Allgemeines Statistisches Archiv 0, 014

c Physica-Verlag 0, ISSN 0002-6018

Unit Root Testing

rgen Wolters and Uwe Hassler
By Ju

Summary: The occurrence of unit roots in economic time series has far reaching consequences for univariate as well as multivariate econometric modelling. Therefore, unit root
tests are nowadays the starting point of most empirical time series studies. The oldest
and most widely used test is due to Dickey and Fuller (1979). Reviewing this test and
variants thereof we focus on the importance of modelling the deterministic component.
In particular, we survey the growing literature on tests accounting for structural shifts.
Finally, further applied aspects are addressed, for instance, how to get the size correct
and obtain good power at the same time.
Keywords: Dickey-Fuller, size and power, deterministic components, structural breaks.
JEL C22.

1. Introduction
A wide variety of economic time series is characterized by trending [Link] raises the important question how to statistically model the
long-run component. In the literature, two different approaches have been
used. The so-called trend stationary model assumes that the long-run component follows a time polynomial, which is often assumed to be linear, and
added to an otherwise stationary autoregressive moving average (ARMA)
process. The difference stationary model assumes that differencing is required to obtain stationarity, i.e. that the first difference of a time series
follows a stationary and invertible ARMA process. This implies that the
level of the time series has a unit root in its autoregressive (AR) part. Unit
root processes are also called integrated of order 1, I(1).
Since the seminal paper by Nelson and Plosser (1982) economists know
that modelling the long-run behaviour by trend or difference stationary
models has far-reaching consequences for the economic interpretation. In a
trend stationary model the effects of shocks are only temporary implying
that the level of the variable is not influenced in the long run. In contrast
a shock has permanent effects in a difference stationary model, meaning
that the level of the variable will be shifted permanently after the shock has
occurred.
Traditional econometrics assumes stationary variables (constant means
and time-independent autocorrelations). This is one of the reasons why
applied economists very often transform non-stationary variables into stationary time series. According to the two above-mentioned models this can
be done by eliminating deterministic trends in the case of a trend stationary model or by taking first differences in the case of a difference stationary model. But what happens if the wrong transformation is applied?
The papers by Chan, Hayya and Ord (1977), Nelson and Kang (1981) and
Received: / Revised:
We thank Mu-Chun Wang for producing the figures, and an anonymous referee for
comments improving the presentation.

UNIT ROOT TESTING

Durlauf and Phillips (1988) investigate this problem. Eliminating the nonstationarity in a trend stationary model by taking first differences has two
effects: one gets rid of the linear trend, but the stationary stochastic part is
overdifferenced, implying spurious short-run cycles. If, on the other hand,
it is tried to eliminate the non-stationarity in a difference stationary model
by taking the residuals of a regression on a constant and on time as explanatory variables, spurious long-run cycles are introduced. These depend
on the number of observations used in the regression. In this case artificial
business cycles are produced that lead to wrong economic interpretations.
Moreover, regressing independent difference stationary processes on each
other leads to the problem of spurious regressions as Granger and Newbold
(1974) have demonstrated in a simulation study. Later on Phillips (1986)
gave the theoretical reasoning for this phenomenon: The usual t-statistics
diverge to infinity in absolute value, while the R2 does not converge to zero,
hence indicating spurious correlation between independent difference stationary processes. Granger (1981) and Engle and Granger (1987) offered a
solution to the spurious regression problem by introducing the concept of
cointegration.
The above discussion clearly indicates that the analysis of non-stationary
time series requires a serious investigation of the trending behaviour. Therefore, formal tests are needed which allow to distinguish between trend stationary and difference stationary behaviour of time series. Such tests have
first been developed by Fuller (1976) and Dickey and Fuller (1979, 1981)
(DF test, or augmented DF test, ADF). In the meantime a lot of extensions
and generalizations have been published which also are presented in different
surveys such as Dickey, Bell and Miller (1986), Diebold and Nerlove (1990),
Campbell and Perron (1991), Hassler (1994), Stock (1994) and Phillips and
Xiao (1998).
Due to page limitations we will present here only the (augmented) DickeyFuller approach for testing the null hypothesis of difference stationarity. The
related semi-parametric approach developed by Phillips (1987) and Phillips
and Perron (1988) is not presented, and the extension to panel unit root tests
is not considered, see Breitung and Pesaran (2005) for a recent overview.
Furthermore, we do not deal with tests for seasonal unit roots as proposed
e.g. by Hylleberg, Engle, Granger and Yoo (1990), or tests having stationarity in the maintained hypothesis as Kwiatkowski, Phillips, Schmidt and
Shin (1992). We rather focus on modelling the deterministic part of the
time series under investigation. This is very important in case of structural
breaks, since neglecting deterministic shifts may result in misleading conclusions.
The paper is structured as follows. In Section 2, Dickey-Fuller unit root
tests are described and discussed. The third section deals with important
applied aspects regarding size and power. Section 4 turns to the handling
of structural breaks.


JURGEN
WOLTERS and UWE HASSLER

2. Dickey-Fuller unit root tests

2.1. Model. For the rest of the paper we assume the following data generating process (DGP):
yt = dt + xt ,

t = 1, . . . , T .

(1)

The observed variable (yt ) is composed of a deterministic component dt and

a purely stochastic component xt . The deterministic part may consist of a
constant, seasonal dummy variables, a linear trend or a step dummy. The
stochastic component is assumed to be a zero mean AR(p) process,
xt = 1 xt1 + 2 xt2 + ... + p xtp + ut ,

(2)

with p 6= 0 and ut being white noise. The AR(p) model in (2) can be
reparameterized as
xt = xt1 +

p1
X

ai xti + ut

(3)

i=1

with
=

p
X

and ai =

j=1

p
X

j , i = 1, ..., p 1.

j=i+1

If the lag polynomial of xt

1 1 z 2 z 2 ... p z p = 0
has a unit root, then it holds
=

p
X

j = 1.

(4)

j=1

Substituting (3) in (1) we get the following expression for the observable
variable yt :
yt = dt dt1

p1
X
i=1

ai dti + yt1 +

p1
X

ai yti + ut .

i=1

Subtracting yt1 on both sides of this equation, we obtain the Augmented

Dickey-Fuller (ADF) regression:
yt = dt dt1

p1
X
i=1

ai dti + ( 1) yt1 +

p1
X

ai yti + ut .

(5)

i=1

Note that in addition to the (lagged) level of the deterministic part its
lagged changes are included, too.

UNIT ROOT TESTING

In the case that dt = c, a constant, the ADF regression is given as

yt = a + ( 1)yt1 +

k
X

ai yti + ut ,

t = k + 2, . . . , T ,

(6)

i=1

with a = (1)c, meaning that under the null hypothesis the process is I(1)
without drift. When the stochastic component xt follows an AR(p) process
then k = p1 in (6). More generally, however, xt may be an ARMA process
with invertible moving average component. In that case, Said and Dickey
(1984) propose to approximate the ARMA stucture by autoregressions of
order k where the lag length k has to grow with the sample size.
In the case of a linear trend, dt = c + mt, we get from (5) as ADF
regression
yt = a + bt + ( 1)yt1 +

k
X

ai yti + ut ,

t = k + 2, . . . , T ,

(7)

i=1

with
a = c(1 ) + m m

k
X

ai and b = m(1 ).

i=1

Under the null of a unit root ( = 1) the trend term in (7) vanishes, while the
constant term contains not only the slope parameter m but also the coefficients ai of the short run dynamic. Under the null it holds that E(yt ) 6= 0,
and hence yt displays a stochastic trend, I(1), as well as a deterministic one
because E(yt ) grows linearly with t. Such series are called integrated with
drift.
2.2. Distribution. The null hypothesis to be tested is that xt and hence
yt is integrated of order one. Under the null hypothesis there is a unit root
in the AR polynomial and we have because of (4):
H0 : 1 = 0.
This hypothesis can be tested directly by estimating (5) with least squares
and using the t-statistic of 1. It is a one-sided test that rejects in favour
of stationarity if 1 is significantly negative. The limiting distribution
was discovered by Fuller (1976) and Dickey and Fuller (1979). It turned out
that it is not centered around zero (but rather shifted to the left) and not
symmetrical. In particular, limiting standard normal theory is invalid for
= 1, while it does apply in case of stationarity (|| < 1). Similarly, the
t-distribution is not a valid guideline for unit root testing in finite samples.
The limiting distribution is very sensitive to the specification of dt . Fuller
(1976) and Dickey and Fuller (1979) consider three cases: No deterministics
(dt = 0), just a constant, and a constant and a linear trend. Critical values for these cases have first been provided by simulation in Fuller (1976,


JURGEN
WOLTERS and UWE HASSLER

normal vs. rho=0.8

rho=0.9 vs rho=0.95

0.5

0.5
normal
rho=0.8

0.4
0.3

0.3

0.2

0.2

0.1

0.1

0
6

rho=0.9
rho=0.95

0.4

0
6

rho=0.99 vs rho=1

normal vs rho=1

0.5

0.5
rho=0.99
rho=1

0.4

0.3

0.2

0.2

0.1

0.1

normal
rho=1

0.4

0.3

0
6

0
6

Figure 1. Estimated density functions (T = 100) and standard normal.

Table 8.5.2, p.373)1 . Nowadays, somewhat more precise critical values are
widely employed, which have been derived using more intensive simulations
by MacKinnon (1991, p.275). For a test with T = 100 at the 5 % level the
critical values are -2.89 and -3.46, respectively, if dt contains only a constant
and if dt includes a constant and a linear trend. These values are larger in
absolute terms than the corresponding critical value of the t-distribution
which is -1.65. Using the incorrect t-distribution the null hypothesis would
be rejected much too often. The decision would wrongly be in favour of stationarity or trendstationarity despite the fact that the time series contains
an I(1) component.
Theoretically, = 1 is a singularity in the parameter space in that for any
|| < 1 limiting normality holds true, while = 1 results in the non-normal
Dickey-Fuller distributions. Some economists blame this as being artificial
and claim that the true equals 1 with probability zero. According to this,
the normal approximation should be applied throughout. In practice however, and that means in finite samples, the distinction between stationarity
and I(1) is not so clear-cut2 . Evans and Savin (1981, p.763) observed that
1 Critical values for polynomials in t up to order 5 are derived in Ouliaris, Park and
Phillips (1989).
2 In fact, Phillips (1987a) presented a unifying local-to-unity theory bridging the gap

from stationarity to I(1) by introducing a time dependent , T = exp Tc .

UNIT ROOT TESTING

for near but below unity this distribution function is very poorly approximated by the limiting normal even for large values of T . Similar evidence
has been collected by Evans and Savin (1984) and Nankervis and Savin
(1985). Evans and Savin (1981, 1984) considered a normalization different
from the t-statistic, while Nankervis and Savin (1985) considered the usual
studentization but did not provide graphs for the t-statistics. Therefore, we
want to add the results of a small Monte Carlo study here3 . The true DGP
is (with y0 = 0)
yt = yt1 + ut , ut N (0, 1),
for t = 1, ..., T = 100. Relying on
yt = a
+ (
1)yt1 + u
t ,
the usual t-statistic is computed. Figure 1 displays density estimates that
are constructed from 10000 replications by smoothing with a normal kernel.
From these results we learn that even for considerably smaller than 1
the standard normal approximation may be a very bad guideline in finite
samples.
3. Size and power considerations
Since the work by Schwert (1989) it has been documented in several papers
that the DF test may be over-sized in situations of practical importance4 .
Hence, proposals how to control for the probability of a type I error have
attracted a lot of attention in the last decade. At the same time DF tests are
blamed for poor power5 , and many papers tackled the problem to increase
power. Several aspects related to those topics are addressed next.
3.1. Lag length selection. The lag length k in (6) and (7) has to be
chosen to ensure that the residuals empirically follow a white noise process.
Said and Dickey (1984) prove that the ADF test in (5) is valid if the true
DGP is an ARMA process of unknown order, provided that the lag length
k in the autoregression increases with the sample size but at a lower rate. A
proof under more general conditions was recently provided by Chang and
Park (2002). In practice, the choice of k is a crucial and difficult exercise.
On the one hand, a growing number of lags reduces the effective sample
while the number of estimated parameters is increased, and this reduction
3

All programming was done by Mu-Chun Wang in MATLAB.

Kim and Schmidt (1993) established experimentally that conditionally heteroskedastic errors have little effect on DF tests as long as there is only moderate heteroskedasticity.
But Valkanov (2005) showed that with strongly heteroskedastic data the use of asymptotic critical DF values leads to grossly oversized tests.
5 Gonzalo and Lee (1996), in contrast, illustrated by means of Monte Carlo experiments
that testing for = 1 results in rejection frequencies very similar to those available if
|| < 1. In case of fractionally integrated alternatives, however, Hassler and Wolters
(1994) showed that the ADF test has little power.
4

JURGEN
WOLTERS and UWE HASSLER

in degrees of freedom will result in a loss of power. On the other hand, k

has to be large enough for the residuals to be approximately uncorrelated in
order for the limiting theory to be valid. Empirical researchers often start
with a maximum lag length kmax and follow a sequential general-to-specific
strategy, i.e. reduce lags until reaching significance at a prespecified level.
Here, significance testing builds on a limiting standard normal distribution.
Alternatively, k may be determined relying on information criteria. The
performance of the two strategies in finite samples has been investigated by
Ng and Perron (1995). In particular, information criteria tend to choose k
too small to get the size correct. Therefore, Ng and Perron (2001) proposed
adequately modified criteria.
The AR approximation is particularly poor in case of moving average
roots close to one. Consider as DGP
yt = a + ut ut1 ,

|| < 1.

With close to one the polynomial 1 L almost cancels with = 1 L,

and the true null of integration will be rejected frequently, resulting in a
test where the empirical size is above the nominal one, cf. Schwert (1989).
In such a situation one may use the procedure proposed by Said and Dickey
(1985) that explicitly takes into account the MA component of a series.

3.2. Deterministic components. When performing unit root tests an

appropriate specification of the deterministics in (5) is of crucial importance. First, consider the case where the true DGP is trend stationary. If the
ADF regression (6) without detrending is applied, then the test has asymptotically no power, which was shown by West (1987) and Perron (1988).
In finite samples the null hypothesis of a unit root is rarely rejected, and
is never rejected in the limit. Hence, we propose to include a linear trend
as in (7) whenever a series is suspicious of a linear trend upon visual inspection6 . Notice that the decision about time as regressor may not build
on the standard t-statistic of the estimate b. Second, assume the other way
round that a detrended test is performed from (7) while the data does not
contain a linear trend. In this situation a test from (6) without detrending
would be more powerful. The effect of ignoring eventual mean shifts in the
DGP when specifying dt will be discussed in the following section, while the
effect of neglected seasonal deterministics will be touched upon in the next
subsection.
The treatment of the deterministic component plays a major role when it
comes to power of unit root tests. Elliott, Rothenberg and Stock (1996) and
Hwang and Schmidt (1996) proposed point optimal unit root tests with
maximum power against a given local alternative T = 1 Tc for some
specified constant c > 0. Power gains are obtained by efficiently removing
6 Ayat and Burridge (2000) investigated a more rigorous sequential procedure to determine the appropriate deterministics when testing for unit roots.

UNIT ROOT TESTING

the deterministic component under the alternative (using Generalized Least

Squares, GLS). Use of GLS, however, amounts to the following procedure,
see also Xiao and Phillips (1998). First, compute quasi-differences of the
observed variable and the deterministic regressors,

c
c
yt1 , c zt = zt 1
zt1 ,
c yt = yt 1
T
T
where zt is a deterministic vector such that dt from (1) is parameterized as
dt = 0 zt . Second, estimate the vector by simply regressing c yt on c zt .
Third, apply the ADF test without deterministics to the residuals from the
second step. The distribution and hence critical values depend on the choice
of c. Yet another approach to obtain more powerful unit root tests has been
advocated by Shin and So (2001). They proposed to estimate by t with
information only up to t, and remove the deterministic component from yt
0
zt1 . Applying an ADF type regression to yt results
as follows: yt = yt t1
in a limiting distribution again different from Dickey-Fuller. Critical values
have been provided by Shin and So (2001, Table II) for the simplest case of
a constant where zt = 1.
Leybourne, Kim and Newbold (2005) explored both asymptotic and
finite-sample properties of five more powerful modifications of the DF test.
They favoured two tests studied by Taylor (2002) and Leybourne (1995).
The latter relies on the usual DF statistic and the test statistic computed
from the reversed time series. The test proposed by Taylor (2002) builds on
recursive detrending by ordinary least squares.

3.3. Span vs. frequency. In practice, it often happens that data are
available only for a fixed time span due to some structural breaks or institutional changes. In such a situation it is often recommended to use data
with higher frequency to increase the number of observations and hence the
power of tests. To that end people often work with monthly data instead of
quarterly or annual observations. Perron (1989) has analyzed the power of
some unit root tests when the sampling interval is varied but the time span
is hold fixed. A general outcome of his computer experiments is that tests
over short time span have low power, which is not significantly enhanced
by choosing a shorter sampling interval. For related results see also Shiller
and Perron (1985).
Moreover, in many cases higher frequency of observations comes at the
price of additional seasonal dynamics that have to be modelled. In case of
deterministic seasonal patterns it is important to remove the seasonality
by including seasonal dummies in the regression. Dickey, Bell and Miller
(1986) prove that the inclusion of seasonal dummies instead of a constant
does not affect the limiting distribution of DF tests, while Demetrescu and
Hassler (2005) demonstrate that neglecting seasonal deterministic results
in tests with low power and bad size properties at the same time. Another
way of removing seasonal deterministics is simply to work with seasonal

JURGEN
WOLTERS and UWE HASSLER

differences, which, however, can not be recommended in general. Hassler and

Demetrescu (2005) argue that seasonal differencing may introduce artificial
persistence into a time series and may hence create spurious unit roots.
Given a fixed time span of data the purpose of unit root testing is not
to investigate the true nature of some abstract economic process but to
describe the degree of persistence in a given sample. Even if difference stationarity is not a plausible theoretical model as T for economic series
such as inflation rates, interest rates or unemployment rates, the unit root
hypothesis may still provide an empirically valid description. In that sense
the significance against H0 : = 1 may be understood as strength of meanreversion in a given sample. Similarly, Juselius (1999, pp.264) argues that
the order of integration of a variable is not in general a property of an
economic variable but a convenient statistical approximation to distinguish
between the short-run, medium-run and long-run variation in the data.
4. Structural breaks
4.1. Ignoring breaks. Consider for the moment a regression with a constant only,
yt = a
+ (
1)yt1 + u
t .

(8)

It is now assumed that the regression (8) is misspecified: The true process
is I(0), but it displays a break in the mean at time T ,
(
xt , t < T
, xt I(0),
(9)
yt =
xt + , t T
where (0, 1), and 6= 0. Given (9), neither the null nor the alternative
hypothesis of the DF test holds true. Perron (1990) proves that converges
to a value that approaches 1 as the break || > 0 is growing. A corresponding
result is found in Perron (1989a) in case of the detrended version of the DF
test,
yt = a
+ b t + (
1)yt1 + u
t .

(10)

This means that for a considerable break 6= 0 the wrong null of a unit
root is hardly rejected. A high probability of a type II error arises because
the DF regression is misspecified in that it does not account for the structural break in the data. See also the intuitive discussion in Rappoport and
Reichlin (1989). Moreover, Perron (1989a) investigates the situation of trend
stationary series with a break:
(
xt + t ,
t < T
yt =
, xt I(0).
xt ( + ) + ( + ) t , t T
For 6= 0 we have a break in the slope of the trend, while there may be an
additional shift in the level ( 6= 0) or not. With this assumption Perron

UNIT ROOT TESTING

(1989a) investigates from the detrended DF test (10). His asymptotic formulae were corrected by Monta
nes and Reyes (1998), but without changing
the empirically relevant fact: Given a trend stationary series with a break
in the linear trend there is little chance to reject the false null hypothesis
of integration. Those results opened a new research avenue aiming at the
discrimination between unit roots and structural shifts as potential causes
of economic persistence.
Quite surprisingly, the opposite feature to that discovered by Perron
(1989a, 1990) has been established by Leybourne, Mills and Newbold (1998):
If a unit root process is subject to a mean shift, then the probability of
rejecting the null of a unit root is not equal to the level of the test7 . The
authors assume
(
xt , t < T
yt =
, xt I(1),
xt + , t T
with 6= 0. Leybourne, Mills and Newbold (1998) prove that the limiting
distribution of the DF statistic depends on if is growing with T . Experimentally, they establish that the empirical level is well above the nominal
one in case of an early break, 0.1; if, however, the break occurs in the
second half of the sample, then the DF test is conservative in the sense that
the rejection frequency is below the nominal level. More generally, Kim,
Leybourne, and Newbold (2004) have shown that the results of the ADF
test (6) including only a constant term is highly unpredictable, if the true
deterministic is a broken trend.
4.2. Correcting for breaks. To avoid spurious unit roots due to structural breaks, Perron (1989a, 1990) suggested to test for integration after
removing structural breaks8 . Unfortunately, this changes the limiting distributions depending on the break fraction (0, 1). To correct for a break
in the level we need the step dummy
(
0, t < T
st () =
.
1, t T
Now, all the deterministic components are removed from the observed series
in a first step by an OLS regression9 ,
yt = a
+
st () ( + b t ) + x
t .
(11)
7 This, however, is only true in finite samples or if the break is growing with the
number of observations. If in contrast the break is finite, then the level of the DF test
is not affected asymptotically, see Amsler and Lee (1995). Still, power considerations
suggest a correction for potential breaks.
8 Perron (1994) provides a very accessible survey.
9 L
utkepohl, M
uller and Saikkonen (2001) and Saikkonen and L
utkepohl (2001) suggested to remove the deterministic components efficiently in the sense of Elliott, Rothenberg and Stock (1996). This approach of working with quasi-differences has the advantage
of yielding limiting distributions independent of . The same property has the test by
Park and Sung (1994).

10

JURGEN
WOLTERS and UWE HASSLER

Next, the zero mean residuals x

t are tested for a unit root. However, the
validity of the asymptotic percentiles requires the inclusion of the impulse
dummy
st () = st () st1 ().
The necessity to include (lagged values of) st () has been recognized only
by Perron and Vogelsang (1992, 1993) although it is well motivated by (5):

xt = (
1)
xt1 +

k
X
i=1

a
i
xti +

k
X

i sti () + u
t .

(12)

i=0

Critical values when testing for = 1 in (12) are found in Perron (1989a,
1990). They depend on the break fraction that is assumed to be known.
Similarly, Perron (1989a) considered modifications of the detrended DF test
allowing for a shift in the slope of the linear time trend.
If is not known the break point can be estimated from the data. Zivot
and Andrews (1992) e.g. suggested to vary and to compute the test statistic ADF () for each regression. Then the potential break point can be de = arg min ADF (). Confer also Banerjee, Lumsdaine and
termined as
Stock (1992) and Christiano (1992).
4.3. Smooth transitions and several breaks. When defining the step
dummy st () we assumed a sudden change at t = T . But even if the cause
of a change occurs instantaneously, its effect most likely evolves gradually over a period of transition. To account for that effect Perron (1989a)
proposed the so-called innovational outlier model, which assumes that the
economy responds to a shock to the trend function the same way as it reacts
to any other shock, Perron (1989a, p.1380). This amounts to adding a step
dummy variable to the augmented Dickey-Fuller regression instead of applying the ADF test after removing all deterministics. Leybourne, Newbold
and Vougas (1998) considered unit root testing in the presence of more general deterministic smooth transition functions, see also Lin and Terasvirta
(1994), however without providing asymptotic theory. Limiting results under smooth transitions have been established in Saikkonen and L
utkepohl
(2001) for known breakpoint, and in Saikkonen and L
utkepohl (2002) in
case the date of the break is not known a priori. For a comparison of related
tests see also Lanne, L
utkepohl and Saikkonen (2002).
Lumsdaine and Papell (1997) allowed for two break points where both
break dates are assumed to be unknown. Park and Sung (1994) dealt with
the case of several breaks, however at known time. Kapetanios (2005) combined both features, i.e. more than two breaks occuring at unknown break
points. Our personal opinion, however, is that the data-driven estimation of
m break dates for a given series should not be a recommended strategy unless it is possible to reveal an economic event or institutional change behind
each eventual break.

UNIT ROOT TESTING

11

Bibliography
Ajat, L., Burridge, P. (2000). Unit root tests in the presence of uncertainty
about the non-stochastic trend. Journal of Econometrics 95 71-96.
Amsler, Ch., Lee, J. (1995). An LM test for a unit root in the presence of a
structural change. Econometric Theory 11 359-368.
Banerjee, A., Lumsdaine, R.L., Stock, J.H. (1992). Recursive and sequential
tests of the unit-root and trend-break hypothesis: theory and international
evidence. Journal of Business & Economic Statistics 10 271-287.
Breitung, J., Pesaran, H. (2005). Unit roots and cointegration in panel data.
mimeo, University of Bonn.
Campbell, J.Y., Perron, P. (1991). Pitfalls and opportunities: What macroeconomists should know about unit roots. In NBER Marcoeconomics Annual
1991 (O.J. Blanchard, S. Fisher, eds.), 141201. MIT Press, Cambridge.
Chan, K.H., Hayya, J.C., Ord, J.K. (1977). A note on trend removal methods:
The case of polynomial regressions versus variate differencing. Econometrica
45 737-744.
Chang, Y., Park, J.Y. (2002). On the asymptotics of ADF tests for unit roots.
Econometric Reviews 21 431-447.
Christiano, L.J. (1992). Searching for a break in GNP. Journal of Business &
Economic Statistics 10 237-250.
Demetrescu, M., Hassler, U. (2005). Effect of neglected deterministic seasonality on unit root tests. Statistical Papers forthcoming .
Dickey, D.A., Fuller, W.A. (1979). Distribution of the estimators for autoregressive time series with a unit root. Journal of the American Statistical Association 74 427-431.
Dickey, D.A., Fuller, W.A. (1981). Likelihood ratio statistics for autoregressive time series with a unit root. Econometrica 49 1057-1072.
Dickey, D.A., Bell, W.R., Miller, R.B. (1986). Unit roots in time series
models: Tests and implications. American Statistician 40 12 - 26.
Diebold, F.X., Nerlove, M. (1990). Unit roots in economic time series: A selective survey. In Advances in Econometrics: Cointegration, Spurious Regressions,
and Unit Roots (T.B. Fomby, G.F. Rhodes, eds.), 3 - 70. JAI Press Greenwich.
Durlauf, S., Phillips. P.C.B. (1988). Trends versus random walks in time series
analysis. Econometrica 56 1333-1354.
Elliott, G., Rothenberg, Th.J., Stock, J.H. (1996). Efficient tests for an
autoregressive unit root. Econometrica 64 813-836.
Engle, R.F., Granger, C.W.J. (1987). Co-integration and error correction:
Representation, estimation, and testing. Econometrica 55 251-276.
Evans, G.B.A., Savin, N.E. (1981). Testing for unit roots: 1. Econometrica 49
753-779.
Evans, G.B.A., Savin, N.E. (1984). Testing for unit roots: 2. Econometrica 52
1241-1270.
Fuller, W.A. (1976). Introduction to Statistical Time Series. Wiley, New York.
Gonzalo, J., Lee, T.-H. (1996). Relative power of t type tests for stationary
and unit root processes. Journal of Time Series Analysis 17 37 - 47.

12

JURGEN
WOLTERS and UWE HASSLER

Granger, C.W.J. (1981). Some properties of time series data and their use in
econometric model specification. Journal of Econometrics 16 121-130.
Granger, C.W.J., Newbold, P. (1974). Spurious regressions in econometrics.
Journal of Econometrics 2 111-120.

Hassler, U. (1994). Einheitswurzeltests - Ein Uberblick.

Allgemeines Statistisches Archiv 78 207-228.
Hassler, U., Demetrescu, M. (2005). Spurious persistence and unit roots due
to seasonal differencing: The case of inflation rates. Journal of Economics and
Statistics 225 413-426.
Hassler, U., Wolters, J. (1994). On the power of unit root tests against fractional alternatives. Economics Letters 45 15.
Hwang, J., Schmidt, P. (1996). Alternative methods of detrending and the
power of unit root tests. Journal of Econometrics 71 227-248.
Hylleberg, S., Engle, R.F., Granger, C.W.J., Yoo, B.S. (1990). Seasonal
integration and cointegration. Journal of Econometrics 44 215-238.
Juselius, K. (1999). Models and relations in economics and econometrics. Journal
of Economic Methodology 6 259-290.
Kapetanios, G. (2005). Unit-root testing against the alternative hypothesis of
up to m structural breaks. Journal of Time Series Analysis 26 123133.
Kim, K., Schmidt, P. (1993). Unit roots tests with conditional heteroskedasticity.
Journal of Econometrics 59 287-300.
Kim, T.W., Leybourne, S., Newbold, P. (2004). Behaviour of Dickey-Fuller
unit-root tests under trend misspecification. Journal of Time Series Analysis
25 755764.
Kwiatkowski, D., Phillips, P.C.B., Schmidt, P., Shin, Y. (1992). Testing
the null hypothesis of stationarity against the alternative of unit root. Journal
of Econometrics 54 159-178.
tkepohl, H., Saikkonen, P. (2002). Comparison of unit root
Lanne, M., Lu
tests for time series with level shifts. Journal of Time Series Analysis 23 667685.
Leybourne, S.J (1995). Testing for unit roots using forward and reverse DickeyFuller regressions. Oxford Bulletin of Economics and Statistics 57 559-571.
Leybourne, S.J., Kim, T.W., Newbold, P. (2005). Examination of some more
powerful modifications of the Dickey-Fuller test. Journal of Time Series Analysis 26 355-369.
Leybourne, S.J., Mills, T.C., Newbold, P. (1998). Spurious rejections by
Dickey-Fuller tests in the presence of a break under the null. Journal of Econometrics 87 191-203.
Leybourne, S., Newbold, P., Vougas, D. (1998). Unit roots and smooth transitions. Journal of Time Series Analysis 19 83-97.
svirta, T. (1994). Testing the constancy of regression parameters
Lin, C.J., Tera
against continuous structural change. Journal of Econometrics 62 211-228.
Lumsdaine, R.L., Papell, D.H. (1997). Multiple trend breaks and the unit-root
hypothesis. The Review of Economics and Statistics 79 212-218.

UNIT ROOT TESTING

13

tkepohl, H., Mu
ller, C., Saikkonen, P. (2001). Unit root tests for time
Lu
series with a structural break when the break point is known. In Nonlinear Statistical Modeling: Essays in Honor of Takeshi Amemiya (C. Hsiao, K.
Morimune, J.L. Powell, eds.) 327-348. Cambridge University Press, Cambridge.
MacKinnon, J.G. (1991). Critical Values for Co-Integration Tests. In Long-Run
Economic Relationships (R.F. Engle and C.W.J. Granger, eds.), 267276. Oxford University Press, Oxford.
e
s, A., Reyes, M. (1998). Effect of a shift in the trend function on the
Montan
Dickey-Fuller unit root test. Econometric Theory 14 355-363.
Nankervis, J.C., Savin, N.E. (1985). Testing the autoregressive parameter with
the t statistic. Journal of Econometrics 27 143-161.
Nelson, C.R., Kang, H. (1981). Spurious periodicity in inappropriately detrended time series. Econometrica 49 741-751.
Nelson, C.R., Plosser, C.I. (1982). Trends and random walks in macro-economic
time series: Some evidence and implications. Journal of Monetary Economics
10 139-162.
Ng, S., Perron, P. (1995). Unit root tests in ARMA models with data-dependent
methods for the selection of the truncation lag. Journal of the American Statistical Association 90 268-281.
Ng, S., Perron, P. (2001). Lag length selection and the construction of unit
root tests with good size and power. Econometrica 69 1519-1554.
Ouliaris, S., Park, J.Y., Phillips, P.C.B. (1989). Testing for a unit root in the
presence of a maintained trend. In Advances in Econometrics and Modelling
(B. Raj, ed.), 728. Kluwer Academic Publishers, Dordrecht.
Park, J.Y., Sung, J. (1994). Testing for unit roots in models with structural
change. Econometric Theory 10 917-936.
Perron, P. (1988). Trends and random walks in macroeconomic time series:
Further evidence from a new approach. Journal of Economic Dynamics and
Control 12 297332.
Perron, P. (1989). Testing for a random walk: A simulation experiment of power
when the sampling interval is varied. In Advances in Econometrics and Modelling (B. Raj, ed.), 4768. Kluwer Academic Publishers, Dordrecht.
Perron, P. (1989a). The great crash, the oil price shock, and the unit root
hypothesis. Econometrica 57 1362-1401.
Perron, P. (1990). Testing for a unit root in a time series with a changing mean.
Journal of Business & Economic Statistics 8 153-162.
Perron, P. (1994). Trend, unit root and structural change in macroeconomic
time series. In Cointegration for the Applied Economist (B.B. Rao, ed.), 113146. St. Martins Press, New York.
Perron, P., Vogelsang, T.J. (1992). Testing for a unit root in a time series
with a changing mean: Corrections and Extensions. Journal of Business &
Economic Statistics 10 467-470.
Perron, P., Vogelsang, T.J. (1993). Erratum. Econometrica 61 248-249.
Phillips, P.C.B. (1986). Understanding spurious regressions in econometrics.
Journal of Econometrics 33 311-340.

14

JURGEN
WOLTERS and UWE HASSLER

Phillips, P.C.B. (1987). Time series regression with a unit root. Econometrica
55 277-301.
Phillips, P.C.B. (1987a). Towards a unified asymptotic theory for autoregression. Biometrika 74 535-47.
Phillips, P.C.B., Perron, P. (1988). Testing for a unit root in time series
regression. Biometrika 75 335-346.
Phillips, P.C.B., Xiao, Z. (1998). A primer on unit root testing. Journal of
Economic Surveys 12 423-469.
Rappoport, P., Reichlin, L. (1989). Segmented trends and non-stationary time
series. The Economic Journal 99 168-177.
Said, S.E., Dickey, D.A. (1984). Testing for unit roots in ARMA(p,q)-models
with unknown p and q. Biometrika 71 599-607.
Said, S.E., Dickey, D.A. (1985). Hypothesis testing in ARIMA(p, 1, q) models.
Journal of the American Statistical Association 80 369-374.
tkepohl, H. (2001). Testing for unit roots in time series with
Saikkonen, P., Lu
level shifts. Allgemeines Statistisches Archiv 85 1-25.
tkepohl, H. (2002). Testing for a unit root in a time series
Saikkonen, P., Lu
with a level shift at unknown time. Econometric Theory 18 313-348.
Schwert, G.W. (1989). Tests for unit roots: A monte carlo investigation. Journal
of Business & Economic Statistics 7 147-158.
Shiller, R.J., Perron, P. (1985). Testing the random walk hypothesis. Economics Letters 18 381-386.
Shin, D.W., So, B.S. (2001). Recursive mean adjustment for unit root tests.
Journal of Time Series Analysis 22 595-612.
Stock, J.H. (1994). Unit roots, structural breaks and trends. In Handbook of
Econometrics, Volume IV (R. F. Engle, D. L. McFadden, eds.), 2739-2841.
Elsevier, Amsterdam et al..
Taylor, A.M.R. (2002). Regression-based unit root tests with recursive mean
adjustment for seasonal and non-seasonal time series. Journal of Business &
Economic Statistics 20 269-281.
West, K.D. (1987). A note on the power of least squares tests for a unit root.
Economics Letters 24 249-252.
Valkanov, R. (2005). Functional central limit theorem approximations and the
distribution of the Dickey-Fuller test with strongly heteroskedastic data. Economics Letters 86 427-433.
Xiao, Z., Phillips, P.C.B. (1998). An ADF coefficient test for a unit root in
ARMA models of unknown order with empirical applications to US economy.
Econometrics Journal 1 27-44.
Zivot, E., Andrews, D.W.K. (1992). Further evidence on the great crash, the
oil-price shock, and the unit-root hypothesis. Journal of Business & Economic
Statistics 10 251-270.

UNIT ROOT TESTING

J
urgen Wolters
Fachbereich Wirtschaftswissenschaft
Freie Universit
at Berlin
Boltzmannstr. 20
14195 Berlin
Deutschland
wolters@[Link]

15

Uwe Hassler
Fachbereich Wirtschaftswissenschaften
Goethe-Universit
at Frankfurt
Gr
afstr. 78, Uni-PF 76
60054 Frankfurt am Main
Deutschland
hassler@[Link]