Advanced Review

Unit root tests

Edward Herranz*

Unit roots are nonstationary autoregressive (AR) or autoregressive moving average

(ARMA) time series processes which may include an intercept and/or a trend. These
processes are used often in economics and finance, but can also be found in other sci-
entific fields. Unit root tests address the null hypothesis of a unit root, and an alterna-
tive hypothesis of a stationary (or trend stationary) time series. Critical values for unit
root tests are typically derived via simulation of limiting distributions expressed as
functionals of Brownian motions. The critical values for the Dickey Fuller unit root
test with a constant and linear trend are derived via simulation in the R language.
Simulation studies are presented showing that linear regressions with unit root pro-
cesses often produce spurious results. Additional simulation studies are reviewed
providing statistical evidence that near-unit roots can often result in spurious cointe-
gration relationships. Various unit root tests are presented, including ones that allow
for structural breaks in intercept and/or trend. Threshold unit root tests are intro-
duced. Simulation studies are used to compare the unit root tests under various sce-
narios. The case where the analyzed time series may have stationary and
nonstationary segments is also considered. © 2017 Wiley Periodicals, Inc.
How to cite this article:
WIREs Comput Stat 2017, e1396. doi: 10.1002/wics.1396

Keywords: Nonstationary time series, unit root tests

INTRODUCTION Unit root tests (URT) address the null hypothesis

of a unit root; in the driftless AR(1) setting (xt =ϕ1xt − 1
U nit roots are nonstationary autoregressive (AR) or
autoregressive moving-average (ARMA) time
series processes; the simplest example is a random walk
+ εt) this statement is equivalent to assuming that
ϕ1 = 1 under the null. The alternative hypothesis is
often taken a stationary process with |ϕ1| < 1, however
such as xt = xt − 1 + εt where εt are random innova-
tions. If we assume that the innovations are independ- we will expand this to be ϕ1 ¼ 6 1 which includes nonsta-
ent and identically distributed with E(εt) = 0 and var tionary processes referred to as explosive time series
(εt) = 1, then E(xt) = x0 and var(xt) = t; the process is with |ϕ1| > 1. We will show using simulation studies
nonstationary because the variance is a function of that unit roots often result in spurious regressions. Addi-
time. These processes are common in economics and tional simulation studies will show that near-unit roots
finance. For example, as was originally proposed in the where ϕ1 = 1 + δ or ϕ1 = 1 − δ for a small positive δ,
seminal work by Fam,1 the logarithm of stock prices is will often lead to spurious cointegration relationships;
often modeled as a random walk: log(St) = log(St − 1) + this can happen when the URT used on the underlying
εt which is equivalent

tomodeling log returns as a sta- time series fail to reject the null hypothesis of a unit root.
tionary process: log St
= εt . Unit root processes can Before we can cover URT in detail, we will discuss
St −1
ARMA models, unit roots and cointegration.
be extended to ARMA processes, where there are
Throughout this review continuous time models
additional lagged innovation terms.
will not be considered, only discrete time models. How-
ever two continuous time processes related to unit roots
*Correspondence to: [Link]@[Link] that are commonly used in many fields would fail to
Finance and Accounting, World Bank Group, USA reject the null hypothesis if a URT would be applied on
Conflict of interest: The author has declared no conflicts of interest them. The first such process is Brownian motion, also
for this article. known as the Wiener process, which can be constructed

© 2017 Wiley Periodicals, Inc. 1 of 20

 Advanced Review [Link]/compstats

as a re-scaled random walk. More specifically, a re- An ARMA(p,q) model is defined in terms of its
scaled sum of mean-zero, variance-one random variables lagged values x(t) and its current and past innova-
εi as follows. Define r 2 [0, 1], the limit of the sum con- tions εt as:
verges in distribution to a Wiener process:
X
p X
q
xt = ϕi xt −i + θi εt −i + εt : ð4Þ
1 X
bnrc
lim pffiffiffi εi ) W ðrÞ: ð1Þ i=1 i=1
n! ∞ n D
i=1
It is commonly assumed that the innovations are an
One of the properties of Brownian motion is that independent identically distributed (i.i.d.) Gaussian
given 0 ≤ t ≤ s, W(s) − W(t) is independent of W(u) white noise series εt  N(0, σ 2), though this is not a
where u ≤ t and has a Gaussian distribution of N(0, requirement.
s − t). So we can easily see that if we were to use a Equation (4) can be rewritten as:
URT on Brownian motion it would fail to reject the
null since we would be testing Wt = ϕ1Wt − 1 + νt X
p X
q

where νt is random and ϕ1 = 1. The other commonly xt − ϕi xt − i = εt + θi εt − i ; ð5Þ

i=1 i=1
used continuous time process is geometric Brownian
motion, which does not allow negative values, and using the back-shift operator we can express
assuming the initial value Xt0 is positive, Equation (5) as polynomial functions:
where t0 = 0:
ϕðLÞxt = θðLÞεt ; ð6Þ
Xt = X0 e μt + σWt ,t 2 ½0, ∞Þ: ð2Þ
where ϕ(L) is the characteristic polynomial of the
If we consider a Taylor expansion of the exponential AR(p) part:
2 3
function e x = 1 + x + x2! + x3! + …, then we can rewrite
Equation (2) as: ϕðzÞ = 1 − ϕð1Þz −… − ϕðpÞz p : ð7Þ

Xt = X0 ð1 + μt + σWt + …Þ = X0 + X0 f ðWt ,μ, σ, tÞ: ð3Þ where xt is stationary only if the roots of ϕ(L) = 0 all
lie outside of the unit circle.
Given that f(Wt, μ, σ, t) is a random function, it is
very likely that if we perform a URT on xt the null
hypothesis of a unit root will fail to be rejected: Xt = Unit Roots
ϕ1X0 + νt where νt is random and ϕ1 = 1. A unit root is defined as an AR or ARMA process
that has 1 as a valid root of the characteristic polyno-
mial equation as in Chan2, p. 29. Time series with
unit roots are nonstationary processes. In the case of
STATIONARY TIME SERIES an AR(1) process if |ϕ1| = 1 there will be a unit root,
however the focus is usually on ϕ1 = 1; when
A weakly stationary time series process is defined ϕ1 = − 1, even though the variance is not constant
as having a constant mean, and an autocovariance the process will exhibit an oscillatory behavior as the
function γ(s, t) that depends on s and t only sign is reversed in every step, which we would argue
through their difference |s − t|. Throughout this is a less pathological case than when ϕ1 = 1. Con-
document a stationary time series is intended to sider the unit root xt = xt − 1 + εt; this is an integrated
mean a weakly stationary time series. A strictly sta- I(1) process which can be turned into a stationary
tionary time series is one where the joint probabil- process via first differences: Δ(xt) = εt. As detailed in
ity distribution does not change with time. A Chan.2 if AR(p) process has all of its characteristic
strictly stationary process is rarely observed in prac- polynomial roots with an absolute value greater than
tice. A nonstationary time series can have a time one, then such as process is defined to be causal, and
varying mean as well as a time varying covariance will also be stationary.
function.
A process that is integrated order n, I(n), is a
process that needs to be differenced n times to URT Hypotheses
become weakly stationary. A weakly stationary proc- URTs of a time series address the null hypothesis that
ess is referred to as an I(0) process. the series is unit root nonstationary; that is, the

2 of 20 © 2017 Wiley Periodicals, Inc.

 WIREs Computational Statistics Unit root tests

hypothesis the series is an integrated order 1 process, Alternative URT Hypothesis

I(1). The alternative hypothesis is that the time series Unless otherwise specified, the alternative hypothesis
is weakly stationary, or trend stationary in the case is assumed to be AR(1) with ϕ1 6¼ 1. When |ϕ1| < 1
we consider a trend under the null. However another this corresponds to a stationary I(0) process. Most of
possibility is that the process becomes explosive the standard URTs, such as the ADF test consider the
when |ϕ1| > 1. alternative hypothesis to be |ϕ1| < 1. However in the
This article focuses on AR(1) models, with inter- case with an explosive AR(1) process with |ϕ1| > 1
cept and linear trend. The Augmented Dickey-Fuller this is a highly nonstationary process. In this article
(ADF) URT handles ARMA(p,q) models by rewriting we will only consider the case when ϕ1 > 1. Testing
them as an AR(1) model with additional components for cointegration with AR(1) explosive series can
of differences; after essentially ‘taking out’ these addi- quite easily lead to a spurious cointegration result as
tional components in a helper regression it is then able will be shown using a simulation study later. This is
to use the critical values of the Dickey-Fuller that were why it is beneficial to have URTs with power against
designed with an AR(1) model. the explosive alternative hypothesis.
When allowing an intercept and a linear trend
Null URT Hypothesis in the time series being tested for unit roots the same
Unless otherwise noted, the null hypothesis is defined one-dimensional state-space model like representa-
as an AR(1) model xt = ϕ1xt − 1 + εt where ϕ1 = 1. tion as under the null is assumed:
This corresponds to a nonstationary I(1) process. This
is unusual in statistics where we would typically con- xt = ϕ1 xt − 1 + εt ,
sider the multiplier to be 0 under the null as in a t-test yt = β0 + β1 t + xt , ð11Þ
for the coefficient of a linear regression model. The rea- ϕ1 6¼ 1:
son ϕ1 = 1 is picked as the null is because many time
series in finance and economics are unit roots. If we allowed a drift and a trend in the AR formula
The case where the null hypothesis in the under the alternative (xt = ϕ1xt − 1 + β0 + β1t + εt),
assuming that x0 = 0, the resulting form would be
AR(1) model is extended to allow an intercept and a Xt
linear trend as follows will also be considered, in a yt = β 0 + β 1 t + ϕt − i εi , which will be stationary if
i=1 1
one-dimensional state-space model like representation: |ϕ1| < 1 and explosive if |ϕ1| > 1.
The case where the URT alternative hypothesis
xt = ϕ1 xt − 1 + εt , allows for structural breaks/changepoints in the
yt = β0 + β1 t + xt , ð8Þ deterministic coefficients β0 and β1 will be reviewed
ϕ1 = 1: later in this article.

Enhancing the URT null hypothesis to allow for

structural breaks/changepoints in the deterministic
Cointegration
coefficients β0 and β1 will also be discussed.
Linear relationships involving integrated nonstation-
If, instead of considering the intercept and trend
ary time series are meaningful only if the time series
in a state-space model framework as before, we had
are cointegrated. There are various definitions of
allowed the trend in the AR formula under the null,
cointegration; as originally proposed by Engle and
this would result in quadratic time trends in the
Granger.3 a cointegrating relationship between two
levels. Consider:
or more time series each having unit roots (I(1) I(1))
is defined to exist if there is a linear combination that
xt = ϕ1 xt − 1 + β0 + β1 t + εt ,
ð9Þ is stationary, i.e., I(0). Two time series xt and yt are
ϕ1 = 1:
cointegrated when there exists a number α1 in the lin-
ear equation yt = α1xt + νt such that νt is a stationary
Assuming that xt = 0, if we solve the recursion, using
process. This approach can be implemented in the
Faulhaber’s formula we would find the following
two variable case using the two-step Engle Granger
expression in the levels of x(t) with a quadratic time
procedure, where the first step consists of using least
trend component:
squares to derive a linear relationship between the
two variables, and the second step consists of using a
t ð t + 1Þ X t
URT on the residuals of the first step’s regression.
xt = β 0 t + β 1 + εi : ð10Þ
2 i=1 When there are more than two time series to check
for cointegration, there are multiple possible

© 2017 Wiley Periodicals, Inc. 3 of 20

 Advanced Review [Link]/compstats

cointegrating relationships, and the Engle Granger residual-based bootstrap and showed with Monte
two step methodology is not sufficiently flexible. In Carlo simulations that the bootstrap corrected size
this case the most common approach is to use the distortions of the asymptotic distribution in finite
Johansen cointegration test, which is based on the samples, and that it had greater power against the
estimation of a pth–order VAR in the k variables. threshold cointegration alternative than conventional
The VAR in the k–vector y is: cointegration tests.
Stigler6 posits that SETAR models are so com-
yt = Π1 yt −1 + Π2 yt −2 + … + Πp yt −p + Ψ Dt + εt : ð12Þ plex that their probability distribution and moments
are only known for simple cases. Estimation for
where Dt is a d –vector of deterministic terms, such more than one threshold is not easy. He points out
as a constant, time trend and seasonal dummies if that in a cointegration model with threshold effects
necessary. The estimation of the rank of the follow- there is no known procedure to test for stationarity
ing matrix Π will approximate the number of possi- with an unknown cointegrating vector. A test to
ble cointegrating relationships if any as follows: determine 2 versus 3 regimes only works in a
restricted case.
 
Π = − I − Π1 − Π2 − …− Πp : ð13Þ Bec et al.8 proposed the BBC URT which
considers a three-regime SETAR model. They
No assumption is made about the rank of Π. In allow a unit root in the middle regime in the mid-
the decomposition Π = αβ0 , α and β are k x dle regime under the alternative hypothesis, but
k matrices. The goal is to determine whether any not for the high and low regimes. Bec et al. used
columns of β0 are statistically indistinguishable the sup-Wald, sup-LM and sup-LR statistics to
from zero vectors. The existence of r cointegrating derive critical values.
vectors reduces the rank of Π by k − r: that is, if
there were r cointegrating relationships between NOTATION AND ABBREVIATIONS
the given variables, then there would be r nonzero
eigenvalues in the dynamic system, and k − r zero All of the simulations presented in this article were
eigenvalues. See Johansen, S. and Juselius, K.4 for developed using the R language. The R language as
details. The methodology is based on canonical well as many extension packages can be down-
correlation analysis. loaded from [Link] for various
computer architectures including Windows, MacOS
and Linux. The following symbols and abbrevia-
Threshold Cointegration tions are used:
Consider a general cointegration framework of two
time series yt and xt: • The number of time steps in a time series is
denoted by l
yt + αxt = zt : ð14Þ
• The number of simulations is indicated by m
5
Balke and Fomby proposed a general equilibrium • The significance level of the statistical test is
model of zt based on a self exciting threshold auto designated by α; it will be typically 0.05
regressive (SETAR) framework such as the following • The seed used for random number generation
with a low, middle and high regimes: in this article is 12345 for each simulation,
8 unless otherwise noted
< μh + ϕh zt −1 + εt , if zt −1 > θH , • Each simulation follows a specified data genera-
zt = μm + ϕm zt − 1 + εt , if θL ≤ zt −1 ≤ θH , ð15Þ tion process (DGP)
:
μl + ϕl zt − 1 + εt , if zt −1 < θL :
• χ ~ N(0, 1): A Gaussian distributed random
6 variable with mean 0 and variance 1
As Stigler points out the common assumed case
where Equation (15) is stable is when ϕh < 1 • χ ~ U(a, b): A uniformly distributed random
and ϕl < 1. variable between a and b where a < b
Seo7 developed a test for the linear no cointe-
gration null hypothesis against threshold cointegra-
IMPORTANCE OF URTS
tion in a threshold vector error correction model.
They used a sup-Wald type test and derived its null URTs are used to determine if time series are likely
asymptotic distribution. They proposed using a nonstationary or not; more specifically in the case of

4 of 20 © 2017 Wiley Periodicals, Inc.

 WIREs Computational Statistics Unit root tests

an AR(1) model, most test the null ϕ1 = 1 in xt = Spurious Cointegration

ϕ1xt − 1 + εt, and the alternative hypothesis is typi-
cally |ϕ1| < 1. However as we will see it is useful to xt = ϕ1a xt − 1 + εt ,
consider the alternative hypothesis to be ϕ1 6¼ 1 yt = ϕ1b yt − 1 + ζt ,
since not only when ϕ1 < 1 do spurious regressions x0 = y0 = 0, ð17Þ
occur, but they also can happen when ϕ1 > 1. We εt  i:i:d  N ð0,1Þ,
will also see that when ϕ1 > 1 many URTs will not ζ t  i:i:d  N ð0,1Þ:
reject the null of a unit root, and almost invariably
a cointegration test of independent variables with We use DGP (17), where xt and yt are independent,
ϕ1 > 1 would result in spurious cointegration to perform simulation tests of spurious cointegration
relationships. where the Johansen cointegration test is performed
on the two series, without considering if they are unit
roots or not.
We can see that in Table 2 for the case of
Spurious Regressions ϕ1a = 1 and ϕ1b = 1 the failure to reject the null hypoth-
Consider the following trivial DGP of linearly inde- esis of no-cointegration is close to the significance
pendent xt and yt series:
level. However for the case of ϕ1a > 1 and/or ϕ1b > 1
the failure to reject the null hypothesis of no-
xt = gxt −1 + εt ,
cointegration are quite high–providing evidence of
yt = hyt −1 + γ t ,
ð16Þ cointegration in 90–100% of the cases which is
x0 = y0 = 0,
entirely spurious and invalid since a valid cointegra-
εt  i:i:d:  N ð0,1Þ, γ t  i:i:d:  N ð0,1Þ:
tion relationship requires that the tested variables are
unit root processes. The same is true if ϕ1a < 0:9
Using the results of the DGP (16) the regression
Yt = βXt + νt is performed in a simulation with and/or ϕ1b < 0:9 but in these cases most of the URT’s
m = 1000 replications. Table 1 summarizes the would likely reject the null hypothesis of a unit root,
regression results, and we see for the median (0.5 so these cases would not be as troublesome as when
quantile) that in the cases of g = 1.01, h = 1.01 and ϕ1a > 1 and/or ϕ1b > 1 where the standard URT’s such
g = 1, h = 1, the regressions appear to be significant as ADF would not reject the null hypothesis of a
even though Yt and Xt are independent of each unit root.
other. Even in the case of g = 0.9, h = 0.9 with an
F-stat p-value of 0.04 leads us to reject the null
hypothesis of β = 0 at the 0.05 significance level. URT SURVEY
Granger9, p. 13, confirms similar findings: that spu-
The following is a survey of some of the most com-
rious regressions can happen for non I(1) processes
monly used URTs, as well as some others. This is not
(with an AR(1) ϕ1 multiplier of 0.9.)
an exhaustive review, as there are too many URTs to
cover in a single article. For more thorough reviews
see Refs 10–12.

TABLE 1 | Regressing yt on xt as Defined in Equation (16) with

l = 1000 and m = 1000 Dickey Fuller URT
2 The original simplest version of the Dickey Fuller test
Adj R Fstat p-*alue
g h Q:0.5 Q:0.5 considered the standard AR(1) model with Gaussian
white noise without any deterministic components;
1.01 1.01 1.0000 0.00
see Ref 13 for details:
1.00 1.00 0.4045 0.00
0.90 0.90 0.034 0.04 yt = ϕ1 yt − 1 + εt ;
0.00 1.00 −0.006 0.50
0.50 1.00 0.003 0.25 εt  i:i:d:  N ð0,1Þ;y0 = 0: ð18Þ
0.50 0.50 −0.002 0.38
0.20 0.20 −0.005 0.47
As detailed in Patterson14, p. 208 ,the two main test
statistics for the Dickey Fuller test are the normalized
0.00 0.00 −0.006 0.50
bias δ^ and the t-type statistic τ^:

© 2017 Wiley Periodicals, Inc. 5 of 20

 Advanced Review [Link]/compstats

TABLE 2 | Proportions of Failures to Reject the Null Hypothesis of X

T ð1
No Cointegration of Pairs of AR(1) Time Series with m = 5000, yt −1 εt =T ) σ 2
Wt dWt ; ð24Þ
D
l = 1000 and α = 0.05 as Defined in Model (17) t=1 0

Johansen X
T ð1
φ1a φ1b r=0 y2t − 1 =T 2 ) σ 2 Wt2 dr: ð25Þ
D 0
t=1
1.010 1.010 1.00
1.000 1.020 1.00
These expressions can be simulated to determine crit-
1.000 1.010 1.00 ical values for these test statistics.
1.000 1.005 0.94 Most standard URTs in the AR(1) framework,
1.000 1.000 0.05 such as the original Dickey-Fuller test, have a null
0.990 0.990 0.17 hypothesis of ϕ1 = 1 and the alternative hypothesis is
|ϕ1| < 1. The case of a negative unit root ϕ1 = − 1
1.000 0.960 0.97  
1.000 0.970 0.80 leads to a distribution of ϕ^1 + 1 that is the mirror
image of the standard Dickey-Fuller distribution as
1.000 0.980 0.42
detailed in Ref 10, p. 55. The situation when |ϕ1| > 1
1.000 0.990 0.12
this leads to an explosive case. As derived in Phillips
1.000 0.900 1.00 and Magadalinos15 the asymptotics of the estimation
1.000 0.200 1.00 error of ϕ1 in this case under certain assumptions is
0.900 0.900 1.00 Cauchy distributed. This distribution is problematic
0.500 0.500 1.00 as it has no finite moments of order greater than or
equal to one.
0.200 0.200 1.00
We can extend the DGP (18) to have a con-
stant and a linear trend; Using OLS these extra
 
δ^ = T ϕ^1 − 1 ; ð19Þ terms can be removed. This will affect the critical
  values of the test statistics. Zivot16 suggests a simu-
ϕ^1 − 1 lation approach based in R. This approach was
τ^ = ; ð20Þ extended for the trend and drift terms for the nor-
σ^ϕ^
malized bias (NB) and t-type (DF) test statistics; the
R code to perform these simulations is detailed in
where ϕ^1 is the standard coefficient estimate as
Listing 1.
derived using ordinary least squares:
Results of running the simulation code from
XT Listing 1 and comparisons with the critical values
yt yt −1 from the R qunitroot() function from the fUnitRoots
ϕ^1 = Xt =T1 ; ð21Þ R package as reported by MacKinnon.17 are detailed
y2
t = 1 t−1 in Listing 2:
XT XT We see in Listing 2 that the simulated values
y ðy − yt − 1 Þ
t = 1 t−1 t
yt − 1 εt
ϕ^1 − 1 = XT = Xt =T 1 : ð22Þ and the ones computed by MacKinnon17 as reported
y2 y2 in the qunitroot() function of the fUnitRoots R pack-
t = 1 t −1 t = 1 t−1
age are relatively close.
The last step in Equation (22) is based on Said and Dickey18 extended the Dickey
εt = yt − yt − 1. Fuller URT for ARMA models and not just
Given these previous equations we can now AR(1) models; this is known as the Augmented
define the normalized bias as a function of the obser- Dickey Fuller (ADF) URTs and is one of the
vations yt: most commonly used URTs in the literature Ref
10, p. 33. The ADF test regression is fitted
XT XT using OLS:
yt − 1 εt yt − 1 εt =T
δ^ = T Xt =T 1 = XtT= 1 : ð23Þ
2 2 =T 2 X
n
t=1
y t −1 t=1
y t −1 Δyt = α + δt + βyt − 1 + γ i Δyt − i + εt ; ð26Þ
i=1
As detailed in Table 7.1 of Ref 14, p. 197 for the
DGP (18) following sample quantities have a limiting where Δ is the difference operator and εt represent 0-
form expressed as functionals of Brownian motion: mean white-noise innovations. Under the null

6 of 20 © 2017 Wiley Periodicals, Inc.

 WIREs Computational Statistics Unit root tests

Listing 1
NB and DF Test Statistic Simulations

hypothesis yt is considered to be I(1) which is equiva- t-ratio of γ i is still significant. For the ADF test
lent to Δyt being I(0) in which case β would be zero. details see Ref 18.
The test statistic is the standard regression t-statistic
^
tβ = s:e:β β^ . A normalized bias test statistic can be used
ðÞ
as well. Elliott, Rothenberg and Stock URTs
The tests’ authors derived the critical values See Ref 20, p. 93, for a detailed overview of the
for these test statistics as they follow a nonstandard Elliott, Rothenberg and Stock (ERS) URTs; the fol-
distribution; they are the same as the standard lowing overview is taken from there. To increase the
Dickey-Fuller critical values which depend on the power of a URT under the null hypothesis of a unit
form of the deterministic components. The lagged root, ERS [1996] proposed a local to unity detrend-
differences allow correcting for serial correlation in ing of the time series. The assumed generating proc-
the innovations. The ADF URT is sensitive to the ess for the series yt is as follows:
choice of the number of lags n. Ng and Perron19
0
proposed an iterative method to determine the lag yt = dt + ut ut = aut − 1 + vt dt = β^ zt ; ð27Þ
length based on choosing the largest lag where the

© 2017 Wiley Periodicals, Inc. 7 of 20

 Advanced Review [Link]/compstats

Listing 2
Critical Values of Simulated DF Statistics

where zt is a deterministic (q × 1) vector and νt is a And a = 1 + c=T where c is a constant, set to −7 in the
stationary zero-mean process. If a = 1 the yt is I(1), case of a constant or −13.5 in the case of a linear
but if |a| < 1 then yt is I(0). trend. The second test type is denoted as the DF-GLS
The authors developed feasible point-optimal test, which is a modified ADF-type test applied to the
tests, which take serial correlation of the error term detrended data without the intercept, which is the t -
into account. The feasible point-optimal test statistic statistic for testing α0 = 0 in the regression:
is defined as:
Δytd = α0 ytd− 1 + α1 Δytd−1 + … + αp Δytd− p + εt ; ð31Þ
Sða = aÞ−aSða = 1Þ
PT = ; ð28Þ
ω^2 ’
where ytd are residuals fromyt − β^ zt
where S(a = a)  and S(a = 1) are the sums of squared Both of these tests are available in the urca pack-
residuals from a least-squares regression of ya on Za age in the [Link]() function. Later in this article we will
with ya = (y1, y2 − ay1,..., yT − ayT − 1) and Za = (z1, see in simulation experiments that these ERS tests
z2 − az1,..., zT − azT − 1) and ya is a T-dimensional become invalid when the initial value y0 is not 0, and
vector and Za is a (T × q) matrix. The estimator for also when the AR(1) ϕ1 multipliers are negative.
the variance of the error process νt is:
Phillips-Perron URT
σ^2
ω^ =
Xν 2 ; ð29Þ Phillips and Perron21 developed a URT that corrects
p
1− i=1
^
αi for serial correlation and heteroskedasticity of the
innovations. Pfaff20, p. 95, reviews the Phillips-Perron
(PP) procedure details. The PP test regression is fitted
where σ^2ν and α^i for i = 1,..., p are from the OLS
using OLS. The first case is with out a linear-trend:
regression:
yt = μ + αyt − 1 + εt ; ð32Þ
Δyt = α0 + α1 Δyt − 1 + … + αp + 1 Δyt − p + νt ð30Þ

8 of 20 © 2017 Wiley Periodicals, Inc.

 WIREs Computational Statistics Unit root tests

λ^ of a time series with the lagged time series using the

Zðα^Þ = T ðα−
^ 1Þ− ; ð33Þ Johansen cointegration test. If the time-series is a unit
myy
root, I(1), we expect only one valid cointegrating rela-
0
s^t α^ λ^ σ^Tl tionship but if there are two cointegrating relation-
Zðτα^Þ = − 1 ð34Þ ships the process is I(0). Given two variables, only
σ^Tl m2 ;
yy
one valid cointegration relationship is possible. For
0 reasonable lags it would be almost impossible for the
  s^t μ^ λ^ σ^Tl my
Z τμ^ = − ð35Þ two time series to fail to reject the null hypothesis of
σ^Tl m12 m12 ;
yy yy no-cointegration since the series and its lagged version
X X  will always be fairly close to each other. The lagged-
with myy = T −2 ðyt −yÞ2 , myy = T −2 y2t , my = series URT is conducted as follows:
P  
T− 3/2 (yt), and λ^ = 0:5 σ^2Tl − s^2 , where s^2 is the sam-
0 • Decide if a constant and/or linear trend is
ple variance of the residuals, λ^ = λ= ^ σ^2 , and tα^, tμ^ are
Tl allowed under the null hypothesis or not.
the student t ratios of α^ μ,
^ respectively. The long run
• In the case a constant and/or linear trend is
variance σ^2Tl is estimated as follows:
allowed under the null hypothesis follow these
X
T X
l X
T steps:
σ^2Tl = T − 1 ε^2t + 2T − 1 wsl ε^t ε^t − s ; ð36Þ • Use a helper regression to estimate and
t=1 s=1 t=s+1
remove these deterministic elements α and β
in xt = yt +α + βt where the underlying AR
where wsl = 1 − s/(l + 1).
process is defined as yt = ϕ1yt − 1 + εt. The
The case of adding a linear trend is now
helper regression to be performed is:
considered:
T Δxt = ϕ1a xt − 1 + C + Dt + εt : ð42Þ
yt = μ + βðt − Þ + αyt − 1 + εt : ð37Þ
2
• The coefficients for the nondifferenced rela-
In this case the test statistics used are:
tionship can be derived as
λ^
Zðα^Þ = T ðα−
^ 1Þ− ; ð38Þ
M ϕ1 = ϕ1a + 1; β = D=ð1− ϕ1 Þ; α
0
s^t α^ λ^ σ^Tl = ðC − βϕ1 Þ=ð1 −ϕ1 Þ: ð43Þ
Zðτα^Þ = − ; ð39Þ
σ^Tl M12
0
  s^t μ^ λ^ σ^Tl my • Remove the deterministic elements as
Z τμ^ = −
12 ; ð40Þ follows:
σTl
^ 1
M2 M + m2y
xt = xt − ðα + βÞ −βt: ð44Þ

 s^t ^ λ^0 σ^ 0:5m − m 
β Tl y ty
Z τβ^ = −  M 12 12 ; ð41Þ • In the case where no constant and no trend is
σTl
^ m yy
12 allowed under the null hypothesis the underly-
ing AR process to be tested is defined as
where my, myy , λ,^ λ^0 and σ^Tl are as defined previously xt = ϕ1xt − 1 + εt
P
and mty = T− 5/2 tyt and tμ^, tβ^, t α^ are the student t • Run the Johansen cointegration test of the time
^ α^ respectively. The scalar M is defined 
ratios of μ,
^ β,    series x*t = fxk + 1 , .. ., xn g and the k-lagged
as M = 1− T − 2 myy − 12m2ty + 12 1 + T − 1 mty my − time series {Lk(xt)} = {x1, x2,..., xn − k} for a
 −1 −2
 2
4 + 6T + 2T my . The critical values used are the small lag, such as k = 3.
same as for the Dickey Fuller URT. • The Johansen test computes a primary and a
secondary test statistic when run with two vari-
ables; the first test statistic is for the null
Lagged-Series URT hypothesis of a cointegration rank of 0. The
Herranz22 proposed a new URT for AR(1) time second test statistic is for a null-hypothesis of a
series referred to as the lagged-series test which cointegration rank of less than or equal to
checks for the number of cointegrating relationships 1. Examine the second cointegration test

© 2017 Wiley Periodicals, Inc. 9 of 20

 Advanced Review [Link]/compstats

statistic of a rank of less than or equal to 1. If where rt is a random  walk and the innovations are
this test statistic is less than or equal to the criti- assumed to be i.i.d. 0, σ 2u . r0 is assumed to be a con-
cal value for the specified significance level α stant level. If ξ < > 0 then the model will contain a
then do not reject the null of I(1). Otherwise, if time trend component in addition to the level. Under
the statistic is greater than the critical value the null hypothesis yt is assumed to be I(0) so that
reject the null of I(1) and accept the alternative σ 2u = 0 and rt = r0 = constant.
of I(0) or an explosive process (|ϕ1| > 1). The test statistic is derived as follows. First yt is
• The Johansen cointegration test has different regressed on a constant or a constant plus a linear trend.
critical values depending on whether the esti- The partial sums St of the residuals ε^t are computed:
mated relationship is assumed to have a con-
X
t
stant, a linear trend or is driftless. We use the St = ε^i , t = 1, .. ., T : ð47Þ
Johansen critical values corresponding to the i=1
driftless case when there are no deterministic
elements. If we allowed for a constant and/or a The test statistic LM is computed as:
trend then the Johansen critical values corre- XT
sponding to the trend case multiplied by 1.04 t=1 t
S2
are used. These critical values were determined LM = ; ð48Þ
σ^ε 2
via various simulation studies, under various
scenarios and time series lengths. where σ^ε 2 is an estimate of the error variance of the
residuals. The authors suggest using a Bartlett win-
The Zivot Andrews (ZA) URT, or many other URTs dow w(s, l) = 1 − s/(l + 1) as an optimal weighting
such as ADF, can be combined with the lagged-series function for estimating the long run variance of the
URT as detailed in Algorithm 1 to increase the statis- errors:
tical power of the ZA test in the case where
AR(1) process is explosive, that is |ϕ1| > 1. The deter- X X
T l
s X T
ministic component can incorporate either an inter- σ^ε 2 = T −1 ε^t 2 + 2T − 1 + ε^t ε^t − 1 : ð49Þ
cept, time trend, or both. The value of C was t=1 s=1
l + 1t =s+1
determined via simulation studies.
The critical values of the test statistic are given in
Algorithm 1 ZA-lagged-series URT Ref 23.
C 20.09299
x0 LaggedSeriesUnitRootTest( tsx, lag, cvalLevel)@Statistic Zivot Andrews URT
x1 [Link](tsx,model=‘trend’)@Statistic Andrews and Zivot24 developed a URT (ZA) that
if x1 > CriticalValue and x0 < C then can handle structural breaks. The test statistic of the
return( I(1) ) /* Unit Root Nonstationary */
ZA test is the Student t ratio just as was the case with
else
return( I(0) or Explosive) /* Stationary or |ϕ1| > 1 */
the PP test. As detailed in Ref 20, p. 110:
h i i
end if
tα^i λ^inf = inf tα^i ðλÞ for i = A,B,C; ð50Þ
λ2Δ

where Δ is a closed subset of (0, 1). Depending on

Kwiatkowski Phillips Schmidt Shin (KPSS) the model, the test statistic is inferred from one of
Stationarity Test these three regression Models:
The KPSS test proposed by Ref 23, has the null   X
k
hypothesis of stationarity around a deterministic yt = μ^A + θ^A DUt λ^ + β^A t + α^A tyt −1 + c^A ^t ;
i Δyt −i + ε
trend and the alternative of a unit root. The follow- i=1
ing definitions and explanations of the KPSS test are ð51Þ
taken from Ref 20, p. 103. The underlying model
considered in the KPSS test is:   X
k
yt = μ^B + γ^B DTt* λ^ + β^B t + α^B tyt −1 + c^Bi Δyt − i + ε^t ;
i=1
yt = ξt + rt + εt ; ð45Þ
ð52Þ
rt = rt − 1 + ut ; ð46Þ

10 of 20 © 2017 Wiley Periodicals, Inc.

 WIREs Computational Statistics Unit root tests

   
yt = μ^C + θ^C DUt λ^ + β^C t + α^C tyt − 1 + γ^B DTt* λ^ Time Series with I(0) and I(1) Segments
Kim28 showed that standard URTs are not consistent
X
k
c^C ^t ; against processes displaying a shift from stationarity
+ i Δyt −i + ε ð53Þ
i=1 (I(0)) to nonstationarity (I(1)) and vice versa. There-
fore new methods are required for differentiating
where DUt(λ) = 1 if t > Tλ and 0 otherwise, and between processes that are I(0) or I(1) for the entire
DTt* ðλÞ = t −λT for t > Tλ and 0 otherwise. The null period v.s. those with a shift from I(0) to I(1) or vice
hypothesis of the ZA URT does not allow structural versa.
breaks. Only breaks are allowed in the alternative Kejriwal et al.29 proposed hybrid testing proce-
hypothesis. Glynn et al.25 criticize this since if there dures that allows ruling out of stable stationary pro-
are breaks under the null (ϕ1 = 1) we can mistakenly cesses or ones that are subject to only stationary
conclude that the series is stationary (with breaks). changes under the null, helping the researcher in
The critical values are derived via simulation of the t- interpreting a rejection as emanating from a switch
test statistic under the null, and in the asymptotic between an I(1) and an I(0) regime. The authors use
case can be derived as a functional of Brownian a combination of their own developed URTs together
motions. with a test with a null of no breaks versus an alterna-
tive with one or more breaks given by Bai and Per-
ron.26 The calculation of the test statistics and the
asymptotic critical values are done using the dynamic
Removing Structural Changes Prior to Unit programming algorithm proposed in Perron
Root Testing and Qu.30
Kejriwal et al.29 used the approach of detrend-
Estimating Change Points ing the data when trends are included prior to using
The Bai and Perron26 procedure for estimating struc- their hybrid testing procedure and they suggest using
tural breaks/change points in a linear model is imple- a sequential procedure developed by Kejriwal et al.29
mented in the breakpoints() function of the that is robust to whether the errors are I(1) or I(0).
strucchange R package (Ref 27, p. 12). The smallest time series length the authors consider
This methodology assumes an underlying linear is 150.
model with a dependent one dimensional variable yt, Herranz22 proposed a testing procedure
a p × 1 covariates vector xt with corresponding coef- referred to as the Hybrid Bai-Perron-ADF procedure
ficient vector β, and the innovations εt: whereby given a time series to consider if the series
follows one of the following four possibilities:
yt = xt β + εt . (54)
• The entire series is stationary (I(0))
If there are m change points in the coefficient
this implies m + 1 regimes. Equation (54) can be • The entire series is nonstationary (I(1))
rewritten as: • The series consists of a stationary segment (I(0))
  followed by a nonstationary segment (I(1))
yt = xt βj + εt j = Tj − 1 + 1, .. ., Tj ,j = 1, .. ., m + 1 : ð55Þ • The series consists of a nonstationary segment
(I(1)) followed by a stationary segment (I(0))
The underlying idea is solving the problem by divid-
ing it into independent optimally solvable sub-pro- The proposed approach is to use the Bai and Per-
blems, whose solutions can be combined to solve the ron26 methodology of estimating structural break
larger problem. In this case a triangular residual sum date/s based on finding the model specification that
of squares (RSS) matrix is computed and stored in minimizes the RSS via a dynamic programming
memory which can be reused over and over again to approach, and then to use the ADF URT to test each
derive the RSS for a segment starting at observation section to determine if it is likely I(0) or I(1). To do
t and ending at t0 with t < t0 . this we use the following regression Model (56) with
This approach is considerably faster than the the Bai and Perron26 procedure which will estimate
brute force approach of computing the RSS for all pos- structural breaks in the coefficients β0, β1, β2:
sible sub-segments. The Bai and Perron26 algorithm
uses only O(T2) least squares operations for a number xðt Þ = β0 + β1 t + β2 xðt − 1Þ + εt : ð56Þ
m of change points. The brute force approach would
require O(Tm) least squares operations.

© 2017 Wiley Periodicals, Inc. 11 of 20

 Advanced Review [Link]/compstats

Listing 3
ADF URTs on IBM Segments

If the Bai and Perron26 structural break meth- We see in Listing 3 that the null hypothesis of a
odology does not find a break date, then we run the unit root cannot be rejected for any of the three seg-
ADF test on the entire time series–and both segments ments, which leads us to conclude they are likely
would have the same result. nonstationary: I(1),I(1),I(1). This finding is consistent
If we assume that if any structural changes in with the common assumption that asset prices follow
linear trend and/or intercept happen they do so at the a random walk. Even though breakpoints are esti-
same time as a structural break in mated using the Bai and Perron31 procedure the ADF
AR(1) autoregressive multiplier ϕ1 then the ADF and tests show that it is likely the AR(1) multiplier ϕ1 is
the Bai and Perron26 methodology were highly accu- 1, even if the drift has a changepoint. Visually we see
rate, then the ADF test would fit the correct linear that there is a level-shift between the first and second
trend and intercept in each segment; in any case segments, and there is likely a linear trend starting in
experiments show that the Bai and Perron26 the third component.
approach can handle changes in β2 across both I(0) I
(0) and I(1) regimes, as well as when there are
changes in the other coefficients.
Figure 1 shows in red vertical bars the esti-
URTS ROBUST TO STRUCTURAL
mated break-points using this procedure with Model CHANGES UNDER THE NULL
(56) on historical IBM prices. When the ADF URT Perron (1989) URT
was run on each of the three segments, the null
Perron32 proposed URTs for processes with a linear
hypothesis of a unit root was never rejected.
trend with a break in intercept with the unit root as a
null hypothesis, and the break model without the
IBM. close unit root as an alternative hypothesis:
200
H0 : yt = yt − 1 + μ1 + ðμ2 −μ1 ÞIft > TB g + βt + et ,
180 H1 : yt = μ1 + ðμ2 −μ1 ÞIft > TB g + βt + et ; ð57Þ

160 where the innovations et are stationary ARMA

processes:

140  
AðLÞet = BðLÞvt ; vt i:i:d 0, σ 2 ; ð58Þ

120 with A(L) and B(L) pth and qth order polynomials, so
the innovation series {et} is taken to be an ARMA
(p,q) process.
Mar 04 Oct 01 Apr 01 Oct 01
2014 2014 2015 2015
And the AR(1) multiplier of yt − 1 is estimated
using OLS (ϕ^1 ). As presented by Ref 10, p. 60 ϕ^1 is
F I G U R E 1 | Estimated changepoints in IBM prices. asymptotically biased towards 1, but is consistent for

12 of 20 © 2017 Wiley Periodicals, Inc.

 WIREs Computational Statistics Unit root tests

Model (57), which makes this test have low statistical for auto-correlated innovations is made by adding
power. lagged terms ΔeSt − j where j = 1,..., k. The location of
Perron32 also considers tests for a process with the break TB is determined by searching across all
breaks in intercept in trend and intercept: breakpoints and picking the one with the most nega-
tive eτ:
H0 : yt = yt − 1 + μ1 + ðμ2 − μ1 ÞIft > TB g

+ ðt − TB Þðβ2 − β1 ÞIft > TB g + β1 t + et , inf eτðe

λÞ = inf eτðλÞ where λ = TB =T : ð63Þ
λ

H1 : yt = μ1 + ðμ2 − μ1 ÞIft > TB g

If the DGP is that in (60), and if the εt satisfy certain
+ ðt − TB Þðβ2 − β1 ÞIft > TB g + β1 t + et : ð59Þ regularity conditions, and TB/T ! λ as T ! ∞, then
under the null hypothesis of β = 1
p
Choi10, p. 60 explains that ϕ^1 ! 1 for Model (58) ð1 − 1=2
1
implying that the URTs for this model are inconsist- inf eτðeλÞ = inf − VðrÞ2 dr ; ð64Þ
ent: a stationary process with a breaking trend can- λ 2 0
not be distinguished from a unit root process with
drift. where VðrÞ represents a demeaned Brownian bridge.

Hybrid Bai-Perron-Zivot-Andrews URT

Lee–Strazicich URTs Herranz22 proposed a new URT that allows struc-
The one-break Lee and Strazicich33 URT procedure tural breaks under the null hypothesis, which we
and the two-break Lee and Strazicich34 URT allow refer to here as the Hybrid Bai-Perron-Zivot-
for structural breaks to be determined endogenously Andrews (HBPZA) URT, and consists of the follow-
from the data and breaks are allowed under both the ing steps:
null and the alternative hypothesis. Consider the sin-
gle break model: 1. Given a time series, first use the Bai-Perron
break-point estimation procedure in Ref 31
yt = δZt , Xt = βXt −1 + εt ; ð60Þ using regression Model (65) to detect changes in
the coefficients. This divides the time series into
where Zt contains exogenous variables. The null k + 1 segments given a total of k breakpoints.
hypothesis is specified by β = 1. 2. For each segment within the time series com-
‘Model A’ is the crash model that allows for a pute the Zivot Andrews (ZA) URT statistic.
single change intercept where Zt = [1, t, Dt]0 where
3. A final test statistic is computed by weighing
Dt = 1 for t ≥ TB + 1, and zero otherwise. TB is the
each sub-test statistic by the segment length.
time of the structural break and δ = (δ1, δ2, δ3)
‘Model C’ allows for a shift in intercept and 4. If there are more than three breakpoints esti-
change in trend slope under the alternative hypothesis mated, only the first three breakpoints are used
where Zt = [1, t, Dt, DTt]0 where DTt = t −TB for t ≥ to determine four segments.
TB + 1, and zero otherwise. TB is the time of the struc-
tural break and δ = (δ1, δ2, δ3). The regression Model (65) is used to detect
The URT statistics are derived from the breakpoints in the coefficients β0, β1, and β2 using
regression: the Bai and Perron31 procedure:

Δyt = δ0 ΔZt + ϕe
St −1 + ut ; ð61Þ xt = β0 + β1 t + β2 xt − 1 + εt : ð65Þ

e
St = yt − Ψ x − Zte
δ,t = 2, …, T; ð62Þ DGP (65) was used to simulate the time series
under the null hypothesis (H0) of a unit root with
where e δ is the coefficients estimated in the regression 1 structural break at time TB to derive the critical
of Δyt on ΔZt and Ψ e x is the restricted MLE of Ψ x = values of the test statistic (via simulation):
Ψ + X0 given by y1 − Z1e δ. The unit root null hypothe-
sis consists of ϕ = 0 and the LM(Lagrange multiplier) yt = ϕ1 yt − 1 + μ1 + ðμ2 −μ1 ÞIft > TB g
t-test statistic eτ = t-statistic testing the null hypothesis + ðt − TB Þðβ2 −β11 ÞIft > TB g + βt + et ,
ϕ = 0. As in the case with the ADF test, a correction

© 2017 Wiley Periodicals, Inc. 13 of 20

 Advanced Review [Link]/compstats

TABLE 3 | Hybrid Bai-Perron-Zivot-Andrews Unit Root Critical

Values from Simulations Box 1

Breaks α = 0.01 α = 0.05 α = 0.1

UNIT ROOTS IN VARIOUS FIELDS
−6.069573 −4.631409 −4.336683
−4.837634 −4.376655 −4.178676 As previously mentioned in the introduction,
many financial time series such as asset prices
are modeled as unit root processes. If you pick
any time series of stock prices, or interest rates,
μ1  Uð − 10;10Þ;μ2  Uð − 10;10Þ;y0  N ð0,1Þ under normal conditions, and run a URT on
them you will likely not be able to reject the
β1  Uð − 3; 3Þ;β2  Uð − 3;3Þ, null hypothesis of a unit root.
The bulk of the literature on unit roots and
H0 : ϕ1 = 1;HA : ϕ1 6¼ 1: ð66Þ cointegration spans mostly the social sciences,
largely focusing on economic and financial
This approach can be extended to two or more applications. However, there are other fields
breakpoints. The HBPZA critical values for 0 and where they may be proving to be useful tools.
1 breakpoints under the null hypothesis of a unit Kipinski et al.37 used Phillips-Perron URTs as
root derived from simulations are detailed in well as other statistical tests to investigate the
Table 3. stationarity of magnetoencephalography (MEG)
or electroencephalography (EEG) time series, and
they found they are largely stationary. However,
Koruek and Ozkaya38 analyzed EEG time series
before, during and after seizures, and they deter-
URTS AND MORE AVAILABLE IN THE mined that short interictal (between seizures)
R LANGUAGE EEG series were nonstationary and could be mod-
Table 4 details URTs available as functions in eled as an ARIMA unit root process.
There have been various studies on testing
R programming language packages; most of them
worldwide zonal temperature anomalies for
were used in simulations detailed in the next section.
unit roots under various assumptions of struc-
Note that the URTs in the fUnitRoots package35 are
tural changes as in Refs 39–41. A three regime
the same as those available in the urca package.20 SETAR model of the Northern Hemisphere tem-
The Lee–Strazicich (LS) URT function ls() used perature anomalies can provide a better model
in this article was written by Raphael Rocha Gouvea that incorporating changepoints than breaks in
and Thais Helena Fernandes Teixeira and was an AR framework. This is the case described in
obtained from the open knowledge repository of the Whyte and Metcalfe42 where they propose a
Institute for Applied Economic Research (Ipea)—a model similar to SETAR which they refer to as
federal public foundation linked to the Secretariat of MTAR using a reconstructed Burgundy temper-
Strategic Affairs of the Presidency of the Republic of ature series of Chuine et al.43 from 1370 to
Brazil.a 1977. The constants are negative but the linear
The [Link]() function of the R forecast trends are positive indicating a diverging
package can be used to fit the optimal ARIMA() relationship further exacerbated by an
model to data. nonstationary(explosive) middle regime.
Three-regime SETAR models of time series can Sobradelo et al.44 used a Dickey-Fuller URT
be defined and fitted using the setar() function of the to provide evidence that a time series of vol-
tsDyn R package.36 For known thresholds the canic eruptions is not stationary. Ramana
SETAR model can be fit using conditional least et al.45 used ADF URTs to show that water
squares; the setar() function performs a grid search levels in reservoirs are likely nonstationary,
to find the thresholds that minimize the RMSE. prior to testing for cointegration.
The Johansen cointegration test used in this Vargas and Kastle46 assumed that geological
article is the [Link]() function of the urca R package magnetic time series are unit roots to consider
a cointegration argument. Oluwole47 used a
by Ref 20. The Bai and Perron26 methodology to
URT to determine the likelihood that a time
determine break dates in time series used in this arti-
series of outbreaks of meningitis was station-
cle is the breakpoints() function of the R strucchange
ary. Skrastins et al.48 used ADF URTs to check
package. for the stationarity of web traffic time series.

14 of 20 © 2017 Wiley Periodicals, Inc.

 WIREs Computational Statistics Unit root tests

TABLE 4 | Unit Root Tests in R

Constant and
Package Function Name Unit Root Test Constant Linear Trend Linear Trend Driftless
tseries [Link] Augmented Dickey Fuller N.A. N.A. Only choice N.A.
fUnitRoots urdfTest Augmented Dickey Fuller ‘c’ N.A. ‘ct’ ‘nc’
fUnitRoots urersTest Elliott Rothenberg Stock ‘c’ N.A. ‘ct’ N.A.
fUnitRoots urkpssTest KPSS for stationarity ‘mu’ N.A. ‘tau’ N.A.
fUnitRoots urppTest Phillips-Perron ‘constant’ ‘trend’ N.A. N.A.
fUnitRoots urspTest Schmidt Phillips N.A. N.A. N.A. N.A.
fUnitRoots urzaTest Zivot Andrews ‘intercept’ ‘trend’ ‘both’ N.A.
tsDyn BBCTest BBC unit root test against a stationary Only choice N.A. N.A. N.A.
3 regime SETAR alternative

URT SIMULATIONS identically distributed Gaussian innovations for vari-

ous values of ϕ1.
URTs on AR(1) with Intercept and a Linear The results can be summarized as follows:
Trend
Table 5 summarizes the results of simulations per- • Data simulated under the alternative hypothesis
formed with DGP (67) with initial value x0 = 20. where ϕ1 > 1: The ERS, ADF and ZA tests do
This process consists of AR(1) processes with a deter- not significantly reject the null hypothesis of
ministic intercept and a linear trend with independent even though ideally they should in this case.
However the two lagged-series based tests reject
TABLE 5 | Proportions of Failures to Reject the Unit Root Null 97% of the time for ϕ1 = 1.04.
Hypothesis of URTs on AR(1) Processes with Intercept U(−1,1) and • Data simulated under the null hypothesis:
Linear Trend U(−1,1) with l = 100, m = 5000, and α = 5 % Hypothesis where ϕ1 = 1: The URTs reject the
and x0 = 20 null hypothesis as expected close to the signifi-
ZA- cance level of 5%: ADF at 4%, and lagged-
Lagged- ERS- ERS- lagged- series and ERS-DFGLS , ZA tests at 5% and
φ1 ADF series Ptest DFGLS series ZA the ERS-Ptest is invalid.
1.1000 1.00 0.03 1.00 1.00 0.21 1.00 • Data simulated under the alternative hypothesis
1.0800 1.00 0.02 1.00 1.00 0.18 1.00 where ϕ1 < 1: The null hypothesis is rejected at
1.0600 1.00 0.02 1.00 0.93 0.18 1.00 a rate of 10% or less as follows from best to
worst:
1.0400 1.00 0.03 0.88 0.98 0.20 1.00
• ADF and lagged-series, ZA and ZA-lagged-
1.0200 0.99 0.74 0.80 0.99 0.91 0.97
series tests: ϕ1 ≤ 0.5
1.0100 0.96 0.94 0.36 0.96 0.96 0.94
• The ERS tests hardly ever reject the null
1.0050 0.96 0.95 0.34 0.95 0.96 0.94 hypothesis in this case. In other tests
1.0025 0.96 0.95 0.35 0.95 0.95 0.94 when the initial value of x0 = 0 the statis-
1.0000 0.96 0.95 0.34 0.95 0.96 0.95 tical power is similar to that of the
0.9950 0.96 0.95 0.34 0.95 0.95 0.95 ADF URT.
0.9800 0.95 0.94 0.34 0.95 0.95 0.94
yt = ϕ1 xt + α + βt,
0.9600 0.94 0.92 0.37 0.96 0.94 0.93
xt = xt − 1 + εt ,
0.9000 0.81 0.75 0.48 1.00 0.81 0.87 x0 = 0;εt  i:i:d:  N ð0,1Þ, ð66Þ
0.7000 0.24 0.18 0.56 1.00 0.24 0.13 t = 1, .. ., l,
0.5000 0.08 0.02 0.58 1.00 0.07 0.00 α  U ð − 1;1Þ,β  Uð − 1; 1Þ:
0.3000 0.04 0.00 0.60 1.00 0.01 0.00
0.1000 0.03 0.00 0.61 1.00 0.00 0.00 Table 6 shows simulations of Model (67) using nega-
tive values of ϕ1. The results for data simulated
0.0000 0.02 0.00 0.61 1.00 0.00 0.00
under the alternative hypothesis where ϕ1 < 0 can be

© 2017 Wiley Periodicals, Inc. 15 of 20

 Advanced Review [Link]/compstats

TABLE 6 | Proportions of Failures to Reject the Null Hypothesis of a processes with one structural break in intercept and
Unit Root of URTs on Driftless AR(1) Processes with φ1 < 0 and one structural break in linear trend with independent
l = 1000, m = 1000, and α = 5 % and x0 = 20 identically distributed Gaussian innovations for vari-
ZA- ous values of ϕ1. The break time is determined at
Lagged- ERS- ERS- lagged- random; it is uniformly distributed, and the same
φ1 ADF series Ptest DFGLS series ZA break time is used for the intercept change and the
−1.0000 0.01 0.00 0.99 0.99 0.00 0.00 linear trend change. It is tested with two different sets
of intercept and trend levels.
−0.9900 0.01 0.00 0.98 0.97 0.00 0.00
−0.9500 0.01 0.00 0.87 0.97 0.00 0.00 yt = xt + α1 + ðα2 −α1 ÞIft > tu g + β1 t
−0.9000 0.01 0.00 0.76 0.98 0.00 0.00
+ ðβ2 − β1 Þðt − tu ÞIft > tu g,
−0.7000 0.01 0.00 0.66 0.99 0.00 0.00
−0.5000 0.01 0.00 0.63 1.00 0.00 0.00 xt = ϕ1 xt −1 + εt ;εt  i:i:d:  N ð0,1Þ,
−0.3000 0.01 0.00 0.62 1.00 0.00 0.00 x0  N ð0,1Þ,
−0.1000 0.02 0.00 0.61 1.00 0.00 0.00
tu  Uð3, l − 2Þ;t = 1, .. ., l: ð68Þ
0.0000 0.02 0.00 0.61 1.00 0.00 0.00
1.0000 0.96 0.95 0.34 0.95 0.96 0.95
The results can be summarized as follows. For
DGP (68) the intercepts and trends are picked uni-
summarized as follows: The null hypothesis is formly random for each simulation as follows:
rejected as follows from best to worst: α1, α2, β1, β2 ~ U(−1,1):

• lagged-series, ZA and ZA-lagged-series fully • Data simulated under the alternative hypothesis
reject the null and the ADF test almost fully for where ϕ1 > 1: The ERS, ADF and ZA tests do
all tested negative values of ϕ1 clearly outper- not significantly reject the null hypothesis of a
forming the ERS tests, which perform quite unit root even though ideally they should in this
poorly in these cases with negative ϕ1. case. However the ZA and the ZA-lagged-series
tests reject 75% of the test cases when ϕ1 = 1.01.
• Data simulated under the null hypothesis where
ϕ1 = 1: Four of the URTs reject the null hypoth-
URTs on AR(1) with Breaks esis as expected close to the significance level of
Table 7 summarizes the results of simulations per- 5%: ADF 3%, ERS-Ptest 2%, lagged-series 2%
formed with DGP (68). This model consists of AR(1) and ERS-DFGLS 2%. However the ZA and the
ZA-lagged-series tests have a rejection rate of
TABLE 7 | Proportions of Failures to Reject Unit Root Null 77% under the null. The ZA related tests are
Hypothesis of URTs on AR(1) Processes with one Structural Break in the worst performing test under the null
Intercept U(−1,1) and Linear Trend U(−1,1) with l = 200, m = 1000,
hypothesis.
x0 ~ N(0, 1) and α = 5 %
• Data simulated under the alternative hypothesis
ZA- where ϕ1 < 1 : The null hypothesis is never sig-
Lagged- ERS- ERS- lagged- nificantly rejected with the ADF, the lagged-
φ1 ADF series Ptest DFGLS series ZA
series and both ERS tests regardless of the value
1.010 0.98 0.96 0.98 0.98 0.25 0.25 of ϕ1. The ZA and the ZA-lagged-series tests
1.005 0.97 0.98 0.98 0.98 0.24 0.24 reject the null hypothesis 100% or more of the
1.000 0.97 0.98 0.98 0.98 0.23 0.23 cases when ϕ1 ≤ 0.0.
0.990 0.97 0.97 0.98 0.98 0.23 0.23
0.980 0.97 0.97 0.97 0.97 0.23 0.23 The results are understandable as the Zivot Andrews
0.970 0.96 0.96 0.96 0.96 0.21 0.21
URT does not allow breaks under the null; it only
allows breaks under the alternative hypothesis. When
0.960 0.95 0.96 0.94 0.95 0.20 0.20
using standard URTs such as the ADF test it is not
0.950 0.95 0.95 0.93 0.94 0.18 0.18 possible to distinguish a stationary process with
0.900 0.92 0.92 0.87 0.88 0.12 0.13 structural breaks in the linear trend from a unit root
0.000 0.88 0.85 0.93 0.88 0.00 0.00 with a drift(trend) as detailed in Ref 10, p. 60.

16 of 20 © 2017 Wiley Periodicals, Inc.

 WIREs Computational Statistics Unit root tests

LS vs Hybrid Bai-Perron-Zivot-Andrews • Data simulated under the null hypothesis where

34
The LS URT and the Hybrid Bai-Perron Zivot- ϕ1 = 1: The HBPZA and LS tests reject the null
Andrews URT proposed in Herranz22 both allow hypothesis less than the significance level of 5%
structural breaks of the intercept and/or linear for both break times of 50 and 25.
trend under the null hypothesis of a unit root. • Data simulated under the alternative hypothesis
Simulations were performed with AR(1) time where ϕ1 < 1: When the break time is 25 and
series with one structural break using Model (68) 50 the HBPZA test significantly outperforms
with α1 = 50, α2 = 1000, β1 = 1, β2 = 3 under vari- the Lee-Strazicich test in this scenario
ous levels of ϕ1. The break times were 25% and when ϕ1 ≤ 0.25.
50% of the time series length. Two sets of tests
were performed, one for each break ratio, with a
time series of length 100 and are detailed in Threshold URTs
Table 8. The three regime SETAR DGP (68) was used for
The summary of results is as follows: simulations under various combinations of the auto-
regressive multipliers ϕ1H , ϕ1M , and ϕ1L corresponding
• Data simulated under the alternative hypothesis to the high(H), middle(M) and low(L) regimes. Two
where ϕ1 = 1.01 : The HBPZA test, nor the LS different sets of symmetric thresholds were used
test significantly reject the null hypothesis of a (θH = 2, θL = − 2) and (θH = 20, θL = − 20). Thresh-
unit root even though ideally they should in old models are complex to fit as there are many fac-
this case. tors that affect the quality of the fit including the
amount of data in each regime, and the number of
TABLE 8 | Proportions of Failures to Reject the Unit Root Null parameters that need to be fit, and the length of the
Hypothesis of URTs on AR(1) Processes with 1 Break in Intercept and time series. If the threshold parameters are known,
Trend with l = 100, m = 1000, and α = 5 %, α1 = 50, α2 = 1000, then conditional least squares can be used to fit the
β1 = 1 and β2 = 3 model.
Bai-Perron- Lee– Breakpoint 8 H
φ1 >
> ϕ1 zt − 1 + εt , if zt −1 > θH ,
Zivot-Andrews Strazicich Proportion < M
ϕ1 zt − 1 + εt , if θL ≤ zt −1 ≤ θH ,
1.0100 0.97 0.99 0.25 zt =
>
> ϕ1 zt − 1 + εt , if zt −1 < θL ,
L ð69Þ
1.0000 0.97 0.99 0.25 :
0.9700 0.96 0.99 0.25 z0  N ð0,1Þ:
0.9500 0.96 0.98 0.25
0.9000 0.96 0.97 0.25 Stigler6 points out that DGP (69) is station-
ary when ϕH < 1 and ϕL < 1. The AR multiplier
0.7000 0.87 0.65 0.25
ϕ1M in the middle regime does not matter much as
0.5000 0.53 0.20 0.25
long as the high and/or low regimes are being acti-
0.4000 0.28 0.18 0.25 vated. If the realization of the process were to always
0.2500 0.05 0.21 0.25 stay in the middle regime because the threshold
0.1000 0.00 0.25 0.25 values were never reached, then for all practical pur-
0.0000 0.00 0.27 0.25 poses this regime would be the important one to con-
1.0100 0.96 0.99 0.50
sider. Simulations were performed under different
combinations of AR multipliers for each high(ϕ1H ),
1.0000 0.96 0.99 0.50
middle(ϕ1M ), and low(ϕ1L ) regimes, for two sets of
0.9700 0.96 0.99 0.50
symmetric thresholds (θL = − 2, θH = 2) and
0.9500 0.96 0.99 0.50
(θL = − 20,
0.9000 0.96 0.97 0.50 θH = 20) and tests were performed for a time series
0.7000 0.83 0.62 0.50 of length 200.
0.5000 0.37 0.22 0.50 The results for tests with time series of length
0.4000 0.18 0.18 0.50 200 are in Table 9 and can be summarized as
follows:
0.2500 0.04 0.22 0.50
0.1000 0.01 0.26 0.50
• Trivial unit root(ϕ1H = ϕ1M = ϕ1L = 1): As expected
0.0000 0.00 0.29 0.50
the tests reject close to the 5% rate of test

© 2017 Wiley Periodicals, Inc. 17 of 20

 Advanced Review [Link]/compstats

TABLE 9 | Proportions of Failures to Reject the Unit Root Null TABLE 10 | Proportions of Failures to Reject the Unit Root Null
Hypothesis of Unit Root Tests on Various Threshold AR(1) Processes Hypothesis on a AR(1) Series with a Break in φ1, l = 100, m = 1000
with l = 200, m = 1000, and α = 0.05, z0 = 0 and α = 0.05

Lagged- Break
θH θL φ1H φ1M φ1L ADF series BBC Proportion φ1A φ1B ADFA ADFB ADFAB
2.0 −2.0 1.00 1.00 1.00 0.94 0.96 0.95 0.50 0.000 0.000 0.06 0.06 0.06
2.0 −2.0 0.00 0.00 0.00 0.00 0.00 0.00 0.50 1.000 1.000 0.94 0.94 0.96
2.0 −2.0 1.00 0.90 1.00 0.90 0.95 0.95 0.50 1.000 0.500 0.93 0.65 0.91
2.0 −2.0 1.00 1.10 1.00 0.96 0.95 0.95 0.25 0.000 0.000 0.06 0.06 0.06
2.0 −2.0 1.01 1.10 1.01 0.99 0.96 0.98 0.25 1.000 1.000 0.94 0.94 0.96
2.0 −2.0 0.90 1.00 0.90 0.01 0.44 0.80 0.25 1.000 0.500 0.79 0.43 0.74
2.0 −2.0 0.90 1.10 0.90 0.02 0.57 0.70
20.0 −20.0 1.00 0.90 1.00 0.01 0.32 0.83
• Fully explosive process (ϕ1H = 1:01, ϕ1M = 1:1, ϕ1L
20.0 −20.0 1.00 1.10 1.00 1.00 0.44 0.54
= 1:01): All tests hardly reject the null hypothesis
20.0 −20.0 1.01 1.10 1.01 1.00 0.95 0.81 of a unit root.
20.0 −20.0 0.90 1.00 0.90 0.94 0.95 0.89
20.0 −20.0 0.90 1.10 0.90 1.00 0.20 0.00 For the previously discussed SETAR simula-
tions not one of the examined URTs outperformed
all of the others in every case. It is clear that the time
series length is a significant factor in these tests, as
cases as dictated by the significance level of well as of course the threshold settings. If the process
0.05: ADF 6%, lagged-series 4% and does not have sufficient data in all thresholds then it
BBC 5%. is impossible to expect the tests to work across all
• Unit root(ϕ1H = ϕ1L = 1, ϕ1M < > 1): When the thresholds.
threshold range is (θL = − 2, θH = 2) all the
tests perform relatively well and reject only Simulations with the Hybrid Bai-Perron
close to 5% of the test cases as expected.
When the threshold range is (θL = − 20,
ADF I(0)/ I(1) Testing Procedure
θH = 20) with ϕ1M = 0:9 the results are signifi-
yt = ϕ1A yt − 1 Ift ≤ TB g + ϕ1B yt − 1 Ift > TB g,
cantly worse where the lagged-series tests reject ð70Þ
68% of the tests of a unit root and the ADF y0  N ð0,1Þ:
test rejects 99% of the cases. However the
BBC only rejects 17% of the cases; the process Simulations performed using the DGP (69) which has
spends considerable more time in the middle breaks only in the AR ϕ1 multiplier with a time series
regime than with the smaller thresholds. When of length l = 100 are summarized in Table 10.
We can see that this new Hybrid Bai-Perron-
ϕ1M = 1:1 the ADF tests do not reject the null
ADF testing procedure is sensitive to the time series
root hypothesis at all, and the lagged-series
length and the location of the structural break. When
and the BBC test reject much more, 56% and
compared to using a single URT on the entire time
46% respectively.
series, this approach can be significantly more accu-
• Stationary process (ϕ1H = ϕ1L = 0:9): When the rate, as can be seen in the case with ϕ1A = 1 and
threshold range is (θL = − 2, θH = 2) only the ϕ1B = 0:5 where we make a greater error if we assume
ADF test sufficiently rejects the null hypothe- that the entire series is homogeneous
sis of a unit root. When the threshold range I(1) or I(0).
is (θL = − 20, θH = 20) when ϕ1M = 1 all of the
tests do not significantly reject the null of a
unit root. When the middle regime is explo- CONCLUSIONS
sive (ϕ1M = 1:1) the BBC tests fully reject all Unit roots can lead to spurious regressions if used in
cases of a unit root, the lagged-series rejects linear regression analyses so we must test our time
80% of the cases and the ADF does not series with URTs beforehand to ensure they are likely
reject any. stationary processes. If with each time series tested

18 of 20 © 2017 Wiley Periodicals, Inc.

 WIREs Computational Statistics Unit root tests

we cannot reject the null hypothesis of a unit root, the Bai and Perron31 structural break detection meth-
then a cointegration analysis should be pursued, odology, and then testing each time series segment with
however we must be aware that near-unit root pro- a URT such as ADF.18 In this paper we only consid-
cesses can result in spurious cointegration. There is ered I(0) and I(1) time series, and did not discuss I
no single URT that is optimal under every scenario. (2) or higher integration orders of time series, nor did
There are complications such as considering whether we consider fractionally integrated time series.b
or not to include deterministic components such as
linear trends or not under the null hypothesis, or NOTES
how long of a lag we should use in a URT. In the a
case where there are changepoints/breaks in the [Link]
b
deterministic components under the null hypothesis In this article it was not possible to review entirely the
we must be aware that most URTs, such as the ADF ever increasing literature on URTs and related theory. A
test,18 will not have sufficient statistical power to good introduction providing some of the tools necessary to
derive asymptotic formulas in URTs is Ref 14. A relatively
ever reject the null hypothesis. In these cases a URT
short yet very comprehensive book on URTs which has
such as the LS49 should be used. been referred to in this article is Ref 10, which includes
We must also realize that most of the URT’s many topics not included in this review such as additional
assume that the time series tested do not exhibit chan- URTs, fractional unit roots, panel unit roots, seasonal unit
gepoints in the level of the auto-regressive multiplier roots and Bayesian inference methods in unit roots. Per-
ϕ1. Consider testing the underlying time series for haps the most extensive and complete books on URTs cov-
breaks in all relevant parameters, including ϕ1 using ering key concepts and extensions Refs 11 and 12.

REFERENCES
1. Fama EF. The behavior of stock market prices. J Bus 11. Patterson K. Unit Root Tests in Time Series Volume 1:
1965, 38:34–105. Key Concepts and Problems. Palgrave Texts in Econo-
2. Chan NH. Time Series Applications to Finance with R metrics. Palgrave Macmillan; 2011. ISBN:
and S-Plus. Hoboken, NJ: John Wiley & Sons, Inc.; 2010. 9780230250246. Available at: [Link]
com/books?id=ST85bwAACAAJ.
3. Engle RF, Granger CWJ. Co-integration and error-cor-
rection: representation, estimation, and testing. Econo- 12. Patterson K. Unit Root Tests in Time Series Volume 2:
metrica 1987, 55:251–276. Extensions and Developments. Palgrave Texts in Econ-
4. Johansen S, Juselius K. Maximum likelihood estima- ometrics. Palgrave Macmillan; 2012. ISBN:
tion and inference on cointegration – with applications 9780230250260. Available at: [Link]
to the demand for money. Oxf Bull Econ Stat 1990, com/books?id=iE9CaLJQMNwC.
52:169–210. 13. Dickey DA, Fuller WA. Likelihood ratio statistics for
5. Balke NS, Fomby TB. Threshold cointegration. Int autoregressive time series with a unit root. Econome-
Econ Rev 1997, 38:627–645. trica 1981, 49:1057–1072.
6. Stigler M. Threshold cointegration: overview and 14. Patterson K. A Primer for Unit Root Testing. Basing-
implementation in R; 2010. Available at: [Link] stoke: Palgrave Macmillan; 2010.
[Link]/CRAN/web/packages/tsDyn/vignettes/
[Link] 15. Phillips PCB, Magadalinos T. Limit Theory for moder-
ate deviations from a unit root under weak depend-
7. Seo B. Bootstrap testing for the null of no cointegration ence. Cowles Foundation Discussion Papers 1517,
in a threshold vector error correction model. J Econ Cowles Foundation for Research in Economics, Yale
2006, 127:129–150. University; 2005. Available at: [Link]
8. Bec F, Ben Salem M, Carrasco M. Detecting mean rever- p/cwl/cwldpp/[Link]
sion in real exchange rates from a multiple regime STAR
16. Zivot E. Unit root tests, Econ584 Class Notes; 2005.
model. RCER Working Papers 509, University of Roches-
ter - Center for Economic Research (RCER); 2004. Availa- 17. MacKinnon JG. Numerical distribution functions for
ble at: [Link] unit root and cointegration tests. J Appl Econ 1996,
9. Granger CWJ. Current Perspectives on Long Memory 11:601–618.
Processes. zsfassung in chines: Sprache; 2000. 18. Said SE, Dickey DA. Testing for unit roots in autore-
10. Choi I. Almost All About Unit Roots. New York: gressive-moving average models of unknown order.
Cambridge University Press; 2010. Biometrika 1984, 71:599–607.

© 2017 Wiley Periodicals, Inc. 19 of 20

 Advanced Review [Link]/compstats

19. Ng S, Perron P. Unit root tests in arma models with 36. Di Narzo A, Di Narzo F, Aznarte JL, Stigler M. tsDyn:
data-dependent methods for the selection of the trun- time series analysis based on dynamical systems the-
cation lag. J Am Stat Assoc 1995, 90:268–281. ory, r package version 0.7; 2009. Available at: http://
20. Pfaff B. Analysis of Integrated and Cointegrated Time [Link]/CRAN/web/packages/tsDyn/vignettes/
Series with R. 2nd ed. New York: Springer; 2008. [Link]
21. Phillips PCB, Perron P. Testing for a unit root in time 37. Kipinski L, Konig R, Sieluzycki C, Kordecki W. Appli-
series regression. Biometrika 1988, 75:335–346. cation of modern tests for stationarity to single-trial
22. Herranz E. New unit root tests to decrease spurious meg data. Biol Cybern 2011, 105:183–195.
results with applications in finance and temperature 38. Koruek M, Ozkaya A. A new method to estimate
anomalies. PhD Thesis, George Mason Univer- short-run and long-run interaction mechanisms in
sity, 2016. interictal state. Digit Signal Process 2010, 20:347–358.
23. Kwiatkowski D, Phillips PCB, Schmidt P, Shin Y. Test- 39. Coggin TD. Using econometric methods to test for
ing the null hypothesis of stationarity against the alter- trends in the hadcrut3 global and hemispheric data.
native of a unit root. J Econ 1992, 54:159–178. Int J Climatol 2012, 32:315–320.
24. Andrews D, Zivot E. Further evidence on the great 40. Romilly P. Time series modelling of global mean tem-
crash, the oil price shock, and the unit-root hypothesis. perature for managerial decision-making. J Environ
J Bus Econ Stat 1992, 10:251–276. Manage 2005, 76:61–70.
25. Glynn J, Perera N, Verma R. Unit root tests and struc-
41. Ivanov MA, Evtimov SN. 1963: the break point of the
tural breaks: a survey with applications. Rev Métodos
northern hemisphere temperature trend during the
Cuantit Para La Econ La Empresa 2007, 55:63–79.
twentieth century. Int J Climatol 2010, 30:61–70.
26. Bai J, Perron P. Estimating and testing linear models
42. Whyte JM, Metcalfe AV. A proposed mildly explosive/
with multiple structural changes. Econometrica 1998,
self-exciting threshold autoregressive model applied to
66:47–78.
climatic time series. In: 19th International Congress on
27. Zeileis A, Leisch F, Hornik K, Kleiber C. Strucchange: Modelling and Simulation, MODSIM2011, Perth,
an r package for testing for structural change in linear Australia, 2011.
regression models. J Stat Softw 2002, 7:1–38.
43. Chuine I, Yiou P, Viovy N, Seguin B, Daux V, Ladurie
28. J. Y. Kim, Inference on segmented cointegration. Econ-
ELR. Historical phenology: grape ripening as a past
omet Theor pp. 620–639 2003), ISSN 1469-4360.
climate indicator. Nature 2004, 432:289–290.
29. Kejriwal M, Perron P, Zhou J. Wald tests for detecting doi:10.1038/432289a.
multiple structural changes in persistence. Economet
44. Sobradelo R, Mart J, Mendoza-Rosas AT, Gómez G.
Theor 2013, 29:289–323.
Volcanic hazard assessment for the canary islands
30. Perron P, Qu Z. Estimating restricted structural change (spain) using extreme value theory. Nat Hazards Earth
models. J Econ 2006, 134:373–399. Syst Sci 2011, 11:2741–2753.
31. Bai J, Perron P. Computation and analysis of multiple
45. Ramana D, Kumar J, Singh R, Chadha R. Co-integra-
structural change models. J Appl Econ 2003, 18:1–22.
tion models for koyna and warna reservoirs, India. Int
32. Perron P. The great crash, the oil price shock, and the J Geosci 2015:1173–1178.
unit root hypothesis. Econometrica 1989,
57:1361–1401. 46. Vargas C, Kastle E. Does the sun trigger earthquakes?
Nature 2012:595–600.
33. Lee J, Strazicich MC. Minimum lm unit root test with
one structural break. Working Papers 04-17, Depart- 47. Oluwole OSA. Climate regimes, El Niño-southern
ment of Economics, Appalachian State Univer- oscillation, and meningococcal meningitis epidemics.
sity; 2004. Front Public Health 2015:187.
34. Lee J, Strazicich MC. Minimum Lagrange multiplier 48. Skrastins A, Lauks G, Jeinskis J. Testing the null
unit root test with two structural breaks. Oxf Bull hypothesis of stationarity of internet traffic. Electron
Econ Stat 2003, 85:1082–1089. Elect Eng 2011:29–32.
35. Diethelm W, et al. fUnitRoots: Trends and Unit Roots, 49. Lee J, Strazicich MC. Break point estimation and spu-
r package version 3010.78, 2013. Available at: https:// rious rejections with endogenous unit root tests. Rev
[Link]/package=fUnitRoots. Econ Stat 2001, 63:535–558.

20 of 20 © 2017 Wiley Periodicals, Inc.