Estimating and Testing Linear Models with Multiple Structural Changes

Author(s): Jushan Bai and Pierre Perron

Source: Econometrica , Jan., 1998, Vol. 66, No. 1 (Jan., 1998), pp. 47-78
Published by: The Econometric Society

Stable URL: [Link]

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@[Link].

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
[Link]

The Econometric Society is collaborating with JSTOR to digitize, preserve and extend access to
Econometrica

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 Econoometrica, Vol. 66, No. 1 (Janualy, 1998), 47-78

ESTIMATING AND TESTING LINEAR MODELS WITH

MULTIPLE STRUCTURAL CHANGES

BY JUSHAN BAI AND PIERRE PERRON1

This paper considers issues related to multiple structural changes, occurring at un-
known dates, in the linear regression model estimated by least squares. The main aspects
are the properties of the estimators, including the estimates of the break dates, and the
construction of tests that allow inference to be made about the presence of structural
change and the number of breaks. We consider the general case of a partial structural
change model where not all parameters are subject to shifts. We study both fixed and
shrinking magnitudes of shifts and obtain the rates of convergence for the estimated
break fractions. We also propose a procedure that allows one to test the null hypothesis
of, say, I changes, versus the alternative hypothesis of I + 1 changes. This is particularly
useful in that it allows a specific to general modeling strategy to consistently determine
the appropriate number of changes present. An estimation strategy for which the location
of the breaks need not be simultaneously determined is discussed. Instead, our method
successively estimates each break point.

KEYWORDS: Asymptotic distribution, change point, rate of convergence, model selec-

tion.

1. INTRODUCTION

THIS PAPER CONSIDERS ISSUES related to multiple structural changes in the linear
regression model estimated by minimizing the sum of squared residuals.
Throughout, we treat the dates of the breaks as unknown variables to be
estimated. The main aspects considered are the properties of the estimators,
including the estimates of the break dates, and the construction of tests that
allow inference to be made about the presence of structural change and the
number of breaks.
Both the statistics and econometrics literature contains a vast amount of work
on issues related to structural change, most of it specifically designed for the
case of a single change.2 The econometric literature has witnessed recently an

I This paper has benefited from the comments of seminar participants at Harvard/MIT, N
western University, McGill University, the University of Copenhagen, the University of Illinois at
Urbana-Champaign, the Universit6 de Montr6al, the University of Sao Paulo, PUC at Rio de
Janeiro, Hitotsubashi University, the Universit6 de Lausanne, the CREST-INSEE, the European
University Institute, the University of York, the University of Pennsylvania, the 1995 Winter
Meeting of the Econometric Society, the 1995 Joint Meeting of the Institute of Mathematical
Statistics and the Canadian Statistical Association, and the Symposium on Nonlinear Dynamics and
Econometrics, Boston. Financial support is acknowledged from the National Science Foundation
under Grant SBR-9414083, the Social Sciences and Humanities Research Council of Canada, the
Natural Sciences and Engineering Research Council of Canada, and the Fonds pour la Formation
de Chercheurs et l'Aide a la Recherche du Qu6bec. Finally, we thank three anonymous referees for
their valuable comments.
2See the surveys of Zacks (1983), Krishnaiah and Miao (1988), and Bhattacharya (1994).

47

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
48 J. BAI AND P. PERRON

upsurge of interest in extending procedures to various models with an unknown

change point. With respect to the problem of testing for structural change,
recent contributions include the comprehensive treatment of Andrews (1993)
and Andrews and Ploberger (1994). This issue has also received a lot of
attention in the debate on unit root versus structural change in the trend
function of a univariate time series (see Perron (1989)). Issues about the
distributional properties of the parameter estimates, in particular those of the
break dates, have also been considered (see Bai (1994a, b)).
In comparison, the literature addressing the issue of multiple structural
changes is relatively sparse. Recent developments include Andrews, Lee, and
Ploberger (1996) who consider optimal tests in the linear model with known
variance. Garcia and Perron (1996) study the sup Wald test for two changes in a
dynamic time series. In an independent study, Liu, Wu, and Zidek (1997)
consider, as we do, multiple shifts in a linear model estimated by least squares.
They study the rate of convergence of the estimated break dates, as well as the
consistency of a modified Schwarz model selection criterion to determine the
number of breaks. Their analysis considers only the so-called pure-structural
change case where all the parameters are subject to shifts. Our assumptions are
much less restrictive than those of Liu, Wu, and Zidek (1997), and our main idea
of argument differs from theirs. Our model allows for general forms of serial
correlation and heteroskedasticity in the errors, lagged dependent variables,
trending regressors, as well as different distributions for the errors and the
regressors across segments. Furthermore, we consider the more general case of
a partial structural change model where not all parameters are subject to shifts.
A partial change model is useful in allowing potential savings in the number of
degrees of freedom, an issue particularly relevant for multiple changes. We
obtain the rates of convergence for the estimated break points not only for fixed
but also for shrinking magnitudes of shifts. The latter is the basis for the
derivation of feasible asymptotic distributions and confidence intervals for the
break dates.
Our study considers, in addition, the important problem of testing for multiple
structural changes for the case with no trending regressors. To that effect, we
present sup Wald type tests for the null hypothesis of no change versus an
alternative hypothesis containing an arbitrary number of changes. We also
propose a test where the alternative specifies an unknown number of changes up
to some maximum and a test of the null hypothesis of, say, 1 changes versus
1 + 1 changes. The latter is useful for a specific to general modeling strategy to
determine the number of changes present. Finally our paper contains a discus-
sion of an estimation strategy for which the locations of the breaks need not be
simultaneously determined. Rather our method successively estimates each
break point.
The rest of this paper is structured as follows. Section 2 discusses the model
and the assumptions imposed on the variables and the errors. Section 3 contains
results about the consistency, the rate of convergence, and the asymptotic
distribution of the estimates of the break dates (as well as other parameters of

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 MULTIPLE STRUCTURAL CHANGES 49

the model). Section 4 proposes test statistics, derives their asymptotic distribu-
tions, and presents critical values. Section 5 discusses sequential methods to
estimate the model without treating all break points simultaneously. All proofs
are collected in an appendix.

2. THE MODEL AND ASSUMPTIONS

Consider the following multiple linear regression with mn breaks (ni + 1

regimes):

(1) Yt=X3?+z j +?u, (t=7T>1 ?,...,T),

for j = 1,..., m + 1 and where we use the convention that To = 0 and T)1 + = T.
In this model, yt is the observed independent variable, xt (p x 1) and zt (q X 1)
are vectors of covariates, and ,B and 8j (j = 1,., i. n + 1) are the corresponding
vectors of coefficients; uti is the disturbance. The indices (T1,.,-T,), or the
break points, are explicitly treated as unknown. The purpose is to estimate the
unknown regression coefficients together with the break points when T observa-
tions on (Yt, xt, Zt) are available. Note that this is a partial structural change
model in the sense that /3 is not subject to shifts and is effectively estimated
using the entire sample. When p = 0, we obtain a pure structural change model
where all the coefficients are subject to change.
The multiple linear regression system (1) may be expressed in matrix form as
Y=X,3?+Z? +U, where Y=(y,.**. XYT), X=(X, ..., XTY, U=(1.... UT)',
6 = (6, 8.,..., 8 ... + I1)' and Z is the matrix which diagonally partitions Z at the
m-partition (T, ... , T,), i.e., Z = diag(Z1, . . . , + 1) with Zi (ZT I + 1-- ZT)
Throughout, we denote the true value of a parameter with a 0 superscript. In
particular, 5 0 = ( 0' 8 . + Jad(T"n..
)' and T2) ae
. . ., are, respectively, the true val-
ues of the parameters 8 and of the break points. The matrix Z0 is the one
which diagonally partitions Z at (T'(, .. ., T,?,). Hence the data-generating proce
is assumed to be

(2) Y = X,3 0 + Z08 0 + U.

The goal is first to estimate the unknown coefficients ( 30, 82,..

TO.,), assuming 8i0 = 5i? I (1 < k < ni). We do not impose the restr
the regression function is continuous at the turning points. For the latter,
readers are referred to Feder (1975) and Gallant and Fuller (1973) for the
special case of a polynomial trend regression. This paper focuses on discrete
shifts. In general, the number of breaks m can be treated as an unknown
variable with true value imn. However, for now, we treat it as known and discuss
methods of estimating it in later sections. We also postpone the problem of
testing for the presence of structural change to Section 4.
The method of estimation considered is that based on the least-squares
principle. For each rn-partition (T1,..., 7T), denoted {Tj}, the associat
squares estimates of /3 and 5 are obtained by minimizing the sum of squared

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 50 J. BAI AND P. PERRON

residuals 7L' 1ZTTi +i[Yt -X /3 -Z8i]2. Let PQTJ) and 6({IT}) denote the
resulting estimates. Substituting them in the objective function and denoting the
resulting sum of squared residuals as ST(T1, ..., T79.), the estimated break points
(T..., T,,) are such that

(3) (XT,...iT,;) = argminT,. T ST(T1,...,T171),

where the minimization is taken over all partitions (T1,.. ., T,,,) such that
Ti- Ti ? 2q. Thus the break-point estimators are global minimizers of the
objective function. Finally, the regression parameter estimates are the associ-
ated least-squares estimates at the estimated m-partition Tj}, i.e. , Tj
and 8 = 5({T.)). Note that the break points need not be obtained via an
exhaustive grid search. We discuss in Bai and Perron (1996) an efficient
algorithm based on the principle of dynamic programming which allows global
minimizers to be obtained using a number of sums of squared residuals that is of
order O(T2) for any m ? 2.
The statistical properties of the resulting estimators are studied in the next
section under the following set of assumptions. As a matter of notation, we let

" > " denote convergence in probability, " " convergence in distribution,
and " " weak converge in the space D[O, 1] under the Skorohod metric (e.g.,
Pollard (1984)).

ASSUMPTION Al: Let wt = (x, Z)', W= (w1,... I, WY, and W? be the diagon
partition of Wat (To,. .., T7) such that W? = diag(W), . . .,j W )+ O. We assume
each i = 1, .. ., m + 1, with T(? = 1 and T, l + 1 = T, that WiJ 'W0/(Ti -Tio 1)
verges in probability to some nonrandom positive definite matrix not necessarily th
same for all i.

ASSUMPTION A2: There exists an 10 > 0 such that for all 1 > lo, the minimum
eigenvalues of Ail = (1/1)ELT%o+ lwtw' and of A* = (I/l) ETo_ w wI are bounded
away from zero (i = 1,..., rn + 1). I

ASSUMPTION A3: The matrix Bkl = EkZt Z is invertible

sion of zt.

The sequence of errors {u,) satisfies one of the following two sets of condi-
tions:

ASSUMPTION A4(i): With {g: i = 1, 2,...) a sequence of increasing o-fields,

assume that {ui,7i) forms a L'-mixingale sequence with r = 4 + 8 for some 8 > 0
(McLeish (1975) and Andrews (1988)). That is, there exist nonnegative constants
{ci: i 2 1) and {I/': j 2 ?0 such that t/j , 0 as j-- oc and for all i ? 1 and j > 0, we
have: (a) EIE(ui ? cif'i', (b) EIui - E(uil 1+)J' < c1 Wj4+I, (c) maxi ci < K
< 0c, (d) EL> = jl ? K+1j < oc for some K > 0. We also assume (e) that the disturbances
Ut are independent of the regressors wS for all t and s.

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 MULTIPLE STRUCTURAL CHANGES 51

Or:

ASSUMPTION A4(ii): Let Y7' = o-field { ***,wt_ ,w

(a) that {lttl is a martingale difference sequence re
< oo for sonie c > 0; (b) T- l [tv]Z Zt -p Q(v) unifor
is positive definite for v > 0 and strictly inicreasing
not inclependent of the regressors {z,J for all t an
defined by (3) is taken over all possible partitio
1. ..m + 1) for some e > 0.

ASSUMPTION A5: Ti = [TA?], where 0 < AO < - < AO, < 1.

Assumption Al is standard for multiple linear regressions. Assumption A2

requires that there be enough observations near the true break points so that
they can be identified. A2 can be weakened as follows. For some c > 0 and
lI1 > lo 0 l(8j?- 860)'Bll > cll(60? - 80)11 for B = (1/l)ETZ-(]+1 ztz
(1/l)LTP_1 ZtZ. In addition, for every e > 0 and 1 = [ET], the minimu
values of Ai,, and A*, are bounded away from zero in probability fo
A3 is imposed because the break points are estimated by a global least-squares
search. If the number of observations in each segment is at least some fixed h
(h ? q, not depending on T), the invertibility requirement in A3 can be
weakened to hold for all combinations (1, k) for which 1 - k > h.
The assumptions stated in A4 pertain to two specific cases related to the
presence or absence of a lagged dependent variable in wt. The conditions
described in part (i) pertain to the case where no lagged dependent variables are
allowed in wt implied by part (e). In this case, the conditions on the residuals a
quite general and allow substantial correlation and heterogeneity. Part (ii) of
Assumption A4 considers the case where lagged dependent variables are al-
lowed as regressors. In this case, no serial correlation is permitted in the errors
{ll)t. This extra generality is obtained at the expense of some restrictions
admissible partitions if a lagged dependent variable is present in the zt. In such
cases, each segment considered to compute global minimizers must contain a
positive fraction of the total sample. This is not constraining from a practical
point of view since e can be arbitrarily small. Note, however, that this restriction
is not necessary if a lagged dependent variable is present only in the xt's. In
both A4(i) and A4(ii), the assumptions are general enough to allow different
distributions for both the regressors and the errors in each segment.
The choice between assumptions A4(i) and A4(ii) can be especially interesting
in the case of dynamic models when the coefficients associated with the lagged
dependent variables are not subject to change. In this case, the investigator can
take the dynamic effects into account either in a direct parametric fashion (e.g.
introducing lagged dependent variables so as to have uncorrelated residuals) or

3 Note that, for the proof of the consistency, A3 could be dispensed using generalized inverses.

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
52 J. BAI AND P. PERRON

using an indirect nonparametric approach (e.g. leaving the dynamics in the

disturbances and applying a nonparametric correction for proper asymptotic
inference).
Assumption A5 is a standard requirement to permit the development of an
asymptotic theory and allows the break points to be asymptotically distinct. It
considers the asymptotic experiments under the assumption that each segment
increases proportionately as the sample size increases. We refer to the quanti-
ties Ao = (AO,..., AO1) as the break fractions and we let Ao = 0 and A
Finally, we assume that polynomial trending regressors are written in the form
of (t/T)' (1 ? 0) or, more generally, written as a continuous function of the time
trend, g(t/T) (see Bai (1994b, 1995) for the case of a single break). The
consistency and rate of convergence of the estimated break points apply to
trending regressors. However, the assumptions in Section 4 for the test statistics
rule out trending regressors.

3. CONSISTENCY AND LIMITING DISTRIBUTIONS

In this section, we analyze the consistency of the estimated break fractions

and their rate of convergence. The latter allows us to derive results about the
asymptotic distribution of the estimates ( /3, ,*..., I + T1,...,7 ). We let
A = (A1, ..., A,11) = (T1/T,..., T,71/T) with corresponding true values AO =
(AO,..., AO). We shall first show that A is consistent for AO and later that the
rate of convergence is T.

3.1. Consistency

The main result of this section is summarized in the following proposition

which states the consistency of A for AO.

PROPOSITION 1: Under Al-A5: Ak p AO, k = 1,...,m.

We outline the main steps of the proof using a few lemmas that are proved in
the appendix. Denote by ut the estimated residuals and by dt the difference
between the fitted and true values. That is, ui =Y -xt/3Zt8k, for t E [Tkl ?
l,Tk] and dt =x'(/-/3) ?z(8k-1j), for t[tk ?1, Tk] n [T + 1, Tj1]
(k,j = 1,. .n.,m + 1). Note that,?in general, dt is defined over (m + 1)2 differ
segments for each of the possible m-partitions {Ti} and {T,?}. Using propertie
projections,
lT lT

(4) -: u2 <-: u2
T t=l T t=1
and using ut = ut -dt,
lT lT lT lT

(5) T lt2=T E ut + d - 2- utdt.

Tt =1 Tt =1 T t1 Tt

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 MULTIPLE STRUCTURAL CHANGES 53

The proof of Proposition 1 simply uses relations (4) and (5) and the associated
limit of T- T T= Iut t. We start with the latter.

LEMMA 1: UnderAl-A 5, we have T- L[ 1l ttdt = o0(1).

Lemma 1 together with (4) and (5) implies that T- lT= Id2 __ 0. The p
follows by showing that this implies A __ AO. More specifically, T- L[= 1d'--j 0
cannot hold if A1 -'p A' for some j. This is stated in the following lemma.

LEMMA 2: Assume Al-A5 hold and that A1 -7 AO for some j; then

T

lim sup P T- 1 Z d2 > CHJI8O1_-,0 0 1112 > ,E,

T ?xt = 1J

for some C > 0 and co >

We are now in the position to prove Proposition 1. Using (5) and Lemmas 1
and 2, and under the supposition that some break date is not consistently
estimated, we have the inequality

T T

T E L 2 2T'Lu7-CII8?-8j 1112 + o0(1)

'1 1

holding with probab

the inequality (4), which holds with probability 1 for all T. Hence, all break
dates are consistently estimated.

3.2. Rates of Convergence

We now consider the rate of convergence of the estimates. We start by

showing that Ak converges to its true value at rate T. More precisely, we have
the following proposition.

PROPOSITION 2: Under Al-A5, for evey r1 > 0, there exists a C < CX, stuch that
forall alarge T, P(IT(Ak - A?)j > C) < -j (k= 1,... m).

It is important to remark that the rate T convergence pertains to the

estimated break fraction Ai and not to Ti, the estimated break date.
latter, our result states that with high probability its deviation from the true
break is bounded by some constant C that is independent of T, i.e. with high
probability, we have ITi - Til < C.
The rate T convergence of the estimated break fractions allows us to obtain
standard root-T asymptotic normality of the estimated coefficients /3 and &. The
relevant results are stated in the following proposition whose proof is similar to
Corollary 1 of Bai (1994b) and is therefore omitted.

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
54 J. BAI AND P. PERRON

PROPOSITION 3: Let 0 =((3, ) and 00 = (,30, S0). Under Al-A,5 T(0-

00) -N(0, V- V- ), witlh V= plim T- lW?'W?, = plim T-1W?0',QW0, and
?2 = E(UU').

Note that when the errors are serially uncorrelated and homoskedastic we
have d = o- 2V and the asymptotic covariance matrix reduces to o2V- 1, which
can be consistently estimated using a consistent estimate of o- 2 When serial
correlation and/or heteroskedasticity is present, a consistent estimate of d) can
be constructed along the lines of Andrews (1991), assuming identical distribu-
tions across segments or allowing the distributions of both the regressors and
the errors to differ.

3.3. Limiting Distributions of Break Dates

Note first that, as in the single break case, the usual limiting distribution of
the break dates obtained specifying fixed magnitude of changes depends on the
exact distribution of the pair {zt, ut}. On the other hand, a strategy that permits
obtaining pivotal statistics is to consider an asymptotic framework where the
magnitudes of the shifts converge to zero as the sample size increases. Even
though the setup is particularly well suited to provide an adequate approxima-
tion to the exact distribution when the shifts are small, it remains adequate even
for moderate shifts. The required conditions are stated in the next assumptions
defined for i = 1, ... , m.

ASSUMPTION A6: Let ATi = - i Assume ATi = vTAi for some Ai

independent of T, where VT > 0 is a scalar satisfying VT -> 0 and TP'/2)
for some O E (0, 1/2). In addition, we assume E zt2 < M and E I u 2 <M
some M < oc and all t.

Note that for a smaller magnitude of shift (small VT), which corresponds to a
smaller X, A6 requires the existence of a higher moment of ut. When VT is
fixed constant, we can choose O arbitrarily close to 1/2. In this case, the
requirement of ElBt 2/0 <M reduces to the existence of 4 + S mo
stated in A4.

PROPOSITION 4: UnderAssumptions A1-A6, we have for k = 1,... m: (i) Ak

A ; and (ii) for every -1 > 0 there exists a C < oo such that for all large T, PI CTV?2
- A?)1 > C) < .

Proposition 4 asserts that the estimated break fractions remain consistent

even in the case where the shifts decrease as the sample size increases. The rate
of convergence is, of course, no longer T but rather TVT. This rate is sufficient
to establish root-T consistency for the estimated regression parameters. This
result will not be presented to save space, and interested readers are referred to

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 MULTIPLE STRUCTURAL CHANGES 55

Bai (1994a, 1994b) for the case of a single break. Proposition 4 allows us to study
the limiting distribution of the estimated break dates. It asserts that we can
restrict the analysis to a "neighborhood" of length C/vT around the true break
dates Tio which makes possible the application of a central limit theorem
this "neighborhood" increases when LT decreases. With the mixing assumptions
on the errors, each segment is asymptotically distinct and the analysis of the
limiting distribution of the break dates is similar to that in the single break case
as analyzed in Bai (1994a, 1994b). We provide, in the rest of this section, a
description of the results when the data are not trending and under the
assumption that the following conditions are satisfied.

ASSUMPTION A7: Let AT1 = T0 -Tio 1; we assume, for i = 1 .. n + 1, tha

AT0 oo
Tik1 + [S ATj] Tit I+ [s ATio]
(a) (ATo)1 ' z zt - sQi, (ATi l , >p sj
t=T?i +1 t=T1 1+1

and
Ti0 ?+[s ATi''] Tj 1?+[SATi ]

(ATO)1 L E(Z,.ZU,.Ut) -1 sf)1

r-=Ti-?+1 t=TiL) 1+1

uniformly in s E [0, 1];

Ti?1 + [s ATi ]

(b) ( ATi?) 1" E z,utit >Bi (S)

t = T7< ?+ 1

where Bi(s) is a multivariate Gaussian process on [0, 1] with mean zero and
covar-iance EBi(s)Bi(u) = min{s, u}lfi.

Now, define for i = 1,..., n: i= A1Qi+I A1A/Qi Ai, i21 = Al2 Ai/IAQi Ai,
,= A2Qi+ Ai/AQi+ 1Ai, and let WP)(s) and W4'(As) be independent Wiener
processes defined on [0, oo), starting at 0 when s = 0. These processes are also
independent across i. Also, define Z(')(s)= - i1WI("( -s)-Is1/2, for s < 0, an
Z(i)(s) = 4ri 4i2WIA(s) - 4JsJ/2, for s > 0. We can state the following result.

PROPOSITION 5: Under- A1-A7 (AQ'A )v2(T - T0) = argmaxsZ(i)(s) (i =

The limiting distribution is the same as that occurring in a single break model.
The density function of argmaxsZ(I)(s) is derived in Bai (1994b) and is nonsym
metric. When the limits Qi, !)i, and o-i2 are the same for adjacent i's, (i = 1, an
X1= Xi 2-X, in which case the limiting distribution reduces to:

(6) Q (T - T0o) =v arg max{W('i(s) - sl/2}

(AQfA i)Ti

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
56 J. BAI AND P. PERRON

which is symmetric about the origin and has distribution function (see Yao
(1987)):

H(x) = 1 + (2-7T Vx2xe-x/8 - (x + 5)P(- x/2)

+ 3exP(- 3Vx:/2),
for x> 0 and H(x) = 1 - H(-x), with >(x) the distribution function of a
standard normal variable. For instance the 95% and 97.5% quantiles are 7.7 and
11.0.
The results discussed above allows easy construction of confidence intervals
for the break dates. All that is needed is to construct consistent estimates of the
various parameters; T` I= 1z, z for Q, T- [ =i2u for C 2 and 5i + ^ - 3
UTiA. When serial correlation is present, D2 can be estimated using a kernel-
based method as discussed in Andrews (1991). Note that when the segments are
not homogeneous, obtaining consistent estimates is still possible using data over
the relevant subsamples only.
The limiting distribution in the case of trending regressors is discussed in Bai
(1994b, 1995) for a single structural change model. His results remain valid for
multiple breaks. We omit the details and refer the reader to those papers.

4. TEST STATISTICS FOR MULTIPLE BREAKS

4.1. A Test of No Br-eak Versus Some Fixed Number of Breaks

We consider the sup F type test of no structural break (m = 0) versus the

alternative hypothesis that there are mn = k breaks. Let (T1, ..., Tk) be a parti-
tion such that Ti = [TAi] (i = 1,-.. , k). Define

(7) ( _T-(k_ 1)q-p )___R' _R(Z_ MxZ) R'I) R_ _

kq ~~~SSRk
where R is the conventional matrix such that (R3 )'= (5 - - 3J - - + )
and Mx =I - X(X'X) 'X'. Here SSRk is the sum of squared residuals under
the alternative hypothesis, which depends on (T,... ., Tk). To carry out the
asymptotic analysis, we need to impose some restrictions on the possible values
of the break dates. In particular, we need to restrict each break date to be
asymptotically distinct and bounded from the boundaries of the sample. To this
effect, we define the following set for some arbitrary small positive number e:
Ae = {(A ..., A); Ai+ 1 - Ai 6, Al ? 6, Ak < 1 - el. The sup F type test statis-
tic is then defined as supFT(k; q) = SUP(A1. Ak)E IIFT(Al ...* Ak; q). It is
generalization of the sup F test considered by Andrews (1993) and others for
the case k = 1. The limiting distribution of the test depends on the nature of the
regressors and the presence or absence of serial correlation and heterogeneity
in the residuals. We consider the case where the following assumptions are
imposed.

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 MULTIPLE STRUCTURAL CHANGES 57

ASSUMPTION A8: T 1 E[Ts]W WI 3P sQ, uniforml

tiLve definite matrix.

Note that A8 precludes the presence of trending regressors. Extensions to the

general case where plimT CT-' 1[TsWtW,W = Q(s), which allows trending r
sors, are beyond the scope of the present paper.

ASSUMPTION A9: The ei7ors {ut) form an atray of martingale differences relati
to {gt7} = u-field {... wt_ w, ... ,t2 1_. Also, E[U2] = cr2 for all t a
T-1/2E[TW iW,t, = oQ1/2W*(r), with W*(r) a (p + q) vector of independent
Wiener processes.

The case where {ut} satisfies the general conditions stated in Assumption A4
is discussed in Section 4.4 below. We show how the results remain valid provided
appropriate modifications are made to account for the effect of serial correla-
tion on the asymptotic distributions. The following proposition is proved in the
appendix.

PROPOSITION 6: Let WJ,( ) be a q-vector of independent Wiener pr-ocesses on

[0, 1]. Under A8-A9 and m = 0, sup FT(k; q) =* sup Fkq= SUP(AdefA AF(Al
Ak; q), with

F(Al ..I Ak; q)

def 1 k [A1W(Ai+1) - Ai+1Wj(Ai)] [AiWC1(Ai+1) - Ai+JIA)]
kq i= 1 AiAi+1(Ai+1 -A )

Note that the asymptotic distribution of the test statistic depends on the value
of e in AE. As e converges to zero, the critical values of the limiting random
variable of supFT(k; q) diverge to infinity. Because the computed test statistic
for a given sample is finite, a small positive value of e can improve the power
significantly; see Andrews (1993) for further details. In what follows, we have
adopted e = 0.05. No critical values for k ? 2 are available except those of
Garcia and Perron (1996) for k = 2 and q = 1.
Asymptotic critical values are obtained via simulations. The Wiener process
WJ,(A) is approximated by the partial sums n-1/2E[nAle with ei i.i.d.
and n = 1,000. The number of replications is 10,000. For each replication, the
supremum of F(A1,..., Ak; q) with respect to (A1,..., Ak) over the set A
obtained via a dynamic programming algorithm. We present, in Table I, critical
values covering cases with up to 9 breaks (k = 1,... , 9) and up to 10 regressors
(q = 1,..., 10) whose coefficients are the object of the test. The values reported
are scaled up by q for comparison purposes. The column corresponding to k = 1
can also be found in Andrews (1993). Because supFT(l; q) < 2sup FT(2; q) <
ksup FT(k; q), the consistency of the supFT(k; q) (k ? 2) follows from Andrews
(1993) who proved the consistency of supFT(l; q) for various alternatives includ-
ing multiple breaks.

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 58 J. BAI AND P. PERRON

TABLE I

ASYMPTOTIC CRITICAL VALUES OF THE MULTIPLE-BREAK TEST.

THE ENTRIES ARE QUANTILES X SUCH THAT P(supFk ? x/q) a.

Number of Breaks, k

q a 1 2 3 4 5 6 7 8 9 UDmax WDmax

1 .90 8.02 7.87 7.07 6.61 6.14 5.74 5.40 5.09 4.81 8.78 9.14
.95 9.63 8.78 7.85 7.21 6.69 6.23 5.86 5.51 5.20 10.17 10.91
.975 11.17 9.81 8.52 7.79 7.22 6.70 6.27 5.92 5.56 11.52 12.53
.99 13.58 10.95 9.37 8.50 7.85 7.21 6.75 6.33 5.98 13.74 15.02
2 .90 11.02 10.48 9.61 8.99 8.50 8.06 7.66 7.32 7.01 11.69 12.33
.95 12.89 11.60 10.46 9.71 9.12 8.65 8.19 7.79 7.46 13.27 14.19
.975 14.53 12.64 11.20 10.29 9.69 9.10 8.64 8.18 7.80 14.69 16.04
.99 16.64 13.78 12.06 11.00 10.28 9.65 9.11 8.66 8.22 16.79 18.11
3 .90 13.43 12.73 11.76 11.04 10.49 10.02 9.59 9.21 8.86 14.05 14.76
.95 15.37 13.84 12.64 11.83 11.15 10.61 10.14 9.71 9.32 15.80 16.82
.975 17.17 14.91 13.44 12.49 11.75 11.13 10.62 10.14 9.72 17.36 18.79
.99 19.25 16.27 14.48 13.40 12.56 11.80 11.22 10.67 10.19 19.38 20.81
4 .90 15.53 14.65 13.63 12.91 12.33 11.79 11.34 10.93 10.55 16.17 16.95
.95 17.60 15.84 14.63 13.71 12.99 12.42 11.91 11.49 11.04 17.88 19.07
.975 19.35 16.85 15.44 14.43 13.64 13.01 12.46 11.94 11.49 19.51 20.89
.99 21.20 18.21 16.43 15.21 14.45 13.70 13.04 12.48 12.02 21.25 22.81
5 .90 17.42 16.45 15.44 14.69 14.05 13.51 13.02 12.59 12.18 17.94 18.85
.95 19.50 17.60 16.40 15.52 14.79 14.19 13.63 13.16 12.70 19.74 20.95
.975 21.47 18.75 17.26 16.13 15.40 14.75 14.19 13.66 13.17 21.57 23.04
.99 23.99 20.18 18.19 17.09 16.14 15.34 14.81 14.26 13.72 24.00 25.46
6 .90 19.38 18.15 17.17 16.39 15.74 15.18 14.63 14.18 13.74 19.92 20.89
.95 21.59 19.61 18.23 17.27 16.50 15.86 15.29 14.77 14.30 21.90 23.27
.975 23.73 20.80 19.15 18.07 17.21 16.49 15.84 15.29 14.78 23.83 25.22
.99 25.95 22.18 20.29 18.93 17.97 17.20 16.54 15.94 15.35 26.07 27.63
7 .90 21.23 19.93 18.75 17.98 17.28 16.69 16.16 15.69 15.24 21.79 22.81
.95 23.50 21.30 19.83 18.91 18.10 17.43 16.83 16.28 15.79 23.77 25.02
.975 25.23 22.54 20.85 19.68 18.79 18.03 17.38 16.79 16.31 25.46 26.92
.99 28.01 24.07 21.89 20.68 19.68 18.81 18.10 17.49 16.96 28.02 29.57
8 .90 22.92 21.56 20.43 19.58 18.84 18.21 17.69 17.19 16.70 23.53 24.55
.95 25.22 23.03 21.48 20.46 19.66 18.97 18.37 17.80 17.30 25.51 26.83
.975 27.21 24.20 22.41 21.29 20.39 19.63 18.98 18.34 17.78 27.32 28.98
.99 29.60 25.66 23.44 22.22 21.22 20.40 19.66 19.03 18.46 29.60 31.32
9 .90 24.75 23.15 21.98 21.12 20.37 19.72 19.13 18.58 18.09 25.19 26.40
.95 27.08 24.55 23.16 22.08 21.22 20.49 19.90 19.29 18.79 27.28 28.78
.975 29.13 25.92 24.14 22.97 21.98 21.28 20.59 19.98 19.39 29.20 30.82
.99 31.66 27.42 25.13 24.01 23.06 22.18 21.35 20.63 19.94 31.72 33.32
10 .90 26.13 24.70 23.48 22.57 21.83 21.16 20.57 20.03 19.55 26.66 27.79
.95 28.49 26.17 24.59 23.59 22.71 21.93 21.34 20.74 20.17 28.75 30.16
.975 30.67 27.52 25.69 24.47 23.45 22.71 21.95 21.34 20.79 30.84 32.46
.99 33.62 29.14 26.90 25.58 24.44 23.49 22.75 22.09 21.47 33.86 35.47

Notes: 1. The test UDmax is defined as max, < k < 5SUP(Al Ak,)C F(Al Ak; q) mtu
given in (9) multiplied by q, and M is chosen to be 5.

4.2. A Double Maximum Test

The test discussed above requires the specification of the number of breaks,
m, under the alternative hypothesis. It is of interest to consider tests of no
structural break against an unknown number of breaks given some upper bound

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 MULTIPLE STRUCTURAL CHANGES 59

M. Consider the following new class of tests, called the doutble maximutm tests:

(8) DmaxFT(M,q,a,... ,a,,)= max a,M sup FT(A1, -, A,1; q),

1?<i1M (A1 ..., A,,,) Ei A

defined for some fixed weights {a1, .. ., am}. Note that the asymptotic distribution
of this class of tests is easily obtained from Proposition 6. Indeed, we have

DmaxFT(M, q,al,....am)=> max a sup F(A* A,; q)-

1<In M (A1 A,

The weights may reflect the imposition of some priors on the likelihood of
various numbers of breaks. Apart from such considerations, precise theoretical
guidelines about their choice remain an open question. An obvious candidate is
to set all weights equal to unity and we label this version of the test as

UDmax FT(M, q) = max1I < I < Msup(A A,,.) e AC FT(Aj, . . , A,,,; q). For
F(A1,.. ., A,,,; q) is the sum of m dependent chi-square random var
degrees of freedom, each one divided by m. This scaling by m can be viewed, in
some sense, as a prior imposed to account for the fact that as ni increases a
fixed sample of data becomes less informative about the hypotheses being
confronted. Since for any fixed q the critical values of the individual tests
SUP(A1... A,,,) eAlFT(Al,.., A,..; q) decrease as m increases, this implies t
marginal p-values decrease with mn and may lead to a test with low power if the
number of breaks is large. One way to alleviate this problem is to consider a set
of weights such that the marginal p-values are equal across values of m. This
implies weights that depend on q and the significance level of the test, say a. To
be more precise, let c(q, a, m) be the asymptotic critical value of the test

SUP(A1. A,,,)e ACFT(Al,. , A,,,; q) for a significance level a. The weights ar

defined as a1 = 1 and for m > 1 as a,M = c(q, a, 1)/c(q, a, m). This version
denoted

c(q, a,l1)
(9) WDmax FT (M, q)= max
1?7l <M c(q, a, m)

x sup FT(Al1,..A,l,;q).
(A1,.,, A I) E e

The last two columns of Table I report the asymptotic critical values of both
tests for M = 5 and e = 0.05. This should be sufficient for most empirical
applications. In any event, the critical values vary little for choices of the upper
bound M larger than 5. The consistency of the tests follows directly from the
consistency of supFT(k; q).

4.3. Test of 1 versus 1 + 1 Breaks

This section considers a test of the null hypothesis of 1 breaks against the
alternative that an additional break exists. Ideally, one would base the test on
the difference between the sum of squared residuals obtained with 1 breaks and
that obtained with 1 + 1 breaks. The limiting distribution of this test statistic is,

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
60 J. BAI AND P. PERRON

however, difficult to obtain. Here, we pursue a different strategy. For the model
with 1 breaks, the estimated break points, denoted by T1,... T, 1', are obtained by
a global minimization of the sum of squared residuals. Our strategy proceeds by
testing each (I + 1) segment (obtained using the estimated partition T,... ,)
for the presence of an additional break. We assume the magnitude of shifts is
fixed (nonshrinking) in this section.
The test amounts to the application of (I + 1) tests of the null hypothesis of
no structural change versus the alternative hypothesis of a single change. It is
applied to each segment containing the observations Ti>l + ito Tj (i = 1
1 + 1) using again the convention that To = 0 and T+1 = T. We conclude for a
rejection in favor of a model with (I + 1) breaks if the overall minimal value of
the sum of squared residuals (over all segments where an additional break is
included) is sufficiently smaller than the sum of squared residuals from the 1
break model. The break date thus selected is the one associated with this overall
minimum. More precisely, the test is defined by

(10) FT(l + 11l) = (ST(Tl ***,T)

- min inf ST(T1,..,T TTi-1J..Ti v

1 < i < I+11 TI 7s Ai,

where

(I11) Ai,= tTi_ I (Ti-Ti _1)< Ti-(Ti-Ti-l

and 6J2 is a consistent estimate of J under the null hypothesis. Note that for
A A A A A A

i=
A A
1,
1 + 1 as ST(T,.. . T, 17, T). We have the following result, proved in the Appendix:

PROPOSITION 7: Under-Assumptions A8-A9 and m = 1: "imT -. ,P(FTU( + 1|1) <

x) = Gq7r(x) + 1 with Gq, (x) the distribution function of sup,, < < 1-WHJ(t )-
,A4'(1)Hl2/( (1 - A)).

The critical values of this test for different values of I can be obtained from
the distribution function Gq, ,(x). A partial tabulation of some percentage po
can be found in DeLong (1981) and Andrews (1993) (see also the first column of
our Table I). However, the grid presented is not fine enough to allow obtaining
the relevant percentage points of Gq,,,(x)1+ 1. Accordingly, we provide a full set
of critical values in Table II calculated with -r = .05. These were obtained using
a simulation method similar to that used for Table I.
Note that 52 is only required to be consistent under the null hypothesis for
the validity of the stated asymptotic distribution. The test may, however, have
better power if 52 is also consistent under the alternative hypothesis. Also, it is
important to note that the results carry through allowing different distributions
across segments for the regressors and the errors. That is, Proposition 7 remains

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 MULTIPLE STRUCTURAL CHANGES 61

TABLE II

ASYMPTOTIC CRITICAL VALUES OF THE SEQUENTIAL TEST FT(l + 111).

THE ENTRIES ARE THE QUANTILES X SUCH THAT Gqq (X)' =

q a 0 1 2 3 4 5 6 7 8 9

1 .90 8.02 9.56 10.45 11.07 11.65 12.07 12.47 12.70 13.07 13.34
.95 9.63 11.14 12.16 12.83 13.45 14.05 14.29 14.50 14.69 14.88
.975 11.17 12.88 14.05 14.50 15.03 15.37 15.56 15.73 16.02 16.39
.99 13.58 15.03 15.62 16.39 16.60 16.90 17.04 17.27 17.32 17.61
2 .90 11.02 12.79 13.72 14.45 14.90 15.35 15.81 16.12 16.44 16.58
.95 12.89 14.50 15.42 16.16 16.61 17.02 17.27 17.55 17.76 17.97
.975 14.53 16.19 17.02 17.55 17.98 18.15 18.46 18.74 18.98 19.22
.99 16.64 17.98 18.66 19.22 20.03 20.87 20.97 21.19 21.43 21.74
3 .90 13.43 15.26 16.38 17.07 17.52 17.91 18.35 18.61 18.92 19.19
.95 15.37 17.15 17.97 18.72 19.23 19.59 19.94 20.31 21.05 21.20
.975 17.17 18.75 19.61 20.31 21.33 21.59 21.78 22.07 22.41 22.73
.99 19.25 21.33 22.01 22.73 23.13 23.48 23.70 23.79 23.84 24.59
4 .90 15.53 17.54 18.55 19.30 19.80 20.15 20.48 20.73 20.94 21.10
.95 17.60 19.33 20.22 20.75 21.15 21.55 21.90 22.27 22.63 22.83
.975 19.35 20.76 21.60 22.27 22.84 23.44 23.74 24.14 24.36 24.54
.99 21.20 22.84 24.04 24.54 24.96 25.36 25.51 25.58 25.63 25.88
5 .90 17.42 19.38 20.46 21.37 21.96 22.47 22.77 23.23 23.56 23.81
.95 19.50 21.43 22.57 23.33 23.90 24.34 24.62 25.14 25.34 25.51
.975 21.47 23.34 24.37 25.14 25.58 25.79 25.96 26.39 26.60 26.84
.99 23.99 25.58 26.32 26.84 27.39 27.86 27.90 28.32 28.38 28.39
6 .90 19.38 21.51 22.81 23.64 24.19 24.59 24.86 25.27 25.53 25.87
.95 21.59 23.72 24.66 25.29 25.89 26.36 26.84 27.10 27.26 27.40
.975 23.73 25.41 26.37 27.10 27.42 28.02 28.39 28.75 29.13 29.44
.99 25.95 27.42 28.60 29.44 30.18 30.52 30.64 30.99 31.25 31.33
7 .90 21.23 23.41 24.51 25.07 25.75 26.30 26.74 27.06 27.46 27.70
.95 23.50 25.17 26.34 27.19 27.96 28.25 28.64 28.84 28.97 29.14
.975 25.23 27.24 28.25 28.84 29.14 29.72 30.41 30.76 31.09 31.43
.99 28.01 29.14 30.61 31.43 32.56 32.75 32.90 33.25 33.25 33.85
8 .90 22.92 25.15 26.38 27.09 27.77 28.15 28.61 28.90 29.19 29.49
.95 25.22 27.18 28.21 28.99 29.54 30.05 30.45 30.79 31.29 31.75
.975 27.21 29.01 30.09 30.79 31.80 32.50 32.81 32.86 33.20 33.60
.99 29.60 31.80 32.84 33.60 34.23 34.57 34.75 35.01 35.50 35.65
9 .90 24.75 26.99 28.11 29.03 29.69 30.18 30.61 30.93 31.14 31.46
.95 27.08 29.10 30.24 30.99 31.48 32.46 32.71 32.89 33.15 33.43
.975 29.13 31.04 32.48 32.89 33.47 33.98 34.25 34.74 34.88 35.07
.99 31.66 33.47 34.60 35.07 35.49 37.08 37.12 37.23 37.47 37.68
10 .90 26.13 28.40 29.68 30.62 31.25 31.81 32.37 32.78 33.09 33.53
.95 28.49 30.65 31.90 32.83 33.57 34.27 34.53 35.01 35.33 35.65
.975 30.67 32.87 34.27 35.01 35.86 36.32 36.65 36.90 37.15 37.41
.99 33.62 35.86 36.68 37.41 38.20 38.70 38.91 39.09 39.11 39.12

valid under A7 instead of A8-A9, provided 52 is replaced by 57 in (44) of the

Appendix.
We next argue that the test based on FT(l + II) is also consistent. If there
are more than I breaks and a model with only I breaks is estimated, there must
be at least one break that is not estimated. Hence, at least one segment contains

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
62 J. BAI AND P. PERRON

a nontrivial break point in the sense that both boundaries of each segment are
separated from the true break point by a positive fraction of the total number of
observations. For this segment, the supFT(l; q) test statistic diverges to infinity
as the sample size increases since it is consistent. Accordingly, the statistic
FT(l + 1l1) (computed for I + 1 segments) also diverges to infinity. This shows
consistency.

4.4. Extensions to Serially Correlated Errors

The tests discussed above can be applied without the imposition of serially
uncorrelated errors as specified in Assumption A9. A simple modification is to
use the following version of the F test instead of that specified in (7):

(12) FT (A ... .,Ak;(q) =( k )'R'(RV()R) R ,

where V(8) is an estimate of the variance covariance matrix of 8 that is robust
to serial correlation and heteroskedasticity; i.e. a consistent estimate of

(13) V(8) = plimT(Z'MxZ) Z'MxQMxZ(Z'MxZ)

Note that it can be constructed allowing identical or different distributions for
the regressors and the errors across segments. In some instances, the form of
the statistic reduces in an interesting way. For example, consider a pure
structural change model (,3 = 0) where the explanatory variables are
plim T- Z'QZ = hu(O)plim T- IZ'Z with h,(O) the spectral density f
the errors ut evaluated at the zero frequency. In that case, we have the
asymptotically equivalent test

FT (Al, , A q) = (52/hU(O))FT(Al . ., A;
with 2= T- Lt=It2 and hj(O) a consistent estimate of hjO). Hence, t
robust version of the test is simply a scaled version of the original statistic. This
is the case, for instance, when testing for a change in mean as in Garcia and
Perron (1996).
The computation of the robust version of the F test (12) can be involved
especially if a data dependent method is used to construct the robust asymptotic
covariance matrix of 8. Since the break fractions are T-consistent even with
correlated errors, an asymptotically equivalent version is to first take the
supremum of the original F test to obtain the break points, i.e. imposing
D = of 2I. The robust version of the test is obtained by evaluating (12) and (13) at
these estimated break dates.

5. SEQUENTIAL METHODS

In this section we discuss issues related to the sequential estimation of the

break points. We start, in Section 5.1 with results about the limit of break point

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 MULTIPLE STRUCTURAL CHANGES 63

estimates in underspecified models. An interesting by-product is a sequential

algorithm to estimate models with an unknown number of breaks discussed in
Section 5.2.

5.1. The Limit of Break Point Estimates in Underspecified Models

In this section, we show that the estimate of the break fraction in a single
structural change regression applied to data that contain two breaks converges
to one of the two true break fractions. In independent work, Chong (1994)
obtains a similar result (see also Bai (1994c) for an earlier exposition). To
present our arguments, we consider a simple three-regime model:

(14) Yt = /1E + t, if [TAj1l]?1<t<[TA.],

for j = 1, 2, 3 and with E, i.i.d.(O, o2-). Assume uL, 0 /-I2, /ut , /3, and A
so there are two break points in the model. Let Ta denote the estimated single
shift point. Our aim is to show that T,,/T is consistent for either Al or A2
depending on the relative magnitudes of the shifts and the spell of each regime.
To verify this claim, we examine the global behavior of S(T), the limit of
T-1ST([TT]). We let ST(O) andST(T) be the sum of squared residuals for the full
sample without a break and ST([TT]) is then well defined for all T ( [0, 1]. It is
not difficult to show that the convergence of T-1ST([TT]) to S(T) is uniform in
T E [0,1]. In particular

(15) 1T([ ])) 2 (1-A7)(A2-Al) 2

(16) ST([ TA2]) -p

Al S(A2)=o2+-A(A 2-

Without loss of generality we consider the case where S(A1) < S(A2); our result
is stated in the following lemma.

LEMMA 3: Suippose that the data are generated by (14) and that S(A1) < S(A2);
the estimated single break point Ta/T is consistent for A1.

The assumption that S(Aj) < S(A2) implies that the first break point is
dominating in terms of the relative magnitudes of shifts and the regime spells.
The above lemma shows that the sum of squared residuals is reduced the most
when the dominating break is identified. Given that T,/T is consistent for A1,
one can use the subsample [Tb, T] to estimate another break point associated
with a minimized sum of squared residuals for this subsample. The resulting
estimate is then consistent for A2. This follows from the same type of argument
because only A2 can be the dominating break in the sample [T, T], even if
T?f < [TA1].

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
64 J. BAI AND P. PERRON

It is relatively straightforward to extend the argument to the case where a

one-break model is fitted to a relationship that exhibits more than two breaks.
The estimate of the break fraction converges to one of the true break fractions,
namely the one which allows the greatest reduction in the sum of squared
residuals. It is also conjectured that a similar result holds when, say an m1 break
model is fitted to a relationship that has m2 breaks (with m2 > nil). Such a
general result is not, however, needed for the arguments that follow.

5.2. Sequential Estimation of the Break Points

The arguments in Section 5.1 showed that T1J/T is consistent for one of the
true break points, the one that allows the greatest reduction in the sum of
squared residuals. Suppose, as above, that this break point is A1, which, in
general, may not be known. In that case, we choose one break point either in
the intervals [1, T] or [Ta, T], such that the sum of squared residuals for all
observations [1 T] is minimized. Let r be this estimator. With probability
tending to 1 as T increases, it is easy to show that the estimated break point r
will be in the interval [ Ta T]. Similarly, if Ta is actually consistent for A2 (t
will be true if S(Al1)> S(A2)), the second estimated break point will be in [1,T
Generally, let (N1, N2) be the ordered version of (T, r) such that N1 < N. Th
(N1/T, N2/T) is consistent for (A1, A2). The preceding argument implies that
we can obtain consistent estimates of A1 and A2 in a sequential way.

5.2.1. Sequential Estimation with a Known Number of Break Points

The above analysis suggests a straightforward sequential algorithm for esti-

mating models with multiple break points. Consider first the case of a known
number of break points, say m. Once the first break point is identified, the
sample is split into two subsamples separated by this first estimated break point.
For each subsample, a one break model is estimated and the second break point
is chosen as that break point (of the two obtained) which allows the greatest
reduction in the sum of squared residuals. The sample is then partitioned in
three regimes and a third break point is selected as the estimate from three
estimated one-break models that allows the greatest reduction in the sum of
squared residuals. This process is continued until the m break points are
selected. It yields consistent estimates of the break points though the estimates
are not guaranteed to be identical to those obtained by global minimization.
Interestingly, it allows the estimation of models with any fixed number of
structural changes using least-squares operations that are only of order 0(T).

5.2.2. Sequential Estimation with an Unknown Number of Breaks

Consider now the case of an unknown number of breaks which is likely to be

of particular relevance in practice. A standard problem is that an improvement

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 MULTIPLE STRUCTURAL CHANGES 65

in the objective function is always possible by allowing more breaks. This

naturally leads to the consideration of a penalty factor for the increased
dimension of a model. Yao (1988) suggests the use of the Bayesian Information
Criterion and Liu, Wu, and Zidek (1997) suggest a modified Schwarz' criterion;
see also Yao and Au (1989). We propose an alternative method directly related
to the sequential procedure outlined above. Start by estimating a model with a
small number of breaks that are thought to be necessaiy (or start with no
break). Then perform parameter-constancy tests for eveiy subsample (those
obtained by cutting off at the estimated breaks), adding a break to a subsample
associated with a rejection using the test FT(l + 11l). This process is repeat
increasing 1 sequentially until the test fails to reject the null hypothesis of no
additional structural changes. A distinct advantage of such a model selection
device over those based on information criteria is that it can easily allow and
take into account the effect of possible serial correlation in the errors.
Note that the application of the test FT(l + ll/) in this sequential context is
different from that discussed earlier. Indeed, the result of Proposition 7 is based
on having the first 1 breaks obtained as global minimizers of the sum of squared
residuals assuming 1 breaks. The limiting distribution of the FT(l + ll/) test in
the sequential setup is the same because rate T convergence still holds, as
shown in Bai (1997), when the break points are obtained sequentially.
With probability approaching 1 as the sample size increases, the number of
breaks determined this way will be no less than the true number. The procedure
does not provide a consistent estimate of the true number of breaks, say mn,
since it implies a nonzero probability of rejection under the null hypothesis
given by the level of the test, say a. However, the asymptotic probability of
selecting a model with a larger number of breaks, say in +j, is given by a1-
which decreases rapidly. Hence, there is no need (with large probability) to
estimate models with more than the true number of breaks. The sequential
procedure could be made consistent by adopting a significance level for the test
FT(l + Ill) that decreases to zero, at a suitable rate, as the sample size increases.
A result to that effect is presented in the next proposition whose proof is similar
to that of Hosoya (1989) and is, therefore, omitted.

PROPOSITION 8: Let 'Ii be the nunmber of bereaks obtain-ed lusinlg the sequienitial

mnethod based on the staltistic FT(l + 111) applied with some size aT, and let nz() be
the true inumber of breatks. If a, conLVeiges to 0 slowly enouigh (for the test bctsed
o01 FT(l + 1 l) to remain conisistenit), theni, itinder Assimptions Al-A5, P(;h = mz()
-I1, as T -^x.

6. CONCLUSIONS

Our analysis has presented a comprehensive treatment of issues related to the

estimation of linear models with multiple structural changes, to tests for the
presence of multiple structural changes and to the determination of the number
of changes present. Our results being asymptotic in nature, there is certainly a

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 66 J. BAI AND P. PERRON

need to evaluate the quality of the approximations and the power of the tests in
finite samples via simulations. We present such a simulation study in a compan-
ion paper, Bai and Perron (1996). Among the topics to be investigated, an
important one appears to be the relative merits of different methods to select
the number of structural changes. There are, of course, many other issues on
the agenda: for instance, extensions of the test procedures to include tests that
are optimal with respect to some criteria and extensions to nonlinear models. In
addition, while the consistency and rate of convergence for the estimated break
points apply to trending regressors, the limiting distributions of the various tests
for structural change remain to be studied in the presence of trending regres-
sors.

Dept. of Economics, E52-247B, Massachusetts Institute of Technology, Cam-

bridge, MA 02139, U.S.A; jbai@[Link]
and
Dept. of Economics, Boston University, 270 Bay State Rd., Boston, MA 02215,
U. S.A.; perron@[Link]

Manuscript received September, 1995; final revision received March, 1997.

MATHEMATICAL APPENDIX

As a matter of notation, for a sequence of matrices BT, we write B

elements is op(1) and likewise for OP(l). For a matrix A, MA =I-PA w
use 11 11 to denote the Euclidean norm, i.e. llxii = (yf X7)1"2 for x E RP. For
vector-induced norm, i.e. IA II = supx , oIIAxII/IIxII. Note that the norm o
root of the maximum eigenvalue of A'A, and thus llAil < [tr(A'A)]1/2. Also, for a projection matrix
P, IIPAII < IlAll. Limits are taken as T, the sample size, increases to infinity. We start with a series of
lemmas that will be used subsequently. Assumption A5 is assumed throughout.

LEMMA A. 1: Let S and V be two matrices having the same number of rows. Then the matrix S'MV S
is nondecreasing as more rows are added to the matrix (S, V).

PROOF: Write S = (S', S)' and V= (V1', V2)'. We need to show that for an arbitrary vector a
(having the same dimension as the number of rows of S and V) a'S'MvSa 2 a'S'Mv1Sl a. Note
that a'S'MvSa (a'S'Mv1Sl a) is the sum of squares of the residuals from a projection of Sa (S a)
on the space spanned by V (V1). The inequality is verified using the fact that the sum of squared
residuals is nondecreasing as the number of observations increases (here the number of rows of Si
and S). See, e.g., Brown, Durbin, and Evans (1975). Q.E.D.

LEMMA A.2: Under A1, SUPT1. T,,(X'MZX/T) 1 = OP(1),

possible partitions such that ITi -1 - Ti I2 q (i = 1. m + 1

PROOF: We have the identity X'MZX=XXMz,X1 + *

(T1,..., T,,) leaves at least one true regime intact. In other
contains (Xi?, Z?) as a submatrix. We have XiMz.X
(X'MZX/T)-' < (Xi0'MzpXi0/T)- 1. This implies II(X'MZX/T
all partitions. The lemma now follows from Assumption A.1. Q.E.D.

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 MULTIPLE STRUCTURAL CHANGES 67

LEMMA A.3: UnderAl, SUPT,.T X MZZ? = OP(T

PROOF: Because Mz is a projection matrix, we h

uniformly over all partitions. The lemma follows fr
IIZt II = O (T1/2). Q.E.D.

LEMMA A.4: Under A4,

suprnemuim with respect
(i = 1 .. ., mi + 1) underAs
under Assumption A 4(ii).

PROOF: Consider first the case where A4(i) is assumed to hold. Because of the independence
between zs and ut, we can treat the z 's as nonstochastic, otherwise conditional argument
used. We shall prove that IU'PzUI= OP(T2a) uniformly in T,,... T7,. Note that U'PZ
summation of the m + 1 terms

H ztz f ' t
Ti+ 1 Ti+ 1 Tj+ 1

for i = 0,..., nz. Thus it suffices to prove that

(17) sup |E t = Op(T)

1<k<I<T t=k

with l-k > q and 4t = (t(k, l) = (A k)1"2z,u, with Ak! = LjkZIZ) No

l<k<l?T t=k k= 1=k+q t=k

T T 1 2s

k=1 =k+q t=k

By the mixingale property, we can write u, = -j= - ,aujt,

for each j, {ujt, _-j} is a sequence of martingale dif
j= aE t=k fjt, where jt =(A k ) -I2z,t ujt. By Minkowski's i

2s2s 11 s 2-
(19) E <, E i-[ =
t =k ( =-w[ t= k

A key point is that for fixed j, k, an

differences. Thus by Burkholder's inequal
depending on q and s, such that

(20) E 2,t ) c( ?(Eiijlti2s)1/S

where the second step follows by Minkow

(Eli ijt i2')12=zt(Akl)'zt(EIu]jt sI). By A4(
(Eiujti25)/25 ? 2ct,fljl < 2(maxici)i111j <
z'(Akl) 1ztK2f1A12. Thus from (20),

(21) E t 2 < C z '(A kl)lz K2ljl ) =sqs

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 68 J. BAI AND P. PERRON

where we have used the fact that L

(19), we have Ell LI=l = lk.S ' Cq sK S
1, this implies, in view of (18), that w
some C1 > 0. Let s = 2 + c/2 (the moment of order 4 + c of izi exists by A4(i)); we can choose an
a E (0, 2/(4 + c)) such that T-2 as + 2 -> 0. This proves (17) and hence the lemma when A4(i)
Consider now the case where A4(ii) is assumed to hold. Here, T-1 L[-t[0,]+ 1zt zQ( - Q(u)
and hence (T 1 L[TL[JrJ]+ T zzt) (Q(u) - Q(G))- uniformly in l) and it such that l) - >
Also, T-1/2L[TvU + 1zit = Op(l) uniformly using a functional central limit theorem for mar
differences. Accordingly, IU'PzUt = Op(l) uniformly in T1,..., T.. and the statement of the
holds with a = 0. Q.E.D.

LEMMA A.5: Under A1-A4, for some a < 1/2, (a) supT. T X PZU = 01)(Ta+ 1/2); (b)
SUPT. Z O'PZU= OP(T a/2).

PROOF: This follows from Lemma A.4, lIXII = O (T 112 ), aind llX'PzUll ? llXj lllPZUjj. Similar
arguments apply for part (b). Q. E.D.

PROOF OF LEMMA 1: By the de

Z' is the diagonal partition of Z at (T1,...,T ,). To prove the lemma, it suffices to show
T- U'X( /3 - o) = op,(l) and T- (UIZ_* - U'Z()8 0) = oP(l). We shall prove a stronger result. Le
(T1. T) be an arbitrary partition and Z be the associated diagonal partition of Z. Also let
/({Tj7) and 6({Tj}) be, respectively, the estimates of 3 and 8 corresponding to this same par
We shall prove

(22) sup - jU'X(: /({7j}) - /? )= OQ_(T- 1/2.) op(l),

1 -T

(23) sup -jU'Z8f({T.})-U'z 508 = 0(Ta)- l op(1),

T1,- T p

where a is given in Lemma A.4. First consider (22). We can rewrite

(24) /3({Tj}) - PC) = (X'MzX) l X'MzZO8O + (X'MzX) Y X'MzU.

The first term is O,lO) uniformly over all partitions by Lemmas A.2 and A.3. The second term
OP(Ta-/2) = 0o () by Lemmas A.2 and A.4 (note that X'MZU = Op(Ta? /2)). Thus 3({Tj}) -
= O(1) uniformly over all partitions. This implies (22) since U'X= - ,0(T /2). Next, from ={T
(ZMAZ) 1Z'MZ Y, MAX = 0, and (2), we obtain

(25) U'Z({T.}) - U'Z780

=U'Z(Z'M Z) Z'M Z08 + U'Z(Z'MXZ) Z'MyU-U'Z08().

From the identity

(Z'MxVz) ZI Z_) + (Z'Z) (Z'X)(XMzX) IX'z(z'z)

the first term of (25) is equal to

(26) U'Z( Z'M Z ) Z'Mx Z6 0 = U'Pz[ MX + X(X'MzX) X'PZ My ]Z 6 ?0.

Because PZ and Mx are projection matrices, IIMXZI <11Z1 Z 0 (T1"2)
Z_X ' II0j 0,(T). Hence, this term is 01,(T a+ /)uiomyoe ly adiXPMZi
IIXII 1z?ll P O/,(T). Hence, this term is Op /-) uniformly over all partitions using
A.4, and A.5. Similar arguments show that the second term of (25) is O0(T2 a). The last
is O (T1l2). Combining these results and noting that 2a < a + 1/2, we have U'Z8({T}
O,(T`+ '/). This implies (23). Q.E.D.

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 MULTIPLE STRUCTURAL CHANGES 69

PIZOOF OF LEMMA 2: If therc exists a break, say A(j)

I J
whiclh cannot be consistenitly estimated, theni
with some positive probability E.1 > 0 there exists a -q > 0 such that no estimated break falls in the
interval [T(A') - -q), T(A'1 + ri)] for a stubsequtence of T (withotut loss of generality, assume this
subsequence is the same as T). SuppOse this interval is classified into the kth regime, namely,

T,; < T(AOjO- ) anid T(A" + n) < T,;. Then ( -I1 =-Xtt ) + .z(8k t- for to [T(A(j1-),Tj1]
and c, d x( ) - pt + Z:(8k - 31 ) for [TAp? + 1, T(A( + -q)]. We have
T

(27) L'W> LEd+E L

1 1

(;iii:2) L-z,x, Lz I I A

where E, extencls over the set T(A - 0 ) < t ? TAj. ancl E, extends ov
+ -q). Let YT and yT be the smallest eigenvalue of the first and seco

ELW + E > >T [T P : 2 + I Wn,j + YT [ii - + II - 8> II]

1 2
rninTi{y7, I ( 11 i0 11' + , - + I 11)

? (12~~~~~
2)m-in{yT~ Tyi)}IIQ
~~~ ( /)iil7 () 1- F'12
T}l6 4- 1

The last inequality follows from

(x - ti)'A(x- a) + (x - b)'A(x- b) 2 (1 2)(a -b)'A(a -b)

for an aibiti-aiv positive definiite natrix A and for all x. Nowv the first matrix in (27) can be wivitteni as

(T-0)(i/ T)11Y i , V >nV,(Tq)AT say. By A2, the smallest eigenvalue of AT is bouncled away
from zero. Thus the smallest eigenvalue of (T-q)AT, T'T iS of the order T-q. The same can be said for
yT. Therefoie, Z Td > TC IjI - 6j? 1K for some C> 0 with probability no less than E( > 0. O.E.D.

PROOF OF PROPOSITION 2: Without loss of generality, we assume there are only thlee breaks
(n = 3) and provide an explicit proof of T-consistency for A, only. The analysis for A, and A. is
virtually the same (and actually simpler) and is thus omitted. For eaclh e > 0, let Ve = {(T1. T, T3);
Ti - Ti7)I < ET). From Proposition 1, P({T1, T1, f- E V1) -* 1. Therefore we only need to examine the
behavior of the sum of squared residuals, ST(TI, T, T3), for those Ti suchi that I Ti - Ti( < E T for all
i. Also using an argument of symmetry, we can, without loss of generality, consider the case T, < TV.
For C > 0, define

VJ0C = {( T, T,, 7'1'7; {'- Tiol < E T, I < i < 3), T-)- T," < - C}.

Thus, 1V(C) c V4. Because ST(T1, T, T,) < ST(TI, 7T, T.,) with probability I, it is enouglh to shiow that
for each r1 > 0, there exist C > 0 and e > 0 such that for large T, P(minl{ST(TI, T, T) -
ST(TI T'', T,)} < 0) < -q, or equivalently,

(28) Pf-min{[ST(Tl ? T, ) -ST(TI. To, ThT)] (7T( - T)} < 0) <71

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 70 J. BAI AND P. PERRON

where the minimum is taken over th

global optimization cannot be achieve
denote SSR1 e ST(Tl, T2, T3), SSR2 =
definition, we have

(29) ST(Tl, T2, T3)-ST(Tl, T20, TO = (SSR1-SSR3)-(SSR2-SSR3).

This latter relation is useful because it allows us to carry the analysis in terms of two problems
involving a single structural change, the first allowing an additional fourth break at time T2? b
T2 and T3 and the second an additional fourth break at time T2 between the T1 and T2?. It is then

easy to derive exact expressions for (29) in terms of estimated coefficients. Let (81*2, 8A, 83
denote the estimator of (8 , 8, 8 2, ?,6 84) based on the partition (T1, T2, T2?, T3) (note 82? is
repeated once). In particular, 82* is an estimate of 82 associated with the regressor (0, ..., 0, zT+
ZT2,0...0)', 8A is an estimate of 82 associated with the regressor Z=(0. 0, ZT2+1.
ZT?,0.--0)', and 8* is an estimate of 82 associated with the regressor (O,....O, zT2+1.ZT3 0,
0)'. Now consider SSR1 - SSR3; we have (e.g., Amemiya (1985, p. 31)),

ST (Tl, T2, T3 -ST (Tl, T2, T2T3) = ( - 8A) [Link](83 - )

where W= (X, Z), with Z the diagonal partition of Z at (T1, T2, T3). Similarly, we have for
SSR2- SSR3,

ST(T1, T2, T3)- ST(T1, T2, T2?, T3) = ( - 8-A) ZAMWZ( 82 - 8A

where W= (X, Z) with Z the diagonal partition of Z at (T1, T2?, T3). Thus

2 ( 83 - M) ZJMWZA( 83 - M) - ( 82- 8,) ZZ 2( 8-

(30) SSR1-SSR2 = (83*MA) ZAMWZA(83 -M-( 8A)Z MWZA(82* -8)

The inequality is due to ZMWZA < ZAZ. From the definition of MW, we have

(31) (SSR1- SSR2)/(T2-2)

2 (53*-M) [ZAZA T2 - T-2 ]83* - 8

_ - 8 )[ WI(TU' -IT]((~
-83 A)[ZA W/(T2?-T2)][W'W/T] [W'Z/T](-83 89

-( 2* )A {ZA AA/(T2 - T2 )} ( 2*- )

- )- (II){--(III).

Consider term (I). Note first that 8,i is close to 8,0 given that, on the
between Ti and Ti? can be controlled and made small by choosing a sm
estimated using observations from the second true regime only, 8A is close to 82? for a large enough
C, on VJ(C). Hence, for large C, large T and small E, (I) is no less than (1/2)(8? - 82)'[Z,ZA/(T2
-T,)](8 - 82) with large probability. Next consider term (II). It is easy to show that on V1(C), 8*
and 8A are Op(l) uniformly. Also on V1(C), (W'W/T)- 1 = Op(1) and ZAW/(T2?-T =Op(l)
(because ZAW involves no more than T2? - T2 observations). Furthermore,

IIW'ZA/TII = IIW'ZA/(T2 - T2 )II(T2 - T2)/T <,EOP(1).

Thus (II) is no larger than EOp(1). Consider finally (III). Because both 8* and 8A are close to 82,
182 - i81 < p with large probability for every p > 0 (this is true for large T, large C, and small E).
Also, because IIZAZA/(T2 - T2)1 = Op(l) uniformly on V1(C), term (III) is no larger than pOp(l).
Hence, the inequality

SSR1- SSR2 _2 1(8o8)' Z AZ ?

(32) >2T 0 2 0(82- 82) - EOp1- pOp(1)

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 MULTIPLE STRUCTURAL CHANGES 71

holds with large probability. By A2,

2T(

ZAZO/(T2T ) = T0' E : Z
-2t=T2+ I

has its minimum eigenvalue bounded away from zero on V1(C). Thus the first term on the right-hand
side of (32) is positive and dominates the other two terms. It follows that with large probability,
(SSRI - SSR2)/(T2 - 7,) > 0. This proves (28) and the proposition. Q.E.D.

PROOF OF PROPOSITION 4(i): The structure of the proof is similar to that of Proposition 1 but
modifications are necessary in view of the fact that T-1 T= 1d - 0 even supposing a break is
consistently estimated when the shifts are shrinking. Using (4) and (5) (without dividing T on both

sides), we can arrive at the desired contradiction if we can show that >T Td2 > 2yT 1Lt0d,
limit as T oc. To do this we show that ET[ Id2 diverges at a faster rate than ET 1ultdt.
We will make use of VT being small to strengthen the result of (22) and (23). We shall drop the

subscript T in 8 i. From 8i0 - 810+ 1 = 0(0?T) under A6, by adding and subtracting terms, we
i?- = O(LT) for all i and j. Now consider (24). The first term on the right-hand side can be
rewritten as (X'MZX)- 'X'Mz(Z0 - Z)8 0 because MZZ = 0. A key to the proof lies in the fact
that (Z? - Z)8? depends on changes in the parameters (i.e. 8i0 - 8-0). In the case of a single change
point, for example, assume T1 < TO; then

(Z -Z)80 = (0,.O-,ZT?+ I. * *ZT;,)O.0)* (80 - 80)-

This implies that X'MZ(Z -Z)80 is at most Op(T)vT. By Lemma A.2,

(X MZX) X'MZ(ZO -Z)80 = Op(v!T).

This implies that, from (24),

({Tj}) - ') = ?p (vT) + O(T`- 1/2)

Thus

(33) U'X( Pu{jl) - '?) = Op(T11/ VT) + OP(T)

over all partitions. Next, we combine the first and the third terms of (25) and rewrite them as

(34) U'Z(Z'MXZ) Z'MX(Z0-Z)80 + U'(Z-Z0)80.

Each term above involves (Z - Z0)80. Using the argument in proving (26) and (33), the
(34) is OP(Ta+ 1/2VT) and the second term is OP(T1/2LT), which is dominated by the
middle term of (25) is OP(T2a), we have

(35) U'Z8({Tj})-U'Zo80 = O,,(Ta+1/22 )?+ o(T2a).

Noting that (35) dominates (33), we have T= 1u,dt = OP(Ta+ l/2VT+T2a).

Next, consider ET 1dI. The proof of Lemma 2 is not changed under shrinking shifts. If ther
exists a change point that cannot be consistently estimated, then

E d> TC,I8r l - 8j III2 > TC'V2'

t= 1

for some C' > 0. Thus T= Id2 > 2E T 1lltdt if Tv2/(Ta+ 1/2LT + T2a) m. This is the case
T"'/2)- aL T -X oc. Under E ILtt 12/ 0 < c of A6, we can choose a such that a < i in Lemma A.4. Thus
T(T/2) a V T X by A6. Q.E.D.

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 72 J. BAI AND P. PERRON

To prove Proposition 4(ii), we first prove a lemma, which generalizes the Hajek and Renyi
inequality to mixingales.

LEMMA A.6: Let 6,,9i,} be a q x 1 L2 mixingale satisfing (a)-(e) of A4(i) withl ui u-eplaced by (i
and r1eplaced by 1 1. Then there exists ani L < - such tlhat, for eveiy c > 0 and n > 0,

( I1k L
P slip k | ,| >c C
k?mn k )= n

t= 1 ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ k

PROOF: Let (jt = E(~t 1t7j) - E( t| 17, )j Then L , _ and so L t L l=.- _

Thus, for each N > 0,

P sup E > supP s 1 ?

N2k2Z)17c t= 1 j= _rz N?k/n1 t= 1

For each j, { j 274}forms a sequence of martingale d

^j~'_=_ aj1 1. The right-hand side above is bounded by

p( supik ||/S
Ejt > ajc
1 N
< , 2 ijill2 + E Ejill)
j= - 17 i 1 i=ii1+E1

the latter bound is due to Hajek and Renyi's inequality for martingale differences. From the
definition of a mixingale and A4(i)(c), Ell _fjl2 < 4c7fr121 < 4K2ff.. Thus the above is bounde
- i=a7j-V'j)(m + LE,, + i2). Since LT, +1i2 ? 2m-1, if we let L
12K2(_J= --a72 ) then the desired upper bound is obtained for a fixed N. Since the boun
not depend on N, the lemma is obtained by letting N -cc*It remains to choose appropriat
such that E jCaI- J121 is bounded. Let P0 = 1 and vj -j 1 K(j 2 1), where K > 0 as given in
Let a1= vj/(1 + 2L> Iv ) and a1-j = aj for 1 ? 0. Then Ejaj = 1. By Assumption A4(i)(d),

qa7jf1, 2, (v~2 Y 2 -2-(1?~ 2K)<'.Q D

EJ 0 + 2 L j + 2 jQ. E. D.
j j=l j =l

PROOF OF PROPOSITION 4(i): We shall maintain all the notations in the proof of Proposition 2.
Define a new set

Vl'Q (C) = {(T1, v , T-); I Ti - Ti <T 1 < i < 3, T2 - T4? < -CIv2
which is a subset of V1. We only need to show that (28) holds when the minimum is taken over
Vet (C). We can prove that, uniformly on the set V1*(C),

(36) 8* i0 = EQ 7(1T) + 0,,(T-1/2) (i = 1,...,4),

(37) 8A 820 = (ZXZ)1 YZAU+?EOp(VT) + Op(T /).

The above is easily seen to be true in the case of pure structural changes. In this case, for example,
A- 8, is given exactly by (ZZ)- [Link]. In the case of partial structural changes, the proof of (36)
and (37) is more complicated. We shall omit the details and give a brief explanation instead. A
detailed proof is available upon request. The term EOp(VT) on the right-hand side of (36) is due to
misspecification in the sense that Ti may not be the same as Tio. However, this misspecification is

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 MULTIPLE STRUCTURAL CHANGES 73

controlled by choosing a small e. The dependence on IoT is explained in part (i). The term
Op(T 1/2) is related to disturbances and this specific rate is due to the fact that each 8.; is
estimated with a positive fraction of the sample for partitions in VQ*(C). The last two terms of (37)
are due to spillover from the misspecification via partial structural changes.
Using (36) and (37), expression (I) in (31) is no smaller than

(6X - 8) [Z0 Z/(T-? _ T)]( 0- 80) - E0((2) - 0,(T- 1/2TT)- -1).

Because the minimum eigenvalue of ZXZ-/(T2- T,) is bounded away from zero on V'(C) for all
large C and &30 - 2 = 0(0T), (I) is no smaller than Av2 - EO7(1v2), where A is a positive consta
Note that EO(IuT) dominates O,,(T-1/'2,) and O,(T-1). The latter two terms also appear in
and (III) of (31) and will be absorbed into EOQ,(72).
I Expression
T (II) in (31) is bounded by EQ 0(i'
pT
Expression (III) is bounded by

(T2( - T2) U'Z (Z' Z )Z U+ E OP(I1)

From [Z'ZA/(T7) -T2)T]- Op(l) on V'(C) for large C, (III) is further bounded by

O,,(1)jj(T|2 - T) IZAUI1 + eO/(L'T)

In summaly,

SSR_I-_SSR, ZAu 2

(38)~~~~~~~~~~~~~~~~2
For every q > 0, we can choose a small e > 0 such that
choose B < -- such that P(10j1)1 > B) < V. Thus

P(min{(SSR1 - SSR7)/(T( - 1 )} < 0)

< 2X1 + P(max{Bjj(T, - T,)l ZjUII2} >AuIT12)

(~~~~~~~

2r (1 m7.,laxC W' - T2) 11 it SI11>[A/(2B)] /2T).

By Lemma A.6 with 7 lit' c = [A/(2B)] 12 VT, and m = CVT2 (applied with data o
i.e. treating T,? as the first observation), the above probability is bounded by

2X1 + (2B/A)L(v2Cv,7 '))= 2X1 + (2B/A)LC- 1 < 3X1

for large C. Q.E.D.

PROOF OF PROPOSITION 6: Note that we can write

FT(Al, , AR; q) = (SSRO-SSRk)/[kq(T-(k+1)q-P)ISSR/],

where SSRo and SSR/k are the sum of squared residuals under the null and alternative hypothese
respectively. We have

(T-(k + )q - p) SSRk /) ap .
Hence, we concentrate on the limit of F* = SSR0 - SSR k. Now, let DU(i, j) (DR(i, j), resp.) be the
sum of squared residuals from the unrestricted (restricted, resp.) model using data from segments i
to j (inclusively), i.e. from observation Ti- I + I to Tj. We can write FT = DR(1, k + 1) - L11DU(i, ),
or

(39) F*= T [D(1, i + 1) - DR(1, i)-DU(i + 1, i + 1)] + DR(, 1)- DU(1, 1).

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 74 J. BAI AND P. PERRON

Let /l3u and ,3R be the estimate of /3

have f3'u = (X'MZX)- 1X'MZY and
Y1,j, U1,j, X1,j, and Zl,j denote the c
belonging to the partition from segmen
vectors or matrices containing elemen
using data on the z's from segment
JJ 1 to
using data on the z's from segment j only in the unrestricted model. We have

81,j = (Zi 1Z1,j) Z1 (Y1, jX1, f3R) and

45 = (z,z.)1 z' (y - Xi ,u)

Using the fact that, under the null hypothesis,

Y=X,/+Z8+U=X,/+Z8+U and

Yj = Xi 3+Z8?+ Uj
(with 8 = (8. 8) a q(k + 1) vector with 8 defined by 81 = 82 = 8k+ 1 8), straightforward
algebra yields

DR(1, j) = 1(I1- Pzlj )(Uj1 -X1 jAT)II2 and

DU(, j) = (I Pzj)(Uj XAT)II2,

where AT = (X'MzX)- 'X'MZU, and AT = (X'MZX)-1X'MZU. Consider the ith element in

summation defining FT* in (39); we have

(40) FT i = D R(j, i + 1) -D R(j, i) - Du( + 1, i + 1)

=1(I - Pz,')(Uli+1 Xli+ AT)II2 II(, _P i)(U1 iX1 iAT)II2

i+ 1 AT)II2.

To simplify the exposition, let Sj = ZU jUl j, Hj = Z Z1 j, Kj -Z X , L

X1 jU1 j. Noting that

U1j+ lUl1i+ 1 = U ,1ul,i + Ui'+ lui+ 1,

1 i+ l i+ 1 =x ,ixl, +xi+ xi+~1

and

U1i+ lXl,i+ 1 = ui;x,i+c + lxi+i1

we deduce that

(41) FT i = -S'+ 1H+' Si+ 1 + S'Hi 'Si + (Si+ 1 - Si)Y[Hi+ 1 - Hi]- '(Si+ 1 - Si)

+ 2Si 1H[i+l'Ki+ AT -2SHi7'KiAT

-2(Si+ 1-Si) [Hi+ 1-Hi] '(Ki+ 1-Ki),TT

+ 2(M+ 1 -M)G(ATT-AT) + (AT -AT) (Li+ 1 - Li)(AT -AT).
Using the stated assumptions, we have the following basic convergence results:

(i) T- 1/2 (X1j, Zl j)'Ul => (B1( Aj), B2(Aj))'

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 MULTIPLE STRUCTURAL CHANGES 75

where B(r) is a (q +p) dimensional vector Brownian motion with covariance matrix

Q [Q ll Q 12
([i) T-1(X1 j,j1 1(XI .,Zl.j) > ( 2A Q
From these two limits, we deduce easily the following results:

(a) T-1/25j= (JB9(Aj);

(b) TH-1H A Q2AQ;

(c) T-1Kj (J2A Q21 ;

(d) T-'Li j (7- 2Aj Q.t1;
(e) T- l 2M. ==> (TB,( Aj );

(f) T19A=T o- -(Ql -Q1,Q2 Q1Q2)(B(1) - Q12Q221B2(1)) .

It remains to consider the limit of T1]2AT. Let A = diag{A1, A2 - Al,. 1 - Ad}, a (k + 1
diagonal matrix. We deduce that

(i) T-1Z'Z --- o2(A 0 Q22);

(ii) T- 'X'Z -- a2(e'A ? Q12) where e' = (1,1,...,1),

a (k + 1) vector;

(iii) T- 1/2Z'U==> uT(B2(A), B2(A) -B,(A1),., B2(1) -B(Ak))B-

We then obtain

(42) Ttl--A =[T- 1X'X- T- X'Z(T- 'Z'Z) -T- 'Z'X]

x [T- XU- T- 1X'Z(T- 'ZZ)1 T- 1 Z'U]

=>( t[Qll -(eW/1 Q12)(A (&Q22)_ ( Ae Q21 A

x [B1() - (e'A Q12)(A 0 Q22) B*]

= or-1[Ql - (e'Ae & Q2 Q- Q )] [BQ(1)-(e' 1QiQ22I)B*]

= - Q12 Q Q122Q]2[B1(1)-Q19Q-,B9(1)] -A.

The second equality follows since e'Ae = 1 and (e' 2 Q12Ql)B* = Q1Q-B2(1). Using the re
stated above we easily deduce that (Mi+ - M)'(AT -AT) =? 0, (AT -AT)A(Li+ - L)(AT
0, and

St+lHi7}lKi+1AT -SiHi KiAT-(Si+ -Si)'[Hi+ 1-Hi]l(Ki+ 1-Ki)AT

=> o-B2(A+1 )Q2,2 Q21A` -(B2(Aj)Q-1 Q21A*

-ff ( B, (Ai + ) )- B,(Ai))Q22lQ21A* =O.

Hence, we are left with

FT, i =-S'+ lHi-',Si+ 1 + S'Hi 'Si + (Si+ 1 - Sj) [Hi+ 1 - Hi] I(Si+ 1 - Si) + (1)

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 76 J. BAI AND P. PERRON

and we deduce, using the fact that B,(A) -j),

(43) FTi -B2(Ai+ J)Q 'Bs(Ai+ 1)/Ai+ + B2(Ai)Q,9'B9 (Ai)/Ai

+( B, (Ai+,I )- B ( Ai))Q - ,1 (B( -i l)B, (Ai))/( Ai + 1-Ai)

=- W - Io(JAi+l )Il/Ai+ I I I W1( Ai )2/Ai

+? oHJJW4(A-) W-Jq(Ai)f2/(Ai+ - Ai)

= c-IIAJAiW(A?+1) - Ai+ W(Ai)J/Ai+Ai(Ai+I - Ai).

Finally, it is easy to verify that DR(1, 1) - DU(1, 1) 0. Note that this convergence result holds
jointly for i = 1.., k; hence

DE11 A-WJ(4A,+) - A,i+ ?WJ(Ai)11

ETAi=IAiAi - A~)
and the result of Proposition 6 follows. Q.E.D.

PROOF OF PROPOSITION 7: For simplicity, we present the arguments in the case of a pure
structural clhange. Let SSR(i,j) be the minimized sum of squared residuals for the segment
containing observations from (i + 1) to j; then we can write

(44) FT( + ) = sup sup {SSR(Ti-1,Ti)-SSR(Ti,T) -SSR(T,T)}/(2-

| <j< 1? T1 TA T Ti

Under Assumptions A8-A9, arguments as in the proof of Proposition 6 show that

(45) u- 2 SUp {SSR(Ti? 1, Ti7 )-SSR(Ti I1, T )-SSR(T,v Ti7)}

IWq( ) - /W?(1)112
sup

where dA? is as defined in (11) with 7i replaced by Ti?. Under the null hypothesis, Propositio
asserts that T7 = 0 + 0,(l). Using this result, we can show that (45) also holds with 7Ti 1 and T
replaced by T> and Ti, respectively. In addition, because over different regimes SSR(-,-) are
computed using nonoverlapping observations, the weak limits in (45) foi different i's are indepen-
dent. Tlhus the limit of (44) is the maximum of I + 1 independent random variables in the form of
(45).

PROOF OF LEMMA 3: We show that S(T) for T E [0,1] has a unique minimum at Al. The function
S(T) has different expressions over [0, 1]. Some algebra reveals that

A1- T
S(T) - S 1At) -A [(I1 Al)( /jl- /2,) + 0 - A,)( T2K ] 7< Al

which is nonnegative. Under the assumption that S( Al) < S( A2), the expression in brackets is
nonzero, so S(T) - S(A1) is strictly positive for T < A1. By symmetry (regarded as reversing the data
order), S(T) - S(A9) is nonnegative for T> A,. Thus for T E [A,, 1],

S(T) -S(A1) = S(T) -S(A,) + S(A2) -S(A1) > S(A2) -S(A1) > 0.

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 MULTIPLE STRUCTURAL CHANGES 77

It remains to consider the case where T E (Al, A,). Ag

A, Al TO - A)(0 - A,)
S(T)- S(A1) = (T- At) [A. K- /J L 2 I A-)(l-A _) 1 2
A , Al (1:-A?) 1 -
2 (sAt) A, At ( I- )' -O

? (T-Al.)[S(A2) -5S(Al)]

where the first inequality follows from [T(I - A,)]/[AM(1 - )] < I and the second inequali
from A./2 ? 1. Thus S(T) - S(A1) is strictly positive for s e (Al, A,) ancd we have sho
has a uniquLe global minimum at A, when S(Al) < S(A). Because 5Th,) ? 5T([TA ]), it
To/T is consistent for A1. Q. E.D.

REFERENCES

AMEMIYA, T. (1985): Advaniced Econiomiietrics. Cambriidge: Harvard University Press.

ANDREWS, D. W. K. (1988): "Laws of Large Numbers for Dependenit Nonidentically Distributed
Random Variables," Econometric Theory, 4, 458-467.
(1991): "Heteroskedasticity and Autocorrelation Consistenit Covariance Matrix Estimiiation,"
Econometrica, 59, 817-858.
(1993): "Tests for Parameter Inistability and Structural Change witlh Unknlownl Clhange
Point," Econonietrica, 61, 821-856.
ANDREWS, D. W. K., I. LEE, AND W. PLOBERGER (1996): "Optimiial Changepoint Tests for Nor-mal
Linear Regression," Joiu'nal of Econ-0om1etrics, 70, 9-38.
ANDREWS, D. W. K., AND W. PLOBERGER (1994): "Optimal Tests wlhen a Nuisaince Paramiieter is
Present Only Under the Alterniative," Econiom1etrica, 62, 1383-1414.
BAI, J. (1994a): "Least Squares Estimationi of a Slhift in Linear Processes,' Joiuviial of Time Series
Analysis, 15, 453-472.
(1994b): "Estimation of Structural Change Based on Wald Type Statistics," Working Paper
No. 94-6, Department of Economics, M.I.T., fortlhcominig in Revieiv of Economics anid Statistics.
(1994c): "GMM Estimation of Multiple Structdral Clhanges," NSF Proposal.
(1995): "Least Absolute Deviationi Estimate of a Slhift," Econometric Theoiv, 11, 403-436.
(1997): 'Estimating Multiple Breaks One at a Time," Econiomiietic Thleomy, 13, 315-352.
BAI, J., AND P. PERRON (1996): "Computationi anid Analysis of Multiple StructuLral Clhange Models,"
Manuscript in Preparation, Universit6 de Montr6al.
BHATTACHARYA, P. K. (1994): 'Some Aspects of Clhanige-poinit Anialysis,' Chliange Poinit Prlobleimis,
IMS Lecture Notes-Monograph Series, Vol. 23, ed. by E. Carlstein, H.-G, Miiller, and D.
Siegmund. Hayward, CA: Institute of Mathematical Statistics, 28-56.
BROWN, R. L., J. DURBIN, AND J. M. EVANS (1975): "Teclhniques for Testinig the Conistanicy of
Regression Relationships Over Time," Journzal of the Roval Statistical Society, Series B, 37,
149-192.
CHONG, T. T-L. (1994): 'Consistency of Clhange-Point Estimators Wlhen the Numliber of Chanige-
Points in Structural Models is Underspecified," Manutscript, Departmenit of Economiiics, Univer-
sity of Roclhester.
DELONG, D. M. (1981): "Crossing Probabilities for a Square Root Boundary by a Bessel Process,"
Conininiiiiicatiomis in Statistics-Tileoy amid Mllethlodis, AI0(21), 2197-2213.
FEDER, P. I. (1975): 'iOn Asymptotic Distributioni Tlheory in Segmented Regression Problem:
Identified Case," Anniials of Statistics, 3, 49-83.
GALLANT, A. R., AND W. A. FULLER (1973): "Fittinlg Segmenited Polyniomial Regression Models
wlhose Join Poinits lhave to be Estimated?" JournlFal of the Anwiericani Statistical Association,
144-147.
GARCIA, R., AND P. PERRON (1996): "Anl Anialysis of the Real Interest Rate unlder Regime Slhifts,'
Revietw of Ecoiionomics aild Statistics, 78, 111-125.

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]
 78 J. BAI AND P. PERRON

HALL, P., AND C. C. HEYDE (1980): Martingale Limit Theory and its Applications. New York:
Academic Press.
HANSEN, B. E. (1991): "Strong Laws for Dependent Heterogeneous Processes," Econometric Theory,
7, 213-221.
HOSOYA, Y. (1989): "Hierarchical Statistical Models and a Generalized Likelihood Ratio Test,"
Joumal of the Royal Statistical Society, Series B, 51, 435-447.
KRISHNAIAH, P. R., AND B. Q. MIAO (1988): "Review about Estimation of Change Points," in
Handbook of Statistics, Vol. 7, ed. by P. R. Krishnaiah and C. R. Rao. New York: Elsevier.
Liu, J., S. Wu, AND J. V. ZIDEK (1997): "On Segmented Multivariate Regressions," Statistica Sinica,
7, 497-525.
McLEISH, D. L. (1975): "A Maximal Inequality and Dependent Strong Laws," The Annals of
Probability, 5, 829-839.
PERRON, P. (1989): "The Great Crash, the Oil Price Shock and the Unit Root Hypothesis,"
Econometrica, 57, 1361-1401.
POLLARD, D. (1984): Convergence of Stochastic Processes. New York: Springer-Verlag.
YAO, Y.-C. (1987): "Approximating the Distribution of the ML Estimate of the Change-Point in a
Sequence of Independent r.v.'s," Annals of Statistics, 3, 1321-1328.
(1988): "Estimating the Number of Change-Points via Schwarz' Criterion," Statistics and
Probability Letters, 6, 181-189.
YAO, Y.-C., AND S. T. Au (1989): "Least Squares Estimation of a Step Function," Sankhya, 51, Ser.
A, 370-381.
ZACKS, S. (1983): "Survey of Classical and Bayesian Approaches to the Change-Point Problem: Fixed
and Sequential Procedures of Testing and Estimation," in Recent Advances in Statistics, ed. by M.
H. Rivzi, J. S. Rustagi, and D. Sigmund. New York: Academic Press, 245-269.

This content downloaded from

[Link] on Fri, 23 Jun 2023 [Link] +00:00
All use subject to [Link]