See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/4828401

CIPS test for Unit Root in Panel Data: further Monte Carlo results

Article  in  Economics Bulletin · March 2008

Source: RePEc

CITATIONS READS
5 898

1 author:

Andrea Cerasa
European Commission
12 PUBLICATIONS   38 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Newcomb-Benford law View project

Modeling View project

All content following this page was uploaded by Andrea Cerasa on 22 January 2016.

The user has requested enhancement of the downloaded file.

 CIPS test for Unit Root in Panel Data: further Monte Carlo
results

Andrea Cerasa
University

Abstract
This paper analyzes, through Monte Carlo experiments, the behaviour of Pesaran’s CIPS test
for the null of a unit root in panel data when (i) the assumption of a single common factor in
the specification of the cross-section dependence is violated and (ii) the autoregressive order
of the residuals is estimated. The simulation analysis points to the single common factor as a
fundamental assumption for a suitable behaviour of the CIPS test and suggests that the test
delivers the best performance when the truncation lag is estimated as a deterministic function
of the sample size.

Citation: Cerasa, Andrea, (2008) "CIPS test for Unit Root in Panel Data: further Monte Carlo results." Economics Bulletin, Vol.
3, No. 16 pp. 1-13
Submitted: July 12, 2007. Accepted: March 19, 2008.
URL: http://economicsbulletin.vanderbilt.edu/2008/volume3/EB-07C20091A.pdf
1 Introduction
In recent years, the issue of testing for unit root in panel data has been a
much debated topic. The literature about the development of such tests was
initially based upon the assumption of cross-sectional independence between
the units and produced the so called ”first generation panel unit root tests”.
However, in several empirical applications, this assumption is likely to be
violated and O’Connell (1998) showed that not considering the possible de-
pendence between units could introduce severe bias in the first generation
panel unit root tests. Hence researchers were interested in developing tests
invariant with respect to the cross-sectional dependence, called ”second gen-
eration unit root tests”.
Among them, Pesaran (2006) proposed the CIPS test, based on a single
common factor specification for the cross-correlation structure. Simulation
results under the assumption of a single common factor and known auto-
correlation order of the residuals, show that the CIPS test performs very
well. Objections can be raised to the empirical relevance of both assump-
tions. While the latter is obvious (in practice the lag order is of course never
known), the former may be acceptable in some cases but not in others. For
instance, the so-called ”convergence clubs” theory implies more than one
common factor (see, inter alia, Ben-David 1994, and Galor 1996).
The aim of this paper is thus to extend Pesaran’s results considering: (i) the
presence of more than one common factor in the Data Generating Process
(DGP) and (ii) empirical estimation of the truncation lag.
We shall now first outline the CIPS test (Section 2) and then present the
results of our Monte Carlo studies (Section 3).

2 Pesaran’s CIPS test

Let us consider the dynamic linear heterogeneous panel data model:

zit = (1 − φi )µi + φi zi,t−1 + uit (1)

where uit has the one common factor structure

uit = γi ft + eit (2)

in which ft ∼ i.i.d.(0, σf2 ) is the unobserved common effect, γi ∼ i.i.d.(0, σγ2 )

the individual factor loading and eit the idiosyncratic component which can
be i.i.d.(0, σi2 ) or, more generally, a stationary autoregressive process. Rewrit-
ing (1) and (2) as
∆zit = αi + βi zi,t−1 + eit

1
Pesaran (2006) proposes to proxy thePcommon factor ft with the cross sec-
tion mean of zit , namely z¯t = N −1 N i=1 zit , and its lagged value(s) z̄t−1 ,
z̄t−2 ,. . . The test for the null of unit root regarding the unit i can now be
based on the t ratio of the OLS estimate of bi in the cross-sectionally aug-
mented Dickey-Fuller (CADF) regression

∆zit = ai + bi zi,t−1 + ci z̄t−1 + di ∆z¯t + eit

A natural test of the null H0 : βi = 0 for all i, against the heterogeneous

alternative H1 : β1 < 0, . . . , βN0 < 0, N0 ≤ N in the whole panel data set, is
given by the average of the individual CADF statistics:
N
X
CIP S(N, T ) = N −1 ti (N, T )
i=1

The distribution of this test is non-standard, even asymptotically; 1%, 5%

and 10% critical values are tabulated by the author for different combinations
of N and T.
In case of serial correlation of the individual-specific error terms, the testing
procedure can be easily extended by adding a suitable number of lagged
values of z¯t and ∆zit in the CADF regression 1 without any change in the
distribution of the statistic.

3 Monte Carlo simulations

3.1 More than one common factor in the DGP
Consider first the presence of more than one common factor in the data.
Adopting the same Data Generating Process (DGP) and simulation design
used by Pesaran (2006), we can replace (2) with
K
X
uit = γij fjt + eit (3)
j=1

where the idiosyncratic component is given by

eit = ρi ei,t−1 + εit (4)

εit ∼ N (0, σi2 ) with σi2 ∼ U ni[0.5, 1.5]; fjt , µi ∼ N (0, 1)

1
For example, for anP AR(p) error specification, the suitable CADF regression is ∆zit =
p Pp
ai + bi zi,t−1 + ci z̄t−1 + j=1 dij ∆z̄t−j j=1 δij ∆zi,t−j

2
and K =1, 2, 3 and 5.While the number of common factor is admittedly
likely to be generally small, it is nevertheless interesting to investigate the
behaviour of the test for a wide range of cases. The cross-section depen-
dence can be low (γij ∼ U ni[0, 0.2]) or high (γij ∼ U ni[−1, 3]), and the serial
correlation absent (ρi = 0), positive (ρi ∼ U ni[0.2, 0.4]), high and positive
(ρi ∼ U ni[0.4, 0.8]2 ) and negative (ρi ∼ U ni[−0.2, −0.4]). Finally, to evalu-
ate the size of the test φi = 1, whereas for the power φi ∼ U ni[0.85; 0.95].
These settings are combined to generate the following DGPs:
DGP 1: no serial correlation, low cross-section dependence;
DGP 2: no serial correlation, high cross-section dependence;
DGP 3: (low) positive serial correlation, high cross-section dependence;
DGP 4: (low) negative serial correlation, high cross-section dependence;
DGP 5: (high) positive serial correlation, high cross-section dependence.
Since the aim of the experiment is to analyze the behaviour of the test when
the assumption of a single common factor is violated, the autoregressive order
of the residuals L is always fixed to the true value for each DGP (that is, 0
for DGP 1 and DGP 2 and 1 for DGP 3, DGP 4 and DGP 5).
As expected, the results (table I) suggest that, when K =1, no or very little
size bias affects the CIPS test, even with T =50. But when K > 1, this
happens only for DGP 1 (low cross-section dependence). In the other cases,
large biases appear even with K =2, which do not worsen as K grows, but
unfortunately do not vanish as the sample size and the number of cross-
sectional units are raised. However, the power does grow with T and N.

3.2 Empirically-selected lag length

Let us now examine the performance of the test when there is a single com-
mon factor and the lag order is empirically estimated. The DGP is now given
by (1), (2) and (4). The settings do not change whereas the autoregressive
order of the residuals is estimated (i) with Campbell and Perron (1991) lag
length selection procedure 3 ; (ii) maximizing the Bayesian Information Cri-
teria (BIC); (iii) as a deterministic function of the sample size, following the
Zivot and Andrews criterion (1992)4 . In the first two cases, the maximum
2
Actually, this last setting was not taken into consideration by Pesaran in his experi-
ments. It has been added because it allows to obtain a more complete description of the
behaviour of the test.
3
The procedure starts by estimating the regression with the maximum number p of lags
admitted. If the coefficient on the last lag considered is significant, the procedure stops,
otherwise the equation is estimated again with (p-1) lags, etc. . . , until only significant lags
are included.
4
Zivot and Andrews (1992) proposed to choose the truncation lag as the largest integer
less than [(T /100)2/9 ]. Thus with T =50 Zivot and Andrews criteria selects L̂=3 and when

3
number of lags admitted is eight, suitable if we are dealing with quarterly
data.
The simulation results (table II) suggest that the introduction of the estima-
tion of the truncation lag in the calculation of the CIPS test has a negative
effect on the size when the DGP is characterized by autocorrelated residu-
als. With the Campbell and Perron and BIC selection methods, depending
on the sign of the autocorrelation, we can have problems of over-rejection
(negative autocorrelation) or under-rejection (positive autocorrelation). The
Zivot and Andrews version of the test, instead, under-rejects the null even
when there is not autocorrelation in the residuals.
Consider first the results on the CIPS test for DGP 1 (low cross-section
dependence) and DGP 2 (high cross-section dependence). The comparison
of the two rows points to what are the consequences on the performance of
the test of an increase in the cross-sectional correlation degree. As we can
see, such consequences are not remarkable, at least when T =100 and N ≤ 50.
In general, when T is not large, an increase of the number of cross-sectional
units N has negative effects on the size of the test. In case of low dependence
(DGP 1), this happens only for the Campbell and Perron version of the test,
but in case of high dependence (DGP 2) the bias in the size of every version
significantly increases as N increases. In particular, when T =50, the 5% size
of the Campbell and Perron version of the test runs from 3.6 (N =20) to 1.6
(N =80), that of the BIC version from 5.3 (N =20) to 2.9 (N =80)and that of
Zivot and Andrews version from 2.7 (N =20) to 1.4 (N =80).
These results appear to be in contrast with the expected asymptotic be-
haviour of the test. However, we should remember that the asymptotic dis-
tribution of the CIPS test is non-standard and the values of the size are
calculated on the basis of the tabulated critical values. Then, the simulation
results suggest that, when the truncation lag is estimated and T is not large
enough, an increasing of N causes a departure of the distribution of the test
from the simulated small sample distribution.
Differently from what happens for the size, when T =50 power improves with
N.
When T =100, the estimated size of the BIC version is not affected anymore
by the number of cross-sectional units. The same does not happen for Zivot
and Andrews and for the Campbell and Perron version in the case of DGP
2. Furthermore, power is satisfactory value in most of cases.
Analyzing and comparing the performance of Campbell and Perron and BIC
procedure, it emerges that both methods estimate the autocorrelation or-
der (which is, for these two DGPs, zero) with high precision, at least when
T =100, L̂=4.

4
T =100. However, the selection procedure based on the Bayesian Information
Criteria is slightly better than the Campbell and Perron Criteria. It is seen,
in fact, that the BIC version of the CIPS test has more power when T =50
than the Campbell and Perron version has and it is less affected by the size
distortion deriving from the growth of N. This is not unexpected since, as Ng
and Perron (1995) showed, the BIC procedure is more parsimonious than the
methods based on sequential t-test as Campbell and Perron, and it works
better when the residuals are not autocorrelated.
Consider next the results of table II devoted to DGP 3, DGP 4 and DGP
5. The Zivot and Andrews version again under-rejects the null and the size
bias does not improve with N and T. Nevertheless, the power is positively
affected by an increase of the sample size and the number of cross-sectional
units.
For what regards the other two versions, a positive autocorrelation of the
residuals (DGP 3 and DGP 5) introduces a clear problem of under-rejection
associated to a very low level of power (at least when T is only 50). It is
interesting to note, when there is high positive serial correlation (DGP 5),
how the size of the test appears to be slightly better compared with the case of
low positive serial correlation (DGP 3). This can be due to the fact that both
Campbell and Perron and BIC procedures perform better when the level of
the autocorrelation is higher. However, this better behaviour in term of size
is joined to a lower level of power. If the residuals are negatively correlated
(DGP 4), the estimated size of the CIPS test is much higher than the nominal
level. In general, an increasing of N leads again to a greater bias in term of
size and to an improvement in term of power, which is somewhat expected
from what we observed for DGP 1 and DGP 2. But it is not expected the
fact that the same tendency is present also in the ”true L” version of the test
when the autocorrelation of the residuals is high and positive (DGP 5).
Differently from what happened for the first two DGPs, an increasing of T
has not always a positive effect on the size. In particular, the estimated
size of the BIC version of the test is further from the nominal level when
T raises to 100. One possible explanation for this negative effect centers on
the fact that the BIC method chooses the truncation lag L which maximizes:
`L (θ̂L )− 12 κL log(T ) where the first expression denotes the maximized value of
the log-likelihood function for the model with autoregressive order L whereas
the second term represents the penalty for having an additional parameter
(κL is, indeed, the number of the parameters that needs to be estimated in the
model). Thus, as T increases, the penalty for having an additional parameter
becomes larger and the method becomes more parsimonious. Hence, even
though asymptotically valid, we can think that BIC method, for T large
but not enough to guarantee a good asymptotical approximation, has the

5
tendency to under-parameterize the model. This has adverse effects in the
size of the CIPS test.
Contrary to what happens for the size, the power is always positively affected
by the increase of the time dimension.
Finally, now the Campbell and Perron version seems to suffer less from the
size distortions than the BIC version, both when the autocorrelation of the
residuals is positive and when is negative. However, this advantage is com-
pensated with a loss in term of power.
The strange behaviour of Campbell and Perron and BIC versions of the CIPS
test is not unexpected. Ng and Perron (2001) indeed show by simulation
that, in case of positive autocorrelation of the residuals, an under-estimation
of the true autocorrelation order leads to under-rejection of the true null for
three individual unit root tests. Whereas, in case of negative autocorrelation,
the under-estimation of the true autocorrelation order causes a strong over-
rejection problem.
Hence, an explanation for the strange behaviour of the test can be provided
by an experiment similar to the one performed by Ng and Perron (2001) and
then estimating the size of the CIPS test with L=0,1,...,8. The results (table
III) confirm that when the true autoregressive order is imposed (that is, 0
for DGP 1 and DGP 2 and 1 for DGP 3, DGP 4 and DGP 5). The estimated
size of the test is close to the nominal level and in line with the estimates
obtained by Pesaran. When the autoregressive order is under-estimated, the
same problems highlighted by Ng and Perron (2001) arise: under-rejection in
case of positive autocorrelation and strong over-rejection in case of negative
autocorrelation.
Therefore, the size problems of Campbell and Perron and BIC versions of the
CIPS test seem to depend on the performance of the two lag length selection
criteria. The probability for both methods to under-estimate the autocor-
relation order can be investigated through a new Monte Carlo experiment.
As expected, the results (table IV), confirm that the BIC procedure is more
parsimonious than that of Campbell and Perron, in accordance with Ng and
Perron (1995). When the true lag L is 0, this means that the BIC method
outperforms that of Campbell and Perron. Moreover, also the previous hint
regarding the penalty term of the BIC procedure is confirmed: on average,
when T is raised, the procedure tends to select a lower order of autocorrela-
tion and then to under-parameterize the model when there is autocorrelation
in the residuals.
The mean truncation lag selected by Campbell and Perron method is around
3 and it is not remarkably influenced by the growth of time dimension. How-
ever, even though on average Campbell and Perron procedure seems to over-
parameterize the model, it selects L̂=0 when the true autoregressive order is

6
1 in a consistent number of cases.
Finally, comparing the rows of table IV devoted to DGP 3 and DGP 5, it
is confirmed that the higher is the level of autocorrelation in the residuals,
the higher the probability is to select the correct truncation lag (that is 1)
and the lower the probability is to under-parameterize the model (that is to
select L̂=0). Despite these improvements both methods continue to select
L̂=0 in a consistent number of cases also when there is high serial correlation
in the residuals.
Concluding, both methods have the tendency to select L̂=0 too often when
the residuals are autocorrelated and this causes large distortions in the CIPS
test size.

4 Conclusion
Monte Carlo evidence suggests that the behavior of the CIPS test is not sat-
isfactory when the assumption of a single common factor in the specification
of the cross-section dependence is violated. Furthermore, the estimation of
the truncation lag L is an important issue: the selection of a wrong L dra-
matically affects the good properties of the test. Experimental results have
shown how the version of the test based on selecting the autoregressive order
according to a deterministic function of T (as that of Zivot and Andrews)
seems to be the most suitable if compared with selection criteria based on In-
formation criteria (as BIC) or on sequential t-test (as Campbell and Perron).
These two methods tend to select L̂=0 too often causing under-rejection in
case positive autocorrelation and strong over-rejection in case of negative
autocorrelation.

7
Table 1: Size and power of a 5% CIPS test when the autoregressive order is
known and the number of common factors K is greater than one.
K =1 K =2 K =3 K =5
N T DGP size power size power size power size power
20 50 1 5.3 61.6 4.9 59.7 5.6 60.0 5.0 60.7
2 5.7 59.3 15.0 58.5 17.2 55.2 14.7 56.2
3 4.7 40.0 9.4 39.4 10.0 41.9 10.3 41.9
4 4.8 50.7 16.0 48.0 16.6 46.1 15.5 46.6
5 3.8 25.0 2.1 16.1 2.3 13.9 2.2 16.9
100 1 5.3 99.9 4.9 99.9 5.4 100.0 5.0 99.9
2 5.5 99.9 14.8 96.7 16.0 95.7 15.5 96.6
3 5.5 98.9 9.6 93.4 9.8 92.2 10.6 93.5
4 5.6 99.7 17.0 90.9 18.2 90.9 15.4 92.8
5 4.2 91.6 1.6 72.0 1.3 67.3 1.7 71.4
50 50 1 5.1 85.3 5.0 84.9 5.4 85.2 5.3 85.6
2 4.9 85.4 19.1 69.6 21.2 67.4 19.5 68.4
3 4.4 56.6 10.4 53.4 12.8 52.9 13.7 54.3
4 4.5 73.9 20.6 59.3 21.2 55.2 18.5 54.6
5 3.4 34.7 1.9 19.8 1.9 18.7 2.6 22.7
100 1 4.6 100.0 4.5 100.0 4.6 100.0 4.9 100.0
2 5.0 100.0 19.1 98.0 20.6 98.4 18.5 98.8
3 4.6 100.0 10.9 97.8 12.1 96.7 13.2 97.7
4 4.2 100.0 19.4 95.0 21.1 94.8 19.5 96.9
5 3.6 99.9 1.4 90.6 1.1 84.3 1.6 87.4
80 50 1 4.7 92.2 4.7 92.1 4.9 92.3 5.1 92.3
2 4.9 92.8 21.2 71.7 22.3 70.7 20.7 70.2
3 4.2 64.0 11.6 57.4 13.5 56.0 15.1 56.6
4 4.5 82.9 22.3 61.6 22.3 57.8 19.6 58.9
5 2.9 38.5 2.1 20.8 1.4 20.2 2.6 25.9
100 1 5.4 100.0 4.8 100.0 4.7 100.0 5.2 100.0
2 5.0 100.0 20.7 98.7 23.8 98.6 20.7 99.5
3 4.8 100.0 11.8 98.2 13.6 97.8 14.9 98.7
4 5.4 100.0 22.8 95.9 23.5 95.8 20.8 97.4
5 3.7 100.0 1.4 94.1 1.1 88.9 1.6 92.1

Notes: the DGPs are generated according to the expressions (1), (3) and (4) with φi =1
∀i in the size case and φi ∼ U ni[0.85, 0.95] for the power case. DGP 1: ρi =0; γij ∼
U ni[0, 0.2]; DGP 2: ρi =0; γij ∼ U ni[−1, 3]; DGP 3: ρi ∼ U ni[0.2, 0.4]; γij ∼ U ni[−1, 3];
DGP 4: ρi ∼ U ni[−0.2, −0.4]; γij ∼ U ni[−1, 3]; DGP 5: ρi ∼ U ni[0.4, 0.8]; γij ∼
U ni[−1, 3].

8
Table 2: Size and power of a 5% CIPS test when the autoregressive order is
estimated.
C. & P. BIC Z. & A. True L
N T DGP size power size power size power size power
20 50 1 4.6 41.4 6.4 61.7 2.3 18.7 5.1 61.0
2 3.6 33.6 5.3 54.5 2.7 22.6 5.4 61.8
3 1.2 11.5 1.2 13.9 3.8 18.1 5.2 39.3
4 16.5 68.8 33.7 91.9 3.0 22.0 6.0 51.1
5 1.6 9.0 1.6 9.9 3.2 14.5 3.9 25.6
100 1 4.6 97.7 6.1 99.9 3.3 85.3 5.1 99.9
2 4.9 97.5 5.4 99.8 4.2 88.7 5.7 99.9
3 1.5 82.1 1.2 79.6 3.4 80.6 4.9 98.8
4 16.3 99.0 37.0 100.0 3.7 90.6 5.1 99.7
5 2.6 64.6 2.2 55.3 3.4 63.4 4.9 91.5
50 50 1 3.2 60.1 6.1 85.8 3.3 26.8 4.7 84.6
2 2.0 46.1 4.5 77.1 2.6 28.3 5.3 85.5
3 0.3 13.1 0.4 15.6 2.9 28.3 4.3 58.2
4 19.5 85.1 40.2 97.9 3.2 30.0 5.1 73.7
5 0.4 7.6 0.5 8.5 1.2 14.3 3.3 36.4
100 1 4.1 100.0 6.3 100.0 2.9 98.9 4.8 100.0
2 3.8 99.9 5.2 100.0 2.6 98.7 5.3 100.0
3 0.6 97.2 0.4 97.4 2.5 97.6 4.0 100.0
4 17.8 100.0 49.6 100.0 1.7 99.6 4.5 100.0
5 1.4 84.6 0.8 76.0 1.7 89.1 4.1 99.9
80 50 1 2.6 65.9 4.7 90.5 2.8 27.5 4.2 92.2
2 1.6 52.3 2.9 82.9 1.4 30.7 4.4 92.2
3 0.1 11.9 0.1 14.8 1.2 23.9 4.3 64.8
4 19.8 89.3 43.0 98.8 2.5 33.9 4.3 83.1
5 0.2 6.3 0.2 6.5 1.1 15.5 2.8 37.8
100 1 4.8 100.0 5.1 100.0 1.7 99.7 5.4 100.0
2 3.1 100.0 4.9 100.0 2.4 99.8 5.1 100.0
3 0.6 99.6 0.1 99.4 2.0 99.3 4.8 100.0
4 20.3 100.0 52.5 100.0 1.1 100.0 5.3 100.0
5 0.7 92.8 0.5 80.5 1.2 93.6 3.7 100.0

Notes: the DGPs are generated according to the expressions (1), (2) and (4) with φi =1 ∀i
in the size case and φi ∼ U ni[0.85, 0.95] for the power case. DGP 1: ρi =0; γi ∼ U ni[0, 0.2];
DGP 2: ρi =0; γi ∼ U ni[−1, 3]; DGP 3: ρi ∼ U ni[0.2, 0.4]; γj ∼ U ni[−1, 3]; DGP 4:
ρi ∼ U ni[−0.2, −0.4]; γi ∼ U ni[−1, 3]; DGP 5: ρi ∼ U ni[0.4, 0.8]; γi ∼ U ni[−1, 3].

9
Table 3: Estimated size of a 5% CIPS test when the truncation lag L is fixed
N T DGP L=0 L=1 L=2 L=3 L=4 L=5 L=6 L=7 L=8
20 50 1 4.3 4.6 2.7 2.3 1.9 1.0 1.1 0.5 0.1
2 6.3 5.0 3.4 2.7 1.7 1.1 1.4 1.4 0.8
3 0.0 5.6 4.2 3.8 2.4 2.0 1.1 1.5 0.8
4 73.9 5.1 3.5 3.0 2.0 1.7 1.4 0.6 0.4
5 0.0 4.8 3.2 3.2 2.5 2.0 1.3 0.5 0.3
100 1 5.3 5.6 4.8 4.0 3.3 3.2 2.2 2.1 1.7
2 6.2 5.8 5.2 5.4 4.2 3.4 3.4 2.7 2.5
3 0.0 4.9 4.1 4.6 3.4 3.7 2.6 2.5 2.1
4 74.3 5.0 3.6 3.9 3.7 3.1 2.6 2.8 2.2
5 0.0 4.0 3.8 4.1 3.4 3.6 2.9 2.7 1.9
50 50 1 4.5 3.6 3.4 3.3 2.0 1.2 0.4 0.1 0.2
2 5.4 4.5 2.7 2.6 1.4 0.7 0.1 0.3 0.0
3 0.1 4.9 2.6 2.9 1.0 1.1 0.7 0.3 0.1
4 85.1 5.1 3.2 3.2 1.3 1.0 0.6 0.7 0.3
5 0.0 2.4 1.5 1.2 0.7 0.5 0.3 0.3 0.2
100 1 5.3 5.3 3.8 3.9 2.9 2.7 2.6 1.8 1.3
2 4.7 4.6 3.2 3.0 2.6 2.5 1.6 1.9 1.4
3 0.0 3.5 3.3 2.7 2.5 2.3 1.1 1.5 1.1
4 89.1 4.3 2.7 2.2 1.7 1.5 1.0 1.2 0.5
5 0.0 3.4 2.2 2.3 1.7 1.4 1.0 1.5 1.0
80 50 1 5.2 4.9 2.3 2.8 0.6 0.6 0.0 0.2 0.0
2 4.3 3.9 1.7 1.4 0.6 0.4 0.1 0.0 0.1
3 0.0 3.9 1.4 1.2 0.5 0.5 0.2 0.2 0.1
4 90.7 5.6 2.2 2.5 0.8 0.4 0.4 0.4 0.0
5 0.0 2.2 1.0 1.1 0.3 0.5 0.2 0.1 0.1
100 1 4.6 3.8 3.4 3.7 1.7 2.0 1.3 1.0 0.5
2 6.1 5.1 3.5 4.0 2.4 1.9 1.8 1.8 1.7
3 0.0 3.7 3.1 2.5 2.0 1.6 1.0 0.9 0.6
4 94.8 4.1 2.1 2.1 1.1 1.3 0.9 1.7 1.0
5 0.0 3.5 2.4 2.1 1.2 1.4 0.9 0.9 0.6

Notes: For the characteristics of the five DGPs see note under table II. In bold type, the
estimated sizes of the CIPS test corresponding to the true autoregressive order L.

10
 Table 4: Performance of Campbell and Perron and BIC lag length selection procedures (N=1 )

Probability of selecting
T DGP L̂=0 L̂=1 L̂=2 L̂=3 L̂=4 L̂=5 L̂=6 L̂=7 L̂=8 Mean L̂
Campbell & Perron 50 1 48.9 5.4 6.0 4.4 5.4 7.2 5.1 8.7 8.9 2.5
2 43.0 5.4 5.8 5.9 5.9 8.0 6.2 10.1 9.7 2.8
3 32.8 17.9 7.7 5.2 5.5 6.4 8.6 7.6 8.3 2.7
4 38.7 13.1 4.4 5.7 5.5 7.5 6.9 9.8 8.4 2.8
5 22.5 26.5 7.4 6.5 6.5 9.0 6.9 7.8 6.9 2.8
100 1 43.9 4.8 4.1 8.0 7.3 7.6 9.8 8.4 6.1 2.7
2 41.8 6.0 5.1 6.1 6.3 6.8 8.1 10.9 8.9 2.9
3 24.7 21.2 5.4 6.1 7.2 7.7 7.8 9.2 10.7 3.1
4 30.5 19.8 5.6 6.2 6.9 6.5 7.9 8.2 8.4 2.8
5 10.8 33.3 9.0 6.7 7.5 7.0 7.6 7.5 10.6 3.2
BIC 50 1 64.7 11.6 6.2 3.1 2.8 2.7 2.6 2.7 3.6 1.2

11
2 68.1 11.2 5.0 3.3 2.9 2.4 1.2 2.5 3.4 1.1
3 52.4 23.3 7.8 4.3 2.9 2.6 2.6 2.0 2.1 1.2
4 55.2 20.8 6.4 4.1 3.2 2.8 1.8 2.1 3.6 1.3
5 29.6 36.9 10.7 6.2 4.3 3.8 2.4 2.8 3.3 1.7
100 1 80.9 9.9 4.9 1.8 0.7 0.7 1.0 0.0 0.1 0.4
2 83.8 8.3 4.0 1.6 1.1 0.3 0.2 0.5 0.2 0.3
3 51.5 36.9 6.7 1.5 1.4 1.2 0.1 0.5 0.2 0.7
4 58.2 30.4 5.3 3.0 0.9 0.9 0.7 0.2 0.4 0.7
5 23.6 54.3 12.7 5.4 1.3 1.7 0.4 0.3 0.3 1.2

Notes: For the characteristics of the five DGPs see note under table II. Moreover, in all cases, φi =1. The results with φi ∼ U ni[0.85, 0.95]
are very similar and then not reported. In bold type, the probability to select the correct truncation lag.
 References
Ben-David, D. (1994) ”Convergence Clubs and diverging economies”, CEPR
Discussion Paper Series 922, London.

Campbell, J. and P. Perron (1991) ”Pitfalls and Opportunities: what Macroe-

conomists should know about Unit Roots”, in Blanchard, O. and S. Fishers
(Eds.), NBER Macroeconomics Annual, MIT Press, Cambridge, MA.

Galor, O. (1996) ”Convergence? Inferences from Theoretical models”, The

Economic Journal, 106, 1056-1069.

Ng, S. and P. Perron (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. and P. Perron (2001) ”Lag Length Selection and the Construction of
Unit Root Tests with Good Size and Power”, Econometrica, 69, 1519-1554.

O’Connell, P. (1998) ”Overvaluation of Purchasing Power Parity”, Journal

of International Economics, 44, 1-19.

Pesaran, M.H. (2006) ”A simple Panel Unit Root test in the presence of
Cross-section Dependence”, Cambridge Working Papers in Economics, 0346.

Zivot, E. and D.W.K. Andrews (1992) ”Further evidence on the great cash,
the oil-price shock and the unit root hypothesis”, Journal of Business and
Economic Statistics, 10, 251-270.

12

View publication stats