 Open Journal of Applied Sciences, 2013, 3, 1-6 doi:10.4236/ojapps.2013.31B1001 Published Online April 2013 (http://www.scirp.org/journal/ojapps) Changepoint Analysis by Modified Empirical Likelihood Method in Two-phase Linear Regression Models Hualing Zhao1, Hanfeng Chen2, Wei Ning2 1Department of Statistics , School of Science, Wuhan University of Technology, Wuhan, Hubei , P.R. of China 2Department of Mathematics and Statistics, Bowling Green State University, Bowling Green, Ohio , USA Email: hualingbo324@126.com, hchen@bgsu.edu, wning@bgsu.edu. Received 2013 ABSTRACT A changepoint in statistical applications refers to an observational time point at which the structure pattern changes during a somewhat long-term experimentation process. In many cases, the change point time and cause are documented and it is reasonably straightforward to statistically adjust (homogenize) the series for the effects of the changepoint. Sadly many changepoint times are und ocumented and the changepo int times themselves are the main purpose of study. In this article, the changepoint analysis in two-phrase linear regression models is developed and discussed. Following Liu and Qian (2010)'s idea in the segmented linear regressio n models, the modified empirical likeliho od ratio statistic i s proposed to test if there exists a changepoint during the long-term experiment and observation. The modified empirical likelihood ratio statistic is computation-friendly and its -value can be easily approximated based on the large sample properties. The procedure is applied to the Old Faithful geyser eruption data in October 1980. Keywords: Changepoint; Extreme-Value Distribution; Modified Empirical Likelihood Ratio; Segmented Linear Regression 1. Introduction In recent years increasing interest has been shown in change- point analysis in two-phrase linear regression models. A changepoint in statistical applications refers to an obser-vational time point at which the structure pattern changes during a somewhat long-term experimentation process. In many cases, the change point time and cause is docu-mented and it is reasonably straightforward to statisti-cally adjust (homogenize) the series for the effects of the changepoint. Sadly many changepoint times are un-documented and the changepoint times themselves are the main interest of study. For example, one of the most important problems in economics is to d etermine as early as possible the starting as well as ending time point of a suspected ongoing recession. In the environmental sci-ences, scientists are of great interest to understand when the global warming started or the Earth's mean surface temperature rise in the past decades should be explained by the normal variability of the Earth's surface tempera-ture over time. (Indeed the official position of the World Natural Health Organization in regards to global warm-ing is that there is no global warming and claims that global warming is nothing more than just another hoax. See their official website: http://www.wnho.net.) The two-phrase linear regression model may be ex-pressed as follows: (1) where pixR are covariates,  and  are p-di- mensional regression parameters, is a putative changepoint at which the liner regression model changes from one phrase to another, and the 1kn are assumed to be independent and identically distributed unobservable measurement errors. The main interest in the two-phrase linear regression model is to determine whether such a change of phrase occurs or not and if it does, when the change happens during the experiment or observation. In the special case of simple linear regression, the model (1) is often called segmental linear regression model. As remarked by Liu and Qian (2010), widespread applica-tions of two-phrase linear regression model (1) have ap-peared in diverse research areas. See, e.g., in environ-mental sciences, Piegors ch and Bailer (1997), in medical science, Simith and Cook (1980), in epidemiology Pastor and Gullar (1998), in econometrics Fiteni (2004) and Koul and Qian (2002) , just for a few. As described above, with responses yi and covariates xi, central to the problem is to determine whether there exists a changepoint during the long-term experiment or obser-vation. In terms of statistical inference, that is to test 0:H versus 1:H Copyright © 2013 SciRes. OJAppS H. L. ZHAO ET AL. 2 Put 1,, ,nyy y,, and .1,,n Let where , . Then the model(1) has the matrix expression: (2) Dong (2004) proposes an empirical likelihood-type Wald statistic to infer the changepoint. More recently, Liu and Qian (2010 ) proposes an interesting and compu-tationally easy empirical likelihood detecting procedure in the segmented linear regression model. In this paper, their ideas are applied to the model (1) to present a modi-fied empirical likelihood ratio statistic to test . The article is organized as follows. The modified em-pirical likelihood ratio test procedure and its computa-tional issues are present and discussed in the next section. The null distribution of the modified empirical likelihood ratio test statistic is studied for large samples and the results are put in the Appendix for interested readers. The modified empirical likelihood method is applied to a real-life data set for changepoint analysis in Section 3. 2. The Modified Empirical Likelihood Method Following Liu an d Qian (2010)'s ideas, the modified em-pirical likelihood method for changepoint analysis in the two-phrase linear regression mode(1) is described as fol-lows: For each given k, estimate the regression parame-ters by least-square methods for each segment, fit the response at via the least-square estimate of the regression parameters for the segment of counter-part, and then construct the empirical likelihood ratio statistic based the fitting residuals. In the notations introduced in the last section, the least-square estimates for and are ; where and . Define (3) for and (4) The modified empirical likelihood ratio statistic is (5) Recall that is the dimension of the covariates, so equal to the number of regression parameters in each phrase. Reject the null hypothesis and as-sert that a changepoint occurs, whenever is signifi-cantly large. It should be noted that the residuals are not the ordinary least-squares fitting residuals but the residuals of fitting at with swapped least-square estimates of the regression parameters. Motivation leading to the modified empirical likelihood ratio statistics is that if and only if , i.e. holds. Through simulation studies, Liu and Qian (2010) investigate whether has an asymptotic Gumbel extreme value distribution under the null hypothesis. We establish the null asymptotic theory of that is given in the Appendix for interested readers. It is proved under regular conditions that if the null hypothesis is true, can be approximated by in probabil-ity with an approximation error in size for some constant , where (6) with It is then shown that for any t, where and . Thus for any t, (7) The above formula indicates that the limiting ex-treme-value distribution has a convergence rate of . For this reason, the authors suggest to use the distribution of under null hypothesis to approximate the p- value of in applications. As the asymptotic null distribution is free of any population distribution, one can easily approximate the p-value of by Monte Carlo methods through simulating the null distribution of . The main advantage of the modified empirical likeli-hood testing procedure based on is its easiness of computation. The package can be used to compute As many researchers remarked (Liu and Qian, 2010; and , 1997), the statistic is sensitive to outliers when k is too small or too close to the sample size. and (1997) proposed the trimmed idea to overcome the problem. Let . Define when it is assumed that as Copyright © 2013 SciRes. OJAppS H. L. ZHAO ET AL. Copyright © 2013 SciRes. OJAppS 3According the asymptotic null distribution discussed in last section, the p-value with the observed is approximately that is very close to 0, leading to the assertion that there exists a change-point during the 270 eruptions of the Old Faithful geyser in October 1980. we have when and are chosen to be constant, . Liu and Qian (2010) suggests to use and . Such a choice clearly satisfies. Another popu-lar choice is , ; see Perron and Vogelsang (1992). In p ar ticu l ar, if for , and where is the greatest integer less than or equal to x, by Corollary A.3.1 of and 4. Acknowledgements The research is partially funded by the Fundamental Re-search Funds for the Central Universities (No. 2011-IV- 116). REFERENCES 3. A Real-Life Example  M. Csorgo and L. Horvnth, “Limit Theorem in Change-Point Analysi,” Wiley Series in Probability and Statistics, John Wiley & Sons: New York, 1997. We now apply the modified empirical likelihood method to the Old Faithful geyser in the Yellowstone National Park of USA. A geyser is a hot spring that occasionally becomes unstable and erupts hot water and steam into air. If we can find the relationship between the duration of the eruptions and the interval to next eruption, then the time of next eruption can be predicted. The data of eruptions of the Old Faithful geyser in October 1980 can be found in Weisberg (2005). Figure 1 is the scatterplot of intervals(y) to the next eruption versus the duration(X) of the eruptions. The scatter plot suggests that the rela-tionship has two phases.  I. Fiteni, “-estimators of Regression Models with Structural Change of Unknown Location,” Journal of Econometrics , Vol. 119, No. 1, 2004, pp. 19-44. doi:10.1016/S0304-4076(03)00153-2  Z. Liu and L. Qian, “Changepoint Estimation in a Seg-mented Linear Regression via Empirical Likelihood,” Communications in Statistics--Simulation and Computa-tion, Vol. 89, 2010, pp. 85-100.  L. H. Koul and L. F. Qian, “Asymptotics of Maximum Likelihood Estimator in a Two-phaselinear Regression Model,”Journal of Statistical Planning and Inference, Vol. 108, No. 1-2, 2002, pp. 99-119. doi:10.1016/S0378-3758(02)00273-2 In this example, and . We adopt and . Thus , so that , and .The function in R package  A. B. Owen, “Empirical Likelihood for Linear Models,” Annals of Statistics, Vol.19, No.19, 1991, pp. 1725-1747. doi:10.1214/aos/1176348368  A. B. Owen, “Empirical Likelihood,” New York: Chap-man & Hall, 2001. doi:10.1201/9781420036152  R. Pastor and E. Guallar, “Use of Two-segmented Logis-tic Regression to Estimate Changepoints in Epidemi-ologic Studies,” American Journal of Epidemiology, Vol. 148, No. 7, 1998, pp. 631-642. doi:10.1093/aje/148.7.631  P. Perron and T. J. Vogelsang, “Testing for a Unit Root in a Time Series with a Changing Mean: Corrections and Extensions,” J. Business Econom. Statist.,Vol. 10,1992, pp. 467-470.  W. W. Piegorsch and A. J. Bailer,“Statistics for Envi-ronmental Biology and Toxicology,” London: Chapman and Hall, 1997.  A. M. F. Smith and D. G. Cook,“Straight Lines with a Change Point: A Bayesian Analysis of Some Renal Transplant Data,” Applied Statistics, Vol. 29, No. 2, 1980，pp. 180-189. doi:10.2307/2986304 Figure 1. Scatter plot of 270 eruptions of the Old Faithful geyser in October 1980 in Yellowstone National Park USA. is used to compute the test statistics and it appears that . Thus  S. Weisberg,“Applied Linear Regeression,” 3th Edition, John Wiley& Sons, Inc., Hoboken, New Jersey, 2005. H. L. ZHAO ET AL. 4 Appendix: Asymptotic Null Distribution The asymptotic null distribution of the modified empirical likelihood ratio test statistic is established under the two-phrase linear regression model (1) that includes the segmented simple linear regression model considered by Liu and Qian (2010) as a special case. With 's, is defined by (4), and by Lagrange multiplier method, , where is the root of (8) According to (5), is defined as follows: (9) Regular conditions needed are listed as follows. Assume C.1 rank =rank for . C.2 There are some , and , and positive-definite matrices , such that as and (10) (11) (12) where , and is the ordinary norm: C.3 There is some such that , and . Assumption C.2 is slightly weaker than C.9 in and (1997, page 204) that assumes . In the two-phrase linear regression model, one is concerned with a slicing rule in the covariate variables. As a result, and may have different limits if existing. In the commonly adapted regression model that 's are an independent and identically distributed sample with for some , it is easily seen that C.2 and C.3 hold in probability one. Theorem 1. Assume that hold and C.1-C.3 are satisfied with some . Then under the null model, (13) for any t, where . The main idea of proof of Theorem 1 is to use Owen (1991)'s arguments to obtain a quadratic approximation to so that the limit (1) follows from that of the classic parametric likelihood ratio test. The crucial step in Owen (1991)'s arguments is to approximate up to an order of uniformly in in r to capture the leadin the Taylor's expansion of ordeing terms . The first lemma gives an order estimate for . Denote and . Lemma 1 Assume that and C.1-C.3 h old. Then both and Proof. Under have the sat me mean. Le an d Under , we can express (14) By C.3, Thus from Lemmin01), a 11.2 Owen (20 implying ) Next, by C.2 and the law of the iterated (15 logarithm, (6) Therefore,1 by C.3 and (16) (17) Similarly, (18) The lemma follows by (14),(15),(17),(18). Lemma 2 . Assume that and and C.1-C.3 hold. Then (a) and in probability, (b) as , we have Furthermore, if Proof. Let , and , Under , Copyright © 2013 SciRes. OJAppS H. L. ZHAO ET AL. 5so that (25)and (26) become (19) By C.2 and the law of iterated logarithm, Thus, (20) Similarly, (21) Combining (19),(20),(21), we have The part (a) is proved. Next consider . We may write (22) By C.2 and the law of iterated logarithm, we have (23) Similarly, (24) By C.2 and the the law of iterated logarithm again, (25) (26) It is clear that for , (27) (28) The part (b) follows from(22-24),(27),(28). The pris complete. oof Lemma 3. Assume that and C.1-C.3 hold. Then for some Proof. Since solves (8), similarly to Owen (2), we consider 001 , So by Lemma 1, for some (29) By Lemma 2, (30) Therefore by (29),(30), (31) Now let By (31) and Lemma 1, itfollows that Using Taylor expansion, (32) where , By Lemma(31) 1 and (33) Therefore, by (31), Lemmas 1 and 2, we couniformly in k, nclude that (34) Copyright © 2013 SciRes. OJAppS H. L. ZHAO ET AL. Copyright © 2013 SciRes. OJAppS 6 The lemma follows from Lemma 2, (32) and (34), with any Proof of Theorem 1. First, we use Lemmas 1, 2 and 3 to obtain a quadratic approximation to , uniformly in .(38) Combining (35),(36),(37),(38) yields that for any Following Owen (2001, page 221)'gu s ar-ments, denote . Using Taylor's expan-sion, , (39) Now applying Taylor expansion (35) , where uniformly in k as argue some d in (33). By Lemmas 1 and 3, for, we have for any (40) Denote i.e., Using the same arguments to the proof oorem 3.1.2 of f The and (1997), we have (36) Next by Lemma 3, for some t, for all Since (37) and , it follows from(40) that The proof is complete.