Minimum LM Unit Root Test with Two Structural Breaks · Minimum LM Unit Root Test with Two...
Transcript of Minimum LM Unit Root Test with Two Structural Breaks · Minimum LM Unit Root Test with Two...
Minimum LM Unit Root Test with Two Structural Breaks
Junsoo LeeMark Strazicich
Department of EconomicsUniversity of Central Florida
July 1999
Abstract
The two-break unit root test of Lumsdaine and Papell (1997) is examined and found tosuffer from bias and spurious rejections in the presence of structural breaks under thenull. A two-break minimum LM unit root test is proposed as a remedy. The two-breakLM test does not suffer from bias and spurious rejections and is mostly invariant to thesize, location, and misspecification of the breaks. We test the Nelson and Plosser (1982)data and find fewer rejections of the unit root than Lumsdaine and Papell.
JEL classification: C12, C15, and C22
Key words: Lagrange Multiplier, Unit Root Test, Structural Break, and EndogenousBreak
Corresponding author: Junsoo Lee, Associate Professor, Department of Economics,University of Central Florida, Orlando, FL, 32816-1400, USA. Telephone: 407-823-2070. Fax: 407-823-3269. E-mail: [email protected].
We thank John List for helpful comments.
1
1. Introduction
Since the influential paper of Perron (1989), researchers have noted the
importance of allowing for a structural break when testing for a unit root. Perron showed
that the ability to reject a unit root decreases when the stationary alternative is true and a
structural break is ignored. Perron used a modified Dickey-Fuller (hereafter DF) unit
root test including dummy variables to account for one known, or “exogenous,” structural
break. Subsequent papers further modify the test to allow for one unknown break point
that is determined “endogenously” from the data. One widely used endogenous
procedure is the “minimum test” of Zivot and Andrews (1992, hereafter ZA), which
selects the break point where the t-statistic testing the null of a unit root is at a minimum.
Given a loss of power when ignoring one structural break in standard unit root
tests, it is logical to expect a similar loss of power when ignoring two, or more, breaks in
the one-break tests. Recent research indicates that many economic time series might
contain more than one structural break.1 Therefore, it may be necessary to allow for
more than one break when testing for a unit root. Recently, Lumsdaine and Papell (1997,
hereafter LP) make a contribution in this direction by extending the ZA test to two
structural breaks.
One critical issue common to these minimum unit root tests is that they typically
assume no breaks under the null, and derive their critical values under this assumption.
Despite their popularity, these tests are invalid if structural breaks occur under the null; as
rejection of the null would not necessarily imply rejection of a unit root per se, but would
1 For example, Ben-David, Lumsdaine, and Papell (1999), Ben-David and Papell (1998), and Papell,Murray, and Ghiblawi (1999) find evidence of more than one structural break in real GDP, per capita realGDP, and unemployment rates among OECD countries.
2
instead imply rejection of a unit root without break. Perron initially allows for a break
under both the null and alternative hypotheses. This is important for his exogenous break
test; otherwise the test statistic will diverge under the null as the size of the break
increases. The same result is true for the endogenous minimum tests. Nunes et al. (1997)
and Lee et al. (1998) provide evidence that assuming no structural break under the null in
the ZA test makes the associated test statistic diverge and leads to spurious rejections
when the data generating process (DGP) contains a break. Lee and Strazicich (1999a)
further investigate the source of these spurious rejections and find that the ZA test most
often selects the break point where bias is maximized. In this paper, we find the same
problems of bias and spurious rejections for the two-break LP unit root test.
To provide a remedy, we propose a “two-break minimum LM test.” The test is
based on the Lagrange Multiplier (LM) unit root test suggested by Schmidt and Phillips
(1992, hereafter SP), and can also be seen as an extension of the one-break minimum LM
test developed in Lee and Strazicich (1999b). The two-break LM test solves the
problems entailed in the LP test: the LM test does not diverge as breaks under the null
increase in size, and is free of bias and spurious rejections. Further, there is no need to
exclude breaks under the null. Whereas, for the two-break LP test it might be necessary
to exclude breaks under the null to make the test statistic invariant to nuisance
parameters, a similar assumption is not required for the two-break LM test. Even with
breaks under the null, the distribution of the two-break LM test statistic is unaffected,
since the test is invariant to break point nuisance parameters. The two-break LM test is
3
also robust to misspecification of the number of breaks under the null; thus providing a
solution to the suggestion raised in LP (p. 212).
The paper proceeds as follows. Section 2 presents the two-break minimum LM
unit root test and compares it to the LP test. Asymptotic properties of the two-break
minimum LM test are discussed. Section 3 compares properties of each test in simulation
experiments. Section 4 examines the Nelson and Plosser (1982) data using the two-break
minimum LM test. Section 5 summarizes and concludes. Throughout the paper, the
symbol “→” denotes weak convergence of the associated probability measure.
2. Test Statistics and Breaks under the Null
Perron previously considered three structural break models as follows. The “crash”
Model A allows for a one-time change in level; the “changing growth” Model B
considers a sudden change in slope of the trend function; and Model C allows for change
in level and trend. We consider the following DGP:
yt = δ'Zt + Xt , Xt = βXt-1 + εt , (1)
where Zt is a vector of exogenous variables, A(L)εt = B(L)ut, and A(L) and B(L) are finite
order polynomials with ut ~ iid (0,σ2). Two structural breaks can be considered from the
above DGP as follows.2 Model A allows for two changes in level and is described by Zt
= [1, t, D1t, D2t]', where Djt = 1 for t ≥ TBj + 1, j=1,2, and zero otherwise. Model C
includes two changes in level and trend, and is described by Zt = [1, t, D1t, D2t, DT1t*,
DT2t*]', where DTjt
* = t for t ≥ TBj + 1, j=1,2, and zero otherwise. The DGP includes
2 Model B is omitted from the discussion that follows, as it is commonly held that most economic timeseries can be described adequately by Model A or C.
4
both the null (β = 1) and alternative (β < 1) models in a consistent manner. For instance,
consider Model A (a similar argument can be applied to Model C); depending on the
value of β, we have:
Null yt = µ0 + d1B1t + d2B 2t + yt-1 + vt , (2a)
Alternative yt = µ1 + γt + d1*D1t + d2*D2t + vt , (2b)
where vt is a stationary error term, and Bjt = 1 for t = TBj + 1, j=1,2, and zero otherwise.
We let d = (d1, d2)′. For Model C, Djt terms are added to (2a) and DTjt* to (2b),
respectively. Note that the null model in (2a) includes dummy variables (Bjt) to allow for
two possible breaks. Nesting both the null and alternative models from (2a) and (2b), we
may consider the two-break augmented unit root test equation of LP as follows:
yt = α0 + α1t + α2B1t + α3B2t + α4D1t + α5D2t + φ yt-1 + ∑j=1
k
cj∆yt-j+ et . (3)
In the exogenous test, the break points are known, and TBj/T → λj as T → ∞, where λ =
(λ1, λ2)′. LP provide the asymptotic distribution of the exogenous t-statistic testing φ = 1
when omitting Bjt terms from (3). We denote the LP exogenous test statistic omitting Bjt
terms as “τ̂*” and when including Bjt terms as “τ̂.” The asymptotic distribution of the LP
exogenous test statistic is found to depend on λ in either case. One critical limitation of
the LP test is that while the test may be valid if the size of breaks under the null is zero,
(i.e., d = 0 in (2a)), it is invalid if d ≠ 0. The asymptotic distribution of the unit root test
statistic is not invariant to d under the null, and the associated t-statistic diverges as d
increases. In this case, it is necessary to include Bjt terms in (3) to insure that the
asymptotic distribution of the test statistic will be invariant to d. This is quite important,
5
as Perron (1989, p. 1393) noted this potential for divergence in his exogenous break test
and therefore included Bt.3
The same divergence problem is found for the endogenous break unit root test of
LP. The authors omit Bjt terms from their testing equation (3), and derive their critical
values assuming d1 = d2 = 0 under the null hypothesis (2a). While this assumption might
be necessary for this type of endogenous test not to depend on the location of the
break(s), this introduces a different problem: the test statistic diverges as the magnitude
of breaks under the null increase, the same as in the exogenous test. Further, unlike in the
exogenous break test, the LP test statistic is shown to diverge even if Bjt terms are
included in the testing equation (3). In sum, ignoring structural breaks under the null can
lead to serious problems.4
Fortunately, the divergence problem of the LP test is not found for the two-break
minimum LM test. Test statistics for the LM unit root test can be obtained according to
the LM (score) principle from the following regression:
∆yt = δ '∆Zt + φ S∼t-1 + ut , (4)
where S∼
t = yt - ψ∼
x - Ztδ∼
, t=2,..,T, δ∼ are coefficients in the regression of ∆yt on ∆Zt, and ψ∼x
is the restricted MLE of ψx (≡ ψ + X0) given by y1 - Z
1δ∼ (see SP). To correct for
3 Perron (1993) and Perron and Vogelsang (1992) rectify their tests by adding the Bt term in order toeliminate dependency of their test statistics on the nuisance parameter d in the additive outlier model.
4 A different, but related, question is what happens in the standard unit root test (without breaks) if the nullhypothesis is true and there is a structural break? This question was initially addressed in Amsler and Lee(1995), who showed that, unlike under the alternative, the standard unit root tests are unaffected byignoring a break under the null. Recently, Leybourne, Mills, and Newbold (1998) show that using astandard Dickey-Fuller test (without break) can lead to spurious rejection of the null if a structural breakoccurs early in the series. Contrary to this, Lee (1999) shows that this spurious rejection problem does notoccur for the LM unit root test of Schmidt and Phillips (1992).
6
autocorrelated errors, one can include augmented terms ∆S∼
t-j, j=1,..,k, in (4), as in the
augmented DF type test. The unit root null hypothesis is described by φ = 0, and the LM
test statistics are given by:
ρ∼ = T· %φ , (5a)
τ∼ = t-statistic testing the null hypothesis φ = 0 . (5b)
Theorem 1. Assume (i) the data are generated according to (1), with Zt = (1, t, D1t, D2t)'
for Model A, and Zt = [1, t, D1t, D2t, DT1t*, DT2t
*]' for Model C; (ii) the innovations εt
satisfy the regularity conditions of Phillips and Perron (1988, p. 336); and (iii) TBj /T →
λj as T → ∞. Then, under the null hypothesis that β = 1:
ρ∼ → - 1 2
σe2
σ2 [⌡⌠0
1 V_(m) 2(r)dr] , (6a)
τ∼ → - 1 2
σ σe
[⌡⌠0
1 V_(m) 2(r)dr]-1/2 , (6b)
where V_(m)(r) is defined for m = A or C, V_(A)(r) is a demeaned Brownian bridge, and
V_(C)(r) = V_(C)(r, λ) is a demeaned and de-breaked Brownian bridge.
Proof is given in the Appendix.
An important implication of Theorem 1 is that the asymptotic null distribution of
the LM statistics in (6) for Model A do not depend on location of the breaks (λj = TBj/T).
Thus, the LM test can allow for two breaks under the null without depending on nuisance
parameters. In addition, the asymptotic distribution of the two-break LM test is the same
as that of the SP test (without breaks), implying that critical values from the SP test can
7
be used for the two-break LM test. This invariance result, initially shown in Amsler and
Lee, holds for a finite number of breaks. Regardless of the presence or absence of
structural breaks, the asymptotic null distribution of the two-break LM test statistic is
unaffected, thereby making the test robust to misspecification of break points under the
null. This result, however, does not hold for Model C. The asymptotic distribution of the
two-break LM test statistic in Model C depends on λ, but unlike in the LP test, the two-
break LM test remains free of spurious rejections.
The minimum LM test uses a grid search to determine the location of two breaks
(TBj) as follows:
LMρ = Inf λ
ρ∼(λ) , (7a)
LMτ = Inf λ
τ∼(λ) . (7b)
The break point estimation scheme is similar to that of the LP test. Yet, in spite of
similar estimation schemes, we shall see that the performance of these tests is quite
different. The asymptotic distribution of the two-break minimum LM test can be
described as follows:
Corollary 1. Under the null hypothesis that β = 1:
LMρ → Inf λ
[ - 1 2
σe2
σ2 {⌡⌠0
1 V_(m )2(r) dr} ] , (8a)
LMτ → Inf λ
[- 1 2
σ σe
⌡⌠0
1 V_(m)2(r) dr]-1/2 . (8b)
To derive critical values for the two-break LM unit root test we generate pseudo-
iid N(0,1) random numbers using the Gauss (version 3.2.12) RNDNS procedure. Critical
values are derived using 50,000 replications for the exogenous break tests, and 5,000
8
replications for the endogenous break tests in samples of T = 100. Critical values are
shown in Table 1 and 2.
3. Simulation Results
This section examines simulation results to compare performance of the minimum
LM test with that of the LP test. While LP consider only testing equations without B1t
and B2t, we additionally consider their test with these terms included. In the discussion
that follows, we denote “LPt*” as the LP two-break test with Bjt terms included in the
testing equation, and “LPt” omitting these terms. Since performance of the LMρ test is
similar, we discuss only the LMτ test. We first examine the exogenous two-break unit
root test, assuming the break points are known, and then proceed to the endogenous break
test. Separate examination of these tests might be useful to investigate more specifically
the effect of using incorrect break points, since using incorrect break points in the LP test
is shown to introduce a bias in estimating the regression parameters. Simulations are
performed for the LM and LP tests using 20,000 replications for the exogenous tests and
2,000 replications for the endogenous tests in samples of T = 100. Throughout, R denotes
the number of structural breaks, λ is a vector denoting location of the breaks, and d is a
vector denoting magnitude of the breaks in the DGP. Re and λe denote the values used in
the tests. All measures of size and power are reported using 5% critical values.
Exogenous Tests
We first examine the exogenous two-break unit root tests in Table 3. In
Experiment A and B, we investigate the effect of different λ and d in Model A. The LMτ
test is clearly invariant to λ and d, thus confirming the invariance results of Theorem 1.
9
As expected, the LPt test shows significant size distortions, which increase with the
magnitude of the breaks. This provides evidence that omitting the Bjt terms from the
exogenous break test introduces a serious problem. While the LPt test appears to have
greater power than the LMτ test in the presence of structural breaks, this result is spurious
and due to the large size distortions. In sum, we can say that the exogenous break LPt
test is simply invalid. While the LPt* test is invariant to d, and therefore free of the
divergence problem, it is still not invariant to λ. In all cases, the LMτ test has greater
power to reject the null than the LPt* test.
In Experiment C, we examine effects of under-specifying the number of breaks
(Re < R). As expected, the LMτ test is mostly invariant to assuming an incorrect number
of breaks under the null, while the LPt test is more seriously affected. Both the LMτ and
LPt* tests lose power under the alternative. This result can be seen as a generalization of
the finding of Perron and Amsler and Lee, indicating that unit root tests lose power when
the number of breaks is underestimated. In Experiment D, we examine effects of
assuming incorrect break points. The LMτ test is again mostly invariant to using incorrect
break points under the null, while other effects are similar to under-specifying the number
of breaks.
Results for Model C are similar to those for Model A, except that the LMτ test is
no longer strictly invariant to λ under the null, but remains invariant to d. The LPt test is
again invalid due to large size distortions and spurious rejections. In Experiment C′ and
D′, we see that the three test statistics all have (mostly negative) size distortions when
break locations are incorrectly estimated or the number of breaks is underestimated.
10
Endogenous Minimum Tests
The two-break endogenous unit root test is examined in Table 4. Experiment E
compares rejection rates for Model A at different break locations and magnitudes.
Overall, the LMτ test performs well with no serious size distortions. Contrary to this, the
LPt and LPt* test suffer significant size distortions and spurious rejections, which
increase with the magnitude of the breaks. Unlike for the exogenous test, the LPt* test is
now as invalid as the LPt test. Thus, regardless of including or excluding Bjt terms in the
testing regression (3), the endogenous two-break LP test exhibits spurious rejections
when d ≠ 0.
The above unexpected result for the LPt* test is closely related to estimating
incorrect break points. Table 4 reports the frequency of estimating break points over a
specified range. For the LPt* test, the frequency of estimating the break points correctly
at TB is virtually zero. Instead, the LPt* test most frequently selects break points
incorrectly at TB-1. This problem becomes more serious as the breaks increase in size.
The reason is that the LPt* test statistic is generally smaller when break points are
misspecified, reaching a minimum at TB-1. Therefore, when the LPt* test searches for the
minimum t-statistic, it most frequently selects the break points incorrectly at TB-1. This
causes the LPt* test t-statistic to diverge and become smaller as |d| increases. This
problem is critical, and is associated with bias in estimating the crucial parameter φ in (3),
corresponding to β in (1). Similar to results in Lee and Strazicich (1999a) for the ZA
one-break test, we shall see below (Table 5) that the bias in estimating β using the LPt*
test is maximized at TB LP test in Table 4 appears to select correct break
11
points most frequently at TB, this outcome is misleading, as spurious rejections are
maximized.
The endogenous break test in Model C is examined in Panel (2) of Table 4. The
minimum LMτ test has somewhat greater size distortions than in Model A, but rejections
are still near 5%. This is not the case for the LP tests. Again, the LPt and LPt* test each
have significant size distortions and spurious rejections when the DGP contains structural
breaks. Apparent success in estimating the break points with the LPt test is again
misleading, as spurious rejections are maximized at TB. Fortunately, the LMτ test remains
free of spurious rejections. Therefore, the LMτ test may still be used in Model C, as long
as critical values are employed corresponding to the break points estimated.5
Bias Effects
In order to see further why using incorrect break points leads to spurious
rejections, we examine possible bias and mean squared error (MSE) in estimating β (or φ)
and σ for Model A in Table 5. An important advantage of the LM test is revealed; bias in
estimating β under the null is small and unaffected by incorrect break points. A similar
result can be seen examining empirical critical values. At the break points selected most
often with the LMτ test (TB), the empirical critical values are invariant to the magnitude of
the breaks and mostly unaffected by incorrect estimation of their location. This is not the
case for the LP tests. Empirical critical values for the LPt and LPt* test depend both on
the magnitude and location of the breaks. The source of the size distortions is revealed;
5 Critical values for the two-break LM test in Model C are provided in Table 2 for a variety of two breakpoint combinations.
12
the LPt* test selects break points where the t-statistic testing for a unit root is minimized
and the bias in estimating β is maximized. Excluding Bjt terms from the testing
regression, the LPt test also selects break points where bias is maximized. Regardless of
including or excluding Bjt terms in the testing regression, the t-statistic in the LPt and
LPt* test is minimized where the bias in estimating β is maximized. Omitting the Bjt
terms in the LPt test only transfers the bias one period from TB-1 to TB. Just as the
exogenous test omitting Bjt terms diverges, the endogenous LPt test also diverges when
selecting break points where bias is maximized, thus both the LPt and LPt* tests are
equally invalid. Except for the break points estimated, the LPt and LPt* tests are quite
similar, since both tend to select break points where bias and spurious rejections are the
greatest as |d| increases. Results for the LPt and LPt* test under the alternative, in terms
of power and break point estimation, are therefore misleading. MSE tells a similar story
for β. Results for σ in terms of bias and MSE are similar to those for β.
4. Empirical Tests
In this section, the two-break minimum LM and LP tests are applied to the Nelson
and Plosser (1982) data. The data comprise fourteen annual time series ranging from
1860 (or later) to 1970 and have the advantage of being examined often in the literature.
All series are in logs except the interest rate. For each test, we determine the number of
augmentation terms, ∆S∼
t-i, i = 1,..,k, in (4) for the LMτ test, and ∆yt-i in (3) for the LPt and
LPt* test, by following the procedure in Perron and LP. Starting from a maximum of k =
8 lagged terms, we examine each combination of two break points over the time interval
13
[.1T, .9T].6 After determining the “optimal k,” we determine the break points where the t-
statistic is a minimum. Throughout, we follow Perron and ZA and assume Model A for
all series except real wages and the S & P 500 stock index, in which we assume Model C.
Overall, we find stronger rejections of the unit root null using the LPt and LPt*
tests than with the LMτ test. At the 5% significance level, the null is rejected for six
series by the LPt* and LPt test, and for four series by the LMτ test.7 For example, while
the null is rejected at the 5% significance level for Real GNP, Nominal GNP, Per-capita
Real GNP, and Employment in the LPt and LPt* tests, the null is rejected at only higher
significance levels using the LMτ test.8 Compared with results in ZA (finite sample
critical values) for one-break, we observe the same number of unit root rejections. This
result may indicate that power to reject the null diminishes if the number of breaks is
under-specified.
To investigate the potential for spurious rejections, we also estimate the size of
structural breaks under the null. The null model is estimated in (2a) using the first
differenced series. Briefly, for each possible combination of TB1 and TB2 in the interval
[.1T, .9T], we again determine the k-augmented terms by using the general to specific
procedure. We then determine the break points where the Schwarz Bayesian Criterion is
6 This “general to specific” procedure looks for significance of the last augmented term. We use the 10%asymptotic normal value of 1.645 on the t-statistic of the last lagged term. The procedure has been shownto perform better than other data-dependant procedures (see, e.g., Ng and Perron, 1995). The trimming ofend points does not affect estimation of the test statistics, but critical values are affected by the trimming.Lumsdaine and Papell (1997) use 1% trimming.
7 Throughout the empirical section, we use the critical values from Table 2 (Model A) and Table 3 (ModelC) in Lumsdaine and Papell (1997) for the LPt and LPt* test statistics (asymptotically equivalent). Forcomparison, LM test critical values were derived using the same sample size and trimming as in Lumsdaineand Papell (T = 125 and 1%). LM test critical values are -4.571, -3.937, and -3.564 for Model A, and -6.281, -5.620, and -5.247 for Model C, at the 1%, 5%, and 10% significance levels, respectively.
14
minimized. Estimated break coefficients in standardized units are shown, along with
other results, in Table 6. Break terms appear significant under the null in most series,
with magnitudes ranging from near 2 to 8. This suggests that even modest size breaks
can lead to different test results, or at least move significance levels in that direction.
Similar to the simulation results in Section 3, we observe that the LPt* test often
selects break points one period before the LPt and LMτ test. For example, the LPt* test
estimates break points for Nominal GNP at 1919 and 1928, while the LPt and LMτ test
estimate break points at 1920 and 1929, and 1920 and 1948, respectively. There is no
evidence that structural breaks occurred in either 1919 or 1928; on the contrary, it is more
reasonable to argue that the correct break points are 1920 and 1929, since Nominal GNP
fell by 24% in 1921 and 12% in 1930. A similar pattern is frequently observed in the
other series. Throughout, the LPt* test tends to select incorrect break points in a
consistent manner when there are significant structural breaks under the null. Omitting
the Btj terms in the LPt test only moves the estimated break points to one period later.
Other than this, the LPt* and LPt tests have similar results throughout. We note that
estimation of the break points can be imprecise and different tests produce different
results. This difficulty may not pose a serious problem for the LM unit root test, since
simulation results indicated the LM test is robust to break point misspecification.
5. Summary and Concluding Remarks
In many time series, allowing for one structural break may be too restrictive. A
unit root test allowing for more than one structural break could, therefore, lead to greater
8 For real wages and the money stock the opposite is the case.
power to reject the null. This paper proposes a two-break minimum LM unit root test as
a remedy to the spurious rejections found for the two-break minimum test of Lumsdaine
critical issue in these Dickey-Fuller type minimum tests is their assumption of no
break(s) under the null. This assumption was found to be detrimental in the presence of
different inference results in unit root tests. On the contrary, the asymptotic null
distribution of the two-break minimum LM unit root test was shown to be invariant to
test, our findings serve to caution researchers inclined to apply endogenous break unit
root tests that assume no breaks under the null.
References
Amsler, C. and J. Lee. (1995). “An LM Test for a Unit-Root in the Presence of aEconometric Theory 11, 359-368.
and Long-Run Growth: Evidence from Two Structural Breaks,” Working Paper,University of Houston.
Process Among the G7 Countries,” Working Paper, University of Houston.
Lee, J. (1999). “The End-Point Issue and the LM Unit Root Test,” Working Paper,
Lee, J., J. List, and M. Strazicich. (1998). “Spurious Rejections with the Minimum UnitRoot Test in the Presence of a Structural Break under the Null,” Working Paper,
Lee, J. and M. Strazicich. (1999a). “Break Point Estimation with Minimum Unit RootTests and Spurious Rejections of the Null,” Working Paper, University of Central
Lee. J. and M. Strazicich. (1999b). “Minimum LM Unit Root Tests,” Working Paper,University of Central Florida.
Fuller Tests in the Presence of a Break Under the Null,” Journal of Econometrics191-203.
Lumsdaine, R. and D. Papell. (1997). “Multiple Trend Breaks and the Unit-RootReview of Economics and Statistics, 212-218.
Time Series,” Journal of Monetary Economics
Ng and Perron. (1995). “Unit Root Tests in ARMA Models with Data-DependentMethods for the Selection of the Truncation Lag,” Association 90, 269-281.
Evidence on the Great Crash and the Unit Root Hypothesis Reconsidered,” Oxford 59, 435-448.
Papell, D., C. Murray, and H. Ghiblawi. (1999). “The Structure of Unemployment,”
17
Perron, P. (1989). “The Great Crash, the Oil Price Shock, and the Unit RootHypothesis,” Econometrica 57, 1361-1401.
Perron, P. (1997). “Further Evidence on Breaking Trend Functions in MacroeconomicVariables,” Journal of Econometrics 80, 355-385.
Perron, P. (1993). “Erratum,” Econometrica 61, 248-249.
Perron, P. and T.J. Vogelsang. (1992), “Testing for a Unit Root in Time Series with aChanging Mean: Corrections and Extensions,” Journal of Business and EconomicStatistics 10, 467-470.
Phillips, P.C.B. and P. Perron. (1988). “Testing for a Unit Root in Time SeriesRegression,” Biometrika 75, 335-346.
Schmidt, P. and P.C.B. Phillips. (1992) “LM Tests for a Unit Root in the Presence ofDeterministic Trends,” Oxford Bulletin of Economics and Statistics 54, 257-287.
Zivot, E. and D. W. K. Andrews. (1992). “Further Evidence on the Great Crash, the Oil-Price Shock and the Unit Root Hypothesis,” Journal of Business and EconomicStatistics 10, 251-270.
18
Table 1
Critical Values of Exogenous LM Tests (T=100)
(1) Model A
1% 5% 10%
τ∼ -3.63 -3.06 -2.77
ρ∼ -23.8 -17.5 -14.6
Note: Critical values for Model A are the same as those in Schmidt and Phillips(1992).
(2) Model C
(i) τ∼
λ2
λ1 .4 .6 .8.2 -4.83, -4.19, -3.89 -4.93, -4.31, -4.00 -4.76, -4.19, -3.88.4 - -4.91, -4.33, -4.03 -4.88, -4.32, -4.03.6 - - -4.84, -4.19, -3.89
(ii) ρρ∼∼
λ2
λ1 .4 .6 .8.2 -38.1, -30.2, -26.4 -39.4, -31.6, -27.9 -37.2, -30.1, -26.3.4 - -39.2, -31.8, -28.1 -38.7, -31.7, -28.1.6 - - -38.3, -30.2, -26.4
Note: Critical values are at the 1%, 5%, and 10% levels, respectively.
19
Table 2
Critical Values of the Endogenous Two-Break Minimum Tests (T = 100)
(1) Model A
1% 5% 10%LMτ -4.545 -3.842 -3.504LMρ -35.73 -26.89 -22.89LPt -6.420 -5.913 -5.587
LPt* -6.400 -5.853 -5.560
Note: The DGP in the simulation does not include breaks. The LP tests areaffected by breaks, while LM tests are invariant to breaks in Model A.
(2) Model C (I)
1% 5% 10%LMτ -5.825 -5.286 -4.989LMρ -52.551 -45.532 -41.664LPt -6.936 -6.386 -6.108
LPt* -6.945 -6.344 -6.064
Note: The DGP in the simulation does not include breaks. Both LP and LM testsare affected by breaks in Model C.
(3) Model C (II)
(i) LMττλ2
λ1 .4 .6 .8.2 -6.16, -5.59, -5.28 -6.40, -5.74, -5.32 -6.33, -5.71, -5.33.4 - -6.46, -5.67, -5.31 -6.42, -5.65, -5.32.6 - - -6.32, -5.73, -5.32
(ii) LMρρλ2
λ1 .4 .6 .8.2 -55.5, -47.9, -44.0 -58.6, -50.0, -44.4 -57.6, -49.6, -44.6.4 -59.3, -49.0, -44.3 -58.8, -48.7, -44.5.6 -57.5, -49.8, -44.4
Note: Critical values are at the 1%, 5%, and 10% levels, respectively.
20
Table 3Rejection Rates of the Exogenous Two-Break Tests (T = 100)
(1) Model A
Exp DGP EstimationUnder the Null
(β = 1.0)Under the Alternative
(β = 0.9)R λ′ d′ Re λe′ τ
∼ τ^ τ^* τ∼ τ^ τ^*
A 0 - - 2 .25, .50 .047 .040 .042 .243 .114 .1132 .25, .75 .047 .040 .040 .239 .110 .1082 .50, .75 .047 .039 .039 .240 .115 .114
B 2 .25, .50 5, 5 2 .25, .50 .047 .487 .042 .243 .763 .11310, 10 2 .25, .50 .047 .955 .042 .243 .998 .113
2 .25, .75 5, 5 2 .25, .75 .047 .485 .040 .239 .757 .10810, 10 2 .25, .75 .047 .956 .040 .239 .997 .108
C 2 .25, .50 5, 5 0 - .053 .005 .005 .125 .010 .010.25, .50 5, 5 1 .25 .045 .109 .013 .151 .201 .033.25, .50 5, 5 1 .50 .045 .092 .012 .137 .127 .023
2 .25, .50 10, 10 0 - .037 .001 .001 .019 .000 .000.25, .50 10, 10 1 .25 .032 .245 .005 .040 .305 .004.25, .50 10, 10 1 .50 .031 .156 .003 .026 .085 .001
D 2 .25, .50 5, 5 2 .25, .75 .046 .151 .028 .147 .250 .05810, 10 2 .25, .75 .033 .292 .010 .038 .359 .010
(2) Model C
Exp DGP EstimationUnder the Null
(β = 1.0)Under the Alternative
(β = 0.9)R λ′ d′ Re λe′ τ
∼ τ^ τ^* τ∼ τ^ τ^*
A′ 0 - - 2 .25, .50 .052 .050 .050 .118 .101 .0972 .25, .75 .048 .052 .050 .115 .101 .0992 .50, .75 .051 .055 .054 .119 .105 .101
B′ 2 .25, .50 5, 5 2 .25, .50 .052 .625 .050 .118 .773 .09710, 10 2 .25, .50 .052 .986 .050 .118 .998 .097
2 .25, .75 5, 5 2 .25, .75 .048 .627 .051 .115 .773 .09910, 10 2 .25, .75 .048 .986 .051 .115 .999 .099
C′ 2 .25, .50 5, 5 0 - .000 .000 .000 .000 .000 .000.25, .50 5, 5 1 .25 .002 .011 .004 .003 .007 .006.25, .50 5, 5 1 .50 .004 .080 .005 .004 .102 .008
2 .25, .50 10, 10 0 - .000 .000 .000 .000 .000 .000.25, .50 10, 10 1 .25 .000 .001 .000 .000 .000 .000.25, .50 10, 10 1 .50 .000 .103 .000 .000 .096 .000
D′ 2 .25, .50 5, 5 2 .25, .75 .015 .045 .015 .014 .031 .01610, 10 2 .25, .75 .000 .006 .000 .000 .001 .000
21
Table 4Rejection Rates and Estimated
Break Points of the Endogenous Break Tests (T = 100)
(1) Model A
Frequency of Estimated Break Points in theRangeExp λ′ d′ Test 5% Rej.
TB-1 TB TB± 10 TB± 30
Under the null (β = 1)
E - 0, 0 LMτ .044 - - - -LPt .046 - - - -
LPt* .038 - - - -.25, .5 5, 5 LMτ .064 .000 .126 .254 .766
LPt .192 .000 .176 .402 .804LPt* .180 .206 .000 .426 .816
.25, .5 10, 10 LMτ .034 .002 .252 .500 .804LPt .748 .000 .818 .882 .976
LPt* .748 .826 .000 .890 .978.25,.75 5, 5 LMτ .052 .000 .028 .152 .646
LPt .170 .008 .162 .392 .752LPt* .166 .190 .000 .408 .760
.2, .3 5, 5 LMτ .040 .000 .266 .438 .668LPt .206 .000 .136 .318 .598
LPt* .174 .148 .000 .286 .586
Under the alternative (β = .9)
F .25, .5 0, 0 LMτ .282 - - - -LPt .098 - - - -
LPt* .078 - - - -.25, .5 5, 5 LMτ .172 .000 .276 .480 .808
LPt .318 .002 .332 .600 .912LPt* .292 .356 .000 .624 .910
.25, .5 10, 10 LMτ .042 .000 .564 .778 .894LPt .954 .000 .972 .990 .998
LPt* .944 .974 .000 .986 .996.25,.75 5, 5 LMτ .204 .004 .132 .304 .734
LPt .298 .016 .304 .594 .820LPt* .284 .318 .000 .616 .824
.2, .3 5, 5 LMτ .130 .000 .364 .580 .738LPt .336 .000 .248 .462 .698
LPt* .302 .262 .000 .434 .702
22
(2) Model C
Frequency of Estimated Break Points in theRangeExp λ′ d′ Test 5% Rej.
TB-1 TB TB± 10 TB± 30
Under the null (β = 1)
E′ - 0, 0 LMτ .080 - - - -LPt .052 - - - -
LPt* .054 - - - -.25, .5 5, 5 LMτ .022 .006 .020 .462 .876
LPt .272 .000 .400 .630 .930LPt* .286 .418 .000 .612 .934
.25, .5 10, 10 LMτ .018 .002 .020 .738 .990LPt .882 .000 .966 .988 .996
LPt* .886 .968 .000 .984 .998.25,.75 5, 5 LMτ .042 .004 .018 .550 .948
LPt .262 .004 .318 .576 .954LPt* .296 .338 .000 .582 .956
.2, .3 5, 5 LMτ .056 .002 .000 .152 .486LPt .146 .000 .152 .248 .528
LPt* .158 .136 .000 .250 .534
Under the alternative (β = .9)
F′ .25, .5 0, 0 LMτ .124 - - - -LPt .098 - - - -
LPt* .096 - - - -.25, .5 5, 5 LMτ .050 .006 .040 .534 .942
LPt .346 .004 .514 .728 .956LPt* .370 .560 .000 .738 .964
.25, .5 10, 10 LMτ .040 .000 .042 .726 1.00LPt .968 .000 .996 1.00 1.00
LPt* .968 .998 .000 1.00 1.00.25,.75 5, 5 LMτ .064 .006 .026 .590 .970
LPt .348 .002 .412 .664 .966LPt* .372 .454 .000 .664 .968
.2, .3 5, 5 LMτ .130 .002 .002 .198 .542LPt .246 .000 .248 .362 .620
LPt* .258 .238 .000 .358 .626
Note: Critical values of the model without breaks are used.
23
Table 5Bias Effects of Using Incorrect Break Points (T = 100)
(Model A)
(a) Under the Null (β = 1)
β σd Test BreakPoint
5%Rej.
Emp.Crit. Bias MSE Bias MSE
5, 5 τ∼
TB -2 .065 -3.16 -.086 .010 .181 .039TB -1 .065 -3.14 -.086 .010 .181 .039
TB .063 -3.14 -.084 .010 -.038 .006TB+1 .065 -3.14 -.086 .010 .181 .039
τ^TB -2 .239 -4.86 -.189 .040 .120 .021TB -1 .376 -5.49 -.210 .050 .107 .018
TB .514 -6.50 -.241 .068 .087 .015TB+1 .027 -3.85 -.141 .028 .159 .032
τ^*TB –2 .380 -5.48 -.215 .052 .097 .016TB –1 .523 -6.56 -248 .072 .075 .013
TB .049 -4.10 -.145 .026 -.070 .010TB+1 .026 -3.84 -.144 .029 .150 .029
10, 10 τ∼
TB –2 .042 -3.00 -.083 .009 .678 .468TB –1 .043 -2.98 -.083 .009 .678 .469
TB .063 -3.14 -.084 .010 -.038 .006TB+1 .042 -3.00 -.083 .009 .678 .468
τ^TB –2 .745 -5.57 -.250 .066 .522 .280TB –1 .903 -7.01 -.312 .103 .460 .221
TB .960 -10.9 -.430 .201 .335 .132TB+1 .009 -3.48 -.141 .037 .651 .432
τ^*TB -2 .907 -6.99 -.318 .107 .450 .212TB -1 .962 -10.99 -.440 .210 .319 .122
TB .049 -4.10 -.145 .026 -.070 .010TB+1 .009 -3.48 -.143 .039 .644 .423
24
(b) Under the Alternative (β = .9)
β σd Test BreakPoint
5%Rej.
Emp.Crit. Bias MSE Bias MSE
5, 5 τ∼
TB -2 .141 -3.45 -.017 .003 .194 .045TB -1 .138 -3.46 -.017 .003 .195 .045
TB .281 -3.86 -.045 .006 -.026 .006TB+1 .141 -3.48 -.017 .003 .194 .045
τ^TB -2 .477 -5.20 -.149 .027 .127 .022TB -1 .661 -5.90 -.182 .039 .108 .018
TB .794 -7.24 -.232 .064 .077 .014TB+1 .061 -4.19 -.111 .021 .176 .038
τ^*TB –2 .659 -5.91 -.186 .040 .098 .016TB –1 .797 -7.27 -.239 .067 .066 .012
TB .137 -4.58 -.114 .019 -.051 .008TB+1 .060 -4.18 -.113 .022 .167 .035
10, 10 τ∼
TB –2 .023 -2.84 .010 .001 .692 .487TB –1 .024 -2.84 .010 .001 .392 .487
TB .281 -3.86 -.045 .006 -.026 .006TB+1 .025 -2.86 .010 .001 .691 .487
τ^TB –2 .949 -5.78 -.191 .039 .515 .273TB –1 .995 -7.37 -.274 .079 .436 .198
TB .999 -12.05 -.440 .206 .260 .085TB+1 .016 -3.67 -.110 .031 .666 .452
τ^*TB -2 .995 -7.33 -.278 .081 .427 .190TB -1 .999 -12.13 -.449 .213 .246 .077
TB .137 -4.58 -.114 .019 -.051 .008TB+1 .015 -3.66 -.113 .032 .659 .443
25
Table 6Empirical Results
LPt* LPt LMτ Null ModelSERIES Model
k̂ T̂BStat.
k̂ T̂BStat.
k̂ T̂BStat.
d̂1*, d̂2* a, b T̂B
Real GNP A 2 19281937
-7.00* 1 19291940
-6.65* 7 19201941
-3.62 3.09, -2.67(2.97, -2.65)
19211929
Nominal GNP A 8 19191928
-7.50* 8 19201929
-7.42* 8 19201948
-3.65 -4.84, -3.461(-4.80, -3.27)
19201931
Per-capita real GNP A 2 19281939
-6.88* 2 19291939
-6.67* 7 19201941
-3.68 3.07, -2.60(2.94, -2.57)
19211929
Industrial Production A 8 19171928
-7.67* 8 19181929
-7.78* 8 19201930
-4.32* -3.73, -4.38(-3.63, -4.13)
19201931
Employment A 8 19281955
-6.80* 8 19291956
-6.83* 7 19201945
-3.91 -2.90, 2.51(-2.73, 2.33)
19311941
Unemployment Rate A 7 19281941
-6.31* 7 19291941
-6.63* 7 19261942
-4.47* -3.43, 1.97(-3.38, 1.83)
19171920
GNP Deflator A 8 19161920
-4.74 1 19291945
-4.64 1 19191922
-3.18 3.88, -8.49(3.73, -7.14)
19171921
CPI A 2 19141944
-4.03 5 19151940
-4.04 4 19161941
-3.92 -2.44, -7.78(-7.13, -2.41)
19201930
Nominal Wage A 7 19141929
-5.85 7 19301949
-5.59 7 19211942
-3.84 -3.75, -2.98(-3.62, -2.89)
19201931
Real Wage C 4 19211940
-6.27 4 19221940
-6.63 8 19221939
-6.24* -3.10, -.57(-2.54, -3.01)
19311945
Money Stock A 8 19291958
-6.22 8 19301958
-6.03 7 19271931
-4.31* -3.54, -3.63(-3.50, -3.50)
19201931
Velocity A 1 18831953
-4.62 1 18841949
-4.77 1 18931947
-2.52 2.33, -2.28(2.32, -2.27)
19411944
Interest Rates A 2 19311957
-1.74 2 19321958
-1.74 3 19491958
-1.58 2.67, -2.50(2.64, -2.45)
19171921
SP500 C 1 19251938
-6.37 1 19241937
-6.12 3 19251941
-5.57 3.12, 3.35(4.82, 2.54)
19281932
Note: * denotes significant at 5%. a: Standardized coefficients (d^
i* = (d^
i/σ^) are reported. b: t-statistics for di = 0 are given in parentheses.
26
Appendix
Proof of Theorem 1
(a) Model A
Amsler and Lee (1995) derive the asymptotic distributions of LM test statistics
with one known, or exogenous, structural break. Here, we consider a more general case
with a finite number of, say, m << T structural breaks. Let Zt = (t, Wt′)′, where Wt =
(D1t,..,Dmt)′ and δ = (δ1, δ2′)′. From the regression of ∆yt on ∆Zt, we obtain δ∼ =
(∆Z′∆Z)−1∆Z′∆y, where ∆Z = (∆Z1, ∆Z2,.., ∆ZT)′, and ∆y = (∆y1, ∆y2,.., ∆yT)′. Following
SP, we define S∼
t = yt - (y1 - Z1δ∼) - Ztδ∼
. Letting St = ∑j=2
t εj and [rT] be the integer part of
rT, r ∈ [0,1], we obtain
T -1/2
S∼
[rT] = T
-½S
[rT] - T -1([rT]-1) T ½(δ∼
1-δ
1) – T -1(W
[rT] -W1)′ T ½(δ∼
2-δ
2 ) . (A.1)
The first term on the right hand side of (A.1) follows T-1/2S[rT] → σW(r). For the second
term, we note T(δ∼1-δ1) = (
1 T i′Μ∆W i)
-1 1 T
i′Μ∆Wε, where Μ∆W=I-∆W(∆W′∆W)-1∆W′.
Here, 1 T i′Μ∆W i → 1, since i′∆W = im′ (1 × m vector of ones), and i′Μ∆W i = T-m-1.
Then,
1 T
i′Μ∆Wε = 1
T ∑j=2
T
εj - 1
T ∑i=1
m
εTbi+1 → σW(1) ,
T -1
([rT]-1) T ½(δ∼
1-δ
1) → σ rW(1) .
We can show that the third term vanishes asymptotically. Since W[rT]
-W1 → im,
T(δ∼2-δ
2) = (
1 T ∆W′Μ1 ∆W)-1 1
T ∆W′Μ1ε = op(1) ,
27
whereas ∆W′Μ1∆W = Im - ImT-1→ Im, and ∆W′Μ1ε = (εTB1+1,..,εTBm+1)′ - imε_. Thus,
combining results, we can show that the terms in (A.1) follow
T -½
S∼
[rT] → σ[W(r) - rW(1)] = σV(r) , (A.2)
where V(r) is a Brownian bridge. This is the same expression as obtained from the usual
SP test ignoring a break (see the equation before (A3.1) in SP, 1992, p. 283). We also
define a demeaned Brownian bridge V_(A)(r) = V(r) –⌡⌠0
1V(r)dr as in SP, and it follows that
ρ∼ and τ∼ have the same asymptotic distributions as the usual SP tests not allowing for
structural breaks.
(b) Model C
We let Zt = (t, Wt*′)′, and Wt* = (D1t,D2t, DT1t*,D2t*)′. We define u
∼1 = 0 and u
∼t = OLS
residuals, t = 2,..,T, from the regression
∆yt = ∆Ztδ + ut . (A.3)
Note that the invariance result does not hold in Model C, and the test statistics in (5)
depend on λ, the parameter indicating the break points. Thus, we define a de-break
Brownian bridge V(C)(λ,r) as a residual process projected onto the subspace generated by
dz(λ,r) = [1, d1(λ1,r), d2(λ2,r)], where dj(λj,r) = 1 if r > λj, for j=1,2, and 0 otherwise.
Here, this expression is free of the effect of bj(λj,r) asymptotically, where bj(λj,r) = 1 if r
= λj, j=1,2, and 0 otherwise, but depends on λ. Then, the residuals from (A.3) follow
1 T
∑j=1
[rT]
u∼
j → V(C)(λ,r) .
Further, as in Perron (1997), we consider the following regression
28
∆yt = δ(λ)'∆Zt(λ) + φ(λ) S∼t-1(λ) + et , t = 2,..,T, (A.4)
where S∼
t(λ) = ∑j=2
t εj - (δ∼(λ)' - δ(λ)')(Zt(λ)- Z1(λ)). We let M∆Ζ(λ) = I - P∆Ζ(λ), where
P∆Ζ(λ) = ∆zT(λ)[∆zT(λ)′∆zT(λ)]-1∆zT(λ), and where ∆zT(λ) = (∆z1,T(λ),..,∆zT,T(λ))′. Pre-
multiplying (A.4) by M∆Ζ(λ), we obtain
M∆Ζ(λ)∆Y = φ(λ) M∆Ζ(λ)S∼
1(λ) + M∆Ζ(λ) e , (A.5)
where ∆Y = (∆y2,.., ∆yT)′, S∼1(λ) = (S∼1(λ),..,S∼T-1(λ))′ and e = (e2,..,eT)′. Then, the τ∼ statistic
in (5a) can be written as
τ∼ = [T -2S∼
1(λ)′ M∆Ζ(λ) S∼
1(λ)] -1/2[T -1S∼
1(λ)′ M∆Ζ(λ) e] / sT(λ) ,
where sT(λ) is the corresponding standard error of the regression. We obtain
T -2S∼
1(λ)′ M∆Ζ(λ) S∼
1(λ) = σ2
⌡⌠0
1[ST(r) - P∆Ζ(λ) ST(r)]2 dr , (A.6)
T -1S∼
1(λ)′ M∆Ζ(λ) e = σ2
⌡⌠0
1ST(r)dST(r) - σ2
⌡⌠0
1P∆Ζ(λ) ST(r)dST(r) . (A.7)
The effect of applying M∆Ζ(λ) or P∆Ζ(λ) to the above expressions is twofold; one is to
demean the process, and the other is to de-trend the structural dummy effect. Then, it is
given that
⌡⌠0
1[ST(r) - P∆Ζ(λ) ST(r)]2 dr = σ2
⌡⌠0
1V_(C)(λ,r)2 dr ,
where V_(C)(λ,r) is a demeaned and de-breaked Brownian bridge. The rest of the proof
follows that of SP, except that V_(C)(λ,r) replaces V_(r) in the expressions in the asymptotic
distribution of the SP statistics.
Proof of Corrolary1
29
The main procedure of the proof is to show continuity of a composite function.
We simply utilize the result of Zivot and Andrews (1992) on continuity of the composite
functional and make a note on corresponding notations. The minimum LMτ statistic can
be expressed as
Inf τ∼(λ∼) = g[ST(r), V_T(λ,r), ⌡⌠0
1ST(r)dST(r), ⌡⌠0
1P∆Ζ(λ) ST(r)dST(r), s2] + op(1) ,
where g = h*[h[H1(•), H2(•), sT(λ)]], with h*(m) = Inf m(•) for any real function m(•),
and h[m1, m2, m3] = m1-1/2m2/m3. The functionals H1 and H2 are defined by (A.6) and
(A.7) for Model C, while the term λ is absent in these expressions for Model A.
Continuity of h* and h is proved in ZA.