3 [LIBRARIES. ^V Jo 1 JMJ.T. LIBRARIES - DEWEY Digitized by the Internet Archive in 2011 with funding from Boston Library Consortium Member Libraries http://www.archive.org/details/noteonspuriousbrOObaij IHB31 @ working paper department of economics A NOTE ON SPURIOUS BREAK AND REGIME SHIFT IN COINTEGRA TING RELA TIONSHIP Jushan Bai May 1996 massachusetts institute of technology 50 memorial drive Cambridge, mass. 02139 A NOTE ON SPURIOUS BREAK AND REGIME SHIFT IN COINTEGRA TING RELA TIONSHIP Jushan 96-13 Bat May 1996 MASSACHUSETTS INSTITUTE JUL 18 1996 LSBRAHltS 96-13 A Note on Spurious Break and Regime Shift in Cointegrating Relationship by Jushan Bai 1 Massachusetts Institute of Technology 2 May, 1996 1 Partial financial support is acknowledged from the National Science Foundation under Grant SBR-9414083. 2 Please send correspondence to Jushan Bai, Department of Economics, E52-274B, Cambridge, MA 02139. Phone: (617) 253-6217. Email: jbaiQmit.edu. MIT, 96-13 Abstract The simulation result of Nunes, Kuan, and Newbold suggests that it is possible to estimate a spurious break for a regression model with 1(1) disturbances. In this note, we provide a rigorous proof for this phenomenon. We also applies to integrated regressors, so that a spurious regression break. However, if is in terms of the integer index rather than This rapid rate of convergence 1(0) regressors. may their finding lead to a spurious two integrated processes are cointegrated with a structural change in the cointegrating relationship, the break point can consistency show that is be consistently estimated. The in terms of the sample fraction. not attainable for stationary or, more generally, for Furthermore, the consistency holds even when magnitudes of breaks are small but do not converge to zero too fast. These consistency results are also obtained for a broken trend model. Key words and phrases. Spurious break, spurious regression, change point, cointe- gration, broken trend. Running head: Spurious Break Introduction 1 Newbold (1995) pointed out that when the disturbances Recently, Nunes, Kuan, and of a regression model follow an 1(1) process there is a tendency to estimate a break point in the middle of the sample, even though a break point does not exist. phenomenon is by the authors and was discovered by a called a "spurious break" simulation experiment. In this note, we provide a may We phenomenon. rigorous proof for this Furthermore, we show that a spurious break occurs a spurious regression This for 1(1) regressors as well, so that lead to a spurious break. then consider the problem in which the dependent variable and the 1(1) re- gressors are cointegrated but the cointegrating relationship undergoes a shift. This a more general notion of cointegration because the cointegrating vector invariant. We ask is not time- the question whether the break point can be consistently estimated. It is shown that the estimated break point converges quickly to the true break The rate of convergence is faster given the same magnitude of A than the corresponding result for pirical applications. may be shift. A em- Cointegration describes a long-run-equilibrium condition. An disturbed by policy regime changes, resulting in a special case which can be expressed This shift is model in rium, so that a different cointegrating vector equilibrium. point. 1(0) regressors, structural change in a cointegrating relationship can be a useful equilibrium is is a shift in the may be needed mean new equilib- to characterize this new level of the long-run equilibrium, as a shift in the intercept of a cointegrating-regression model. exhibited graphically as a change in the "gap" between two cointegrated series. Although not a concern which allows Hansen Ericsson, and Hendry is and Phillips (1993), (1996). that one we point out that testing for cointegration change has been studied by a number of authors; for a structural (1992), Quintos relationship of this paper, may One see, e.g., Gregory and Hansen (1996), and Campos, implication of a structural change in a cointegrating not be able to reject the null of no cointegration if conventional tests are used, even though a long-run relationship between two series does, in fact, exist. This situation calls for use of the test statistics proposed by the aforementioned authors. Spurious Break 2 Consider the model: Vt ( = x'tPi + x't /32 +e { { et t Let $\{k) be the least squares estimator of t = 1,2,...,* t = k + l,...,T based on the f3\ first T— $2(k) be the least squares estimator of #2 based on the last k observations, and k observations, i.e., k Pi( k ) = {J2 x t x 't) T &(*) = . E ( (J2 x tyt), _, **<)' sum E ( t=k+l Define the T . *&)• t=k+l of squared residuals for the full k sample as T 2 E sT (k) = Y,(yt-x'Mk)) + (yt-x'Mk)) t=k+l t=\ and define the break point estimator k = as , argmin 1 < < T S T (A:). fc Finally, let AT where < A < A < 1. = min{A : The behavior A = of argminue[AiI] 5T ([Tu]) Aj is considered for two cases: 1(0) and 1(1) error processes. For an 1(0) error process, let Q(X) and R(\) be defined as Kuan, and Newbold (1995, hereafter NKN), the limit of the limit of Dj DT (Y/t=i x t x't )D^ 1 In for some Q> More and [A3] of Nunes, Q(X) is an appropriate scaling matrix Dj, and R(\) is respectively. specifically, Xrp \ 1 52t=i x t^t- strictly increasing. for in [Al] The and The matrix Q(X) process R(X) -R(A) is is is Gaussian. a Brownian motion. assumed When x to t be positive- definite and is stationary, Q(X) = \Q we assume In this section, Theorem their there is no break, i.e. = j3\ NKN /?2- show that 3.1b) A T -i argmax Ae[A TiM(A) \e[A,A] where M(X) For an 1(1) error process e t - [i?(l) R(X)]'[Q(1) NKN 1 ), for some a > - we assume that T~ 2 , > being a standard Wiener process, and c [A3'] in (1) a stochastic process given by is M(A) = R{\)'Q{X)- X R{\) + W(u) (see Q^m) - J2 t =i £ 2 /2(A)]. (2) c/ A W^ 2 (u)g?u with =>• a constant. Further assume (see 0, [XA] T-*/ where G(A) is Gaussian process. 2 /^ 1/2 £> ^ G(A) (3) <£t NKN prove that A T -i argmax Ae[AjX] M*(A) (4) where M*(A) = 1 G(A)'g(A)" G(A) Examining (2) and (5) we + [G(l) find that, said is a random Andrews M(A) —> oo as A (1993). In their - ^(A)]- in the — > or Remark However, 1, thus At 1 (p. to characterize the limiting behavior of 742), — {0, 1}, if A for A near Their original assumption current form allows process. j/< is stated in terms of yt — or not diverge to infinity as A decreases to zero or increases to The (5) 1(0) or 1(1), the for /(0) error process e t NKN pointed out M*(X) is G(A)]. Not much further can be . they find that M*(A) behaves differently from M(X). More 1 ^!) - absence of a break, the estimated , [A, A]. 1 whether the error process variable with support in [A A] on the compact interval prove that G(X)]'[Q(l) Namely, results are essentially the same. break point At - rather than et, , and A NKN — > 1; further also see that they were unable 1. Through simulation, specifically, M*(X) does 1. which applies to yt being 1(1). to depend on deterministic regressors as well as on an additive 1(1) error we In the following, shall prove that M*(X) uniformly bounded in probability over of T' a MT ([TA]), where a k t t=l T k _j (E <=1 x t) Note that M*(X) [0, 1]. [0, 1] and is the limiting process is defined in (3) and is k , M£(k)= (J2 £ a well defined process on is {J2 x £ t) x<x + t E { T , e t x t) E ( t=k+l t=l T -1 x*x ( t=k+l E x < e <) t=k+l (6) We assume that a > shall 2 because this When regressor (e.g., a constant, or a trend). is possible that Theorem a For 1 = This case 1. the a when x contains a nonzero mean true is xt t is a 7(0) process with zero mean, it not considered in this paper. is defined in (3), assume a > 2. We have = Op (l). sup M*(A) (7) Ae(o,i) Proof of Theorem P, we have z' Pz < For an arbitrary vector z and an arbitrary projection matrix 1. z'z. Apply this inequality to Mj{k) to obtain T MT(k)<J2 £ a > Since T~ 2 2, J2t=i £ t M*(A), To is we have T~ a M$(k) < T~ 2 h as a hmit, T~ a Mj{k) = is rule out the possibility that and A = 1. As the and limit of for A is = obtained from 0. (5) (8) [1,T]. Moreover, because uniformly bounded in probability. Thus — Xt 1. > {0,1}, we need 0, to further M*(A) is not defined yet at M*(0) should be defined as = 0,Q(A) is = (9) 0, and C(A)'g(A)- 1 G(A) the limit of the first term T~ a Now . HEWE^r'E^) < r^e? < r- ^ 2 i=l examine the = GilYQil^Gil) Note that the term G(A)'Q(A)- 1 G(A) i=l its limit, (0, 1). Strictly speaking, by taking G{X) the right hand side divided by E for all k M*(X) when A —> M*(0) which £f=1 e] uniformly bounded in probability for A £ behavior of M*(A) for A near A forallJfc€[l,T]. t t=l t=\ 4=1 22 £t = of (6) on which converges to zero any given in probability for 1 follows that G(A)'(5(A)- G(A) -> in probability as A -> 0. we can is the limit of M*(A) as A —* 0. A — next show that the at » 1, or = M*(l) 2 (i) With probability M*(0) If = [TX] with A Thus the of > 0. It definition of (9) define, as the limit of maximum — M*(X) is M*(A) when not attained 1. Theorem fiij We M*(0). Similarly, k k, or for = 1, M*(l) < M*(A), < /or every A < 1. (10) G{X) has a nonsingular covariance function, then with probability M*(0) To prove Theorem Lemma 1 = M*(l) < M*(A), we need the 2, For arbitrary < /or every following A < 1 1. (11) lemma. positive-definite matrices A and B with - y) A>B (p x p), and arbitrary vectors x and y (p x 1), we have x x'A~ x Proof of Lemma 1 : to prove equal to -z'Hz with the first x y - (x - y)\A - B)~ l (x < 0. (12) Define the matrix ( ^ ^ ~ It suffices - y'B~ H to for z' p rows = (/, {A (A-B)-i-A-* -(A-B)-* -(A-B)' {A-By' + B1 be positive-semidefinite because the (x',y'). Let D= (A \ J' left - B)' + B' > 1 1 (A — B)~ l D~*) and second p rows [l6) 1 0. (0, /). hand Let side of (12) C is be a matrix Using the identity - B)- - A- = (A- B)- D~ {A - B) -l 1 1 1 1 to obtain C'HC = Thus C'HC lemma. is positive-semidefinite, so diag(0, is D) > 0. H because C has full rank. This proves the Proof of Theorem The 2. M*(0) < M*(A) inequality equivalent to is 1 1 1 G(l)'g(l)- G(l)-G(A)'g(A)- G(A)-[G(l)-G(A)]'[Q(l)-Q(A)]- [G(l)-G(A)]<0. Clearly, part = x let £ = Theorem of = G(l), and y H be defined G(A). in (13). (G(l)', G(A)')'. Let > with Ai the (i) > A2 maximum • • • 2 follows rj so does the is implying — 2 first with probability The above NKN. Let (ii). Then M*(0) < M*(A) is letting A = A= equivalent to A 2p where X^s are the eigenvalues of H. Since < component is B= Q{1), B = Q(l) and -£'#£ < T be an orthogonal matrix such that T'HT eigenvalue of i/ A1 /; by 1 , Thus T£ £. Lemma Next, consider -?Ht = where from = H Q(X) and 0, diag(A l5 > Q(X), and where ..., H A 2p ) ^ 0, positive. It follows that is -(rO'diag(A 1 ,...,A 2p )r^ When of T£. < -r) 2 ^ G(X) has a nonsingular covariance matrix, a vector of normal variables with a nonsingular covariance matrix, with probability 1 because P(rj 2 = 0) = 0. That is, —£'H£ < 1. analysis applies to 1(1) regressors as well, which is not considered by Let yt where both x and e f are t 1(1), so that = x't fi_+ e t we have a spurious regression. Assume [TX] T- 2 J2x ^Q(X) t x't (14) t=i where Q(X) i.e. Q(u) is a stochastic matrix with Q(X) — Q(v) > with probability 1 > for v < (a.s.) u. and Q(X) is strictly increasing, Also assume [TX] T- 2 J2x t et ^G(X) (15) t=i where G(-) is a vector of random processes, and for each A, G(X) possesses a density function and a nonsingular covariance matrix. Theorem 3 The results of Theorem 1 and Theorem 2 apply to 1(1) regressors satis- fying (14) and (15), so that spurious regression leads to a spurious break. Proof of Theorem proof of (7) under the new only need to note that if definite with probability the proof is imply with (3) DT = T A—B A, B, and 1, 2 I and To prove . we are stochastic matrices that are positive- then inequality (12) holds with probability We conjecture that responsible for is Theorem 1. The rest of . 2. grated may its a spurious break will not occur because an 1(1) error occurrence. Shift in Cointegrating Relationships In the previous section we show that two integrated processes that are not cointe- What happens when give rise to a spurious break. the two processes are cointegrated but the cointegrating relationship undergoes a shift? be consistently estimated in the presence of The Can the shift point 1(1) regressors? issue of a structural change in cointegrating relationships of considerable is A interest. Cointegration describes a system's long-run equilibrium condition. may have A structural change model allows us to describe such a system. Consider, {ai+jix + e t C*2 where x t = Xt-i + et with x = Yl'jLo Q-jVt-j sequences with + 8 (8 is we assume that a shift in the fc mean > = t and Var(e t ) = finite 4 t= t + 72Z< + £t processes such that e t addition, and et > = = 0. [Tr ] for (16) + l,...,T fc We assume Y^jLo^j(,t-ji moments, and 0) 1,2, ...,k some r G X2jj| a j| (0, 1). < e t are e t are 1(0) linear where rj t = - a 2 ,7i - 72)'. £ t are When 71 = 72, Let X t = (l,x t )'. f3i = (0:1,71)' When and < /3 2 = is 71 i.i.d. co. In but «i of the long-run equilibrium. This intercept shift a shift in the cointegrating relationship. Let (ati and oo and J2jj\bj\ visualized as a change in the "gap" between two cointegrated series. is system multiple long-run equilibria with an occasional shift from one equilibrium to another. and 8 (10), t Regime there The remains an open question whether a spurious break arises when y t and x are process there = 2. a identical to the previous proof because inequality virtually identical to that of is cointegrated. 3 setting (15) a pure mathematical inequality and holds for arbitrary x t (8) is It and First, (14) 3. ^ a2 , often ^ 72, (02,72)' evident that the larger the magnitude of a It is The break. magnitude it is to identify the rate of convergence for the estimated break point depends not only on the on the magnitude of the regressors. of the shift in the coefficients, but also For stationary where the easier shift, X t the rate depends on the "effective magnitude of shift", 8'Q6/cr^, , Q = E{X X[). t In this case, can be shown that k it = + O p (l). k This the is best rate that can be achieved for stationary regressors; see, e.g., Bai (1994, 1995). not consistent for ko, although in terms of the sample fraction, Furthermore, k itself is k/T converges at a rate of k becomes consistent for converges to as t —» infinity. A: More T to r With . specifically, — 00 as > break point more precisely than with Lumsdaine and Stock (1994) For the 4 Let k denote the least if T —* 72, then a(t) = k ) = such that P(\k - k Because k > M) < \ P(k + 4. — > Based on {k : — \k that the event {k £ / ' = ko -> + Op (l), > 00, one can estimate the this fact, = ko + O p (l); we can establish 1. squares estimator ofko- Assuming that ko ) — o~\ = [Tto]. we have 1. for any > e 0, there exist an M < 00 Thus t. = P{\k-ko\>M) + P(\k-ko\<M,k^ko) ko) < Dm = 'E{XtX[)8 In particular, k shifted- cointegrating relationship of (16) with 71 7^ 72, Proof of Theorem show that even 6' 00. Consequently, for a proof. P(k = k where ^ 71 1(0) regressors. a more interesting result. Namely, P(k Theorem shall This follows because the "effective magnitude of shift" - 00. In particular, a(ko) see Bai, we 1(1) regressors, k Dm} \ e + P(keD M < M, k / implies that k }. (17) ) By definition, {mmk e D M Sj(k) Srik) < < J2t=i £ t- YlJ=i £ ?}- ^ follows Thus T P(k e Dm) v ' <P( mm keD M \ ST (k)<J2tf)j (18) ' We show that the right hand side of (18) converges to zero for every given M. is a finite set, it is sufficient to show that for 8 each k G Dm, P{ST(k) < Since Ylt=i £ t) Dm ~~* 0- We < prove this for k Y1<k = = = = Xi,k £i,fc ^*:,A:o Furthermore, Then for k The sum < M let k , kQ The . > case of k ko (yi,...,yk)', Ykj (Xi,...,Xk)', Xk,r (ei, £k,T .-., £*:)', (Xk+l,..-,Xk is = = = = X^ T )', = I - X^X^xjk)^ x similar. Define 1Je (y*+i,-,w)' pGfc+i,...,Xr)' (£fc+i,...,£x)' (Xk+l,...,Xko ,0,...,0)'. M and 2 = I - X'KT {X'KT X k T )- Xk T 1 , . , we have of squared residuals Sr(k) given by is = n'^F!,, + y iT M2 n, T ' Sr(fc) fc = £i, fc M l£l , * 2^'X T 'M2 e,, T + e^Maejb.T + fc fc + 8' X^' 2 Xt<T 8 T = E ~ et e i,*- x i,*(^J,*- x u)~ ^i,fc £ i,fc _ r £ fc,T^ fc,T(-^it,r^A:,T)~ X'kT e k ,T k +28' J2 X et-28X'kM Xk,k 1 (X'ktT Xk,T)~ Xk,Te k ,T t t=k+l ko + S '( E X Xt)t - X XWo( Xk,T Xk,T)~ $' t 'k < ko 1 X'k jCo Xk,k 8 t=k+l T = Yl £2 + t + aT ir + cT + dT + eT + /x t=\ We have used the fact that J2t=k+i x Xv XkT F° r each k G Dm, t = EtU+i 'e k ,T it is xt£t and XkT 'Xk ,T = X'kM Xk easy to see that ax, 6j, dx, and /x are all ,k = p (l). Thus «?r(A;)-E £ ? For each A; < the ko, for k = ko — T .2 St(*)-E*? t=l 1, **£« + is the P (T) sum right E *'( xtX?)8 + p (l) (19) t=k+l t=fc+l term on the first whereas the second term example, E = 2^ (=1 hand side of (19) and dominates involves only one in is bounded by magnitude the summand, and first Op (y/T), term. For = 26'Xko e ko +S'(X'kQ = 28 A s ko X ko + Op (l) )6 +28 2 (^e)e ko +8l + 2 e )+ 8 l(l2 e (^ i=\ i=l 28 l 8 2 z=l 0p (62 Y:e The any + ) > e °v( l ) ( 2 0) j:e y (8 2 i term can be rewritten last and converges for l + >) as T£f(-!=Ei_i e,) 2 , which dominates the to positive infinity with probability approaching 1. first term This implies that 0, P(ST (k)-^e 2 <0)<e t t=i for large T. Combining with for all large T When for ko, from and (18), only the intercept has a break magnitude of (20). we have, for every e > 0, P{k ^ k ) < 2e . even though k/T "effective (17) When 82 converges to still shift" stays = 0, (at\ / a2 ji , To at = 72), k is a rate of T. no longer consistent This is because the bounded. The lack of consistency can also be seen ST (k) - ELi e? = 2< ^ £ *o + Sl + P (1), which cannot be guaranteed to be positive. se. The underlying reason for k being consistent for ko Roughly speaking, = if k + O p (l) ko is not the 1(1) regressor per any holds, then for set of regressors that the second term (19) converges to infinity and dominates the consistent for ko- In particular, this we is first X t such term, k will be true for polynomials regressions. For simplicity, consider the following broken-trend model: {oi +7i< + £t oc 2 where et is + -y 2 t a linear process such that < +e e< t = Yl'jLo t = 1,2, t = k ..., Ar (21) + l,...,T a j r]t-j with and Y^]LoJ\ a j\ < r\ t a sequence of martingale differences and Theorem 5 For the broken trend model (21), assume 71 =^72- sup,- Er\\ 00, squares estimator of ko with ko P(k = = Then [Tto] &o) —* °°- as 1, 10 T— > 00 Let k be the least MIT LIBRARIES 3 9080 00988 2306 Proof of Theorem Let 5. X t The (19) can be copied here. = Then the proof (1, t)'. hand right side (19) X (with probability 1) for the newly defined k = — ko for t still of Theorem 4 up to equation converges to positive infinity each k £ Dm- For example, for 1, StW-^s? = 26'Xko e ko +6'(X'ko = 2<5ie feo X ko )S + O p (l) 6\ + 28l 82 k t=i + = O p {6 2 k + ) This implies that, for each k infinity. Although Theorem 5 = ko + O p (l) (but k DM , ko and Tj2 Thus, if the model is cast in the t for > t written as is ko)- If < 0) ^ we ^r, not unex- it is it is proved that rewrite the broken trend model then the new slope coefficients The magnitude . ^ and 0. of shift will of Bai (1994, 1995), be T(~/ 2 one — 71). essentially is the consistency of k for ko shift. In this sense, not surprising. The results of Theorem 4 and Theorem of shift to converge to zero. Let 8 2< j Theorem ^2,T ~~> 0, the model 2 still hold even (21), 6. assume T8 2< j In fact, the > c > we allow the magnitude > 00. assume y/T82 ,T —* 00 • For Under these assumptions, even consistent for ko- That of shift, is, P(k we — 1. have k = — still if ko) » and Bai, Lumsdaine and Stock (1994). 2 Consider the assumed magnitude of shift trend model, 62,t/VT is still — (16), Under the assumed magnitude see Bai (1994, 1995) if = 72 — 71. estimated break point k Proof of Theorem + O p (l); 5 6 For the cointegrated regression model the broken trend k k framework assuming an unbounded magnitude of is £? not explicitly presented in the literature, < for (22) P {1) and converges to positive first P(ST (k) - ELi (21) with the linear trend expressed in the format become Tji + 2 not consistent for is 2 ) (82 T) t* In Bai (1994, 1995), the linear trend pected. k is <E (8 2 k ) term above dominates the follows that the second + 2 (8 2 k = Op (62 Tt ) + It + 26 2 k e ko is sufficient is stronger than necessary. For example, for the broken- for k = k to be consistent for ko- 11 ko + Op (l). But the assumption is necessary for cointegrating regression (16). All that infinity with probability 1. As is we before, X ST (k) - £ right -4= hand model is Si=i ei is to prove (19) converges to positive consider k £ = O p (VT62 T ^= Vi ,=i k — 1. , + {VT6 2 T random converges to a normal By , variable (20), ~ ± e^ k i e t) side of (23) converges to positive infinity. similar = k i e? t=l Because needed Vi (23) i=i and VT82J —* The proof for the oo, the broken trend and thus omitted. References [1] Bai, J. (1994) Estimation of Structural Working Paper No. [2] 94-6, Department Change based on Wald Type Statistics. of Economics, M.I.T. Bai, J. (1995) Least Absolute Deviation Estimation of a Shift. Econometric The- ory 11, 403-436. [3] Bai, J., R.L. Lumsdaine, and J.H. Stock (1994) Testing for and Dating Breaks in Time Integrated and Cointegrated Series. Manuscript, Kennedy School of Gov- ernment, Harvard University. [4] Gregory, A.W. and B.E. Hansen models with regime [5] Hansen, B.E. (1992) Tests cesses. [6] shifts. (1996) Residual-based tests for cointegration in Journal of Econometrics 70 99-126. for parameter instability Journal of Business and Economic Statistics Nunes, L.C., CM. Kuan, P. in regressions with 1(1) pro- 10, 321-335. Newbold (1995) Spurious Break. Econometric The- ory, 11, 736-749. [7] Quintos, C.E. and P.B.C. Phillips (1993) Parameter constancy in cointegrating regressions. Empirical Economics 18, 675-703. 12