westerlund 2006

Upload: meone99

Post on 02-Jun-2018

221 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/10/2019 Westerlund 2006

    1/33

    PDFlib PLOP: PDF Linearization, Optimization, Protection

    Page inserted by evaluation versionwww.pdflib.com [email protected]

  • 8/10/2019 Westerlund 2006

    2/33

    Testing for Panel Cointegration with Multiple

    Structural Breaks*

    Joakim Westerlund

    Department of Economics, Lund University, Lund, Sweden(e-mail: [email protected])

    Abstract

    This paper proposes a Lagrange multiplier (LM) test for the null hypothesis of

    cointegration that allows for the possibility of multiple structural breaks in

    both the level and trend of a cointegrated panel regression. The test is general

    enough to allow for endogenous regressors, serial correlation and an unknown

    number of breaks that may be located at different dates for differentindividuals. We derive the limiting distribution of the test and conduct a

    small Monte Carlo study to investigate its finite sample properties. In our

    empirical application to the solvency of the current account, we find evidence

    of cointegration between saving and investment once a level break is

    accommodated.

    I. Introduction

    There is a vast literature concerned with extending time-series cointegrationor, rather, non-cointegration tests, to panel data with both a time-series

    dimension T and a cross-sectional dimension N. However, there is only a

    handful of studies that have attempted to develop panel data tests under the

    maintained hypothesis of cointegration. The most noticeable study in this field

    is that of McCoskey and Kao (1998). In their study, the authors extend to

    *The author thanks Anindya Banerjee and two anonymous referees for many valuable commentsand suggestions. The author would also like to thank David Edgerton, Johan Lyhagen, MichaelBergman and seminar participants at Lund University. Financial support from the Jan Wallander and

    Tom Hedelius Foundation, research grant number P2005-0117:1, is gratefully acknowledged. Theusual disclaimer applies.

    JEL Classification numbers: C12, C32, C33, F21.

    OXFORD BULLETIN OF ECONOMICS AND STATISTICS, 68, 1 (2006) 0305-9049

    101 Blackwell Publishing Ltd, 2006. Published by Blackwell Publishing Ltd, 9600 Garsington Road, Oxford OX4 2DQ, UK

    and 350 Main Street, Malden, MA 02148, USA.

  • 8/10/2019 Westerlund 2006

    3/33

    panel data the univariate LM test for cointegration proposed by Harris and

    Inder (1994) and Shin (1994), which is the locally best unbiased invariant test

    of the error covariance matrix of a linear regression model. McCoskey andKao (1998) show that the proposed test has a limiting normal distribution

    under the null hypothesis of cointegration as T tends to infinity and then N

    sequentially. Because of its good finite sample properties and the attractive-

    ness of the null hypothesis of cointegration rather than the opposite, the test

    has become very popular in empirical research (e.g. see, McCoskey and Kao,

    2001).

    Having been derived using sequential limit arguments wherebyTis passed

    to infinity prior to N, the LM test may be motivated in research situations

    involvingTwhich is substantially larger thanN. However, researchers need to

    be aware that the probability of a structural break in the cointegration relation

    increases with the length of the time-series dimension of the panel. As shown

    by Hao (1996), this alters the limiting distribution of the test as the

    deterministic component of the cointegration regression needs to be modified

    to collect for the presence of the structural break. Erroneous omission of

    structural breaks is therefore likely to lead to size distortions and deceptive

    inference when testing for cointegration. One study that explicitly addresses

    this concern when testing the hypothesis of a unit root in panel data is that of

    Im, Lee and Tieslau (2005), which generalizes the univariate LM unit root test

    of Amsler and Lee (1995). However, to our knowledge there are no such studythat deals with the hypothesis of cointegration in the presence of structural

    change in panel data.

    In this paper, we propose a simple test for the null hypothesis of

    cointegration that accommodate for structural change in the deterministic

    component of a cointegrated panel regression. The test is based on the LM

    test of McCoskey and Kao (1998) and it is able to accommodate for an

    unknown number of breaks in both the constant and trend of the individual

    regressions, which may be located at different dates for different individuals.

    Test statistics are derived when the locations of the breaks are knowna priori and when they are determined endogenously from the data. Test

    statistics are also derived when there is no break but the deterministic

    component includes individual-specific constant and trend terms.

    Using sequential limit arguments, it is shown that the test has a limiting

    normal distribution, i.e. free of nuisance parameters under the null hypothesis.

    In particular, it is shown that the limiting distribution is invariant with respect

    to both the number and the locations of the breaks and that there is no need to

    compute different critical values for all possible patterns of break points as in,

    e.g. Bartley, Lee and Strazicich (2001). This invariance property makes the

    test computationally very convenient. We also evaluate the small-sample

    performance of the test via Monte Carlo simulations. The results suggest that

    102 Bulletin

    Blackwell Publishing Ltd 2006

  • 8/10/2019 Westerlund 2006

    4/33

    the test has small size distortions and reasonable power. In our empirical

    application, we reexamine the data set employed by Ho (2002) and Taylor

    (2002) to assess the solvency of the current account. Contrary to much of thefindings presented in the earlier literature, we find evidence suggesting that

    saving and investment are cointegrated once a level break in the cointegration

    vector is accommodated.

    The paper proceeds as follows. In section II, we present the LM test

    statistic under the assumption that the locations of the structural breaks are

    known. Sections III and IV describes the asymptotic distribution of the test,

    while section V relaxes the assumption of known breaks. Section VI is then

    devoted to the Monte Carlo study, whereas section VII relates itself with the

    empirical application. Section VIII concludes the paper. For notational

    convenience, the Bownian motion Bi(r) defined on the unit interval

    r2 [0, 1] will be written as only Bi and integrals such asR1

    0Wirdr will

    be written asR1

    0Wi,

    Rr0

    Wisds asRr

    0Wi and

    R10

    WirdWir asR1

    0 WidWi. The

    symbol is used to signify weak convergence of the associated probability

    measure, [z] is used to denote the largest integer less than zandkzk denotesthe Euclidian norm tr(z0z)1/2 ofz.

    II. The panel LM test with breaksConsider the multidimensional time-series variableyit, which is observable for

    i 1, . . . ,N cross-sectional and t 1, . . . , T time-series observations. Thedata generating process (DGP) for yit is given by the following system of

    equations

    yitz0itcij x

    0itbi eit; 1

    eitrit uit; 2

    ritrit1 /iuit; 3

    where xit xit)1 + vit is a K-dimensional vector of regressors and zit is avector of deterministic components. The corresponding vectors of param-

    eters are denoted bi and cij respectively. The index j 1, . . . , Mi + 1 isused to denote the structural breaks. There can be at most Mi such breaks,

    or Mi + 1 regimes, that are located at the dates Ti1, . . . , TiMi, where Ti0 1and TiMi+1 T. Furthermore, the initial value of rit is assumed to be zero,which entails no loss of generality as long as zit includes an individual-

    specific intercept. For convenience in constructing the test and in deriving

    its asymptotic distribution, we assume that the vector wit uit; v0it0 iscross-sectionally independent and that it follows a general linear process

    103Panel test with multiple structural breaks

    Blackwell Publishing Ltd 2006

  • 8/10/2019 Westerlund 2006

    5/33

    whose parameters satisfy the summability conditions of the following

    assumption.

    Assumption 1(Error process). (i) The vectorswijandwktare independent forallj,tandi 6 k; (ii) The vectorwitsatisfies wit Ci(L)it, where L is the lagoperator, CiL

    P1j0CijL

    j;P1

    j0 kCijC0ijk< 1 and it is distributed

    independent and identically distributed (i.i.d.) (0, IK+1); (iii) the lower right

    K Ksubmatrix ofXi Ci(1)Ci(1)0 is positive definite.

    In addition to this, we make the following assumptions regarding the

    structural break model.

    Assumption 2 (Break model). (i) The locations of the structural breaks are

    given as a fixed fraction kij2 (0, 1) ofTsuch thatTij [kijT] and kij)1 < kijfor j 1, . . . , Mi; (ii) Both Mi and kij are assumed to be known.

    Assumption 1 (i) states that the members of the panel are independent

    of each other. This assumption is convenient as it will allow us to

    apply standard central limit theory in a relatively straightforward manner.

    For many applications, such as the one considered for this paper, however,

    the independence assumption may be quite restrictive and section VII

    therefore suggests some possibilities for dealing with the case of cross-

    sectionally correlated data. However, for the present we maintain Assump-

    tion 1 (i).As for the time-series dimension, Assumption 1 (ii) states that the standard

    invariance principle is assumed to hold for each individuali as Tgrows large,

    which ensures that the partial sum process constructed from witconverges to

    the vector Brownian motion Bi. Thus, if SiT PTr

    t1wit, then T)1/2SiT

    Bi Ci(1)Wi as T! 1 for a fixed N, where Bi (Bi1, Bi20)0 is a vector

    Brownian motion andWi (Wi1, Wi20)0 is a vector standard Brownian motion

    with covariance matrix equal to identity. The covariance matrix of Bi is

    defined as

    Xi limT!1

    T1E SiTS0iT

    x2i11 x0i21

    xi21 Xi22

    :

    This matrix can be regarded as the long-run covariance matrix of wit in the

    sense that it captures both the contemporaneous variances and covariances

    as well as the covariances at all lags and leads. For later discussion, we

    also define the long-run variance of uit conditional on vit as x2i1:2

    x2i11 x0i21X

    1i22xi21. Assumption 1 (iii) requires Xi22being a positive definite

    matrix, which precludes the possibility of cointegration within xitin the event

    that we have multiple regressors. The off-diagonal elements ofXicaptures the

    dependence between the first differentiated regressors and the equilibrium

    104 Bulletin

    Blackwell Publishing Ltd 2006

  • 8/10/2019 Westerlund 2006

    6/33

  • 8/10/2019 Westerlund 2006

    7/33

    non-empty. Thus, we require that N1/N! d as N! 1, where d 2 (0, 1].The panel LM test statistic for this hypothesis is given as follows.

    Definition 1 (The panel LM test statistic). Letx2i1:2 x2i11 x0i21X

    1i22xi21

    and SitPt

    kTij11eik; where e

    it is any efficient estimate of eit. The panel

    LM test statistic is defined as follows

    ZM XNi1

    XMi1j1

    XTijtTij11

    Tij Tij12x2i1:2S

    2it: 4

    Remark 1. The statistic is written as an explicit function of M

    (M1,. . .

    , MN)0

    to denote that it has been constructed for a certain number ofbreaks for each cross-section and that its asymptotic distribution depend on it.

    Remark 2. To compute the statistic, we need to obtain an efficient estimator

    eitof the error term eit. Under Assumption 1 this can be achieved using either

    the dynamic ordinary least square (DOLS) estimator of Saikkonen (1991) or

    the fully modified OLS (FMOLS) estimator of Phillips and Hansen (1990).

    The estimation is carried out separately for each subsample ranging from Tij)1toTijtime-series observations. For Case 1, the regression is fitted without any

    deterministic components. For Cases 2 and 4, the regression is fitted with an

    individual-specific intercept as the deterministic component, while, for Cases3 and 5, it is fitted with both individual-specific intercepts and trends. In

    addition, for Cases 13, we have Mi 0 for all i suggesting that theestimation is carried out using the entire sample.

    Remark 3. To be able to construct x2i1:2, we need to obtain a consistent

    estimatorXi ofXi. To this end, because we are allowing for general forms of

    serial dependence within the individual regressions over time, consistency

    may be achieved using any semiparametrical kernel estimator of the following

    form

    XiT1

    Xkjk

    1 j

    k 1

    XTtj1

    witw0itj:

    As OLS is consistent even with endogenous regressors, Xi may be

    constructed using wit eit; v0it

    0, where eitdenotes the OLS estimate of eit.

    The weight function 1 ) j/(1 + k) is the Bartlett kernel, which ensures the

    positive semidefiniteness ofXi. For consistency of the test, it is necessary

    that the bandwidth parameter k satisfies the property that k! 1 and k/T! 0 as T! 1. The choice k [T1/3] is sufficient. As the long-runconditional variance of eit is assumed to be constant, the estimation of x2i1:2may be carried out using the full length of the time-series dimension with

    106 Bulletin

    Blackwell Publishing Ltd 2006

  • 8/10/2019 Westerlund 2006

    8/33

    dummy variables to account for the breaks. However, in practice, it may be

    more convenient to estimate not only the parameters but also the variances

    separately for each segment, which does not alter the limit distribution of thetest. Besides being easy to implement, this approach has the additional

    advantage that it makes the test robust to shifts in the variance if we assume

    that they are located at the same dates as the shifts in the deterministic

    component.

    III. Asymptotic distribution

    In this section, we derive the asymptotic distribution of the test statistic under

    the null hypothesis. To this effect, we positQ and R, respectively, to be the

    expected value and the variance of the following vector standard Brownian

    motion functional

    Fi

    Z 10

    V2i where Vi Wi1:2

    Z r0

    Wi2

    Z 10

    Wi2 W0i2

    1Z 10

    Wi2dWi1:2:

    The process Wi1:2 x1i1:2Bi1:2 with Bi1:2 Bi1 x

    0i21X

    1i22Bi2 is a scalar

    standard Brownian motion, that is independent of Wi2. The vector standard

    Brownian motion Wi2 may be written as Wi2 z0; W0i2

    0, where z is the

    appropriately scaled limit of the deterministic componentzit. Specifically,z{[} in Case 1,z 1 in Cases 2 and 4, and z (1, r)0 in Cases 3 and 5. Thefunctional Fi is referred to as a generalized Cramer-von Mises distribution

    with 1 d.f. This functional has the additive property that the sum of Mi + 1

    independent generalized Cramer-von Mises distributions with 1 d.f. each is

    Cramer-von Mises with Mi + 1 d.f. The expected value of such a sum is equal

    to (Mi + 1)Q and the variance is equal to (Mi + 1)2R. Let us define H and R as

    the cross-sectional averages of these expectations and variances for each

    individual. As the following theorem indicates, when the test statistic is

    normalized by the appropriate order ofN, then the asymptotic distribution will

    depend only on the known values of H and R.

    Theorem 1(Asymptotic distribution). Under Assumptions 1 and 2, and the

    null hypothesis of cointegration, as T! 1and then N! 1

    N1=2ZM N1=2H )N0; R: 5

    Remark 4. The proof of Theorem 1 is provided in the Appendix. It uses the

    sequential limit theory developed by Phillips and Moon (1999), which means

    that in deriving the asymptotic distribution of the test, we first take the limit as

    T! 1 followed sequentially by the limit as N! 1. The sequential limittheory greatly simplifies the derivation of the limit distribution of the test but it

    107Panel test with multiple structural breaks

    Blackwell Publishing Ltd 2006

  • 8/10/2019 Westerlund 2006

    9/33

    is not the most general method. In fact, by using the method of joint limits, it

    is possible allow for both Tand Nto approach infinity concurrently rather in

    sequence. As pointed out by Phillips and Moon (1999), this implies that thejoint limit distribution characterizes the limit distribution for any monotonic

    expansion rate ofTrelative to N. The authors also provide a set of conditions

    that are required for sequential convergence to imply joint convergence.

    Remark 5. In our case, the sequential limit method substantially simplifies the

    derivation of the limit distribution for at least two reasons. One reason is that it

    allows us to control the effects of the nuisance parameters associated with the

    locations of the structural breaks and the serial correlation properties of the

    data in the first step as T! 1by virtue of the standard invariance principle.

    To get an intuition on this, note that each of the subsample quantitiesPTijtTij11

    Tij Tij12x2i1:2S

    2it comprised of Z(M) can be treated as a test

    statistic that is not subject to structural change. As shown in the Appendix,

    each of these subsample statistics converges to a generalized Cramer-von

    Mises distribution with 1 d.f. as T! 1. Hence, by the additive property ofFi, the intermediate limit distribution of Z(M) for each individual can be

    described entirely in terms of a Cramer-von Mises distribution with Mi + 1

    d.f. This means that the limit distribution is invariant not only with respect to

    the serial correlation properties of the data but also with respect to the

    locations of the structural breaks and that it only depends onMi, the number of

    breaks for each cross-section, which is assumed to be known. Another reason

    why sequential limits simplify the derivation is that, as the members of the

    panel are assumed to be independent of each other, the statistic may be treated

    as a sum ofNindependent drawings from a set of underlying distributions to

    which standard LindbergFeller central limit arguments may be applied to

    obtain a limiting normal distribution.

    Remark 6. To be able to compute the standardized statistic in equation (5),

    we need to obtain numerical values of the moments Q and R. One standard

    way in which moments have been obtained in situations where no closed formexpressions exist is to use Monte Carlo simulation of the limiting distribution

    of the test. Table 1 presents the simulated values of Q and R using this

    method. To obtain the moments, we make 10,000 draws ofK+ 1 independent

    random walks of lengthT 1,000. By using these random walks as simulatedBrownian motions, we can construct approximations of the functional Fi,

    which are then used to compute the moments. The moments apply in general

    for any deterministic specification and for any number of regressors when the

    slope parameters are estimated separately for each member of the panel.

    However, the specific values ofQ and R depend on the particular model, such

    as whether individual-specific intercepts or trends have been included in the

    estimation, and onK, the number of regressors. Accordingly, Table 1 gives the

    108 Bulletin

    Blackwell Publishing Ltd 2006

  • 8/10/2019 Westerlund 2006

    10/33

    moments for each of the five deterministic specifications and for up to five

    regressors. To obtain H based on the simulated expectations in Table 1,

    simply multiply the appropriate expected value byMi + 1 for each individual

    and then take the cross-sectional average. Similarly, to obtain R, multiply the

    appropriate variance by the square ofMi + 1 and then take the cross-sectional

    average. Once the standardized test statistic in equation (5) has been

    computed, it should be compared with the right tail of the normal distribution.Thus, large positive values indicate that the null hypothesis should be rejected.

    Remark 7. The fact that the moments of Table 1 does not depend on the

    locations of the structural breaks is a great advantage. For example, the

    univariate cointegration test proposed by Bartley et al. (2001) allows for a single

    break to affect the deterministic component of the cointegration vector.

    However, this allowance makes the test asymptotically dependent on the

    location of the break. To accommodate for this, the authors simulate critical

    values for a few equally spaced break points, which can be used to perform the

    test. There are at least two problems with this approach. The first problem is thatassuming equally spaced breaks is not appropriate in general in which case

    critical values for a possible continuum of breakpoints may be required. The

    second problem is that, even with discrete and equally spaced breaks, one faces

    the complication factor of having to consult a table for the critical values for

    every break location, which may be a very tedious undertaking in itself. With

    multiple breaks, this approach becomes practically infeasible as constructing

    tables for every possible pattern ofMibreaks would be extremely cumbersome.

    Remark 8. Note that the breaks are allowed under both the null and alternative

    hypotheses. This is very convenient as it facilitates a straightforward

    interpretation of the test result. To illustrate this, let us consider the univariate

    TABLE 1

    Simulated asymptotic moments

    Moment Case K 1 K 2 K 3 K 4 K 5

    Q 1 0.35500 0.26963 0.20584 0.16688 0.13770

    2/4 0.11601 0.08464 0.06539 0.05295 0.04442

    3/5 0.05530 0.04686 0.04063 0.03532 0.03133

    R 1 0.18017 0.11260 0.06310 0.04072 0.02670

    2/4 0.01151 0.00559 0.00266 0.00144 0.00086

    3/5 0.00115 0.00078 0.00056 0.00036 0.00027

    Notes: Case 1 refers to the regression without any deterministic intercepts or trends, Case 2 refersto the regression with an individual-specific intercept, Case 3 refers to the regression with individual-specific intercepts and trends, Case 4 refers to the regression with breaks in the intercept and Case 5

    refers to the regression with breaks in both intercept and trend. The value Krefers to the number ofregressors excluding any fitted deterministic intercepts or trends.

    109Panel test with multiple structural breaks

    Blackwell Publishing Ltd 2006

  • 8/10/2019 Westerlund 2006

    11/33

    unit root test of Zivot and Andrews (1992), which allows for a single unknown

    break to affect the level and trend of the series. The problem with this test is that

    the break is only permitted under the alternative hypothesis of stationarity. Thus,a rejection of the null does not necessarily imply a rejection of a unit rootper se

    but rather a rejection of a unit root without breaks. This outcome calls for a

    careful interpretation of the test result in applied work. Particularly, in the

    presence of breaks under the null, researchers might incorrectly conclude that a

    rejection of the null indicates evidence of stationarity with a break, when in fact

    the series non-stationary with breaks.

    IV. Response surface regressionsPreliminary results suggest that the simulated asymptotic moments may not be

    borne out fully in realistically small samples and that moments for small

    values ofTmay be required in order to allow for accurate testing. To this end,

    in order to present the results more succinctly, we use response surface

    regressions. We experimented with a variety of specifications and we opted for

    the following linear regression model

    yi d1 d2T1=2

    i d3T1

    i d4T2

    i ei; 6

    where yi is the simulated moment based on a sample of Ti time-seriesobservations and ei is a heteroskedastic error term that reflects the

    simulation uncertainty and the approximation of the true regression form

    by using the quadratic one in equation (6). The intercept is an estimate of

    the asymptotic moment of the test statistic and the remaining para-

    meters determine the shape of the response surface for finite values of T.

    Hence, by including T1=2

    i , T1

    i and T2

    i as explanatory variables, we are

    in effect allowing the small-sample moments of the test statistic to differ

    from the asymptotic ones. The choice of regressors to include was dictated

    by the size of the difference between the estimated intercept and theappropriate asymptotic moment, by the overall fit of the regressions and by

    the significance of the parameters themselves. Adding powers larger than

    two of Ti never seemed to be necessary so the same specification applies to

    all moments.

    The estimation proceeds as follows. For each combination ofKand T, we

    generate 1,000 test statistics under the null hypothesis with all regression

    parameters set equal to one. We use both FMOLS and DOLS estimation. The

    FMOLS is performed using the Bartlett kernel with the bandwidth parameter

    set equal to [T1/3] and the DOLS is performed using [0.5T1/3] lags and

    leads of the first differences of the regressors. We generate moments

    for T2 {20, 50, 60, 70, 100, 200, 500, 1,000} time-series observations and

    110 Bulletin

    Blackwell Publishing Ltd 2006

  • 8/10/2019 Westerlund 2006

    12/33

    up to five regressors. Most samples are relatively small as these seem to

    provide more information about the shape of the response surfaces. However,

    a few large values of T are also included to ensure that the estimates ofthe asymptotic moments are sufficiently accurate. Moments for all three

    deterministic cases are extracted and stored. We then perform 100 replications

    of each experiment, which means that there is a total of 800 observations

    available for each of the response surface regressions.

    As the regression error is heteroskedastic by construction, we follow

    MacKinnon (1996) and apply the generalized method of moments estimator

    proposed by Cragg (1983). To this effect, letZdenote the matrix of right-hand

    side variables and let Y denote the vector of simulated moments for each

    experiment. Also, suppose that e denotes the residual obtained from the OLS

    fit ofYonZ, then Vis the diagonal matrix with the principal diagonal equal to

    that of ee0. The Cragg (1983) estimator may now be written as

    d Z0WW0 V W1W0 Z1Z0WW0 V W1W0 Y;

    where d is the vector of estimated response surface parameters and W is a

    matrix comprised of one dummy variable for each value ofT. The results are

    presented in Table 2 for the FMOLS-based test and in Table 3 for the DOLS-

    based test. These permit a quick computation of the approximate moments for

    a particular value of T. As a measure of regression uncertainty, the tablereports the estimated standard error for each regression. These reflect both the

    simulation uncertainty from estimating the moment rather than knowing it and

    the approximation error of using a quadratic regression form rather than the

    true one. The standard error is always smaller than 0.03, assuring reasonably

    high precision in the estimates and low regression uncertainty. Another

    measure of the quality of the estimated response surfaces is the size of the

    differences between the estimated intercepts and the moments reported in

    Table 1 based on simulating the asymptotic distribution of the test. It was

    observed that the intercepts appear to be remarkably precise estimates of theasymptotic moments.

    V. Unknown break points

    In the previous sections, we have assumed that both the number and the

    locations of the structural breaks are known. This is not necessary. In fact, if

    there is no a priori knowledge about the breaks, one may prefer to treat them

    as endogenous variables that need to be estimated from the data. For this

    purpose, we suggest using the proposal of Bai and Perron (1998, 2003), which

    obtains the location of the breaks by globally minimizing the sum of squared

    residuals as follows

    111Panel test with multiple structural breaks

    Blackwell Publishing Ltd 2006

  • 8/10/2019 Westerlund 2006

    13/33

    TABLE 2

    FMOLS-based response surface moments

    Case K

    Expected value Variance

    Intercept T)1/2 T)1 T)2 SE Intercept T )1/2 T)

    1 1 0.35264 0.05048 )0.61249 12.82076 0.00407 0.18995 )0.42167 0

    1 2 0.26144 0.03296 )0.91888 17.62065 0.00383 0.09720 0.11888 )2

    1 3 0.20695 )0.12335 )0.20455 15.76128 0.00576 0.06441 )0.12325 )0

    1 4 0.16916 )0.26228 0.82432 8.03949 0.00679 0.04111 )0.13058 0

    1 5 0.13943 )0.24630 0.75407 11.69389 0.00890 0.02693 )0.10471 0

    2 1 0.11708 )

    0.05532 0.31233 3.14865 0.00421 0.01130 )

    0.02838 )

    02 2 0.08557 )0.06494 0.24917 6.69821 0.00571 0.00581 )0.02252 0

    2 3 0.06730 )0.11746 0.60610 6.52534 0.00806 0.00309 )0.02008 0

    2 4 0.05391 )0.09801 0.49008 11.09135 0.01095 0.00161 )0.00995 0

    2 5 0.04451 )0.09041 0.53038 13.72616 0.01394 0.00095 )0.00646 0

    3 1 0.05515 0.01680 0.21483 4.98491 0.00794 0.00125 )0.00527 0

    3 2 0.04749 )0.02300 0.39897 6.46768 0.00968 0.00085 )0.00385 0

    3 3 0.04103 )0.03676 0.45948 9.02230 0.01186 0.00057 )0.00232 )0

    3 4 0.03558 )0.04270 0.53224 11.23719 0.01436 0.00036 )0.00087 )0

    3 5 0.03073 )0.02145 0.41501 15.37253 0.01708 0.00025 )0.00023 )0

    Notes: See Table 1 for an explanation of the different cases. The table gives simulated finite and asymptotic mome

    appropriate moment is obtained from the table by calculating the fitted value of the corresponding regression.

    Blackwell

    PublishingLtd2006

  • 8/10/2019 Westerlund 2006

    14/33

    TABLE 3DOLS-based response surface moments

    Case K

    Expected value Variance

    Intercept T)1/2 T)1 T)2 SE Intercept T )1/2 T)

    1 1 0.36132 )0.22961 1.57112 )16.31658 0.00211 0.19043 )0.44026 0.

    1 2 0.27401 )0.31577 1.99793 )15.55645 0.00149 0.11030 )0.24182 0.

    1 3 0.22226 )0.63793 4.13834 )30.95020 0.00373 0.07280 )0.41392 2.

    1 4 0.18901 )0.82065 5.10891 )33.87714 0.00639 0.05053 )0.39326 2.

    1 5 0.14815 )0.40345 1.59114 29.82585 0.00997 0.03162 )0.22870 1.

    2 1 0.11625 0.00344 0.25486 10.42508 0.00626 0.01159 )0.03922 0.

    2 2 0.08775 )0.06769 0.69281 9.15342 0.00706 0.00642 )0.03858 0.

    2 3 0.06583 0.02019 )0.05966 22.41660 0.00804 0.00301 )0.01180 0.

    2 4 0.05238 0.03925 )0.27469 30.42694 0.00951 0.00177 )0.01178 0.

    2 5 0.04255 0.08621 )0.72910 41.76848 0.01105 0.00109 )0.00809 0.

    3 1 0.05347 0.10835 )0.31741 19.11540 0.00821 0.00124 )0.00398 0.

    3 2 0.04542 0.10271 )0.39739 23.86645 0.00878 0.00091 )0.00446 0.

    3 3 0.03855 0.11087 )0.54675 30.28444 0.00987 0.00062 )0.00303 0.

    3 4 0.03248 0.14376 )0.88995 39.98488 0.01128 0.00044 )0.00217 0.

    3 5 0.02710 0.19392 )1.38020 52.16569 0.01289 0.00027 )0.00027 )0.

    Notes: See Table 1 for an explanation of the different cases. The table gives simulated finite and asymptotic mom

    appropriate moment is obtained from the table by calculating the fitted value of the corresponding regression.

    BlackwellPublishingLtd2006

  • 8/10/2019 Westerlund 2006

    15/33

    Tiarg minTi X

    Mi1

    j1 XTij

    tTij1

    1

    yit z0itcij x

    0itbi

    2; 7

    where Ti Ti1; . . . ; TiMi 0

    is the vector of estimated break points, cij andbiare

    the estimates of the cointegration parameters based on the partition Ti (Ti1, . . . , TiMi)

    0 and s is a trimming parameter such thatkij) kij)1 > s, which

    imposes a minimum length for each subsample. Because the minimization is

    taken over all possible partitions of permissable length, the break-point

    estimators are said to be global minimizers. The estimation is performed in

    two separate steps. In the first step, the global minimizers of the sum of

    squared residuals are estimated and stored together with the associated optimal

    break partitions for each possible number of breaksMi 1, . . . , J, whereJissome predetermined upper boundary. In the second step, the number of breaks

    are estimated using an information criterion.

    The purpose of the first step is to estimate the unknown regression parameters

    together with the unknown break points when Tobservations are available for

    each individual. This can be achieved using the dynamic programming

    algorithm developed by Bai and Perron (2003). However, in our case, as the

    slope parameters bi are not subject to shift and are estimated using the full length

    of the time-series dimension, the algorithm cannot be applied directly. This is so

    because the estimate ofbiassociated with the minimum of the sum of squared

    residuals depends on the optimal break partition Ti that we are trying to estimate,which means that we cannot concentrate outbi from the objective function.

    Therefore, the minimization of the sum of squared residuals cannot be performed

    with respect toTidirectly but must be carried out iteratively.

    The iterative procedure suggested by Bai and Perron (2003) proceeds in the

    following fashion. Given a starting value for bi, initiate the procedure by

    minimizing the objective function with respect to cijand Ti while keeping bifixed. This requires an evaluation of the optimal break partition for all

    permissable subsamples that admits the possibility ofMibreaks. Becausebiis

    held fixed, this stage amounts to minimize the objective function of a purestructural change model to which the dynamic programming algorithm apply.

    The next stage is to minimize with respect to cijand bi simultaneously while

    keeping Ti fixed. Then, iterate until the marginal decrease in the objective

    function converges or until the number of iterations reaches some predetermined

    upper boundary. Each iteration assures a decrease in the objective function.

    The first step yields estimated break partitions and sum of squared

    residuals for each number of breaks that lies in the interval [1, J]. The second

    step uses these sum of squared residuals to estimate the number of

    breaks to include in the cointegrated regression. To this end, we follow the

    recommendation of Bai and Perron (2003) and use the Schwarz Bayesian

    information criterion. Also, as the case with no breaks is permissable, the sum

    114 Bulletin

    Blackwell Publishing Ltd 2006

  • 8/10/2019 Westerlund 2006

    16/33

    of squared residuals obtained from the first step are compared with those

    obtained for the corresponding model configuration without any break. The

    estimated break vector ^

    Ti is then obtained as the partition associated with theparticular number of breaks that minimizes the criterion.

    The two steps described above are repeatedNtimes, which produces a vector

    of estimated break points for each individual in the sample. The LM statistic can

    then be constructed using Tiin place ofTifor each individual. As shown by Bai

    and Perron (1998), the above procedure yields consistent estimators of both kijandMifor each individual as Tgrows large. This is very convenient because it

    implies that the sequential limit distribution of the statistic derived in section III

    will be unaltered even if we treat the locations of the breaks as unknown.

    VI. Monte Carlo simulations

    To examine the small-sample properties of the test, we conduct Monte Carlo

    experiments similar to the design of McCoskey and Kao (1998). The DGP can

    be summarized using the following system of equations

    yitz0itcij xitbi rit uit;

    xitxit1 vit;

    ritrit1 /uit:

    The error vector wit (uit, vit)0 is generated as an MA(1) process satisfyingwit eit+ heit)1, whereh is the moving average parameter, eit N(0, V) andVis a positive definite matrix with V11 V22 1.

    1 For the initiation ofxit,ritandeit, we use the value zero. The data were generated for 1,000 panels with

    N cross-sectional and T + 50 time-series observations. The first 50 observa-

    tions for each cross-section is then disregarded in order to reduce the effect of

    the initial condition.

    The parametrization of the DGP is as follows. For simplicity, make the

    assumption that there is a single break such thatki1 k1 for all i and k1

    (0.3, 0.5, 0.7).2

    We also assume that there is a single regressor with slopebi N(1, 1). For the deterministic component, we have five differentconfigurations, each of which correspond to one of our five model specifica-

    tions. For Cases 13, we have M 0 so cij ci1 is constant for all i.Specifically, ci1 0 in Case 1, ci1 N(1, 1) in Case 2 and ci1 N(1, I2) inCase 3. For Case 4, we have cij N(lj, 1), whilecij N(lj, I2) for Case 5. In

    both cases, the break parameters are generated withl11 andd l2) l1,where the parameterd (1, 2, 3) is used to control the size of the break.

    1For some simulation results for the panel LM test with autoregressive errors, see Westerlund

    (2005).2Some simulations results for the case with two structural breaks are available from the authorupon request.

    115Panel test with multiple structural breaks

    Blackwell Publishing Ltd 2006

  • 8/10/2019 Westerlund 2006

    17/33

    The parameter/ determines whether the null hypothesis is true or not. For

    brevity, we make the assumption that this parameter takes on a common value

    for all individuals. Therefore, under the null hypothesis, we have / 0,while/ (0.05, 0.1) under the alternative hypothesis. The parameters h andV introduces nuisance in the DGP. The effect of the MA(1) component is

    governed by h ()0.5, 0, 0.5), while the degree of endogeneity is capturedby the off-diagonal elementsV12and V21ofV. For brevity, we only present the

    results for the case when V12 V21 0.4.Simulations have been carried out using both DOLS and FMOLS estimation.

    However, the results were very similar and we therefore only present the results

    for the FMOLS estimator, which employs a semiparametric correction to

    account for the presence of endogenous regressors. To this end, we use the

    Bartlett kernel with the bandwidth parameter chosen as a fixed function ofTsuch

    that [T1/3]. In estimating the unknown break points, we follow the recommen-

    dation of Bai and Perron (2003) and use the trimming parameters 0.15. Theconvergence criterion in the iterative procedure is set to 0.0001 and the

    maximum number of iterations allowed is 50. The maximum number of breaks

    considered isJ 5. The test results are based on the moments obtained usingthe estimated response surfaces in Table 2. For brevity, we only report the size

    and size-adjusted power of a nominal 5% level test.

    We begin by considering the performance of the test for the case with no

    structural breaks. The results are presented in Table 4. Turning first to theresults on the empirical size we see that the test generally maintains the

    nominal level well when h 0 but it is distorted when h 6 0. For positivevalues of h, the test has a size above the nominal level. Conversely, for

    negative values ofh, the test has a size below the nominal level. The larger the

    absolute value of h, the larger the distortion. Yet, negative values of h in

    general led to more severe distortions than positive. In addition, while

    decreasing in T, we see that the distortions have a tendency of accumulating

    and to become more serious as Nincreases.

    The results on the size-adjusted power suggest that the test performs wellwith rejection frequencies that are close to one in most experiments. The test

    of Case 1 has the highest power and the test of Case 3 has the lowest power.

    Thus, the introduction of deterministic intercept and trend terms that need to

    be estimated affects the test by reducing its power. Moreover, the power

    increases as both N and T grows, which is presumably a reflection of

    consistency of the test. As expected, the rate at which this happens depend

    strongly on the influence of the unit root component in the errors as indicated

    by the value taken by /.

    For Cases 4 and 5 with known breaks, the simulations were carried out

    using d 1 for the size of the shifts. The results are reported in Table 5. Inagreement with the results for the models with no break, we see that the test is

    116 Bulletin

    Blackwell Publishing Ltd 2006

  • 8/10/2019 Westerlund 2006

    18/33

    TABLE 4

    Size and size-adjusted power for Cases 13

    Case /

    N 10, T 100 N 20, T 100 N 10, T 200

    h 0 h )0.5 h 0.5 h 0 h )0.5 h 0.5 h 0 h )0.5 h 0.5

    1 0.00 0.078 0.108 0.002 0.050 0.124 0.000 0.070 0.094 0.000

    1 0.05 0.836 0.870 0.910 0.988 0.982 0.992 0.994 0.992 0.994

    1 0.10 0.980 0.978 0.998 1.000 1.000 1.000 1.000 1.000 1.000

    2 0.00 0.074 0.120 0.002 0.066 0.118 0.000 0.076 0.108 0.002

    2 0.05 0.598 0.602 0.556 0.804 0.786 0.814 0.962 0.978 0.982

    2 0.10 0.952 0.946 0.946 0.998 0.994 1.000 1.000 1.000 1.000

    3 0.00 0.058 0.148 0.000 0.066 0.138 0.000 0.080 0.146 0.000

    3 0.05 0.244 0.218 0.258 0.334 0.388 0.414 0.782 0.814 0.784

    3 0.10 0.716 0.694 0.768 0.920 0.934 0.950 0.996 0.998 0.996

    Notes: The value/ refers to the weight being attached to the unit root component in the regression errors and h refers toTable 1 for an explanation of the different cases.

    BlackwellPublishingLtd2006

  • 8/10/2019 Westerlund 2006

    19/33

    correctly sized when h 0 but there is a problem with size distortions whenh6 0. As expected from the asymptotic theory, we see that the performance

    under the null is unaffected by the location of the structural break. One notableexception is for Case 5 where the results suggest that the test can be severely

    distorted unless the time-series dimension of the individual subsamples before

    and after each break is sufficiently large. In fact, based on the results obtained

    from the simulations, we recommend not using the test in this case unless the

    size of each subsample is T 100 or greater.As for the power of the test, the results indicate that the performance is good

    in general. Notably, the loss of power caused by estimating the parameters of the

    model for each subsample separately rather than over the entire length of the

    time-series dimension as in Cases 2 and 3 need not be large and is effectively

    non-existing in some experiments. These results appear to be rather robust and

    applies regardless of the location of the structural break. As in the no break case,

    we also observe large gains in power as N,Tand/ increases.

    The results on the size and power for the DGP with unknown breaks are

    presented in Table 6. In this case, the data were generated with k1 0.5 fixedand we instead vary the magnitude of the breaks. The results under the null

    hypothesis indicate that the test maintain the nominal size well and that there are

    no large differences in the performance depending on whether the breaks are

    treated as known or not. As for the performance under the alternative

    hypothesis, we find that the power of the test decreases considerably in someexperiments when compared with the case with known breaks, which accords

    with the results reported by Kurozumi (2002) in the context of stationarity

    testing in time-series data. One reason for this might be that the estimated

    number of breaks tend to lie above its true value, which means that the

    alternative model becomes close to the null hypothesis, as is the case when more

    breaks are being estimated, thus causing a loss of power. However, given the

    smallness of the weight attached to the random walk component in the errors,

    the power still appears to be reasonable in most experiments. In addition, it was

    observed that the power increases as the size of the break becomes larger. Aswill be argued momentarily, a larger break is easier to detect, which suggests

    that the power increases as the precision of the estimated breaks improves.

    By treating the locations of the breaks as unknown, the test results can also

    be analysed in terms of the accuracy of the break estimates. To this end, Bai

    and Perron (1998) show that, although consistent as T increases, the rate at

    which the estimated break fractions converge to their true values depend on

    the magnitude of the breaks, which seems reasonable as a smaller break is

    more difficult to discern. Bartley et al. (2001) analyses this issue in small

    samples and found that Ti can be estimated fairly well when the size of the

    break is larger than two in absolute value. To examine the accuracy of the

    estimated breaks in our case, Table 7 presents the correct selection frequencies

    118 Bulletin

    Blackwell Publishing Ltd 2006

  • 8/10/2019 Westerlund 2006

    20/33

    TABLE 5

    Size and size-adjusted power for Cases 4 and 5 with known breaks

    Case k1 /

    N 10, T 100 N 20, T 100 N 10, T 200

    h 0 h )0.5 h 0.5 h 0 h )0.5 h 0.5 h 0 h )0.5 h 0

    4 0.3 0.00 0.022 0.044 0.002 0.018 0.066 0.000 0.020 0.066 0.000

    4 0.3 0.05 0.246 0.246 0.256 0.356 0.316 0.344 0.784 0.766 0.822

    4 0.3 0.10 0.720 0.758 0.710 0.926 0.904 0.934 0.994 0.994 0.998

    4 0.5 0.00 0.016 0.054 0.000 0.018 0.060 0.000 0.030 0.052 0.000

    4 0.5 0.05 0.244 0.258 0.234 0.340 0.354 0.334 0.760 0.796 0.826

    4 0.5 0.10 0.716 0.706 0.706 0.922 0.910 0.896 0.996 0.994 0.998

    4 0.7 0.00 0.012 0.044 0.000 0.020 0.056 0.000 0.010 0.042 0.000

    4 0.7 0.05 0.276 0.262 0.230 0.438 0.442 0.462 0.858 0.818 0.854 4 0.7 0.10 0.768 0.730 0.704 0.930 0.936 0.952 0.996 1.000 1.000

    5 0.3 0.00 0.200 0.230 0.042 0.454 0.534 0.110 0.026 0.074 0.000

    5 0.3 0.05 0.090 0.170 0.098 0.132 0.144 0.110 0.372 0.410 0.480

    5 0.3 0.10 0.256 0.388 0.286 0.510 0.504 0.408 0.946 0.944 0.954

    5 0.5 0.00 0.042 0.124 0.000 0.072 0.186 0.000 0.012 0.074 0.000

    5 0.5 0.05 0.240 0.234 0.210 0.364 0.340 0.338 0.776 0.772 0.798

    5 0.5 0.10 0.718 0.702 0.668 0.920 0.910 0.920 1.000 1.000 0.998

    5 0.7 0.00 0.154 0.222 0.044 0.430 0.510 0.130 0.030 0.088 0.000

    5 0.7 0.05 0.298 0.304 0.234 0.460 0.428 0.456 0.816 0.792 0.846

    5 0.7 0.10 0.752 0.706 0.752 0.940 0.932 0.936 0.996 0.996 1.000

    Notes: The valuek1refers to the location of the structural break, the value / refers to the weight being attached to the uerrors and h refers to the moving average parameter. See Table 1 for an explanation of the different cases.

    BlackwellPublishingLtd2006

  • 8/10/2019 Westerlund 2006

    21/33

    TABLE 6

    Size and size-adjusted power for Cases 4 and 5 with unknown breaks

    Case d /

    N 10, T 100 N 20, T 100 N 10, T 200

    h 0 h )0.5 h 0.5 h 0 h )0.5 h 0.5 h 0 h )0.5 h 0.

    4 1 0.00 0.018 0.000 0.146 0.012 0.006 0.142 0.020 0.000 0.150

    4 1 0.05 0.083 0.073 0.147 0.173 0.103 0.153 0.280 0.187 0.170

    4 1 0.10 0.127 0.157 0.153 0.253 0.163 0.220 0.507 0.600 0.560

    4 2 0.00 0.007 0.003 0.030 0.007 0.007 0.015 0.018 0.010 0.037

    4 2 0.05 0.093 0.087 0.110 0.133 0.073 0.110 0.273 0.267 0.320

    4 2 0.10 0.197 0.173 0.330 0.390 0.207 0.437 0.643 0.700 0.880

    4 3 0.00 0.013 0.003 0.007 0.010 0.007 0.010 0.013 0.007 0.013

    4 3 0.05 0.173 0.117 0.150 0.207 0.127 0.100 0.427 0.297 0.493

    4 3 0.10 0.310 0.233 0.427 0.533 0.383 0.550 0.747 0.780 0.943

    5 1 0.00 0.040 0.070 0.000 0.070 0.140 0.000 0.017 0.020 0.000

    5 1 0.05 0.120 0.083 0.093 0.097 0.160 0.110 0.300 0.207 0.347

    5 1 0.10 0.267 0.140 0.210 0.387 0.327 0.380 0.750 0.523 0.943

    5 2 0.00 0.037 0.047 0.000 0.080 0.133 0.000 0.020 0.007 0.000

    5 2 0.05 0.120 0.120 0.087 0.103 0.163 0.107 0.263 0.210 0.303

    5 2 0.10 0.260 0.207 0.250 0.310 0.263 0.353 0.763 0.533 0.900

    5 3 0.00 0.057 0.080 0.000 0.070 0.153 0.000 0.007 0.017 0.000

    5 3 0.05 0.087 0.083 0.107 0.107 0.100 0.100 0.293 0.200 0.383

    5 3 0.10 0.167 0.153 0.307 0.353 0.297 0.390 0.713 0.660 0.960

    Notes: See Table 1 for an explanation of the different cases and Table 5 for an explanation of the various parameters. structural break.

    Blackwell

    PublishingLtd2006

  • 8/10/2019 Westerlund 2006

    22/33

    for both the number and the locations of the estimated break fractions obtained

    under the null hypothesis with k1 0.5 held fixed. The results indicate that

    the breaks can be estimated with high accuracy in most cases and that theprecision of the estimates increase as both T and d increases, which is

    consistent with the asymptotic results presented by Bai and Perron (1998).

    Overall, the simulations leads us to the conclusion that the test performs

    well in general with reasonable power and small size distortions in most

    experiments. The results also suggest that the performance of the test in Cases

    4 and 5 is unaffected by the locations of the breaks but the performance of the

    test with shifts in both intercept and trend can be poor unless the time series of

    each subsample is sufficiently long. We have also examined the effects of a

    misspecified model where the true DGP includes a structural break that is not

    considered by the researcher. In this case, the rejection frequencies are

    effectively one in all experiments and the results are therefore not presented.

    Thus, serious size distortions are likely to arise when an existing structural

    breaks are omitted.

    VII. An application to the current account

    In this section, we make an attempt to demonstrate empirically the usefulness

    of our test by revisiting the data set used by Taylor (2002) and Ho (2002) to

    assess the solvency of the current account. To this end, let Iitbe the ratio ofinvestment to gross domestic product (GDP) at current local prices and letSit

    be the corresponding saving rate. The starting point of our investigation is the

    following regression

    Iitai biSit eit: 8

    Recent theoretical and empirical work suggests that saving and investment as a

    fraction of GDP are non-stationary variables (e.g. see Coakley, Kulasi and

    Smith, 1996; Levy, 2000). The implication being that the current account must

    be stationary as debt cannot explode. Because the current account is identicallythe difference between saving and investment, this suggests that saving and

    investment as a fraction of GDP should be cointegrated. Yet, for a theory so

    straightforward and widely accepted, the solvency of the current account has

    proven extremely difficult to establish empirically (e.g. see Leachman, 1991;

    de Haan and Siermann 1994; Lemmen and Eijffinger, 1995; Ho, 2002).

    Although there are many explanations for these empirical findings, this

    section focuses on two explanations that has attracted much attention recently.

    The first explanation is that conventional time-series cointegration tests may

    have low power against persistent alternatives because of the short-sample

    periods usually employed (e.g. see de Haan and Siermann, 1994; Ho, 2002).

    The second explanation is that the empirical relationship between saving and

    121Panel test with multiple structural breaks

    Blackwell Publishing Ltd 2006

  • 8/10/2019 Westerlund 2006

    23/33

    TABLE 7

    Correct selection frequencies for Cases 4 and 5 with unknown breaks

    Case d Frequency

    N 10, T 100 N 20, T 100 N 10, T 200

    h 0 h )0.5 h 0.5 h 0 h )0.5 h 0.5 h 0 h )0.5 h 0

    4 1 Number 0.647 0.567 0.613 0.641 0.578 0.609 0.725 0.658 0.686

    4 1 Location 0.292 0.177 0.327 0.281 0.171 0.324 0.305 0.197 0.363

    4 2 Number 0.786 0.669 0.810 0.791 0.663 0.798 0.856 0.734 0.845

    4 2 Location 0.549 0.380 0.579 0.541 0.382 0.580 0.570 0.424 0.605

    4 3 Number 0.846 0.733 0.874 0.856 0.706 0.868 0.901 0.757 0.8934 3 Location 0.669 0.522 0.716 0.672 0.511 0.710 0.709 0.534 0.726

    5 1 Number 0.954 0.762 0.987 0.960 0.741 0.987 0.979 0.828 0.993

    5 1 Location 0.944 0.749 0.979 0.950 0.729 0.979 0.974 0.823 0.989

    5 2 Number 0.956 0.730 0.996 0.961 0.747 0.996 0.987 0.822 0.999

    5 2 Location 0.949 0.724 0.990 0.956 0.740 0.990 0.981 0.818 0.994

    5 3 Number 0.961 0.747 0.997 0.963 0.745 0.997 0.987 0.818 0.999

    5 3 Location 0.957 0.742 0.994 0.959 0.741 0.993 0.981 0.817 0.998

    Notes: The number frequency specifies the frequency count of correctly chosen number of breaks. The location frequencorrectly chosen break locations when the estimated number of breaks is equal to its true value. See Table 1 for an explanatifor an explanation of the various parameters. The value drefers to the size of the structural break.

    Blackwell

    PublishingLtd2006

  • 8/10/2019 Westerlund 2006

    24/33

    investment that is being estimated need not be invariant to policy regime

    changes, which may lead to structural shifts in both saving and investment and

    hence in the relationship between them (e.g. see Alexakis and Apergis, 1994;Sarno and Taylor, 1998; Corbin, 2004; Ozmen and Parmaksiz, 2004).

    However, while very reasonable and potentially appealing, these explan-

    ations have been found to be empirically inadequate and far from convincing,

    and this section therefore offers an alternative route. The idea is that, to

    provide any robust evidence on the cointegration between saving and

    investment, one needs to consider not the low power and the presence of

    structural change separately but simultaneously. The intuition is simple. First,

    while increasing the length of the sample may be justified from a power point

    of view, it also increases the probability of a break occurring somewhere in the

    sample. On the other hand, modifying existing tests so as to accommodate for

    structural change is typically very costly in terms of power. Therefore, it is the

    joint consideration of both these aspects that is likely to be key when testing

    the solvency of the current account. One way to accomplish this is to use the

    panel LM test developed in the previous sections.

    The data that we use consists of cross-sectional observations on saving and

    investment as fractions of GDP sampled at annual frequency between 1883

    and 1992 for 15 Organization for Economic Co-operation and Development

    (OECD) countries plus Argentina.3 Before applying our test, we examine

    whether the variables are stationary or not and whether there exist any structuralbreaks or not. To test if the variables are non-stationary, we employ the panel data

    unit root test recently proposed by Im et al. (2005), which allows for the

    possibility of a single heterogeneous shift in the level of the series.4 The

    calculated values of the statistic for the saving and investment variables are

    )0.560 and)1.330 respectively. Thus, as the critical region of the test is given by

    the left tail of the normal distribution, the null hypothesis of a unit root test cannot

    be rejected. This conclusion is supported by the individual Schmidt and Phillips

    (1992) and Amsler and Lee (1995) unit root test statistics presented in Table 8.

    The first does not allow for any break, while the second allows for a single breakto affect the level of each series. The results suggest that we cannot reject the null

    of a unit root at the 10% level of significance for any of the countries.

    To test the presence of structural change in the estimated regressions, we

    employ for each individual the B(s) statistic of Hao and Inder (1996), and

    the MeanF and SupF statistics of Hansen (1992). The results presented in

    Table 8 suggest that at least six of the individual cointegration regressions

    3For a detailed description of the data, see Taylor (2002).4In calculating the value of the test statistic, we use a lag augmentation of three to account for the

    effects of serial correlation. The location of the break is determined endogenously via grid search atthe minimum of the statistics. To this end, we use a trimming parameter of 0.15 to eliminate theendpoints.

    123Panel test with multiple structural breaks

    Blackwell Publishing Ltd 2006

  • 8/10/2019 Westerlund 2006

    25/33

    TABLE 8

    Country specific tests

    Country

    Unit root saving Unit root investment Cointegration tests Stability tests

    Z(s) ~s Z(s) ~s Zt Zt B(s) MeanF Sup

    Argentina )1.942 )1.935 )3.349 )3.320 )4.657 )3.900 0.748 10.349 13

    Australia )2.806 )2.854 )2.511 )2.068 )6.877 )6.397 0.667 1.741 4

    Canada )2.374 )2.408 )2.144 )2.093 )3.991 )3.041 1.363 3.434 14

    Finland )

    2.266 )

    2.959 )

    2.117 )

    2.166 )

    5.576 )

    4.580 1.033 7.939 14Spain )1.957 )2.353 )2.143 )2.682 )5.386 )4.838 1.091 2.298 11

    Sweden )1.671 )1.855 )2.487 )2.588 )5.462 )4.393 0.920 7.622 15

    UK )2.178 )2.238 )2.446 )2.942 )4.457 )2.261 1.990 12.636 17

    USA )2.157 )2.895 )2.462 )3.273 )3.971 )3.364 0.798 3.266 6

    Notes: In constructing the test statistics, we use three lags, a bandwidth of [ T1/3] and a trimming parameter of 0.15. The the estimated break points used in computing the panel LM statistic. The maximum number of breaks allowed is five. Tfollows: )3.72 for the ~s statistic of Amsler and Lee (1995) and the Z(s) statistic of Schmidt and Phillips (1992); )3.065Ouliaris (1990); )4.34 for theZt statistic of Gregory and Hansen (1996); 1.0477 for theB(s) statistic of Hao and Inder (199SupF statistics of Hansen (1992) respectively.

    Blackwell

    PublishingLtd2006

  • 8/10/2019 Westerlund 2006

    26/33

    cannot be regarded as stable and there is a need to allow for structural change. To

    this effect, we now employ the panel LM statistic. The calculated value for Case

    2 with only an individual-specific constant term is 9.025 for the FMOLS-basedtest and 8.173 for the DOLS-based test. Thus, if we ignore the possibility of

    structural change, then we find no evidence of cointegration. By contrast, if

    we allow for a level shift in each regression, then the calculated values of

    the statistics are 2.138 and 1.414 for the FMOLS- and DOLS-based test

    respectively. Thus, in this case we cannot reject the null hypothesis on the 1%

    level, which suggests that the variables are cointegrated around a broken

    intercept. To reinforce this assertion, Table 8 presents some confirmatory

    evidence for each country. The Ztstatistic of Phillips and Ouliaris (1990) do not

    allow for any break, while theZ

    t

    statistic of Gregory and Hansen (1996) is able

    to accommodate for a single break in the intercept of each regression. As seen

    from the table, the null of no cointegration is rejected on six occasions at the 10%

    level for both statistics.

    Table 8 also reports the estimated break points obtained by using the

    procedure described in section V. At least two breaks are found for each

    country with all breaks occurring during the period 190675. From an

    historical point of view, this seems very reasonable. First, there are essentially

    no breaks prior to the advent of the First World War, which agrees with

    the stability of the classical gold standard regime. Secondly, there is a

    preponderance of breaks occurring between 1913 and 1949. This accordsapproximately with the interwar period and seems consistent with the findings

    of Levy (2000), Hoffmann (2004) and Corbin (2004). The break in 1913 for

    Argentina is also expected given the Bearing Crisis. Thirdly, there are several

    breaks occurring between the years 1957 and 1975. This coincides with the oil

    price shocks of that period, the breakdown of the Bretton Woods system and

    the establishment of the Exchange Rate Mechanism.

    An important caveat worth noting is that so far all forms of cross-sectional

    dependency has been ignored. If this assumption is violated, then the panel

    LM test depends on various nuisance parameters associated with the cross-sectional correlation properties of the data, which means that the test no longer

    has a limiting normal distribution. A limited form of cross-sectional

    dependence can be permitted by using data that has been demeaned with

    respect to a common time effects, which does not affect the distribution of the

    test. Therefore, to accommodate some form of cross-sectional dependency, we

    applied the test to the cross-sectionally demeaned data. Using this approach

    the calculated values of the FMOLS- and DOLS-based test statistics are 1.783

    and 1.632 respectively. Thus, in agreement with the earlier results, we cannot

    reject the null on the 1% level of significance.

    However, in general, with dynamic feedback effects that runs from one

    cross-section to another, and which are not common across the members of the

    125Panel test with multiple structural breaks

    Blackwell Publishing Ltd 2006

  • 8/10/2019 Westerlund 2006

    27/33

    panel, then a common time effects will not account for all dependencies. One

    possible solution to this problem is to employ the bootstrap approach, which

    makes inference possible even under very general forms of cross-sectionaldependence. The bootstrap opted for this section proceeds as follows.

    1 Obtain Ti for each i using the procedure outlined in section V.

    2 Estimate cij and bi for each i by FMOLS or DOLS using Tobserva-

    tions and dummy variables to account for the breaks. Obtain

    et e1t; . . . ; e

    Nt

    0.

    3 Compute the centred residual

    ~et et T

    1 XT

    t1

    et:

    4 Let wt ~e0t; v

    0t

    0, where vt v

    01t; . . . ; v

    0Nt

    0. Generate the bootstrap

    innovations wt ~e0t ; v

    0t

    0by resampling the vectorwt with replace-

    ment.5

    5 Generate the bootstrap sample yit and xit with x

    i0 0 recursively as

    xitxit1 v

    it;

    yitz0itcij x

    0itbi e

    it:

    6 Compute the standardized panel LM test statistic for each bootstrap

    sample while treating Ti as known.

    We apply the bootstrap approach to the saving and investment data using

    5,000 bootstrap replications. The 1% critical value obtained from the

    bootstrap distribution is 2.218 for the FMOLS-based test and 1.647 for the

    DOLS-based test. Hence, the null is not rejected on the 1% level.

    Nevertheless, it is tempting, particularly given the results from the more

    straightforward setting of the previous sections, to conclude that the effects of

    cross-sectional correlation are small and that the critical value from the normal

    distribution in fact apply. To check whether this is indeed the case, we

    performed a small set of Monte Carlo simulations based on the DGP of

    section VI but now with the error uit having the factor specification uitpft+ eit, where (eit, vit, ft)

    0 N(0, I3). With this modification, the correlationbetweenuitand uktfor i 6 kis given by p

    2/(1 + p2). Thus, a non-zero value

    on the parameterp induces correlation between the members of the panel.

    Table 9 presents some results on the size of a nominal 5% level test for this

    DGP using both the bootstrapped critical value and the critical value obtained

    from the normal distribution. It is observed that the bootstrap test performs

    5By resampling wtrather than ~et, we can preserve not only the cross-sectional correlation structureofeitbut also any endogenous effects that may run across the individual regressions of the system.

    126 Bulletin

    Blackwell Publishing Ltd 2006

  • 8/10/2019 Westerlund 2006

    28/33

    well and the test based on the normal distribution appears to be remark-

    able robust to moderate degrees of cross-sectional correlation. To get an

    appreciation of the size of the problem in the saving and investment data, weestimate the correlation matrix of the equilibrium errors. The largest

    correlation is 0.402, suggesting that the simulation results for the DGP with

    p 1 should be relevant in which case Table 9 indicates that the effects ofcross-sectional dependence on the test should be small and that inference

    based on the normal distribution should not be misleading.

    The bootstrap approach can accommodate for all types of cross-sectional

    dependencies that are absorbed in the long-run covariance matrix of the

    equilibrium errors and the first differences of the regressors. Another source of

    cross-sectional dependence is cointegration that runs between the membersof the panel. Banerjee, Marcellino and Osbat (2004) uses Monte Carlo

    simulations to study this issue in small samples and found that the

    consequences of cross-sectional cointegration on existing panel cointegration

    tests can be quite severe. The authors suggest testing for the presence of cross-

    sectional cointegration by using the procedure developed by Gonzalo and

    Granger (1995), which involves first extracting the common trends from each

    cross-section and then testing for cointegration among these trends. When we

    employ this approach to the saving and investment data, we end up marginally

    rejecting the null of no cointegration on the 1% level on one occasion. Hence,

    there appears to be no severe violation of the assumption of no cross-sectional

    cointegration.

    TABLE 9

    Size for Case 4 with cross-sectional dependence

    p N T

    FMOLS DOLS

    ZB(M) Z(M) Z B(M) Z(M)

    1 10 100 0.024 0.028 0.038 0.030

    2 10 100 0.036 0.074 0.040 0.040

    6 10 100 0.024 0.076 0.046 0.068

    1 20 100 0.038 0.062 0.044 0.046

    2 20 100 0.026 0.120 0.046 0.082

    6 20 100 0.028 0.146 0.036 0.090

    1 10 200 0.024 0.030 0.032 0.030

    2 10 200 0.034 0.072 0.032 0.050

    6 10 200 0.026 0.080 0.036 0.0801 20 200 0.038 0.064 0.036 0.048

    2 20 200 0.032 0.108 0.040 0.118

    6 20 200 0.024 0.116 0.024 0.122

    Notes: FMOLS, fully modified ordinary least square; DOLS, dynamic ordinary least square.The valuep refers to the loading parameter in the common time effects specification. We use Z(M)

    and ZB(M) to refer to the test based on the normal and bootstrapped 5% critical value.

    127Panel test with multiple structural breaks

    Blackwell Publishing Ltd 2006

  • 8/10/2019 Westerlund 2006

    29/33

    VIII. Conclusions

    In this paper, we extend the panel LM cointegration test proposed by

    McCoskey and Kao (1998) to allow for multiple structural breaks in both level

    and trend of the cointegration regression. Test statistics are derived when both

    the number and the locations of the breaks are known and when they are

    determined endogenously from the data. Test statistics are also derived when

    the cointegration regression is not subject to structural change but the

    deterministic component includes individual-specific intercepts and trends.

    Using sequential limit theory, we are able to show that the test has a limiting

    normal distribution that is free of nuisance parameters under the null

    hypothesis. In particular, we show that the limiting distribution is invariant

    with respect to both the number and the locations of the breaks and that thereis no need to compute different critical values for all possible patterns of break

    points, which makes the test computationally very convenient. Results form

    Monte Carlo experiments suggest that the test has small size distortions and

    reasonable power. To demonstrate the practical usefulness of the new test, we

    reexamine the data set employed by Ho (2002) and Taylor (2002) to assess the

    solvency of the current account. Contrary to most of the findings presented in

    the earlier literature, we find evidence suggesting that saving and investment

    are cointegrated once a level break in the cointegration regression is

    accommodated.

    Final Manuscript Received: May 2005

    References

    Alexakis, P. and Apergis, N. (1994). The Feldstein-Horioka puzzle and exchange rate regimes:

    evidence from cointegration tests, Journal of Policy Modeling, Vol. 16, pp. 459472.

    Amsler, C. and Lee, J. (1995). An LM test for a unit root in the presence of a structural break,

    Econometric Theory, Vol. 11, pp. 359368.

    Bai, J. and Perron, P. (1998). Estimating and testing linear models with multiple structuralchanges,Econometrica, Vol. 66, pp. 4778.

    Bai, J. and Perron, P. (2003). Computation and analysis of multiple structural change models,

    Journal of Applied Econometrics, Vol. 18, pp. 122.

    Banerjee, A., Marcellino, M. and Osbat, C. (2004). Some cautions on the use of panel

    methods for integrated series of macroeconomic data, Econometrics Journal, Vol. 7,

    pp. 322340.

    Bartley, W. A., Lee, J. and Strazicich, M. C. (2001). Testing the null of cointegration in the

    presence of a structural break, Economics Letters, Vol. 73, pp. 315323.

    Coakley, J., Kulasi, F. and Smith, R. (1996). Current account solvency and the Feldstein-

    Horioka puzzle, The Economic Journal, Vol. 106, pp. 620627.

    Corbin, A. (2004). Capital mobility and adjustment of the current account imbalances: abounds testing approach to cointegration in 12 countries, International Journal of Finance

    and Economics, Vol. 9, pp. 257276.

    128 Bulletin

    Blackwell Publishing Ltd 2006

  • 8/10/2019 Westerlund 2006

    30/33

    Cragg, J. G. (1983). More efficient estimation in the presence of heteroskedasticity of

    unknown form, Econometrica, Vol. 51, pp. 751763.

    Gonzalo, J. and Granger, C. (1995). Estimation of common long-memory components in

    cointegrated systems, Journal of Business and Economic Statistics, Vol. 13, pp. 2735.

    Gregory, A. M. and Hansen, B. E. (1996). Residual-based tests for cointegration in models

    with regime shifts, Journal of Econometrics, Vol. 70, pp. 99126.

    de Haan, J. and Siermann, C. L. J. (1994). Saving, investment and capital mobility: a comment

    on Leachman, Open Economies Review, Vol. 5, pp. 517.

    Hansen, B. E. (1992). Tests for parameter instability in regressions with I(1) processes,

    Journal of Business and Economic Statistics, Vol. 10, pp. 4559.

    Hao, K. (1996). Testing for structural change in cointegrated regression models: some com-

    parisons and generalizations, Econometric Reviews, Vol. 15, pp. 401429.

    Hao, K. and Inder, B. (1996). Diagnostic test for structural change in cointegrated regression

    models, Economics Letters, Vol. 50, pp. 179187.

    Harris, D. and Inder, B. (1994). A Test of the Null Hypothesis of Cointegration, NonstationaryTime Series Analysis and Cointegration, in Hargreaves C. P. (ed.),Nonstationary Time Series

    Analysis and Cointegration, Oxford University Press, New York, NY, pp. 133152.

    Ho, T. (2002). A Panel cointegration approach to the saving-investment correlation, Empirical

    Economics, Vol. 27, pp. 91100.

    Hoffmann, M. (2004). International capital mobility in the long run and the short run: can we

    still learn from saving-investment data?, Journal of International Money and Finance,

    Vol. 23, pp. 113131.

    Im, K. S., Lee, J. and Tieslau, M. (2005). Panel LM unit root tests with level shifts, Oxford

    Bulletin of Economics and Statistics, Vol. 67, pp. 393419.

    Kurozumi, E. (2002). Testing for stationarity with a break,Journal of Econometrics, Vol. 108,

    pp. 6399.Leachman, L. (1991). Saving, investment and capital mobility among OECD countries, Open

    Economies Review, Vol. 2, pp. 137163.

    Lemmen, J. J. G. and Eijffinger, S. C. W. (1995). The quantity approach to financial integ-

    ration: the Feldstein-Horioka criterion revisited, Open Economies Review, Vol. 6, pp. 145

    165.

    Levy, D. (2000). Investment-saving comovement and capital mobility: evidence from century

    long U.S. time series, Review of Economic Dynamics, Vol. 3, pp. 100136.

    MacKinnon, J. G. (1996). Numerical distribution functions for unit root and cointegration

    tests, Journal of Applied Econometrics, Vol. 11, pp. 601618.

    McCoskey, S. and Kao, C. (1998). A residual-based test of the null of cointegration in panel

    data,Econometric Reviews, Vol. 17, pp. 5784.McCoskey, S. and Kao, C. (2001). A Monte Carlo comparison of tests for cointegration in

    panel data, Journal of Propagation in Probability and Statistics, Vol. 1, pp. 165198.

    Park, J. P. and Phillips, P. C. B. (1988). Statistical inference in regressions with integrated

    regressors: part I, Econometric Theory, Vol. 4, pp. 468497.

    Phillips, P. C. B. and Hansen, B. E. (1990). Statistical inference in instrumental variables

    regression with I(1) process, Review of Economics Studies, Vol. 57, pp. 99125.

    Phillips, P. C. B. and Moon, H. R. (1999). Linear regression limit theory of nonstationary panel

    data,Econometrica, Vol. 67, pp. 10571111.

    Phillips, P. C. B. and Ouliaris, S. (1990). Asymptotic properties of residual based tests for

    cointegration,Econometrica, Vol. 58, pp. 165193.

    Saikkonen, P. (1991). Asymptotic efficient estimation of cointegration regressions, Econo-metric Theory, Vol. 7, pp. 121.

    129Panel test with multiple structural breaks

    Blackwell Publishing Ltd 2006

  • 8/10/2019 Westerlund 2006

    31/33

    Sarno, L. and Taylor, M. P. (1998). Exchange controls, international capital flows and saving-

    investment correlations in the UK: an empirical investigation, Weltwirtschaftliches Archiv,

    Vol. 134, pp. 6997.

    Schmidt, P. and Phillips, P. C. B. (1992). LM tests for a unit root in the presence of deter-

    ministic trends, Oxford Bulletin of Economics and Statistics, Vol. 54, pp. 257287.

    Shin, Y. (1994). A residual-based test of the null of cointegration against the alternative of no

    cointegration,Econometric Theory, Vol. 10, pp. 91115.

    Taylor, A. M. (2002). A century of current account dynamics, Journal of International Money

    and Finance, Vol. 21, pp. 725748.

    Ozmen, E. and Parmaksiz, K. (2004). Policy regime change and the Feldstein-Horioka puzzle:

    the UK evidence, Journal of Policy Modeling, Vol. 25, pp. 137149.

    Westerlund, J. (2005). A panel CUSUM test of the null of cointegration, Oxford Bulletin of

    Economics and Statistics, Vol. 62, pp. 231262.

    Zivot, E. and Andrews, D. W. K. (1992). Further evidence on the great crash, the oil-price

    shock and the unit root hypothesis, Journal of Business and Economic Statistics, Vol. 10,pp. 251270.

    Appendix

    Mathematical proofs

    In this appendix, we prove the asymptotic distribution for the DOLS estimator.

    The proof uses the methods of Shin (1994) so only essential details will be

    given.

    Proof of Theorem 1. Under the null hypothesis, the DGP for the most general

    case with endogenous regressors and breaks in both the level and trend may be

    written as the following truncated regression

    yitX0

    itdijXH

    kH

    v0itk/ik uit; A1

    where Xit z0it;x

    0it

    0, dij (cij, bi)

    0 and uit Pjkj>Hv0itk/ik u

    it is a

    stationary error term comprised of the projection of uit

    onto all the lags and

    leads above the truncation point H and an orthogonal term uit. The DOLS

    estimator is obtained by performing OLS on (A1). Consider the DOLS

    estimator of segmentjusing the subsamplet Tij)1, . . . , Tijand letdi denotethe OLS estimate ofdijbased on this subsample. The fitted regression may be

    written in the following fashion

    yitX0

    itdiXH

    kH

    v0itk/ik uit: A2

    Assumption 2 (i) ensures that each subsample approaches to infinity when

    T! 1. Hence, there will be no loss of generality by analysing each subsampleas a sample of length T that is not subject to structural change. To this end,

    130 Bulletin

    Blackwell Publishing Ltd 2006

  • 8/10/2019 Westerlund 2006

    32/33

    consider the partial sum process SiT PTr

    t1 uitobtained from equation (A2).

    If we let D1 diag(T)1/2, T)3/2, T)1IK) and D2 diag(T

    )1, T)2, T)3/2IK),

    then this sum may be written as

    T1=2SiT T1=2

    XTrt1

    yit T1=2

    XTrt1

    X0itdi T1=2

    XTrt1

    XHkH

    v0itk/ik

    T1=2XTrt1

    uit T1=2

    XTrt1

    X0itdi di

    T1=2 XTr

    t1X

    H

    kH

    v0itk/ik /ik

    T1=2XTrt1

    uit T1=2

    XTrt1

    Xjkj>H

    v0itk/ij

    D12XTrt1

    X0itD11 di di T

    1=2XTrt1

    XHkH

    v0itk/ik /ik:

    Under the regulatory conditions of Saikkonen (1991), the second and

    fourth terms are op(1) and can be disregarded. Moreover, by the multi-variate invariance principle, we have D12

    PTrt1Xit)

    Rr0B0i2 and

    T1=2PTr

    t1uit)Bi1:2. Also, by Theorem 3.3 of Park and Phillips (1988),

    as T! 1

    D11 di di )

    Z 10

    Bi2B0i2

    1Z 10

    Bi2dBi1:2:

    Using these weak convergence results, and as Wi1:2 x1i1:2Bi1:2 and Wi2

    X1=2

    i22 B

    i2, the limit of T)1/2S

    iTas T! 1 may now be readily obtained as

    follows

    T1=2SiT )Bi1:2

    Z r0

    B0i2

    Z 10

    Bi2B0i2

    1Z 10

    Bi2dBi1:2

    xi1:2Wi1:2 xi1:2

    Z r0

    W0i2

    Z 10

    Wi2 W0i2

    1Z 10

    Wi2dWi1:2

    xi1:2Vi: A3

    Given that the restriction placed on the bandwidth expansion rate is satisfied,

    then x2i1:2 is consistent forx2i1:2 as T! 1. Hence, we obtain

    131Panel test with multiple structural breaks

    Blackwell Publishing Ltd 2006

  • 8/10/2019 Westerlund 2006

    33/33

    T2x2i1:2 XT

    t1

    S2iT )Fi Z 1

    0

    V2i : A4

    Consequently, if we define Rij PTij

    tTij11Tij Tij1

    2x2i1:2S

    2it, then

    it follows that Rij Fij, where Fi1, . . . , FiMi+1 are Mi + 1 independent

    generalized Cramer-von Mises distributions. Using this result, we can prove

    the intermediate limit ofZ(M) as follows

    ZM XNi1

    XMi1j1

    XTijtTij11

    Tij Tij12x2i1:2S

    2it

    XNi1

    XMi1j1 R

    ij

    )XNi1

    XMi1j1

    Fij: A5

    Thus, as the limiting distribution passing T! 1 is i.i.d. over the cross-section,it follows thatE(Fij) Q andEF

    2ij R for alliandj. Consequently, if we let

    FiPMi1

    j1 Fij, thenE(Fi) (Mi + 1)Q andEF2

    i Mi 12R. Let H and

    R be the cross-sectional averages of the expected value and the variance ofFi

    respectively. To derive the sequential limiting distribution ofZ(M), first defineRi PMi1

    j1 Rij and note thatN1PN

    i1Ri converges to H in probability as

    T! 1 and then N! 1 sequentially by the law of large numbers. Next,expand the statistic in the following manner

    N1=2ZM N1=2H N1=2 N1XNi1

    Ri H

    !: A6

    Assume that the Lindeberg condition is satisfied, then N1=2ZM N1=2 H )N0; R as T! 1 prior to N by direct application of the

    Lindeberg-Feller central limit theorem. This establishes the limit distributionof the panel LM statistic as required for the proof. n

    132 Bulletin