon}_{t}){x}_{t1}+{\epsilon}_{t}, where \{{\epsilon}_{t}\} is for i.i.d. random variables.
5 Proofs of the main results
In order to prove the main results, we need some auxiliary lemmas.
Lemma 5.1 Suppose that (C_{1})(C_{3}) hold. Then, under {\mathrm{H}}_{0}, oneorder generalized random coefficient autoregressive model (1.1) is βmixing with geometric decaying order and E{Y}_{t}^{3}<\mathrm{\infty}.
Proof The proof can be found in Carrasco and Chen [7]. □
Lemma 5.2 Suppose that (C_{1})(C_{3}) hold. Then, under {\mathrm{H}}_{0}, for the oneorder generalized random coefficient autoregressive model (1.1), there exists a positive constant M such that, for any 0<d<3,
Proof This lemma can be proved by Lemma 5.1 and Theorem 5 in Kuelbs and Philipp [23]. □
Lemma 5.3 Let \{{Y}_{i},i\ge 1\} be a stationary ergodic stochastic sequence with E({Y}_{i}{Y}_{1},{Y}_{2},\dots ,{Y}_{i1})=0 a.s. for all i\ge 2 and E{Y}_{1}^{2}=1. Then {lim\hspace{0.17em}sup}_{n\to \mathrm{\infty}}{\sum}_{i=1}^{n}\frac{{Y}_{i}}{{(2nloglogn)}^{\frac{1}{2}}}=1 a.s.
Proof The proof can be found in Stout [24]. □
Proof of Lemma 2.3 Note that
By the ergodic theorem, we have
and
These, together with (5.1), imply that {\stackrel{\u02c6}{\mathrm{\Upsilon}}}_{k}\stackrel{\text{a.s.}}{\to}\mathrm{\Upsilon}, and thus the proof of Lemma 2.3 is complete. □
Proof of Theorem 2.1 We apply Lemma 2.1 to prove Theorem 2.1.
Firstly we decompose {\stackrel{\u02c6}{\varphi}}_{k} into the sum of martingale differences and a negligible term. Note that
Next, we verify that they meet the conditions of Lemma 2.1.
Observe that {(E{Y}_{1}^{2})}^{1}\{({Y}_{t}{\varphi}_{0}{Y}_{t1}){Y}_{t1},t\ge 1\} is a sequence of mean zero stationary martingale difference. Thus, there exists a 1dimensional standard Brownian motion \{{W}_{1}(s),s\ge 0\} such that, for some 0<\lambda <\frac{1}{2},
In what follows, we prove that
By Lemma 4.3, we have
From Lemma 4.2 we know that there exists a positive constant {M}_{1} such that, for any 0<d<3,
from which, together with (5.6), we have
Further, note that
By (5.8), we prove (5.5). Thus, by Lemma 2.1, we prove Theorem 2.1. □
References
Hwang SY, Basawa IV: Parameter estimation for generalized random coefficient autoregressive processes. J. Stat. Plan. Inference 1998, 68: 323–327. 10.1016/S03783758(97)00147X
Tong H: A note on a Markov bilinear stochastic process in discrete time. J. Time Ser. Anal. 1981, 2: 279–284. 10.1111/j.14679892.1981.tb00326.x
Feigin PD, Tweedie RL: Random coefficient autoregressive processes: a Markov chain analysis of stationarity and finiteness of moments. J. Time Ser. Anal. 1985, 6: 1–14. 10.1111/j.14679892.1985.tb00394.x
Hwang SY, Basawa IV: Asymptotic optimal inference for a class of nonlinear time series models. Stoch. Process. Appl. 1993, 46: 91–113. 10.1016/03044149(93)90086J
Hwang SY, Basawa IV: The local asymptotic normality of a class of generalized random coefficient autoregressive processes. Stat. Probab. Lett. 1997, 34: 165–170.
Lee S: Coefficient constancy test in a random coefficient autoregressive model. J. Stat. Plan. Inference 1998, 74: 93–101. 10.1016/S03783758(98)000950
Carrasco M, Chen X: β Mixing and moment properties of RCA models with application to GARCH(p,q) . C. R. Acad. Sci., Sér. I Math. 2000, 331: 85–90.
Page ES: Continuous inspection schemes. Biometrika 1954, 41: 100–114. 10.1093/biomet/41.12.100
Page ES: A test for a change in a parameter occurring at an unknown point. Biometrika 1955, 42: 523–527. 10.1093/biomet/42.34.523
Page ES: On problems in which a change in parameter occurs at an unknown point. Biometrika 1957, 44: 248–252. 10.1093/biomet/44.12.248
Hinkley DV: Inference about the changepoint from cumulative sum tests. Biometrika 1971, 58: 509–523. 10.1093/biomet/58.3.509
Brown RL, Durbin J, Evans JM: Techniques for testing the constancy of regression relationships over time. J. R. Stat. Soc., Ser. B 1975, 37: 149–163.
Inclan C, Tiao GC: Use of cumulative sums of squares for retrospective detection of changes of variances. J. Am. Stat. Assoc. 1994, 89: 913–923.
Wichern DW, Miller RB, Hsu DA: Changes of variance in firstorder autoregressive time series models  with an application. J. R. Stat. Soc., Ser. C 1976, 25: 248–256.
Lee S, Ha J, Na O, Na S: The cusum test for parameter change in time series models. Scand. J. Stat. 2003, 30: 781–796.
Lee S, Park S: The monitoring test for the stability of regression models with nonstationary regressors. Econ. Lett. 2009, 105: 250–252. 10.1016/j.econlet.2009.08.013
Gombay E, Serban D:Monitoring parameter change in AR(p) time series models. J. Multivar. Anal. 2009, 100: 715–725. 10.1016/j.jmva.2008.08.005
Kang J, Lee S: Parameter change test for random coefficient integervalued autoregressive processes with application to polio data analysis. J. Time Ser. Anal. 2009, 30: 239–258. 10.1111/j.14679892.2009.00608.x
Na O, Lee J, Lee S: Monitoring parameter change in time series models. Stat. Methods Appl. 2011, 20: 171–199. 10.1007/s1026001101623
Eberlein E: On strong invariance principles under dependence assumptions. Ann. Probab. 1986, 14: 260–270. 10.1214/aop/1176992626
Kuelbs J, Philipp W: Almost sure invariance principles for partial sums of mixing B valued random variables. Ann. Probab. 1980, 8: 1003–1036. 10.1214/aop/1176994565
Chu CSJ, Stinchcombe M, White H: Monitoring structural change. Econometrica 1996, 64: 1045–1065. 10.2307/2171955
Kuelbs J, Philipp W: Almost sure invariance principles for partial sums of mixing B valued random variables. Ann. Probab. 1980, 8: 1003–1036. 10.1214/aop/1176994565
Stout WF: The HartmanWintner law of the iterated logarithm for martingales. Ann. Math. Stat. 1970, 41: 2158–2160. 10.1214/aoms/1177696721
Acknowledgements
This work is supported by National Natural Science Foundation of China (Nos. 11271155, 11001105, 11071126, 10926156, 11071269), Specialized Research Fund for the Doctoral Program of Higher Education (Nos. 20110061110003, 20090061120037), Scientific Research Fund of Jilin University (Nos. 201100011, 200903278), the Science and Technology Development Program of Jilin Province (201201082), Jilin Province Social Science Fund (2012B115) and Jilin Province Natural Science Foundation (20101596, 20130101066JC).
Author information
Authors and Affiliations
Corresponding author
Additional information
Competing interests
The authors declare that they have no competing interests.
Authors’ contributions
All authors contributed equally to the writing of this paper. All authors read and approved the final manuscript.
Authors’ original submitted files for images
Below are the links to the authors’ original submitted files for images.
Rights and permissions
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (https://creativecommons.org/licenses/by/4.0), which permits use, duplication, adaptation, distribution, and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
About this article
Cite this article
Zhao, ZW., Wang, DH. & Peng, CX. Test for parameter changes in generalized random coefficient autoregressive model. J Inequal Appl 2014, 309 (2014). https://doi.org/10.1186/1029242X2014309
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/1029242X2014309
Keywords
 monitoring parameter changes
 conditional leastsquares estimation
 asymptotic variance
 generalized random coefficient autoregressive model